fs: kill block_prepare_write
[linux-2.6.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/jbd2.h>
29 #include <linux/highuid.h>
30 #include <linux/pagemap.h>
31 #include <linux/quotaops.h>
32 #include <linux/string.h>
33 #include <linux/buffer_head.h>
34 #include <linux/writeback.h>
35 #include <linux/pagevec.h>
36 #include <linux/mpage.h>
37 #include <linux/namei.h>
38 #include <linux/uio.h>
39 #include <linux/bio.h>
40 #include <linux/workqueue.h>
41 #include <linux/kernel.h>
42 #include <linux/slab.h>
43
44 #include "ext4_jbd2.h"
45 #include "xattr.h"
46 #include "acl.h"
47 #include "ext4_extents.h"
48
49 #include <trace/events/ext4.h>
50
51 #define MPAGE_DA_EXTENT_TAIL 0x01
52
53 static inline int ext4_begin_ordered_truncate(struct inode *inode,
54                                               loff_t new_size)
55 {
56         return jbd2_journal_begin_ordered_truncate(
57                                         EXT4_SB(inode->i_sb)->s_journal,
58                                         &EXT4_I(inode)->jinode,
59                                         new_size);
60 }
61
62 static void ext4_invalidatepage(struct page *page, unsigned long offset);
63
64 /*
65  * Test whether an inode is a fast symlink.
66  */
67 static int ext4_inode_is_fast_symlink(struct inode *inode)
68 {
69         int ea_blocks = EXT4_I(inode)->i_file_acl ?
70                 (inode->i_sb->s_blocksize >> 9) : 0;
71
72         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
73 }
74
75 /*
76  * Work out how many blocks we need to proceed with the next chunk of a
77  * truncate transaction.
78  */
79 static unsigned long blocks_for_truncate(struct inode *inode)
80 {
81         ext4_lblk_t needed;
82
83         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
84
85         /* Give ourselves just enough room to cope with inodes in which
86          * i_blocks is corrupt: we've seen disk corruptions in the past
87          * which resulted in random data in an inode which looked enough
88          * like a regular file for ext4 to try to delete it.  Things
89          * will go a bit crazy if that happens, but at least we should
90          * try not to panic the whole kernel. */
91         if (needed < 2)
92                 needed = 2;
93
94         /* But we need to bound the transaction so we don't overflow the
95          * journal. */
96         if (needed > EXT4_MAX_TRANS_DATA)
97                 needed = EXT4_MAX_TRANS_DATA;
98
99         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
100 }
101
102 /*
103  * Truncate transactions can be complex and absolutely huge.  So we need to
104  * be able to restart the transaction at a conventient checkpoint to make
105  * sure we don't overflow the journal.
106  *
107  * start_transaction gets us a new handle for a truncate transaction,
108  * and extend_transaction tries to extend the existing one a bit.  If
109  * extend fails, we need to propagate the failure up and restart the
110  * transaction in the top-level truncate loop. --sct
111  */
112 static handle_t *start_transaction(struct inode *inode)
113 {
114         handle_t *result;
115
116         result = ext4_journal_start(inode, blocks_for_truncate(inode));
117         if (!IS_ERR(result))
118                 return result;
119
120         ext4_std_error(inode->i_sb, PTR_ERR(result));
121         return result;
122 }
123
124 /*
125  * Try to extend this transaction for the purposes of truncation.
126  *
127  * Returns 0 if we managed to create more room.  If we can't create more
128  * room, and the transaction must be restarted we return 1.
129  */
130 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
131 {
132         if (!ext4_handle_valid(handle))
133                 return 0;
134         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
135                 return 0;
136         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
137                 return 0;
138         return 1;
139 }
140
141 /*
142  * Restart the transaction associated with *handle.  This does a commit,
143  * so before we call here everything must be consistently dirtied against
144  * this transaction.
145  */
146 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
147                                  int nblocks)
148 {
149         int ret;
150
151         /*
152          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
153          * moment, get_block can be called only for blocks inside i_size since
154          * page cache has been already dropped and writes are blocked by
155          * i_mutex. So we can safely drop the i_data_sem here.
156          */
157         BUG_ON(EXT4_JOURNAL(inode) == NULL);
158         jbd_debug(2, "restarting handle %p\n", handle);
159         up_write(&EXT4_I(inode)->i_data_sem);
160         ret = ext4_journal_restart(handle, blocks_for_truncate(inode));
161         down_write(&EXT4_I(inode)->i_data_sem);
162         ext4_discard_preallocations(inode);
163
164         return ret;
165 }
166
167 /*
168  * Called at the last iput() if i_nlink is zero.
169  */
170 void ext4_evict_inode(struct inode *inode)
171 {
172         handle_t *handle;
173         int err;
174
175         if (inode->i_nlink) {
176                 truncate_inode_pages(&inode->i_data, 0);
177                 goto no_delete;
178         }
179
180         if (!is_bad_inode(inode))
181                 dquot_initialize(inode);
182
183         if (ext4_should_order_data(inode))
184                 ext4_begin_ordered_truncate(inode, 0);
185         truncate_inode_pages(&inode->i_data, 0);
186
187         if (is_bad_inode(inode))
188                 goto no_delete;
189
190         handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
191         if (IS_ERR(handle)) {
192                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
193                 /*
194                  * If we're going to skip the normal cleanup, we still need to
195                  * make sure that the in-core orphan linked list is properly
196                  * cleaned up.
197                  */
198                 ext4_orphan_del(NULL, inode);
199                 goto no_delete;
200         }
201
202         if (IS_SYNC(inode))
203                 ext4_handle_sync(handle);
204         inode->i_size = 0;
205         err = ext4_mark_inode_dirty(handle, inode);
206         if (err) {
207                 ext4_warning(inode->i_sb,
208                              "couldn't mark inode dirty (err %d)", err);
209                 goto stop_handle;
210         }
211         if (inode->i_blocks)
212                 ext4_truncate(inode);
213
214         /*
215          * ext4_ext_truncate() doesn't reserve any slop when it
216          * restarts journal transactions; therefore there may not be
217          * enough credits left in the handle to remove the inode from
218          * the orphan list and set the dtime field.
219          */
220         if (!ext4_handle_has_enough_credits(handle, 3)) {
221                 err = ext4_journal_extend(handle, 3);
222                 if (err > 0)
223                         err = ext4_journal_restart(handle, 3);
224                 if (err != 0) {
225                         ext4_warning(inode->i_sb,
226                                      "couldn't extend journal (err %d)", err);
227                 stop_handle:
228                         ext4_journal_stop(handle);
229                         ext4_orphan_del(NULL, inode);
230                         goto no_delete;
231                 }
232         }
233
234         /*
235          * Kill off the orphan record which ext4_truncate created.
236          * AKPM: I think this can be inside the above `if'.
237          * Note that ext4_orphan_del() has to be able to cope with the
238          * deletion of a non-existent orphan - this is because we don't
239          * know if ext4_truncate() actually created an orphan record.
240          * (Well, we could do this if we need to, but heck - it works)
241          */
242         ext4_orphan_del(handle, inode);
243         EXT4_I(inode)->i_dtime  = get_seconds();
244
245         /*
246          * One subtle ordering requirement: if anything has gone wrong
247          * (transaction abort, IO errors, whatever), then we can still
248          * do these next steps (the fs will already have been marked as
249          * having errors), but we can't free the inode if the mark_dirty
250          * fails.
251          */
252         if (ext4_mark_inode_dirty(handle, inode))
253                 /* If that failed, just do the required in-core inode clear. */
254                 ext4_clear_inode(inode);
255         else
256                 ext4_free_inode(handle, inode);
257         ext4_journal_stop(handle);
258         return;
259 no_delete:
260         ext4_clear_inode(inode);        /* We must guarantee clearing of inode... */
261 }
262
263 typedef struct {
264         __le32  *p;
265         __le32  key;
266         struct buffer_head *bh;
267 } Indirect;
268
269 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
270 {
271         p->key = *(p->p = v);
272         p->bh = bh;
273 }
274
275 /**
276  *      ext4_block_to_path - parse the block number into array of offsets
277  *      @inode: inode in question (we are only interested in its superblock)
278  *      @i_block: block number to be parsed
279  *      @offsets: array to store the offsets in
280  *      @boundary: set this non-zero if the referred-to block is likely to be
281  *             followed (on disk) by an indirect block.
282  *
283  *      To store the locations of file's data ext4 uses a data structure common
284  *      for UNIX filesystems - tree of pointers anchored in the inode, with
285  *      data blocks at leaves and indirect blocks in intermediate nodes.
286  *      This function translates the block number into path in that tree -
287  *      return value is the path length and @offsets[n] is the offset of
288  *      pointer to (n+1)th node in the nth one. If @block is out of range
289  *      (negative or too large) warning is printed and zero returned.
290  *
291  *      Note: function doesn't find node addresses, so no IO is needed. All
292  *      we need to know is the capacity of indirect blocks (taken from the
293  *      inode->i_sb).
294  */
295
296 /*
297  * Portability note: the last comparison (check that we fit into triple
298  * indirect block) is spelled differently, because otherwise on an
299  * architecture with 32-bit longs and 8Kb pages we might get into trouble
300  * if our filesystem had 8Kb blocks. We might use long long, but that would
301  * kill us on x86. Oh, well, at least the sign propagation does not matter -
302  * i_block would have to be negative in the very beginning, so we would not
303  * get there at all.
304  */
305
306 static int ext4_block_to_path(struct inode *inode,
307                               ext4_lblk_t i_block,
308                               ext4_lblk_t offsets[4], int *boundary)
309 {
310         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
311         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
312         const long direct_blocks = EXT4_NDIR_BLOCKS,
313                 indirect_blocks = ptrs,
314                 double_blocks = (1 << (ptrs_bits * 2));
315         int n = 0;
316         int final = 0;
317
318         if (i_block < direct_blocks) {
319                 offsets[n++] = i_block;
320                 final = direct_blocks;
321         } else if ((i_block -= direct_blocks) < indirect_blocks) {
322                 offsets[n++] = EXT4_IND_BLOCK;
323                 offsets[n++] = i_block;
324                 final = ptrs;
325         } else if ((i_block -= indirect_blocks) < double_blocks) {
326                 offsets[n++] = EXT4_DIND_BLOCK;
327                 offsets[n++] = i_block >> ptrs_bits;
328                 offsets[n++] = i_block & (ptrs - 1);
329                 final = ptrs;
330         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
331                 offsets[n++] = EXT4_TIND_BLOCK;
332                 offsets[n++] = i_block >> (ptrs_bits * 2);
333                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
334                 offsets[n++] = i_block & (ptrs - 1);
335                 final = ptrs;
336         } else {
337                 ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
338                              i_block + direct_blocks +
339                              indirect_blocks + double_blocks, inode->i_ino);
340         }
341         if (boundary)
342                 *boundary = final - 1 - (i_block & (ptrs - 1));
343         return n;
344 }
345
346 static int __ext4_check_blockref(const char *function, unsigned int line,
347                                  struct inode *inode,
348                                  __le32 *p, unsigned int max)
349 {
350         struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
351         __le32 *bref = p;
352         unsigned int blk;
353
354         while (bref < p+max) {
355                 blk = le32_to_cpu(*bref++);
356                 if (blk &&
357                     unlikely(!ext4_data_block_valid(EXT4_SB(inode->i_sb),
358                                                     blk, 1))) {
359                         es->s_last_error_block = cpu_to_le64(blk);
360                         ext4_error_inode(inode, function, line, blk,
361                                          "invalid block");
362                         return -EIO;
363                 }
364         }
365         return 0;
366 }
367
368
369 #define ext4_check_indirect_blockref(inode, bh)                         \
370         __ext4_check_blockref(__func__, __LINE__, inode,                \
371                               (__le32 *)(bh)->b_data,                   \
372                               EXT4_ADDR_PER_BLOCK((inode)->i_sb))
373
374 #define ext4_check_inode_blockref(inode)                                \
375         __ext4_check_blockref(__func__, __LINE__, inode,                \
376                               EXT4_I(inode)->i_data,                    \
377                               EXT4_NDIR_BLOCKS)
378
379 /**
380  *      ext4_get_branch - read the chain of indirect blocks leading to data
381  *      @inode: inode in question
382  *      @depth: depth of the chain (1 - direct pointer, etc.)
383  *      @offsets: offsets of pointers in inode/indirect blocks
384  *      @chain: place to store the result
385  *      @err: here we store the error value
386  *
387  *      Function fills the array of triples <key, p, bh> and returns %NULL
388  *      if everything went OK or the pointer to the last filled triple
389  *      (incomplete one) otherwise. Upon the return chain[i].key contains
390  *      the number of (i+1)-th block in the chain (as it is stored in memory,
391  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
392  *      number (it points into struct inode for i==0 and into the bh->b_data
393  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
394  *      block for i>0 and NULL for i==0. In other words, it holds the block
395  *      numbers of the chain, addresses they were taken from (and where we can
396  *      verify that chain did not change) and buffer_heads hosting these
397  *      numbers.
398  *
399  *      Function stops when it stumbles upon zero pointer (absent block)
400  *              (pointer to last triple returned, *@err == 0)
401  *      or when it gets an IO error reading an indirect block
402  *              (ditto, *@err == -EIO)
403  *      or when it reads all @depth-1 indirect blocks successfully and finds
404  *      the whole chain, all way to the data (returns %NULL, *err == 0).
405  *
406  *      Need to be called with
407  *      down_read(&EXT4_I(inode)->i_data_sem)
408  */
409 static Indirect *ext4_get_branch(struct inode *inode, int depth,
410                                  ext4_lblk_t  *offsets,
411                                  Indirect chain[4], int *err)
412 {
413         struct super_block *sb = inode->i_sb;
414         Indirect *p = chain;
415         struct buffer_head *bh;
416
417         *err = 0;
418         /* i_data is not going away, no lock needed */
419         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
420         if (!p->key)
421                 goto no_block;
422         while (--depth) {
423                 bh = sb_getblk(sb, le32_to_cpu(p->key));
424                 if (unlikely(!bh))
425                         goto failure;
426
427                 if (!bh_uptodate_or_lock(bh)) {
428                         if (bh_submit_read(bh) < 0) {
429                                 put_bh(bh);
430                                 goto failure;
431                         }
432                         /* validate block references */
433                         if (ext4_check_indirect_blockref(inode, bh)) {
434                                 put_bh(bh);
435                                 goto failure;
436                         }
437                 }
438
439                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
440                 /* Reader: end */
441                 if (!p->key)
442                         goto no_block;
443         }
444         return NULL;
445
446 failure:
447         *err = -EIO;
448 no_block:
449         return p;
450 }
451
452 /**
453  *      ext4_find_near - find a place for allocation with sufficient locality
454  *      @inode: owner
455  *      @ind: descriptor of indirect block.
456  *
457  *      This function returns the preferred place for block allocation.
458  *      It is used when heuristic for sequential allocation fails.
459  *      Rules are:
460  *        + if there is a block to the left of our position - allocate near it.
461  *        + if pointer will live in indirect block - allocate near that block.
462  *        + if pointer will live in inode - allocate in the same
463  *          cylinder group.
464  *
465  * In the latter case we colour the starting block by the callers PID to
466  * prevent it from clashing with concurrent allocations for a different inode
467  * in the same block group.   The PID is used here so that functionally related
468  * files will be close-by on-disk.
469  *
470  *      Caller must make sure that @ind is valid and will stay that way.
471  */
472 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
473 {
474         struct ext4_inode_info *ei = EXT4_I(inode);
475         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
476         __le32 *p;
477         ext4_fsblk_t bg_start;
478         ext4_fsblk_t last_block;
479         ext4_grpblk_t colour;
480         ext4_group_t block_group;
481         int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
482
483         /* Try to find previous block */
484         for (p = ind->p - 1; p >= start; p--) {
485                 if (*p)
486                         return le32_to_cpu(*p);
487         }
488
489         /* No such thing, so let's try location of indirect block */
490         if (ind->bh)
491                 return ind->bh->b_blocknr;
492
493         /*
494          * It is going to be referred to from the inode itself? OK, just put it
495          * into the same cylinder group then.
496          */
497         block_group = ei->i_block_group;
498         if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
499                 block_group &= ~(flex_size-1);
500                 if (S_ISREG(inode->i_mode))
501                         block_group++;
502         }
503         bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
504         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
505
506         /*
507          * If we are doing delayed allocation, we don't need take
508          * colour into account.
509          */
510         if (test_opt(inode->i_sb, DELALLOC))
511                 return bg_start;
512
513         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
514                 colour = (current->pid % 16) *
515                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
516         else
517                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
518         return bg_start + colour;
519 }
520
521 /**
522  *      ext4_find_goal - find a preferred place for allocation.
523  *      @inode: owner
524  *      @block:  block we want
525  *      @partial: pointer to the last triple within a chain
526  *
527  *      Normally this function find the preferred place for block allocation,
528  *      returns it.
529  *      Because this is only used for non-extent files, we limit the block nr
530  *      to 32 bits.
531  */
532 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
533                                    Indirect *partial)
534 {
535         ext4_fsblk_t goal;
536
537         /*
538          * XXX need to get goal block from mballoc's data structures
539          */
540
541         goal = ext4_find_near(inode, partial);
542         goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
543         return goal;
544 }
545
546 /**
547  *      ext4_blks_to_allocate: Look up the block map and count the number
548  *      of direct blocks need to be allocated for the given branch.
549  *
550  *      @branch: chain of indirect blocks
551  *      @k: number of blocks need for indirect blocks
552  *      @blks: number of data blocks to be mapped.
553  *      @blocks_to_boundary:  the offset in the indirect block
554  *
555  *      return the total number of blocks to be allocate, including the
556  *      direct and indirect blocks.
557  */
558 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
559                                  int blocks_to_boundary)
560 {
561         unsigned int count = 0;
562
563         /*
564          * Simple case, [t,d]Indirect block(s) has not allocated yet
565          * then it's clear blocks on that path have not allocated
566          */
567         if (k > 0) {
568                 /* right now we don't handle cross boundary allocation */
569                 if (blks < blocks_to_boundary + 1)
570                         count += blks;
571                 else
572                         count += blocks_to_boundary + 1;
573                 return count;
574         }
575
576         count++;
577         while (count < blks && count <= blocks_to_boundary &&
578                 le32_to_cpu(*(branch[0].p + count)) == 0) {
579                 count++;
580         }
581         return count;
582 }
583
584 /**
585  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
586  *      @indirect_blks: the number of blocks need to allocate for indirect
587  *                      blocks
588  *
589  *      @new_blocks: on return it will store the new block numbers for
590  *      the indirect blocks(if needed) and the first direct block,
591  *      @blks:  on return it will store the total number of allocated
592  *              direct blocks
593  */
594 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
595                              ext4_lblk_t iblock, ext4_fsblk_t goal,
596                              int indirect_blks, int blks,
597                              ext4_fsblk_t new_blocks[4], int *err)
598 {
599         struct ext4_allocation_request ar;
600         int target, i;
601         unsigned long count = 0, blk_allocated = 0;
602         int index = 0;
603         ext4_fsblk_t current_block = 0;
604         int ret = 0;
605
606         /*
607          * Here we try to allocate the requested multiple blocks at once,
608          * on a best-effort basis.
609          * To build a branch, we should allocate blocks for
610          * the indirect blocks(if not allocated yet), and at least
611          * the first direct block of this branch.  That's the
612          * minimum number of blocks need to allocate(required)
613          */
614         /* first we try to allocate the indirect blocks */
615         target = indirect_blks;
616         while (target > 0) {
617                 count = target;
618                 /* allocating blocks for indirect blocks and direct blocks */
619                 current_block = ext4_new_meta_blocks(handle, inode,
620                                                         goal, &count, err);
621                 if (*err)
622                         goto failed_out;
623
624                 if (unlikely(current_block + count > EXT4_MAX_BLOCK_FILE_PHYS)) {
625                         EXT4_ERROR_INODE(inode,
626                                          "current_block %llu + count %lu > %d!",
627                                          current_block, count,
628                                          EXT4_MAX_BLOCK_FILE_PHYS);
629                         *err = -EIO;
630                         goto failed_out;
631                 }
632
633                 target -= count;
634                 /* allocate blocks for indirect blocks */
635                 while (index < indirect_blks && count) {
636                         new_blocks[index++] = current_block++;
637                         count--;
638                 }
639                 if (count > 0) {
640                         /*
641                          * save the new block number
642                          * for the first direct block
643                          */
644                         new_blocks[index] = current_block;
645                         printk(KERN_INFO "%s returned more blocks than "
646                                                 "requested\n", __func__);
647                         WARN_ON(1);
648                         break;
649                 }
650         }
651
652         target = blks - count ;
653         blk_allocated = count;
654         if (!target)
655                 goto allocated;
656         /* Now allocate data blocks */
657         memset(&ar, 0, sizeof(ar));
658         ar.inode = inode;
659         ar.goal = goal;
660         ar.len = target;
661         ar.logical = iblock;
662         if (S_ISREG(inode->i_mode))
663                 /* enable in-core preallocation only for regular files */
664                 ar.flags = EXT4_MB_HINT_DATA;
665
666         current_block = ext4_mb_new_blocks(handle, &ar, err);
667         if (unlikely(current_block + ar.len > EXT4_MAX_BLOCK_FILE_PHYS)) {
668                 EXT4_ERROR_INODE(inode,
669                                  "current_block %llu + ar.len %d > %d!",
670                                  current_block, ar.len,
671                                  EXT4_MAX_BLOCK_FILE_PHYS);
672                 *err = -EIO;
673                 goto failed_out;
674         }
675
676         if (*err && (target == blks)) {
677                 /*
678                  * if the allocation failed and we didn't allocate
679                  * any blocks before
680                  */
681                 goto failed_out;
682         }
683         if (!*err) {
684                 if (target == blks) {
685                         /*
686                          * save the new block number
687                          * for the first direct block
688                          */
689                         new_blocks[index] = current_block;
690                 }
691                 blk_allocated += ar.len;
692         }
693 allocated:
694         /* total number of blocks allocated for direct blocks */
695         ret = blk_allocated;
696         *err = 0;
697         return ret;
698 failed_out:
699         for (i = 0; i < index; i++)
700                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
701         return ret;
702 }
703
704 /**
705  *      ext4_alloc_branch - allocate and set up a chain of blocks.
706  *      @inode: owner
707  *      @indirect_blks: number of allocated indirect blocks
708  *      @blks: number of allocated direct blocks
709  *      @offsets: offsets (in the blocks) to store the pointers to next.
710  *      @branch: place to store the chain in.
711  *
712  *      This function allocates blocks, zeroes out all but the last one,
713  *      links them into chain and (if we are synchronous) writes them to disk.
714  *      In other words, it prepares a branch that can be spliced onto the
715  *      inode. It stores the information about that chain in the branch[], in
716  *      the same format as ext4_get_branch() would do. We are calling it after
717  *      we had read the existing part of chain and partial points to the last
718  *      triple of that (one with zero ->key). Upon the exit we have the same
719  *      picture as after the successful ext4_get_block(), except that in one
720  *      place chain is disconnected - *branch->p is still zero (we did not
721  *      set the last link), but branch->key contains the number that should
722  *      be placed into *branch->p to fill that gap.
723  *
724  *      If allocation fails we free all blocks we've allocated (and forget
725  *      their buffer_heads) and return the error value the from failed
726  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
727  *      as described above and return 0.
