ext4,jbd2: convert tracepoints to use major/minor numbers
[linux-2.6.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/jbd2.h>
29 #include <linux/highuid.h>
30 #include <linux/pagemap.h>
31 #include <linux/quotaops.h>
32 #include <linux/string.h>
33 #include <linux/buffer_head.h>
34 #include <linux/writeback.h>
35 #include <linux/pagevec.h>
36 #include <linux/mpage.h>
37 #include <linux/namei.h>
38 #include <linux/uio.h>
39 #include <linux/bio.h>
40 #include <linux/workqueue.h>
41 #include <linux/kernel.h>
42 #include <linux/slab.h>
43
44 #include "ext4_jbd2.h"
45 #include "xattr.h"
46 #include "acl.h"
47 #include "ext4_extents.h"
48
49 #include <trace/events/ext4.h>
50
51 #define MPAGE_DA_EXTENT_TAIL 0x01
52
53 static inline int ext4_begin_ordered_truncate(struct inode *inode,
54                                               loff_t new_size)
55 {
56         return jbd2_journal_begin_ordered_truncate(
57                                         EXT4_SB(inode->i_sb)->s_journal,
58                                         &EXT4_I(inode)->jinode,
59                                         new_size);
60 }
61
62 static void ext4_invalidatepage(struct page *page, unsigned long offset);
63 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
64                                    struct buffer_head *bh_result, int create);
65 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
66 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
67 static int __ext4_journalled_writepage(struct page *page, unsigned int len);
68 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
69
70 /*
71  * Test whether an inode is a fast symlink.
72  */
73 static int ext4_inode_is_fast_symlink(struct inode *inode)
74 {
75         int ea_blocks = EXT4_I(inode)->i_file_acl ?
76                 (inode->i_sb->s_blocksize >> 9) : 0;
77
78         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
79 }
80
81 /*
82  * Work out how many blocks we need to proceed with the next chunk of a
83  * truncate transaction.
84  */
85 static unsigned long blocks_for_truncate(struct inode *inode)
86 {
87         ext4_lblk_t needed;
88
89         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
90
91         /* Give ourselves just enough room to cope with inodes in which
92          * i_blocks is corrupt: we've seen disk corruptions in the past
93          * which resulted in random data in an inode which looked enough
94          * like a regular file for ext4 to try to delete it.  Things
95          * will go a bit crazy if that happens, but at least we should
96          * try not to panic the whole kernel. */
97         if (needed < 2)
98                 needed = 2;
99
100         /* But we need to bound the transaction so we don't overflow the
101          * journal. */
102         if (needed > EXT4_MAX_TRANS_DATA)
103                 needed = EXT4_MAX_TRANS_DATA;
104
105         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
106 }
107
108 /*
109  * Truncate transactions can be complex and absolutely huge.  So we need to
110  * be able to restart the transaction at a conventient checkpoint to make
111  * sure we don't overflow the journal.
112  *
113  * start_transaction gets us a new handle for a truncate transaction,
114  * and extend_transaction tries to extend the existing one a bit.  If
115  * extend fails, we need to propagate the failure up and restart the
116  * transaction in the top-level truncate loop. --sct
117  */
118 static handle_t *start_transaction(struct inode *inode)
119 {
120         handle_t *result;
121
122         result = ext4_journal_start(inode, blocks_for_truncate(inode));
123         if (!IS_ERR(result))
124                 return result;
125
126         ext4_std_error(inode->i_sb, PTR_ERR(result));
127         return result;
128 }
129
130 /*
131  * Try to extend this transaction for the purposes of truncation.
132  *
133  * Returns 0 if we managed to create more room.  If we can't create more
134  * room, and the transaction must be restarted we return 1.
135  */
136 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
137 {
138         if (!ext4_handle_valid(handle))
139                 return 0;
140         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
141                 return 0;
142         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
143                 return 0;
144         return 1;
145 }
146
147 /*
148  * Restart the transaction associated with *handle.  This does a commit,
149  * so before we call here everything must be consistently dirtied against
150  * this transaction.
151  */
152 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
153                                  int nblocks)
154 {
155         int ret;
156
157         /*
158          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
159          * moment, get_block can be called only for blocks inside i_size since
160          * page cache has been already dropped and writes are blocked by
161          * i_mutex. So we can safely drop the i_data_sem here.
162          */
163         BUG_ON(EXT4_JOURNAL(inode) == NULL);
164         jbd_debug(2, "restarting handle %p\n", handle);
165         up_write(&EXT4_I(inode)->i_data_sem);
166         ret = ext4_journal_restart(handle, blocks_for_truncate(inode));
167         down_write(&EXT4_I(inode)->i_data_sem);
168         ext4_discard_preallocations(inode);
169
170         return ret;
171 }
172
173 /*
174  * Called at the last iput() if i_nlink is zero.
175  */
176 void ext4_evict_inode(struct inode *inode)
177 {
178         handle_t *handle;
179         int err;
180
181         if (inode->i_nlink) {
182                 truncate_inode_pages(&inode->i_data, 0);
183                 goto no_delete;
184         }
185
186         if (!is_bad_inode(inode))
187                 dquot_initialize(inode);
188
189         if (ext4_should_order_data(inode))
190                 ext4_begin_ordered_truncate(inode, 0);
191         truncate_inode_pages(&inode->i_data, 0);
192
193         if (is_bad_inode(inode))
194                 goto no_delete;
195
196         handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
197         if (IS_ERR(handle)) {
198                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
199                 /*
200                  * If we're going to skip the normal cleanup, we still need to
201                  * make sure that the in-core orphan linked list is properly
202                  * cleaned up.
203                  */
204                 ext4_orphan_del(NULL, inode);
205                 goto no_delete;
206         }
207
208         if (IS_SYNC(inode))
209                 ext4_handle_sync(handle);
210         inode->i_size = 0;
211         err = ext4_mark_inode_dirty(handle, inode);
212         if (err) {
213                 ext4_warning(inode->i_sb,
214                              "couldn't mark inode dirty (err %d)", err);
215                 goto stop_handle;
216         }
217         if (inode->i_blocks)
218                 ext4_truncate(inode);
219
220         /*
221          * ext4_ext_truncate() doesn't reserve any slop when it
222          * restarts journal transactions; therefore there may not be
223          * enough credits left in the handle to remove the inode from
224          * the orphan list and set the dtime field.
225          */
226         if (!ext4_handle_has_enough_credits(handle, 3)) {
227                 err = ext4_journal_extend(handle, 3);
228                 if (err > 0)
229                         err = ext4_journal_restart(handle, 3);
230                 if (err != 0) {
231                         ext4_warning(inode->i_sb,
232                                      "couldn't extend journal (err %d)", err);
233                 stop_handle:
234                         ext4_journal_stop(handle);
235                         ext4_orphan_del(NULL, inode);
236                         goto no_delete;
237                 }
238         }
239
240         /*
241          * Kill off the orphan record which ext4_truncate created.
242          * AKPM: I think this can be inside the above `if'.
243          * Note that ext4_orphan_del() has to be able to cope with the
244          * deletion of a non-existent orphan - this is because we don't
245          * know if ext4_truncate() actually created an orphan record.
246          * (Well, we could do this if we need to, but heck - it works)
247          */
248         ext4_orphan_del(handle, inode);
249         EXT4_I(inode)->i_dtime  = get_seconds();
250
251         /*
252          * One subtle ordering requirement: if anything has gone wrong
253          * (transaction abort, IO errors, whatever), then we can still
254          * do these next steps (the fs will already have been marked as
255          * having errors), but we can't free the inode if the mark_dirty
256          * fails.
257          */
258         if (ext4_mark_inode_dirty(handle, inode))
259                 /* If that failed, just do the required in-core inode clear. */
260                 ext4_clear_inode(inode);
261         else
262                 ext4_free_inode(handle, inode);
263         ext4_journal_stop(handle);
264         return;
265 no_delete:
266         ext4_clear_inode(inode);        /* We must guarantee clearing of inode... */
267 }
268
269 typedef struct {
270         __le32  *p;
271         __le32  key;
272         struct buffer_head *bh;
273 } Indirect;
274
275 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
276 {
277         p->key = *(p->p = v);
278         p->bh = bh;
279 }
280
281 /**
282  *      ext4_block_to_path - parse the block number into array of offsets
283  *      @inode: inode in question (we are only interested in its superblock)
284  *      @i_block: block number to be parsed
285  *      @offsets: array to store the offsets in
286  *      @boundary: set this non-zero if the referred-to block is likely to be
287  *             followed (on disk) by an indirect block.
288  *
289  *      To store the locations of file's data ext4 uses a data structure common
290  *      for UNIX filesystems - tree of pointers anchored in the inode, with
291  *      data blocks at leaves and indirect blocks in intermediate nodes.
292  *      This function translates the block number into path in that tree -
293  *      return value is the path length and @offsets[n] is the offset of
294  *      pointer to (n+1)th node in the nth one. If @block is out of range
295  *      (negative or too large) warning is printed and zero returned.
296  *
297  *      Note: function doesn't find node addresses, so no IO is needed. All
298  *      we need to know is the capacity of indirect blocks (taken from the
299  *      inode->i_sb).
300  */
301
302 /*
303  * Portability note: the last comparison (check that we fit into triple
304  * indirect block) is spelled differently, because otherwise on an
305  * architecture with 32-bit longs and 8Kb pages we might get into trouble
306  * if our filesystem had 8Kb blocks. We might use long long, but that would
307  * kill us on x86. Oh, well, at least the sign propagation does not matter -
308  * i_block would have to be negative in the very beginning, so we would not
309  * get there at all.
310  */
311
312 static int ext4_block_to_path(struct inode *inode,
313                               ext4_lblk_t i_block,
314                               ext4_lblk_t offsets[4], int *boundary)
315 {
316         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
317         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
318         const long direct_blocks = EXT4_NDIR_BLOCKS,
319                 indirect_blocks = ptrs,
320                 double_blocks = (1 << (ptrs_bits * 2));
321         int n = 0;
322         int final = 0;
323
324         if (i_block < direct_blocks) {
325                 offsets[n++] = i_block;
326                 final = direct_blocks;
327         } else if ((i_block -= direct_blocks) < indirect_blocks) {
328                 offsets[n++] = EXT4_IND_BLOCK;
329                 offsets[n++] = i_block;
330                 final = ptrs;
331         } else if ((i_block -= indirect_blocks) < double_blocks) {
332                 offsets[n++] = EXT4_DIND_BLOCK;
333                 offsets[n++] = i_block >> ptrs_bits;
334                 offsets[n++] = i_block & (ptrs - 1);
335                 final = ptrs;
336         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
337                 offsets[n++] = EXT4_TIND_BLOCK;
338                 offsets[n++] = i_block >> (ptrs_bits * 2);
339                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
340                 offsets[n++] = i_block & (ptrs - 1);
341                 final = ptrs;
342         } else {
343                 ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
344                              i_block + direct_blocks +
345                              indirect_blocks + double_blocks, inode->i_ino);
346         }
347         if (boundary)
348                 *boundary = final - 1 - (i_block & (ptrs - 1));
349         return n;
350 }
351
352 static int __ext4_check_blockref(const char *function, unsigned int line,
353                                  struct inode *inode,
354                                  __le32 *p, unsigned int max)
355 {
356         struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
357         __le32 *bref = p;
358         unsigned int blk;
359
360         while (bref < p+max) {
361                 blk = le32_to_cpu(*bref++);
362                 if (blk &&
363                     unlikely(!ext4_data_block_valid(EXT4_SB(inode->i_sb),
364                                                     blk, 1))) {
365                         es->s_last_error_block = cpu_to_le64(blk);
366                         ext4_error_inode(inode, function, line, blk,
367                                          "invalid block");
368                         return -EIO;
369                 }
370         }
371         return 0;
372 }
373
374
375 #define ext4_check_indirect_blockref(inode, bh)                         \
376         __ext4_check_blockref(__func__, __LINE__, inode,                \
377                               (__le32 *)(bh)->b_data,                   \
378                               EXT4_ADDR_PER_BLOCK((inode)->i_sb))
379
380 #define ext4_check_inode_blockref(inode)                                \
381         __ext4_check_blockref(__func__, __LINE__, inode,                \
382                               EXT4_I(inode)->i_data,                    \
383                               EXT4_NDIR_BLOCKS)
384
385 /**
386  *      ext4_get_branch - read the chain of indirect blocks leading to data
387  *      @inode: inode in question
388  *      @depth: depth of the chain (1 - direct pointer, etc.)
389  *      @offsets: offsets of pointers in inode/indirect blocks
390  *      @chain: place to store the result
391  *      @err: here we store the error value
392  *
393  *      Function fills the array of triples <key, p, bh> and returns %NULL
394  *      if everything went OK or the pointer to the last filled triple
395  *      (incomplete one) otherwise. Upon the return chain[i].key contains
396  *      the number of (i+1)-th block in the chain (as it is stored in memory,
397  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
398  *      number (it points into struct inode for i==0 and into the bh->b_data
399  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
400  *      block for i>0 and NULL for i==0. In other words, it holds the block
401  *      numbers of the chain, addresses they were taken from (and where we can
402  *      verify that chain did not change) and buffer_heads hosting these
403  *      numbers.
404  *
405  *      Function stops when it stumbles upon zero pointer (absent block)
406  *              (pointer to last triple returned, *@err == 0)
407  *      or when it gets an IO error reading an indirect block
408  *              (ditto, *@err == -EIO)
409  *      or when it reads all @depth-1 indirect blocks successfully and finds
410  *      the whole chain, all way to the data (returns %NULL, *err == 0).
411  *
412  *      Need to be called with
413  *      down_read(&EXT4_I(inode)->i_data_sem)
414  */
415 static Indirect *ext4_get_branch(struct inode *inode, int depth,
416                                  ext4_lblk_t  *offsets,
417                                  Indirect chain[4], int *err)
418 {
419         struct super_block *sb = inode->i_sb;
420         Indirect *p = chain;
421         struct buffer_head *bh;
422
423         *err = 0;
424         /* i_data is not going away, no lock needed */
425         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
426         if (!p->key)
427                 goto no_block;
428         while (--depth) {
429                 bh = sb_getblk(sb, le32_to_cpu(p->key));
430                 if (unlikely(!bh))
431                         goto failure;
432
433                 if (!bh_uptodate_or_lock(bh)) {
434                         if (bh_submit_read(bh) < 0) {
435                                 put_bh(bh);
436                                 goto failure;
437                         }
438                         /* validate block references */
439                         if (ext4_check_indirect_blockref(inode, bh)) {
440                                 put_bh(bh);
441                                 goto failure;
442                         }
443                 }
444
445                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
446                 /* Reader: end */
447                 if (!p->key)
448                         goto no_block;
449         }
450         return NULL;
451
452 failure:
453         *err = -EIO;
454 no_block:
455         return p;
456 }
457
458 /**
459  *      ext4_find_near - find a place for allocation with sufficient locality
460  *      @inode: owner
461  *      @ind: descriptor of indirect block.
462  *
463  *      This function returns the preferred place for block allocation.
464  *      It is used when heuristic for sequential allocation fails.
465  *      Rules are:
466  *        + if there is a block to the left of our position - allocate near it.
467  *        + if pointer will live in indirect block - allocate near that block.
468  *        + if pointer will live in inode - allocate in the same
469  *          cylinder group.
470  *
471  * In the latter case we colour the starting block by the callers PID to
472  * prevent it from clashing with concurrent allocations for a different inode
473  * in the same block group.   The PID is used here so that functionally related
474  * files will be close-by on-disk.
475  *
476  *      Caller must make sure that @ind is valid and will stay that way.
477  */
478 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
479 {
480         struct ext4_inode_info *ei = EXT4_I(inode);
481         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
482         __le32 *p;
483         ext4_fsblk_t bg_start;
484         ext4_fsblk_t last_block;
485         ext4_grpblk_t colour;
486         ext4_group_t block_group;
487         int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
488
489         /* Try to find previous block */
490         for (p = ind->p - 1; p >= start; p--) {
491                 if (*p)
492                         return le32_to_cpu(*p);
493         }
494
495         /* No such thing, so let's try location of indirect block */
496         if (ind->bh)
497                 return ind->bh->b_blocknr;
498
499         /*
500          * It is going to be referred to from the inode itself? OK, just put it
501          * into the same cylinder group then.
502          */
503         block_group = ei->i_block_group;
504         if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
505                 block_group &= ~(flex_size-1);
506                 if (S_ISREG(inode->i_mode))
507                         block_group++;
508         }
509         bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
510         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
511
512         /*
513          * If we are doing delayed allocation, we don't need take
514          * colour into account.
515          */
516         if (test_opt(inode->i_sb, DELALLOC))
517                 return bg_start;
518
519         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
520                 colour = (current->pid % 16) *
521                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
522         else
523                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
524         return bg_start + colour;
525 }
526
527 /**
528  *      ext4_find_goal - find a preferred place for allocation.
529  *      @inode: owner
530  *      @block:  block we want
531  *      @partial: pointer to the last triple within a chain
532  *
533  *      Normally this function find the preferred place for block allocation,
534  *      returns it.
535  *      Because this is only used for non-extent files, we limit the block nr
536  *      to 32 bits.
537  */
538 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
539                                    Indirect *partial)
540 {
541         ext4_fsblk_t goal;
542
543         /*
544          * XXX need to get goal block from mballoc's data structures
545          */
546
547         goal = ext4_find_near(inode, partial);
548         goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
549         return goal;
550 }
551
552 /**
553  *      ext4_blks_to_allocate: Look up the block map and count the number
554  *      of direct blocks need to be allocated for the given branch.
555  *
556  *      @branch: chain of indirect blocks
557  *      @k: number of blocks need for indirect blocks
558  *      @blks: number of data blocks to be mapped.
559  *      @blocks_to_boundary:  the offset in the indirect block
560  *
561  *      return the total number of blocks to be allocate, including the
562  *      direct and indirect blocks.
563  */
564 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
565                                  int blocks_to_boundary)
566 {
567         unsigned int count = 0;
568
569         /*
570          * Simple case, [t,d]Indirect block(s) has not allocated yet
571          * then it's clear blocks on that path have not allocated
572          */
573         if (k > 0) {
574                 /* right now we don't handle cross boundary allocation */
575                 if (blks < blocks_to_boundary + 1)
576                         count += blks;
577                 else
578                         count += blocks_to_boundary + 1;
579                 return count;
580         }
581
582         count++;
583         while (count < blks && count <= blocks_to_boundary &&
584                 le32_to_cpu(*(branch[0].p + count)) == 0) {
585                 count++;
586         }
587         return count;
588 }
589
590 /**
591  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
592  *      @indirect_blks: the number of blocks need to allocate for indirect
593  *                      blocks
594  *
595  *      @new_blocks: on return it will store the new block numbers for
596  *      the indirect blocks(if needed) and the first direct block,
597  *      @blks:  on return it will store the total number of allocated
598  *              direct blocks
599  */
600 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
601                              ext4_lblk_t iblock, ext4_fsblk_t goal,
602                              int indirect_blks, int blks,
603                              ext4_fsblk_t new_blocks[4], int *err)
604 {
605         struct ext4_allocation_request ar;
606         int target, i;
607         unsigned long count = 0, blk_allocated = 0;
608         int index = 0;
609         ext4_fsblk_t current_block = 0;
610         int ret = 0;
611
612         /*
613          * Here we try to allocate the requested multiple blocks at once,
614          * on a best-effort basis.
615          * To build a branch, we should allocate blocks for
616          * the indirect blocks(if not allocated yet), and at least
617          * the first direct block of this branch.  That's the
618          * minimum number of blocks need to allocate(required)
619          */
620         /* first we try to allocate the indirect blocks */
621         target = indirect_blks;
622         while (target > 0) {
623                 count = target;
624                 /* allocating blocks for indirect blocks and direct blocks */
625                 current_block = ext4_new_meta_blocks(handle, inode,
626                                                         goal, &count, err);
627                 if (*err)
628                         goto failed_out;
629
630                 if (unlikely(current_block + count > EXT4_MAX_BLOCK_FILE_PHYS)) {
631                         EXT4_ERROR_INODE(inode,
632                                          "current_block %llu + count %lu > %d!",
633                                          current_block, count,
634                                          EXT4_MAX_BLOCK_FILE_PHYS);
635                         *err = -EIO;
636                         goto failed_out;
637                 }
638
639                 target -= count;
640                 /* allocate blocks for indirect blocks */
641                 while (index < indirect_blks && count) {
642                         new_blocks[index++] = current_block++;
643                         count--;
644                 }
645                 if (count > 0) {
646                         /*
647                          * save the new block number
648                          * for the first direct block
649                          */
650                         new_blocks[index] = current_block;
651                         printk(KERN_INFO "%s returned more blocks than "
652                                                 "requested\n", __func__);
653                         WARN_ON(1);
654                         break;
655                 }
656         }
657
658         target = blks - count ;
659         blk_allocated = count;
660         if (!target)
661                 goto allocated;
662         /* Now allocate data blocks */
663         memset(&ar, 0, sizeof(ar));
664         ar.inode = inode;
665         ar.goal = goal;
666         ar.len = target;
667         ar.logical = iblock;
668         if (S_ISREG(inode->i_mode))
669                 /* enable in-core preallocation only for regular files */
670                 ar.flags = EXT4_MB_HINT_DATA;
671
672         current_block = ext4_mb_new_blocks(handle, &ar, err);
673         if (unlikely(current_block + ar.len > EXT4_MAX_BLOCK_FILE_PHYS)) {
674                 EXT4_ERROR_INODE(inode,
675                                  "current_block %llu + ar.len %d > %d!",
676                                  current_block, ar.len,
677                                  EXT4_MAX_BLOCK_FILE_PHYS);
678                 *err = -EIO;
679                 goto failed_out;
680         }
681
682         if (*err && (target == blks)) {
683                 /*
684                  * if the allocation failed and we didn't allocate
685                  * any blocks before
686                  */
687                 goto failed_out;
688         }
689         if (!*err) {
690                 if (target == blks) {
691                         /*
692                          * save the new block number
693                          * for the first direct block
694                          */
695                         new_blocks[index] = current_block;
696                 }
697                 blk_allocated += ar.len;
698         }
699 allocated:
700         /* total number of blocks allocated for direct blocks */
701         ret = blk_allocated;
702         *err = 0;
703         return ret;
704 failed_out:
705         for (i = 0; i < index; i++)
706                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
707         return ret;
708 }
709
710 /**
711  *      ext4_alloc_branch - allocate and set up a chain of blocks.
712  *      @inode: owner
713  *      @indirect_blks: number of allocated indirect blocks
714  *      @blks: number of allocated direct blocks
715  *      @offsets: offsets (in the blocks) to store the pointers to next.
716  *      @branch: place to store the chain in.
717  *
718  *      This function allocates blocks, zeroes out all but the last one,
719  *      links them into chain and (if we are synchronous) writes them to disk.
720  *      In other words, it prepares a branch that can be spliced onto the
721  *      inode. It stores the information about that chain in the branch[], in
722  *      the same format as ext4_get_branch() would do. We are calling it after
723  *      we had read the existing part of chain and partial points to the last
724  *      triple of that (one with zero ->key). Upon the exit we have the same
725  *      picture as after the successful ext4_get_block(), except that in one
726  *      place chain is disconnected - *branch->p is still zero (we did not
727  *      set the last link), but branch->key contains the number that should
728  *      be placed into *branch->p to fill that gap.
