ext4: inline walk_page_buffers() into mpage_da_submit_io
[linux-2.6.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/jbd2.h>
29 #include <linux/highuid.h>
30 #include <linux/pagemap.h>
31 #include <linux/quotaops.h>
32 #include <linux/string.h>
33 #include <linux/buffer_head.h>
34 #include <linux/writeback.h>
35 #include <linux/pagevec.h>
36 #include <linux/mpage.h>
37 #include <linux/namei.h>
38 #include <linux/uio.h>
39 #include <linux/bio.h>
40 #include <linux/workqueue.h>
41 #include <linux/kernel.h>
42 #include <linux/slab.h>
43
44 #include "ext4_jbd2.h"
45 #include "xattr.h"
46 #include "acl.h"
47 #include "ext4_extents.h"
48
49 #include <trace/events/ext4.h>
50
51 #define MPAGE_DA_EXTENT_TAIL 0x01
52
53 static inline int ext4_begin_ordered_truncate(struct inode *inode,
54                                               loff_t new_size)
55 {
56         return jbd2_journal_begin_ordered_truncate(
57                                         EXT4_SB(inode->i_sb)->s_journal,
58                                         &EXT4_I(inode)->jinode,
59                                         new_size);
60 }
61
62 static void ext4_invalidatepage(struct page *page, unsigned long offset);
63 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, block_start;
2015         struct buffer_head *bh, *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
2068                         bh = page_bufs = page_buffers(page);
2069                         block_start = 0;
2070                         do {
2071                                 /* redirty page if block allocation undone */
2072                                 if (!bh || buffer_delay(bh) ||
2073                                     buffer_unwritten(bh))
2074                                         goto redirty_page;
2075                                 bh = bh->b_this_page;
2076                                 block_start += bh->b_size;
2077                         } while ((bh != page_bufs) && (block_start < len));
2078
2079                         if (commit_write)
2080                                 /* mark the buffer_heads as dirty & uptodate */
2081                                 block_commit_write(page, 0, len);
2082
2083                         if (journal_data && PageChecked(page))
2084                                 err = __ext4_journalled_writepage(page, len);
2085                         else if (buffer_uninit(page_bufs)) {
2086                                 ext4_set_bh_endio(page_bufs, inode);
2087                                 err = block_write_full_page_endio(page,
2088                                         noalloc_get_block_write,
2089                                         mpd->wbc, ext4_end_io_buffer_write);
2090                         } else
2091                                 err = block_write_full_page(page,
2092                                             noalloc_get_block_write, mpd->wbc);
2093
2094                         if (!err)
2095                                 mpd->pages_written++;
2096                         /*
2097                          * In error case, we have to continue because
2098                          * remaining pages are still locked
2099                          */
2100                         if (ret == 0)
2101                                 ret = err;
2102                 }
2103                 pagevec_release(&pvec);
2104         }
2105         return ret;
2106 }
2107
2108 /*
2109  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
2110  *
2111  * the function goes through all passed space and put actual disk
2112  * block numbers into buffer heads, dropping BH_Delay and BH_Unwritten
2113  */
2114 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd,
2115                                  struct ext4_map_blocks *map)
2116 {
2117         struct inode *inode = mpd->inode;
2118         struct address_space *mapping = inode->i_mapping;
2119         int blocks = map->m_len;
2120         sector_t pblock = map->m_pblk, cur_logical;
2121         struct buffer_head *head, *bh;
2122         pgoff_t index, end;
2123         struct pagevec pvec;
2124         int nr_pages, i;
2125
2126         index = map->m_lblk >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2127         end = (map->m_lblk + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2128         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2129
2130         pagevec_init(&pvec, 0);
2131
2132         while (index <= end) {
2133                 /* XXX: optimize tail */
2134                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2135                 if (nr_pages == 0)
2136                         break;
2137                 for (i = 0; i < nr_pages; i++) {
2138                         struct page *page = pvec.pages[i];
2139
2140                         index = page->index;
2141                         if (index > end)
2142                                 break;
2143                         index++;
2144
2145                         BUG_ON(!PageLocked(page));
2146                         BUG_ON(PageWriteback(page));
2147                         BUG_ON(!page_has_buffers(page));
2148
2149                         bh = page_buffers(page);
2150                         head = bh;
2151
2152                         /* skip blocks out of the range */
2153                         do {
2154                                 if (cur_logical >= map->m_lblk)
2155                                         break;
2156                                 cur_logical++;
2157                         } while ((bh = bh->b_this_page) != head);
2158
2159                         do {
2160                                 if (cur_logical > map->m_lblk + (blocks - 1))
2161                                         break;
2162
2163                                 if (buffer_delay(bh) || buffer_unwritten(bh)) {
2164
2165                                         BUG_ON(bh->b_bdev != inode->i_sb->s_bdev);
2166
2167                                         if (buffer_delay(bh)) {
2168                                                 clear_buffer_delay(bh);
2169                                                 bh->b_blocknr = pblock;
2170                                         } else {
2171                                                 /*
2172                                                  * unwritten already should have
2173                                                  * blocknr assigned. Verify that
2174                                                  */
2175                                                 clear_buffer_unwritten(bh);
2176                                                 BUG_ON(bh->b_blocknr != pblock);
2177                                         }
2178
2179                                 } else if (buffer_mapped(bh))
2180                                         BUG_ON(bh->b_blocknr != pblock);
2181
2182                                 if (map->m_flags & EXT4_MAP_UNINIT)
2183                                         set_buffer_uninit(bh);
2184                                 cur_logical++;
2185                                 pblock++;
2186                         } while ((bh = bh->b_this_page) != head);
2187                 }
2188                 pagevec_release(&pvec);
2189         }
2190 }
2191
2192
2193 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2194                                         sector_t logical, long blk_cnt)
2195 {
2196         int nr_pages, i;
2197         pgoff_t index, end;
2198         struct pagevec pvec;
2199         struct inode *inode = mpd->inode;
2200         struct address_space *mapping = inode->i_mapping;
2201
2202         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2203         end   = (logical + blk_cnt - 1) >>
2204                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2205         while (index <= end) {
2206                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2207                 if (nr_pages == 0)
2208                         break;
2209                 for (i = 0; i < nr_pages; i++) {
2210                         struct page *page = pvec.pages[i];
2211                         if (page->index > end)
2212                                 break;
2213                         BUG_ON(!PageLocked(page));
2214                         BUG_ON(PageWriteback(page));
2215                         block_invalidatepage(page, 0);
2216                         ClearPageUptodate(page);
2217                         unlock_page(page);
2218                 }
2219                 index = pvec.pages[nr_pages - 1]->index + 1;
2220                 pagevec_release(&pvec);
2221         }
2222         return;
2223 }
2224
2225 static void ext4_print_free_blocks(struct inode *inode)
2226 {
2227         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2228         printk(KERN_CRIT "Total free blocks count %lld\n",
2229                ext4_count_free_blocks(inode->i_sb));
2230         printk(KERN_CRIT "Free/Dirty block details\n");
2231         printk(KERN_CRIT "free_blocks=%lld\n",
2232                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2233         printk(KERN_CRIT "dirty_blocks=%lld\n",
2234                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2235         printk(KERN_CRIT "Block reservation details\n");
2236         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2237                EXT4_I(inode)->i_reserved_data_blocks);
2238         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2239                EXT4_I(inode)->i_reserved_meta_blocks);
2240         return;
2241 }
2242
2243 /*
2244  * mpage_da_map_and_submit - go through given space, map them
2245  *       if necessary, and then submit them for I/O
2246  *
2247  * @mpd - bh describing space
2248  *
2249  * The function skips space we know is already mapped to disk blocks.
2250  *
2251  */
2252 static void mpage_da_map_and_submit(struct mpage_da_data *mpd)
2253 {
2254         int err, blks, get_blocks_flags;
2255         struct ext4_map_blocks map;
2256         sector_t next = mpd->b_blocknr;
2257         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2258         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2259         handle_t *handle = NULL;
2260
2261         /*
2262          * If the blocks are mapped already, or we couldn't accumulate
2263          * any blocks, then proceed immediately to the submission stage.
2264          */
2265         if ((mpd->b_size == 0) ||
2266             ((mpd->b_state  & (1 << BH_Mapped)) &&
2267              !(mpd->b_state & (1 << BH_Delay)) &&
2268              !(mpd->b_state & (1 << BH_Unwritten))))
2269                 goto submit_io;
2270
2271         handle = ext4_journal_current_handle();
2272         BUG_ON(!handle);
2273
2274         /*
2275          * Call ext4_map_blocks() to allocate any delayed allocation
2276          * blocks, or to convert an uninitialized extent to be
2277          * initialized (in the case where we have written into
2278          * one or more preallocated blocks).
2279          *
2280          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2281          * indicate that we are on the delayed allocation path.  This
2282          * affects functions in many different parts of the allocation
2283          * call path.  This flag exists primarily because we don't
2284          * want to change *many* call functions, so ext4_map_blocks()
2285          * will set the magic i_delalloc_reserved_flag once the
2286          * inode's allocation semaphore is taken.
2287          *
2288          * If the blocks in questions were delalloc blocks, set
2289          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2290          * variables are updated after the blocks have been allocated.
2291          */
2292         map.m_lblk = next;
2293         map.m_len = max_blocks;
2294         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
2295         if (ext4_should_dioread_nolock(mpd->inode))
2296                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2297         if (mpd->b_state & (1 << BH_Delay))
2298                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2299
2300         blks = ext4_map_blocks(handle, mpd->inode, &map, get_blocks_flags);
2301         if (blks < 0) {
2302                 struct super_block *sb = mpd->inode->i_sb;
2303
2304                 err = blks;
2305                 /*
2306                  * If get block returns EAGAIN or ENOSPC and there
2307                  * appears to be free blocks we will call
2308                  * ext4_writepage() for all of the pages which will
2309                  * just redirty the pages.
2310                  */
2311                 if (err == -EAGAIN)
2312                         goto submit_io;
2313
2314                 if (err == -ENOSPC &&
2315                     ext4_count_free_blocks(sb)) {
2316                         mpd->retval = err;
2317                         goto submit_io;
2318                 }
2319
2320                 /*
2321                  * get block failure will cause us to loop in
2322                  * writepages, because a_ops->writepage won't be able
2323                  * to make progress. The page will be redirtied by
2324                  * writepage and writepages will again try to write
2325                  * the same.
2326                  */
2327                 if (!(EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2328                         ext4_msg(sb, KERN_CRIT,
2329                                  "delayed block allocation failed for inode %lu "
2330                                  "at logical offset %llu with max blocks %zd "
2331                                  "with error %d", mpd->inode->i_ino,
2332                                  (unsigned long long) next,
2333                                  mpd->b_size >> mpd->inode->i_blkbits, err);
2334                         ext4_msg(sb, KERN_CRIT,
2335                                 "This should not happen!! Data will be lost\n");
2336                         if (err == -ENOSPC)
2337                                 ext4_print_free_blocks(mpd->inode);
2338                 }
2339                 /* invalidate all the pages */
2340                 ext4_da_block_invalidatepages(mpd, next,
2341                                 mpd->b_size >> mpd->inode->i_blkbits);
2342                 return;
2343         }
2344         BUG_ON(blks == 0);
2345
2346         if (map.m_flags & EXT4_MAP_NEW) {
2347                 struct block_device *bdev = mpd->inode->i_sb->s_bdev;
2348                 int i;
2349
2350                 for (i = 0; i < map.m_len; i++)
2351                         unmap_underlying_metadata(bdev, map.m_pblk + i);
2352         }
2353
2354         /*
2355          * If blocks are delayed marked, we need to
2356          * put actual blocknr and drop delayed bit
2357          */
2358         if ((mpd->b_state & (1 << BH_Delay)) ||
2359             (mpd->b_state & (1 << BH_Unwritten)))
2360                 mpage_put_bnr_to_bhs(mpd, &map);
2361
2362         if (ext4_should_order_data(mpd->inode)) {
2363                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2364                 if (err)
2365                         /* This only happens if the journal is aborted */
2366                         return;
2367         }
2368
2369         /*
2370          * Update on-disk size along with block allocation.
2371          */
2372         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2373         if (disksize > i_size_read(mpd->inode))
2374                 disksize = i_size_read(mpd->inode);
2375         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2376                 ext4_update_i_disksize(mpd->inode, disksize);
2377                 err = ext4_mark_inode_dirty(handle, mpd->inode);
2378                 if (err)
2379                         ext4_error(mpd->inode->i_sb,
2380                                    "Failed to mark inode %lu dirty",
2381                                    mpd->inode->i_ino);
2382         }
2383
2384 submit_io:
2385         mpage_da_submit_io(mpd);
2386         mpd->io_done = 1;
2387 }
2388
2389 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2390                 (1 << BH_Delay) | (1 << BH_Unwritten))
2391
2392 /*
2393  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2394  *
2395  * @mpd->lbh - extent of blocks
2396  * @logical - logical number of the block in the file
2397  * @bh - bh of the block (used to access block's state)
2398  *
2399  * the function is used to collect contig. blocks in same state
2400  */
2401 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2402                                    sector_t logical, size_t b_size,
2403                                    unsigned long b_state)
2404 {
2405         sector_t next;
2406         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2407
2408         /*
2409          * XXX Don't go larger than mballoc is willing to allocate
2410          * This is a stopgap solution.  We eventually need to fold
2411          * mpage_da_submit_io() into this function and then call
2412          * ext4_map_blocks() multiple times in a loop
2413          */
2414         if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
2415                 goto flush_it;
2416
2417         /* check if thereserved journal credits might overflow */
2418         if (!(ext4_test_inode_flag(mpd->inode, EXT4_INODE_EXTENTS))) {
2419                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2420                         /*
2421                          * With non-extent format we are limited by the journal
2422                          * credit available.  Total credit needed to insert
2423                          * nrblocks contiguous blocks is dependent on the
2424                          * nrblocks.  So limit nrblocks.
2425                          */
2426                         goto flush_it;
2427                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2428                                 EXT4_MAX_TRANS_DATA) {
2429                         /*
2430                          * Adding the new buffer_head would make it cross the
2431                          * allowed limit for which we have journal credit
2432                          * reserved. So limit the new bh->b_size
2433                          */
2434                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2435                                                 mpd->inode->i_blkbits;
2436                         /* we will do mpage_da_submit_io in the next loop */
2437                 }
2438         }
2439         /*
2440          * First block in the extent
2441          */
2442         if (mpd->b_size == 0) {
2443                 mpd->b_blocknr = logical;
2444                 mpd->b_size = b_size;
2445                 mpd->b_state = b_state & BH_FLAGS;
2446                 return;
2447         }
2448
2449         next = mpd->b_blocknr + nrblocks;
2450         /*
2451          * Can we merge the block to our big extent?
2452          */
2453         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2454                 mpd->b_size += b_size;
2455                 return;
2456         }
2457
2458 flush_it:
2459         /*
2460          * We couldn't merge the block to our extent, so we
2461          * need to flush current  extent and start new one
2462          */
2463         mpage_da_map_and_submit(mpd);
2464         return;
2465 }
2466
2467 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
2468 {
2469         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
2470 }
2471
2472 /*
2473  * __mpage_da_writepage - finds extent of pages and blocks
2474  *
2475  * @page: page to consider
2476  * @wbc: not used, we just follow rules
2477  * @data: context
2478  *
2479  * The function finds extents of pages and scan them for all blocks.
