ext4: don't scan/accumulate more pages than mballoc will allocate
[linux-2.6.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/jbd2.h>
29 #include <linux/highuid.h>
30 #include <linux/pagemap.h>
31 #include <linux/quotaops.h>
32 #include <linux/string.h>
33 #include <linux/buffer_head.h>
34 #include <linux/writeback.h>
35 #include <linux/pagevec.h>
36 #include <linux/mpage.h>
37 #include <linux/namei.h>
38 #include <linux/uio.h>
39 #include <linux/bio.h>
40 #include <linux/workqueue.h>
41 #include <linux/kernel.h>
42 #include <linux/slab.h>
43
44 #include "ext4_jbd2.h"
45 #include "xattr.h"
46 #include "acl.h"
47 #include "ext4_extents.h"
48
49 #include <trace/events/ext4.h>
50
51 #define MPAGE_DA_EXTENT_TAIL 0x01
52
53 static inline int ext4_begin_ordered_truncate(struct inode *inode,
54                                               loff_t new_size)
55 {
56         return jbd2_journal_begin_ordered_truncate(
57                                         EXT4_SB(inode->i_sb)->s_journal,
58                                         &EXT4_I(inode)->jinode,
59                                         new_size);
60 }
61
62 static void ext4_invalidatepage(struct page *page, unsigned long offset);
63
64 /*
65  * Test whether an inode is a fast symlink.
66  */
67 static int ext4_inode_is_fast_symlink(struct inode *inode)
68 {
69         int ea_blocks = EXT4_I(inode)->i_file_acl ?
70                 (inode->i_sb->s_blocksize >> 9) : 0;
71
72         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
73 }
74
75 /*
76  * Work out how many blocks we need to proceed with the next chunk of a
77  * truncate transaction.
78  */
79 static unsigned long blocks_for_truncate(struct inode *inode)
80 {
81         ext4_lblk_t needed;
82
83         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
84
85         /* Give ourselves just enough room to cope with inodes in which
86          * i_blocks is corrupt: we've seen disk corruptions in the past
87          * which resulted in random data in an inode which looked enough
88          * like a regular file for ext4 to try to delete it.  Things
89          * will go a bit crazy if that happens, but at least we should
90          * try not to panic the whole kernel. */
91         if (needed < 2)
92                 needed = 2;
93
94         /* But we need to bound the transaction so we don't overflow the
95          * journal. */
96         if (needed > EXT4_MAX_TRANS_DATA)
97                 needed = EXT4_MAX_TRANS_DATA;
98
99         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
100 }
101
102 /*
103  * Truncate transactions can be complex and absolutely huge.  So we need to
104  * be able to restart the transaction at a conventient checkpoint to make
105  * sure we don't overflow the journal.
106  *
107  * start_transaction gets us a new handle for a truncate transaction,
108  * and extend_transaction tries to extend the existing one a bit.  If
109  * extend fails, we need to propagate the failure up and restart the
110  * transaction in the top-level truncate loop. --sct
111  */
112 static handle_t *start_transaction(struct inode *inode)
113 {
114         handle_t *result;
115
116         result = ext4_journal_start(inode, blocks_for_truncate(inode));
117         if (!IS_ERR(result))
118                 return result;
119
120         ext4_std_error(inode->i_sb, PTR_ERR(result));
121         return result;
122 }
123
124 /*
125  * Try to extend this transaction for the purposes of truncation.
126  *
127  * Returns 0 if we managed to create more room.  If we can't create more
128  * room, and the transaction must be restarted we return 1.
129  */
130 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
131 {
132         if (!ext4_handle_valid(handle))
133                 return 0;
134         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
135                 return 0;
136         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
137                 return 0;
138         return 1;
139 }
140
141 /*
142  * Restart the transaction associated with *handle.  This does a commit,
143  * so before we call here everything must be consistently dirtied against
144  * this transaction.
145  */
146 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
147                                  int nblocks)
148 {
149         int ret;
150
151         /*
152          * Drop i_data_sem to avoid deadlock with ext4_get_blocks At this
153          * moment, get_block can be called only for blocks inside i_size since
154          * page cache has been already dropped and writes are blocked by
155          * i_mutex. So we can safely drop the i_data_sem here.
156          */
157         BUG_ON(EXT4_JOURNAL(inode) == NULL);
158         jbd_debug(2, "restarting handle %p\n", handle);
159         up_write(&EXT4_I(inode)->i_data_sem);
160         ret = ext4_journal_restart(handle, blocks_for_truncate(inode));
161         down_write(&EXT4_I(inode)->i_data_sem);
162         ext4_discard_preallocations(inode);
163
164         return ret;
165 }
166
167 /*
168  * Called at the last iput() if i_nlink is zero.
169  */
170 void ext4_delete_inode(struct inode *inode)
171 {
172         handle_t *handle;
173         int err;
174
175         if (!is_bad_inode(inode))
176                 dquot_initialize(inode);
177
178         if (ext4_should_order_data(inode))
179                 ext4_begin_ordered_truncate(inode, 0);
180         truncate_inode_pages(&inode->i_data, 0);
181
182         if (is_bad_inode(inode))
183                 goto no_delete;
184
185         handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
186         if (IS_ERR(handle)) {
187                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
188                 /*
189                  * If we're going to skip the normal cleanup, we still need to
190                  * make sure that the in-core orphan linked list is properly
191                  * cleaned up.
192                  */
193                 ext4_orphan_del(NULL, inode);
194                 goto no_delete;
195         }
196
197         if (IS_SYNC(inode))
198                 ext4_handle_sync(handle);
199         inode->i_size = 0;
200         err = ext4_mark_inode_dirty(handle, inode);
201         if (err) {
202                 ext4_warning(inode->i_sb,
203                              "couldn't mark inode dirty (err %d)", err);
204                 goto stop_handle;
205         }
206         if (inode->i_blocks)
207                 ext4_truncate(inode);
208
209         /*
210          * ext4_ext_truncate() doesn't reserve any slop when it
211          * restarts journal transactions; therefore there may not be
212          * enough credits left in the handle to remove the inode from
213          * the orphan list and set the dtime field.
214          */
215         if (!ext4_handle_has_enough_credits(handle, 3)) {
216                 err = ext4_journal_extend(handle, 3);
217                 if (err > 0)
218                         err = ext4_journal_restart(handle, 3);
219                 if (err != 0) {
220                         ext4_warning(inode->i_sb,
221                                      "couldn't extend journal (err %d)", err);
222                 stop_handle:
223                         ext4_journal_stop(handle);
224                         goto no_delete;
225                 }
226         }
227
228         /*
229          * Kill off the orphan record which ext4_truncate created.
230          * AKPM: I think this can be inside the above `if'.
231          * Note that ext4_orphan_del() has to be able to cope with the
232          * deletion of a non-existent orphan - this is because we don't
233          * know if ext4_truncate() actually created an orphan record.
234          * (Well, we could do this if we need to, but heck - it works)
235          */
236         ext4_orphan_del(handle, inode);
237         EXT4_I(inode)->i_dtime  = get_seconds();
238
239         /*
240          * One subtle ordering requirement: if anything has gone wrong
241          * (transaction abort, IO errors, whatever), then we can still
242          * do these next steps (the fs will already have been marked as
243          * having errors), but we can't free the inode if the mark_dirty
244          * fails.
245          */
246         if (ext4_mark_inode_dirty(handle, inode))
247                 /* If that failed, just do the required in-core inode clear. */
248                 clear_inode(inode);
249         else
250                 ext4_free_inode(handle, inode);
251         ext4_journal_stop(handle);
252         return;
253 no_delete:
254         clear_inode(inode);     /* We must guarantee clearing of inode... */
255 }
256
257 typedef struct {
258         __le32  *p;
259         __le32  key;
260         struct buffer_head *bh;
261 } Indirect;
262
263 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
264 {
265         p->key = *(p->p = v);
266         p->bh = bh;
267 }
268
269 /**
270  *      ext4_block_to_path - parse the block number into array of offsets
271  *      @inode: inode in question (we are only interested in its superblock)
272  *      @i_block: block number to be parsed
273  *      @offsets: array to store the offsets in
274  *      @boundary: set this non-zero if the referred-to block is likely to be
275  *             followed (on disk) by an indirect block.
276  *
277  *      To store the locations of file's data ext4 uses a data structure common
278  *      for UNIX filesystems - tree of pointers anchored in the inode, with
279  *      data blocks at leaves and indirect blocks in intermediate nodes.
280  *      This function translates the block number into path in that tree -
281  *      return value is the path length and @offsets[n] is the offset of
282  *      pointer to (n+1)th node in the nth one. If @block is out of range
283  *      (negative or too large) warning is printed and zero returned.
284  *
285  *      Note: function doesn't find node addresses, so no IO is needed. All
286  *      we need to know is the capacity of indirect blocks (taken from the
287  *      inode->i_sb).
288  */
289
290 /*
291  * Portability note: the last comparison (check that we fit into triple
292  * indirect block) is spelled differently, because otherwise on an
293  * architecture with 32-bit longs and 8Kb pages we might get into trouble
294  * if our filesystem had 8Kb blocks. We might use long long, but that would
295  * kill us on x86. Oh, well, at least the sign propagation does not matter -
296  * i_block would have to be negative in the very beginning, so we would not
297  * get there at all.
298  */
299
300 static int ext4_block_to_path(struct inode *inode,
301                               ext4_lblk_t i_block,
302                               ext4_lblk_t offsets[4], int *boundary)
303 {
304         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
305         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
306         const long direct_blocks = EXT4_NDIR_BLOCKS,
307                 indirect_blocks = ptrs,
308                 double_blocks = (1 << (ptrs_bits * 2));
309         int n = 0;
310         int final = 0;
311
312         if (i_block < direct_blocks) {
313                 offsets[n++] = i_block;
314                 final = direct_blocks;
315         } else if ((i_block -= direct_blocks) < indirect_blocks) {
316                 offsets[n++] = EXT4_IND_BLOCK;
317                 offsets[n++] = i_block;
318                 final = ptrs;
319         } else if ((i_block -= indirect_blocks) < double_blocks) {
320                 offsets[n++] = EXT4_DIND_BLOCK;
321                 offsets[n++] = i_block >> ptrs_bits;
322                 offsets[n++] = i_block & (ptrs - 1);
323                 final = ptrs;
324         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
325                 offsets[n++] = EXT4_TIND_BLOCK;
326                 offsets[n++] = i_block >> (ptrs_bits * 2);
327                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
328                 offsets[n++] = i_block & (ptrs - 1);
329                 final = ptrs;
330         } else {
331                 ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
332                              i_block + direct_blocks +
333                              indirect_blocks + double_blocks, inode->i_ino);
334         }
335         if (boundary)
336                 *boundary = final - 1 - (i_block & (ptrs - 1));
337         return n;
338 }
339
340 static int __ext4_check_blockref(const char *function, struct inode *inode,
341                                  __le32 *p, unsigned int max)
342 {
343         __le32 *bref = p;
344         unsigned int blk;
345
346         while (bref < p+max) {
347                 blk = le32_to_cpu(*bref++);
348                 if (blk &&
349                     unlikely(!ext4_data_block_valid(EXT4_SB(inode->i_sb),
350                                                     blk, 1))) {
351                         __ext4_error(inode->i_sb, function,
352                                    "invalid block reference %u "
353                                    "in inode #%lu", blk, inode->i_ino);
354                         return -EIO;
355                 }
356         }
357         return 0;
358 }
359
360
361 #define ext4_check_indirect_blockref(inode, bh)                         \
362         __ext4_check_blockref(__func__, inode, (__le32 *)(bh)->b_data,  \
363                               EXT4_ADDR_PER_BLOCK((inode)->i_sb))
364
365 #define ext4_check_inode_blockref(inode)                                \
366         __ext4_check_blockref(__func__, inode, EXT4_I(inode)->i_data,   \
367                               EXT4_NDIR_BLOCKS)
368
369 /**
370  *      ext4_get_branch - read the chain of indirect blocks leading to data
371  *      @inode: inode in question
372  *      @depth: depth of the chain (1 - direct pointer, etc.)
373  *      @offsets: offsets of pointers in inode/indirect blocks
374  *      @chain: place to store the result
375  *      @err: here we store the error value
376  *
377  *      Function fills the array of triples <key, p, bh> and returns %NULL
378  *      if everything went OK or the pointer to the last filled triple
379  *      (incomplete one) otherwise. Upon the return chain[i].key contains
380  *      the number of (i+1)-th block in the chain (as it is stored in memory,
381  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
382  *      number (it points into struct inode for i==0 and into the bh->b_data
383  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
384  *      block for i>0 and NULL for i==0. In other words, it holds the block
385  *      numbers of the chain, addresses they were taken from (and where we can
386  *      verify that chain did not change) and buffer_heads hosting these
387  *      numbers.
388  *
389  *      Function stops when it stumbles upon zero pointer (absent block)
390  *              (pointer to last triple returned, *@err == 0)
391  *      or when it gets an IO error reading an indirect block
392  *              (ditto, *@err == -EIO)
393  *      or when it reads all @depth-1 indirect blocks successfully and finds
394  *      the whole chain, all way to the data (returns %NULL, *err == 0).
395  *
396  *      Need to be called with
397  *      down_read(&EXT4_I(inode)->i_data_sem)
398  */
399 static Indirect *ext4_get_branch(struct inode *inode, int depth,
400                                  ext4_lblk_t  *offsets,
401                                  Indirect chain[4], int *err)
402 {
403         struct super_block *sb = inode->i_sb;
404         Indirect *p = chain;
405         struct buffer_head *bh;
406
407         *err = 0;
408         /* i_data is not going away, no lock needed */
409         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
410         if (!p->key)
411                 goto no_block;
412         while (--depth) {
413                 bh = sb_getblk(sb, le32_to_cpu(p->key));
414                 if (unlikely(!bh))
415                         goto failure;
416
417                 if (!bh_uptodate_or_lock(bh)) {
418                         if (bh_submit_read(bh) < 0) {
419                                 put_bh(bh);
420                                 goto failure;
421                         }
422                         /* validate block references */
423                         if (ext4_check_indirect_blockref(inode, bh)) {
424                                 put_bh(bh);
425                                 goto failure;
426                         }
427                 }
428
429                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
430                 /* Reader: end */
431                 if (!p->key)
432                         goto no_block;
433         }
434         return NULL;
435
436 failure:
437         *err = -EIO;
438 no_block:
439         return p;
440 }
441
442 /**
443  *      ext4_find_near - find a place for allocation with sufficient locality
444  *      @inode: owner
445  *      @ind: descriptor of indirect block.
446  *
447  *      This function returns the preferred place for block allocation.
448  *      It is used when heuristic for sequential allocation fails.
449  *      Rules are:
450  *        + if there is a block to the left of our position - allocate near it.
451  *        + if pointer will live in indirect block - allocate near that block.
452  *        + if pointer will live in inode - allocate in the same
453  *          cylinder group.
454  *
455  * In the latter case we colour the starting block by the callers PID to
456  * prevent it from clashing with concurrent allocations for a different inode
457  * in the same block group.   The PID is used here so that functionally related
458  * files will be close-by on-disk.
459  *
460  *      Caller must make sure that @ind is valid and will stay that way.
461  */
462 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
463 {
464         struct ext4_inode_info *ei = EXT4_I(inode);
465         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
466         __le32 *p;
467         ext4_fsblk_t bg_start;
468         ext4_fsblk_t last_block;
469         ext4_grpblk_t colour;
470         ext4_group_t block_group;
471         int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
472
473         /* Try to find previous block */
474         for (p = ind->p - 1; p >= start; p--) {
475                 if (*p)
476                         return le32_to_cpu(*p);
477         }
478
479         /* No such thing, so let's try location of indirect block */
480         if (ind->bh)
481                 return ind->bh->b_blocknr;
482
483         /*
484          * It is going to be referred to from the inode itself? OK, just put it
485          * into the same cylinder group then.
486          */
487         block_group = ei->i_block_group;
488         if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
489                 block_group &= ~(flex_size-1);
490                 if (S_ISREG(inode->i_mode))
491                         block_group++;
492         }
493         bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
494         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
495
496         /*
497          * If we are doing delayed allocation, we don't need take
498          * colour into account.
499          */
500         if (test_opt(inode->i_sb, DELALLOC))
501                 return bg_start;
502
503         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
504                 colour = (current->pid % 16) *
505                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
506         else
507                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
508         return bg_start + colour;
509 }
510
511 /**
512  *      ext4_find_goal - find a preferred place for allocation.
513  *      @inode: owner
514  *      @block:  block we want
515  *      @partial: pointer to the last triple within a chain
516  *
517  *      Normally this function find the preferred place for block allocation,
518  *      returns it.
519  *      Because this is only used for non-extent files, we limit the block nr
520  *      to 32 bits.
521  */
522 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
523                                    Indirect *partial)
524 {
525         ext4_fsblk_t goal;
526
527         /*
528          * XXX need to get goal block from mballoc's data structures
529          */
530
531         goal = ext4_find_near(inode, partial);
532         goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
533         return goal;
534 }
535
536 /**
537  *      ext4_blks_to_allocate: Look up the block map and count the number
538  *      of direct blocks need to be allocated for the given branch.
539  *
540  *      @branch: chain of indirect blocks
541  *      @k: number of blocks need for indirect blocks
542  *      @blks: number of data blocks to be mapped.
543  *      @blocks_to_boundary:  the offset in the indirect block
544  *
545  *      return the total number of blocks to be allocate, including the
546  *      direct and indirect blocks.
547  */
548 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
549                                  int blocks_to_boundary)
550 {
551         unsigned int count = 0;
552
553         /*
554          * Simple case, [t,d]Indirect block(s) has not allocated yet
555          * then it's clear blocks on that path have not allocated
556          */
557         if (k > 0) {
558                 /* right now we don't handle cross boundary allocation */
559                 if (blks < blocks_to_boundary + 1)
560                         count += blks;
561                 else
562                         count += blocks_to_boundary + 1;
563                 return count;
564         }
565
566         count++;
567         while (count < blks && count <= blocks_to_boundary &&
568                 le32_to_cpu(*(branch[0].p + count)) == 0) {
569                 count++;
570         }
571         return count;
572 }
573
574 /**
575  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
576  *      @indirect_blks: the number of blocks need to allocate for indirect
577  *                      blocks
578  *
579  *      @new_blocks: on return it will store the new block numbers for
580  *      the indirect blocks(if needed) and the first direct block,
581  *      @blks:  on return it will store the total number of allocated
582  *              direct blocks
583  */
584 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
585                              ext4_lblk_t iblock, ext4_fsblk_t goal,
586                              int indirect_blks, int blks,
587                              ext4_fsblk_t new_blocks[4], int *err)
588 {
589         struct ext4_allocation_request ar;
590         int target, i;
591         unsigned long count = 0, blk_allocated = 0;
592         int index = 0;
593         ext4_fsblk_t current_block = 0;
594         int ret = 0;
595
596         /*
597          * Here we try to allocate the requested multiple blocks at once,
598          * on a best-effort basis.
599          * To build a branch, we should allocate blocks for
600          * the indirect blocks(if not allocated yet), and at least
601          * the first direct block of this branch.  That's the
602          * minimum number of blocks need to allocate(required)
603          */
604         /* first we try to allocate the indirect blocks */
605         target = indirect_blks;
606         while (target > 0) {
607                 count = target;
608                 /* allocating blocks for indirect blocks and direct blocks */
609                 current_block = ext4_new_meta_blocks(handle, inode,
610                                                         goal, &count, err);
611                 if (*err)
612                         goto failed_out;
613
614                 if (unlikely(current_block + count > EXT4_MAX_BLOCK_FILE_PHYS)) {
615                         EXT4_ERROR_INODE(inode,
616                                          "current_block %llu + count %lu > %d!",
617                                          current_block, count,
618                                          EXT4_MAX_BLOCK_FILE_PHYS);
619                         *err = -EIO;
620                         goto failed_out;
621                 }
622
623                 target -= count;
624                 /* allocate blocks for indirect blocks */
625                 while (index < indirect_blks && count) {
626                         new_blocks[index++] = current_block++;
627                         count--;
628                 }
629                 if (count > 0) {
630                         /*
631                          * save the new block number
632                          * for the first direct block
633                          */
634                         new_blocks[index] = current_block;
635                         printk(KERN_INFO "%s returned more blocks than "
636                                                 "requested\n", __func__);
637                         WARN_ON(1);
638                         break;
639                 }
640         }
641
642         target = blks - count ;
643         blk_allocated = count;
644         if (!target)
645                 goto allocated;
646         /* Now allocate data blocks */
647         memset(&ar, 0, sizeof(ar));
648         ar.inode = inode;
649         ar.goal = goal;
650         ar.len = target;
651         ar.logical = iblock;
652         if (S_ISREG(inode->i_mode))
653                 /* enable in-core preallocation only for regular files */
654                 ar.flags = EXT4_MB_HINT_DATA;
655
656         current_block = ext4_mb_new_blocks(handle, &ar, err);
657         if (unlikely(current_block + ar.len > EXT4_MAX_BLOCK_FILE_PHYS)) {
658                 EXT4_ERROR_INODE(inode,
659                                  "current_block %llu + ar.len %d > %d!",
660                                  current_block, ar.len,
661                                  EXT4_MAX_BLOCK_FILE_PHYS);
662                 *err = -EIO;
663                 goto failed_out;
664         }
665
666         if (*err && (target == blks)) {
667                 /*
668                  * if the allocation failed and we didn't allocate
669                  * any blocks before
670                  */
671                 goto failed_out;
672         }
673         if (!*err) {
674                 if (target == blks) {
675                         /*
676                          * save the new block number
677                          * for the first direct block
678                          */
679                         new_blocks[index] = current_block;
680                 }
681                 blk_allocated += ar.len;
682         }
683 allocated:
684         /* total number of blocks allocated for direct blocks */
685         ret = blk_allocated;
686         *err = 0;
687         return ret;
688 failed_out:
689         for (i = 0; i < index; i++)
690                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
691         return ret;
692 }
693
694 /**
695  *      ext4_alloc_branch - allocate and set up a chain of blocks.
696  *      @inode: owner
697  *      @indirect_blks: number of allocated indirect blocks
698  *      @blks: number of allocated direct blocks
699  *      @offsets: offsets (in the blocks) to store the pointers to next.
700  *      @branch: place to store the chain in.
701  *
702  *      This function allocates blocks, zeroes out all but the last one,
703  *      links them into chain and (if we are synchronous) writes them to disk.
704  *      In other words, it prepares a branch that can be spliced onto the
705  *      inode. It stores the information about that chain in the branch[], in
706  *      the same format as ext4_get_branch() would do. We are calling it after
707  *      we had read the existing part of chain and partial points to the last
708  *      triple of that (one with zero ->key). Upon the exit we have the same
709  *      picture as after the successful ext4_get_block(), except that in one
710  *      place chain is disconnected - *branch->p is still zero (we did not
711  *      set the last link), but branch->key contains the number that should
712  *      be placed into *branch->p to fill that gap.
713  *
714  *      If allocation fails we free all blocks we've allocated (and forget
715  *      their buffer_heads) and return the error value the from failed
716  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
717  *      as described above and return 0.
718  */
719 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
720                              ext4_lblk_t iblock, int indirect_blks,
721                              int *blks, ext4_fsblk_t goal,
722                              ext4_lblk_t *offsets, Indirect *branch)
723 {
724         int blocksize = inode->i_sb->s_blocksize;
725         int i, n = 0;
726         int err = 0;
727         struct buffer_head *bh;
728         int num;
729         ext4_fsblk_t new_blocks[4];
730         ext4_fsblk_t current_block;
731
732         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
733                                 *blks, new_blocks, &err);
734         if (err)
735                 return err;
736
737         branch[0].key = cpu_to_le32(new_blocks[0]);
738         /*
739          * metadata blocks and data blocks are allocated.
