97a0c35219ae4bf2a61a5fe6bf38621b38681184
[pandora-kernel.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 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
64                                    struct buffer_head *bh_result, int create);
65 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
66 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
67 static int __ext4_journalled_writepage(struct page *page, unsigned int len);
68 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
69
70 /*
71  * Test whether an inode is a fast symlink.
72  */
73 static int ext4_inode_is_fast_symlink(struct inode *inode)
74 {
75         int ea_blocks = EXT4_I(inode)->i_file_acl ?
76                 (inode->i_sb->s_blocksize >> 9) : 0;
77
78         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
79 }
80
81 /*
82  * Work out how many blocks we need to proceed with the next chunk of a
83  * truncate transaction.
84  */
85 static unsigned long blocks_for_truncate(struct inode *inode)
86 {
87         ext4_lblk_t needed;
88
89         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
90
91         /* Give ourselves just enough room to cope with inodes in which
92          * i_blocks is corrupt: we've seen disk corruptions in the past
93          * which resulted in random data in an inode which looked enough
94          * like a regular file for ext4 to try to delete it.  Things
95          * will go a bit crazy if that happens, but at least we should
96          * try not to panic the whole kernel. */
97         if (needed < 2)
98                 needed = 2;
99
100         /* But we need to bound the transaction so we don't overflow the
101          * journal. */
102         if (needed > EXT4_MAX_TRANS_DATA)
103                 needed = EXT4_MAX_TRANS_DATA;
104
105         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
106 }
107
108 /*
109  * Truncate transactions can be complex and absolutely huge.  So we need to
110  * be able to restart the transaction at a conventient checkpoint to make
111  * sure we don't overflow the journal.
112  *
113  * start_transaction gets us a new handle for a truncate transaction,
114  * and extend_transaction tries to extend the existing one a bit.  If
115  * extend fails, we need to propagate the failure up and restart the
116  * transaction in the top-level truncate loop. --sct
117  */
118 static handle_t *start_transaction(struct inode *inode)
119 {
120         handle_t *result;
121
122         result = ext4_journal_start(inode, blocks_for_truncate(inode));
123         if (!IS_ERR(result))
124                 return result;
125
126         ext4_std_error(inode->i_sb, PTR_ERR(result));
127         return result;
128 }
129
130 /*
131  * Try to extend this transaction for the purposes of truncation.
132  *
133  * Returns 0 if we managed to create more room.  If we can't create more
134  * room, and the transaction must be restarted we return 1.
135  */
136 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
137 {
138         if (!ext4_handle_valid(handle))
139                 return 0;
140         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
141                 return 0;
142         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
143                 return 0;
144         return 1;
145 }
146
147 /*
148  * Restart the transaction associated with *handle.  This does a commit,
149  * so before we call here everything must be consistently dirtied against
150  * this transaction.
151  */
152 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
153                                  int nblocks)
154 {
155         int ret;
156
157         /*
158          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
159          * moment, get_block can be called only for blocks inside i_size since
160          * page cache has been already dropped and writes are blocked by
161          * i_mutex. So we can safely drop the i_data_sem here.
162          */
163         BUG_ON(EXT4_JOURNAL(inode) == NULL);
164         jbd_debug(2, "restarting handle %p\n", handle);
165         up_write(&EXT4_I(inode)->i_data_sem);
166         ret = ext4_journal_restart(handle, blocks_for_truncate(inode));
167         down_write(&EXT4_I(inode)->i_data_sem);
168         ext4_discard_preallocations(inode);
169
170         return ret;
171 }
172
173 /*
174  * Called at the last iput() if i_nlink is zero.
175  */
176 void ext4_evict_inode(struct inode *inode)
177 {
178         handle_t *handle;
179         int err;
180
181         if (inode->i_nlink) {
182                 truncate_inode_pages(&inode->i_data, 0);
183                 goto no_delete;
184         }
185
186         if (!is_bad_inode(inode))
187                 dquot_initialize(inode);
188
189         if (ext4_should_order_data(inode))
190                 ext4_begin_ordered_truncate(inode, 0);
191         truncate_inode_pages(&inode->i_data, 0);
192
193         if (is_bad_inode(inode))
194                 goto no_delete;
195
196         handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
197         if (IS_ERR(handle)) {
198                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
199                 /*
200                  * If we're going to skip the normal cleanup, we still need to
201                  * make sure that the in-core orphan linked list is properly
202                  * cleaned up.
203                  */
204                 ext4_orphan_del(NULL, inode);
205                 goto no_delete;
206         }
207
208         if (IS_SYNC(inode))
209                 ext4_handle_sync(handle);
210         inode->i_size = 0;
211         err = ext4_mark_inode_dirty(handle, inode);
212         if (err) {
213                 ext4_warning(inode->i_sb,
214                              "couldn't mark inode dirty (err %d)", err);
215                 goto stop_handle;
216         }
217         if (inode->i_blocks)
218                 ext4_truncate(inode);
219
220         /*
221          * ext4_ext_truncate() doesn't reserve any slop when it
222          * restarts journal transactions; therefore there may not be
223          * enough credits left in the handle to remove the inode from
224          * the orphan list and set the dtime field.
225          */
226         if (!ext4_handle_has_enough_credits(handle, 3)) {
227                 err = ext4_journal_extend(handle, 3);
228                 if (err > 0)
229                         err = ext4_journal_restart(handle, 3);
230                 if (err != 0) {
231                         ext4_warning(inode->i_sb,
232                                      "couldn't extend journal (err %d)", err);
233                 stop_handle:
234                         ext4_journal_stop(handle);
235                         ext4_orphan_del(NULL, inode);
236                         goto no_delete;
237                 }
238         }
239
240         /*
241          * Kill off the orphan record which ext4_truncate created.
242          * AKPM: I think this can be inside the above `if'.
243          * Note that ext4_orphan_del() has to be able to cope with the
244          * deletion of a non-existent orphan - this is because we don't
245          * know if ext4_truncate() actually created an orphan record.
246          * (Well, we could do this if we need to, but heck - it works)
247          */
248         ext4_orphan_del(handle, inode);
249         EXT4_I(inode)->i_dtime  = get_seconds();
250
251         /*
252          * One subtle ordering requirement: if anything has gone wrong
253          * (transaction abort, IO errors, whatever), then we can still
254          * do these next steps (the fs will already have been marked as
255          * having errors), but we can't free the inode if the mark_dirty
256          * fails.
257          */
258         if (ext4_mark_inode_dirty(handle, inode))
259                 /* If that failed, just do the required in-core inode clear. */
260                 ext4_clear_inode(inode);
261         else
262                 ext4_free_inode(handle, inode);
263         ext4_journal_stop(handle);
264         return;
265 no_delete:
266         ext4_clear_inode(inode);        /* We must guarantee clearing of inode... */
267 }
268
269 typedef struct {
270         __le32  *p;
271         __le32  key;
272         struct buffer_head *bh;
273 } Indirect;
274
275 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
276 {
277         p->key = *(p->p = v);
278         p->bh = bh;
279 }
280
281 /**
282  *      ext4_block_to_path - parse the block number into array of offsets
283  *      @inode: inode in question (we are only interested in its superblock)
284  *      @i_block: block number to be parsed
285  *      @offsets: array to store the offsets in
286  *      @boundary: set this non-zero if the referred-to block is likely to be
287  *             followed (on disk) by an indirect block.
288  *
289  *      To store the locations of file's data ext4 uses a data structure common
290  *      for UNIX filesystems - tree of pointers anchored in the inode, with
291  *      data blocks at leaves and indirect blocks in intermediate nodes.
292  *      This function translates the block number into path in that tree -
293  *      return value is the path length and @offsets[n] is the offset of
294  *      pointer to (n+1)th node in the nth one. If @block is out of range
295  *      (negative or too large) warning is printed and zero returned.
296  *
297  *      Note: function doesn't find node addresses, so no IO is needed. All
298  *      we need to know is the capacity of indirect blocks (taken from the
299  *      inode->i_sb).
300  */
301
302 /*
303  * Portability note: the last comparison (check that we fit into triple
304  * indirect block) is spelled differently, because otherwise on an
305  * architecture with 32-bit longs and 8Kb pages we might get into trouble
306  * if our filesystem had 8Kb blocks. We might use long long, but that would
307  * kill us on x86. Oh, well, at least the sign propagation does not matter -
308  * i_block would have to be negative in the very beginning, so we would not
309  * get there at all.
310  */
311
312 static int ext4_block_to_path(struct inode *inode,
313                               ext4_lblk_t i_block,
314                               ext4_lblk_t offsets[4], int *boundary)
315 {
316         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
317         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
318         const long direct_blocks = EXT4_NDIR_BLOCKS,
319                 indirect_blocks = ptrs,
320                 double_blocks = (1 << (ptrs_bits * 2));
321         int n = 0;
322         int final = 0;
323
324         if (i_block < direct_blocks) {
325                 offsets[n++] = i_block;
326                 final = direct_blocks;
327         } else if ((i_block -= direct_blocks) < indirect_blocks) {
328                 offsets[n++] = EXT4_IND_BLOCK;
329                 offsets[n++] = i_block;
330                 final = ptrs;
331         } else if ((i_block -= indirect_blocks) < double_blocks) {
332                 offsets[n++] = EXT4_DIND_BLOCK;
333                 offsets[n++] = i_block >> ptrs_bits;
334                 offsets[n++] = i_block & (ptrs - 1);
335                 final = ptrs;
336         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
337                 offsets[n++] = EXT4_TIND_BLOCK;
338                 offsets[n++] = i_block >> (ptrs_bits * 2);
339                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
340                 offsets[n++] = i_block & (ptrs - 1);
341                 final = ptrs;
342         } else {
343                 ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
344                              i_block + direct_blocks +
345                              indirect_blocks + double_blocks, inode->i_ino);
346         }
347         if (boundary)
348                 *boundary = final - 1 - (i_block & (ptrs - 1));
349         return n;
350 }
351
352 static int __ext4_check_blockref(const char *function, unsigned int line,
353                                  struct inode *inode,
354                                  __le32 *p, unsigned int max)
355 {
356         struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
357         __le32 *bref = p;
358         unsigned int blk;
359
360         while (bref < p+max) {
361                 blk = le32_to_cpu(*bref++);
362                 if (blk &&
363                     unlikely(!ext4_data_block_valid(EXT4_SB(inode->i_sb),
364                                                     blk, 1))) {
365                         es->s_last_error_block = cpu_to_le64(blk);
366                         ext4_error_inode(inode, function, line, blk,
367                                          "invalid block");
368                         return -EIO;
369                 }
370         }
371         return 0;
372 }
373
374
375 #define ext4_check_indirect_blockref(inode, bh)                         \
376         __ext4_check_blockref(__func__, __LINE__, inode,                \
377                               (__le32 *)(bh)->b_data,                   \
378                               EXT4_ADDR_PER_BLOCK((inode)->i_sb))
379
380 #define ext4_check_inode_blockref(inode)                                \
381         __ext4_check_blockref(__func__, __LINE__, inode,                \
382                               EXT4_I(inode)->i_data,                    \
383                               EXT4_NDIR_BLOCKS)
384
385 /**
386  *      ext4_get_branch - read the chain of indirect blocks leading to data
387  *      @inode: inode in question
388  *      @depth: depth of the chain (1 - direct pointer, etc.)
389  *      @offsets: offsets of pointers in inode/indirect blocks
390  *      @chain: place to store the result
391  *      @err: here we store the error value
392  *
393  *      Function fills the array of triples <key, p, bh> and returns %NULL
394  *      if everything went OK or the pointer to the last filled triple
395  *      (incomplete one) otherwise. Upon the return chain[i].key contains
396  *      the number of (i+1)-th block in the chain (as it is stored in memory,
397  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
398  *      number (it points into struct inode for i==0 and into the bh->b_data
399  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
400  *      block for i>0 and NULL for i==0. In other words, it holds the block
401  *      numbers of the chain, addresses they were taken from (and where we can
402  *      verify that chain did not change) and buffer_heads hosting these
403  *      numbers.
404  *
405  *      Function stops when it stumbles upon zero pointer (absent block)
406  *              (pointer to last triple returned, *@err == 0)
407  *      or when it gets an IO error reading an indirect block
408  *              (ditto, *@err == -EIO)
409  *      or when it reads all @depth-1 indirect blocks successfully and finds
410  *      the whole chain, all way to the data (returns %NULL, *err == 0).