2480  */
2481 static int __mpage_da_writepage(struct page *page,
2482                                 struct writeback_control *wbc, void *data)
2483 {
2484         struct mpage_da_data *mpd = data;
2485         struct inode *inode = mpd->inode;
2486         struct buffer_head *bh, *head;
2487         sector_t logical;
2488
2489         /*
2490          * Can we merge this page to current extent?
2491          */
2492         if (mpd->next_page != page->index) {
2493                 /*
2494                  * Nope, we can't. So, we map non-allocated blocks
2495                  * and start IO on them
2496                  */
2497                 if (mpd->next_page != mpd->first_page) {
2498                         mpage_da_map_and_submit(mpd);
2499                         /*
2500                          * skip rest of the page in the page_vec
2501                          */
2502                         redirty_page_for_writepage(wbc, page);
2503                         unlock_page(page);
2504                         return MPAGE_DA_EXTENT_TAIL;
2505                 }
2506
2507                 /*
2508                  * Start next extent of pages ...
2509                  */
2510                 mpd->first_page = page->index;
2511
2512                 /*
2513                  * ... and blocks
2514                  */
2515                 mpd->b_size = 0;
2516                 mpd->b_state = 0;
2517                 mpd->b_blocknr = 0;
2518         }
2519
2520         mpd->next_page = page->index + 1;
2521         logical = (sector_t) page->index <<
2522                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2523
2524         if (!page_has_buffers(page)) {
2525                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2526                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2527                 if (mpd->io_done)
2528                         return MPAGE_DA_EXTENT_TAIL;
2529         } else {
2530                 /*
2531                  * Page with regular buffer heads, just add all dirty ones
2532                  */
2533                 head = page_buffers(page);
2534                 bh = head;
2535                 do {
2536                         BUG_ON(buffer_locked(bh));
2537                         /*
2538                          * We need to try to allocate
2539                          * unmapped blocks in the same page.
2540                          * Otherwise we won't make progress
2541                          * with the page in ext4_writepage
2542                          */
2543                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2544                                 mpage_add_bh_to_extent(mpd, logical,
2545                                                        bh->b_size,
2546                                                        bh->b_state);
2547                                 if (mpd->io_done)
2548                                         return MPAGE_DA_EXTENT_TAIL;
2549                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2550                                 /*
2551                                  * mapped dirty buffer. We need to update
2552                                  * the b_state because we look at
2553                                  * b_state in mpage_da_map_blocks. We don't
2554                                  * update b_size because if we find an
2555                                  * unmapped buffer_head later we need to
2556                                  * use the b_state flag of that buffer_head.
2557                                  */
2558                                 if (mpd->b_size == 0)
2559                                         mpd->b_state = bh->b_state & BH_FLAGS;
2560                         }
2561                         logical++;
2562                 } while ((bh = bh->b_this_page) != head);
2563         }
2564
2565         return 0;
2566 }
2567
2568 /*
2569  * This is a special get_blocks_t callback which is used by
2570  * ext4_da_write_begin().  It will either return mapped block or
2571  * reserve space for a single block.
2572  *
2573  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2574  * We also have b_blocknr = -1 and b_bdev initialized properly
2575  *
2576  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2577  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2578  * initialized properly.
2579  */
2580 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2581                                   struct buffer_head *bh, int create)
2582 {
2583         struct ext4_map_blocks map;
2584         int ret = 0;
2585         sector_t invalid_block = ~((sector_t) 0xffff);
2586
2587         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2588                 invalid_block = ~0;
2589
2590         BUG_ON(create == 0);
2591         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
2592
2593         map.m_lblk = iblock;
2594         map.m_len = 1;
2595
2596         /*
2597          * first, we need to know whether the block is allocated already
2598          * preallocated blocks are unmapped but should treated
2599          * the same as allocated blocks.
2600          */
2601         ret = ext4_map_blocks(NULL, inode, &map, 0);
2602         if (ret < 0)
2603                 return ret;
2604         if (ret == 0) {
2605                 if (buffer_delay(bh))
2606                         return 0; /* Not sure this could or should happen */
2607                 /*
2608                  * XXX: __block_prepare_write() unmaps passed block,
2609                  * is it OK?
2610                  */
2611                 ret = ext4_da_reserve_space(inode, iblock);
2612                 if (ret)
2613                         /* not enough space to reserve */
2614                         return ret;
2615
2616                 map_bh(bh, inode->i_sb, invalid_block);
2617                 set_buffer_new(bh);
2618                 set_buffer_delay(bh);
2619                 return 0;
2620         }
2621
2622         map_bh(bh, inode->i_sb, map.m_pblk);
2623         bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
2624
2625         if (buffer_unwritten(bh)) {
2626                 /* A delayed write to unwritten bh should be marked
2627                  * new and mapped.  Mapped ensures that we don't do
2628                  * get_block multiple times when we write to the same
2629                  * offset and new ensures that we do proper zero out
2630                  * for partial write.
2631                  */
2632                 set_buffer_new(bh);
2633                 set_buffer_mapped(bh);
2634         }
2635         return 0;
2636 }
2637
2638 /*
2639  * This function is used as a standard get_block_t calback function
2640  * when there is no desire to allocate any blocks.  It is used as a
2641  * callback function for block_prepare_write() and block_write_full_page().
2642  * These functions should only try to map a single block at a time.
2643  *
2644  * Since this function doesn't do block allocations even if the caller
2645  * requests it by passing in create=1, it is critically important that
2646  * any caller checks to make sure that any buffer heads are returned
2647  * by this function are either all already mapped or marked for
2648  * delayed allocation before calling  block_write_full_page().  Otherwise,
2649  * b_blocknr could be left unitialized, and the page write functions will
2650  * be taken by surprise.
2651  */
2652 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2653                                    struct buffer_head *bh_result, int create)
2654 {
2655         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2656         return _ext4_get_block(inode, iblock, bh_result, 0);
2657 }
2658
2659 static int bget_one(handle_t *handle, struct buffer_head *bh)
2660 {
2661         get_bh(bh);
2662         return 0;
2663 }
2664
2665 static int bput_one(handle_t *handle, struct buffer_head *bh)
2666 {
2667         put_bh(bh);
2668         return 0;
2669 }
2670
2671 static int __ext4_journalled_writepage(struct page *page,
2672                                        unsigned int len)
2673 {
2674         struct address_space *mapping = page->mapping;
2675         struct inode *inode = mapping->host;
2676         struct buffer_head *page_bufs;
2677         handle_t *handle = NULL;
2678         int ret = 0;
2679         int err;
2680
2681         ClearPageChecked(page);
2682         page_bufs = page_buffers(page);
2683         BUG_ON(!page_bufs);
2684         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2685         /* As soon as we unlock the page, it can go away, but we have
2686          * references to buffers so we are safe */
2687         unlock_page(page);
2688
2689         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2690         if (IS_ERR(handle)) {
2691                 ret = PTR_ERR(handle);
2692                 goto out;
2693         }
2694
2695         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2696                                 do_journal_get_write_access);
2697
2698         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2699                                 write_end_fn);
2700         if (ret == 0)
2701                 ret = err;
2702         err = ext4_journal_stop(handle);
2703         if (!ret)
2704                 ret = err;
2705
2706         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2707         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2708 out:
2709         return ret;
2710 }
2711
2712 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
2713 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
2714
2715 /*
2716  * Note that we don't need to start a transaction unless we're journaling data
2717  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2718  * need to file the inode to the transaction's list in ordered mode because if
2719  * we are writing back data added by write(), the inode is already there and if
2720  * we are writing back data modified via mmap(), noone guarantees in which
2721  * transaction the data will hit the disk. In case we are journaling data, we
2722  * cannot start transaction directly because transaction start ranks above page
2723  * lock so we have to do some magic.
2724  *
2725  * This function can get called via...
2726  *   - ext4_da_writepages after taking page lock (have journal handle)
2727  *   - journal_submit_inode_data_buffers (no journal handle)
2728  *   - shrink_page_list via pdflush (no journal handle)
2729  *   - grab_page_cache when doing write_begin (have journal handle)
2730  *
2731  * We don't do any block allocation in this function. If we have page with
2732  * multiple blocks we need to write those buffer_heads that are mapped. This
2733  * is important for mmaped based write. So if we do with blocksize 1K
2734  * truncate(f, 1024);
2735  * a = mmap(f, 0, 4096);
2736  * a[0] = 'a';
2737  * truncate(f, 4096);
2738  * we have in the page first buffer_head mapped via page_mkwrite call back
2739  * but other bufer_heads would be unmapped but dirty(dirty done via the
2740  * do_wp_page). So writepage should write the first block. If we modify
2741  * the mmap area beyond 1024 we will again get a page_fault and the
2742  * page_mkwrite callback will do the block allocation and mark the
2743  * buffer_heads mapped.
2744  *
2745  * We redirty the page if we have any buffer_heads that is either delay or
2746  * unwritten in the page.
2747  *
2748  * We can get recursively called as show below.
2749  *
2750  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2751  *              ext4_writepage()
2752  *
2753  * But since we don't do any block allocation we should not deadlock.
2754  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2755  */
2756 static int ext4_writepage(struct page *page,
2757                           struct writeback_control *wbc)
2758 {
2759         int ret = 0, commit_write = 0;
2760         loff_t size;
2761         unsigned int len;
2762         struct buffer_head *page_bufs = NULL;
2763         struct inode *inode = page->mapping->host;
2764
2765         trace_ext4_writepage(inode, page);
2766         size = i_size_read(inode);
2767         if (page->index == size >> PAGE_CACHE_SHIFT)
2768                 len = size & ~PAGE_CACHE_MASK;
2769         else
2770                 len = PAGE_CACHE_SIZE;
2771
2772         /*
2773          * If the page does not have buffers (for whatever reason),
2774          * try to create them using block_prepare_write.  If this
2775          * fails, redirty the page and move on.
2776          */
2777         if (!page_buffers(page)) {
2778                 if (block_prepare_write(page, 0, len,
2779                                         noalloc_get_block_write)) {
2780                 redirty_page:
2781                         redirty_page_for_writepage(wbc, page);
2782                         unlock_page(page);
2783                         return 0;
2784                 }
2785                 commit_write = 1;
2786         }
2787         page_bufs = page_buffers(page);
2788         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2789                               ext4_bh_delay_or_unwritten)) {
2790                 /*
2791                  * We don't want to do block allocation So redirty the
2792                  * page and return We may reach here when we do a
2793                  * journal commit via
2794                  * journal_submit_inode_data_buffers.  If we don't
2795                  * have mapping block we just ignore them. We can also
2796                  * reach here via shrink_page_list
2797                  */
2798                 goto redirty_page;
2799         }
2800         if (commit_write)
2801                 /* now mark the buffer_heads as dirty and uptodate */
2802                 block_commit_write(page, 0, len);
2803
2804         if (PageChecked(page) && ext4_should_journal_data(inode))
2805                 /*
2806                  * It's mmapped pagecache.  Add buffers and journal it.  There
2807                  * doesn't seem much point in redirtying the page here.
2808                  */
2809                 return __ext4_journalled_writepage(page, len);
2810
2811         if (buffer_uninit(page_bufs)) {
2812                 ext4_set_bh_endio(page_bufs, inode);
2813                 ret = block_write_full_page_endio(page, noalloc_get_block_write,
2814                                             wbc, ext4_end_io_buffer_write);
2815         } else
2816                 ret = block_write_full_page(page, noalloc_get_block_write,
2817                                             wbc);
2818
2819         return ret;
2820 }
2821
2822 /*
2823  * This is called via ext4_da_writepages() to
2824  * calulate the total number of credits to reserve to fit
2825  * a single extent allocation into a single transaction,
2826  * ext4_da_writpeages() will loop calling this before
2827  * the block allocation.
2828  */
2829
2830 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2831 {
2832         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2833
2834         /*
2835          * With non-extent format the journal credit needed to
2836          * insert nrblocks contiguous block is dependent on
2837          * number of contiguous block. So we will limit
2838          * number of contiguous block to a sane value
2839          */
2840         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) &&
2841             (max_blocks > EXT4_MAX_TRANS_DATA))
2842                 max_blocks = EXT4_MAX_TRANS_DATA;
2843
2844         return ext4_chunk_trans_blocks(inode, max_blocks);
2845 }
2846
2847 /*
2848  * write_cache_pages_da - walk the list of dirty pages of the given
2849  * address space and call the callback function (which usually writes
2850  * the pages).
2851  *
2852  * This is a forked version of write_cache_pages().  Differences:
2853  *      Range cyclic is ignored.
2854  *      no_nrwrite_index_update is always presumed true
2855  */
2856 static int write_cache_pages_da(struct address_space *mapping,
2857                                 struct writeback_control *wbc,
2858                                 struct mpage_da_data *mpd)
2859 {
2860         int ret = 0;
2861         int done = 0;
2862         struct pagevec pvec;
2863         int nr_pages;
2864         pgoff_t index;
2865         pgoff_t end;            /* Inclusive */
2866         long nr_to_write = wbc->nr_to_write;
2867
2868         pagevec_init(&pvec, 0);
2869         index = wbc->range_start >> PAGE_CACHE_SHIFT;
2870         end = wbc->range_end >> PAGE_CACHE_SHIFT;
2871
2872         while (!done && (index <= end)) {
2873                 int i;
2874
2875                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2876                               PAGECACHE_TAG_DIRTY,
2877                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2878                 if (nr_pages == 0)
2879                         break;
2880
2881                 for (i = 0; i < nr_pages; i++) {
2882                         struct page *page = pvec.pages[i];
2883
2884                         /*
2885                          * At this point, the page may be truncated or
2886                          * invalidated (changing page->mapping to NULL), or
2887                          * even swizzled back from swapper_space to tmpfs file
2888                          * mapping. However, page->index will not change
2889                          * because we have a reference on the page.
2890                          */
2891                         if (page->index > end) {
2892                                 done = 1;
2893                                 break;
2894                         }
2895
2896                         lock_page(page);
2897
2898                         /*
2899                          * Page truncated or invalidated. We can freely skip it
2900                          * then, even for data integrity operations: the page
2901                          * has disappeared concurrently, so there could be no
2902                          * real expectation of this data interity operation
2903                          * even if there is now a new, dirty page at the same
2904                          * pagecache address.
2905                          */
2906                         if (unlikely(page->mapping != mapping)) {
2907 continue_unlock:
2908                                 unlock_page(page);
2909                                 continue;
2910                         }
2911
2912                         if (!PageDirty(page)) {
2913                                 /* someone wrote it for us */
2914                                 goto continue_unlock;
2915                         }
2916
2917                         if (PageWriteback(page)) {
2918                                 if (wbc->sync_mode != WB_SYNC_NONE)
2919                                         wait_on_page_writeback(page);
2920                                 else
2921                                         goto continue_unlock;
2922                         }
2923
2924                         BUG_ON(PageWriteback(page));
2925                         if (!clear_page_dirty_for_io(page))
2926                                 goto continue_unlock;
2927
2928                         ret = __mpage_da_writepage(page, wbc, mpd);
2929                         if (unlikely(ret)) {
2930                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2931                                         unlock_page(page);
2932                                         ret = 0;
2933                                 } else {
2934                                         done = 1;
2935                                         break;
2936                                 }
2937                         }
2938
2939                         if (nr_to_write > 0) {
2940                                 nr_to_write--;
2941                                 if (nr_to_write == 0 &&
2942                                     wbc->sync_mode == WB_SYNC_NONE) {
2943                                         /*
2944                                          * We stop writing back only if we are
2945                                          * not doing integrity sync. In case of
2946                                          * integrity sync we have to keep going
2947                                          * because someone may be concurrently
2948                                          * dirtying pages, and we might have
2949                                          * synced a lot of newly appeared dirty
2950                                          * pages, but have not synced all of the
2951                                          * old dirty pages.