728  */
729 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
730                              ext4_lblk_t iblock, int indirect_blks,
731                              int *blks, ext4_fsblk_t goal,
732                              ext4_lblk_t *offsets, Indirect *branch)
733 {
734         int blocksize = inode->i_sb->s_blocksize;
735         int i, n = 0;
736         int err = 0;
737         struct buffer_head *bh;
738         int num;
739         ext4_fsblk_t new_blocks[4];
740         ext4_fsblk_t current_block;
741
742         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
743                                 *blks, new_blocks, &err);
744         if (err)
745                 return err;
746
747         branch[0].key = cpu_to_le32(new_blocks[0]);
748         /*
749          * metadata blocks and data blocks are allocated.
750          */
751         for (n = 1; n <= indirect_blks;  n++) {
752                 /*
753                  * Get buffer_head for parent block, zero it out
754                  * and set the pointer to new one, then send
755                  * parent to disk.
756                  */
757                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
758                 branch[n].bh = bh;
759                 lock_buffer(bh);
760                 BUFFER_TRACE(bh, "call get_create_access");
761                 err = ext4_journal_get_create_access(handle, bh);
762                 if (err) {
763                         /* Don't brelse(bh) here; it's done in
764                          * ext4_journal_forget() below */
765                         unlock_buffer(bh);
766                         goto failed;
767                 }
768
769                 memset(bh->b_data, 0, blocksize);
770                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
771                 branch[n].key = cpu_to_le32(new_blocks[n]);
772                 *branch[n].p = branch[n].key;
773                 if (n == indirect_blks) {
774                         current_block = new_blocks[n];
775                         /*
776                          * End of chain, update the last new metablock of
777                          * the chain to point to the new allocated
778                          * data blocks numbers
779                          */
780                         for (i = 1; i < num; i++)
781                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
782                 }
783                 BUFFER_TRACE(bh, "marking uptodate");
784                 set_buffer_uptodate(bh);
785                 unlock_buffer(bh);
786
787                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
788                 err = ext4_handle_dirty_metadata(handle, inode, bh);
789                 if (err)
790                         goto failed;
791         }
792         *blks = num;
793         return err;
794 failed:
795         /* Allocation failed, free what we already allocated */
796         ext4_free_blocks(handle, inode, 0, new_blocks[0], 1, 0);
797         for (i = 1; i <= n ; i++) {
798                 /*
799                  * branch[i].bh is newly allocated, so there is no
800                  * need to revoke the block, which is why we don't
801                  * need to set EXT4_FREE_BLOCKS_METADATA.
802                  */
803                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1,
804                                  EXT4_FREE_BLOCKS_FORGET);
805         }
806         for (i = n+1; i < indirect_blks; i++)
807                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
808
809         ext4_free_blocks(handle, inode, 0, new_blocks[i], num, 0);
810
811         return err;
812 }
813
814 /**
815  * ext4_splice_branch - splice the allocated branch onto inode.
816  * @inode: owner
817  * @block: (logical) number of block we are adding
818  * @chain: chain of indirect blocks (with a missing link - see
819  *      ext4_alloc_branch)
820  * @where: location of missing link
821  * @num:   number of indirect blocks we are adding
822  * @blks:  number of direct blocks we are adding
823  *
824  * This function fills the missing link and does all housekeeping needed in
825  * inode (->i_blocks, etc.). In case of success we end up with the full
826  * chain to new block and return 0.
827  */
828 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
829                               ext4_lblk_t block, Indirect *where, int num,
830                               int blks)
831 {
832         int i;
833         int err = 0;
834         ext4_fsblk_t current_block;
835
836         /*
837          * If we're splicing into a [td]indirect block (as opposed to the
838          * inode) then we need to get write access to the [td]indirect block
839          * before the splice.
840          */
841         if (where->bh) {
842                 BUFFER_TRACE(where->bh, "get_write_access");
843                 err = ext4_journal_get_write_access(handle, where->bh);
844                 if (err)
845                         goto err_out;
846         }
847         /* That's it */
848
849         *where->p = where->key;
850
851         /*
852          * Update the host buffer_head or inode to point to more just allocated
853          * direct blocks blocks
854          */
855         if (num == 0 && blks > 1) {
856                 current_block = le32_to_cpu(where->key) + 1;
857                 for (i = 1; i < blks; i++)
858                         *(where->p + i) = cpu_to_le32(current_block++);
859         }
860
861         /* We are done with atomic stuff, now do the rest of housekeeping */
862         /* had we spliced it onto indirect block? */
863         if (where->bh) {
864                 /*
865                  * If we spliced it onto an indirect block, we haven't
866                  * altered the inode.  Note however that if it is being spliced
867                  * onto an indirect block at the very end of the file (the
868                  * file is growing) then we *will* alter the inode to reflect
869                  * the new i_size.  But that is not done here - it is done in
870                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
871                  */
872                 jbd_debug(5, "splicing indirect only\n");
873                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
874                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
875                 if (err)
876                         goto err_out;
877         } else {
878                 /*
879                  * OK, we spliced it into the inode itself on a direct block.
880                  */
881                 ext4_mark_inode_dirty(handle, inode);
882                 jbd_debug(5, "splicing direct\n");
883         }
884         return err;
885
886 err_out:
887         for (i = 1; i <= num; i++) {
888                 /*
889                  * branch[i].bh is newly allocated, so there is no
890                  * need to revoke the block, which is why we don't
891                  * need to set EXT4_FREE_BLOCKS_METADATA.
892                  */
893                 ext4_free_blocks(handle, inode, where[i].bh, 0, 1,
894                                  EXT4_FREE_BLOCKS_FORGET);
895         }
896         ext4_free_blocks(handle, inode, 0, le32_to_cpu(where[num].key),
897                          blks, 0);
898
899         return err;
900 }
901
902 /*
903  * The ext4_ind_map_blocks() function handles non-extents inodes
904  * (i.e., using the traditional indirect/double-indirect i_blocks
905  * scheme) for ext4_map_blocks().
906  *
907  * Allocation strategy is simple: if we have to allocate something, we will
908  * have to go the whole way to leaf. So let's do it before attaching anything
909  * to tree, set linkage between the newborn blocks, write them if sync is
910  * required, recheck the path, free and repeat if check fails, otherwise
911  * set the last missing link (that will protect us from any truncate-generated
912  * removals - all blocks on the path are immune now) and possibly force the
913  * write on the parent block.
914  * That has a nice additional property: no special recovery from the failed
915  * allocations is needed - we simply release blocks and do not touch anything
916  * reachable from inode.
917  *
918  * `handle' can be NULL if create == 0.
919  *
920  * return > 0, # of blocks mapped or allocated.
921  * return = 0, if plain lookup failed.
922  * return < 0, error case.
923  *
924  * The ext4_ind_get_blocks() function should be called with
925  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
926  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
927  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
928  * blocks.
929  */
930 static int ext4_ind_map_blocks(handle_t *handle, struct inode *inode,
931                                struct ext4_map_blocks *map,
932                                int flags)
933 {
934         int err = -EIO;
935         ext4_lblk_t offsets[4];
936         Indirect chain[4];
937         Indirect *partial;
938         ext4_fsblk_t goal;
939         int indirect_blks;
940         int blocks_to_boundary = 0;
941         int depth;
942         int count = 0;
943         ext4_fsblk_t first_block = 0;
944
945         J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
946         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
947         depth = ext4_block_to_path(inode, map->m_lblk, offsets,
948                                    &blocks_to_boundary);
949
950         if (depth == 0)
951                 goto out;
952
953         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
954
955         /* Simplest case - block found, no allocation needed */
956         if (!partial) {
957                 first_block = le32_to_cpu(chain[depth - 1].key);
958                 count++;
959                 /*map more blocks*/
960                 while (count < map->m_len && count <= blocks_to_boundary) {
961                         ext4_fsblk_t blk;
962
963                         blk = le32_to_cpu(*(chain[depth-1].p + count));
964
965                         if (blk == first_block + count)
966                                 count++;
967                         else
968                                 break;
969                 }
970                 goto got_it;
971         }
972
973         /* Next simple case - plain lookup or failed read of indirect block */
974         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
975                 goto cleanup;
976
977         /*
978          * Okay, we need to do block allocation.
979         */
980         goal = ext4_find_goal(inode, map->m_lblk, partial);
981
982         /* the number of blocks need to allocate for [d,t]indirect blocks */
983         indirect_blks = (chain + depth) - partial - 1;
984
985         /*
986          * Next look up the indirect map to count the totoal number of
987          * direct blocks to allocate for this branch.
988          */
989         count = ext4_blks_to_allocate(partial, indirect_blks,
990                                       map->m_len, blocks_to_boundary);
991         /*
992          * Block out ext4_truncate while we alter the tree
993          */
994         err = ext4_alloc_branch(handle, inode, map->m_lblk, indirect_blks,
995                                 &count, goal,
996                                 offsets + (partial - chain), partial);
997
998         /*
999          * The ext4_splice_branch call will free and forget any buffers
1000          * on the new chain if there is a failure, but that risks using
1001          * up transaction credits, especially for bitmaps where the
1002          * credits cannot be returned.  Can we handle this somehow?  We
1003          * may need to return -EAGAIN upwards in the worst case.  --sct
1004          */
1005         if (!err)
1006                 err = ext4_splice_branch(handle, inode, map->m_lblk,
1007                                          partial, indirect_blks, count);
1008         if (err)
1009                 goto cleanup;
1010
1011         map->m_flags |= EXT4_MAP_NEW;
1012
1013         ext4_update_inode_fsync_trans(handle, inode, 1);
1014 got_it:
1015         map->m_flags |= EXT4_MAP_MAPPED;
1016         map->m_pblk = le32_to_cpu(chain[depth-1].key);
1017         map->m_len = count;
1018         if (count > blocks_to_boundary)
1019                 map->m_flags |= EXT4_MAP_BOUNDARY;
1020         err = count;
1021         /* Clean up and exit */
1022         partial = chain + depth - 1;    /* the whole chain */
1023 cleanup:
1024         while (partial > chain) {
1025                 BUFFER_TRACE(partial->bh, "call brelse");
1026                 brelse(partial->bh);
1027                 partial--;
1028         }
1029 out:
1030         return err;
1031 }
1032
1033 #ifdef CONFIG_QUOTA
1034 qsize_t *ext4_get_reserved_space(struct inode *inode)
1035 {
1036         return &EXT4_I(inode)->i_reserved_quota;
1037 }
1038 #endif
1039
1040 /*
1041  * Calculate the number of metadata blocks need to reserve
1042  * to allocate a new block at @lblocks for non extent file based file
1043  */
1044 static int ext4_indirect_calc_metadata_amount(struct inode *inode,
1045                                               sector_t lblock)
1046 {
1047         struct ext4_inode_info *ei = EXT4_I(inode);
1048         sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
1049         int blk_bits;
1050
1051         if (lblock < EXT4_NDIR_BLOCKS)
1052                 return 0;
1053
1054         lblock -= EXT4_NDIR_BLOCKS;
1055
1056         if (ei->i_da_metadata_calc_len &&
1057             (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
1058                 ei->i_da_metadata_calc_len++;
1059                 return 0;
1060         }
1061         ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
1062         ei->i_da_metadata_calc_len = 1;
1063         blk_bits = order_base_2(lblock);
1064         return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
1065 }
1066
1067 /*
1068  * Calculate the number of metadata blocks need to reserve
1069  * to allocate a block located at @lblock
1070  */
1071 static int ext4_calc_metadata_amount(struct inode *inode, sector_t lblock)
1072 {
1073         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1074                 return ext4_ext_calc_metadata_amount(inode, lblock);
1075
1076         return ext4_indirect_calc_metadata_amount(inode, lblock);
1077 }
1078
1079 /*
1080  * Called with i_data_sem down, which is important since we can call
1081  * ext4_discard_preallocations() from here.
1082  */
1083 void ext4_da_update_reserve_space(struct inode *inode,
1084                                         int used, int quota_claim)
1085 {
1086         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1087         struct ext4_inode_info *ei = EXT4_I(inode);
1088
1089         spin_lock(&ei->i_block_reservation_lock);
1090         trace_ext4_da_update_reserve_space(inode, used);
1091         if (unlikely(used > ei->i_reserved_data_blocks)) {
1092                 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
1093                          "with only %d reserved data blocks\n",
1094                          __func__, inode->i_ino, used,
1095                          ei->i_reserved_data_blocks);
1096                 WARN_ON(1);
1097                 used = ei->i_reserved_data_blocks;
1098         }
1099
1100         /* Update per-inode reservations */
1101         ei->i_reserved_data_blocks -= used;
1102         ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
1103         percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1104                            used + ei->i_allocated_meta_blocks);
1105         ei->i_allocated_meta_blocks = 0;
1106
1107         if (ei->i_reserved_data_blocks == 0) {
1108                 /*
1109                  * We can release all of the reserved metadata blocks
1110                  * only when we have written all of the delayed
1111                  * allocation blocks.
1112                  */
1113                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1114                                    ei->i_reserved_meta_blocks);
1115                 ei->i_reserved_meta_blocks = 0;
1116                 ei->i_da_metadata_calc_len = 0;
1117         }
1118         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1119
1120         /* Update quota subsystem for data blocks */
1121         if (quota_claim)
1122                 dquot_claim_block(inode, used);
1123         else {
1124                 /*
1125                  * We did fallocate with an offset that is already delayed
1126                  * allocated. So on delayed allocated writeback we should
1127                  * not re-claim the quota for fallocated blocks.
1128                  */
1129                 dquot_release_reservation_block(inode, used);
1130         }
1131
1132         /*
1133          * If we have done all the pending block allocations and if
1134          * there aren't any writers on the inode, we can discard the
1135          * inode's preallocations.
1136          */
1137         if ((ei->i_reserved_data_blocks == 0) &&
1138             (atomic_read(&inode->i_writecount) == 0))
1139                 ext4_discard_preallocations(inode);
1140 }
1141
1142 static int __check_block_validity(struct inode *inode, const char *func,
1143                                 unsigned int line,
1144                                 struct ext4_map_blocks *map)
1145 {
1146         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
1147                                    map->m_len)) {
1148                 ext4_error_inode(inode, func, line, map->m_pblk,
1149                                  "lblock %lu mapped to illegal pblock "
1150                                  "(length %d)", (unsigned long) map->m_lblk,
1151                                  map->m_len);
1152                 return -EIO;
1153         }
1154         return 0;
1155 }
1156
1157 #define check_block_validity(inode, map)        \
1158         __check_block_validity((inode), __func__, __LINE__, (map))
1159
1160 /*
1161  * Return the number of contiguous dirty pages in a given inode
1162  * starting at page frame idx.
1163  */
1164 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
1165                                     unsigned int max_pages)
1166 {
1167         struct address_space *mapping = inode->i_mapping;
1168         pgoff_t index;
1169         struct pagevec pvec;
1170         pgoff_t num = 0;
1171         int i, nr_pages, done = 0;
1172
1173         if (max_pages == 0)
1174                 return 0;
1175         pagevec_init(&pvec, 0);
1176         while (!done) {
1177                 index = idx;
1178                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
1179                                               PAGECACHE_TAG_DIRTY,
1180                                               (pgoff_t)PAGEVEC_SIZE);
1181                 if (nr_pages == 0)
1182                         break;
1183                 for (i = 0; i < nr_pages; i++) {
1184                         struct page *page = pvec.pages[i];
1185                         struct buffer_head *bh, *head;
1186
1187                         lock_page(page);
1188                         if (unlikely(page->mapping != mapping) ||
1189                             !PageDirty(page) ||
1190                             PageWriteback(page) ||
1191                             page->index != idx) {
1192                                 done = 1;
1193                                 unlock_page(page);
1194                                 break;
1195                         }
1196                         if (page_has_buffers(page)) {
1197                                 bh = head = page_buffers(page);
1198                                 do {
1199                                         if (!buffer_delay(bh) &&
1200                                             !buffer_unwritten(bh))
1201                                                 done = 1;
1202                                         bh = bh->b_this_page;
1203                                 } while (!done && (bh != head));
1204                         }
1205                         unlock_page(page);
1206                         if (done)
1207                                 break;
1208                         idx++;
1209                         num++;
1210                         if (num >= max_pages)
1211                                 break;
1212                 }
1213                 pagevec_release(&pvec);
1214         }
1215         return num;
1216 }
1217
1218 /*
1219  * The ext4_map_blocks() function tries to look up the requested blocks,
1220  * and returns if the blocks are already mapped.
1221  *
1222  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1223  * and store the allocated blocks in the result buffer head and mark it
1224  * mapped.
1225  *
1226  * If file type is extents based, it will call ext4_ext_map_blocks(),
1227  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
1228  * based files
1229  *
1230  * On success, it returns the number of blocks being mapped or allocate.
1231  * if create==0 and the blocks are pre-allocated and uninitialized block,
1232  * the result buffer head is unmapped. If the create ==1, it will make sure
1233  * the buffer head is mapped.
1234  *
1235  * It returns 0 if plain look up failed (blocks have not been allocated), in
1236  * that casem, buffer head is unmapped
1237  *
1238  * It returns the error in case of allocation failure.
1239  */
1240 int ext4_map_blocks(handle_t *handle, struct inode *inode,
1241                     struct ext4_map_blocks *map, int flags)
1242 {
1243         int retval;
1244
1245         map->m_flags = 0;
1246         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
1247                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
1248                   (unsigned long) map->m_lblk);
1249         /*
1250          * Try to see if we can get the block without requesting a new
1251          * file system block.
1252          */
1253         down_read((&EXT4_I(inode)->i_data_sem));
1254         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
1255                 retval = ext4_ext_map_blocks(handle, inode, map, 0);
1256         } else {
1257                 retval = ext4_ind_map_blocks(handle, inode, map, 0);
1258         }
1259         up_read((&EXT4_I(inode)->i_data_sem));
1260
1261         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
1262                 int ret = check_block_validity(inode, map);
1263                 if (ret != 0)
1264                         return ret;
1265         }
1266
1267         /* If it is only a block(s) look up */
1268         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1269                 return retval;
1270
1271         /*
1272          * Returns if the blocks have already allocated
1273          *
1274          * Note that if blocks have been preallocated
1275          * ext4_ext_get_block() returns th create = 0
1276          * with buffer head unmapped.
1277          */
1278         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
1279                 return retval;
1280
1281         /*
1282          * When we call get_blocks without the create flag, the
1283          * BH_Unwritten flag could have gotten set if the blocks
1284          * requested were part of a uninitialized extent.  We need to
1285          * clear this flag now that we are committed to convert all or
1286          * part of the uninitialized extent to be an initialized
1287          * extent.  This is because we need to avoid the combination
1288          * of BH_Unwritten and BH_Mapped flags being simultaneously
1289          * set on the buffer_head.
1290          */
1291         map->m_flags &= ~EXT4_MAP_UNWRITTEN;
1292
1293         /*
1294          * New blocks allocate and/or writing to uninitialized extent
1295          * will possibly result in updating i_data, so we take
1296          * the write lock of i_data_sem, and call get_blocks()
1297          * with create == 1 flag.
1298          */
1299         down_write((&EXT4_I(inode)->i_data_sem));
1300
1301         /*
1302          * if the caller is from delayed allocation writeout path
1303          * we have already reserved fs blocks for allocation
1304          * let the underlying get_block() function know to
1305          * avoid double accounting
1306          */
1307         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1308                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1309         /*
1310          * We need to check for EXT4 here because migrate
1311          * could have changed the inode type in between
1312          */
1313         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
1314                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
1315         } else {
1316                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
1317
1318                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
1319                         /*
1320                          * We allocated new blocks which will result in
1321                          * i_data's format changing.  Force the migrate
1322                          * to fail by clearing migrate flags
1323                          */
1324                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
1325                 }
1326
1327                 /*
1328                  * Update reserved blocks/metadata blocks after successful
1329                  * block allocation which had been deferred till now. We don't
1330                  * support fallocate for non extent files. So we can update
1331                  * reserve space here.
1332                  */
1333                 if ((retval > 0) &&
1334                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
1335                         ext4_da_update_reserve_space(inode, retval, 1);
1336         }
1337         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1338                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1339
1340         up_write((&EXT4_I(inode)->i_data_sem));
1341         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
1342                 int ret = check_block_validity(inode, map);
1343                 if (ret != 0)
1344                         return ret;
1345         }
1346         return retval;
1347 }
1348
1349 /* Maximum number of blocks we map for direct IO at once. */
1350 #define DIO_MAX_BLOCKS 4096
1351
1352 static int _ext4_get_block(struct inode *inode, sector_t iblock,
1353                            struct buffer_head *bh, int flags)
1354 {
1355         handle_t *handle = ext4_journal_current_handle();
1356         struct ext4_map_blocks map;
1357         int ret = 0, started = 0;
1358         int dio_credits;
1359
1360         map.m_lblk = iblock;
1361         map.m_len = bh->b_size >> inode->i_blkbits;
1362
1363         if (flags && !handle) {
1364                 /* Direct IO write... */
1365                 if (map.m_len > DIO_MAX_BLOCKS)
1366                         map.m_len = DIO_MAX_BLOCKS;
1367                 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
1368                 handle = ext4_journal_start(inode, dio_credits);
1369                 if (IS_ERR(handle)) {
1370                         ret = PTR_ERR(handle);
1371                         return ret;
1372                 }
1373                 started = 1;
1374         }
1375
1376         ret = ext4_map_blocks(handle, inode, &map, flags);
1377         if (ret > 0) {
1378                 map_bh(bh, inode->i_sb, map.m_pblk);
1379                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
1380                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
1381                 ret = 0;
1382         }
1383         if (started)
1384                 ext4_journal_stop(handle);
1385         return ret;
1386 }
1387
1388 int ext4_get_block(struct inode *inode, sector_t iblock,
1389                    struct buffer_head *bh, int create)
1390 {
1391         return _ext4_get_block(inode, iblock, bh,
1392                                create ? EXT4_GET_BLOCKS_CREATE : 0);
1393 }
1394
1395 /*
1396  * `handle' can be NULL if create is zero
1397  */
1398 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1399                                 ext4_lblk_t block, int create, int *errp)
1400 {
1401         struct ext4_map_blocks map;
1402         struct buffer_head *bh;
1403         int fatal = 0, err;
1404
1405         J_ASSERT(handle != NULL || create == 0);
1406
1407         map.m_lblk = block;
1408         map.m_len = 1;
1409         err = ext4_map_blocks(handle, inode, &map,
1410                               create ? EXT4_GET_BLOCKS_CREATE : 0);
1411
1412         if (err < 0)
1413                 *errp = err;
1414         if (err <= 0)
1415                 return NULL;
1416         *errp = 0;
1417
1418         bh = sb_getblk(inode->i_sb, map.m_pblk);
1419         if (!bh) {
1420                 *errp = -EIO;
1421                 return NULL;
1422         }
1423         if (map.m_flags & EXT4_MAP_NEW) {
1424                 J_ASSERT(create != 0);
1425                 J_ASSERT(handle != NULL);
1426
1427                 /*
1428                  * Now that we do not always journal data, we should
1429                  * keep in mind whether this should always journal the
1430                  * new buffer as metadata.  For now, regular file
1431                  * writes use ext4_get_block instead, so it's not a
1432                  * problem.