740          */
741         for (n = 1; n <= indirect_blks;  n++) {
742                 /*
743                  * Get buffer_head for parent block, zero it out
744                  * and set the pointer to new one, then send
745                  * parent to disk.
746                  */
747                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
748                 branch[n].bh = bh;
749                 lock_buffer(bh);
750                 BUFFER_TRACE(bh, "call get_create_access");
751                 err = ext4_journal_get_create_access(handle, bh);
752                 if (err) {
753                         /* Don't brelse(bh) here; it's done in
754                          * ext4_journal_forget() below */
755                         unlock_buffer(bh);
756                         goto failed;
757                 }
758
759                 memset(bh->b_data, 0, blocksize);
760                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
761                 branch[n].key = cpu_to_le32(new_blocks[n]);
762                 *branch[n].p = branch[n].key;
763                 if (n == indirect_blks) {
764                         current_block = new_blocks[n];
765                         /*
766                          * End of chain, update the last new metablock of
767                          * the chain to point to the new allocated
768                          * data blocks numbers
769                          */
770                         for (i = 1; i < num; i++)
771                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
772                 }
773                 BUFFER_TRACE(bh, "marking uptodate");
774                 set_buffer_uptodate(bh);
775                 unlock_buffer(bh);
776
777                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
778                 err = ext4_handle_dirty_metadata(handle, inode, bh);
779                 if (err)
780                         goto failed;
781         }
782         *blks = num;
783         return err;
784 failed:
785         /* Allocation failed, free what we already allocated */
786         ext4_free_blocks(handle, inode, 0, new_blocks[0], 1, 0);
787         for (i = 1; i <= n ; i++) {
788                 /* 
789                  * branch[i].bh is newly allocated, so there is no
790                  * need to revoke the block, which is why we don't
791                  * need to set EXT4_FREE_BLOCKS_METADATA.
792                  */
793                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1,
794                                  EXT4_FREE_BLOCKS_FORGET);
795         }
796         for (i = n+1; i < indirect_blks; i++)
797                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
798
799         ext4_free_blocks(handle, inode, 0, new_blocks[i], num, 0);
800
801         return err;
802 }
803
804 /**
805  * ext4_splice_branch - splice the allocated branch onto inode.
806  * @inode: owner
807  * @block: (logical) number of block we are adding
808  * @chain: chain of indirect blocks (with a missing link - see
809  *      ext4_alloc_branch)
810  * @where: location of missing link
811  * @num:   number of indirect blocks we are adding
812  * @blks:  number of direct blocks we are adding
813  *
814  * This function fills the missing link and does all housekeeping needed in
815  * inode (->i_blocks, etc.). In case of success we end up with the full
816  * chain to new block and return 0.
817  */
818 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
819                               ext4_lblk_t block, Indirect *where, int num,
820                               int blks)
821 {
822         int i;
823         int err = 0;
824         ext4_fsblk_t current_block;
825
826         /*
827          * If we're splicing into a [td]indirect block (as opposed to the
828          * inode) then we need to get write access to the [td]indirect block
829          * before the splice.
830          */
831         if (where->bh) {
832                 BUFFER_TRACE(where->bh, "get_write_access");
833                 err = ext4_journal_get_write_access(handle, where->bh);
834                 if (err)
835                         goto err_out;
836         }
837         /* That's it */
838
839         *where->p = where->key;
840
841         /*
842          * Update the host buffer_head or inode to point to more just allocated
843          * direct blocks blocks
844          */
845         if (num == 0 && blks > 1) {
846                 current_block = le32_to_cpu(where->key) + 1;
847                 for (i = 1; i < blks; i++)
848                         *(where->p + i) = cpu_to_le32(current_block++);
849         }
850
851         /* We are done with atomic stuff, now do the rest of housekeeping */
852         /* had we spliced it onto indirect block? */
853         if (where->bh) {
854                 /*
855                  * If we spliced it onto an indirect block, we haven't
856                  * altered the inode.  Note however that if it is being spliced
857                  * onto an indirect block at the very end of the file (the
858                  * file is growing) then we *will* alter the inode to reflect
859                  * the new i_size.  But that is not done here - it is done in
860                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
861                  */
862                 jbd_debug(5, "splicing indirect only\n");
863                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
864                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
865                 if (err)
866                         goto err_out;
867         } else {
868                 /*
869                  * OK, we spliced it into the inode itself on a direct block.
870                  */
871                 ext4_mark_inode_dirty(handle, inode);
872                 jbd_debug(5, "splicing direct\n");
873         }
874         return err;
875
876 err_out:
877         for (i = 1; i <= num; i++) {
878                 /* 
879                  * branch[i].bh is newly allocated, so there is no
880                  * need to revoke the block, which is why we don't
881                  * need to set EXT4_FREE_BLOCKS_METADATA.
882                  */
883                 ext4_free_blocks(handle, inode, where[i].bh, 0, 1,
884                                  EXT4_FREE_BLOCKS_FORGET);
885         }
886         ext4_free_blocks(handle, inode, 0, le32_to_cpu(where[num].key),
887                          blks, 0);
888
889         return err;
890 }
891
892 /*
893  * The ext4_ind_get_blocks() function handles non-extents inodes
894  * (i.e., using the traditional indirect/double-indirect i_blocks
895  * scheme) for ext4_get_blocks().
896  *
897  * Allocation strategy is simple: if we have to allocate something, we will
898  * have to go the whole way to leaf. So let's do it before attaching anything
899  * to tree, set linkage between the newborn blocks, write them if sync is
900  * required, recheck the path, free and repeat if check fails, otherwise
901  * set the last missing link (that will protect us from any truncate-generated
902  * removals - all blocks on the path are immune now) and possibly force the
903  * write on the parent block.
904  * That has a nice additional property: no special recovery from the failed
905  * allocations is needed - we simply release blocks and do not touch anything
906  * reachable from inode.
907  *
908  * `handle' can be NULL if create == 0.
909  *
910  * return > 0, # of blocks mapped or allocated.
911  * return = 0, if plain lookup failed.
912  * return < 0, error case.
913  *
914  * The ext4_ind_get_blocks() function should be called with
915  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
916  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
917  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
918  * blocks.
919  */
920 static int ext4_ind_get_blocks(handle_t *handle, struct inode *inode,
921                                ext4_lblk_t iblock, unsigned int maxblocks,
922                                struct buffer_head *bh_result,
923                                int flags)
924 {
925         int err = -EIO;
926         ext4_lblk_t offsets[4];
927         Indirect chain[4];
928         Indirect *partial;
929         ext4_fsblk_t goal;
930         int indirect_blks;
931         int blocks_to_boundary = 0;
932         int depth;
933         int count = 0;
934         ext4_fsblk_t first_block = 0;
935
936         J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
937         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
938         depth = ext4_block_to_path(inode, iblock, offsets,
939                                    &blocks_to_boundary);
940
941         if (depth == 0)
942                 goto out;
943
944         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
945
946         /* Simplest case - block found, no allocation needed */
947         if (!partial) {
948                 first_block = le32_to_cpu(chain[depth - 1].key);
949                 clear_buffer_new(bh_result);
950                 count++;
951                 /*map more blocks*/
952                 while (count < maxblocks && count <= blocks_to_boundary) {
953                         ext4_fsblk_t blk;
954
955                         blk = le32_to_cpu(*(chain[depth-1].p + count));
956
957                         if (blk == first_block + count)
958                                 count++;
959                         else
960                                 break;
961                 }
962                 goto got_it;
963         }
964
965         /* Next simple case - plain lookup or failed read of indirect block */
966         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
967                 goto cleanup;
968
969         /*
970          * Okay, we need to do block allocation.
971         */
972         goal = ext4_find_goal(inode, iblock, partial);
973
974         /* the number of blocks need to allocate for [d,t]indirect blocks */
975         indirect_blks = (chain + depth) - partial - 1;
976
977         /*
978          * Next look up the indirect map to count the totoal number of
979          * direct blocks to allocate for this branch.
980          */
981         count = ext4_blks_to_allocate(partial, indirect_blks,
982                                         maxblocks, blocks_to_boundary);
983         /*
984          * Block out ext4_truncate while we alter the tree
985          */
986         err = ext4_alloc_branch(handle, inode, iblock, indirect_blks,
987                                 &count, goal,
988                                 offsets + (partial - chain), partial);
989
990         /*
991          * The ext4_splice_branch call will free and forget any buffers
992          * on the new chain if there is a failure, but that risks using
993          * up transaction credits, especially for bitmaps where the
994          * credits cannot be returned.  Can we handle this somehow?  We
995          * may need to return -EAGAIN upwards in the worst case.  --sct
996          */
997         if (!err)
998                 err = ext4_splice_branch(handle, inode, iblock,
999                                          partial, indirect_blks, count);
1000         if (err)
1001                 goto cleanup;
1002
1003         set_buffer_new(bh_result);
1004
1005         ext4_update_inode_fsync_trans(handle, inode, 1);
1006 got_it:
1007         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
1008         if (count > blocks_to_boundary)
1009                 set_buffer_boundary(bh_result);
1010         err = count;
1011         /* Clean up and exit */
1012         partial = chain + depth - 1;    /* the whole chain */
1013 cleanup:
1014         while (partial > chain) {
1015                 BUFFER_TRACE(partial->bh, "call brelse");
1016                 brelse(partial->bh);
1017                 partial--;
1018         }
1019         BUFFER_TRACE(bh_result, "returned");
1020 out:
1021         return err;
1022 }
1023
1024 #ifdef CONFIG_QUOTA
1025 qsize_t *ext4_get_reserved_space(struct inode *inode)
1026 {
1027         return &EXT4_I(inode)->i_reserved_quota;
1028 }
1029 #endif
1030
1031 /*
1032  * Calculate the number of metadata blocks need to reserve
1033  * to allocate a new block at @lblocks for non extent file based file
1034  */
1035 static int ext4_indirect_calc_metadata_amount(struct inode *inode,
1036                                               sector_t lblock)
1037 {
1038         struct ext4_inode_info *ei = EXT4_I(inode);
1039         sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
1040         int blk_bits;
1041
1042         if (lblock < EXT4_NDIR_BLOCKS)
1043                 return 0;
1044
1045         lblock -= EXT4_NDIR_BLOCKS;
1046
1047         if (ei->i_da_metadata_calc_len &&
1048             (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
1049                 ei->i_da_metadata_calc_len++;
1050                 return 0;
1051         }
1052         ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
1053         ei->i_da_metadata_calc_len = 1;
1054         blk_bits = order_base_2(lblock);
1055         return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
1056 }
1057
1058 /*
1059  * Calculate the number of metadata blocks need to reserve
1060  * to allocate a block located at @lblock
1061  */
1062 static int ext4_calc_metadata_amount(struct inode *inode, sector_t lblock)
1063 {
1064         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
1065                 return ext4_ext_calc_metadata_amount(inode, lblock);
1066
1067         return ext4_indirect_calc_metadata_amount(inode, lblock);
1068 }
1069
1070 /*
1071  * Called with i_data_sem down, which is important since we can call
1072  * ext4_discard_preallocations() from here.
1073  */
1074 void ext4_da_update_reserve_space(struct inode *inode,
1075                                         int used, int quota_claim)
1076 {
1077         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1078         struct ext4_inode_info *ei = EXT4_I(inode);
1079         int mdb_free = 0, allocated_meta_blocks = 0;
1080
1081         spin_lock(&ei->i_block_reservation_lock);
1082         trace_ext4_da_update_reserve_space(inode, used);
1083         if (unlikely(used > ei->i_reserved_data_blocks)) {
1084                 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
1085                          "with only %d reserved data blocks\n",
1086                          __func__, inode->i_ino, used,
1087                          ei->i_reserved_data_blocks);
1088                 WARN_ON(1);
1089                 used = ei->i_reserved_data_blocks;
1090         }
1091
1092         /* Update per-inode reservations */
1093         ei->i_reserved_data_blocks -= used;
1094         used += ei->i_allocated_meta_blocks;
1095         ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
1096         allocated_meta_blocks = ei->i_allocated_meta_blocks;
1097         ei->i_allocated_meta_blocks = 0;
1098         percpu_counter_sub(&sbi->s_dirtyblocks_counter, used);
1099
1100         if (ei->i_reserved_data_blocks == 0) {
1101                 /*
1102                  * We can release all of the reserved metadata blocks
1103                  * only when we have written all of the delayed
1104                  * allocation blocks.
1105                  */
1106                 mdb_free = ei->i_reserved_meta_blocks;
1107                 ei->i_reserved_meta_blocks = 0;
1108                 ei->i_da_metadata_calc_len = 0;
1109                 percpu_counter_sub(&sbi->s_dirtyblocks_counter, mdb_free);
1110         }
1111         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1112
1113         /* Update quota subsystem */
1114         if (quota_claim) {
1115                 dquot_claim_block(inode, used);
1116                 if (mdb_free)
1117                         dquot_release_reservation_block(inode, mdb_free);
1118         } else {
1119                 /*
1120                  * We did fallocate with an offset that is already delayed
1121                  * allocated. So on delayed allocated writeback we should
1122                  * not update the quota for allocated blocks. But then
1123                  * converting an fallocate region to initialized region would
1124                  * have caused a metadata allocation. So claim quota for
1125                  * that
1126                  */
1127                 if (allocated_meta_blocks)
1128                         dquot_claim_block(inode, allocated_meta_blocks);
1129                 dquot_release_reservation_block(inode, mdb_free + used -
1130                                                 allocated_meta_blocks);
1131         }
1132
1133         /*
1134          * If we have done all the pending block allocations and if
1135          * there aren't any writers on the inode, we can discard the
1136          * inode's preallocations.
1137          */
1138         if ((ei->i_reserved_data_blocks == 0) &&
1139             (atomic_read(&inode->i_writecount) == 0))
1140                 ext4_discard_preallocations(inode);
1141 }
1142
1143 static int check_block_validity(struct inode *inode, const char *msg,
1144                                 sector_t logical, sector_t phys, int len)
1145 {
1146         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), phys, len)) {
1147                 __ext4_error(inode->i_sb, msg,
1148                            "inode #%lu logical block %llu mapped to %llu "
1149                            "(size %d)", inode->i_ino,
1150                            (unsigned long long) logical,
1151                            (unsigned long long) phys, len);
1152                 return -EIO;
1153         }
1154         return 0;
1155 }
1156
1157 /*
1158  * Return the number of contiguous dirty pages in a given inode
1159  * starting at page frame idx.
1160  */
1161 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
1162                                     unsigned int max_pages)
1163 {
1164         struct address_space *mapping = inode->i_mapping;
1165         pgoff_t index;
1166         struct pagevec pvec;
1167         pgoff_t num = 0;
1168         int i, nr_pages, done = 0;
1169
1170         if (max_pages == 0)
1171                 return 0;
1172         pagevec_init(&pvec, 0);
1173         while (!done) {
1174                 index = idx;
1175                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
1176                                               PAGECACHE_TAG_DIRTY,
1177                                               (pgoff_t)PAGEVEC_SIZE);
1178                 if (nr_pages == 0)
1179                         break;
1180                 for (i = 0; i < nr_pages; i++) {
1181                         struct page *page = pvec.pages[i];
1182                         struct buffer_head *bh, *head;
1183
1184                         lock_page(page);
1185                         if (unlikely(page->mapping != mapping) ||
1186                             !PageDirty(page) ||
1187                             PageWriteback(page) ||
1188                             page->index != idx) {
1189                                 done = 1;
1190                                 unlock_page(page);
1191                                 break;
1192                         }
1193                         if (page_has_buffers(page)) {
1194                                 bh = head = page_buffers(page);
1195                                 do {
1196                                         if (!buffer_delay(bh) &&
1197                                             !buffer_unwritten(bh))
1198                                                 done = 1;
1199                                         bh = bh->b_this_page;
1200                                 } while (!done && (bh != head));
1201                         }
1202                         unlock_page(page);
1203                         if (done)
1204                                 break;
1205                         idx++;
1206                         num++;
1207                         if (num >= max_pages)
1208                                 break;
1209                 }
1210                 pagevec_release(&pvec);
1211         }
1212         return num;
1213 }
1214
1215 /*
1216  * The ext4_get_blocks() function tries to look up the requested blocks,
1217  * and returns if the blocks are already mapped.
1218  *
1219  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1220  * and store the allocated blocks in the result buffer head and mark it
1221  * mapped.
1222  *
1223  * If file type is extents based, it will call ext4_ext_get_blocks(),
1224  * Otherwise, call with ext4_ind_get_blocks() to handle indirect mapping
1225  * based files
1226  *
1227  * On success, it returns the number of blocks being mapped or allocate.
1228  * if create==0 and the blocks are pre-allocated and uninitialized block,
1229  * the result buffer head is unmapped. If the create ==1, it will make sure
1230  * the buffer head is mapped.
1231  *
1232  * It returns 0 if plain look up failed (blocks have not been allocated), in
1233  * that casem, buffer head is unmapped
1234  *
1235  * It returns the error in case of allocation failure.
1236  */
1237 int ext4_get_blocks(handle_t *handle, struct inode *inode, sector_t block,
1238                     unsigned int max_blocks, struct buffer_head *bh,
1239                     int flags)
1240 {
1241         int retval;
1242
1243         clear_buffer_mapped(bh);
1244         clear_buffer_unwritten(bh);
1245
1246         ext_debug("ext4_get_blocks(): inode %lu, flag %d, max_blocks %u,"
1247                   "logical block %lu\n", inode->i_ino, flags, max_blocks,
1248                   (unsigned long)block);
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_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1255                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1256                                 bh, 0);
1257         } else {
1258                 retval = ext4_ind_get_blocks(handle, inode, block, max_blocks,
1259                                              bh, 0);
1260         }
1261         up_read((&EXT4_I(inode)->i_data_sem));
1262
1263         if (retval > 0 && buffer_mapped(bh)) {
1264                 int ret = check_block_validity(inode, "file system corruption",
1265                                                block, bh->b_blocknr, retval);
1266                 if (ret != 0)
1267                         return ret;
1268         }
1269
1270         /* If it is only a block(s) look up */
1271         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1272                 return retval;
1273
1274         /*
1275          * Returns if the blocks have already allocated
1276          *
1277          * Note that if blocks have been preallocated
1278          * ext4_ext_get_block() returns th create = 0
1279          * with buffer head unmapped.
1280          */
1281         if (retval > 0 && buffer_mapped(bh))
1282                 return retval;
1283
1284         /*
1285          * When we call get_blocks without the create flag, the
1286          * BH_Unwritten flag could have gotten set if the blocks
1287          * requested were part of a uninitialized extent.  We need to
1288          * clear this flag now that we are committed to convert all or
1289          * part of the uninitialized extent to be an initialized
1290          * extent.  This is because we need to avoid the combination
1291          * of BH_Unwritten and BH_Mapped flags being simultaneously
1292          * set on the buffer_head.
1293          */
1294         clear_buffer_unwritten(bh);
1295
1296         /*
1297          * New blocks allocate and/or writing to uninitialized extent
1298          * will possibly result in updating i_data, so we take
1299          * the write lock of i_data_sem, and call get_blocks()
1300          * with create == 1 flag.
1301          */
1302         down_write((&EXT4_I(inode)->i_data_sem));
1303
1304         /*
1305          * if the caller is from delayed allocation writeout path
1306          * we have already reserved fs blocks for allocation
1307          * let the underlying get_block() function know to
1308          * avoid double accounting
1309          */
1310         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1311                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1312         /*
1313          * We need to check for EXT4 here because migrate
1314          * could have changed the inode type in between
1315          */
1316         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1317                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1318                                               bh, flags);
1319         } else {
1320                 retval = ext4_ind_get_blocks(handle, inode, block,
1321                                              max_blocks, bh, flags);
1322
1323                 if (retval > 0 && buffer_new(bh)) {
1324                         /*
1325                          * We allocated new blocks which will result in
1326                          * i_data's format changing.  Force the migrate
1327                          * to fail by clearing migrate flags
1328                          */
1329                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
1330                 }
1331
1332                 /*
1333                  * Update reserved blocks/metadata blocks after successful
1334                  * block allocation which had been deferred till now. We don't
1335                  * support fallocate for non extent files. So we can update
1336                  * reserve space here.
1337                  */
1338                 if ((retval > 0) &&
1339                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
1340                         ext4_da_update_reserve_space(inode, retval, 1);
1341         }
1342         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1343                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1344
1345         up_write((&EXT4_I(inode)->i_data_sem));
1346         if (retval > 0 && buffer_mapped(bh)) {
1347                 int ret = check_block_validity(inode, "file system "
1348                                                "corruption after allocation",
1349                                                block, bh->b_blocknr, retval);
1350                 if (ret != 0)
1351                         return ret;
1352         }
1353         return retval;
1354 }
1355
1356 /* Maximum number of blocks we map for direct IO at once. */
1357 #define DIO_MAX_BLOCKS 4096
1358
1359 int ext4_get_block(struct inode *inode, sector_t iblock,
1360                    struct buffer_head *bh_result, int create)
1361 {
1362         handle_t *handle = ext4_journal_current_handle();
1363         int ret = 0, started = 0;
1364         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1365         int dio_credits;
1366
1367         if (create && !handle) {
1368                 /* Direct IO write... */
1369                 if (max_blocks > DIO_MAX_BLOCKS)
1370                         max_blocks = DIO_MAX_BLOCKS;
1371                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
1372                 handle = ext4_journal_start(inode, dio_credits);
1373                 if (IS_ERR(handle)) {
1374                         ret = PTR_ERR(handle);
1375                         goto out;
1376                 }
1377                 started = 1;
1378         }
1379
1380         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
1381                               create ? EXT4_GET_BLOCKS_CREATE : 0);
1382         if (ret > 0) {
1383                 bh_result->b_size = (ret << inode->i_blkbits);
1384                 ret = 0;
1385         }
1386         if (started)
1387                 ext4_journal_stop(handle);
1388 out:
1389         return ret;
1390 }
1391
1392 /*
1393  * `handle' can be NULL if create is zero
1394  */
1395 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1396                                 ext4_lblk_t block, int create, int *errp)
1397 {
1398         struct buffer_head dummy;
1399         int fatal = 0, err;
1400         int flags = 0;
1401
1402         J_ASSERT(handle != NULL || create == 0);
1403
1404         dummy.b_state = 0;
1405         dummy.b_blocknr = -1000;
1406         buffer_trace_init(&dummy.b_history);
1407         if (create)
1408                 flags |= EXT4_GET_BLOCKS_CREATE;
1409         err = ext4_get_blocks(handle, inode, block, 1, &dummy, flags);
1410         /*
1411          * ext4_get_blocks() returns number of blocks mapped. 0 in
1412          * case of a HOLE.
1413          */
1414         if (err > 0) {
1415                 if (err > 1)
1416                         WARN_ON(1);
1417                 err = 0;
1418         }
1419         *errp = err;
1420         if (!err && buffer_mapped(&dummy)) {
1421                 struct buffer_head *bh;
1422                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1423                 if (!bh) {
1424                         *errp = -EIO;
1425                         goto err;
1426                 }
1427                 if (buffer_new(&dummy)) {
1428                         J_ASSERT(create != 0);
1429                         J_ASSERT(handle != NULL);
1430
1431                         /*
1432                          * Now that we do not always journal data, we should
1433                          * keep in mind whether this should always journal the
1434                          * new buffer as metadata.  For now, regular file
1435                          * writes use ext4_get_block instead, so it's not a
1436                          * problem.