2952                                          */
2953                                         done = 1;
2954                                         break;
2955                                 }
2956                         }
2957                 }
2958                 pagevec_release(&pvec);
2959                 cond_resched();
2960         }
2961         return ret;
2962 }
2963
2964
2965 static int ext4_da_writepages(struct address_space *mapping,
2966                               struct writeback_control *wbc)
2967 {
2968         pgoff_t index;
2969         int range_whole = 0;
2970         handle_t *handle = NULL;
2971         struct mpage_da_data mpd;
2972         struct inode *inode = mapping->host;
2973         int pages_written = 0;
2974         long pages_skipped;
2975         unsigned int max_pages;
2976         int range_cyclic, cycled = 1, io_done = 0;
2977         int needed_blocks, ret = 0;
2978         long desired_nr_to_write, nr_to_writebump = 0;
2979         loff_t range_start = wbc->range_start;
2980         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2981
2982         trace_ext4_da_writepages(inode, wbc);
2983
2984         /*
2985          * No pages to write? This is mainly a kludge to avoid starting
2986          * a transaction for special inodes like journal inode on last iput()
2987          * because that could violate lock ordering on umount
2988          */
2989         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2990                 return 0;
2991
2992         /*
2993          * If the filesystem has aborted, it is read-only, so return
2994          * right away instead of dumping stack traces later on that
2995          * will obscure the real source of the problem.  We test
2996          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2997          * the latter could be true if the filesystem is mounted
2998          * read-only, and in that case, ext4_da_writepages should
2999          * *never* be called, so if that ever happens, we would want
3000          * the stack trace.
3001          */
3002         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
3003                 return -EROFS;
3004
3005         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
3006                 range_whole = 1;
3007
3008         range_cyclic = wbc->range_cyclic;
3009         if (wbc->range_cyclic) {
3010                 index = mapping->writeback_index;
3011                 if (index)
3012                         cycled = 0;
3013                 wbc->range_start = index << PAGE_CACHE_SHIFT;
3014                 wbc->range_end  = LLONG_MAX;
3015                 wbc->range_cyclic = 0;
3016         } else
3017                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3018
3019         /*
3020          * This works around two forms of stupidity.  The first is in
3021          * the writeback code, which caps the maximum number of pages
3022          * written to be 1024 pages.  This is wrong on multiple
3023          * levels; different architectues have a different page size,
3024          * which changes the maximum amount of data which gets
3025          * written.  Secondly, 4 megabytes is way too small.  XFS
3026          * forces this value to be 16 megabytes by multiplying
3027          * nr_to_write parameter by four, and then relies on its
3028          * allocator to allocate larger extents to make them
3029          * contiguous.  Unfortunately this brings us to the second
3030          * stupidity, which is that ext4's mballoc code only allocates
3031          * at most 2048 blocks.  So we force contiguous writes up to
3032          * the number of dirty blocks in the inode, or
3033          * sbi->max_writeback_mb_bump whichever is smaller.
3034          */
3035         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
3036         if (!range_cyclic && range_whole) {
3037                 if (wbc->nr_to_write == LONG_MAX)
3038                         desired_nr_to_write = wbc->nr_to_write;
3039                 else
3040                         desired_nr_to_write = wbc->nr_to_write * 8;
3041         } else
3042                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
3043                                                            max_pages);
3044         if (desired_nr_to_write > max_pages)
3045                 desired_nr_to_write = max_pages;
3046
3047         if (wbc->nr_to_write < desired_nr_to_write) {
3048                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
3049                 wbc->nr_to_write = desired_nr_to_write;
3050         }
3051
3052         mpd.wbc = wbc;
3053         mpd.inode = mapping->host;
3054
3055         pages_skipped = wbc->pages_skipped;
3056
3057 retry:
3058         while (!ret && wbc->nr_to_write > 0) {
3059
3060                 /*
3061                  * we  insert one extent at a time. So we need
3062                  * credit needed for single extent allocation.
3063                  * journalled mode is currently not supported
3064                  * by delalloc
3065                  */
3066                 BUG_ON(ext4_should_journal_data(inode));
3067                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
3068
3069                 /* start a new transaction*/
3070                 handle = ext4_journal_start(inode, needed_blocks);
3071                 if (IS_ERR(handle)) {
3072                         ret = PTR_ERR(handle);
3073                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
3074                                "%ld pages, ino %lu; err %d", __func__,
3075                                 wbc->nr_to_write, inode->i_ino, ret);
3076                         goto out_writepages;
3077                 }
3078
3079                 /*
3080                  * Now call __mpage_da_writepage to find the next
3081                  * contiguous region of logical blocks that need
3082                  * blocks to be allocated by ext4.  We don't actually
3083                  * submit the blocks for I/O here, even though
3084                  * write_cache_pages thinks it will, and will set the
3085                  * pages as clean for write before calling
3086                  * __mpage_da_writepage().
3087                  */
3088                 mpd.b_size = 0;
3089                 mpd.b_state = 0;
3090                 mpd.b_blocknr = 0;
3091                 mpd.first_page = 0;
3092                 mpd.next_page = 0;
3093                 mpd.io_done = 0;
3094                 mpd.pages_written = 0;
3095                 mpd.retval = 0;
3096                 ret = write_cache_pages_da(mapping, wbc, &mpd);
3097                 /*
3098                  * If we have a contiguous extent of pages and we
3099                  * haven't done the I/O yet, map the blocks and submit
3100                  * them for I/O.
3101                  */
3102                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
3103                         mpage_da_map_and_submit(&mpd);
3104                         ret = MPAGE_DA_EXTENT_TAIL;
3105                 }
3106                 trace_ext4_da_write_pages(inode, &mpd);
3107                 wbc->nr_to_write -= mpd.pages_written;
3108
3109                 ext4_journal_stop(handle);
3110
3111                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
3112                         /* commit the transaction which would
3113                          * free blocks released in the transaction
3114                          * and try again
3115                          */
3116                         jbd2_journal_force_commit_nested(sbi->s_journal);
3117                         wbc->pages_skipped = pages_skipped;
3118                         ret = 0;
3119                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
3120                         /*
3121                          * got one extent now try with
3122                          * rest of the pages
3123                          */
3124                         pages_written += mpd.pages_written;
3125                         wbc->pages_skipped = pages_skipped;
3126                         ret = 0;
3127                         io_done = 1;
3128                 } else if (wbc->nr_to_write)
3129                         /*
3130                          * There is no more writeout needed
3131                          * or we requested for a noblocking writeout
3132                          * and we found the device congested
3133                          */
3134                         break;
3135         }
3136         if (!io_done && !cycled) {
3137                 cycled = 1;
3138                 index = 0;
3139                 wbc->range_start = index << PAGE_CACHE_SHIFT;
3140                 wbc->range_end  = mapping->writeback_index - 1;
3141                 goto retry;
3142         }
3143         if (pages_skipped != wbc->pages_skipped)
3144                 ext4_msg(inode->i_sb, KERN_CRIT,
3145                          "This should not happen leaving %s "
3146                          "with nr_to_write = %ld ret = %d",
3147                          __func__, wbc->nr_to_write, ret);
3148
3149         /* Update index */
3150         index += pages_written;
3151         wbc->range_cyclic = range_cyclic;
3152         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
3153                 /*
3154                  * set the writeback_index so that range_cyclic
3155                  * mode will write it back later
3156                  */
3157                 mapping->writeback_index = index;
3158
3159 out_writepages:
3160         wbc->nr_to_write -= nr_to_writebump;
3161         wbc->range_start = range_start;
3162         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
3163         return ret;
3164 }
3165
3166 #define FALL_BACK_TO_NONDELALLOC 1
3167 static int ext4_nonda_switch(struct super_block *sb)
3168 {
3169         s64 free_blocks, dirty_blocks;
3170         struct ext4_sb_info *sbi = EXT4_SB(sb);
3171
3172         /*
3173          * switch to non delalloc mode if we are running low
3174          * on free block. The free block accounting via percpu
3175          * counters can get slightly wrong with percpu_counter_batch getting
3176          * accumulated on each CPU without updating global counters
3177          * Delalloc need an accurate free block accounting. So switch
3178          * to non delalloc when we are near to error range.
3179          */
3180         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
3181         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
3182         if (2 * free_blocks < 3 * dirty_blocks ||
3183                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
3184                 /*
3185                  * free block count is less than 150% of dirty blocks
3186                  * or free blocks is less than watermark
3187                  */
3188                 return 1;
3189         }
3190         /*
3191          * Even if we don't switch but are nearing capacity,
3192          * start pushing delalloc when 1/2 of free blocks are dirty.
3193          */
3194         if (free_blocks < 2 * dirty_blocks)
3195                 writeback_inodes_sb_if_idle(sb);
3196
3197         return 0;
3198 }
3199
3200 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3201                                loff_t pos, unsigned len, unsigned flags,
3202                                struct page **pagep, void **fsdata)
3203 {
3204         int ret, retries = 0;
3205         struct page *page;
3206         pgoff_t index;
3207         struct inode *inode = mapping->host;
3208         handle_t *handle;
3209
3210         index = pos >> PAGE_CACHE_SHIFT;
3211
3212         if (ext4_nonda_switch(inode->i_sb)) {
3213                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3214                 return ext4_write_begin(file, mapping, pos,
3215                                         len, flags, pagep, fsdata);
3216         }
3217         *fsdata = (void *)0;
3218         trace_ext4_da_write_begin(inode, pos, len, flags);
3219 retry:
3220         /*
3221          * With delayed allocation, we don't log the i_disksize update
3222          * if there is delayed block allocation. But we still need
3223          * to journalling the i_disksize update if writes to the end
3224          * of file which has an already mapped buffer.
3225          */
3226         handle = ext4_journal_start(inode, 1);
3227         if (IS_ERR(handle)) {
3228                 ret = PTR_ERR(handle);
3229                 goto out;
3230         }
3231         /* We cannot recurse into the filesystem as the transaction is already
3232          * started */
3233         flags |= AOP_FLAG_NOFS;
3234
3235         page = grab_cache_page_write_begin(mapping, index, flags);
3236         if (!page) {
3237                 ext4_journal_stop(handle);
3238                 ret = -ENOMEM;
3239                 goto out;
3240         }
3241         *pagep = page;
3242
3243         ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
3244         if (ret < 0) {
3245                 unlock_page(page);
3246                 ext4_journal_stop(handle);
3247                 page_cache_release(page);
3248                 /*
3249                  * block_write_begin may have instantiated a few blocks
3250                  * outside i_size.  Trim these off again. Don't need
3251                  * i_size_read because we hold i_mutex.
3252                  */
3253                 if (pos + len > inode->i_size)
3254                         ext4_truncate_failed_write(inode);
3255         }
3256
3257         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3258                 goto retry;
3259 out:
3260         return ret;
3261 }
3262
3263 /*
3264  * Check if we should update i_disksize
3265  * when write to the end of file but not require block allocation
3266  */
3267 static int ext4_da_should_update_i_disksize(struct page *page,
3268                                             unsigned long offset)
3269 {
3270         struct buffer_head *bh;
3271         struct inode *inode = page->mapping->host;
3272         unsigned int idx;
3273         int i;
3274
3275         bh = page_buffers(page);
3276         idx = offset >> inode->i_blkbits;
3277
3278         for (i = 0; i < idx; i++)
3279                 bh = bh->b_this_page;
3280
3281         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3282                 return 0;
3283         return 1;
3284 }
3285
3286 static int ext4_da_write_end(struct file *file,
3287                              struct address_space *mapping,
3288                              loff_t pos, unsigned len, unsigned copied,
3289                              struct page *page, void *fsdata)
3290 {
3291         struct inode *inode = mapping->host;
3292         int ret = 0, ret2;
3293         handle_t *handle = ext4_journal_current_handle();
3294         loff_t new_i_size;
3295         unsigned long start, end;
3296         int write_mode = (int)(unsigned long)fsdata;
3297
3298         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
3299                 if (ext4_should_order_data(inode)) {
3300                         return ext4_ordered_write_end(file, mapping, pos,
3301                                         len, copied, page, fsdata);
3302                 } else if (ext4_should_writeback_data(inode)) {
3303                         return ext4_writeback_write_end(file, mapping, pos,
3304                                         len, copied, page, fsdata);
3305                 } else {
3306                         BUG();
3307                 }
3308         }
3309
3310         trace_ext4_da_write_end(inode, pos, len, copied);
3311         start = pos & (PAGE_CACHE_SIZE - 1);
3312         end = start + copied - 1;
3313
3314         /*
3315          * generic_write_end() will run mark_inode_dirty() if i_size
3316          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3317          * into that.
3318          */
3319
3320         new_i_size = pos + copied;
3321         if (new_i_size > EXT4_I(inode)->i_disksize) {
3322                 if (ext4_da_should_update_i_disksize(page, end)) {
3323                         down_write(&EXT4_I(inode)->i_data_sem);
3324                         if (new_i_size > EXT4_I(inode)->i_disksize) {
3325                                 /*
3326                                  * Updating i_disksize when extending file
3327                                  * without needing block allocation
3328                                  */
3329                                 if (ext4_should_order_data(inode))
3330                                         ret = ext4_jbd2_file_inode(handle,
3331                                                                    inode);
3332
3333                                 EXT4_I(inode)->i_disksize = new_i_size;
3334                         }
3335                         up_write(&EXT4_I(inode)->i_data_sem);
3336                         /* We need to mark inode dirty even if
3337                          * new_i_size is less that inode->i_size
3338                          * bu greater than i_disksize.(hint delalloc)
3339                          */
3340                         ext4_mark_inode_dirty(handle, inode);
3341                 }
3342         }
3343         ret2 = generic_write_end(file, mapping, pos, len, copied,
3344                                                         page, fsdata);
3345         copied = ret2;
3346         if (ret2 < 0)
3347                 ret = ret2;
3348         ret2 = ext4_journal_stop(handle);
3349         if (!ret)
3350                 ret = ret2;
3351
3352         return ret ? ret : copied;
3353 }
3354
3355 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
3356 {
3357         /*
3358          * Drop reserved blocks
3359          */
3360         BUG_ON(!PageLocked(page));
3361         if (!page_has_buffers(page))
3362                 goto out;
3363
3364         ext4_da_page_release_reservation(page, offset);
3365
3366 out:
3367         ext4_invalidatepage(page, offset);
3368
3369         return;
3370 }
3371
3372 /*
3373  * Force all delayed allocation blocks to be allocated for a given inode.