1433                  */
1434                 lock_buffer(bh);
1435                 BUFFER_TRACE(bh, "call get_create_access");
1436                 fatal = ext4_journal_get_create_access(handle, bh);
1437                 if (!fatal && !buffer_uptodate(bh)) {
1438                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1439                         set_buffer_uptodate(bh);
1440                 }
1441                 unlock_buffer(bh);
1442                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1443                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1444                 if (!fatal)
1445                         fatal = err;
1446         } else {
1447                 BUFFER_TRACE(bh, "not a new buffer");
1448         }
1449         if (fatal) {
1450                 *errp = fatal;
1451                 brelse(bh);
1452                 bh = NULL;
1453         }
1454         return bh;
1455 }
1456
1457 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1458                                ext4_lblk_t block, int create, int *err)
1459 {
1460         struct buffer_head *bh;
1461
1462         bh = ext4_getblk(handle, inode, block, create, err);
1463         if (!bh)
1464                 return bh;
1465         if (buffer_uptodate(bh))
1466                 return bh;
1467         ll_rw_block(READ_META, 1, &bh);
1468         wait_on_buffer(bh);
1469         if (buffer_uptodate(bh))
1470                 return bh;
1471         put_bh(bh);
1472         *err = -EIO;
1473         return NULL;
1474 }
1475
1476 static int walk_page_buffers(handle_t *handle,
1477                              struct buffer_head *head,
1478                              unsigned from,
1479                              unsigned to,
1480                              int *partial,
1481                              int (*fn)(handle_t *handle,
1482                                        struct buffer_head *bh))
1483 {
1484         struct buffer_head *bh;
1485         unsigned block_start, block_end;
1486         unsigned blocksize = head->b_size;
1487         int err, ret = 0;
1488         struct buffer_head *next;
1489
1490         for (bh = head, block_start = 0;
1491              ret == 0 && (bh != head || !block_start);
1492              block_start = block_end, bh = next) {
1493                 next = bh->b_this_page;
1494                 block_end = block_start + blocksize;
1495                 if (block_end <= from || block_start >= to) {
1496                         if (partial && !buffer_uptodate(bh))
1497                                 *partial = 1;
1498                         continue;
1499                 }
1500                 err = (*fn)(handle, bh);
1501                 if (!ret)
1502                         ret = err;
1503         }
1504         return ret;
1505 }
1506
1507 /*
1508  * To preserve ordering, it is essential that the hole instantiation and
1509  * the data write be encapsulated in a single transaction.  We cannot
1510  * close off a transaction and start a new one between the ext4_get_block()
1511  * and the commit_write().  So doing the jbd2_journal_start at the start of
1512  * prepare_write() is the right place.
1513  *
1514  * Also, this function can nest inside ext4_writepage() ->
1515  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1516  * has generated enough buffer credits to do the whole page.  So we won't
1517  * block on the journal in that case, which is good, because the caller may
1518  * be PF_MEMALLOC.
1519  *
1520  * By accident, ext4 can be reentered when a transaction is open via
1521  * quota file writes.  If we were to commit the transaction while thus
1522  * reentered, there can be a deadlock - we would be holding a quota
1523  * lock, and the commit would never complete if another thread had a
1524  * transaction open and was blocking on the quota lock - a ranking
1525  * violation.
1526  *
1527  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1528  * will _not_ run commit under these circumstances because handle->h_ref
1529  * is elevated.  We'll still have enough credits for the tiny quotafile
1530  * write.
1531  */
1532 static int do_journal_get_write_access(handle_t *handle,
1533                                        struct buffer_head *bh)
1534 {
1535         int dirty = buffer_dirty(bh);
1536         int ret;
1537
1538         if (!buffer_mapped(bh) || buffer_freed(bh))
1539                 return 0;
1540         /*
1541          * __block_write_begin() could have dirtied some buffers. Clean
1542          * the dirty bit as jbd2_journal_get_write_access() could complain
1543          * otherwise about fs integrity issues. Setting of the dirty bit
1544          * by __block_write_begin() isn't a real problem here as we clear
1545          * the bit before releasing a page lock and thus writeback cannot
1546          * ever write the buffer.
1547          */
1548         if (dirty)
1549                 clear_buffer_dirty(bh);
1550         ret = ext4_journal_get_write_access(handle, bh);
1551         if (!ret && dirty)
1552                 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1553         return ret;
1554 }
1555
1556 /*
1557  * Truncate blocks that were not used by write. We have to truncate the
1558  * pagecache as well so that corresponding buffers get properly unmapped.
1559  */
1560 static void ext4_truncate_failed_write(struct inode *inode)
1561 {
1562         truncate_inode_pages(inode->i_mapping, inode->i_size);
1563         ext4_truncate(inode);
1564 }
1565
1566 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
1567                    struct buffer_head *bh_result, int create);
1568 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1569                             loff_t pos, unsigned len, unsigned flags,
1570                             struct page **pagep, void **fsdata)
1571 {
1572         struct inode *inode = mapping->host;
1573         int ret, needed_blocks;
1574         handle_t *handle;
1575         int retries = 0;
1576         struct page *page;
1577         pgoff_t index;
1578         unsigned from, to;
1579
1580         trace_ext4_write_begin(inode, pos, len, flags);
1581         /*
1582          * Reserve one block more for addition to orphan list in case
1583          * we allocate blocks but write fails for some reason
1584          */
1585         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1586         index = pos >> PAGE_CACHE_SHIFT;
1587         from = pos & (PAGE_CACHE_SIZE - 1);
1588         to = from + len;
1589
1590 retry:
1591         handle = ext4_journal_start(inode, needed_blocks);
1592         if (IS_ERR(handle)) {
1593                 ret = PTR_ERR(handle);
1594                 goto out;
1595         }
1596
1597         /* We cannot recurse into the filesystem as the transaction is already
1598          * started */
1599         flags |= AOP_FLAG_NOFS;
1600
1601         page = grab_cache_page_write_begin(mapping, index, flags);
1602         if (!page) {
1603                 ext4_journal_stop(handle);
1604                 ret = -ENOMEM;
1605                 goto out;
1606         }
1607         *pagep = page;
1608
1609         if (ext4_should_dioread_nolock(inode))
1610                 ret = __block_write_begin(page, pos, len, ext4_get_block_write);
1611         else
1612                 ret = __block_write_begin(page, pos, len, ext4_get_block);
1613
1614         if (!ret && ext4_should_journal_data(inode)) {
1615                 ret = walk_page_buffers(handle, page_buffers(page),
1616                                 from, to, NULL, do_journal_get_write_access);
1617         }
1618
1619         if (ret) {
1620                 unlock_page(page);
1621                 page_cache_release(page);
1622                 /*
1623                  * __block_write_begin may have instantiated a few blocks
1624                  * outside i_size.  Trim these off again. Don't need
1625                  * i_size_read because we hold i_mutex.
1626                  *
1627                  * Add inode to orphan list in case we crash before
1628                  * truncate finishes
1629                  */
1630                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1631                         ext4_orphan_add(handle, inode);
1632
1633                 ext4_journal_stop(handle);
1634                 if (pos + len > inode->i_size) {
1635                         ext4_truncate_failed_write(inode);
1636                         /*
1637                          * If truncate failed early the inode might
1638                          * still be on the orphan list; we need to
1639                          * make sure the inode is removed from the
1640                          * orphan list in that case.
1641                          */
1642                         if (inode->i_nlink)
1643                                 ext4_orphan_del(NULL, inode);
1644                 }
1645         }
1646
1647         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1648                 goto retry;
1649 out:
1650         return ret;
1651 }
1652
1653 /* For write_end() in data=journal mode */
1654 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1655 {
1656         if (!buffer_mapped(bh) || buffer_freed(bh))
1657                 return 0;
1658         set_buffer_uptodate(bh);
1659         return ext4_handle_dirty_metadata(handle, NULL, bh);
1660 }
1661
1662 static int ext4_generic_write_end(struct file *file,
1663                                   struct address_space *mapping,
1664                                   loff_t pos, unsigned len, unsigned copied,
1665                                   struct page *page, void *fsdata)
1666 {
1667         int i_size_changed = 0;
1668         struct inode *inode = mapping->host;
1669         handle_t *handle = ext4_journal_current_handle();
1670
1671         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1672
1673         /*
1674          * No need to use i_size_read() here, the i_size
1675          * cannot change under us because we hold i_mutex.
1676          *
1677          * But it's important to update i_size while still holding page lock:
1678          * page writeout could otherwise come in and zero beyond i_size.
1679          */
1680         if (pos + copied > inode->i_size) {
1681                 i_size_write(inode, pos + copied);
1682                 i_size_changed = 1;
1683         }
1684
1685         if (pos + copied >  EXT4_I(inode)->i_disksize) {
1686                 /* We need to mark inode dirty even if
1687                  * new_i_size is less that inode->i_size
1688                  * bu greater than i_disksize.(hint delalloc)
1689                  */
1690                 ext4_update_i_disksize(inode, (pos + copied));
1691                 i_size_changed = 1;
1692         }
1693         unlock_page(page);
1694         page_cache_release(page);
1695
1696         /*
1697          * Don't mark the inode dirty under page lock. First, it unnecessarily
1698          * makes the holding time of page lock longer. Second, it forces lock
1699          * ordering of page lock and transaction start for journaling
1700          * filesystems.
1701          */
1702         if (i_size_changed)
1703                 ext4_mark_inode_dirty(handle, inode);
1704
1705         return copied;
1706 }
1707
1708 /*
1709  * We need to pick up the new inode size which generic_commit_write gave us
1710  * `file' can be NULL - eg, when called from page_symlink().
1711  *
1712  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1713  * buffers are managed internally.
1714  */
1715 static int ext4_ordered_write_end(struct file *file,
1716                                   struct address_space *mapping,
1717                                   loff_t pos, unsigned len, unsigned copied,
1718                                   struct page *page, void *fsdata)
1719 {
1720         handle_t *handle = ext4_journal_current_handle();
1721         struct inode *inode = mapping->host;
1722         int ret = 0, ret2;
1723
1724         trace_ext4_ordered_write_end(inode, pos, len, copied);
1725         ret = ext4_jbd2_file_inode(handle, inode);
1726
1727         if (ret == 0) {
1728                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1729                                                         page, fsdata);
1730                 copied = ret2;
1731                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1732                         /* if we have allocated more blocks and copied
1733                          * less. We will have blocks allocated outside
1734                          * inode->i_size. So truncate them
1735                          */
1736                         ext4_orphan_add(handle, inode);
1737                 if (ret2 < 0)
1738                         ret = ret2;
1739         }
1740         ret2 = ext4_journal_stop(handle);
1741         if (!ret)
1742                 ret = ret2;
1743
1744         if (pos + len > inode->i_size) {
1745                 ext4_truncate_failed_write(inode);
1746                 /*
1747                  * If truncate failed early the inode might still be
1748                  * on the orphan list; we need to make sure the inode
1749                  * is removed from the orphan list in that case.
1750                  */
1751                 if (inode->i_nlink)
1752                         ext4_orphan_del(NULL, inode);
1753         }
1754
1755
1756         return ret ? ret : copied;
1757 }
1758
1759 static int ext4_writeback_write_end(struct file *file,
1760                                     struct address_space *mapping,
1761                                     loff_t pos, unsigned len, unsigned copied,
1762                                     struct page *page, void *fsdata)
1763 {
1764         handle_t *handle = ext4_journal_current_handle();
1765         struct inode *inode = mapping->host;
1766         int ret = 0, ret2;
1767
1768         trace_ext4_writeback_write_end(inode, pos, len, copied);
1769         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1770                                                         page, fsdata);
1771         copied = ret2;
1772         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1773                 /* if we have allocated more blocks and copied
1774                  * less. We will have blocks allocated outside
1775                  * inode->i_size. So truncate them
1776                  */
1777                 ext4_orphan_add(handle, inode);
1778
1779         if (ret2 < 0)
1780                 ret = ret2;
1781
1782         ret2 = ext4_journal_stop(handle);
1783         if (!ret)
1784                 ret = ret2;
1785
1786         if (pos + len > inode->i_size) {
1787                 ext4_truncate_failed_write(inode);
1788                 /*
1789                  * If truncate failed early the inode might still be
1790                  * on the orphan list; we need to make sure the inode
1791                  * is removed from the orphan list in that case.
1792                  */
1793                 if (inode->i_nlink)
1794                         ext4_orphan_del(NULL, inode);
1795         }
1796
1797         return ret ? ret : copied;
1798 }
1799
1800 static int ext4_journalled_write_end(struct file *file,
1801                                      struct address_space *mapping,
1802                                      loff_t pos, unsigned len, unsigned copied,
1803                                      struct page *page, void *fsdata)
1804 {
1805         handle_t *handle = ext4_journal_current_handle();
1806         struct inode *inode = mapping->host;
1807         int ret = 0, ret2;
1808         int partial = 0;
1809         unsigned from, to;
1810         loff_t new_i_size;
1811
1812         trace_ext4_journalled_write_end(inode, pos, len, copied);
1813         from = pos & (PAGE_CACHE_SIZE - 1);
1814         to = from + len;
1815
1816         if (copied < len) {
1817                 if (!PageUptodate(page))
1818                         copied = 0;
1819                 page_zero_new_buffers(page, from+copied, to);
1820         }
1821
1822         ret = walk_page_buffers(handle, page_buffers(page), from,
1823                                 to, &partial, write_end_fn);
1824         if (!partial)
1825                 SetPageUptodate(page);
1826         new_i_size = pos + copied;
1827         if (new_i_size > inode->i_size)
1828                 i_size_write(inode, pos+copied);
1829         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1830         if (new_i_size > EXT4_I(inode)->i_disksize) {
1831                 ext4_update_i_disksize(inode, new_i_size);
1832                 ret2 = ext4_mark_inode_dirty(handle, inode);
1833                 if (!ret)
1834                         ret = ret2;
1835         }
1836
1837         unlock_page(page);
1838         page_cache_release(page);
1839         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1840                 /* if we have allocated more blocks and copied
1841                  * less. We will have blocks allocated outside
1842                  * inode->i_size. So truncate them
1843                  */
1844                 ext4_orphan_add(handle, inode);
1845
1846         ret2 = ext4_journal_stop(handle);
1847         if (!ret)
1848                 ret = ret2;
1849         if (pos + len > inode->i_size) {
1850                 ext4_truncate_failed_write(inode);
1851                 /*
1852                  * If truncate failed early the inode might still be
1853                  * on the orphan list; we need to make sure the inode
1854                  * is removed from the orphan list in that case.
1855                  */
1856                 if (inode->i_nlink)
1857                         ext4_orphan_del(NULL, inode);
1858         }
1859
1860         return ret ? ret : copied;
1861 }
1862
1863 /*
1864  * Reserve a single block located at lblock
1865  */
1866 static int ext4_da_reserve_space(struct inode *inode, sector_t lblock)
1867 {
1868         int retries = 0;
1869         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1870         struct ext4_inode_info *ei = EXT4_I(inode);
1871         unsigned long md_needed;
1872         int ret;
1873
1874         /*
1875          * recalculate the amount of metadata blocks to reserve
1876          * in order to allocate nrblocks
1877          * worse case is one extent per block
1878          */
1879 repeat:
1880         spin_lock(&ei->i_block_reservation_lock);
1881         md_needed = ext4_calc_metadata_amount(inode, lblock);
1882         trace_ext4_da_reserve_space(inode, md_needed);
1883         spin_unlock(&ei->i_block_reservation_lock);
1884
1885         /*
1886          * We will charge metadata quota at writeout time; this saves
1887          * us from metadata over-estimation, though we may go over by
1888          * a small amount in the end.  Here we just reserve for data.
1889          */
1890         ret = dquot_reserve_block(inode, 1);
1891         if (ret)
1892                 return ret;
1893         /*
1894          * We do still charge estimated metadata to the sb though;
1895          * we cannot afford to run out of free blocks.
1896          */
1897         if (ext4_claim_free_blocks(sbi, md_needed + 1)) {
1898                 dquot_release_reservation_block(inode, 1);
1899                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1900                         yield();
1901                         goto repeat;
1902                 }
1903                 return -ENOSPC;
1904         }
1905         spin_lock(&ei->i_block_reservation_lock);
1906         ei->i_reserved_data_blocks++;
1907         ei->i_reserved_meta_blocks += md_needed;
1908         spin_unlock(&ei->i_block_reservation_lock);
1909
1910         return 0;       /* success */
1911 }
1912
1913 static void ext4_da_release_space(struct inode *inode, int to_free)
1914 {
1915         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1916         struct ext4_inode_info *ei = EXT4_I(inode);
1917
1918         if (!to_free)
1919                 return;         /* Nothing to release, exit */
1920
1921         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1922
1923         trace_ext4_da_release_space(inode, to_free);
1924         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1925                 /*
1926                  * if there aren't enough reserved blocks, then the
1927                  * counter is messed up somewhere.  Since this
1928                  * function is called from invalidate page, it's
1929                  * harmless to return without any action.
1930                  */
1931                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1932                          "ino %lu, to_free %d with only %d reserved "
1933                          "data blocks\n", inode->i_ino, to_free,
1934                          ei->i_reserved_data_blocks);
1935                 WARN_ON(1);
1936                 to_free = ei->i_reserved_data_blocks;
1937         }
1938         ei->i_reserved_data_blocks -= to_free;
1939
1940         if (ei->i_reserved_data_blocks == 0) {
1941                 /*
1942                  * We can release all of the reserved metadata blocks
1943                  * only when we have written all of the delayed
1944                  * allocation blocks.
1945                  */
1946                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1947                                    ei->i_reserved_meta_blocks);
1948                 ei->i_reserved_meta_blocks = 0;
1949                 ei->i_da_metadata_calc_len = 0;
1950         }
1951
1952         /* update fs dirty data blocks counter */
1953         percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1954
1955         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1956
1957         dquot_release_reservation_block(inode, to_free);
1958 }
1959
1960 static void ext4_da_page_release_reservation(struct page *page,
1961                                              unsigned long offset)
1962 {
1963         int to_release = 0;
1964         struct buffer_head *head, *bh;
1965         unsigned int curr_off = 0;
1966
1967         head = page_buffers(page);
1968         bh = head;
1969         do {
1970                 unsigned int next_off = curr_off + bh->b_size;
1971
1972                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1973                         to_release++;
1974                         clear_buffer_delay(bh);
1975                 }
1976                 curr_off = next_off;
1977         } while ((bh = bh->b_this_page) != head);
1978         ext4_da_release_space(page->mapping->host, to_release);
1979 }
1980
1981 /*
1982  * Delayed allocation stuff
1983  */
1984
1985 /*
1986  * mpage_da_submit_io - walks through extent of pages and try to write
1987  * them with writepage() call back
1988  *
1989  * @mpd->inode: inode
1990  * @mpd->first_page: first page of the extent
1991  * @mpd->next_page: page after the last page of the extent
1992  *
1993  * By the time mpage_da_submit_io() is called we expect all blocks
1994  * to be allocated. this may be wrong if allocation failed.
1995  *
1996  * As pages are already locked by write_cache_pages(), we can't use it
1997  */
1998 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1999 {
2000         long pages_skipped;
2001         struct pagevec pvec;
2002         unsigned long index, end;
2003         int ret = 0, err, nr_pages, i;
2004         struct inode *inode = mpd->inode;
2005         struct address_space *mapping = inode->i_mapping;
2006
2007         BUG_ON(mpd->next_page <= mpd->first_page);
2008         /*
2009          * We need to start from the first_page to the next_page - 1
2010          * to make sure we also write the mapped dirty buffer_heads.
2011          * If we look at mpd->b_blocknr we would only be looking
2012          * at the currently mapped buffer_heads.
2013          */
2014         index = mpd->first_page;
2015         end = mpd->next_page - 1;
2016
2017         pagevec_init(&pvec, 0);
2018         while (index <= end) {
2019                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2020                 if (nr_pages == 0)
2021                         break;
2022                 for (i = 0; i < nr_pages; i++) {
2023                         struct page *page = pvec.pages[i];
2024
2025                         index = page->index;
2026                         if (index > end)
2027                                 break;
2028                         index++;
2029
2030                         BUG_ON(!PageLocked(page));
2031                         BUG_ON(PageWriteback(page));
2032
2033                         pages_skipped = mpd->wbc->pages_skipped;
2034                         err = mapping->a_ops->writepage(page, mpd->wbc);
2035                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
2036                                 /*
2037                                  * have successfully written the page
2038                                  * without skipping the same
2039                                  */
2040                                 mpd->pages_written++;
2041                         /*
2042                          * In error case, we have to continue because
2043                          * remaining pages are still locked
2044                          * XXX: unlock and re-dirty them?
2045                          */
2046                         if (ret == 0)
2047                                 ret = err;
2048                 }
2049                 pagevec_release(&pvec);
2050         }
2051         return ret;
2052 }
2053
2054 /*
2055  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
2056  *
2057  * the function goes through all passed space and put actual disk
2058  * block numbers into buffer heads, dropping BH_Delay and BH_Unwritten
2059  */
2060 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd,
2061                                  struct ext4_map_blocks *map)
2062 {
2063         struct inode *inode = mpd->inode;
2064         struct address_space *mapping = inode->i_mapping;
2065         int blocks = map->m_len;
2066         sector_t pblock = map->m_pblk, cur_logical;
2067         struct buffer_head *head, *bh;
2068         pgoff_t index, end;
2069         struct pagevec pvec;
2070         int nr_pages, i;
2071
2072         index = map->m_lblk >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2073         end = (map->m_lblk + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2074         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2075
2076         pagevec_init(&pvec, 0);
2077
2078         while (index <= end) {
2079                 /* XXX: optimize tail */
2080                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2081                 if (nr_pages == 0)
2082                         break;
2083                 for (i = 0; i < nr_pages; i++) {
2084                         struct page *page = pvec.pages[i];
2085
2086                         index = page->index;
2087                         if (index > end)
2088                                 break;
2089                         index++;
2090
2091                         BUG_ON(!PageLocked(page));
2092                         BUG_ON(PageWriteback(page));
2093                         BUG_ON(!page_has_buffers(page));
2094
2095                         bh = page_buffers(page);
2096                         head = bh;
2097
2098                         /* skip blocks out of the range */
2099                         do {
2100                                 if (cur_logical >= map->m_lblk)
2101                                         break;
2102                                 cur_logical++;
2103                         } while ((bh = bh->b_this_page) != head);
2104
2105                         do {
2106                                 if (cur_logical >= map->m_lblk + blocks)
2107                                         break;
2108
2109                                 if (buffer_delay(bh) || buffer_unwritten(bh)) {
2110
2111                                         BUG_ON(bh->b_bdev != inode->i_sb->s_bdev);
2112
2113                                         if (buffer_delay(bh)) {
2114                                                 clear_buffer_delay(bh);
2115                                                 bh->b_blocknr = pblock;
2116                                         } else {
2117                                                 /*
2118                                                  * unwritten already should have
2119                                                  * blocknr assigned. Verify that
2120                                                  */
2121                                                 clear_buffer_unwritten(bh);
2122                                                 BUG_ON(bh->b_blocknr != pblock);
2123                                         }
2124
2125                                 } else if (buffer_mapped(bh))
2126                                         BUG_ON(bh->b_blocknr != pblock);
2127
2128                                 if (map->m_flags & EXT4_MAP_UNINIT)
2129                                         set_buffer_uninit(bh);
2130                                 cur_logical++;
2131                                 pblock++;
2132                         } while ((bh = bh->b_this_page) != head);
2133                 }
2134                 pagevec_release(&pvec);
2135         }
2136 }
2137
2138
2139 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2140                                         sector_t logical, long blk_cnt)
2141 {
2142         int nr_pages, i;
2143         pgoff_t index, end;
2144         struct pagevec pvec;
2145         struct inode *inode = mpd->inode;
2146         struct address_space *mapping = inode->i_mapping;
2147
2148         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2149         end   = (logical + blk_cnt - 1) >>
2150                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2151         while (index <= end) {
2152                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2153                 if (nr_pages == 0)
2154                         break;
2155                 for (i = 0; i < nr_pages; i++) {
2156                         struct page *page = pvec.pages[i];
2157                         if (page->index > end)
2158                                 break;
2159                         BUG_ON(!PageLocked(page));
2160                         BUG_ON(PageWriteback(page));
2161                         block_invalidatepage(page, 0);
2162                         ClearPageUptodate(page);
2163                         unlock_page(page);
2164                 }
2165                 index = pvec.pages[nr_pages - 1]->index + 1;
2166                 pagevec_release(&pvec);
2167         }
2168         return;
2169 }
2170
2171 static void ext4_print_free_blocks(struct inode *inode)
2172 {
2173         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2174         printk(KERN_CRIT "Total free blocks count %lld\n",
2175                ext4_count_free_blocks(inode->i_sb));
2176         printk(KERN_CRIT "Free/Dirty block details\n");
2177         printk(KERN_CRIT "free_blocks=%lld\n",
2178                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2179         printk(KERN_CRIT "dirty_blocks=%lld\n",
2180                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2181         printk(KERN_CRIT "Block reservation details\n");
2182         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2183                EXT4_I(inode)->i_reserved_data_blocks);
2184         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2185                EXT4_I(inode)->i_reserved_meta_blocks);
2186         return;
2187 }
2188
2189 /*
2190  * mpage_da_map_blocks - go through given space
2191  *
2192  * @mpd - bh describing space
2193  *
2194  * The function skips space we know is already mapped to disk blocks.