729  *
730  *      If allocation fails we free all blocks we've allocated (and forget
731  *      their buffer_heads) and return the error value the from failed
732  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
733  *      as described above and return 0.
734  */
735 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
736                              ext4_lblk_t iblock, int indirect_blks,
737                              int *blks, ext4_fsblk_t goal,
738                              ext4_lblk_t *offsets, Indirect *branch)
739 {
740         int blocksize = inode->i_sb->s_blocksize;
741         int i, n = 0;
742         int err = 0;
743         struct buffer_head *bh;
744         int num;
745         ext4_fsblk_t new_blocks[4];
746         ext4_fsblk_t current_block;
747
748         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
749                                 *blks, new_blocks, &err);
750         if (err)
751                 return err;
752
753         branch[0].key = cpu_to_le32(new_blocks[0]);
754         /*
755          * metadata blocks and data blocks are allocated.
756          */
757         for (n = 1; n <= indirect_blks;  n++) {
758                 /*
759                  * Get buffer_head for parent block, zero it out
760                  * and set the pointer to new one, then send
761                  * parent to disk.
762                  */
763                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
764                 if (unlikely(!bh)) {
765                         err = -EIO;
766                         goto failed;
767                 }
768
769                 branch[n].bh = bh;
770                 lock_buffer(bh);
771                 BUFFER_TRACE(bh, "call get_create_access");
772                 err = ext4_journal_get_create_access(handle, bh);
773                 if (err) {
774                         /* Don't brelse(bh) here; it's done in
775                          * ext4_journal_forget() below */
776                         unlock_buffer(bh);
777                         goto failed;
778                 }
779
780                 memset(bh->b_data, 0, blocksize);
781                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
782                 branch[n].key = cpu_to_le32(new_blocks[n]);
783                 *branch[n].p = branch[n].key;
784                 if (n == indirect_blks) {
785                         current_block = new_blocks[n];
786                         /*
787                          * End of chain, update the last new metablock of
788                          * the chain to point to the new allocated
789                          * data blocks numbers
790                          */
791                         for (i = 1; i < num; i++)
792                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
793                 }
794                 BUFFER_TRACE(bh, "marking uptodate");
795                 set_buffer_uptodate(bh);
796                 unlock_buffer(bh);
797
798                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
799                 err = ext4_handle_dirty_metadata(handle, inode, bh);
800                 if (err)
801                         goto failed;
802         }
803         *blks = num;
804         return err;
805 failed:
806         /* Allocation failed, free what we already allocated */
807         ext4_free_blocks(handle, inode, 0, new_blocks[0], 1, 0);
808         for (i = 1; i <= n ; i++) {
809                 /*
810                  * branch[i].bh is newly allocated, so there is no
811                  * need to revoke the block, which is why we don't
812                  * need to set EXT4_FREE_BLOCKS_METADATA.
813                  */
814                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1,
815                                  EXT4_FREE_BLOCKS_FORGET);
816         }
817         for (i = n+1; i < indirect_blks; i++)
818                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
819
820         ext4_free_blocks(handle, inode, 0, new_blocks[i], num, 0);
821
822         return err;
823 }
824
825 /**
826  * ext4_splice_branch - splice the allocated branch onto inode.
827  * @inode: owner
828  * @block: (logical) number of block we are adding
829  * @chain: chain of indirect blocks (with a missing link - see
830  *      ext4_alloc_branch)
831  * @where: location of missing link
832  * @num:   number of indirect blocks we are adding
833  * @blks:  number of direct blocks we are adding
834  *
835  * This function fills the missing link and does all housekeeping needed in
836  * inode (->i_blocks, etc.). In case of success we end up with the full
837  * chain to new block and return 0.
838  */
839 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
840                               ext4_lblk_t block, Indirect *where, int num,
841                               int blks)
842 {
843         int i;
844         int err = 0;
845         ext4_fsblk_t current_block;
846
847         /*
848          * If we're splicing into a [td]indirect block (as opposed to the
849          * inode) then we need to get write access to the [td]indirect block
850          * before the splice.
851          */
852         if (where->bh) {
853                 BUFFER_TRACE(where->bh, "get_write_access");
854                 err = ext4_journal_get_write_access(handle, where->bh);
855                 if (err)
856                         goto err_out;
857         }
858         /* That's it */
859
860         *where->p = where->key;
861
862         /*
863          * Update the host buffer_head or inode to point to more just allocated
864          * direct blocks blocks
865          */
866         if (num == 0 && blks > 1) {
867                 current_block = le32_to_cpu(where->key) + 1;
868                 for (i = 1; i < blks; i++)
869                         *(where->p + i) = cpu_to_le32(current_block++);
870         }
871
872         /* We are done with atomic stuff, now do the rest of housekeeping */
873         /* had we spliced it onto indirect block? */
874         if (where->bh) {
875                 /*
876                  * If we spliced it onto an indirect block, we haven't
877                  * altered the inode.  Note however that if it is being spliced
878                  * onto an indirect block at the very end of the file (the
879                  * file is growing) then we *will* alter the inode to reflect
880                  * the new i_size.  But that is not done here - it is done in
881                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
882                  */
883                 jbd_debug(5, "splicing indirect only\n");
884                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
885                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
886                 if (err)
887                         goto err_out;
888         } else {
889                 /*
890                  * OK, we spliced it into the inode itself on a direct block.
891                  */
892                 ext4_mark_inode_dirty(handle, inode);
893                 jbd_debug(5, "splicing direct\n");
894         }
895         return err;
896
897 err_out:
898         for (i = 1; i <= num; i++) {
899                 /*
900                  * branch[i].bh is newly allocated, so there is no
901                  * need to revoke the block, which is why we don't
902                  * need to set EXT4_FREE_BLOCKS_METADATA.
903                  */
904                 ext4_free_blocks(handle, inode, where[i].bh, 0, 1,
905                                  EXT4_FREE_BLOCKS_FORGET);
906         }
907         ext4_free_blocks(handle, inode, 0, le32_to_cpu(where[num].key),
908                          blks, 0);
909
910         return err;
911 }
912
913 /*
914  * The ext4_ind_map_blocks() function handles non-extents inodes
915  * (i.e., using the traditional indirect/double-indirect i_blocks
916  * scheme) for ext4_map_blocks().
917  *
918  * Allocation strategy is simple: if we have to allocate something, we will
919  * have to go the whole way to leaf. So let's do it before attaching anything
920  * to tree, set linkage between the newborn blocks, write them if sync is
921  * required, recheck the path, free and repeat if check fails, otherwise
922  * set the last missing link (that will protect us from any truncate-generated
923  * removals - all blocks on the path are immune now) and possibly force the
924  * write on the parent block.
925  * That has a nice additional property: no special recovery from the failed
926  * allocations is needed - we simply release blocks and do not touch anything
927  * reachable from inode.
928  *
929  * `handle' can be NULL if create == 0.
930  *
931  * return > 0, # of blocks mapped or allocated.
932  * return = 0, if plain lookup failed.
933  * return < 0, error case.
934  *
935  * The ext4_ind_get_blocks() function should be called with
936  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
937  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
938  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
939  * blocks.
940  */
941 static int ext4_ind_map_blocks(handle_t *handle, struct inode *inode,
942                                struct ext4_map_blocks *map,
943                                int flags)
944 {
945         int err = -EIO;
946         ext4_lblk_t offsets[4];
947         Indirect chain[4];
948         Indirect *partial;
949         ext4_fsblk_t goal;
950         int indirect_blks;
951         int blocks_to_boundary = 0;
952         int depth;
953         int count = 0;
954         ext4_fsblk_t first_block = 0;
955
956         J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
957         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
958         depth = ext4_block_to_path(inode, map->m_lblk, offsets,
959                                    &blocks_to_boundary);
960
961         if (depth == 0)
962                 goto out;
963
964         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
965
966         /* Simplest case - block found, no allocation needed */
967         if (!partial) {
968                 first_block = le32_to_cpu(chain[depth - 1].key);
969                 count++;
970                 /*map more blocks*/
971                 while (count < map->m_len && count <= blocks_to_boundary) {
972                         ext4_fsblk_t blk;
973
974                         blk = le32_to_cpu(*(chain[depth-1].p + count));
975
976                         if (blk == first_block + count)
977                                 count++;
978                         else
979                                 break;
980                 }
981                 goto got_it;
982         }
983
984         /* Next simple case - plain lookup or failed read of indirect block */
985         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
986                 goto cleanup;
987
988         /*
989          * Okay, we need to do block allocation.
990         */
991         goal = ext4_find_goal(inode, map->m_lblk, partial);
992
993         /* the number of blocks need to allocate for [d,t]indirect blocks */
994         indirect_blks = (chain + depth) - partial - 1;
995
996         /*
997          * Next look up the indirect map to count the totoal number of
998          * direct blocks to allocate for this branch.
999          */
1000         count = ext4_blks_to_allocate(partial, indirect_blks,
1001                                       map->m_len, blocks_to_boundary);
1002         /*
1003          * Block out ext4_truncate while we alter the tree
1004          */
1005         err = ext4_alloc_branch(handle, inode, map->m_lblk, indirect_blks,
1006                                 &count, goal,
1007                                 offsets + (partial - chain), partial);
1008
1009         /*
1010          * The ext4_splice_branch call will free and forget any buffers
1011          * on the new chain if there is a failure, but that risks using
1012          * up transaction credits, especially for bitmaps where the
1013          * credits cannot be returned.  Can we handle this somehow?  We
1014          * may need to return -EAGAIN upwards in the worst case.  --sct
1015          */
1016         if (!err)
1017                 err = ext4_splice_branch(handle, inode, map->m_lblk,
1018                                          partial, indirect_blks, count);
1019         if (err)
1020                 goto cleanup;
1021
1022         map->m_flags |= EXT4_MAP_NEW;
1023
1024         ext4_update_inode_fsync_trans(handle, inode, 1);
1025 got_it:
1026         map->m_flags |= EXT4_MAP_MAPPED;
1027         map->m_pblk = le32_to_cpu(chain[depth-1].key);
1028         map->m_len = count;
1029         if (count > blocks_to_boundary)
1030                 map->m_flags |= EXT4_MAP_BOUNDARY;
1031         err = count;
1032         /* Clean up and exit */
1033         partial = chain + depth - 1;    /* the whole chain */
1034 cleanup:
1035         while (partial > chain) {
1036                 BUFFER_TRACE(partial->bh, "call brelse");
1037                 brelse(partial->bh);
1038                 partial--;
1039         }
1040 out:
1041         return err;
1042 }
1043
1044 #ifdef CONFIG_QUOTA
1045 qsize_t *ext4_get_reserved_space(struct inode *inode)
1046 {
1047         return &EXT4_I(inode)->i_reserved_quota;
1048 }
1049 #endif
1050
1051 /*
1052  * Calculate the number of metadata blocks need to reserve
1053  * to allocate a new block at @lblocks for non extent file based file
1054  */
1055 static int ext4_indirect_calc_metadata_amount(struct inode *inode,
1056                                               sector_t lblock)
1057 {
1058         struct ext4_inode_info *ei = EXT4_I(inode);
1059         sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
1060         int blk_bits;
1061
1062         if (lblock < EXT4_NDIR_BLOCKS)
1063                 return 0;
1064
1065         lblock -= EXT4_NDIR_BLOCKS;
1066
1067         if (ei->i_da_metadata_calc_len &&
1068             (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
1069                 ei->i_da_metadata_calc_len++;
1070                 return 0;
1071         }
1072         ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
1073         ei->i_da_metadata_calc_len = 1;
1074         blk_bits = order_base_2(lblock);
1075         return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
1076 }
1077
1078 /*
1079  * Calculate the number of metadata blocks need to reserve
1080  * to allocate a block located at @lblock
1081  */
1082 static int ext4_calc_metadata_amount(struct inode *inode, sector_t lblock)
1083 {
1084         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1085                 return ext4_ext_calc_metadata_amount(inode, lblock);
1086
1087         return ext4_indirect_calc_metadata_amount(inode, lblock);
1088 }
1089
1090 /*
1091  * Called with i_data_sem down, which is important since we can call
1092  * ext4_discard_preallocations() from here.
1093  */
1094 void ext4_da_update_reserve_space(struct inode *inode,
1095                                         int used, int quota_claim)
1096 {
1097         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1098         struct ext4_inode_info *ei = EXT4_I(inode);
1099
1100         spin_lock(&ei->i_block_reservation_lock);
1101         trace_ext4_da_update_reserve_space(inode, used);
1102         if (unlikely(used > ei->i_reserved_data_blocks)) {
1103                 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
1104                          "with only %d reserved data blocks\n",
1105                          __func__, inode->i_ino, used,
1106                          ei->i_reserved_data_blocks);
1107                 WARN_ON(1);
1108                 used = ei->i_reserved_data_blocks;
1109         }
1110
1111         /* Update per-inode reservations */
1112         ei->i_reserved_data_blocks -= used;
1113         ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
1114         percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1115                            used + ei->i_allocated_meta_blocks);
1116         ei->i_allocated_meta_blocks = 0;
1117
1118         if (ei->i_reserved_data_blocks == 0) {
1119                 /*
1120                  * We can release all of the reserved metadata blocks
1121                  * only when we have written all of the delayed
1122                  * allocation blocks.
1123                  */
1124                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1125                                    ei->i_reserved_meta_blocks);
1126                 ei->i_reserved_meta_blocks = 0;
1127                 ei->i_da_metadata_calc_len = 0;
1128         }
1129         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1130
1131         /* Update quota subsystem for data blocks */
1132         if (quota_claim)
1133                 dquot_claim_block(inode, used);
1134         else {
1135                 /*
1136                  * We did fallocate with an offset that is already delayed
1137                  * allocated. So on delayed allocated writeback we should
1138                  * not re-claim the quota for fallocated blocks.
1139                  */
1140                 dquot_release_reservation_block(inode, used);
1141         }
1142
1143         /*
1144          * If we have done all the pending block allocations and if
1145          * there aren't any writers on the inode, we can discard the
1146          * inode's preallocations.
1147          */
1148         if ((ei->i_reserved_data_blocks == 0) &&
1149             (atomic_read(&inode->i_writecount) == 0))
1150                 ext4_discard_preallocations(inode);
1151 }
1152
1153 static int __check_block_validity(struct inode *inode, const char *func,
1154                                 unsigned int line,
1155                                 struct ext4_map_blocks *map)
1156 {
1157         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
1158                                    map->m_len)) {
1159                 ext4_error_inode(inode, func, line, map->m_pblk,
1160                                  "lblock %lu mapped to illegal pblock "
1161                                  "(length %d)", (unsigned long) map->m_lblk,
1162                                  map->m_len);
1163                 return -EIO;
1164         }
1165         return 0;
1166 }
1167
1168 #define check_block_validity(inode, map)        \
1169         __check_block_validity((inode), __func__, __LINE__, (map))
1170
1171 /*
1172  * Return the number of contiguous dirty pages in a given inode
1173  * starting at page frame idx.
1174  */
1175 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
1176                                     unsigned int max_pages)
1177 {
1178         struct address_space *mapping = inode->i_mapping;
1179         pgoff_t index;
1180         struct pagevec pvec;
1181         pgoff_t num = 0;
1182         int i, nr_pages, done = 0;
1183
1184         if (max_pages == 0)
1185                 return 0;
1186         pagevec_init(&pvec, 0);
1187         while (!done) {
1188                 index = idx;
1189                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
1190                                               PAGECACHE_TAG_DIRTY,
1191                                               (pgoff_t)PAGEVEC_SIZE);
1192                 if (nr_pages == 0)
1193                         break;
1194                 for (i = 0; i < nr_pages; i++) {
1195                         struct page *page = pvec.pages[i];
1196                         struct buffer_head *bh, *head;
1197
1198                         lock_page(page);
1199                         if (unlikely(page->mapping != mapping) ||
1200                             !PageDirty(page) ||
1201                             PageWriteback(page) ||
1202                             page->index != idx) {
1203                                 done = 1;
1204                                 unlock_page(page);
1205                                 break;
1206                         }
1207                         if (page_has_buffers(page)) {
1208                                 bh = head = page_buffers(page);
1209                                 do {
1210                                         if (!buffer_delay(bh) &&
1211                                             !buffer_unwritten(bh))
1212                                                 done = 1;
1213                                         bh = bh->b_this_page;
1214                                 } while (!done && (bh != head));
1215                         }
1216                         unlock_page(page);
1217                         if (done)
1218                                 break;
1219                         idx++;
1220                         num++;
1221                         if (num >= max_pages) {
1222                                 done = 1;
1223                                 break;
1224                         }
1225                 }
1226                 pagevec_release(&pvec);
1227         }
1228         return num;
1229 }
1230
1231 /*
1232  * The ext4_map_blocks() function tries to look up the requested blocks,
1233  * and returns if the blocks are already mapped.
1234  *
1235  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1236  * and store the allocated blocks in the result buffer head and mark it
1237  * mapped.
1238  *
1239  * If file type is extents based, it will call ext4_ext_map_blocks(),
1240  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
1241  * based files
1242  *
1243  * On success, it returns the number of blocks being mapped or allocate.
1244  * if create==0 and the blocks are pre-allocated and uninitialized block,
1245  * the result buffer head is unmapped. If the create ==1, it will make sure
1246  * the buffer head is mapped.
1247  *
1248  * It returns 0 if plain look up failed (blocks have not been allocated), in
1249  * that casem, buffer head is unmapped
1250  *
1251  * It returns the error in case of allocation failure.
1252  */
1253 int ext4_map_blocks(handle_t *handle, struct inode *inode,
1254                     struct ext4_map_blocks *map, int flags)
1255 {
1256         int retval;
1257
1258         map->m_flags = 0;
1259         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
1260                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
1261                   (unsigned long) map->m_lblk);
1262         /*
1263          * Try to see if we can get the block without requesting a new
1264          * file system block.
1265          */
1266         down_read((&EXT4_I(inode)->i_data_sem));
1267         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
1268                 retval = ext4_ext_map_blocks(handle, inode, map, 0);
1269         } else {
1270                 retval = ext4_ind_map_blocks(handle, inode, map, 0);
1271         }
1272         up_read((&EXT4_I(inode)->i_data_sem));
1273
1274         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
1275                 int ret = check_block_validity(inode, map);
1276                 if (ret != 0)
1277                         return ret;
1278         }
1279
1280         /* If it is only a block(s) look up */
1281         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1282                 return retval;
1283
1284         /*
1285          * Returns if the blocks have already allocated
1286          *
1287          * Note that if blocks have been preallocated
1288          * ext4_ext_get_block() returns th create = 0
1289          * with buffer head unmapped.
1290          */
1291         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
1292                 return retval;
1293
1294         /*
1295          * When we call get_blocks without the create flag, the
1296          * BH_Unwritten flag could have gotten set if the blocks
1297          * requested were part of a uninitialized extent.  We need to
1298          * clear this flag now that we are committed to convert all or
1299          * part of the uninitialized extent to be an initialized
1300          * extent.  This is because we need to avoid the combination
1301          * of BH_Unwritten and BH_Mapped flags being simultaneously
1302          * set on the buffer_head.
1303          */
1304         map->m_flags &= ~EXT4_MAP_UNWRITTEN;
1305
1306         /*
1307          * New blocks allocate and/or writing to uninitialized extent
1308          * will possibly result in updating i_data, so we take
1309          * the write lock of i_data_sem, and call get_blocks()
1310          * with create == 1 flag.
1311          */
1312         down_write((&EXT4_I(inode)->i_data_sem));
1313
1314         /*
1315          * if the caller is from delayed allocation writeout path
1316          * we have already reserved fs blocks for allocation
1317          * let the underlying get_block() function know to
1318          * avoid double accounting
1319          */
1320         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1321                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1322         /*
1323          * We need to check for EXT4 here because migrate
1324          * could have changed the inode type in between
1325          */
1326         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
1327                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
1328         } else {
1329                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
1330
1331                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
1332                         /*
1333                          * We allocated new blocks which will result in
1334                          * i_data's format changing.  Force the migrate
1335                          * to fail by clearing migrate flags
1336                          */
1337                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
1338                 }
1339
1340                 /*
1341                  * Update reserved blocks/metadata blocks after successful
1342                  * block allocation which had been deferred till now. We don't
1343                  * support fallocate for non extent files. So we can update
1344                  * reserve space here.
1345                  */
1346                 if ((retval > 0) &&
1347                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
1348                         ext4_da_update_reserve_space(inode, retval, 1);
1349         }
1350         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1351                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1352
1353         up_write((&EXT4_I(inode)->i_data_sem));
1354         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
1355                 int ret = check_block_validity(inode, map);
1356                 if (ret != 0)
1357                         return ret;
1358         }
1359         return retval;
1360 }
1361
1362 /* Maximum number of blocks we map for direct IO at once. */
1363 #define DIO_MAX_BLOCKS 4096
1364
1365 static int _ext4_get_block(struct inode *inode, sector_t iblock,
1366                            struct buffer_head *bh, int flags)
1367 {
1368         handle_t *handle = ext4_journal_current_handle();
1369         struct ext4_map_blocks map;
1370         int ret = 0, started = 0;
1371         int dio_credits;
1372
1373         map.m_lblk = iblock;
1374         map.m_len = bh->b_size >> inode->i_blkbits;
1375
1376         if (flags && !handle) {
1377                 /* Direct IO write... */
1378                 if (map.m_len > DIO_MAX_BLOCKS)
1379                         map.m_len = DIO_MAX_BLOCKS;
1380                 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
1381                 handle = ext4_journal_start(inode, dio_credits);
1382                 if (IS_ERR(handle)) {
1383                         ret = PTR_ERR(handle);
1384                         return ret;
1385                 }
1386                 started = 1;
1387         }
1388
1389         ret = ext4_map_blocks(handle, inode, &map, flags);
1390         if (ret > 0) {
1391                 map_bh(bh, inode->i_sb, map.m_pblk);
1392                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
1393                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
1394                 ret = 0;
1395         }
1396         if (started)
1397                 ext4_journal_stop(handle);
1398         return ret;
1399 }
1400
1401 int ext4_get_block(struct inode *inode, sector_t iblock,
1402                    struct buffer_head *bh, int create)
1403 {
1404         return _ext4_get_block(inode, iblock, bh,
1405                                create ? EXT4_GET_BLOCKS_CREATE : 0);
1406 }
1407
1408 /*
1409  * `handle' can be NULL if create is zero
1410  */
1411 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1412                                 ext4_lblk_t block, int create, int *errp)
1413 {
1414         struct ext4_map_blocks map;
1415         struct buffer_head *bh;
1416         int fatal = 0, err;
1417
1418         J_ASSERT(handle != NULL || create == 0);
1419
1420         map.m_lblk = block;
1421         map.m_len = 1;
1422         err = ext4_map_blocks(handle, inode, &map,
1423                               create ? EXT4_GET_BLOCKS_CREATE : 0);
1424
1425         if (err < 0)
1426                 *errp = err;
1427         if (err <= 0)
1428                 return NULL;
1429         *errp = 0;
1430
1431         bh = sb_getblk(inode->i_sb, map.m_pblk);
1432         if (!bh) {
1433                 *errp = -EIO;
1434                 return NULL;
1435         }
1436         if (map.m_flags & EXT4_MAP_NEW) {
1437                 J_ASSERT(create != 0);
1438                 J_ASSERT(handle != NULL);
1439
1440                 /*
1441                  * Now that we do not always journal data, we should
1442                  * keep in mind whether this should always journal the
1443                  * new buffer as metadata.  For now, regular file
1444                  * writes use ext4_get_block instead, so it's not a
1445                  * problem.
