ext4: remove vestiges of nobh support
[linux-2.6.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/jbd2.h>
29 #include <linux/highuid.h>
30 #include <linux/pagemap.h>
31 #include <linux/quotaops.h>
32 #include <linux/string.h>
33 #include <linux/buffer_head.h>
34 #include <linux/writeback.h>
35 #include <linux/pagevec.h>
36 #include <linux/mpage.h>
37 #include <linux/namei.h>
38 #include <linux/uio.h>
39 #include <linux/bio.h>
40 #include <linux/workqueue.h>
41 #include <linux/kernel.h>
42 #include <linux/slab.h>
43
44 #include "ext4_jbd2.h"
45 #include "xattr.h"
46 #include "acl.h"
47 #include "ext4_extents.h"
48
49 #include <trace/events/ext4.h>
50
51 #define MPAGE_DA_EXTENT_TAIL 0x01
52
53 static inline int ext4_begin_ordered_truncate(struct inode *inode,
54                                               loff_t new_size)
55 {
56         return jbd2_journal_begin_ordered_truncate(
57                                         EXT4_SB(inode->i_sb)->s_journal,
58                                         &EXT4_I(inode)->jinode,
59                                         new_size);
60 }
61
62 static void ext4_invalidatepage(struct page *page, unsigned long offset);
63
64 /*
65  * Test whether an inode is a fast symlink.
66  */
67 static int ext4_inode_is_fast_symlink(struct inode *inode)
68 {
69         int ea_blocks = EXT4_I(inode)->i_file_acl ?
70                 (inode->i_sb->s_blocksize >> 9) : 0;
71
72         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
73 }
74
75 /*
76  * Work out how many blocks we need to proceed with the next chunk of a
77  * truncate transaction.
78  */
79 static unsigned long blocks_for_truncate(struct inode *inode)
80 {
81         ext4_lblk_t needed;
82
83         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
84
85         /* Give ourselves just enough room to cope with inodes in which
86          * i_blocks is corrupt: we've seen disk corruptions in the past
87          * which resulted in random data in an inode which looked enough
88          * like a regular file for ext4 to try to delete it.  Things
89          * will go a bit crazy if that happens, but at least we should
90          * try not to panic the whole kernel. */
91         if (needed < 2)
92                 needed = 2;
93
94         /* But we need to bound the transaction so we don't overflow the
95          * journal. */
96         if (needed > EXT4_MAX_TRANS_DATA)
97                 needed = EXT4_MAX_TRANS_DATA;
98
99         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
100 }
101
102 /*
103  * Truncate transactions can be complex and absolutely huge.  So we need to
104  * be able to restart the transaction at a conventient checkpoint to make
105  * sure we don't overflow the journal.
106  *
107  * start_transaction gets us a new handle for a truncate transaction,
108  * and extend_transaction tries to extend the existing one a bit.  If
109  * extend fails, we need to propagate the failure up and restart the
110  * transaction in the top-level truncate loop. --sct
111  */
112 static handle_t *start_transaction(struct inode *inode)
113 {
114         handle_t *result;
115
116         result = ext4_journal_start(inode, blocks_for_truncate(inode));
117         if (!IS_ERR(result))
118                 return result;
119
120         ext4_std_error(inode->i_sb, PTR_ERR(result));
121         return result;
122 }
123
124 /*
125  * Try to extend this transaction for the purposes of truncation.
126  *
127  * Returns 0 if we managed to create more room.  If we can't create more
128  * room, and the transaction must be restarted we return 1.
129  */
130 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
131 {
132         if (!ext4_handle_valid(handle))
133                 return 0;
134         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
135                 return 0;
136         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
137                 return 0;
138         return 1;
139 }
140
141 /*
142  * Restart the transaction associated with *handle.  This does a commit,
143  * so before we call here everything must be consistently dirtied against
144  * this transaction.
145  */
146 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
147                                  int nblocks)
148 {
149         int ret;
150
151         /*
152          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
153          * moment, get_block can be called only for blocks inside i_size since
154          * page cache has been already dropped and writes are blocked by
155          * i_mutex. So we can safely drop the i_data_sem here.
156          */
157         BUG_ON(EXT4_JOURNAL(inode) == NULL);
158         jbd_debug(2, "restarting handle %p\n", handle);
159         up_write(&EXT4_I(inode)->i_data_sem);
160         ret = ext4_journal_restart(handle, blocks_for_truncate(inode));
161         down_write(&EXT4_I(inode)->i_data_sem);
162         ext4_discard_preallocations(inode);
163
164         return ret;
165 }
166
167 /*
168  * Called at the last iput() if i_nlink is zero.
169  */
170 void ext4_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(function, inode,
352                                          "invalid block reference %u", blk);
353                         return -EIO;
354                 }
355         }
356         return 0;
357 }
358
359
360 #define ext4_check_indirect_blockref(inode, bh)                         \
361         __ext4_check_blockref(__func__, inode, (__le32 *)(bh)->b_data,  \
362                               EXT4_ADDR_PER_BLOCK((inode)->i_sb))
363
364 #define ext4_check_inode_blockref(inode)                                \
365         __ext4_check_blockref(__func__, inode, EXT4_I(inode)->i_data,   \
366                               EXT4_NDIR_BLOCKS)
367
368 /**
369  *      ext4_get_branch - read the chain of indirect blocks leading to data
370  *      @inode: inode in question
371  *      @depth: depth of the chain (1 - direct pointer, etc.)
372  *      @offsets: offsets of pointers in inode/indirect blocks
373  *      @chain: place to store the result
374  *      @err: here we store the error value
375  *
376  *      Function fills the array of triples <key, p, bh> and returns %NULL
377  *      if everything went OK or the pointer to the last filled triple
378  *      (incomplete one) otherwise. Upon the return chain[i].key contains
379  *      the number of (i+1)-th block in the chain (as it is stored in memory,
380  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
381  *      number (it points into struct inode for i==0 and into the bh->b_data
382  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
383  *      block for i>0 and NULL for i==0. In other words, it holds the block
384  *      numbers of the chain, addresses they were taken from (and where we can
385  *      verify that chain did not change) and buffer_heads hosting these
386  *      numbers.
387  *
388  *      Function stops when it stumbles upon zero pointer (absent block)
389  *              (pointer to last triple returned, *@err == 0)
390  *      or when it gets an IO error reading an indirect block
391  *              (ditto, *@err == -EIO)
392  *      or when it reads all @depth-1 indirect blocks successfully and finds
393  *      the whole chain, all way to the data (returns %NULL, *err == 0).
394  *
395  *      Need to be called with
396  *      down_read(&EXT4_I(inode)->i_data_sem)
397  */
398 static Indirect *ext4_get_branch(struct inode *inode, int depth,
399                                  ext4_lblk_t  *offsets,
400                                  Indirect chain[4], int *err)
401 {
402         struct super_block *sb = inode->i_sb;
403         Indirect *p = chain;
404         struct buffer_head *bh;
405
406         *err = 0;
407         /* i_data is not going away, no lock needed */
408         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
409         if (!p->key)
410                 goto no_block;
411         while (--depth) {
412                 bh = sb_getblk(sb, le32_to_cpu(p->key));
413                 if (unlikely(!bh))
414                         goto failure;
415
416                 if (!bh_uptodate_or_lock(bh)) {
417                         if (bh_submit_read(bh) < 0) {
418                                 put_bh(bh);
419                                 goto failure;
420                         }
421                         /* validate block references */
422                         if (ext4_check_indirect_blockref(inode, bh)) {
423                                 put_bh(bh);
424                                 goto failure;
425                         }
426                 }
427
428                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
429                 /* Reader: end */
430                 if (!p->key)
431                         goto no_block;
432         }
433         return NULL;
434
435 failure:
436         *err = -EIO;
437 no_block:
438         return p;
439 }
440
441 /**
442  *      ext4_find_near - find a place for allocation with sufficient locality
443  *      @inode: owner
444  *      @ind: descriptor of indirect block.
445  *
446  *      This function returns the preferred place for block allocation.
447  *      It is used when heuristic for sequential allocation fails.
448  *      Rules are:
449  *        + if there is a block to the left of our position - allocate near it.
450  *        + if pointer will live in indirect block - allocate near that block.
451  *        + if pointer will live in inode - allocate in the same
452  *          cylinder group.
453  *
454  * In the latter case we colour the starting block by the callers PID to
455  * prevent it from clashing with concurrent allocations for a different inode
456  * in the same block group.   The PID is used here so that functionally related
457  * files will be close-by on-disk.
458  *
459  *      Caller must make sure that @ind is valid and will stay that way.
460  */
461 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
462 {
463         struct ext4_inode_info *ei = EXT4_I(inode);
464         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
465         __le32 *p;
466         ext4_fsblk_t bg_start;
467         ext4_fsblk_t last_block;
468         ext4_grpblk_t colour;
469         ext4_group_t block_group;
470         int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
471
472         /* Try to find previous block */
473         for (p = ind->p - 1; p >= start; p--) {
474                 if (*p)
475                         return le32_to_cpu(*p);
476         }
477
478         /* No such thing, so let's try location of indirect block */
479         if (ind->bh)
480                 return ind->bh->b_blocknr;
481
482         /*
483          * It is going to be referred to from the inode itself? OK, just put it
484          * into the same cylinder group then.
485          */
486         block_group = ei->i_block_group;
487         if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
488                 block_group &= ~(flex_size-1);
489                 if (S_ISREG(inode->i_mode))
490                         block_group++;
491         }
492         bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
493         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
494
495         /*
496          * If we are doing delayed allocation, we don't need take
497          * colour into account.
498          */
499         if (test_opt(inode->i_sb, DELALLOC))
500                 return bg_start;
501
502         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
503                 colour = (current->pid % 16) *
504                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
505         else
506                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
507         return bg_start + colour;
508 }
509
510 /**
511  *      ext4_find_goal - find a preferred place for allocation.
512  *      @inode: owner
513  *      @block:  block we want
514  *      @partial: pointer to the last triple within a chain
515  *
516  *      Normally this function find the preferred place for block allocation,
517  *      returns it.
518  *      Because this is only used for non-extent files, we limit the block nr
519  *      to 32 bits.
520  */
521 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
522                                    Indirect *partial)
523 {
524         ext4_fsblk_t goal;
525
526         /*
527          * XXX need to get goal block from mballoc's data structures
528          */
529
530         goal = ext4_find_near(inode, partial);
531         goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
532         return goal;
533 }
534
535 /**
536  *      ext4_blks_to_allocate: Look up the block map and count the number
537  *      of direct blocks need to be allocated for the given branch.
538  *
539  *      @branch: chain of indirect blocks
540  *      @k: number of blocks need for indirect blocks
541  *      @blks: number of data blocks to be mapped.
542  *      @blocks_to_boundary:  the offset in the indirect block
543  *
544  *      return the total number of blocks to be allocate, including the
545  *      direct and indirect blocks.
546  */
547 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
548                                  int blocks_to_boundary)
549 {
550         unsigned int count = 0;
551
552         /*
553          * Simple case, [t,d]Indirect block(s) has not allocated yet
554          * then it's clear blocks on that path have not allocated
555          */
556         if (k > 0) {
557                 /* right now we don't handle cross boundary allocation */
558                 if (blks < blocks_to_boundary + 1)
559                         count += blks;
560                 else
561                         count += blocks_to_boundary + 1;
562                 return count;
563         }
564
565         count++;
566         while (count < blks && count <= blocks_to_boundary &&
567                 le32_to_cpu(*(branch[0].p + count)) == 0) {
568                 count++;
569         }
570         return count;
571 }
572
573 /**
574  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
575  *      @indirect_blks: the number of blocks need to allocate for indirect
576  *                      blocks
577  *
578  *      @new_blocks: on return it will store the new block numbers for
579  *      the indirect blocks(if needed) and the first direct block,
580  *      @blks:  on return it will store the total number of allocated
581  *              direct blocks
582  */
583 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
584                              ext4_lblk_t iblock, ext4_fsblk_t goal,
585                              int indirect_blks, int blks,
586                              ext4_fsblk_t new_blocks[4], int *err)
587 {
588         struct ext4_allocation_request ar;
589         int target, i;
590         unsigned long count = 0, blk_allocated = 0;
591         int index = 0;
592         ext4_fsblk_t current_block = 0;
593         int ret = 0;
594
595         /*
596          * Here we try to allocate the requested multiple blocks at once,
597          * on a best-effort basis.
598          * To build a branch, we should allocate blocks for
599          * the indirect blocks(if not allocated yet), and at least
600          * the first direct block of this branch.  That's the
601          * minimum number of blocks need to allocate(required)
602          */
603         /* first we try to allocate the indirect blocks */
604         target = indirect_blks;
605         while (target > 0) {
606                 count = target;
607                 /* allocating blocks for indirect blocks and direct blocks */
608                 current_block = ext4_new_meta_blocks(handle, inode,
609                                                         goal, &count, err);
610                 if (*err)
611                         goto failed_out;
612
613                 if (unlikely(current_block + count > EXT4_MAX_BLOCK_FILE_PHYS)) {
614                         EXT4_ERROR_INODE(inode,
615                                          "current_block %llu + count %lu > %d!",
616                                          current_block, count,
617                                          EXT4_MAX_BLOCK_FILE_PHYS);
618                         *err = -EIO;
619                         goto failed_out;
620                 }
621
622                 target -= count;
623                 /* allocate blocks for indirect blocks */
624                 while (index < indirect_blks && count) {
625                         new_blocks[index++] = current_block++;
626                         count--;
627                 }
628                 if (count > 0) {
629                         /*
630                          * save the new block number
631                          * for the first direct block
632                          */
633                         new_blocks[index] = current_block;
634                         printk(KERN_INFO "%s returned more blocks than "
635                                                 "requested\n", __func__);
636                         WARN_ON(1);
637                         break;
638                 }
639         }
640
641         target = blks - count ;
642         blk_allocated = count;
643         if (!target)
644                 goto allocated;
645         /* Now allocate data blocks */
646         memset(&ar, 0, sizeof(ar));
647         ar.inode = inode;
648         ar.goal = goal;
649         ar.len = target;
650         ar.logical = iblock;
651         if (S_ISREG(inode->i_mode))
652                 /* enable in-core preallocation only for regular files */
653                 ar.flags = EXT4_MB_HINT_DATA;
654
655         current_block = ext4_mb_new_blocks(handle, &ar, err);
656         if (unlikely(current_block + ar.len > EXT4_MAX_BLOCK_FILE_PHYS)) {
657                 EXT4_ERROR_INODE(inode,
658                                  "current_block %llu + ar.len %d > %d!",
659                                  current_block, ar.len,
660                                  EXT4_MAX_BLOCK_FILE_PHYS);
661                 *err = -EIO;
662                 goto failed_out;
663         }
664
665         if (*err && (target == blks)) {
666                 /*
667                  * if the allocation failed and we didn't allocate
668                  * any blocks before
669                  */
670                 goto failed_out;
671         }
672         if (!*err) {
673                 if (target == blks) {
674                         /*
675                          * save the new block number
676                          * for the first direct block
677                          */
678                         new_blocks[index] = current_block;
679                 }
680                 blk_allocated += ar.len;
681         }
682 allocated:
683         /* total number of blocks allocated for direct blocks */
684         ret = blk_allocated;
685         *err = 0;
686         return ret;
687 failed_out:
688         for (i = 0; i < index; i++)
689                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
690         return ret;
691 }
692
693 /**
694  *      ext4_alloc_branch - allocate and set up a chain of blocks.
695  *      @inode: owner
696  *      @indirect_blks: number of allocated indirect blocks
697  *      @blks: number of allocated direct blocks
698  *      @offsets: offsets (in the blocks) to store the pointers to next.
699  *      @branch: place to store the chain in.
700  *
701  *      This function allocates blocks, zeroes out all but the last one,
702  *      links them into chain and (if we are synchronous) writes them to disk.
703  *      In other words, it prepares a branch that can be spliced onto the
704  *      inode. It stores the information about that chain in the branch[], in
705  *      the same format as ext4_get_branch() would do. We are calling it after
706  *      we had read the existing part of chain and partial points to the last
707  *      triple of that (one with zero ->key). Upon the exit we have the same
708  *      picture as after the successful ext4_get_block(), except that in one
709  *      place chain is disconnected - *branch->p is still zero (we did not
710  *      set the last link), but branch->key contains the number that should
711  *      be placed into *branch->p to fill that gap.
712  *
713  *      If allocation fails we free all blocks we've allocated (and forget
714  *      their buffer_heads) and return the error value the from failed
715  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
716  *      as described above and return 0.
717  */
718 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
719                              ext4_lblk_t iblock, int indirect_blks,
720                              int *blks, ext4_fsblk_t goal,
721                              ext4_lblk_t *offsets, Indirect *branch)
722 {
723         int blocksize = inode->i_sb->s_blocksize;
724         int i, n = 0;
725         int err = 0;
726         struct buffer_head *bh;
727         int num;
728         ext4_fsblk_t new_blocks[4];
729         ext4_fsblk_t current_block;
730
731         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
732                                 *blks, new_blocks, &err);
733         if (err)
734                 return err;
735
736         branch[0].key = cpu_to_le32(new_blocks[0]);
737         /*
738          * metadata blocks and data blocks are allocated.
739          */
740         for (n = 1; n <= indirect_blks;  n++) {
741                 /*
742                  * Get buffer_head for parent block, zero it out
743                  * and set the pointer to new one, then send
744                  * parent to disk.
745                  */
746                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
747                 branch[n].bh = bh;
748                 lock_buffer(bh);
749                 BUFFER_TRACE(bh, "call get_create_access");
750                 err = ext4_journal_get_create_access(handle, bh);
751                 if (err) {
752                         /* Don't brelse(bh) here; it's done in
753                          * ext4_journal_forget() below */
754                         unlock_buffer(bh);
755                         goto failed;
756                 }
757
758                 memset(bh->b_data, 0, blocksize);
759                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
760                 branch[n].key = cpu_to_le32(new_blocks[n]);
761                 *branch[n].p = branch[n].key;
762                 if (n == indirect_blks) {
763                         current_block = new_blocks[n];
764                         /*
765                          * End of chain, update the last new metablock of
766                          * the chain to point to the new allocated
767                          * data blocks numbers
768                          */
769                         for (i = 1; i < num; i++)
770                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
771                 }
772                 BUFFER_TRACE(bh, "marking uptodate");
773                 set_buffer_uptodate(bh);
774                 unlock_buffer(bh);
775
776                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
777                 err = ext4_handle_dirty_metadata(handle, inode, bh);
778                 if (err)
779                         goto failed;
780         }
781         *blks = num;
782         return err;
783 failed:
784         /* Allocation failed, free what we already allocated */
785         ext4_free_blocks(handle, inode, 0, new_blocks[0], 1, 0);
786         for (i = 1; i <= n ; i++) {
787                 /*
788                  * branch[i].bh is newly allocated, so there is no
789                  * need to revoke the block, which is why we don't
790                  * need to set EXT4_FREE_BLOCKS_METADATA.
791                  */
792                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1,
793                                  EXT4_FREE_BLOCKS_FORGET);
794         }
795         for (i = n+1; i < indirect_blks; i++)
796                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
797
798         ext4_free_blocks(handle, inode, 0, new_blocks[i], num, 0);
799
800         return err;
801 }
802
803 /**
804  * ext4_splice_branch - splice the allocated branch onto inode.
805  * @inode: owner
806  * @block: (logical) number of block we are adding
807  * @chain: chain of indirect blocks (with a missing link - see
808  *      ext4_alloc_branch)
809  * @where: location of missing link
810  * @num:   number of indirect blocks we are adding
811  * @blks:  number of direct blocks we are adding
812  *
813  * This function fills the missing link and does all housekeeping needed in
814  * inode (->i_blocks, etc.). In case of success we end up with the full
815  * chain to new block and return 0.
816  */
817 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
818                               ext4_lblk_t block, Indirect *where, int num,
819                               int blks)
820 {
821         int i;
822         int err = 0;
823         ext4_fsblk_t current_block;
824
825         /*
826          * If we're splicing into a [td]indirect block (as opposed to the
827          * inode) then we need to get write access to the [td]indirect block
828          * before the splice.
829          */
830         if (where->bh) {
831                 BUFFER_TRACE(where->bh, "get_write_access");
832                 err = ext4_journal_get_write_access(handle, where->bh);
833                 if (err)
834                         goto err_out;
835         }
836         /* That's it */
837
838         *where->p = where->key;
839
840         /*
841          * Update the host buffer_head or inode to point to more just allocated
842          * direct blocks blocks
843          */
844         if (num == 0 && blks > 1) {
845                 current_block = le32_to_cpu(where->key) + 1;
846                 for (i = 1; i < blks; i++)
847                         *(where->p + i) = cpu_to_le32(current_block++);
848         }
849
850         /* We are done with atomic stuff, now do the rest of housekeeping */
851         /* had we spliced it onto indirect block? */
852         if (where->bh) {
853                 /*
854                  * If we spliced it onto an indirect block, we haven't
855                  * altered the inode.  Note however that if it is being spliced
856                  * onto an indirect block at the very end of the file (the
857                  * file is growing) then we *will* alter the inode to reflect
858                  * the new i_size.  But that is not done here - it is done in
859                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
860                  */
861                 jbd_debug(5, "splicing indirect only\n");
862                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
863                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
864                 if (err)
865                         goto err_out;
866         } else {
867                 /*
868                  * OK, we spliced it into the inode itself on a direct block.
869                  */
870                 ext4_mark_inode_dirty(handle, inode);
871                 jbd_debug(5, "splicing direct\n");
872         }
873         return err;
874
875 err_out:
876         for (i = 1; i <= num; i++) {
877                 /*
878                  * branch[i].bh is newly allocated, so there is no
879                  * need to revoke the block, which is why we don't
880                  * need to set EXT4_FREE_BLOCKS_METADATA.
881                  */
882                 ext4_free_blocks(handle, inode, where[i].bh, 0, 1,
883                                  EXT4_FREE_BLOCKS_FORGET);
884         }
885         ext4_free_blocks(handle, inode, 0, le32_to_cpu(where[num].key),
886                          blks, 0);
887
888         return err;
889 }
890
891 /*
892  * The ext4_ind_map_blocks() function handles non-extents inodes
893  * (i.e., using the traditional indirect/double-indirect i_blocks
894  * scheme) for ext4_map_blocks().
895  *
896  * Allocation strategy is simple: if we have to allocate something, we will
897  * have to go the whole way to leaf. So let's do it before attaching anything
898  * to tree, set linkage between the newborn blocks, write them if sync is
899  * required, recheck the path, free and repeat if check fails, otherwise
900  * set the last missing link (that will protect us from any truncate-generated
901  * removals - all blocks on the path are immune now) and possibly force the
902  * write on the parent block.