411  *
412  *      Need to be called with
413  *      down_read(&EXT4_I(inode)->i_data_sem)
414  */
415 static Indirect *ext4_get_branch(struct inode *inode, int depth,
416                                  ext4_lblk_t  *offsets,
417                                  Indirect chain[4], int *err)
418 {
419         struct super_block *sb = inode->i_sb;
420         Indirect *p = chain;
421         struct buffer_head *bh;
422
423         *err = 0;
424         /* i_data is not going away, no lock needed */
425         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
426         if (!p->key)
427                 goto no_block;
428         while (--depth) {
429                 bh = sb_getblk(sb, le32_to_cpu(p->key));
430                 if (unlikely(!bh))
431                         goto failure;
432
433                 if (!bh_uptodate_or_lock(bh)) {
434                         if (bh_submit_read(bh) < 0) {
435                                 put_bh(bh);
436                                 goto failure;
437                         }
438                         /* validate block references */
439                         if (ext4_check_indirect_blockref(inode, bh)) {
440                                 put_bh(bh);
441                                 goto failure;
442                         }
443                 }
444
445                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
446                 /* Reader: end */
447                 if (!p->key)
448                         goto no_block;
449         }
450         return NULL;
451
452 failure:
453         *err = -EIO;
454 no_block:
455         return p;
456 }
457
458 /**
459  *      ext4_find_near - find a place for allocation with sufficient locality
460  *      @inode: owner
461  *      @ind: descriptor of indirect block.
462  *
463  *      This function returns the preferred place for block allocation.
464  *      It is used when heuristic for sequential allocation fails.
465  *      Rules are:
466  *        + if there is a block to the left of our position - allocate near it.
467  *        + if pointer will live in indirect block - allocate near that block.
468  *        + if pointer will live in inode - allocate in the same
469  *          cylinder group.
470  *
471  * In the latter case we colour the starting block by the callers PID to
472  * prevent it from clashing with concurrent allocations for a different inode
473  * in the same block group.   The PID is used here so that functionally related
474  * files will be close-by on-disk.
475  *
476  *      Caller must make sure that @ind is valid and will stay that way.
477  */
478 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
479 {
480         struct ext4_inode_info *ei = EXT4_I(inode);
481         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
482         __le32 *p;
483         ext4_fsblk_t bg_start;
484         ext4_fsblk_t last_block;
485         ext4_grpblk_t colour;
486         ext4_group_t block_group;
487         int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
488
489         /* Try to find previous block */
490         for (p = ind->p - 1; p >= start; p--) {
491                 if (*p)
492                         return le32_to_cpu(*p);
493         }
494
495         /* No such thing, so let's try location of indirect block */
496         if (ind->bh)
497                 return ind->bh->b_blocknr;
498
499         /*
500          * It is going to be referred to from the inode itself? OK, just put it
501          * into the same cylinder group then.
502          */
503         block_group = ei->i_block_group;
504         if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
505                 block_group &= ~(flex_size-1);
506                 if (S_ISREG(inode->i_mode))
507                         block_group++;
508         }
509         bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
510         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
511
512         /*
513          * If we are doing delayed allocation, we don't need take
514          * colour into account.
515          */
516         if (test_opt(inode->i_sb, DELALLOC))
517                 return bg_start;
518
519         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
520                 colour = (current->pid % 16) *
521                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
522         else
523                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
524         return bg_start + colour;
525 }
526
527 /**
528  *      ext4_find_goal - find a preferred place for allocation.
529  *      @inode: owner
530  *      @block:  block we want
531  *      @partial: pointer to the last triple within a chain
532  *
533  *      Normally this function find the preferred place for block allocation,
534  *      returns it.
535  *      Because this is only used for non-extent files, we limit the block nr
536  *      to 32 bits.
537  */
538 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
539                                    Indirect *partial)
540 {
541         ext4_fsblk_t goal;
542
543         /*
544          * XXX need to get goal block from mballoc's data structures
545          */
546
547         goal = ext4_find_near(inode, partial);
548         goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
549         return goal;
550 }
551
552 /**
553  *      ext4_blks_to_allocate: Look up the block map and count the number
554  *      of direct blocks need to be allocated for the given branch.
555  *
556  *      @branch: chain of indirect blocks
557  *      @k: number of blocks need for indirect blocks
558  *      @blks: number of data blocks to be mapped.
559  *      @blocks_to_boundary:  the offset in the indirect block
560  *
561  *      return the total number of blocks to be allocate, including the
562  *      direct and indirect blocks.
563  */
564 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
565                                  int blocks_to_boundary)
566 {
567         unsigned int count = 0;
568
569         /*
570          * Simple case, [t,d]Indirect block(s) has not allocated yet
571          * then it's clear blocks on that path have not allocated
572          */
573         if (k > 0) {
574                 /* right now we don't handle cross boundary allocation */
575                 if (blks < blocks_to_boundary + 1)
576                         count += blks;
577                 else
578                         count += blocks_to_boundary + 1;
579                 return count;
580         }
581
582         count++;
583         while (count < blks && count <= blocks_to_boundary &&
584                 le32_to_cpu(*(branch[0].p + count)) == 0) {
585                 count++;
586         }
587         return count;
588 }
589
590 /**
591  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
592  *      @indirect_blks: the number of blocks need to allocate for indirect
593  *                      blocks
594  *
595  *      @new_blocks: on return it will store the new block numbers for
596  *      the indirect blocks(if needed) and the first direct block,
597  *      @blks:  on return it will store the total number of allocated
598  *              direct blocks
599  */
600 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
601                              ext4_lblk_t iblock, ext4_fsblk_t goal,
602                              int indirect_blks, int blks,
603                              ext4_fsblk_t new_blocks[4], int *err)
604 {
605         struct ext4_allocation_request ar;
606         int target, i;
607         unsigned long count = 0, blk_allocated = 0;
608         int index = 0;
609         ext4_fsblk_t current_block = 0;
610         int ret = 0;
611
612         /*
613          * Here we try to allocate the requested multiple blocks at once,
614          * on a best-effort basis.
615          * To build a branch, we should allocate blocks for
616          * the indirect blocks(if not allocated yet), and at least
617          * the first direct block of this branch.  That's the
618          * minimum number of blocks need to allocate(required)
619          */
620         /* first we try to allocate the indirect blocks */
621         target = indirect_blks;
622         while (target > 0) {
623                 count = target;
624                 /* allocating blocks for indirect blocks and direct blocks */
625                 current_block = ext4_new_meta_blocks(handle, inode,
626                                                         goal, &count, err);
627                 if (*err)
628                         goto failed_out;
629
630                 if (unlikely(current_block + count > EXT4_MAX_BLOCK_FILE_PHYS)) {
631                         EXT4_ERROR_INODE(inode,
632                                          "current_block %llu + count %lu > %d!",
633                                          current_block, count,
634                                          EXT4_MAX_BLOCK_FILE_PHYS);
635                         *err = -EIO;
636                         goto failed_out;
637                 }
638
639                 target -= count;
640                 /* allocate blocks for indirect blocks */
641                 while (index < indirect_blks && count) {
642                         new_blocks[index++] = current_block++;
643                         count--;
644                 }
645                 if (count > 0) {
646                         /*
647                          * save the new block number
648                          * for the first direct block
649                          */
650                         new_blocks[index] = current_block;
651                         printk(KERN_INFO "%s returned more blocks than "
652                                                 "requested\n", __func__);
653                         WARN_ON(1);
654                         break;
655                 }
656         }
657
658         target = blks - count ;
659         blk_allocated = count;
660         if (!target)
661                 goto allocated;
662         /* Now allocate data blocks */
663         memset(&ar, 0, sizeof(ar));
664         ar.inode = inode;
665         ar.goal = goal;
666         ar.len = target;
667         ar.logical = iblock;
668         if (S_ISREG(inode->i_mode))
669                 /* enable in-core preallocation only for regular files */
670                 ar.flags = EXT4_MB_HINT_DATA;
671
672         current_block = ext4_mb_new_blocks(handle, &ar, err);
673         if (unlikely(current_block + ar.len > EXT4_MAX_BLOCK_FILE_PHYS)) {
674                 EXT4_ERROR_INODE(inode,
675                                  "current_block %llu + ar.len %d > %d!",
676                                  current_block, ar.len,
677                                  EXT4_MAX_BLOCK_FILE_PHYS);
678                 *err = -EIO;
679                 goto failed_out;
680         }
681
682         if (*err && (target == blks)) {
683                 /*
684                  * if the allocation failed and we didn't allocate
685                  * any blocks before
686                  */
687                 goto failed_out;
688         }
689         if (!*err) {
690                 if (target == blks) {
691                         /*
692                          * save the new block number
693                          * for the first direct block
694                          */
695                         new_blocks[index] = current_block;
696                 }
697                 blk_allocated += ar.len;
698         }
699 allocated:
700         /* total number of blocks allocated for direct blocks */
701         ret = blk_allocated;
702         *err = 0;
703         return ret;
704 failed_out:
705         for (i = 0; i < index; i++)
706                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
707         return ret;
708 }
709
710 /**
711  *      ext4_alloc_branch - allocate and set up a chain of blocks.
712  *      @inode: owner
713  *      @indirect_blks: number of allocated indirect blocks
714  *      @blks: number of allocated direct blocks
715  *      @offsets: offsets (in the blocks) to store the pointers to next.
716  *      @branch: place to store the chain in.
717  *
718  *      This function allocates blocks, zeroes out all but the last one,
719  *      links them into chain and (if we are synchronous) writes them to disk.
720  *      In other words, it prepares a branch that can be spliced onto the
721  *      inode. It stores the information about that chain in the branch[], in
722  *      the same format as ext4_get_branch() would do. We are calling it after
723  *      we had read the existing part of chain and partial points to the last
724  *      triple of that (one with zero ->key). Upon the exit we have the same
725  *      picture as after the successful ext4_get_block(), except that in one
726  *      place chain is disconnected - *branch->p is still zero (we did not
727  *      set the last link), but branch->key contains the number that should
728  *      be placed into *branch->p to fill that gap.
729  *
730  *      If allocation fails we free all blocks we've allocated (and forget
731  *      their buffer_heads) and return the error value the from failed
732  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
733  *      as described above and return 0.
734  */
735 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
736                              ext4_lblk_t iblock, int indirect_blks,
737                              int *blks, ext4_fsblk_t goal,
738                              ext4_lblk_t *offsets, Indirect *branch)
739 {
740         int blocksize = inode->i_sb->s_blocksize;
741         int i, n = 0;
742         int err = 0;
743         struct buffer_head *bh;
744         int num;
745         ext4_fsblk_t new_blocks[4];
746         ext4_fsblk_t current_block;
747
748         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
749                                 *blks, new_blocks, &err);
750         if (err)
751                 return err;
752
753         branch[0].key = cpu_to_le32(new_blocks[0]);
754         /*
755          * metadata blocks and data blocks are allocated.
756          */
757         for (n = 1; n <= indirect_blks;  n++) {
758                 /*
759                  * Get buffer_head for parent block, zero it out
760                  * and set the pointer to new one, then send
761                  * parent to disk.
762                  */
763                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
764                 branch[n].bh = bh;
765                 lock_buffer(bh);
766                 BUFFER_TRACE(bh, "call get_create_access");
767                 err = ext4_journal_get_create_access(handle, bh);
768                 if (err) {
769                         /* Don't brelse(bh) here; it's done in
770                          * ext4_journal_forget() below */
771                         unlock_buffer(bh);
772                         goto failed;
773                 }
774
775                 memset(bh->b_data, 0, blocksize);
776                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
777                 branch[n].key = cpu_to_le32(new_blocks[n]);
778                 *branch[n].p = branch[n].key;
779                 if (n == indirect_blks) {
780                         current_block = new_blocks[n];
781                         /*
782                          * End of chain, update the last new metablock of
783                          * the chain to point to the new allocated
784                          * data blocks numbers
785                          */
786                         for (i = 1; i < num; i++)
787                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
788                 }
789                 BUFFER_TRACE(bh, "marking uptodate");
790                 set_buffer_uptodate(bh);
791                 unlock_buffer(bh);
792
793                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
794                 err = ext4_handle_dirty_metadata(handle, inode, bh);
795                 if (err)
796                         goto failed;
797         }
798         *blks = num;
799         return err;
800 failed:
801         /* Allocation failed, free what we already allocated */
802         ext4_free_blocks(handle, inode, 0, new_blocks[0], 1, 0);
803         for (i = 1; i <= n ; i++) {
804                 /*
805                  * branch[i].bh is newly allocated, so there is no
806                  * need to revoke the block, which is why we don't
807                  * need to set EXT4_FREE_BLOCKS_METADATA.
808                  */
809                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1,
810                                  EXT4_FREE_BLOCKS_FORGET);
811         }
812         for (i = n+1; i < indirect_blks; i++)
813                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
814
815         ext4_free_blocks(handle, inode, 0, new_blocks[i], num, 0);
816
817         return err;
818 }
819
820 /**
821  * ext4_splice_branch - splice the allocated branch onto inode.