1437                          */
1438                         lock_buffer(bh);
1439                         BUFFER_TRACE(bh, "call get_create_access");
1440                         fatal = ext4_journal_get_create_access(handle, bh);
1441                         if (!fatal && !buffer_uptodate(bh)) {
1442                                 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1443                                 set_buffer_uptodate(bh);
1444                         }
1445                         unlock_buffer(bh);
1446                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1447                         err = ext4_handle_dirty_metadata(handle, inode, bh);
1448                         if (!fatal)
1449                                 fatal = err;
1450                 } else {
1451                         BUFFER_TRACE(bh, "not a new buffer");
1452                 }
1453                 if (fatal) {
1454                         *errp = fatal;
1455                         brelse(bh);
1456                         bh = NULL;
1457                 }
1458                 return bh;
1459         }
1460 err:
1461         return NULL;
1462 }
1463
1464 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1465                                ext4_lblk_t block, int create, int *err)
1466 {
1467         struct buffer_head *bh;
1468
1469         bh = ext4_getblk(handle, inode, block, create, err);
1470         if (!bh)
1471                 return bh;
1472         if (buffer_uptodate(bh))
1473                 return bh;
1474         ll_rw_block(READ_META, 1, &bh);
1475         wait_on_buffer(bh);
1476         if (buffer_uptodate(bh))
1477                 return bh;
1478         put_bh(bh);
1479         *err = -EIO;
1480         return NULL;
1481 }
1482
1483 static int walk_page_buffers(handle_t *handle,
1484                              struct buffer_head *head,
1485                              unsigned from,
1486                              unsigned to,
1487                              int *partial,
1488                              int (*fn)(handle_t *handle,
1489                                        struct buffer_head *bh))
1490 {
1491         struct buffer_head *bh;
1492         unsigned block_start, block_end;
1493         unsigned blocksize = head->b_size;
1494         int err, ret = 0;
1495         struct buffer_head *next;
1496
1497         for (bh = head, block_start = 0;
1498              ret == 0 && (bh != head || !block_start);
1499              block_start = block_end, bh = next) {
1500                 next = bh->b_this_page;
1501                 block_end = block_start + blocksize;
1502                 if (block_end <= from || block_start >= to) {
1503                         if (partial && !buffer_uptodate(bh))
1504                                 *partial = 1;
1505                         continue;
1506                 }
1507                 err = (*fn)(handle, bh);
1508                 if (!ret)
1509                         ret = err;
1510         }
1511         return ret;
1512 }
1513
1514 /*
1515  * To preserve ordering, it is essential that the hole instantiation and
1516  * the data write be encapsulated in a single transaction.  We cannot
1517  * close off a transaction and start a new one between the ext4_get_block()
1518  * and the commit_write().  So doing the jbd2_journal_start at the start of
1519  * prepare_write() is the right place.
1520  *
1521  * Also, this function can nest inside ext4_writepage() ->
1522  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1523  * has generated enough buffer credits to do the whole page.  So we won't
1524  * block on the journal in that case, which is good, because the caller may
1525  * be PF_MEMALLOC.
1526  *
1527  * By accident, ext4 can be reentered when a transaction is open via
1528  * quota file writes.  If we were to commit the transaction while thus
1529  * reentered, there can be a deadlock - we would be holding a quota
1530  * lock, and the commit would never complete if another thread had a
1531  * transaction open and was blocking on the quota lock - a ranking
1532  * violation.
1533  *
1534  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1535  * will _not_ run commit under these circumstances because handle->h_ref
1536  * is elevated.  We'll still have enough credits for the tiny quotafile
1537  * write.
1538  */
1539 static int do_journal_get_write_access(handle_t *handle,
1540                                        struct buffer_head *bh)
1541 {
1542         if (!buffer_mapped(bh) || buffer_freed(bh))
1543                 return 0;
1544         return ext4_journal_get_write_access(handle, bh);
1545 }
1546
1547 /*
1548  * Truncate blocks that were not used by write. We have to truncate the
1549  * pagecache as well so that corresponding buffers get properly unmapped.
1550  */
1551 static void ext4_truncate_failed_write(struct inode *inode)
1552 {
1553         truncate_inode_pages(inode->i_mapping, inode->i_size);
1554         ext4_truncate(inode);
1555 }
1556
1557 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
1558                    struct buffer_head *bh_result, int create);
1559 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1560                             loff_t pos, unsigned len, unsigned flags,
1561                             struct page **pagep, void **fsdata)
1562 {
1563         struct inode *inode = mapping->host;
1564         int ret, needed_blocks;
1565         handle_t *handle;
1566         int retries = 0;
1567         struct page *page;
1568         pgoff_t index;
1569         unsigned from, to;
1570
1571         trace_ext4_write_begin(inode, pos, len, flags);
1572         /*
1573          * Reserve one block more for addition to orphan list in case
1574          * we allocate blocks but write fails for some reason
1575          */
1576         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1577         index = pos >> PAGE_CACHE_SHIFT;
1578         from = pos & (PAGE_CACHE_SIZE - 1);
1579         to = from + len;
1580
1581 retry:
1582         handle = ext4_journal_start(inode, needed_blocks);
1583         if (IS_ERR(handle)) {
1584                 ret = PTR_ERR(handle);
1585                 goto out;
1586         }
1587
1588         /* We cannot recurse into the filesystem as the transaction is already
1589          * started */
1590         flags |= AOP_FLAG_NOFS;
1591
1592         page = grab_cache_page_write_begin(mapping, index, flags);
1593         if (!page) {
1594                 ext4_journal_stop(handle);
1595                 ret = -ENOMEM;
1596                 goto out;
1597         }
1598         *pagep = page;
1599
1600         if (ext4_should_dioread_nolock(inode))
1601                 ret = block_write_begin(file, mapping, pos, len, flags, pagep,
1602                                 fsdata, ext4_get_block_write);
1603         else
1604                 ret = block_write_begin(file, mapping, pos, len, flags, pagep,
1605                                 fsdata, ext4_get_block);
1606
1607         if (!ret && ext4_should_journal_data(inode)) {
1608                 ret = walk_page_buffers(handle, page_buffers(page),
1609                                 from, to, NULL, do_journal_get_write_access);
1610         }
1611
1612         if (ret) {
1613                 unlock_page(page);
1614                 page_cache_release(page);
1615                 /*
1616                  * block_write_begin may have instantiated a few blocks
1617                  * outside i_size.  Trim these off again. Don't need
1618                  * i_size_read because we hold i_mutex.
1619                  *
1620                  * Add inode to orphan list in case we crash before
1621                  * truncate finishes
1622                  */
1623                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1624                         ext4_orphan_add(handle, inode);
1625
1626                 ext4_journal_stop(handle);
1627                 if (pos + len > inode->i_size) {
1628                         ext4_truncate_failed_write(inode);
1629                         /*
1630                          * If truncate failed early the inode might
1631                          * still be on the orphan list; we need to
1632                          * make sure the inode is removed from the
1633                          * orphan list in that case.
1634                          */
1635                         if (inode->i_nlink)
1636                                 ext4_orphan_del(NULL, inode);
1637                 }
1638         }
1639
1640         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1641                 goto retry;
1642 out:
1643         return ret;
1644 }
1645
1646 /* For write_end() in data=journal mode */
1647 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1648 {
1649         if (!buffer_mapped(bh) || buffer_freed(bh))
1650                 return 0;
1651         set_buffer_uptodate(bh);
1652         return ext4_handle_dirty_metadata(handle, NULL, bh);
1653 }
1654
1655 static int ext4_generic_write_end(struct file *file,
1656                                   struct address_space *mapping,
1657                                   loff_t pos, unsigned len, unsigned copied,
1658                                   struct page *page, void *fsdata)
1659 {
1660         int i_size_changed = 0;
1661         struct inode *inode = mapping->host;
1662         handle_t *handle = ext4_journal_current_handle();
1663
1664         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1665
1666         /*
1667          * No need to use i_size_read() here, the i_size
1668          * cannot change under us because we hold i_mutex.
1669          *
1670          * But it's important to update i_size while still holding page lock:
1671          * page writeout could otherwise come in and zero beyond i_size.
1672          */
1673         if (pos + copied > inode->i_size) {
1674                 i_size_write(inode, pos + copied);
1675                 i_size_changed = 1;
1676         }
1677
1678         if (pos + copied >  EXT4_I(inode)->i_disksize) {
1679                 /* We need to mark inode dirty even if
1680                  * new_i_size is less that inode->i_size
1681                  * bu greater than i_disksize.(hint delalloc)
1682                  */
1683                 ext4_update_i_disksize(inode, (pos + copied));
1684                 i_size_changed = 1;
1685         }
1686         unlock_page(page);
1687         page_cache_release(page);
1688
1689         /*
1690          * Don't mark the inode dirty under page lock. First, it unnecessarily
1691          * makes the holding time of page lock longer. Second, it forces lock
1692          * ordering of page lock and transaction start for journaling
1693          * filesystems.
1694          */
1695         if (i_size_changed)
1696                 ext4_mark_inode_dirty(handle, inode);
1697
1698         return copied;
1699 }
1700
1701 /*
1702  * We need to pick up the new inode size which generic_commit_write gave us
1703  * `file' can be NULL - eg, when called from page_symlink().
1704  *
1705  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1706  * buffers are managed internally.
1707  */
1708 static int ext4_ordered_write_end(struct file *file,
1709                                   struct address_space *mapping,
1710                                   loff_t pos, unsigned len, unsigned copied,
1711                                   struct page *page, void *fsdata)
1712 {
1713         handle_t *handle = ext4_journal_current_handle();
1714         struct inode *inode = mapping->host;
1715         int ret = 0, ret2;
1716
1717         trace_ext4_ordered_write_end(inode, pos, len, copied);
1718         ret = ext4_jbd2_file_inode(handle, inode);
1719
1720         if (ret == 0) {
1721                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1722                                                         page, fsdata);
1723                 copied = ret2;
1724                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1725                         /* if we have allocated more blocks and copied
1726                          * less. We will have blocks allocated outside
1727                          * inode->i_size. So truncate them
1728                          */
1729                         ext4_orphan_add(handle, inode);
1730                 if (ret2 < 0)
1731                         ret = ret2;
1732         }
1733         ret2 = ext4_journal_stop(handle);
1734         if (!ret)
1735                 ret = ret2;
1736
1737         if (pos + len > inode->i_size) {
1738                 ext4_truncate_failed_write(inode);
1739                 /*
1740                  * If truncate failed early the inode might still be
1741                  * on the orphan list; we need to make sure the inode
1742                  * is removed from the orphan list in that case.
1743                  */
1744                 if (inode->i_nlink)
1745                         ext4_orphan_del(NULL, inode);
1746         }
1747
1748
1749         return ret ? ret : copied;
1750 }
1751
1752 static int ext4_writeback_write_end(struct file *file,
1753                                     struct address_space *mapping,
1754                                     loff_t pos, unsigned len, unsigned copied,
1755                                     struct page *page, void *fsdata)
1756 {
1757         handle_t *handle = ext4_journal_current_handle();
1758         struct inode *inode = mapping->host;
1759         int ret = 0, ret2;
1760
1761         trace_ext4_writeback_write_end(inode, pos, len, copied);
1762         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1763                                                         page, fsdata);
1764         copied = ret2;
1765         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1766                 /* if we have allocated more blocks and copied
1767                  * less. We will have blocks allocated outside
1768                  * inode->i_size. So truncate them
1769                  */
1770                 ext4_orphan_add(handle, inode);
1771
1772         if (ret2 < 0)
1773                 ret = ret2;
1774
1775         ret2 = ext4_journal_stop(handle);
1776         if (!ret)
1777                 ret = ret2;
1778
1779         if (pos + len > inode->i_size) {
1780                 ext4_truncate_failed_write(inode);
1781                 /*
1782                  * If truncate failed early the inode might still be
1783                  * on the orphan list; we need to make sure the inode
1784                  * is removed from the orphan list in that case.
1785                  */
1786                 if (inode->i_nlink)
1787                         ext4_orphan_del(NULL, inode);
1788         }
1789
1790         return ret ? ret : copied;
1791 }
1792
1793 static int ext4_journalled_write_end(struct file *file,
1794                                      struct address_space *mapping,
1795                                      loff_t pos, unsigned len, unsigned copied,
1796                                      struct page *page, void *fsdata)
1797 {
1798         handle_t *handle = ext4_journal_current_handle();
1799         struct inode *inode = mapping->host;
1800         int ret = 0, ret2;
1801         int partial = 0;
1802         unsigned from, to;
1803         loff_t new_i_size;
1804
1805         trace_ext4_journalled_write_end(inode, pos, len, copied);
1806         from = pos & (PAGE_CACHE_SIZE - 1);
1807         to = from + len;
1808
1809         if (copied < len) {
1810                 if (!PageUptodate(page))
1811                         copied = 0;
1812                 page_zero_new_buffers(page, from+copied, to);
1813         }
1814
1815         ret = walk_page_buffers(handle, page_buffers(page), from,
1816                                 to, &partial, write_end_fn);
1817         if (!partial)
1818                 SetPageUptodate(page);
1819         new_i_size = pos + copied;
1820         if (new_i_size > inode->i_size)
1821                 i_size_write(inode, pos+copied);
1822         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1823         if (new_i_size > EXT4_I(inode)->i_disksize) {
1824                 ext4_update_i_disksize(inode, new_i_size);
1825                 ret2 = ext4_mark_inode_dirty(handle, inode);
1826                 if (!ret)
1827                         ret = ret2;
1828         }
1829
1830         unlock_page(page);
1831         page_cache_release(page);
1832         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1833                 /* if we have allocated more blocks and copied
1834                  * less. We will have blocks allocated outside
1835                  * inode->i_size. So truncate them
1836                  */
1837                 ext4_orphan_add(handle, inode);
1838
1839         ret2 = ext4_journal_stop(handle);
1840         if (!ret)
1841                 ret = ret2;
1842         if (pos + len > inode->i_size) {
1843                 ext4_truncate_failed_write(inode);
1844                 /*
1845                  * If truncate failed early the inode might still be
1846                  * on the orphan list; we need to make sure the inode
1847                  * is removed from the orphan list in that case.
1848                  */
1849                 if (inode->i_nlink)
1850                         ext4_orphan_del(NULL, inode);
1851         }
1852
1853         return ret ? ret : copied;
1854 }
1855
1856 /*
1857  * Reserve a single block located at lblock
1858  */
1859 static int ext4_da_reserve_space(struct inode *inode, sector_t lblock)
1860 {
1861         int retries = 0;
1862         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1863         struct ext4_inode_info *ei = EXT4_I(inode);
1864         unsigned long md_needed, md_reserved;
1865         int ret;
1866
1867         /*
1868          * recalculate the amount of metadata blocks to reserve
1869          * in order to allocate nrblocks
1870          * worse case is one extent per block
1871          */
1872 repeat:
1873         spin_lock(&ei->i_block_reservation_lock);
1874         md_reserved = ei->i_reserved_meta_blocks;
1875         md_needed = ext4_calc_metadata_amount(inode, lblock);
1876         trace_ext4_da_reserve_space(inode, md_needed);
1877         spin_unlock(&ei->i_block_reservation_lock);
1878
1879         /*
1880          * Make quota reservation here to prevent quota overflow
1881          * later. Real quota accounting is done at pages writeout
1882          * time.
1883          */
1884         ret = dquot_reserve_block(inode, md_needed + 1);
1885         if (ret)
1886                 return ret;
1887
1888         if (ext4_claim_free_blocks(sbi, md_needed + 1)) {
1889                 dquot_release_reservation_block(inode, md_needed + 1);
1890                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1891                         yield();
1892                         goto repeat;
1893                 }
1894                 return -ENOSPC;
1895         }
1896         spin_lock(&ei->i_block_reservation_lock);
1897         ei->i_reserved_data_blocks++;
1898         ei->i_reserved_meta_blocks += md_needed;
1899         spin_unlock(&ei->i_block_reservation_lock);
1900
1901         return 0;       /* success */
1902 }
1903
1904 static void ext4_da_release_space(struct inode *inode, int to_free)
1905 {
1906         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1907         struct ext4_inode_info *ei = EXT4_I(inode);
1908
1909         if (!to_free)
1910                 return;         /* Nothing to release, exit */
1911
1912         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1913
1914         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1915                 /*
1916                  * if there aren't enough reserved blocks, then the
1917                  * counter is messed up somewhere.  Since this
1918                  * function is called from invalidate page, it's
1919                  * harmless to return without any action.
1920                  */
1921                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1922                          "ino %lu, to_free %d with only %d reserved "
1923                          "data blocks\n", inode->i_ino, to_free,
1924                          ei->i_reserved_data_blocks);
1925                 WARN_ON(1);
1926                 to_free = ei->i_reserved_data_blocks;
1927         }
1928         ei->i_reserved_data_blocks -= to_free;
1929
1930         if (ei->i_reserved_data_blocks == 0) {
1931                 /*
1932                  * We can release all of the reserved metadata blocks
1933                  * only when we have written all of the delayed
1934                  * allocation blocks.
1935                  */
1936                 to_free += ei->i_reserved_meta_blocks;
1937                 ei->i_reserved_meta_blocks = 0;
1938                 ei->i_da_metadata_calc_len = 0;
1939         }
1940
1941         /* update fs dirty blocks counter */
1942         percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1943
1944         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1945
1946         dquot_release_reservation_block(inode, to_free);
1947 }
1948
1949 static void ext4_da_page_release_reservation(struct page *page,
1950                                              unsigned long offset)
1951 {
1952         int to_release = 0;
1953         struct buffer_head *head, *bh;
1954         unsigned int curr_off = 0;
1955
1956         head = page_buffers(page);
1957         bh = head;
1958         do {
1959                 unsigned int next_off = curr_off + bh->b_size;
1960
1961                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1962                         to_release++;
1963                         clear_buffer_delay(bh);
1964                 }
1965                 curr_off = next_off;
1966         } while ((bh = bh->b_this_page) != head);
1967         ext4_da_release_space(page->mapping->host, to_release);
1968 }
1969
1970 /*
1971  * Delayed allocation stuff
1972  */
1973
1974 /*
1975  * mpage_da_submit_io - walks through extent of pages and try to write
1976  * them with writepage() call back
1977  *
1978  * @mpd->inode: inode
1979  * @mpd->first_page: first page of the extent
1980  * @mpd->next_page: page after the last page of the extent
1981  *
1982  * By the time mpage_da_submit_io() is called we expect all blocks
1983  * to be allocated. this may be wrong if allocation failed.
1984  *
1985  * As pages are already locked by write_cache_pages(), we can't use it
1986  */
1987 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1988 {
1989         long pages_skipped;
1990         struct pagevec pvec;
1991         unsigned long index, end;
1992         int ret = 0, err, nr_pages, i;
1993         struct inode *inode = mpd->inode;
1994         struct address_space *mapping = inode->i_mapping;
1995
1996         BUG_ON(mpd->next_page <= mpd->first_page);
1997         /*
1998          * We need to start from the first_page to the next_page - 1
1999          * to make sure we also write the mapped dirty buffer_heads.
2000          * If we look at mpd->b_blocknr we would only be looking
2001          * at the currently mapped buffer_heads.
2002          */
2003         index = mpd->first_page;
2004         end = mpd->next_page - 1;
2005
2006         pagevec_init(&pvec, 0);
2007         while (index <= end) {
2008                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2009                 if (nr_pages == 0)
2010                         break;
2011                 for (i = 0; i < nr_pages; i++) {
2012                         struct page *page = pvec.pages[i];
2013
2014                         index = page->index;
2015                         if (index > end)
2016                                 break;
2017                         index++;
2018
2019                         BUG_ON(!PageLocked(page));
2020                         BUG_ON(PageWriteback(page));
2021
2022                         pages_skipped = mpd->wbc->pages_skipped;
2023                         err = mapping->a_ops->writepage(page, mpd->wbc);
2024                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
2025                                 /*
2026                                  * have successfully written the page
2027                                  * without skipping the same
2028                                  */
2029                                 mpd->pages_written++;
2030                         /*
2031                          * In error case, we have to continue because
2032                          * remaining pages are still locked
2033                          * XXX: unlock and re-dirty them?
2034                          */
2035                         if (ret == 0)
2036                                 ret = err;
2037                 }
2038                 pagevec_release(&pvec);
2039         }
2040         return ret;
2041 }
2042
2043 /*
2044  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
2045  *
2046  * @mpd->inode - inode to walk through
2047  * @exbh->b_blocknr - first block on a disk
2048  * @exbh->b_size - amount of space in bytes
2049  * @logical - first logical block to start assignment with
2050  *
2051  * the function goes through all passed space and put actual disk
2052  * block numbers into buffer heads, dropping BH_Delay and BH_Unwritten
2053  */
2054 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
2055                                  struct buffer_head *exbh)
2056 {
2057         struct inode *inode = mpd->inode;
2058         struct address_space *mapping = inode->i_mapping;
2059         int blocks = exbh->b_size >> inode->i_blkbits;
2060         sector_t pblock = exbh->b_blocknr, cur_logical;
2061         struct buffer_head *head, *bh;
2062         pgoff_t index, end;
2063         struct pagevec pvec;
2064         int nr_pages, i;
2065
2066         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2067         end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2068         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2069
2070         pagevec_init(&pvec, 0);
2071
2072         while (index <= end) {
2073                 /* XXX: optimize tail */
2074                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2075                 if (nr_pages == 0)
2076                         break;
2077                 for (i = 0; i < nr_pages; i++) {
2078                         struct page *page = pvec.pages[i];
2079
2080                         index = page->index;
2081                         if (index > end)
2082                                 break;
2083                         index++;
2084
2085                         BUG_ON(!PageLocked(page));
2086                         BUG_ON(PageWriteback(page));
2087                         BUG_ON(!page_has_buffers(page));
2088
2089                         bh = page_buffers(page);
2090                         head = bh;
2091
2092                         /* skip blocks out of the range */
2093                         do {
2094                                 if (cur_logical >= logical)
2095                                         break;
2096                                 cur_logical++;
2097                         } while ((bh = bh->b_this_page) != head);
2098
2099                         do {
2100                                 if (cur_logical >= logical + blocks)
2101                                         break;
2102
2103                                 if (buffer_delay(bh) ||
2104                                                 buffer_unwritten(bh)) {
2105
2106                                         BUG_ON(bh->b_bdev != inode->i_sb->s_bdev);
2107
2108                                         if (buffer_delay(bh)) {
2109                                                 clear_buffer_delay(bh);
2110                                                 bh->b_blocknr = pblock;
2111                                         } else {
2112                                                 /*
2113                                                  * unwritten already should have
2114                                                  * blocknr assigned. Verify that
2115                                                  */
2116                                                 clear_buffer_unwritten(bh);
2117                                                 BUG_ON(bh->b_blocknr != pblock);
2118                                         }
2119
2120                                 } else if (buffer_mapped(bh))
2121                                         BUG_ON(bh->b_blocknr != pblock);
2122
2123                                 if (buffer_uninit(exbh))
2124                                         set_buffer_uninit(bh);
2125                                 cur_logical++;
2126                                 pblock++;
2127                         } while ((bh = bh->b_this_page) != head);
2128                 }
2129                 pagevec_release(&pvec);
2130         }
2131 }
2132
2133
2134 /*
2135  * __unmap_underlying_blocks - just a helper function to unmap
2136  * set of blocks described by @bh
2137  */
2138 static inline void __unmap_underlying_blocks(struct inode *inode,
2139                                              struct buffer_head *bh)
2140 {
2141         struct block_device *bdev = inode->i_sb->s_bdev;
2142         int blocks, i;
2143
2144         blocks = bh->b_size >> inode->i_blkbits;
2145         for (i = 0; i < blocks; i++)
2146                 unmap_underlying_metadata(bdev, bh->b_blocknr + i);
2147 }
2148
2149 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2150                                         sector_t logical, long blk_cnt)
2151 {
2152         int nr_pages, i;
2153         pgoff_t index, end;
2154         struct pagevec pvec;
2155         struct inode *inode = mpd->inode;
2156         struct address_space *mapping = inode->i_mapping;
2157
2158         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2159         end   = (logical + blk_cnt - 1) >>
2160                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2161         while (index <= end) {
2162                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2163                 if (nr_pages == 0)
2164                         break;
2165                 for (i = 0; i < nr_pages; i++) {
2166                         struct page *page = pvec.pages[i];
2167                         if (page->index > end)
2168                                 break;
2169                         BUG_ON(!PageLocked(page));
2170                         BUG_ON(PageWriteback(page));
2171                         block_invalidatepage(page, 0);
2172                         ClearPageUptodate(page);
2173                         unlock_page(page);
2174                 }
2175                 index = pvec.pages[nr_pages - 1]->index + 1;
2176                 pagevec_release(&pvec);
2177         }
2178         return;
2179 }
2180
2181 static void ext4_print_free_blocks(struct inode *inode)
2182 {
2183         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2184         printk(KERN_CRIT "Total free blocks count %lld\n",
2185                ext4_count_free_blocks(inode->i_sb));
2186         printk(KERN_CRIT "Free/Dirty block details\n");
2187         printk(KERN_CRIT "free_blocks=%lld\n",
2188                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2189         printk(KERN_CRIT "dirty_blocks=%lld\n",
2190                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2191         printk(KERN_CRIT "Block reservation details\n");
2192         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2193                EXT4_I(inode)->i_reserved_data_blocks);
2194         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2195                EXT4_I(inode)->i_reserved_meta_blocks);
2196         return;
2197 }
2198
2199 /*
2200  * mpage_da_map_blocks - go through given space
2201  *
2202  * @mpd - bh describing space
2203  *
2204  * The function skips space we know is already mapped to disk blocks.