1446                  */
1447                 lock_buffer(bh);
1448                 BUFFER_TRACE(bh, "call get_create_access");
1449                 fatal = ext4_journal_get_create_access(handle, bh);
1450                 if (!fatal && !buffer_uptodate(bh)) {
1451                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1452                         set_buffer_uptodate(bh);
1453                 }
1454                 unlock_buffer(bh);
1455                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1456                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1457                 if (!fatal)
1458                         fatal = err;
1459         } else {
1460                 BUFFER_TRACE(bh, "not a new buffer");
1461         }
1462         if (fatal) {
1463                 *errp = fatal;
1464                 brelse(bh);
1465                 bh = NULL;
1466         }
1467         return bh;
1468 }
1469
1470 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1471                                ext4_lblk_t block, int create, int *err)
1472 {
1473         struct buffer_head *bh;
1474
1475         bh = ext4_getblk(handle, inode, block, create, err);
1476         if (!bh)
1477                 return bh;
1478         if (buffer_uptodate(bh))
1479                 return bh;
1480         ll_rw_block(READ_META, 1, &bh);
1481         wait_on_buffer(bh);
1482         if (buffer_uptodate(bh))
1483                 return bh;
1484         put_bh(bh);
1485         *err = -EIO;
1486         return NULL;
1487 }
1488
1489 static int walk_page_buffers(handle_t *handle,
1490                              struct buffer_head *head,
1491                              unsigned from,
1492                              unsigned to,
1493                              int *partial,
1494                              int (*fn)(handle_t *handle,
1495                                        struct buffer_head *bh))
1496 {
1497         struct buffer_head *bh;
1498         unsigned block_start, block_end;
1499         unsigned blocksize = head->b_size;
1500         int err, ret = 0;
1501         struct buffer_head *next;
1502
1503         for (bh = head, block_start = 0;
1504              ret == 0 && (bh != head || !block_start);
1505              block_start = block_end, bh = next) {
1506                 next = bh->b_this_page;
1507                 block_end = block_start + blocksize;
1508                 if (block_end <= from || block_start >= to) {
1509                         if (partial && !buffer_uptodate(bh))
1510                                 *partial = 1;
1511                         continue;
1512                 }
1513                 err = (*fn)(handle, bh);
1514                 if (!ret)
1515                         ret = err;
1516         }
1517         return ret;
1518 }
1519
1520 /*
1521  * To preserve ordering, it is essential that the hole instantiation and
1522  * the data write be encapsulated in a single transaction.  We cannot
1523  * close off a transaction and start a new one between the ext4_get_block()
1524  * and the commit_write().  So doing the jbd2_journal_start at the start of
1525  * prepare_write() is the right place.
1526  *
1527  * Also, this function can nest inside ext4_writepage() ->
1528  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1529  * has generated enough buffer credits to do the whole page.  So we won't
1530  * block on the journal in that case, which is good, because the caller may
1531  * be PF_MEMALLOC.
1532  *
1533  * By accident, ext4 can be reentered when a transaction is open via
1534  * quota file writes.  If we were to commit the transaction while thus
1535  * reentered, there can be a deadlock - we would be holding a quota
1536  * lock, and the commit would never complete if another thread had a
1537  * transaction open and was blocking on the quota lock - a ranking
1538  * violation.
1539  *
1540  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1541  * will _not_ run commit under these circumstances because handle->h_ref
1542  * is elevated.  We'll still have enough credits for the tiny quotafile
1543  * write.
1544  */
1545 static int do_journal_get_write_access(handle_t *handle,
1546                                        struct buffer_head *bh)
1547 {
1548         int dirty = buffer_dirty(bh);
1549         int ret;
1550
1551         if (!buffer_mapped(bh) || buffer_freed(bh))
1552                 return 0;
1553         /*
1554          * __block_prepare_write() could have dirtied some buffers. Clean
1555          * the dirty bit as jbd2_journal_get_write_access() could complain
1556          * otherwise about fs integrity issues. Setting of the dirty bit
1557          * by __block_prepare_write() isn't a real problem here as we clear
1558          * the bit before releasing a page lock and thus writeback cannot
1559          * ever write the buffer.
1560          */
1561         if (dirty)
1562                 clear_buffer_dirty(bh);
1563         ret = ext4_journal_get_write_access(handle, bh);
1564         if (!ret && dirty)
1565                 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1566         return ret;
1567 }
1568
1569 /*
1570  * Truncate blocks that were not used by write. We have to truncate the
1571  * pagecache as well so that corresponding buffers get properly unmapped.
1572  */
1573 static void ext4_truncate_failed_write(struct inode *inode)
1574 {
1575         truncate_inode_pages(inode->i_mapping, inode->i_size);
1576         ext4_truncate(inode);
1577 }
1578
1579 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
1580                    struct buffer_head *bh_result, int create);
1581 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1582                             loff_t pos, unsigned len, unsigned flags,
1583                             struct page **pagep, void **fsdata)
1584 {
1585         struct inode *inode = mapping->host;
1586         int ret, needed_blocks;
1587         handle_t *handle;
1588         int retries = 0;
1589         struct page *page;
1590         pgoff_t index;
1591         unsigned from, to;
1592
1593         trace_ext4_write_begin(inode, pos, len, flags);
1594         /*
1595          * Reserve one block more for addition to orphan list in case
1596          * we allocate blocks but write fails for some reason
1597          */
1598         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1599         index = pos >> PAGE_CACHE_SHIFT;
1600         from = pos & (PAGE_CACHE_SIZE - 1);
1601         to = from + len;
1602
1603 retry:
1604         handle = ext4_journal_start(inode, needed_blocks);
1605         if (IS_ERR(handle)) {
1606                 ret = PTR_ERR(handle);
1607                 goto out;
1608         }
1609
1610         /* We cannot recurse into the filesystem as the transaction is already
1611          * started */
1612         flags |= AOP_FLAG_NOFS;
1613
1614         page = grab_cache_page_write_begin(mapping, index, flags);
1615         if (!page) {
1616                 ext4_journal_stop(handle);
1617                 ret = -ENOMEM;
1618                 goto out;
1619         }
1620         *pagep = page;
1621
1622         if (ext4_should_dioread_nolock(inode))
1623                 ret = __block_write_begin(page, pos, len, ext4_get_block_write);
1624         else
1625                 ret = __block_write_begin(page, pos, len, ext4_get_block);
1626
1627         if (!ret && ext4_should_journal_data(inode)) {
1628                 ret = walk_page_buffers(handle, page_buffers(page),
1629                                 from, to, NULL, do_journal_get_write_access);
1630         }
1631
1632         if (ret) {
1633                 unlock_page(page);
1634                 page_cache_release(page);
1635                 /*
1636                  * __block_write_begin may have instantiated a few blocks
1637                  * outside i_size.  Trim these off again. Don't need
1638                  * i_size_read because we hold i_mutex.
1639                  *
1640                  * Add inode to orphan list in case we crash before
1641                  * truncate finishes
1642                  */
1643                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1644                         ext4_orphan_add(handle, inode);
1645
1646                 ext4_journal_stop(handle);
1647                 if (pos + len > inode->i_size) {
1648                         ext4_truncate_failed_write(inode);
1649                         /*
1650                          * If truncate failed early the inode might
1651                          * still be on the orphan list; we need to
1652                          * make sure the inode is removed from the
1653                          * orphan list in that case.
1654                          */
1655                         if (inode->i_nlink)
1656                                 ext4_orphan_del(NULL, inode);
1657                 }
1658         }
1659
1660         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1661                 goto retry;
1662 out:
1663         return ret;
1664 }
1665
1666 /* For write_end() in data=journal mode */
1667 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1668 {
1669         if (!buffer_mapped(bh) || buffer_freed(bh))
1670                 return 0;
1671         set_buffer_uptodate(bh);
1672         return ext4_handle_dirty_metadata(handle, NULL, bh);
1673 }
1674
1675 static int ext4_generic_write_end(struct file *file,
1676                                   struct address_space *mapping,
1677                                   loff_t pos, unsigned len, unsigned copied,
1678                                   struct page *page, void *fsdata)
1679 {
1680         int i_size_changed = 0;
1681         struct inode *inode = mapping->host;
1682         handle_t *handle = ext4_journal_current_handle();
1683
1684         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1685
1686         /*
1687          * No need to use i_size_read() here, the i_size
1688          * cannot change under us because we hold i_mutex.
1689          *
1690          * But it's important to update i_size while still holding page lock:
1691          * page writeout could otherwise come in and zero beyond i_size.
1692          */
1693         if (pos + copied > inode->i_size) {
1694                 i_size_write(inode, pos + copied);
1695                 i_size_changed = 1;
1696         }
1697
1698         if (pos + copied >  EXT4_I(inode)->i_disksize) {
1699                 /* We need to mark inode dirty even if
1700                  * new_i_size is less that inode->i_size
1701                  * bu greater than i_disksize.(hint delalloc)
1702                  */
1703                 ext4_update_i_disksize(inode, (pos + copied));
1704                 i_size_changed = 1;
1705         }
1706         unlock_page(page);
1707         page_cache_release(page);
1708
1709         /*
1710          * Don't mark the inode dirty under page lock. First, it unnecessarily
1711          * makes the holding time of page lock longer. Second, it forces lock
1712          * ordering of page lock and transaction start for journaling
1713          * filesystems.
1714          */
1715         if (i_size_changed)
1716                 ext4_mark_inode_dirty(handle, inode);
1717
1718         return copied;
1719 }
1720
1721 /*
1722  * We need to pick up the new inode size which generic_commit_write gave us
1723  * `file' can be NULL - eg, when called from page_symlink().
1724  *
1725  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1726  * buffers are managed internally.
1727  */
1728 static int ext4_ordered_write_end(struct file *file,
1729                                   struct address_space *mapping,
1730                                   loff_t pos, unsigned len, unsigned copied,
1731                                   struct page *page, void *fsdata)
1732 {
1733         handle_t *handle = ext4_journal_current_handle();
1734         struct inode *inode = mapping->host;
1735         int ret = 0, ret2;
1736
1737         trace_ext4_ordered_write_end(inode, pos, len, copied);
1738         ret = ext4_jbd2_file_inode(handle, inode);
1739
1740         if (ret == 0) {
1741                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1742                                                         page, fsdata);
1743                 copied = ret2;
1744                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1745                         /* if we have allocated more blocks and copied
1746                          * less. We will have blocks allocated outside
1747                          * inode->i_size. So truncate them
1748                          */
1749                         ext4_orphan_add(handle, inode);
1750                 if (ret2 < 0)
1751                         ret = ret2;
1752         }
1753         ret2 = ext4_journal_stop(handle);
1754         if (!ret)
1755                 ret = ret2;
1756
1757         if (pos + len > inode->i_size) {
1758                 ext4_truncate_failed_write(inode);
1759                 /*
1760                  * If truncate failed early the inode might still be
1761                  * on the orphan list; we need to make sure the inode
1762                  * is removed from the orphan list in that case.
1763                  */
1764                 if (inode->i_nlink)
1765                         ext4_orphan_del(NULL, inode);
1766         }
1767
1768
1769         return ret ? ret : copied;
1770 }
1771
1772 static int ext4_writeback_write_end(struct file *file,
1773                                     struct address_space *mapping,
1774                                     loff_t pos, unsigned len, unsigned copied,
1775                                     struct page *page, void *fsdata)
1776 {
1777         handle_t *handle = ext4_journal_current_handle();
1778         struct inode *inode = mapping->host;
1779         int ret = 0, ret2;
1780
1781         trace_ext4_writeback_write_end(inode, pos, len, copied);
1782         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1783                                                         page, fsdata);
1784         copied = ret2;
1785         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1786                 /* if we have allocated more blocks and copied
1787                  * less. We will have blocks allocated outside
1788                  * inode->i_size. So truncate them
1789                  */
1790                 ext4_orphan_add(handle, inode);
1791
1792         if (ret2 < 0)
1793                 ret = ret2;
1794
1795         ret2 = ext4_journal_stop(handle);
1796         if (!ret)
1797                 ret = ret2;
1798
1799         if (pos + len > inode->i_size) {
1800                 ext4_truncate_failed_write(inode);
1801                 /*
1802                  * If truncate failed early the inode might still be
1803                  * on the orphan list; we need to make sure the inode
1804                  * is removed from the orphan list in that case.
1805                  */
1806                 if (inode->i_nlink)
1807                         ext4_orphan_del(NULL, inode);
1808         }
1809
1810         return ret ? ret : copied;
1811 }
1812
1813 static int ext4_journalled_write_end(struct file *file,
1814                                      struct address_space *mapping,
1815                                      loff_t pos, unsigned len, unsigned copied,
1816                                      struct page *page, void *fsdata)
1817 {
1818         handle_t *handle = ext4_journal_current_handle();
1819         struct inode *inode = mapping->host;
1820         int ret = 0, ret2;
1821         int partial = 0;
1822         unsigned from, to;
1823         loff_t new_i_size;
1824
1825         trace_ext4_journalled_write_end(inode, pos, len, copied);
1826         from = pos & (PAGE_CACHE_SIZE - 1);
1827         to = from + len;
1828
1829         if (copied < len) {
1830                 if (!PageUptodate(page))
1831                         copied = 0;
1832                 page_zero_new_buffers(page, from+copied, to);
1833         }
1834
1835         ret = walk_page_buffers(handle, page_buffers(page), from,
1836                                 to, &partial, write_end_fn);
1837         if (!partial)
1838                 SetPageUptodate(page);
1839         new_i_size = pos + copied;
1840         if (new_i_size > inode->i_size)
1841                 i_size_write(inode, pos+copied);
1842         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1843         if (new_i_size > EXT4_I(inode)->i_disksize) {
1844                 ext4_update_i_disksize(inode, new_i_size);
1845                 ret2 = ext4_mark_inode_dirty(handle, inode);
1846                 if (!ret)
1847                         ret = ret2;
1848         }
1849
1850         unlock_page(page);
1851         page_cache_release(page);
1852         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1853                 /* if we have allocated more blocks and copied
1854                  * less. We will have blocks allocated outside
1855                  * inode->i_size. So truncate them
1856                  */
1857                 ext4_orphan_add(handle, inode);
1858
1859         ret2 = ext4_journal_stop(handle);
1860         if (!ret)
1861                 ret = ret2;
1862         if (pos + len > inode->i_size) {
1863                 ext4_truncate_failed_write(inode);
1864                 /*
1865                  * If truncate failed early the inode might still be
1866                  * on the orphan list; we need to make sure the inode
1867                  * is removed from the orphan list in that case.
1868                  */
1869                 if (inode->i_nlink)
1870                         ext4_orphan_del(NULL, inode);
1871         }
1872
1873         return ret ? ret : copied;
1874 }
1875
1876 /*
1877  * Reserve a single block located at lblock
1878  */
1879 static int ext4_da_reserve_space(struct inode *inode, sector_t lblock)
1880 {
1881         int retries = 0;
1882         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1883         struct ext4_inode_info *ei = EXT4_I(inode);
1884         unsigned long md_needed;
1885         int ret;
1886
1887         /*
1888          * recalculate the amount of metadata blocks to reserve
1889          * in order to allocate nrblocks
1890          * worse case is one extent per block
1891          */
1892 repeat:
1893         spin_lock(&ei->i_block_reservation_lock);
1894         md_needed = ext4_calc_metadata_amount(inode, lblock);
1895         trace_ext4_da_reserve_space(inode, md_needed);
1896         spin_unlock(&ei->i_block_reservation_lock);
1897
1898         /*
1899          * We will charge metadata quota at writeout time; this saves
1900          * us from metadata over-estimation, though we may go over by
1901          * a small amount in the end.  Here we just reserve for data.
1902          */
1903         ret = dquot_reserve_block(inode, 1);
1904         if (ret)
1905                 return ret;
1906         /*
1907          * We do still charge estimated metadata to the sb though;
1908          * we cannot afford to run out of free blocks.
1909          */
1910         if (ext4_claim_free_blocks(sbi, md_needed + 1)) {
1911                 dquot_release_reservation_block(inode, 1);
1912                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1913                         yield();
1914                         goto repeat;
1915                 }
1916                 return -ENOSPC;
1917         }
1918         spin_lock(&ei->i_block_reservation_lock);
1919         ei->i_reserved_data_blocks++;
1920         ei->i_reserved_meta_blocks += md_needed;
1921         spin_unlock(&ei->i_block_reservation_lock);
1922
1923         return 0;       /* success */
1924 }
1925
1926 static void ext4_da_release_space(struct inode *inode, int to_free)
1927 {
1928         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1929         struct ext4_inode_info *ei = EXT4_I(inode);
1930
1931         if (!to_free)
1932                 return;         /* Nothing to release, exit */
1933
1934         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1935
1936         trace_ext4_da_release_space(inode, to_free);
1937         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1938                 /*
1939                  * if there aren't enough reserved blocks, then the
1940                  * counter is messed up somewhere.  Since this
1941                  * function is called from invalidate page, it's
1942                  * harmless to return without any action.
1943                  */
1944                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1945                          "ino %lu, to_free %d with only %d reserved "
1946                          "data blocks\n", inode->i_ino, to_free,
1947                          ei->i_reserved_data_blocks);
1948                 WARN_ON(1);
1949                 to_free = ei->i_reserved_data_blocks;
1950         }
1951         ei->i_reserved_data_blocks -= to_free;
1952
1953         if (ei->i_reserved_data_blocks == 0) {
1954                 /*
1955                  * We can release all of the reserved metadata blocks
1956                  * only when we have written all of the delayed
1957                  * allocation blocks.
1958                  */
1959                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1960                                    ei->i_reserved_meta_blocks);
1961                 ei->i_reserved_meta_blocks = 0;
1962                 ei->i_da_metadata_calc_len = 0;
1963         }
1964
1965         /* update fs dirty data blocks counter */
1966         percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1967
1968         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1969
1970         dquot_release_reservation_block(inode, to_free);
1971 }
1972
1973 static void ext4_da_page_release_reservation(struct page *page,
1974                                              unsigned long offset)
1975 {
1976         int to_release = 0;
1977         struct buffer_head *head, *bh;
1978         unsigned int curr_off = 0;
1979
1980         head = page_buffers(page);
1981         bh = head;
1982         do {
1983                 unsigned int next_off = curr_off + bh->b_size;
1984
1985                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1986                         to_release++;
1987                         clear_buffer_delay(bh);
1988                 }
1989                 curr_off = next_off;
1990         } while ((bh = bh->b_this_page) != head);
1991         ext4_da_release_space(page->mapping->host, to_release);
1992 }
1993
1994 /*
1995  * Delayed allocation stuff
1996  */
1997
1998 /*
1999  * mpage_da_submit_io - walks through extent of pages and try to write
2000  * them with writepage() call back
2001  *
2002  * @mpd->inode: inode
2003  * @mpd->first_page: first page of the extent
2004  * @mpd->next_page: page after the last page of the extent
2005  *
2006  * By the time mpage_da_submit_io() is called we expect all blocks
2007  * to be allocated. this may be wrong if allocation failed.
2008  *
2009  * As pages are already locked by write_cache_pages(), we can't use it
2010  */
2011 static int mpage_da_submit_io(struct mpage_da_data *mpd,
2012                               struct ext4_map_blocks *map)
2013 {
2014         struct pagevec pvec;
2015         unsigned long index, end;
2016         int ret = 0, err, nr_pages, i;
2017         struct inode *inode = mpd->inode;
2018         struct address_space *mapping = inode->i_mapping;
2019         loff_t size = i_size_read(inode);
2020         unsigned int len, block_start;
2021         struct buffer_head *bh, *page_bufs = NULL;
2022         int journal_data = ext4_should_journal_data(inode);
2023         sector_t pblock = 0, cur_logical = 0;
2024         struct ext4_io_submit io_submit;
2025
2026         BUG_ON(mpd->next_page <= mpd->first_page);
2027         memset(&io_submit, 0, sizeof(io_submit));
2028         /*
2029          * We need to start from the first_page to the next_page - 1
2030          * to make sure we also write the mapped dirty buffer_heads.
2031          * If we look at mpd->b_blocknr we would only be looking
2032          * at the currently mapped buffer_heads.
2033          */
2034         index = mpd->first_page;
2035         end = mpd->next_page - 1;
2036
2037         pagevec_init(&pvec, 0);
2038         while (index <= end) {
2039                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2040                 if (nr_pages == 0)
2041                         break;
2042                 for (i = 0; i < nr_pages; i++) {
2043                         int commit_write = 0, redirty_page = 0;
2044                         struct page *page = pvec.pages[i];
2045
2046                         index = page->index;
2047                         if (index > end)
2048                                 break;
2049
2050                         if (index == size >> PAGE_CACHE_SHIFT)
2051                                 len = size & ~PAGE_CACHE_MASK;
2052                         else
2053                                 len = PAGE_CACHE_SIZE;
2054                         if (map) {
2055                                 cur_logical = index << (PAGE_CACHE_SHIFT -
2056                                                         inode->i_blkbits);
2057                                 pblock = map->m_pblk + (cur_logical -
2058                                                         map->m_lblk);
2059                         }
2060                         index++;
2061
2062                         BUG_ON(!PageLocked(page));
2063                         BUG_ON(PageWriteback(page));
2064
2065                         /*
2066                          * If the page does not have buffers (for
2067                          * whatever reason), try to create them using
2068                          * block_prepare_write.  If this fails,
2069                          * redirty the page and move on.
2070                          */
2071                         if (!page_has_buffers(page)) {
2072                                 if (block_prepare_write(page, 0, len,
2073                                                 noalloc_get_block_write)) {
2074                                 redirty_page:
2075                                         redirty_page_for_writepage(mpd->wbc,
2076                                                                    page);
2077                                         unlock_page(page);
2078                                         continue;
2079                                 }
2080                                 commit_write = 1;
2081                         }
2082
2083                         bh = page_bufs = page_buffers(page);
2084                         block_start = 0;
2085                         do {
2086                                 if (!bh)
2087                                         goto redirty_page;
2088                                 if (map && (cur_logical >= map->m_lblk) &&
2089                                     (cur_logical <= (map->m_lblk +
2090                                                      (map->m_len - 1)))) {
2091                                         if (buffer_delay(bh)) {
2092                                                 clear_buffer_delay(bh);
2093                                                 bh->b_blocknr = pblock;
2094                                         }
2095                                         if (buffer_unwritten(bh) ||
2096                                             buffer_mapped(bh))
2097                                                 BUG_ON(bh->b_blocknr != pblock);
2098                                         if (map->m_flags & EXT4_MAP_UNINIT)
2099                                                 set_buffer_uninit(bh);
2100                                         clear_buffer_unwritten(bh);
2101                                 }
2102
2103                                 /* redirty page if block allocation undone */
2104                                 if (buffer_delay(bh) || buffer_unwritten(bh))
2105                                         redirty_page = 1;
2106                                 bh = bh->b_this_page;
2107                                 block_start += bh->b_size;
2108                                 cur_logical++;
2109                                 pblock++;
2110                         } while (bh != page_bufs);
2111
2112                         if (redirty_page)
2113                                 goto redirty_page;
2114
2115                         if (commit_write)
2116                                 /* mark the buffer_heads as dirty & uptodate */
2117                                 block_commit_write(page, 0, len);
2118
2119                         /*
2120                          * Delalloc doesn't support data journalling,
2121                          * but eventually maybe we'll lift this
2122                          * restriction.