3374  */
3375 int ext4_alloc_da_blocks(struct inode *inode)
3376 {
3377         trace_ext4_alloc_da_blocks(inode);
3378
3379         if (!EXT4_I(inode)->i_reserved_data_blocks &&
3380             !EXT4_I(inode)->i_reserved_meta_blocks)
3381                 return 0;
3382
3383         /*
3384          * We do something simple for now.  The filemap_flush() will
3385          * also start triggering a write of the data blocks, which is
3386          * not strictly speaking necessary (and for users of
3387          * laptop_mode, not even desirable).  However, to do otherwise
3388          * would require replicating code paths in:
3389          *
3390          * ext4_da_writepages() ->
3391          *    write_cache_pages() ---> (via passed in callback function)
3392          *        __mpage_da_writepage() -->
3393          *           mpage_add_bh_to_extent()
3394          *           mpage_da_map_blocks()
3395          *
3396          * The problem is that write_cache_pages(), located in
3397          * mm/page-writeback.c, marks pages clean in preparation for
3398          * doing I/O, which is not desirable if we're not planning on
3399          * doing I/O at all.
3400          *
3401          * We could call write_cache_pages(), and then redirty all of
3402          * the pages by calling redirty_page_for_writeback() but that
3403          * would be ugly in the extreme.  So instead we would need to
3404          * replicate parts of the code in the above functions,
3405          * simplifying them becuase we wouldn't actually intend to
3406          * write out the pages, but rather only collect contiguous
3407          * logical block extents, call the multi-block allocator, and
3408          * then update the buffer heads with the block allocations.
3409          *
3410          * For now, though, we'll cheat by calling filemap_flush(),
3411          * which will map the blocks, and start the I/O, but not
3412          * actually wait for the I/O to complete.
3413          */
3414         return filemap_flush(inode->i_mapping);
3415 }
3416
3417 /*
3418  * bmap() is special.  It gets used by applications such as lilo and by
3419  * the swapper to find the on-disk block of a specific piece of data.
3420  *
3421  * Naturally, this is dangerous if the block concerned is still in the
3422  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3423  * filesystem and enables swap, then they may get a nasty shock when the
3424  * data getting swapped to that swapfile suddenly gets overwritten by
3425  * the original zero's written out previously to the journal and
3426  * awaiting writeback in the kernel's buffer cache.
3427  *
3428  * So, if we see any bmap calls here on a modified, data-journaled file,
3429  * take extra steps to flush any blocks which might be in the cache.
3430  */
3431 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3432 {
3433         struct inode *inode = mapping->host;
3434         journal_t *journal;
3435         int err;
3436
3437         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3438                         test_opt(inode->i_sb, DELALLOC)) {
3439                 /*
3440                  * With delalloc we want to sync the file
3441                  * so that we can make sure we allocate
3442                  * blocks for file
3443                  */
3444                 filemap_write_and_wait(mapping);
3445         }
3446
3447         if (EXT4_JOURNAL(inode) &&
3448             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3449                 /*
3450                  * This is a REALLY heavyweight approach, but the use of
3451                  * bmap on dirty files is expected to be extremely rare:
3452                  * only if we run lilo or swapon on a freshly made file
3453                  * do we expect this to happen.
3454                  *
3455                  * (bmap requires CAP_SYS_RAWIO so this does not
3456                  * represent an unprivileged user DOS attack --- we'd be
3457                  * in trouble if mortal users could trigger this path at
3458                  * will.)
3459                  *
3460                  * NB. EXT4_STATE_JDATA is not set on files other than
3461                  * regular files.  If somebody wants to bmap a directory
3462                  * or symlink and gets confused because the buffer
3463                  * hasn't yet been flushed to disk, they deserve
3464                  * everything they get.
3465                  */
3466
3467                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3468                 journal = EXT4_JOURNAL(inode);
3469                 jbd2_journal_lock_updates(journal);
3470                 err = jbd2_journal_flush(journal);
3471                 jbd2_journal_unlock_updates(journal);
3472
3473                 if (err)
3474                         return 0;
3475         }
3476
3477         return generic_block_bmap(mapping, block, ext4_get_block);
3478 }
3479
3480 static int ext4_readpage(struct file *file, struct page *page)
3481 {
3482         return mpage_readpage(page, ext4_get_block);
3483 }
3484
3485 static int
3486 ext4_readpages(struct file *file, struct address_space *mapping,
3487                 struct list_head *pages, unsigned nr_pages)
3488 {
3489         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3490 }
3491
3492 static void ext4_free_io_end(ext4_io_end_t *io)
3493 {
3494         BUG_ON(!io);
3495         if (io->page)
3496                 put_page(io->page);
3497         iput(io->inode);
3498         kfree(io);
3499 }
3500
3501 static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
3502 {
3503         struct buffer_head *head, *bh;
3504         unsigned int curr_off = 0;
3505
3506         if (!page_has_buffers(page))
3507                 return;
3508         head = bh = page_buffers(page);
3509         do {
3510                 if (offset <= curr_off && test_clear_buffer_uninit(bh)
3511                                         && bh->b_private) {
3512                         ext4_free_io_end(bh->b_private);
3513                         bh->b_private = NULL;
3514                         bh->b_end_io = NULL;
3515                 }
3516                 curr_off = curr_off + bh->b_size;
3517                 bh = bh->b_this_page;
3518         } while (bh != head);
3519 }
3520
3521 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3522 {
3523         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3524
3525         /*
3526          * free any io_end structure allocated for buffers to be discarded
3527          */
3528         if (ext4_should_dioread_nolock(page->mapping->host))
3529                 ext4_invalidatepage_free_endio(page, offset);
3530         /*
3531          * If it's a full truncate we just forget about the pending dirtying
3532          */
3533         if (offset == 0)
3534                 ClearPageChecked(page);
3535
3536         if (journal)
3537                 jbd2_journal_invalidatepage(journal, page, offset);
3538         else
3539                 block_invalidatepage(page, offset);
3540 }
3541
3542 static int ext4_releasepage(struct page *page, gfp_t wait)
3543 {
3544         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3545
3546         WARN_ON(PageChecked(page));
3547         if (!page_has_buffers(page))
3548                 return 0;
3549         if (journal)
3550                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3551         else
3552                 return try_to_free_buffers(page);
3553 }
3554
3555 /*
3556  * O_DIRECT for ext3 (or indirect map) based files
3557  *
3558  * If the O_DIRECT write will extend the file then add this inode to the
3559  * orphan list.  So recovery will truncate it back to the original size
3560  * if the machine crashes during the write.
3561  *
3562  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3563  * crashes then stale disk data _may_ be exposed inside the file. But current
3564  * VFS code falls back into buffered path in that case so we are safe.
3565  */
3566 static ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
3567                               const struct iovec *iov, loff_t offset,
3568                               unsigned long nr_segs)
3569 {
3570         struct file *file = iocb->ki_filp;
3571         struct inode *inode = file->f_mapping->host;
3572         struct ext4_inode_info *ei = EXT4_I(inode);
3573         handle_t *handle;
3574         ssize_t ret;
3575         int orphan = 0;
3576         size_t count = iov_length(iov, nr_segs);
3577         int retries = 0;
3578
3579         if (rw == WRITE) {
3580                 loff_t final_size = offset + count;
3581
3582                 if (final_size > inode->i_size) {
3583                         /* Credits for sb + inode write */
3584                         handle = ext4_journal_start(inode, 2);
3585                         if (IS_ERR(handle)) {
3586                                 ret = PTR_ERR(handle);
3587                                 goto out;
3588                         }
3589                         ret = ext4_orphan_add(handle, inode);
3590                         if (ret) {
3591                                 ext4_journal_stop(handle);
3592                                 goto out;
3593                         }
3594                         orphan = 1;
3595                         ei->i_disksize = inode->i_size;
3596                         ext4_journal_stop(handle);
3597                 }
3598         }
3599
3600 retry:
3601         if (rw == READ && ext4_should_dioread_nolock(inode))
3602                 ret = __blockdev_direct_IO(rw, iocb, inode,
3603                                  inode->i_sb->s_bdev, iov,
3604                                  offset, nr_segs,
3605                                  ext4_get_block, NULL, NULL, 0);
3606         else {
3607                 ret = blockdev_direct_IO(rw, iocb, inode,
3608                                  inode->i_sb->s_bdev, iov,
3609                                  offset, nr_segs,
3610                                  ext4_get_block, NULL);
3611
3612                 if (unlikely((rw & WRITE) && ret < 0)) {
3613                         loff_t isize = i_size_read(inode);
3614                         loff_t end = offset + iov_length(iov, nr_segs);
3615
3616                         if (end > isize)
3617                                 vmtruncate(inode, isize);
3618                 }
3619         }
3620         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3621                 goto retry;
3622
3623         if (orphan) {
3624                 int err;
3625
3626                 /* Credits for sb + inode write */
3627                 handle = ext4_journal_start(inode, 2);
3628                 if (IS_ERR(handle)) {
3629                         /* This is really bad luck. We've written the data
3630                          * but cannot extend i_size. Bail out and pretend
3631                          * the write failed... */
3632                         ret = PTR_ERR(handle);
3633                         if (inode->i_nlink)
3634                                 ext4_orphan_del(NULL, inode);
3635
3636                         goto out;
3637                 }
3638                 if (inode->i_nlink)
3639                         ext4_orphan_del(handle, inode);
3640                 if (ret > 0) {
3641                         loff_t end = offset + ret;
3642                         if (end > inode->i_size) {
3643                                 ei->i_disksize = end;
3644                                 i_size_write(inode, end);
3645                                 /*
3646                                  * We're going to return a positive `ret'
3647                                  * here due to non-zero-length I/O, so there's
3648                                  * no way of reporting error returns from
3649                                  * ext4_mark_inode_dirty() to userspace.  So
3650                                  * ignore it.
3651                                  */
3652                                 ext4_mark_inode_dirty(handle, inode);
3653                         }
3654                 }
3655                 err = ext4_journal_stop(handle);
3656                 if (ret == 0)
3657                         ret = err;
3658         }
3659 out:
3660         return ret;
3661 }
3662
3663 /*
3664  * ext4_get_block used when preparing for a DIO write or buffer write.
3665  * We allocate an uinitialized extent if blocks haven't been allocated.
3666  * The extent will be converted to initialized after the IO is complete.
3667  */
3668 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
3669                    struct buffer_head *bh_result, int create)
3670 {
3671         ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
3672                    inode->i_ino, create);
3673         return _ext4_get_block(inode, iblock, bh_result,
3674                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
3675 }
3676
3677 static void dump_completed_IO(struct inode * inode)
3678 {
3679 #ifdef  EXT4_DEBUG
3680         struct list_head *cur, *before, *after;
3681         ext4_io_end_t *io, *io0, *io1;
3682         unsigned long flags;
3683
3684         if (list_empty(&EXT4_I(inode)->i_completed_io_list)){
3685                 ext4_debug("inode %lu completed_io list is empty\n", inode->i_ino);
3686                 return;
3687         }
3688
3689         ext4_debug("Dump inode %lu completed_io list \n", inode->i_ino);
3690         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3691         list_for_each_entry(io, &EXT4_I(inode)->i_completed_io_list, list){
3692                 cur = &io->list;
3693                 before = cur->prev;
3694                 io0 = container_of(before, ext4_io_end_t, list);
3695                 after = cur->next;
3696                 io1 = container_of(after, ext4_io_end_t, list);
3697
3698                 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
3699                             io, inode->i_ino, io0, io1);
3700         }
3701         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3702 #endif
3703 }
3704
3705 /*
3706  * check a range of space and convert unwritten extents to written.
3707  */
3708 static int ext4_end_io_nolock(ext4_io_end_t *io)
3709 {
3710         struct inode *inode = io->inode;
3711         loff_t offset = io->offset;
3712         ssize_t size = io->size;
3713         int ret = 0;
3714
3715         ext4_debug("ext4_end_io_nolock: io 0x%p from inode %lu,list->next 0x%p,"
3716                    "list->prev 0x%p\n",
3717                    io, inode->i_ino, io->list.next, io->list.prev);
3718
3719         if (list_empty(&io->list))
3720                 return ret;
3721
3722         if (io->flag != EXT4_IO_UNWRITTEN)
3723                 return ret;
3724
3725         ret = ext4_convert_unwritten_extents(inode, offset, size);
3726         if (ret < 0) {
3727                 printk(KERN_EMERG "%s: failed to convert unwritten"
3728                         "extents to written extents, error is %d"
3729                         " io is still on inode %lu aio dio list\n",
3730                        __func__, ret, inode->i_ino);
3731                 return ret;
3732         }
3733
3734         if (io->iocb)
3735                 aio_complete(io->iocb, io->result, 0);
3736         /* clear the DIO AIO unwritten flag */
3737         io->flag = 0;
3738         return ret;
3739 }
3740
3741 /*
3742  * work on completed aio dio IO, to convert unwritten extents to extents
3743  */
3744 static void ext4_end_io_work(struct work_struct *work)
3745 {
3746         ext4_io_end_t           *io = container_of(work, ext4_io_end_t, work);
3747         struct inode            *inode = io->inode;
3748         struct ext4_inode_info  *ei = EXT4_I(inode);
3749         unsigned long           flags;
3750         int                     ret;
3751
3752         mutex_lock(&inode->i_mutex);
3753         ret = ext4_end_io_nolock(io);
3754         if (ret < 0) {
3755                 mutex_unlock(&inode->i_mutex);
3756                 return;
3757         }
3758
3759         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3760         if (!list_empty(&io->list))
3761                 list_del_init(&io->list);
3762         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3763         mutex_unlock(&inode->i_mutex);
3764         ext4_free_io_end(io);
3765 }
3766
3767 /*
3768  * This function is called from ext4_sync_file().
3769  *
3770  * When IO is completed, the work to convert unwritten extents to
3771  * written is queued on workqueue but may not get immediately
3772  * scheduled. When fsync is called, we need to ensure the
3773  * conversion is complete before fsync returns.
3774  * The inode keeps track of a list of pending/completed IO that
3775  * might needs to do the conversion. This function walks through
3776  * the list and convert the related unwritten extents for completed IO
3777  * to written.
3778  * The function return the number of pending IOs on success.
3779  */
3780 int flush_completed_IO(struct inode *inode)
3781 {
3782         ext4_io_end_t *io;
3783         struct ext4_inode_info *ei = EXT4_I(inode);
3784         unsigned long flags;
3785         int ret = 0;
3786         int ret2 = 0;
3787
3788         if (list_empty(&ei->i_completed_io_list))
3789                 return ret;
3790
3791         dump_completed_IO(inode);
3792         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3793         while (!list_empty(&ei->i_completed_io_list)){
3794                 io = list_entry(ei->i_completed_io_list.next,
3795                                 ext4_io_end_t, list);
3796                 /*
3797                  * Calling ext4_end_io_nolock() to convert completed
3798                  * IO to written.
3799                  *
3800                  * When ext4_sync_file() is called, run_queue() may already
3801                  * about to flush the work corresponding to this io structure.
3802                  * It will be upset if it founds the io structure related
3803                  * to the work-to-be schedule is freed.
3804                  *
3805                  * Thus we need to keep the io structure still valid here after
3806                  * convertion finished. The io structure has a flag to
3807                  * avoid double converting from both fsync and background work
3808                  * queue work.