2205  *
2206  */
2207 static int mpage_da_map_blocks(struct mpage_da_data *mpd)
2208 {
2209         int err, blks, get_blocks_flags;
2210         struct buffer_head new;
2211         sector_t next = mpd->b_blocknr;
2212         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2213         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2214         handle_t *handle = NULL;
2215
2216         /*
2217          * We consider only non-mapped and non-allocated blocks
2218          */
2219         if ((mpd->b_state  & (1 << BH_Mapped)) &&
2220                 !(mpd->b_state & (1 << BH_Delay)) &&
2221                 !(mpd->b_state & (1 << BH_Unwritten)))
2222                 return 0;
2223
2224         /*
2225          * If we didn't accumulate anything to write simply return
2226          */
2227         if (!mpd->b_size)
2228                 return 0;
2229
2230         handle = ext4_journal_current_handle();
2231         BUG_ON(!handle);
2232
2233         /*
2234          * Call ext4_get_blocks() to allocate any delayed allocation
2235          * blocks, or to convert an uninitialized extent to be
2236          * initialized (in the case where we have written into
2237          * one or more preallocated blocks).
2238          *
2239          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2240          * indicate that we are on the delayed allocation path.  This
2241          * affects functions in many different parts of the allocation
2242          * call path.  This flag exists primarily because we don't
2243          * want to change *many* call functions, so ext4_get_blocks()
2244          * will set the magic i_delalloc_reserved_flag once the
2245          * inode's allocation semaphore is taken.
2246          *
2247          * If the blocks in questions were delalloc blocks, set
2248          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2249          * variables are updated after the blocks have been allocated.
2250          */
2251         new.b_state = 0;
2252         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
2253         if (ext4_should_dioread_nolock(mpd->inode))
2254                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2255         if (mpd->b_state & (1 << BH_Delay))
2256                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2257
2258         blks = ext4_get_blocks(handle, mpd->inode, next, max_blocks,
2259                                &new, get_blocks_flags);
2260         if (blks < 0) {
2261                 err = blks;
2262                 /*
2263                  * If get block returns with error we simply
2264                  * return. Later writepage will redirty the page and
2265                  * writepages will find the dirty page again
2266                  */
2267                 if (err == -EAGAIN)
2268                         return 0;
2269
2270                 if (err == -ENOSPC &&
2271                     ext4_count_free_blocks(mpd->inode->i_sb)) {
2272                         mpd->retval = err;
2273                         return 0;
2274                 }
2275
2276                 /*
2277                  * get block failure will cause us to loop in
2278                  * writepages, because a_ops->writepage won't be able
2279                  * to make progress. The page will be redirtied by
2280                  * writepage and writepages will again try to write
2281                  * the same.
2282                  */
2283                 ext4_msg(mpd->inode->i_sb, KERN_CRIT,
2284                          "delayed block allocation failed for inode %lu at "
2285                          "logical offset %llu with max blocks %zd with "
2286                          "error %d\n", mpd->inode->i_ino,
2287                          (unsigned long long) next,
2288                          mpd->b_size >> mpd->inode->i_blkbits, err);
2289                 printk(KERN_CRIT "This should not happen!!  "
2290                        "Data will be lost\n");
2291                 if (err == -ENOSPC) {
2292                         ext4_print_free_blocks(mpd->inode);
2293                 }
2294                 /* invalidate all the pages */
2295                 ext4_da_block_invalidatepages(mpd, next,
2296                                 mpd->b_size >> mpd->inode->i_blkbits);
2297                 return err;
2298         }
2299         BUG_ON(blks == 0);
2300
2301         new.b_size = (blks << mpd->inode->i_blkbits);
2302
2303         if (buffer_new(&new))
2304                 __unmap_underlying_blocks(mpd->inode, &new);
2305
2306         /*
2307          * If blocks are delayed marked, we need to
2308          * put actual blocknr and drop delayed bit
2309          */
2310         if ((mpd->b_state & (1 << BH_Delay)) ||
2311             (mpd->b_state & (1 << BH_Unwritten)))
2312                 mpage_put_bnr_to_bhs(mpd, next, &new);
2313
2314         if (ext4_should_order_data(mpd->inode)) {
2315                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2316                 if (err)
2317                         return err;
2318         }
2319
2320         /*
2321          * Update on-disk size along with block allocation.
2322          */
2323         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2324         if (disksize > i_size_read(mpd->inode))
2325                 disksize = i_size_read(mpd->inode);
2326         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2327                 ext4_update_i_disksize(mpd->inode, disksize);
2328                 return ext4_mark_inode_dirty(handle, mpd->inode);
2329         }
2330
2331         return 0;
2332 }
2333
2334 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2335                 (1 << BH_Delay) | (1 << BH_Unwritten))
2336
2337 /*
2338  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2339  *
2340  * @mpd->lbh - extent of blocks
2341  * @logical - logical number of the block in the file
2342  * @bh - bh of the block (used to access block's state)
2343  *
2344  * the function is used to collect contig. blocks in same state
2345  */
2346 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2347                                    sector_t logical, size_t b_size,
2348                                    unsigned long b_state)
2349 {
2350         sector_t next;
2351         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2352
2353         /*
2354          * XXX Don't go larger than mballoc is willing to allocate
2355          * This is a stopgap solution.  We eventually need to fold
2356          * mpage_da_submit_io() into this function and then call
2357          * ext4_get_blocks() multiple times in a loop
2358          */
2359         if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
2360                 goto flush_it;
2361
2362         /* check if thereserved journal credits might overflow */
2363         if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
2364                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2365                         /*
2366                          * With non-extent format we are limited by the journal
2367                          * credit available.  Total credit needed to insert
2368                          * nrblocks contiguous blocks is dependent on the
2369                          * nrblocks.  So limit nrblocks.
2370                          */
2371                         goto flush_it;
2372                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2373                                 EXT4_MAX_TRANS_DATA) {
2374                         /*
2375                          * Adding the new buffer_head would make it cross the
2376                          * allowed limit for which we have journal credit
2377                          * reserved. So limit the new bh->b_size
2378                          */
2379                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2380                                                 mpd->inode->i_blkbits;
2381                         /* we will do mpage_da_submit_io in the next loop */
2382                 }
2383         }
2384         /*
2385          * First block in the extent
2386          */
2387         if (mpd->b_size == 0) {
2388                 mpd->b_blocknr = logical;
2389                 mpd->b_size = b_size;
2390                 mpd->b_state = b_state & BH_FLAGS;
2391                 return;
2392         }
2393
2394         next = mpd->b_blocknr + nrblocks;
2395         /*
2396          * Can we merge the block to our big extent?
2397          */
2398         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2399                 mpd->b_size += b_size;
2400                 return;
2401         }
2402
2403 flush_it:
2404         /*
2405          * We couldn't merge the block to our extent, so we
2406          * need to flush current  extent and start new one
2407          */
2408         if (mpage_da_map_blocks(mpd) == 0)
2409                 mpage_da_submit_io(mpd);
2410         mpd->io_done = 1;
2411         return;
2412 }
2413
2414 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
2415 {
2416         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
2417 }
2418
2419 /*
2420  * __mpage_da_writepage - finds extent of pages and blocks
2421  *
2422  * @page: page to consider
2423  * @wbc: not used, we just follow rules
2424  * @data: context
2425  *
2426  * The function finds extents of pages and scan them for all blocks.
2427  */
2428 static int __mpage_da_writepage(struct page *page,
2429                                 struct writeback_control *wbc, void *data)
2430 {
2431         struct mpage_da_data *mpd = data;
2432         struct inode *inode = mpd->inode;
2433         struct buffer_head *bh, *head;
2434         sector_t logical;
2435
2436         if (mpd->io_done) {
2437                 /*
2438                  * Rest of the page in the page_vec
2439                  * redirty then and skip then. We will
2440                  * try to write them again after
2441                  * starting a new transaction
2442                  */
2443                 redirty_page_for_writepage(wbc, page);
2444                 unlock_page(page);
2445                 return MPAGE_DA_EXTENT_TAIL;
2446         }
2447         /*
2448          * Can we merge this page to current extent?
2449          */
2450         if (mpd->next_page != page->index) {
2451                 /*
2452                  * Nope, we can't. So, we map non-allocated blocks
2453                  * and start IO on them using writepage()
2454                  */
2455                 if (mpd->next_page != mpd->first_page) {
2456                         if (mpage_da_map_blocks(mpd) == 0)
2457                                 mpage_da_submit_io(mpd);
2458                         /*
2459                          * skip rest of the page in the page_vec
2460                          */
2461                         mpd->io_done = 1;
2462                         redirty_page_for_writepage(wbc, page);
2463                         unlock_page(page);
2464                         return MPAGE_DA_EXTENT_TAIL;
2465                 }
2466
2467                 /*
2468                  * Start next extent of pages ...
2469                  */
2470                 mpd->first_page = page->index;
2471
2472                 /*
2473                  * ... and blocks
2474                  */
2475                 mpd->b_size = 0;
2476                 mpd->b_state = 0;
2477                 mpd->b_blocknr = 0;
2478         }
2479
2480         mpd->next_page = page->index + 1;
2481         logical = (sector_t) page->index <<
2482                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2483
2484         if (!page_has_buffers(page)) {
2485                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2486                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2487                 if (mpd->io_done)
2488                         return MPAGE_DA_EXTENT_TAIL;
2489         } else {
2490                 /*
2491                  * Page with regular buffer heads, just add all dirty ones
2492                  */
2493                 head = page_buffers(page);
2494                 bh = head;
2495                 do {
2496                         BUG_ON(buffer_locked(bh));
2497                         /*
2498                          * We need to try to allocate
2499                          * unmapped blocks in the same page.
2500                          * Otherwise we won't make progress
2501                          * with the page in ext4_writepage
2502                          */
2503                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2504                                 mpage_add_bh_to_extent(mpd, logical,
2505                                                        bh->b_size,
2506                                                        bh->b_state);
2507                                 if (mpd->io_done)
2508                                         return MPAGE_DA_EXTENT_TAIL;
2509                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2510                                 /*
2511                                  * mapped dirty buffer. We need to update
2512                                  * the b_state because we look at
2513                                  * b_state in mpage_da_map_blocks. We don't
2514                                  * update b_size because if we find an
2515                                  * unmapped buffer_head later we need to
2516                                  * use the b_state flag of that buffer_head.
2517                                  */
2518                                 if (mpd->b_size == 0)
2519                                         mpd->b_state = bh->b_state & BH_FLAGS;
2520                         }
2521                         logical++;
2522                 } while ((bh = bh->b_this_page) != head);
2523         }
2524
2525         return 0;
2526 }
2527
2528 /*
2529  * This is a special get_blocks_t callback which is used by
2530  * ext4_da_write_begin().  It will either return mapped block or
2531  * reserve space for a single block.
2532  *
2533  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2534  * We also have b_blocknr = -1 and b_bdev initialized properly
2535  *
2536  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2537  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2538  * initialized properly.
2539  */
2540 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2541                                   struct buffer_head *bh_result, int create)
2542 {
2543         int ret = 0;
2544         sector_t invalid_block = ~((sector_t) 0xffff);
2545
2546         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2547                 invalid_block = ~0;
2548
2549         BUG_ON(create == 0);
2550         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2551
2552         /*
2553          * first, we need to know whether the block is allocated already
2554          * preallocated blocks are unmapped but should treated
2555          * the same as allocated blocks.
2556          */
2557         ret = ext4_get_blocks(NULL, inode, iblock, 1,  bh_result, 0);
2558         if ((ret == 0) && !buffer_delay(bh_result)) {
2559                 /* the block isn't (pre)allocated yet, let's reserve space */
2560                 /*
2561                  * XXX: __block_prepare_write() unmaps passed block,
2562                  * is it OK?
2563                  */
2564                 ret = ext4_da_reserve_space(inode, iblock);
2565                 if (ret)
2566                         /* not enough space to reserve */
2567                         return ret;
2568
2569                 map_bh(bh_result, inode->i_sb, invalid_block);
2570                 set_buffer_new(bh_result);
2571                 set_buffer_delay(bh_result);
2572         } else if (ret > 0) {
2573                 bh_result->b_size = (ret << inode->i_blkbits);
2574                 if (buffer_unwritten(bh_result)) {
2575                         /* A delayed write to unwritten bh should
2576                          * be marked new and mapped.  Mapped ensures
2577                          * that we don't do get_block multiple times
2578                          * when we write to the same offset and new
2579                          * ensures that we do proper zero out for
2580                          * partial write.
2581                          */
2582                         set_buffer_new(bh_result);
2583                         set_buffer_mapped(bh_result);
2584                 }
2585                 ret = 0;
2586         }
2587
2588         return ret;
2589 }
2590
2591 /*
2592  * This function is used as a standard get_block_t calback function
2593  * when there is no desire to allocate any blocks.  It is used as a
2594  * callback function for block_prepare_write(), nobh_writepage(), and
2595  * block_write_full_page().  These functions should only try to map a
2596  * single block at a time.
2597  *
2598  * Since this function doesn't do block allocations even if the caller
2599  * requests it by passing in create=1, it is critically important that
2600  * any caller checks to make sure that any buffer heads are returned
2601  * by this function are either all already mapped or marked for
2602  * delayed allocation before calling nobh_writepage() or
2603  * block_write_full_page().  Otherwise, b_blocknr could be left
2604  * unitialized, and the page write functions will be taken by
2605  * surprise.
2606  */
2607 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2608                                    struct buffer_head *bh_result, int create)
2609 {
2610         int ret = 0;
2611         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2612
2613         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2614
2615         /*
2616          * we don't want to do block allocation in writepage
2617          * so call get_block_wrap with create = 0
2618          */
2619         ret = ext4_get_blocks(NULL, inode, iblock, max_blocks, bh_result, 0);
2620         if (ret > 0) {
2621                 bh_result->b_size = (ret << inode->i_blkbits);
2622                 ret = 0;
2623         }
2624         return ret;
2625 }
2626
2627 static int bget_one(handle_t *handle, struct buffer_head *bh)
2628 {
2629         get_bh(bh);
2630         return 0;
2631 }
2632
2633 static int bput_one(handle_t *handle, struct buffer_head *bh)
2634 {
2635         put_bh(bh);
2636         return 0;
2637 }
2638
2639 static int __ext4_journalled_writepage(struct page *page,
2640                                        unsigned int len)
2641 {
2642         struct address_space *mapping = page->mapping;
2643         struct inode *inode = mapping->host;
2644         struct buffer_head *page_bufs;
2645         handle_t *handle = NULL;
2646         int ret = 0;
2647         int err;
2648
2649         page_bufs = page_buffers(page);
2650         BUG_ON(!page_bufs);
2651         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2652         /* As soon as we unlock the page, it can go away, but we have
2653          * references to buffers so we are safe */
2654         unlock_page(page);
2655
2656         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2657         if (IS_ERR(handle)) {
2658                 ret = PTR_ERR(handle);
2659                 goto out;
2660         }
2661
2662         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2663                                 do_journal_get_write_access);
2664
2665         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2666                                 write_end_fn);
2667         if (ret == 0)
2668                 ret = err;
2669         err = ext4_journal_stop(handle);
2670         if (!ret)
2671                 ret = err;
2672
2673         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2674         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2675 out:
2676         return ret;
2677 }
2678
2679 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
2680 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
2681
2682 /*
2683  * Note that we don't need to start a transaction unless we're journaling data
2684  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2685  * need to file the inode to the transaction's list in ordered mode because if
2686  * we are writing back data added by write(), the inode is already there and if
2687  * we are writing back data modified via mmap(), noone guarantees in which
2688  * transaction the data will hit the disk. In case we are journaling data, we
2689  * cannot start transaction directly because transaction start ranks above page
2690  * lock so we have to do some magic.
2691  *
2692  * This function can get called via...
2693  *   - ext4_da_writepages after taking page lock (have journal handle)
2694  *   - journal_submit_inode_data_buffers (no journal handle)
2695  *   - shrink_page_list via pdflush (no journal handle)
2696  *   - grab_page_cache when doing write_begin (have journal handle)
2697  *
2698  * We don't do any block allocation in this function. If we have page with
2699  * multiple blocks we need to write those buffer_heads that are mapped. This
2700  * is important for mmaped based write. So if we do with blocksize 1K
2701  * truncate(f, 1024);
2702  * a = mmap(f, 0, 4096);
2703  * a[0] = 'a';
2704  * truncate(f, 4096);
2705  * we have in the page first buffer_head mapped via page_mkwrite call back
2706  * but other bufer_heads would be unmapped but dirty(dirty done via the
2707  * do_wp_page). So writepage should write the first block. If we modify
2708  * the mmap area beyond 1024 we will again get a page_fault and the
2709  * page_mkwrite callback will do the block allocation and mark the
2710  * buffer_heads mapped.
2711  *
2712  * We redirty the page if we have any buffer_heads that is either delay or
2713  * unwritten in the page.
2714  *
2715  * We can get recursively called as show below.
2716  *
2717  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2718  *              ext4_writepage()
2719  *
2720  * But since we don't do any block allocation we should not deadlock.
2721  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2722  */
2723 static int ext4_writepage(struct page *page,
2724                           struct writeback_control *wbc)
2725 {
2726         int ret = 0;
2727         loff_t size;
2728         unsigned int len;
2729         struct buffer_head *page_bufs = NULL;
2730         struct inode *inode = page->mapping->host;
2731
2732         trace_ext4_writepage(inode, page);
2733         size = i_size_read(inode);
2734         if (page->index == size >> PAGE_CACHE_SHIFT)
2735                 len = size & ~PAGE_CACHE_MASK;
2736         else
2737                 len = PAGE_CACHE_SIZE;
2738
2739         if (page_has_buffers(page)) {
2740                 page_bufs = page_buffers(page);
2741                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2742                                         ext4_bh_delay_or_unwritten)) {
2743                         /*
2744                          * We don't want to do  block allocation
2745                          * So redirty the page and return
2746                          * We may reach here when we do a journal commit
2747                          * via journal_submit_inode_data_buffers.
2748                          * If we don't have mapping block we just ignore
2749                          * them. We can also reach here via shrink_page_list
2750                          */
2751                         redirty_page_for_writepage(wbc, page);
2752                         unlock_page(page);
2753                         return 0;
2754                 }
2755         } else {
2756                 /*
2757                  * The test for page_has_buffers() is subtle:
2758                  * We know the page is dirty but it lost buffers. That means
2759                  * that at some moment in time after write_begin()/write_end()
2760                  * has been called all buffers have been clean and thus they
2761                  * must have been written at least once. So they are all
2762                  * mapped and we can happily proceed with mapping them
2763                  * and writing the page.
2764                  *
2765                  * Try to initialize the buffer_heads and check whether
2766                  * all are mapped and non delay. We don't want to
2767                  * do block allocation here.
2768                  */
2769                 ret = block_prepare_write(page, 0, len,
2770                                           noalloc_get_block_write);
2771                 if (!ret) {
2772                         page_bufs = page_buffers(page);
2773                         /* check whether all are mapped and non delay */
2774                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2775                                                 ext4_bh_delay_or_unwritten)) {
2776                                 redirty_page_for_writepage(wbc, page);
2777                                 unlock_page(page);
2778                                 return 0;
2779                         }
2780                 } else {
2781                         /*
2782                          * We can't do block allocation here
2783                          * so just redity the page and unlock
2784                          * and return
2785                          */
2786                         redirty_page_for_writepage(wbc, page);
2787                         unlock_page(page);
2788                         return 0;
2789                 }
2790                 /* now mark the buffer_heads as dirty and uptodate */
2791                 block_commit_write(page, 0, len);
2792         }
2793
2794         if (PageChecked(page) && ext4_should_journal_data(inode)) {
2795                 /*
2796                  * It's mmapped pagecache.  Add buffers and journal it.  There
2797                  * doesn't seem much point in redirtying the page here.
2798                  */
2799                 ClearPageChecked(page);
2800                 return __ext4_journalled_writepage(page, len);
2801         }
2802
2803         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2804                 ret = nobh_writepage(page, noalloc_get_block_write, wbc);
2805         else if (page_bufs && buffer_uninit(page_bufs)) {
2806                 ext4_set_bh_endio(page_bufs, inode);
2807                 ret = block_write_full_page_endio(page, noalloc_get_block_write,
2808                                             wbc, ext4_end_io_buffer_write);
2809         } else
2810                 ret = block_write_full_page(page, noalloc_get_block_write,
2811                                             wbc);
2812
2813         return ret;
2814 }
2815
2816 /*
2817  * This is called via ext4_da_writepages() to
2818  * calulate the total number of credits to reserve to fit
2819  * a single extent allocation into a single transaction,
2820  * ext4_da_writpeages() will loop calling this before
2821  * the block allocation.
