ext4: stop looping in ext4_num_dirty_pages when max_pages reached
[linux-2.6.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/jbd2.h>
29 #include <linux/highuid.h>
30 #include <linux/pagemap.h>
31 #include <linux/quotaops.h>
32 #include <linux/string.h>
33 #include <linux/buffer_head.h>
34 #include <linux/writeback.h>
35 #include <linux/pagevec.h>
36 #include <linux/mpage.h>
37 #include <linux/namei.h>
38 #include <linux/uio.h>
39 #include <linux/bio.h>
40 #include <linux/workqueue.h>
41 #include <linux/kernel.h>
42 #include <linux/slab.h>
43
44 #include "ext4_jbd2.h"
45 #include "xattr.h"
46 #include "acl.h"
47 #include "ext4_extents.h"
48
49 #include <trace/events/ext4.h>
50
51 #define MPAGE_DA_EXTENT_TAIL 0x01
52
53 static inline int ext4_begin_ordered_truncate(struct inode *inode,
54                                               loff_t new_size)
55 {
56         return jbd2_journal_begin_ordered_truncate(
57                                         EXT4_SB(inode->i_sb)->s_journal,
58                                         &EXT4_I(inode)->jinode,
59                                         new_size);
60 }
61
62 static void ext4_invalidatepage(struct page *page, unsigned long offset);
63
64 /*
65  * Test whether an inode is a fast symlink.
66  */
67 static int ext4_inode_is_fast_symlink(struct inode *inode)
68 {
69         int ea_blocks = EXT4_I(inode)->i_file_acl ?
70                 (inode->i_sb->s_blocksize >> 9) : 0;
71
72         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
73 }
74
75 /*
76  * Work out how many blocks we need to proceed with the next chunk of a
77  * truncate transaction.
78  */
79 static unsigned long blocks_for_truncate(struct inode *inode)
80 {
81         ext4_lblk_t needed;
82
83         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
84
85         /* Give ourselves just enough room to cope with inodes in which
86          * i_blocks is corrupt: we've seen disk corruptions in the past
87          * which resulted in random data in an inode which looked enough
88          * like a regular file for ext4 to try to delete it.  Things
89          * will go a bit crazy if that happens, but at least we should
90          * try not to panic the whole kernel. */
91         if (needed < 2)
92                 needed = 2;
93
94         /* But we need to bound the transaction so we don't overflow the
95          * journal. */
96         if (needed > EXT4_MAX_TRANS_DATA)
97                 needed = EXT4_MAX_TRANS_DATA;
98
99         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
100 }
101
102 /*
103  * Truncate transactions can be complex and absolutely huge.  So we need to
104  * be able to restart the transaction at a conventient checkpoint to make
105  * sure we don't overflow the journal.
106  *
107  * start_transaction gets us a new handle for a truncate transaction,
108  * and extend_transaction tries to extend the existing one a bit.  If
109  * extend fails, we need to propagate the failure up and restart the
110  * transaction in the top-level truncate loop. --sct
111  */
112 static handle_t *start_transaction(struct inode *inode)
113 {
114         handle_t *result;
115
116         result = ext4_journal_start(inode, blocks_for_truncate(inode));
117         if (!IS_ERR(result))
118                 return result;
119
120         ext4_std_error(inode->i_sb, PTR_ERR(result));
121         return result;
122 }
123
124 /*
125  * Try to extend this transaction for the purposes of truncation.
126  *
127  * Returns 0 if we managed to create more room.  If we can't create more
128  * room, and the transaction must be restarted we return 1.
129  */
130 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
131 {
132         if (!ext4_handle_valid(handle))
133                 return 0;
134         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
135                 return 0;
136         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
137                 return 0;
138         return 1;
139 }
140
141 /*
142  * Restart the transaction associated with *handle.  This does a commit,
143  * so before we call here everything must be consistently dirtied against
144  * this transaction.
145  */
146 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
147                                  int nblocks)
148 {
149         int ret;
150
151         /*
152          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
153          * moment, get_block can be called only for blocks inside i_size since
154          * page cache has been already dropped and writes are blocked by
155          * i_mutex. So we can safely drop the i_data_sem here.
156          */
157         BUG_ON(EXT4_JOURNAL(inode) == NULL);
158         jbd_debug(2, "restarting handle %p\n", handle);
159         up_write(&EXT4_I(inode)->i_data_sem);
160         ret = ext4_journal_restart(handle, blocks_for_truncate(inode));
161         down_write(&EXT4_I(inode)->i_data_sem);
162         ext4_discard_preallocations(inode);
163
164         return ret;
165 }
166
167 /*
168  * Called at the last iput() if i_nlink is zero.
169  */
170 void ext4_evict_inode(struct inode *inode)
171 {
172         handle_t *handle;
173         int err;
174
175         if (inode->i_nlink) {
176                 truncate_inode_pages(&inode->i_data, 0);
177                 goto no_delete;
178         }
179
180         if (!is_bad_inode(inode))
181                 dquot_initialize(inode);
182
183         if (ext4_should_order_data(inode))
184                 ext4_begin_ordered_truncate(inode, 0);
185         truncate_inode_pages(&inode->i_data, 0);
186
187         if (is_bad_inode(inode))
188                 goto no_delete;
189
190         handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
191         if (IS_ERR(handle)) {
192                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
193                 /*
194                  * If we're going to skip the normal cleanup, we still need to
195                  * make sure that the in-core orphan linked list is properly
196                  * cleaned up.
197                  */
198                 ext4_orphan_del(NULL, inode);
199                 goto no_delete;
200         }
201
202         if (IS_SYNC(inode))
203                 ext4_handle_sync(handle);
204         inode->i_size = 0;
205         err = ext4_mark_inode_dirty(handle, inode);
206         if (err) {
207                 ext4_warning(inode->i_sb,
208                              "couldn't mark inode dirty (err %d)", err);
209                 goto stop_handle;
210         }
211         if (inode->i_blocks)
212                 ext4_truncate(inode);
213
214         /*
215          * ext4_ext_truncate() doesn't reserve any slop when it
216          * restarts journal transactions; therefore there may not be
217          * enough credits left in the handle to remove the inode from
218          * the orphan list and set the dtime field.
219          */
220         if (!ext4_handle_has_enough_credits(handle, 3)) {
221                 err = ext4_journal_extend(handle, 3);
222                 if (err > 0)
223                         err = ext4_journal_restart(handle, 3);
224                 if (err != 0) {
225                         ext4_warning(inode->i_sb,
226                                      "couldn't extend journal (err %d)", err);
227                 stop_handle:
228                         ext4_journal_stop(handle);
229                         ext4_orphan_del(NULL, inode);
230                         goto no_delete;
231                 }
232         }
233
234         /*
235          * Kill off the orphan record which ext4_truncate created.
236          * AKPM: I think this can be inside the above `if'.
237          * Note that ext4_orphan_del() has to be able to cope with the
238          * deletion of a non-existent orphan - this is because we don't
239          * know if ext4_truncate() actually created an orphan record.
240          * (Well, we could do this if we need to, but heck - it works)
241          */
242         ext4_orphan_del(handle, inode);
243         EXT4_I(inode)->i_dtime  = get_seconds();
244
245         /*
246          * One subtle ordering requirement: if anything has gone wrong
247          * (transaction abort, IO errors, whatever), then we can still
248          * do these next steps (the fs will already have been marked as
249          * having errors), but we can't free the inode if the mark_dirty
250          * fails.
251          */
252         if (ext4_mark_inode_dirty(handle, inode))
253                 /* If that failed, just do the required in-core inode clear. */
254                 ext4_clear_inode(inode);
255         else
256                 ext4_free_inode(handle, inode);
257         ext4_journal_stop(handle);
258         return;
259 no_delete:
260         ext4_clear_inode(inode);        /* We must guarantee clearing of inode... */
261 }
262
263 typedef struct {
264         __le32  *p;
265         __le32  key;
266         struct buffer_head *bh;
267 } Indirect;
268
269 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
270 {
271         p->key = *(p->p = v);
272         p->bh = bh;
273 }
274
275 /**
276  *      ext4_block_to_path - parse the block number into array of offsets
277  *      @inode: inode in question (we are only interested in its superblock)
278  *      @i_block: block number to be parsed
279  *      @offsets: array to store the offsets in
280  *      @boundary: set this non-zero if the referred-to block is likely to be
281  *             followed (on disk) by an indirect block.
282  *
283  *      To store the locations of file's data ext4 uses a data structure common
284  *      for UNIX filesystems - tree of pointers anchored in the inode, with
285  *      data blocks at leaves and indirect blocks in intermediate nodes.
286  *      This function translates the block number into path in that tree -
287  *      return value is the path length and @offsets[n] is the offset of
288  *      pointer to (n+1)th node in the nth one. If @block is out of range
289  *      (negative or too large) warning is printed and zero returned.
290  *
291  *      Note: function doesn't find node addresses, so no IO is needed. All
292  *      we need to know is the capacity of indirect blocks (taken from the
293  *      inode->i_sb).
294  */
295
296 /*
297  * Portability note: the last comparison (check that we fit into triple
298  * indirect block) is spelled differently, because otherwise on an
299  * architecture with 32-bit longs and 8Kb pages we might get into trouble
300  * if our filesystem had 8Kb blocks. We might use long long, but that would
301  * kill us on x86. Oh, well, at least the sign propagation does not matter -
302  * i_block would have to be negative in the very beginning, so we would not
303  * get there at all.
304  */
305
306 static int ext4_block_to_path(struct inode *inode,
307                               ext4_lblk_t i_block,
308                               ext4_lblk_t offsets[4], int *boundary)
309 {
310         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
311         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
312         const long direct_blocks = EXT4_NDIR_BLOCKS,
313                 indirect_blocks = ptrs,
314                 double_blocks = (1 << (ptrs_bits * 2));
315         int n = 0;
316         int final = 0;
317
318         if (i_block < direct_blocks) {
319                 offsets[n++] = i_block;
320                 final = direct_blocks;
321         } else if ((i_block -= direct_blocks) < indirect_blocks) {
322                 offsets[n++] = EXT4_IND_BLOCK;
323                 offsets[n++] = i_block;
324                 final = ptrs;
325         } else if ((i_block -= indirect_blocks) < double_blocks) {
326                 offsets[n++] = EXT4_DIND_BLOCK;
327                 offsets[n++] = i_block >> ptrs_bits;
328                 offsets[n++] = i_block & (ptrs - 1);
329                 final = ptrs;
330         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
331                 offsets[n++] = EXT4_TIND_BLOCK;
332                 offsets[n++] = i_block >> (ptrs_bits * 2);
333                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
334                 offsets[n++] = i_block & (ptrs - 1);
335                 final = ptrs;
336         } else {
337                 ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
338                              i_block + direct_blocks +
339                              indirect_blocks + double_blocks, inode->i_ino);
340         }
341         if (boundary)
342                 *boundary = final - 1 - (i_block & (ptrs - 1));
343         return n;
344 }
345
346 static int __ext4_check_blockref(const char *function, unsigned int line,
347                                  struct inode *inode,
348                                  __le32 *p, unsigned int max)
349 {
350         struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
351         __le32 *bref = p;
352         unsigned int blk;
353
354         while (bref < p+max) {
355                 blk = le32_to_cpu(*bref++);
356                 if (blk &&
357                     unlikely(!ext4_data_block_valid(EXT4_SB(inode->i_sb),
358                                                     blk, 1))) {
359                         es->s_last_error_block = cpu_to_le64(blk);
360                         ext4_error_inode(inode, function, line, blk,
361                                          "invalid block");
362                         return -EIO;
363                 }
364         }
365         return 0;
366 }
367
368
369 #define ext4_check_indirect_blockref(inode, bh)                         \
370         __ext4_check_blockref(__func__, __LINE__, inode,                \
371                               (__le32 *)(bh)->b_data,                   \
372                               EXT4_ADDR_PER_BLOCK((inode)->i_sb))
373
374 #define ext4_check_inode_blockref(inode)                                \
375         __ext4_check_blockref(__func__, __LINE__, inode,                \
376                               EXT4_I(inode)->i_data,                    \
377                               EXT4_NDIR_BLOCKS)
378
379 /**
380  *      ext4_get_branch - read the chain of indirect blocks leading to data
381  *      @inode: inode in question
382  *      @depth: depth of the chain (1 - direct pointer, etc.)
383  *      @offsets: offsets of pointers in inode/indirect blocks
384  *      @chain: place to store the result
385  *      @err: here we store the error value
386  *
387  *      Function fills the array of triples <key, p, bh> and returns %NULL
388  *      if everything went OK or the pointer to the last filled triple
389  *      (incomplete one) otherwise. Upon the return chain[i].key contains
390  *      the number of (i+1)-th block in the chain (as it is stored in memory,
391  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
392  *      number (it points into struct inode for i==0 and into the bh->b_data
393  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
394  *      block for i>0 and NULL for i==0. In other words, it holds the block
395  *      numbers of the chain, addresses they were taken from (and where we can
396  *      verify that chain did not change) and buffer_heads hosting these
397  *      numbers.
398  *
399  *      Function stops when it stumbles upon zero pointer (absent block)
400  *              (pointer to last triple returned, *@err == 0)
401  *      or when it gets an IO error reading an indirect block
402  *              (ditto, *@err == -EIO)
403  *      or when it reads all @depth-1 indirect blocks successfully and finds
404  *      the whole chain, all way to the data (returns %NULL, *err == 0).