822  * @inode: owner
823  * @block: (logical) number of block we are adding
824  * @chain: chain of indirect blocks (with a missing link - see
825  *      ext4_alloc_branch)
826  * @where: location of missing link
827  * @num:   number of indirect blocks we are adding
828  * @blks:  number of direct blocks we are adding
829  *
830  * This function fills the missing link and does all housekeeping needed in
831  * inode (->i_blocks, etc.). In case of success we end up with the full
832  * chain to new block and return 0.
833  */
834 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
835                               ext4_lblk_t block, Indirect *where, int num,
836                               int blks)
837 {
838         int i;
839         int err = 0;
840         ext4_fsblk_t current_block;
841
842         /*
843          * If we're splicing into a [td]indirect block (as opposed to the
844          * inode) then we need to get write access to the [td]indirect block
845          * before the splice.
846          */
847         if (where->bh) {
848                 BUFFER_TRACE(where->bh, "get_write_access");
849                 err = ext4_journal_get_write_access(handle, where->bh);
850                 if (err)
851                         goto err_out;
852         }
853         /* That's it */
854
855         *where->p = where->key;
856
857         /*
858          * Update the host buffer_head or inode to point to more just allocated
859          * direct blocks blocks
860          */
861         if (num == 0 && blks > 1) {
862                 current_block = le32_to_cpu(where->key) + 1;
863                 for (i = 1; i < blks; i++)
864                         *(where->p + i) = cpu_to_le32(current_block++);
865         }
866
867         /* We are done with atomic stuff, now do the rest of housekeeping */
868         /* had we spliced it onto indirect block? */
869         if (where->bh) {
870                 /*
871                  * If we spliced it onto an indirect block, we haven't
872                  * altered the inode.  Note however that if it is being spliced
873                  * onto an indirect block at the very end of the file (the
874                  * file is growing) then we *will* alter the inode to reflect
875                  * the new i_size.  But that is not done here - it is done in
876                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
877                  */
878                 jbd_debug(5, "splicing indirect only\n");
879                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
880                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
881                 if (err)
882                         goto err_out;
883         } else {
884                 /*
885                  * OK, we spliced it into the inode itself on a direct block.
886                  */
887                 ext4_mark_inode_dirty(handle, inode);
888                 jbd_debug(5, "splicing direct\n");
889         }
890         return err;
891
892 err_out:
893         for (i = 1; i <= num; i++) {
894                 /*
895                  * branch[i].bh is newly allocated, so there is no
896                  * need to revoke the block, which is why we don't
897                  * need to set EXT4_FREE_BLOCKS_METADATA.
898                  */
899                 ext4_free_blocks(handle, inode, where[i].bh, 0, 1,
900                                  EXT4_FREE_BLOCKS_FORGET);
901         }
902         ext4_free_blocks(handle, inode, 0, le32_to_cpu(where[num].key),
903                          blks, 0);
904
905         return err;
906 }
907
908 /*
909  * The ext4_ind_map_blocks() function handles non-extents inodes
910  * (i.e., using the traditional indirect/double-indirect i_blocks
911  * scheme) for ext4_map_blocks().
912  *
913  * Allocation strategy is simple: if we have to allocate something, we will
914  * have to go the whole way to leaf. So let's do it before attaching anything
915  * to tree, set linkage between the newborn blocks, write them if sync is
916  * required, recheck the path, free and repeat if check fails, otherwise
917  * set the last missing link (that will protect us from any truncate-generated
918  * removals - all blocks on the path are immune now) and possibly force the
919  * write on the parent block.
920  * That has a nice additional property: no special recovery from the failed
921  * allocations is needed - we simply release blocks and do not touch anything
922  * reachable from inode.
923  *
924  * `handle' can be NULL if create == 0.
925  *
926  * return > 0, # of blocks mapped or allocated.
927  * return = 0, if plain lookup failed.
928  * return < 0, error case.
929  *
930  * The ext4_ind_get_blocks() function should be called with
931  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
932  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
933  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
934  * blocks.
935  */
936 static int ext4_ind_map_blocks(handle_t *handle, struct inode *inode,
937                                struct ext4_map_blocks *map,
938                                int flags)
939 {
940         int err = -EIO;
941         ext4_lblk_t offsets[4];
942         Indirect chain[4];
943         Indirect *partial;
944         ext4_fsblk_t goal;
945         int indirect_blks;
946         int blocks_to_boundary = 0;
947         int depth;
948         int count = 0;
949         ext4_fsblk_t first_block = 0;
950
951         J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
952         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
953         depth = ext4_block_to_path(inode, map->m_lblk, offsets,
954                                    &blocks_to_boundary);
955
956         if (depth == 0)
957                 goto out;
958
959         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
960
961         /* Simplest case - block found, no allocation needed */
962         if (!partial) {
963                 first_block = le32_to_cpu(chain[depth - 1].key);
964                 count++;
965                 /*map more blocks*/
966                 while (count < map->m_len && count <= blocks_to_boundary) {
967                         ext4_fsblk_t blk;
968
969                         blk = le32_to_cpu(*(chain[depth-1].p + count));
970
971                         if (blk == first_block + count)
972                                 count++;
973                         else
974                                 break;
975                 }
976                 goto got_it;
977         }
978
979         /* Next simple case - plain lookup or failed read of indirect block */
980         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
981                 goto cleanup;
982
983         /*
984          * Okay, we need to do block allocation.
985         */
986         goal = ext4_find_goal(inode, map->m_lblk, partial);
987
988         /* the number of blocks need to allocate for [d,t]indirect blocks */
989         indirect_blks = (chain + depth) - partial - 1;
990
991         /*
992          * Next look up the indirect map to count the totoal number of
993          * direct blocks to allocate for this branch.
994          */
995         count = ext4_blks_to_allocate(partial, indirect_blks,
996                                       map->m_len, blocks_to_boundary);
997         /*
998          * Block out ext4_truncate while we alter the tree
999          */
1000         err = ext4_alloc_branch(handle, inode, map->m_lblk, indirect_blks,
1001                                 &count, goal,
1002                                 offsets + (partial - chain), partial);
1003
1004         /*
1005          * The ext4_splice_branch call will free and forget any buffers
1006          * on the new chain if there is a failure, but that risks using
1007          * up transaction credits, especially for bitmaps where the
1008          * credits cannot be returned.  Can we handle this somehow?  We
1009          * may need to return -EAGAIN upwards in the worst case.  --sct
1010          */
1011         if (!err)
1012                 err = ext4_splice_branch(handle, inode, map->m_lblk,
1013                                          partial, indirect_blks, count);
1014         if (err)
1015                 goto cleanup;
1016
1017         map->m_flags |= EXT4_MAP_NEW;
1018
1019         ext4_update_inode_fsync_trans(handle, inode, 1);
1020 got_it:
1021         map->m_flags |= EXT4_MAP_MAPPED;
1022         map->m_pblk = le32_to_cpu(chain[depth-1].key);
1023         map->m_len = count;
1024         if (count > blocks_to_boundary)
1025                 map->m_flags |= EXT4_MAP_BOUNDARY;
1026         err = count;
1027         /* Clean up and exit */
1028         partial = chain + depth - 1;    /* the whole chain */
1029 cleanup:
1030         while (partial > chain) {
1031                 BUFFER_TRACE(partial->bh, "call brelse");
1032                 brelse(partial->bh);
1033                 partial--;
1034         }
1035 out:
1036         return err;
1037 }
1038
1039 #ifdef CONFIG_QUOTA
1040 qsize_t *ext4_get_reserved_space(struct inode *inode)
1041 {
1042         return &EXT4_I(inode)->i_reserved_quota;
1043 }
1044 #endif
1045
1046 /*
1047  * Calculate the number of metadata blocks need to reserve
1048  * to allocate a new block at @lblocks for non extent file based file
1049  */
1050 static int ext4_indirect_calc_metadata_amount(struct inode *inode,
1051                                               sector_t lblock)
1052 {
1053         struct ext4_inode_info *ei = EXT4_I(inode);
1054         sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
1055         int blk_bits;
1056
1057         if (lblock < EXT4_NDIR_BLOCKS)
1058                 return 0;
1059
1060         lblock -= EXT4_NDIR_BLOCKS;
1061
1062         if (ei->i_da_metadata_calc_len &&
1063             (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
1064                 ei->i_da_metadata_calc_len++;
1065                 return 0;
1066         }
1067         ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
1068         ei->i_da_metadata_calc_len = 1;
1069         blk_bits = order_base_2(lblock);
1070         return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
1071 }
1072
1073 /*
1074  * Calculate the number of metadata blocks need to reserve
1075  * to allocate a block located at @lblock
1076  */
1077 static int ext4_calc_metadata_amount(struct inode *inode, sector_t lblock)
1078 {
1079         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1080                 return ext4_ext_calc_metadata_amount(inode, lblock);
1081
1082         return ext4_indirect_calc_metadata_amount(inode, lblock);
1083 }
1084
1085 /*
1086  * Called with i_data_sem down, which is important since we can call
1087  * ext4_discard_preallocations() from here.
1088  */
1089 void ext4_da_update_reserve_space(struct inode *inode,
1090                                         int used, int quota_claim)
1091 {
1092         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1093         struct ext4_inode_info *ei = EXT4_I(inode);
1094
1095         spin_lock(&ei->i_block_reservation_lock);
1096         trace_ext4_da_update_reserve_space(inode, used);
1097         if (unlikely(used > ei->i_reserved_data_blocks)) {
1098                 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
1099                          "with only %d reserved data blocks\n",
1100                          __func__, inode->i_ino, used,
1101                          ei->i_reserved_data_blocks);
1102                 WARN_ON(1);
1103                 used = ei->i_reserved_data_blocks;
1104         }
1105
1106         /* Update per-inode reservations */
1107         ei->i_reserved_data_blocks -= used;
1108         ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
1109         percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1110                            used + ei->i_allocated_meta_blocks);
1111         ei->i_allocated_meta_blocks = 0;
1112
1113         if (ei->i_reserved_data_blocks == 0) {
1114                 /*
1115                  * We can release all of the reserved metadata blocks
1116                  * only when we have written all of the delayed
1117                  * allocation blocks.
1118                  */
1119                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1120                                    ei->i_reserved_meta_blocks);
1121                 ei->i_reserved_meta_blocks = 0;
1122                 ei->i_da_metadata_calc_len = 0;
1123         }
1124         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1125
1126         /* Update quota subsystem for data blocks */
1127         if (quota_claim)
1128                 dquot_claim_block(inode, used);
1129         else {
1130                 /*
1131                  * We did fallocate with an offset that is already delayed
1132                  * allocated. So on delayed allocated writeback we should
1133                  * not re-claim the quota for fallocated blocks.
1134                  */
1135                 dquot_release_reservation_block(inode, used);
1136         }
1137
1138         /*
1139          * If we have done all the pending block allocations and if
1140          * there aren't any writers on the inode, we can discard the
1141          * inode's preallocations.
1142          */
1143         if ((ei->i_reserved_data_blocks == 0) &&
1144             (atomic_read(&inode->i_writecount) == 0))
1145                 ext4_discard_preallocations(inode);
1146 }
1147
1148 static int __check_block_validity(struct inode *inode, const char *func,
1149                                 unsigned int line,
1150                                 struct ext4_map_blocks *map)
1151 {
1152         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
1153                                    map->m_len)) {
1154                 ext4_error_inode(inode, func, line, map->m_pblk,
1155                                  "lblock %lu mapped to illegal pblock "
1156                                  "(length %d)", (unsigned long) map->m_lblk,
1157                                  map->m_len);
1158                 return -EIO;
1159         }
1160         return 0;
1161 }
1162
1163 #define check_block_validity(inode, map)        \
1164         __check_block_validity((inode), __func__, __LINE__, (map))
1165
1166 /*
1167  * Return the number of contiguous dirty pages in a given inode
1168  * starting at page frame idx.
1169  */
1170 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
1171                                     unsigned int max_pages)
1172 {
1173         struct address_space *mapping = inode->i_mapping;
1174         pgoff_t index;
1175         struct pagevec pvec;
1176         pgoff_t num = 0;
1177         int i, nr_pages, done = 0;
1178
1179         if (max_pages == 0)
1180                 return 0;
1181         pagevec_init(&pvec, 0);
1182         while (!done) {
1183                 index = idx;
1184                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
1185                                               PAGECACHE_TAG_DIRTY,
1186                                               (pgoff_t)PAGEVEC_SIZE);
1187                 if (nr_pages == 0)
1188                         break;
1189                 for (i = 0; i < nr_pages; i++) {
1190                         struct page *page = pvec.pages[i];
1191                         struct buffer_head *bh, *head;
1192
1193                         lock_page(page);
1194                         if (unlikely(page->mapping != mapping) ||
1195                             !PageDirty(page) ||
1196                             PageWriteback(page) ||
1197                             page->index != idx) {
1198                                 done = 1;
1199                                 unlock_page(page);
1200                                 break;
1201                         }
1202                         if (page_has_buffers(page)) {
1203                                 bh = head = page_buffers(page);
1204                                 do {
1205                                         if (!buffer_delay(bh) &&
1206                                             !buffer_unwritten(bh))
1207                                                 done = 1;
1208                                         bh = bh->b_this_page;
1209                                 } while (!done && (bh != head));
1210                         }
1211                         unlock_page(page);
1212                         if (done)
1213                                 break;
1214                         idx++;
1215                         num++;
1216                         if (num >= max_pages) {
1217                                 done = 1;
1218                                 break;
1219                         }
1220                 }
1221                 pagevec_release(&pvec);
1222         }
1223         return num;
1224 }
1225
1226 /*
1227  * The ext4_map_blocks() function tries to look up the requested blocks,
1228  * and returns if the blocks are already mapped.