2123                          */
2124                         if (unlikely(journal_data && PageChecked(page)))
2125                                 err = __ext4_journalled_writepage(page, len);
2126                         else
2127                                 err = ext4_bio_write_page(&io_submit, page,
2128                                                           len, mpd->wbc);
2129
2130                         if (!err)
2131                                 mpd->pages_written++;
2132                         /*
2133                          * In error case, we have to continue because
2134                          * remaining pages are still locked
2135                          */
2136                         if (ret == 0)
2137                                 ret = err;
2138                 }
2139                 pagevec_release(&pvec);
2140         }
2141         ext4_io_submit(&io_submit);
2142         return ret;
2143 }
2144
2145 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2146                                         sector_t logical, long blk_cnt)
2147 {
2148         int nr_pages, i;
2149         pgoff_t index, end;
2150         struct pagevec pvec;
2151         struct inode *inode = mpd->inode;
2152         struct address_space *mapping = inode->i_mapping;
2153
2154         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2155         end   = (logical + blk_cnt - 1) >>
2156                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2157         while (index <= end) {
2158                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2159                 if (nr_pages == 0)
2160                         break;
2161                 for (i = 0; i < nr_pages; i++) {
2162                         struct page *page = pvec.pages[i];
2163                         if (page->index > end)
2164                                 break;
2165                         BUG_ON(!PageLocked(page));
2166                         BUG_ON(PageWriteback(page));
2167                         block_invalidatepage(page, 0);
2168                         ClearPageUptodate(page);
2169                         unlock_page(page);
2170                 }
2171                 index = pvec.pages[nr_pages - 1]->index + 1;
2172                 pagevec_release(&pvec);
2173         }
2174         return;
2175 }
2176
2177 static void ext4_print_free_blocks(struct inode *inode)
2178 {
2179         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2180         printk(KERN_CRIT "Total free blocks count %lld\n",
2181                ext4_count_free_blocks(inode->i_sb));
2182         printk(KERN_CRIT "Free/Dirty block details\n");
2183         printk(KERN_CRIT "free_blocks=%lld\n",
2184                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2185         printk(KERN_CRIT "dirty_blocks=%lld\n",
2186                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2187         printk(KERN_CRIT "Block reservation details\n");
2188         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2189                EXT4_I(inode)->i_reserved_data_blocks);
2190         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2191                EXT4_I(inode)->i_reserved_meta_blocks);
2192         return;
2193 }
2194
2195 /*
2196  * mpage_da_map_and_submit - go through given space, map them
2197  *       if necessary, and then submit them for I/O
2198  *
2199  * @mpd - bh describing space
2200  *
2201  * The function skips space we know is already mapped to disk blocks.
2202  *
2203  */
2204 static void mpage_da_map_and_submit(struct mpage_da_data *mpd)
2205 {
2206         int err, blks, get_blocks_flags;
2207         struct ext4_map_blocks map, *mapp = NULL;
2208         sector_t next = mpd->b_blocknr;
2209         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2210         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2211         handle_t *handle = NULL;
2212
2213         /*
2214          * If the blocks are mapped already, or we couldn't accumulate
2215          * any blocks, then proceed immediately to the submission stage.
2216          */
2217         if ((mpd->b_size == 0) ||
2218             ((mpd->b_state  & (1 << BH_Mapped)) &&
2219              !(mpd->b_state & (1 << BH_Delay)) &&
2220              !(mpd->b_state & (1 << BH_Unwritten))))
2221                 goto submit_io;
2222
2223         handle = ext4_journal_current_handle();
2224         BUG_ON(!handle);
2225
2226         /*
2227          * Call ext4_map_blocks() to allocate any delayed allocation
2228          * blocks, or to convert an uninitialized extent to be
2229          * initialized (in the case where we have written into
2230          * one or more preallocated blocks).
2231          *
2232          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2233          * indicate that we are on the delayed allocation path.  This
2234          * affects functions in many different parts of the allocation
2235          * call path.  This flag exists primarily because we don't
2236          * want to change *many* call functions, so ext4_map_blocks()
2237          * will set the magic i_delalloc_reserved_flag once the
2238          * inode's allocation semaphore is taken.
2239          *
2240          * If the blocks in questions were delalloc blocks, set
2241          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2242          * variables are updated after the blocks have been allocated.
2243          */
2244         map.m_lblk = next;
2245         map.m_len = max_blocks;
2246         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
2247         if (ext4_should_dioread_nolock(mpd->inode))
2248                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2249         if (mpd->b_state & (1 << BH_Delay))
2250                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2251
2252         blks = ext4_map_blocks(handle, mpd->inode, &map, get_blocks_flags);
2253         if (blks < 0) {
2254                 struct super_block *sb = mpd->inode->i_sb;
2255
2256                 err = blks;
2257                 /*
2258                  * If get block returns EAGAIN or ENOSPC and there
2259                  * appears to be free blocks we will call
2260                  * ext4_writepage() for all of the pages which will
2261                  * just redirty the pages.
2262                  */
2263                 if (err == -EAGAIN)
2264                         goto submit_io;
2265
2266                 if (err == -ENOSPC &&
2267                     ext4_count_free_blocks(sb)) {
2268                         mpd->retval = err;
2269                         goto submit_io;
2270                 }
2271
2272                 /*
2273                  * get block failure will cause us to loop in
2274                  * writepages, because a_ops->writepage won't be able
2275                  * to make progress. The page will be redirtied by
2276                  * writepage and writepages will again try to write
2277                  * the same.
2278                  */
2279                 if (!(EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2280                         ext4_msg(sb, KERN_CRIT,
2281                                  "delayed block allocation failed for inode %lu "
2282                                  "at logical offset %llu with max blocks %zd "
2283                                  "with error %d", mpd->inode->i_ino,
2284                                  (unsigned long long) next,
2285                                  mpd->b_size >> mpd->inode->i_blkbits, err);
2286                         ext4_msg(sb, KERN_CRIT,
2287                                 "This should not happen!! Data will be lost\n");
2288                         if (err == -ENOSPC)
2289                                 ext4_print_free_blocks(mpd->inode);
2290                 }
2291                 /* invalidate all the pages */
2292                 ext4_da_block_invalidatepages(mpd, next,
2293                                 mpd->b_size >> mpd->inode->i_blkbits);
2294                 return;
2295         }
2296         BUG_ON(blks == 0);
2297
2298         mapp = &map;
2299         if (map.m_flags & EXT4_MAP_NEW) {
2300                 struct block_device *bdev = mpd->inode->i_sb->s_bdev;
2301                 int i;
2302
2303                 for (i = 0; i < map.m_len; i++)
2304                         unmap_underlying_metadata(bdev, map.m_pblk + i);
2305         }
2306
2307         if (ext4_should_order_data(mpd->inode)) {
2308                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2309                 if (err)
2310                         /* This only happens if the journal is aborted */
2311                         return;
2312         }
2313
2314         /*
2315          * Update on-disk size along with block allocation.
2316          */
2317         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2318         if (disksize > i_size_read(mpd->inode))
2319                 disksize = i_size_read(mpd->inode);
2320         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2321                 ext4_update_i_disksize(mpd->inode, disksize);
2322                 err = ext4_mark_inode_dirty(handle, mpd->inode);
2323                 if (err)
2324                         ext4_error(mpd->inode->i_sb,
2325                                    "Failed to mark inode %lu dirty",
2326                                    mpd->inode->i_ino);
2327         }
2328
2329 submit_io:
2330         mpage_da_submit_io(mpd, mapp);
2331         mpd->io_done = 1;
2332 }
2333
2334 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2335                 (1 << BH_Delay) | (1 << BH_Unwritten))
2336
2337 /*
2338  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2339  *
2340  * @mpd->lbh - extent of blocks
2341  * @logical - logical number of the block in the file
2342  * @bh - bh of the block (used to access block's state)
2343  *
2344  * the function is used to collect contig. blocks in same state
2345  */
2346 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2347                                    sector_t logical, size_t b_size,
2348                                    unsigned long b_state)
2349 {
2350         sector_t next;
2351         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2352
2353         /*
2354          * XXX Don't go larger than mballoc is willing to allocate
2355          * This is a stopgap solution.  We eventually need to fold
2356          * mpage_da_submit_io() into this function and then call
2357          * ext4_map_blocks() multiple times in a loop
2358          */
2359         if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
2360                 goto flush_it;
2361
2362         /* check if thereserved journal credits might overflow */
2363         if (!(ext4_test_inode_flag(mpd->inode, EXT4_INODE_EXTENTS))) {
2364                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2365                         /*
2366                          * With non-extent format we are limited by the journal
2367                          * credit available.  Total credit needed to insert
2368                          * nrblocks contiguous blocks is dependent on the
2369                          * nrblocks.  So limit nrblocks.
2370                          */
2371                         goto flush_it;
2372                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2373                                 EXT4_MAX_TRANS_DATA) {
2374                         /*
2375                          * Adding the new buffer_head would make it cross the
2376                          * allowed limit for which we have journal credit
2377                          * reserved. So limit the new bh->b_size
2378                          */
2379                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2380                                                 mpd->inode->i_blkbits;
2381                         /* we will do mpage_da_submit_io in the next loop */
2382                 }
2383         }
2384         /*
2385          * First block in the extent
2386          */
2387         if (mpd->b_size == 0) {
2388                 mpd->b_blocknr = logical;
2389                 mpd->b_size = b_size;
2390                 mpd->b_state = b_state & BH_FLAGS;
2391                 return;
2392         }
2393
2394         next = mpd->b_blocknr + nrblocks;
2395         /*
2396          * Can we merge the block to our big extent?
2397          */
2398         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2399                 mpd->b_size += b_size;
2400                 return;
2401         }
2402
2403 flush_it:
2404         /*
2405          * We couldn't merge the block to our extent, so we
2406          * need to flush current  extent and start new one
2407          */
2408         mpage_da_map_and_submit(mpd);
2409         return;
2410 }
2411
2412 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
2413 {
2414         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
2415 }
2416
2417 /*
2418  * __mpage_da_writepage - finds extent of pages and blocks
2419  *
2420  * @page: page to consider
2421  * @wbc: not used, we just follow rules
2422  * @data: context
2423  *
2424  * The function finds extents of pages and scan them for all blocks.
2425  */
2426 static int __mpage_da_writepage(struct page *page,
2427                                 struct writeback_control *wbc,
2428                                 struct mpage_da_data *mpd)
2429 {
2430         struct inode *inode = mpd->inode;
2431         struct buffer_head *bh, *head;
2432         sector_t logical;
2433
2434         /*
2435          * Can we merge this page to current extent?
2436          */
2437         if (mpd->next_page != page->index) {
2438                 /*
2439                  * Nope, we can't. So, we map non-allocated blocks
2440                  * and start IO on them
2441                  */
2442                 if (mpd->next_page != mpd->first_page) {
2443                         mpage_da_map_and_submit(mpd);
2444                         /*
2445                          * skip rest of the page in the page_vec
2446                          */
2447                         redirty_page_for_writepage(wbc, page);
2448                         unlock_page(page);
2449                         return MPAGE_DA_EXTENT_TAIL;
2450                 }
2451
2452                 /*
2453                  * Start next extent of pages ...
2454                  */
2455                 mpd->first_page = page->index;
2456
2457                 /*
2458                  * ... and blocks
2459                  */
2460                 mpd->b_size = 0;
2461                 mpd->b_state = 0;
2462                 mpd->b_blocknr = 0;
2463         }
2464
2465         mpd->next_page = page->index + 1;
2466         logical = (sector_t) page->index <<
2467                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2468
2469         if (!page_has_buffers(page)) {
2470                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2471                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2472                 if (mpd->io_done)
2473                         return MPAGE_DA_EXTENT_TAIL;
2474         } else {
2475                 /*
2476                  * Page with regular buffer heads, just add all dirty ones
2477                  */
2478                 head = page_buffers(page);
2479                 bh = head;
2480                 do {
2481                         BUG_ON(buffer_locked(bh));
2482                         /*
2483                          * We need to try to allocate
2484                          * unmapped blocks in the same page.
2485                          * Otherwise we won't make progress
2486                          * with the page in ext4_writepage
2487                          */
2488                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2489                                 mpage_add_bh_to_extent(mpd, logical,
2490                                                        bh->b_size,
2491                                                        bh->b_state);
2492                                 if (mpd->io_done)
2493                                         return MPAGE_DA_EXTENT_TAIL;
2494                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2495                                 /*
2496                                  * mapped dirty buffer. We need to update
2497                                  * the b_state because we look at
2498                                  * b_state in mpage_da_map_blocks. We don't
2499                                  * update b_size because if we find an
2500                                  * unmapped buffer_head later we need to
2501                                  * use the b_state flag of that buffer_head.
2502                                  */
2503                                 if (mpd->b_size == 0)
2504                                         mpd->b_state = bh->b_state & BH_FLAGS;
2505                         }
2506                         logical++;
2507                 } while ((bh = bh->b_this_page) != head);
2508         }
2509
2510         return 0;
2511 }
2512
2513 /*
2514  * This is a special get_blocks_t callback which is used by
2515  * ext4_da_write_begin().  It will either return mapped block or
2516  * reserve space for a single block.
2517  *
2518  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2519  * We also have b_blocknr = -1 and b_bdev initialized properly
2520  *
2521  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2522  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2523  * initialized properly.
2524  */
2525 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2526                                   struct buffer_head *bh, int create)
2527 {
2528         struct ext4_map_blocks map;
2529         int ret = 0;
2530         sector_t invalid_block = ~((sector_t) 0xffff);
2531
2532         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2533                 invalid_block = ~0;
2534
2535         BUG_ON(create == 0);
2536         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
2537
2538         map.m_lblk = iblock;
2539         map.m_len = 1;
2540
2541         /*
2542          * first, we need to know whether the block is allocated already
2543          * preallocated blocks are unmapped but should treated
2544          * the same as allocated blocks.
2545          */
2546         ret = ext4_map_blocks(NULL, inode, &map, 0);
2547         if (ret < 0)
2548                 return ret;
2549         if (ret == 0) {
2550                 if (buffer_delay(bh))
2551                         return 0; /* Not sure this could or should happen */
2552                 /*
2553                  * XXX: __block_prepare_write() unmaps passed block,
2554                  * is it OK?
2555                  */
2556                 ret = ext4_da_reserve_space(inode, iblock);
2557                 if (ret)
2558                         /* not enough space to reserve */
2559                         return ret;
2560
2561                 map_bh(bh, inode->i_sb, invalid_block);
2562                 set_buffer_new(bh);
2563                 set_buffer_delay(bh);
2564                 return 0;
2565         }
2566
2567         map_bh(bh, inode->i_sb, map.m_pblk);
2568         bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
2569
2570         if (buffer_unwritten(bh)) {
2571                 /* A delayed write to unwritten bh should be marked
2572                  * new and mapped.  Mapped ensures that we don't do
2573                  * get_block multiple times when we write to the same
2574                  * offset and new ensures that we do proper zero out
2575                  * for partial write.
2576                  */
2577                 set_buffer_new(bh);
2578                 set_buffer_mapped(bh);
2579         }
2580         return 0;
2581 }
2582
2583 /*
2584  * This function is used as a standard get_block_t calback function
2585  * when there is no desire to allocate any blocks.  It is used as a
2586  * callback function for block_prepare_write() and block_write_full_page().
2587  * These functions should only try to map a single block at a time.
2588  *
2589  * Since this function doesn't do block allocations even if the caller
2590  * requests it by passing in create=1, it is critically important that
2591  * any caller checks to make sure that any buffer heads are returned
2592  * by this function are either all already mapped or marked for
2593  * delayed allocation before calling  block_write_full_page().  Otherwise,
2594  * b_blocknr could be left unitialized, and the page write functions will
2595  * be taken by surprise.
2596  */
2597 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2598                                    struct buffer_head *bh_result, int create)
2599 {
2600         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2601         return _ext4_get_block(inode, iblock, bh_result, 0);
2602 }
2603
2604 static int bget_one(handle_t *handle, struct buffer_head *bh)
2605 {
2606         get_bh(bh);
2607         return 0;
2608 }
2609
2610 static int bput_one(handle_t *handle, struct buffer_head *bh)
2611 {
2612         put_bh(bh);
2613         return 0;
2614 }
2615
2616 static int __ext4_journalled_writepage(struct page *page,
2617                                        unsigned int len)
2618 {
2619         struct address_space *mapping = page->mapping;
2620         struct inode *inode = mapping->host;
2621         struct buffer_head *page_bufs;
2622         handle_t *handle = NULL;
2623         int ret = 0;
2624         int err;
2625
2626         ClearPageChecked(page);
2627         page_bufs = page_buffers(page);
2628         BUG_ON(!page_bufs);
2629         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2630         /* As soon as we unlock the page, it can go away, but we have
2631          * references to buffers so we are safe */
2632         unlock_page(page);
2633
2634         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2635         if (IS_ERR(handle)) {
2636                 ret = PTR_ERR(handle);
2637                 goto out;
2638         }
2639
2640         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2641                                 do_journal_get_write_access);
2642
2643         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2644                                 write_end_fn);
2645         if (ret == 0)
2646                 ret = err;
2647         err = ext4_journal_stop(handle);
2648         if (!ret)
2649                 ret = err;
2650
2651         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2652         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2653 out:
2654         return ret;
2655 }
2656
2657 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
2658 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
2659
2660 /*
2661  * Note that we don't need to start a transaction unless we're journaling data
2662  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2663  * need to file the inode to the transaction's list in ordered mode because if
2664  * we are writing back data added by write(), the inode is already there and if
2665  * we are writing back data modified via mmap(), noone guarantees in which
2666  * transaction the data will hit the disk. In case we are journaling data, we
2667  * cannot start transaction directly because transaction start ranks above page
2668  * lock so we have to do some magic.
2669  *
2670  * This function can get called via...
2671  *   - ext4_da_writepages after taking page lock (have journal handle)
2672  *   - journal_submit_inode_data_buffers (no journal handle)
2673  *   - shrink_page_list via pdflush (no journal handle)
2674  *   - grab_page_cache when doing write_begin (have journal handle)
2675  *
2676  * We don't do any block allocation in this function. If we have page with
2677  * multiple blocks we need to write those buffer_heads that are mapped. This
2678  * is important for mmaped based write. So if we do with blocksize 1K
2679  * truncate(f, 1024);
2680  * a = mmap(f, 0, 4096);
2681  * a[0] = 'a';
2682  * truncate(f, 4096);
2683  * we have in the page first buffer_head mapped via page_mkwrite call back
2684  * but other bufer_heads would be unmapped but dirty(dirty done via the
2685  * do_wp_page). So writepage should write the first block. If we modify
2686  * the mmap area beyond 1024 we will again get a page_fault and the
2687  * page_mkwrite callback will do the block allocation and mark the
2688  * buffer_heads mapped.
2689  *
2690  * We redirty the page if we have any buffer_heads that is either delay or
2691  * unwritten in the page.
2692  *
2693  * We can get recursively called as show below.
2694  *
2695  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2696  *              ext4_writepage()
2697  *
2698  * But since we don't do any block allocation we should not deadlock.
2699  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2700  */
2701 static int ext4_writepage(struct page *page,
2702                           struct writeback_control *wbc)
2703 {
2704         int ret = 0, commit_write = 0;
2705         loff_t size;
2706         unsigned int len;
2707         struct buffer_head *page_bufs = NULL;
2708         struct inode *inode = page->mapping->host;
2709
2710         trace_ext4_writepage(inode, page);
2711         size = i_size_read(inode);
2712         if (page->index == size >> PAGE_CACHE_SHIFT)
2713                 len = size & ~PAGE_CACHE_MASK;
2714         else
2715                 len = PAGE_CACHE_SIZE;
2716
2717         /*
2718          * If the page does not have buffers (for whatever reason),
2719          * try to create them using block_prepare_write.  If this
2720          * fails, redirty the page and move on.
2721          */
2722         if (!page_buffers(page)) {
2723                 if (block_prepare_write(page, 0, len,
2724                                         noalloc_get_block_write)) {
2725                 redirty_page:
2726                         redirty_page_for_writepage(wbc, page);
2727                         unlock_page(page);
2728                         return 0;
2729                 }
2730                 commit_write = 1;
2731         }
2732         page_bufs = page_buffers(page);
2733         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2734                               ext4_bh_delay_or_unwritten)) {
2735                 /*
2736                  * We don't want to do block allocation So redirty the
2737                  * page and return We may reach here when we do a
2738                  * journal commit via
2739                  * journal_submit_inode_data_buffers.  If we don't
2740                  * have mapping block we just ignore them. We can also
2741                  * reach here via shrink_page_list
2742                  */
2743                 goto redirty_page;
2744         }
2745         if (commit_write)
2746                 /* now mark the buffer_heads as dirty and uptodate */
2747                 block_commit_write(page, 0, len);
2748
2749         if (PageChecked(page) && ext4_should_journal_data(inode))
2750                 /*
2751                  * It's mmapped pagecache.  Add buffers and journal it.  There
2752                  * doesn't seem much point in redirtying the page here.
2753                  */
2754                 return __ext4_journalled_writepage(page, len);
2755
2756         if (buffer_uninit(page_bufs)) {
2757                 ext4_set_bh_endio(page_bufs, inode);
2758                 ret = block_write_full_page_endio(page, noalloc_get_block_write,
2759                                             wbc, ext4_end_io_buffer_write);
2760         } else
2761                 ret = block_write_full_page(page, noalloc_get_block_write,
2762                                             wbc);
2763
2764         return ret;
2765 }
2766
2767 /*
2768  * This is called via ext4_da_writepages() to
2769  * calulate the total number of credits to reserve to fit
2770  * a single extent allocation into a single transaction,
2771  * ext4_da_writpeages() will loop calling this before
2772  * the block allocation.
2773  */
2774
2775 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2776 {
2777         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2778
2779         /*
2780          * With non-extent format the journal credit needed to
2781          * insert nrblocks contiguous block is dependent on
2782          * number of contiguous block. So we will limit
2783          * number of contiguous block to a sane value
2784          */
2785         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) &&
2786             (max_blocks > EXT4_MAX_TRANS_DATA))
2787                 max_blocks = EXT4_MAX_TRANS_DATA;
2788
2789         return ext4_chunk_trans_blocks(inode, max_blocks);
2790 }
2791
2792 /*
2793  * write_cache_pages_da - walk the list of dirty pages of the given
2794  * address space and call the callback function (which usually writes
2795  * the pages).
2796  *
2797  * This is a forked version of write_cache_pages().  Differences:
2798  *      Range cyclic is ignored.