405  *
406  *      Need to be called with
407  *      down_read(&EXT4_I(inode)->i_data_sem)
408  */
409 static Indirect *ext4_get_branch(struct inode *inode, int depth,
410                                  ext4_lblk_t  *offsets,
411                                  Indirect chain[4], int *err)
412 {
413         struct super_block *sb = inode->i_sb;
414         Indirect *p = chain;
415         struct buffer_head *bh;
416
417         *err = 0;
418         /* i_data is not going away, no lock needed */
419         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
420         if (!p->key)
421                 goto no_block;
422         while (--depth) {
423                 bh = sb_getblk(sb, le32_to_cpu(p->key));
424                 if (unlikely(!bh))
425                         goto failure;
426
427                 if (!bh_uptodate_or_lock(bh)) {
428                         if (bh_submit_read(bh) < 0) {
429                                 put_bh(bh);
430                                 goto failure;
431                         }
432                         /* validate block references */
433                         if (ext4_check_indirect_blockref(inode, bh)) {
434                                 put_bh(bh);
435                                 goto failure;
436                         }
437                 }
438
439                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
440                 /* Reader: end */
441                 if (!p->key)
442                         goto no_block;
443         }
444         return NULL;
445
446 failure:
447         *err = -EIO;
448 no_block:
449         return p;
450 }
451
452 /**
453  *      ext4_find_near - find a place for allocation with sufficient locality
454  *      @inode: owner
455  *      @ind: descriptor of indirect block.
456  *
457  *      This function returns the preferred place for block allocation.
458  *      It is used when heuristic for sequential allocation fails.
459  *      Rules are:
460  *        + if there is a block to the left of our position - allocate near it.
461  *        + if pointer will live in indirect block - allocate near that block.
462  *        + if pointer will live in inode - allocate in the same
463  *          cylinder group.
464  *
465  * In the latter case we colour the starting block by the callers PID to
466  * prevent it from clashing with concurrent allocations for a different inode
467  * in the same block group.   The PID is used here so that functionally related
468  * files will be close-by on-disk.
469  *
470  *      Caller must make sure that @ind is valid and will stay that way.
471  */
472 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
473 {
474         struct ext4_inode_info *ei = EXT4_I(inode);
475         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
476         __le32 *p;
477         ext4_fsblk_t bg_start;
478         ext4_fsblk_t last_block;
479         ext4_grpblk_t colour;
480         ext4_group_t block_group;
481         int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
482
483         /* Try to find previous block */
484         for (p = ind->p - 1; p >= start; p--) {
485                 if (*p)
486                         return le32_to_cpu(*p);
487         }
488
489         /* No such thing, so let's try location of indirect block */
490         if (ind->bh)
491                 return ind->bh->b_blocknr;
492
493         /*
494          * It is going to be referred to from the inode itself? OK, just put it
495          * into the same cylinder group then.
496          */
497         block_group = ei->i_block_group;
498         if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
499                 block_group &= ~(flex_size-1);
500                 if (S_ISREG(inode->i_mode))
501                         block_group++;
502         }
503         bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
504         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
505
506         /*
507          * If we are doing delayed allocation, we don't need take
508          * colour into account.
509          */
510         if (test_opt(inode->i_sb, DELALLOC))
511                 return bg_start;
512
513         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
514                 colour = (current->pid % 16) *
515                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
516         else
517                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
518         return bg_start + colour;
519 }
520
521 /**
522  *      ext4_find_goal - find a preferred place for allocation.
523  *      @inode: owner
524  *      @block:  block we want
525  *      @partial: pointer to the last triple within a chain
526  *
527  *      Normally this function find the preferred place for block allocation,
528  *      returns it.
529  *      Because this is only used for non-extent files, we limit the block nr
530  *      to 32 bits.
531  */
532 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
533                                    Indirect *partial)
534 {
535         ext4_fsblk_t goal;
536
537         /*
538          * XXX need to get goal block from mballoc's data structures
539          */
540
541         goal = ext4_find_near(inode, partial);
542         goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
543         return goal;
544 }
545
546 /**
547  *      ext4_blks_to_allocate: Look up the block map and count the number
548  *      of direct blocks need to be allocated for the given branch.
549  *
550  *      @branch: chain of indirect blocks
551  *      @k: number of blocks need for indirect blocks
552  *      @blks: number of data blocks to be mapped.
553  *      @blocks_to_boundary:  the offset in the indirect block
554  *
555  *      return the total number of blocks to be allocate, including the
556  *      direct and indirect blocks.
557  */
558 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
559                                  int blocks_to_boundary)
560 {
561         unsigned int count = 0;
562
563         /*
564          * Simple case, [t,d]Indirect block(s) has not allocated yet
565          * then it's clear blocks on that path have not allocated
566          */
567         if (k > 0) {
568                 /* right now we don't handle cross boundary allocation */
569                 if (blks < blocks_to_boundary + 1)
570                         count += blks;
571                 else
572                         count += blocks_to_boundary + 1;
573                 return count;
574         }
575
576         count++;
577         while (count < blks && count <= blocks_to_boundary &&
578                 le32_to_cpu(*(branch[0].p + count)) == 0) {
579                 count++;
580         }
581         return count;
582 }
583
584 /**
585  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
586  *      @indirect_blks: the number of blocks need to allocate for indirect
587  *                      blocks
588  *
589  *      @new_blocks: on return it will store the new block numbers for
590  *      the indirect blocks(if needed) and the first direct block,
591  *      @blks:  on return it will store the total number of allocated
592  *              direct blocks
593  */
594 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
595                              ext4_lblk_t iblock, ext4_fsblk_t goal,
596                              int indirect_blks, int blks,
597                              ext4_fsblk_t new_blocks[4], int *err)
598 {
599         struct ext4_allocation_request ar;
600         int target, i;
601         unsigned long count = 0, blk_allocated = 0;
602         int index = 0;
603         ext4_fsblk_t current_block = 0;
604         int ret = 0;
605
606         /*
607          * Here we try to allocate the requested multiple blocks at once,
608          * on a best-effort basis.
609          * To build a branch, we should allocate blocks for
610          * the indirect blocks(if not allocated yet), and at least
611          * the first direct block of this branch.  That's the
612          * minimum number of blocks need to allocate(required)
613          */
614         /* first we try to allocate the indirect blocks */
615         target = indirect_blks;
616         while (target > 0) {
617                 count = target;
618                 /* allocating blocks for indirect blocks and direct blocks */
619                 current_block = ext4_new_meta_blocks(handle, inode,
620                                                         goal, &count, err);
621                 if (*err)
622                         goto failed_out;
623
624                 if (unlikely(current_block + count > EXT4_MAX_BLOCK_FILE_PHYS)) {
625                         EXT4_ERROR_INODE(inode,
626                                          "current_block %llu + count %lu > %d!",
627                                          current_block, count,
628                                          EXT4_MAX_BLOCK_FILE_PHYS);
629                         *err = -EIO;
630                         goto failed_out;
631                 }
632
633                 target -= count;
634                 /* allocate blocks for indirect blocks */
635                 while (index < indirect_blks && count) {
636                         new_blocks[index++] = current_block++;
637                         count--;
638                 }
639                 if (count > 0) {
640                         /*
641                          * save the new block number
642                          * for the first direct block
643                          */
644                         new_blocks[index] = current_block;
645                         printk(KERN_INFO "%s returned more blocks than "
646                                                 "requested\n", __func__);
647                         WARN_ON(1);
648                         break;
649                 }
650         }
651
652         target = blks - count ;
653         blk_allocated = count;
654         if (!target)
655                 goto allocated;
656         /* Now allocate data blocks */
657         memset(&ar, 0, sizeof(ar));
658         ar.inode = inode;
659         ar.goal = goal;
660         ar.len = target;
661         ar.logical = iblock;
662         if (S_ISREG(inode->i_mode))
663                 /* enable in-core preallocation only for regular files */
664                 ar.flags = EXT4_MB_HINT_DATA;
665
666         current_block = ext4_mb_new_blocks(handle, &ar, err);
667         if (unlikely(current_block + ar.len > EXT4_MAX_BLOCK_FILE_PHYS)) {
668                 EXT4_ERROR_INODE(inode,
669                                  "current_block %llu + ar.len %d > %d!",
670                                  current_block, ar.len,
671                                  EXT4_MAX_BLOCK_FILE_PHYS);
672                 *err = -EIO;
673                 goto failed_out;
674         }
675
676         if (*err && (target == blks)) {
677                 /*
678                  * if the allocation failed and we didn't allocate
679                  * any blocks before
680                  */
681                 goto failed_out;
682         }
683         if (!*err) {
684                 if (target == blks) {
685                         /*
686                          * save the new block number
687                          * for the first direct block
688                          */
689                         new_blocks[index] = current_block;
690                 }
691                 blk_allocated += ar.len;
692         }
693 allocated:
694         /* total number of blocks allocated for direct blocks */
695         ret = blk_allocated;
696         *err = 0;
697         return ret;
698 failed_out:
699         for (i = 0; i < index; i++)
700                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
701         return ret;
702 }
703
704 /**
705  *      ext4_alloc_branch - allocate and set up a chain of blocks.
706  *      @inode: owner
707  *      @indirect_blks: number of allocated indirect blocks
708  *      @blks: number of allocated direct blocks
709  *      @offsets: offsets (in the blocks) to store the pointers to next.
710  *      @branch: place to store the chain in.
711  *
712  *      This function allocates blocks, zeroes out all but the last one,
713  *      links them into chain and (if we are synchronous) writes them to disk.
714  *      In other words, it prepares a branch that can be spliced onto the
715  *      inode. It stores the information about that chain in the branch[], in
716  *      the same format as ext4_get_branch() would do. We are calling it after
717  *      we had read the existing part of chain and partial points to the last
718  *      triple of that (one with zero ->key). Upon the exit we have the same
719  *      picture as after the successful ext4_get_block(), except that in one
720  *      place chain is disconnected - *branch->p is still zero (we did not
721  *      set the last link), but branch->key contains the number that should
722  *      be placed into *branch->p to fill that gap.
723  *
724  *      If allocation fails we free all blocks we've allocated (and forget
725  *      their buffer_heads) and return the error value the from failed
726  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
727  *      as described above and return 0.
728  */
729 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
730                              ext4_lblk_t iblock, int indirect_blks,
731                              int *blks, ext4_fsblk_t goal,
732                              ext4_lblk_t *offsets, Indirect *branch)
733 {
734         int blocksize = inode->i_sb->s_blocksize;
735         int i, n = 0;
736         int err = 0;
737         struct buffer_head *bh;
738         int num;
739         ext4_fsblk_t new_blocks[4];
740         ext4_fsblk_t current_block;
741
742         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
743                                 *blks, new_blocks, &err);
744         if (err)
745                 return err;
746
747         branch[0].key = cpu_to_le32(new_blocks[0]);
748         /*
749          * metadata blocks and data blocks are allocated.
750          */
751         for (n = 1; n <= indirect_blks;  n++) {
752                 /*
753                  * Get buffer_head for parent block, zero it out
754                  * and set the pointer to new one, then send
755                  * parent to disk.
756                  */
757                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
758                 branch[n].bh = bh;
759                 lock_buffer(bh);
760                 BUFFER_TRACE(bh, "call get_create_access");
761                 err = ext4_journal_get_create_access(handle, bh);
762                 if (err) {
763                         /* Don't brelse(bh) here; it's done in
764                          * ext4_journal_forget() below */
765                         unlock_buffer(bh);
766                         goto failed;
767                 }
768
769                 memset(bh->b_data, 0, blocksize);
770                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
771                 branch[n].key = cpu_to_le32(new_blocks[n]);
772                 *branch[n].p = branch[n].key;
773                 if (n == indirect_blks) {
774                         current_block = new_blocks[n];
775                         /*
776                          * End of chain, update the last new metablock of
777                          * the chain to point to the new allocated
778                          * data blocks numbers
779                          */
780                         for (i = 1; i < num; i++)
781                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
782                 }
783                 BUFFER_TRACE(bh, "marking uptodate");
784                 set_buffer_uptodate(bh);
785                 unlock_buffer(bh);
786
787                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
788                 err = ext4_handle_dirty_metadata(handle, inode, bh);
789                 if (err)
790                         goto failed;
791         }
792         *blks = num;
793         return err;
794 failed:
795         /* Allocation failed, free what we already allocated */
796         ext4_free_blocks(handle, inode, 0, new_blocks[0], 1, 0);
797         for (i = 1; i <= n ; i++) {
798                 /*
799                  * branch[i].bh is newly allocated, so there is no
800                  * need to revoke the block, which is why we don't
801                  * need to set EXT4_FREE_BLOCKS_METADATA.
802                  */
803                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1,
804                                  EXT4_FREE_BLOCKS_FORGET);
805         }
806         for (i = n+1; i < indirect_blks; i++)
807                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
808
809         ext4_free_blocks(handle, inode, 0, new_blocks[i], num, 0);
810
811         return err;
812 }
813
814 /**
815  * ext4_splice_branch - splice the allocated branch onto inode.
816  * @inode: owner
817  * @block: (logical) number of block we are adding
818  * @chain: chain of indirect blocks (with a missing link - see
819  *      ext4_alloc_branch)
820  * @where: location of missing link
821  * @num:   number of indirect blocks we are adding
822  * @blks:  number of direct blocks we are adding
823  *
824  * This function fills the missing link and does all housekeeping needed in
825  * inode (->i_blocks, etc.). In case of success we end up with the full
826  * chain to new block and return 0.