3809                  */
3810                 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3811                 ret = ext4_end_io_nolock(io);
3812                 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3813                 if (ret < 0)
3814                         ret2 = ret;
3815                 else
3816                         list_del_init(&io->list);
3817         }
3818         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3819         return (ret2 < 0) ? ret2 : 0;
3820 }
3821
3822 static ext4_io_end_t *ext4_init_io_end (struct inode *inode, gfp_t flags)
3823 {
3824         ext4_io_end_t *io = NULL;
3825
3826         io = kmalloc(sizeof(*io), flags);
3827
3828         if (io) {
3829                 igrab(inode);
3830                 io->inode = inode;
3831                 io->flag = 0;
3832                 io->offset = 0;
3833                 io->size = 0;
3834                 io->page = NULL;
3835                 io->iocb = NULL;
3836                 io->result = 0;
3837                 INIT_WORK(&io->work, ext4_end_io_work);
3838                 INIT_LIST_HEAD(&io->list);
3839         }
3840
3841         return io;
3842 }
3843
3844 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3845                             ssize_t size, void *private, int ret,
3846                             bool is_async)
3847 {
3848         ext4_io_end_t *io_end = iocb->private;
3849         struct workqueue_struct *wq;
3850         unsigned long flags;
3851         struct ext4_inode_info *ei;
3852
3853         /* if not async direct IO or dio with 0 bytes write, just return */
3854         if (!io_end || !size)
3855                 goto out;
3856
3857         ext_debug("ext4_end_io_dio(): io_end 0x%p"
3858                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
3859                   iocb->private, io_end->inode->i_ino, iocb, offset,
3860                   size);
3861
3862         /* if not aio dio with unwritten extents, just free io and return */
3863         if (io_end->flag != EXT4_IO_UNWRITTEN){
3864                 ext4_free_io_end(io_end);
3865                 iocb->private = NULL;
3866 out:
3867                 if (is_async)
3868                         aio_complete(iocb, ret, 0);
3869                 return;
3870         }
3871
3872         io_end->offset = offset;
3873         io_end->size = size;
3874         if (is_async) {
3875                 io_end->iocb = iocb;
3876                 io_end->result = ret;
3877         }
3878         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
3879
3880         /* Add the io_end to per-inode completed aio dio list*/
3881         ei = EXT4_I(io_end->inode);
3882         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3883         list_add_tail(&io_end->list, &ei->i_completed_io_list);
3884         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3885
3886         /* queue the work to convert unwritten extents to written */
3887         queue_work(wq, &io_end->work);
3888         iocb->private = NULL;
3889 }
3890
3891 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
3892 {
3893         ext4_io_end_t *io_end = bh->b_private;
3894         struct workqueue_struct *wq;
3895         struct inode *inode;
3896         unsigned long flags;
3897
3898         if (!test_clear_buffer_uninit(bh) || !io_end)
3899                 goto out;
3900
3901         if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
3902                 printk("sb umounted, discard end_io request for inode %lu\n",
3903                         io_end->inode->i_ino);
3904                 ext4_free_io_end(io_end);
3905                 goto out;
3906         }
3907
3908         io_end->flag = EXT4_IO_UNWRITTEN;
3909         inode = io_end->inode;
3910
3911         /* Add the io_end to per-inode completed io list*/
3912         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3913         list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
3914         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3915
3916         wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
3917         /* queue the work to convert unwritten extents to written */
3918         queue_work(wq, &io_end->work);
3919 out:
3920         bh->b_private = NULL;
3921         bh->b_end_io = NULL;
3922         clear_buffer_uninit(bh);
3923         end_buffer_async_write(bh, uptodate);
3924 }
3925
3926 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
3927 {
3928         ext4_io_end_t *io_end;
3929         struct page *page = bh->b_page;
3930         loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
3931         size_t size = bh->b_size;
3932
3933 retry:
3934         io_end = ext4_init_io_end(inode, GFP_ATOMIC);
3935         if (!io_end) {
3936                 if (printk_ratelimit())
3937                         printk(KERN_WARNING "%s: allocation fail\n", __func__);
3938                 schedule();
3939                 goto retry;
3940         }
3941         io_end->offset = offset;
3942         io_end->size = size;
3943         /*
3944          * We need to hold a reference to the page to make sure it
3945          * doesn't get evicted before ext4_end_io_work() has a chance
3946          * to convert the extent from written to unwritten.
3947          */
3948         io_end->page = page;
3949         get_page(io_end->page);
3950
3951         bh->b_private = io_end;
3952         bh->b_end_io = ext4_end_io_buffer_write;
3953         return 0;
3954 }
3955
3956 /*
3957  * For ext4 extent files, ext4 will do direct-io write to holes,
3958  * preallocated extents, and those write extend the file, no need to
3959  * fall back to buffered IO.
3960  *
3961  * For holes, we fallocate those blocks, mark them as unintialized
3962  * If those blocks were preallocated, we mark sure they are splited, but
3963  * still keep the range to write as unintialized.
3964  *
3965  * The unwrritten extents will be converted to written when DIO is completed.
3966  * For async direct IO, since the IO may still pending when return, we
3967  * set up an end_io call back function, which will do the convertion
3968  * when async direct IO completed.
3969  *
3970  * If the O_DIRECT write will extend the file then add this inode to the
3971  * orphan list.  So recovery will truncate it back to the original size
3972  * if the machine crashes during the write.
3973  *
3974  */
3975 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
3976                               const struct iovec *iov, loff_t offset,
3977                               unsigned long nr_segs)
3978 {
3979         struct file *file = iocb->ki_filp;
3980         struct inode *inode = file->f_mapping->host;
3981         ssize_t ret;
3982         size_t count = iov_length(iov, nr_segs);
3983
3984         loff_t final_size = offset + count;
3985         if (rw == WRITE && final_size <= inode->i_size) {
3986                 /*
3987                  * We could direct write to holes and fallocate.
3988                  *
3989                  * Allocated blocks to fill the hole are marked as uninitialized
3990                  * to prevent paralel buffered read to expose the stale data
3991                  * before DIO complete the data IO.
3992                  *
3993                  * As to previously fallocated extents, ext4 get_block
3994                  * will just simply mark the buffer mapped but still
3995                  * keep the extents uninitialized.
3996                  *
3997                  * for non AIO case, we will convert those unwritten extents
3998                  * to written after return back from blockdev_direct_IO.
3999                  *
4000                  * for async DIO, the conversion needs to be defered when
4001                  * the IO is completed. The ext4 end_io callback function
4002                  * will be called to take care of the conversion work.
4003                  * Here for async case, we allocate an io_end structure to
4004                  * hook to the iocb.
4005                  */
4006                 iocb->private = NULL;
4007                 EXT4_I(inode)->cur_aio_dio = NULL;
4008                 if (!is_sync_kiocb(iocb)) {
4009                         iocb->private = ext4_init_io_end(inode, GFP_NOFS);
4010                         if (!iocb->private)
4011                                 return -ENOMEM;
4012                         /*
4013                          * we save the io structure for current async
4014                          * direct IO, so that later ext4_map_blocks()
4015                          * could flag the io structure whether there
4016                          * is a unwritten extents needs to be converted
4017                          * when IO is completed.
4018                          */
4019                         EXT4_I(inode)->cur_aio_dio = iocb->private;
4020                 }
4021
4022                 ret = blockdev_direct_IO(rw, iocb, inode,
4023                                          inode->i_sb->s_bdev, iov,
4024                                          offset, nr_segs,
4025                                          ext4_get_block_write,
4026                                          ext4_end_io_dio);
4027                 if (iocb->private)
4028                         EXT4_I(inode)->cur_aio_dio = NULL;
4029                 /*
4030                  * The io_end structure takes a reference to the inode,
4031                  * that structure needs to be destroyed and the
4032                  * reference to the inode need to be dropped, when IO is
4033                  * complete, even with 0 byte write, or failed.
4034                  *
4035                  * In the successful AIO DIO case, the io_end structure will be
4036                  * desctroyed and the reference to the inode will be dropped
4037                  * after the end_io call back function is called.
4038                  *
4039                  * In the case there is 0 byte write, or error case, since
4040                  * VFS direct IO won't invoke the end_io call back function,
4041                  * we need to free the end_io structure here.
4042                  */
4043                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
4044                         ext4_free_io_end(iocb->private);
4045                         iocb->private = NULL;
4046                 } else if (ret > 0 && ext4_test_inode_state(inode,
4047                                                 EXT4_STATE_DIO_UNWRITTEN)) {
4048                         int err;
4049                         /*
4050                          * for non AIO case, since the IO is already
4051                          * completed, we could do the convertion right here
4052                          */
4053                         err = ext4_convert_unwritten_extents(inode,
4054                                                              offset, ret);
4055                         if (err < 0)
4056                                 ret = err;
4057                         ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
4058                 }
4059                 return ret;
4060         }
4061
4062         /* for write the the end of file case, we fall back to old way */
4063         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
4064 }
4065
4066 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
4067                               const struct iovec *iov, loff_t offset,
4068                               unsigned long nr_segs)
4069 {
4070         struct file *file = iocb->ki_filp;
4071         struct inode *inode = file->f_mapping->host;
4072
4073         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4074                 return ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
4075
4076         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
4077 }
4078
4079 /*
4080  * Pages can be marked dirty completely asynchronously from ext4's journalling
4081  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
4082  * much here because ->set_page_dirty is called under VFS locks.  The page is
4083  * not necessarily locked.
4084  *
4085  * We cannot just dirty the page and leave attached buffers clean, because the
4086  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
4087  * or jbddirty because all the journalling code will explode.
4088  *
4089  * So what we do is to mark the page "pending dirty" and next time writepage
4090  * is called, propagate that into the buffers appropriately.
4091  */
4092 static int ext4_journalled_set_page_dirty(struct page *page)
4093 {
4094         SetPageChecked(page);
4095         return __set_page_dirty_nobuffers(page);
4096 }
4097
4098 static const struct address_space_operations ext4_ordered_aops = {
4099         .readpage               = ext4_readpage,
4100         .readpages              = ext4_readpages,
4101         .writepage              = ext4_writepage,
4102         .sync_page              = block_sync_page,
4103         .write_begin            = ext4_write_begin,
4104         .write_end              = ext4_ordered_write_end,
4105         .bmap                   = ext4_bmap,
4106         .invalidatepage         = ext4_invalidatepage,
4107         .releasepage            = ext4_releasepage,
4108         .direct_IO              = ext4_direct_IO,
4109         .migratepage            = buffer_migrate_page,
4110         .is_partially_uptodate  = block_is_partially_uptodate,
4111         .error_remove_page      = generic_error_remove_page,
4112 };
4113
4114 static const struct address_space_operations ext4_writeback_aops = {
4115         .readpage               = ext4_readpage,
4116         .readpages              = ext4_readpages,
4117         .writepage              = ext4_writepage,
4118         .sync_page              = block_sync_page,
4119         .write_begin            = ext4_write_begin,
4120         .write_end              = ext4_writeback_write_end,
4121         .bmap                   = ext4_bmap,
4122         .invalidatepage         = ext4_invalidatepage,
4123         .releasepage            = ext4_releasepage,
4124         .direct_IO              = ext4_direct_IO,
4125         .migratepage            = buffer_migrate_page,
4126         .is_partially_uptodate  = block_is_partially_uptodate,
4127         .error_remove_page      = generic_error_remove_page,
4128 };
4129
4130 static const struct address_space_operations ext4_journalled_aops = {
4131         .readpage               = ext4_readpage,
4132         .readpages              = ext4_readpages,
4133         .writepage              = ext4_writepage,
4134         .sync_page              = block_sync_page,
4135         .write_begin            = ext4_write_begin,
4136         .write_end              = ext4_journalled_write_end,
4137         .set_page_dirty         = ext4_journalled_set_page_dirty,
4138         .bmap                   = ext4_bmap,
4139         .invalidatepage         = ext4_invalidatepage,
4140         .releasepage            = ext4_releasepage,
4141         .is_partially_uptodate  = block_is_partially_uptodate,
4142         .error_remove_page      = generic_error_remove_page,
4143 };
4144
4145 static const struct address_space_operations ext4_da_aops = {
4146         .readpage               = ext4_readpage,
4147         .readpages              = ext4_readpages,
4148         .writepage              = ext4_writepage,
4149         .writepages             = ext4_da_writepages,
4150         .sync_page              = block_sync_page,
4151         .write_begin            = ext4_da_write_begin,
4152         .write_end              = ext4_da_write_end,
4153         .bmap                   = ext4_bmap,
4154         .invalidatepage         = ext4_da_invalidatepage,
4155         .releasepage            = ext4_releasepage,
4156         .direct_IO              = ext4_direct_IO,
4157         .migratepage            = buffer_migrate_page,
4158         .is_partially_uptodate  = block_is_partially_uptodate,
4159         .error_remove_page      = generic_error_remove_page,
4160 };
4161
4162 void ext4_set_aops(struct inode *inode)
4163 {
4164         if (ext4_should_order_data(inode) &&
4165                 test_opt(inode->i_sb, DELALLOC))
4166                 inode->i_mapping->a_ops = &ext4_da_aops;
4167         else if (ext4_should_order_data(inode))
4168                 inode->i_mapping->a_ops = &ext4_ordered_aops;
4169         else if (ext4_should_writeback_data(inode) &&
4170                  test_opt(inode->i_sb, DELALLOC))
4171                 inode->i_mapping->a_ops = &ext4_da_aops;
4172         else if (ext4_should_writeback_data(inode))
4173                 inode->i_mapping->a_ops = &ext4_writeback_aops;
4174         else
4175                 inode->i_mapping->a_ops = &ext4_journalled_aops;
4176 }
4177
4178 /*
4179  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
4180  * up to the end of the block which corresponds to `from'.
4181  * This required during truncate. We need to physically zero the tail end
4182  * of that block so it doesn't yield old data if the file is later grown.