2799  *      no_nrwrite_index_update is always presumed true
2800  */
2801 static int write_cache_pages_da(struct address_space *mapping,
2802                                 struct writeback_control *wbc,
2803                                 struct mpage_da_data *mpd,
2804                                 pgoff_t *done_index)
2805 {
2806         int ret = 0;
2807         int done = 0;
2808         struct pagevec pvec;
2809         unsigned nr_pages;
2810         pgoff_t index;
2811         pgoff_t end;            /* Inclusive */
2812         long nr_to_write = wbc->nr_to_write;
2813         int tag;
2814
2815         pagevec_init(&pvec, 0);
2816         index = wbc->range_start >> PAGE_CACHE_SHIFT;
2817         end = wbc->range_end >> PAGE_CACHE_SHIFT;
2818
2819         if (wbc->sync_mode == WB_SYNC_ALL)
2820                 tag = PAGECACHE_TAG_TOWRITE;
2821         else
2822                 tag = PAGECACHE_TAG_DIRTY;
2823
2824         *done_index = index;
2825         while (!done && (index <= end)) {
2826                 int i;
2827
2828                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2829                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2830                 if (nr_pages == 0)
2831                         break;
2832
2833                 for (i = 0; i < nr_pages; i++) {
2834                         struct page *page = pvec.pages[i];
2835
2836                         /*
2837                          * At this point, the page may be truncated or
2838                          * invalidated (changing page->mapping to NULL), or
2839                          * even swizzled back from swapper_space to tmpfs file
2840                          * mapping. However, page->index will not change
2841                          * because we have a reference on the page.
2842                          */
2843                         if (page->index > end) {
2844                                 done = 1;
2845                                 break;
2846                         }
2847
2848                         *done_index = page->index + 1;
2849
2850                         lock_page(page);
2851
2852                         /*
2853                          * Page truncated or invalidated. We can freely skip it
2854                          * then, even for data integrity operations: the page
2855                          * has disappeared concurrently, so there could be no
2856                          * real expectation of this data interity operation
2857                          * even if there is now a new, dirty page at the same
2858                          * pagecache address.
2859                          */
2860                         if (unlikely(page->mapping != mapping)) {
2861 continue_unlock:
2862                                 unlock_page(page);
2863                                 continue;
2864                         }
2865
2866                         if (!PageDirty(page)) {
2867                                 /* someone wrote it for us */
2868                                 goto continue_unlock;
2869                         }
2870
2871                         if (PageWriteback(page)) {
2872                                 if (wbc->sync_mode != WB_SYNC_NONE)
2873                                         wait_on_page_writeback(page);
2874                                 else
2875                                         goto continue_unlock;
2876                         }
2877
2878                         BUG_ON(PageWriteback(page));
2879                         if (!clear_page_dirty_for_io(page))
2880                                 goto continue_unlock;
2881
2882                         ret = __mpage_da_writepage(page, wbc, mpd);
2883                         if (unlikely(ret)) {
2884                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2885                                         unlock_page(page);
2886                                         ret = 0;
2887                                 } else {
2888                                         done = 1;
2889                                         break;
2890                                 }
2891                         }
2892
2893                         if (nr_to_write > 0) {
2894                                 nr_to_write--;
2895                                 if (nr_to_write == 0 &&
2896                                     wbc->sync_mode == WB_SYNC_NONE) {
2897                                         /*
2898                                          * We stop writing back only if we are
2899                                          * not doing integrity sync. In case of
2900                                          * integrity sync we have to keep going
2901                                          * because someone may be concurrently
2902                                          * dirtying pages, and we might have
2903                                          * synced a lot of newly appeared dirty
2904                                          * pages, but have not synced all of the
2905                                          * old dirty pages.
2906                                          */
2907                                         done = 1;
2908                                         break;
2909                                 }
2910                         }
2911                 }
2912                 pagevec_release(&pvec);
2913                 cond_resched();
2914         }
2915         return ret;
2916 }
2917
2918
2919 static int ext4_da_writepages(struct address_space *mapping,
2920                               struct writeback_control *wbc)
2921 {
2922         pgoff_t index;
2923         int range_whole = 0;
2924         handle_t *handle = NULL;
2925         struct mpage_da_data mpd;
2926         struct inode *inode = mapping->host;
2927         int pages_written = 0;
2928         long pages_skipped;
2929         unsigned int max_pages;
2930         int range_cyclic, cycled = 1, io_done = 0;
2931         int needed_blocks, ret = 0;
2932         long desired_nr_to_write, nr_to_writebump = 0;
2933         loff_t range_start = wbc->range_start;
2934         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2935         pgoff_t done_index = 0;
2936         pgoff_t end;
2937
2938         trace_ext4_da_writepages(inode, wbc);
2939
2940         /*
2941          * No pages to write? This is mainly a kludge to avoid starting
2942          * a transaction for special inodes like journal inode on last iput()
2943          * because that could violate lock ordering on umount
2944          */
2945         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2946                 return 0;
2947
2948         /*
2949          * If the filesystem has aborted, it is read-only, so return
2950          * right away instead of dumping stack traces later on that
2951          * will obscure the real source of the problem.  We test
2952          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2953          * the latter could be true if the filesystem is mounted
2954          * read-only, and in that case, ext4_da_writepages should
2955          * *never* be called, so if that ever happens, we would want
2956          * the stack trace.
2957          */
2958         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2959                 return -EROFS;
2960
2961         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2962                 range_whole = 1;
2963
2964         range_cyclic = wbc->range_cyclic;
2965         if (wbc->range_cyclic) {
2966                 index = mapping->writeback_index;
2967                 if (index)
2968                         cycled = 0;
2969                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2970                 wbc->range_end  = LLONG_MAX;
2971                 wbc->range_cyclic = 0;
2972                 end = -1;
2973         } else {
2974                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2975                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2976         }
2977
2978         /*
2979          * This works around two forms of stupidity.  The first is in
2980          * the writeback code, which caps the maximum number of pages
2981          * written to be 1024 pages.  This is wrong on multiple
2982          * levels; different architectues have a different page size,
2983          * which changes the maximum amount of data which gets
2984          * written.  Secondly, 4 megabytes is way too small.  XFS
2985          * forces this value to be 16 megabytes by multiplying
2986          * nr_to_write parameter by four, and then relies on its
2987          * allocator to allocate larger extents to make them
2988          * contiguous.  Unfortunately this brings us to the second
2989          * stupidity, which is that ext4's mballoc code only allocates
2990          * at most 2048 blocks.  So we force contiguous writes up to
2991          * the number of dirty blocks in the inode, or
2992          * sbi->max_writeback_mb_bump whichever is smaller.
2993          */
2994         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2995         if (!range_cyclic && range_whole) {
2996                 if (wbc->nr_to_write == LONG_MAX)
2997                         desired_nr_to_write = wbc->nr_to_write;
2998                 else
2999                         desired_nr_to_write = wbc->nr_to_write * 8;
3000         } else
3001                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
3002                                                            max_pages);
3003         if (desired_nr_to_write > max_pages)
3004                 desired_nr_to_write = max_pages;
3005
3006         if (wbc->nr_to_write < desired_nr_to_write) {
3007                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
3008                 wbc->nr_to_write = desired_nr_to_write;
3009         }
3010
3011         mpd.wbc = wbc;
3012         mpd.inode = mapping->host;
3013
3014         pages_skipped = wbc->pages_skipped;
3015
3016 retry:
3017         if (wbc->sync_mode == WB_SYNC_ALL)
3018                 tag_pages_for_writeback(mapping, index, end);
3019
3020         while (!ret && wbc->nr_to_write > 0) {
3021
3022                 /*
3023                  * we  insert one extent at a time. So we need
3024                  * credit needed for single extent allocation.
3025                  * journalled mode is currently not supported
3026                  * by delalloc
3027                  */
3028                 BUG_ON(ext4_should_journal_data(inode));
3029                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
3030
3031                 /* start a new transaction*/
3032                 handle = ext4_journal_start(inode, needed_blocks);
3033                 if (IS_ERR(handle)) {
3034                         ret = PTR_ERR(handle);
3035                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
3036                                "%ld pages, ino %lu; err %d", __func__,
3037                                 wbc->nr_to_write, inode->i_ino, ret);
3038                         goto out_writepages;
3039                 }
3040
3041                 /*
3042                  * Now call __mpage_da_writepage to find the next
3043                  * contiguous region of logical blocks that need
3044                  * blocks to be allocated by ext4.  We don't actually
3045                  * submit the blocks for I/O here, even though
3046                  * write_cache_pages thinks it will, and will set the
3047                  * pages as clean for write before calling
3048                  * __mpage_da_writepage().
3049                  */
3050                 mpd.b_size = 0;
3051                 mpd.b_state = 0;
3052                 mpd.b_blocknr = 0;
3053                 mpd.first_page = 0;
3054                 mpd.next_page = 0;
3055                 mpd.io_done = 0;
3056                 mpd.pages_written = 0;
3057                 mpd.retval = 0;
3058                 ret = write_cache_pages_da(mapping, wbc, &mpd, &done_index);
3059                 /*
3060                  * If we have a contiguous extent of pages and we
3061                  * haven't done the I/O yet, map the blocks and submit
3062                  * them for I/O.
3063                  */
3064                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
3065                         mpage_da_map_and_submit(&mpd);
3066                         ret = MPAGE_DA_EXTENT_TAIL;
3067                 }
3068                 trace_ext4_da_write_pages(inode, &mpd);
3069                 wbc->nr_to_write -= mpd.pages_written;
3070
3071                 ext4_journal_stop(handle);
3072
3073                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
3074                         /* commit the transaction which would
3075                          * free blocks released in the transaction
3076                          * and try again
3077                          */
3078                         jbd2_journal_force_commit_nested(sbi->s_journal);
3079                         wbc->pages_skipped = pages_skipped;
3080                         ret = 0;
3081                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
3082                         /*
3083                          * got one extent now try with
3084                          * rest of the pages
3085                          */
3086                         pages_written += mpd.pages_written;
3087                         wbc->pages_skipped = pages_skipped;
3088                         ret = 0;
3089                         io_done = 1;
3090                 } else if (wbc->nr_to_write)
3091                         /*
3092                          * There is no more writeout needed
3093                          * or we requested for a noblocking writeout
3094                          * and we found the device congested
3095                          */
3096                         break;
3097         }
3098         if (!io_done && !cycled) {
3099                 cycled = 1;
3100                 index = 0;
3101                 wbc->range_start = index << PAGE_CACHE_SHIFT;
3102                 wbc->range_end  = mapping->writeback_index - 1;
3103                 goto retry;
3104         }
3105         if (pages_skipped != wbc->pages_skipped)
3106                 ext4_msg(inode->i_sb, KERN_CRIT,
3107                          "This should not happen leaving %s "
3108                          "with nr_to_write = %ld ret = %d",
3109                          __func__, wbc->nr_to_write, ret);
3110
3111         /* Update index */
3112         wbc->range_cyclic = range_cyclic;
3113         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
3114                 /*
3115                  * set the writeback_index so that range_cyclic
3116                  * mode will write it back later
3117                  */
3118                 mapping->writeback_index = done_index;
3119
3120 out_writepages:
3121         wbc->nr_to_write -= nr_to_writebump;
3122         wbc->range_start = range_start;
3123         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
3124         return ret;
3125 }
3126
3127 #define FALL_BACK_TO_NONDELALLOC 1
3128 static int ext4_nonda_switch(struct super_block *sb)
3129 {
3130         s64 free_blocks, dirty_blocks;
3131         struct ext4_sb_info *sbi = EXT4_SB(sb);
3132
3133         /*
3134          * switch to non delalloc mode if we are running low
3135          * on free block. The free block accounting via percpu
3136          * counters can get slightly wrong with percpu_counter_batch getting
3137          * accumulated on each CPU without updating global counters
3138          * Delalloc need an accurate free block accounting. So switch
3139          * to non delalloc when we are near to error range.
3140          */
3141         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
3142         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
3143         if (2 * free_blocks < 3 * dirty_blocks ||
3144                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
3145                 /*
3146                  * free block count is less than 150% of dirty blocks
3147                  * or free blocks is less than watermark
3148                  */
3149                 return 1;
3150         }
3151         /*
3152          * Even if we don't switch but are nearing capacity,
3153          * start pushing delalloc when 1/2 of free blocks are dirty.
3154          */
3155         if (free_blocks < 2 * dirty_blocks)
3156                 writeback_inodes_sb_if_idle(sb);
3157
3158         return 0;
3159 }
3160
3161 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3162                                loff_t pos, unsigned len, unsigned flags,
3163                                struct page **pagep, void **fsdata)
3164 {
3165         int ret, retries = 0;
3166         struct page *page;
3167         pgoff_t index;
3168         struct inode *inode = mapping->host;
3169         handle_t *handle;
3170
3171         index = pos >> PAGE_CACHE_SHIFT;
3172
3173         if (ext4_nonda_switch(inode->i_sb)) {
3174                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3175                 return ext4_write_begin(file, mapping, pos,
3176                                         len, flags, pagep, fsdata);
3177         }
3178         *fsdata = (void *)0;
3179         trace_ext4_da_write_begin(inode, pos, len, flags);
3180 retry:
3181         /*
3182          * With delayed allocation, we don't log the i_disksize update
3183          * if there is delayed block allocation. But we still need
3184          * to journalling the i_disksize update if writes to the end
3185          * of file which has an already mapped buffer.
3186          */
3187         handle = ext4_journal_start(inode, 1);
3188         if (IS_ERR(handle)) {
3189                 ret = PTR_ERR(handle);
3190                 goto out;
3191         }
3192         /* We cannot recurse into the filesystem as the transaction is already
3193          * started */
3194         flags |= AOP_FLAG_NOFS;
3195
3196         page = grab_cache_page_write_begin(mapping, index, flags);
3197         if (!page) {
3198                 ext4_journal_stop(handle);
3199                 ret = -ENOMEM;
3200                 goto out;
3201         }
3202         *pagep = page;
3203
3204         ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
3205         if (ret < 0) {
3206                 unlock_page(page);
3207                 ext4_journal_stop(handle);
3208                 page_cache_release(page);
3209                 /*
3210                  * block_write_begin may have instantiated a few blocks
3211                  * outside i_size.  Trim these off again. Don't need
3212                  * i_size_read because we hold i_mutex.
3213                  */
3214                 if (pos + len > inode->i_size)
3215                         ext4_truncate_failed_write(inode);
3216         }
3217
3218         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3219                 goto retry;
3220 out:
3221         return ret;
3222 }
3223
3224 /*
3225  * Check if we should update i_disksize
3226  * when write to the end of file but not require block allocation
3227  */
3228 static int ext4_da_should_update_i_disksize(struct page *page,
3229                                             unsigned long offset)
3230 {
3231         struct buffer_head *bh;
3232         struct inode *inode = page->mapping->host;
3233         unsigned int idx;
3234         int i;
3235
3236         bh = page_buffers(page);
3237         idx = offset >> inode->i_blkbits;
3238
3239         for (i = 0; i < idx; i++)
3240                 bh = bh->b_this_page;
3241
3242         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3243                 return 0;
3244         return 1;
3245 }
3246
3247 static int ext4_da_write_end(struct file *file,
3248                              struct address_space *mapping,
3249                              loff_t pos, unsigned len, unsigned copied,
3250                              struct page *page, void *fsdata)
3251 {
3252         struct inode *inode = mapping->host;
3253         int ret = 0, ret2;
3254         handle_t *handle = ext4_journal_current_handle();
3255         loff_t new_i_size;
3256         unsigned long start, end;
3257         int write_mode = (int)(unsigned long)fsdata;
3258
3259         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
3260                 if (ext4_should_order_data(inode)) {
3261                         return ext4_ordered_write_end(file, mapping, pos,
3262                                         len, copied, page, fsdata);
3263                 } else if (ext4_should_writeback_data(inode)) {
3264                         return ext4_writeback_write_end(file, mapping, pos,
3265                                         len, copied, page, fsdata);
3266                 } else {
3267                         BUG();
3268                 }
3269         }
3270
3271         trace_ext4_da_write_end(inode, pos, len, copied);
3272         start = pos & (PAGE_CACHE_SIZE - 1);
3273         end = start + copied - 1;
3274
3275         /*
3276          * generic_write_end() will run mark_inode_dirty() if i_size
3277          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3278          * into that.
3279          */
3280
3281         new_i_size = pos + copied;
3282         if (new_i_size > EXT4_I(inode)->i_disksize) {
3283                 if (ext4_da_should_update_i_disksize(page, end)) {
3284                         down_write(&EXT4_I(inode)->i_data_sem);
3285                         if (new_i_size > EXT4_I(inode)->i_disksize) {
3286                                 /*
3287                                  * Updating i_disksize when extending file
3288                                  * without needing block allocation
3289                                  */
3290                                 if (ext4_should_order_data(inode))
3291                                         ret = ext4_jbd2_file_inode(handle,
3292                                                                    inode);
3293
3294                                 EXT4_I(inode)->i_disksize = new_i_size;
3295                         }
3296                         up_write(&EXT4_I(inode)->i_data_sem);
3297                         /* We need to mark inode dirty even if
3298                          * new_i_size is less that inode->i_size
3299                          * bu greater than i_disksize.(hint delalloc)
3300                          */
3301                         ext4_mark_inode_dirty(handle, inode);
3302                 }
3303         }
3304         ret2 = generic_write_end(file, mapping, pos, len, copied,
3305                                                         page, fsdata);
3306         copied = ret2;
3307         if (ret2 < 0)
3308                 ret = ret2;
3309         ret2 = ext4_journal_stop(handle);
3310         if (!ret)
3311                 ret = ret2;
3312
3313         return ret ? ret : copied;
3314 }
3315
3316 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
3317 {
3318         /*
3319          * Drop reserved blocks
3320          */
3321         BUG_ON(!PageLocked(page));
3322         if (!page_has_buffers(page))
3323                 goto out;
3324
3325         ext4_da_page_release_reservation(page, offset);
3326
3327 out:
3328         ext4_invalidatepage(page, offset);
3329
3330         return;
3331 }
3332
3333 /*
3334  * Force all delayed allocation blocks to be allocated for a given inode.
3335  */
3336 int ext4_alloc_da_blocks(struct inode *inode)
3337 {
3338         trace_ext4_alloc_da_blocks(inode);
3339
3340         if (!EXT4_I(inode)->i_reserved_data_blocks &&
3341             !EXT4_I(inode)->i_reserved_meta_blocks)
3342                 return 0;
3343
3344         /*
3345          * We do something simple for now.  The filemap_flush() will
3346          * also start triggering a write of the data blocks, which is
3347          * not strictly speaking necessary (and for users of
3348          * laptop_mode, not even desirable).  However, to do otherwise
3349          * would require replicating code paths in:
3350          *
3351          * ext4_da_writepages() ->
3352          *    write_cache_pages() ---> (via passed in callback function)
3353          *        __mpage_da_writepage() -->
3354          *           mpage_add_bh_to_extent()
3355          *           mpage_da_map_blocks()
3356          *
3357          * The problem is that write_cache_pages(), located in
3358          * mm/page-writeback.c, marks pages clean in preparation for
3359          * doing I/O, which is not desirable if we're not planning on
3360          * doing I/O at all.
3361          *
3362          * We could call write_cache_pages(), and then redirty all of
3363          * the pages by calling redirty_page_for_writeback() but that
3364          * would be ugly in the extreme.  So instead we would need to
3365          * replicate parts of the code in the above functions,
3366          * simplifying them becuase we wouldn't actually intend to
3367          * write out the pages, but rather only collect contiguous
3368          * logical block extents, call the multi-block allocator, and
3369          * then update the buffer heads with the block allocations.
3370          *
3371          * For now, though, we'll cheat by calling filemap_flush(),
3372          * which will map the blocks, and start the I/O, but not
3373          * actually wait for the I/O to complete.
3374          */
3375         return filemap_flush(inode->i_mapping);
3376 }
3377
3378 /*
3379  * bmap() is special.  It gets used by applications such as lilo and by
3380  * the swapper to find the on-disk block of a specific piece of data.
3381  *
3382  * Naturally, this is dangerous if the block concerned is still in the
3383  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3384  * filesystem and enables swap, then they may get a nasty shock when the
3385  * data getting swapped to that swapfile suddenly gets overwritten by
3386  * the original zero's written out previously to the journal and
3387  * awaiting writeback in the kernel's buffer cache.
3388  *
3389  * So, if we see any bmap calls here on a modified, data-journaled file,
3390  * take extra steps to flush any blocks which might be in the cache.
3391  */
3392 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3393 {
3394         struct inode *inode = mapping->host;
3395         journal_t *journal;
3396         int err;
3397
3398         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3399                         test_opt(inode->i_sb, DELALLOC)) {
3400                 /*
3401                  * With delalloc we want to sync the file
3402                  * so that we can make sure we allocate
3403                  * blocks for file
3404                  */
3405                 filemap_write_and_wait(mapping);
3406         }
3407
3408         if (EXT4_JOURNAL(inode) &&
3409             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3410                 /*
3411                  * This is a REALLY heavyweight approach, but the use of
3412                  * bmap on dirty files is expected to be extremely rare:
3413                  * only if we run lilo or swapon on a freshly made file
3414                  * do we expect this to happen.
3415                  *
3416                  * (bmap requires CAP_SYS_RAWIO so this does not
3417                  * represent an unprivileged user DOS attack --- we'd be
3418                  * in trouble if mortal users could trigger this path at
3419                  * will.)
3420                  *
3421                  * NB. EXT4_STATE_JDATA is not set on files other than
3422                  * regular files.  If somebody wants to bmap a directory
3423                  * or symlink and gets confused because the buffer
3424                  * hasn't yet been flushed to disk, they deserve
3425                  * everything they get.
3426                  */
3427
3428                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3429                 journal = EXT4_JOURNAL(inode);
3430                 jbd2_journal_lock_updates(journal);
3431                 err = jbd2_journal_flush(journal);
3432                 jbd2_journal_unlock_updates(journal);
3433
3434                 if (err)
3435                         return 0;
3436         }
3437
3438         return generic_block_bmap(mapping, block, ext4_get_block);
3439 }
3440
3441 static int ext4_readpage(struct file *file, struct page *page)
3442 {
3443         return mpage_readpage(page, ext4_get_block);
3444 }
3445
3446 static int
3447 ext4_readpages(struct file *file, struct address_space *mapping,
3448                 struct list_head *pages, unsigned nr_pages)
3449 {
3450         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3451 }
3452
3453 static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
3454 {
3455         struct buffer_head *head, *bh;
3456         unsigned int curr_off = 0;
3457
3458         if (!page_has_buffers(page))
3459                 return;
3460         head = bh = page_buffers(page);
3461         do {
3462                 if (offset <= curr_off && test_clear_buffer_uninit(bh)
3463                                         && bh->b_private) {
3464                         ext4_free_io_end(bh->b_private);
3465                         bh->b_private = NULL;
3466                         bh->b_end_io = NULL;
3467                 }
3468                 curr_off = curr_off + bh->b_size;
3469                 bh = bh->b_this_page;
3470         } while (bh != head);
3471 }
3472
3473 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3474 {
3475         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3476
3477         /*
3478          * free any io_end structure allocated for buffers to be discarded
3479          */
3480         if (ext4_should_dioread_nolock(page->mapping->host))
3481                 ext4_invalidatepage_free_endio(page, offset);
3482         /*
3483          * If it's a full truncate we just forget about the pending dirtying
3484          */
3485         if (offset == 0)
3486                 ClearPageChecked(page);
3487
3488         if (journal)
3489                 jbd2_journal_invalidatepage(journal, page, offset);
3490         else
3491                 block_invalidatepage(page, offset);
3492 }
3493
3494 static int ext4_releasepage(struct page *page, gfp_t wait)
3495 {
3496         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3497
3498         WARN_ON(PageChecked(page));
3499         if (!page_has_buffers(page))
3500                 return 0;
3501         if (journal)
3502                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3503         else
3504                 return try_to_free_buffers(page);
3505 }
3506
3507 /*
3508  * O_DIRECT for ext3 (or indirect map) based files
3509  *
3510  * If the O_DIRECT write will extend the file then add this inode to the
3511  * orphan list.  So recovery will truncate it back to the original size
3512  * if the machine crashes during the write.