1229  *
1230  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1231  * and store the allocated blocks in the result buffer head and mark it
1232  * mapped.
1233  *
1234  * If file type is extents based, it will call ext4_ext_map_blocks(),
1235  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
1236  * based files
1237  *
1238  * On success, it returns the number of blocks being mapped or allocate.
1239  * if create==0 and the blocks are pre-allocated and uninitialized block,
1240  * the result buffer head is unmapped. If the create ==1, it will make sure
1241  * the buffer head is mapped.
1242  *
1243  * It returns 0 if plain look up failed (blocks have not been allocated), in
1244  * that casem, buffer head is unmapped
1245  *
1246  * It returns the error in case of allocation failure.
1247  */
1248 int ext4_map_blocks(handle_t *handle, struct inode *inode,
1249                     struct ext4_map_blocks *map, int flags)
1250 {
1251         int retval;
1252
1253         map->m_flags = 0;
1254         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
1255                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
1256                   (unsigned long) map->m_lblk);
1257         /*
1258          * Try to see if we can get the block without requesting a new
1259          * file system block.
1260          */
1261         down_read((&EXT4_I(inode)->i_data_sem));
1262         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
1263                 retval = ext4_ext_map_blocks(handle, inode, map, 0);
1264         } else {
1265                 retval = ext4_ind_map_blocks(handle, inode, map, 0);
1266         }
1267         up_read((&EXT4_I(inode)->i_data_sem));
1268
1269         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
1270                 int ret = check_block_validity(inode, map);
1271                 if (ret != 0)
1272                         return ret;
1273         }
1274
1275         /* If it is only a block(s) look up */
1276         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1277                 return retval;
1278
1279         /*
1280          * Returns if the blocks have already allocated
1281          *
1282          * Note that if blocks have been preallocated
1283          * ext4_ext_get_block() returns th create = 0
1284          * with buffer head unmapped.
1285          */
1286         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
1287                 return retval;
1288
1289         /*
1290          * When we call get_blocks without the create flag, the
1291          * BH_Unwritten flag could have gotten set if the blocks
1292          * requested were part of a uninitialized extent.  We need to
1293          * clear this flag now that we are committed to convert all or
1294          * part of the uninitialized extent to be an initialized
1295          * extent.  This is because we need to avoid the combination
1296          * of BH_Unwritten and BH_Mapped flags being simultaneously
1297          * set on the buffer_head.
1298          */
1299         map->m_flags &= ~EXT4_MAP_UNWRITTEN;
1300
1301         /*
1302          * New blocks allocate and/or writing to uninitialized extent
1303          * will possibly result in updating i_data, so we take
1304          * the write lock of i_data_sem, and call get_blocks()
1305          * with create == 1 flag.
1306          */
1307         down_write((&EXT4_I(inode)->i_data_sem));
1308
1309         /*
1310          * if the caller is from delayed allocation writeout path
1311          * we have already reserved fs blocks for allocation
1312          * let the underlying get_block() function know to
1313          * avoid double accounting
1314          */
1315         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1316                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1317         /*
1318          * We need to check for EXT4 here because migrate
1319          * could have changed the inode type in between
1320          */
1321         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
1322                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
1323         } else {
1324                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
1325
1326                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
1327                         /*
1328                          * We allocated new blocks which will result in
1329                          * i_data's format changing.  Force the migrate
1330                          * to fail by clearing migrate flags
1331                          */
1332                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
1333                 }
1334
1335                 /*
1336                  * Update reserved blocks/metadata blocks after successful
1337                  * block allocation which had been deferred till now. We don't
1338                  * support fallocate for non extent files. So we can update
1339                  * reserve space here.
1340                  */
1341                 if ((retval > 0) &&
1342                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
1343                         ext4_da_update_reserve_space(inode, retval, 1);
1344         }
1345         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1346                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1347
1348         up_write((&EXT4_I(inode)->i_data_sem));
1349         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
1350                 int ret = check_block_validity(inode, map);
1351                 if (ret != 0)
1352                         return ret;
1353         }
1354         return retval;
1355 }
1356
1357 /* Maximum number of blocks we map for direct IO at once. */
1358 #define DIO_MAX_BLOCKS 4096
1359
1360 static int _ext4_get_block(struct inode *inode, sector_t iblock,
1361                            struct buffer_head *bh, int flags)
1362 {
1363         handle_t *handle = ext4_journal_current_handle();
1364         struct ext4_map_blocks map;
1365         int ret = 0, started = 0;
1366         int dio_credits;
1367
1368         map.m_lblk = iblock;
1369         map.m_len = bh->b_size >> inode->i_blkbits;
1370
1371         if (flags && !handle) {
1372                 /* Direct IO write... */
1373                 if (map.m_len > DIO_MAX_BLOCKS)
1374                         map.m_len = DIO_MAX_BLOCKS;
1375                 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
1376                 handle = ext4_journal_start(inode, dio_credits);
1377                 if (IS_ERR(handle)) {
1378                         ret = PTR_ERR(handle);
1379                         return ret;
1380                 }
1381                 started = 1;
1382         }
1383
1384         ret = ext4_map_blocks(handle, inode, &map, flags);
1385         if (ret > 0) {
1386                 map_bh(bh, inode->i_sb, map.m_pblk);
1387                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
1388                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
1389                 ret = 0;
1390         }
1391         if (started)
1392                 ext4_journal_stop(handle);
1393         return ret;
1394 }
1395
1396 int ext4_get_block(struct inode *inode, sector_t iblock,
1397                    struct buffer_head *bh, int create)
1398 {
1399         return _ext4_get_block(inode, iblock, bh,
1400                                create ? EXT4_GET_BLOCKS_CREATE : 0);
1401 }
1402
1403 /*
1404  * `handle' can be NULL if create is zero
1405  */
1406 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1407                                 ext4_lblk_t block, int create, int *errp)
1408 {
1409         struct ext4_map_blocks map;
1410         struct buffer_head *bh;
1411         int fatal = 0, err;
1412
1413         J_ASSERT(handle != NULL || create == 0);
1414
1415         map.m_lblk = block;
1416         map.m_len = 1;
1417         err = ext4_map_blocks(handle, inode, &map,
1418                               create ? EXT4_GET_BLOCKS_CREATE : 0);
1419
1420         if (err < 0)
1421                 *errp = err;
1422         if (err <= 0)
1423                 return NULL;
1424         *errp = 0;
1425
1426         bh = sb_getblk(inode->i_sb, map.m_pblk);
1427         if (!bh) {
1428                 *errp = -EIO;
1429                 return NULL;
1430         }
1431         if (map.m_flags & EXT4_MAP_NEW) {
1432                 J_ASSERT(create != 0);
1433                 J_ASSERT(handle != NULL);
1434
1435                 /*
1436                  * Now that we do not always journal data, we should
1437                  * keep in mind whether this should always journal the
1438                  * new buffer as metadata.  For now, regular file
1439                  * writes use ext4_get_block instead, so it's not a
1440                  * problem.
1441                  */
1442                 lock_buffer(bh);
1443                 BUFFER_TRACE(bh, "call get_create_access");
1444                 fatal = ext4_journal_get_create_access(handle, bh);
1445                 if (!fatal && !buffer_uptodate(bh)) {
1446                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1447                         set_buffer_uptodate(bh);
1448                 }
1449                 unlock_buffer(bh);
1450                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1451                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1452                 if (!fatal)
1453                         fatal = err;
1454         } else {
1455                 BUFFER_TRACE(bh, "not a new buffer");
1456         }
1457         if (fatal) {
1458                 *errp = fatal;
1459                 brelse(bh);
1460                 bh = NULL;
1461         }
1462         return bh;
1463 }
1464
1465 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1466                                ext4_lblk_t block, int create, int *err)
1467 {
1468         struct buffer_head *bh;
1469
1470         bh = ext4_getblk(handle, inode, block, create, err);
1471         if (!bh)
1472                 return bh;
1473         if (buffer_uptodate(bh))
1474                 return bh;
1475         ll_rw_block(READ_META, 1, &bh);
1476         wait_on_buffer(bh);
1477         if (buffer_uptodate(bh))
1478                 return bh;
1479         put_bh(bh);
1480         *err = -EIO;
1481         return NULL;
1482 }
1483
1484 static int walk_page_buffers(handle_t *handle,
1485                              struct buffer_head *head,
1486                              unsigned from,
1487                              unsigned to,
1488                              int *partial,
1489                              int (*fn)(handle_t *handle,
1490                                        struct buffer_head *bh))
1491 {
1492         struct buffer_head *bh;
1493         unsigned block_start, block_end;
1494         unsigned blocksize = head->b_size;
1495         int err, ret = 0;
1496         struct buffer_head *next;
1497
1498         for (bh = head, block_start = 0;
1499              ret == 0 && (bh != head || !block_start);
1500              block_start = block_end, bh = next) {
1501                 next = bh->b_this_page;
1502                 block_end = block_start + blocksize;
1503                 if (block_end <= from || block_start >= to) {
1504                         if (partial && !buffer_uptodate(bh))
1505                                 *partial = 1;
1506                         continue;
1507                 }
1508                 err = (*fn)(handle, bh);
1509                 if (!ret)
1510                         ret = err;
1511         }
1512         return ret;
1513 }
1514
1515 /*
1516  * To preserve ordering, it is essential that the hole instantiation and
1517  * the data write be encapsulated in a single transaction.  We cannot
1518  * close off a transaction and start a new one between the ext4_get_block()
1519  * and the commit_write().  So doing the jbd2_journal_start at the start of
1520  * prepare_write() is the right place.
1521  *
1522  * Also, this function can nest inside ext4_writepage() ->
1523  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1524  * has generated enough buffer credits to do the whole page.  So we won't
1525  * block on the journal in that case, which is good, because the caller may
1526  * be PF_MEMALLOC.
1527  *
1528  * By accident, ext4 can be reentered when a transaction is open via
1529  * quota file writes.  If we were to commit the transaction while thus
1530  * reentered, there can be a deadlock - we would be holding a quota
1531  * lock, and the commit would never complete if another thread had a
1532  * transaction open and was blocking on the quota lock - a ranking
1533  * violation.
1534  *
1535  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1536  * will _not_ run commit under these circumstances because handle->h_ref
1537  * is elevated.  We'll still have enough credits for the tiny quotafile
1538  * write.
1539  */
1540 static int do_journal_get_write_access(handle_t *handle,
1541                                        struct buffer_head *bh)
1542 {
1543         int dirty = buffer_dirty(bh);
1544         int ret;
1545
1546         if (!buffer_mapped(bh) || buffer_freed(bh))
1547                 return 0;
1548         /*
1549          * __block_prepare_write() could have dirtied some buffers. Clean
1550          * the dirty bit as jbd2_journal_get_write_access() could complain
1551          * otherwise about fs integrity issues. Setting of the dirty bit
1552          * by __block_prepare_write() isn't a real problem here as we clear
1553          * the bit before releasing a page lock and thus writeback cannot
1554          * ever write the buffer.
1555          */
1556         if (dirty)
1557                 clear_buffer_dirty(bh);
1558         ret = ext4_journal_get_write_access(handle, bh);
1559         if (!ret && dirty)
1560                 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1561         return ret;
1562 }
1563
1564 /*
1565  * Truncate blocks that were not used by write. We have to truncate the
1566  * pagecache as well so that corresponding buffers get properly unmapped.