903  * That has a nice additional property: no special recovery from the failed
904  * allocations is needed - we simply release blocks and do not touch anything
905  * reachable from inode.
906  *
907  * `handle' can be NULL if create == 0.
908  *
909  * return > 0, # of blocks mapped or allocated.
910  * return = 0, if plain lookup failed.
911  * return < 0, error case.
912  *
913  * The ext4_ind_get_blocks() function should be called with
914  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
915  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
916  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
917  * blocks.
918  */
919 static int ext4_ind_map_blocks(handle_t *handle, struct inode *inode,
920                                struct ext4_map_blocks *map,
921                                int flags)
922 {
923         int err = -EIO;
924         ext4_lblk_t offsets[4];
925         Indirect chain[4];
926         Indirect *partial;
927         ext4_fsblk_t goal;
928         int indirect_blks;
929         int blocks_to_boundary = 0;
930         int depth;
931         int count = 0;
932         ext4_fsblk_t first_block = 0;
933
934         J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
935         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
936         depth = ext4_block_to_path(inode, map->m_lblk, offsets,
937                                    &blocks_to_boundary);
938
939         if (depth == 0)
940                 goto out;
941
942         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
943
944         /* Simplest case - block found, no allocation needed */
945         if (!partial) {
946                 first_block = le32_to_cpu(chain[depth - 1].key);
947                 count++;
948                 /*map more blocks*/
949                 while (count < map->m_len && count <= blocks_to_boundary) {
950                         ext4_fsblk_t blk;
951
952                         blk = le32_to_cpu(*(chain[depth-1].p + count));
953
954                         if (blk == first_block + count)
955                                 count++;
956                         else
957                                 break;
958                 }
959                 goto got_it;
960         }
961
962         /* Next simple case - plain lookup or failed read of indirect block */
963         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
964                 goto cleanup;
965
966         /*
967          * Okay, we need to do block allocation.
968         */
969         goal = ext4_find_goal(inode, map->m_lblk, partial);
970
971         /* the number of blocks need to allocate for [d,t]indirect blocks */
972         indirect_blks = (chain + depth) - partial - 1;
973
974         /*
975          * Next look up the indirect map to count the totoal number of
976          * direct blocks to allocate for this branch.
977          */
978         count = ext4_blks_to_allocate(partial, indirect_blks,
979                                       map->m_len, blocks_to_boundary);
980         /*
981          * Block out ext4_truncate while we alter the tree
982          */
983         err = ext4_alloc_branch(handle, inode, map->m_lblk, indirect_blks,
984                                 &count, goal,
985                                 offsets + (partial - chain), partial);
986
987         /*
988          * The ext4_splice_branch call will free and forget any buffers
989          * on the new chain if there is a failure, but that risks using
990          * up transaction credits, especially for bitmaps where the
991          * credits cannot be returned.  Can we handle this somehow?  We
992          * may need to return -EAGAIN upwards in the worst case.  --sct
993          */
994         if (!err)
995                 err = ext4_splice_branch(handle, inode, map->m_lblk,
996                                          partial, indirect_blks, count);
997         if (err)
998                 goto cleanup;
999
1000         map->m_flags |= EXT4_MAP_NEW;
1001
1002         ext4_update_inode_fsync_trans(handle, inode, 1);
1003 got_it:
1004         map->m_flags |= EXT4_MAP_MAPPED;
1005         map->m_pblk = le32_to_cpu(chain[depth-1].key);
1006         map->m_len = count;
1007         if (count > blocks_to_boundary)
1008                 map->m_flags |= EXT4_MAP_BOUNDARY;
1009         err = count;
1010         /* Clean up and exit */
1011         partial = chain + depth - 1;    /* the whole chain */
1012 cleanup:
1013         while (partial > chain) {
1014                 BUFFER_TRACE(partial->bh, "call brelse");
1015                 brelse(partial->bh);
1016                 partial--;
1017         }
1018 out:
1019         return err;
1020 }
1021
1022 #ifdef CONFIG_QUOTA
1023 qsize_t *ext4_get_reserved_space(struct inode *inode)
1024 {
1025         return &EXT4_I(inode)->i_reserved_quota;
1026 }
1027 #endif
1028
1029 /*
1030  * Calculate the number of metadata blocks need to reserve
1031  * to allocate a new block at @lblocks for non extent file based file
1032  */
1033 static int ext4_indirect_calc_metadata_amount(struct inode *inode,
1034                                               sector_t lblock)
1035 {
1036         struct ext4_inode_info *ei = EXT4_I(inode);
1037         sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
1038         int blk_bits;
1039
1040         if (lblock < EXT4_NDIR_BLOCKS)
1041                 return 0;
1042
1043         lblock -= EXT4_NDIR_BLOCKS;
1044
1045         if (ei->i_da_metadata_calc_len &&
1046             (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
1047                 ei->i_da_metadata_calc_len++;
1048                 return 0;
1049         }
1050         ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
1051         ei->i_da_metadata_calc_len = 1;
1052         blk_bits = order_base_2(lblock);
1053         return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
1054 }
1055
1056 /*
1057  * Calculate the number of metadata blocks need to reserve
1058  * to allocate a block located at @lblock
1059  */
1060 static int ext4_calc_metadata_amount(struct inode *inode, sector_t lblock)
1061 {
1062         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1063                 return ext4_ext_calc_metadata_amount(inode, lblock);
1064
1065         return ext4_indirect_calc_metadata_amount(inode, lblock);
1066 }
1067
1068 /*
1069  * Called with i_data_sem down, which is important since we can call
1070  * ext4_discard_preallocations() from here.
1071  */
1072 void ext4_da_update_reserve_space(struct inode *inode,
1073                                         int used, int quota_claim)
1074 {
1075         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1076         struct ext4_inode_info *ei = EXT4_I(inode);
1077
1078         spin_lock(&ei->i_block_reservation_lock);
1079         trace_ext4_da_update_reserve_space(inode, used);
1080         if (unlikely(used > ei->i_reserved_data_blocks)) {
1081                 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
1082                          "with only %d reserved data blocks\n",
1083                          __func__, inode->i_ino, used,
1084                          ei->i_reserved_data_blocks);
1085                 WARN_ON(1);
1086                 used = ei->i_reserved_data_blocks;
1087         }
1088
1089         /* Update per-inode reservations */
1090         ei->i_reserved_data_blocks -= used;
1091         ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
1092         percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1093                            used + ei->i_allocated_meta_blocks);
1094         ei->i_allocated_meta_blocks = 0;
1095
1096         if (ei->i_reserved_data_blocks == 0) {
1097                 /*
1098                  * We can release all of the reserved metadata blocks
1099                  * only when we have written all of the delayed
1100                  * allocation blocks.
1101                  */
1102                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1103                                    ei->i_reserved_meta_blocks);
1104                 ei->i_reserved_meta_blocks = 0;
1105                 ei->i_da_metadata_calc_len = 0;
1106         }
1107         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1108
1109         /* Update quota subsystem for data blocks */
1110         if (quota_claim)
1111                 dquot_claim_block(inode, used);
1112         else {
1113                 /*
1114                  * We did fallocate with an offset that is already delayed
1115                  * allocated. So on delayed allocated writeback we should
1116                  * not re-claim the quota for fallocated blocks.
1117                  */
1118                 dquot_release_reservation_block(inode, used);
1119         }
1120
1121         /*
1122          * If we have done all the pending block allocations and if
1123          * there aren't any writers on the inode, we can discard the
1124          * inode's preallocations.
1125          */
1126         if ((ei->i_reserved_data_blocks == 0) &&
1127             (atomic_read(&inode->i_writecount) == 0))
1128                 ext4_discard_preallocations(inode);
1129 }
1130
1131 static int check_block_validity(struct inode *inode, const char *func,
1132                                 struct ext4_map_blocks *map)
1133 {
1134         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
1135                                    map->m_len)) {
1136                 ext4_error_inode(func, inode,
1137                            "lblock %lu mapped to illegal pblock %llu "
1138                            "(length %d)", (unsigned long) map->m_lblk,
1139                                  map->m_pblk, map->m_len);
1140                 return -EIO;
1141         }
1142         return 0;
1143 }
1144
1145 /*
1146  * Return the number of contiguous dirty pages in a given inode
1147  * starting at page frame idx.
1148  */
1149 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
1150                                     unsigned int max_pages)
1151 {
1152         struct address_space *mapping = inode->i_mapping;
1153         pgoff_t index;
1154         struct pagevec pvec;
1155         pgoff_t num = 0;
1156         int i, nr_pages, done = 0;
1157
1158         if (max_pages == 0)
1159                 return 0;
1160         pagevec_init(&pvec, 0);
1161         while (!done) {
1162                 index = idx;
1163                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
1164                                               PAGECACHE_TAG_DIRTY,
1165                                               (pgoff_t)PAGEVEC_SIZE);
1166                 if (nr_pages == 0)
1167                         break;
1168                 for (i = 0; i < nr_pages; i++) {
1169                         struct page *page = pvec.pages[i];
1170                         struct buffer_head *bh, *head;
1171
1172                         lock_page(page);
1173                         if (unlikely(page->mapping != mapping) ||
1174                             !PageDirty(page) ||
1175                             PageWriteback(page) ||
1176                             page->index != idx) {
1177                                 done = 1;
1178                                 unlock_page(page);
1179                                 break;
1180                         }
1181                         if (page_has_buffers(page)) {
1182                                 bh = head = page_buffers(page);
1183                                 do {
1184                                         if (!buffer_delay(bh) &&
1185                                             !buffer_unwritten(bh))
1186                                                 done = 1;
1187                                         bh = bh->b_this_page;
1188                                 } while (!done && (bh != head));
1189                         }
1190                         unlock_page(page);
1191                         if (done)
1192                                 break;
1193                         idx++;
1194                         num++;
1195                         if (num >= max_pages)
1196                                 break;
1197                 }
1198                 pagevec_release(&pvec);
1199         }
1200         return num;
1201 }
1202
1203 /*
1204  * The ext4_map_blocks() function tries to look up the requested blocks,
1205  * and returns if the blocks are already mapped.
1206  *
1207  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1208  * and store the allocated blocks in the result buffer head and mark it
1209  * mapped.
1210  *
1211  * If file type is extents based, it will call ext4_ext_map_blocks(),
1212  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
1213  * based files
1214  *
1215  * On success, it returns the number of blocks being mapped or allocate.
1216  * if create==0 and the blocks are pre-allocated and uninitialized block,
1217  * the result buffer head is unmapped. If the create ==1, it will make sure
1218  * the buffer head is mapped.
1219  *
1220  * It returns 0 if plain look up failed (blocks have not been allocated), in
1221  * that casem, buffer head is unmapped
1222  *
1223  * It returns the error in case of allocation failure.
1224  */
1225 int ext4_map_blocks(handle_t *handle, struct inode *inode,
1226                     struct ext4_map_blocks *map, int flags)
1227 {
1228         int retval;
1229
1230         map->m_flags = 0;
1231         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
1232                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
1233                   (unsigned long) map->m_lblk);
1234         /*
1235          * Try to see if we can get the block without requesting a new
1236          * file system block.
1237          */
1238         down_read((&EXT4_I(inode)->i_data_sem));
1239         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
1240                 retval = ext4_ext_map_blocks(handle, inode, map, 0);
1241         } else {
1242                 retval = ext4_ind_map_blocks(handle, inode, map, 0);
1243         }
1244         up_read((&EXT4_I(inode)->i_data_sem));
1245
1246         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
1247                 int ret = check_block_validity(inode, __func__, map);
1248                 if (ret != 0)
1249                         return ret;
1250         }
1251
1252         /* If it is only a block(s) look up */
1253         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1254                 return retval;
1255
1256         /*
1257          * Returns if the blocks have already allocated
1258          *
1259          * Note that if blocks have been preallocated
1260          * ext4_ext_get_block() returns th create = 0
1261          * with buffer head unmapped.
1262          */
1263         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
1264                 return retval;
1265
1266         /*
1267          * When we call get_blocks without the create flag, the
1268          * BH_Unwritten flag could have gotten set if the blocks
1269          * requested were part of a uninitialized extent.  We need to
1270          * clear this flag now that we are committed to convert all or
1271          * part of the uninitialized extent to be an initialized
1272          * extent.  This is because we need to avoid the combination
1273          * of BH_Unwritten and BH_Mapped flags being simultaneously
1274          * set on the buffer_head.
1275          */
1276         map->m_flags &= ~EXT4_MAP_UNWRITTEN;
1277
1278         /*
1279          * New blocks allocate and/or writing to uninitialized extent
1280          * will possibly result in updating i_data, so we take
1281          * the write lock of i_data_sem, and call get_blocks()
1282          * with create == 1 flag.
1283          */
1284         down_write((&EXT4_I(inode)->i_data_sem));
1285
1286         /*
1287          * if the caller is from delayed allocation writeout path
1288          * we have already reserved fs blocks for allocation
1289          * let the underlying get_block() function know to
1290          * avoid double accounting
1291          */
1292         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1293                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1294         /*
1295          * We need to check for EXT4 here because migrate
1296          * could have changed the inode type in between
1297          */
1298         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
1299                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
1300         } else {
1301                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
1302
1303                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
1304                         /*
1305                          * We allocated new blocks which will result in
1306                          * i_data's format changing.  Force the migrate
1307                          * to fail by clearing migrate flags
1308                          */
1309                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
1310                 }
1311
1312                 /*
1313                  * Update reserved blocks/metadata blocks after successful
1314                  * block allocation which had been deferred till now. We don't
1315                  * support fallocate for non extent files. So we can update
1316                  * reserve space here.
1317                  */
1318                 if ((retval > 0) &&
1319                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
1320                         ext4_da_update_reserve_space(inode, retval, 1);
1321         }
1322         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1323                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1324
1325         up_write((&EXT4_I(inode)->i_data_sem));
1326         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
1327                 int ret = check_block_validity(inode,
1328                                                "ext4_map_blocks_after_alloc",
1329                                                map);
1330                 if (ret != 0)
1331                         return ret;
1332         }
1333         return retval;
1334 }
1335
1336 /* Maximum number of blocks we map for direct IO at once. */
1337 #define DIO_MAX_BLOCKS 4096
1338
1339 static int _ext4_get_block(struct inode *inode, sector_t iblock,
1340                            struct buffer_head *bh, int flags)
1341 {
1342         handle_t *handle = ext4_journal_current_handle();
1343         struct ext4_map_blocks map;
1344         int ret = 0, started = 0;
1345         int dio_credits;
1346
1347         map.m_lblk = iblock;
1348         map.m_len = bh->b_size >> inode->i_blkbits;
1349
1350         if (flags && !handle) {
1351                 /* Direct IO write... */
1352                 if (map.m_len > DIO_MAX_BLOCKS)
1353                         map.m_len = DIO_MAX_BLOCKS;
1354                 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
1355                 handle = ext4_journal_start(inode, dio_credits);
1356                 if (IS_ERR(handle)) {
1357                         ret = PTR_ERR(handle);
1358                         return ret;
1359                 }
1360                 started = 1;
1361         }
1362
1363         ret = ext4_map_blocks(handle, inode, &map, flags);
1364         if (ret > 0) {
1365                 map_bh(bh, inode->i_sb, map.m_pblk);
1366                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
1367                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
1368                 ret = 0;
1369         }
1370         if (started)
1371                 ext4_journal_stop(handle);
1372         return ret;
1373 }
1374
1375 int ext4_get_block(struct inode *inode, sector_t iblock,
1376                    struct buffer_head *bh, int create)
1377 {
1378         return _ext4_get_block(inode, iblock, bh,
1379                                create ? EXT4_GET_BLOCKS_CREATE : 0);
1380 }
1381
1382 /*
1383  * `handle' can be NULL if create is zero
1384  */
1385 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1386                                 ext4_lblk_t block, int create, int *errp)
1387 {
1388         struct ext4_map_blocks map;
1389         struct buffer_head *bh;
1390         int fatal = 0, err;
1391
1392         J_ASSERT(handle != NULL || create == 0);
1393
1394         map.m_lblk = block;
1395         map.m_len = 1;
1396         err = ext4_map_blocks(handle, inode, &map,
1397                               create ? EXT4_GET_BLOCKS_CREATE : 0);
1398
1399         if (err < 0)
1400                 *errp = err;
1401         if (err <= 0)
1402                 return NULL;
1403         *errp = 0;
1404
1405         bh = sb_getblk(inode->i_sb, map.m_pblk);
1406         if (!bh) {
1407                 *errp = -EIO;
1408                 return NULL;
1409         }
1410         if (map.m_flags & EXT4_MAP_NEW) {
1411                 J_ASSERT(create != 0);
1412                 J_ASSERT(handle != NULL);
1413
1414                 /*
1415                  * Now that we do not always journal data, we should
1416                  * keep in mind whether this should always journal the
1417                  * new buffer as metadata.  For now, regular file
1418                  * writes use ext4_get_block instead, so it's not a
1419                  * problem.
1420                  */
1421                 lock_buffer(bh);
1422                 BUFFER_TRACE(bh, "call get_create_access");
1423                 fatal = ext4_journal_get_create_access(handle, bh);
1424                 if (!fatal && !buffer_uptodate(bh)) {
1425                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1426                         set_buffer_uptodate(bh);
1427                 }
1428                 unlock_buffer(bh);
1429                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1430                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1431                 if (!fatal)
1432                         fatal = err;
1433         } else {
1434                 BUFFER_TRACE(bh, "not a new buffer");
1435         }
1436         if (fatal) {
1437                 *errp = fatal;
1438                 brelse(bh);
1439                 bh = NULL;
1440         }
1441         return bh;
1442 }
1443
1444 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1445                                ext4_lblk_t block, int create, int *err)
1446 {
1447         struct buffer_head *bh;
1448
1449         bh = ext4_getblk(handle, inode, block, create, err);
1450         if (!bh)
1451                 return bh;
1452         if (buffer_uptodate(bh))
1453                 return bh;
1454         ll_rw_block(READ_META, 1, &bh);
1455         wait_on_buffer(bh);
1456         if (buffer_uptodate(bh))
1457                 return bh;
1458         put_bh(bh);
1459         *err = -EIO;
1460         return NULL;
1461 }
1462
1463 static int walk_page_buffers(handle_t *handle,
1464                              struct buffer_head *head,
1465                              unsigned from,
1466                              unsigned to,
1467                              int *partial,
1468                              int (*fn)(handle_t *handle,
1469                                        struct buffer_head *bh))
1470 {
1471         struct buffer_head *bh;
1472         unsigned block_start, block_end;
1473         unsigned blocksize = head->b_size;
1474         int err, ret = 0;
1475         struct buffer_head *next;
1476
1477         for (bh = head, block_start = 0;
1478              ret == 0 && (bh != head || !block_start);
1479              block_start = block_end, bh = next) {
1480                 next = bh->b_this_page;
1481                 block_end = block_start + blocksize;
1482                 if (block_end <= from || block_start >= to) {
1483                         if (partial && !buffer_uptodate(bh))
1484                                 *partial = 1;
1485                         continue;
1486                 }
1487                 err = (*fn)(handle, bh);
1488                 if (!ret)
1489                         ret = err;
1490         }
1491         return ret;
1492 }
1493
1494 /*
1495  * To preserve ordering, it is essential that the hole instantiation and
1496  * the data write be encapsulated in a single transaction.  We cannot
1497  * close off a transaction and start a new one between the ext4_get_block()
1498  * and the commit_write().  So doing the jbd2_journal_start at the start of
1499  * prepare_write() is the right place.
1500  *
1501  * Also, this function can nest inside ext4_writepage() ->
1502  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1503  * has generated enough buffer credits to do the whole page.  So we won't
1504  * block on the journal in that case, which is good, because the caller may
1505  * be PF_MEMALLOC.
1506  *
1507  * By accident, ext4 can be reentered when a transaction is open via
1508  * quota file writes.  If we were to commit the transaction while thus
1509  * reentered, there can be a deadlock - we would be holding a quota
1510  * lock, and the commit would never complete if another thread had a
1511  * transaction open and was blocking on the quota lock - a ranking
1512  * violation.
1513  *
1514  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1515  * will _not_ run commit under these circumstances because handle->h_ref
1516  * is elevated.  We'll still have enough credits for the tiny quotafile
1517  * write.
1518  */
1519 static int do_journal_get_write_access(handle_t *handle,
1520                                        struct buffer_head *bh)
1521 {
1522         if (!buffer_mapped(bh) || buffer_freed(bh))
1523                 return 0;
1524         return ext4_journal_get_write_access(handle, bh);
1525 }
1526
1527 /*
1528  * Truncate blocks that were not used by write. We have to truncate the
1529  * pagecache as well so that corresponding buffers get properly unmapped.
1530  */
1531 static void ext4_truncate_failed_write(struct inode *inode)
1532 {
1533         truncate_inode_pages(inode->i_mapping, inode->i_size);
1534         ext4_truncate(inode);
1535 }
1536
1537 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
1538                    struct buffer_head *bh_result, int create);
1539 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1540                             loff_t pos, unsigned len, unsigned flags,
1541                             struct page **pagep, void **fsdata)
1542 {
1543         struct inode *inode = mapping->host;
1544         int ret, needed_blocks;
1545         handle_t *handle;
1546         int retries = 0;
1547         struct page *page;
1548         pgoff_t index;
1549         unsigned from, to;
1550
1551         trace_ext4_write_begin(inode, pos, len, flags);
1552         /*
1553          * Reserve one block more for addition to orphan list in case
1554          * we allocate blocks but write fails for some reason
1555          */
1556         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1557         index = pos >> PAGE_CACHE_SHIFT;
1558         from = pos & (PAGE_CACHE_SIZE - 1);
1559         to = from + len;
1560
1561 retry:
1562         handle = ext4_journal_start(inode, needed_blocks);
1563         if (IS_ERR(handle)) {
1564                 ret = PTR_ERR(handle);
1565                 goto out;
1566         }
1567
1568         /* We cannot recurse into the filesystem as the transaction is already
1569          * started */
1570         flags |= AOP_FLAG_NOFS;
1571
1572         page = grab_cache_page_write_begin(mapping, index, flags);
1573         if (!page) {
1574                 ext4_journal_stop(handle);
1575                 ret = -ENOMEM;
1576                 goto out;
1577         }
1578         *pagep = page;
1579
1580         if (ext4_should_dioread_nolock(inode))
1581                 ret = block_write_begin(file, mapping, pos, len, flags, pagep,
1582                                 fsdata, ext4_get_block_write);
1583         else
1584                 ret = block_write_begin(file, mapping, pos, len, flags, pagep,
1585                                 fsdata, ext4_get_block);
1586
1587         if (!ret && ext4_should_journal_data(inode)) {
1588                 ret = walk_page_buffers(handle, page_buffers(page),
1589                                 from, to, NULL, do_journal_get_write_access);
1590         }
1591
1592         if (ret) {
1593                 unlock_page(page);
1594                 page_cache_release(page);
1595                 /*
1596                  * block_write_begin may have instantiated a few blocks
1597                  * outside i_size.  Trim these off again. Don't need
1598                  * i_size_read because we hold i_mutex.