3513  *
3514  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3515  * crashes then stale disk data _may_ be exposed inside the file. But current
3516  * VFS code falls back into buffered path in that case so we are safe.
3517  */
3518 static ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
3519                               const struct iovec *iov, loff_t offset,
3520                               unsigned long nr_segs)
3521 {
3522         struct file *file = iocb->ki_filp;
3523         struct inode *inode = file->f_mapping->host;
3524         struct ext4_inode_info *ei = EXT4_I(inode);
3525         handle_t *handle;
3526         ssize_t ret;
3527         int orphan = 0;
3528         size_t count = iov_length(iov, nr_segs);
3529         int retries = 0;
3530
3531         if (rw == WRITE) {
3532                 loff_t final_size = offset + count;
3533
3534                 if (final_size > inode->i_size) {
3535                         /* Credits for sb + inode write */
3536                         handle = ext4_journal_start(inode, 2);
3537                         if (IS_ERR(handle)) {
3538                                 ret = PTR_ERR(handle);
3539                                 goto out;
3540                         }
3541                         ret = ext4_orphan_add(handle, inode);
3542                         if (ret) {
3543                                 ext4_journal_stop(handle);
3544                                 goto out;
3545                         }
3546                         orphan = 1;
3547                         ei->i_disksize = inode->i_size;
3548                         ext4_journal_stop(handle);
3549                 }
3550         }
3551
3552 retry:
3553         if (rw == READ && ext4_should_dioread_nolock(inode))
3554                 ret = __blockdev_direct_IO(rw, iocb, inode,
3555                                  inode->i_sb->s_bdev, iov,
3556                                  offset, nr_segs,
3557                                  ext4_get_block, NULL, NULL, 0);
3558         else {
3559                 ret = blockdev_direct_IO(rw, iocb, inode,
3560                                  inode->i_sb->s_bdev, iov,
3561                                  offset, nr_segs,
3562                                  ext4_get_block, NULL);
3563
3564                 if (unlikely((rw & WRITE) && ret < 0)) {
3565                         loff_t isize = i_size_read(inode);
3566                         loff_t end = offset + iov_length(iov, nr_segs);
3567
3568                         if (end > isize)
3569                                 vmtruncate(inode, isize);
3570                 }
3571         }
3572         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3573                 goto retry;
3574
3575         if (orphan) {
3576                 int err;
3577
3578                 /* Credits for sb + inode write */
3579                 handle = ext4_journal_start(inode, 2);
3580                 if (IS_ERR(handle)) {
3581                         /* This is really bad luck. We've written the data
3582                          * but cannot extend i_size. Bail out and pretend
3583                          * the write failed... */
3584                         ret = PTR_ERR(handle);
3585                         if (inode->i_nlink)
3586                                 ext4_orphan_del(NULL, inode);
3587
3588                         goto out;
3589                 }
3590                 if (inode->i_nlink)
3591                         ext4_orphan_del(handle, inode);
3592                 if (ret > 0) {
3593                         loff_t end = offset + ret;
3594                         if (end > inode->i_size) {
3595                                 ei->i_disksize = end;
3596                                 i_size_write(inode, end);
3597                                 /*
3598                                  * We're going to return a positive `ret'
3599                                  * here due to non-zero-length I/O, so there's
3600                                  * no way of reporting error returns from
3601                                  * ext4_mark_inode_dirty() to userspace.  So
3602                                  * ignore it.
3603                                  */
3604                                 ext4_mark_inode_dirty(handle, inode);
3605                         }
3606                 }
3607                 err = ext4_journal_stop(handle);
3608                 if (ret == 0)
3609                         ret = err;
3610         }
3611 out:
3612         return ret;
3613 }
3614
3615 /*
3616  * ext4_get_block used when preparing for a DIO write or buffer write.
3617  * We allocate an uinitialized extent if blocks haven't been allocated.
3618  * The extent will be converted to initialized after the IO is complete.
3619  */
3620 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
3621                    struct buffer_head *bh_result, int create)
3622 {
3623         ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
3624                    inode->i_ino, create);
3625         return _ext4_get_block(inode, iblock, bh_result,
3626                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
3627 }
3628
3629 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3630                             ssize_t size, void *private, int ret,
3631                             bool is_async)
3632 {
3633         ext4_io_end_t *io_end = iocb->private;
3634         struct workqueue_struct *wq;
3635         unsigned long flags;
3636         struct ext4_inode_info *ei;
3637
3638         /* if not async direct IO or dio with 0 bytes write, just return */
3639         if (!io_end || !size)
3640                 goto out;
3641
3642         ext_debug("ext4_end_io_dio(): io_end 0x%p"
3643                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
3644                   iocb->private, io_end->inode->i_ino, iocb, offset,
3645                   size);
3646
3647         /* if not aio dio with unwritten extents, just free io and return */
3648         if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
3649                 ext4_free_io_end(io_end);
3650                 iocb->private = NULL;
3651 out:
3652                 if (is_async)
3653                         aio_complete(iocb, ret, 0);
3654                 return;
3655         }
3656
3657         io_end->offset = offset;
3658         io_end->size = size;
3659         if (is_async) {
3660                 io_end->iocb = iocb;
3661                 io_end->result = ret;
3662         }
3663         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
3664
3665         /* Add the io_end to per-inode completed aio dio list*/
3666         ei = EXT4_I(io_end->inode);
3667         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3668         list_add_tail(&io_end->list, &ei->i_completed_io_list);
3669         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3670
3671         /* queue the work to convert unwritten extents to written */
3672         queue_work(wq, &io_end->work);
3673         iocb->private = NULL;
3674 }
3675
3676 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
3677 {
3678         ext4_io_end_t *io_end = bh->b_private;
3679         struct workqueue_struct *wq;
3680         struct inode *inode;
3681         unsigned long flags;
3682
3683         if (!test_clear_buffer_uninit(bh) || !io_end)
3684                 goto out;
3685
3686         if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
3687                 printk("sb umounted, discard end_io request for inode %lu\n",
3688                         io_end->inode->i_ino);
3689                 ext4_free_io_end(io_end);
3690                 goto out;
3691         }
3692
3693         io_end->flag = EXT4_IO_END_UNWRITTEN;
3694         inode = io_end->inode;
3695
3696         /* Add the io_end to per-inode completed io list*/
3697         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3698         list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
3699         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3700
3701         wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
3702         /* queue the work to convert unwritten extents to written */
3703         queue_work(wq, &io_end->work);
3704 out:
3705         bh->b_private = NULL;
3706         bh->b_end_io = NULL;
3707         clear_buffer_uninit(bh);
3708         end_buffer_async_write(bh, uptodate);
3709 }
3710
3711 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
3712 {
3713         ext4_io_end_t *io_end;
3714         struct page *page = bh->b_page;
3715         loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
3716         size_t size = bh->b_size;
3717
3718 retry:
3719         io_end = ext4_init_io_end(inode, GFP_ATOMIC);
3720         if (!io_end) {
3721                 if (printk_ratelimit())
3722                         printk(KERN_WARNING "%s: allocation fail\n", __func__);
3723                 schedule();
3724                 goto retry;
3725         }
3726         io_end->offset = offset;
3727         io_end->size = size;
3728         /*
3729          * We need to hold a reference to the page to make sure it
3730          * doesn't get evicted before ext4_end_io_work() has a chance
3731          * to convert the extent from written to unwritten.
3732          */
3733         io_end->page = page;
3734         get_page(io_end->page);
3735
3736         bh->b_private = io_end;
3737         bh->b_end_io = ext4_end_io_buffer_write;
3738         return 0;
3739 }
3740
3741 /*
3742  * For ext4 extent files, ext4 will do direct-io write to holes,
3743  * preallocated extents, and those write extend the file, no need to
3744  * fall back to buffered IO.
3745  *
3746  * For holes, we fallocate those blocks, mark them as unintialized
3747  * If those blocks were preallocated, we mark sure they are splited, but
3748  * still keep the range to write as unintialized.
3749  *
3750  * The unwrritten extents will be converted to written when DIO is completed.
3751  * For async direct IO, since the IO may still pending when return, we
3752  * set up an end_io call back function, which will do the convertion
3753  * when async direct IO completed.
3754  *
3755  * If the O_DIRECT write will extend the file then add this inode to the
3756  * orphan list.  So recovery will truncate it back to the original size
3757  * if the machine crashes during the write.
3758  *
3759  */
3760 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
3761                               const struct iovec *iov, loff_t offset,
3762                               unsigned long nr_segs)
3763 {
3764         struct file *file = iocb->ki_filp;
3765         struct inode *inode = file->f_mapping->host;
3766         ssize_t ret;
3767         size_t count = iov_length(iov, nr_segs);
3768
3769         loff_t final_size = offset + count;
3770         if (rw == WRITE && final_size <= inode->i_size) {
3771                 /*
3772                  * We could direct write to holes and fallocate.
3773                  *
3774                  * Allocated blocks to fill the hole are marked as uninitialized
3775                  * to prevent paralel buffered read to expose the stale data
3776                  * before DIO complete the data IO.
3777                  *
3778                  * As to previously fallocated extents, ext4 get_block
3779                  * will just simply mark the buffer mapped but still
3780                  * keep the extents uninitialized.
3781                  *
3782                  * for non AIO case, we will convert those unwritten extents
3783                  * to written after return back from blockdev_direct_IO.
3784                  *
3785                  * for async DIO, the conversion needs to be defered when
3786                  * the IO is completed. The ext4 end_io callback function
3787                  * will be called to take care of the conversion work.
3788                  * Here for async case, we allocate an io_end structure to
3789                  * hook to the iocb.
3790                  */
3791                 iocb->private = NULL;
3792                 EXT4_I(inode)->cur_aio_dio = NULL;
3793                 if (!is_sync_kiocb(iocb)) {
3794                         iocb->private = ext4_init_io_end(inode, GFP_NOFS);
3795                         if (!iocb->private)
3796                                 return -ENOMEM;
3797                         /*
3798                          * we save the io structure for current async
3799                          * direct IO, so that later ext4_map_blocks()
3800                          * could flag the io structure whether there
3801                          * is a unwritten extents needs to be converted
3802                          * when IO is completed.
3803                          */
3804                         EXT4_I(inode)->cur_aio_dio = iocb->private;
3805                 }
3806
3807                 ret = blockdev_direct_IO(rw, iocb, inode,
3808                                          inode->i_sb->s_bdev, iov,
3809                                          offset, nr_segs,
3810                                          ext4_get_block_write,
3811                                          ext4_end_io_dio);
3812                 if (iocb->private)
3813                         EXT4_I(inode)->cur_aio_dio = NULL;
3814                 /*
3815                  * The io_end structure takes a reference to the inode,
3816                  * that structure needs to be destroyed and the
3817                  * reference to the inode need to be dropped, when IO is
3818                  * complete, even with 0 byte write, or failed.
3819                  *
3820                  * In the successful AIO DIO case, the io_end structure will be
3821                  * desctroyed and the reference to the inode will be dropped
3822                  * after the end_io call back function is called.
3823                  *
3824                  * In the case there is 0 byte write, or error case, since
3825                  * VFS direct IO won't invoke the end_io call back function,
3826                  * we need to free the end_io structure here.
3827                  */
3828                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
3829                         ext4_free_io_end(iocb->private);
3830                         iocb->private = NULL;
3831                 } else if (ret > 0 && ext4_test_inode_state(inode,
3832                                                 EXT4_STATE_DIO_UNWRITTEN)) {
3833                         int err;
3834                         /*
3835                          * for non AIO case, since the IO is already
3836                          * completed, we could do the convertion right here
3837                          */
3838                         err = ext4_convert_unwritten_extents(inode,
3839                                                              offset, ret);
3840                         if (err < 0)
3841                                 ret = err;
3842                         ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3843                 }
3844                 return ret;
3845         }
3846
3847         /* for write the the end of file case, we fall back to old way */
3848         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3849 }
3850
3851 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3852                               const struct iovec *iov, loff_t offset,
3853                               unsigned long nr_segs)
3854 {
3855         struct file *file = iocb->ki_filp;
3856         struct inode *inode = file->f_mapping->host;
3857
3858         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3859                 return ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
3860
3861         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3862 }
3863
3864 /*
3865  * Pages can be marked dirty completely asynchronously from ext4's journalling
3866  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3867  * much here because ->set_page_dirty is called under VFS locks.  The page is
3868  * not necessarily locked.
3869  *
3870  * We cannot just dirty the page and leave attached buffers clean, because the
3871  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3872  * or jbddirty because all the journalling code will explode.
3873  *
3874  * So what we do is to mark the page "pending dirty" and next time writepage
3875  * is called, propagate that into the buffers appropriately.
3876  */
3877 static int ext4_journalled_set_page_dirty(struct page *page)
3878 {
3879         SetPageChecked(page);
3880         return __set_page_dirty_nobuffers(page);
3881 }
3882
3883 static const struct address_space_operations ext4_ordered_aops = {
3884         .readpage               = ext4_readpage,
3885         .readpages              = ext4_readpages,
3886         .writepage              = ext4_writepage,
3887         .sync_page              = block_sync_page,
3888         .write_begin            = ext4_write_begin,
3889         .write_end              = ext4_ordered_write_end,
3890         .bmap                   = ext4_bmap,
3891         .invalidatepage         = ext4_invalidatepage,
3892         .releasepage            = ext4_releasepage,
3893         .direct_IO              = ext4_direct_IO,
3894         .migratepage            = buffer_migrate_page,
3895         .is_partially_uptodate  = block_is_partially_uptodate,
3896         .error_remove_page      = generic_error_remove_page,
3897 };
3898
3899 static const struct address_space_operations ext4_writeback_aops = {
3900         .readpage               = ext4_readpage,
3901         .readpages              = ext4_readpages,
3902         .writepage              = ext4_writepage,
3903         .sync_page              = block_sync_page,
3904         .write_begin            = ext4_write_begin,
3905         .write_end              = ext4_writeback_write_end,
3906         .bmap                   = ext4_bmap,
3907         .invalidatepage         = ext4_invalidatepage,
3908         .releasepage            = ext4_releasepage,
3909         .direct_IO              = ext4_direct_IO,
3910         .migratepage            = buffer_migrate_page,
3911         .is_partially_uptodate  = block_is_partially_uptodate,
3912         .error_remove_page      = generic_error_remove_page,
3913 };
3914
3915 static const struct address_space_operations ext4_journalled_aops = {
3916         .readpage               = ext4_readpage,
3917         .readpages              = ext4_readpages,
3918         .writepage              = ext4_writepage,
3919         .sync_page              = block_sync_page,
3920         .write_begin            = ext4_write_begin,
3921         .write_end              = ext4_journalled_write_end,
3922         .set_page_dirty         = ext4_journalled_set_page_dirty,
3923         .bmap                   = ext4_bmap,
3924         .invalidatepage         = ext4_invalidatepage,
3925         .releasepage            = ext4_releasepage,
3926         .is_partially_uptodate  = block_is_partially_uptodate,
3927         .error_remove_page      = generic_error_remove_page,
3928 };
3929
3930 static const struct address_space_operations ext4_da_aops = {
3931         .readpage               = ext4_readpage,
3932         .readpages              = ext4_readpages,
3933         .writepage              = ext4_writepage,
3934         .writepages             = ext4_da_writepages,
3935         .sync_page              = block_sync_page,
3936         .write_begin            = ext4_da_write_begin,
3937         .write_end              = ext4_da_write_end,
3938         .bmap                   = ext4_bmap,
3939         .invalidatepage         = ext4_da_invalidatepage,
3940         .releasepage            = ext4_releasepage,
3941         .direct_IO              = ext4_direct_IO,
3942         .migratepage            = buffer_migrate_page,
3943         .is_partially_uptodate  = block_is_partially_uptodate,
3944         .error_remove_page      = generic_error_remove_page,
3945 };
3946
3947 void ext4_set_aops(struct inode *inode)
3948 {
3949         if (ext4_should_order_data(inode) &&
3950                 test_opt(inode->i_sb, DELALLOC))
3951                 inode->i_mapping->a_ops = &ext4_da_aops;
3952         else if (ext4_should_order_data(inode))
3953                 inode->i_mapping->a_ops = &ext4_ordered_aops;
3954         else if (ext4_should_writeback_data(inode) &&
3955                  test_opt(inode->i_sb, DELALLOC))
3956                 inode->i_mapping->a_ops = &ext4_da_aops;
3957         else if (ext4_should_writeback_data(inode))
3958                 inode->i_mapping->a_ops = &ext4_writeback_aops;
3959         else
3960                 inode->i_mapping->a_ops = &ext4_journalled_aops;
3961 }
3962
3963 /*
3964  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3965  * up to the end of the block which corresponds to `from'.
3966  * This required during truncate. We need to physically zero the tail end
3967  * of that block so it doesn't yield old data if the file is later grown.
3968  */
3969 int ext4_block_truncate_page(handle_t *handle,
3970                 struct address_space *mapping, loff_t from)
3971 {
3972         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3973         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3974         unsigned blocksize, length, pos;
3975         ext4_lblk_t iblock;
3976         struct inode *inode = mapping->host;
3977         struct buffer_head *bh;
3978         struct page *page;
3979         int err = 0;
3980
3981         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
3982                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
3983         if (!page)
3984                 return -EINVAL;
3985
3986         blocksize = inode->i_sb->s_blocksize;
3987         length = blocksize - (offset & (blocksize - 1));
3988         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3989
3990         if (!page_has_buffers(page))
3991                 create_empty_buffers(page, blocksize, 0);
3992
3993         /* Find the buffer that contains "offset" */
3994         bh = page_buffers(page);
3995         pos = blocksize;
3996         while (offset >= pos) {
3997                 bh = bh->b_this_page;
3998                 iblock++;
3999                 pos += blocksize;
4000         }
4001
4002         err = 0;
4003         if (buffer_freed(bh)) {
4004                 BUFFER_TRACE(bh, "freed: skip");
4005                 goto unlock;
4006         }
4007
4008         if (!buffer_mapped(bh)) {
4009                 BUFFER_TRACE(bh, "unmapped");
4010                 ext4_get_block(inode, iblock, bh, 0);
4011                 /* unmapped? It's a hole - nothing to do */
4012                 if (!buffer_mapped(bh)) {
4013                         BUFFER_TRACE(bh, "still unmapped");
4014                         goto unlock;
4015                 }
4016         }
4017
4018         /* Ok, it's mapped. Make sure it's up-to-date */
4019         if (PageUptodate(page))
4020                 set_buffer_uptodate(bh);
4021
4022         if (!buffer_uptodate(bh)) {
4023                 err = -EIO;
4024                 ll_rw_block(READ, 1, &bh);
4025                 wait_on_buffer(bh);
4026                 /* Uhhuh. Read error. Complain and punt. */
4027                 if (!buffer_uptodate(bh))
4028                         goto unlock;
4029         }
4030
4031         if (ext4_should_journal_data(inode)) {
4032                 BUFFER_TRACE(bh, "get write access");
4033                 err = ext4_journal_get_write_access(handle, bh);
4034                 if (err)
4035                         goto unlock;
4036         }
4037
4038         zero_user(page, offset, length);
4039
4040         BUFFER_TRACE(bh, "zeroed end of block");
4041
4042         err = 0;
4043         if (ext4_should_journal_data(inode)) {
4044                 err = ext4_handle_dirty_metadata(handle, inode, bh);
4045         } else {
4046                 if (ext4_should_order_data(inode))
4047                         err = ext4_jbd2_file_inode(handle, inode);
4048                 mark_buffer_dirty(bh);
4049         }
4050
4051 unlock:
4052         unlock_page(page);
4053         page_cache_release(page);
4054         return err;
4055 }
4056
4057 /*
4058  * Probably it should be a library function... search for first non-zero word
4059  * or memcmp with zero_page, whatever is better for particular architecture.
4060  * Linus?
4061  */
4062 static inline int all_zeroes(__le32 *p, __le32 *q)
4063 {
4064         while (p < q)
4065                 if (*p++)
4066                         return 0;
4067         return 1;
4068 }
4069
4070 /**
4071  *      ext4_find_shared - find the indirect blocks for partial truncation.
4072  *      @inode:   inode in question
4073  *      @depth:   depth of the affected branch
4074  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
4075  *      @chain:   place to store the pointers to partial indirect blocks
4076  *      @top:     place to the (detached) top of branch
4077  *
4078  *      This is a helper function used by ext4_truncate().
4079  *
4080  *      When we do truncate() we may have to clean the ends of several
4081  *      indirect blocks but leave the blocks themselves alive. Block is
4082  *      partially truncated if some data below the new i_size is refered
4083  *      from it (and it is on the path to the first completely truncated
4084  *      data block, indeed).  We have to free the top of that path along
4085  *      with everything to the right of the path. Since no allocation
4086  *      past the truncation point is possible until ext4_truncate()
4087  *      finishes, we may safely do the latter, but top of branch may
4088  *      require special attention - pageout below the truncation point
4089  *      might try to populate it.
4090  *
4091  *      We atomically detach the top of branch from the tree, store the
4092  *      block number of its root in *@top, pointers to buffer_heads of
4093  *      partially truncated blocks - in @chain[].bh and pointers to
4094  *      their last elements that should not be removed - in
4095  *      @chain[].p. Return value is the pointer to last filled element
4096  *      of @chain.
4097  *
4098  *      The work left to caller to do the actual freeing of subtrees:
4099  *              a) free the subtree starting from *@top
4100  *              b) free the subtrees whose roots are stored in
4101  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
4102  *              c) free the subtrees growing from the inode past the @chain[0].
4103  *                      (no partially truncated stuff there).  */
4104
4105 static Indirect *ext4_find_shared(struct inode *inode, int depth,
4106                                   ext4_lblk_t offsets[4], Indirect chain[4],
4107                                   __le32 *top)
4108 {
4109         Indirect *partial, *p;
4110         int k, err;
4111
4112         *top = 0;
4113         /* Make k index the deepest non-null offset + 1 */
4114         for (k = depth; k > 1 && !offsets[k-1]; k--)
4115                 ;
4116         partial = ext4_get_branch(inode, k, offsets, chain, &err);
4117         /* Writer: pointers */
4118         if (!partial)
4119                 partial = chain + k-1;
4120         /*
4121          * If the branch acquired continuation since we've looked at it -
4122          * fine, it should all survive and (new) top doesn't belong to us.
4123          */
4124         if (!partial->key && *partial->p)
4125                 /* Writer: end */
4126                 goto no_top;
4127         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
4128                 ;
4129         /*
4130          * OK, we've found the last block that must survive. The rest of our
4131          * branch should be detached before unlocking. However, if that rest
4132          * of branch is all ours and does not grow immediately from the inode
4133          * it's easier to cheat and just decrement partial->p.
4134          */
4135         if (p == chain + k - 1 && p > chain) {
4136                 p->p--;
4137         } else {
4138                 *top = *p->p;
4139                 /* Nope, don't do this in ext4.  Must leave the tree intact */
4140 #if 0
4141                 *p->p = 0;
4142 #endif
4143         }
4144         /* Writer: end */
4145
4146         while (partial > p) {
4147                 brelse(partial->bh);
4148                 partial--;
4149         }
4150 no_top:
4151         return partial;
4152 }
4153
4154 /*
4155  * Zero a number of block pointers in either an inode or an indirect block.