1567  */
1568 static void ext4_truncate_failed_write(struct inode *inode)
1569 {
1570         truncate_inode_pages(inode->i_mapping, inode->i_size);
1571         ext4_truncate(inode);
1572 }
1573
1574 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
1575                    struct buffer_head *bh_result, int create);
1576 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1577                             loff_t pos, unsigned len, unsigned flags,
1578                             struct page **pagep, void **fsdata)
1579 {
1580         struct inode *inode = mapping->host;
1581         int ret, needed_blocks;
1582         handle_t *handle;
1583         int retries = 0;
1584         struct page *page;
1585         pgoff_t index;
1586         unsigned from, to;
1587
1588         trace_ext4_write_begin(inode, pos, len, flags);
1589         /*
1590          * Reserve one block more for addition to orphan list in case
1591          * we allocate blocks but write fails for some reason
1592          */
1593         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1594         index = pos >> PAGE_CACHE_SHIFT;
1595         from = pos & (PAGE_CACHE_SIZE - 1);
1596         to = from + len;
1597
1598 retry:
1599         handle = ext4_journal_start(inode, needed_blocks);
1600         if (IS_ERR(handle)) {
1601                 ret = PTR_ERR(handle);
1602                 goto out;
1603         }
1604
1605         /* We cannot recurse into the filesystem as the transaction is already
1606          * started */
1607         flags |= AOP_FLAG_NOFS;
1608
1609         page = grab_cache_page_write_begin(mapping, index, flags);
1610         if (!page) {
1611                 ext4_journal_stop(handle);
1612                 ret = -ENOMEM;
1613                 goto out;
1614         }
1615         *pagep = page;
1616
1617         if (ext4_should_dioread_nolock(inode))
1618                 ret = __block_write_begin(page, pos, len, ext4_get_block_write);
1619         else
1620                 ret = __block_write_begin(page, pos, len, ext4_get_block);
1621
1622         if (!ret && ext4_should_journal_data(inode)) {
1623                 ret = walk_page_buffers(handle, page_buffers(page),
1624                                 from, to, NULL, do_journal_get_write_access);
1625         }
1626
1627         if (ret) {
1628                 unlock_page(page);
1629                 page_cache_release(page);
1630                 /*
1631                  * __block_write_begin may have instantiated a few blocks
1632                  * outside i_size.  Trim these off again. Don't need
1633                  * i_size_read because we hold i_mutex.
1634                  *
1635                  * Add inode to orphan list in case we crash before
1636                  * truncate finishes
1637                  */
1638                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1639                         ext4_orphan_add(handle, inode);
1640
1641                 ext4_journal_stop(handle);
1642                 if (pos + len > inode->i_size) {
1643                         ext4_truncate_failed_write(inode);
1644                         /*
1645                          * If truncate failed early the inode might
1646                          * still be on the orphan list; we need to
1647                          * make sure the inode is removed from the
1648                          * orphan list in that case.
1649                          */
1650                         if (inode->i_nlink)
1651                                 ext4_orphan_del(NULL, inode);
1652                 }
1653         }
1654
1655         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1656                 goto retry;
1657 out:
1658         return ret;
1659 }
1660
1661 /* For write_end() in data=journal mode */
1662 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1663 {
1664         if (!buffer_mapped(bh) || buffer_freed(bh))
1665                 return 0;
1666         set_buffer_uptodate(bh);
1667         return ext4_handle_dirty_metadata(handle, NULL, bh);
1668 }
1669
1670 static int ext4_generic_write_end(struct file *file,
1671                                   struct address_space *mapping,
1672                                   loff_t pos, unsigned len, unsigned copied,
1673                                   struct page *page, void *fsdata)
1674 {
1675         int i_size_changed = 0;
1676         struct inode *inode = mapping->host;
1677         handle_t *handle = ext4_journal_current_handle();
1678
1679         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1680
1681         /*
1682          * No need to use i_size_read() here, the i_size
1683          * cannot change under us because we hold i_mutex.
1684          *
1685          * But it's important to update i_size while still holding page lock:
1686          * page writeout could otherwise come in and zero beyond i_size.
1687          */
1688         if (pos + copied > inode->i_size) {
1689                 i_size_write(inode, pos + copied);
1690                 i_size_changed = 1;
1691         }
1692
1693         if (pos + copied >  EXT4_I(inode)->i_disksize) {
1694                 /* We need to mark inode dirty even if
1695                  * new_i_size is less that inode->i_size
1696                  * bu greater than i_disksize.(hint delalloc)
1697                  */
1698                 ext4_update_i_disksize(inode, (pos + copied));
1699                 i_size_changed = 1;
1700         }
1701         unlock_page(page);
1702         page_cache_release(page);
1703
1704         /*
1705          * Don't mark the inode dirty under page lock. First, it unnecessarily
1706          * makes the holding time of page lock longer. Second, it forces lock
1707          * ordering of page lock and transaction start for journaling
1708          * filesystems.
1709          */
1710         if (i_size_changed)
1711                 ext4_mark_inode_dirty(handle, inode);
1712
1713         return copied;
1714 }
1715
1716 /*
1717  * We need to pick up the new inode size which generic_commit_write gave us
1718  * `file' can be NULL - eg, when called from page_symlink().
1719  *
1720  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1721  * buffers are managed internally.
1722  */
1723 static int ext4_ordered_write_end(struct file *file,
1724                                   struct address_space *mapping,
1725                                   loff_t pos, unsigned len, unsigned copied,
1726                                   struct page *page, void *fsdata)
1727 {
1728         handle_t *handle = ext4_journal_current_handle();
1729         struct inode *inode = mapping->host;
1730         int ret = 0, ret2;
1731
1732         trace_ext4_ordered_write_end(inode, pos, len, copied);
1733         ret = ext4_jbd2_file_inode(handle, inode);
1734
1735         if (ret == 0) {
1736                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1737                                                         page, fsdata);
1738                 copied = ret2;
1739                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1740                         /* if we have allocated more blocks and copied
1741                          * less. We will have blocks allocated outside
1742                          * inode->i_size. So truncate them
1743                          */
1744                         ext4_orphan_add(handle, inode);
1745                 if (ret2 < 0)
1746                         ret = ret2;
1747         }
1748         ret2 = ext4_journal_stop(handle);
1749         if (!ret)
1750                 ret = ret2;
1751
1752         if (pos + len > inode->i_size) {
1753                 ext4_truncate_failed_write(inode);
1754                 /*
1755                  * If truncate failed early the inode might still be
1756                  * on the orphan list; we need to make sure the inode
1757                  * is removed from the orphan list in that case.
1758                  */
1759                 if (inode->i_nlink)
1760                         ext4_orphan_del(NULL, inode);
1761         }
1762
1763
1764         return ret ? ret : copied;
1765 }
1766
1767 static int ext4_writeback_write_end(struct file *file,
1768                                     struct address_space *mapping,
1769                                     loff_t pos, unsigned len, unsigned copied,
1770                                     struct page *page, void *fsdata)
1771 {
1772         handle_t *handle = ext4_journal_current_handle();
1773         struct inode *inode = mapping->host;
1774         int ret = 0, ret2;
1775
1776         trace_ext4_writeback_write_end(inode, pos, len, copied);
1777         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1778                                                         page, fsdata);
1779         copied = ret2;
1780         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1781                 /* if we have allocated more blocks and copied
1782                  * less. We will have blocks allocated outside
1783                  * inode->i_size. So truncate them
1784                  */
1785                 ext4_orphan_add(handle, inode);
1786
1787         if (ret2 < 0)
1788                 ret = ret2;
1789
1790         ret2 = ext4_journal_stop(handle);
1791         if (!ret)
1792                 ret = ret2;
1793
1794         if (pos + len > inode->i_size) {
1795                 ext4_truncate_failed_write(inode);
1796                 /*
1797                  * If truncate failed early the inode might still be
1798                  * on the orphan list; we need to make sure the inode
1799                  * is removed from the orphan list in that case.
1800                  */
1801                 if (inode->i_nlink)
1802                         ext4_orphan_del(NULL, inode);
1803         }
1804
1805         return ret ? ret : copied;
1806 }
1807
1808 static int ext4_journalled_write_end(struct file *file,
1809                                      struct address_space *mapping,
1810                                      loff_t pos, unsigned len, unsigned copied,
1811                                      struct page *page, void *fsdata)
1812 {
1813         handle_t *handle = ext4_journal_current_handle();
1814         struct inode *inode = mapping->host;
1815         int ret = 0, ret2;
1816         int partial = 0;
1817         unsigned from, to;
1818         loff_t new_i_size;
1819
1820         trace_ext4_journalled_write_end(inode, pos, len, copied);
1821         from = pos & (PAGE_CACHE_SIZE - 1);
1822         to = from + len;
1823
1824         if (copied < len) {
1825                 if (!PageUptodate(page))
1826                         copied = 0;
1827                 page_zero_new_buffers(page, from+copied, to);
1828         }
1829
1830         ret = walk_page_buffers(handle, page_buffers(page), from,
1831                                 to, &partial, write_end_fn);
1832         if (!partial)
1833                 SetPageUptodate(page);
1834         new_i_size = pos + copied;
1835         if (new_i_size > inode->i_size)
1836                 i_size_write(inode, pos+copied);
1837         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1838         if (new_i_size > EXT4_I(inode)->i_disksize) {
1839                 ext4_update_i_disksize(inode, new_i_size);
1840                 ret2 = ext4_mark_inode_dirty(handle, inode);
1841                 if (!ret)
1842                         ret = ret2;
1843         }
1844
1845         unlock_page(page);
1846         page_cache_release(page);
1847         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1848                 /* if we have allocated more blocks and copied
1849                  * less. We will have blocks allocated outside
1850                  * inode->i_size. So truncate them
1851                  */
1852                 ext4_orphan_add(handle, inode);
1853
1854         ret2 = ext4_journal_stop(handle);
1855         if (!ret)
1856                 ret = ret2;
1857         if (pos + len > inode->i_size) {
1858                 ext4_truncate_failed_write(inode);
1859                 /*
1860                  * If truncate failed early the inode might still be
1861                  * on the orphan list; we need to make sure the inode
1862                  * is removed from the orphan list in that case.
1863                  */
1864                 if (inode->i_nlink)
1865                         ext4_orphan_del(NULL, inode);
1866         }
1867
1868         return ret ? ret : copied;
1869 }
1870
1871 /*
1872  * Reserve a single block located at lblock
1873  */
1874 static int ext4_da_reserve_space(struct inode *inode, sector_t lblock)
1875 {
1876         int retries = 0;
1877         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1878         struct ext4_inode_info *ei = EXT4_I(inode);
1879         unsigned long md_needed;
1880         int ret;
1881
1882         /*
1883          * recalculate the amount of metadata blocks to reserve
1884          * in order to allocate nrblocks
1885          * worse case is one extent per block
1886          */
1887 repeat:
1888         spin_lock(&ei->i_block_reservation_lock);
1889         md_needed = ext4_calc_metadata_amount(inode, lblock);
1890         trace_ext4_da_reserve_space(inode, md_needed);
1891         spin_unlock(&ei->i_block_reservation_lock);
1892
1893         /*
1894          * We will charge metadata quota at writeout time; this saves
1895          * us from metadata over-estimation, though we may go over by
1896          * a small amount in the end.  Here we just reserve for data.
1897          */
1898         ret = dquot_reserve_block(inode, 1);
1899         if (ret)
1900                 return ret;
1901         /*
1902          * We do still charge estimated metadata to the sb though;
1903          * we cannot afford to run out of free blocks.
1904          */
1905         if (ext4_claim_free_blocks(sbi, md_needed + 1)) {
1906                 dquot_release_reservation_block(inode, 1);
1907                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1908                         yield();
1909                         goto repeat;
1910                 }
1911                 return -ENOSPC;
1912         }
1913         spin_lock(&ei->i_block_reservation_lock);
1914         ei->i_reserved_data_blocks++;
1915         ei->i_reserved_meta_blocks += md_needed;
1916         spin_unlock(&ei->i_block_reservation_lock);
1917
1918         return 0;       /* success */
1919 }
1920
1921 static void ext4_da_release_space(struct inode *inode, int to_free)
1922 {
1923         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1924         struct ext4_inode_info *ei = EXT4_I(inode);
1925
1926         if (!to_free)
1927                 return;         /* Nothing to release, exit */
1928
1929         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1930
1931         trace_ext4_da_release_space(inode, to_free);
1932         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1933                 /*
1934                  * if there aren't enough reserved blocks, then the
1935                  * counter is messed up somewhere.  Since this
1936                  * function is called from invalidate page, it's
1937                  * harmless to return without any action.
1938                  */
1939                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1940                          "ino %lu, to_free %d with only %d reserved "
1941                          "data blocks\n", inode->i_ino, to_free,
1942                          ei->i_reserved_data_blocks);
1943                 WARN_ON(1);
1944                 to_free = ei->i_reserved_data_blocks;
1945         }
1946         ei->i_reserved_data_blocks -= to_free;
1947
1948         if (ei->i_reserved_data_blocks == 0) {
1949                 /*
1950                  * We can release all of the reserved metadata blocks
1951                  * only when we have written all of the delayed
1952                  * allocation blocks.