827  */
828 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
829                               ext4_lblk_t block, Indirect *where, int num,
830                               int blks)
831 {
832         int i;
833         int err = 0;
834         ext4_fsblk_t current_block;
835
836         /*
837          * If we're splicing into a [td]indirect block (as opposed to the
838          * inode) then we need to get write access to the [td]indirect block
839          * before the splice.
840          */
841         if (where->bh) {
842                 BUFFER_TRACE(where->bh, "get_write_access");
843                 err = ext4_journal_get_write_access(handle, where->bh);
844                 if (err)
845                         goto err_out;
846         }
847         /* That's it */
848
849         *where->p = where->key;
850
851         /*
852          * Update the host buffer_head or inode to point to more just allocated
853          * direct blocks blocks
854          */
855         if (num == 0 && blks > 1) {
856                 current_block = le32_to_cpu(where->key) + 1;
857                 for (i = 1; i < blks; i++)
858                         *(where->p + i) = cpu_to_le32(current_block++);
859         }
860
861         /* We are done with atomic stuff, now do the rest of housekeeping */
862         /* had we spliced it onto indirect block? */
863         if (where->bh) {
864                 /*
865                  * If we spliced it onto an indirect block, we haven't
866                  * altered the inode.  Note however that if it is being spliced
867                  * onto an indirect block at the very end of the file (the
868                  * file is growing) then we *will* alter the inode to reflect
869                  * the new i_size.  But that is not done here - it is done in
870                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
871                  */
872                 jbd_debug(5, "splicing indirect only\n");
873                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
874                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
875                 if (err)
876                         goto err_out;
877         } else {
878                 /*
879                  * OK, we spliced it into the inode itself on a direct block.
880                  */
881                 ext4_mark_inode_dirty(handle, inode);
882                 jbd_debug(5, "splicing direct\n");
883         }
884         return err;
885
886 err_out:
887         for (i = 1; i <= num; i++) {
888                 /*
889                  * branch[i].bh is newly allocated, so there is no
890                  * need to revoke the block, which is why we don't
891                  * need to set EXT4_FREE_BLOCKS_METADATA.
892                  */
893                 ext4_free_blocks(handle, inode, where[i].bh, 0, 1,
894                                  EXT4_FREE_BLOCKS_FORGET);
895         }
896         ext4_free_blocks(handle, inode, 0, le32_to_cpu(where[num].key),
897                          blks, 0);
898
899         return err;
900 }
901
902 /*
903  * The ext4_ind_map_blocks() function handles non-extents inodes
904  * (i.e., using the traditional indirect/double-indirect i_blocks
905  * scheme) for ext4_map_blocks().
906  *
907  * Allocation strategy is simple: if we have to allocate something, we will
908  * have to go the whole way to leaf. So let's do it before attaching anything
909  * to tree, set linkage between the newborn blocks, write them if sync is
910  * required, recheck the path, free and repeat if check fails, otherwise
911  * set the last missing link (that will protect us from any truncate-generated
912  * removals - all blocks on the path are immune now) and possibly force the
913  * write on the parent block.
914  * That has a nice additional property: no special recovery from the failed
915  * allocations is needed - we simply release blocks and do not touch anything
916  * reachable from inode.
917  *
918  * `handle' can be NULL if create == 0.
919  *
920  * return > 0, # of blocks mapped or allocated.
921  * return = 0, if plain lookup failed.
922  * return < 0, error case.
923  *
924  * The ext4_ind_get_blocks() function should be called with
925  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
926  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
927  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
928  * blocks.
929  */
930 static int ext4_ind_map_blocks(handle_t *handle, struct inode *inode,
931                                struct ext4_map_blocks *map,
932                                int flags)
933 {
934         int err = -EIO;
935         ext4_lblk_t offsets[4];
936         Indirect chain[4];
937         Indirect *partial;
938         ext4_fsblk_t goal;
939         int indirect_blks;
940         int blocks_to_boundary = 0;
941         int depth;
942         int count = 0;
943         ext4_fsblk_t first_block = 0;
944
945         J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
946         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
947         depth = ext4_block_to_path(inode, map->m_lblk, offsets,
948                                    &blocks_to_boundary);
949
950         if (depth == 0)
951                 goto out;
952
953         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
954
955         /* Simplest case - block found, no allocation needed */
956         if (!partial) {
957                 first_block = le32_to_cpu(chain[depth - 1].key);
958                 count++;
959                 /*map more blocks*/
960                 while (count < map->m_len && count <= blocks_to_boundary) {
961                         ext4_fsblk_t blk;
962
963                         blk = le32_to_cpu(*(chain[depth-1].p + count));
964
965                         if (blk == first_block + count)
966                                 count++;
967                         else
968                                 break;
969                 }
970                 goto got_it;
971         }
972
973         /* Next simple case - plain lookup or failed read of indirect block */
974         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
975                 goto cleanup;
976
977         /*
978          * Okay, we need to do block allocation.
979         */
980         goal = ext4_find_goal(inode, map->m_lblk, partial);
981
982         /* the number of blocks need to allocate for [d,t]indirect blocks */
983         indirect_blks = (chain + depth) - partial - 1;
984
985         /*
986          * Next look up the indirect map to count the totoal number of
987          * direct blocks to allocate for this branch.
988          */
989         count = ext4_blks_to_allocate(partial, indirect_blks,
990                                       map->m_len, blocks_to_boundary);
991         /*
992          * Block out ext4_truncate while we alter the tree
993          */
994         err = ext4_alloc_branch(handle, inode, map->m_lblk, indirect_blks,
995                                 &count, goal,
996                                 offsets + (partial - chain), partial);
997
998         /*
999          * The ext4_splice_branch call will free and forget any buffers
1000          * on the new chain if there is a failure, but that risks using
1001          * up transaction credits, especially for bitmaps where the
1002          * credits cannot be returned.  Can we handle this somehow?  We
1003          * may need to return -EAGAIN upwards in the worst case.  --sct
1004          */
1005         if (!err)
1006                 err = ext4_splice_branch(handle, inode, map->m_lblk,
1007                                          partial, indirect_blks, count);
1008         if (err)
1009                 goto cleanup;
1010
1011         map->m_flags |= EXT4_MAP_NEW;
1012
1013         ext4_update_inode_fsync_trans(handle, inode, 1);
1014 got_it:
1015         map->m_flags |= EXT4_MAP_MAPPED;
1016         map->m_pblk = le32_to_cpu(chain[depth-1].key);
1017         map->m_len = count;
1018         if (count > blocks_to_boundary)
1019                 map->m_flags |= EXT4_MAP_BOUNDARY;
1020         err = count;
1021         /* Clean up and exit */
1022         partial = chain + depth - 1;    /* the whole chain */
1023 cleanup:
1024         while (partial > chain) {
1025                 BUFFER_TRACE(partial->bh, "call brelse");
1026                 brelse(partial->bh);
1027                 partial--;
1028         }
1029 out:
1030         return err;
1031 }
1032
1033 #ifdef CONFIG_QUOTA
1034 qsize_t *ext4_get_reserved_space(struct inode *inode)
1035 {
1036         return &EXT4_I(inode)->i_reserved_quota;
1037 }
1038 #endif
1039
1040 /*
1041  * Calculate the number of metadata blocks need to reserve
1042  * to allocate a new block at @lblocks for non extent file based file
1043  */
1044 static int ext4_indirect_calc_metadata_amount(struct inode *inode,
1045                                               sector_t lblock)
1046 {
1047         struct ext4_inode_info *ei = EXT4_I(inode);
1048         sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
1049         int blk_bits;
1050
1051         if (lblock < EXT4_NDIR_BLOCKS)
1052                 return 0;
1053
1054         lblock -= EXT4_NDIR_BLOCKS;
1055
1056         if (ei->i_da_metadata_calc_len &&
1057             (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
1058                 ei->i_da_metadata_calc_len++;
1059                 return 0;
1060         }
1061         ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
1062         ei->i_da_metadata_calc_len = 1;
1063         blk_bits = order_base_2(lblock);
1064         return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
1065 }
1066
1067 /*
1068  * Calculate the number of metadata blocks need to reserve
1069  * to allocate a block located at @lblock
1070  */
1071 static int ext4_calc_metadata_amount(struct inode *inode, sector_t lblock)
1072 {
1073         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1074                 return ext4_ext_calc_metadata_amount(inode, lblock);
1075
1076         return ext4_indirect_calc_metadata_amount(inode, lblock);
1077 }
1078
1079 /*
1080  * Called with i_data_sem down, which is important since we can call
1081  * ext4_discard_preallocations() from here.
1082  */
1083 void ext4_da_update_reserve_space(struct inode *inode,
1084                                         int used, int quota_claim)
1085 {
1086         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1087         struct ext4_inode_info *ei = EXT4_I(inode);
1088
1089         spin_lock(&ei->i_block_reservation_lock);
1090         trace_ext4_da_update_reserve_space(inode, used);
1091         if (unlikely(used > ei->i_reserved_data_blocks)) {
1092                 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
1093                          "with only %d reserved data blocks\n",
1094                          __func__, inode->i_ino, used,
1095                          ei->i_reserved_data_blocks);
1096                 WARN_ON(1);
1097                 used = ei->i_reserved_data_blocks;
1098         }
1099
1100         /* Update per-inode reservations */
1101         ei->i_reserved_data_blocks -= used;
1102         ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
1103         percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1104                            used + ei->i_allocated_meta_blocks);
1105         ei->i_allocated_meta_blocks = 0;
1106
1107         if (ei->i_reserved_data_blocks == 0) {
1108                 /*
1109                  * We can release all of the reserved metadata blocks
1110                  * only when we have written all of the delayed
1111                  * allocation blocks.
1112                  */
1113                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1114                                    ei->i_reserved_meta_blocks);
1115                 ei->i_reserved_meta_blocks = 0;
1116                 ei->i_da_metadata_calc_len = 0;
1117         }
1118         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1119
1120         /* Update quota subsystem for data blocks */
1121         if (quota_claim)
1122                 dquot_claim_block(inode, used);
1123         else {
1124                 /*
1125                  * We did fallocate with an offset that is already delayed
1126                  * allocated. So on delayed allocated writeback we should
1127                  * not re-claim the quota for fallocated blocks.
1128                  */
1129                 dquot_release_reservation_block(inode, used);
1130         }
1131
1132         /*
1133          * If we have done all the pending block allocations and if
1134          * there aren't any writers on the inode, we can discard the
1135          * inode's preallocations.
1136          */
1137         if ((ei->i_reserved_data_blocks == 0) &&
1138             (atomic_read(&inode->i_writecount) == 0))
1139                 ext4_discard_preallocations(inode);
1140 }
1141
1142 static int __check_block_validity(struct inode *inode, const char *func,
1143                                 unsigned int line,
1144                                 struct ext4_map_blocks *map)
1145 {
1146         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
1147                                    map->m_len)) {
1148                 ext4_error_inode(inode, func, line, map->m_pblk,
1149                                  "lblock %lu mapped to illegal pblock "
1150                                  "(length %d)", (unsigned long) map->m_lblk,
1151                                  map->m_len);
1152                 return -EIO;
1153         }
1154         return 0;
1155 }
1156
1157 #define check_block_validity(inode, map)        \
1158         __check_block_validity((inode), __func__, __LINE__, (map))
1159
1160 /*
1161  * Return the number of contiguous dirty pages in a given inode
1162  * starting at page frame idx.
1163  */
1164 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
1165                                     unsigned int max_pages)
1166 {
1167         struct address_space *mapping = inode->i_mapping;
1168         pgoff_t index;
1169         struct pagevec pvec;
1170         pgoff_t num = 0;
1171         int i, nr_pages, done = 0;
1172
1173         if (max_pages == 0)
1174                 return 0;
1175         pagevec_init(&pvec, 0);
1176         while (!done) {
1177                 index = idx;
1178                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
1179                                               PAGECACHE_TAG_DIRTY,
1180                                               (pgoff_t)PAGEVEC_SIZE);
1181                 if (nr_pages == 0)
1182                         break;
1183                 for (i = 0; i < nr_pages; i++) {
1184                         struct page *page = pvec.pages[i];
1185                         struct buffer_head *bh, *head;
1186
1187                         lock_page(page);
1188                         if (unlikely(page->mapping != mapping) ||
1189                             !PageDirty(page) ||
1190                             PageWriteback(page) ||
1191                             page->index != idx) {
1192                                 done = 1;
1193                                 unlock_page(page);
1194                                 break;
1195                         }
1196                         if (page_has_buffers(page)) {
1197                                 bh = head = page_buffers(page);
1198                                 do {
1199                                         if (!buffer_delay(bh) &&
1200                                             !buffer_unwritten(bh))
1201                                                 done = 1;
1202                                         bh = bh->b_this_page;
1203                                 } while (!done && (bh != head));
1204                         }
1205                         unlock_page(page);
1206                         if (done)
1207                                 break;
1208                         idx++;
1209                         num++;
1210                         if (num >= max_pages) {
1211                                 done = 1;
1212                                 break;
1213                         }
1214                 }
1215                 pagevec_release(&pvec);
1216         }
1217         return num;
1218 }
1219
1220 /*
1221  * The ext4_map_blocks() function tries to look up the requested blocks,
1222  * and returns if the blocks are already mapped.