2822  */
2823
2824 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2825 {
2826         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2827
2828         /*
2829          * With non-extent format the journal credit needed to
2830          * insert nrblocks contiguous block is dependent on
2831          * number of contiguous block. So we will limit
2832          * number of contiguous block to a sane value
2833          */
2834         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) &&
2835             (max_blocks > EXT4_MAX_TRANS_DATA))
2836                 max_blocks = EXT4_MAX_TRANS_DATA;
2837
2838         return ext4_chunk_trans_blocks(inode, max_blocks);
2839 }
2840
2841 static int ext4_da_writepages(struct address_space *mapping,
2842                               struct writeback_control *wbc)
2843 {
2844         pgoff_t index;
2845         int range_whole = 0;
2846         handle_t *handle = NULL;
2847         struct mpage_da_data mpd;
2848         struct inode *inode = mapping->host;
2849         int no_nrwrite_index_update;
2850         int pages_written = 0;
2851         long pages_skipped;
2852         unsigned int max_pages;
2853         int range_cyclic, cycled = 1, io_done = 0;
2854         int needed_blocks, ret = 0;
2855         long desired_nr_to_write, nr_to_writebump = 0;
2856         loff_t range_start = wbc->range_start;
2857         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2858
2859         trace_ext4_da_writepages(inode, wbc);
2860
2861         /*
2862          * No pages to write? This is mainly a kludge to avoid starting
2863          * a transaction for special inodes like journal inode on last iput()
2864          * because that could violate lock ordering on umount
2865          */
2866         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2867                 return 0;
2868
2869         /*
2870          * If the filesystem has aborted, it is read-only, so return
2871          * right away instead of dumping stack traces later on that
2872          * will obscure the real source of the problem.  We test
2873          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2874          * the latter could be true if the filesystem is mounted
2875          * read-only, and in that case, ext4_da_writepages should
2876          * *never* be called, so if that ever happens, we would want
2877          * the stack trace.
2878          */
2879         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2880                 return -EROFS;
2881
2882         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2883                 range_whole = 1;
2884
2885         range_cyclic = wbc->range_cyclic;
2886         if (wbc->range_cyclic) {
2887                 index = mapping->writeback_index;
2888                 if (index)
2889                         cycled = 0;
2890                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2891                 wbc->range_end  = LLONG_MAX;
2892                 wbc->range_cyclic = 0;
2893         } else
2894                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2895
2896         /*
2897          * This works around two forms of stupidity.  The first is in
2898          * the writeback code, which caps the maximum number of pages
2899          * written to be 1024 pages.  This is wrong on multiple
2900          * levels; different architectues have a different page size,
2901          * which changes the maximum amount of data which gets
2902          * written.  Secondly, 4 megabytes is way too small.  XFS
2903          * forces this value to be 16 megabytes by multiplying
2904          * nr_to_write parameter by four, and then relies on its
2905          * allocator to allocate larger extents to make them
2906          * contiguous.  Unfortunately this brings us to the second
2907          * stupidity, which is that ext4's mballoc code only allocates
2908          * at most 2048 blocks.  So we force contiguous writes up to
2909          * the number of dirty blocks in the inode, or
2910          * sbi->max_writeback_mb_bump whichever is smaller.
2911          */
2912         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2913         if (!range_cyclic && range_whole)
2914                 desired_nr_to_write = wbc->nr_to_write * 8;
2915         else
2916                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2917                                                            max_pages);
2918         if (desired_nr_to_write > max_pages)
2919                 desired_nr_to_write = max_pages;
2920
2921         if (wbc->nr_to_write < desired_nr_to_write) {
2922                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2923                 wbc->nr_to_write = desired_nr_to_write;
2924         }
2925
2926         mpd.wbc = wbc;
2927         mpd.inode = mapping->host;
2928
2929         /*
2930          * we don't want write_cache_pages to update
2931          * nr_to_write and writeback_index
2932          */
2933         no_nrwrite_index_update = wbc->no_nrwrite_index_update;
2934         wbc->no_nrwrite_index_update = 1;
2935         pages_skipped = wbc->pages_skipped;
2936
2937 retry:
2938         while (!ret && wbc->nr_to_write > 0) {
2939
2940                 /*
2941                  * we  insert one extent at a time. So we need
2942                  * credit needed for single extent allocation.
2943                  * journalled mode is currently not supported
2944                  * by delalloc
2945                  */
2946                 BUG_ON(ext4_should_journal_data(inode));
2947                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2948
2949                 /* start a new transaction*/
2950                 handle = ext4_journal_start(inode, needed_blocks);
2951                 if (IS_ERR(handle)) {
2952                         ret = PTR_ERR(handle);
2953                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2954                                "%ld pages, ino %lu; err %d\n", __func__,
2955                                 wbc->nr_to_write, inode->i_ino, ret);
2956                         goto out_writepages;
2957                 }
2958
2959                 /*
2960                  * Now call __mpage_da_writepage to find the next
2961                  * contiguous region of logical blocks that need
2962                  * blocks to be allocated by ext4.  We don't actually
2963                  * submit the blocks for I/O here, even though
2964                  * write_cache_pages thinks it will, and will set the
2965                  * pages as clean for write before calling
2966                  * __mpage_da_writepage().
2967                  */
2968                 mpd.b_size = 0;
2969                 mpd.b_state = 0;
2970                 mpd.b_blocknr = 0;
2971                 mpd.first_page = 0;
2972                 mpd.next_page = 0;
2973                 mpd.io_done = 0;
2974                 mpd.pages_written = 0;
2975                 mpd.retval = 0;
2976                 ret = write_cache_pages(mapping, wbc, __mpage_da_writepage,
2977                                         &mpd);
2978                 /*
2979                  * If we have a contiguous extent of pages and we
2980                  * haven't done the I/O yet, map the blocks and submit
2981                  * them for I/O.
2982                  */
2983                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2984                         if (mpage_da_map_blocks(&mpd) == 0)
2985                                 mpage_da_submit_io(&mpd);
2986                         mpd.io_done = 1;
2987                         ret = MPAGE_DA_EXTENT_TAIL;
2988                 }
2989                 trace_ext4_da_write_pages(inode, &mpd);
2990                 wbc->nr_to_write -= mpd.pages_written;
2991
2992                 ext4_journal_stop(handle);
2993
2994                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2995                         /* commit the transaction which would
2996                          * free blocks released in the transaction
2997                          * and try again
2998                          */
2999                         jbd2_journal_force_commit_nested(sbi->s_journal);
3000                         wbc->pages_skipped = pages_skipped;
3001                         ret = 0;
3002                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
3003                         /*
3004                          * got one extent now try with
3005                          * rest of the pages
3006                          */
3007                         pages_written += mpd.pages_written;
3008                         wbc->pages_skipped = pages_skipped;
3009                         ret = 0;
3010                         io_done = 1;
3011                 } else if (wbc->nr_to_write)
3012                         /*
3013                          * There is no more writeout needed
3014                          * or we requested for a noblocking writeout
3015                          * and we found the device congested
3016                          */
3017                         break;
3018         }
3019         if (!io_done && !cycled) {
3020                 cycled = 1;
3021                 index = 0;
3022                 wbc->range_start = index << PAGE_CACHE_SHIFT;
3023                 wbc->range_end  = mapping->writeback_index - 1;
3024                 goto retry;
3025         }
3026         if (pages_skipped != wbc->pages_skipped)
3027                 ext4_msg(inode->i_sb, KERN_CRIT,
3028                          "This should not happen leaving %s "
3029                          "with nr_to_write = %ld ret = %d\n",
3030                          __func__, wbc->nr_to_write, ret);
3031
3032         /* Update index */
3033         index += pages_written;
3034         wbc->range_cyclic = range_cyclic;
3035         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
3036                 /*
3037                  * set the writeback_index so that range_cyclic
3038                  * mode will write it back later
3039                  */
3040                 mapping->writeback_index = index;
3041
3042 out_writepages:
3043         if (!no_nrwrite_index_update)
3044                 wbc->no_nrwrite_index_update = 0;
3045         wbc->nr_to_write -= nr_to_writebump;
3046         wbc->range_start = range_start;
3047         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
3048         return ret;
3049 }
3050
3051 #define FALL_BACK_TO_NONDELALLOC 1
3052 static int ext4_nonda_switch(struct super_block *sb)
3053 {
3054         s64 free_blocks, dirty_blocks;
3055         struct ext4_sb_info *sbi = EXT4_SB(sb);
3056
3057         /*
3058          * switch to non delalloc mode if we are running low
3059          * on free block. The free block accounting via percpu
3060          * counters can get slightly wrong with percpu_counter_batch getting
3061          * accumulated on each CPU without updating global counters
3062          * Delalloc need an accurate free block accounting. So switch
3063          * to non delalloc when we are near to error range.
3064          */
3065         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
3066         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
3067         if (2 * free_blocks < 3 * dirty_blocks ||
3068                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
3069                 /*
3070                  * free block count is less than 150% of dirty blocks
3071                  * or free blocks is less than watermark
3072                  */
3073                 return 1;
3074         }
3075         /*
3076          * Even if we don't switch but are nearing capacity,
3077          * start pushing delalloc when 1/2 of free blocks are dirty.
3078          */
3079         if (free_blocks < 2 * dirty_blocks)
3080                 writeback_inodes_sb_if_idle(sb);
3081
3082         return 0;
3083 }
3084
3085 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3086                                loff_t pos, unsigned len, unsigned flags,
3087                                struct page **pagep, void **fsdata)
3088 {
3089         int ret, retries = 0, quota_retries = 0;
3090         struct page *page;
3091         pgoff_t index;
3092         unsigned from, to;
3093         struct inode *inode = mapping->host;
3094         handle_t *handle;
3095
3096         index = pos >> PAGE_CACHE_SHIFT;
3097         from = pos & (PAGE_CACHE_SIZE - 1);
3098         to = from + len;
3099
3100         if (ext4_nonda_switch(inode->i_sb)) {
3101                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3102                 return ext4_write_begin(file, mapping, pos,
3103                                         len, flags, pagep, fsdata);
3104         }
3105         *fsdata = (void *)0;
3106         trace_ext4_da_write_begin(inode, pos, len, flags);
3107 retry:
3108         /*
3109          * With delayed allocation, we don't log the i_disksize update
3110          * if there is delayed block allocation. But we still need
3111          * to journalling the i_disksize update if writes to the end
3112          * of file which has an already mapped buffer.
3113          */
3114         handle = ext4_journal_start(inode, 1);
3115         if (IS_ERR(handle)) {
3116                 ret = PTR_ERR(handle);
3117                 goto out;
3118         }
3119         /* We cannot recurse into the filesystem as the transaction is already
3120          * started */
3121         flags |= AOP_FLAG_NOFS;
3122
3123         page = grab_cache_page_write_begin(mapping, index, flags);
3124         if (!page) {
3125                 ext4_journal_stop(handle);
3126                 ret = -ENOMEM;
3127                 goto out;
3128         }
3129         *pagep = page;
3130
3131         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
3132                                 ext4_da_get_block_prep);
3133         if (ret < 0) {
3134                 unlock_page(page);
3135                 ext4_journal_stop(handle);
3136                 page_cache_release(page);
3137                 /*
3138                  * block_write_begin may have instantiated a few blocks
3139                  * outside i_size.  Trim these off again. Don't need
3140                  * i_size_read because we hold i_mutex.
3141                  */
3142                 if (pos + len > inode->i_size)
3143                         ext4_truncate_failed_write(inode);
3144         }
3145
3146         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3147                 goto retry;
3148
3149         if ((ret == -EDQUOT) &&
3150             EXT4_I(inode)->i_reserved_meta_blocks &&
3151             (quota_retries++ < 3)) {
3152                 /*
3153                  * Since we often over-estimate the number of meta
3154                  * data blocks required, we may sometimes get a
3155                  * spurios out of quota error even though there would
3156                  * be enough space once we write the data blocks and
3157                  * find out how many meta data blocks were _really_
3158                  * required.  So try forcing the inode write to see if
3159                  * that helps.
3160                  */
3161                 write_inode_now(inode, (quota_retries == 3));
3162                 goto retry;
3163         }
3164 out:
3165         return ret;
3166 }
3167
3168 /*
3169  * Check if we should update i_disksize
3170  * when write to the end of file but not require block allocation
3171  */
3172 static int ext4_da_should_update_i_disksize(struct page *page,
3173                                             unsigned long offset)
3174 {
3175         struct buffer_head *bh;
3176         struct inode *inode = page->mapping->host;
3177         unsigned int idx;
3178         int i;
3179
3180         bh = page_buffers(page);
3181         idx = offset >> inode->i_blkbits;
3182
3183         for (i = 0; i < idx; i++)
3184                 bh = bh->b_this_page;
3185
3186         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3187                 return 0;
3188         return 1;
3189 }
3190
3191 static int ext4_da_write_end(struct file *file,
3192                              struct address_space *mapping,
3193                              loff_t pos, unsigned len, unsigned copied,
3194                              struct page *page, void *fsdata)
3195 {
3196         struct inode *inode = mapping->host;
3197         int ret = 0, ret2;
3198         handle_t *handle = ext4_journal_current_handle();
3199         loff_t new_i_size;
3200         unsigned long start, end;
3201         int write_mode = (int)(unsigned long)fsdata;
3202
3203         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
3204                 if (ext4_should_order_data(inode)) {
3205                         return ext4_ordered_write_end(file, mapping, pos,
3206                                         len, copied, page, fsdata);
3207                 } else if (ext4_should_writeback_data(inode)) {
3208                         return ext4_writeback_write_end(file, mapping, pos,
3209                                         len, copied, page, fsdata);
3210                 } else {
3211                         BUG();
3212                 }
3213         }
3214
3215         trace_ext4_da_write_end(inode, pos, len, copied);
3216         start = pos & (PAGE_CACHE_SIZE - 1);
3217         end = start + copied - 1;
3218
3219         /*
3220          * generic_write_end() will run mark_inode_dirty() if i_size
3221          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3222          * into that.
3223          */
3224
3225         new_i_size = pos + copied;
3226         if (new_i_size > EXT4_I(inode)->i_disksize) {
3227                 if (ext4_da_should_update_i_disksize(page, end)) {
3228                         down_write(&EXT4_I(inode)->i_data_sem);
3229                         if (new_i_size > EXT4_I(inode)->i_disksize) {
3230                                 /*
3231                                  * Updating i_disksize when extending file
3232                                  * without needing block allocation
3233                                  */
3234                                 if (ext4_should_order_data(inode))
3235                                         ret = ext4_jbd2_file_inode(handle,
3236                                                                    inode);
3237
3238                                 EXT4_I(inode)->i_disksize = new_i_size;
3239                         }
3240                         up_write(&EXT4_I(inode)->i_data_sem);
3241                         /* We need to mark inode dirty even if
3242                          * new_i_size is less that inode->i_size
3243                          * bu greater than i_disksize.(hint delalloc)
3244                          */
3245                         ext4_mark_inode_dirty(handle, inode);
3246                 }
3247         }
3248         ret2 = generic_write_end(file, mapping, pos, len, copied,
3249                                                         page, fsdata);
3250         copied = ret2;
3251         if (ret2 < 0)
3252                 ret = ret2;
3253         ret2 = ext4_journal_stop(handle);
3254         if (!ret)
3255                 ret = ret2;
3256
3257         return ret ? ret : copied;
3258 }
3259
3260 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
3261 {
3262         /*
3263          * Drop reserved blocks
3264          */
3265         BUG_ON(!PageLocked(page));
3266         if (!page_has_buffers(page))
3267                 goto out;
3268
3269         ext4_da_page_release_reservation(page, offset);
3270
3271 out:
3272         ext4_invalidatepage(page, offset);
3273
3274         return;
3275 }
3276
3277 /*
3278  * Force all delayed allocation blocks to be allocated for a given inode.
3279  */
3280 int ext4_alloc_da_blocks(struct inode *inode)
3281 {
3282         trace_ext4_alloc_da_blocks(inode);
3283
3284         if (!EXT4_I(inode)->i_reserved_data_blocks &&
3285             !EXT4_I(inode)->i_reserved_meta_blocks)
3286                 return 0;
3287
3288         /*
3289          * We do something simple for now.  The filemap_flush() will
3290          * also start triggering a write of the data blocks, which is
3291          * not strictly speaking necessary (and for users of
3292          * laptop_mode, not even desirable).  However, to do otherwise
3293          * would require replicating code paths in:
3294          *
3295          * ext4_da_writepages() ->
3296          *    write_cache_pages() ---> (via passed in callback function)
3297          *        __mpage_da_writepage() -->
3298          *           mpage_add_bh_to_extent()
3299          *           mpage_da_map_blocks()
3300          *
3301          * The problem is that write_cache_pages(), located in
3302          * mm/page-writeback.c, marks pages clean in preparation for
3303          * doing I/O, which is not desirable if we're not planning on
3304          * doing I/O at all.
3305          *
3306          * We could call write_cache_pages(), and then redirty all of
3307          * the pages by calling redirty_page_for_writeback() but that
3308          * would be ugly in the extreme.  So instead we would need to
3309          * replicate parts of the code in the above functions,
3310          * simplifying them becuase we wouldn't actually intend to
3311          * write out the pages, but rather only collect contiguous
3312          * logical block extents, call the multi-block allocator, and
3313          * then update the buffer heads with the block allocations.
3314          *
3315          * For now, though, we'll cheat by calling filemap_flush(),
3316          * which will map the blocks, and start the I/O, but not
3317          * actually wait for the I/O to complete.
3318          */
3319         return filemap_flush(inode->i_mapping);
3320 }
3321
3322 /*
3323  * bmap() is special.  It gets used by applications such as lilo and by
3324  * the swapper to find the on-disk block of a specific piece of data.
3325  *
3326  * Naturally, this is dangerous if the block concerned is still in the
3327  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3328  * filesystem and enables swap, then they may get a nasty shock when the
3329  * data getting swapped to that swapfile suddenly gets overwritten by
3330  * the original zero's written out previously to the journal and
3331  * awaiting writeback in the kernel's buffer cache.
3332  *
3333  * So, if we see any bmap calls here on a modified, data-journaled file,
3334  * take extra steps to flush any blocks which might be in the cache.
3335  */
3336 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3337 {
3338         struct inode *inode = mapping->host;
3339         journal_t *journal;
3340         int err;
3341
3342         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3343                         test_opt(inode->i_sb, DELALLOC)) {
3344                 /*
3345                  * With delalloc we want to sync the file
3346                  * so that we can make sure we allocate
3347                  * blocks for file
3348                  */
3349                 filemap_write_and_wait(mapping);
3350         }
3351
3352         if (EXT4_JOURNAL(inode) &&
3353             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3354                 /*
3355                  * This is a REALLY heavyweight approach, but the use of
3356                  * bmap on dirty files is expected to be extremely rare:
3357                  * only if we run lilo or swapon on a freshly made file
3358                  * do we expect this to happen.
3359                  *
3360                  * (bmap requires CAP_SYS_RAWIO so this does not
3361                  * represent an unprivileged user DOS attack --- we'd be
3362                  * in trouble if mortal users could trigger this path at
3363                  * will.)
3364                  *
3365                  * NB. EXT4_STATE_JDATA is not set on files other than
3366                  * regular files.  If somebody wants to bmap a directory
3367                  * or symlink and gets confused because the buffer
3368                  * hasn't yet been flushed to disk, they deserve
3369                  * everything they get.
3370                  */
3371
3372                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3373                 journal = EXT4_JOURNAL(inode);
3374                 jbd2_journal_lock_updates(journal);
3375                 err = jbd2_journal_flush(journal);
3376                 jbd2_journal_unlock_updates(journal);
3377
3378                 if (err)
3379                         return 0;
3380         }
3381
3382         return generic_block_bmap(mapping, block, ext4_get_block);
3383 }
3384
3385 static int ext4_readpage(struct file *file, struct page *page)
3386 {
3387         return mpage_readpage(page, ext4_get_block);
3388 }
3389
3390 static int
3391 ext4_readpages(struct file *file, struct address_space *mapping,
3392                 struct list_head *pages, unsigned nr_pages)
3393 {
3394         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3395 }
3396
3397 static void ext4_free_io_end(ext4_io_end_t *io)
3398 {
3399         BUG_ON(!io);
3400         if (io->page)
3401                 put_page(io->page);
3402         iput(io->inode);
3403         kfree(io);
3404 }
3405
3406 static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
3407 {
3408         struct buffer_head *head, *bh;
3409         unsigned int curr_off = 0;
3410
3411         if (!page_has_buffers(page))
3412                 return;
3413         head = bh = page_buffers(page);
3414         do {
3415                 if (offset <= curr_off && test_clear_buffer_uninit(bh)
3416                                         && bh->b_private) {
3417                         ext4_free_io_end(bh->b_private);
3418                         bh->b_private = NULL;
3419                         bh->b_end_io = NULL;
3420                 }
3421                 curr_off = curr_off + bh->b_size;
3422                 bh = bh->b_this_page;
3423         } while (bh != head);
3424 }
3425
3426 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3427 {
3428         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3429
3430         /*
3431          * free any io_end structure allocated for buffers to be discarded
3432          */
3433         if (ext4_should_dioread_nolock(page->mapping->host))
3434                 ext4_invalidatepage_free_endio(page, offset);
3435         /*
3436          * If it's a full truncate we just forget about the pending dirtying
3437          */
3438         if (offset == 0)
3439                 ClearPageChecked(page);
3440
3441         if (journal)
3442                 jbd2_journal_invalidatepage(journal, page, offset);
3443         else
3444                 block_invalidatepage(page, offset);
3445 }
3446
3447 static int ext4_releasepage(struct page *page, gfp_t wait)
3448 {
3449         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3450
3451         WARN_ON(PageChecked(page));
3452         if (!page_has_buffers(page))
3453                 return 0;
3454         if (journal)
3455                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3456         else
3457                 return try_to_free_buffers(page);
3458 }
3459
3460 /*
3461  * O_DIRECT for ext3 (or indirect map) based files
3462  *
3463  * If the O_DIRECT write will extend the file then add this inode to the
3464  * orphan list.  So recovery will truncate it back to the original size
3465  * if the machine crashes during the write.
3466  *
3467  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3468  * crashes then stale disk data _may_ be exposed inside the file. But current
3469  * VFS code falls back into buffered path in that case so we are safe.
3470  */
3471 static ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
3472                               const struct iovec *iov, loff_t offset,
3473                               unsigned long nr_segs)
3474 {
3475         struct file *file = iocb->ki_filp;
3476         struct inode *inode = file->f_mapping->host;
3477         struct ext4_inode_info *ei = EXT4_I(inode);
3478         handle_t *handle;
3479         ssize_t ret;
3480         int orphan = 0;
3481         size_t count = iov_length(iov, nr_segs);
3482         int retries = 0;
3483
3484         if (rw == WRITE) {
3485                 loff_t final_size = offset + count;
3486
3487                 if (final_size > inode->i_size) {
3488                         /* Credits for sb + inode write */
3489                         handle = ext4_journal_start(inode, 2);
3490                         if (IS_ERR(handle)) {
3491                                 ret = PTR_ERR(handle);
3492                                 goto out;
3493                         }
3494                         ret = ext4_orphan_add(handle, inode);
3495                         if (ret) {
3496                                 ext4_journal_stop(handle);
3497                                 goto out;
3498                         }
3499                         orphan = 1;
3500                         ei->i_disksize = inode->i_size;
3501                         ext4_journal_stop(handle);
3502                 }
3503         }
3504
3505 retry:
3506         if (rw == READ && ext4_should_dioread_nolock(inode))
3507                 ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
3508                                  inode->i_sb->s_bdev, iov,
3509                                  offset, nr_segs,
3510                                  ext4_get_block, NULL);
3511         else
3512                 ret = blockdev_direct_IO(rw, iocb, inode,
3513                                  inode->i_sb->s_bdev, iov,
3514                                  offset, nr_segs,
3515                                  ext4_get_block, NULL);
3516         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3517                 goto retry;
3518
3519         if (orphan) {
3520                 int err;
3521
3522                 /* Credits for sb + inode write */
3523                 handle = ext4_journal_start(inode, 2);
3524                 if (IS_ERR(handle)) {
3525                         /* This is really bad luck. We've written the data
3526                          * but cannot extend i_size. Bail out and pretend
3527                          * the write failed... */
3528                         ret = PTR_ERR(handle);
3529                         if (inode->i_nlink)
3530                                 ext4_orphan_del(NULL, inode);
3531
3532                         goto out;
3533                 }
3534                 if (inode->i_nlink)
3535                         ext4_orphan_del(handle, inode);
3536                 if (ret > 0) {
3537                         loff_t end = offset + ret;
3538                         if (end > inode->i_size) {
3539                                 ei->i_disksize = end;
3540                                 i_size_write(inode, end);
3541                                 /*
3542                                  * We're going to return a positive `ret'
3543                                  * here due to non-zero-length I/O, so there's
3544                                  * no way of reporting error returns from
3545                                  * ext4_mark_inode_dirty() to userspace.  So
3546                                  * ignore it.