1599                  *
1600                  * Add inode to orphan list in case we crash before
1601                  * truncate finishes
1602                  */
1603                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1604                         ext4_orphan_add(handle, inode);
1605
1606                 ext4_journal_stop(handle);
1607                 if (pos + len > inode->i_size) {
1608                         ext4_truncate_failed_write(inode);
1609                         /*
1610                          * If truncate failed early the inode might
1611                          * still be on the orphan list; we need to
1612                          * make sure the inode is removed from the
1613                          * orphan list in that case.
1614                          */
1615                         if (inode->i_nlink)
1616                                 ext4_orphan_del(NULL, inode);
1617                 }
1618         }
1619
1620         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1621                 goto retry;
1622 out:
1623         return ret;
1624 }
1625
1626 /* For write_end() in data=journal mode */
1627 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1628 {
1629         if (!buffer_mapped(bh) || buffer_freed(bh))
1630                 return 0;
1631         set_buffer_uptodate(bh);
1632         return ext4_handle_dirty_metadata(handle, NULL, bh);
1633 }
1634
1635 static int ext4_generic_write_end(struct file *file,
1636                                   struct address_space *mapping,
1637                                   loff_t pos, unsigned len, unsigned copied,
1638                                   struct page *page, void *fsdata)
1639 {
1640         int i_size_changed = 0;
1641         struct inode *inode = mapping->host;
1642         handle_t *handle = ext4_journal_current_handle();
1643
1644         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1645
1646         /*
1647          * No need to use i_size_read() here, the i_size
1648          * cannot change under us because we hold i_mutex.
1649          *
1650          * But it's important to update i_size while still holding page lock:
1651          * page writeout could otherwise come in and zero beyond i_size.
1652          */
1653         if (pos + copied > inode->i_size) {
1654                 i_size_write(inode, pos + copied);
1655                 i_size_changed = 1;
1656         }
1657
1658         if (pos + copied >  EXT4_I(inode)->i_disksize) {
1659                 /* We need to mark inode dirty even if
1660                  * new_i_size is less that inode->i_size
1661                  * bu greater than i_disksize.(hint delalloc)
1662                  */
1663                 ext4_update_i_disksize(inode, (pos + copied));
1664                 i_size_changed = 1;
1665         }
1666         unlock_page(page);
1667         page_cache_release(page);
1668
1669         /*
1670          * Don't mark the inode dirty under page lock. First, it unnecessarily
1671          * makes the holding time of page lock longer. Second, it forces lock
1672          * ordering of page lock and transaction start for journaling
1673          * filesystems.
1674          */
1675         if (i_size_changed)
1676                 ext4_mark_inode_dirty(handle, inode);
1677
1678         return copied;
1679 }
1680
1681 /*
1682  * We need to pick up the new inode size which generic_commit_write gave us
1683  * `file' can be NULL - eg, when called from page_symlink().
1684  *
1685  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1686  * buffers are managed internally.
1687  */
1688 static int ext4_ordered_write_end(struct file *file,
1689                                   struct address_space *mapping,
1690                                   loff_t pos, unsigned len, unsigned copied,
1691                                   struct page *page, void *fsdata)
1692 {
1693         handle_t *handle = ext4_journal_current_handle();
1694         struct inode *inode = mapping->host;
1695         int ret = 0, ret2;
1696
1697         trace_ext4_ordered_write_end(inode, pos, len, copied);
1698         ret = ext4_jbd2_file_inode(handle, inode);
1699
1700         if (ret == 0) {
1701                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1702                                                         page, fsdata);
1703                 copied = ret2;
1704                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1705                         /* if we have allocated more blocks and copied
1706                          * less. We will have blocks allocated outside
1707                          * inode->i_size. So truncate them
1708                          */
1709                         ext4_orphan_add(handle, inode);
1710                 if (ret2 < 0)
1711                         ret = ret2;
1712         }
1713         ret2 = ext4_journal_stop(handle);
1714         if (!ret)
1715                 ret = ret2;
1716
1717         if (pos + len > inode->i_size) {
1718                 ext4_truncate_failed_write(inode);
1719                 /*
1720                  * If truncate failed early the inode might still be
1721                  * on the orphan list; we need to make sure the inode
1722                  * is removed from the orphan list in that case.
1723                  */
1724                 if (inode->i_nlink)
1725                         ext4_orphan_del(NULL, inode);
1726         }
1727
1728
1729         return ret ? ret : copied;
1730 }
1731
1732 static int ext4_writeback_write_end(struct file *file,
1733                                     struct address_space *mapping,
1734                                     loff_t pos, unsigned len, unsigned copied,
1735                                     struct page *page, void *fsdata)
1736 {
1737         handle_t *handle = ext4_journal_current_handle();
1738         struct inode *inode = mapping->host;
1739         int ret = 0, ret2;
1740
1741         trace_ext4_writeback_write_end(inode, pos, len, copied);
1742         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1743                                                         page, fsdata);
1744         copied = ret2;
1745         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1746                 /* if we have allocated more blocks and copied
1747                  * less. We will have blocks allocated outside
1748                  * inode->i_size. So truncate them
1749                  */
1750                 ext4_orphan_add(handle, inode);
1751
1752         if (ret2 < 0)
1753                 ret = ret2;
1754
1755         ret2 = ext4_journal_stop(handle);
1756         if (!ret)
1757                 ret = ret2;
1758
1759         if (pos + len > inode->i_size) {
1760                 ext4_truncate_failed_write(inode);
1761                 /*
1762                  * If truncate failed early the inode might still be
1763                  * on the orphan list; we need to make sure the inode
1764                  * is removed from the orphan list in that case.
1765                  */
1766                 if (inode->i_nlink)
1767                         ext4_orphan_del(NULL, inode);
1768         }
1769
1770         return ret ? ret : copied;
1771 }
1772
1773 static int ext4_journalled_write_end(struct file *file,
1774                                      struct address_space *mapping,
1775                                      loff_t pos, unsigned len, unsigned copied,
1776                                      struct page *page, void *fsdata)
1777 {
1778         handle_t *handle = ext4_journal_current_handle();
1779         struct inode *inode = mapping->host;
1780         int ret = 0, ret2;
1781         int partial = 0;
1782         unsigned from, to;
1783         loff_t new_i_size;
1784
1785         trace_ext4_journalled_write_end(inode, pos, len, copied);
1786         from = pos & (PAGE_CACHE_SIZE - 1);
1787         to = from + len;
1788
1789         if (copied < len) {
1790                 if (!PageUptodate(page))
1791                         copied = 0;
1792                 page_zero_new_buffers(page, from+copied, to);
1793         }
1794
1795         ret = walk_page_buffers(handle, page_buffers(page), from,
1796                                 to, &partial, write_end_fn);
1797         if (!partial)
1798                 SetPageUptodate(page);
1799         new_i_size = pos + copied;
1800         if (new_i_size > inode->i_size)
1801                 i_size_write(inode, pos+copied);
1802         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1803         if (new_i_size > EXT4_I(inode)->i_disksize) {
1804                 ext4_update_i_disksize(inode, new_i_size);
1805                 ret2 = ext4_mark_inode_dirty(handle, inode);
1806                 if (!ret)
1807                         ret = ret2;
1808         }
1809
1810         unlock_page(page);
1811         page_cache_release(page);
1812         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1813                 /* if we have allocated more blocks and copied
1814                  * less. We will have blocks allocated outside
1815                  * inode->i_size. So truncate them
1816                  */
1817                 ext4_orphan_add(handle, inode);
1818
1819         ret2 = ext4_journal_stop(handle);
1820         if (!ret)
1821                 ret = ret2;
1822         if (pos + len > inode->i_size) {
1823                 ext4_truncate_failed_write(inode);
1824                 /*
1825                  * If truncate failed early the inode might still be
1826                  * on the orphan list; we need to make sure the inode
1827                  * is removed from the orphan list in that case.
1828                  */
1829                 if (inode->i_nlink)
1830                         ext4_orphan_del(NULL, inode);
1831         }
1832
1833         return ret ? ret : copied;
1834 }
1835
1836 /*
1837  * Reserve a single block located at lblock
1838  */
1839 static int ext4_da_reserve_space(struct inode *inode, sector_t lblock)
1840 {
1841         int retries = 0;
1842         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1843         struct ext4_inode_info *ei = EXT4_I(inode);
1844         unsigned long md_needed;
1845         int ret;
1846
1847         /*
1848          * recalculate the amount of metadata blocks to reserve
1849          * in order to allocate nrblocks
1850          * worse case is one extent per block
1851          */
1852 repeat:
1853         spin_lock(&ei->i_block_reservation_lock);
1854         md_needed = ext4_calc_metadata_amount(inode, lblock);
1855         trace_ext4_da_reserve_space(inode, md_needed);
1856         spin_unlock(&ei->i_block_reservation_lock);
1857
1858         /*
1859          * We will charge metadata quota at writeout time; this saves
1860          * us from metadata over-estimation, though we may go over by
1861          * a small amount in the end.  Here we just reserve for data.
1862          */
1863         ret = dquot_reserve_block(inode, 1);
1864         if (ret)
1865                 return ret;
1866         /*
1867          * We do still charge estimated metadata to the sb though;
1868          * we cannot afford to run out of free blocks.
1869          */
1870         if (ext4_claim_free_blocks(sbi, md_needed + 1)) {
1871                 dquot_release_reservation_block(inode, 1);
1872                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1873                         yield();
1874                         goto repeat;
1875                 }
1876                 return -ENOSPC;
1877         }
1878         spin_lock(&ei->i_block_reservation_lock);
1879         ei->i_reserved_data_blocks++;
1880         ei->i_reserved_meta_blocks += md_needed;
1881         spin_unlock(&ei->i_block_reservation_lock);
1882
1883         return 0;       /* success */
1884 }
1885
1886 static void ext4_da_release_space(struct inode *inode, int to_free)
1887 {
1888         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1889         struct ext4_inode_info *ei = EXT4_I(inode);
1890
1891         if (!to_free)
1892                 return;         /* Nothing to release, exit */
1893
1894         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1895
1896         trace_ext4_da_release_space(inode, to_free);
1897         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1898                 /*
1899                  * if there aren't enough reserved blocks, then the
1900                  * counter is messed up somewhere.  Since this
1901                  * function is called from invalidate page, it's
1902                  * harmless to return without any action.
1903                  */
1904                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1905                          "ino %lu, to_free %d with only %d reserved "
1906                          "data blocks\n", inode->i_ino, to_free,
1907                          ei->i_reserved_data_blocks);
1908                 WARN_ON(1);
1909                 to_free = ei->i_reserved_data_blocks;
1910         }
1911         ei->i_reserved_data_blocks -= to_free;
1912
1913         if (ei->i_reserved_data_blocks == 0) {
1914                 /*
1915                  * We can release all of the reserved metadata blocks
1916                  * only when we have written all of the delayed
1917                  * allocation blocks.
1918                  */
1919                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1920                                    ei->i_reserved_meta_blocks);
1921                 ei->i_reserved_meta_blocks = 0;
1922                 ei->i_da_metadata_calc_len = 0;
1923         }
1924
1925         /* update fs dirty data blocks counter */
1926         percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1927
1928         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1929
1930         dquot_release_reservation_block(inode, to_free);
1931 }
1932
1933 static void ext4_da_page_release_reservation(struct page *page,
1934                                              unsigned long offset)
1935 {
1936         int to_release = 0;
1937         struct buffer_head *head, *bh;
1938         unsigned int curr_off = 0;
1939
1940         head = page_buffers(page);
1941         bh = head;
1942         do {
1943                 unsigned int next_off = curr_off + bh->b_size;
1944
1945                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1946                         to_release++;
1947                         clear_buffer_delay(bh);
1948                 }
1949                 curr_off = next_off;
1950         } while ((bh = bh->b_this_page) != head);
1951         ext4_da_release_space(page->mapping->host, to_release);
1952 }
1953
1954 /*
1955  * Delayed allocation stuff
1956  */
1957
1958 /*
1959  * mpage_da_submit_io - walks through extent of pages and try to write
1960  * them with writepage() call back
1961  *
1962  * @mpd->inode: inode
1963  * @mpd->first_page: first page of the extent
1964  * @mpd->next_page: page after the last page of the extent
1965  *
1966  * By the time mpage_da_submit_io() is called we expect all blocks
1967  * to be allocated. this may be wrong if allocation failed.
1968  *
1969  * As pages are already locked by write_cache_pages(), we can't use it
1970  */
1971 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1972 {
1973         long pages_skipped;
1974         struct pagevec pvec;
1975         unsigned long index, end;
1976         int ret = 0, err, nr_pages, i;
1977         struct inode *inode = mpd->inode;
1978         struct address_space *mapping = inode->i_mapping;
1979
1980         BUG_ON(mpd->next_page <= mpd->first_page);
1981         /*
1982          * We need to start from the first_page to the next_page - 1
1983          * to make sure we also write the mapped dirty buffer_heads.
1984          * If we look at mpd->b_blocknr we would only be looking
1985          * at the currently mapped buffer_heads.
1986          */
1987         index = mpd->first_page;
1988         end = mpd->next_page - 1;
1989
1990         pagevec_init(&pvec, 0);
1991         while (index <= end) {
1992                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1993                 if (nr_pages == 0)
1994                         break;
1995                 for (i = 0; i < nr_pages; i++) {
1996                         struct page *page = pvec.pages[i];
1997
1998                         index = page->index;
1999                         if (index > end)
2000                                 break;
2001                         index++;
2002
2003                         BUG_ON(!PageLocked(page));
2004                         BUG_ON(PageWriteback(page));
2005
2006                         pages_skipped = mpd->wbc->pages_skipped;
2007                         err = mapping->a_ops->writepage(page, mpd->wbc);
2008                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
2009                                 /*
2010                                  * have successfully written the page
2011                                  * without skipping the same
2012                                  */
2013                                 mpd->pages_written++;
2014                         /*
2015                          * In error case, we have to continue because
2016                          * remaining pages are still locked
2017                          * XXX: unlock and re-dirty them?
2018                          */
2019                         if (ret == 0)
2020                                 ret = err;
2021                 }
2022                 pagevec_release(&pvec);
2023         }
2024         return ret;
2025 }
2026
2027 /*
2028  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
2029  *
2030  * the function goes through all passed space and put actual disk
2031  * block numbers into buffer heads, dropping BH_Delay and BH_Unwritten
2032  */
2033 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd,
2034                                  struct ext4_map_blocks *map)
2035 {
2036         struct inode *inode = mpd->inode;
2037         struct address_space *mapping = inode->i_mapping;
2038         int blocks = map->m_len;
2039         sector_t pblock = map->m_pblk, cur_logical;
2040         struct buffer_head *head, *bh;
2041         pgoff_t index, end;
2042         struct pagevec pvec;
2043         int nr_pages, i;
2044
2045         index = map->m_lblk >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2046         end = (map->m_lblk + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2047         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2048
2049         pagevec_init(&pvec, 0);
2050
2051         while (index <= end) {
2052                 /* XXX: optimize tail */
2053                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2054                 if (nr_pages == 0)
2055                         break;
2056                 for (i = 0; i < nr_pages; i++) {
2057                         struct page *page = pvec.pages[i];
2058
2059                         index = page->index;
2060                         if (index > end)
2061                                 break;
2062                         index++;
2063
2064                         BUG_ON(!PageLocked(page));
2065                         BUG_ON(PageWriteback(page));
2066                         BUG_ON(!page_has_buffers(page));
2067
2068                         bh = page_buffers(page);
2069                         head = bh;
2070
2071                         /* skip blocks out of the range */
2072                         do {
2073                                 if (cur_logical >= map->m_lblk)
2074                                         break;
2075                                 cur_logical++;
2076                         } while ((bh = bh->b_this_page) != head);
2077
2078                         do {
2079                                 if (cur_logical >= map->m_lblk + blocks)
2080                                         break;
2081
2082                                 if (buffer_delay(bh) || buffer_unwritten(bh)) {
2083
2084                                         BUG_ON(bh->b_bdev != inode->i_sb->s_bdev);
2085
2086                                         if (buffer_delay(bh)) {
2087                                                 clear_buffer_delay(bh);
2088                                                 bh->b_blocknr = pblock;
2089                                         } else {
2090                                                 /*
2091                                                  * unwritten already should have
2092                                                  * blocknr assigned. Verify that
2093                                                  */
2094                                                 clear_buffer_unwritten(bh);
2095                                                 BUG_ON(bh->b_blocknr != pblock);
2096                                         }
2097
2098                                 } else if (buffer_mapped(bh))
2099                                         BUG_ON(bh->b_blocknr != pblock);
2100
2101                                 if (map->m_flags & EXT4_MAP_UNINIT)
2102                                         set_buffer_uninit(bh);
2103                                 cur_logical++;
2104                                 pblock++;
2105                         } while ((bh = bh->b_this_page) != head);
2106                 }
2107                 pagevec_release(&pvec);
2108         }
2109 }
2110
2111
2112 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2113                                         sector_t logical, long blk_cnt)
2114 {
2115         int nr_pages, i;
2116         pgoff_t index, end;
2117         struct pagevec pvec;
2118         struct inode *inode = mpd->inode;
2119         struct address_space *mapping = inode->i_mapping;
2120
2121         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2122         end   = (logical + blk_cnt - 1) >>
2123                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2124         while (index <= end) {
2125                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2126                 if (nr_pages == 0)
2127                         break;
2128                 for (i = 0; i < nr_pages; i++) {
2129                         struct page *page = pvec.pages[i];
2130                         if (page->index > end)
2131                                 break;
2132                         BUG_ON(!PageLocked(page));
2133                         BUG_ON(PageWriteback(page));
2134                         block_invalidatepage(page, 0);
2135                         ClearPageUptodate(page);
2136                         unlock_page(page);
2137                 }
2138                 index = pvec.pages[nr_pages - 1]->index + 1;
2139                 pagevec_release(&pvec);
2140         }
2141         return;
2142 }
2143
2144 static void ext4_print_free_blocks(struct inode *inode)
2145 {
2146         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2147         printk(KERN_CRIT "Total free blocks count %lld\n",
2148                ext4_count_free_blocks(inode->i_sb));
2149         printk(KERN_CRIT "Free/Dirty block details\n");
2150         printk(KERN_CRIT "free_blocks=%lld\n",
2151                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2152         printk(KERN_CRIT "dirty_blocks=%lld\n",
2153                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2154         printk(KERN_CRIT "Block reservation details\n");
2155         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2156                EXT4_I(inode)->i_reserved_data_blocks);
2157         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2158                EXT4_I(inode)->i_reserved_meta_blocks);
2159         return;
2160 }
2161
2162 /*
2163  * mpage_da_map_blocks - go through given space
2164  *
2165  * @mpd - bh describing space
2166  *
2167  * The function skips space we know is already mapped to disk blocks.
2168  *
2169  */
2170 static int mpage_da_map_blocks(struct mpage_da_data *mpd)
2171 {
2172         int err, blks, get_blocks_flags;
2173         struct ext4_map_blocks map;
2174         sector_t next = mpd->b_blocknr;
2175         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2176         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2177         handle_t *handle = NULL;
2178
2179         /*
2180          * We consider only non-mapped and non-allocated blocks
2181          */
2182         if ((mpd->b_state  & (1 << BH_Mapped)) &&
2183                 !(mpd->b_state & (1 << BH_Delay)) &&
2184                 !(mpd->b_state & (1 << BH_Unwritten)))
2185                 return 0;
2186
2187         /*
2188          * If we didn't accumulate anything to write simply return
2189          */
2190         if (!mpd->b_size)
2191                 return 0;
2192
2193         handle = ext4_journal_current_handle();
2194         BUG_ON(!handle);
2195
2196         /*
2197          * Call ext4_get_blocks() to allocate any delayed allocation
2198          * blocks, or to convert an uninitialized extent to be
2199          * initialized (in the case where we have written into
2200          * one or more preallocated blocks).
2201          *
2202          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2203          * indicate that we are on the delayed allocation path.  This
2204          * affects functions in many different parts of the allocation
2205          * call path.  This flag exists primarily because we don't
2206          * want to change *many* call functions, so ext4_get_blocks()
2207          * will set the magic i_delalloc_reserved_flag once the
2208          * inode's allocation semaphore is taken.
2209          *
2210          * If the blocks in questions were delalloc blocks, set
2211          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2212          * variables are updated after the blocks have been allocated.
2213          */
2214         map.m_lblk = next;
2215         map.m_len = max_blocks;
2216         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
2217         if (ext4_should_dioread_nolock(mpd->inode))
2218                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2219         if (mpd->b_state & (1 << BH_Delay))
2220                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2221
2222         blks = ext4_map_blocks(handle, mpd->inode, &map, get_blocks_flags);
2223         if (blks < 0) {
2224                 err = blks;
2225                 /*
2226                  * If get block returns with error we simply
2227                  * return. Later writepage will redirty the page and
2228                  * writepages will find the dirty page again
2229                  */
2230                 if (err == -EAGAIN)
2231                         return 0;
2232
2233                 if (err == -ENOSPC &&
2234                     ext4_count_free_blocks(mpd->inode->i_sb)) {
2235                         mpd->retval = err;
2236                         return 0;
2237                 }
2238
2239                 /*
2240                  * get block failure will cause us to loop in
2241                  * writepages, because a_ops->writepage won't be able
2242                  * to make progress. The page will be redirtied by
2243                  * writepage and writepages will again try to write
2244                  * the same.