1223  *
1224  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1225  * and store the allocated blocks in the result buffer head and mark it
1226  * mapped.
1227  *
1228  * If file type is extents based, it will call ext4_ext_map_blocks(),
1229  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
1230  * based files
1231  *
1232  * On success, it returns the number of blocks being mapped or allocate.
1233  * if create==0 and the blocks are pre-allocated and uninitialized block,
1234  * the result buffer head is unmapped. If the create ==1, it will make sure
1235  * the buffer head is mapped.
1236  *
1237  * It returns 0 if plain look up failed (blocks have not been allocated), in
1238  * that casem, buffer head is unmapped
1239  *
1240  * It returns the error in case of allocation failure.
1241  */
1242 int ext4_map_blocks(handle_t *handle, struct inode *inode,
1243                     struct ext4_map_blocks *map, int flags)
1244 {
1245         int retval;
1246
1247         map->m_flags = 0;
1248         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
1249                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
1250                   (unsigned long) map->m_lblk);
1251         /*
1252          * Try to see if we can get the block without requesting a new
1253          * file system block.
1254          */
1255         down_read((&EXT4_I(inode)->i_data_sem));
1256         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
1257                 retval = ext4_ext_map_blocks(handle, inode, map, 0);
1258         } else {
1259                 retval = ext4_ind_map_blocks(handle, inode, map, 0);
1260         }
1261         up_read((&EXT4_I(inode)->i_data_sem));
1262
1263         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
1264                 int ret = check_block_validity(inode, map);
1265                 if (ret != 0)
1266                         return ret;
1267         }
1268
1269         /* If it is only a block(s) look up */
1270         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1271                 return retval;
1272
1273         /*
1274          * Returns if the blocks have already allocated
1275          *
1276          * Note that if blocks have been preallocated
1277          * ext4_ext_get_block() returns th create = 0
1278          * with buffer head unmapped.
1279          */
1280         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
1281                 return retval;
1282
1283         /*
1284          * When we call get_blocks without the create flag, the
1285          * BH_Unwritten flag could have gotten set if the blocks
1286          * requested were part of a uninitialized extent.  We need to
1287          * clear this flag now that we are committed to convert all or
1288          * part of the uninitialized extent to be an initialized
1289          * extent.  This is because we need to avoid the combination
1290          * of BH_Unwritten and BH_Mapped flags being simultaneously
1291          * set on the buffer_head.
1292          */
1293         map->m_flags &= ~EXT4_MAP_UNWRITTEN;
1294
1295         /*
1296          * New blocks allocate and/or writing to uninitialized extent
1297          * will possibly result in updating i_data, so we take
1298          * the write lock of i_data_sem, and call get_blocks()
1299          * with create == 1 flag.
1300          */
1301         down_write((&EXT4_I(inode)->i_data_sem));
1302
1303         /*
1304          * if the caller is from delayed allocation writeout path
1305          * we have already reserved fs blocks for allocation
1306          * let the underlying get_block() function know to
1307          * avoid double accounting
1308          */
1309         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1310                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1311         /*
1312          * We need to check for EXT4 here because migrate
1313          * could have changed the inode type in between
1314          */
1315         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
1316                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
1317         } else {
1318                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
1319
1320                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
1321                         /*
1322                          * We allocated new blocks which will result in
1323                          * i_data's format changing.  Force the migrate
1324                          * to fail by clearing migrate flags
1325                          */
1326                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
1327                 }
1328
1329                 /*
1330                  * Update reserved blocks/metadata blocks after successful
1331                  * block allocation which had been deferred till now. We don't
1332                  * support fallocate for non extent files. So we can update
1333                  * reserve space here.
1334                  */
1335                 if ((retval > 0) &&
1336                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
1337                         ext4_da_update_reserve_space(inode, retval, 1);
1338         }
1339         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1340                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1341
1342         up_write((&EXT4_I(inode)->i_data_sem));
1343         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
1344                 int ret = check_block_validity(inode, map);
1345                 if (ret != 0)
1346                         return ret;
1347         }
1348         return retval;
1349 }
1350
1351 /* Maximum number of blocks we map for direct IO at once. */
1352 #define DIO_MAX_BLOCKS 4096
1353
1354 static int _ext4_get_block(struct inode *inode, sector_t iblock,
1355                            struct buffer_head *bh, int flags)
1356 {
1357         handle_t *handle = ext4_journal_current_handle();
1358         struct ext4_map_blocks map;
1359         int ret = 0, started = 0;
1360         int dio_credits;
1361
1362         map.m_lblk = iblock;
1363         map.m_len = bh->b_size >> inode->i_blkbits;
1364
1365         if (flags && !handle) {
1366                 /* Direct IO write... */
1367                 if (map.m_len > DIO_MAX_BLOCKS)
1368                         map.m_len = DIO_MAX_BLOCKS;
1369                 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
1370                 handle = ext4_journal_start(inode, dio_credits);
1371                 if (IS_ERR(handle)) {
1372                         ret = PTR_ERR(handle);
1373                         return ret;
1374                 }
1375                 started = 1;
1376         }
1377
1378         ret = ext4_map_blocks(handle, inode, &map, flags);
1379         if (ret > 0) {
1380                 map_bh(bh, inode->i_sb, map.m_pblk);
1381                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
1382                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
1383                 ret = 0;
1384         }
1385         if (started)
1386                 ext4_journal_stop(handle);
1387         return ret;
1388 }
1389
1390 int ext4_get_block(struct inode *inode, sector_t iblock,
1391                    struct buffer_head *bh, int create)
1392 {
1393         return _ext4_get_block(inode, iblock, bh,
1394                                create ? EXT4_GET_BLOCKS_CREATE : 0);
1395 }
1396
1397 /*
1398  * `handle' can be NULL if create is zero
1399  */
1400 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1401                                 ext4_lblk_t block, int create, int *errp)
1402 {
1403         struct ext4_map_blocks map;
1404         struct buffer_head *bh;
1405         int fatal = 0, err;
1406
1407         J_ASSERT(handle != NULL || create == 0);
1408
1409         map.m_lblk = block;
1410         map.m_len = 1;
1411         err = ext4_map_blocks(handle, inode, &map,
1412                               create ? EXT4_GET_BLOCKS_CREATE : 0);
1413
1414         if (err < 0)
1415                 *errp = err;
1416         if (err <= 0)
1417                 return NULL;
1418         *errp = 0;
1419
1420         bh = sb_getblk(inode->i_sb, map.m_pblk);
1421         if (!bh) {
1422                 *errp = -EIO;
1423                 return NULL;
1424         }
1425         if (map.m_flags & EXT4_MAP_NEW) {
1426                 J_ASSERT(create != 0);
1427                 J_ASSERT(handle != NULL);
1428
1429                 /*
1430                  * Now that we do not always journal data, we should
1431                  * keep in mind whether this should always journal the
1432                  * new buffer as metadata.  For now, regular file
1433                  * writes use ext4_get_block instead, so it's not a
1434                  * problem.
1435                  */
1436                 lock_buffer(bh);
1437                 BUFFER_TRACE(bh, "call get_create_access");
1438                 fatal = ext4_journal_get_create_access(handle, bh);
1439                 if (!fatal && !buffer_uptodate(bh)) {
1440                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1441                         set_buffer_uptodate(bh);
1442                 }
1443                 unlock_buffer(bh);
1444                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1445                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1446                 if (!fatal)
1447                         fatal = err;
1448         } else {
1449                 BUFFER_TRACE(bh, "not a new buffer");
1450         }
1451         if (fatal) {
1452                 *errp = fatal;
1453                 brelse(bh);
1454                 bh = NULL;
1455         }
1456         return bh;
1457 }
1458
1459 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1460                                ext4_lblk_t block, int create, int *err)
1461 {
1462         struct buffer_head *bh;
1463
1464         bh = ext4_getblk(handle, inode, block, create, err);
1465         if (!bh)
1466                 return bh;
1467         if (buffer_uptodate(bh))
1468                 return bh;
1469         ll_rw_block(READ_META, 1, &bh);
1470         wait_on_buffer(bh);
1471         if (buffer_uptodate(bh))
1472                 return bh;
1473         put_bh(bh);
1474         *err = -EIO;
1475         return NULL;
1476 }
1477
1478 static int walk_page_buffers(handle_t *handle,
1479                              struct buffer_head *head,
1480                              unsigned from,
1481                              unsigned to,
1482                              int *partial,
1483                              int (*fn)(handle_t *handle,
1484                                        struct buffer_head *bh))
1485 {
1486         struct buffer_head *bh;
1487         unsigned block_start, block_end;
1488         unsigned blocksize = head->b_size;
1489         int err, ret = 0;
1490         struct buffer_head *next;
1491
1492         for (bh = head, block_start = 0;
1493              ret == 0 && (bh != head || !block_start);
1494              block_start = block_end, bh = next) {
1495                 next = bh->b_this_page;
1496                 block_end = block_start + blocksize;
1497                 if (block_end <= from || block_start >= to) {
1498                         if (partial && !buffer_uptodate(bh))
1499                                 *partial = 1;
1500                         continue;
1501                 }
1502                 err = (*fn)(handle, bh);
1503                 if (!ret)
1504                         ret = err;
1505         }
1506         return ret;
1507 }
1508
1509 /*
1510  * To preserve ordering, it is essential that the hole instantiation and
1511  * the data write be encapsulated in a single transaction.  We cannot
1512  * close off a transaction and start a new one between the ext4_get_block()
1513  * and the commit_write().  So doing the jbd2_journal_start at the start of
1514  * prepare_write() is the right place.
1515  *
1516  * Also, this function can nest inside ext4_writepage() ->
1517  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1518  * has generated enough buffer credits to do the whole page.  So we won't
1519  * block on the journal in that case, which is good, because the caller may
1520  * be PF_MEMALLOC.
1521  *
1522  * By accident, ext4 can be reentered when a transaction is open via
1523  * quota file writes.  If we were to commit the transaction while thus
1524  * reentered, there can be a deadlock - we would be holding a quota
1525  * lock, and the commit would never complete if another thread had a
1526  * transaction open and was blocking on the quota lock - a ranking
1527  * violation.
1528  *
1529  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1530  * will _not_ run commit under these circumstances because handle->h_ref
1531  * is elevated.  We'll still have enough credits for the tiny quotafile
1532  * write.
1533  */
1534 static int do_journal_get_write_access(handle_t *handle,
1535                                        struct buffer_head *bh)
1536 {
1537         int dirty = buffer_dirty(bh);
1538         int ret;
1539
1540         if (!buffer_mapped(bh) || buffer_freed(bh))
1541                 return 0;
1542         /*
1543          * __block_prepare_write() could have dirtied some buffers. Clean
1544          * the dirty bit as jbd2_journal_get_write_access() could complain
1545          * otherwise about fs integrity issues. Setting of the dirty bit
1546          * by __block_prepare_write() isn't a real problem here as we clear
1547          * the bit before releasing a page lock and thus writeback cannot
1548          * ever write the buffer.
1549          */
1550         if (dirty)
1551                 clear_buffer_dirty(bh);
1552         ret = ext4_journal_get_write_access(handle, bh);
1553         if (!ret && dirty)
1554                 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1555         return ret;
1556 }
1557
1558 /*
1559  * Truncate blocks that were not used by write. We have to truncate the
1560  * pagecache as well so that corresponding buffers get properly unmapped.
1561  */
1562 static void ext4_truncate_failed_write(struct inode *inode)
1563 {
1564         truncate_inode_pages(inode->i_mapping, inode->i_size);
1565         ext4_truncate(inode);
1566 }
1567
1568 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
1569                    struct buffer_head *bh_result, int create);
1570 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1571                             loff_t pos, unsigned len, unsigned flags,
1572                             struct page **pagep, void **fsdata)
1573 {
1574         struct inode *inode = mapping->host;
1575         int ret, needed_blocks;
1576         handle_t *handle;
1577         int retries = 0;
1578         struct page *page;
1579         pgoff_t index;
1580         unsigned from, to;
1581
1582         trace_ext4_write_begin(inode, pos, len, flags);
1583         /*
1584          * Reserve one block more for addition to orphan list in case
1585          * we allocate blocks but write fails for some reason
1586          */
1587         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1588         index = pos >> PAGE_CACHE_SHIFT;
1589         from = pos & (PAGE_CACHE_SIZE - 1);
1590         to = from + len;
1591
1592 retry:
1593         handle = ext4_journal_start(inode, needed_blocks);
1594         if (IS_ERR(handle)) {
1595                 ret = PTR_ERR(handle);
1596                 goto out;
1597         }
1598
1599         /* We cannot recurse into the filesystem as the transaction is already
1600          * started */
1601         flags |= AOP_FLAG_NOFS;
1602
1603         page = grab_cache_page_write_begin(mapping, index, flags);
1604         if (!page) {
1605                 ext4_journal_stop(handle);
1606                 ret = -ENOMEM;
1607                 goto out;
1608         }
1609         *pagep = page;
1610
1611         if (ext4_should_dioread_nolock(inode))
1612                 ret = __block_write_begin(page, pos, len, ext4_get_block_write);
1613         else
1614                 ret = __block_write_begin(page, pos, len, ext4_get_block);
1615
1616         if (!ret && ext4_should_journal_data(inode)) {
1617                 ret = walk_page_buffers(handle, page_buffers(page),
1618                                 from, to, NULL, do_journal_get_write_access);
1619         }
1620
1621         if (ret) {
1622                 unlock_page(page);
1623                 page_cache_release(page);
1624                 /*
1625                  * __block_write_begin may have instantiated a few blocks
1626                  * outside i_size.  Trim these off again. Don't need
1627                  * i_size_read because we hold i_mutex.