2195  *
2196  */
2197 static int mpage_da_map_blocks(struct mpage_da_data *mpd)
2198 {
2199         int err, blks, get_blocks_flags;
2200         struct ext4_map_blocks map;
2201         sector_t next = mpd->b_blocknr;
2202         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2203         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2204         handle_t *handle = NULL;
2205
2206         /*
2207          * We consider only non-mapped and non-allocated blocks
2208          */
2209         if ((mpd->b_state  & (1 << BH_Mapped)) &&
2210                 !(mpd->b_state & (1 << BH_Delay)) &&
2211                 !(mpd->b_state & (1 << BH_Unwritten)))
2212                 return 0;
2213
2214         /*
2215          * If we didn't accumulate anything to write simply return
2216          */
2217         if (!mpd->b_size)
2218                 return 0;
2219
2220         handle = ext4_journal_current_handle();
2221         BUG_ON(!handle);
2222
2223         /*
2224          * Call ext4_map_blocks() to allocate any delayed allocation
2225          * blocks, or to convert an uninitialized extent to be
2226          * initialized (in the case where we have written into
2227          * one or more preallocated blocks).
2228          *
2229          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2230          * indicate that we are on the delayed allocation path.  This
2231          * affects functions in many different parts of the allocation
2232          * call path.  This flag exists primarily because we don't
2233          * want to change *many* call functions, so ext4_map_blocks()
2234          * will set the magic i_delalloc_reserved_flag once the
2235          * inode's allocation semaphore is taken.
2236          *
2237          * If the blocks in questions were delalloc blocks, set
2238          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2239          * variables are updated after the blocks have been allocated.
2240          */
2241         map.m_lblk = next;
2242         map.m_len = max_blocks;
2243         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
2244         if (ext4_should_dioread_nolock(mpd->inode))
2245                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2246         if (mpd->b_state & (1 << BH_Delay))
2247                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2248
2249         blks = ext4_map_blocks(handle, mpd->inode, &map, get_blocks_flags);
2250         if (blks < 0) {
2251                 struct super_block *sb = mpd->inode->i_sb;
2252
2253                 err = blks;
2254                 /*
2255                  * If get block returns with error we simply
2256                  * return. Later writepage will redirty the page and
2257                  * writepages will find the dirty page again
2258                  */
2259                 if (err == -EAGAIN)
2260                         return 0;
2261
2262                 if (err == -ENOSPC &&
2263                     ext4_count_free_blocks(sb)) {
2264                         mpd->retval = err;
2265                         return 0;
2266                 }
2267
2268                 /*
2269                  * get block failure will cause us to loop in
2270                  * writepages, because a_ops->writepage won't be able
2271                  * to make progress. The page will be redirtied by
2272                  * writepage and writepages will again try to write
2273                  * the same.
2274                  */
2275                 if (!(EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2276                         ext4_msg(sb, KERN_CRIT,
2277                                  "delayed block allocation failed for inode %lu "
2278                                  "at logical offset %llu with max blocks %zd "
2279                                  "with error %d", mpd->inode->i_ino,
2280                                  (unsigned long long) next,
2281                                  mpd->b_size >> mpd->inode->i_blkbits, err);
2282                         ext4_msg(sb, KERN_CRIT,
2283                                 "This should not happen!! Data will be lost\n");
2284                         if (err == -ENOSPC)
2285                                 ext4_print_free_blocks(mpd->inode);
2286                 }
2287                 /* invalidate all the pages */
2288                 ext4_da_block_invalidatepages(mpd, next,
2289                                 mpd->b_size >> mpd->inode->i_blkbits);
2290                 return err;
2291         }
2292         BUG_ON(blks == 0);
2293
2294         if (map.m_flags & EXT4_MAP_NEW) {
2295                 struct block_device *bdev = mpd->inode->i_sb->s_bdev;
2296                 int i;
2297
2298                 for (i = 0; i < map.m_len; i++)
2299                         unmap_underlying_metadata(bdev, map.m_pblk + i);
2300         }
2301
2302         /*
2303          * If blocks are delayed marked, we need to
2304          * put actual blocknr and drop delayed bit
2305          */
2306         if ((mpd->b_state & (1 << BH_Delay)) ||
2307             (mpd->b_state & (1 << BH_Unwritten)))
2308                 mpage_put_bnr_to_bhs(mpd, &map);
2309
2310         if (ext4_should_order_data(mpd->inode)) {
2311                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2312                 if (err)
2313                         return err;
2314         }
2315
2316         /*
2317          * Update on-disk size along with block allocation.
2318          */
2319         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2320         if (disksize > i_size_read(mpd->inode))
2321                 disksize = i_size_read(mpd->inode);
2322         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2323                 ext4_update_i_disksize(mpd->inode, disksize);
2324                 return ext4_mark_inode_dirty(handle, mpd->inode);
2325         }
2326
2327         return 0;
2328 }
2329
2330 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2331                 (1 << BH_Delay) | (1 << BH_Unwritten))
2332
2333 /*
2334  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2335  *
2336  * @mpd->lbh - extent of blocks
2337  * @logical - logical number of the block in the file
2338  * @bh - bh of the block (used to access block's state)
2339  *
2340  * the function is used to collect contig. blocks in same state
2341  */
2342 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2343                                    sector_t logical, size_t b_size,
2344                                    unsigned long b_state)
2345 {
2346         sector_t next;
2347         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2348
2349         /*
2350          * XXX Don't go larger than mballoc is willing to allocate
2351          * This is a stopgap solution.  We eventually need to fold
2352          * mpage_da_submit_io() into this function and then call
2353          * ext4_map_blocks() multiple times in a loop
2354          */
2355         if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
2356                 goto flush_it;
2357
2358         /* check if thereserved journal credits might overflow */
2359         if (!(ext4_test_inode_flag(mpd->inode, EXT4_INODE_EXTENTS))) {
2360                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2361                         /*
2362                          * With non-extent format we are limited by the journal
2363                          * credit available.  Total credit needed to insert
2364                          * nrblocks contiguous blocks is dependent on the
2365                          * nrblocks.  So limit nrblocks.
2366                          */
2367                         goto flush_it;
2368                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2369                                 EXT4_MAX_TRANS_DATA) {
2370                         /*
2371                          * Adding the new buffer_head would make it cross the
2372                          * allowed limit for which we have journal credit
2373                          * reserved. So limit the new bh->b_size
2374                          */
2375                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2376                                                 mpd->inode->i_blkbits;
2377                         /* we will do mpage_da_submit_io in the next loop */
2378                 }
2379         }
2380         /*
2381          * First block in the extent
2382          */
2383         if (mpd->b_size == 0) {
2384                 mpd->b_blocknr = logical;
2385                 mpd->b_size = b_size;
2386                 mpd->b_state = b_state & BH_FLAGS;
2387                 return;
2388         }
2389
2390         next = mpd->b_blocknr + nrblocks;
2391         /*
2392          * Can we merge the block to our big extent?
2393          */
2394         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2395                 mpd->b_size += b_size;
2396                 return;
2397         }
2398
2399 flush_it:
2400         /*
2401          * We couldn't merge the block to our extent, so we
2402          * need to flush current  extent and start new one
2403          */
2404         if (mpage_da_map_blocks(mpd) == 0)
2405                 mpage_da_submit_io(mpd);
2406         mpd->io_done = 1;
2407         return;
2408 }
2409
2410 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
2411 {
2412         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
2413 }
2414
2415 /*
2416  * __mpage_da_writepage - finds extent of pages and blocks
2417  *
2418  * @page: page to consider
2419  * @wbc: not used, we just follow rules
2420  * @data: context
2421  *
2422  * The function finds extents of pages and scan them for all blocks.
2423  */
2424 static int __mpage_da_writepage(struct page *page,
2425                                 struct writeback_control *wbc, void *data)
2426 {
2427         struct mpage_da_data *mpd = data;
2428         struct inode *inode = mpd->inode;
2429         struct buffer_head *bh, *head;
2430         sector_t logical;
2431
2432         /*
2433          * Can we merge this page to current extent?
2434          */
2435         if (mpd->next_page != page->index) {
2436                 /*
2437                  * Nope, we can't. So, we map non-allocated blocks
2438                  * and start IO on them using writepage()
2439                  */
2440                 if (mpd->next_page != mpd->first_page) {
2441                         if (mpage_da_map_blocks(mpd) == 0)
2442                                 mpage_da_submit_io(mpd);
2443                         /*
2444                          * skip rest of the page in the page_vec
2445                          */
2446                         mpd->io_done = 1;
2447                         redirty_page_for_writepage(wbc, page);
2448                         unlock_page(page);
2449                         return MPAGE_DA_EXTENT_TAIL;
2450                 }
2451
2452                 /*
2453                  * Start next extent of pages ...
2454                  */
2455                 mpd->first_page = page->index;
2456
2457                 /*
2458                  * ... and blocks
2459                  */
2460                 mpd->b_size = 0;
2461                 mpd->b_state = 0;
2462                 mpd->b_blocknr = 0;
2463         }
2464
2465         mpd->next_page = page->index + 1;
2466         logical = (sector_t) page->index <<
2467                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2468
2469         if (!page_has_buffers(page)) {
2470                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2471                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2472                 if (mpd->io_done)
2473                         return MPAGE_DA_EXTENT_TAIL;
2474         } else {
2475                 /*
2476                  * Page with regular buffer heads, just add all dirty ones
2477                  */
2478                 head = page_buffers(page);
2479                 bh = head;
2480                 do {
2481                         BUG_ON(buffer_locked(bh));
2482                         /*
2483                          * We need to try to allocate
2484                          * unmapped blocks in the same page.
2485                          * Otherwise we won't make progress
2486                          * with the page in ext4_writepage
2487                          */
2488                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2489                                 mpage_add_bh_to_extent(mpd, logical,
2490                                                        bh->b_size,
2491                                                        bh->b_state);
2492                                 if (mpd->io_done)
2493                                         return MPAGE_DA_EXTENT_TAIL;
2494                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2495                                 /*
2496                                  * mapped dirty buffer. We need to update
2497                                  * the b_state because we look at
2498                                  * b_state in mpage_da_map_blocks. We don't
2499                                  * update b_size because if we find an
2500                                  * unmapped buffer_head later we need to
2501                                  * use the b_state flag of that buffer_head.
2502                                  */
2503                                 if (mpd->b_size == 0)
2504                                         mpd->b_state = bh->b_state & BH_FLAGS;
2505                         }
2506                         logical++;
2507                 } while ((bh = bh->b_this_page) != head);
2508         }
2509
2510         return 0;
2511 }
2512
2513 /*
2514  * This is a special get_blocks_t callback which is used by
2515  * ext4_da_write_begin().  It will either return mapped block or
2516  * reserve space for a single block.
2517  *
2518  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2519  * We also have b_blocknr = -1 and b_bdev initialized properly
2520  *
2521  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2522  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2523  * initialized properly.
2524  */
2525 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2526                                   struct buffer_head *bh, int create)
2527 {
2528         struct ext4_map_blocks map;
2529         int ret = 0;
2530         sector_t invalid_block = ~((sector_t) 0xffff);
2531
2532         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2533                 invalid_block = ~0;
2534
2535         BUG_ON(create == 0);
2536         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
2537
2538         map.m_lblk = iblock;
2539         map.m_len = 1;
2540
2541         /*
2542          * first, we need to know whether the block is allocated already
2543          * preallocated blocks are unmapped but should treated
2544          * the same as allocated blocks.
2545          */
2546         ret = ext4_map_blocks(NULL, inode, &map, 0);
2547         if (ret < 0)
2548                 return ret;
2549         if (ret == 0) {
2550                 if (buffer_delay(bh))
2551                         return 0; /* Not sure this could or should happen */
2552                 /*
2553                  * XXX: __block_write_begin() unmaps passed block, is it OK?
2554                  */
2555                 ret = ext4_da_reserve_space(inode, iblock);
2556                 if (ret)
2557                         /* not enough space to reserve */
2558                         return ret;
2559
2560                 map_bh(bh, inode->i_sb, invalid_block);
2561                 set_buffer_new(bh);
2562                 set_buffer_delay(bh);
2563                 return 0;
2564         }
2565
2566         map_bh(bh, inode->i_sb, map.m_pblk);
2567         bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
2568
2569         if (buffer_unwritten(bh)) {
2570                 /* A delayed write to unwritten bh should be marked
2571                  * new and mapped.  Mapped ensures that we don't do
2572                  * get_block multiple times when we write to the same
2573                  * offset and new ensures that we do proper zero out
2574                  * for partial write.
2575                  */
2576                 set_buffer_new(bh);
2577                 set_buffer_mapped(bh);
2578         }
2579         return 0;
2580 }
2581
2582 /*
2583  * This function is used as a standard get_block_t calback function
2584  * when there is no desire to allocate any blocks.  It is used as a
2585  * callback function for block_write_begin() and block_write_full_page().
2586  * These functions should only try to map a single block at a time.
2587  *
2588  * Since this function doesn't do block allocations even if the caller
2589  * requests it by passing in create=1, it is critically important that
2590  * any caller checks to make sure that any buffer heads are returned
2591  * by this function are either all already mapped or marked for
2592  * delayed allocation before calling  block_write_full_page().  Otherwise,
2593  * b_blocknr could be left unitialized, and the page write functions will
2594  * be taken by surprise.
2595  */
2596 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2597                                    struct buffer_head *bh_result, int create)
2598 {
2599         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2600         return _ext4_get_block(inode, iblock, bh_result, 0);
2601 }
2602
2603 static int bget_one(handle_t *handle, struct buffer_head *bh)
2604 {
2605         get_bh(bh);
2606         return 0;
2607 }
2608
2609 static int bput_one(handle_t *handle, struct buffer_head *bh)
2610 {
2611         put_bh(bh);
2612         return 0;
2613 }
2614
2615 static int __ext4_journalled_writepage(struct page *page,
2616                                        unsigned int len)
2617 {
2618         struct address_space *mapping = page->mapping;
2619         struct inode *inode = mapping->host;
2620         struct buffer_head *page_bufs;
2621         handle_t *handle = NULL;
2622         int ret = 0;
2623         int err;
2624
2625         page_bufs = page_buffers(page);
2626         BUG_ON(!page_bufs);
2627         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2628         /* As soon as we unlock the page, it can go away, but we have
2629          * references to buffers so we are safe */
2630         unlock_page(page);
2631
2632         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2633         if (IS_ERR(handle)) {
2634                 ret = PTR_ERR(handle);
2635                 goto out;
2636         }
2637
2638         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2639                                 do_journal_get_write_access);
2640
2641         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2642                                 write_end_fn);
2643         if (ret == 0)
2644                 ret = err;
2645         err = ext4_journal_stop(handle);
2646         if (!ret)
2647                 ret = err;
2648
2649         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2650         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2651 out:
2652         return ret;
2653 }
2654
2655 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
2656 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
2657
2658 /*
2659  * Note that we don't need to start a transaction unless we're journaling data
2660  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2661  * need to file the inode to the transaction's list in ordered mode because if
2662  * we are writing back data added by write(), the inode is already there and if
2663  * we are writing back data modified via mmap(), noone guarantees in which
2664  * transaction the data will hit the disk. In case we are journaling data, we
2665  * cannot start transaction directly because transaction start ranks above page
2666  * lock so we have to do some magic.
2667  *
2668  * This function can get called via...
2669  *   - ext4_da_writepages after taking page lock (have journal handle)
2670  *   - journal_submit_inode_data_buffers (no journal handle)
2671  *   - shrink_page_list via pdflush (no journal handle)
2672  *   - grab_page_cache when doing write_begin (have journal handle)
2673  *
2674  * We don't do any block allocation in this function. If we have page with
2675  * multiple blocks we need to write those buffer_heads that are mapped. This
2676  * is important for mmaped based write. So if we do with blocksize 1K
2677  * truncate(f, 1024);
2678  * a = mmap(f, 0, 4096);
2679  * a[0] = 'a';
2680  * truncate(f, 4096);
2681  * we have in the page first buffer_head mapped via page_mkwrite call back
2682  * but other bufer_heads would be unmapped but dirty(dirty done via the
2683  * do_wp_page). So writepage should write the first block. If we modify
2684  * the mmap area beyond 1024 we will again get a page_fault and the
2685  * page_mkwrite callback will do the block allocation and mark the
2686  * buffer_heads mapped.
2687  *
2688  * We redirty the page if we have any buffer_heads that is either delay or
2689  * unwritten in the page.
2690  *
2691  * We can get recursively called as show below.
2692  *
2693  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2694  *              ext4_writepage()
2695  *
2696  * But since we don't do any block allocation we should not deadlock.
2697  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2698  */
2699 static int ext4_writepage(struct page *page,
2700                           struct writeback_control *wbc)
2701 {
2702         int ret = 0;
2703         loff_t size;
2704         unsigned int len;
2705         struct buffer_head *page_bufs = NULL;
2706         struct inode *inode = page->mapping->host;
2707
2708         trace_ext4_writepage(inode, page);
2709         size = i_size_read(inode);
2710         if (page->index == size >> PAGE_CACHE_SHIFT)
2711                 len = size & ~PAGE_CACHE_MASK;
2712         else
2713                 len = PAGE_CACHE_SIZE;
2714
2715         if (page_has_buffers(page)) {
2716                 page_bufs = page_buffers(page);
2717                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2718                                         ext4_bh_delay_or_unwritten)) {
2719                         /*
2720                          * We don't want to do  block allocation
2721                          * So redirty the page and return
2722                          * We may reach here when we do a journal commit
2723                          * via journal_submit_inode_data_buffers.
2724                          * If we don't have mapping block we just ignore
2725                          * them. We can also reach here via shrink_page_list
2726                          */
2727                         redirty_page_for_writepage(wbc, page);
2728                         unlock_page(page);
2729                         return 0;
2730                 }
2731         } else {
2732                 /*
2733                  * The test for page_has_buffers() is subtle:
2734                  * We know the page is dirty but it lost buffers. That means
2735                  * that at some moment in time after write_begin()/write_end()
2736                  * has been called all buffers have been clean and thus they
2737                  * must have been written at least once. So they are all
2738                  * mapped and we can happily proceed with mapping them
2739                  * and writing the page.
2740                  *
2741                  * Try to initialize the buffer_heads and check whether
2742                  * all are mapped and non delay. We don't want to
2743                  * do block allocation here.
2744                  */
2745                 ret = __block_write_begin(page, 0, len,
2746                                           noalloc_get_block_write);
2747                 if (!ret) {
2748                         page_bufs = page_buffers(page);
2749                         /* check whether all are mapped and non delay */
2750                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2751                                                 ext4_bh_delay_or_unwritten)) {
2752                                 redirty_page_for_writepage(wbc, page);
2753                                 unlock_page(page);
2754                                 return 0;
2755                         }
2756                 } else {
2757                         /*
2758                          * We can't do block allocation here
2759                          * so just redity the page and unlock
2760                          * and return
2761                          */
2762                         redirty_page_for_writepage(wbc, page);
2763                         unlock_page(page);
2764                         return 0;
2765                 }
2766                 /* now mark the buffer_heads as dirty and uptodate */
2767                 block_commit_write(page, 0, len);
2768         }
2769
2770         if (PageChecked(page) && ext4_should_journal_data(inode)) {
2771                 /*
2772                  * It's mmapped pagecache.  Add buffers and journal it.  There
2773                  * doesn't seem much point in redirtying the page here.
2774                  */
2775                 ClearPageChecked(page);
2776                 return __ext4_journalled_writepage(page, len);
2777         }
2778
2779         if (page_bufs && buffer_uninit(page_bufs)) {
2780                 ext4_set_bh_endio(page_bufs, inode);
2781                 ret = block_write_full_page_endio(page, noalloc_get_block_write,
2782                                             wbc, ext4_end_io_buffer_write);
2783         } else
2784                 ret = block_write_full_page(page, noalloc_get_block_write,
2785                                             wbc);
2786
2787         return ret;
2788 }
2789
2790 /*
2791  * This is called via ext4_da_writepages() to
2792  * calulate the total number of credits to reserve to fit
2793  * a single extent allocation into a single transaction,
2794  * ext4_da_writpeages() will loop calling this before
2795  * the block allocation.
2796  */
2797
2798 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2799 {
2800         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2801
2802         /*
2803          * With non-extent format the journal credit needed to
2804          * insert nrblocks contiguous block is dependent on
2805          * number of contiguous block. So we will limit
2806          * number of contiguous block to a sane value
2807          */
2808         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) &&
2809             (max_blocks > EXT4_MAX_TRANS_DATA))
2810                 max_blocks = EXT4_MAX_TRANS_DATA;
2811
2812         return ext4_chunk_trans_blocks(inode, max_blocks);
2813 }
2814
2815 /*
2816  * write_cache_pages_da - walk the list of dirty pages of the given
2817  * address space and call the callback function (which usually writes
2818  * the pages).
2819  *
2820  * This is a forked version of write_cache_pages().  Differences:
2821  *      Range cyclic is ignored.
2822  *      no_nrwrite_index_update is always presumed true
2823  */
2824 static int write_cache_pages_da(struct address_space *mapping,
2825                                 struct writeback_control *wbc,
2826                                 struct mpage_da_data *mpd)
2827 {
2828         int ret = 0;
2829         int done = 0;
2830         struct pagevec pvec;
2831         int nr_pages;
2832         pgoff_t index;
2833         pgoff_t end;            /* Inclusive */
2834         long nr_to_write = wbc->nr_to_write;
2835
2836         pagevec_init(&pvec, 0);
2837         index = wbc->range_start >> PAGE_CACHE_SHIFT;
2838         end = wbc->range_end >> PAGE_CACHE_SHIFT;
2839
2840         while (!done && (index <= end)) {
2841                 int i;
2842
2843                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2844                               PAGECACHE_TAG_DIRTY,
2845                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2846                 if (nr_pages == 0)
2847                         break;
2848
2849                 for (i = 0; i < nr_pages; i++) {
2850                         struct page *page = pvec.pages[i];
2851
2852                         /*
2853                          * At this point, the page may be truncated or
2854                          * invalidated (changing page->mapping to NULL), or
2855                          * even swizzled back from swapper_space to tmpfs file
2856                          * mapping. However, page->index will not change
2857                          * because we have a reference on the page.