1953                  */
1954                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1955                                    ei->i_reserved_meta_blocks);
1956                 ei->i_reserved_meta_blocks = 0;
1957                 ei->i_da_metadata_calc_len = 0;
1958         }
1959
1960         /* update fs dirty data blocks counter */
1961         percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1962
1963         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1964
1965         dquot_release_reservation_block(inode, to_free);
1966 }
1967
1968 static void ext4_da_page_release_reservation(struct page *page,
1969                                              unsigned long offset)
1970 {
1971         int to_release = 0;
1972         struct buffer_head *head, *bh;
1973         unsigned int curr_off = 0;
1974
1975         head = page_buffers(page);
1976         bh = head;
1977         do {
1978                 unsigned int next_off = curr_off + bh->b_size;
1979
1980                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1981                         to_release++;
1982                         clear_buffer_delay(bh);
1983                 }
1984                 curr_off = next_off;
1985         } while ((bh = bh->b_this_page) != head);
1986         ext4_da_release_space(page->mapping->host, to_release);
1987 }
1988
1989 /*
1990  * Delayed allocation stuff
1991  */
1992
1993 /*
1994  * mpage_da_submit_io - walks through extent of pages and try to write
1995  * them with writepage() call back
1996  *
1997  * @mpd->inode: inode
1998  * @mpd->first_page: first page of the extent
1999  * @mpd->next_page: page after the last page of the extent
2000  *
2001  * By the time mpage_da_submit_io() is called we expect all blocks
2002  * to be allocated. this may be wrong if allocation failed.
2003  *
2004  * As pages are already locked by write_cache_pages(), we can't use it
2005  */
2006 static int mpage_da_submit_io(struct mpage_da_data *mpd)
2007 {
2008         struct pagevec pvec;
2009         unsigned long index, end;
2010         int ret = 0, err, nr_pages, i;
2011         struct inode *inode = mpd->inode;
2012         struct address_space *mapping = inode->i_mapping;
2013         loff_t size = i_size_read(inode);
2014         unsigned int len;
2015         struct buffer_head *page_bufs = NULL;
2016         int journal_data = ext4_should_journal_data(inode);
2017
2018         BUG_ON(mpd->next_page <= mpd->first_page);
2019         /*
2020          * We need to start from the first_page to the next_page - 1
2021          * to make sure we also write the mapped dirty buffer_heads.
2022          * If we look at mpd->b_blocknr we would only be looking
2023          * at the currently mapped buffer_heads.
2024          */
2025         index = mpd->first_page;
2026         end = mpd->next_page - 1;
2027
2028         pagevec_init(&pvec, 0);
2029         while (index <= end) {
2030                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2031                 if (nr_pages == 0)
2032                         break;
2033                 for (i = 0; i < nr_pages; i++) {
2034                         int commit_write = 0;
2035                         struct page *page = pvec.pages[i];
2036
2037                         index = page->index;
2038                         if (index > end)
2039                                 break;
2040
2041                         if (index == size >> PAGE_CACHE_SHIFT)
2042                                 len = size & ~PAGE_CACHE_MASK;
2043                         else
2044                                 len = PAGE_CACHE_SIZE;
2045                         index++;
2046
2047                         BUG_ON(!PageLocked(page));
2048                         BUG_ON(PageWriteback(page));
2049
2050                         /*
2051                          * If the page does not have buffers (for
2052                          * whatever reason), try to create them using
2053                          * block_prepare_write.  If this fails,
2054                          * redirty the page and move on.
2055                          */
2056                         if (!page_has_buffers(page)) {
2057                                 if (block_prepare_write(page, 0, len,
2058                                                 noalloc_get_block_write)) {
2059                                 redirty_page:
2060                                         redirty_page_for_writepage(mpd->wbc,
2061                                                                    page);
2062                                         unlock_page(page);
2063                                         continue;
2064                                 }
2065                                 commit_write = 1;
2066                         }
2067                         page_bufs = page_buffers(page);
2068                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2069                                               ext4_bh_delay_or_unwritten)) {
2070                                 /*
2071                                  * We couldn't do block allocation for
2072                                  * some reason.
2073                                  */
2074                                 goto redirty_page;
2075                         }
2076
2077                         if (commit_write)
2078                                 /* mark the buffer_heads as dirty & uptodate */
2079                                 block_commit_write(page, 0, len);
2080
2081                         if (journal_data && PageChecked(page))
2082                                 err = __ext4_journalled_writepage(page, len);
2083                         else if (buffer_uninit(page_bufs)) {
2084                                 ext4_set_bh_endio(page_bufs, inode);
2085                                 err = block_write_full_page_endio(page,
2086                                         noalloc_get_block_write,
2087                                         mpd->wbc, ext4_end_io_buffer_write);
2088                         } else
2089                                 err = block_write_full_page(page,
2090                                             noalloc_get_block_write, mpd->wbc);
2091
2092                         if (!err)
2093                                 mpd->pages_written++;
2094                         /*
2095                          * In error case, we have to continue because
2096                          * remaining pages are still locked
2097                          */
2098                         if (ret == 0)
2099                                 ret = err;
2100                 }
2101                 pagevec_release(&pvec);
2102         }
2103         return ret;
2104 }
2105
2106 /*
2107  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
2108  *
2109  * the function goes through all passed space and put actual disk
2110  * block numbers into buffer heads, dropping BH_Delay and BH_Unwritten
2111  */
2112 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd,
2113                                  struct ext4_map_blocks *map)
2114 {
2115         struct inode *inode = mpd->inode;
2116         struct address_space *mapping = inode->i_mapping;
2117         int blocks = map->m_len;
2118         sector_t pblock = map->m_pblk, cur_logical;
2119         struct buffer_head *head, *bh;
2120         pgoff_t index, end;
2121         struct pagevec pvec;
2122         int nr_pages, i;
2123
2124         index = map->m_lblk >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2125         end = (map->m_lblk + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2126         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2127
2128         pagevec_init(&pvec, 0);
2129
2130         while (index <= end) {
2131                 /* XXX: optimize tail */
2132                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2133                 if (nr_pages == 0)
2134                         break;
2135                 for (i = 0; i < nr_pages; i++) {
2136                         struct page *page = pvec.pages[i];
2137
2138                         index = page->index;
2139                         if (index > end)
2140                                 break;
2141                         index++;
2142
2143                         BUG_ON(!PageLocked(page));
2144                         BUG_ON(PageWriteback(page));
2145                         BUG_ON(!page_has_buffers(page));
2146
2147                         bh = page_buffers(page);
2148                         head = bh;
2149
2150                         /* skip blocks out of the range */
2151                         do {
2152                                 if (cur_logical >= map->m_lblk)
2153                                         break;
2154                                 cur_logical++;
2155                         } while ((bh = bh->b_this_page) != head);
2156
2157                         do {
2158                                 if (cur_logical > map->m_lblk + (blocks - 1))
2159                                         break;
2160
2161                                 if (buffer_delay(bh) || buffer_unwritten(bh)) {
2162
2163                                         BUG_ON(bh->b_bdev != inode->i_sb->s_bdev);
2164
2165                                         if (buffer_delay(bh)) {
2166                                                 clear_buffer_delay(bh);
2167                                                 bh->b_blocknr = pblock;
2168                                         } else {
2169                                                 /*
2170                                                  * unwritten already should have
2171                                                  * blocknr assigned. Verify that
2172                                                  */
2173                                                 clear_buffer_unwritten(bh);
2174                                                 BUG_ON(bh->b_blocknr != pblock);
2175                                         }
2176
2177                                 } else if (buffer_mapped(bh))
2178                                         BUG_ON(bh->b_blocknr != pblock);
2179
2180                                 if (map->m_flags & EXT4_MAP_UNINIT)
2181                                         set_buffer_uninit(bh);
2182                                 cur_logical++;
2183                                 pblock++;
2184                         } while ((bh = bh->b_this_page) != head);
2185                 }
2186                 pagevec_release(&pvec);
2187         }
2188 }
2189
2190
2191 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2192                                         sector_t logical, long blk_cnt)
2193 {
2194         int nr_pages, i;
2195         pgoff_t index, end;
2196         struct pagevec pvec;
2197         struct inode *inode = mpd->inode;
2198         struct address_space *mapping = inode->i_mapping;
2199
2200         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2201         end   = (logical + blk_cnt - 1) >>
2202                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2203         while (index <= end) {
2204                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2205                 if (nr_pages == 0)
2206                         break;
2207                 for (i = 0; i < nr_pages; i++) {
2208                         struct page *page = pvec.pages[i];
2209                         if (page->index > end)
2210                                 break;
2211                         BUG_ON(!PageLocked(page));
2212                         BUG_ON(PageWriteback(page));
2213                         block_invalidatepage(page, 0);
2214                         ClearPageUptodate(page);
2215                         unlock_page(page);
2216                 }
2217                 index = pvec.pages[nr_pages - 1]->index + 1;
2218                 pagevec_release(&pvec);
2219         }
2220         return;
2221 }
2222
2223 static void ext4_print_free_blocks(struct inode *inode)
2224 {
2225         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2226         printk(KERN_CRIT "Total free blocks count %lld\n",
2227                ext4_count_free_blocks(inode->i_sb));
2228         printk(KERN_CRIT "Free/Dirty block details\n");
2229         printk(KERN_CRIT "free_blocks=%lld\n",
2230                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2231         printk(KERN_CRIT "dirty_blocks=%lld\n",
2232                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2233         printk(KERN_CRIT "Block reservation details\n");
2234         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2235                EXT4_I(inode)->i_reserved_data_blocks);
2236         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2237                EXT4_I(inode)->i_reserved_meta_blocks);
2238         return;
2239 }
2240
2241 /*
2242  * mpage_da_map_and_submit - go through given space, map them
2243  *       if necessary, and then submit them for I/O
2244  *
2245  * @mpd - bh describing space
2246  *
2247  * The function skips space we know is already mapped to disk blocks.
2248  *
2249  */
2250 static void mpage_da_map_and_submit(struct mpage_da_data *mpd)
2251 {
2252         int err, blks, get_blocks_flags;
2253         struct ext4_map_blocks map;
2254         sector_t next = mpd->b_blocknr;
2255         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2256         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2257         handle_t *handle = NULL;
2258
2259         /*
2260          * If the blocks are mapped already, or we couldn't accumulate
2261          * any blocks, then proceed immediately to the submission stage.
2262          */
2263         if ((mpd->b_size == 0) ||
2264             ((mpd->b_state  & (1 << BH_Mapped)) &&
2265              !(mpd->b_state & (1 << BH_Delay)) &&
2266              !(mpd->b_state & (1 << BH_Unwritten))))
2267                 goto submit_io;
2268
2269         handle = ext4_journal_current_handle();
2270         BUG_ON(!handle);
2271
2272         /*
2273          * Call ext4_map_blocks() to allocate any delayed allocation
2274          * blocks, or to convert an uninitialized extent to be
2275          * initialized (in the case where we have written into
2276          * one or more preallocated blocks).
2277          *
2278          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2279          * indicate that we are on the delayed allocation path.  This
2280          * affects functions in many different parts of the allocation
2281          * call path.  This flag exists primarily because we don't
2282          * want to change *many* call functions, so ext4_map_blocks()
2283          * will set the magic i_delalloc_reserved_flag once the
2284          * inode's allocation semaphore is taken.
2285          *
2286          * If the blocks in questions were delalloc blocks, set
2287          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2288          * variables are updated after the blocks have been allocated.
2289          */
2290         map.m_lblk = next;
2291         map.m_len = max_blocks;
2292         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
2293         if (ext4_should_dioread_nolock(mpd->inode))
2294                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2295         if (mpd->b_state & (1 << BH_Delay))
2296                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2297
2298         blks = ext4_map_blocks(handle, mpd->inode, &map, get_blocks_flags);
2299         if (blks < 0) {
2300                 struct super_block *sb = mpd->inode->i_sb;
2301
2302                 err = blks;
2303                 /*
2304                  * If get block returns EAGAIN or ENOSPC and there
2305                  * appears to be free blocks we will call
2306                  * ext4_writepage() for all of the pages which will
2307                  * just redirty the pages.
2308                  */
2309                 if (err == -EAGAIN)
2310                         goto submit_io;
2311
2312                 if (err == -ENOSPC &&
2313                     ext4_count_free_blocks(sb)) {
2314                         mpd->retval = err;
2315                         goto submit_io;
2316                 }
2317
2318                 /*
2319                  * get block failure will cause us to loop in
2320                  * writepages, because a_ops->writepage won't be able
2321                  * to make progress. The page will be redirtied by
2322                  * writepage and writepages will again try to write
2323                  * the same.