1628                  *
1629                  * Add inode to orphan list in case we crash before
1630                  * truncate finishes
1631                  */
1632                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1633                         ext4_orphan_add(handle, inode);
1634
1635                 ext4_journal_stop(handle);
1636                 if (pos + len > inode->i_size) {
1637                         ext4_truncate_failed_write(inode);
1638                         /*
1639                          * If truncate failed early the inode might
1640                          * still be on the orphan list; we need to
1641                          * make sure the inode is removed from the
1642                          * orphan list in that case.
1643                          */
1644                         if (inode->i_nlink)
1645                                 ext4_orphan_del(NULL, inode);
1646                 }
1647         }
1648
1649         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1650                 goto retry;
1651 out:
1652         return ret;
1653 }
1654
1655 /* For write_end() in data=journal mode */
1656 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1657 {
1658         if (!buffer_mapped(bh) || buffer_freed(bh))
1659                 return 0;
1660         set_buffer_uptodate(bh);
1661         return ext4_handle_dirty_metadata(handle, NULL, bh);
1662 }
1663
1664 static int ext4_generic_write_end(struct file *file,
1665                                   struct address_space *mapping,
1666                                   loff_t pos, unsigned len, unsigned copied,
1667                                   struct page *page, void *fsdata)
1668 {
1669         int i_size_changed = 0;
1670         struct inode *inode = mapping->host;
1671         handle_t *handle = ext4_journal_current_handle();
1672
1673         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1674
1675         /*
1676          * No need to use i_size_read() here, the i_size
1677          * cannot change under us because we hold i_mutex.
1678          *
1679          * But it's important to update i_size while still holding page lock:
1680          * page writeout could otherwise come in and zero beyond i_size.
1681          */
1682         if (pos + copied > inode->i_size) {
1683                 i_size_write(inode, pos + copied);
1684                 i_size_changed = 1;
1685         }
1686
1687         if (pos + copied >  EXT4_I(inode)->i_disksize) {
1688                 /* We need to mark inode dirty even if
1689                  * new_i_size is less that inode->i_size
1690                  * bu greater than i_disksize.(hint delalloc)
1691                  */
1692                 ext4_update_i_disksize(inode, (pos + copied));
1693                 i_size_changed = 1;
1694         }
1695         unlock_page(page);
1696         page_cache_release(page);
1697
1698         /*
1699          * Don't mark the inode dirty under page lock. First, it unnecessarily
1700          * makes the holding time of page lock longer. Second, it forces lock
1701          * ordering of page lock and transaction start for journaling
1702          * filesystems.
1703          */
1704         if (i_size_changed)
1705                 ext4_mark_inode_dirty(handle, inode);
1706
1707         return copied;
1708 }
1709
1710 /*
1711  * We need to pick up the new inode size which generic_commit_write gave us
1712  * `file' can be NULL - eg, when called from page_symlink().
1713  *
1714  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1715  * buffers are managed internally.
1716  */
1717 static int ext4_ordered_write_end(struct file *file,
1718                                   struct address_space *mapping,
1719                                   loff_t pos, unsigned len, unsigned copied,
1720                                   struct page *page, void *fsdata)
1721 {
1722         handle_t *handle = ext4_journal_current_handle();
1723         struct inode *inode = mapping->host;
1724         int ret = 0, ret2;
1725
1726         trace_ext4_ordered_write_end(inode, pos, len, copied);
1727         ret = ext4_jbd2_file_inode(handle, inode);
1728
1729         if (ret == 0) {
1730                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1731                                                         page, fsdata);
1732                 copied = ret2;
1733                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1734                         /* if we have allocated more blocks and copied
1735                          * less. We will have blocks allocated outside
1736                          * inode->i_size. So truncate them
1737                          */
1738                         ext4_orphan_add(handle, inode);
1739                 if (ret2 < 0)
1740                         ret = ret2;
1741         }
1742         ret2 = ext4_journal_stop(handle);
1743         if (!ret)
1744                 ret = ret2;
1745
1746         if (pos + len > inode->i_size) {
1747                 ext4_truncate_failed_write(inode);
1748                 /*
1749                  * If truncate failed early the inode might still be
1750                  * on the orphan list; we need to make sure the inode
1751                  * is removed from the orphan list in that case.
1752                  */
1753                 if (inode->i_nlink)
1754                         ext4_orphan_del(NULL, inode);
1755         }
1756
1757
1758         return ret ? ret : copied;
1759 }
1760
1761 static int ext4_writeback_write_end(struct file *file,
1762                                     struct address_space *mapping,
1763                                     loff_t pos, unsigned len, unsigned copied,
1764                                     struct page *page, void *fsdata)
1765 {
1766         handle_t *handle = ext4_journal_current_handle();
1767         struct inode *inode = mapping->host;
1768         int ret = 0, ret2;
1769
1770         trace_ext4_writeback_write_end(inode, pos, len, copied);
1771         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1772                                                         page, fsdata);
1773         copied = ret2;
1774         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1775                 /* if we have allocated more blocks and copied
1776                  * less. We will have blocks allocated outside
1777                  * inode->i_size. So truncate them
1778                  */
1779                 ext4_orphan_add(handle, inode);
1780
1781         if (ret2 < 0)
1782                 ret = ret2;
1783
1784         ret2 = ext4_journal_stop(handle);
1785         if (!ret)
1786                 ret = ret2;
1787
1788         if (pos + len > inode->i_size) {
1789                 ext4_truncate_failed_write(inode);
1790                 /*
1791                  * If truncate failed early the inode might still be
1792                  * on the orphan list; we need to make sure the inode
1793                  * is removed from the orphan list in that case.
1794                  */
1795                 if (inode->i_nlink)
1796                         ext4_orphan_del(NULL, inode);
1797         }
1798
1799         return ret ? ret : copied;
1800 }
1801
1802 static int ext4_journalled_write_end(struct file *file,
1803                                      struct address_space *mapping,
1804                                      loff_t pos, unsigned len, unsigned copied,
1805                                      struct page *page, void *fsdata)
1806 {
1807         handle_t *handle = ext4_journal_current_handle();
1808         struct inode *inode = mapping->host;
1809         int ret = 0, ret2;
1810         int partial = 0;
1811         unsigned from, to;
1812         loff_t new_i_size;
1813
1814         trace_ext4_journalled_write_end(inode, pos, len, copied);
1815         from = pos & (PAGE_CACHE_SIZE - 1);
1816         to = from + len;
1817
1818         if (copied < len) {
1819                 if (!PageUptodate(page))
1820                         copied = 0;
1821                 page_zero_new_buffers(page, from+copied, to);
1822         }
1823
1824         ret = walk_page_buffers(handle, page_buffers(page), from,
1825                                 to, &partial, write_end_fn);
1826         if (!partial)
1827                 SetPageUptodate(page);
1828         new_i_size = pos + copied;
1829         if (new_i_size > inode->i_size)
1830                 i_size_write(inode, pos+copied);
1831         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1832         if (new_i_size > EXT4_I(inode)->i_disksize) {
1833                 ext4_update_i_disksize(inode, new_i_size);
1834                 ret2 = ext4_mark_inode_dirty(handle, inode);
1835                 if (!ret)
1836                         ret = ret2;
1837         }
1838
1839         unlock_page(page);
1840         page_cache_release(page);
1841         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1842                 /* if we have allocated more blocks and copied
1843                  * less. We will have blocks allocated outside
1844                  * inode->i_size. So truncate them
1845                  */
1846                 ext4_orphan_add(handle, inode);
1847
1848         ret2 = ext4_journal_stop(handle);
1849         if (!ret)
1850                 ret = ret2;
1851         if (pos + len > inode->i_size) {
1852                 ext4_truncate_failed_write(inode);
1853                 /*
1854                  * If truncate failed early the inode might still be
1855                  * on the orphan list; we need to make sure the inode
1856                  * is removed from the orphan list in that case.
1857                  */
1858                 if (inode->i_nlink)
1859                         ext4_orphan_del(NULL, inode);
1860         }
1861
1862         return ret ? ret : copied;
1863 }
1864
1865 /*
1866  * Reserve a single block located at lblock
1867  */
1868 static int ext4_da_reserve_space(struct inode *inode, sector_t lblock)
1869 {
1870         int retries = 0;
1871         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1872         struct ext4_inode_info *ei = EXT4_I(inode);
1873         unsigned long md_needed;
1874         int ret;
1875
1876         /*
1877          * recalculate the amount of metadata blocks to reserve
1878          * in order to allocate nrblocks
1879          * worse case is one extent per block
1880          */
1881 repeat:
1882         spin_lock(&ei->i_block_reservation_lock);
1883         md_needed = ext4_calc_metadata_amount(inode, lblock);
1884         trace_ext4_da_reserve_space(inode, md_needed);
1885         spin_unlock(&ei->i_block_reservation_lock);
1886
1887         /*
1888          * We will charge metadata quota at writeout time; this saves
1889          * us from metadata over-estimation, though we may go over by
1890          * a small amount in the end.  Here we just reserve for data.
1891          */
1892         ret = dquot_reserve_block(inode, 1);
1893         if (ret)
1894                 return ret;
1895         /*
1896          * We do still charge estimated metadata to the sb though;
1897          * we cannot afford to run out of free blocks.
1898          */
1899         if (ext4_claim_free_blocks(sbi, md_needed + 1)) {
1900                 dquot_release_reservation_block(inode, 1);
1901                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1902                         yield();
1903                         goto repeat;
1904                 }
1905                 return -ENOSPC;
1906         }
1907         spin_lock(&ei->i_block_reservation_lock);
1908         ei->i_reserved_data_blocks++;
1909         ei->i_reserved_meta_blocks += md_needed;
1910         spin_unlock(&ei->i_block_reservation_lock);
1911
1912         return 0;       /* success */
1913 }
1914
1915 static void ext4_da_release_space(struct inode *inode, int to_free)
1916 {
1917         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1918         struct ext4_inode_info *ei = EXT4_I(inode);
1919
1920         if (!to_free)
1921                 return;         /* Nothing to release, exit */
1922
1923         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1924
1925         trace_ext4_da_release_space(inode, to_free);
1926         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1927                 /*
1928                  * if there aren't enough reserved blocks, then the
1929                  * counter is messed up somewhere.  Since this
1930                  * function is called from invalidate page, it's
1931                  * harmless to return without any action.
1932                  */
1933                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1934                          "ino %lu, to_free %d with only %d reserved "
1935                          "data blocks\n", inode->i_ino, to_free,
1936                          ei->i_reserved_data_blocks);
1937                 WARN_ON(1);
1938                 to_free = ei->i_reserved_data_blocks;
1939         }
1940         ei->i_reserved_data_blocks -= to_free;
1941
1942         if (ei->i_reserved_data_blocks == 0) {
1943                 /*
1944                  * We can release all of the reserved metadata blocks
1945                  * only when we have written all of the delayed
1946                  * allocation blocks.
1947                  */
1948                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1949                                    ei->i_reserved_meta_blocks);
1950                 ei->i_reserved_meta_blocks = 0;
1951                 ei->i_da_metadata_calc_len = 0;
1952         }
1953
1954         /* update fs dirty data blocks counter */
1955         percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1956
1957         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1958
1959         dquot_release_reservation_block(inode, to_free);
1960 }
1961
1962 static void ext4_da_page_release_reservation(struct page *page,
1963                                              unsigned long offset)
1964 {
1965         int to_release = 0;
1966         struct buffer_head *head, *bh;
1967         unsigned int curr_off = 0;
1968
1969         head = page_buffers(page);
1970         bh = head;
1971         do {
1972                 unsigned int next_off = curr_off + bh->b_size;
1973
1974                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1975                         to_release++;
1976                         clear_buffer_delay(bh);
1977                 }
1978                 curr_off = next_off;
1979         } while ((bh = bh->b_this_page) != head);
1980         ext4_da_release_space(page->mapping->host, to_release);
1981 }
1982
1983 /*
1984  * Delayed allocation stuff
1985  */
1986
1987 /*
1988  * mpage_da_submit_io - walks through extent of pages and try to write
1989  * them with writepage() call back
1990  *
1991  * @mpd->inode: inode
1992  * @mpd->first_page: first page of the extent
1993  * @mpd->next_page: page after the last page of the extent
1994  *
1995  * By the time mpage_da_submit_io() is called we expect all blocks
1996  * to be allocated. this may be wrong if allocation failed.
1997  *
1998  * As pages are already locked by write_cache_pages(), we can't use it
1999  */
2000 static int mpage_da_submit_io(struct mpage_da_data *mpd)
2001 {
2002         long pages_skipped;
2003         struct pagevec pvec;
2004         unsigned long index, end;
2005         int ret = 0, err, nr_pages, i;
2006         struct inode *inode = mpd->inode;
2007         struct address_space *mapping = inode->i_mapping;
2008
2009         BUG_ON(mpd->next_page <= mpd->first_page);
2010         /*
2011          * We need to start from the first_page to the next_page - 1
2012          * to make sure we also write the mapped dirty buffer_heads.