4156  * If we restart the transaction we must again get write access to the
4157  * indirect block for further modification.
4158  *
4159  * We release `count' blocks on disk, but (last - first) may be greater
4160  * than `count' because there can be holes in there.
4161  */
4162 static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
4163                              struct buffer_head *bh,
4164                              ext4_fsblk_t block_to_free,
4165                              unsigned long count, __le32 *first,
4166                              __le32 *last)
4167 {
4168         __le32 *p;
4169         int     flags = EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_VALIDATED;
4170
4171         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
4172                 flags |= EXT4_FREE_BLOCKS_METADATA;
4173
4174         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
4175                                    count)) {
4176                 EXT4_ERROR_INODE(inode, "attempt to clear invalid "
4177                                  "blocks %llu len %lu",
4178                                  (unsigned long long) block_to_free, count);
4179                 return 1;
4180         }
4181
4182         if (try_to_extend_transaction(handle, inode)) {
4183                 if (bh) {
4184                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4185                         ext4_handle_dirty_metadata(handle, inode, bh);
4186                 }
4187                 ext4_mark_inode_dirty(handle, inode);
4188                 ext4_truncate_restart_trans(handle, inode,
4189                                             blocks_for_truncate(inode));
4190                 if (bh) {
4191                         BUFFER_TRACE(bh, "retaking write access");
4192                         ext4_journal_get_write_access(handle, bh);
4193                 }
4194         }
4195
4196         for (p = first; p < last; p++)
4197                 *p = 0;
4198
4199         ext4_free_blocks(handle, inode, 0, block_to_free, count, flags);
4200         return 0;
4201 }
4202
4203 /**
4204  * ext4_free_data - free a list of data blocks
4205  * @handle:     handle for this transaction
4206  * @inode:      inode we are dealing with
4207  * @this_bh:    indirect buffer_head which contains *@first and *@last
4208  * @first:      array of block numbers
4209  * @last:       points immediately past the end of array
4210  *
4211  * We are freeing all blocks refered from that array (numbers are stored as
4212  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
4213  *
4214  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
4215  * blocks are contiguous then releasing them at one time will only affect one
4216  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
4217  * actually use a lot of journal space.
4218  *
4219  * @this_bh will be %NULL if @first and @last point into the inode's direct
4220  * block pointers.
4221  */
4222 static void ext4_free_data(handle_t *handle, struct inode *inode,
4223                            struct buffer_head *this_bh,
4224                            __le32 *first, __le32 *last)
4225 {
4226         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
4227         unsigned long count = 0;            /* Number of blocks in the run */
4228         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
4229                                                corresponding to
4230                                                block_to_free */
4231         ext4_fsblk_t nr;                    /* Current block # */
4232         __le32 *p;                          /* Pointer into inode/ind
4233                                                for current block */
4234         int err;
4235
4236         if (this_bh) {                          /* For indirect block */
4237                 BUFFER_TRACE(this_bh, "get_write_access");
4238                 err = ext4_journal_get_write_access(handle, this_bh);
4239                 /* Important: if we can't update the indirect pointers
4240                  * to the blocks, we can't free them. */
4241                 if (err)
4242                         return;
4243         }
4244
4245         for (p = first; p < last; p++) {
4246                 nr = le32_to_cpu(*p);
4247                 if (nr) {
4248                         /* accumulate blocks to free if they're contiguous */
4249                         if (count == 0) {
4250                                 block_to_free = nr;
4251                                 block_to_free_p = p;
4252                                 count = 1;
4253                         } else if (nr == block_to_free + count) {
4254                                 count++;
4255                         } else {
4256                                 if (ext4_clear_blocks(handle, inode, this_bh,
4257                                                       block_to_free, count,
4258                                                       block_to_free_p, p))
4259                                         break;
4260                                 block_to_free = nr;
4261                                 block_to_free_p = p;
4262                                 count = 1;
4263                         }
4264                 }
4265         }
4266
4267         if (count > 0)
4268                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
4269                                   count, block_to_free_p, p);
4270
4271         if (this_bh) {
4272                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
4273
4274                 /*
4275                  * The buffer head should have an attached journal head at this
4276                  * point. However, if the data is corrupted and an indirect
4277                  * block pointed to itself, it would have been detached when
4278                  * the block was cleared. Check for this instead of OOPSing.
4279                  */
4280                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
4281                         ext4_handle_dirty_metadata(handle, inode, this_bh);
4282                 else
4283                         EXT4_ERROR_INODE(inode,
4284                                          "circular indirect block detected at "
4285                                          "block %llu",
4286                                 (unsigned long long) this_bh->b_blocknr);
4287         }
4288 }
4289
4290 /**
4291  *      ext4_free_branches - free an array of branches
4292  *      @handle: JBD handle for this transaction
4293  *      @inode: inode we are dealing with
4294  *      @parent_bh: the buffer_head which contains *@first and *@last
4295  *      @first: array of block numbers
4296  *      @last:  pointer immediately past the end of array
4297  *      @depth: depth of the branches to free
4298  *
4299  *      We are freeing all blocks refered from these branches (numbers are
4300  *      stored as little-endian 32-bit) and updating @inode->i_blocks
4301  *      appropriately.
4302  */
4303 static void ext4_free_branches(handle_t *handle, struct inode *inode,
4304                                struct buffer_head *parent_bh,
4305                                __le32 *first, __le32 *last, int depth)
4306 {
4307         ext4_fsblk_t nr;
4308         __le32 *p;
4309
4310         if (ext4_handle_is_aborted(handle))
4311                 return;
4312
4313         if (depth--) {
4314                 struct buffer_head *bh;
4315                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4316                 p = last;
4317                 while (--p >= first) {
4318                         nr = le32_to_cpu(*p);
4319                         if (!nr)
4320                                 continue;               /* A hole */
4321
4322                         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
4323                                                    nr, 1)) {
4324                                 EXT4_ERROR_INODE(inode,
4325                                                  "invalid indirect mapped "
4326                                                  "block %lu (level %d)",
4327                                                  (unsigned long) nr, depth);
4328                                 break;
4329                         }
4330
4331                         /* Go read the buffer for the next level down */
4332                         bh = sb_bread(inode->i_sb, nr);
4333
4334                         /*
4335                          * A read failure? Report error and clear slot
4336                          * (should be rare).
4337                          */
4338                         if (!bh) {
4339                                 EXT4_ERROR_INODE_BLOCK(inode, nr,
4340                                                        "Read failure");
4341                                 continue;
4342                         }
4343
4344                         /* This zaps the entire block.  Bottom up. */
4345                         BUFFER_TRACE(bh, "free child branches");
4346                         ext4_free_branches(handle, inode, bh,
4347                                         (__le32 *) bh->b_data,
4348                                         (__le32 *) bh->b_data + addr_per_block,
4349                                         depth);
4350
4351                         /*
4352                          * Everything below this this pointer has been
4353                          * released.  Now let this top-of-subtree go.
4354                          *
4355                          * We want the freeing of this indirect block to be
4356                          * atomic in the journal with the updating of the
4357                          * bitmap block which owns it.  So make some room in
4358                          * the journal.
4359                          *
4360                          * We zero the parent pointer *after* freeing its
4361                          * pointee in the bitmaps, so if extend_transaction()
4362                          * for some reason fails to put the bitmap changes and
4363                          * the release into the same transaction, recovery
4364                          * will merely complain about releasing a free block,
4365                          * rather than leaking blocks.
4366                          */
4367                         if (ext4_handle_is_aborted(handle))
4368                                 return;
4369                         if (try_to_extend_transaction(handle, inode)) {
4370                                 ext4_mark_inode_dirty(handle, inode);
4371                                 ext4_truncate_restart_trans(handle, inode,
4372                                             blocks_for_truncate(inode));
4373                         }
4374
4375                         /*
4376                          * The forget flag here is critical because if
4377                          * we are journaling (and not doing data
4378                          * journaling), we have to make sure a revoke
4379                          * record is written to prevent the journal
4380                          * replay from overwriting the (former)
4381                          * indirect block if it gets reallocated as a
4382                          * data block.  This must happen in the same
4383                          * transaction where the data blocks are
4384                          * actually freed.
4385                          */
4386                         ext4_free_blocks(handle, inode, 0, nr, 1,
4387                                          EXT4_FREE_BLOCKS_METADATA|
4388                                          EXT4_FREE_BLOCKS_FORGET);
4389
4390                         if (parent_bh) {
4391                                 /*
4392                                  * The block which we have just freed is
4393                                  * pointed to by an indirect block: journal it
4394                                  */
4395                                 BUFFER_TRACE(parent_bh, "get_write_access");
4396                                 if (!ext4_journal_get_write_access(handle,
4397                                                                    parent_bh)){
4398                                         *p = 0;
4399                                         BUFFER_TRACE(parent_bh,
4400                                         "call ext4_handle_dirty_metadata");
4401                                         ext4_handle_dirty_metadata(handle,
4402                                                                    inode,
4403                                                                    parent_bh);
4404                                 }
4405                         }
4406                 }
4407         } else {
4408                 /* We have reached the bottom of the tree. */
4409                 BUFFER_TRACE(parent_bh, "free data blocks");
4410                 ext4_free_data(handle, inode, parent_bh, first, last);
4411         }
4412 }
4413
4414 int ext4_can_truncate(struct inode *inode)
4415 {
4416         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4417                 return 0;
4418         if (S_ISREG(inode->i_mode))
4419                 return 1;
4420         if (S_ISDIR(inode->i_mode))
4421                 return 1;
4422         if (S_ISLNK(inode->i_mode))
4423                 return !ext4_inode_is_fast_symlink(inode);
4424         return 0;
4425 }
4426
4427 /*
4428  * ext4_truncate()
4429  *
4430  * We block out ext4_get_block() block instantiations across the entire
4431  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4432  * simultaneously on behalf of the same inode.
4433  *
4434  * As we work through the truncate and commmit bits of it to the journal there
4435  * is one core, guiding principle: the file's tree must always be consistent on
4436  * disk.  We must be able to restart the truncate after a crash.
4437  *
4438  * The file's tree may be transiently inconsistent in memory (although it
4439  * probably isn't), but whenever we close off and commit a journal transaction,
4440  * the contents of (the filesystem + the journal) must be consistent and
4441  * restartable.  It's pretty simple, really: bottom up, right to left (although
4442  * left-to-right works OK too).
4443  *
4444  * Note that at recovery time, journal replay occurs *before* the restart of
4445  * truncate against the orphan inode list.
4446  *
4447  * The committed inode has the new, desired i_size (which is the same as
4448  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4449  * that this inode's truncate did not complete and it will again call
4450  * ext4_truncate() to have another go.  So there will be instantiated blocks
4451  * to the right of the truncation point in a crashed ext4 filesystem.  But
4452  * that's fine - as long as they are linked from the inode, the post-crash
4453  * ext4_truncate() run will find them and release them.
4454  */
4455 void ext4_truncate(struct inode *inode)
4456 {
4457         handle_t *handle;
4458         struct ext4_inode_info *ei = EXT4_I(inode);
4459         __le32 *i_data = ei->i_data;
4460         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4461         struct address_space *mapping = inode->i_mapping;
4462         ext4_lblk_t offsets[4];
4463         Indirect chain[4];
4464         Indirect *partial;
4465         __le32 nr = 0;
4466         int n;
4467         ext4_lblk_t last_block;
4468         unsigned blocksize = inode->i_sb->s_blocksize;
4469
4470         if (!ext4_can_truncate(inode))
4471                 return;
4472
4473         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4474
4475         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4476                 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4477
4478         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
4479                 ext4_ext_truncate(inode);
4480                 return;
4481         }
4482
4483         handle = start_transaction(inode);
4484         if (IS_ERR(handle))
4485                 return;         /* AKPM: return what? */
4486
4487         last_block = (inode->i_size + blocksize-1)
4488                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4489
4490         if (inode->i_size & (blocksize - 1))
4491                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4492                         goto out_stop;
4493
4494         n = ext4_block_to_path(inode, last_block, offsets, NULL);
4495         if (n == 0)
4496                 goto out_stop;  /* error */
4497
4498         /*
4499          * OK.  This truncate is going to happen.  We add the inode to the
4500          * orphan list, so that if this truncate spans multiple transactions,
4501          * and we crash, we will resume the truncate when the filesystem
4502          * recovers.  It also marks the inode dirty, to catch the new size.
4503          *
4504          * Implication: the file must always be in a sane, consistent
4505          * truncatable state while each transaction commits.
4506          */
4507         if (ext4_orphan_add(handle, inode))
4508                 goto out_stop;
4509
4510         /*
4511          * From here we block out all ext4_get_block() callers who want to
4512          * modify the block allocation tree.
4513          */
4514         down_write(&ei->i_data_sem);
4515
4516         ext4_discard_preallocations(inode);
4517
4518         /*
4519          * The orphan list entry will now protect us from any crash which
4520          * occurs before the truncate completes, so it is now safe to propagate
4521          * the new, shorter inode size (held for now in i_size) into the
4522          * on-disk inode. We do this via i_disksize, which is the value which
4523          * ext4 *really* writes onto the disk inode.
4524          */
4525         ei->i_disksize = inode->i_size;
4526
4527         if (n == 1) {           /* direct blocks */
4528                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4529                                i_data + EXT4_NDIR_BLOCKS);
4530                 goto do_indirects;
4531         }
4532
4533         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4534         /* Kill the top of shared branch (not detached) */
4535         if (nr) {
4536                 if (partial == chain) {
4537                         /* Shared branch grows from the inode */
4538                         ext4_free_branches(handle, inode, NULL,
4539                                            &nr, &nr+1, (chain+n-1) - partial);
4540                         *partial->p = 0;
4541                         /*
4542                          * We mark the inode dirty prior to restart,
4543                          * and prior to stop.  No need for it here.
4544                          */
4545                 } else {
4546                         /* Shared branch grows from an indirect block */
4547                         BUFFER_TRACE(partial->bh, "get_write_access");
4548                         ext4_free_branches(handle, inode, partial->bh,
4549                                         partial->p,
4550                                         partial->p+1, (chain+n-1) - partial);
4551                 }
4552         }
4553         /* Clear the ends of indirect blocks on the shared branch */
4554         while (partial > chain) {
4555                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4556                                    (__le32*)partial->bh->b_data+addr_per_block,
4557                                    (chain+n-1) - partial);
4558                 BUFFER_TRACE(partial->bh, "call brelse");
4559                 brelse(partial->bh);
4560                 partial--;
4561         }
4562 do_indirects:
4563         /* Kill the remaining (whole) subtrees */
4564         switch (offsets[0]) {
4565         default:
4566                 nr = i_data[EXT4_IND_BLOCK];
4567                 if (nr) {
4568                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4569                         i_data[EXT4_IND_BLOCK] = 0;
4570                 }
4571         case EXT4_IND_BLOCK:
4572                 nr = i_data[EXT4_DIND_BLOCK];
4573                 if (nr) {
4574                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4575                         i_data[EXT4_DIND_BLOCK] = 0;
4576                 }
4577         case EXT4_DIND_BLOCK:
4578                 nr = i_data[EXT4_TIND_BLOCK];
4579                 if (nr) {
4580                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4581                         i_data[EXT4_TIND_BLOCK] = 0;
4582                 }
4583         case EXT4_TIND_BLOCK:
4584                 ;
4585         }
4586
4587         up_write(&ei->i_data_sem);
4588         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4589         ext4_mark_inode_dirty(handle, inode);
4590
4591         /*
4592          * In a multi-transaction truncate, we only make the final transaction
4593          * synchronous
4594          */
4595         if (IS_SYNC(inode))
4596                 ext4_handle_sync(handle);
4597 out_stop:
4598         /*
4599          * If this was a simple ftruncate(), and the file will remain alive
4600          * then we need to clear up the orphan record which we created above.
4601          * However, if this was a real unlink then we were called by
4602          * ext4_delete_inode(), and we allow that function to clean up the
4603          * orphan info for us.
4604          */
4605         if (inode->i_nlink)
4606                 ext4_orphan_del(handle, inode);
4607
4608         ext4_journal_stop(handle);
4609 }
4610
4611 /*
4612  * ext4_get_inode_loc returns with an extra refcount against the inode's
4613  * underlying buffer_head on success. If 'in_mem' is true, we have all
4614  * data in memory that is needed to recreate the on-disk version of this
4615  * inode.
4616  */
4617 static int __ext4_get_inode_loc(struct inode *inode,
4618                                 struct ext4_iloc *iloc, int in_mem)
4619 {
4620         struct ext4_group_desc  *gdp;
4621         struct buffer_head      *bh;
4622         struct super_block      *sb = inode->i_sb;
4623         ext4_fsblk_t            block;
4624         int                     inodes_per_block, inode_offset;
4625
4626         iloc->bh = NULL;
4627         if (!ext4_valid_inum(sb, inode->i_ino))
4628                 return -EIO;
4629
4630         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4631         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4632         if (!gdp)
4633                 return -EIO;
4634
4635         /*
4636          * Figure out the offset within the block group inode table
4637          */
4638         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4639         inode_offset = ((inode->i_ino - 1) %
4640                         EXT4_INODES_PER_GROUP(sb));
4641         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4642         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4643
4644         bh = sb_getblk(sb, block);
4645         if (!bh) {
4646                 EXT4_ERROR_INODE_BLOCK(inode, block,
4647                                        "unable to read itable block");
4648                 return -EIO;
4649         }
4650         if (!buffer_uptodate(bh)) {
4651                 lock_buffer(bh);
4652
4653                 /*
4654                  * If the buffer has the write error flag, we have failed
4655                  * to write out another inode in the same block.  In this
4656                  * case, we don't have to read the block because we may
4657                  * read the old inode data successfully.
4658                  */
4659                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4660                         set_buffer_uptodate(bh);
4661
4662                 if (buffer_uptodate(bh)) {
4663                         /* someone brought it uptodate while we waited */
4664                         unlock_buffer(bh);
4665                         goto has_buffer;
4666                 }
4667
4668                 /*
4669                  * If we have all information of the inode in memory and this
4670                  * is the only valid inode in the block, we need not read the
4671                  * block.
4672                  */
4673                 if (in_mem) {
4674                         struct buffer_head *bitmap_bh;
4675                         int i, start;
4676
4677                         start = inode_offset & ~(inodes_per_block - 1);
4678
4679                         /* Is the inode bitmap in cache? */
4680                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4681                         if (!bitmap_bh)
4682                                 goto make_io;
4683
4684                         /*
4685                          * If the inode bitmap isn't in cache then the
4686                          * optimisation may end up performing two reads instead
4687                          * of one, so skip it.
4688                          */
4689                         if (!buffer_uptodate(bitmap_bh)) {
4690                                 brelse(bitmap_bh);
4691                                 goto make_io;
4692                         }
4693                         for (i = start; i < start + inodes_per_block; i++) {
4694                                 if (i == inode_offset)
4695                                         continue;
4696                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4697                                         break;
4698                         }
4699                         brelse(bitmap_bh);
4700                         if (i == start + inodes_per_block) {
4701                                 /* all other inodes are free, so skip I/O */
4702                                 memset(bh->b_data, 0, bh->b_size);
4703                                 set_buffer_uptodate(bh);
4704                                 unlock_buffer(bh);
4705                                 goto has_buffer;
4706                         }
4707                 }
4708
4709 make_io:
4710                 /*
4711                  * If we need to do any I/O, try to pre-readahead extra
4712                  * blocks from the inode table.
4713                  */
4714                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4715                         ext4_fsblk_t b, end, table;
4716                         unsigned num;
4717
4718                         table = ext4_inode_table(sb, gdp);
4719                         /* s_inode_readahead_blks is always a power of 2 */
4720                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4721                         if (table > b)
4722                                 b = table;
4723                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4724                         num = EXT4_INODES_PER_GROUP(sb);
4725                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4726                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4727                                 num -= ext4_itable_unused_count(sb, gdp);
4728                         table += num / inodes_per_block;
4729                         if (end > table)
4730                                 end = table;
4731                         while (b <= end)
4732                                 sb_breadahead(sb, b++);
4733                 }
4734
4735                 /*
4736                  * There are other valid inodes in the buffer, this inode
4737                  * has in-inode xattrs, or we don't have this inode in memory.
4738                  * Read the block from disk.
4739                  */
4740                 get_bh(bh);
4741                 bh->b_end_io = end_buffer_read_sync;
4742                 submit_bh(READ_META, bh);
4743                 wait_on_buffer(bh);
4744                 if (!buffer_uptodate(bh)) {
4745                         EXT4_ERROR_INODE_BLOCK(inode, block,
4746                                                "unable to read itable block");
4747                         brelse(bh);
4748                         return -EIO;
4749                 }
4750         }
4751 has_buffer:
4752         iloc->bh = bh;
4753         return 0;
4754 }
4755
4756 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4757 {
4758         /* We have all inode data except xattrs in memory here. */
4759         return __ext4_get_inode_loc(inode, iloc,
4760                 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
4761 }
4762
4763 void ext4_set_inode_flags(struct inode *inode)
4764 {
4765         unsigned int flags = EXT4_I(inode)->i_flags;
4766
4767         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4768         if (flags & EXT4_SYNC_FL)
4769                 inode->i_flags |= S_SYNC;
4770         if (flags & EXT4_APPEND_FL)
4771                 inode->i_flags |= S_APPEND;
4772         if (flags & EXT4_IMMUTABLE_FL)
4773                 inode->i_flags |= S_IMMUTABLE;
4774         if (flags & EXT4_NOATIME_FL)
4775                 inode->i_flags |= S_NOATIME;
4776         if (flags & EXT4_DIRSYNC_FL)
4777                 inode->i_flags |= S_DIRSYNC;
4778 }
4779
4780 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4781 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4782 {
4783         unsigned int vfs_fl;
4784         unsigned long old_fl, new_fl;
4785
4786         do {
4787                 vfs_fl = ei->vfs_inode.i_flags;
4788                 old_fl = ei->i_flags;
4789                 new_fl = old_fl & ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4790                                 EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|
4791                                 EXT4_DIRSYNC_FL);
4792                 if (vfs_fl & S_SYNC)
4793                         new_fl |= EXT4_SYNC_FL;
4794                 if (vfs_fl & S_APPEND)
4795                         new_fl |= EXT4_APPEND_FL;
4796                 if (vfs_fl & S_IMMUTABLE)
4797                         new_fl |= EXT4_IMMUTABLE_FL;
4798                 if (vfs_fl & S_NOATIME)
4799                         new_fl |= EXT4_NOATIME_FL;
4800                 if (vfs_fl & S_DIRSYNC)
4801                         new_fl |= EXT4_DIRSYNC_FL;
4802         } while (cmpxchg(&ei->i_flags, old_fl, new_fl) != old_fl);
4803 }
4804
4805 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4806                                   struct ext4_inode_info *ei)
4807 {
4808         blkcnt_t i_blocks ;
4809         struct inode *inode = &(ei->vfs_inode);
4810         struct super_block *sb = inode->i_sb;
4811
4812         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4813                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4814                 /* we are using combined 48 bit field */
4815                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4816                                         le32_to_cpu(raw_inode->i_blocks_lo);
4817                 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
4818                         /* i_blocks represent file system block size */
4819                         return i_blocks  << (inode->i_blkbits - 9);
4820                 } else {
4821                         return i_blocks;
4822                 }
4823         } else {
4824                 return le32_to_cpu(raw_inode->i_blocks_lo);
4825         }
4826 }
4827
4828 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4829 {
4830         struct ext4_iloc iloc;
4831         struct ext4_inode *raw_inode;
4832         struct ext4_inode_info *ei;
4833         struct inode *inode;
4834         journal_t *journal = EXT4_SB(sb)->s_journal;
4835         long ret;
4836         int block;
4837
4838         inode = iget_locked(sb, ino);
4839         if (!inode)
4840                 return ERR_PTR(-ENOMEM);
4841         if (!(inode->i_state & I_NEW))
4842                 return inode;
4843
4844         ei = EXT4_I(inode);
4845         iloc.bh = 0;
4846
4847         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4848         if (ret < 0)
4849                 goto bad_inode;
4850         raw_inode = ext4_raw_inode(&iloc);
4851         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4852         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4853         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4854         if (!(test_opt(inode->i_sb, NO_UID32))) {
4855                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4856                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4857         }
4858         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
4859
4860         ei->i_state_flags = 0;
4861         ei->i_dir_start_lookup = 0;
4862         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4863         /* We now have enough fields to check if the inode was active or not.