2324                  */
2325                 if (!(EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2326                         ext4_msg(sb, KERN_CRIT,
2327                                  "delayed block allocation failed for inode %lu "
2328                                  "at logical offset %llu with max blocks %zd "
2329                                  "with error %d", mpd->inode->i_ino,
2330                                  (unsigned long long) next,
2331                                  mpd->b_size >> mpd->inode->i_blkbits, err);
2332                         ext4_msg(sb, KERN_CRIT,
2333                                 "This should not happen!! Data will be lost\n");
2334                         if (err == -ENOSPC)
2335                                 ext4_print_free_blocks(mpd->inode);
2336                 }
2337                 /* invalidate all the pages */
2338                 ext4_da_block_invalidatepages(mpd, next,
2339                                 mpd->b_size >> mpd->inode->i_blkbits);
2340                 return;
2341         }
2342         BUG_ON(blks == 0);
2343
2344         if (map.m_flags & EXT4_MAP_NEW) {
2345                 struct block_device *bdev = mpd->inode->i_sb->s_bdev;
2346                 int i;
2347
2348                 for (i = 0; i < map.m_len; i++)
2349                         unmap_underlying_metadata(bdev, map.m_pblk + i);
2350         }
2351
2352         /*
2353          * If blocks are delayed marked, we need to
2354          * put actual blocknr and drop delayed bit
2355          */
2356         if ((mpd->b_state & (1 << BH_Delay)) ||
2357             (mpd->b_state & (1 << BH_Unwritten)))
2358                 mpage_put_bnr_to_bhs(mpd, &map);
2359
2360         if (ext4_should_order_data(mpd->inode)) {
2361                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2362                 if (err)
2363                         /* This only happens if the journal is aborted */
2364                         return;
2365         }
2366
2367         /*
2368          * Update on-disk size along with block allocation.
2369          */
2370         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2371         if (disksize > i_size_read(mpd->inode))
2372                 disksize = i_size_read(mpd->inode);
2373         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2374                 ext4_update_i_disksize(mpd->inode, disksize);
2375                 err = ext4_mark_inode_dirty(handle, mpd->inode);
2376                 if (err)
2377                         ext4_error(mpd->inode->i_sb,
2378                                    "Failed to mark inode %lu dirty",
2379                                    mpd->inode->i_ino);
2380         }
2381
2382 submit_io:
2383         mpage_da_submit_io(mpd);
2384         mpd->io_done = 1;
2385 }
2386
2387 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2388                 (1 << BH_Delay) | (1 << BH_Unwritten))
2389
2390 /*
2391  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2392  *
2393  * @mpd->lbh - extent of blocks
2394  * @logical - logical number of the block in the file
2395  * @bh - bh of the block (used to access block's state)
2396  *
2397  * the function is used to collect contig. blocks in same state
2398  */
2399 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2400                                    sector_t logical, size_t b_size,
2401                                    unsigned long b_state)
2402 {
2403         sector_t next;
2404         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2405
2406         /*
2407          * XXX Don't go larger than mballoc is willing to allocate
2408          * This is a stopgap solution.  We eventually need to fold
2409          * mpage_da_submit_io() into this function and then call
2410          * ext4_map_blocks() multiple times in a loop
2411          */
2412         if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
2413                 goto flush_it;
2414
2415         /* check if thereserved journal credits might overflow */
2416         if (!(ext4_test_inode_flag(mpd->inode, EXT4_INODE_EXTENTS))) {
2417                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2418                         /*
2419                          * With non-extent format we are limited by the journal
2420                          * credit available.  Total credit needed to insert
2421                          * nrblocks contiguous blocks is dependent on the
2422                          * nrblocks.  So limit nrblocks.
2423                          */
2424                         goto flush_it;
2425                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2426                                 EXT4_MAX_TRANS_DATA) {
2427                         /*
2428                          * Adding the new buffer_head would make it cross the
2429                          * allowed limit for which we have journal credit
2430                          * reserved. So limit the new bh->b_size
2431                          */
2432                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2433                                                 mpd->inode->i_blkbits;
2434                         /* we will do mpage_da_submit_io in the next loop */
2435                 }
2436         }
2437         /*
2438          * First block in the extent
2439          */
2440         if (mpd->b_size == 0) {
2441                 mpd->b_blocknr = logical;
2442                 mpd->b_size = b_size;
2443                 mpd->b_state = b_state & BH_FLAGS;
2444                 return;
2445         }
2446
2447         next = mpd->b_blocknr + nrblocks;
2448         /*
2449          * Can we merge the block to our big extent?
2450          */
2451         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2452                 mpd->b_size += b_size;
2453                 return;
2454         }
2455
2456 flush_it:
2457         /*
2458          * We couldn't merge the block to our extent, so we
2459          * need to flush current  extent and start new one
2460          */
2461         mpage_da_map_and_submit(mpd);
2462         return;
2463 }
2464
2465 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
2466 {
2467         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
2468 }
2469
2470 /*
2471  * __mpage_da_writepage - finds extent of pages and blocks
2472  *
2473  * @page: page to consider
2474  * @wbc: not used, we just follow rules
2475  * @data: context
2476  *
2477  * The function finds extents of pages and scan them for all blocks.
2478  */
2479 static int __mpage_da_writepage(struct page *page,
2480                                 struct writeback_control *wbc, void *data)
2481 {
2482         struct mpage_da_data *mpd = data;
2483         struct inode *inode = mpd->inode;
2484         struct buffer_head *bh, *head;
2485         sector_t logical;
2486
2487         /*
2488          * Can we merge this page to current extent?
2489          */
2490         if (mpd->next_page != page->index) {
2491                 /*
2492                  * Nope, we can't. So, we map non-allocated blocks
2493                  * and start IO on them
2494                  */
2495                 if (mpd->next_page != mpd->first_page) {
2496                         mpage_da_map_and_submit(mpd);
2497                         /*
2498                          * skip rest of the page in the page_vec
2499                          */
2500                         redirty_page_for_writepage(wbc, page);
2501                         unlock_page(page);
2502                         return MPAGE_DA_EXTENT_TAIL;
2503                 }
2504
2505                 /*
2506                  * Start next extent of pages ...
2507                  */
2508                 mpd->first_page = page->index;
2509
2510                 /*
2511                  * ... and blocks
2512                  */
2513                 mpd->b_size = 0;
2514                 mpd->b_state = 0;
2515                 mpd->b_blocknr = 0;
2516         }
2517
2518         mpd->next_page = page->index + 1;
2519         logical = (sector_t) page->index <<
2520                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2521
2522         if (!page_has_buffers(page)) {
2523                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2524                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2525                 if (mpd->io_done)
2526                         return MPAGE_DA_EXTENT_TAIL;
2527         } else {
2528                 /*
2529                  * Page with regular buffer heads, just add all dirty ones
2530                  */
2531                 head = page_buffers(page);
2532                 bh = head;
2533                 do {
2534                         BUG_ON(buffer_locked(bh));
2535                         /*
2536                          * We need to try to allocate
2537                          * unmapped blocks in the same page.
2538                          * Otherwise we won't make progress
2539                          * with the page in ext4_writepage
2540                          */
2541                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2542                                 mpage_add_bh_to_extent(mpd, logical,
2543                                                        bh->b_size,
2544                                                        bh->b_state);
2545                                 if (mpd->io_done)
2546                                         return MPAGE_DA_EXTENT_TAIL;
2547                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2548                                 /*
2549                                  * mapped dirty buffer. We need to update
2550                                  * the b_state because we look at
2551                                  * b_state in mpage_da_map_blocks. We don't
2552                                  * update b_size because if we find an
2553                                  * unmapped buffer_head later we need to
2554                                  * use the b_state flag of that buffer_head.
2555                                  */
2556                                 if (mpd->b_size == 0)
2557                                         mpd->b_state = bh->b_state & BH_FLAGS;
2558                         }
2559                         logical++;
2560                 } while ((bh = bh->b_this_page) != head);
2561         }
2562
2563         return 0;
2564 }
2565
2566 /*
2567  * This is a special get_blocks_t callback which is used by
2568  * ext4_da_write_begin().  It will either return mapped block or
2569  * reserve space for a single block.
2570  *
2571  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2572  * We also have b_blocknr = -1 and b_bdev initialized properly
2573  *
2574  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2575  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2576  * initialized properly.
2577  */
2578 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2579                                   struct buffer_head *bh, int create)
2580 {
2581         struct ext4_map_blocks map;
2582         int ret = 0;
2583         sector_t invalid_block = ~((sector_t) 0xffff);
2584
2585         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2586                 invalid_block = ~0;
2587
2588         BUG_ON(create == 0);
2589         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
2590
2591         map.m_lblk = iblock;
2592         map.m_len = 1;
2593
2594         /*
2595          * first, we need to know whether the block is allocated already
2596          * preallocated blocks are unmapped but should treated
2597          * the same as allocated blocks.
2598          */
2599         ret = ext4_map_blocks(NULL, inode, &map, 0);
2600         if (ret < 0)
2601                 return ret;
2602         if (ret == 0) {
2603                 if (buffer_delay(bh))
2604                         return 0; /* Not sure this could or should happen */
2605                 /*
2606                  * XXX: __block_prepare_write() unmaps passed block,
2607                  * is it OK?
2608                  */
2609                 ret = ext4_da_reserve_space(inode, iblock);
2610                 if (ret)
2611                         /* not enough space to reserve */
2612                         return ret;
2613
2614                 map_bh(bh, inode->i_sb, invalid_block);
2615                 set_buffer_new(bh);
2616                 set_buffer_delay(bh);
2617                 return 0;
2618         }
2619
2620         map_bh(bh, inode->i_sb, map.m_pblk);
2621         bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
2622
2623         if (buffer_unwritten(bh)) {
2624                 /* A delayed write to unwritten bh should be marked
2625                  * new and mapped.  Mapped ensures that we don't do
2626                  * get_block multiple times when we write to the same
2627                  * offset and new ensures that we do proper zero out
2628                  * for partial write.
2629                  */
2630                 set_buffer_new(bh);
2631                 set_buffer_mapped(bh);
2632         }
2633         return 0;
2634 }
2635
2636 /*
2637  * This function is used as a standard get_block_t calback function
2638  * when there is no desire to allocate any blocks.  It is used as a
2639  * callback function for block_prepare_write() and block_write_full_page().
2640  * These functions should only try to map a single block at a time.
2641  *
2642  * Since this function doesn't do block allocations even if the caller
2643  * requests it by passing in create=1, it is critically important that
2644  * any caller checks to make sure that any buffer heads are returned
2645  * by this function are either all already mapped or marked for
2646  * delayed allocation before calling  block_write_full_page().  Otherwise,
2647  * b_blocknr could be left unitialized, and the page write functions will
2648  * be taken by surprise.
2649  */
2650 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2651                                    struct buffer_head *bh_result, int create)
2652 {
2653         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2654         return _ext4_get_block(inode, iblock, bh_result, 0);
2655 }
2656
2657 static int bget_one(handle_t *handle, struct buffer_head *bh)
2658 {
2659         get_bh(bh);
2660         return 0;
2661 }
2662
2663 static int bput_one(handle_t *handle, struct buffer_head *bh)
2664 {
2665         put_bh(bh);
2666         return 0;
2667 }
2668
2669 static int __ext4_journalled_writepage(struct page *page,
2670                                        unsigned int len)
2671 {
2672         struct address_space *mapping = page->mapping;
2673         struct inode *inode = mapping->host;
2674         struct buffer_head *page_bufs;
2675         handle_t *handle = NULL;
2676         int ret = 0;
2677         int err;
2678
2679         ClearPageChecked(page);
2680         page_bufs = page_buffers(page);
2681         BUG_ON(!page_bufs);
2682         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2683         /* As soon as we unlock the page, it can go away, but we have
2684          * references to buffers so we are safe */
2685         unlock_page(page);
2686
2687         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2688         if (IS_ERR(handle)) {
2689                 ret = PTR_ERR(handle);
2690                 goto out;
2691         }
2692
2693         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2694                                 do_journal_get_write_access);
2695
2696         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2697                                 write_end_fn);
2698         if (ret == 0)
2699                 ret = err;
2700         err = ext4_journal_stop(handle);
2701         if (!ret)
2702                 ret = err;
2703
2704         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2705         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2706 out:
2707         return ret;
2708 }
2709
2710 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
2711 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
2712
2713 /*
2714  * Note that we don't need to start a transaction unless we're journaling data
2715  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2716  * need to file the inode to the transaction's list in ordered mode because if
2717  * we are writing back data added by write(), the inode is already there and if
2718  * we are writing back data modified via mmap(), noone guarantees in which
2719  * transaction the data will hit the disk. In case we are journaling data, we
2720  * cannot start transaction directly because transaction start ranks above page
2721  * lock so we have to do some magic.
2722  *
2723  * This function can get called via...
2724  *   - ext4_da_writepages after taking page lock (have journal handle)
2725  *   - journal_submit_inode_data_buffers (no journal handle)
2726  *   - shrink_page_list via pdflush (no journal handle)
2727  *   - grab_page_cache when doing write_begin (have journal handle)
2728  *
2729  * We don't do any block allocation in this function. If we have page with
2730  * multiple blocks we need to write those buffer_heads that are mapped. This
2731  * is important for mmaped based write. So if we do with blocksize 1K
2732  * truncate(f, 1024);
2733  * a = mmap(f, 0, 4096);
2734  * a[0] = 'a';
2735  * truncate(f, 4096);
2736  * we have in the page first buffer_head mapped via page_mkwrite call back
2737  * but other bufer_heads would be unmapped but dirty(dirty done via the
2738  * do_wp_page). So writepage should write the first block. If we modify
2739  * the mmap area beyond 1024 we will again get a page_fault and the
2740  * page_mkwrite callback will do the block allocation and mark the
2741  * buffer_heads mapped.