2013          * If we look at mpd->b_blocknr we would only be looking
2014          * at the currently mapped buffer_heads.
2015          */
2016         index = mpd->first_page;
2017         end = mpd->next_page - 1;
2018
2019         pagevec_init(&pvec, 0);
2020         while (index <= end) {
2021                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2022                 if (nr_pages == 0)
2023                         break;
2024                 for (i = 0; i < nr_pages; i++) {
2025                         struct page *page = pvec.pages[i];
2026
2027                         index = page->index;
2028                         if (index > end)
2029                                 break;
2030                         index++;
2031
2032                         BUG_ON(!PageLocked(page));
2033                         BUG_ON(PageWriteback(page));
2034
2035                         pages_skipped = mpd->wbc->pages_skipped;
2036                         err = mapping->a_ops->writepage(page, mpd->wbc);
2037                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
2038                                 /*
2039                                  * have successfully written the page
2040                                  * without skipping the same
2041                                  */
2042                                 mpd->pages_written++;
2043                         /*
2044                          * In error case, we have to continue because
2045                          * remaining pages are still locked
2046                          * XXX: unlock and re-dirty them?
2047                          */
2048                         if (ret == 0)
2049                                 ret = err;
2050                 }
2051                 pagevec_release(&pvec);
2052         }
2053         return ret;
2054 }
2055
2056 /*
2057  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
2058  *
2059  * the function goes through all passed space and put actual disk
2060  * block numbers into buffer heads, dropping BH_Delay and BH_Unwritten
2061  */
2062 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd,
2063                                  struct ext4_map_blocks *map)
2064 {
2065         struct inode *inode = mpd->inode;
2066         struct address_space *mapping = inode->i_mapping;
2067         int blocks = map->m_len;
2068         sector_t pblock = map->m_pblk, cur_logical;
2069         struct buffer_head *head, *bh;
2070         pgoff_t index, end;
2071         struct pagevec pvec;
2072         int nr_pages, i;
2073
2074         index = map->m_lblk >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2075         end = (map->m_lblk + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2076         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2077
2078         pagevec_init(&pvec, 0);
2079
2080         while (index <= end) {
2081                 /* XXX: optimize tail */
2082                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2083                 if (nr_pages == 0)
2084                         break;
2085                 for (i = 0; i < nr_pages; i++) {
2086                         struct page *page = pvec.pages[i];
2087
2088                         index = page->index;
2089                         if (index > end)
2090                                 break;
2091                         index++;
2092
2093                         BUG_ON(!PageLocked(page));
2094                         BUG_ON(PageWriteback(page));
2095                         BUG_ON(!page_has_buffers(page));
2096
2097                         bh = page_buffers(page);
2098                         head = bh;
2099
2100                         /* skip blocks out of the range */
2101                         do {
2102                                 if (cur_logical >= map->m_lblk)
2103                                         break;
2104                                 cur_logical++;
2105                         } while ((bh = bh->b_this_page) != head);
2106
2107                         do {
2108                                 if (cur_logical >= map->m_lblk + blocks)
2109                                         break;
2110
2111                                 if (buffer_delay(bh) || buffer_unwritten(bh)) {
2112
2113                                         BUG_ON(bh->b_bdev != inode->i_sb->s_bdev);
2114
2115                                         if (buffer_delay(bh)) {
2116                                                 clear_buffer_delay(bh);
2117                                                 bh->b_blocknr = pblock;
2118                                         } else {
2119                                                 /*
2120                                                  * unwritten already should have
2121                                                  * blocknr assigned. Verify that
2122                                                  */
2123                                                 clear_buffer_unwritten(bh);
2124                                                 BUG_ON(bh->b_blocknr != pblock);
2125                                         }
2126
2127                                 } else if (buffer_mapped(bh))
2128                                         BUG_ON(bh->b_blocknr != pblock);
2129
2130                                 if (map->m_flags & EXT4_MAP_UNINIT)
2131                                         set_buffer_uninit(bh);
2132                                 cur_logical++;
2133                                 pblock++;
2134                         } while ((bh = bh->b_this_page) != head);
2135                 }
2136                 pagevec_release(&pvec);
2137         }
2138 }
2139
2140
2141 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2142                                         sector_t logical, long blk_cnt)
2143 {
2144         int nr_pages, i;
2145         pgoff_t index, end;
2146         struct pagevec pvec;
2147         struct inode *inode = mpd->inode;
2148         struct address_space *mapping = inode->i_mapping;
2149
2150         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2151         end   = (logical + blk_cnt - 1) >>
2152                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2153         while (index <= end) {
2154                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2155                 if (nr_pages == 0)
2156                         break;
2157                 for (i = 0; i < nr_pages; i++) {
2158                         struct page *page = pvec.pages[i];
2159                         if (page->index > end)
2160                                 break;
2161                         BUG_ON(!PageLocked(page));
2162                         BUG_ON(PageWriteback(page));
2163                         block_invalidatepage(page, 0);
2164                         ClearPageUptodate(page);
2165                         unlock_page(page);
2166                 }
2167                 index = pvec.pages[nr_pages - 1]->index + 1;
2168                 pagevec_release(&pvec);
2169         }
2170         return;
2171 }
2172
2173 static void ext4_print_free_blocks(struct inode *inode)
2174 {
2175         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2176         printk(KERN_CRIT "Total free blocks count %lld\n",
2177                ext4_count_free_blocks(inode->i_sb));
2178         printk(KERN_CRIT "Free/Dirty block details\n");
2179         printk(KERN_CRIT "free_blocks=%lld\n",
2180                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2181         printk(KERN_CRIT "dirty_blocks=%lld\n",
2182                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2183         printk(KERN_CRIT "Block reservation details\n");
2184         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2185                EXT4_I(inode)->i_reserved_data_blocks);
2186         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2187                EXT4_I(inode)->i_reserved_meta_blocks);
2188         return;
2189 }
2190
2191 /*
2192  * mpage_da_map_blocks - go through given space
2193  *
2194  * @mpd - bh describing space
2195  *
2196  * The function skips space we know is already mapped to disk blocks.
2197  *
2198  */
2199 static int mpage_da_map_blocks(struct mpage_da_data *mpd)
2200 {
2201         int err, blks, get_blocks_flags;
2202         struct ext4_map_blocks map;
2203         sector_t next = mpd->b_blocknr;
2204         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2205         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2206         handle_t *handle = NULL;
2207
2208         /*
2209          * We consider only non-mapped and non-allocated blocks
2210          */
2211         if ((mpd->b_state  & (1 << BH_Mapped)) &&
2212                 !(mpd->b_state & (1 << BH_Delay)) &&
2213                 !(mpd->b_state & (1 << BH_Unwritten)))
2214                 return 0;
2215
2216         /*
2217          * If we didn't accumulate anything to write simply return
2218          */
2219         if (!mpd->b_size)
2220                 return 0;
2221
2222         handle = ext4_journal_current_handle();
2223         BUG_ON(!handle);
2224
2225         /*
2226          * Call ext4_map_blocks() to allocate any delayed allocation
2227          * blocks, or to convert an uninitialized extent to be
2228          * initialized (in the case where we have written into
2229          * one or more preallocated blocks).
2230          *
2231          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2232          * indicate that we are on the delayed allocation path.  This
2233          * affects functions in many different parts of the allocation
2234          * call path.  This flag exists primarily because we don't
2235          * want to change *many* call functions, so ext4_map_blocks()
2236          * will set the magic i_delalloc_reserved_flag once the
2237          * inode's allocation semaphore is taken.
2238          *
2239          * If the blocks in questions were delalloc blocks, set
2240          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2241          * variables are updated after the blocks have been allocated.
2242          */
2243         map.m_lblk = next;
2244         map.m_len = max_blocks;
2245         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
2246         if (ext4_should_dioread_nolock(mpd->inode))
2247                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2248         if (mpd->b_state & (1 << BH_Delay))
2249                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2250
2251         blks = ext4_map_blocks(handle, mpd->inode, &map, get_blocks_flags);
2252         if (blks < 0) {
2253                 struct super_block *sb = mpd->inode->i_sb;
2254
2255                 err = blks;
2256                 /*
2257                  * If get block returns with error we simply
2258                  * return. Later writepage will redirty the page and
2259                  * writepages will find the dirty page again
2260                  */
2261                 if (err == -EAGAIN)
2262                         return 0;
2263
2264                 if (err == -ENOSPC &&
2265                     ext4_count_free_blocks(sb)) {
2266                         mpd->retval = err;
2267                         return 0;
2268                 }
2269
2270                 /*
2271                  * get block failure will cause us to loop in
2272                  * writepages, because a_ops->writepage won't be able
2273                  * to make progress. The page will be redirtied by
2274                  * writepage and writepages will again try to write
2275                  * the same.
2276                  */
2277                 if (!(EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2278                         ext4_msg(sb, KERN_CRIT,
2279                                  "delayed block allocation failed for inode %lu "
2280                                  "at logical offset %llu with max blocks %zd "
2281                                  "with error %d", mpd->inode->i_ino,
2282                                  (unsigned long long) next,
2283                                  mpd->b_size >> mpd->inode->i_blkbits, err);
2284                         ext4_msg(sb, KERN_CRIT,
2285                                 "This should not happen!! Data will be lost\n");
2286                         if (err == -ENOSPC)
2287                                 ext4_print_free_blocks(mpd->inode);
2288                 }
2289                 /* invalidate all the pages */
2290                 ext4_da_block_invalidatepages(mpd, next,
2291                                 mpd->b_size >> mpd->inode->i_blkbits);
2292                 return err;
2293         }
2294         BUG_ON(blks == 0);
2295
2296         if (map.m_flags & EXT4_MAP_NEW) {
2297                 struct block_device *bdev = mpd->inode->i_sb->s_bdev;
2298                 int i;
2299
2300                 for (i = 0; i < map.m_len; i++)
2301                         unmap_underlying_metadata(bdev, map.m_pblk + i);
2302         }
2303
2304         /*
2305          * If blocks are delayed marked, we need to
2306          * put actual blocknr and drop delayed bit
2307          */
2308         if ((mpd->b_state & (1 << BH_Delay)) ||
2309             (mpd->b_state & (1 << BH_Unwritten)))
2310                 mpage_put_bnr_to_bhs(mpd, &map);
2311
2312         if (ext4_should_order_data(mpd->inode)) {
2313                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2314                 if (err)
2315                         return err;
2316         }
2317
2318         /*
2319          * Update on-disk size along with block allocation.
2320          */
2321         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2322         if (disksize > i_size_read(mpd->inode))
2323                 disksize = i_size_read(mpd->inode);
2324         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2325                 ext4_update_i_disksize(mpd->inode, disksize);
2326                 return ext4_mark_inode_dirty(handle, mpd->inode);
2327         }
2328
2329         return 0;
2330 }
2331
2332 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2333                 (1 << BH_Delay) | (1 << BH_Unwritten))
2334
2335 /*
2336  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2337  *
2338  * @mpd->lbh - extent of blocks
2339  * @logical - logical number of the block in the file
2340  * @bh - bh of the block (used to access block's state)
2341  *
2342  * the function is used to collect contig. blocks in same state
2343  */
2344 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2345                                    sector_t logical, size_t b_size,
2346                                    unsigned long b_state)
2347 {
2348         sector_t next;
2349         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2350
2351         /*
2352          * XXX Don't go larger than mballoc is willing to allocate
2353          * This is a stopgap solution.  We eventually need to fold
2354          * mpage_da_submit_io() into this function and then call
2355          * ext4_map_blocks() multiple times in a loop
2356          */
2357         if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
2358                 goto flush_it;
2359
2360         /* check if thereserved journal credits might overflow */
2361         if (!(ext4_test_inode_flag(mpd->inode, EXT4_INODE_EXTENTS))) {
2362                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2363                         /*
2364                          * With non-extent format we are limited by the journal
2365                          * credit available.  Total credit needed to insert
2366                          * nrblocks contiguous blocks is dependent on the
2367                          * nrblocks.  So limit nrblocks.
2368                          */
2369                         goto flush_it;
2370                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2371                                 EXT4_MAX_TRANS_DATA) {
2372                         /*
2373                          * Adding the new buffer_head would make it cross the
2374                          * allowed limit for which we have journal credit
2375                          * reserved. So limit the new bh->b_size
2376                          */
2377                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2378                                                 mpd->inode->i_blkbits;
2379                         /* we will do mpage_da_submit_io in the next loop */
2380                 }
2381         }
2382         /*
2383          * First block in the extent
2384          */
2385         if (mpd->b_size == 0) {
2386                 mpd->b_blocknr = logical;
2387                 mpd->b_size = b_size;
2388                 mpd->b_state = b_state & BH_FLAGS;
2389                 return;
2390         }
2391
2392         next = mpd->b_blocknr + nrblocks;
2393         /*
2394          * Can we merge the block to our big extent?