4183  */
4184 int ext4_block_truncate_page(handle_t *handle,
4185                 struct address_space *mapping, loff_t from)
4186 {
4187         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
4188         unsigned offset = from & (PAGE_CACHE_SIZE-1);
4189         unsigned blocksize, length, pos;
4190         ext4_lblk_t iblock;
4191         struct inode *inode = mapping->host;
4192         struct buffer_head *bh;
4193         struct page *page;
4194         int err = 0;
4195
4196         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
4197                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
4198         if (!page)
4199                 return -EINVAL;
4200
4201         blocksize = inode->i_sb->s_blocksize;
4202         length = blocksize - (offset & (blocksize - 1));
4203         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
4204
4205         if (!page_has_buffers(page))
4206                 create_empty_buffers(page, blocksize, 0);
4207
4208         /* Find the buffer that contains "offset" */
4209         bh = page_buffers(page);
4210         pos = blocksize;
4211         while (offset >= pos) {
4212                 bh = bh->b_this_page;
4213                 iblock++;
4214                 pos += blocksize;
4215         }
4216
4217         err = 0;
4218         if (buffer_freed(bh)) {
4219                 BUFFER_TRACE(bh, "freed: skip");
4220                 goto unlock;
4221         }
4222
4223         if (!buffer_mapped(bh)) {
4224                 BUFFER_TRACE(bh, "unmapped");
4225                 ext4_get_block(inode, iblock, bh, 0);
4226                 /* unmapped? It's a hole - nothing to do */
4227                 if (!buffer_mapped(bh)) {
4228                         BUFFER_TRACE(bh, "still unmapped");
4229                         goto unlock;
4230                 }
4231         }
4232
4233         /* Ok, it's mapped. Make sure it's up-to-date */
4234         if (PageUptodate(page))
4235                 set_buffer_uptodate(bh);
4236
4237         if (!buffer_uptodate(bh)) {
4238                 err = -EIO;
4239                 ll_rw_block(READ, 1, &bh);
4240                 wait_on_buffer(bh);
4241                 /* Uhhuh. Read error. Complain and punt. */
4242                 if (!buffer_uptodate(bh))
4243                         goto unlock;
4244         }
4245
4246         if (ext4_should_journal_data(inode)) {
4247                 BUFFER_TRACE(bh, "get write access");
4248                 err = ext4_journal_get_write_access(handle, bh);
4249                 if (err)
4250                         goto unlock;
4251         }
4252
4253         zero_user(page, offset, length);
4254
4255         BUFFER_TRACE(bh, "zeroed end of block");
4256
4257         err = 0;
4258         if (ext4_should_journal_data(inode)) {
4259                 err = ext4_handle_dirty_metadata(handle, inode, bh);
4260         } else {
4261                 if (ext4_should_order_data(inode))
4262                         err = ext4_jbd2_file_inode(handle, inode);
4263                 mark_buffer_dirty(bh);
4264         }
4265
4266 unlock:
4267         unlock_page(page);
4268         page_cache_release(page);
4269         return err;
4270 }
4271
4272 /*
4273  * Probably it should be a library function... search for first non-zero word
4274  * or memcmp with zero_page, whatever is better for particular architecture.
4275  * Linus?
4276  */
4277 static inline int all_zeroes(__le32 *p, __le32 *q)
4278 {
4279         while (p < q)
4280                 if (*p++)
4281                         return 0;
4282         return 1;
4283 }
4284
4285 /**
4286  *      ext4_find_shared - find the indirect blocks for partial truncation.
4287  *      @inode:   inode in question
4288  *      @depth:   depth of the affected branch
4289  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
4290  *      @chain:   place to store the pointers to partial indirect blocks
4291  *      @top:     place to the (detached) top of branch
4292  *
4293  *      This is a helper function used by ext4_truncate().
4294  *
4295  *      When we do truncate() we may have to clean the ends of several
4296  *      indirect blocks but leave the blocks themselves alive. Block is
4297  *      partially truncated if some data below the new i_size is refered
4298  *      from it (and it is on the path to the first completely truncated
4299  *      data block, indeed).  We have to free the top of that path along
4300  *      with everything to the right of the path. Since no allocation
4301  *      past the truncation point is possible until ext4_truncate()
4302  *      finishes, we may safely do the latter, but top of branch may
4303  *      require special attention - pageout below the truncation point
4304  *      might try to populate it.
4305  *
4306  *      We atomically detach the top of branch from the tree, store the
4307  *      block number of its root in *@top, pointers to buffer_heads of
4308  *      partially truncated blocks - in @chain[].bh and pointers to
4309  *      their last elements that should not be removed - in
4310  *      @chain[].p. Return value is the pointer to last filled element
4311  *      of @chain.
4312  *
4313  *      The work left to caller to do the actual freeing of subtrees:
4314  *              a) free the subtree starting from *@top
4315  *              b) free the subtrees whose roots are stored in
4316  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
4317  *              c) free the subtrees growing from the inode past the @chain[0].
4318  *                      (no partially truncated stuff there).  */
4319
4320 static Indirect *ext4_find_shared(struct inode *inode, int depth,
4321                                   ext4_lblk_t offsets[4], Indirect chain[4],
4322                                   __le32 *top)
4323 {
4324         Indirect *partial, *p;
4325         int k, err;
4326
4327         *top = 0;
4328         /* Make k index the deepest non-null offset + 1 */
4329         for (k = depth; k > 1 && !offsets[k-1]; k--)
4330                 ;
4331         partial = ext4_get_branch(inode, k, offsets, chain, &err);
4332         /* Writer: pointers */
4333         if (!partial)
4334                 partial = chain + k-1;
4335         /*
4336          * If the branch acquired continuation since we've looked at it -
4337          * fine, it should all survive and (new) top doesn't belong to us.
4338          */
4339         if (!partial->key && *partial->p)
4340                 /* Writer: end */
4341                 goto no_top;
4342         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
4343                 ;
4344         /*
4345          * OK, we've found the last block that must survive. The rest of our
4346          * branch should be detached before unlocking. However, if that rest
4347          * of branch is all ours and does not grow immediately from the inode
4348          * it's easier to cheat and just decrement partial->p.
4349          */
4350         if (p == chain + k - 1 && p > chain) {
4351                 p->p--;
4352         } else {
4353                 *top = *p->p;
4354                 /* Nope, don't do this in ext4.  Must leave the tree intact */
4355 #if 0
4356                 *p->p = 0;
4357 #endif
4358         }
4359         /* Writer: end */
4360
4361         while (partial > p) {
4362                 brelse(partial->bh);
4363                 partial--;
4364         }
4365 no_top:
4366         return partial;
4367 }
4368
4369 /*
4370  * Zero a number of block pointers in either an inode or an indirect block.
4371  * If we restart the transaction we must again get write access to the
4372  * indirect block for further modification.
4373  *
4374  * We release `count' blocks on disk, but (last - first) may be greater
4375  * than `count' because there can be holes in there.
4376  */
4377 static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
4378                              struct buffer_head *bh,
4379                              ext4_fsblk_t block_to_free,
4380                              unsigned long count, __le32 *first,
4381                              __le32 *last)
4382 {
4383         __le32 *p;
4384         int     flags = EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_VALIDATED;
4385
4386         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
4387                 flags |= EXT4_FREE_BLOCKS_METADATA;
4388
4389         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
4390                                    count)) {
4391                 EXT4_ERROR_INODE(inode, "attempt to clear invalid "
4392                                  "blocks %llu len %lu",
4393                                  (unsigned long long) block_to_free, count);
4394                 return 1;
4395         }
4396
4397         if (try_to_extend_transaction(handle, inode)) {
4398                 if (bh) {
4399                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4400                         ext4_handle_dirty_metadata(handle, inode, bh);
4401                 }
4402                 ext4_mark_inode_dirty(handle, inode);
4403                 ext4_truncate_restart_trans(handle, inode,
4404                                             blocks_for_truncate(inode));
4405                 if (bh) {
4406                         BUFFER_TRACE(bh, "retaking write access");
4407                         ext4_journal_get_write_access(handle, bh);
4408                 }
4409         }
4410
4411         for (p = first; p < last; p++)
4412                 *p = 0;
4413
4414         ext4_free_blocks(handle, inode, 0, block_to_free, count, flags);
4415         return 0;
4416 }
4417
4418 /**
4419  * ext4_free_data - free a list of data blocks
4420  * @handle:     handle for this transaction
4421  * @inode:      inode we are dealing with
4422  * @this_bh:    indirect buffer_head which contains *@first and *@last
4423  * @first:      array of block numbers
4424  * @last:       points immediately past the end of array
4425  *
4426  * We are freeing all blocks refered from that array (numbers are stored as
4427  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
4428  *
4429  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
4430  * blocks are contiguous then releasing them at one time will only affect one
4431  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
4432  * actually use a lot of journal space.
4433  *
4434  * @this_bh will be %NULL if @first and @last point into the inode's direct
4435  * block pointers.
4436  */
4437 static void ext4_free_data(handle_t *handle, struct inode *inode,
4438                            struct buffer_head *this_bh,
4439                            __le32 *first, __le32 *last)
4440 {
4441         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
4442         unsigned long count = 0;            /* Number of blocks in the run */
4443         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
4444                                                corresponding to
4445                                                block_to_free */
4446         ext4_fsblk_t nr;                    /* Current block # */
4447         __le32 *p;                          /* Pointer into inode/ind
4448                                                for current block */
4449         int err;
4450
4451         if (this_bh) {                          /* For indirect block */
4452                 BUFFER_TRACE(this_bh, "get_write_access");
4453                 err = ext4_journal_get_write_access(handle, this_bh);
4454                 /* Important: if we can't update the indirect pointers
4455                  * to the blocks, we can't free them. */
4456                 if (err)
4457                         return;
4458         }
4459
4460         for (p = first; p < last; p++) {
4461                 nr = le32_to_cpu(*p);
4462                 if (nr) {
4463                         /* accumulate blocks to free if they're contiguous */
4464                         if (count == 0) {
4465                                 block_to_free = nr;
4466                                 block_to_free_p = p;
4467                                 count = 1;
4468                         } else if (nr == block_to_free + count) {
4469                                 count++;
4470                         } else {
4471                                 if (ext4_clear_blocks(handle, inode, this_bh,
4472                                                       block_to_free, count,
4473                                                       block_to_free_p, p))
4474                                         break;
4475                                 block_to_free = nr;
4476                                 block_to_free_p = p;
4477                                 count = 1;
4478                         }
4479                 }
4480         }
4481
4482         if (count > 0)
4483                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
4484                                   count, block_to_free_p, p);
4485
4486         if (this_bh) {
4487                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
4488
4489                 /*
4490                  * The buffer head should have an attached journal head at this
4491                  * point. However, if the data is corrupted and an indirect
4492                  * block pointed to itself, it would have been detached when
4493                  * the block was cleared. Check for this instead of OOPSing.
4494                  */
4495                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
4496                         ext4_handle_dirty_metadata(handle, inode, this_bh);
4497                 else
4498                         EXT4_ERROR_INODE(inode,
4499                                          "circular indirect block detected at "
4500                                          "block %llu",
4501                                 (unsigned long long) this_bh->b_blocknr);
4502         }
4503 }
4504
4505 /**
4506  *      ext4_free_branches - free an array of branches
4507  *      @handle: JBD handle for this transaction
4508  *      @inode: inode we are dealing with
4509  *      @parent_bh: the buffer_head which contains *@first and *@last
4510  *      @first: array of block numbers
4511  *      @last:  pointer immediately past the end of array
4512  *      @depth: depth of the branches to free
4513  *
4514  *      We are freeing all blocks refered from these branches (numbers are
4515  *      stored as little-endian 32-bit) and updating @inode->i_blocks
4516  *      appropriately.
4517  */
4518 static void ext4_free_branches(handle_t *handle, struct inode *inode,
4519                                struct buffer_head *parent_bh,
4520                                __le32 *first, __le32 *last, int depth)
4521 {
4522         ext4_fsblk_t nr;
4523         __le32 *p;
4524
4525         if (ext4_handle_is_aborted(handle))
4526                 return;
4527
4528         if (depth--) {
4529                 struct buffer_head *bh;
4530                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4531                 p = last;
4532                 while (--p >= first) {
4533                         nr = le32_to_cpu(*p);
4534                         if (!nr)
4535                                 continue;               /* A hole */
4536
4537                         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
4538                                                    nr, 1)) {
4539                                 EXT4_ERROR_INODE(inode,
4540                                                  "invalid indirect mapped "
4541                                                  "block %lu (level %d)",
4542                                                  (unsigned long) nr, depth);
4543                                 break;
4544                         }
4545
4546                         /* Go read the buffer for the next level down */
4547                         bh = sb_bread(inode->i_sb, nr);
4548
4549                         /*
4550                          * A read failure? Report error and clear slot
4551                          * (should be rare).
4552                          */
4553                         if (!bh) {
4554                                 EXT4_ERROR_INODE_BLOCK(inode, nr,
4555                                                        "Read failure");
4556                                 continue;
4557                         }
4558
4559                         /* This zaps the entire block.  Bottom up. */
4560                         BUFFER_TRACE(bh, "free child branches");
4561                         ext4_free_branches(handle, inode, bh,
4562                                         (__le32 *) bh->b_data,
4563                                         (__le32 *) bh->b_data + addr_per_block,
4564                                         depth);
4565
4566                         /*
4567                          * Everything below this this pointer has been
4568                          * released.  Now let this top-of-subtree go.
4569                          *
4570                          * We want the freeing of this indirect block to be
4571                          * atomic in the journal with the updating of the
4572                          * bitmap block which owns it.  So make some room in
4573                          * the journal.
4574                          *
4575                          * We zero the parent pointer *after* freeing its
4576                          * pointee in the bitmaps, so if extend_transaction()
4577                          * for some reason fails to put the bitmap changes and
4578                          * the release into the same transaction, recovery
4579                          * will merely complain about releasing a free block,
4580                          * rather than leaking blocks.
4581                          */
4582                         if (ext4_handle_is_aborted(handle))
4583                                 return;
4584                         if (try_to_extend_transaction(handle, inode)) {
4585                                 ext4_mark_inode_dirty(handle, inode);
4586                                 ext4_truncate_restart_trans(handle, inode,
4587                                             blocks_for_truncate(inode));
4588                         }
4589
4590                         /*
4591                          * The forget flag here is critical because if
4592                          * we are journaling (and not doing data
4593                          * journaling), we have to make sure a revoke
4594                          * record is written to prevent the journal
4595                          * replay from overwriting the (former)
4596                          * indirect block if it gets reallocated as a
4597                          * data block.  This must happen in the same
4598                          * transaction where the data blocks are
4599                          * actually freed.
4600                          */
4601                         ext4_free_blocks(handle, inode, 0, nr, 1,
4602                                          EXT4_FREE_BLOCKS_METADATA|
4603                                          EXT4_FREE_BLOCKS_FORGET);
4604
4605                         if (parent_bh) {
4606                                 /*
4607                                  * The block which we have just freed is
4608                                  * pointed to by an indirect block: journal it
4609                                  */
4610                                 BUFFER_TRACE(parent_bh, "get_write_access");
4611                                 if (!ext4_journal_get_write_access(handle,
4612                                                                    parent_bh)){
4613                                         *p = 0;
4614                                         BUFFER_TRACE(parent_bh,
4615                                         "call ext4_handle_dirty_metadata");
4616                                         ext4_handle_dirty_metadata(handle,
4617                                                                    inode,
4618                                                                    parent_bh);
4619                                 }
4620                         }
4621                 }
4622         } else {
4623                 /* We have reached the bottom of the tree. */
4624                 BUFFER_TRACE(parent_bh, "free data blocks");
4625                 ext4_free_data(handle, inode, parent_bh, first, last);
4626         }
4627 }
4628
4629 int ext4_can_truncate(struct inode *inode)
4630 {
4631         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4632                 return 0;
4633         if (S_ISREG(inode->i_mode))
4634                 return 1;
4635         if (S_ISDIR(inode->i_mode))
4636                 return 1;
4637         if (S_ISLNK(inode->i_mode))
4638                 return !ext4_inode_is_fast_symlink(inode);
4639         return 0;
4640 }
4641
4642 /*
4643  * ext4_truncate()
4644  *
4645  * We block out ext4_get_block() block instantiations across the entire
4646  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4647  * simultaneously on behalf of the same inode.
4648  *
4649  * As we work through the truncate and commmit bits of it to the journal there
4650  * is one core, guiding principle: the file's tree must always be consistent on
4651  * disk.  We must be able to restart the truncate after a crash.