2858                          */
2859                         if (page->index > end) {
2860                                 done = 1;
2861                                 break;
2862                         }
2863
2864                         lock_page(page);
2865
2866                         /*
2867                          * Page truncated or invalidated. We can freely skip it
2868                          * then, even for data integrity operations: the page
2869                          * has disappeared concurrently, so there could be no
2870                          * real expectation of this data interity operation
2871                          * even if there is now a new, dirty page at the same
2872                          * pagecache address.
2873                          */
2874                         if (unlikely(page->mapping != mapping)) {
2875 continue_unlock:
2876                                 unlock_page(page);
2877                                 continue;
2878                         }
2879
2880                         if (!PageDirty(page)) {
2881                                 /* someone wrote it for us */
2882                                 goto continue_unlock;
2883                         }
2884
2885                         if (PageWriteback(page)) {
2886                                 if (wbc->sync_mode != WB_SYNC_NONE)
2887                                         wait_on_page_writeback(page);
2888                                 else
2889                                         goto continue_unlock;
2890                         }
2891
2892                         BUG_ON(PageWriteback(page));
2893                         if (!clear_page_dirty_for_io(page))
2894                                 goto continue_unlock;
2895
2896                         ret = __mpage_da_writepage(page, wbc, mpd);
2897                         if (unlikely(ret)) {
2898                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2899                                         unlock_page(page);
2900                                         ret = 0;
2901                                 } else {
2902                                         done = 1;
2903                                         break;
2904                                 }
2905                         }
2906
2907                         if (nr_to_write > 0) {
2908                                 nr_to_write--;
2909                                 if (nr_to_write == 0 &&
2910                                     wbc->sync_mode == WB_SYNC_NONE) {
2911                                         /*
2912                                          * We stop writing back only if we are
2913                                          * not doing integrity sync. In case of
2914                                          * integrity sync we have to keep going
2915                                          * because someone may be concurrently
2916                                          * dirtying pages, and we might have
2917                                          * synced a lot of newly appeared dirty
2918                                          * pages, but have not synced all of the
2919                                          * old dirty pages.
2920                                          */
2921                                         done = 1;
2922                                         break;
2923                                 }
2924                         }
2925                 }
2926                 pagevec_release(&pvec);
2927                 cond_resched();
2928         }
2929         return ret;
2930 }
2931
2932
2933 static int ext4_da_writepages(struct address_space *mapping,
2934                               struct writeback_control *wbc)
2935 {
2936         pgoff_t index;
2937         int range_whole = 0;
2938         handle_t *handle = NULL;
2939         struct mpage_da_data mpd;
2940         struct inode *inode = mapping->host;
2941         int pages_written = 0;
2942         long pages_skipped;
2943         unsigned int max_pages;
2944         int range_cyclic, cycled = 1, io_done = 0;
2945         int needed_blocks, ret = 0;
2946         long desired_nr_to_write, nr_to_writebump = 0;
2947         loff_t range_start = wbc->range_start;
2948         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2949
2950         trace_ext4_da_writepages(inode, wbc);
2951
2952         /*
2953          * No pages to write? This is mainly a kludge to avoid starting
2954          * a transaction for special inodes like journal inode on last iput()
2955          * because that could violate lock ordering on umount
2956          */
2957         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2958                 return 0;
2959
2960         /*
2961          * If the filesystem has aborted, it is read-only, so return
2962          * right away instead of dumping stack traces later on that
2963          * will obscure the real source of the problem.  We test
2964          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2965          * the latter could be true if the filesystem is mounted
2966          * read-only, and in that case, ext4_da_writepages should
2967          * *never* be called, so if that ever happens, we would want
2968          * the stack trace.
2969          */
2970         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2971                 return -EROFS;
2972
2973         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2974                 range_whole = 1;
2975
2976         range_cyclic = wbc->range_cyclic;
2977         if (wbc->range_cyclic) {
2978                 index = mapping->writeback_index;
2979                 if (index)
2980                         cycled = 0;
2981                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2982                 wbc->range_end  = LLONG_MAX;
2983                 wbc->range_cyclic = 0;
2984         } else
2985                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2986
2987         /*
2988          * This works around two forms of stupidity.  The first is in
2989          * the writeback code, which caps the maximum number of pages
2990          * written to be 1024 pages.  This is wrong on multiple
2991          * levels; different architectues have a different page size,
2992          * which changes the maximum amount of data which gets
2993          * written.  Secondly, 4 megabytes is way too small.  XFS
2994          * forces this value to be 16 megabytes by multiplying
2995          * nr_to_write parameter by four, and then relies on its
2996          * allocator to allocate larger extents to make them
2997          * contiguous.  Unfortunately this brings us to the second
2998          * stupidity, which is that ext4's mballoc code only allocates
2999          * at most 2048 blocks.  So we force contiguous writes up to
3000          * the number of dirty blocks in the inode, or
3001          * sbi->max_writeback_mb_bump whichever is smaller.
3002          */
3003         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
3004         if (!range_cyclic && range_whole)
3005                 desired_nr_to_write = wbc->nr_to_write * 8;
3006         else
3007                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
3008                                                            max_pages);
3009         if (desired_nr_to_write > max_pages)
3010                 desired_nr_to_write = max_pages;
3011
3012         if (wbc->nr_to_write < desired_nr_to_write) {
3013                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
3014                 wbc->nr_to_write = desired_nr_to_write;
3015         }
3016
3017         mpd.wbc = wbc;
3018         mpd.inode = mapping->host;
3019
3020         pages_skipped = wbc->pages_skipped;
3021
3022 retry:
3023         while (!ret && wbc->nr_to_write > 0) {
3024
3025                 /*
3026                  * we  insert one extent at a time. So we need
3027                  * credit needed for single extent allocation.
3028                  * journalled mode is currently not supported
3029                  * by delalloc
3030                  */
3031                 BUG_ON(ext4_should_journal_data(inode));
3032                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
3033
3034                 /* start a new transaction*/
3035                 handle = ext4_journal_start(inode, needed_blocks);
3036                 if (IS_ERR(handle)) {
3037                         ret = PTR_ERR(handle);
3038                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
3039                                "%ld pages, ino %lu; err %d", __func__,
3040                                 wbc->nr_to_write, inode->i_ino, ret);
3041                         goto out_writepages;
3042                 }
3043
3044                 /*
3045                  * Now call __mpage_da_writepage to find the next
3046                  * contiguous region of logical blocks that need
3047                  * blocks to be allocated by ext4.  We don't actually
3048                  * submit the blocks for I/O here, even though
3049                  * write_cache_pages thinks it will, and will set the
3050                  * pages as clean for write before calling
3051                  * __mpage_da_writepage().
3052                  */
3053                 mpd.b_size = 0;
3054                 mpd.b_state = 0;
3055                 mpd.b_blocknr = 0;
3056                 mpd.first_page = 0;
3057                 mpd.next_page = 0;
3058                 mpd.io_done = 0;
3059                 mpd.pages_written = 0;
3060                 mpd.retval = 0;
3061                 ret = write_cache_pages_da(mapping, wbc, &mpd);
3062                 /*
3063                  * If we have a contiguous extent of pages and we
3064                  * haven't done the I/O yet, map the blocks and submit
3065                  * them for I/O.
3066                  */
3067                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
3068                         if (mpage_da_map_blocks(&mpd) == 0)
3069                                 mpage_da_submit_io(&mpd);
3070                         mpd.io_done = 1;
3071                         ret = MPAGE_DA_EXTENT_TAIL;
3072                 }
3073                 trace_ext4_da_write_pages(inode, &mpd);
3074                 wbc->nr_to_write -= mpd.pages_written;
3075
3076                 ext4_journal_stop(handle);
3077
3078                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
3079                         /* commit the transaction which would
3080                          * free blocks released in the transaction
3081                          * and try again
3082                          */
3083                         jbd2_journal_force_commit_nested(sbi->s_journal);
3084                         wbc->pages_skipped = pages_skipped;
3085                         ret = 0;
3086                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
3087                         /*
3088                          * got one extent now try with
3089                          * rest of the pages
3090                          */
3091                         pages_written += mpd.pages_written;
3092                         wbc->pages_skipped = pages_skipped;
3093                         ret = 0;
3094                         io_done = 1;
3095                 } else if (wbc->nr_to_write)
3096                         /*
3097                          * There is no more writeout needed
3098                          * or we requested for a noblocking writeout
3099                          * and we found the device congested
3100                          */
3101                         break;
3102         }
3103         if (!io_done && !cycled) {
3104                 cycled = 1;
3105                 index = 0;
3106                 wbc->range_start = index << PAGE_CACHE_SHIFT;
3107                 wbc->range_end  = mapping->writeback_index - 1;
3108                 goto retry;
3109         }
3110         if (pages_skipped != wbc->pages_skipped)
3111                 ext4_msg(inode->i_sb, KERN_CRIT,
3112                          "This should not happen leaving %s "
3113                          "with nr_to_write = %ld ret = %d",
3114                          __func__, wbc->nr_to_write, ret);
3115
3116         /* Update index */
3117         index += pages_written;
3118         wbc->range_cyclic = range_cyclic;
3119         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
3120                 /*
3121                  * set the writeback_index so that range_cyclic
3122                  * mode will write it back later
3123                  */
3124                 mapping->writeback_index = index;
3125
3126 out_writepages:
3127         wbc->nr_to_write -= nr_to_writebump;
3128         wbc->range_start = range_start;
3129         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
3130         return ret;
3131 }
3132
3133 #define FALL_BACK_TO_NONDELALLOC 1
3134 static int ext4_nonda_switch(struct super_block *sb)
3135 {
3136         s64 free_blocks, dirty_blocks;
3137         struct ext4_sb_info *sbi = EXT4_SB(sb);
3138
3139         /*
3140          * switch to non delalloc mode if we are running low
3141          * on free block. The free block accounting via percpu
3142          * counters can get slightly wrong with percpu_counter_batch getting
3143          * accumulated on each CPU without updating global counters
3144          * Delalloc need an accurate free block accounting. So switch
3145          * to non delalloc when we are near to error range.
3146          */
3147         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
3148         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
3149         if (2 * free_blocks < 3 * dirty_blocks ||
3150                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
3151                 /*
3152                  * free block count is less than 150% of dirty blocks
3153                  * or free blocks is less than watermark
3154                  */
3155                 return 1;
3156         }
3157         /*
3158          * Even if we don't switch but are nearing capacity,
3159          * start pushing delalloc when 1/2 of free blocks are dirty.
3160          */
3161         if (free_blocks < 2 * dirty_blocks)
3162                 writeback_inodes_sb_if_idle(sb);
3163
3164         return 0;
3165 }
3166
3167 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3168                                loff_t pos, unsigned len, unsigned flags,
3169                                struct page **pagep, void **fsdata)
3170 {
3171         int ret, retries = 0;
3172         struct page *page;
3173         pgoff_t index;
3174         struct inode *inode = mapping->host;
3175         handle_t *handle;
3176
3177         index = pos >> PAGE_CACHE_SHIFT;
3178
3179         if (ext4_nonda_switch(inode->i_sb)) {
3180                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3181                 return ext4_write_begin(file, mapping, pos,
3182                                         len, flags, pagep, fsdata);
3183         }
3184         *fsdata = (void *)0;
3185         trace_ext4_da_write_begin(inode, pos, len, flags);
3186 retry:
3187         /*
3188          * With delayed allocation, we don't log the i_disksize update
3189          * if there is delayed block allocation. But we still need
3190          * to journalling the i_disksize update if writes to the end
3191          * of file which has an already mapped buffer.
3192          */
3193         handle = ext4_journal_start(inode, 1);
3194         if (IS_ERR(handle)) {
3195                 ret = PTR_ERR(handle);
3196                 goto out;
3197         }
3198         /* We cannot recurse into the filesystem as the transaction is already
3199          * started */
3200         flags |= AOP_FLAG_NOFS;
3201
3202         page = grab_cache_page_write_begin(mapping, index, flags);
3203         if (!page) {
3204                 ext4_journal_stop(handle);
3205                 ret = -ENOMEM;
3206                 goto out;
3207         }
3208         *pagep = page;
3209
3210         ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
3211         if (ret < 0) {
3212                 unlock_page(page);
3213                 ext4_journal_stop(handle);
3214                 page_cache_release(page);
3215                 /*
3216                  * block_write_begin may have instantiated a few blocks
3217                  * outside i_size.  Trim these off again. Don't need
3218                  * i_size_read because we hold i_mutex.
3219                  */
3220                 if (pos + len > inode->i_size)
3221                         ext4_truncate_failed_write(inode);
3222         }
3223
3224         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3225                 goto retry;
3226 out:
3227         return ret;
3228 }
3229
3230 /*
3231  * Check if we should update i_disksize
3232  * when write to the end of file but not require block allocation
3233  */
3234 static int ext4_da_should_update_i_disksize(struct page *page,
3235                                             unsigned long offset)
3236 {
3237         struct buffer_head *bh;
3238         struct inode *inode = page->mapping->host;
3239         unsigned int idx;
3240         int i;
3241
3242         bh = page_buffers(page);
3243         idx = offset >> inode->i_blkbits;
3244
3245         for (i = 0; i < idx; i++)
3246                 bh = bh->b_this_page;
3247
3248         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3249                 return 0;
3250         return 1;
3251 }
3252
3253 static int ext4_da_write_end(struct file *file,
3254                              struct address_space *mapping,
3255                              loff_t pos, unsigned len, unsigned copied,
3256                              struct page *page, void *fsdata)
3257 {
3258         struct inode *inode = mapping->host;
3259         int ret = 0, ret2;
3260         handle_t *handle = ext4_journal_current_handle();
3261         loff_t new_i_size;
3262         unsigned long start, end;
3263         int write_mode = (int)(unsigned long)fsdata;
3264
3265         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
3266                 if (ext4_should_order_data(inode)) {
3267                         return ext4_ordered_write_end(file, mapping, pos,
3268                                         len, copied, page, fsdata);
3269                 } else if (ext4_should_writeback_data(inode)) {
3270                         return ext4_writeback_write_end(file, mapping, pos,
3271                                         len, copied, page, fsdata);
3272                 } else {
3273                         BUG();
3274                 }
3275         }
3276
3277         trace_ext4_da_write_end(inode, pos, len, copied);
3278         start = pos & (PAGE_CACHE_SIZE - 1);
3279         end = start + copied - 1;
3280
3281         /*
3282          * generic_write_end() will run mark_inode_dirty() if i_size
3283          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3284          * into that.
3285          */
3286
3287         new_i_size = pos + copied;
3288         if (new_i_size > EXT4_I(inode)->i_disksize) {
3289                 if (ext4_da_should_update_i_disksize(page, end)) {
3290                         down_write(&EXT4_I(inode)->i_data_sem);
3291                         if (new_i_size > EXT4_I(inode)->i_disksize) {
3292                                 /*
3293                                  * Updating i_disksize when extending file
3294                                  * without needing block allocation
3295                                  */
3296                                 if (ext4_should_order_data(inode))
3297                                         ret = ext4_jbd2_file_inode(handle,
3298                                                                    inode);
3299
3300                                 EXT4_I(inode)->i_disksize = new_i_size;
3301                         }
3302                         up_write(&EXT4_I(inode)->i_data_sem);
3303                         /* We need to mark inode dirty even if
3304                          * new_i_size is less that inode->i_size
3305                          * bu greater than i_disksize.(hint delalloc)
3306                          */
3307                         ext4_mark_inode_dirty(handle, inode);
3308                 }
3309         }
3310         ret2 = generic_write_end(file, mapping, pos, len, copied,
3311                                                         page, fsdata);
3312         copied = ret2;
3313         if (ret2 < 0)
3314                 ret = ret2;
3315         ret2 = ext4_journal_stop(handle);
3316         if (!ret)
3317                 ret = ret2;
3318
3319         return ret ? ret : copied;
3320 }
3321
3322 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
3323 {
3324         /*
3325          * Drop reserved blocks
3326          */
3327         BUG_ON(!PageLocked(page));
3328         if (!page_has_buffers(page))
3329                 goto out;
3330
3331         ext4_da_page_release_reservation(page, offset);
3332
3333 out:
3334         ext4_invalidatepage(page, offset);
3335
3336         return;
3337 }
3338
3339 /*
3340  * Force all delayed allocation blocks to be allocated for a given inode.
3341  */
3342 int ext4_alloc_da_blocks(struct inode *inode)
3343 {
3344         trace_ext4_alloc_da_blocks(inode);
3345
3346         if (!EXT4_I(inode)->i_reserved_data_blocks &&
3347             !EXT4_I(inode)->i_reserved_meta_blocks)
3348                 return 0;
3349
3350         /*
3351          * We do something simple for now.  The filemap_flush() will
3352          * also start triggering a write of the data blocks, which is
3353          * not strictly speaking necessary (and for users of
3354          * laptop_mode, not even desirable).  However, to do otherwise
3355          * would require replicating code paths in:
3356          *
3357          * ext4_da_writepages() ->
3358          *    write_cache_pages() ---> (via passed in callback function)
3359          *        __mpage_da_writepage() -->
3360          *           mpage_add_bh_to_extent()
3361          *           mpage_da_map_blocks()
3362          *
3363          * The problem is that write_cache_pages(), located in
3364          * mm/page-writeback.c, marks pages clean in preparation for
3365          * doing I/O, which is not desirable if we're not planning on
3366          * doing I/O at all.
3367          *
3368          * We could call write_cache_pages(), and then redirty all of
3369          * the pages by calling redirty_page_for_writeback() but that
3370          * would be ugly in the extreme.  So instead we would need to
3371          * replicate parts of the code in the above functions,
3372          * simplifying them becuase we wouldn't actually intend to
3373          * write out the pages, but rather only collect contiguous
3374          * logical block extents, call the multi-block allocator, and
3375          * then update the buffer heads with the block allocations.
3376          *
3377          * For now, though, we'll cheat by calling filemap_flush(),
3378          * which will map the blocks, and start the I/O, but not
3379          * actually wait for the I/O to complete.
3380          */
3381         return filemap_flush(inode->i_mapping);
3382 }
3383
3384 /*
3385  * bmap() is special.  It gets used by applications such as lilo and by
3386  * the swapper to find the on-disk block of a specific piece of data.
3387  *
3388  * Naturally, this is dangerous if the block concerned is still in the
3389  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3390  * filesystem and enables swap, then they may get a nasty shock when the
3391  * data getting swapped to that swapfile suddenly gets overwritten by
3392  * the original zero's written out previously to the journal and
3393  * awaiting writeback in the kernel's buffer cache.
3394  *
3395  * So, if we see any bmap calls here on a modified, data-journaled file,
3396  * take extra steps to flush any blocks which might be in the cache.
3397  */
3398 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3399 {
3400         struct inode *inode = mapping->host;
3401         journal_t *journal;
3402         int err;
3403
3404         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3405                         test_opt(inode->i_sb, DELALLOC)) {
3406                 /*
3407                  * With delalloc we want to sync the file
3408                  * so that we can make sure we allocate
3409                  * blocks for file
3410                  */
3411                 filemap_write_and_wait(mapping);
3412         }
3413
3414         if (EXT4_JOURNAL(inode) &&
3415             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3416                 /*
3417                  * This is a REALLY heavyweight approach, but the use of
3418                  * bmap on dirty files is expected to be extremely rare:
3419                  * only if we run lilo or swapon on a freshly made file
3420                  * do we expect this to happen.
3421                  *
3422                  * (bmap requires CAP_SYS_RAWIO so this does not
3423                  * represent an unprivileged user DOS attack --- we'd be
3424                  * in trouble if mortal users could trigger this path at
3425                  * will.)
3426                  *
3427                  * NB. EXT4_STATE_JDATA is not set on files other than
3428                  * regular files.  If somebody wants to bmap a directory
3429                  * or symlink and gets confused because the buffer
3430                  * hasn't yet been flushed to disk, they deserve
3431                  * everything they get.
3432                  */
3433
3434                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3435                 journal = EXT4_JOURNAL(inode);
3436                 jbd2_journal_lock_updates(journal);
3437                 err = jbd2_journal_flush(journal);
3438                 jbd2_journal_unlock_updates(journal);
3439
3440                 if (err)
3441                         return 0;
3442         }
3443
3444         return generic_block_bmap(mapping, block, ext4_get_block);
3445 }
3446
3447 static int ext4_readpage(struct file *file, struct page *page)
3448 {
3449         return mpage_readpage(page, ext4_get_block);
3450 }
3451
3452 static int
3453 ext4_readpages(struct file *file, struct address_space *mapping,
3454                 struct list_head *pages, unsigned nr_pages)
3455 {
3456         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3457 }
3458
3459 static void ext4_free_io_end(ext4_io_end_t *io)
3460 {
3461         BUG_ON(!io);
3462         if (io->page)
3463                 put_page(io->page);
3464         iput(io->inode);
3465         kfree(io);
3466 }
3467
3468 static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
3469 {
3470         struct buffer_head *head, *bh;
3471         unsigned int curr_off = 0;
3472
3473         if (!page_has_buffers(page))
3474                 return;
3475         head = bh = page_buffers(page);
3476         do {
3477                 if (offset <= curr_off && test_clear_buffer_uninit(bh)
3478                                         && bh->b_private) {
3479                         ext4_free_io_end(bh->b_private);
3480                         bh->b_private = NULL;
3481                         bh->b_end_io = NULL;
3482                 }
3483                 curr_off = curr_off + bh->b_size;
3484                 bh = bh->b_this_page;
3485         } while (bh != head);
3486 }
3487
3488 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3489 {
3490         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3491
3492         /*
3493          * free any io_end structure allocated for buffers to be discarded
3494          */
3495         if (ext4_should_dioread_nolock(page->mapping->host))
3496                 ext4_invalidatepage_free_endio(page, offset);
3497         /*
3498          * If it's a full truncate we just forget about the pending dirtying
3499          */
3500         if (offset == 0)
3501                 ClearPageChecked(page);
3502
3503         if (journal)
3504                 jbd2_journal_invalidatepage(journal, page, offset);
3505         else
3506                 block_invalidatepage(page, offset);
3507 }
3508
3509 static int ext4_releasepage(struct page *page, gfp_t wait)
3510 {
3511         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3512
3513         WARN_ON(PageChecked(page));
3514         if (!page_has_buffers(page))
3515                 return 0;
3516         if (journal)
3517                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3518         else
3519                 return try_to_free_buffers(page);
3520 }
3521
3522 /*
3523  * O_DIRECT for ext3 (or indirect map) based files
3524  *
3525  * If the O_DIRECT write will extend the file then add this inode to the
3526  * orphan list.  So recovery will truncate it back to the original size
3527  * if the machine crashes during the write.
3528  *
3529  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3530  * crashes then stale disk data _may_ be exposed inside the file. But current
3531  * VFS code falls back into buffered path in that case so we are safe.