2245                  */
2246                 ext4_msg(mpd->inode->i_sb, KERN_CRIT,
2247                          "delayed block allocation failed for inode %lu at "
2248                          "logical offset %llu with max blocks %zd with "
2249                          "error %d", mpd->inode->i_ino,
2250                          (unsigned long long) next,
2251                          mpd->b_size >> mpd->inode->i_blkbits, err);
2252                 printk(KERN_CRIT "This should not happen!!  "
2253                        "Data will be lost\n");
2254                 if (err == -ENOSPC) {
2255                         ext4_print_free_blocks(mpd->inode);
2256                 }
2257                 /* invalidate all the pages */
2258                 ext4_da_block_invalidatepages(mpd, next,
2259                                 mpd->b_size >> mpd->inode->i_blkbits);
2260                 return err;
2261         }
2262         BUG_ON(blks == 0);
2263
2264         if (map.m_flags & EXT4_MAP_NEW) {
2265                 struct block_device *bdev = mpd->inode->i_sb->s_bdev;
2266                 int i;
2267
2268                 for (i = 0; i < map.m_len; i++)
2269                         unmap_underlying_metadata(bdev, map.m_pblk + i);
2270         }
2271
2272         /*
2273          * If blocks are delayed marked, we need to
2274          * put actual blocknr and drop delayed bit
2275          */
2276         if ((mpd->b_state & (1 << BH_Delay)) ||
2277             (mpd->b_state & (1 << BH_Unwritten)))
2278                 mpage_put_bnr_to_bhs(mpd, &map);
2279
2280         if (ext4_should_order_data(mpd->inode)) {
2281                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2282                 if (err)
2283                         return err;
2284         }
2285
2286         /*
2287          * Update on-disk size along with block allocation.
2288          */
2289         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2290         if (disksize > i_size_read(mpd->inode))
2291                 disksize = i_size_read(mpd->inode);
2292         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2293                 ext4_update_i_disksize(mpd->inode, disksize);
2294                 return ext4_mark_inode_dirty(handle, mpd->inode);
2295         }
2296
2297         return 0;
2298 }
2299
2300 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2301                 (1 << BH_Delay) | (1 << BH_Unwritten))
2302
2303 /*
2304  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2305  *
2306  * @mpd->lbh - extent of blocks
2307  * @logical - logical number of the block in the file
2308  * @bh - bh of the block (used to access block's state)
2309  *
2310  * the function is used to collect contig. blocks in same state
2311  */
2312 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2313                                    sector_t logical, size_t b_size,
2314                                    unsigned long b_state)
2315 {
2316         sector_t next;
2317         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2318
2319         /*
2320          * XXX Don't go larger than mballoc is willing to allocate
2321          * This is a stopgap solution.  We eventually need to fold
2322          * mpage_da_submit_io() into this function and then call
2323          * ext4_get_blocks() multiple times in a loop
2324          */
2325         if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
2326                 goto flush_it;
2327
2328         /* check if thereserved journal credits might overflow */
2329         if (!(ext4_test_inode_flag(mpd->inode, EXT4_INODE_EXTENTS))) {
2330                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2331                         /*
2332                          * With non-extent format we are limited by the journal
2333                          * credit available.  Total credit needed to insert
2334                          * nrblocks contiguous blocks is dependent on the
2335                          * nrblocks.  So limit nrblocks.
2336                          */
2337                         goto flush_it;
2338                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2339                                 EXT4_MAX_TRANS_DATA) {
2340                         /*
2341                          * Adding the new buffer_head would make it cross the
2342                          * allowed limit for which we have journal credit
2343                          * reserved. So limit the new bh->b_size
2344                          */
2345                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2346                                                 mpd->inode->i_blkbits;
2347                         /* we will do mpage_da_submit_io in the next loop */
2348                 }
2349         }
2350         /*
2351          * First block in the extent
2352          */
2353         if (mpd->b_size == 0) {
2354                 mpd->b_blocknr = logical;
2355                 mpd->b_size = b_size;
2356                 mpd->b_state = b_state & BH_FLAGS;
2357                 return;
2358         }
2359
2360         next = mpd->b_blocknr + nrblocks;
2361         /*
2362          * Can we merge the block to our big extent?
2363          */
2364         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2365                 mpd->b_size += b_size;
2366                 return;
2367         }
2368
2369 flush_it:
2370         /*
2371          * We couldn't merge the block to our extent, so we
2372          * need to flush current  extent and start new one
2373          */
2374         if (mpage_da_map_blocks(mpd) == 0)
2375                 mpage_da_submit_io(mpd);
2376         mpd->io_done = 1;
2377         return;
2378 }
2379
2380 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
2381 {
2382         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
2383 }
2384
2385 /*
2386  * __mpage_da_writepage - finds extent of pages and blocks
2387  *
2388  * @page: page to consider
2389  * @wbc: not used, we just follow rules
2390  * @data: context
2391  *
2392  * The function finds extents of pages and scan them for all blocks.
2393  */
2394 static int __mpage_da_writepage(struct page *page,
2395                                 struct writeback_control *wbc, void *data)
2396 {
2397         struct mpage_da_data *mpd = data;
2398         struct inode *inode = mpd->inode;
2399         struct buffer_head *bh, *head;
2400         sector_t logical;
2401
2402         /*
2403          * Can we merge this page to current extent?
2404          */
2405         if (mpd->next_page != page->index) {
2406                 /*
2407                  * Nope, we can't. So, we map non-allocated blocks
2408                  * and start IO on them using writepage()
2409                  */
2410                 if (mpd->next_page != mpd->first_page) {
2411                         if (mpage_da_map_blocks(mpd) == 0)
2412                                 mpage_da_submit_io(mpd);
2413                         /*
2414                          * skip rest of the page in the page_vec
2415                          */
2416                         mpd->io_done = 1;
2417                         redirty_page_for_writepage(wbc, page);
2418                         unlock_page(page);
2419                         return MPAGE_DA_EXTENT_TAIL;
2420                 }
2421
2422                 /*
2423                  * Start next extent of pages ...
2424                  */
2425                 mpd->first_page = page->index;
2426
2427                 /*
2428                  * ... and blocks
2429                  */
2430                 mpd->b_size = 0;
2431                 mpd->b_state = 0;
2432                 mpd->b_blocknr = 0;
2433         }
2434
2435         mpd->next_page = page->index + 1;
2436         logical = (sector_t) page->index <<
2437                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2438
2439         if (!page_has_buffers(page)) {
2440                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2441                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2442                 if (mpd->io_done)
2443                         return MPAGE_DA_EXTENT_TAIL;
2444         } else {
2445                 /*
2446                  * Page with regular buffer heads, just add all dirty ones
2447                  */
2448                 head = page_buffers(page);
2449                 bh = head;
2450                 do {
2451                         BUG_ON(buffer_locked(bh));
2452                         /*
2453                          * We need to try to allocate
2454                          * unmapped blocks in the same page.
2455                          * Otherwise we won't make progress
2456                          * with the page in ext4_writepage
2457                          */
2458                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2459                                 mpage_add_bh_to_extent(mpd, logical,
2460                                                        bh->b_size,
2461                                                        bh->b_state);
2462                                 if (mpd->io_done)
2463                                         return MPAGE_DA_EXTENT_TAIL;
2464                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2465                                 /*
2466                                  * mapped dirty buffer. We need to update
2467                                  * the b_state because we look at
2468                                  * b_state in mpage_da_map_blocks. We don't
2469                                  * update b_size because if we find an
2470                                  * unmapped buffer_head later we need to
2471                                  * use the b_state flag of that buffer_head.
2472                                  */
2473                                 if (mpd->b_size == 0)
2474                                         mpd->b_state = bh->b_state & BH_FLAGS;
2475                         }
2476                         logical++;
2477                 } while ((bh = bh->b_this_page) != head);
2478         }
2479
2480         return 0;
2481 }
2482
2483 /*
2484  * This is a special get_blocks_t callback which is used by
2485  * ext4_da_write_begin().  It will either return mapped block or
2486  * reserve space for a single block.
2487  *
2488  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2489  * We also have b_blocknr = -1 and b_bdev initialized properly
2490  *
2491  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2492  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2493  * initialized properly.
2494  */
2495 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2496                                   struct buffer_head *bh, int create)
2497 {
2498         struct ext4_map_blocks map;
2499         int ret = 0;
2500         sector_t invalid_block = ~((sector_t) 0xffff);
2501
2502         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2503                 invalid_block = ~0;
2504
2505         BUG_ON(create == 0);
2506         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
2507
2508         map.m_lblk = iblock;
2509         map.m_len = 1;
2510
2511         /*
2512          * first, we need to know whether the block is allocated already
2513          * preallocated blocks are unmapped but should treated
2514          * the same as allocated blocks.
2515          */
2516         ret = ext4_map_blocks(NULL, inode, &map, 0);
2517         if (ret < 0)
2518                 return ret;
2519         if (ret == 0) {
2520                 if (buffer_delay(bh))
2521                         return 0; /* Not sure this could or should happen */
2522                 /*
2523                  * XXX: __block_prepare_write() unmaps passed block,
2524                  * is it OK?
2525                  */
2526                 ret = ext4_da_reserve_space(inode, iblock);
2527                 if (ret)
2528                         /* not enough space to reserve */
2529                         return ret;
2530
2531                 map_bh(bh, inode->i_sb, invalid_block);
2532                 set_buffer_new(bh);
2533                 set_buffer_delay(bh);
2534                 return 0;
2535         }
2536
2537         map_bh(bh, inode->i_sb, map.m_pblk);
2538         bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
2539
2540         if (buffer_unwritten(bh)) {
2541                 /* A delayed write to unwritten bh should be marked
2542                  * new and mapped.  Mapped ensures that we don't do
2543                  * get_block multiple times when we write to the same
2544                  * offset and new ensures that we do proper zero out
2545                  * for partial write.
2546                  */
2547                 set_buffer_new(bh);
2548                 set_buffer_mapped(bh);
2549         }
2550         return 0;
2551 }
2552
2553 /*
2554  * This function is used as a standard get_block_t calback function
2555  * when there is no desire to allocate any blocks.  It is used as a
2556  * callback function for block_prepare_write() and block_write_full_page().
2557  * These functions should only try to map a single block at a time.
2558  *
2559  * Since this function doesn't do block allocations even if the caller
2560  * requests it by passing in create=1, it is critically important that
2561  * any caller checks to make sure that any buffer heads are returned
2562  * by this function are either all already mapped or marked for
2563  * delayed allocation before calling  block_write_full_page().  Otherwise,
2564  * b_blocknr could be left unitialized, and the page write functions will
2565  * be taken by surprise.
2566  */
2567 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2568                                    struct buffer_head *bh_result, int create)
2569 {
2570         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2571         return _ext4_get_block(inode, iblock, bh_result, 0);
2572 }
2573
2574 static int bget_one(handle_t *handle, struct buffer_head *bh)
2575 {
2576         get_bh(bh);
2577         return 0;
2578 }
2579
2580 static int bput_one(handle_t *handle, struct buffer_head *bh)
2581 {
2582         put_bh(bh);
2583         return 0;
2584 }
2585
2586 static int __ext4_journalled_writepage(struct page *page,
2587                                        unsigned int len)
2588 {
2589         struct address_space *mapping = page->mapping;
2590         struct inode *inode = mapping->host;
2591         struct buffer_head *page_bufs;
2592         handle_t *handle = NULL;
2593         int ret = 0;
2594         int err;
2595
2596         page_bufs = page_buffers(page);
2597         BUG_ON(!page_bufs);
2598         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2599         /* As soon as we unlock the page, it can go away, but we have
2600          * references to buffers so we are safe */
2601         unlock_page(page);
2602
2603         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2604         if (IS_ERR(handle)) {
2605                 ret = PTR_ERR(handle);
2606                 goto out;
2607         }
2608
2609         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2610                                 do_journal_get_write_access);
2611
2612         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2613                                 write_end_fn);
2614         if (ret == 0)
2615                 ret = err;
2616         err = ext4_journal_stop(handle);
2617         if (!ret)
2618                 ret = err;
2619
2620         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2621         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2622 out:
2623         return ret;
2624 }
2625
2626 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
2627 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
2628
2629 /*
2630  * Note that we don't need to start a transaction unless we're journaling data
2631  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2632  * need to file the inode to the transaction's list in ordered mode because if
2633  * we are writing back data added by write(), the inode is already there and if
2634  * we are writing back data modified via mmap(), noone guarantees in which
2635  * transaction the data will hit the disk. In case we are journaling data, we
2636  * cannot start transaction directly because transaction start ranks above page
2637  * lock so we have to do some magic.
2638  *
2639  * This function can get called via...
2640  *   - ext4_da_writepages after taking page lock (have journal handle)
2641  *   - journal_submit_inode_data_buffers (no journal handle)
2642  *   - shrink_page_list via pdflush (no journal handle)
2643  *   - grab_page_cache when doing write_begin (have journal handle)
2644  *
2645  * We don't do any block allocation in this function. If we have page with
2646  * multiple blocks we need to write those buffer_heads that are mapped. This
2647  * is important for mmaped based write. So if we do with blocksize 1K
2648  * truncate(f, 1024);
2649  * a = mmap(f, 0, 4096);
2650  * a[0] = 'a';
2651  * truncate(f, 4096);
2652  * we have in the page first buffer_head mapped via page_mkwrite call back
2653  * but other bufer_heads would be unmapped but dirty(dirty done via the
2654  * do_wp_page). So writepage should write the first block. If we modify
2655  * the mmap area beyond 1024 we will again get a page_fault and the
2656  * page_mkwrite callback will do the block allocation and mark the
2657  * buffer_heads mapped.
2658  *
2659  * We redirty the page if we have any buffer_heads that is either delay or
2660  * unwritten in the page.
2661  *
2662  * We can get recursively called as show below.
2663  *
2664  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2665  *              ext4_writepage()
2666  *
2667  * But since we don't do any block allocation we should not deadlock.
2668  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2669  */
2670 static int ext4_writepage(struct page *page,
2671                           struct writeback_control *wbc)
2672 {
2673         int ret = 0;
2674         loff_t size;
2675         unsigned int len;
2676         struct buffer_head *page_bufs = NULL;
2677         struct inode *inode = page->mapping->host;
2678
2679         trace_ext4_writepage(inode, page);
2680         size = i_size_read(inode);
2681         if (page->index == size >> PAGE_CACHE_SHIFT)
2682                 len = size & ~PAGE_CACHE_MASK;
2683         else
2684                 len = PAGE_CACHE_SIZE;
2685
2686         if (page_has_buffers(page)) {
2687                 page_bufs = page_buffers(page);
2688                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2689                                         ext4_bh_delay_or_unwritten)) {
2690                         /*
2691                          * We don't want to do  block allocation
2692                          * So redirty the page and return
2693                          * We may reach here when we do a journal commit
2694                          * via journal_submit_inode_data_buffers.
2695                          * If we don't have mapping block we just ignore
2696                          * them. We can also reach here via shrink_page_list
2697                          */
2698                         redirty_page_for_writepage(wbc, page);
2699                         unlock_page(page);
2700                         return 0;
2701                 }
2702         } else {
2703                 /*
2704                  * The test for page_has_buffers() is subtle:
2705                  * We know the page is dirty but it lost buffers. That means
2706                  * that at some moment in time after write_begin()/write_end()
2707                  * has been called all buffers have been clean and thus they
2708                  * must have been written at least once. So they are all
2709                  * mapped and we can happily proceed with mapping them
2710                  * and writing the page.
2711                  *
2712                  * Try to initialize the buffer_heads and check whether
2713                  * all are mapped and non delay. We don't want to
2714                  * do block allocation here.
2715                  */
2716                 ret = block_prepare_write(page, 0, len,
2717                                           noalloc_get_block_write);
2718                 if (!ret) {
2719                         page_bufs = page_buffers(page);
2720                         /* check whether all are mapped and non delay */
2721                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2722                                                 ext4_bh_delay_or_unwritten)) {
2723                                 redirty_page_for_writepage(wbc, page);
2724                                 unlock_page(page);
2725                                 return 0;
2726                         }
2727                 } else {
2728                         /*
2729                          * We can't do block allocation here
2730                          * so just redity the page and unlock
2731                          * and return
2732                          */
2733                         redirty_page_for_writepage(wbc, page);
2734                         unlock_page(page);
2735                         return 0;
2736                 }
2737                 /* now mark the buffer_heads as dirty and uptodate */
2738                 block_commit_write(page, 0, len);
2739         }
2740
2741         if (PageChecked(page) && ext4_should_journal_data(inode)) {
2742                 /*
2743                  * It's mmapped pagecache.  Add buffers and journal it.  There
2744                  * doesn't seem much point in redirtying the page here.
2745                  */
2746                 ClearPageChecked(page);
2747                 return __ext4_journalled_writepage(page, len);
2748         }
2749
2750         if (page_bufs && buffer_uninit(page_bufs)) {
2751                 ext4_set_bh_endio(page_bufs, inode);
2752                 ret = block_write_full_page_endio(page, noalloc_get_block_write,
2753                                             wbc, ext4_end_io_buffer_write);
2754         } else
2755                 ret = block_write_full_page(page, noalloc_get_block_write,
2756                                             wbc);
2757
2758         return ret;
2759 }
2760
2761 /*
2762  * This is called via ext4_da_writepages() to
2763  * calulate the total number of credits to reserve to fit
2764  * a single extent allocation into a single transaction,
2765  * ext4_da_writpeages() will loop calling this before
2766  * the block allocation.
2767  */
2768
2769 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2770 {
2771         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2772
2773         /*
2774          * With non-extent format the journal credit needed to
2775          * insert nrblocks contiguous block is dependent on
2776          * number of contiguous block. So we will limit
2777          * number of contiguous block to a sane value
2778          */
2779         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) &&
2780             (max_blocks > EXT4_MAX_TRANS_DATA))
2781                 max_blocks = EXT4_MAX_TRANS_DATA;
2782
2783         return ext4_chunk_trans_blocks(inode, max_blocks);
2784 }
2785
2786 /*
2787  * write_cache_pages_da - walk the list of dirty pages of the given
2788  * address space and call the callback function (which usually writes
2789  * the pages).
2790  *
2791  * This is a forked version of write_cache_pages().  Differences:
2792  *      Range cyclic is ignored.
2793  *      no_nrwrite_index_update is always presumed true
2794  */
2795 static int write_cache_pages_da(struct address_space *mapping,
2796                                 struct writeback_control *wbc,
2797                                 struct mpage_da_data *mpd)
2798 {
2799         int ret = 0;
2800         int done = 0;
2801         struct pagevec pvec;
2802         int nr_pages;
2803         pgoff_t index;
2804         pgoff_t end;            /* Inclusive */
2805         long nr_to_write = wbc->nr_to_write;
2806
2807         pagevec_init(&pvec, 0);
2808         index = wbc->range_start >> PAGE_CACHE_SHIFT;
2809         end = wbc->range_end >> PAGE_CACHE_SHIFT;
2810
2811         while (!done && (index <= end)) {
2812                 int i;
2813
2814                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2815                               PAGECACHE_TAG_DIRTY,
2816                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2817                 if (nr_pages == 0)
2818                         break;
2819
2820                 for (i = 0; i < nr_pages; i++) {
2821                         struct page *page = pvec.pages[i];
2822
2823                         /*
2824                          * At this point, the page may be truncated or
2825                          * invalidated (changing page->mapping to NULL), or
2826                          * even swizzled back from swapper_space to tmpfs file
2827                          * mapping. However, page->index will not change
2828                          * because we have a reference on the page.
2829                          */
2830                         if (page->index > end) {
2831                                 done = 1;
2832                                 break;
2833                         }
2834
2835                         lock_page(page);
2836
2837                         /*
2838                          * Page truncated or invalidated. We can freely skip it
2839                          * then, even for data integrity operations: the page
2840                          * has disappeared concurrently, so there could be no
2841                          * real expectation of this data interity operation
2842                          * even if there is now a new, dirty page at the same
2843                          * pagecache address.
2844                          */
2845                         if (unlikely(page->mapping != mapping)) {
2846 continue_unlock:
2847                                 unlock_page(page);
2848                                 continue;
2849                         }
2850
2851                         if (!PageDirty(page)) {
2852                                 /* someone wrote it for us */
2853                                 goto continue_unlock;
2854                         }
2855
2856                         if (PageWriteback(page)) {
2857                                 if (wbc->sync_mode != WB_SYNC_NONE)
2858                                         wait_on_page_writeback(page);
2859                                 else
2860                                         goto continue_unlock;
2861                         }
2862
2863                         BUG_ON(PageWriteback(page));
2864                         if (!clear_page_dirty_for_io(page))
2865                                 goto continue_unlock;
2866
2867                         ret = __mpage_da_writepage(page, wbc, mpd);
2868                         if (unlikely(ret)) {
2869                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2870                                         unlock_page(page);
2871                                         ret = 0;
2872                                 } else {
2873                                         done = 1;
2874                                         break;
2875                                 }
2876                         }
2877
2878                         if (nr_to_write > 0) {
2879                                 nr_to_write--;
2880                                 if (nr_to_write == 0 &&
2881                                     wbc->sync_mode == WB_SYNC_NONE) {
2882                                         /*
2883                                          * We stop writing back only if we are
2884                                          * not doing integrity sync. In case of
2885                                          * integrity sync we have to keep going
2886                                          * because someone may be concurrently
2887                                          * dirtying pages, and we might have
2888                                          * synced a lot of newly appeared dirty
2889                                          * pages, but have not synced all of the
2890                                          * old dirty pages.
2891                                          */
2892                                         done = 1;
2893                                         break;
2894                                 }
2895                         }
2896                 }
2897                 pagevec_release(&pvec);
2898                 cond_resched();
2899         }
2900         return ret;
2901 }
2902
2903
2904 static int ext4_da_writepages(struct address_space *mapping,
2905                               struct writeback_control *wbc)
2906 {
2907         pgoff_t index;
2908         int range_whole = 0;
2909         handle_t *handle = NULL;
2910         struct mpage_da_data mpd;
2911         struct inode *inode = mapping->host;
2912         int pages_written = 0;
2913         long pages_skipped;
2914         unsigned int max_pages;
2915         int range_cyclic, cycled = 1, io_done = 0;
2916         int needed_blocks, ret = 0;
2917         long desired_nr_to_write, nr_to_writebump = 0;
2918         loff_t range_start = wbc->range_start;
2919         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2920
2921         trace_ext4_da_writepages(inode, wbc);
2922
2923         /*
2924          * No pages to write? This is mainly a kludge to avoid starting
2925          * a transaction for special inodes like journal inode on last iput()
2926          * because that could violate lock ordering on umount
2927          */
2928         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2929                 return 0;
2930
2931         /*
2932          * If the filesystem has aborted, it is read-only, so return
2933          * right away instead of dumping stack traces later on that
2934          * will obscure the real source of the problem.  We test
2935          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2936          * the latter could be true if the filesystem is mounted
2937          * read-only, and in that case, ext4_da_writepages should
2938          * *never* be called, so if that ever happens, we would want
2939          * the stack trace.