4864          * This is needed because nfsd might try to access dead inodes
4865          * the test is that same one that e2fsck uses
4866          * NeilBrown 1999oct15
4867          */
4868         if (inode->i_nlink == 0) {
4869                 if (inode->i_mode == 0 ||
4870                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
4871                         /* this inode is deleted */
4872                         ret = -ESTALE;
4873                         goto bad_inode;
4874                 }
4875                 /* The only unlinked inodes we let through here have
4876                  * valid i_mode and are being read by the orphan
4877                  * recovery code: that's fine, we're about to complete
4878                  * the process of deleting those. */
4879         }
4880         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4881         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4882         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4883         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
4884                 ei->i_file_acl |=
4885                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4886         inode->i_size = ext4_isize(raw_inode);
4887         ei->i_disksize = inode->i_size;
4888 #ifdef CONFIG_QUOTA
4889         ei->i_reserved_quota = 0;
4890 #endif
4891         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4892         ei->i_block_group = iloc.block_group;
4893         ei->i_last_alloc_group = ~0;
4894         /*
4895          * NOTE! The in-memory inode i_data array is in little-endian order
4896          * even on big-endian machines: we do NOT byteswap the block numbers!
4897          */
4898         for (block = 0; block < EXT4_N_BLOCKS; block++)
4899                 ei->i_data[block] = raw_inode->i_block[block];
4900         INIT_LIST_HEAD(&ei->i_orphan);
4901
4902         /*
4903          * Set transaction id's of transactions that have to be committed
4904          * to finish f[data]sync. We set them to currently running transaction
4905          * as we cannot be sure that the inode or some of its metadata isn't
4906          * part of the transaction - the inode could have been reclaimed and
4907          * now it is reread from disk.
4908          */
4909         if (journal) {
4910                 transaction_t *transaction;
4911                 tid_t tid;
4912
4913                 read_lock(&journal->j_state_lock);
4914                 if (journal->j_running_transaction)
4915                         transaction = journal->j_running_transaction;
4916                 else
4917                         transaction = journal->j_committing_transaction;
4918                 if (transaction)
4919                         tid = transaction->t_tid;
4920                 else
4921                         tid = journal->j_commit_sequence;
4922                 read_unlock(&journal->j_state_lock);
4923                 ei->i_sync_tid = tid;
4924                 ei->i_datasync_tid = tid;
4925         }
4926
4927         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4928                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4929                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4930                     EXT4_INODE_SIZE(inode->i_sb)) {
4931                         ret = -EIO;
4932                         goto bad_inode;
4933                 }
4934                 if (ei->i_extra_isize == 0) {
4935                         /* The extra space is currently unused. Use it. */
4936                         ei->i_extra_isize = sizeof(struct ext4_inode) -
4937                                             EXT4_GOOD_OLD_INODE_SIZE;
4938                 } else {
4939                         __le32 *magic = (void *)raw_inode +
4940                                         EXT4_GOOD_OLD_INODE_SIZE +
4941                                         ei->i_extra_isize;
4942                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
4943                                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
4944                 }
4945         } else
4946                 ei->i_extra_isize = 0;
4947
4948         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4949         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4950         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4951         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4952
4953         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4954         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4955                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4956                         inode->i_version |=
4957                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4958         }
4959
4960         ret = 0;
4961         if (ei->i_file_acl &&
4962             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
4963                 EXT4_ERROR_INODE(inode, "bad extended attribute block %llu",
4964                                  ei->i_file_acl);
4965                 ret = -EIO;
4966                 goto bad_inode;
4967         } else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
4968                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4969                     (S_ISLNK(inode->i_mode) &&
4970                      !ext4_inode_is_fast_symlink(inode)))
4971                         /* Validate extent which is part of inode */
4972                         ret = ext4_ext_check_inode(inode);
4973         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4974                    (S_ISLNK(inode->i_mode) &&
4975                     !ext4_inode_is_fast_symlink(inode))) {
4976                 /* Validate block references which are part of inode */
4977                 ret = ext4_check_inode_blockref(inode);
4978         }
4979         if (ret)
4980                 goto bad_inode;
4981
4982         if (S_ISREG(inode->i_mode)) {
4983                 inode->i_op = &ext4_file_inode_operations;
4984                 inode->i_fop = &ext4_file_operations;
4985                 ext4_set_aops(inode);
4986         } else if (S_ISDIR(inode->i_mode)) {
4987                 inode->i_op = &ext4_dir_inode_operations;
4988                 inode->i_fop = &ext4_dir_operations;
4989         } else if (S_ISLNK(inode->i_mode)) {
4990                 if (ext4_inode_is_fast_symlink(inode)) {
4991                         inode->i_op = &ext4_fast_symlink_inode_operations;
4992                         nd_terminate_link(ei->i_data, inode->i_size,
4993                                 sizeof(ei->i_data) - 1);
4994                 } else {
4995                         inode->i_op = &ext4_symlink_inode_operations;
4996                         ext4_set_aops(inode);
4997                 }
4998         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
4999               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
5000                 inode->i_op = &ext4_special_inode_operations;
5001                 if (raw_inode->i_block[0])
5002                         init_special_inode(inode, inode->i_mode,
5003                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
5004                 else
5005                         init_special_inode(inode, inode->i_mode,
5006                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
5007         } else {
5008                 ret = -EIO;
5009                 EXT4_ERROR_INODE(inode, "bogus i_mode (%o)", inode->i_mode);
5010                 goto bad_inode;
5011         }
5012         brelse(iloc.bh);
5013         ext4_set_inode_flags(inode);
5014         unlock_new_inode(inode);
5015         return inode;
5016
5017 bad_inode:
5018         brelse(iloc.bh);
5019         iget_failed(inode);
5020         return ERR_PTR(ret);
5021 }
5022
5023 static int ext4_inode_blocks_set(handle_t *handle,
5024                                 struct ext4_inode *raw_inode,
5025                                 struct ext4_inode_info *ei)
5026 {
5027         struct inode *inode = &(ei->vfs_inode);
5028         u64 i_blocks = inode->i_blocks;
5029         struct super_block *sb = inode->i_sb;
5030
5031         if (i_blocks <= ~0U) {
5032                 /*
5033                  * i_blocks can be represnted in a 32 bit variable
5034                  * as multiple of 512 bytes
5035                  */
5036                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5037                 raw_inode->i_blocks_high = 0;
5038                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5039                 return 0;
5040         }
5041         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
5042                 return -EFBIG;
5043
5044         if (i_blocks <= 0xffffffffffffULL) {
5045                 /*
5046                  * i_blocks can be represented in a 48 bit variable
5047                  * as multiple of 512 bytes
5048                  */
5049                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5050                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5051                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5052         } else {
5053                 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5054                 /* i_block is stored in file system block size */
5055                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
5056                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5057                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5058         }
5059         return 0;
5060 }
5061
5062 /*
5063  * Post the struct inode info into an on-disk inode location in the
5064  * buffer-cache.  This gobbles the caller's reference to the
5065  * buffer_head in the inode location struct.
5066  *
5067  * The caller must have write access to iloc->bh.
5068  */
5069 static int ext4_do_update_inode(handle_t *handle,
5070                                 struct inode *inode,
5071                                 struct ext4_iloc *iloc)
5072 {
5073         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5074         struct ext4_inode_info *ei = EXT4_I(inode);
5075         struct buffer_head *bh = iloc->bh;
5076         int err = 0, rc, block;
5077
5078         /* For fields not not tracking in the in-memory inode,
5079          * initialise them to zero for new inodes. */
5080         if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5081                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5082
5083         ext4_get_inode_flags(ei);
5084         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5085         if (!(test_opt(inode->i_sb, NO_UID32))) {
5086                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
5087                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
5088 /*
5089  * Fix up interoperability with old kernels. Otherwise, old inodes get
5090  * re-used with the upper 16 bits of the uid/gid intact
5091  */
5092                 if (!ei->i_dtime) {
5093                         raw_inode->i_uid_high =
5094                                 cpu_to_le16(high_16_bits(inode->i_uid));
5095                         raw_inode->i_gid_high =
5096                                 cpu_to_le16(high_16_bits(inode->i_gid));
5097                 } else {
5098                         raw_inode->i_uid_high = 0;
5099                         raw_inode->i_gid_high = 0;
5100                 }
5101         } else {
5102                 raw_inode->i_uid_low =
5103                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
5104                 raw_inode->i_gid_low =
5105                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
5106                 raw_inode->i_uid_high = 0;
5107                 raw_inode->i_gid_high = 0;
5108         }
5109         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5110
5111         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5112         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5113         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5114         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5115
5116         if (ext4_inode_blocks_set(handle, raw_inode, ei))
5117                 goto out_brelse;
5118         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5119         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
5120         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
5121             cpu_to_le32(EXT4_OS_HURD))
5122                 raw_inode->i_file_acl_high =
5123                         cpu_to_le16(ei->i_file_acl >> 32);
5124         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5125         ext4_isize_set(raw_inode, ei->i_disksize);
5126         if (ei->i_disksize > 0x7fffffffULL) {
5127                 struct super_block *sb = inode->i_sb;
5128                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
5129                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
5130                                 EXT4_SB(sb)->s_es->s_rev_level ==
5131                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
5132                         /* If this is the first large file
5133                          * created, add a flag to the superblock.
5134                          */
5135                         err = ext4_journal_get_write_access(handle,
5136                                         EXT4_SB(sb)->s_sbh);
5137                         if (err)
5138                                 goto out_brelse;
5139                         ext4_update_dynamic_rev(sb);
5140                         EXT4_SET_RO_COMPAT_FEATURE(sb,
5141                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
5142                         sb->s_dirt = 1;
5143                         ext4_handle_sync(handle);
5144                         err = ext4_handle_dirty_metadata(handle, NULL,
5145                                         EXT4_SB(sb)->s_sbh);
5146                 }
5147         }
5148         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5149         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5150                 if (old_valid_dev(inode->i_rdev)) {
5151                         raw_inode->i_block[0] =
5152                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
5153                         raw_inode->i_block[1] = 0;
5154                 } else {
5155                         raw_inode->i_block[0] = 0;
5156                         raw_inode->i_block[1] =
5157                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
5158                         raw_inode->i_block[2] = 0;
5159                 }
5160         } else
5161                 for (block = 0; block < EXT4_N_BLOCKS; block++)
5162                         raw_inode->i_block[block] = ei->i_data[block];
5163
5164         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
5165         if (ei->i_extra_isize) {
5166                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5167                         raw_inode->i_version_hi =
5168                         cpu_to_le32(inode->i_version >> 32);
5169                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
5170         }
5171
5172         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5173         rc = ext4_handle_dirty_metadata(handle, NULL, bh);
5174         if (!err)
5175                 err = rc;
5176         ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5177
5178         ext4_update_inode_fsync_trans(handle, inode, 0);
5179 out_brelse:
5180         brelse(bh);
5181         ext4_std_error(inode->i_sb, err);
5182         return err;
5183 }
5184
5185 /*
5186  * ext4_write_inode()
5187  *
5188  * We are called from a few places:
5189  *
5190  * - Within generic_file_write() for O_SYNC files.
5191  *   Here, there will be no transaction running. We wait for any running
5192  *   trasnaction to commit.
5193  *
5194  * - Within sys_sync(), kupdate and such.
5195  *   We wait on commit, if tol to.
5196  *
5197  * - Within prune_icache() (PF_MEMALLOC == true)
5198  *   Here we simply return.  We can't afford to block kswapd on the
5199  *   journal commit.
5200  *
5201  * In all cases it is actually safe for us to return without doing anything,
5202  * because the inode has been copied into a raw inode buffer in
5203  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
5204  * knfsd.
5205  *
5206  * Note that we are absolutely dependent upon all inode dirtiers doing the
5207  * right thing: they *must* call mark_inode_dirty() after dirtying info in
5208  * which we are interested.
5209  *
5210  * It would be a bug for them to not do this.  The code:
5211  *
5212  *      mark_inode_dirty(inode)
5213  *      stuff();
5214  *      inode->i_size = expr;
5215  *
5216  * is in error because a kswapd-driven write_inode() could occur while
5217  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
5218  * will no longer be on the superblock's dirty inode list.
5219  */
5220 int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5221 {
5222         int err;
5223
5224         if (current->flags & PF_MEMALLOC)
5225                 return 0;
5226
5227         if (EXT4_SB(inode->i_sb)->s_journal) {
5228                 if (ext4_journal_current_handle()) {
5229                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5230                         dump_stack();
5231                         return -EIO;
5232                 }
5233
5234                 if (wbc->sync_mode != WB_SYNC_ALL)
5235                         return 0;
5236
5237                 err = ext4_force_commit(inode->i_sb);
5238         } else {
5239                 struct ext4_iloc iloc;
5240
5241                 err = __ext4_get_inode_loc(inode, &iloc, 0);
5242                 if (err)
5243                         return err;
5244                 if (wbc->sync_mode == WB_SYNC_ALL)
5245                         sync_dirty_buffer(iloc.bh);
5246                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5247                         EXT4_ERROR_INODE_BLOCK(inode, iloc.bh->b_blocknr,
5248                                          "IO error syncing inode");
5249                         err = -EIO;
5250                 }
5251                 brelse(iloc.bh);
5252         }
5253         return err;
5254 }
5255
5256 /*
5257  * ext4_setattr()
5258  *
5259  * Called from notify_change.
5260  *
5261  * We want to trap VFS attempts to truncate the file as soon as
5262  * possible.  In particular, we want to make sure that when the VFS
5263  * shrinks i_size, we put the inode on the orphan list and modify
5264  * i_disksize immediately, so that during the subsequent flushing of
5265  * dirty pages and freeing of disk blocks, we can guarantee that any
5266  * commit will leave the blocks being flushed in an unused state on
5267  * disk.  (On recovery, the inode will get truncated and the blocks will
5268  * be freed, so we have a strong guarantee that no future commit will
5269  * leave these blocks visible to the user.)
5270  *
5271  * Another thing we have to assure is that if we are in ordered mode
5272  * and inode is still attached to the committing transaction, we must
5273  * we start writeout of all the dirty pages which are being truncated.
5274  * This way we are sure that all the data written in the previous
5275  * transaction are already on disk (truncate waits for pages under
5276  * writeback).
5277  *
5278  * Called with inode->i_mutex down.
5279  */
5280 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5281 {
5282         struct inode *inode = dentry->d_inode;
5283         int error, rc = 0;
5284         int orphan = 0;
5285         const unsigned int ia_valid = attr->ia_valid;
5286
5287         error = inode_change_ok(inode, attr);
5288         if (error)
5289                 return error;
5290
5291         if (is_quota_modification(inode, attr))
5292                 dquot_initialize(inode);
5293         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
5294                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
5295                 handle_t *handle;
5296
5297                 /* (user+group)*(old+new) structure, inode write (sb,
5298                  * inode block, ? - but truncate inode update has it) */
5299                 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
5300                                         EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
5301                 if (IS_ERR(handle)) {
5302                         error = PTR_ERR(handle);
5303                         goto err_out;
5304                 }
5305                 error = dquot_transfer(inode, attr);
5306                 if (error) {
5307                         ext4_journal_stop(handle);
5308                         return error;
5309                 }
5310                 /* Update corresponding info in inode so that everything is in
5311                  * one transaction */
5312                 if (attr->ia_valid & ATTR_UID)
5313                         inode->i_uid = attr->ia_uid;
5314                 if (attr->ia_valid & ATTR_GID)
5315                         inode->i_gid = attr->ia_gid;
5316                 error = ext4_mark_inode_dirty(handle, inode);
5317                 ext4_journal_stop(handle);
5318         }
5319
5320         if (attr->ia_valid & ATTR_SIZE) {
5321                 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
5322                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5323
5324                         if (attr->ia_size > sbi->s_bitmap_maxbytes)
5325                                 return -EFBIG;
5326                 }
5327         }
5328
5329         if (S_ISREG(inode->i_mode) &&
5330             attr->ia_valid & ATTR_SIZE &&
5331             (attr->ia_size < inode->i_size ||
5332              (ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS)))) {
5333                 handle_t *handle;
5334
5335                 handle = ext4_journal_start(inode, 3);
5336                 if (IS_ERR(handle)) {
5337                         error = PTR_ERR(handle);
5338                         goto err_out;
5339                 }
5340                 if (ext4_handle_valid(handle)) {
5341                         error = ext4_orphan_add(handle, inode);
5342                         orphan = 1;
5343                 }
5344                 EXT4_I(inode)->i_disksize = attr->ia_size;
5345                 rc = ext4_mark_inode_dirty(handle, inode);
5346                 if (!error)
5347                         error = rc;
5348                 ext4_journal_stop(handle);
5349
5350                 if (ext4_should_order_data(inode)) {
5351                         error = ext4_begin_ordered_truncate(inode,
5352                                                             attr->ia_size);
5353                         if (error) {
5354                                 /* Do as much error cleanup as possible */
5355                                 handle = ext4_journal_start(inode, 3);
5356                                 if (IS_ERR(handle)) {
5357                                         ext4_orphan_del(NULL, inode);
5358                                         goto err_out;
5359                                 }
5360                                 ext4_orphan_del(handle, inode);
5361                                 orphan = 0;
5362                                 ext4_journal_stop(handle);
5363                                 goto err_out;
5364                         }
5365                 }
5366                 /* ext4_truncate will clear the flag */
5367                 if ((ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS)))
5368                         ext4_truncate(inode);
5369         }
5370
5371         if ((attr->ia_valid & ATTR_SIZE) &&
5372             attr->ia_size != i_size_read(inode))
5373                 rc = vmtruncate(inode, attr->ia_size);
5374
5375         if (!rc) {
5376                 setattr_copy(inode, attr);
5377                 mark_inode_dirty(inode);
5378         }
5379
5380         /*
5381          * If the call to ext4_truncate failed to get a transaction handle at
5382          * all, we need to clean up the in-core orphan list manually.
5383          */
5384         if (orphan && inode->i_nlink)
5385                 ext4_orphan_del(NULL, inode);
5386
5387         if (!rc && (ia_valid & ATTR_MODE))
5388                 rc = ext4_acl_chmod(inode);
5389
5390 err_out:
5391         ext4_std_error(inode->i_sb, error);
5392         if (!error)
5393                 error = rc;
5394         return error;
5395 }
5396
5397 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
5398                  struct kstat *stat)
5399 {
5400         struct inode *inode;
5401         unsigned long delalloc_blocks;
5402
5403         inode = dentry->d_inode;
5404         generic_fillattr(inode, stat);
5405
5406         /*
5407          * We can't update i_blocks if the block allocation is delayed
5408          * otherwise in the case of system crash before the real block
5409          * allocation is done, we will have i_blocks inconsistent with
5410          * on-disk file blocks.
5411          * We always keep i_blocks updated together with real
5412          * allocation. But to not confuse with user, stat
5413          * will return the blocks that include the delayed allocation
5414          * blocks for this file.
5415          */
5416         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
5417         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
5418         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
5419
5420         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
5421         return 0;
5422 }
5423
5424 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
5425                                       int chunk)
5426 {
5427         int indirects;
5428
5429         /* if nrblocks are contiguous */
5430         if (chunk) {
5431                 /*
5432                  * With N contiguous data blocks, it need at most
5433                  * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) indirect blocks
5434                  * 2 dindirect blocks
5435                  * 1 tindirect block
5436                  */
5437                 indirects = nrblocks / EXT4_ADDR_PER_BLOCK(inode->i_sb);
5438                 return indirects + 3;
5439         }
5440         /*
5441          * if nrblocks are not contiguous, worse case, each block touch
5442          * a indirect block, and each indirect block touch a double indirect
5443          * block, plus a triple indirect block
5444          */
5445         indirects = nrblocks * 2 + 1;
5446         return indirects;
5447 }
5448
5449 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5450 {
5451         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5452                 return ext4_indirect_trans_blocks(inode, nrblocks, chunk);
5453         return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
5454 }
5455
5456 /*
5457  * Account for index blocks, block groups bitmaps and block group
5458  * descriptor blocks if modify datablocks and index blocks
5459  * worse case, the indexs blocks spread over different block groups
5460  *
5461  * If datablocks are discontiguous, they are possible to spread over
5462  * different block groups too. If they are contiuguous, with flexbg,
5463  * they could still across block group boundary.
5464  *
5465  * Also account for superblock, inode, quota and xattr blocks
5466  */
5467 static int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5468 {
5469         ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
5470         int gdpblocks;
5471         int idxblocks;
5472         int ret = 0;
5473
5474         /*
5475          * How many index blocks need to touch to modify nrblocks?
5476          * The "Chunk" flag indicating whether the nrblocks is
5477          * physically contiguous on disk
5478          *
5479          * For Direct IO and fallocate, they calls get_block to allocate
5480          * one single extent at a time, so they could set the "Chunk" flag
5481          */
5482         idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
5483
5484         ret = idxblocks;
5485
5486         /*
5487          * Now let's see how many group bitmaps and group descriptors need
5488          * to account
5489          */
5490         groups = idxblocks;
5491         if (chunk)
5492                 groups += 1;
5493         else
5494                 groups += nrblocks;
5495
5496         gdpblocks = groups;
5497         if (groups > ngroups)
5498                 groups = ngroups;
5499         if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
5500                 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
5501
5502         /* bitmaps and block group descriptor blocks */
5503         ret += groups + gdpblocks;
5504
5505         /* Blocks for super block, inode, quota and xattr blocks */
5506         ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
5507
5508         return ret;
5509 }
5510
5511 /*
5512  * Calulate the total number of credits to reserve to fit
5513  * the modification of a single pages into a single transaction,
5514  * which may include multiple chunks of block allocations.
5515  *
5516  * This could be called via ext4_write_begin()
5517  *
5518  * We need to consider the worse case, when
5519  * one new block per extent.
5520  */
5521 int ext4_writepage_trans_blocks(struct inode *inode)
5522 {
5523         int bpp = ext4_journal_blocks_per_page(inode);
5524         int ret;
5525
5526         ret = ext4_meta_trans_blocks(inode, bpp, 0);
5527
5528         /* Account for data blocks for journalled mode */
5529