4652  *
4653  * The file's tree may be transiently inconsistent in memory (although it
4654  * probably isn't), but whenever we close off and commit a journal transaction,
4655  * the contents of (the filesystem + the journal) must be consistent and
4656  * restartable.  It's pretty simple, really: bottom up, right to left (although
4657  * left-to-right works OK too).
4658  *
4659  * Note that at recovery time, journal replay occurs *before* the restart of
4660  * truncate against the orphan inode list.
4661  *
4662  * The committed inode has the new, desired i_size (which is the same as
4663  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4664  * that this inode's truncate did not complete and it will again call
4665  * ext4_truncate() to have another go.  So there will be instantiated blocks
4666  * to the right of the truncation point in a crashed ext4 filesystem.  But
4667  * that's fine - as long as they are linked from the inode, the post-crash
4668  * ext4_truncate() run will find them and release them.
4669  */
4670 void ext4_truncate(struct inode *inode)
4671 {
4672         handle_t *handle;
4673         struct ext4_inode_info *ei = EXT4_I(inode);
4674         __le32 *i_data = ei->i_data;
4675         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4676         struct address_space *mapping = inode->i_mapping;
4677         ext4_lblk_t offsets[4];
4678         Indirect chain[4];
4679         Indirect *partial;
4680         __le32 nr = 0;
4681         int n;
4682         ext4_lblk_t last_block;
4683         unsigned blocksize = inode->i_sb->s_blocksize;
4684
4685         if (!ext4_can_truncate(inode))
4686                 return;
4687
4688         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4689
4690         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4691                 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4692
4693         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
4694                 ext4_ext_truncate(inode);
4695                 return;
4696         }
4697
4698         handle = start_transaction(inode);
4699         if (IS_ERR(handle))
4700                 return;         /* AKPM: return what? */
4701
4702         last_block = (inode->i_size + blocksize-1)
4703                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4704
4705         if (inode->i_size & (blocksize - 1))
4706                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4707                         goto out_stop;
4708
4709         n = ext4_block_to_path(inode, last_block, offsets, NULL);
4710         if (n == 0)
4711                 goto out_stop;  /* error */
4712
4713         /*
4714          * OK.  This truncate is going to happen.  We add the inode to the
4715          * orphan list, so that if this truncate spans multiple transactions,
4716          * and we crash, we will resume the truncate when the filesystem
4717          * recovers.  It also marks the inode dirty, to catch the new size.
4718          *
4719          * Implication: the file must always be in a sane, consistent
4720          * truncatable state while each transaction commits.
4721          */
4722         if (ext4_orphan_add(handle, inode))
4723                 goto out_stop;
4724
4725         /*
4726          * From here we block out all ext4_get_block() callers who want to
4727          * modify the block allocation tree.
4728          */
4729         down_write(&ei->i_data_sem);
4730
4731         ext4_discard_preallocations(inode);
4732
4733         /*
4734          * The orphan list entry will now protect us from any crash which
4735          * occurs before the truncate completes, so it is now safe to propagate
4736          * the new, shorter inode size (held for now in i_size) into the
4737          * on-disk inode. We do this via i_disksize, which is the value which
4738          * ext4 *really* writes onto the disk inode.
4739          */
4740         ei->i_disksize = inode->i_size;
4741
4742         if (n == 1) {           /* direct blocks */
4743                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4744                                i_data + EXT4_NDIR_BLOCKS);
4745                 goto do_indirects;
4746         }
4747
4748         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4749         /* Kill the top of shared branch (not detached) */
4750         if (nr) {
4751                 if (partial == chain) {
4752                         /* Shared branch grows from the inode */
4753                         ext4_free_branches(handle, inode, NULL,
4754                                            &nr, &nr+1, (chain+n-1) - partial);
4755                         *partial->p = 0;
4756                         /*
4757                          * We mark the inode dirty prior to restart,
4758                          * and prior to stop.  No need for it here.
4759                          */
4760                 } else {
4761                         /* Shared branch grows from an indirect block */
4762                         BUFFER_TRACE(partial->bh, "get_write_access");
4763                         ext4_free_branches(handle, inode, partial->bh,
4764                                         partial->p,
4765                                         partial->p+1, (chain+n-1) - partial);
4766                 }
4767         }
4768         /* Clear the ends of indirect blocks on the shared branch */
4769         while (partial > chain) {
4770                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4771                                    (__le32*)partial->bh->b_data+addr_per_block,
4772                                    (chain+n-1) - partial);
4773                 BUFFER_TRACE(partial->bh, "call brelse");
4774                 brelse(partial->bh);
4775                 partial--;
4776         }
4777 do_indirects:
4778         /* Kill the remaining (whole) subtrees */
4779         switch (offsets[0]) {
4780         default:
4781                 nr = i_data[EXT4_IND_BLOCK];
4782                 if (nr) {
4783                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4784                         i_data[EXT4_IND_BLOCK] = 0;
4785                 }
4786         case EXT4_IND_BLOCK:
4787                 nr = i_data[EXT4_DIND_BLOCK];
4788                 if (nr) {
4789                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4790                         i_data[EXT4_DIND_BLOCK] = 0;
4791                 }
4792         case EXT4_DIND_BLOCK:
4793                 nr = i_data[EXT4_TIND_BLOCK];
4794                 if (nr) {
4795                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4796                         i_data[EXT4_TIND_BLOCK] = 0;
4797                 }
4798         case EXT4_TIND_BLOCK:
4799                 ;
4800         }
4801
4802         up_write(&ei->i_data_sem);
4803         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4804         ext4_mark_inode_dirty(handle, inode);
4805
4806         /*
4807          * In a multi-transaction truncate, we only make the final transaction
4808          * synchronous
4809          */
4810         if (IS_SYNC(inode))
4811                 ext4_handle_sync(handle);
4812 out_stop:
4813         /*
4814          * If this was a simple ftruncate(), and the file will remain alive
4815          * then we need to clear up the orphan record which we created above.
4816          * However, if this was a real unlink then we were called by
4817          * ext4_delete_inode(), and we allow that function to clean up the
4818          * orphan info for us.
4819          */
4820         if (inode->i_nlink)
4821                 ext4_orphan_del(handle, inode);
4822
4823         ext4_journal_stop(handle);
4824 }
4825
4826 /*
4827  * ext4_get_inode_loc returns with an extra refcount against the inode's
4828  * underlying buffer_head on success. If 'in_mem' is true, we have all
4829  * data in memory that is needed to recreate the on-disk version of this
4830  * inode.
4831  */
4832 static int __ext4_get_inode_loc(struct inode *inode,
4833                                 struct ext4_iloc *iloc, int in_mem)
4834 {
4835         struct ext4_group_desc  *gdp;
4836         struct buffer_head      *bh;
4837         struct super_block      *sb = inode->i_sb;
4838         ext4_fsblk_t            block;
4839         int                     inodes_per_block, inode_offset;
4840
4841         iloc->bh = NULL;
4842         if (!ext4_valid_inum(sb, inode->i_ino))
4843                 return -EIO;
4844
4845         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4846         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4847         if (!gdp)
4848                 return -EIO;
4849
4850         /*
4851          * Figure out the offset within the block group inode table
4852          */
4853         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4854         inode_offset = ((inode->i_ino - 1) %
4855                         EXT4_INODES_PER_GROUP(sb));
4856         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4857         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4858
4859         bh = sb_getblk(sb, block);
4860         if (!bh) {
4861                 EXT4_ERROR_INODE_BLOCK(inode, block,
4862                                        "unable to read itable block");
4863                 return -EIO;
4864         }
4865         if (!buffer_uptodate(bh)) {
4866                 lock_buffer(bh);
4867
4868                 /*
4869                  * If the buffer has the write error flag, we have failed
4870                  * to write out another inode in the same block.  In this
4871                  * case, we don't have to read the block because we may
4872                  * read the old inode data successfully.
4873                  */
4874                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4875                         set_buffer_uptodate(bh);
4876
4877                 if (buffer_uptodate(bh)) {
4878                         /* someone brought it uptodate while we waited */
4879                         unlock_buffer(bh);
4880                         goto has_buffer;
4881                 }
4882
4883                 /*
4884                  * If we have all information of the inode in memory and this
4885                  * is the only valid inode in the block, we need not read the
4886                  * block.
4887                  */
4888                 if (in_mem) {
4889                         struct buffer_head *bitmap_bh;
4890                         int i, start;
4891
4892                         start = inode_offset & ~(inodes_per_block - 1);
4893
4894                         /* Is the inode bitmap in cache? */
4895                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4896                         if (!bitmap_bh)
4897                                 goto make_io;
4898
4899                         /*
4900                          * If the inode bitmap isn't in cache then the
4901                          * optimisation may end up performing two reads instead
4902                          * of one, so skip it.
4903                          */
4904                         if (!buffer_uptodate(bitmap_bh)) {
4905                                 brelse(bitmap_bh);
4906                                 goto make_io;
4907                         }
4908                         for (i = start; i < start + inodes_per_block; i++) {
4909                                 if (i == inode_offset)
4910                                         continue;
4911                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4912                                         break;
4913                         }
4914                         brelse(bitmap_bh);
4915                         if (i == start + inodes_per_block) {
4916                                 /* all other inodes are free, so skip I/O */
4917                                 memset(bh->b_data, 0, bh->b_size);
4918                                 set_buffer_uptodate(bh);
4919                                 unlock_buffer(bh);
4920                                 goto has_buffer;
4921                         }
4922                 }
4923
4924 make_io:
4925                 /*
4926                  * If we need to do any I/O, try to pre-readahead extra
4927                  * blocks from the inode table.
4928                  */
4929                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4930                         ext4_fsblk_t b, end, table;
4931                         unsigned num;
4932
4933                         table = ext4_inode_table(sb, gdp);
4934                         /* s_inode_readahead_blks is always a power of 2 */
4935                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4936                         if (table > b)
4937                                 b = table;
4938                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4939                         num = EXT4_INODES_PER_GROUP(sb);
4940                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4941                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4942                                 num -= ext4_itable_unused_count(sb, gdp);
4943                         table += num / inodes_per_block;
4944                         if (end > table)
4945                                 end = table;
4946                         while (b <= end)
4947                                 sb_breadahead(sb, b++);
4948                 }
4949
4950                 /*
4951                  * There are other valid inodes in the buffer, this inode
4952                  * has in-inode xattrs, or we don't have this inode in memory.
4953                  * Read the block from disk.
4954                  */
4955                 get_bh(bh);
4956                 bh->b_end_io = end_buffer_read_sync;
4957                 submit_bh(READ_META, bh);
4958                 wait_on_buffer(bh);
4959                 if (!buffer_uptodate(bh)) {
4960                         EXT4_ERROR_INODE_BLOCK(inode, block,
4961                                                "unable to read itable block");
4962                         brelse(bh);
4963                         return -EIO;
4964                 }
4965         }
4966 has_buffer:
4967         iloc->bh = bh;
4968         return 0;
4969 }
4970
4971 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4972 {
4973         /* We have all inode data except xattrs in memory here. */
4974         return __ext4_get_inode_loc(inode, iloc,
4975                 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
4976 }
4977
4978 void ext4_set_inode_flags(struct inode *inode)
4979 {
4980         unsigned int flags = EXT4_I(inode)->i_flags;
4981
4982         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4983         if (flags & EXT4_SYNC_FL)
4984                 inode->i_flags |= S_SYNC;
4985         if (flags & EXT4_APPEND_FL)
4986                 inode->i_flags |= S_APPEND;
4987         if (flags & EXT4_IMMUTABLE_FL)
4988                 inode->i_flags |= S_IMMUTABLE;
4989         if (flags & EXT4_NOATIME_FL)
4990                 inode->i_flags |= S_NOATIME;
4991         if (flags & EXT4_DIRSYNC_FL)
4992                 inode->i_flags |= S_DIRSYNC;
4993 }
4994
4995 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4996 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4997 {
4998         unsigned int vfs_fl;
4999         unsigned long old_fl, new_fl;
5000
5001         do {
5002                 vfs_fl = ei->vfs_inode.i_flags;
5003                 old_fl = ei->i_flags;
5004                 new_fl = old_fl & ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
5005                                 EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|
5006                                 EXT4_DIRSYNC_FL);
5007                 if (vfs_fl & S_SYNC)
5008                         new_fl |= EXT4_SYNC_FL;
5009                 if (vfs_fl & S_APPEND)
5010                         new_fl |= EXT4_APPEND_FL;
5011                 if (vfs_fl & S_IMMUTABLE)
5012                         new_fl |= EXT4_IMMUTABLE_FL;
5013                 if (vfs_fl & S_NOATIME)
5014                         new_fl |= EXT4_NOATIME_FL;
5015                 if (vfs_fl & S_DIRSYNC)
5016                         new_fl |= EXT4_DIRSYNC_FL;
5017         } while (cmpxchg(&ei->i_flags, old_fl, new_fl) != old_fl);
5018 }
5019
5020 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
5021                                   struct ext4_inode_info *ei)
5022 {
5023         blkcnt_t i_blocks ;
5024         struct inode *inode = &(ei->vfs_inode);
5025         struct super_block *sb = inode->i_sb;
5026
5027         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
5028                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
5029                 /* we are using combined 48 bit field */
5030                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
5031                                         le32_to_cpu(raw_inode->i_blocks_lo);
5032                 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
5033                         /* i_blocks represent file system block size */
5034                         return i_blocks  << (inode->i_blkbits - 9);
5035                 } else {
5036                         return i_blocks;
5037                 }
5038         } else {
5039                 return le32_to_cpu(raw_inode->i_blocks_lo);
5040         }
5041 }
5042
5043 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
5044 {
5045         struct ext4_iloc iloc;
5046         struct ext4_inode *raw_inode;
5047         struct ext4_inode_info *ei;
5048         struct inode *inode;
5049         journal_t *journal = EXT4_SB(sb)->s_journal;
5050         long ret;
5051         int block;
5052
5053         inode = iget_locked(sb, ino);
5054         if (!inode)
5055                 return ERR_PTR(-ENOMEM);
5056         if (!(inode->i_state & I_NEW))
5057                 return inode;
5058
5059         ei = EXT4_I(inode);
5060         iloc.bh = 0;
5061
5062         ret = __ext4_get_inode_loc(inode, &iloc, 0);
5063         if (ret < 0)
5064                 goto bad_inode;
5065         raw_inode = ext4_raw_inode(&iloc);
5066         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
5067         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
5068         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
5069         if (!(test_opt(inode->i_sb, NO_UID32))) {
5070                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
5071                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
5072         }
5073         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
5074
5075         ei->i_state_flags = 0;
5076         ei->i_dir_start_lookup = 0;
5077         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
5078         /* We now have enough fields to check if the inode was active or not.