2940          */
2941         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2942                 return -EROFS;
2943
2944         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2945                 range_whole = 1;
2946
2947         range_cyclic = wbc->range_cyclic;
2948         if (wbc->range_cyclic) {
2949                 index = mapping->writeback_index;
2950                 if (index)
2951                         cycled = 0;
2952                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2953                 wbc->range_end  = LLONG_MAX;
2954                 wbc->range_cyclic = 0;
2955         } else
2956                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2957
2958         /*
2959          * This works around two forms of stupidity.  The first is in
2960          * the writeback code, which caps the maximum number of pages
2961          * written to be 1024 pages.  This is wrong on multiple
2962          * levels; different architectues have a different page size,
2963          * which changes the maximum amount of data which gets
2964          * written.  Secondly, 4 megabytes is way too small.  XFS
2965          * forces this value to be 16 megabytes by multiplying
2966          * nr_to_write parameter by four, and then relies on its
2967          * allocator to allocate larger extents to make them
2968          * contiguous.  Unfortunately this brings us to the second
2969          * stupidity, which is that ext4's mballoc code only allocates
2970          * at most 2048 blocks.  So we force contiguous writes up to
2971          * the number of dirty blocks in the inode, or
2972          * sbi->max_writeback_mb_bump whichever is smaller.
2973          */
2974         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2975         if (!range_cyclic && range_whole)
2976                 desired_nr_to_write = wbc->nr_to_write * 8;
2977         else
2978                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2979                                                            max_pages);
2980         if (desired_nr_to_write > max_pages)
2981                 desired_nr_to_write = max_pages;
2982
2983         if (wbc->nr_to_write < desired_nr_to_write) {
2984                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2985                 wbc->nr_to_write = desired_nr_to_write;
2986         }
2987
2988         mpd.wbc = wbc;
2989         mpd.inode = mapping->host;
2990
2991         pages_skipped = wbc->pages_skipped;
2992
2993 retry:
2994         while (!ret && wbc->nr_to_write > 0) {
2995
2996                 /*
2997                  * we  insert one extent at a time. So we need
2998                  * credit needed for single extent allocation.
2999                  * journalled mode is currently not supported
3000                  * by delalloc
3001                  */
3002                 BUG_ON(ext4_should_journal_data(inode));
3003                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
3004
3005                 /* start a new transaction*/
3006                 handle = ext4_journal_start(inode, needed_blocks);
3007                 if (IS_ERR(handle)) {
3008                         ret = PTR_ERR(handle);
3009                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
3010                                "%ld pages, ino %lu; err %d", __func__,
3011                                 wbc->nr_to_write, inode->i_ino, ret);
3012                         goto out_writepages;
3013                 }
3014
3015                 /*
3016                  * Now call __mpage_da_writepage to find the next
3017                  * contiguous region of logical blocks that need
3018                  * blocks to be allocated by ext4.  We don't actually
3019                  * submit the blocks for I/O here, even though
3020                  * write_cache_pages thinks it will, and will set the
3021                  * pages as clean for write before calling
3022                  * __mpage_da_writepage().
3023                  */
3024                 mpd.b_size = 0;
3025                 mpd.b_state = 0;
3026                 mpd.b_blocknr = 0;
3027                 mpd.first_page = 0;
3028                 mpd.next_page = 0;
3029                 mpd.io_done = 0;
3030                 mpd.pages_written = 0;
3031                 mpd.retval = 0;
3032                 ret = write_cache_pages_da(mapping, wbc, &mpd);
3033                 /*
3034                  * If we have a contiguous extent of pages and we
3035                  * haven't done the I/O yet, map the blocks and submit
3036                  * them for I/O.
3037                  */
3038                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
3039                         if (mpage_da_map_blocks(&mpd) == 0)
3040                                 mpage_da_submit_io(&mpd);
3041                         mpd.io_done = 1;
3042                         ret = MPAGE_DA_EXTENT_TAIL;
3043                 }
3044                 trace_ext4_da_write_pages(inode, &mpd);
3045                 wbc->nr_to_write -= mpd.pages_written;
3046
3047                 ext4_journal_stop(handle);
3048
3049                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
3050                         /* commit the transaction which would
3051                          * free blocks released in the transaction
3052                          * and try again
3053                          */
3054                         jbd2_journal_force_commit_nested(sbi->s_journal);
3055                         wbc->pages_skipped = pages_skipped;
3056                         ret = 0;
3057                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
3058                         /*
3059                          * got one extent now try with
3060                          * rest of the pages
3061                          */
3062                         pages_written += mpd.pages_written;
3063                         wbc->pages_skipped = pages_skipped;
3064                         ret = 0;
3065                         io_done = 1;
3066                 } else if (wbc->nr_to_write)
3067                         /*
3068                          * There is no more writeout needed
3069                          * or we requested for a noblocking writeout
3070                          * and we found the device congested
3071                          */
3072                         break;
3073         }
3074         if (!io_done && !cycled) {
3075                 cycled = 1;
3076                 index = 0;
3077                 wbc->range_start = index << PAGE_CACHE_SHIFT;
3078                 wbc->range_end  = mapping->writeback_index - 1;
3079                 goto retry;
3080         }
3081         if (pages_skipped != wbc->pages_skipped)
3082                 ext4_msg(inode->i_sb, KERN_CRIT,
3083                          "This should not happen leaving %s "
3084                          "with nr_to_write = %ld ret = %d",
3085                          __func__, wbc->nr_to_write, ret);
3086
3087         /* Update index */
3088         index += pages_written;
3089         wbc->range_cyclic = range_cyclic;
3090         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
3091                 /*
3092                  * set the writeback_index so that range_cyclic
3093                  * mode will write it back later
3094                  */
3095                 mapping->writeback_index = index;
3096
3097 out_writepages:
3098         wbc->nr_to_write -= nr_to_writebump;
3099         wbc->range_start = range_start;
3100         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
3101         return ret;
3102 }
3103
3104 #define FALL_BACK_TO_NONDELALLOC 1
3105 static int ext4_nonda_switch(struct super_block *sb)
3106 {
3107         s64 free_blocks, dirty_blocks;
3108         struct ext4_sb_info *sbi = EXT4_SB(sb);
3109
3110         /*
3111          * switch to non delalloc mode if we are running low
3112          * on free block. The free block accounting via percpu
3113          * counters can get slightly wrong with percpu_counter_batch getting
3114          * accumulated on each CPU without updating global counters
3115          * Delalloc need an accurate free block accounting. So switch
3116          * to non delalloc when we are near to error range.
3117          */
3118         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
3119         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
3120         if (2 * free_blocks < 3 * dirty_blocks ||
3121                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
3122                 /*
3123                  * free block count is less than 150% of dirty blocks
3124                  * or free blocks is less than watermark
3125                  */
3126                 return 1;
3127         }
3128         /*
3129          * Even if we don't switch but are nearing capacity,
3130          * start pushing delalloc when 1/2 of free blocks are dirty.
3131          */
3132         if (free_blocks < 2 * dirty_blocks)
3133                 writeback_inodes_sb_if_idle(sb);
3134
3135         return 0;
3136 }
3137
3138 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3139                                loff_t pos, unsigned len, unsigned flags,
3140                                struct page **pagep, void **fsdata)
3141 {
3142         int ret, retries = 0;
3143         struct page *page;
3144         pgoff_t index;
3145         struct inode *inode = mapping->host;
3146         handle_t *handle;
3147
3148         index = pos >> PAGE_CACHE_SHIFT;
3149
3150         if (ext4_nonda_switch(inode->i_sb)) {
3151                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3152                 return ext4_write_begin(file, mapping, pos,
3153                                         len, flags, pagep, fsdata);
3154         }
3155         *fsdata = (void *)0;
3156         trace_ext4_da_write_begin(inode, pos, len, flags);
3157 retry:
3158         /*
3159          * With delayed allocation, we don't log the i_disksize update
3160          * if there is delayed block allocation. But we still need
3161          * to journalling the i_disksize update if writes to the end
3162          * of file which has an already mapped buffer.
3163          */
3164         handle = ext4_journal_start(inode, 1);
3165         if (IS_ERR(handle)) {
3166                 ret = PTR_ERR(handle);
3167                 goto out;
3168         }
3169         /* We cannot recurse into the filesystem as the transaction is already
3170          * started */
3171         flags |= AOP_FLAG_NOFS;
3172
3173         page = grab_cache_page_write_begin(mapping, index, flags);
3174         if (!page) {
3175                 ext4_journal_stop(handle);
3176                 ret = -ENOMEM;
3177                 goto out;
3178         }
3179         *pagep = page;
3180
3181         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
3182                                 ext4_da_get_block_prep);
3183         if (ret < 0) {
3184                 unlock_page(page);
3185                 ext4_journal_stop(handle);
3186                 page_cache_release(page);
3187                 /*
3188                  * block_write_begin may have instantiated a few blocks
3189                  * outside i_size.  Trim these off again. Don't need
3190                  * i_size_read because we hold i_mutex.
3191                  */
3192                 if (pos + len > inode->i_size)
3193                         ext4_truncate_failed_write(inode);
3194         }
3195
3196         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3197                 goto retry;
3198 out:
3199         return ret;
3200 }
3201
3202 /*
3203  * Check if we should update i_disksize
3204  * when write to the end of file but not require block allocation
3205  */
3206 static int ext4_da_should_update_i_disksize(struct page *page,
3207                                             unsigned long offset)
3208 {
3209         struct buffer_head *bh;
3210         struct inode *inode = page->mapping->host;
3211         unsigned int idx;
3212         int i;
3213
3214         bh = page_buffers(page);
3215         idx = offset >> inode->i_blkbits;
3216
3217         for (i = 0; i < idx; i++)
3218                 bh = bh->b_this_page;
3219
3220         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3221                 return 0;
3222         return 1;
3223 }
3224
3225 static int ext4_da_write_end(struct file *file,
3226                              struct address_space *mapping,
3227                              loff_t pos, unsigned len, unsigned copied,
3228                              struct page *page, void *fsdata)
3229 {
3230         struct inode *inode = mapping->host;
3231         int ret = 0, ret2;
3232         handle_t *handle = ext4_journal_current_handle();
3233         loff_t new_i_size;
3234         unsigned long start, end;
3235         int write_mode = (int)(unsigned long)fsdata;
3236
3237         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
3238                 if (ext4_should_order_data(inode)) {
3239                         return ext4_ordered_write_end(file, mapping, pos,
3240                                         len, copied, page, fsdata);
3241                 } else if (ext4_should_writeback_data(inode)) {
3242                         return ext4_writeback_write_end(file, mapping, pos,
3243                                         len, copied, page, fsdata);
3244                 } else {
3245                         BUG();
3246                 }
3247         }
3248
3249         trace_ext4_da_write_end(inode, pos, len, copied);
3250         start = pos & (PAGE_CACHE_SIZE - 1);
3251         end = start + copied - 1;
3252
3253         /*
3254          * generic_write_end() will run mark_inode_dirty() if i_size
3255          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3256          * into that.
3257          */
3258
3259         new_i_size = pos + copied;
3260         if (new_i_size > EXT4_I(inode)->i_disksize) {
3261                 if (ext4_da_should_update_i_disksize(page, end)) {
3262                         down_write(&EXT4_I(inode)->i_data_sem);
3263                         if (new_i_size > EXT4_I(inode)->i_disksize) {
3264                                 /*
3265                                  * Updating i_disksize when extending file
3266                                  * without needing block allocation
3267                                  */
3268                                 if (ext4_should_order_data(inode))
3269                                         ret = ext4_jbd2_file_inode(handle,
3270                                                                    inode);
3271
3272                                 EXT4_I(inode)->i_disksize = new_i_size;
3273                         }
3274                         up_write(&EXT4_I(inode)->i_data_sem);
3275                         /* We need to mark inode dirty even if
3276                          * new_i_size is less that inode->i_size
3277                          * bu greater than i_disksize.(hint delalloc)
3278                          */
3279                         ext4_mark_inode_dirty(handle, inode);
3280                 }
3281         }
3282         ret2 = generic_write_end(file, mapping, pos, len, copied,
3283                                                         page, fsdata);
3284         copied = ret2;
3285         if (ret2 < 0)
3286                 ret = ret2;
3287         ret2 = ext4_journal_stop(handle);
3288         if (!ret)
3289                 ret = ret2;
3290
3291         return ret ? ret : copied;
3292 }
3293
3294 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
3295 {
3296         /*
3297          * Drop reserved blocks
3298          */
3299         BUG_ON(!PageLocked(page));
3300         if (!page_has_buffers(page))
3301                 goto out;
3302
3303         ext4_da_page_release_reservation(page, offset);
3304
3305 out:
3306         ext4_invalidatepage(page, offset);
3307
3308         return;
3309 }
3310
3311 /*
3312  * Force all delayed allocation blocks to be allocated for a given inode.
3313  */
3314 int ext4_alloc_da_blocks(struct inode *inode)
3315 {
3316         trace_ext4_alloc_da_blocks(inode);
3317
3318         if (!EXT4_I(inode)->i_reserved_data_blocks &&
3319             !EXT4_I(inode)->i_reserved_meta_blocks)
3320                 return 0;
3321
3322         /*
3323          * We do something simple for now.  The filemap_flush() will
3324          * also start triggering a write of the data blocks, which is
3325          * not strictly speaking necessary (and for users of
3326          * laptop_mode, not even desirable).  However, to do otherwise
3327          * would require replicating code paths in:
3328          *
3329          * ext4_da_writepages() ->
3330          *    write_cache_pages() ---> (via passed in callback function)
3331          *        __mpage_da_writepage() -->
3332          *           mpage_add_bh_to_extent()
3333          *           mpage_da_map_blocks()
3334          *
3335          * The problem is that write_cache_pages(), located in
3336          * mm/page-writeback.c, marks pages clean in preparation for
3337          * doing I/O, which is not desirable if we're not planning on
3338          * doing I/O at all.
3339          *
3340          * We could call write_cache_pages(), and then redirty all of
3341          * the pages by calling redirty_page_for_writeback() but that
3342          * would be ugly in the extreme.  So instead we would need to
3343          * replicate parts of the code in the above functions,
3344          * simplifying them becuase we wouldn't actually intend to
3345          * write out the pages, but rather only collect contiguous
3346          * logical block extents, call the multi-block allocator, and
3347          * then update the buffer heads with the block allocations.
3348          *
3349          * For now, though, we'll cheat by calling filemap_flush(),
3350          * which will map the blocks, and start the I/O, but not
3351          * actually wait for the I/O to complete.
3352          */
3353         return filemap_flush(inode->i_mapping);
3354 }
3355
3356 /*
3357  * bmap() is special.  It gets used by applications such as lilo and by
3358  * the swapper to find the on-disk block of a specific piece of data.
3359  *
3360  * Naturally, this is dangerous if the block concerned is still in the
3361  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3362  * filesystem and enables swap, then they may get a nasty shock when the
3363  * data getting swapped to that swapfile suddenly gets overwritten by
3364  * the original zero's written out previously to the journal and
3365  * awaiting writeback in the kernel's buffer cache.
3366  *
3367  * So, if we see any bmap calls here on a modified, data-journaled file,
3368  * take extra steps to flush any blocks which might be in the cache.
3369  */
3370 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3371 {
3372         struct inode *inode = mapping->host;
3373         journal_t *journal;
3374         int err;
3375
3376         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3377                         test_opt(inode->i_sb, DELALLOC)) {
3378                 /*
3379                  * With delalloc we want to sync the file
3380                  * so that we can make sure we allocate
3381                  * blocks for file
3382                  */
3383                 filemap_write_and_wait(mapping);
3384         }
3385
3386         if (EXT4_JOURNAL(inode) &&
3387             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3388                 /*
3389                  * This is a REALLY heavyweight approach, but the use of
3390                  * bmap on dirty files is expected to be extremely rare:
3391                  * only if we run lilo or swapon on a freshly made file
3392                  * do we expect this to happen.
3393                  *
3394                  * (bmap requires CAP_SYS_RAWIO so this does not
3395                  * represent an unprivileged user DOS attack --- we'd be
3396                  * in trouble if mortal users could trigger this path at
3397                  * will.)
3398                  *
3399                  * NB. EXT4_STATE_JDATA is not set on files other than
3400                  * regular files.  If somebody wants to bmap a directory
3401                  * or symlink and gets confused because the buffer
3402                  * hasn't yet been flushed to disk, they deserve
3403                  * everything they get.
3404                  */
3405
3406                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3407                 journal = EXT4_JOURNAL(inode);
3408                 jbd2_journal_lock_updates(journal);
3409                 err = jbd2_journal_flush(journal);
3410                 jbd2_journal_unlock_updates(journal);
3411
3412                 if (err)
3413                         return 0;
3414         }
3415
3416         return generic_block_bmap(mapping, block, ext4_get_block);
3417 }
3418
3419 static int ext4_readpage(struct file *file, struct page *page)
3420 {
3421         return mpage_readpage(page, ext4_get_block);
3422 }
3423
3424 static int
3425 ext4_readpages(struct file *file, struct address_space *mapping,
3426                 struct list_head *pages, unsigned nr_pages)
3427 {
3428         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3429 }
3430
3431 static void ext4_free_io_end(ext4_io_end_t *io)
3432 {
3433         BUG_ON(!io);
3434         if (io->page)
3435                 put_page(io->page);
3436         iput(io->inode);
3437         kfree(io);
3438 }
3439
3440 static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
3441 {
3442         struct buffer_head *head, *bh;
3443         unsigned int curr_off = 0;
3444
3445         if (!page_has_buffers(page))
3446                 return;
3447         head = bh = page_buffers(page);
3448         do {
3449                 if (offset <= curr_off && test_clear_buffer_uninit(bh)
3450                                         && bh->b_private) {
3451                         ext4_free_io_end(bh->b_private);
3452                         bh->b_private = NULL;
3453                         bh->b_end_io = NULL;
3454                 }
3455                 curr_off = curr_off + bh->b_size;
3456                 bh = bh->b_this_page;
3457         } while (bh != head);
3458 }
3459
3460 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3461 {
3462         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3463
3464         /*
3465          * free any io_end structure allocated for buffers to be discarded
3466          */
3467         if (ext4_should_dioread_nolock(page->mapping->host))
3468                 ext4_invalidatepage_free_endio(page, offset);
3469         /*
3470          * If it's a full truncate we just forget about the pending dirtying
3471          */
3472         if (offset == 0)
3473                 ClearPageChecked(page);
3474
3475         if (journal)
3476                 jbd2_journal_invalidatepage(journal, page, offset);
3477         else
3478                 block_invalidatepage(page, offset);
3479 }
3480
3481 static int ext4_releasepage(struct page *page, gfp_t wait)
3482 {
3483         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3484
3485         WARN_ON(PageChecked(page));
3486         if (!page_has_buffers(page))
3487                 return 0;
3488         if (journal)
3489                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3490         else
3491                 return try_to_free_buffers(page);
3492 }
3493
3494 /*
3495  * O_DIRECT for ext3 (or indirect map) based files
3496  *
3497  * If the O_DIRECT write will extend the file then add this inode to the
3498  * orphan list.  So recovery will truncate it back to the original size
3499  * if the machine crashes during the write.
3500  *
3501  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3502  * crashes then stale disk data _may_ be exposed inside the file. But current
3503  * VFS code falls back into buffered path in that case so we are safe.
3504  */
3505 static ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
3506                               const struct iovec *iov, loff_t offset,
3507                               unsigned long nr_segs)
3508 {
3509         struct file *file = iocb->ki_filp;
3510         struct inode *inode = file->f_mapping->host;
3511         struct ext4_inode_info *ei = EXT4_I(inode);
3512         handle_t *handle;
3513         ssize_t ret;
3514         int orphan = 0;
3515         size_t count = iov_length(iov, nr_segs);
3516         int retries = 0;
3517
3518         if (rw == WRITE) {
3519                 loff_t final_size = offset + count;
3520
3521                 if (final_size > inode->i_size) {
3522                         /* Credits for sb + inode write */
3523                         handle = ext4_journal_start(inode, 2);
3524                         if (IS_ERR(handle)) {
3525                                 ret = PTR_ERR(handle);
3526                                 goto out;
3527                         }
3528                         ret = ext4_orphan_add(handle, inode);
3529                         if (ret) {
3530                                 ext4_journal_stop(handle);
3531                                 goto out;
3532                         }
3533                         orphan = 1;
3534                         ei->i_disksize = inode->i_size;
3535                         ext4_journal_stop(handle);
3536                 }
3537         }
3538
3539 retry:
3540         if (rw == READ && ext4_should_dioread_nolock(inode))
3541                 ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
3542                                  inode->i_sb->s_bdev, iov,
3543                                  offset, nr_segs,
3544                                  ext4_get_block, NULL);
3545         else
3546                 ret = blockdev_direct_IO(rw, iocb, inode,
3547                                  inode->i_sb->s_bdev, iov,
3548                                  offset, nr_segs,
3549                                  ext4_get_block, NULL);
3550         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3551                 goto retry;
3552
3553         if (orphan) {
3554                 int err;
3555
3556                 /* Credits for sb + inode write */
3557                 handle = ext4_journal_start(inode, 2);
3558                 if (IS_ERR(handle)) {
3559                         /* This is really bad luck. We've written the data
3560                          * but cannot extend i_size. Bail out and pretend
3561                          * the write failed... */
3562                         ret = PTR_ERR(handle);
3563                         if (inode->i_nlink)
3564                                 ext4_orphan_del(NULL, inode);
3565
3566                         goto out;
3567                 }
3568                 if (inode->i_nlink)
3569                         ext4_orphan_del(handle, inode);
3570                 if (ret > 0) {
3571                         loff_t end = offset + ret;
3572                         if (end > inode->i_size) {
3573                                 ei->i_disksize = end;
3574                                 i_size_write(inode, end);
3575                                 /*
3576                                  * We're going to return a positive `ret'
3577                                  * here due to non-zero-length I/O, so there's
3578                                  * no way of reporting error returns from
3579                                  * ext4_mark_inode_dirty() to userspace.  So
3580                                  * ignore it.
3581                                  */
3582                                 ext4_mark_inode_dirty(handle, inode);
3583                         }
3584                 }
3585                 err = ext4_journal_stop(handle);
3586                 if (ret == 0)
3587                         ret = err;
3588         }
3589 out:
3590         return ret;
3591 }
3592
3593 /*
3594  * ext4_get_block used when preparing for a DIO write or buffer write.
3595  * We allocate an uinitialized extent if blocks haven't been allocated.
3596  * The extent will be converted to initialized after the IO is complete.