2742  *
2743  * We redirty the page if we have any buffer_heads that is either delay or
2744  * unwritten in the page.
2745  *
2746  * We can get recursively called as show below.
2747  *
2748  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2749  *              ext4_writepage()
2750  *
2751  * But since we don't do any block allocation we should not deadlock.
2752  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2753  */
2754 static int ext4_writepage(struct page *page,
2755                           struct writeback_control *wbc)
2756 {
2757         int ret = 0, commit_write = 0;
2758         loff_t size;
2759         unsigned int len;
2760         struct buffer_head *page_bufs = NULL;
2761         struct inode *inode = page->mapping->host;
2762
2763         trace_ext4_writepage(inode, page);
2764         size = i_size_read(inode);
2765         if (page->index == size >> PAGE_CACHE_SHIFT)
2766                 len = size & ~PAGE_CACHE_MASK;
2767         else
2768                 len = PAGE_CACHE_SIZE;
2769
2770         /*
2771          * If the page does not have buffers (for whatever reason),
2772          * try to create them using block_prepare_write.  If this
2773          * fails, redirty the page and move on.
2774          */
2775         if (!page_buffers(page)) {
2776                 if (block_prepare_write(page, 0, len,
2777                                         noalloc_get_block_write)) {
2778                 redirty_page:
2779                         redirty_page_for_writepage(wbc, page);
2780                         unlock_page(page);
2781                         return 0;
2782                 }
2783                 commit_write = 1;
2784         }
2785         page_bufs = page_buffers(page);
2786         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2787                               ext4_bh_delay_or_unwritten)) {
2788                 /*
2789                  * We don't want to do block allocation So redirty the
2790                  * page and return We may reach here when we do a
2791                  * journal commit via
2792                  * journal_submit_inode_data_buffers.  If we don't
2793                  * have mapping block we just ignore them. We can also
2794                  * reach here via shrink_page_list
2795                  */
2796                 goto redirty_page;
2797         }
2798         if (commit_write)
2799                 /* now mark the buffer_heads as dirty and uptodate */
2800                 block_commit_write(page, 0, len);
2801
2802         if (PageChecked(page) && ext4_should_journal_data(inode))
2803                 /*
2804                  * It's mmapped pagecache.  Add buffers and journal it.  There
2805                  * doesn't seem much point in redirtying the page here.
2806                  */
2807                 return __ext4_journalled_writepage(page, len);
2808
2809         if (buffer_uninit(page_bufs)) {
2810                 ext4_set_bh_endio(page_bufs, inode);
2811                 ret = block_write_full_page_endio(page, noalloc_get_block_write,
2812                                             wbc, ext4_end_io_buffer_write);
2813         } else
2814                 ret = block_write_full_page(page, noalloc_get_block_write,
2815                                             wbc);
2816
2817         return ret;
2818 }
2819
2820 /*
2821  * This is called via ext4_da_writepages() to
2822  * calulate the total number of credits to reserve to fit
2823  * a single extent allocation into a single transaction,
2824  * ext4_da_writpeages() will loop calling this before
2825  * the block allocation.
2826  */
2827
2828 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2829 {
2830         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2831
2832         /*
2833          * With non-extent format the journal credit needed to
2834          * insert nrblocks contiguous block is dependent on
2835          * number of contiguous block. So we will limit
2836          * number of contiguous block to a sane value
2837          */
2838         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) &&
2839             (max_blocks > EXT4_MAX_TRANS_DATA))
2840                 max_blocks = EXT4_MAX_TRANS_DATA;
2841
2842         return ext4_chunk_trans_blocks(inode, max_blocks);
2843 }
2844
2845 /*
2846  * write_cache_pages_da - walk the list of dirty pages of the given
2847  * address space and call the callback function (which usually writes
2848  * the pages).
2849  *
2850  * This is a forked version of write_cache_pages().  Differences:
2851  *      Range cyclic is ignored.
2852  *      no_nrwrite_index_update is always presumed true
2853  */
2854 static int write_cache_pages_da(struct address_space *mapping,
2855                                 struct writeback_control *wbc,
2856                                 struct mpage_da_data *mpd)
2857 {
2858         int ret = 0;
2859         int done = 0;
2860         struct pagevec pvec;
2861         int nr_pages;
2862         pgoff_t index;
2863         pgoff_t end;            /* Inclusive */
2864         long nr_to_write = wbc->nr_to_write;
2865
2866         pagevec_init(&pvec, 0);
2867         index = wbc->range_start >> PAGE_CACHE_SHIFT;
2868         end = wbc->range_end >> PAGE_CACHE_SHIFT;
2869
2870         while (!done && (index <= end)) {
2871                 int i;
2872
2873                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2874                               PAGECACHE_TAG_DIRTY,
2875                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2876                 if (nr_pages == 0)
2877                         break;
2878
2879                 for (i = 0; i < nr_pages; i++) {
2880                         struct page *page = pvec.pages[i];
2881
2882                         /*
2883                          * At this point, the page may be truncated or
2884                          * invalidated (changing page->mapping to NULL), or
2885                          * even swizzled back from swapper_space to tmpfs file
2886                          * mapping. However, page->index will not change
2887                          * because we have a reference on the page.
2888                          */
2889                         if (page->index > end) {
2890                                 done = 1;
2891                                 break;
2892                         }
2893
2894                         lock_page(page);
2895
2896                         /*
2897                          * Page truncated or invalidated. We can freely skip it
2898                          * then, even for data integrity operations: the page
2899                          * has disappeared concurrently, so there could be no
2900                          * real expectation of this data interity operation
2901                          * even if there is now a new, dirty page at the same
2902                          * pagecache address.
2903                          */
2904                         if (unlikely(page->mapping != mapping)) {
2905 continue_unlock:
2906                                 unlock_page(page);
2907                                 continue;
2908                         }
2909
2910                         if (!PageDirty(page)) {
2911                                 /* someone wrote it for us */
2912                                 goto continue_unlock;
2913                         }
2914
2915                         if (PageWriteback(page)) {
2916                                 if (wbc->sync_mode != WB_SYNC_NONE)
2917                                         wait_on_page_writeback(page);
2918                                 else
2919                                         goto continue_unlock;
2920                         }
2921
2922                         BUG_ON(PageWriteback(page));
2923                         if (!clear_page_dirty_for_io(page))
2924                                 goto continue_unlock;
2925
2926                         ret = __mpage_da_writepage(page, wbc, mpd);
2927                         if (unlikely(ret)) {
2928                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2929                                         unlock_page(page);
2930                                         ret = 0;
2931                                 } else {
2932                                         done = 1;
2933                                         break;
2934                                 }
2935                         }
2936
2937                         if (nr_to_write > 0) {
2938                                 nr_to_write--;
2939                                 if (nr_to_write == 0 &&
2940                                     wbc->sync_mode == WB_SYNC_NONE) {
2941                                         /*
2942                                          * We stop writing back only if we are
2943                                          * not doing integrity sync. In case of
2944                                          * integrity sync we have to keep going
2945                                          * because someone may be concurrently
2946                                          * dirtying pages, and we might have
2947                                          * synced a lot of newly appeared dirty
2948                                          * pages, but have not synced all of the
2949                                          * old dirty pages.
2950                                          */
2951                                         done = 1;
2952                                         break;
2953                                 }
2954                         }
2955                 }
2956                 pagevec_release(&pvec);
2957                 cond_resched();
2958         }
2959         return ret;
2960 }
2961
2962
2963 static int ext4_da_writepages(struct address_space *mapping,
2964                               struct writeback_control *wbc)
2965 {
2966         pgoff_t index;
2967         int range_whole = 0;
2968         handle_t *handle = NULL;
2969         struct mpage_da_data mpd;
2970         struct inode *inode = mapping->host;
2971         int pages_written = 0;
2972         long pages_skipped;
2973         unsigned int max_pages;
2974         int range_cyclic, cycled = 1, io_done = 0;
2975         int needed_blocks, ret = 0;
2976         long desired_nr_to_write, nr_to_writebump = 0;
2977         loff_t range_start = wbc->range_start;
2978         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2979
2980         trace_ext4_da_writepages(inode, wbc);
2981
2982         /*
2983          * No pages to write? This is mainly a kludge to avoid starting
2984          * a transaction for special inodes like journal inode on last iput()
2985          * because that could violate lock ordering on umount
2986          */
2987         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2988                 return 0;
2989
2990         /*
2991          * If the filesystem has aborted, it is read-only, so return
2992          * right away instead of dumping stack traces later on that
2993          * will obscure the real source of the problem.  We test
2994          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2995          * the latter could be true if the filesystem is mounted
2996          * read-only, and in that case, ext4_da_writepages should
2997          * *never* be called, so if that ever happens, we would want
2998          * the stack trace.
2999          */
3000         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
3001                 return -EROFS;
3002
3003         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
3004                 range_whole = 1;
3005
3006         range_cyclic = wbc->range_cyclic;
3007         if (wbc->range_cyclic) {
3008                 index = mapping->writeback_index;
3009                 if (index)
3010                         cycled = 0;
3011                 wbc->range_start = index << PAGE_CACHE_SHIFT;
3012                 wbc->range_end  = LLONG_MAX;
3013                 wbc->range_cyclic = 0;
3014         } else
3015                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3016
3017         /*
3018          * This works around two forms of stupidity.  The first is in
3019          * the writeback code, which caps the maximum number of pages
3020          * written to be 1024 pages.  This is wrong on multiple
3021          * levels; different architectues have a different page size,
3022          * which changes the maximum amount of data which gets
3023          * written.  Secondly, 4 megabytes is way too small.  XFS
3024          * forces this value to be 16 megabytes by multiplying
3025          * nr_to_write parameter by four, and then relies on its
3026          * allocator to allocate larger extents to make them
3027          * contiguous.  Unfortunately this brings us to the second
3028          * stupidity, which is that ext4's mballoc code only allocates
3029          * at most 2048 blocks.  So we force contiguous writes up to
3030          * the number of dirty blocks in the inode, or
3031          * sbi->max_writeback_mb_bump whichever is smaller.
3032          */
3033         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
3034         if (!range_cyclic && range_whole) {
3035                 if (wbc->nr_to_write == LONG_MAX)
3036                         desired_nr_to_write = wbc->nr_to_write;
3037                 else
3038                         desired_nr_to_write = wbc->nr_to_write * 8;
3039         } else
3040                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
3041                                                            max_pages);
3042         if (desired_nr_to_write > max_pages)
3043                 desired_nr_to_write = max_pages;
3044
3045         if (wbc->nr_to_write < desired_nr_to_write) {
3046                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
3047                 wbc->nr_to_write = desired_nr_to_write;
3048         }
3049
3050         mpd.wbc = wbc;
3051         mpd.inode = mapping->host;
3052
3053         pages_skipped = wbc->pages_skipped;
3054
3055 retry:
3056         while (!ret && wbc->nr_to_write > 0) {
3057
3058                 /*
3059                  * we  insert one extent at a time. So we need
3060                  * credit needed for single extent allocation.
3061                  * journalled mode is currently not supported
3062                  * by delalloc
3063                  */
3064                 BUG_ON(ext4_should_journal_data(inode));
3065                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
3066
3067                 /* start a new transaction*/
3068                 handle = ext4_journal_start(inode, needed_blocks);
3069                 if (IS_ERR(handle)) {
3070                         ret = PTR_ERR(handle);
3071                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
3072                                "%ld pages, ino %lu; err %d", __func__,
3073                                 wbc->nr_to_write, inode->i_ino, ret);
3074                         goto out_writepages;
3075                 }
3076
3077                 /*
3078                  * Now call __mpage_da_writepage to find the next
3079                  * contiguous region of logical blocks that need
3080                  * blocks to be allocated by ext4.  We don't actually
3081                  * submit the blocks for I/O here, even though
3082                  * write_cache_pages thinks it will, and will set the
3083                  * pages as clean for write before calling
3084                  * __mpage_da_writepage().
3085                  */
3086                 mpd.b_size = 0;
3087                 mpd.b_state = 0;
3088                 mpd.b_blocknr = 0;
3089                 mpd.first_page = 0;
3090                 mpd.next_page = 0;
3091                 mpd.io_done = 0;
3092                 mpd.pages_written = 0;
3093                 mpd.retval = 0;
3094                 ret = write_cache_pages_da(mapping, wbc, &mpd);
3095     &n