3547                                  */
3548                                 ext4_mark_inode_dirty(handle, inode);
3549                         }
3550                 }
3551                 err = ext4_journal_stop(handle);
3552                 if (ret == 0)
3553                         ret = err;
3554         }
3555 out:
3556         return ret;
3557 }
3558
3559 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
3560                    struct buffer_head *bh_result, int create)
3561 {
3562         handle_t *handle = ext4_journal_current_handle();
3563         int ret = 0;
3564         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
3565         int dio_credits;
3566         int started = 0;
3567
3568         ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
3569                    inode->i_ino, create);
3570         /*
3571          * ext4_get_block in prepare for a DIO write or buffer write.
3572          * We allocate an uinitialized extent if blocks haven't been allocated.
3573          * The extent will be converted to initialized after IO complete.
3574          */
3575         create = EXT4_GET_BLOCKS_IO_CREATE_EXT;
3576
3577         if (!handle) {
3578                 if (max_blocks > DIO_MAX_BLOCKS)
3579                         max_blocks = DIO_MAX_BLOCKS;
3580                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
3581                 handle = ext4_journal_start(inode, dio_credits);
3582                 if (IS_ERR(handle)) {
3583                         ret = PTR_ERR(handle);
3584                         goto out;
3585                 }
3586                 started = 1;
3587         }
3588
3589         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
3590                               create);
3591         if (ret > 0) {
3592                 bh_result->b_size = (ret << inode->i_blkbits);
3593                 ret = 0;
3594         }
3595         if (started)
3596                 ext4_journal_stop(handle);
3597 out:
3598         return ret;
3599 }
3600
3601 static void dump_completed_IO(struct inode * inode)
3602 {
3603 #ifdef  EXT4_DEBUG
3604         struct list_head *cur, *before, *after;
3605         ext4_io_end_t *io, *io0, *io1;
3606         unsigned long flags;
3607
3608         if (list_empty(&EXT4_I(inode)->i_completed_io_list)){
3609                 ext4_debug("inode %lu completed_io list is empty\n", inode->i_ino);
3610                 return;
3611         }
3612
3613         ext4_debug("Dump inode %lu completed_io list \n", inode->i_ino);
3614         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3615         list_for_each_entry(io, &EXT4_I(inode)->i_completed_io_list, list){
3616                 cur = &io->list;
3617                 before = cur->prev;
3618                 io0 = container_of(before, ext4_io_end_t, list);
3619                 after = cur->next;
3620                 io1 = container_of(after, ext4_io_end_t, list);
3621
3622                 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
3623                             io, inode->i_ino, io0, io1);
3624         }
3625         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3626 #endif
3627 }
3628
3629 /*
3630  * check a range of space and convert unwritten extents to written.
3631  */
3632 static int ext4_end_io_nolock(ext4_io_end_t *io)
3633 {
3634         struct inode *inode = io->inode;
3635         loff_t offset = io->offset;
3636         ssize_t size = io->size;
3637         int ret = 0;
3638
3639         ext4_debug("ext4_end_io_nolock: io 0x%p from inode %lu,list->next 0x%p,"
3640                    "list->prev 0x%p\n",
3641                    io, inode->i_ino, io->list.next, io->list.prev);
3642
3643         if (list_empty(&io->list))
3644                 return ret;
3645
3646         if (io->flag != EXT4_IO_UNWRITTEN)
3647                 return ret;
3648
3649         ret = ext4_convert_unwritten_extents(inode, offset, size);
3650         if (ret < 0) {
3651                 printk(KERN_EMERG "%s: failed to convert unwritten"
3652                         "extents to written extents, error is %d"
3653                         " io is still on inode %lu aio dio list\n",
3654                        __func__, ret, inode->i_ino);
3655                 return ret;
3656         }
3657
3658         /* clear the DIO AIO unwritten flag */
3659         io->flag = 0;
3660         return ret;
3661 }
3662
3663 /*
3664  * work on completed aio dio IO, to convert unwritten extents to extents
3665  */
3666 static void ext4_end_io_work(struct work_struct *work)
3667 {
3668         ext4_io_end_t           *io = container_of(work, ext4_io_end_t, work);
3669         struct inode            *inode = io->inode;
3670         struct ext4_inode_info  *ei = EXT4_I(inode);
3671         unsigned long           flags;
3672         int                     ret;
3673
3674         mutex_lock(&inode->i_mutex);
3675         ret = ext4_end_io_nolock(io);
3676         if (ret < 0) {
3677                 mutex_unlock(&inode->i_mutex);
3678                 return;
3679         }
3680
3681         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3682         if (!list_empty(&io->list))
3683                 list_del_init(&io->list);
3684         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3685         mutex_unlock(&inode->i_mutex);
3686         ext4_free_io_end(io);
3687 }
3688
3689 /*
3690  * This function is called from ext4_sync_file().
3691  *
3692  * When IO is completed, the work to convert unwritten extents to
3693  * written is queued on workqueue but may not get immediately
3694  * scheduled. When fsync is called, we need to ensure the
3695  * conversion is complete before fsync returns.
3696  * The inode keeps track of a list of pending/completed IO that
3697  * might needs to do the conversion. This function walks through
3698  * the list and convert the related unwritten extents for completed IO
3699  * to written.
3700  * The function return the number of pending IOs on success.
3701  */
3702 int flush_completed_IO(struct inode *inode)
3703 {
3704         ext4_io_end_t *io;
3705         struct ext4_inode_info *ei = EXT4_I(inode);
3706         unsigned long flags;
3707         int ret = 0;
3708         int ret2 = 0;
3709
3710         if (list_empty(&ei->i_completed_io_list))
3711                 return ret;
3712
3713         dump_completed_IO(inode);
3714         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3715         while (!list_empty(&ei->i_completed_io_list)){
3716                 io = list_entry(ei->i_completed_io_list.next,
3717                                 ext4_io_end_t, list);
3718                 /*
3719                  * Calling ext4_end_io_nolock() to convert completed
3720                  * IO to written.
3721                  *
3722                  * When ext4_sync_file() is called, run_queue() may already
3723                  * about to flush the work corresponding to this io structure.
3724                  * It will be upset if it founds the io structure related
3725                  * to the work-to-be schedule is freed.
3726                  *
3727                  * Thus we need to keep the io structure still valid here after
3728                  * convertion finished. The io structure has a flag to
3729                  * avoid double converting from both fsync and background work
3730                  * queue work.
3731                  */
3732                 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3733                 ret = ext4_end_io_nolock(io);
3734                 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3735                 if (ret < 0)
3736                         ret2 = ret;
3737                 else
3738                         list_del_init(&io->list);
3739         }
3740         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3741         return (ret2 < 0) ? ret2 : 0;
3742 }
3743
3744 static ext4_io_end_t *ext4_init_io_end (struct inode *inode, gfp_t flags)
3745 {
3746         ext4_io_end_t *io = NULL;
3747
3748         io = kmalloc(sizeof(*io), flags);
3749
3750         if (io) {
3751                 igrab(inode);
3752                 io->inode = inode;
3753                 io->flag = 0;
3754                 io->offset = 0;
3755                 io->size = 0;
3756                 io->page = NULL;
3757                 INIT_WORK(&io->work, ext4_end_io_work);
3758                 INIT_LIST_HEAD(&io->list);
3759         }
3760
3761         return io;
3762 }
3763
3764 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3765                             ssize_t size, void *private)
3766 {
3767         ext4_io_end_t *io_end = iocb->private;
3768         struct workqueue_struct *wq;
3769         unsigned long flags;
3770         struct ext4_inode_info *ei;
3771
3772         /* if not async direct IO or dio with 0 bytes write, just return */
3773         if (!io_end || !size)
3774                 return;
3775
3776         ext_debug("ext4_end_io_dio(): io_end 0x%p"
3777                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
3778                   iocb->private, io_end->inode->i_ino, iocb, offset,
3779                   size);
3780
3781         /* if not aio dio with unwritten extents, just free io and return */
3782         if (io_end->flag != EXT4_IO_UNWRITTEN){
3783                 ext4_free_io_end(io_end);
3784                 iocb->private = NULL;
3785                 return;
3786         }
3787
3788         io_end->offset = offset;
3789         io_end->size = size;
3790         io_end->flag = EXT4_IO_UNWRITTEN;
3791         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
3792
3793         /* queue the work to convert unwritten extents to written */
3794         queue_work(wq, &io_end->work);
3795
3796         /* Add the io_end to per-inode completed aio dio list*/
3797         ei = EXT4_I(io_end->inode);
3798         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3799         list_add_tail(&io_end->list, &ei->i_completed_io_list);
3800         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3801         iocb->private = NULL;
3802 }
3803
3804 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
3805 {
3806         ext4_io_end_t *io_end = bh->b_private;
3807         struct workqueue_struct *wq;
3808         struct inode *inode;
3809         unsigned long flags;
3810
3811         if (!test_clear_buffer_uninit(bh) || !io_end)
3812                 goto out;
3813
3814         if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
3815                 printk("sb umounted, discard end_io request for inode %lu\n",
3816                         io_end->inode->i_ino);
3817                 ext4_free_io_end(io_end);
3818                 goto out;
3819         }
3820
3821         io_end->flag = EXT4_IO_UNWRITTEN;
3822         inode = io_end->inode;
3823
3824         /* Add the io_end to per-inode completed io list*/
3825         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3826         list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
3827         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3828
3829         wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
3830         /* queue the work to convert unwritten extents to written */
3831         queue_work(wq, &io_end->work);
3832 out:
3833         bh->b_private = NULL;
3834         bh->b_end_io = NULL;
3835         clear_buffer_uninit(bh);
3836         end_buffer_async_write(bh, uptodate);
3837 }
3838
3839 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
3840 {
3841         ext4_io_end_t *io_end;
3842         struct page *page = bh->b_page;
3843         loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
3844         size_t size = bh->b_size;
3845
3846 retry:
3847         io_end = ext4_init_io_end(inode, GFP_ATOMIC);
3848         if (!io_end) {
3849                 if (printk_ratelimit())
3850                         printk(KERN_WARNING "%s: allocation fail\n", __func__);
3851                 schedule();
3852                 goto retry;
3853         }
3854         io_end->offset = offset;
3855         io_end->size = size;
3856         /*
3857          * We need to hold a reference to the page to make sure it
3858          * doesn't get evicted before ext4_end_io_work() has a chance
3859          * to convert the extent from written to unwritten.
3860          */
3861         io_end->page = page;
3862         get_page(io_end->page);
3863
3864         bh->b_private = io_end;
3865         bh->b_end_io = ext4_end_io_buffer_write;
3866         return 0;
3867 }
3868
3869 /*
3870  * For ext4 extent files, ext4 will do direct-io write to holes,
3871  * preallocated extents, and those write extend the file, no need to
3872  * fall back to buffered IO.
3873  *
3874  * For holes, we fallocate those blocks, mark them as unintialized
3875  * If those blocks were preallocated, we mark sure they are splited, but
3876  * still keep the range to write as unintialized.
3877  *
3878  * The unwrritten extents will be converted to written when DIO is completed.
3879  * For async direct IO, since the IO may still pending when return, we
3880  * set up an end_io call back function, which will do the convertion
3881  * when async direct IO completed.
3882  *
3883  * If the O_DIRECT write will extend the file then add this inode to the
3884  * orphan list.  So recovery will truncate it back to the original size
3885  * if the machine crashes during the write.
3886  *
3887  */
3888 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
3889                               const struct iovec *iov, loff_t offset,
3890                               unsigned long nr_segs)
3891 {
3892         struct file *file = iocb->ki_filp;
3893         struct inode *inode = file->f_mapping->host;
3894         ssize_t ret;
3895         size_t count = iov_length(iov, nr_segs);
3896
3897         loff_t final_size = offset + count;
3898         if (rw == WRITE && final_size <= inode->i_size) {
3899                 /*
3900                  * We could direct write to holes and fallocate.
3901                  *
3902                  * Allocated blocks to fill the hole are marked as uninitialized
3903                  * to prevent paralel buffered read to expose the stale data
3904                  * before DIO complete the data IO.
3905                  *
3906                  * As to previously fallocated extents, ext4 get_block
3907                  * will just simply mark the buffer mapped but still
3908                  * keep the extents uninitialized.
3909                  *
3910                  * for non AIO case, we will convert those unwritten extents
3911                  * to written after return back from blockdev_direct_IO.
3912                  *
3913                  * for async DIO, the conversion needs to be defered when
3914                  * the IO is completed. The ext4 end_io callback function
3915                  * will be called to take care of the conversion work.
3916                  * Here for async case, we allocate an io_end structure to
3917                  * hook to the iocb.
3918                  */
3919                 iocb->private = NULL;
3920                 EXT4_I(inode)->cur_aio_dio = NULL;
3921                 if (!is_sync_kiocb(iocb)) {
3922                         iocb->private = ext4_init_io_end(inode, GFP_NOFS);
3923                         if (!iocb->private)
3924                                 return -ENOMEM;
3925                         /*
3926                          * we save the io structure for current async
3927                          * direct IO, so that later ext4_get_blocks()
3928                          * could flag the io structure whether there
3929                          * is a unwritten extents needs to be converted
3930                          * when IO is completed.
3931                          */
3932                         EXT4_I(inode)->cur_aio_dio = iocb->private;
3933                 }
3934
3935                 ret = blockdev_direct_IO(rw, iocb, inode,
3936                                          inode->i_sb->s_bdev, iov,
3937                                          offset, nr_segs,
3938                                          ext4_get_block_write,
3939                                          ext4_end_io_dio);
3940                 if (iocb->private)
3941                         EXT4_I(inode)->cur_aio_dio = NULL;
3942                 /*
3943                  * The io_end structure takes a reference to the inode,
3944                  * that structure needs to be destroyed and the
3945                  * reference to the inode need to be dropped, when IO is
3946                  * complete, even with 0 byte write, or failed.
3947                  *
3948                  * In the successful AIO DIO case, the io_end structure will be
3949                  * desctroyed and the reference to the inode will be dropped
3950                  * after the end_io call back function is called.
3951                  *
3952                  * In the case there is 0 byte write, or error case, since
3953                  * VFS direct IO won't invoke the end_io call back function,
3954                  * we need to free the end_io structure here.
3955                  */
3956                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
3957                         ext4_free_io_end(iocb->private);
3958                         iocb->private = NULL;
3959                 } else if (ret > 0 && ext4_test_inode_state(inode,
3960                                                 EXT4_STATE_DIO_UNWRITTEN)) {
3961                         int err;
3962                         /*
3963                          * for non AIO case, since the IO is already
3964                          * completed, we could do the convertion right here
3965                          */
3966                         err = ext4_convert_unwritten_extents(inode,
3967                                                              offset, ret);
3968                         if (err < 0)
3969                                 ret = err;
3970                         ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3971                 }
3972                 return ret;
3973         }
3974
3975         /* for write the the end of file case, we fall back to old way */
3976         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3977 }
3978
3979 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3980                               const struct iovec *iov, loff_t offset,
3981                               unsigned long nr_segs)
3982 {
3983         struct file *file = iocb->ki_filp;
3984         struct inode *inode = file->f_mapping->host;
3985
3986         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
3987                 return ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
3988
3989         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3990 }
3991
3992 /*
3993  * Pages can be marked dirty completely asynchronously from ext4's journalling
3994  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3995  * much here because ->set_page_dirty is called under VFS locks.  The page is
3996  * not necessarily locked.
3997  *
3998  * We cannot just dirty the page and leave attached buffers clean, because the
3999  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
4000  * or jbddirty because all the journalling code will explode.
4001  *
4002  * So what we do is to mark the page "pending dirty" and next time writepage
4003  * is called, propagate that into the buffers appropriately.
4004  */
4005 static int ext4_journalled_set_page_dirty(struct page *page)
4006 {
4007         SetPageChecked(page);
4008         return __set_page_dirty_nobuffers(page);
4009 }
4010
4011 static const struct address_space_operations ext4_ordered_aops = {
4012         .readpage               = ext4_readpage,
4013         .readpages              = ext4_readpages,
4014         .writepage              = ext4_writepage,
4015         .sync_page              = block_sync_page,
4016         .write_begin            = ext4_write_begin,
4017         .write_end              = ext4_ordered_write_end,
4018         .bmap                   = ext4_bmap,
4019         .invalidatepage         = ext4_invalidatepage,
4020         .releasepage            = ext4_releasepage,
4021         .direct_IO              = ext4_direct_IO,
4022         .migratepage            = buffer_migrate_page,
4023         .is_partially_uptodate  = block_is_partially_uptodate,
4024         .error_remove_page      = generic_error_remove_page,
4025 };
4026
4027 static const struct address_space_operations ext4_writeback_aops = {
4028         .readpage               = ext4_readpage,
4029         .readpages              = ext4_readpages,
4030         .writepage              = ext4_writepage,
4031         .sync_page              = block_sync_page,
4032         .write_begin            = ext4_write_begin,
4033         .write_end              = ext4_writeback_write_end,
4034         .bmap                   = ext4_bmap,
4035         .invalidatepage         = ext4_invalidatepage,
4036         .releasepage            = ext4_releasepage,
4037         .direct_IO              = ext4_direct_IO,
4038         .migratepage            = buffer_migrate_page,
4039         .is_partially_uptodate  = block_is_partially_uptodate,
4040         .error_remove_page      = generic_error_remove_page,
4041 };
4042
4043 static const struct address_space_operations ext4_journalled_aops = {
4044         .readpage               = ext4_readpage,
4045         .readpages              = ext4_readpages,
4046         .writepage              = ext4_writepage,
4047         .sync_page              = block_sync_page,
4048         .write_begin            = ext4_write_begin,
4049         .write_end              = ext4_journalled_write_end,
4050         .set_page_dirty         = ext4_journalled_set_page_dirty,
4051         .bmap                   = ext4_bmap,
4052         .invalidatepage         = ext4_invalidatepage,
4053         .releasepage            = ext4_releasepage,
4054         .is_partially_uptodate  = block_is_partially_uptodate,
4055         .error_remove_page      = generic_error_remove_page,
4056 };
4057
4058 static const struct address_space_operations ext4_da_aops = {
4059         .readpage               = ext4_readpage,
4060         .readpages              = ext4_readpages,
4061         .writepage              = ext4_writepage,
4062         .writepages             = ext4_da_writepages,
4063         .sync_page              = block_sync_page,
4064         .write_begin            = ext4_da_write_begin,
4065         .write_end              = ext4_da_write_end,
4066         .bmap                   = ext4_bmap,
4067         .invalidatepage         = ext4_da_invalidatepage,
4068         .releasepage            = ext4_releasepage,
4069         .direct_IO              = ext4_direct_IO,
4070         .migratepage            = buffer_migrate_page,
4071         .is_partially_uptodate  = block_is_partially_uptodate,
4072         .error_remove_page      = generic_error_remove_page,
4073 };
4074
4075 void ext4_set_aops(struct inode *inode)
4076 {
4077         if (ext4_should_order_data(inode) &&
4078                 test_opt(inode->i_sb, DELALLOC))
4079                 inode->i_mapping->a_ops = &ext4_da_aops;
4080         else if (ext4_should_order_data(inode))
4081                 inode->i_mapping->a_ops = &ext4_ordered_aops;
4082         else if (ext4_should_writeback_data(inode) &&
4083                  test_opt(inode->i_sb, DELALLOC))
4084                 inode->i_mapping->a_ops = &ext4_da_aops;
4085         else if (ext4_should_writeback_data(inode))
4086                 inode->i_mapping->a_ops = &ext4_writeback_aops;
4087         else
4088                 inode->i_mapping->a_ops = &ext4_journalled_aops;
4089 }
4090
4091 /*
4092  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
4093  * up to the end of the block which corresponds to `from'.
4094  * This required during truncate. We need to physically zero the tail end
4095  * of that block so it doesn't yield old data if the file is later grown.
4096  */
4097 int ext4_block_truncate_page(handle_t *handle,
4098                 struct address_space *mapping, loff_t from)
4099 {
4100         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
4101         unsigned offset = from & (PAGE_CACHE_SIZE-1);
4102         unsigned blocksize, length, pos;
4103         ext4_lblk_t iblock;
4104         struct inode *inode = mapping->host;
4105         struct buffer_head *bh;
4106         struct page *page;
4107         int err = 0;
4108
4109         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
4110                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
4111         if (!page)
4112                 return -EINVAL;
4113
4114         blocksize = inode->i_sb->s_blocksize;
4115         length = blocksize - (offset & (blocksize - 1));
4116         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
4117
4118         /*
4119          * For "nobh" option,  we can only work if we don't need to
4120          * read-in the page - otherwise we create buffers to do the IO.
4121          */
4122         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
4123              ext4_should_writeback_data(inode) && PageUptodate(page)) {
4124                 zero_user(page, offset, length);
4125                 set_page_dirty(page);
4126                 goto unlock;
4127         }
4128
4129         if (!page_has_buffers(page))
4130                 create_empty_buffers(page, blocksize, 0);
4131
4132         /* Find the buffer that contains "offset" */
4133         bh = page_buffers(page);
4134         pos = blocksize;
4135         while (offset >= pos) {
4136                 bh = bh->b_this_page;
4137                 iblock++;
4138                 pos += blocksize;
4139         }
4140
4141         err = 0;
4142         if (buffer_freed(bh)) {
4143                 BUFFER_TRACE(bh, "freed: skip");
4144                 goto unlock;
4145         }
4146
4147         if (!buffer_mapped(bh)) {
4148                 BUFFER_TRACE(bh, "unmapped");
4149                 ext4_get_block(inode, iblock, bh, 0);
4150                 /* unmapped? It's a hole - nothing to do */
4151                 if (!buffer_mapped(bh)) {
4152                         BUFFER_TRACE(bh, "still unmapped");
4153                         goto unlock;
4154                 }
4155         }
4156
4157         /* Ok, it's mapped. Make sure it's up-to-date */
4158         if (PageUptodate(page))
4159                 set_buffer_uptodate(bh);
4160
4161         if (!buffer_uptodate(bh)) {
4162                 err = -EIO;
4163                 ll_rw_block(READ, 1, &bh);
4164                 wait_on_buffer(bh);
4165                 /* Uhhuh. Read error. Complain and punt. */
4166                 if (!buffer_uptodate(bh))
4167                         goto unlock;
4168         }
4169
4170         if (ext4_should_journal_data(inode)) {
4171                 BUFFER_TRACE(bh, "get write access");
4172                 err = ext4_journal_get_write_access(handle, bh);
4173                 if (err)
4174                         goto unlock;
4175         }
4176
4177         zero_user(page, offset, length);
4178
4179         BUFFER_TRACE(bh, "zeroed end of block");
4180
4181         err = 0;
4182         if (ext4_should_journal_data(inode)) {
4183                 err = ext4_handle_dirty_metadata(handle, inode, bh);
4184         } else {
4185                 if (ext4_should_order_data(inode))
4186                         err = ext4_jbd2_file_inode(handle, inode);
4187                 mark_buffer_dirty(bh);
4188         }
4189
4190 unlock:
4191         unlock_page(page);
4192         page_cache_release(page);
4193         return err;
4194 }
4195
4196 /*
4197  * Probably it should be a library function... search for first non-zero word
4198  * or memcmp with zero_page, whatever is better for particular architecture.