3597  */
3598 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
3599                    struct buffer_head *bh_result, int create)
3600 {
3601         ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
3602                    inode->i_ino, create);
3603         return _ext4_get_block(inode, iblock, bh_result,
3604                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
3605 }
3606
3607 static void dump_completed_IO(struct inode * inode)
3608 {
3609 #ifdef  EXT4_DEBUG
3610         struct list_head *cur, *before, *after;
3611         ext4_io_end_t *io, *io0, *io1;
3612         unsigned long flags;
3613
3614         if (list_empty(&EXT4_I(inode)->i_completed_io_list)){
3615                 ext4_debug("inode %lu completed_io list is empty\n", inode->i_ino);
3616                 return;
3617         }
3618
3619         ext4_debug("Dump inode %lu completed_io list \n", inode->i_ino);
3620         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3621         list_for_each_entry(io, &EXT4_I(inode)->i_completed_io_list, list){
3622                 cur = &io->list;
3623                 before = cur->prev;
3624                 io0 = container_of(before, ext4_io_end_t, list);
3625                 after = cur->next;
3626                 io1 = container_of(after, ext4_io_end_t, list);
3627
3628                 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
3629                             io, inode->i_ino, io0, io1);
3630         }
3631         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3632 #endif
3633 }
3634
3635 /*
3636  * check a range of space and convert unwritten extents to written.
3637  */
3638 static int ext4_end_io_nolock(ext4_io_end_t *io)
3639 {
3640         struct inode *inode = io->inode;
3641         loff_t offset = io->offset;
3642         ssize_t size = io->size;
3643         int ret = 0;
3644
3645         ext4_debug("ext4_end_io_nolock: io 0x%p from inode %lu,list->next 0x%p,"
3646                    "list->prev 0x%p\n",
3647                    io, inode->i_ino, io->list.next, io->list.prev);
3648
3649         if (list_empty(&io->list))
3650                 return ret;
3651
3652         if (io->flag != EXT4_IO_UNWRITTEN)
3653                 return ret;
3654
3655         ret = ext4_convert_unwritten_extents(inode, offset, size);
3656         if (ret < 0) {
3657                 printk(KERN_EMERG "%s: failed to convert unwritten"
3658                         "extents to written extents, error is %d"
3659                         " io is still on inode %lu aio dio list\n",
3660                        __func__, ret, inode->i_ino);
3661                 return ret;
3662         }
3663
3664         /* clear the DIO AIO unwritten flag */
3665         io->flag = 0;
3666         return ret;
3667 }
3668
3669 /*
3670  * work on completed aio dio IO, to convert unwritten extents to extents
3671  */
3672 static void ext4_end_io_work(struct work_struct *work)
3673 {
3674         ext4_io_end_t           *io = container_of(work, ext4_io_end_t, work);
3675         struct inode            *inode = io->inode;
3676         struct ext4_inode_info  *ei = EXT4_I(inode);
3677         unsigned long           flags;
3678         int                     ret;
3679
3680         mutex_lock(&inode->i_mutex);
3681         ret = ext4_end_io_nolock(io);
3682         if (ret < 0) {
3683                 mutex_unlock(&inode->i_mutex);
3684                 return;
3685         }
3686
3687         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3688         if (!list_empty(&io->list))
3689                 list_del_init(&io->list);
3690         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3691         mutex_unlock(&inode->i_mutex);
3692         ext4_free_io_end(io);
3693 }
3694
3695 /*
3696  * This function is called from ext4_sync_file().
3697  *
3698  * When IO is completed, the work to convert unwritten extents to
3699  * written is queued on workqueue but may not get immediately
3700  * scheduled. When fsync is called, we need to ensure the
3701  * conversion is complete before fsync returns.
3702  * The inode keeps track of a list of pending/completed IO that
3703  * might needs to do the conversion. This function walks through
3704  * the list and convert the related unwritten extents for completed IO
3705  * to written.
3706  * The function return the number of pending IOs on success.
3707  */
3708 int flush_completed_IO(struct inode *inode)
3709 {
3710         ext4_io_end_t *io;
3711         struct ext4_inode_info *ei = EXT4_I(inode);
3712         unsigned long flags;
3713         int ret = 0;
3714         int ret2 = 0;
3715
3716         if (list_empty(&ei->i_completed_io_list))
3717                 return ret;
3718
3719         dump_completed_IO(inode);
3720         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3721         while (!list_empty(&ei->i_completed_io_list)){
3722                 io = list_entry(ei->i_completed_io_list.next,
3723                                 ext4_io_end_t, list);
3724                 /*
3725                  * Calling ext4_end_io_nolock() to convert completed
3726                  * IO to written.
3727                  *
3728                  * When ext4_sync_file() is called, run_queue() may already
3729                  * about to flush the work corresponding to this io structure.
3730                  * It will be upset if it founds the io structure related
3731                  * to the work-to-be schedule is freed.
3732                  *
3733                  * Thus we need to keep the io structure still valid here after
3734                  * convertion finished. The io structure has a flag to
3735                  * avoid double converting from both fsync and background work
3736                  * queue work.
3737                  */
3738                 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3739                 ret = ext4_end_io_nolock(io);
3740                 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3741                 if (ret < 0)
3742                         ret2 = ret;
3743                 else
3744                         list_del_init(&io->list);
3745         }
3746         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3747         return (ret2 < 0) ? ret2 : 0;
3748 }
3749
3750 static ext4_io_end_t *ext4_init_io_end (struct inode *inode, gfp_t flags)
3751 {
3752         ext4_io_end_t *io = NULL;
3753
3754         io = kmalloc(sizeof(*io), flags);
3755
3756         if (io) {
3757                 igrab(inode);
3758                 io->inode = inode;
3759                 io->flag = 0;
3760                 io->offset = 0;
3761                 io->size = 0;
3762                 io->page = NULL;
3763                 INIT_WORK(&io->work, ext4_end_io_work);
3764                 INIT_LIST_HEAD(&io->list);
3765         }
3766
3767         return io;
3768 }
3769
3770 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3771                             ssize_t size, void *private)
3772 {
3773         ext4_io_end_t *io_end = iocb->private;
3774         struct workqueue_struct *wq;
3775         unsigned long flags;
3776         struct ext4_inode_info *ei;
3777
3778         /* if not async direct IO or dio with 0 bytes write, just return */
3779         if (!io_end || !size)
3780                 return;
3781
3782         ext_debug("ext4_end_io_dio(): io_end 0x%p"
3783                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
3784                   iocb->private, io_end->inode->i_ino, iocb, offset,
3785                   size);
3786
3787         /* if not aio dio with unwritten extents, just free io and return */
3788         if (io_end->flag != EXT4_IO_UNWRITTEN){
3789                 ext4_free_io_end(io_end);
3790                 iocb->private = NULL;
3791                 return;
3792         }
3793
3794         io_end->offset = offset;
3795         io_end->size = size;
3796         io_end->flag = EXT4_IO_UNWRITTEN;
3797         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
3798
3799         /* queue the work to convert unwritten extents to written */
3800         queue_work(wq, &io_end->work);
3801
3802         /* Add the io_end to per-inode completed aio dio list*/
3803         ei = EXT4_I(io_end->inode);
3804         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3805         list_add_tail(&io_end->list, &ei->i_completed_io_list);
3806         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3807         iocb->private = NULL;
3808 }
3809
3810 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
3811 {
3812         ext4_io_end_t *io_end = bh->b_private;
3813         struct workqueue_struct *wq;
3814         struct inode *inode;
3815         unsigned long flags;
3816
3817         if (!test_clear_buffer_uninit(bh) || !io_end)
3818                 goto out;
3819
3820         if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
3821                 printk("sb umounted, discard end_io request for inode %lu\n",
3822                         io_end->inode->i_ino);
3823                 ext4_free_io_end(io_end);
3824                 goto out;
3825         }
3826
3827         io_end->flag = EXT4_IO_UNWRITTEN;
3828         inode = io_end->inode;
3829
3830         /* Add the io_end to per-inode completed io list*/
3831         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3832         list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
3833         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3834
3835         wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
3836         /* queue the work to convert unwritten extents to written */
3837         queue_work(wq, &io_end->work);
3838 out:
3839         bh->b_private = NULL;
3840         bh->b_end_io = NULL;
3841         clear_buffer_uninit(bh);
3842         end_buffer_async_write(bh, uptodate);
3843 }
3844
3845 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
3846 {
3847         ext4_io_end_t *io_end;
3848         struct page *page = bh->b_page;
3849         loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
3850         size_t size = bh->b_size;
3851
3852 retry:
3853         io_end = ext4_init_io_end(inode, GFP_ATOMIC);
3854         if (!io_end) {
3855                 if (printk_ratelimit())
3856                         printk(KERN_WARNING "%s: allocation fail\n", __func__);
3857                 schedule();
3858                 goto retry;
3859         }
3860         io_end->offset = offset;
3861         io_end->size = size;
3862         /*
3863          * We need to hold a reference to the page to make sure it
3864          * doesn't get evicted before ext4_end_io_work() has a chance
3865          * to convert the extent from written to unwritten.
3866          */
3867         io_end->page = page;
3868         get_page(io_end->page);
3869
3870         bh->b_private = io_end;
3871         bh->b_end_io = ext4_end_io_buffer_write;
3872         return 0;
3873 }
3874
3875 /*
3876  * For ext4 extent files, ext4 will do direct-io write to holes,
3877  * preallocated extents, and those write extend the file, no need to
3878  * fall back to buffered IO.
3879  *
3880  * For holes, we fallocate those blocks, mark them as unintialized
3881  * If those blocks were preallocated, we mark sure they are splited, but
3882  * still keep the range to write as unintialized.
3883  *
3884  * The unwrritten extents will be converted to written when DIO is completed.
3885  * For async direct IO, since the IO may still pending when return, we
3886  * set up an end_io call back function, which will do the convertion
3887  * when async direct IO completed.
3888  *
3889  * If the O_DIRECT write will extend the file then add this inode to the
3890  * orphan list.  So recovery will truncate it back to the original size
3891  * if the machine crashes during the write.
3892  *
3893  */
3894 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
3895                               const struct iovec *iov, loff_t offset,
3896                               unsigned long nr_segs)
3897 {
3898         struct file *file = iocb->ki_filp;
3899         struct inode *inode = file->f_mapping->host;
3900         ssize_t ret;
3901         size_t count = iov_length(iov, nr_segs);
3902
3903         loff_t final_size = offset + count;
3904         if (rw == WRITE && final_size <= inode->i_size) {
3905                 /*
3906                  * We could direct write to holes and fallocate.
3907                  *
3908                  * Allocated blocks to fill the hole are marked as uninitialized
3909                  * to prevent paralel buffered read to expose the stale data
3910                  * before DIO complete the data IO.
3911                  *
3912                  * As to previously fallocated extents, ext4 get_block
3913                  * will just simply mark the buffer mapped but still
3914                  * keep the extents uninitialized.
3915                  *
3916                  * for non AIO case, we will convert those unwritten extents
3917                  * to written after return back from blockdev_direct_IO.
3918                  *
3919                  * for async DIO, the conversion needs to be defered when
3920                  * the IO is completed. The ext4 end_io callback function
3921                  * will be called to take care of the conversion work.
3922                  * Here for async case, we allocate an io_end structure to
3923                  * hook to the iocb.
3924                  */
3925                 iocb->private = NULL;
3926                 EXT4_I(inode)->cur_aio_dio = NULL;
3927                 if (!is_sync_kiocb(iocb)) {
3928                         iocb->private = ext4_init_io_end(inode, GFP_NOFS);
3929                         if (!iocb->private)
3930                                 return -ENOMEM;
3931                         /*
3932                          * we save the io structure for current async
3933                          * direct IO, so that later ext4_get_blocks()
3934                          * could flag the io structure whether there
3935                          * is a unwritten extents needs to be converted
3936                          * when IO is completed.
3937                          */
3938                         EXT4_I(inode)->cur_aio_dio = iocb->private;
3939                 }
3940
3941                 ret = blockdev_direct_IO(rw, iocb, inode,
3942                                          inode->i_sb->s_bdev, iov,
3943                                          offset, nr_segs,
3944                                          ext4_get_block_write,
3945                                          ext4_end_io_dio);
3946                 if (iocb->private)
3947                         EXT4_I(inode)->cur_aio_dio = NULL;
3948                 /*
3949                  * The io_end structure takes a reference to the inode,
3950                  * that structure needs to be destroyed and the
3951                  * reference to the inode need to be dropped, when IO is
3952                  * complete, even with 0 byte write, or failed.
3953                  *
3954                  * In the successful AIO DIO case, the io_end structure will be
3955                  * desctroyed and the reference to the inode will be dropped
3956                  * after the end_io call back function is called.
3957                  *
3958                  * In the case there is 0 byte write, or error case, since
3959                  * VFS direct IO won't invoke the end_io call back function,
3960                  * we need to free the end_io structure here.
3961                  */
3962                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
3963                         ext4_free_io_end(iocb->private);
3964                         iocb->private = NULL;
3965                 } else if (ret > 0 && ext4_test_inode_state(inode,
3966                                                 EXT4_STATE_DIO_UNWRITTEN)) {
3967                         int err;
3968                         /*
3969                          * for non AIO case, since the IO is already
3970                          * completed, we could do the convertion right here
3971                          */
3972                         err = ext4_convert_unwritten_extents(inode,
3973                                                              offset, ret);
3974                         if (err < 0)
3975                                 ret = err;
3976                         ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3977                 }
3978                 return ret;
3979         }
3980
3981         /* for write the the end of file case, we fall back to old way */
3982         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3983 }
3984
3985 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3986                               const struct iovec *iov, loff_t offset,
3987                               unsigned long nr_segs)
3988 {
3989         struct file *file = iocb->ki_filp;
3990         struct inode *inode = file->f_mapping->host;
3991
3992         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3993                 return ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
3994
3995         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3996 }
3997
3998 /*
3999  * Pages can be marked dirty completely asynchronously from ext4's journalling
4000  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
4001  * much here because ->set_page_dirty is called under VFS locks.  The page is
4002  * not necessarily locked.
4003  *
4004  * We cannot just dirty the page and leave attached buffers clean, because the
4005  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
4006  * or jbddirty because all the journalling code will explode.
4007  *
4008  * So what we do is to mark the page "pending dirty" and next time writepage
4009  * is called, propagate that into the buffers appropriately.
4010  */
4011 static int ext4_journalled_set_page_dirty(struct page *page)
4012 {
4013         SetPageChecked(page);
4014         return __set_page_dirty_nobuffers(page);
4015 }
4016
4017 static const struct address_space_operations ext4_ordered_aops = {
4018         .readpage               = ext4_readpage,
4019         .readpages              = ext4_readpages,
4020         .writepage              = ext4_writepage,
4021         .sync_page              = block_sync_page,
4022         .write_begin            = ext4_write_begin,
4023         .write_end              = ext4_ordered_write_end,
4024         .bmap                   = ext4_bmap,
4025         .invalidatepage         = ext4_invalidatepage,
4026         .releasepage            = ext4_releasepage,
4027         .direct_IO              = ext4_direct_IO,
4028         .migratepage            = buffer_migrate_page,
4029         .is_partially_uptodate  = block_is_partially_uptodate,
4030         .error_remove_page      = generic_error_remove_page,
4031 };
4032
4033 static const struct address_space_operations ext4_writeback_aops = {
4034         .readpage               = ext4_readpage,
4035         .readpages              = ext4_readpages,
4036         .writepage              = ext4_writepage,
4037         .sync_page              = block_sync_page,
4038         .write_begin            = ext4_write_begin,
4039         .write_end              = ext4_writeback_write_end,
4040         .bmap                   = ext4_bmap,
4041         .invalidatepage         = ext4_invalidatepage,
4042         .releasepage            = ext4_releasepage,
4043         .direct_IO              = ext4_direct_IO,
4044         .migratepage            = buffer_migrate_page,
4045         .is_partially_uptodate  = block_is_partially_uptodate,
4046         .error_remove_page      = generic_error_remove_page,
4047 };
4048
4049 static const struct address_space_operations ext4_journalled_aops = {
4050         .readpage               = ext4_readpage,
4051         .readpages              = ext4_readpages,
4052         .writepage              = ext4_writepage,
4053         .sync_page              = block_sync_page,
4054         .write_begin            = ext4_write_begin,
4055         .write_end              = ext4_journalled_write_end,
4056         .set_page_dirty         = ext4_journalled_set_page_dirty,
4057         .bmap                   = ext4_bmap,
4058         .invalidatepage         = ext4_invalidatepage,
4059         .releasepage            = ext4_releasepage,
4060         .is_partially_uptodate  = block_is_partially_uptodate,
4061         .error_remove_page      = generic_error_remove_page,
4062 };
4063
4064 static const struct address_space_operations ext4_da_aops = {
4065         .readpage               = ext4_readpage,
4066         .readpages              = ext4_readpages,
4067         .writepage              = ext4_writepage,
4068         .writepages             = ext4_da_writepages,
4069         .sync_page              = block_sync_page,
4070         .write_begin            = ext4_da_write_begin,
4071         .write_end              = ext4_da_write_end,
4072         .bmap                   = ext4_bmap,
4073         .invalidatepage         = ext4_da_invalidatepage,
4074         .releasepage            = ext4_releasepage,
4075         .direct_IO              = ext4_direct_IO,
4076         .migratepage            = buffer_migrate_page,
4077         .is_partially_uptodate  = block_is_partially_uptodate,
4078         .error_remove_page      = generic_error_remove_page,
4079 };
4080
4081 void ext4_set_aops(struct inode *inode)
4082 {
4083         if (ext4_should_order_data(inode) &&
4084                 test_opt(inode->i_sb, DELALLOC))
4085                 inode->i_mapping->a_ops = &ext4_da_aops;
4086         else if (ext4_should_order_data(inode))
4087                 inode->i_mapping->a_ops = &ext4_ordered_aops;
4088         else if (ext4_should_writeback_data(inode) &&
4089                  test_opt(inode->i_sb, DELALLOC))
4090                 inode->i_mapping->a_ops = &ext4_da_aops;
4091         else if (ext4_should_writeback_data(inode))
4092                 inode->i_mapping->a_ops = &ext4_writeback_aops;
4093         else
4094                 inode->i_mapping->a_ops = &ext4_journalled_aops;
4095 }
4096
4097 /*
4098  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
4099  * up to the end of the block which corresponds to `from'.
4100  * This required during truncate. We need to physically zero the tail end
4101  * of that block so it doesn't yield old data if the file is later grown.
4102  */
4103 int ext4_block_truncate_page(handle_t *handle,
4104                 struct address_space *mapping, loff_t from)
4105 {
4106         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
4107         unsigned offset = from & (PAGE_CACHE_SIZE-1);
4108         unsigned blocksize, length, pos;
4109         ext4_lblk_t iblock;
4110         struct inode *inode = mapping->host;
4111         struct buffer_head *bh;
4112         struct page *page;
4113         int err = 0;
4114
4115         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
4116                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
4117         if (!page)
4118                 return -EINVAL;
4119
4120         blocksize = inode->i_sb->s_blocksize;
4121         length = blocksize - (offset & (blocksize - 1));
4122         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
4123
4124         if (!page_has_buffers(page))
4125                 create_empty_buffers(page, blocksize, 0);
4126
4127         /* Find the buffer that contains "offset" */
4128         bh = page_buffers(page);
4129         pos = blocksize;
4130         while (offset >= pos) {
4131                 bh = bh->b_this_page;
4132                 iblock++;
4133                 pos += blocksize;
4134         }
4135
4136         err = 0;
4137         if (buffer_freed(bh)) {
4138                 BUFFER_TRACE(bh, "freed: skip");
4139                 goto unlock;
4140         }
4141
4142         if (!buffer_mapped(bh)) {
4143                 BUFFER_TRACE(bh, "unmapped");
4144                 ext4_get_block(inode, iblock, bh, 0);
4145                 /* unmapped? It's a hole - nothing to do */
4146                 if (!buffer_mapped(bh)) {
4147                         BUFFER_TRACE(bh, "still unmapped");
4148                         goto unlock;
4149                 }
4150         }
4151
4152         /* Ok, it's mapped. Make sure it's up-to-date */
4153         if (PageUptodate(page))
4154                 set_buffer_uptodate(bh);
4155
4156         if (!buffer_uptodate(bh)) {
4157                 err = -EIO;
4158                 ll_rw_block(READ, 1, &bh);
4159                 wait_on_buffer(bh);
4160                 /* Uhhuh. Read error. Complain and punt. */
4161                 if (!buffer_uptodate(bh))
4162                         goto unlock;
4163         }
4164
4165         if (ext4_should_journal_data(inode)) {
4166                 BUFFER_TRACE(bh, "get write access");
4167                 err = ext4_journal_get_write_access(handle, bh);
4168                 if (err)
4169                         goto unlock;
4170         }
4171
4172         zero_user(page, offset, length);
4173
4174         BUFFER_TRACE(bh, "zeroed end of block");
4175
4176         err = 0;
4177         if (ext4_should_journal_data(inode)) {
4178                 err = ext4_handle_dirty_metadata(handle, inode, bh);
4179         } else {
4180                 if (ext4_should_order_data(inode))
4181                         err = ext4_jbd2_file_inode(handle, inode);
4182                 mark_buffer_dirty(bh);
4183         }
4184
4185 unlock:
4186         unlock_page(page);
4187         page_cache_release(page);
4188         return err;
4189 }
4190
4191 /*
4192  * Probably it should be a library function... search for first non-zero word
4193  * or memcmp with zero_page, whatever is better for particular architecture.
4194  * Linus?
4195  */
4196 static inline int all_zeroes(__le32 *p, __le32 *q)
4197 {
4198         while (p < q)
4199                 if (*p++)
4200                         return 0;
4201         return 1;
4202 }
4203
4204 /**
4205  *      ext4_find_shared - find the indirect blocks for partial truncation.
4206  *      @inode:   inode in question
4207  *      @depth:   depth of the affected branch
4208  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
4209  *      @chain:   place to store the pointers to partial indirect blocks
4210  *      @top:     place to the (detached) top of branch
4211  *
4212  *      This is a helper function used by ext4_truncate().
4213  *
4214  *      When we do truncate() we may have to clean the ends of several
4215  *      indirect blocks but leave the blocks themselves alive. Block is
4216  *      partially truncated if some data below the new i_size is refered
4217  *      from it (and it is on the path to the first completely truncated
4218  *      data block, indeed).  We have to free the top of that path along
4219  *      with everything to the right of the path. Since no allocation
4220  *      past the truncation point is possible until ext4_truncate()
4221  *      finishes, we may safely do the latter, but top of branch may
4222  *      require special attention - pageout below the truncation point
4223  *      might try to populate it.