2395          */
2396         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2397                 mpd->b_size += b_size;
2398                 return;
2399         }
2400
2401 flush_it:
2402         /*
2403          * We couldn't merge the block to our extent, so we
2404          * need to flush current  extent and start new one
2405          */
2406         if (mpage_da_map_blocks(mpd) == 0)
2407                 mpage_da_submit_io(mpd);
2408         mpd->io_done = 1;
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, void *data)
2428 {
2429         struct mpage_da_data *mpd = data;
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 using writepage()
2441                  */
2442                 if (mpd->next_page != mpd->first_page) {
2443                         if (mpage_da_map_blocks(mpd) == 0)
2444                                 mpage_da_submit_io(mpd);
2445                         /*
2446                          * skip rest of the page in the page_vec
2447                          */
2448                         mpd->io_done = 1;
2449                         redirty_page_for_writepage(wbc, page);
2450                         unlock_page(page);
2451                         return MPAGE_DA_EXTENT_TAIL;
2452                 }
2453
2454                 /*
2455                  * Start next extent of pages ...
2456                  */
2457                 mpd->first_page = page->index;
2458
2459                 /*
2460                  * ... and blocks
2461                  */
2462                 mpd->b_size = 0;
2463                 mpd->b_state = 0;
2464                 mpd->b_blocknr = 0;
2465         }
2466
2467         mpd->next_page = page->index + 1;
2468         logical = (sector_t) page->index <<
2469                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2470
2471         if (!page_has_buffers(page)) {
2472                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2473                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2474                 if (mpd->io_done)
2475                         return MPAGE_DA_EXTENT_TAIL;
2476         } else {
2477                 /*
2478                  * Page with regular buffer heads, just add all dirty ones
2479                  */
2480                 head = page_buffers(page);
2481                 bh = head;
2482                 do {
2483                         BUG_ON(buffer_locked(bh));
2484                         /*
2485                          * We need to try to allocate
2486                          * unmapped blocks in the same page.
2487                          * Otherwise we won't make progress
2488                          * with the page in ext4_writepage
2489                          */
2490                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2491                                 mpage_add_bh_to_extent(mpd, logical,
2492                                                        bh->b_size,
2493                                                        bh->b_state);
2494                                 if (mpd->io_done)
2495                                         return MPAGE_DA_EXTENT_TAIL;
2496                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2497                                 /*
2498                                  * mapped dirty buffer. We need to update
2499                                  * the b_state because we look at
2500                                  * b_state in mpage_da_map_blocks. We don't
2501                                  * update b_size because if we find an
2502                                  * unmapped buffer_head later we need to
2503                                  * use the b_state flag of that buffer_head.
2504                                  */
2505                                 if (mpd->b_size == 0)
2506                                         mpd->b_state = bh->b_state & BH_FLAGS;
2507                         }
2508                         logical++;
2509                 } while ((bh = bh->b_this_page) != head);
2510         }
2511
2512         return 0;
2513 }
2514
2515 /*
2516  * This is a special get_blocks_t callback which is used by
2517  * ext4_da_write_begin().  It will either return mapped block or
2518  * reserve space for a single block.
2519  *
2520  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2521  * We also have b_blocknr = -1 and b_bdev initialized properly
2522  *
2523  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2524  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2525  * initialized properly.
2526  */
2527 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2528                                   struct buffer_head *bh, int create)
2529 {
2530         struct ext4_map_blocks map;
2531         int ret = 0;
2532         sector_t invalid_block = ~((sector_t) 0xffff);
2533
2534         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2535                 invalid_block = ~0;
2536
2537         BUG_ON(create == 0);
2538         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
2539
2540         map.m_lblk = iblock;
2541         map.m_len = 1;
2542
2543         /*
2544          * first, we need to know whether the block is allocated already
2545          * preallocated blocks are unmapped but should treated
2546          * the same as allocated blocks.
2547          */
2548         ret = ext4_map_blocks(NULL, inode, &map, 0);
2549         if (ret < 0)
2550                 return ret;
2551         if (ret == 0) {
2552                 if (buffer_delay(bh))
2553                         return 0; /* Not sure this could or should happen */
2554                 /*
2555                  * XXX: __block_prepare_write() unmaps passed block,
2556                  * is it OK?
2557                  */
2558                 ret = ext4_da_reserve_space(inode, iblock);
2559                 if (ret)
2560                         /* not enough space to reserve */
2561                         return ret;
2562
2563                 map_bh(bh, inode->i_sb, invalid_block);
2564                 set_buffer_new(bh);
2565                 set_buffer_delay(bh);
2566                 return 0;
2567         }
2568
2569         map_bh(bh, inode->i_sb, map.m_pblk);
2570         bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
2571
2572         if (buffer_unwritten(bh)) {
2573                 /* A delayed write to unwritten bh should be marked
2574                  * new and mapped.  Mapped ensures that we don't do
2575                  * get_block multiple times when we write to the same
2576                  * offset and new ensures that we do proper zero out
2577                  * for partial write.
2578                  */
2579                 set_buffer_new(bh);
2580                 set_buffer_mapped(bh);
2581         }
2582         return 0;
2583 }
2584
2585 /*
2586  * This function is used as a standard get_block_t calback function
2587  * when there is no desire to allocate any blocks.  It is used as a
2588  * callback function for block_prepare_write() and block_write_full_page().
2589  * These functions should only try to map a single block at a time.
2590  *
2591  * Since this function doesn't do block allocations even if the caller
2592  * requests it by passing in create=1, it is critically important that
2593  * any caller checks to make sure that any buffer heads are returned
2594  * by this function are either all already mapped or marked for
2595  * delayed allocation before calling  block_write_full_page().  Otherwise,
2596  * b_blocknr could be left unitialized, and the page write functions will
2597  * be taken by surprise.
2598  */
2599 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2600                                    struct buffer_head *bh_result, int create)
2601 {
2602         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2603         return _ext4_get_block(inode, iblock, bh_result, 0);
2604 }
2605
2606 static int bget_one(handle_t *handle, struct buffer_head *bh)
2607 {
2608         get_bh(bh);
2609         return 0;
2610 }
2611
2612 static int bput_one(handle_t *handle, struct buffer_head *bh)
2613 {
2614         put_bh(bh);
2615         return 0;
2616 }
2617
2618 static int __ext4_journalled_writepage(struct page *page,
2619                                        unsigned int len)
2620 {
2621         struct address_space *mapping = page->mapping;
2622         struct inode *inode = mapping->host;
2623         struct buffer_head *page_bufs;
2624         handle_t *handle = NULL;
2625         int ret = 0;
2626         int err;
2627
2628         page_bufs = page_buffers(page);
2629         BUG_ON(!page_bufs);
2630         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2631         /* As soon as we unlock the page, it can go away, but we have
2632          * references to buffers so we are safe */
2633         unlock_page(page);
2634
2635         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2636         if (IS_ERR(handle)) {
2637                 ret = PTR_ERR(handle);
2638                 goto out;
2639         }
2640
2641         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2642                                 do_journal_get_write_access);
2643
2644         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2645                                 write_end_fn);
2646         if (ret == 0)
2647                 ret = err;
2648         err = ext4_journal_stop(handle);
2649         if (!ret)
2650                 ret = err;
2651
2652         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2653         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2654 out:
2655         return ret;
2656 }
2657
2658 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
2659 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
2660
2661 /*
2662  * Note that we don't need to start a transaction unless we're journaling data
2663  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2664  * need to file the inode to the transaction's list in ordered mode because if
2665  * we are writing back data added by write(), the inode is already there and if
2666  * we are writing back data modified via mmap(), noone guarantees in which
2667  * transaction the data will hit the disk. In case we are journaling data, we
2668  * cannot start transaction directly because transaction start ranks above page
2669  * lock so we have to do some magic.
2670  *
2671  * This function can get called via...
2672  *   - ext4_da_writepages after taking page lock (have journal handle)
2673  *   - journal_submit_inode_data_buffers (no journal handle)
2674  *   - shrink_page_list via pdflush (no journal handle)
2675  *   - grab_page_cache when doing write_begin (have journal handle)
2676  *
2677  * We don't do any block allocation in this function. If we have page with
2678  * multiple blocks we need to write those buffer_heads that are mapped. This
2679  * is important for mmaped based write. So if we do with blocksize 1K
2680  * truncate(f, 1024);
2681  * a = mmap(f, 0, 4096);
2682  * a[0] = 'a';
2683  * truncate(f, 4096);
2684  * we have in the page first buffer_head mapped via page_mkwrite call back
2685  * but other bufer_heads would be unmapped but dirty(dirty done via the
2686  * do_wp_page). So writepage should write the first block. If we modify
2687  * the mmap area beyond 1024 we will again get a page_fault and the
2688  * page_mkwrite callback will do the block allocation and mark the
2689  * buffer_heads mapped.
2690  *
2691  * We redirty the page if we have any buffer_heads that is either delay or
2692  * unwritten in the page.
2693  *
2694  * We can get recursively called as show below.
2695  *
2696  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2697  *              ext4_writepage()
2698  *
2699  * But since we don't do any block allocation we should not deadlock.
2700  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2701  */
2702 static int ext4_writepage(struct page *page,
2703                           struct writeback_control *wbc)
2704 {
2705         int ret = 0;
2706         loff_t size;
2707         unsigned int len;
2708         struct buffer_head *page_bufs = NULL;
2709         struct inode *inode = page->mapping->host;
2710
2711         trace_ext4_writepage(inode, page);
2712         size = i_size_read(inode);
2713         if (page->index == size >> PAGE_CACHE_SHIFT)
2714                 len = size & ~PAGE_CACHE_MASK;
2715         else
2716                 len = PAGE_CACHE_SIZE;
2717
2718         if (page_has_buffers(page)) {
2719                 page_bufs = page_buffers(page);
2720                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2721                                         ext4_bh_delay_or_unwritten)) {
2722                         /*
2723                          * We don't want to do  block allocation
2724                          * So redirty the page and return
2725                          * We may reach here when we do a journal commit
2726                          * via journal_submit_inode_data_buffers.
2727                          * If we don't have mapping block we just ignore
2728                          * them. We can also reach here via shrink_page_list
2729                          */
2730                         redirty_page_for_writepage(wbc, page);
2731                         unlock_page(page);
2732                         return 0;
2733                 }
2734         } else {
2735                 /*
2736                  * The test for page_has_buffers() is subtle:
2737                  * We know the page is dirty but it lost buffers. That means
2738                  * that at some moment in time after write_begin()/write_end()
2739                  * has been called all buffers have been clean and thus they
2740                  * must have been written at least once. So they are all
2741                  * mapped and we can happily proceed with mapping them
2742                  * and writing the page.
2743                  *
2744                  * Try to initialize the buffer_heads and check whether
2745                  * all are mapped and non delay. We don't want to
2746                  * do block allocation here.
2747                  */
2748                 ret = block_prepare_write(page, 0, len,
2749                                           noalloc_get_block_write);
2750                 if (!ret) {
2751                         page_bufs = page_buffers(page);
2752                         /* check whether all are mapped and non delay */
2753                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2754                                                 ext4_bh_delay_or_unwritten)) {
2755                                 redirty_page_for_writepage(wbc, page);
2756                                 unlock_page(page);
2757                                 return 0;
2758                         }
2759                 } else {
2760                         /*
2761                          * We can't do block allocation here
2762                          * so just redity the page and unlock
2763                          * and return
2764                          */
2765                         redirty_page_for_writepage(wbc, page);
2766                         unlock_page(page);
2767                         return 0;
2768                 }
2769                 /* now mark the buffer_heads as dirty and uptodate */
2770                 block_commit_write(page, 0, len);
2771         }
2772
2773         if (PageChecked(page) && ext4_should_journal_data(inode)) {
2774                 /*
2775                  * It's mmapped pagecache.  Add buffers and journal it.  There
2776                  * doesn't seem much point in redirtying the page here.
2777                  */
2778                 ClearPageChecked(page);
2779                 return __ext4_journalled_writepage(page, len);
2780         }
2781
2782         if (page_bufs && buffer_uninit(page_bufs)) {
2783                 ext4_set_bh_endio(page_bufs, inode);
2784                 ret = block_write_full_page_endio(page, noalloc_get_block_write,
2785                                             wbc, ext4_end_io_buffer_write);
2786         } else
2787                 ret = block_write_full_page(page, noalloc_get_block_write,
2788                                             wbc);
2789
2790         return ret;
2791 }
2792
2793 /*
2794  * This is called via ext4_da_writepages() to
2795  * calulate the total number of credits to reserve to fit
2796  * a single extent allocation into a single transaction,
2797  * ext4_da_writpeages() will loop calling this before
2798  * the block allocation.
2799  */
2800
2801 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2802 {
2803         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2804
2805         /*
2806          * With non-extent format the journal credit needed to
2807          * insert nrblocks contiguous block is dependent on
2808          * number of contiguous block. So we will limit
2809          * number of contiguous block to a sane value
2810          */
2811         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) &&
2812             (max_blocks > EXT4_MAX_TRANS_DATA))
2813                 max_blocks = EXT4_MAX_TRANS_DATA;
2814
2815         return ext4_chunk_trans_blocks(inode, max_blocks);
2816 }
2817
2818 /*
2819  * write_cache_pages_da - walk the list of dirty pages of the given
2820  * address space and call the callback function (which usually writes
2821  * the pages).
2822  *
2823  * This is a forked version of write_cache_pages().  Differences:
2824  *      Range cyclic is ignored.