3532  */
3533 static ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
3534                               const struct iovec *iov, loff_t offset,
3535                               unsigned long nr_segs)
3536 {
3537         struct file *file = iocb->ki_filp;
3538         struct inode *inode = file->f_mapping->host;
3539         struct ext4_inode_info *ei = EXT4_I(inode);
3540         handle_t *handle;
3541         ssize_t ret;
3542         int orphan = 0;
3543         size_t count = iov_length(iov, nr_segs);
3544         int retries = 0;
3545
3546         if (rw == WRITE) {
3547                 loff_t final_size = offset + count;
3548
3549                 if (final_size > inode->i_size) {
3550                         /* Credits for sb + inode write */
3551                         handle = ext4_journal_start(inode, 2);
3552                         if (IS_ERR(handle)) {
3553                                 ret = PTR_ERR(handle);
3554                                 goto out;
3555                         }
3556                         ret = ext4_orphan_add(handle, inode);
3557                         if (ret) {
3558                                 ext4_journal_stop(handle);
3559                                 goto out;
3560                         }
3561                         orphan = 1;
3562                         ei->i_disksize = inode->i_size;
3563                         ext4_journal_stop(handle);
3564                 }
3565         }
3566
3567 retry:
3568         if (rw == READ && ext4_should_dioread_nolock(inode))
3569                 ret = __blockdev_direct_IO(rw, iocb, inode,
3570                                  inode->i_sb->s_bdev, iov,
3571                                  offset, nr_segs,
3572                                  ext4_get_block, NULL, NULL, 0);
3573         else {
3574                 ret = blockdev_direct_IO(rw, iocb, inode,
3575                                  inode->i_sb->s_bdev, iov,
3576                                  offset, nr_segs,
3577                                  ext4_get_block, NULL);
3578
3579                 if (unlikely((rw & WRITE) && ret < 0)) {
3580                         loff_t isize = i_size_read(inode);
3581                         loff_t end = offset + iov_length(iov, nr_segs);
3582
3583                         if (end > isize)
3584                                 vmtruncate(inode, isize);
3585                 }
3586         }
3587         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3588                 goto retry;
3589
3590         if (orphan) {
3591                 int err;
3592
3593                 /* Credits for sb + inode write */
3594                 handle = ext4_journal_start(inode, 2);
3595                 if (IS_ERR(handle)) {
3596                         /* This is really bad luck. We've written the data
3597                          * but cannot extend i_size. Bail out and pretend
3598                          * the write failed... */
3599                         ret = PTR_ERR(handle);
3600                         if (inode->i_nlink)
3601                                 ext4_orphan_del(NULL, inode);
3602
3603                         goto out;
3604                 }
3605                 if (inode->i_nlink)
3606                         ext4_orphan_del(handle, inode);
3607                 if (ret > 0) {
3608                         loff_t end = offset + ret;
3609                         if (end > inode->i_size) {
3610                                 ei->i_disksize = end;
3611                                 i_size_write(inode, end);
3612                                 /*
3613                                  * We're going to return a positive `ret'
3614                                  * here due to non-zero-length I/O, so there's
3615                                  * no way of reporting error returns from
3616                                  * ext4_mark_inode_dirty() to userspace.  So
3617                                  * ignore it.
3618                                  */
3619                                 ext4_mark_inode_dirty(handle, inode);
3620                         }
3621                 }
3622                 err = ext4_journal_stop(handle);
3623                 if (ret == 0)
3624                         ret = err;
3625         }
3626 out:
3627         return ret;
3628 }
3629
3630 /*
3631  * ext4_get_block used when preparing for a DIO write or buffer write.
3632  * We allocate an uinitialized extent if blocks haven't been allocated.
3633  * The extent will be converted to initialized after the IO is complete.
3634  */
3635 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
3636                    struct buffer_head *bh_result, int create)
3637 {
3638         ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
3639                    inode->i_ino, create);
3640         return _ext4_get_block(inode, iblock, bh_result,
3641                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
3642 }
3643
3644 static void dump_completed_IO(struct inode * inode)
3645 {
3646 #ifdef  EXT4_DEBUG
3647         struct list_head *cur, *before, *after;
3648         ext4_io_end_t *io, *io0, *io1;
3649         unsigned long flags;
3650
3651         if (list_empty(&EXT4_I(inode)->i_completed_io_list)){
3652                 ext4_debug("inode %lu completed_io list is empty\n", inode->i_ino);
3653                 return;
3654         }
3655
3656         ext4_debug("Dump inode %lu completed_io list \n", inode->i_ino);
3657         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3658         list_for_each_entry(io, &EXT4_I(inode)->i_completed_io_list, list){
3659                 cur = &io->list;
3660                 before = cur->prev;
3661                 io0 = container_of(before, ext4_io_end_t, list);
3662                 after = cur->next;
3663                 io1 = container_of(after, ext4_io_end_t, list);
3664
3665                 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
3666                             io, inode->i_ino, io0, io1);
3667         }
3668         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3669 #endif
3670 }
3671
3672 /*
3673  * check a range of space and convert unwritten extents to written.
3674  */
3675 static int ext4_end_io_nolock(ext4_io_end_t *io)
3676 {
3677         struct inode *inode = io->inode;
3678         loff_t offset = io->offset;
3679         ssize_t size = io->size;
3680         int ret = 0;
3681
3682         ext4_debug("ext4_end_io_nolock: io 0x%p from inode %lu,list->next 0x%p,"
3683                    "list->prev 0x%p\n",
3684                    io, inode->i_ino, io->list.next, io->list.prev);
3685
3686         if (list_empty(&io->list))
3687                 return ret;
3688
3689         if (io->flag != EXT4_IO_UNWRITTEN)
3690                 return ret;
3691
3692         ret = ext4_convert_unwritten_extents(inode, offset, size);
3693         if (ret < 0) {
3694                 printk(KERN_EMERG "%s: failed to convert unwritten"
3695                         "extents to written extents, error is %d"
3696                         " io is still on inode %lu aio dio list\n",
3697                        __func__, ret, inode->i_ino);
3698                 return ret;
3699         }
3700
3701         if (io->iocb)
3702                 aio_complete(io->iocb, io->result, 0);
3703         /* clear the DIO AIO unwritten flag */
3704         io->flag = 0;
3705         return ret;
3706 }
3707
3708 /*
3709  * work on completed aio dio IO, to convert unwritten extents to extents
3710  */
3711 static void ext4_end_io_work(struct work_struct *work)
3712 {
3713         ext4_io_end_t           *io = container_of(work, ext4_io_end_t, work);
3714         struct inode            *inode = io->inode;
3715         struct ext4_inode_info  *ei = EXT4_I(inode);
3716         unsigned long           flags;
3717         int                     ret;
3718
3719         mutex_lock(&inode->i_mutex);
3720         ret = ext4_end_io_nolock(io);
3721         if (ret < 0) {
3722                 mutex_unlock(&inode->i_mutex);
3723                 return;
3724         }
3725
3726         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3727         if (!list_empty(&io->list))
3728                 list_del_init(&io->list);
3729         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3730         mutex_unlock(&inode->i_mutex);
3731         ext4_free_io_end(io);
3732 }
3733
3734 /*
3735  * This function is called from ext4_sync_file().
3736  *
3737  * When IO is completed, the work to convert unwritten extents to
3738  * written is queued on workqueue but may not get immediately
3739  * scheduled. When fsync is called, we need to ensure the
3740  * conversion is complete before fsync returns.
3741  * The inode keeps track of a list of pending/completed IO that
3742  * might needs to do the conversion. This function walks through
3743  * the list and convert the related unwritten extents for completed IO
3744  * to written.
3745  * The function return the number of pending IOs on success.
3746  */
3747 int flush_completed_IO(struct inode *inode)
3748 {
3749         ext4_io_end_t *io;
3750         struct ext4_inode_info *ei = EXT4_I(inode);
3751         unsigned long flags;
3752         int ret = 0;
3753         int ret2 = 0;
3754
3755         if (list_empty(&ei->i_completed_io_list))
3756                 return ret;
3757
3758         dump_completed_IO(inode);
3759         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3760         while (!list_empty(&ei->i_completed_io_list)){
3761                 io = list_entry(ei->i_completed_io_list.next,
3762                                 ext4_io_end_t, list);
3763                 /*
3764                  * Calling ext4_end_io_nolock() to convert completed
3765                  * IO to written.
3766                  *
3767                  * When ext4_sync_file() is called, run_queue() may already
3768                  * about to flush the work corresponding to this io structure.
3769                  * It will be upset if it founds the io structure related
3770                  * to the work-to-be schedule is freed.
3771                  *
3772                  * Thus we need to keep the io structure still valid here after
3773                  * convertion finished. The io structure has a flag to
3774                  * avoid double converting from both fsync and background work
3775                  * queue work.
3776                  */
3777                 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3778                 ret = ext4_end_io_nolock(io);
3779                 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3780                 if (ret < 0)
3781                         ret2 = ret;
3782                 else
3783                         list_del_init(&io->list);
3784         }
3785         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3786         return (ret2 < 0) ? ret2 : 0;
3787 }
3788
3789 static ext4_io_end_t *ext4_init_io_end (struct inode *inode, gfp_t flags)
3790 {
3791         ext4_io_end_t *io = NULL;
3792
3793         io = kmalloc(sizeof(*io), flags);
3794
3795         if (io) {
3796                 igrab(inode);
3797                 io->inode = inode;
3798                 io->flag = 0;
3799                 io->offset = 0;
3800                 io->size = 0;
3801                 io->page = NULL;
3802                 io->iocb = NULL;
3803                 io->result = 0;
3804                 INIT_WORK(&io->work, ext4_end_io_work);
3805                 INIT_LIST_HEAD(&io->list);
3806         }
3807
3808         return io;
3809 }
3810
3811 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3812                             ssize_t size, void *private, int ret,
3813                             bool is_async)
3814 {
3815         ext4_io_end_t *io_end = iocb->private;
3816         struct workqueue_struct *wq;
3817         unsigned long flags;
3818         struct ext4_inode_info *ei;
3819
3820         /* if not async direct IO or dio with 0 bytes write, just return */
3821         if (!io_end || !size)
3822                 goto out;
3823
3824         ext_debug("ext4_end_io_dio(): io_end 0x%p"
3825                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
3826                   iocb->private, io_end->inode->i_ino, iocb, offset,
3827                   size);
3828
3829         /* if not aio dio with unwritten extents, just free io and return */
3830         if (io_end->flag != EXT4_IO_UNWRITTEN){
3831                 ext4_free_io_end(io_end);
3832                 iocb->private = NULL;
3833 out:
3834                 if (is_async)
3835                         aio_complete(iocb, ret, 0);
3836                 return;
3837         }
3838
3839         io_end->offset = offset;
3840         io_end->size = size;
3841         if (is_async) {
3842                 io_end->iocb = iocb;
3843                 io_end->result = ret;
3844         }
3845         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
3846
3847         /* queue the work to convert unwritten extents to written */
3848         queue_work(wq, &io_end->work);
3849
3850         /* Add the io_end to per-inode completed aio dio list*/
3851         ei = EXT4_I(io_end->inode);
3852         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3853         list_add_tail(&io_end->list, &ei->i_completed_io_list);
3854         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3855         iocb->private = NULL;
3856 }
3857
3858 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
3859 {
3860         ext4_io_end_t *io_end = bh->b_private;
3861         struct workqueue_struct *wq;
3862         struct inode *inode;
3863         unsigned long flags;
3864
3865         if (!test_clear_buffer_uninit(bh) || !io_end)
3866                 goto out;
3867
3868         if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
3869                 printk("sb umounted, discard end_io request for inode %lu\n",
3870                         io_end->inode->i_ino);
3871                 ext4_free_io_end(io_end);
3872                 goto out;
3873         }
3874
3875         io_end->flag = EXT4_IO_UNWRITTEN;
3876         inode = io_end->inode;
3877
3878         /* Add the io_end to per-inode completed io list*/
3879         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3880         list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
3881         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3882
3883         wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
3884         /* queue the work to convert unwritten extents to written */
3885         queue_work(wq, &io_end->work);
3886 out:
3887         bh->b_private = NULL;
3888         bh->b_end_io = NULL;
3889         clear_buffer_uninit(bh);
3890         end_buffer_async_write(bh, uptodate);
3891 }
3892
3893 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
3894 {
3895         ext4_io_end_t *io_end;
3896         struct page *page = bh->b_page;
3897         loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
3898         size_t size = bh->b_size;
3899
3900 retry:
3901         io_end = ext4_init_io_end(inode, GFP_ATOMIC);
3902         if (!io_end) {
3903                 if (printk_ratelimit())
3904                         printk(KERN_WARNING "%s: allocation fail\n", __func__);
3905                 schedule();
3906                 goto retry;
3907         }
3908         io_end->offset = offset;
3909         io_end->size = size;
3910         /*
3911          * We need to hold a reference to the page to make sure it
3912          * doesn't get evicted before ext4_end_io_work() has a chance
3913          * to convert the extent from written to unwritten.
3914          */
3915         io_end->page = page;
3916         get_page(io_end->page);
3917
3918         bh->b_private = io_end;
3919         bh->b_end_io = ext4_end_io_buffer_write;
3920         return 0;
3921 }
3922
3923 /*
3924  * For ext4 extent files, ext4 will do direct-io write to holes,
3925  * preallocated extents, and those write extend the file, no need to
3926  * fall back to buffered IO.
3927  *
3928  * For holes, we fallocate those blocks, mark them as unintialized
3929  * If those blocks were preallocated, we mark sure they are splited, but
3930  * still keep the range to write as unintialized.
3931  *
3932  * The unwrritten extents will be converted to written when DIO is completed.
3933  * For async direct IO, since the IO may still pending when return, we
3934  * set up an end_io call back function, which will do the convertion
3935  * when async direct IO completed.
3936  *
3937  * If the O_DIRECT write will extend the file then add this inode to the
3938  * orphan list.  So recovery will truncate it back to the original size
3939  * if the machine crashes during the write.
3940  *
3941  */
3942 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
3943                               const struct iovec *iov, loff_t offset,
3944                               unsigned long nr_segs)
3945 {
3946         struct file *file = iocb->ki_filp;
3947         struct inode *inode = file->f_mapping->host;
3948         ssize_t ret;
3949         size_t count = iov_length(iov, nr_segs);
3950
3951         loff_t final_size = offset + count;
3952         if (rw == WRITE && final_size <= inode->i_size) {
3953                 /*
3954                  * We could direct write to holes and fallocate.
3955                  *
3956                  * Allocated blocks to fill the hole are marked as uninitialized
3957                  * to prevent paralel buffered read to expose the stale data
3958                  * before DIO complete the data IO.
3959                  *
3960                  * As to previously fallocated extents, ext4 get_block
3961                  * will just simply mark the buffer mapped but still
3962                  * keep the extents uninitialized.
3963                  *
3964                  * for non AIO case, we will convert those unwritten extents
3965                  * to written after return back from blockdev_direct_IO.
3966                  *
3967                  * for async DIO, the conversion needs to be defered when
3968                  * the IO is completed. The ext4 end_io callback function
3969                  * will be called to take care of the conversion work.
3970                  * Here for async case, we allocate an io_end structure to
3971                  * hook to the iocb.
3972                  */
3973                 iocb->private = NULL;
3974                 EXT4_I(inode)->cur_aio_dio = NULL;
3975                 if (!is_sync_kiocb(iocb)) {
3976                         iocb->private = ext4_init_io_end(inode, GFP_NOFS);
3977                         if (!iocb->private)
3978                                 return -ENOMEM;
3979                         /*
3980                          * we save the io structure for current async
3981                          * direct IO, so that later ext4_map_blocks()
3982                          * could flag the io structure whether there
3983                          * is a unwritten extents needs to be converted
3984                          * when IO is completed.
3985                          */
3986                         EXT4_I(inode)->cur_aio_dio = iocb->private;
3987                 }
3988
3989                 ret = blockdev_direct_IO(rw, iocb, inode,
3990                                          inode->i_sb->s_bdev, iov,
3991                                          offset, nr_segs,
3992                                          ext4_get_block_write,
3993                                          ext4_end_io_dio);
3994                 if (iocb->private)
3995                         EXT4_I(inode)->cur_aio_dio = NULL;
3996                 /*
3997                  * The io_end structure takes a reference to the inode,
3998                  * that structure needs to be destroyed and the
3999                  * reference to the inode need to be dropped, when IO is
4000                  * complete, even with 0 byte write, or failed.
4001                  *
4002                  * In the successful AIO DIO case, the io_end structure will be
4003                  * desctroyed and the reference to the inode will be dropped
4004                  * after the end_io call back function is called.
4005                  *
4006                  * In the case there is 0 byte write, or error case, since
4007                  * VFS direct IO won't invoke the end_io call back function,
4008                  * we need to free the end_io structure here.
4009                  */
4010                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
4011                         ext4_free_io_end(iocb->private);
4012                         iocb->private = NULL;
4013                 } else if (ret > 0 && ext4_test_inode_state(inode,
4014                                                 EXT4_STATE_DIO_UNWRITTEN)) {
4015                         int err;
4016                         /*
4017                          * for non AIO case, since the IO is already
4018                          * completed, we could do the convertion right here
4019                          */
4020                         err = ext4_convert_unwritten_extents(inode,
4021                                                              offset, ret);
4022                         if (err < 0)
4023                                 ret = err;
4024                         ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
4025                 }
4026                 return ret;
4027         }
4028
4029         /* for write the the end of file case, we fall back to old way */
4030         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
4031 }
4032
4033 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
4034                               const struct iovec *iov, loff_t offset,
4035                               unsigned long nr_segs)
4036 {
4037         struct file *file = iocb->ki_filp;
4038         struct inode *inode = file->f_mapping->host;
4039
4040         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4041                 return ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
4042
4043         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
4044 }
4045
4046 /*
4047  * Pages can be marked dirty completely asynchronously from ext4's journalling
4048  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
4049  * much here because ->set_page_dirty is called under VFS locks.  The page is
4050  * not necessarily locked.
4051  *
4052  * We cannot just dirty the page and leave attached buffers clean, because the
4053  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
4054  * or jbddirty because all the journalling code will explode.
4055  *
4056  * So what we do is to mark the page "pending dirty" and next time writepage
4057  * is called, propagate that into the buffers appropriately.
4058  */
4059 static int ext4_journalled_set_page_dirty(struct page *page)
4060 {
4061         SetPageChecked(page);
4062         return __set_page_dirty_nobuffers(page);
4063 }
4064
4065 static const struct address_space_operations ext4_ordered_aops = {
4066         .readpage               = ext4_readpage,
4067         .readpages              = ext4_readpages,
4068         .writepage              = ext4_writepage,
4069         .sync_page              = block_sync_page,
4070         .write_begin            = ext4_write_begin,
4071         .write_end              = ext4_ordered_write_end,
4072         .bmap                   = ext4_bmap,
4073         .invalidatepage         = ext4_invalidatepage,
4074         .releasepage            = ext4_releasepage,
4075         .direct_IO              = ext4_direct_IO,
4076         .migratepage            = buffer_migrate_page,
4077         .is_partially_uptodate  = block_is_partially_uptodate,
4078         .error_remove_page      = generic_error_remove_page,
4079 };
4080
4081 static const struct address_space_operations ext4_writeback_aops = {
4082         .readpage               = ext4_readpage,
4083         .readpages              = ext4_readpages,
4084         .writepage              = ext4_writepage,
4085         .sync_page              = block_sync_page,
4086         .write_begin            = ext4_write_begin,
4087         .write_end              = ext4_writeback_write_end,
4088         .bmap                   = ext4_bmap,
4089         .invalidatepage         = ext4_invalidatepage,
4090         .releasepage            = ext4_releasepage,
4091         .direct_IO              = ext4_direct_IO,
4092         .migratepage            = buffer_migrate_page,
4093         .is_partially_uptodate  = block_is_partially_uptodate,
4094         .error_remove_page      = generic_error_remove_page,
4095 };
4096
4097 static const struct address_space_operations ext4_journalled_aops = {
4098         .readpage               = ext4_readpage,
4099         .readpages              = ext4_readpages,
4100         .writepage              = ext4_writepage,
4101         .sync_page              = block_sync_page,
4102         .write_begin            = ext4_write_begin,
4103         .write_end              = ext4_journalled_write_end,
4104         .set_page_dirty         = ext4_journalled_set_page_dirty,
4105         .bmap                   = ext4_bmap,
4106         .invalidatepage         = ext4_invalidatepage,
4107         .releasepage            = ext4_releasepage,
4108         .is_partially_uptodate  = block_is_partially_uptodate,
4109         .error_remove_page      = generic_error_remove_page,
4110 };
4111
4112 static const struct address_space_operations ext4_da_aops = {
4113         .readpage               = ext4_readpage,
4114         .readpages              = ext4_readpages,
4115         .writepage              = ext4_writepage,
4116         .writepages             = ext4_da_writepages,
4117         .sync_page              = block_sync_page,
4118         .write_begin            = ext4_da_write_begin,
4119         .write_end              = ext4_da_write_end,
4120         .bmap                   = ext4_bmap,
4121         .invalidatepage         = ext4_da_invalidatepage,
4122         .releasepage            = ext4_releasepage,
4123         .direct_IO              = ext4_direct_IO,
4124         .migratepage            = buffer_migrate_page,
4125         .is_partially_uptodate  = block_is_partially_uptodate,
4126         .error_remove_page      = generic_error_remove_page,
4127 };
4128
4129 void ext4_set_aops(struct inode *inode)
4130 {
4131         if (ext4_should_order_data(inode) &&
4132                 test_opt(inode->i_sb, DELALLOC))
4133                 inode->i_mapping->a_ops = &ext4_da_aops;
4134         else if (ext4_should_order_data(inode))
4135                 inode->i_mapping->a_ops = &ext4_ordered_aops;
4136         else if (ext4_should_writeback_data(inode) &&
4137                  test_opt(inode->i_sb, DELALLOC))
4138                 inode->i_mapping->a_ops = &ext4_da_aops;
4139         else if (ext4_should_writeback_data(inode))
4140                 inode->i_mapping->a_ops = &ext4_writeback_aops;
4141         else
4142                 inode->i_mapping->a_ops = &ext4_journalled_aops;
4143 }
4144
4145 /*
4146  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
4147  * up to the end of the block which corresponds to `from'.
4148  * This required during truncate. We need to physically zero the tail end
4149  * of that block so it doesn't yield old data if the file is later grown.