4224  *
4225  *      We atomically detach the top of branch from the tree, store the
4226  *      block number of its root in *@top, pointers to buffer_heads of
4227  *      partially truncated blocks - in @chain[].bh and pointers to
4228  *      their last elements that should not be removed - in
4229  *      @chain[].p. Return value is the pointer to last filled element
4230  *      of @chain.
4231  *
4232  *      The work left to caller to do the actual freeing of subtrees:
4233  *              a) free the subtree starting from *@top
4234  *              b) free the subtrees whose roots are stored in
4235  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
4236  *              c) free the subtrees growing from the inode past the @chain[0].
4237  *                      (no partially truncated stuff there).  */
4238
4239 static Indirect *ext4_find_shared(struct inode *inode, int depth,
4240                                   ext4_lblk_t offsets[4], Indirect chain[4],
4241                                   __le32 *top)
4242 {
4243         Indirect *partial, *p;
4244         int k, err;
4245
4246         *top = 0;
4247         /* Make k index the deepest non-null offset + 1 */
4248         for (k = depth; k > 1 && !offsets[k-1]; k--)
4249                 ;
4250         partial = ext4_get_branch(inode, k, offsets, chain, &err);
4251         /* Writer: pointers */
4252         if (!partial)
4253                 partial = chain + k-1;
4254         /*
4255          * If the branch acquired continuation since we've looked at it -
4256          * fine, it should all survive and (new) top doesn't belong to us.
4257          */
4258         if (!partial->key && *partial->p)
4259                 /* Writer: end */
4260                 goto no_top;
4261         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
4262                 ;
4263         /*
4264          * OK, we've found the last block that must survive. The rest of our
4265          * branch should be detached before unlocking. However, if that rest
4266          * of branch is all ours and does not grow immediately from the inode
4267          * it's easier to cheat and just decrement partial->p.
4268          */
4269         if (p == chain + k - 1 && p > chain) {
4270                 p->p--;
4271         } else {
4272                 *top = *p->p;
4273                 /* Nope, don't do this in ext4.  Must leave the tree intact */
4274 #if 0
4275                 *p->p = 0;
4276 #endif
4277         }
4278         /* Writer: end */
4279
4280         while (partial > p) {
4281                 brelse(partial->bh);
4282                 partial--;
4283         }
4284 no_top:
4285         return partial;
4286 }
4287
4288 /*
4289  * Zero a number of block pointers in either an inode or an indirect block.
4290  * If we restart the transaction we must again get write access to the
4291  * indirect block for further modification.
4292  *
4293  * We release `count' blocks on disk, but (last - first) may be greater
4294  * than `count' because there can be holes in there.
4295  */
4296 static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
4297                              struct buffer_head *bh,
4298                              ext4_fsblk_t block_to_free,
4299                              unsigned long count, __le32 *first,
4300                              __le32 *last)
4301 {
4302         __le32 *p;
4303         int     flags = EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_VALIDATED;
4304
4305         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
4306                 flags |= EXT4_FREE_BLOCKS_METADATA;
4307
4308         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
4309                                    count)) {
4310                 EXT4_ERROR_INODE(inode, "attempt to clear invalid "
4311                                  "blocks %llu len %lu",
4312                                  (unsigned long long) block_to_free, count);
4313                 return 1;
4314         }
4315
4316         if (try_to_extend_transaction(handle, inode)) {
4317                 if (bh) {
4318                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4319                         ext4_handle_dirty_metadata(handle, inode, bh);
4320                 }
4321                 ext4_mark_inode_dirty(handle, inode);
4322                 ext4_truncate_restart_trans(handle, inode,
4323                                             blocks_for_truncate(inode));
4324                 if (bh) {
4325                         BUFFER_TRACE(bh, "retaking write access");
4326                         ext4_journal_get_write_access(handle, bh);
4327                 }
4328         }
4329
4330         for (p = first; p < last; p++)
4331                 *p = 0;
4332
4333         ext4_free_blocks(handle, inode, 0, block_to_free, count, flags);
4334         return 0;
4335 }
4336
4337 /**
4338  * ext4_free_data - free a list of data blocks
4339  * @handle:     handle for this transaction
4340  * @inode:      inode we are dealing with
4341  * @this_bh:    indirect buffer_head which contains *@first and *@last
4342  * @first:      array of block numbers
4343  * @last:       points immediately past the end of array
4344  *
4345  * We are freeing all blocks refered from that array (numbers are stored as
4346  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
4347  *
4348  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
4349  * blocks are contiguous then releasing them at one time will only affect one
4350  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
4351  * actually use a lot of journal space.
4352  *
4353  * @this_bh will be %NULL if @first and @last point into the inode's direct
4354  * block pointers.
4355  */
4356 static void ext4_free_data(handle_t *handle, struct inode *inode,
4357                            struct buffer_head *this_bh,
4358                            __le32 *first, __le32 *last)
4359 {
4360         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
4361         unsigned long count = 0;            /* Number of blocks in the run */
4362         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
4363                                                corresponding to
4364                                                block_to_free */
4365         ext4_fsblk_t nr;                    /* Current block # */
4366         __le32 *p;                          /* Pointer into inode/ind
4367                                                for current block */
4368         int err;
4369
4370         if (this_bh) {                          /* For indirect block */
4371                 BUFFER_TRACE(this_bh, "get_write_access");
4372                 err = ext4_journal_get_write_access(handle, this_bh);
4373                 /* Important: if we can't update the indirect pointers
4374                  * to the blocks, we can't free them. */
4375                 if (err)
4376                         return;
4377         }
4378
4379         for (p = first; p < last; p++) {
4380                 nr = le32_to_cpu(*p);
4381                 if (nr) {
4382                         /* accumulate blocks to free if they're contiguous */
4383                         if (count == 0) {
4384                                 block_to_free = nr;
4385                                 block_to_free_p = p;
4386                                 count = 1;
4387                         } else if (nr == block_to_free + count) {
4388                                 count++;
4389                         } else {
4390                                 if (ext4_clear_blocks(handle, inode, this_bh,
4391                                                       block_to_free, count,
4392                                                       block_to_free_p, p))
4393                                         break;
4394                                 block_to_free = nr;
4395                                 block_to_free_p = p;
4396                                 count = 1;
4397                         }
4398                 }
4399         }
4400
4401         if (count > 0)
4402                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
4403                                   count, block_to_free_p, p);
4404
4405         if (this_bh) {
4406                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
4407
4408                 /*
4409                  * The buffer head should have an attached journal head at this
4410                  * point. However, if the data is corrupted and an indirect
4411                  * block pointed to itself, it would have been detached when
4412                  * the block was cleared. Check for this instead of OOPSing.
4413                  */
4414                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
4415                         ext4_handle_dirty_metadata(handle, inode, this_bh);
4416                 else
4417                         EXT4_ERROR_INODE(inode,
4418                                          "circular indirect block detected at "
4419                                          "block %llu",
4420                                 (unsigned long long) this_bh->b_blocknr);
4421         }
4422 }
4423
4424 /**
4425  *      ext4_free_branches - free an array of branches
4426  *      @handle: JBD handle for this transaction
4427  *      @inode: inode we are dealing with
4428  *      @parent_bh: the buffer_head which contains *@first and *@last
4429  *      @first: array of block numbers
4430  *      @last:  pointer immediately past the end of array
4431  *      @depth: depth of the branches to free
4432  *
4433  *      We are freeing all blocks refered from these branches (numbers are
4434  *      stored as little-endian 32-bit) and updating @inode->i_blocks
4435  *      appropriately.
4436  */
4437 static void ext4_free_branches(handle_t *handle, struct inode *inode,
4438                                struct buffer_head *parent_bh,
4439                                __le32 *first, __le32 *last, int depth)
4440 {
4441         ext4_fsblk_t nr;
4442         __le32 *p;
4443
4444         if (ext4_handle_is_aborted(handle))
4445                 return;
4446
4447         if (depth--) {
4448                 struct buffer_head *bh;
4449                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4450                 p = last;
4451                 while (--p >= first) {
4452                         nr = le32_to_cpu(*p);
4453                         if (!nr)
4454                                 continue;               /* A hole */
4455
4456                         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
4457                                                    nr, 1)) {
4458                                 EXT4_ERROR_INODE(inode,
4459                                                  "invalid indirect mapped "
4460                                                  "block %lu (level %d)",
4461                                                  (unsigned long) nr, depth);
4462                                 break;
4463                         }
4464
4465                         /* Go read the buffer for the next level down */
4466                         bh = sb_bread(inode->i_sb, nr);
4467
4468                         /*
4469                          * A read failure? Report error and clear slot
4470                          * (should be rare).
4471                          */
4472                         if (!bh) {
4473                                 EXT4_ERROR_INODE(inode,
4474                                                  "Read failure block=%llu",
4475                                                  (unsigned long long) nr);
4476                                 continue;
4477                         }
4478
4479                         /* This zaps the entire block.  Bottom up. */
4480                         BUFFER_TRACE(bh, "free child branches");
4481                         ext4_free_branches(handle, inode, bh,
4482                                         (__le32 *) bh->b_data,
4483                                         (__le32 *) bh->b_data + addr_per_block,
4484                                         depth);
4485
4486                         /*
4487                          * We've probably journalled the indirect block several
4488                          * times during the truncate.  But it's no longer
4489                          * needed and we now drop it from the transaction via
4490                          * jbd2_journal_revoke().
4491                          *
4492                          * That's easy if it's exclusively part of this
4493                          * transaction.  But if it's part of the committing
4494                          * transaction then jbd2_journal_forget() will simply
4495                          * brelse() it.  That means that if the underlying
4496                          * block is reallocated in ext4_get_block(),
4497                          * unmap_underlying_metadata() will find this block
4498                          * and will try to get rid of it.  damn, damn.
4499                          *
4500                          * If this block has already been committed to the
4501                          * journal, a revoke record will be written.  And
4502                          * revoke records must be emitted *before* clearing
4503                          * this block's bit in the bitmaps.
4504                          */
4505                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
4506
4507                         /*
4508                          * Everything below this this pointer has been
4509                          * released.  Now let this top-of-subtree go.
4510                          *
4511                          * We want the freeing of this indirect block to be
4512                          * atomic in the journal with the updating of the
4513                          * bitmap block which owns it.  So make some room in
4514                          * the journal.
4515                          *
4516                          * We zero the parent pointer *after* freeing its
4517                          * pointee in the bitmaps, so if extend_transaction()
4518                          * for some reason fails to put the bitmap changes and
4519                          * the release into the same transaction, recovery
4520                          * will merely complain about releasing a free block,
4521                          * rather than leaking blocks.
4522                          */
4523                         if (ext4_handle_is_aborted(handle))
4524                                 return;
4525                         if (try_to_extend_transaction(handle, inode)) {
4526                                 ext4_mark_inode_dirty(handle, inode);
4527                                 ext4_truncate_restart_trans(handle, inode,
4528                                             blocks_for_truncate(inode));
4529                         }
4530
4531                         ext4_free_blocks(handle, inode, 0, nr, 1,
4532                                          EXT4_FREE_BLOCKS_METADATA);
4533
4534                         if (parent_bh) {
4535                                 /*
4536                                  * The block which we have just freed is
4537                                  * pointed to by an indirect block: journal it
4538                                  */
4539                                 BUFFER_TRACE(parent_bh, "get_write_access");
4540                                 if (!ext4_journal_get_write_access(handle,
4541                                                                    parent_bh)){
4542                                         *p = 0;
4543                                         BUFFER_TRACE(parent_bh,
4544                                         "call ext4_handle_dirty_metadata");
4545                                         ext4_handle_dirty_metadata(handle,
4546                                                                    inode,
4547                                                                    parent_bh);
4548                                 }
4549                         }
4550                 }
4551         } else {
4552                 /* We have reached the bottom of the tree. */
4553                 BUFFER_TRACE(parent_bh, "free data blocks");
4554                 ext4_free_data(handle, inode, parent_bh, first, last);
4555         }
4556 }
4557
4558 int ext4_can_truncate(struct inode *inode)
4559 {
4560         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4561                 return 0;
4562         if (S_ISREG(inode->i_mode))
4563                 return 1;
4564         if (S_ISDIR(inode->i_mode))
4565                 return 1;
4566         if (S_ISLNK(inode->i_mode))
4567                 return !ext4_inode_is_fast_symlink(inode);
4568         return 0;
4569 }
4570
4571 /*
4572  * ext4_truncate()
4573  *
4574  * We block out ext4_get_block() block instantiations across the entire
4575  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4576  * simultaneously on behalf of the same inode.
4577  *
4578  * As we work through the truncate and commmit bits of it to the journal there
4579  * is one core, guiding principle: the file's tree must always be consistent on
4580  * disk.  We must be able to restart the truncate after a crash.
4581  *
4582  * The file's tree may be transiently inconsistent in memory (although it
4583  * probably isn't), but whenever we close off and commit a journal transaction,
4584  * the contents of (the filesystem + the journal) must be consistent and
4585  * restartable.  It's pretty simple, really: bottom up, right to left (although
4586  * left-to-right works OK too).
4587  *
4588  * Note that at recovery time, journal replay occurs *before* the restart of
4589  * truncate against the orphan inode list.
4590  *
4591  * The committed inode has the new, desired i_size (which is the same as
4592  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4593  * that this inode's truncate did not complete and it will again call
4594  * ext4_truncate() to have another go.  So there will be instantiated blocks
4595  * to the right of the truncation point in a crashed ext4 filesystem.  But
4596  * that's fine - as long as they are linked from the inode, the post-crash
4597  * ext4_truncate() run will find them and release them.
4598  */
4599 void ext4_truncate(struct inode *inode)
4600 {
4601         handle_t *handle;
4602         struct ext4_inode_info *ei = EXT4_I(inode);
4603         __le32 *i_data = ei->i_data;
4604         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4605         struct address_space *mapping = inode->i_mapping;
4606         ext4_lblk_t offsets[4];
4607         Indirect chain[4];
4608         Indirect *partial;
4609         __le32 nr = 0;
4610         int n;
4611         ext4_lblk_t last_block;
4612         unsigned blocksize = inode->i_sb->s_blocksize;
4613
4614         if (!ext4_can_truncate(inode))
4615                 return;
4616
4617         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4618
4619         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4620                 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4621
4622         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
4623                 ext4_ext_truncate(inode);
4624                 return;
4625         }
4626
4627         handle = start_transaction(inode);
4628         if (IS_ERR(handle))
4629                 return;         /* AKPM: return what? */
4630
4631         last_block = (inode->i_size + blocksize-1)
4632                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4633
4634         if (inode->i_size & (blocksize - 1))
4635                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4636                         goto out_stop;
4637
4638         n = ext4_block_to_path(inode, last_block, offsets, NULL);
4639         if (n == 0)
4640                 goto out_stop;  /* error */
4641
4642         /*
4643          * OK.  This truncate is going to happen.  We add the inode to the
4644          * orphan list, so that if this truncate spans multiple transactions,
4645          * and we crash, we will resume the truncate when the filesystem
4646          * recovers.  It also marks the inode dirty, to catch the new size.
4647          *
4648          * Implication: the file must always be in a sane, consistent
4649          * truncatable state while each transaction commits.
4650          */
4651         if (ext4_orphan_add(handle, inode))
4652                 goto out_stop;
4653
4654         /*
4655          * From here we block out all ext4_get_block() callers who want to
4656          * modify the block allocation tree.
4657          */
4658         down_write(&ei->i_data_sem);
4659
4660         ext4_discard_preallocations(inode);
4661
4662         /*
4663          * The orphan list entry will now protect us from any crash which
4664          * occurs before the truncate completes, so it is now safe to propagate
4665          * the new, shorter inode size (held for now in i_size) into the
4666          * on-disk inode. We do this via i_disksize, which is the value which
4667          * ext4 *really* writes onto the disk inode.
4668          */
4669         ei->i_disksize = inode->i_size;
4670
4671         if (n == 1) {           /* direct blocks */
4672                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4673                                i_data + EXT4_NDIR_BLOCKS);
4674                 goto do_indirects;
4675         }
4676
4677         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4678         /* Kill the top of shared branch (not detached) */
4679         if (nr) {
4680                 if (partial == chain) {
4681                         /* Shared branch grows from the inode */
4682                         ext4_free_branches(handle, inode, NULL,
4683                                            &nr, &nr+1, (chain+n-1) - partial);
4684                         *partial->p = 0;
4685                         /*
4686                          * We mark the inode dirty prior to restart,
4687                          * and prior to stop.  No need for it here.
4688                          */
4689                 } else {
4690                         /* Shared branch grows from an indirect block */
4691                         BUFFER_TRACE(partial->bh, "get_write_access");
4692                         ext4_free_branches(handle, inode, partial->bh,
4693                                         partial->p,
4694                                         partial->p+1, (chain+n-1) - partial);
4695                 }
4696         }
4697         /* Clear the ends of indirect blocks on the shared branch */
4698         while (partial > chain) {
4699                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4700                                    (__le32*)partial->bh->b_data+addr_per_block,
4701                                    (chain+n-1) - partial);
4702                 BUFFER_TRACE(partial->bh, "call brelse");
4703                 brelse(partial->bh);
4704                 partial--;
4705         }
4706 do_indirects:
4707         /* Kill the remaining (whole) subtrees */
4708         switch (offsets[0]) {
4709         default:
4710                 nr = i_data[EXT4_IND_BLOCK];
4711                 if (nr) {
4712                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4713                         i_data[EXT4_IND_BLOCK] = 0;
4714                 }
4715         case EXT4_IND_BLOCK:
4716                 nr = i_data[EXT4_DIND_BLOCK];
4717                 if (nr) {
4718                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4719                         i_data[EXT4_DIND_BLOCK] = 0;
4720                 }
4721         case EXT4_DIND_BLOCK:
4722                 nr = i_data[EXT4_TIND_BLOCK];
4723                 if (nr) {
4724                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4725                         i_data[EXT4_TIND_BLOCK] = 0;
4726                 }
4727         case EXT4_TIND_BLOCK:
4728                 ;
4729         }
4730
4731         up_write(&ei->i_data_sem);
4732         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4733         ext4_mark_inode_dirty(handle, inode);
4734
4735         /*
4736          * In a multi-transaction truncate, we only make the final transaction
4737          * synchronous
4738          */
4739         if (IS_SYNC(inode))
4740                 ext4_handle_sync(handle);
4741 out_stop:
4742         /*
4743          * If this was a simple ftruncate(), and the file will remain alive
4744          * then we need to clear up the orphan record which we created above.
4745          * However, if this was a real unlink then we were called by
4746          * ext4_delete_inode(), and we allow that function to clean up the
4747          * orphan info for us.
4748          */
4749         if (inode->i_nlink)
4750                 ext4_orphan_del(handle, inode);
4751
4752         ext4_journal_stop(handle);
4753 }
4754
4755 /*
4756  * ext4_get_inode_loc returns with an extra refcount against the inode's
4757  * underlying buffer_head on success. If 'in_mem' is true, we have all
4758  * data in memory that is needed to recreate the on-disk version of this
4759  * inode.
4760  */
4761 static int __ext4_get_inode_loc(struct inode *inode,
4762                                 struct ext4_iloc *iloc, int in_mem)
4763 {
4764         struct ext4_group_desc  *gdp;
4765         struct buffer_head      *bh;
4766         struct super_block      *sb = inode->i_sb;
4767         ext4_fsblk_t            block;
4768         int                     inodes_per_block, inode_offset;
4769
4770         iloc->bh = NULL;
4771         if (!ext4_valid_inum(sb, inode->i_ino))
4772                 return -EIO;
4773
4774         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4775         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4776         if (!gdp)
4777                 return -EIO;
4778
4779         /*
4780          * Figure out the offset within the block group inode table
4781          */
4782         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4783         inode_offset = ((inode->i_ino - 1) %
4784                         EXT4_INODES_PER_GROUP(sb));
4785         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4786         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4787
4788         bh = sb_getblk(sb, block);
4789         if (!bh) {
4790                 EXT4_ERROR_INODE(inode, "unable to read inode block - "
4791                                  "block %llu", block);
4792                 return -EIO;
4793         }
4794         if (!buffer_uptodate(bh)) {
4795                 lock_buffer(bh);
4796
4797                 /*
4798                  * If the buffer has the write error flag, we have failed
4799                  * to write out another inode in the same block.  In this
4800                  * case, we don't have to read the block because we may
4801                  * read the old inode data successfully.
4802                  */
4803                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4804                         set_buffer_uptodate(bh);
4805
4806                 if (buffer_uptodate(bh)) {
4807                         /* someone brought it uptodate while we waited */
4808                         unlock_buffer(bh);
4809                         goto has_buffer;
4810                 }
4811
4812                 /*
4813                  * If we have all information of the inode in memory and this
4814                  * is the only valid inode in the block, we need not read the
4815                  * block.
4816                  */
4817                 if (in_mem) {
4818                         struct buffer_head *bitmap_bh;
4819                         int i, start;
4820
4821                         start = inode_offset & ~(inodes_per_block - 1);
4822
4823                         /* Is the inode bitmap in cache? */
4824                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4825                         if (!bitmap_bh)
4826                                 goto make_io;
4827
4828                         /*
4829                          * If the inode bitmap isn't in cache then the
4830                          * optimisation may end up performing two reads instead
4831                          * of one, so skip it.
4832                          */
4833                         if (!buffer_uptodate(bitmap_bh)) {
4834                                 brelse(bitmap_bh);
4835                                 goto make_io;
4836                         }
4837                         for (i = start; i < start + inodes_per_block; i++) {
4838                                 if (i == inode_offset)
4839                                         continue;
4840                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4841                                         break;
4842                         }
4843                         brelse(bitmap_bh);
4844                         if (i == start + inodes_per_block) {
4845                                 /* all other inodes are free, so skip I/O */
4846                                 memset(bh->b_data, 0, bh->b_size);
4847                                 set_buffer_uptodate(bh);
4848                                 unlock_buffer(bh);
4849                                 goto has_buffer;
4850                         }
4851                 }
4852
4853 make_io:
4854                 /*
4855                  * If we need to do any I/O, try to pre-readahead extra
4856                  * blocks from the inode table.
4857                  */
4858                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4859                         ext4_fsblk_t b, end, table;
4860                         unsigned num;
4861
4862                         table = ext4_inode_table(sb, gdp);
4863                         /* s_inode_readahead_blks is always a power of 2 */
4864                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4865                         if (table > b)
4866                                 b = table;
4867                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4868                         num = EXT4_INODES_PER_GROUP(sb);
4869                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4870                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4871                                 num -= ext4_itable_unused_count(sb, gdp);
4872                         table += num / inodes_per_block;
4873                         if (end > table)
4874                                 end = table;
4875                         while (b <= end)
4876                                 sb_breadahead(sb, b++);
4877                 }
4878
4879                 /*
4880                  * There are other valid inodes in the buffer, this inode
4881                  * has in-inode xattrs, or we don't have this inode in memory.