5079          * This is needed because nfsd might try to access dead inodes
5080          * the test is that same one that e2fsck uses
5081          * NeilBrown 1999oct15
5082          */
5083         if (inode->i_nlink == 0) {
5084                 if (inode->i_mode == 0 ||
5085                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
5086                         /* this inode is deleted */
5087                         ret = -ESTALE;
5088                         goto bad_inode;
5089                 }
5090                 /* The only unlinked inodes we let through here have
5091                  * valid i_mode and are being read by the orphan
5092                  * recovery code: that's fine, we're about to complete
5093                  * the process of deleting those. */
5094         }
5095         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
5096         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
5097         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
5098         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
5099                 ei->i_file_acl |=
5100                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
5101         inode->i_size = ext4_isize(raw_inode);
5102         ei->i_disksize = inode->i_size;
5103 #ifdef CONFIG_QUOTA
5104         ei->i_reserved_quota = 0;
5105 #endif
5106         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
5107         ei->i_block_group = iloc.block_group;
5108         ei->i_last_alloc_group = ~0;
5109         /*
5110          * NOTE! The in-memory inode i_data array is in little-endian order
5111          * even on big-endian machines: we do NOT byteswap the block numbers!
5112          */
5113         for (block = 0; block < EXT4_N_BLOCKS; block++)
5114                 ei->i_data[block] = raw_inode->i_block[block];
5115         INIT_LIST_HEAD(&ei->i_orphan);
5116
5117         /*
5118          * Set transaction id's of transactions that have to be committed
5119          * to finish f[data]sync. We set them to currently running transaction
5120          * as we cannot be sure that the inode or some of its metadata isn't
5121          * part of the transaction - the inode could have been reclaimed and
5122          * now it is reread from disk.
5123          */
5124         if (journal) {
5125                 transaction_t *transaction;
5126                 tid_t tid;
5127
5128                 read_lock(&journal->j_state_lock);
5129                 if (journal->j_running_transaction)
5130                         transaction = journal->j_running_transaction;
5131                 else
5132                         transaction = journal->j_committing_transaction;
5133                 if (transaction)
5134                         tid = transaction->t_tid;
5135                 else
5136                         tid = journal->j_commit_sequence;
5137                 read_unlock(&journal->j_state_lock);
5138                 ei->i_sync_tid = tid;
5139                 ei->i_datasync_tid = tid;
5140         }
5141
5142         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5143                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
5144                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
5145                     EXT4_INODE_SIZE(inode->i_sb)) {
5146                         ret = -EIO;
5147                         goto bad_inode;
5148                 }
5149                 if (ei->i_extra_isize == 0) {
5150                         /* The extra space is currently unused. Use it. */
5151                         ei->i_extra_isize = sizeof(struct ext4_inode) -
5152                                             EXT4_GOOD_OLD_INODE_SIZE;
5153                 } else {
5154                         __le32 *magic = (void *)raw_inode +
5155                                         EXT4_GOOD_OLD_INODE_SIZE +
5156                                         ei->i_extra_isize;
5157                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
5158                                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
5159                 }
5160         } else
5161                 ei->i_extra_isize = 0;
5162
5163         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
5164         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
5165         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
5166         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
5167
5168         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
5169         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5170                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5171                         inode->i_version |=
5172                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
5173         }
5174
5175         ret = 0;
5176         if (ei->i_file_acl &&
5177             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
5178                 EXT4_ERROR_INODE(inode, "bad extended attribute block %llu",
5179                                  ei->i_file_acl);
5180                 ret = -EIO;
5181                 goto bad_inode;
5182         } else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5183                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5184                     (S_ISLNK(inode->i_mode) &&
5185                      !ext4_inode_is_fast_symlink(inode)))
5186                         /* Validate extent which is part of inode */
5187                         ret = ext4_ext_check_inode(inode);
5188         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5189                    (S_ISLNK(inode->i_mode) &&
5190                     !ext4_inode_is_fast_symlink(inode))) {
5191                 /* Validate block references which are part of inode */
5192                 ret = ext4_check_inode_blockref(inode);
5193         }
5194         if (ret)
5195                 goto bad_inode;
5196
5197         if (S_ISREG(inode->i_mode)) {
5198                 inode->i_op = &ext4_file_inode_operations;
5199                 inode->i_fop = &ext4_file_operations;
5200                 ext4_set_aops(inode);
5201         } else if (S_ISDIR(inode->i_mode)) {
5202                 inode->i_op = &ext4_dir_inode_operations;
5203                 inode->i_fop = &ext4_dir_operations;
5204         } else if (S_ISLNK(inode->i_mode)) {
5205                 if (ext4_inode_is_fast_symlink(inode)) {
5206                         inode->i_op = &ext4_fast_symlink_inode_operations;
5207                         nd_terminate_link(ei->i_data, inode->i_size,
5208                                 sizeof(ei->i_data) - 1);
5209                 } else {
5210                         inode->i_op = &ext4_symlink_inode_operations;
5211                         ext4_set_aops(inode);
5212                 }
5213         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
5214               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
5215                 inode->i_op = &ext4_special_inode_operations;
5216                 if (raw_inode->i_block[0])
5217                         init_special_inode(inode, inode->i_mode,
5218                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
5219                 else
5220                         init_special_inode(inode, inode->i_mode,
5221                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
5222         } else {
5223                 ret = -EIO;
5224                 EXT4_ERROR_INODE(inode, "bogus i_mode (%o)", inode->i_mode);
5225                 goto bad_inode;
5226         }
5227         brelse(iloc.bh);
5228         ext4_set_inode_flags(inode);
5229         unlock_new_inode(inode);
5230         return inode;
5231
5232 bad_inode:
5233         brelse(iloc.bh);
5234         iget_failed(inode);
5235         return ERR_PTR(ret);
5236 }
5237
5238 static int ext4_inode_blocks_set(handle_t *handle,
5239                                 struct ext4_inode *raw_inode,
5240                                 struct ext4_inode_info *ei)
5241 {
5242         struct inode *inode = &(ei->vfs_inode);
5243         u64 i_blocks = inode->i_blocks;
5244         struct super_block *sb = inode->i_sb;
5245
5246         if (i_blocks <= ~0U) {
5247                 /*
5248                  * i_blocks can be represnted in a 32 bit variable
5249                  * as multiple of 512 bytes
5250                  */
5251                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5252                 raw_inode->i_blocks_high = 0;
5253                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5254                 return 0;
5255         }
5256         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
5257                 return -EFBIG;
5258
5259         if (i_blocks <= 0xffffffffffffULL) {
5260                 /*
5261                  * i_blocks can be represented in a 48 bit variable
5262                  * as multiple of 512 bytes
5263                  */
5264                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5265                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5266                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5267         } else {
5268                 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5269                 /* i_block is stored in file system block size */
5270                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
5271                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5272                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5273         }
5274         return 0;
5275 }
5276
5277 /*
5278  * Post the struct inode info into an on-disk inode location in the
5279  * buffer-cache.  This gobbles the caller's reference to the
5280  * buffer_head in the inode location struct.
5281  *
5282  * The caller must have write access to iloc->bh.
5283  */
5284 static int ext4_do_update_inode(handle_t *handle,
5285                                 struct inode *inode,
5286                                 struct ext4_iloc *iloc)
5287 {
5288         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5289         struct ext4_inode_info *ei = EXT4_I(inode);
5290         struct buffer_head *bh = iloc->bh;
5291         int err = 0, rc, block;
5292
5293         /* For fields not not tracking in the in-memory inode,
5294          * initialise them to zero for new inodes. */
5295         if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5296                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5297
5298         ext4_get_inode_flags(ei);
5299         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5300         if (!(test_opt(inode->i_sb, NO_UID32))) {
5301                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
5302                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
5303 /*
5304  * Fix up interoperability with old kernels. Otherwise, old inodes get
5305  * re-used with the upper 16 bits of the uid/gid intact
5306  */
5307                 if (!ei->i_dtime) {
5308                         raw_inode->i_uid_high =
5309                                 cpu_to_le16(high_16_bits(inode->i_uid));
5310                         raw_inode->i_gid_high =
5311                                 cpu_to_le16(high_16_bits(inode->i_gid));
5312                 } else {
5313                         raw_inode->i_uid_high = 0;
5314                         raw_inode->i_gid_high = 0;
5315                 }
5316         } else {
5317                 raw_inode->i_uid_low =
5318                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
5319                 raw_inode->i_gid_low =
5320                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
5321                 raw_inode->i_uid_high = 0;
5322                 raw_inode->i_gid_high = 0;
5323         }
5324         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5325
5326         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5327         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5328         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5329         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5330
5331         if (ext4_inode_blocks_set(handle, raw_inode, ei))
5332                 goto out_brelse;
5333         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5334         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
5335         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
5336             cpu_to_le32(EXT4_OS_HURD))
5337                 raw_inode->i_file_acl_high =
5338                         cpu_to_le16(ei->i_file_acl >> 32);
5339         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5340         ext4_isize_set(raw_inode, ei->i_disksize);
5341         if (ei->i_disksize > 0x7fffffffULL) {
5342                 struct super_block *sb = inode->i_sb;
5343                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
5344                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
5345                                 EXT4_SB(sb)->s_es->s_rev_level ==
5346                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
5347                         /* If this is the first large file
5348                          * created, add a flag to the superblock.
5349                          */
5350                         err = ext4_journal_get_write_access(handle,
5351                                         EXT4_SB(sb)->s_sbh);
5352                         if (err)
5353                                 goto out_brelse;
5354                         ext4_update_dynamic_rev(sb);
5355                         EXT4_SET_RO_COMPAT_FEATURE(sb,
5356                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
5357                         sb->s_dirt = 1;
5358                         ext4_handle_sync(handle);
5359                         err = ext4_handle_dirty_metadata(handle, NULL,
5360                                         EXT4_SB(sb)->s_sbh);
5361                 }
5362         }
5363         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5364         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5365                 if (old_valid_dev(inode->i_rdev)) {
5366                         raw_inode->i_block[0] =
5367                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
5368                         raw_inode->i_block[1] = 0;
5369                 } else {
5370                         raw_inode->i_block[0] = 0;
5371                         raw_inode->i_block[1] =
5372                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
5373                         raw_inode->i_block[2] = 0;
5374                 }
5375         } else
5376                 for (block = 0; block < EXT4_N_BLOCKS; block++)
5377                         raw_inode->i_block[block] = ei->i_data[block];
5378
5379         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
5380         if (ei->i_extra_isize) {
5381                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5382                         raw_inode->i_version_hi =
5383                         cpu_to_le32(inode->i_version >> 32);
5384                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
5385         }
5386
5387         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5388         rc = ext4_handle_dirty_metadata(handle, NULL, bh);
5389         if (!err)
5390                 err = rc;
5391         ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5392
5393         ext4_update_inode_fsync_trans(handle, inode, 0);
5394 out_brelse:
5395         brelse(bh);
5396         ext4_std_error(inode->i_sb, err);
5397         return err;
5398 }
5399
5400 /*
5401  * ext4_write_inode()
5402  *
5403  * We are called from a few places:
5404  *
5405  * - Within generic_file_write() for O_SYNC files.
5406  *   Here, there will be no transaction running. We wait for any running
5407  *   trasnaction to commit.
5408  *
5409  * - Within sys_sync(), kupdate and such.
5410  *   We wait on commit, if tol to.
5411  *
5412  * - Within prune_icache() (PF_MEMALLOC == true)
5413  *   Here we simply return.  We can't afford to block kswapd on the
5414  *   journal commit.
5415  *
5416  * In all cases it is actually safe for us to return without doing anything,
5417  * because the inode has been copied into a raw inode buffer in
5418  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
5419  * knfsd.
5420  *
5421  * Note that we are absolutely dependent upon all inode dirtiers doing the
5422  * right thing: they *must* call mark_inode_dirty() after dirtying info in
5423  * which we are interested.
5424  *
5425  * It would be a bug for them to not do this.  The code:
5426  *
5427  *      mark_inode_dirty(inode)
5428  *      stuff();
5429  *      inode->i_size = expr;
5430  *
5431  * is in error because a kswapd-driven write_inode() could occur while
5432  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
5433  * will no longer be on the superblock's dirty inode list.
5434  */
5435 int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5436 {
5437         int err;
5438
5439         if (current->flags & PF_MEMALLOC)
5440                 return 0;
5441
5442         if (EXT4_SB(inode->i_sb)->s_journal) {
5443                 if (ext4_journal_current_handle()) {
5444                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5445                         dump_stack();
5446                         return -EIO;
5447                 }
5448
5449                 if (wbc->sync_mode != WB_SYNC_ALL)
5450                         return 0;
5451
5452                 err = ext4_force_commit(inode->i_sb);
5453         } else {
5454                 struct ext4_iloc iloc;
5455
5456                 err = __ext4_get_inode_loc(inode, &iloc, 0);
5457                 if (err)
5458                         return err;
5459                 if (wbc->sync_mode == WB_SYNC_ALL)
5460                         sync_dirty_buffer(iloc.bh);
5461                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5462                         EXT4_ERROR_INODE_BLOCK(inode, iloc.bh->b_blocknr,
5463                                          "IO error syncing inode");
5464                         err = -EIO;
5465                 }
5466                 brelse(iloc.bh);
5467         }
5468         return err;
5469 }
5470
5471 /*
5472  * ext4_setattr()
5473  *
5474  * Called from notify_change.
5475  *
5476  * We want to trap VFS attempts to truncate the file as soon as
5477  * possible.  In particular, we want to make sure that when the VFS
5478  * shrinks i_size, we put the inode on the orphan list and modify
5479  * i_disksize immediately, so that during the subsequent flushing of
5480  * dirty pages and freeing of disk blocks, we can guarantee that any
5481  * commit will leave the blocks being flushed in an unused state on
5482  * disk.  (On recovery, the inode will get truncated and the blocks will
5483  * be freed, so we have a strong guarantee that no future commit will
5484  * leave these blocks visible to the user.)
5485  *
5486  * Another thing we have to assure is that if we are in ordered mode
5487  * and inode is still attached to the committing transaction, we must
5488  * we start writeout of all the dirty pages which are being truncated.
5489  * This way we are sure that all the data written in the previous
5490  * transaction are already on disk (truncate waits for pages under
5491  * writeback).
5492  *
5493  * Called with inode->i_mutex down.
5494  */
5495 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5496 {
5497         struct inode *inode = dentry->d_inode;
5498         int error, rc = 0;
5499         const unsigned int ia_valid = attr->ia_valid;
5500
5501         error = inode_change_ok(inode, attr);
5502         if (error)
5503                 return error;
5504
5505         if (is_quota_modification(inode, attr))
5506                 dquot_initialize(inode);
5507         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
5508                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
5509                 handle_t *handle;
5510
5511                 /* (user+group)*(old+new) structure, inode write (sb,
5512                  * inode block, ? - but truncate inode update has it) */
5513                 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
5514                                         EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
5515                 if (IS_ERR(handle)) {
5516                         error = PTR_ERR(handle);
5517                         goto err_out;
5518                 }
5519                 error = dquot_transfer(inode, attr);
5520                 if (error) {
5521                         ext4_journal_stop(handle);
5522                         return error;
5523                 }
5524                 /* Update corresponding info in inode so that everything is in
5525                  * one transaction */
5526                 if (attr->ia_valid & ATTR_UID)
5527                         inode->i_uid = attr->ia_uid;
5528                 if (attr->ia_valid & ATTR_GID)
5529                         inode->i_gid = attr->ia_gid;
5530                 error = ext4_mark_inode_dirty(handle, inode);
5531                 ext4_journal_stop(handle);
5532         }
5533
5534         if (attr->ia_valid & ATTR_SIZE) {
5535            &nbs