4199  * Linus?
4200  */
4201 static inline int all_zeroes(__le32 *p, __le32 *q)
4202 {
4203         while (p < q)
4204                 if (*p++)
4205                         return 0;
4206         return 1;
4207 }
4208
4209 /**
4210  *      ext4_find_shared - find the indirect blocks for partial truncation.
4211  *      @inode:   inode in question
4212  *      @depth:   depth of the affected branch
4213  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
4214  *      @chain:   place to store the pointers to partial indirect blocks
4215  *      @top:     place to the (detached) top of branch
4216  *
4217  *      This is a helper function used by ext4_truncate().
4218  *
4219  *      When we do truncate() we may have to clean the ends of several
4220  *      indirect blocks but leave the blocks themselves alive. Block is
4221  *      partially truncated if some data below the new i_size is refered
4222  *      from it (and it is on the path to the first completely truncated
4223  *      data block, indeed).  We have to free the top of that path along
4224  *      with everything to the right of the path. Since no allocation
4225  *      past the truncation point is possible until ext4_truncate()
4226  *      finishes, we may safely do the latter, but top of branch may
4227  *      require special attention - pageout below the truncation point
4228  *      might try to populate it.
4229  *
4230  *      We atomically detach the top of branch from the tree, store the
4231  *      block number of its root in *@top, pointers to buffer_heads of
4232  *      partially truncated blocks - in @chain[].bh and pointers to
4233  *      their last elements that should not be removed - in
4234  *      @chain[].p. Return value is the pointer to last filled element
4235  *      of @chain.
4236  *
4237  *      The work left to caller to do the actual freeing of subtrees:
4238  *              a) free the subtree starting from *@top
4239  *              b) free the subtrees whose roots are stored in
4240  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
4241  *              c) free the subtrees growing from the inode past the @chain[0].
4242  *                      (no partially truncated stuff there).  */
4243
4244 static Indirect *ext4_find_shared(struct inode *inode, int depth,
4245                                   ext4_lblk_t offsets[4], Indirect chain[4],
4246                                   __le32 *top)
4247 {
4248         Indirect *partial, *p;
4249         int k, err;
4250
4251         *top = 0;
4252         /* Make k index the deepest non-null offset + 1 */
4253         for (k = depth; k > 1 && !offsets[k-1]; k--)
4254                 ;
4255         partial = ext4_get_branch(inode, k, offsets, chain, &err);
4256         /* Writer: pointers */
4257         if (!partial)
4258                 partial = chain + k-1;
4259         /*
4260          * If the branch acquired continuation since we've looked at it -
4261          * fine, it should all survive and (new) top doesn't belong to us.
4262          */
4263         if (!partial->key && *partial->p)
4264                 /* Writer: end */
4265                 goto no_top;
4266         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
4267                 ;
4268         /*
4269          * OK, we've found the last block that must survive. The rest of our
4270          * branch should be detached before unlocking. However, if that rest
4271          * of branch is all ours and does not grow immediately from the inode
4272          * it's easier to cheat and just decrement partial->p.
4273          */
4274         if (p == chain + k - 1 && p > chain) {
4275                 p->p--;
4276         } else {
4277                 *top = *p->p;
4278                 /* Nope, don't do this in ext4.  Must leave the tree intact */
4279 #if 0
4280                 *p->p = 0;
4281 #endif
4282         }
4283         /* Writer: end */
4284
4285         while (partial > p) {
4286                 brelse(partial->bh);
4287                 partial--;
4288         }
4289 no_top:
4290         return partial;
4291 }
4292
4293 /*
4294  * Zero a number of block pointers in either an inode or an indirect block.
4295  * If we restart the transaction we must again get write access to the
4296  * indirect block for further modification.
4297  *
4298  * We release `count' blocks on disk, but (last - first) may be greater
4299  * than `count' because there can be holes in there.
4300  */
4301 static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
4302                              struct buffer_head *bh,
4303                              ext4_fsblk_t block_to_free,
4304                              unsigned long count, __le32 *first,
4305                              __le32 *last)
4306 {
4307         __le32 *p;
4308         int     flags = EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_VALIDATED;
4309
4310         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
4311                 flags |= EXT4_FREE_BLOCKS_METADATA;
4312
4313         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
4314                                    count)) {
4315                 ext4_error(inode->i_sb, "inode #%lu: "
4316                            "attempt to clear blocks %llu len %lu, invalid",
4317                            inode->i_ino, (unsigned long long) block_to_free,
4318                            count);
4319                 return 1;
4320         }
4321
4322         if (try_to_extend_transaction(handle, inode)) {
4323                 if (bh) {
4324                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4325                         ext4_handle_dirty_metadata(handle, inode, bh);
4326                 }
4327                 ext4_mark_inode_dirty(handle, inode);
4328                 ext4_truncate_restart_trans(handle, inode,
4329                                             blocks_for_truncate(inode));
4330                 if (bh) {
4331                         BUFFER_TRACE(bh, "retaking write access");
4332                         ext4_journal_get_write_access(handle, bh);
4333                 }
4334         }
4335
4336         for (p = first; p < last; p++)
4337                 *p = 0;
4338
4339         ext4_free_blocks(handle, inode, 0, block_to_free, count, flags);
4340         return 0;
4341 }
4342
4343 /**
4344  * ext4_free_data - free a list of data blocks
4345  * @handle:     handle for this transaction
4346  * @inode:      inode we are dealing with
4347  * @this_bh:    indirect buffer_head which contains *@first and *@last
4348  * @first:      array of block numbers
4349  * @last:       points immediately past the end of array
4350  *
4351  * We are freeing all blocks refered from that array (numbers are stored as
4352  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
4353  *
4354  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
4355  * blocks are contiguous then releasing them at one time will only affect one
4356  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
4357  * actually use a lot of journal space.
4358  *
4359  * @this_bh will be %NULL if @first and @last point into the inode's direct
4360  * block pointers.
4361  */
4362 static void ext4_free_data(handle_t *handle, struct inode *inode,
4363                            struct buffer_head *this_bh,
4364                            __le32 *first, __le32 *last)
4365 {
4366         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
4367         unsigned long count = 0;            /* Number of blocks in the run */
4368         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
4369                                                corresponding to
4370                                                block_to_free */
4371         ext4_fsblk_t nr;                    /* Current block # */
4372         __le32 *p;                          /* Pointer into inode/ind
4373                                                for current block */
4374         int err;
4375
4376         if (this_bh) {                          /* For indirect block */
4377                 BUFFER_TRACE(this_bh, "get_write_access");
4378                 err = ext4_journal_get_write_access(handle, this_bh);
4379                 /* Important: if we can't update the indirect pointers
4380                  * to the blocks, we can't free them. */
4381                 if (err)
4382                         return;
4383         }
4384
4385         for (p = first; p < last; p++) {
4386                 nr = le32_to_cpu(*p);
4387                 if (nr) {
4388                         /* accumulate blocks to free if they're contiguous */
4389                         if (count == 0) {
4390                                 block_to_free = nr;
4391                                 block_to_free_p = p;
4392                                 count = 1;
4393                         } else if (nr == block_to_free + count) {
4394                                 count++;
4395                         } else {
4396                                 if (ext4_clear_blocks(handle, inode, this_bh,
4397                                                       block_to_free, count,
4398                                                       block_to_free_p, p))
4399                                         break;
4400                                 block_to_free = nr;
4401                                 block_to_free_p = p;
4402                                 count = 1;
4403                         }
4404                 }
4405         }
4406
4407         if (count > 0)
4408                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
4409                                   count, block_to_free_p, p);
4410
4411         if (this_bh) {
4412                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
4413
4414                 /*
4415                  * The buffer head should have an attached journal head at this
4416                  * point. However, if the data is corrupted and an indirect
4417                  * block pointed to itself, it would have been detached when
4418                  * the block was cleared. Check for this instead of OOPSing.
4419                  */
4420                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
4421                         ext4_handle_dirty_metadata(handle, inode, this_bh);
4422                 else
4423                         ext4_error(inode->i_sb,
4424                                    "circular indirect block detected, "
4425                                    "inode=%lu, block=%llu",
4426                                    inode->i_ino,
4427                                    (unsigned long long) this_bh->b_blocknr);
4428         }
4429 }
4430
4431 /**
4432  *      ext4_free_branches - free an array of branches
4433  *      @handle: JBD handle for this transaction
4434  *      @inode: inode we are dealing with
4435  *      @parent_bh: the buffer_head which contains *@first and *@last
4436  *      @first: array of block numbers
4437  *      @last:  pointer immediately past the end of array
4438  *      @depth: depth of the branches to free
4439  *
4440  *      We are freeing all blocks refered from these branches (numbers are
4441  *      stored as little-endian 32-bit) and updating @inode->i_blocks
4442  *      appropriately.
4443  */
4444 static void ext4_free_branches(handle_t *handle, struct inode *inode,
4445                                struct buffer_head *parent_bh,
4446                                __le32 *first, __le32 *last, int depth)
4447 {
4448         ext4_fsblk_t nr;
4449         __le32 *p;
4450
4451         if (ext4_handle_is_aborted(handle))
4452                 return;
4453
4454         if (depth--) {
4455                 struct buffer_head *bh;
4456                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4457                 p = last;
4458                 while (--p >= first) {
4459                         nr = le32_to_cpu(*p);
4460                         if (!nr)
4461                                 continue;               /* A hole */
4462
4463                         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
4464                                                    nr, 1)) {
4465                                 ext4_error(inode->i_sb,
4466                                            "indirect mapped block in inode "
4467                                            "#%lu invalid (level %d, blk #%lu)",
4468                                            inode->i_ino, depth,
4469                                            (unsigned long) nr);
4470                                 break;
4471                         }
4472
4473                         /* Go read the buffer for the next level down */
4474                         bh = sb_bread(inode->i_sb, nr);
4475
4476                         /*
4477                          * A read failure? Report error and clear slot
4478                          * (should be rare).
4479                          */
4480                         if (!bh) {
4481                                 ext4_error(inode->i_sb,
4482                                            "Read failure, inode=%lu, block=%llu",
4483                                            inode->i_ino, nr);
4484                                 continue;
4485                         }
4486
4487                         /* This zaps the entire block.  Bottom up. */
4488                         BUFFER_TRACE(bh, "free child branches");
4489                         ext4_free_branches(handle, inode, bh,
4490                                         (__le32 *) bh->b_data,
4491                                         (__le32 *) bh->b_data + addr_per_block,
4492                                         depth);
4493
4494                         /*
4495                          * We've probably journalled the indirect block several
4496                          * times during the truncate.  But it's no longer
4497                          * needed and we now drop it from the transaction via
4498                          * jbd2_journal_revoke().
4499                          *
4500                          * That's easy if it's exclusively part of this
4501                          * transaction.  But if it's part of the committing
4502                          * transaction then jbd2_journal_forget() will simply
4503                          * brelse() it.  That means that if the underlying
4504                          * block is reallocated in ext4_get_block(),
4505                          * unmap_underlying_metadata() will find this block
4506                          * and will try to get rid of it.  damn, damn.
4507                          *
4508                          * If this block has already been committed to the
4509                          * journal, a revoke record will be written.  And
4510                          * revoke records must be emitted *before* clearing
4511                          * this block's bit in the bitmaps.
4512                          */
4513                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
4514
4515                         /*
4516                          * Everything below this this pointer has been
4517                          * released.  Now let this top-of-subtree go.
4518                          *
4519                          * We want the freeing of this indirect block to be
4520                          * atomic in the journal with the updating of the
4521                          * bitmap block which owns it.  So make some room in
4522                          * the journal.
4523                          *
4524                          * We zero the parent pointer *after* freeing its
4525                          * pointee in the bitmaps, so if extend_transaction()
4526                          * for some reason fails to put the bitmap changes and
4527                          * the release into the same transaction, recovery
4528                          * will merely complain about releasing a free block,
4529                          * rather than leaking blocks.
4530                          */
4531                         if (ext4_handle_is_aborted(handle))
4532                                 return;
4533                         if (try_to_extend_transaction(handle, inode)) {
4534                                 ext4_mark_inode_dirty(handle, inode);
4535                                 ext4_truncate_restart_trans(handle, inode,
4536                                             blocks_for_truncate(inode));
4537                         }
4538
4539                         ext4_free_blocks(handle, inode, 0, nr, 1,
4540                                          EXT4_FREE_BLOCKS_METADATA);
4541
4542                         if (parent_bh) {
4543                                 /*
4544                                  * The block which we have just freed is
4545                                  * pointed to by an indirect block: journal it
4546                                  */
4547                                 BUFFER_TRACE(parent_bh, "get_write_access");
4548                                 if (!ext4_journal_get_write_access(handle,
4549                                                                    parent_bh)){
4550                                         *p = 0;
4551                                         BUFFER_TRACE(parent_bh,
4552                                         "call ext4_handle_dirty_metadata");
4553                                         ext4_handle_dirty_metadata(handle,
4554                                                                    inode,
4555                                                                    parent_bh);
4556                                 }
4557                         }
4558                 }
4559         } else {
4560                 /* We have reached the bottom of the tree. */
4561                 BUFFER_TRACE(parent_bh, "free data blocks");
4562                 ext4_free_data(handle, inode, parent_bh, first, last);
4563         }
4564 }
4565
4566 int ext4_can_truncate(struct inode *inode)
4567 {
4568         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4569                 return 0;
4570         if (S_ISREG(inode->i_mode))
4571                 return 1;
4572         if (S_ISDIR(inode->i_mode))
4573                 return 1;
4574         if (S_ISLNK(inode->i_mode))
4575                 return !ext4_inode_is_fast_symlink(inode);
4576         return 0;
4577 }
4578
4579 /*
4580  * ext4_truncate()
4581  *
4582  * We block out ext4_get_block() block instantiations across the entire
4583  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4584  * simultaneously on behalf of the same inode.
4585  *
4586  * As we work through the truncate and commmit bits of it to the journal there
4587  * is one core, guiding principle: the file's tree must always be consistent on
4588  * disk.  We must be able to restart the truncate after a crash.
4589  *
4590  * The file's tree may be transiently inconsistent in memory (although it
4591  * probably isn't), but whenever we close off and commit a journal transaction,
4592  * the contents of (the filesystem + the journal) must be consistent and
4593  * restartable.  It's pretty simple, really: bottom up, right to left (although
4594  * left-to-right works OK too).
4595  *
4596  * Note that at recovery time, journal replay occurs *before* the restart of
4597  * truncate against the orphan inode list.
4598  *
4599  * The committed inode has the new, desired i_size (which is the same as
4600  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4601  * that this inode's truncate did not complete and it will again call
4602  * ext4_truncate() to have another go.  So there will be instantiated blocks
4603  * to the right of the truncation point in a crashed ext4 filesystem.  But
4604  * that's fine - as long as they are linked from the inode, the post-crash
4605  * ext4_truncate() run will find them and release them.
4606  */
4607 void ext4_truncate(struct inode *inode)
4608 {
4609         handle_t *handle;
4610         struct ext4_inode_info *ei = EXT4_I(inode);
4611         __le32 *i_data = ei->i_data;
4612         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4613         struct address_space *mapping = inode->i_mapping;
4614         ext4_lblk_t offsets[4];
4615         Indirect chain[4];
4616         Indirect *partial;
4617         __le32 nr = 0;
4618         int n;
4619         ext4_lblk_t last_block;
4620         unsigned blocksize = inode->i_sb->s_blocksize;
4621
4622         if (!ext4_can_truncate(inode))
4623                 return;
4624
4625         EXT4_I(inode)->i_flags &= ~EXT4_EOFBLOCKS_FL;
4626
4627         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4628                 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4629
4630         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
4631                 ext4_ext_truncate(inode);
4632                 return;
4633         }
4634
4635         handle = start_transaction(inode);
4636         if (IS_ERR(handle))
4637                 return;         /* AKPM: return what? */
4638
4639         last_block = (inode->i_size + blocksize-1)
4640                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4641
4642         if (inode->i_size & (blocksize - 1))
4643                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4644                         goto out_stop;
4645
4646         n = ext4_block_to_path(inode, last_block, offsets, NULL);
4647         if (n == 0)
4648                 goto out_stop;  /* error */
4649
4650         /*
4651          * OK.  This truncate is going to happen.  We add the inode to the
4652          * orphan list, so that if this truncate spans multiple transactions,
4653          * and we crash, we will resume the truncate when the filesystem
4654          * recovers.  It also marks the inode dirty, to catch the new size.
4655          *
4656          * Implication: the file must always be in a sane, consistent
4657          * truncatable state while each transaction commits.
4658          */
4659         if (ext4_orphan_add(handle, inode))
4660                 goto out_stop;
4661
4662         /*
4663          * From here we block out all ext4_get_block() callers who want to
4664          * modify the block allocation tree.
4665          */
4666         down_write(&ei->i_data_sem);
4667
4668         ext4_discard_preallocations(inode);
4669
4670         /*
4671          * The orphan list entry will now protect us from any crash which
4672          * occurs before the truncate completes, so it is now safe to propagate
4673          * the new, shorter inode size (held for now in i_size) into the
4674          * on-disk inode. We do this via i_disksize, which is the value which
4675          * ext4 *really* writes onto the disk inode.
4676          */
4677         ei->i_disksize = inode->i_size;
4678
4679         if (n == 1) {           /* direct blocks */
4680                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4681                                i_data + EXT4_NDIR_BLOCKS);
4682                 goto do_indirects;
4683         }
4684
4685         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4686         /* Kill the top of shared branch (not detached) */
4687         if (nr) {
4688                 if (partial == chain) {
4689                         /* Shared branch grows from the inode */
4690                         ext4_free_branches(handle, inode, NULL,
4691                                            &nr, &nr+1, (chain+n-1) - partial);
4692                         *partial->p = 0;
4693                         /*
4694                          * We mark the inode dirty prior to restart,
4695                          * and prior to stop.  No need for it here.
4696                          */
4697                 } else {
4698                         /* Shared branch grows from an indirect block */
4699                         BUFFER_TRACE(partial->bh, "get_write_access");
4700                         ext4_free_branches(handle, inode, partial->bh,
4701                                         partial->p,
4702                                         partial->p+1, (chain+n-1) - partial);
4703                 }
4704         }
4705         /* Clear the ends of indirect blocks on the shared branch */
4706         while (partial > chain) {
4707                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4708                                    (__le32*)partial->bh->b_data+addr_per_block,
4709                                    (chain+n-1) - partial);
4710                 BUFFER_TRACE(partial->bh, "call brelse");
4711                 brelse(partial->bh);
4712                 partial--;
4713         }
4714 do_indirects:
4715         /* Kill the remaining (whole) subtrees */
4716         switch (offsets[0]) {
4717         default:
4718                 nr = i_data[EXT4_IND_BLOCK];
4719                 if (nr) {
4720                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4721                         i_data[EXT4_IND_BLOCK] = 0;
4722                 }
4723         case EXT4_IND_BLOCK:
4724                 nr = i_data[EXT4_DIND_BLOCK];
4725                 if (nr) {
4726                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4727                         i_data[EXT4_DIND_BLOCK] = 0;
4728                 }
4729         case EXT4_DIND_BLOCK:
4730                 nr = i_data[EXT4_TIND_BLOCK];
4731                 if (nr) {
4732                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4733                         i_data[EXT4_TIND_BLOCK] = 0;
4734                 }
4735         case EXT4_TIND_BLOCK:
4736                 ;
4737         }
4738
4739         up_write(&ei->i_data_sem);
4740         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4741         ext4_mark_inode_dirty(handle, inode);
4742
4743         /*
4744          * In a multi-transaction truncate, we only make the final transaction
4745          * synchronous
4746          */
4747         if (IS_SYNC(inode))
4748                 ext4_handle_sync(handle);
4749 out_stop:
4750         /*
4751          * If this was a simple ftruncate(), and the file will remain alive
4752          * then we need to clear up the orphan record which we created above.
4753          * However, if this was a real unlink then we were called by
4754          * ext4_delete_inode(), and we allow that function to clean up the
4755          * orphan info for us.
4756          */
4757         if (inode->i_nlink)
4758                 ext4_orphan_del(handle, inode);
4759
4760         ext4_journal_stop(handle);
4761 }
4762
4763 /*
4764  * ext4_get_inode_loc returns with an extra refcount against the inode's
4765  * underlying buffer_head on success. If 'in_mem' is true, we have all
4766  * data in memory that is needed to recreate the on-disk version of this
4767  * inode.
4768  */
4769 static int __ext4_get_inode_loc(struct inode *inode,
4770                                 struct ext4_iloc *iloc, int in_mem)
4771 {
4772         struct ext4_group_desc  *gdp;
4773         struct buffer_head      *bh;
4774         struct super_block      *sb = inode->i_sb;
4775         ext4_fsblk_t            block;
4776         int                     inodes_per_block, inode_offset;
4777
4778         iloc->bh = NULL;
4779         if (!ext4_valid_inum(sb, inode->i_ino))
4780                 return -EIO;
4781
4782         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4783         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4784         if (!gdp)
4785                 return -EIO;
4786
4787         /*
4788          * Figure out the offset within the block group inode table
4789          */
4790         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4791         inode_offset = ((inode->i_ino - 1) %
4792                         EXT4_INODES_PER_GROUP(sb));
4793         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4794         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4795
4796         bh = sb_getblk(sb, block);
4797         if (!bh) {
4798                 ext4_error(sb, "unable to read inode block - "
4799                            "inode=%lu, block=%llu", inode->i_ino, block);
4800                 return -EIO;
4801         }
4802         if (!buffer_uptodate(bh)) {
4803                 lock_buffer(bh);
4804
4805                 /*
4806                  * If the buffer has the write error flag, we have failed
4807                  * to write out another inode in the same block.  In this
4808                  * case, we don't have to read the block because we may
4809                  * read the old inode data successfully.
4810                  */
4811                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4812                         set_buffer_uptodate(bh);
4813
4814                 if (buffer_uptodate(bh)) {
4815                         /* someone brought it uptodate while we waited */
4816                         unlock_buffer(bh);
4817                         goto has_buffer;
4818                 }
4819
4820                 /*
4821                  * If we have all information of the inode in memory and this
4822                  * is the only valid inode in the block, we need not read the
4823                  * block.
4824                  */
4825                 if (in_mem) {
4826                         struct buffer_head *bitmap_bh;
4827                         int i, start;
4828
4829                         start = inode_offset & ~(inodes_per_block - 1);
4830
4831                         /* Is the inode bitmap in cache? */
4832                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4833                         if (!bitmap_bh)
4834                                 goto make_io;
4835
4836                         /*
4837                          * If the inode bitmap isn't in cache then the
4838                          * optimisation may end up performing two reads instead
4839                          * of one, so skip it.
4840                          */
4841                         if (!buffer_uptodate(bitmap_bh)) {
4842                                 brelse(bitmap_bh);
4843                                 goto make_io;
4844                         }
4845                         for (i = start; i < start + inodes_per_block; i++) {
4846                                 if (i == inode_offset)
4847                                         continue;
4848                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4849                                         break;
4850                         }
4851                         brelse(bitmap_bh);
4852                         if (i == start + inodes_per_block) {
4853                                 /* all other inodes are free, so skip I/O */
4854                                 memset(bh->b_data, 0, bh->b_size);
4855                                 set_buffer_uptodate(bh);
4856                                 unlock_buffer(bh);
4857                                 goto has_buffer;
4858                         }
4859                 }
4860
4861 make_io:
4862                 /*
4863                  * If we need to do any I/O, try to pre-readahead extra
4864                  * blocks from the inode table.