4882                  * Read the block from disk.
4883                  */
4884                 get_bh(bh);
4885                 bh->b_end_io = end_buffer_read_sync;
4886                 submit_bh(READ_META, bh);
4887                 wait_on_buffer(bh);
4888                 if (!buffer_uptodate(bh)) {
4889                         EXT4_ERROR_INODE(inode, "unable to read inode "
4890                                          "block %llu", block);
4891                         brelse(bh);
4892                         return -EIO;
4893                 }
4894         }
4895 has_buffer:
4896         iloc->bh = bh;
4897         return 0;
4898 }
4899
4900 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4901 {
4902         /* We have all inode data except xattrs in memory here. */
4903         return __ext4_get_inode_loc(inode, iloc,
4904                 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
4905 }
4906
4907 void ext4_set_inode_flags(struct inode *inode)
4908 {
4909         unsigned int flags = EXT4_I(inode)->i_flags;
4910
4911         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4912         if (flags & EXT4_SYNC_FL)
4913                 inode->i_flags |= S_SYNC;
4914         if (flags & EXT4_APPEND_FL)
4915                 inode->i_flags |= S_APPEND;
4916         if (flags & EXT4_IMMUTABLE_FL)
4917                 inode->i_flags |= S_IMMUTABLE;
4918         if (flags & EXT4_NOATIME_FL)
4919                 inode->i_flags |= S_NOATIME;
4920         if (flags & EXT4_DIRSYNC_FL)
4921                 inode->i_flags |= S_DIRSYNC;
4922 }
4923
4924 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4925 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4926 {
4927         unsigned int vfs_fl;
4928         unsigned long old_fl, new_fl;
4929
4930         do {
4931                 vfs_fl = ei->vfs_inode.i_flags;
4932                 old_fl = ei->i_flags;
4933                 new_fl = old_fl & ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4934                                 EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|
4935                                 EXT4_DIRSYNC_FL);
4936                 if (vfs_fl & S_SYNC)
4937                         new_fl |= EXT4_SYNC_FL;
4938                 if (vfs_fl & S_APPEND)
4939                         new_fl |= EXT4_APPEND_FL;
4940                 if (vfs_fl & S_IMMUTABLE)
4941                         new_fl |= EXT4_IMMUTABLE_FL;
4942                 if (vfs_fl & S_NOATIME)
4943                         new_fl |= EXT4_NOATIME_FL;
4944                 if (vfs_fl & S_DIRSYNC)
4945                         new_fl |= EXT4_DIRSYNC_FL;
4946         } while (cmpxchg(&ei->i_flags, old_fl, new_fl) != old_fl);
4947 }
4948
4949 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4950                                   struct ext4_inode_info *ei)
4951 {
4952         blkcnt_t i_blocks ;
4953         struct inode *inode = &(ei->vfs_inode);
4954         struct super_block *sb = inode->i_sb;
4955
4956         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4957                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4958                 /* we are using combined 48 bit field */
4959                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4960                                         le32_to_cpu(raw_inode->i_blocks_lo);
4961                 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
4962                         /* i_blocks represent file system block size */
4963                         return i_blocks  << (inode->i_blkbits - 9);
4964                 } else {
4965                         return i_blocks;
4966                 }
4967         } else {
4968                 return le32_to_cpu(raw_inode->i_blocks_lo);
4969         }
4970 }
4971
4972 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4973 {
4974         struct ext4_iloc iloc;
4975         struct ext4_inode *raw_inode;
4976         struct ext4_inode_info *ei;
4977         struct inode *inode;
4978         journal_t *journal = EXT4_SB(sb)->s_journal;
4979         long ret;
4980         int block;
4981
4982         inode = iget_locked(sb, ino);
4983         if (!inode)
4984                 return ERR_PTR(-ENOMEM);
4985         if (!(inode->i_state & I_NEW))
4986                 return inode;
4987
4988         ei = EXT4_I(inode);
4989         iloc.bh = 0;
4990
4991         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4992         if (ret < 0)
4993                 goto bad_inode;
4994         raw_inode = ext4_raw_inode(&iloc);
4995         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4996         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4997         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4998         if (!(test_opt(inode->i_sb, NO_UID32))) {
4999                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
5000                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
5001         }
5002         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
5003
5004         ei->i_state_flags = 0;
5005         ei->i_dir_start_lookup = 0;
5006         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
5007         /* We now have enough fields to check if the inode was active or not.
5008          * This is needed because nfsd might try to access dead inodes
5009          * the test is that same one that e2fsck uses
5010          * NeilBrown 1999oct15
5011          */
5012         if (inode->i_nlink == 0) {
5013                 if (inode->i_mode == 0 ||
5014                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
5015                         /* this inode is deleted */
5016                         ret = -ESTALE;
5017                         goto bad_inode;
5018                 }
5019                 /* The only unlinked inodes we let through here have
5020                  * valid i_mode and are being read by the orphan
5021                  * recovery code: that's fine, we're about to complete
5022                  * the process of deleting those. */
5023         }
5024         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
5025         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
5026         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
5027         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
5028                 ei->i_file_acl |=
5029                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
5030         inode->i_size = ext4_isize(raw_inode);
5031         ei->i_disksize = inode->i_size;
5032 #ifdef CONFIG_QUOTA
5033         ei->i_reserved_quota = 0;
5034 #endif
5035         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
5036         ei->i_block_group = iloc.block_group;
5037         ei->i_last_alloc_group = ~0;
5038         /*
5039          * NOTE! The in-memory inode i_data array is in little-endian order
5040          * even on big-endian machines: we do NOT byteswap the block numbers!
5041          */
5042         for (block = 0; block < EXT4_N_BLOCKS; block++)
5043                 ei->i_data[block] = raw_inode->i_block[block];
5044         INIT_LIST_HEAD(&ei->i_orphan);
5045
5046         /*
5047          * Set transaction id's of transactions that have to be committed
5048          * to finish f[data]sync. We set them to currently running transaction
5049          * as we cannot be sure that the inode or some of its metadata isn't
5050          * part of the transaction - the inode could have been reclaimed and
5051          * now it is reread from disk.
5052          */
5053         if (journal) {
5054                 transaction_t *transaction;
5055                 tid_t tid;
5056
5057                 spin_lock(&journal->j_state_lock);
5058                 if (journal->j_running_transaction)
5059                         transaction = journal->j_running_transaction;
5060                 else
5061                         transaction = journal->j_committing_transaction;
5062                 if (transaction)
5063                         tid = transaction->t_tid;
5064                 else
5065                         tid = journal->j_commit_sequence;
5066                 spin_unlock(&journal->j_state_lock);
5067                 ei->i_sync_tid = tid;
5068                 ei->i_datasync_tid = tid;
5069         }
5070
5071         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5072                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
5073                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
5074                     EXT4_INODE_SIZE(inode->i_sb)) {
5075                         ret = -EIO;
5076                         goto bad_inode;
5077                 }
5078                 if (ei->i_extra_isize == 0) {
5079                         /* The extra space is currently unused. Use it. */
5080                         ei->i_extra_isize = sizeof(struct ext4_inode) -
5081                                             EXT4_GOOD_OLD_INODE_SIZE;
5082                 } else {
5083                         __le32 *magic = (void *)raw_inode +
5084                                         EXT4_GOOD_OLD_INODE_SIZE +
5085                                         ei->i_extra_isize;
5086                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
5087                                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
5088                 }
5089         } else
5090                 ei->i_extra_isize = 0;
5091
5092         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
5093         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
5094         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
5095         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
5096
5097         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
5098         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5099                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5100                         inode->i_version |=
5101                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
5102         }
5103
5104         ret = 0;
5105         if (ei->i_file_acl &&
5106             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
5107                 EXT4_ERROR_INODE(inode, "bad extended attribute block %llu",
5108                                  ei->i_file_acl);
5109                 ret = -EIO;
5110                 goto bad_inode;
5111         } else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5112                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5113                     (S_ISLNK(inode->i_mode) &&
5114                      !ext4_inode_is_fast_symlink(inode)))
5115                         /* Validate extent which is part of inode */
5116                         ret = ext4_ext_check_inode(inode);
5117         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5118                    (S_ISLNK(inode->i_mode) &&
5119                     !ext4_inode_is_fast_symlink(inode))) {
5120                 /* Validate block references which are part of inode */
5121                 ret = ext4_check_inode_blockref(inode);
5122         }
5123         if (ret)
5124                 goto bad_inode;
5125
5126         if (S_ISREG(inode->i_mode)) {
5127                 inode->i_op = &ext4_file_inode_operations;
5128                 inode->i_fop = &ext4_file_operations;
5129                 ext4_set_aops(inode);
5130         } else if (S_ISDIR(inode->i_mode)) {
5131                 inode->i_op = &ext4_dir_inode_operations;
5132                 inode->i_fop = &ext4_dir_operations;
5133         } else if (S_ISLNK(inode->i_mode)) {
5134                 if (ext4_inode_is_fast_symlink(inode)) {
5135                         inode->i_op = &ext4_fast_symlink_inode_operations;
5136                         nd_terminate_link(ei->i_data, inode->i_size,
5137                                 sizeof(ei->i_data) - 1);
5138                 } else {
5139                         inode->i_op = &ext4_symlink_inode_operations;
5140                         ext4_set_aops(inode);
5141                 }
5142         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
5143               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
5144                 inode->i_op = &ext4_special_inode_operations;
5145                 if (raw_inode->i_block[0])
5146                         init_special_inode(inode, inode->i_mode,
5147                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
5148                 else
5149                         init_special_inode(inode, inode->i_mode,
5150                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
5151         } else {
5152                 ret = -EIO;
5153                 EXT4_ERROR_INODE(inode, "bogus i_mode (%o)", inode->i_mode);
5154                 goto bad_inode;
5155         }
5156         brelse(iloc.bh);
5157         ext4_set_inode_flags(inode);
5158         unlock_new_inode(inode);
5159         return inode;
5160
5161 bad_inode:
5162         brelse(iloc.bh);
5163         iget_failed(inode);
5164         return ERR_PTR(ret);
5165 }
5166
5167 static int ext4_inode_blocks_set(handle_t *handle,
5168                                 struct ext4_inode *raw_inode,
5169                                 struct ext4_inode_info *ei)
5170 {
5171         struct inode *inode = &(ei->vfs_inode);
5172         u64 i_blocks = inode->i_blocks;
5173         struct super_block *sb = inode->i_sb;
5174
5175         if (i_blocks <= ~0U) {
5176                 /*
5177                  * i_blocks can be represnted in a 32 bit variable
5178                  * as multiple of 512 bytes
5179                  */
5180                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5181                 raw_inode->i_blocks_high = 0;
5182                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5183                 return 0;
5184         }
5185         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
5186                 return -EFBIG;
5187
5188         if (i_blocks <= 0xffffffffffffULL) {
5189                 /*
5190                  * i_blocks can be represented in a 48 bit variable
5191                  * as multiple of 512 bytes
5192                  */
5193                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5194                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5195                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5196         } else {
5197                 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5198                 /* i_block is stored in file system block size */
5199                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
5200                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5201                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5202         }
5203         return 0;
5204 }
5205
5206 /*
5207  * Post the struct inode info into an on-disk inode location in the
5208  * buffer-cache.  This gobbles the caller's reference to the
5209  * buffer_head in the inode location struct.
5210  *
5211  * The caller must have write access to iloc->bh.
5212  */
5213 static int ext4_do_update_inode(handle_t *handle,
5214                                 struct inode *inode,
5215                                 struct ext4_iloc *iloc)
5216 {
5217         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5218         struct ext4_inode_info *ei = EXT4_I(inode);
5219         struct buffer_head *bh = iloc->bh;
5220         int err = 0, rc, block;
5221
5222         /* For fields not not tracking in the in-memory inode,
5223          * initialise them to zero for new inodes. */
5224         if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5225                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5226
5227         ext4_get_inode_flags(ei);
5228         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5229         if (!(test_opt(inode->i_sb, NO_UID32))) {
5230                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
5231                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
5232 /*
5233  * Fix up interoperability with old kernels. Otherwise, old inodes get
5234  * re-used with the upper 16 bits of the uid/gid intact
5235  */
5236                 if (!ei->i_dtime) {
5237                         raw_inode->i_uid_high =
5238                                 cpu_to_le16(high_16_bits(inode->i_uid));
5239                         raw_inode->i_gid_high =
5240                                 cpu_to_le16(high_16_bits(inode->i_gid));
5241                 } else {
5242                         raw_inode->i_uid_high = 0;
5243                         raw_inode->i_gid_high = 0;
5244                 }
5245         } else {
5246                 raw_inode->i_uid_low =
5247                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
5248                 raw_inode->i_gid_low =
5249                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
5250                 raw_inode->i_uid_high = 0;
5251                 raw_inode->i_gid_high = 0;
5252         }
5253         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5254
5255         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5256         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5257         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5258         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5259
5260         if (ext4_inode_blocks_set(handle, raw_inode, ei))
5261                 goto out_brelse;
5262         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5263         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
5264         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
5265             cpu_to_le32(EXT4_OS_HURD))
5266                 raw_inode->i_file_acl_high =
5267                         cpu_to_le16(ei->i_file_acl >> 32);
5268         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5269         ext4_isize_set(raw_inode, ei->i_disksize);
5270         if (ei->i_disksize > 0x7fffffffULL) {
5271                 struct super_block *sb = inode->i_sb;
5272                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
5273                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
5274                                 EXT4_SB(sb)->s_es->s_rev_level ==
5275                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
5276                         /* If this is the first large file
5277                          * created, add a flag to the superblock.
5278                          */
5279                         err = ext4_journal_get_write_access(handle,
5280                                         EXT4_SB(sb)->s_sbh);
5281                         if (err)
5282                                 goto out_brelse;
5283                         ext4_update_dynamic_rev(sb);
5284                         EXT4_SET_RO_COMPAT_FEATURE(sb,
5285                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
5286                         sb->s_dirt = 1;
5287                         ext4_handle_sync(handle);
5288                         err = ext4_handle_dirty_metadata(handle, NULL,
5289                                         EXT4_SB(sb)->s_sbh);
5290                 }
5291         }
5292         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5293         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5294                 if (old_valid_dev(inode->i_rdev)) {
5295                         raw_inode->i_block[0] =
5296                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
5297                         raw_inode->i_block[1] = 0;
5298                 } else {
5299                         raw_inode->i_block[0] = 0;
5300                         raw_inode->i_block[1] =
5301                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
5302                         raw_inode->i_block[2] = 0;
5303                 }
5304         } else
5305                 for (block = 0; block < EXT4_N_BLOCKS; block++)
5306                         raw_inode->i_block[block] = ei->i_data[block];
5307
5308         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
5309         if (ei->i_extra_isize) {
5310                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5311                         raw_inode->i_version_hi =
5312                         cpu_to_le32(inode->i_version >> 32);
5313                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
5314         }
5315
5316         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5317         rc = ext4_handle_dirty_metadata(handle, NULL, bh);
5318         if (!err)
5319                 err = rc;
5320         ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5321
5322         ext4_update_inode_fsync_trans(handle, inode, 0);
5323 out_brelse:
5324         brelse(bh);
5325         ext4_std_error(inode->i_sb, err);
5326         return err;
5327 }
5328
5329 /*
5330  * ext4_write_inode()
5331  *
5332  * We are called from a few places:
5333  *
5334  * - Within generic_file_write() for O_SYNC files.
5335  *   Here, there will be no transaction running. We wait for any running
5336  *   trasnaction to commit.
5337  *
5338  * - Within sys_sync(), kupdate and such.
5339  *   We wait on commit, if tol to.
5340  *
5341  * - Within prune_icache() (PF_MEMALLOC == true)
5342  *   Here we simply return.  We can't afford to block kswapd on the
5343  *   journal commit.
5344  *
5345  * In all cases it is actually safe for us to return without doing anything,
5346  * because the inode has been copied into a raw inode buffer in
5347  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
5348  * knfsd.
5349  *
5350  * Note that we are absolutely dependent upon all inode dirtiers doing the
5351  * right thing: they *must* call mark_inode_dirty() after dirtying info in
5352  * which we are interested.
5353  *
5354  * It would be a bug for them to not do this.  The code:
5355  *
5356  *      mark_inode_dirty(inode)
5357  *      stuff();
5358  *      inode->i_size = expr;
5359  *
5360  * is in error because a kswapd-driven write_inode() could occur while
5361  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
5362  * will no longer be on the superblock's dirty inode list.
5363  */
5364 int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5365 {
5366         int err;
5367
5368         if (current->flags & PF_MEMALLOC)
5369                 return 0;
5370
5371         if (EXT4_SB(inode->i_sb)->s_journal) {
5372                 if (ext4_journal_current_handle()) {
5373                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5374                         dump_stack();
5375                         return -EIO;
5376                 }
5377
5378                 if (wbc->sync_mode != WB_SYNC_ALL)
5379                         return 0;
5380
5381                 err = ext4_force_commit(inode->i_sb);
5382         } else {
5383                 struct ext4_iloc iloc;
5384
5385                 err = __ext4_get_inode_loc(inode, &iloc, 0);
5386                 if (err)
5387                         return err;
5388                 if (wbc->sync_mode == WB_SYNC_ALL)
5389                         sync_dirty_buffer(iloc.bh);
5390                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5391                         EXT4_ERROR_INODE(inode,
5392                                 "IO error syncing inode (block=%llu)",
5393                                 (unsigned long long) iloc.bh->b_blocknr);
5394                         err = -EIO;
5395                 }
5396                 brelse(iloc.bh);
5397         }
5398         return err;
5399 }
5400
5401 /*
5402  * ext4_setattr()
5403  *
5404  * Called from notify_change.
5405  *
5406  * We want to trap VFS attempts to truncate the file as soon as
5407  * possible.  In particular, we want to make sure that when the VFS
5408  * shrinks i_size, we put the inode on the orphan list and modify
5409  * i_disksize immediately, so that during the subsequent flushing of
5410  * dirty pages and freeing of disk blocks, we can guarantee that any
5411  * commit will leave the blocks being flushed in an unused state on
5412  * disk.  (On recovery, the inode will get truncated and the blocks will
5413  * be freed, so we have a strong guarantee that no future commit will
5414  * leave these blocks visible to the user.)
5415  *
5416  * Another thing we have to assure is that if we are in ordered mode
5417  * and inode is still attached to the committing transaction, we must
5418  * we start writeout of all the dirty pages which are being truncated.
5419  * This way we are sure that all the data written in the previous
5420  * transaction are already on disk (truncate waits for pages under
5421  * writeback).
5422  *
5423  * Called with inode->i_mutex down.
5424  */
5425 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5426 {
5427         struct inode *inode = dentry->d_inode;
5428         int error, rc = 0;
5429         const unsigned int ia_valid = attr->ia_valid;
5430
5431         error = inode_change_ok(inode, attr);
5432         if (error)
5433                 return error;
5434
5435         if (is_quota_modification(inode, attr))
5436                 dquot_initialize(inode);
5437         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
5438                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
5439                 handle_t *handle;
5440
5441                 /* (user+group)*(old+new) structure, inode write (sb,
5442                  * inode block, ? - but truncate inode update has it) */
5443                 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
5444                                         EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
5445                 if (IS_ERR(handle)) {
5446                         error = PTR_ERR(handle);
5447                         goto err_out;
5448                 }
5449                 error = dquot_transfer(inode, attr);
5450                 if (error) {
5451                         ext4_journal_stop(handle);
5452                         return error;
5453                 }
5454                 /* Update corresponding info in inode so that everything is in
5455                  * one transaction */
5456                 if (attr->ia_valid & ATTR_UID)
5457                         inode->i_uid = attr->ia_uid;
5458                 if (attr->ia_valid & ATTR_GID)
5459                         inode->i_gid = attr->ia_gid;
5460                 error = ext4_mark_inode_dirty(handle, inode);
5461                 ext4_journal_stop(handle);
5462         }
5463
5464         if (attr->ia_valid & ATTR_SIZE) {
5465                 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
5466                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5467
5468                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
5469                                 error = -EFBIG;
5470                                 goto err_out;
5471                         }
5472                 }
5473         }
5474
5475         if (S_ISREG(inode->i_mode) &&
5476             attr->ia_valid & ATTR_SIZE &&
5477             (attr->ia_size < inode->i_size ||
5478              (ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS)))) {
5479                 handle_t *handle;
5480
5481                 handle = ext4_journal_start(inode, 3);
5482                 if (IS_ERR(handle)) {
5483                         error = PTR_ERR(handle);
5484                         goto err_out;
5485                 }
5486
5487                 error = ext4_orphan_add(handle, inode);
5488                 EXT4_I(inode)->i_disksize = attr->ia_size;
5489                 rc = ext4_mark_inode_dirty(handle, inode);
5490                 if (!error)
5491                         error = rc;
5492                 ext4_journal_stop(handle);
5493
5494                 if (ext4_should_order_data(inode)) {
5495                         error = ext4_begin_ordered_truncate(inode,
5496                                                             attr->ia_size);
5497                         if (error) {
5498                                 /* Do as much error cleanup as possible */
5499                                 handle = ext4_journal_start(inode, 3);
5500                                 if (IS_ERR(handle)) {
5501                                         ext4_orphan_del(NULL, inode);
5502                                         goto err_out;
5503                                 }
5504                                 ext4_orphan_del(handle, inode);
5505                                 ext4_journal_stop(handle);
5506                                 goto err_out;
5507                         }
5508                 }
5509                 /* ext4_truncate will clear the flag */
5510                 if ((ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS)))
5511                         ext4_truncate(inode);
5512         }
5513
5514         rc = inode_setattr(inode, attr);
5515
5516         /* If inode_setattr's call to ext4_truncate failed to get a
5517          * transaction handle at all, we need to clean up the in-core
5518          * orphan list manually. */
5519         if (inode->i_nlink)
5520                 ext4_orphan_del(NULL, inode);
5521
5522         if (!rc && (ia_valid & ATTR_MODE))
5523                 rc = ext4_acl_chmod(inode);
5524
5525 err_out:
5526         ext4_std_error(inode->i_sb, error);
5527         if (!error)
5528                 error = rc;
5529         return error;
5530 }
5531
5532 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
5533                  struct kstat *stat)
5534 {
5535         struct inode *inode;
5536         unsigned long delalloc_blocks;
5537
5538         inode