4150  */
4151 int ext4_block_truncate_page(handle_t *handle,
4152                 struct address_space *mapping, loff_t from)
4153 {
4154         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
4155         unsigned offset = from & (PAGE_CACHE_SIZE-1);
4156         unsigned blocksize, length, pos;
4157         ext4_lblk_t iblock;
4158         struct inode *inode = mapping->host;
4159         struct buffer_head *bh;
4160         struct page *page;
4161         int err = 0;
4162
4163         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
4164                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
4165         if (!page)
4166                 return -EINVAL;
4167
4168         blocksize = inode->i_sb->s_blocksize;
4169         length = blocksize - (offset & (blocksize - 1));
4170         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
4171
4172         if (!page_has_buffers(page))
4173                 create_empty_buffers(page, blocksize, 0);
4174
4175         /* Find the buffer that contains "offset" */
4176         bh = page_buffers(page);
4177         pos = blocksize;
4178         while (offset >= pos) {
4179                 bh = bh->b_this_page;
4180                 iblock++;
4181                 pos += blocksize;
4182         }
4183
4184         err = 0;
4185         if (buffer_freed(bh)) {
4186                 BUFFER_TRACE(bh, "freed: skip");
4187                 goto unlock;
4188         }
4189
4190         if (!buffer_mapped(bh)) {
4191                 BUFFER_TRACE(bh, "unmapped");
4192                 ext4_get_block(inode, iblock, bh, 0);
4193                 /* unmapped? It's a hole - nothing to do */
4194                 if (!buffer_mapped(bh)) {
4195                         BUFFER_TRACE(bh, "still unmapped");
4196                         goto unlock;
4197                 }
4198         }
4199
4200         /* Ok, it's mapped. Make sure it's up-to-date */
4201         if (PageUptodate(page))
4202                 set_buffer_uptodate(bh);
4203
4204         if (!buffer_uptodate(bh)) {
4205                 err = -EIO;
4206                 ll_rw_block(READ, 1, &bh);
4207                 wait_on_buffer(bh);
4208                 /* Uhhuh. Read error. Complain and punt. */
4209                 if (!buffer_uptodate(bh))
4210                         goto unlock;
4211         }
4212
4213         if (ext4_should_journal_data(inode)) {
4214                 BUFFER_TRACE(bh, "get write access");
4215                 err = ext4_journal_get_write_access(handle, bh);
4216                 if (err)
4217                         goto unlock;
4218         }
4219
4220         zero_user(page, offset, length);
4221
4222         BUFFER_TRACE(bh, "zeroed end of block");
4223
4224         err = 0;
4225         if (ext4_should_journal_data(inode)) {
4226                 err = ext4_handle_dirty_metadata(handle, inode, bh);
4227         } else {
4228                 if (ext4_should_order_data(inode))
4229                         err = ext4_jbd2_file_inode(handle, inode);
4230                 mark_buffer_dirty(bh);
4231         }
4232
4233 unlock:
4234         unlock_page(page);
4235         page_cache_release(page);
4236         return err;
4237 }
4238
4239 /*
4240  * Probably it should be a library function... search for first non-zero word
4241  * or memcmp with zero_page, whatever is better for particular architecture.
4242  * Linus?
4243  */
4244 static inline int all_zeroes(__le32 *p, __le32 *q)
4245 {
4246         while (p < q)
4247                 if (*p++)
4248                         return 0;
4249         return 1;
4250 }
4251
4252 /**
4253  *      ext4_find_shared - find the indirect blocks for partial truncation.
4254  *      @inode:   inode in question
4255  *      @depth:   depth of the affected branch
4256  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
4257  *      @chain:   place to store the pointers to partial indirect blocks
4258  *      @top:     place to the (detached) top of branch
4259  *
4260  *      This is a helper function used by ext4_truncate().
4261  *
4262  *      When we do truncate() we may have to clean the ends of several
4263  *      indirect blocks but leave the blocks themselves alive. Block is
4264  *      partially truncated if some data below the new i_size is refered
4265  *      from it (and it is on the path to the first completely truncated
4266  *      data block, indeed).  We have to free the top of that path along
4267  *      with everything to the right of the path. Since no allocation
4268  *      past the truncation point is possible until ext4_truncate()
4269  *      finishes, we may safely do the latter, but top of branch may
4270  *      require special attention - pageout below the truncation point
4271  *      might try to populate it.
4272  *
4273  *      We atomically detach the top of branch from the tree, store the
4274  *      block number of its root in *@top, pointers to buffer_heads of
4275  *      partially truncated blocks - in @chain[].bh and pointers to
4276  *      their last elements that should not be removed - in
4277  *      @chain[].p. Return value is the pointer to last filled element
4278  *      of @chain.
4279  *
4280  *      The work left to caller to do the actual freeing of subtrees:
4281  *              a) free the subtree starting from *@top
4282  *              b) free the subtrees whose roots are stored in
4283  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
4284  *              c) free the subtrees growing from the inode past the @chain[0].
4285  *                      (no partially truncated stuff there).  */
4286
4287 static Indirect *ext4_find_shared(struct inode *inode, int depth,
4288                                   ext4_lblk_t offsets[4], Indirect chain[4],
4289                                   __le32 *top)
4290 {
4291         Indirect *partial, *p;
4292         int k, err;
4293
4294         *top = 0;
4295         /* Make k index the deepest non-null offset + 1 */
4296         for (k = depth; k > 1 && !offsets[k-1]; k--)
4297                 ;
4298         partial = ext4_get_branch(inode, k, offsets, chain, &err);
4299         /* Writer: pointers */
4300         if (!partial)
4301                 partial = chain + k-1;
4302         /*
4303          * If the branch acquired continuation since we've looked at it -
4304          * fine, it should all survive and (new) top doesn't belong to us.
4305          */
4306         if (!partial->key && *partial->p)
4307                 /* Writer: end */
4308                 goto no_top;
4309         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
4310                 ;
4311         /*
4312          * OK, we've found the last block that must survive. The rest of our
4313          * branch should be detached before unlocking. However, if that rest
4314          * of branch is all ours and does not grow immediately from the inode
4315          * it's easier to cheat and just decrement partial->p.
4316          */
4317         if (p == chain + k - 1 && p > chain) {
4318                 p->p--;
4319         } else {
4320                 *top = *p->p;
4321                 /* Nope, don't do this in ext4.  Must leave the tree intact */
4322 #if 0
4323                 *p->p = 0;
4324 #endif
4325         }
4326         /* Writer: end */
4327
4328         while (partial > p) {
4329                 brelse(partial->bh);
4330                 partial--;
4331         }
4332 no_top:
4333         return partial;
4334 }
4335
4336 /*
4337  * Zero a number of block pointers in either an inode or an indirect block.
4338  * If we restart the transaction we must again get write access to the
4339  * indirect block for further modification.
4340  *
4341  * We release `count' blocks on disk, but (last - first) may be greater
4342  * than `count' because there can be holes in there.
4343  */
4344 static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
4345                              struct buffer_head *bh,
4346                              ext4_fsblk_t block_to_free,
4347                              unsigned long count, __le32 *first,
4348                              __le32 *last)
4349 {
4350         __le32 *p;
4351         int     flags = EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_VALIDATED;
4352
4353         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
4354                 flags |= EXT4_FREE_BLOCKS_METADATA;
4355
4356         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
4357                                    count)) {
4358                 EXT4_ERROR_INODE(inode, "attempt to clear invalid "
4359                                  "blocks %llu len %lu",
4360                                  (unsigned long long) block_to_free, count);
4361                 return 1;
4362         }
4363
4364         if (try_to_extend_transaction(handle, inode)) {
4365                 if (bh) {
4366                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4367                         ext4_handle_dirty_metadata(handle, inode, bh);
4368                 }
4369                 ext4_mark_inode_dirty(handle, inode);
4370                 ext4_truncate_restart_trans(handle, inode,
4371                                             blocks_for_truncate(inode));
4372                 if (bh) {
4373                         BUFFER_TRACE(bh, "retaking write access");
4374                         ext4_journal_get_write_access(handle, bh);
4375                 }
4376         }
4377
4378         for (p = first; p < last; p++)
4379                 *p = 0;
4380
4381         ext4_free_blocks(handle, inode, 0, block_to_free, count, flags);
4382         return 0;
4383 }
4384
4385 /**
4386  * ext4_free_data - free a list of data blocks
4387  * @handle:     handle for this transaction
4388  * @inode:      inode we are dealing with
4389  * @this_bh:    indirect buffer_head which contains *@first and *@last
4390  * @first:      array of block numbers
4391  * @last:       points immediately past the end of array
4392  *
4393  * We are freeing all blocks refered from that array (numbers are stored as
4394  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
4395  *
4396  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
4397  * blocks are contiguous then releasing them at one time will only affect one
4398  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
4399  * actually use a lot of journal space.
4400  *
4401  * @this_bh will be %NULL if @first and @last point into the inode's direct
4402  * block pointers.
4403  */
4404 static void ext4_free_data(handle_t *handle, struct inode *inode,
4405                            struct buffer_head *this_bh,
4406                            __le32 *first, __le32 *last)
4407 {
4408         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
4409         unsigned long count = 0;            /* Number of blocks in the run */
4410         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
4411                                                corresponding to
4412                                                block_to_free */
4413         ext4_fsblk_t nr;                    /* Current block # */
4414         __le32 *p;                          /* Pointer into inode/ind
4415                                                for current block */
4416         int err;
4417
4418         if (this_bh) {                          /* For indirect block */
4419                 BUFFER_TRACE(this_bh, "get_write_access");
4420                 err = ext4_journal_get_write_access(handle, this_bh);
4421                 /* Important: if we can't update the indirect pointers
4422                  * to the blocks, we can't free them. */
4423                 if (err)
4424                         return;
4425         }
4426
4427         for (p = first; p < last; p++) {
4428                 nr = le32_to_cpu(*p);
4429                 if (nr) {
4430                         /* accumulate blocks to free if they're contiguous */
4431                         if (count == 0) {
4432                                 block_to_free = nr;
4433                                 block_to_free_p = p;
4434                                 count = 1;
4435                         } else if (nr == block_to_free + count) {
4436                                 count++;
4437                         } else {
4438                                 if (ext4_clear_blocks(handle, inode, this_bh,
4439                                                       block_to_free, count,
4440                                                       block_to_free_p, p))
4441                                         break;
4442                                 block_to_free = nr;
4443                                 block_to_free_p = p;
4444                                 count = 1;
4445                         }
4446                 }
4447         }
4448
4449         if (count > 0)
4450                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
4451                                   count, block_to_free_p, p);
4452
4453         if (this_bh) {
4454                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
4455
4456                 /*
4457                  * The buffer head should have an attached journal head at this
4458                  * point. However, if the data is corrupted and an indirect
4459                  * block pointed to itself, it would have been detached when
4460                  * the block was cleared. Check for this instead of OOPSing.
4461                  */
4462                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
4463                         ext4_handle_dirty_metadata(handle, inode, this_bh);
4464                 else
4465                         EXT4_ERROR_INODE(inode,
4466                                          "circular indirect block detected at "
4467                                          "block %llu",
4468                                 (unsigned long long) this_bh->b_blocknr);
4469         }
4470 }
4471
4472 /**
4473  *      ext4_free_branches - free an array of branches
4474  *      @handle: JBD handle for this transaction
4475  *      @inode: inode we are dealing with
4476  *      @parent_bh: the buffer_head which contains *@first and *@last
4477  *      @first: array of block numbers
4478  *      @last:  pointer immediately past the end of array
4479  *      @depth: depth of the branches to free
4480  *
4481  *      We are freeing all blocks refered from these branches (numbers are
4482  *      stored as little-endian 32-bit) and updating @inode->i_blocks
4483  *      appropriately.
4484  */
4485 static void ext4_free_branches(handle_t *handle, struct inode *inode,
4486                                struct buffer_head *parent_bh,
4487                                __le32 *first, __le32 *last, int depth)
4488 {
4489         ext4_fsblk_t nr;
4490         __le32 *p;
4491
4492         if (ext4_handle_is_aborted(handle))
4493                 return;
4494
4495         if (depth--) {
4496                 struct buffer_head *bh;
4497                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4498                 p = last;
4499                 while (--p >= first) {
4500                         nr = le32_to_cpu(*p);
4501                         if (!nr)
4502                                 continue;               /* A hole */
4503
4504                         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
4505                                                    nr, 1)) {
4506                                 EXT4_ERROR_INODE(inode,
4507                                                  "invalid indirect mapped "
4508                                                  "block %lu (level %d)",
4509                                                  (unsigned long) nr, depth);
4510                                 break;
4511                         }
4512
4513                         /* Go read the buffer for the next level down */
4514                         bh = sb_bread(inode->i_sb, nr);
4515
4516                         /*
4517                          * A read failure? Report error and clear slot
4518                          * (should be rare).
4519                          */
4520                         if (!bh) {
4521                                 EXT4_ERROR_INODE_BLOCK(inode, nr,
4522                                                        "Read failure");
4523                                 continue;
4524                         }
4525
4526                         /* This zaps the entire block.  Bottom up. */
4527                         BUFFER_TRACE(bh, "free child branches");
4528                         ext4_free_branches(handle, inode, bh,
4529                                         (__le32 *) bh->b_data,
4530                                         (__le32 *) bh->b_data + addr_per_block,
4531                                         depth);
4532
4533                         /*
4534                          * Everything below this this pointer has been
4535                          * released.  Now let this top-of-subtree go.
4536                          *
4537                          * We want the freeing of this indirect block to be
4538                          * atomic in the journal with the updating of the
4539                          * bitmap block which owns it.  So make some room in
4540                          * the journal.
4541                          *
4542                          * We zero the parent pointer *after* freeing its
4543                          * pointee in the bitmaps, so if extend_transaction()
4544                          * for some reason fails to put the bitmap changes and
4545                          * the release into the same transaction, recovery
4546                          * will merely complain about releasing a free block,
4547                          * rather than leaking blocks.
4548                          */
4549                         if (ext4_handle_is_aborted(handle))
4550                                 return;
4551                         if (try_to_extend_transaction(handle, inode)) {
4552                                 ext4_mark_inode_dirty(handle, inode);
4553                                 ext4_truncate_restart_trans(handle, inode,
4554                                             blocks_for_truncate(inode));
4555                         }
4556
4557                         /*
4558                          * The forget flag here is critical because if
4559                          * we are journaling (and not doing data
4560                          * journaling), we have to make sure a revoke
4561                          * record is written to prevent the journal
4562                          * replay from overwriting the (former)
4563                          * indirect block if it gets reallocated as a
4564                          * data block.  This must happen in the same
4565                          * transaction where the data blocks are
4566                          * actually freed.
4567                          */
4568                         ext4_free_blocks(handle, inode, 0, nr, 1,
4569                                          EXT4_FREE_BLOCKS_METADATA|
4570                                          EXT4_FREE_BLOCKS_FORGET);
4571
4572                         if (parent_bh) {
4573                                 /*
4574                                  * The block which we have just freed is
4575                                  * pointed to by an indirect block: journal it
4576                                  */
4577                                 BUFFER_TRACE(parent_bh, "get_write_access");
4578                                 if (!ext4_journal_get_write_access(handle,
4579                                                                    parent_bh)){
4580                                         *p = 0;
4581                                         BUFFER_TRACE(parent_bh,
4582                                         "call ext4_handle_dirty_metadata");
4583                                         ext4_handle_dirty_metadata(handle,
4584                                                                    inode,
4585                                                                    parent_bh);
4586                                 }
4587                         }
4588                 }
4589         } else {
4590                 /* We have reached the bottom of the tree. */
4591                 BUFFER_TRACE(parent_bh, "free data blocks");
4592                 ext4_free_data(handle, inode, parent_bh, first, last);
4593         }
4594 }
4595
4596 int ext4_can_truncate(struct inode *inode)
4597 {
4598         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4599                 return 0;
4600         if (S_ISREG(inode->i_mode))
4601                 return 1;
4602         if (S_ISDIR(inode->i_mode))
4603                 return 1;
4604         if (S_ISLNK(inode->i_mode))
4605                 return !ext4_inode_is_fast_symlink(inode);
4606         return 0;
4607 }
4608
4609 /*
4610  * ext4_truncate()
4611  *
4612  * We block out ext4_get_block() block instantiations across the entire
4613  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4614  * simultaneously on behalf of the same inode.
4615  *
4616  * As we work through the truncate and commmit bits of it to the journal there
4617  * is one core, guiding principle: the file's tree must always be consistent on
4618  * disk.  We must be able to restart the truncate after a crash.
4619  *
4620  * The file's tree may be transiently inconsistent in memory (although it
4621  * probably isn't), but whenever we close off and commit a journal transaction,
4622  * the contents of (the filesystem + the journal) must be consistent and
4623  * restartable.  It's pretty simple, really: bottom up, right to left (although
4624  * left-to-right works OK too).
4625  *
4626  * Note that at recovery time, journal replay occurs *before* the restart of
4627  * truncate against the orphan inode list.
4628  *
4629  * The committed inode has the new, desired i_size (which is the same as
4630  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4631  * that this inode's truncate did not complete and it will again call
4632  * ext4_truncate() to have another go.  So there will be instantiated blocks
4633  * to the right of the truncation point in a crashed ext4 filesystem.  But
4634  * that's fine - as long as they are linked from the inode, the post-crash
4635  * ext4_truncate() run will find them and release them.
4636  */
4637 void ext4_truncate(struct inode *inode)
4638 {
4639         handle_t *handle;
4640         struct ext4_inode_info *ei = EXT4_I(inode);
4641         __le32 *i_data = ei->i_data;
4642         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4643         struct address_space *mapping = inode->i_mapping;
4644         ext4_lblk_t offsets[4];
4645         Indirect chain[4];
4646         Indirect *partial;
4647         __le32 nr = 0;
4648         int n;
4649         ext4_lblk_t last_block;
4650         unsigned blocksize = inode->i_sb->s_blocksize;
4651
4652         if (!ext4_can_truncate(inode))
4653                 return;
4654
4655         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4656
4657         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4658                 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4659
4660         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
4661                 ext4_ext_truncate(inode);
4662                 return;
4663         }
4664
4665         handle = start_transaction(inode);
4666         if (IS_ERR(handle))
4667                 return;         /* AKPM: return what? */
4668
4669         last_block = (inode->i_size + blocksize-1)
4670                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4671
4672         if (inode->i_size & (blocksize - 1))
4673                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4674                         goto out_stop;
4675
4676         n = ext4_block_to_path(inode, last_block, offsets, NULL);
4677         if (n == 0)
4678                 goto out_stop;  /* error */
4679
4680         /*
4681          * OK.  This truncate is going to happen.  We add the inode to the
4682          * orphan list, so that if this truncate spans multiple transactions,
4683          * and we crash, we will resume the truncate when the filesystem
4684          * recovers.  It also marks the inode dirty, to catch the new size.
4685          *
4686          * Implication: the file must always be in a sane, consistent
4687          * truncatable state while each transaction commits.
4688          */
4689         if (ext4_orphan_add(handle, inode))
4690                 goto out_stop;
4691
4692         /*
4693          * From here we block out all ext4_get_block() callers who want to
4694          * modify the block allocation tree.
4695          */
4696         down_write(&ei->i_data_sem);
4697
4698         ext4_discard_preallocations(inode);
4699
4700         /*
4701          * The orphan list entry will now protect us from any crash which
4702          * occurs before the truncate completes, so it is now safe to propagate
4703          * the new, shorter inode size (held for now in i_size) into the
4704          * on-disk inode. We do this via i_disksize, which is the value which
4705          * ext4 *really* writes onto the disk inode.
4706          */
4707         ei->i_disksize = inode->i_size;
4708
4709         if (n == 1) {           /* direct blocks */
4710                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4711                                i_data + EXT4_NDIR_BLOCKS);
4712                 goto do_indirects;
4713         }
4714
4715         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4716         /* Kill the top of shared branch (not detached) */
4717         if (nr) {
4718                 if (partial == chain) {
4719                         /* Shared branch grows from the inode */
4720                         ext4_free_branches(handle, inode, NULL,
4721                                            &nr, &nr+1, (chain+n-1) - partial);
4722                         *partial->p = 0;
4723                         /*
4724                          * We mark the inode dirty prior to restart,
4725                          * and prior to stop.  No need for it here.
4726                          */
4727                 } else {
4728                         /* Shared branch grows from an indirect block */
4729                         BUFFER_TRACE(partial->bh, "get_write_access");
4730                         ext4_free_branches(handle, inode, partial->bh,
4731                                         partial->p,
4732                                         partial->p+1, (chain+n-1) - partial);
4733                 }
4734         }
4735         /* Clear the ends of indirect blocks on the shared branch */
4736         while (partial > chain) {
4737                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4738                                    (__le32*)partial->bh->b_data+addr_per_block,
4739                                    (chain+n-1) - partial);
4740                 BUFFER_TRACE(partial->bh, "call brelse");
4741                 brelse(partial->bh);
4742                 partial--;
4743         }
4744 do_indirects:
4745         /* Kill the remaining (whole) subtrees */
4746         switch (offsets[0]) {
4747         default:
4748                 nr = i_data[EXT4_IND_BLOCK];
4749                 if (nr) {
4750                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4751                         i_data[EXT4_IND_BLOCK] = 0;
4752                 }
4753         case EXT4_IND_BLOCK:
4754                 nr = i_data[EXT4_DIND_BLOCK];
4755                 if (nr) {
4756                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4757                         i_data[EXT4_DIND_BLOCK] = 0;
4758                 }
4759         case EXT4_DIND_BLOCK:
4760                 nr = i_data[EXT4_TIND_BLOCK];
4761                 if (nr) {
4762                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4763                         i_data[EXT4_TIND_BLOCK] = 0;
4764                 }
4765         case EXT4_TIND_BLOCK:
4766                 ;
4767         }
4768
4769         up_write(&ei->i_data_sem);
4770         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4771         ext4_mark_inode_dirty(handle, inode);
4772
4773         /*
4774          * In a multi-transaction truncate, we only make the final transaction
4775          * synchronous
4776          */
4777         if (IS_SYNC(inode))
4778                 ext4_handle_sync(handle);
4779 out_stop:
4780         /*
4781          * If this was a simple ftruncate(), and the file will remain alive
4782          * then we need to clear up the orphan record which we created above.
4783          * However, if this was a real unlink then we were called by
4784          * ext4_delete_inode(), and we allow that function to clean up the
4785          * orphan info for us.
4786          */
4787         if (inode->i_nlink)
4788                 ext4_orphan_del(handle, inode);
4789
4790         ext4_journal_stop(handle);
4791 }
4792
4793 /*
4794  * ext4_get_inode_loc returns with an extra refcount against the inode's
4795  * underlying buffer_head on success. If 'in_mem' is true, we have all
4796  * data in memory that is needed to recreate the on-disk version of this
4797  * inode.
4798  */
4799 static int __ext4_get_inode_loc(struct inode *inode,
4800                                 struct ext4_iloc *iloc, int in_mem)
4801 {
4802         struct ext4_group_desc  *gdp;
4803         struct buffer_head      *bh;
4804         struct super_block      *sb = inode->i_sb;
4805         ext4_fsblk_t            block;
4806         int                     inodes_per_block, inode_offset;
4807
4808         iloc->bh = NULL;
4809         if (!ext4_valid_inum(sb, inode->i_ino))
4810                 return -EIO;
4811
4812         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4813         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4814         if (!gdp)
4815                 return -EIO;
4816
4817         /*
4818          * Figure out the offset within the block group inode table
4819          */
4820         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4821         inode_offset = ((inode->i_ino - 1) %
4822                         EXT4_INODES_PER_GROUP(sb));
4823         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4824         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4825
4826         bh = sb_getblk(sb, block);
4827         if (!bh) {
4828                 EXT4_ERROR_INODE_BLOCK(inode, block,
4829                                        "unable to read itable block");
4830                 return -EIO;
4831         }
4832         if (!buffer_uptodate(bh)) {
4833                 lock_buffer(bh);
4834
4835                 /*
4836                  * If the buffer has the write error flag, we have failed
4837                  * to write out another inode in the same block.  In this
4838                  * case, we don't have to read the block because we may
4839                  * read the old inode data successfully.
4840                  */
4841                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4842                         set_buffer_uptodate(bh);
4843
4844                 if (buffer_uptodate(bh)) {
4845                         /* someone brought it uptodate while we waited */
4846                         unlock_buffer(bh);
4847                         goto has_buffer;
4848                 }
4849
4850                 /*
4851                  * If we have all information of the inode in memory and this
4852                  * is the only valid inode in the block, we need not read the
4853                  * block.