4865                  */
4866                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4867                         ext4_fsblk_t b, end, table;
4868                         unsigned num;
4869
4870                         table = ext4_inode_table(sb, gdp);
4871                         /* s_inode_readahead_blks is always a power of 2 */
4872                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4873                         if (table > b)
4874                                 b = table;
4875                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4876                         num = EXT4_INODES_PER_GROUP(sb);
4877                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4878                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4879                                 num -= ext4_itable_unused_count(sb, gdp);
4880                         table += num / inodes_per_block;
4881                         if (end > table)
4882                                 end = table;
4883                         while (b <= end)
4884                                 sb_breadahead(sb, b++);
4885                 }
4886
4887                 /*
4888                  * There are other valid inodes in the buffer, this inode
4889                  * has in-inode xattrs, or we don't have this inode in memory.
4890                  * Read the block from disk.
4891                  */
4892                 get_bh(bh);
4893                 bh->b_end_io = end_buffer_read_sync;
4894                 submit_bh(READ_META, bh);
4895                 wait_on_buffer(bh);
4896                 if (!buffer_uptodate(bh)) {
4897                         ext4_error(sb, "unable to read inode block - inode=%lu,"
4898                                    " block=%llu", inode->i_ino, block);
4899                         brelse(bh);
4900                         return -EIO;
4901                 }
4902         }
4903 has_buffer:
4904         iloc->bh = bh;
4905         return 0;
4906 }
4907
4908 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4909 {
4910         /* We have all inode data except xattrs in memory here. */
4911         return __ext4_get_inode_loc(inode, iloc,
4912                 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
4913 }
4914
4915 void ext4_set_inode_flags(struct inode *inode)
4916 {
4917         unsigned int flags = EXT4_I(inode)->i_flags;
4918
4919         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4920         if (flags & EXT4_SYNC_FL)
4921                 inode->i_flags |= S_SYNC;
4922         if (flags & EXT4_APPEND_FL)
4923                 inode->i_flags |= S_APPEND;
4924         if (flags & EXT4_IMMUTABLE_FL)
4925                 inode->i_flags |= S_IMMUTABLE;
4926         if (flags & EXT4_NOATIME_FL)
4927                 inode->i_flags |= S_NOATIME;
4928         if (flags & EXT4_DIRSYNC_FL)
4929                 inode->i_flags |= S_DIRSYNC;
4930 }
4931
4932 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4933 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4934 {
4935         unsigned int flags = ei->vfs_inode.i_flags;
4936
4937         ei->i_flags &= ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4938                         EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|EXT4_DIRSYNC_FL);
4939         if (flags & S_SYNC)
4940                 ei->i_flags |= EXT4_SYNC_FL;
4941         if (flags & S_APPEND)
4942                 ei->i_flags |= EXT4_APPEND_FL;
4943         if (flags & S_IMMUTABLE)
4944                 ei->i_flags |= EXT4_IMMUTABLE_FL;
4945         if (flags & S_NOATIME)
4946                 ei->i_flags |= EXT4_NOATIME_FL;
4947         if (flags & S_DIRSYNC)
4948                 ei->i_flags |= EXT4_DIRSYNC_FL;
4949 }
4950
4951 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4952                                   struct ext4_inode_info *ei)
4953 {
4954         blkcnt_t i_blocks ;
4955         struct inode *inode = &(ei->vfs_inode);
4956         struct super_block *sb = inode->i_sb;
4957
4958         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4959                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4960                 /* we are using combined 48 bit field */
4961                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4962                                         le32_to_cpu(raw_inode->i_blocks_lo);
4963                 if (ei->i_flags & EXT4_HUGE_FILE_FL) {
4964                         /* i_blocks represent file system block size */
4965                         return i_blocks  << (inode->i_blkbits - 9);
4966                 } else {
4967                         return i_blocks;
4968                 }
4969         } else {
4970                 return le32_to_cpu(raw_inode->i_blocks_lo);
4971         }
4972 }
4973
4974 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4975 {
4976         struct ext4_iloc iloc;
4977         struct ext4_inode *raw_inode;
4978         struct ext4_inode_info *ei;
4979         struct inode *inode;
4980         journal_t *journal = EXT4_SB(sb)->s_journal;
4981         long ret;
4982         int block;
4983
4984         inode = iget_locked(sb, ino);
4985         if (!inode)
4986                 return ERR_PTR(-ENOMEM);
4987         if (!(inode->i_state & I_NEW))
4988                 return inode;
4989
4990         ei = EXT4_I(inode);
4991         iloc.bh = 0;
4992
4993         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4994         if (ret < 0)
4995                 goto bad_inode;
4996         raw_inode = ext4_raw_inode(&iloc);
4997         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4998         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4999         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
5000         if (!(test_opt(inode->i_sb, NO_UID32))) {
5001                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
5002                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
5003         }
5004         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
5005
5006         ei->i_state_flags = 0;
5007         ei->i_dir_start_lookup = 0;
5008         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
5009         /* We now have enough fields to check if the inode was active or not.
5010          * This is needed because nfsd might try to access dead inodes
5011          * the test is that same one that e2fsck uses
5012          * NeilBrown 1999oct15
5013          */
5014         if (inode->i_nlink == 0) {
5015                 if (inode->i_mode == 0 ||
5016                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
5017                         /* this inode is deleted */
5018                         ret = -ESTALE;
5019                         goto bad_inode;
5020                 }
5021                 /* The only unlinked inodes we let through here have
5022                  * valid i_mode and are being read by the orphan
5023                  * recovery code: that's fine, we're about to complete
5024                  * the process of deleting those. */
5025         }
5026         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
5027         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
5028         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
5029         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
5030                 ei->i_file_acl |=
5031                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
5032         inode->i_size = ext4_isize(raw_inode);
5033         ei->i_disksize = inode->i_size;
5034 #ifdef CONFIG_QUOTA
5035         ei->i_reserved_quota = 0;
5036 #endif
5037         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
5038         ei->i_block_group = iloc.block_group;
5039         ei->i_last_alloc_group = ~0;
5040         /*
5041          * NOTE! The in-memory inode i_data array is in little-endian order
5042          * even on big-endian machines: we do NOT byteswap the block numbers!
5043          */
5044         for (block = 0; block < EXT4_N_BLOCKS; block++)
5045                 ei->i_data[block] = raw_inode->i_block[block];
5046         INIT_LIST_HEAD(&ei->i_orphan);
5047
5048         /*
5049          * Set transaction id's of transactions that have to be committed
5050          * to finish f[data]sync. We set them to currently running transaction
5051          * as we cannot be sure that the inode or some of its metadata isn't
5052          * part of the transaction - the inode could have been reclaimed and
5053          * now it is reread from disk.
5054          */
5055         if (journal) {
5056                 transaction_t *transaction;
5057                 tid_t tid;
5058
5059                 spin_lock(&journal->j_state_lock);
5060                 if (journal->j_running_transaction)
5061                         transaction = journal->j_running_transaction;
5062                 else
5063                         transaction = journal->j_committing_transaction;
5064                 if (transaction)
5065                         tid = transaction->t_tid;
5066                 else
5067                         tid = journal->j_commit_sequence;
5068                 spin_unlock(&journal->j_state_lock);
5069                 ei->i_sync_tid = tid;
5070                 ei->i_datasync_tid = tid;
5071         }
5072
5073         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5074                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
5075                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
5076                     EXT4_INODE_SIZE(inode->i_sb)) {
5077                         ret = -EIO;
5078                         goto bad_inode;
5079                 }
5080                 if (ei->i_extra_isize == 0) {
5081                         /* The extra space is currently unused. Use it. */
5082                         ei->i_extra_isize = sizeof(struct ext4_inode) -
5083                                             EXT4_GOOD_OLD_INODE_SIZE;
5084                 } else {
5085                         __le32 *magic = (void *)raw_inode +
5086                                         EXT4_GOOD_OLD_INODE_SIZE +
5087                                         ei->i_extra_isize;
5088                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
5089                                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
5090                 }
5091         } else
5092                 ei->i_extra_isize = 0;
5093
5094         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
5095         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
5096         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
5097         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
5098
5099         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
5100         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5101                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5102                         inode->i_version |=
5103                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
5104         }
5105
5106         ret = 0;
5107         if (ei->i_file_acl &&
5108             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
5109                 ext4_error(sb, "bad extended attribute block %llu inode #%lu",
5110                            ei->i_file_acl, inode->i_ino);
5111                 ret = -EIO;
5112                 goto bad_inode;
5113         } else if (ei->i_flags & EXT4_EXTENTS_FL) {
5114                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5115                     (S_ISLNK(inode->i_mode) &&
5116                      !ext4_inode_is_fast_symlink(inode)))
5117                         /* Validate extent which is part of inode */
5118                         ret = ext4_ext_check_inode(inode);
5119         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5120                    (S_ISLNK(inode->i_mode) &&
5121                     !ext4_inode_is_fast_symlink(inode))) {
5122                 /* Validate block references which are part of inode */
5123                 ret = ext4_check_inode_blockref(inode);
5124         }
5125         if (ret)
5126                 goto bad_inode;
5127
5128         if (S_ISREG(inode->i_mode)) {
5129                 inode->i_op = &ext4_file_inode_operations;
5130                 inode->i_fop = &ext4_file_operations;
5131                 ext4_set_aops(inode);
5132         } else if (S_ISDIR(inode->i_mode)) {
5133                 inode->i_op = &ext4_dir_inode_operations;
5134                 inode->i_fop = &ext4_dir_operations;
5135         } else if (S_ISLNK(inode->i_mode)) {
5136                 if (ext4_inode_is_fast_symlink(inode)) {
5137                         inode->i_op = &ext4_fast_symlink_inode_operations;
5138                         nd_terminate_link(ei->i_data, inode->i_size,
5139                                 sizeof(ei->i_data) - 1);
5140                 } else {
5141                         inode->i_op = &ext4_symlink_inode_operations;
5142                         ext4_set_aops(inode);
5143                 }
5144         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
5145               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
5146                 inode->i_op = &ext4_special_inode_operations;
5147                 if (raw_inode->i_block[0])
5148                         init_special_inode(inode, inode->i_mode,
5149                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
5150                 else
5151                         init_special_inode(inode, inode->i_mode,
5152                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
5153         } else {
5154                 ret = -EIO;
5155                 ext4_error(inode->i_sb, "bogus i_mode (%o) for inode=%lu",
5156                            inode->i_mode, inode->i_ino);
5157                 goto bad_inode;
5158         }
5159         brelse(iloc.bh);
5160         ext4_set_inode_flags(inode);
5161         unlock_new_inode(inode);
5162         return inode;
5163
5164 bad_inode:
5165         brelse(iloc.bh);
5166         iget_failed(inode);
5167         return ERR_PTR(ret);
5168 }
5169
5170 static int ext4_inode_blocks_set(handle_t *handle,
5171                                 struct ext4_inode *raw_inode,
5172                                 struct ext4_inode_info *ei)
5173 {
5174         struct inode *inode = &(ei->vfs_inode);
5175         u64 i_blocks = inode->i_blocks;
5176         struct super_block *sb = inode->i_sb;
5177
5178         if (i_blocks <= ~0U) {
5179                 /*
5180                  * i_blocks can be represnted in a 32 bit variable
5181                  * as multiple of 512 bytes
5182                  */
5183                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5184                 raw_inode->i_blocks_high = 0;
5185                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
5186                 return 0;
5187         }
5188         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
5189                 return -EFBIG;
5190
5191         if (i_blocks <= 0xffffffffffffULL) {
5192                 /*
5193                  * i_blocks can be represented in a 48 bit variable
5194                  * as multiple of 512 bytes
5195                  */
5196                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5197                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5198                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
5199         } else {
5200                 ei->i_flags |= EXT4_HUGE_FILE_FL;
5201                 /* i_block is stored in file system block size */
5202                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
5203                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5204                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5205         }
5206         return 0;
5207 }
5208
5209 /*
5210  * Post the struct inode info into an on-disk inode location in the
5211  * buffer-cache.  This gobbles the caller's reference to the
5212  * buffer_head in the inode location struct.
5213  *
5214  * The caller must have write access to iloc->bh.
5215  */
5216 static int ext4_do_update_inode(handle_t *handle,
5217                                 struct inode *inode,
5218                                 struct ext4_iloc *iloc)
5219 {
5220         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5221         struct ext4_inode_info *ei = EXT4_I(inode);
5222         struct buffer_head *bh = iloc->bh;
5223         int err = 0, rc, block;
5224
5225         /* For fields not not tracking in the in-memory inode,
5226          * initialise them to zero for new inodes. */
5227         if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5228                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5229
5230         ext4_get_inode_flags(ei);
5231         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5232         if (!(test_opt(inode->i_sb, NO_UID32))) {
5233                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
5234                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
5235 /*
5236  * Fix up interoperability with old kernels. Otherwise, old inodes get
5237  * re-used with the upper 16 bits of the uid/gid intact
5238  */
5239                 if (!ei->i_dtime) {
5240                         raw_inode->i_uid_high =
5241                                 cpu_to_le16(high_16_bits(inode->i_uid));
5242                         raw_inode->i_gid_high =
5243                                 cpu_to_le16(high_16_bits(inode->i_gid));
5244                 } else {
5245                         raw_inode->i_uid_high = 0;
5246                         raw_inode->i_gid_high = 0;
5247                 }
5248         } else {
5249                 raw_inode->i_uid_low =
5250                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
5251                 raw_inode->i_gid_low =
5252                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
5253                 raw_inode->i_uid_high = 0;
5254                 raw_inode->i_gid_high = 0;
5255         }
5256         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5257
5258         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5259         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5260         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5261         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5262
5263         if (ext4_inode_blocks_set(handle, raw_inode, ei))
5264                 goto out_brelse;
5265         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5266         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
5267         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
5268             cpu_to_le32(EXT4_OS_HURD))
5269                 raw_inode->i_file_acl_high =
5270                         cpu_to_le16(ei->i_file_acl >> 32);
5271         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5272         ext4_isize_set(raw_inode, ei->i_disksize);
5273         if (ei->i_disksize > 0x7fffffffULL) {
5274                 struct super_block *sb = inode->i_sb;
5275                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
5276                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
5277                                 EXT4_SB(sb)->s_es->s_rev_level ==
5278                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
5279                         /* If this is the first large file
5280                          * created, add a flag to the superblock.
5281                          */
5282                         err = ext4_journal_get_write_access(handle,
5283                                         EXT4_SB(sb)->s_sbh);
5284                         if (err)
5285                                 goto out_brelse;
5286                         ext4_update_dynamic_rev(sb);
5287                         EXT4_SET_RO_COMPAT_FEATURE(sb,
5288                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
5289                         sb->s_dirt = 1;
5290                         ext4_handle_sync(handle);
5291                         err = ext4_handle_dirty_metadata(handle, NULL,
5292                                         EXT4_SB(sb)->s_sbh);
5293                 }
5294         }
5295         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5296         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5297                 if (old_valid_dev(inode->i_rdev)) {
5298                         raw_inode->i_block[0] =
5299                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
5300                         raw_inode->i_block[1] = 0;
5301                 } else {
5302                         raw_inode->i_block[0] = 0;
5303                         raw_inode->i_block[1] =
5304                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
5305                         raw_inode->i_block[2] = 0;
5306                 }
5307         } else
5308                 for (block = 0; block < EXT4_N_BLOCKS; block++)
5309                         raw_inode->i_block[block] = ei->i_data[block];
5310
5311         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
5312         if (ei->i_extra_isize) {
5313                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5314                         raw_inode->i_version_hi =
5315                         cpu_to_le32(inode->i_version >> 32);
5316                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
5317         }
5318
5319         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5320         rc = ext4_handle_dirty_metadata(handle, NULL, bh);
5321         if (!err)
5322                 err = rc;
5323         ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5324
5325         ext4_update_inode_fsync_trans(handle, inode, 0);
5326 out_brelse:
5327         brelse(bh);
5328         ext4_std_error(inode->i_sb, err);
5329         return err;
5330 }
5331
5332 /*
5333  * ext4_write_inode()
5334  *
5335  * We are called from a few places:
5336  *
5337  * - Within generic_file_write() for O_SYNC files.
5338  *   Here, there will be no transaction running. We wait for any running
5339  *   trasnaction to commit.
5340  *
5341  * - Within sys_sync(), kupdate and such.
5342  *   We wait on commit, if tol to.
5343  *
5344  * - Within prune_icache() (PF_MEMALLOC == true)
5345  *   Here we simply return.  We can't afford to block kswapd on the
5346  *   journal commit.
5347  *
5348  * In all cases it is actually safe for us to return without doing anything,
5349  * because the inode has been copied into a raw inode buffer in
5350  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
5351  * knfsd.
5352  *
5353  * Note that we are absolutely dependent upon all inode dirtiers doing the
5354  * right thing: they *must* call mark_inode_dirty() after dirtying info in
5355  * which we are interested.
5356  *
5357  * It would be a bug for them to not do this.  The code:
5358  *
5359  *      mark_inode_dirty(inode)
5360  *      stuff();
5361  *      inode->i_size = expr;
5362  *
5363  * is in error because a kswapd-driven write_inode() could occur while
5364  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
5365  * will no longer be on the superblock's dirty inode list.
5366  */
5367 int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5368 {
5369         int err;
5370
5371         if (current->flags & PF_MEMALLOC)
5372                 return 0;
5373
5374         if (EXT4_SB(inode->i_sb)->s_journal) {
5375                 if (ext4_journal_current_handle()) {
5376                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5377                         dump_stack();
5378                         return -EIO;
5379                 }
5380
5381                 if (wbc->sync_mode != WB_SYNC_ALL)
5382                         return 0;
5383
5384                 err = ext4_force_commit(inode->i_sb);
5385         } else {
5386                 struct ext4_iloc iloc;
5387
5388                 err = __ext4_get_inode_loc(inode, &iloc, 0);
5389                 if (err)
5390                         return err;
5391                 if (wbc->sync_mode == WB_SYNC_ALL)
5392                         sync_dirty_buffer(iloc.bh);
5393                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5394                         ext4_error(inode->i_sb, "IO error syncing inode, "
5395                                    "inode=%lu, block=%llu", inode->i_ino,
5396                                    (unsigned long long)iloc.bh->b_blocknr);
5397                         err = -EIO;
5398                 }
5399                 brelse(iloc.bh);
5400         }
5401         return err;
5402 }
5403
5404 /*
5405  * ext4_setattr()
5406  *
5407  * Called from notify_change.
5408  *
5409  * We want to trap VFS attempts to truncate the file as soon as
5410  * possible.  In particular, we want to make sure that when the VFS
5411  * shrinks i_size, we put the inode on the orphan list and modify
5412  * i_disksize immediately, so that during the subsequent flushing of
5413  * dirty pages and freeing of disk blocks, we can guarantee that any
5414  * commit will leave the blocks being flushed in an unused state on
5415  * disk.  (On recovery, the inode will get truncated and the blocks will
5416  * be freed, so we have a strong guarantee that no future commit will
5417  * leave these blocks visible to the user.)
5418  *
5419  * Another thing we have to assure is that if we are in ordered mode
5420  * and inode is still attached to the committing transaction, we must
5421  * we start writeout of all the dirty pages which are being truncated.
5422  * This way we are sure that all the data written in the previous
5423  * transaction are already on disk (truncate waits for pages under
5424  * writeback).
5425  *
5426  * Called with inode->i_mutex down.
5427  */
5428 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5429 {
5430         struct inode *inode = dentry->d_inode;
5431         int error, rc = 0;
5432         const unsigned int ia_valid = attr->ia_valid;
5433
5434         error = inode_change_ok(inode, attr);
5435         if (error)
5436                 return error;
5437
5438         if (ia_valid & ATTR_SIZE)
5439                 dquot_initialize(inode);
5440         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
5441                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
5442                 handle_t *handle;
5443
5444                 /* (user+group)*(old+new) structure, inode write (sb,
5445                  * inode block, ? - but truncate inode update has it) */
5446                 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
5447                                         EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
5448                 if (IS_ERR(handle)) {
5449                         error = PTR_ERR(handle);
5450                         goto err_out;
5451                 }
5452                 error = dquot_transfer(inode, attr);
5453                 if (error) {
5454                         ext4_journal_stop(handle);
5455                         return error;
5456                 }
5457                 /* Update corresponding info in inode so that everything is in
5458                  * one transaction */
5459                 if (attr->ia_valid & ATTR_UID)
5460                         inode->i_uid = attr->ia_uid;
5461                 if (attr->ia_valid & ATTR_GID)
5462                         inode->i_gid = attr->ia_gid;
5463                 error = ext4_mark_inode_dirty(handle, inode);
5464                 ext4_journal_stop(handle);
5465         }
5466
5467         if (attr->ia_valid & ATTR_SIZE) {
5468                 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)) {
5469                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5470
5471                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
5472                                 error = -EFBIG;
5473                                 goto err_out;
5474                         }
5475                 }
5476         }
5477
5478         if (S_ISREG(inode->i_mode) &&
5479             attr->ia_valid & ATTR_SIZE &&
5480             (attr->ia_size < inode->i_size ||
5481              (EXT4_I(inode)->i_flags & EXT4_EOFBLOCKS_FL))) {
5482                 handle_t *handle;
5483
5484                 handle = ext4_journal_start(inode, 3);
5485                 if (IS_ERR(handle)) {
5486                         error = PTR_ERR(handle);
5487                         goto err_out;
5488                 }
5489
5490                 error = ext4_orphan_add(handle, inode);
5491                 EXT4_I(inode)->i_disksize = attr->ia_size;
5492                 rc = ext4_mark_inode_dirty(handle, inode);
5493                 if (!error)
5494                         error = rc;
5495                 ext4_journal_stop(handle);
5496
5497                 if (ext4_should_order_data(inode)) {
5498                         error = ext4_begin_ordered_truncate(inode,
5499                                                             attr->ia_size);
5500                         if (error) {
5501                                 /* Do as much error cleanup as possible */
5502                                 handle = ext4_journal_start(inode, 3);
5503                                 if (IS_ERR(handle)) {
5504                                         ext4_orphan_del(NULL, inode);
5505                                         goto err_out;
5506                                 }
5507                                 ext4_orphan_del(handle, inode);
5508                                 ext4_journal_stop(handle);
5509                                 goto err_out;
5510                         }
5511                 }
5512                 /* ext4_truncate will clear the flag */
5513                 if ((EXT4_I(inode)->i_flags & EXT4_EOFBLOCKS_FL))
5514                         ext4_truncate(inode);
5515         }
5516
5517         rc = inode_setattr(inode, attr);
5518
5519         /* If inode_setattr's call to ext4_truncate failed to get a
5520          * transaction handle at all, we need to clean up the in-core
5521          * orphan list manually. */
5522         if (inode->i_nlink)
5523                 ext4_orphan_del(NULL, inode);
5524
5525         if (!rc && (ia_valid & ATTR_MODE))
5526                 rc = ext4_acl_chmod(inode);
5527
5528 err_out:
5529         ext4_std_error(inode->i_sb, error);
5530         if (!error)
5531                 error = rc;
5532         return error;
5533 }
5534
5535 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
5536                  struct kstat *stat)
5537 {
5538         struct inode *inode;
5539         unsigned long delalloc_blocks;
5540
5541         inode = dentry->d_inode;
5542         generic_fillattr(inode, stat);
5543
5544         /*
5545          * We can't update i_blocks if the block allocation is delayed
5546          * otherwise in the case of system crash before the real block
5547          * allocation is done, we will have i_blocks inconsistent with
5548          * on-disk file blocks.
5549          * We always keep i_blocks updated together with real
5550          * allocation. But to not confuse with user, stat