4854                  */
4855                 if (in_mem) {
4856                         struct buffer_head *bitmap_bh;
4857                         int i, start;
4858
4859                         start = inode_offset & ~(inodes_per_block - 1);
4860
4861                         /* Is the inode bitmap in cache? */
4862                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4863                         if (!bitmap_bh)
4864                                 goto make_io;
4865
4866                         /*
4867                          * If the inode bitmap isn't in cache then the
4868                          * optimisation may end up performing two reads instead
4869                          * of one, so skip it.
4870                          */
4871                         if (!buffer_uptodate(bitmap_bh)) {
4872                                 brelse(bitmap_bh);
4873                                 goto make_io;
4874                         }
4875                         for (i = start; i < start + inodes_per_block; i++) {
4876                                 if (i == inode_offset)
4877                                         continue;
4878                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4879                                         break;
4880                         }
4881                         brelse(bitmap_bh);
4882                         if (i == start + inodes_per_block) {
4883                                 /* all other inodes are free, so skip I/O */
4884                                 memset(bh->b_data, 0, bh->b_size);
4885                                 set_buffer_uptodate(bh);
4886                                 unlock_buffer(bh);
4887                                 goto has_buffer;
4888                         }
4889                 }
4890
4891 make_io:
4892                 /*
4893                  * If we need to do any I/O, try to pre-readahead extra
4894                  * blocks from the inode table.
4895                  */
4896                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4897                         ext4_fsblk_t b, end, table;
4898                         unsigned num;
4899
4900                         table = ext4_inode_table(sb, gdp);
4901                         /* s_inode_readahead_blks is always a power of 2 */
4902                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4903                         if (table > b)
4904                                 b = table;
4905                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4906                         num = EXT4_INODES_PER_GROUP(sb);
4907                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4908                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4909                                 num -= ext4_itable_unused_count(sb, gdp);
4910                         table += num / inodes_per_block;
4911                         if (end > table)
4912                                 end = table;
4913                         while (b <= end)
4914                                 sb_breadahead(sb, b++);
4915                 }
4916
4917                 /*
4918                  * There are other valid inodes in the buffer, this inode
4919                  * has in-inode xattrs, or we don't have this inode in memory.
4920                  * Read the block from disk.
4921                  */
4922                 get_bh(bh);
4923                 bh->b_end_io = end_buffer_read_sync;
4924                 submit_bh(READ_META, bh);
4925                 wait_on_buffer(bh);
4926                 if (!buffer_uptodate(bh)) {
4927                         EXT4_ERROR_INODE_BLOCK(inode, block,
4928                                                "unable to read itable block");
4929                         brelse(bh);
4930                         return -EIO;
4931                 }
4932         }
4933 has_buffer:
4934         iloc->bh = bh;
4935         return 0;
4936 }
4937
4938 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4939 {
4940         /* We have all inode data except xattrs in memory here. */
4941         return __ext4_get_inode_loc(inode, iloc,
4942                 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
4943 }
4944
4945 void ext4_set_inode_flags(struct inode *inode)
4946 {
4947         unsigned int flags = EXT4_I(inode)->i_flags;
4948
4949         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4950         if (flags & EXT4_SYNC_FL)
4951                 inode->i_flags |= S_SYNC;
4952         if (flags & EXT4_APPEND_FL)
4953                 inode->i_flags |= S_APPEND;
4954         if (flags & EXT4_IMMUTABLE_FL)
4955                 inode->i_flags |= S_IMMUTABLE;
4956         if (flags & EXT4_NOATIME_FL)
4957                 inode->i_flags |= S_NOATIME;
4958         if (flags & EXT4_DIRSYNC_FL)
4959                 inode->i_flags |= S_DIRSYNC;
4960 }
4961
4962 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4963 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4964 {
4965         unsigned int vfs_fl;
4966         unsigned long old_fl, new_fl;
4967
4968         do {
4969                 vfs_fl = ei->vfs_inode.i_flags;
4970                 old_fl = ei->i_flags;
4971                 new_fl = old_fl & ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4972                                 EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|
4973                                 EXT4_DIRSYNC_FL);
4974                 if (vfs_fl & S_SYNC)
4975                         new_fl |= EXT4_SYNC_FL;
4976                 if (vfs_fl & S_APPEND)
4977                         new_fl |= EXT4_APPEND_FL;
4978                 if (vfs_fl & S_IMMUTABLE)
4979                         new_fl |= EXT4_IMMUTABLE_FL;
4980                 if (vfs_fl & S_NOATIME)
4981                         new_fl |= EXT4_NOATIME_FL;
4982                 if (vfs_fl & S_DIRSYNC)
4983                         new_fl |= EXT4_DIRSYNC_FL;
4984         } while (cmpxchg(&ei->i_flags, old_fl, new_fl) != old_fl);
4985 }
4986
4987 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4988                                   struct ext4_inode_info *ei)
4989 {
4990         blkcnt_t i_blocks ;
4991         struct inode *inode = &(ei->vfs_inode);
4992         struct super_block *sb = inode->i_sb;
4993
4994         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4995                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4996                 /* we are using combined 48 bit field */
4997                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4998                                         le32_to_cpu(raw_inode->i_blocks_lo);
4999                 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
5000                         /* i_blocks represent file system block size */
5001                         return i_blocks  << (inode->i_blkbits - 9);
5002                 } else {
5003                         return i_blocks;
5004                 }
5005         } else {
5006                 return le32_to_cpu(raw_inode->i_blocks_lo);
5007         }
5008 }
5009
5010 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
5011 {
5012         struct ext4_iloc iloc;
5013         struct ext4_inode *raw_inode;
5014         struct ext4_inode_info *ei;
5015         struct inode *inode;
5016         journal_t *journal = EXT4_SB(sb)->s_journal;
5017         long ret;
5018         int block;
5019
5020         inode = iget_locked(sb, ino);
5021         if (!inode)
5022                 return ERR_PTR(-ENOMEM);
5023         if (!(inode->i_state & I_NEW))
5024                 return inode;
5025
5026         ei = EXT4_I(inode);
5027         iloc.bh = 0;
5028
5029         ret = __ext4_get_inode_loc(inode, &iloc, 0);
5030         if (ret < 0)
5031                 goto bad_inode;
5032         raw_inode = ext4_raw_inode(&iloc);
5033         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
5034         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
5035         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
5036         if (!(test_opt(inode->i_sb, NO_UID32))) {
5037                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
5038                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
5039         }
5040         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
5041
5042         ei->i_state_flags = 0;
5043         ei->i_dir_start_lookup = 0;
5044         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
5045         /* We now have enough fields to check if the inode was active or not.
5046          * This is needed because nfsd might try to access dead inodes
5047          * the test is that same one that e2fsck uses
5048          * NeilBrown 1999oct15
5049          */
5050         if (inode->i_nlink == 0) {
5051                 if (inode->i_mode == 0 ||
5052                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
5053                         /* this inode is deleted */
5054                         ret = -ESTALE;
5055                         goto bad_inode;
5056                 }
5057                 /* The only unlinked inodes we let through here have
5058                  * valid i_mode and are being read by the orphan
5059                  * recovery code: that's fine, we're about to complete
5060                  * the process of deleting those. */
5061         }
5062         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
5063         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
5064         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
5065         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
5066                 ei->i_file_acl |=
5067                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
5068         inode->i_size = ext4_isize(raw_inode);
5069         ei->i_disksize = inode->i_size;
5070 #ifdef CONFIG_QUOTA
5071         ei->i_reserved_quota = 0;
5072 #endif
5073         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
5074         ei->i_block_group = iloc.block_group;
5075         ei->i_last_alloc_group = ~0;
5076         /*
5077          * NOTE! The in-memory inode i_data array is in little-endian order
5078          * even on big-endian machines: we do NOT byteswap the block numbers!
5079          */
5080         for (block = 0; block < EXT4_N_BLOCKS; block++)
5081                 ei->i_data[block] = raw_inode->i_block[block];
5082         INIT_LIST_HEAD(&ei->i_orphan);
5083
5084         /*
5085          * Set transaction id's of transactions that have to be committed
5086          * to finish f[data]sync. We set them to currently running transaction
5087          * as we cannot be sure that the inode or some of its metadata isn't
5088          * part of the transaction - the inode could have been reclaimed and
5089          * now it is reread from disk.
5090          */
5091         if (journal) {
5092                 transaction_t *transaction;
5093                 tid_t tid;
5094
5095                 read_lock(&journal->j_state_lock);
5096                 if (journal->j_running_transaction)
5097                         transaction = journal->j_running_transaction;
5098                 else
5099                         transaction = journal->j_committing_transaction;
5100                 if (transaction)
5101                         tid = transaction->t_tid;
5102                 else
5103                         tid = journal->j_commit_sequence;
5104                 read_unlock(&journal->j_state_lock);
5105                 ei->i_sync_tid = tid;
5106                 ei->i_datasync_tid = tid;
5107         }
5108
5109         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5110                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
5111                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
5112                     EXT4_INODE_SIZE(inode->i_sb)) {
5113                         ret = -EIO;
5114                         goto bad_inode;
5115                 }
5116                 if (ei->i_extra_isize == 0) {
5117                         /* The extra space is currently unused. Use it. */
5118                         ei->i_extra_isize = sizeof(struct ext4_inode) -
5119                                             EXT4_GOOD_OLD_INODE_SIZE;
5120                 } else {
5121                         __le32 *magic = (void *)raw_inode +
5122                                         EXT4_GOOD_OLD_INODE_SIZE +
5123                                         ei->i_extra_isize;
5124                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
5125                                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
5126                 }
5127         } else
5128                 ei->i_extra_isize = 0;
5129
5130         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
5131         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
5132         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
5133         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
5134
5135         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
5136         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5137                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5138                         inode->i_version |=
5139                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
5140         }
5141
5142         ret = 0;
5143         if (ei->i_file_acl &&
5144             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
5145                 EXT4_ERROR_INODE(inode, "bad extended attribute block %llu",
5146                                  ei->i_file_acl);
5147                 ret = -EIO;
5148                 goto bad_inode;
5149         } else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5150                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5151                     (S_ISLNK(inode->i_mode) &&
5152                      !ext4_inode_is_fast_symlink(inode)))
5153                         /* Validate extent which is part of inode */
5154                         ret = ext4_ext_check_inode(inode);
5155         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5156                    (S_ISLNK(inode->i_mode) &&
5157                     !ext4_inode_is_fast_symlink(inode))) {
5158                 /* Validate block references which are part of inode */
5159                 ret = ext4_check_inode_blockref(inode);
5160         }
5161         if (ret)
5162                 goto bad_inode;
5163
5164         if (S_ISREG(inode->i_mode)) {
5165                 inode->i_op = &ext4_file_inode_operations;
5166                 inode->i_fop = &ext4_file_operations;
5167                 ext4_set_aops(inode);
5168         } else if (S_ISDIR(inode->i_mode)) {
5169                 inode->i_op = &ext4_dir_inode_operations;
5170                 inode->i_fop = &ext4_dir_operations;
5171         } else if (S_ISLNK(inode->i_mode)) {
5172                 if (ext4_inode_is_fast_symlink(inode)) {
5173                         inode->i_op = &ext4_fast_symlink_inode_operations;
5174                         nd_terminate_link(ei->i_data, inode->i_size,
5175                                 sizeof(ei->i_data) - 1);
5176                 } else {
5177                         inode->i_op = &ext4_symlink_inode_operations;
5178                         ext4_set_aops(inode);
5179                 }
5180         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
5181               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
5182                 inode->i_op = &ext4_special_inode_operations;
5183                 if (raw_inode->i_block[0])
5184                         init_special_inode(inode, inode->i_mode,
5185                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
5186                 else
5187                         init_special_inode(inode, inode->i_mode,
5188                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
5189         } else {
5190                 ret = -EIO;
5191                 EXT4_ERROR_INODE(inode, "bogus i_mode (%o)", inode->i_mode);
5192                 goto bad_inode;
5193         }
5194         brelse(iloc.bh);
5195         ext4_set_inode_flags(inode);
5196         unlock_new_inode(inode);
5197         return inode;
5198
5199 bad_inode:
5200         brelse(iloc.bh);
5201         iget_failed(inode);
5202         return ERR_PTR(ret);
5203 }
5204
5205 static int ext4_inode_blocks_set(handle_t *handle,
5206                                 struct ext4_inode *raw_inode,
5207                                 struct ext4_inode_info *ei)
5208 {
5209         struct inode *inode = &(ei->vfs_inode);
5210         u64 i_blocks = inode->i_blocks;
5211         struct super_block *sb = inode->i_sb;
5212
5213         if (i_blocks <= ~0U) {
5214                 /*
5215                  * i_blocks can be represnted in a 32 bit variable
5216                  * as multiple of 512 bytes
5217                  */
5218                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5219                 raw_inode->i_blocks_high = 0;
5220                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5221                 return 0;
5222         }
5223         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
5224                 return -EFBIG;
5225
5226         if (i_blocks <= 0xffffffffffffULL) {
5227                 /*
5228                  * i_blocks can be represented in a 48 bit variable
5229                  * as multiple of 512 bytes
5230                  */
5231                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5232                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5233                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5234         } else {
5235                 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5236                 /* i_block is stored in file system block size */
5237                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
5238                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5239                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5240         }
5241         return 0;
5242 }
5243
5244 /*
5245  * Post the struct inode info into an on-disk inode location in the
5246  * buffer-cache.  This gobbles the caller's reference to the
5247  * buffer_head in the inode location struct.
5248  *
5249  * The caller must have write access to iloc->bh.
5250  */
5251 static int ext4_do_update_inode(handle_t *handle,
5252                                 struct inode *inode,
5253                                 struct ext4_iloc *iloc)
5254 {
5255         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5256         struct ext4_inode_info *ei = EXT4_I(inode);
5257         struct buffer_head *bh = iloc->bh;
5258         int err = 0, rc, block;
5259
5260         /* For fields not not tracking in the in-memory inode,
5261          * initialise them to zero for new inodes. */
5262         if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5263                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5264
5265         ext4_get_inode_flags(ei);
5266         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5267         if (!(test_opt(inode->i_sb, NO_UID32))) {
5268                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
5269                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
5270 /*
5271  * Fix up interoperability with old kernels. Otherwise, old inodes get
5272  * re-used with the upper 16 bits of the uid/gid intact
5273  */
5274                 if (!ei->i_dtime) {
5275                         raw_inode->i_uid_high =
5276                                 cpu_to_le16(high_16_bits(inode->i_uid));
5277                         raw_inode->i_gid_high =
5278                                 cpu_to_le16(high_16_bits(inode->i_gid));
5279                 } else {
5280                         raw_inode->i_uid_high = 0;
5281                         raw_inode->i_gid_high = 0;
5282                 }
5283         } else {
5284                 raw_inode->i_uid_low =
5285                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
5286                 raw_inode->i_gid_low =
5287                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
5288                 raw_inode->i_uid_high = 0;
5289                 raw_inode->i_gid_high = 0;
5290         }
5291         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5292
5293         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5294         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5295         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5296         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5297
5298         if (ext4_inode_blocks_set(handle, raw_inode, ei))
5299                 goto out_brelse;
5300         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5301         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
5302         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
5303             cpu_to_le32(EXT4_OS_HURD))
5304                 raw_inode->i_file_acl_high =
5305                         cpu_to_le16(ei->i_file_acl >> 32);
5306         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5307         ext4_isize_set(raw_inode, ei->i_disksize);
5308         if (ei->i_disksize > 0x7fffffffULL) {
5309                 struct super_block *sb = inode->i_sb;
5310                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
5311                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
5312                                 EXT4_SB(sb)->s_es->s_rev_level ==
5313                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
5314                         /* If this is the first large file
5315                          * created, add a flag to the superblock.
5316                          */
5317                         err = ext4_journal_get_write_access(handle,
5318                                         EXT4_SB(sb)->s_sbh);
5319                         if (err)
5320                                 goto out_brelse;
5321                         ext4_update_dynamic_rev(sb);
5322                         EXT4_SET_RO_COMPAT_FEATURE(sb,
5323                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
5324                         sb->s_dirt = 1;
5325                         ext4_handle_sync(handle);
5326                         err = ext4_handle_dirty_metadata(handle, NULL,
5327                                         EXT4_SB(sb)->s_sbh);
5328                 }
5329         }
5330         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5331         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5332                 if (old_valid_dev(inode->i_rdev)) {
5333                         raw_inode->i_block[0] =
5334                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
5335                         raw_inode->i_block[1] = 0;
5336                 } else {
5337                         raw_inode->i_block[0] = 0;
5338                         raw_inode->i_block[1] =
5339                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
5340                         raw_inode->i_block[2] = 0;
5341                 }
5342         } else
5343                 for (block = 0; block < EXT4_N_BLOCKS; block++)
5344                         raw_inode->i_block[block] = ei->i_data[block];
5345
5346         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
5347         if (ei->i_extra_isize) {
5348                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5349                         raw_inode->i_version_hi =
5350                         cpu_to_le32(inode->i_version >> 32);
5351                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
5352         }
5353
5354         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5355         rc = ext4_handle_dirty_metadata(handle, NULL, bh);
5356         if (!err)
5357                 err = rc;
5358         ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5359
5360         ext4_update_inode_fsync_trans(handle, inode, 0);
5361 out_brelse:
5362         brelse(bh);
5363         ext4_std_error(inode->i_sb, err);
5364         return err;
5365 }
5366
5367 /*
5368  * ext4_write_inode()
5369  *
5370  * We are called from a few places:
5371  *
5372  * - Within generic_file_write() for O_SYNC files.
5373  *   Here, there will be no transaction running. We wait for any running
5374  *   trasnaction to commit.
5375  *
5376  * - Within sys_sync(), kupdate and such.
5377  *   We wait on commit, if tol to.
5378  *
5379  * - Within prune_icache() (PF_MEMALLOC == true)
5380  *   Here we simply return.  We can't afford to block kswapd on the
5381  *   journal commit.
5382  *
5383  * In all cases it is actually safe for us to return without doing anything,
5384  * because the inode has been copied into a raw inode buffer in
5385  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
5386  * knfsd.
5387  *
5388  * Note that we are absolutely dependent upon all inode dirtiers doing the
5389  * right thing: they *must* call mark_inode_dirty() after dirtying info in
5390  * which we are interested.
5391  *
5392  * It would be a bug for them to not do this.  The code:
5393  *
5394  *      mark_inode_dirty(inode)
5395  *      stuff();
5396  *      inode->i_size = expr;
5397  *
5398  * is in error because a kswapd-driven write_inode() could occur while
5399  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
5400  * will no longer be on the superblock's dirty inode list.
5401  */
5402 int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5403 {
5404         int err;
5405
5406         if (current->flags & PF_MEMALLOC)
5407                 return 0;
5408
5409         if (EXT4_SB(inode->i_sb)->s_journal) {
5410                 if (ext4_journal_current_handle()) {
5411                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5412                         dump_stack();
5413                         return -EIO;
5414                 }
5415
5416                 if (wbc->sync_mode != WB_SYNC_ALL)
5417                         return 0;
5418
5419                 err = ext4_force_commit(inode->i_sb);
5420         } else {
5421                 struct ext4_iloc iloc;
5422
5423                 err = __ext4_get_inode_loc(inode, &iloc, 0);
5424                 if (err)
5425                         return err;
5426                 if (wbc->sync_mode == WB_SYNC_ALL)
5427                         sync_dirty_buffer(iloc.bh);
5428                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5429                         EXT4_ERROR_INODE_BLOCK(inode, iloc.bh->b_blocknr,
5430                                          "IO error syncing inode");
5431                         err = -EIO;
5432                 }
5433                 brelse(iloc.bh);
5434         }
5435         return err;
5436 }
5437
5438 /*
5439  * ext4_setattr()
5440  *
5441  * Called from notify_change.
5442  *
5443  * We want to trap VFS attempts to truncate the file as soon as
5444  * possible.  In particular, we want to make sure that when the VFS
5445  * shrinks i_size, we put the inode on the orphan list and modify
5446  * i_disksize immediately, so that during the subsequent flushing of
5447  * dirty pages and freeing of disk blocks, we can guarantee that any
5448  * commit will leave the blocks being flushed in an unused state on
5449  * disk.  (On recovery, the inode will get truncated and the blocks will
5450  * be freed, so we have a strong guarantee that no future commit will
5451  * leave these blocks visible to the user.)
5452  *
5453  * Another thing we have to assure is that if we are in ordered mode
5454  * and inode is still attached to the committing transaction, we must
5455  * we start writeout of all the dirty pages which are being truncated.
5456  * This way we are sure that all the data written in the previous
5457  * transaction are already on disk (truncate waits for pages under
5458  * writeback).
5459  *
5460  * Called with inode->i_mutex down.
5461  */
5462 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5463 {
5464         struct inode *inode = dentry->d_inode;
5465         int error, rc = 0;
5466         const unsigned int ia_valid = attr->ia_valid;
5467
5468         error = inode_change_ok(inode, attr);
5469         if (error)
5470                 return error;
5471
5472         if (is_quota_modification(inode, attr))
5473                 dquot_initialize(inode);
5474         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
5475                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
5476                 handle_t *handle;
5477
5478                 /* (user+group)*(old+new) structure, inode write (sb,
5479                  * inode block, ? - but truncate inode update has it) */
5480                 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
5481                                         EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
5482                 if (IS_ERR(handle)) {
5483                         error = PTR_ERR(handle);
5484                         goto err_out;
5485                 }
5486                 error = dquot_transfer(inode, attr);
5487                 if (error) {
5488                         ext4_journal_stop(handle);
5489                         return error;
5490                 }
5491                 /* Update corresponding info in inode so that everything is in
5492                  * one transaction */
5493                 if (attr->ia_valid & ATTR_UID)
5494                         inode->i_uid = attr->ia_uid;
5495                 if (attr->ia_valid & ATTR_GID)
5496                         inode->i_gid = attr->ia_gid;
5497                 error = ext4_mark_inode_dirty(handle, inode);
5498                 ext4_journal_stop(handle);
5499         }
5500
5501         if (attr->ia_valid & ATTR_SIZE) {
5502                 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
5503                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5504
5505                         if (attr->ia_size > sbi->s_bitmap_maxbytes)
5506                                 return -EFBIG;
5507                 }
5508         }
5509
5510         if (S_ISREG(inode->i_mode) &&
5511             attr->ia_valid & ATTR_SIZE &&
5512             (attr->ia_size < inode->i_size ||
5513              (ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS)))) {
5514                 handle_t *handle;
5515
5516                 handle = ext4_journal_start(inode, 3);
5517                 if (IS_ERR(handle)) {
5518                         error = PTR_ERR(handle);
5519                         goto err_out;
5520                 }
5521
5522                 error = ext4_orphan_add(handle, inode);
5523                 EXT4_I(inode)->i_disksize = attr->ia_size;
5524                 rc = ext4_mark_inode_dirty(handle, inode);
5525                 if (!error)
5526                         error = rc;
5527                 ext4_journal_stop(handle);
5528
5529                 if (ext4_should_order_data(inode)) {
5530                         error = ext4_begin_ordered_truncate(inode,
5531                                                             attr->ia_size);
5532                         if (error) {
5533                                 /* Do as much error cleanup as possible */
5534                                 handle = ext4_journal_start(inode, 3);
5535                                 if (IS_ERR(handle)) {
5536                                         ext4_orphan_del(NULL, inode);
5537                                         goto err_out;
5538