2825  *      no_nrwrite_index_update is always presumed true
2826  */
2827 static int write_cache_pages_da(struct address_space *mapping,
2828                                 struct writeback_control *wbc,
2829                                 struct mpage_da_data *mpd)
2830 {
2831         int ret = 0;
2832         int done = 0;
2833         struct pagevec pvec;
2834         int nr_pages;
2835         pgoff_t index;
2836         pgoff_t end;            /* Inclusive */
2837         long nr_to_write = wbc->nr_to_write;
2838
2839         pagevec_init(&pvec, 0);
2840         index = wbc->range_start >> PAGE_CACHE_SHIFT;
2841         end = wbc->range_end >> PAGE_CACHE_SHIFT;
2842
2843         while (!done && (index <= end)) {
2844                 int i;
2845
2846                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2847                               PAGECACHE_TAG_DIRTY,
2848                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2849                 if (nr_pages == 0)
2850                         break;
2851
2852                 for (i = 0; i < nr_pages; i++) {
2853                         struct page *page = pvec.pages[i];
2854
2855                         /*
2856                          * At this point, the page may be truncated or
2857                          * invalidated (changing page->mapping to NULL), or
2858                          * even swizzled back from swapper_space to tmpfs file
2859                          * mapping. However, page->index will not change
2860                          * because we have a reference on the page.
2861                          */
2862                         if (page->index > end) {
2863                                 done = 1;
2864                                 break;
2865                         }
2866
2867                         lock_page(page);
2868
2869                         /*
2870                          * Page truncated or invalidated. We can freely skip it
2871                          * then, even for data integrity operations: the page
2872                          * has disappeared concurrently, so there could be no
2873                          * real expectation of this data interity operation
2874                          * even if there is now a new, dirty page at the same
2875                          * pagecache address.
2876                          */
2877                         if (unlikely(page->mapping != mapping)) {
2878 continue_unlock:
2879                                 unlock_page(page);
2880                                 continue;
2881                         }
2882
2883                         if (!PageDirty(page)) {
2884                                 /* someone wrote it for us */
2885                                 goto continue_unlock;
2886                         }
2887
2888                         if (PageWriteback(page)) {
2889                                 if (wbc->sync_mode != WB_SYNC_NONE)
2890                                         wait_on_page_writeback(page);
2891                                 else
2892                                         goto continue_unlock;
2893                         }
2894
2895                         BUG_ON(PageWriteback(page));
2896                         if (!clear_page_dirty_for_io(page))
2897                                 goto continue_unlock;
2898
2899                         ret = __mpage_da_writepage(page, wbc, mpd);
2900                         if (unlikely(ret)) {
2901                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2902                                         unlock_page(page);
2903                                         ret = 0;
2904                                 } else {
2905                                         done = 1;
2906                                         break;
2907                                 }
2908                         }
2909
2910                         if (nr_to_write > 0) {
2911                                 nr_to_write--;
2912                                 if (nr_to_write == 0 &&
2913                                     wbc->sync_mode == WB_SYNC_NONE) {
2914                                         /*
2915                                          * We stop writing back only if we are
2916                                          * not doing integrity sync. In case of
2917                                          * integrity sync we have to keep going
2918                                          * because someone may be concurrently
2919                                          * dirtying pages, and we might have
2920                                          * synced a lot of newly appeared dirty
2921                                          * pages, but have not synced all of the
2922                                          * old dirty pages.
2923                                          */
2924                                         done = 1;
2925                                         break;
2926                                 }
2927                         }
2928                 }
2929                 pagevec_release(&pvec);
2930                 cond_resched();
2931         }
2932         return ret;
2933 }
2934
2935
2936 static int ext4_da_writepages(struct address_space *mapping,
2937                               struct writeback_control *wbc)
2938 {
2939         pgoff_t index;
2940         int range_whole = 0;
2941         handle_t *handle = NULL;
2942         struct mpage_da_data mpd;
2943         struct inode *inode = mapping->host;
2944         int pages_written = 0;
2945         long pages_skipped;
2946         unsigned int max_pages;
2947         int range_cyclic, cycled = 1, io_done = 0;
2948         int needed_blocks, ret = 0;
2949         long desired_nr_to_write, nr_to_writebump = 0;
2950         loff_t range_start = wbc->range_start;
2951         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2952
2953         trace_ext4_da_writepages(inode, wbc);
2954
2955         /*
2956          * No pages to write? This is mainly a kludge to avoid starting
2957          * a transaction for special inodes like journal inode on last iput()
2958          * because that could violate lock ordering on umount
2959          */
2960         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2961                 return 0;
2962
2963         /*
2964          * If the filesystem has aborted, it is read-only, so return
2965          * right away instead of dumping stack traces later on that
2966          * will obscure the real source of the problem.  We test
2967          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2968          * the latter could be true if the filesystem is mounted
2969          * read-only, and in that case, ext4_da_writepages should
2970          * *never* be called, so if that ever happens, we would want
2971          * the stack trace.
2972          */
2973         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2974                 return -EROFS;
2975
2976         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2977                 range_whole = 1;
2978
2979         range_cyclic = wbc->range_cyclic;
2980         if (wbc->range_cyclic) {
2981                 index = mapping->writeback_index;
2982                 if (index)
2983                         cycled = 0;
2984                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2985                 wbc->range_end  = LLONG_MAX;
2986                 wbc->range_cyclic = 0;
2987         } else
2988                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2989
2990         /*
2991          * This works around two forms of stupidity.  The first is in
2992          * the writeback code, which caps the maximum number of pages
2993          * written to be 1024 pages.  This is wrong on multiple
2994          * levels; different architectues have a different page size,
2995          * which changes the maximum amount of data which gets
2996          * written.  Secondly, 4 megabytes is way too small.  XFS
2997          * forces this value to be 16 megabytes by multiplying
2998          * nr_to_write parameter by four, and then relies on its
2999          * allocator to allocate larger extents to make them
3000          * contiguous.  Unfortunately this brings us to the second
3001          * stupidity, which is that ext4's mballoc code only allocates
3002          * at most 2048 blocks.  So we force contiguous writes up to
3003          * the number of dirty blocks in the inode, or
3004          * sbi->max_writeback_mb_bump whichever is smaller.
3005          */
3006         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
3007         if (!range_cyclic && range_whole)
3008                 desired_nr_to_write = wbc->nr_to_write * 8;
3009         else
3010                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
3011                                                            max_pages);
3012         if (desired_nr_to_write > max_pages)
3013                 desired_nr_to_write = max_pages;
3014
3015         if (wbc->nr_to_write < desired_nr_to_write) {
3016                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
3017                 wbc->nr_to_write = desired_nr_to_write;
3018         }
3019
3020         mpd.wbc = wbc;
3021         mpd.inode = mapping->host;
3022
3023         pages_skipped = wbc->pages_skipped;
3024
3025 retry:
3026         while (!ret && wbc->nr_to_write > 0) {
3027
3028                 /*
3029                  * we  insert one extent at a time. So we need
3030                  * credit needed for single extent allocation.
3031                  * journalled mode is currently not supported
3032                  * by delalloc
3033                  */
3034                 BUG_ON(ext4_should_journal_data(inode));
3035                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
3036
3037                 /* start a new transaction*/
3038                 handle = ext4_journal_start(inode, needed_blocks);
3039                 if (IS_ERR(handle)) {
3040                         ret = PTR_ERR(handle);
3041                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
3042                                "%ld pages, ino %lu; err %d", __func__,
3043                                 wbc->nr_to_write, inode->i_ino, ret);
3044                         goto out_writepages;
3045                 }
3046
3047                 /*
3048                  * Now call __mpage_da_writepage to find the next
3049                  * contiguous region of logical blocks that need
3050                  * blocks to be allocated by ext4.  We don't actually
3051                  * submit the blocks for I/O here, even though
3052                  * write_cache_pages thinks it will, and will set the
3053                  * pages as clean for write before calling
3054                  * __mpage_da_writepage().
3055                  */
3056                 mpd.b_size = 0;
3057                 mpd.b_state = 0;
3058                 mpd.b_blocknr = 0;
3059                 mpd.first_page = 0;
3060                 mpd.next_page = 0;
3061                 mpd.io_done = 0;
3062                 mpd.pages_written = 0;
3063                 mpd.retval = 0;
3064                 ret = write_cache_pages_da(mapping, wbc, &mpd);
3065                 /*
3066                  * If we have a contiguous extent of pages and we
3067                  * haven't done the I/O yet, map the blocks and submit
3068                  * them for I/O.
3069                  */
3070                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
3071                         if (mpage_da_map_blocks(&mpd) == 0)
3072                                 mpage_da_submit_io(&mpd);
3073                         mpd.io_done = 1;
3074                         ret = MPAGE_DA_EXTENT_TAIL;
3075                 }
3076                 trace_ext4_da_write_pages(inode, &mpd);
3077                 wbc->nr_to_write -= mpd.pages_written;
3078
3079                 ext4_journal_stop(handle);
3080
3081                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
3082                         /* commit the transaction which would
3083                          * free blocks released in the transaction
3084                          * and try again
3085                          */
3086                         jbd2_journal_force_commit_nested(sbi->s_journal);
3087                         wbc->pages_skipped = pages_skipped;
3088                         ret = 0;
3089                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
3090                         /*
3091                          * got one extent now try with
3092                          * rest of the pages
3093                          */
3094                         pages_written += mpd.pages_written;
3095                         wbc->pages_skipped = pages_skipped;
3096                         ret = 0;
3097                         io_done = 1;
3098                 } else if (wbc->nr_to_write)
3099                         /*
3100                          * There is no more writeout needed
3101                          * or we requested for a noblocking writeout
3102                          * and we found the device congested
3103                          */
3104                         break;
3105         }
3106         if (!io_done && !cycled) {
3107                 cycled = 1;
3108                 index = 0;
3109                 wbc->range_start = index << PAGE_CACHE_SHIFT;
3110                 wbc->range_end  = mapping->writeback_index - 1;
3111                 goto retry;
3112         }
3113         if (pages_skipped != wbc->pages_skipped)
3114                 ext4_msg(inode->i_sb, KERN_CRIT,
3115                          "This should not happen leaving %s "
3116                          "with nr_to_write = %ld ret = %d",
3117                          __func__, wbc->nr_to_write, ret);
3118
3119         /* Update index */
3120         index += pages_written;
3121         wbc->range_cyclic = range_cyclic;
3122         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
3123                 /*
3124                  * set the writeback_index so that range_cyclic
3125                  * mode will write it back later
3126                  */
3127                 mapping->writeback_index = index;
3128
3129 out_writepages:
3130         wbc->nr_to_write -= nr_to_writebump;
3131         wbc->range_start = range_start;
3132         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
3133         return ret;
3134 }
3135
3136 #define FALL_BACK_TO_NONDELALLOC 1
3137 static int ext4_nonda_switch(struct super_block *sb)
3138 {
3139         s64 free_blocks, dirty_blocks;
3140         struct ext4_sb_info *sbi = EXT4_SB(sb);
3141
3142         /*
3143          * switch to non delalloc mode if we are running low
3144          * on free block. The free block accounting via percpu
3145          * counters can get slightly wrong with percpu_counter_batch getting
3146          * accumulated on each CPU without updating global counters
3147          * Delalloc need an accurate free block accounting. So switch
3148          * to non delalloc when we are near to error range.
3149          */
3150         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
3151         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
3152         if (2 * free_blocks < 3 * dirty_blocks ||
3153                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
3154                 /*
3155                  * free block count is less than 150% of dirty blocks
3156                  * or free blocks is less than watermark
3157                  */
3158                 return 1;
3159         }
3160         /*
3161          * Even if we don't switch but are nearing capacity,
3162          * start pushing delalloc when 1/2 of free blocks are dirty.
3163          */
3164         if (free_blocks < 2 * dirty_blocks)
3165                 writeback_inodes_sb_if_idle(sb);
3166
3167         return 0;
3168 }
3169
3170 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3171                                loff_t pos, unsigned len, unsigned flags,
3172                                struct page **pagep, void **fsdata)
3173 {
3174         int ret, retries = 0;
3175         struct page *page;
3176         pgoff_t index;
3177         struct inode *inode = mapping->host;
3178         handle_t *handle;
3179
3180         index = pos >> PAGE_CACHE_SHIFT;
3181
3182         if (ext4_nonda_switch(inode->i_sb)) {
3183                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3184                 return ext4_write_begin(file, mapping, pos,
3185                                         len, flags, pagep, fsdata);
3186         }
3187         *fsdata = (void *)0;
3188         trace_ext4_da_write_begin(inode, pos, len, flags);
3189 retry:
3190         /*
3191          * With delayed allocation, we don't log the i_disksize update
3192          * if there is delayed block allocation. But we still need
3193          * to journalling the i_disksize update if writes to the end
3194          * of file which has an already mapped buffer.
3195          */
3196         handle = ext4_journal_start(inode, 1);
3197         if (IS_ERR(handle)) {
3198                 ret = PTR_ERR(handle);
3199                 goto out;
3200         }
3201         /* We cannot recurse into the filesystem as the transaction is already
3202          * started */
3203         flags |= AOP_FLAG_NOFS;
3204
3205         page = grab_cache_page_write_begin(mapping, index, flags);
3206         if (!page) {
3207                 ext4_journal_stop(handle);
3208                 ret = -ENOMEM;
3209                 goto out;
3210         }
3211         *pagep = page;
3212
3213         ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
3214         if (ret < 0) {
3215                 unlock_page(page);
3216                 ext4_journal_stop(handle);
3217                 page_cache_release(page);
3218                 /*
3219                  * block_write_begin may have instantiated a few blocks
3220                  * outside i_size.  Trim these off again. Don't need
3221                  * i_size_read because we hold i_mutex.
3222                  */
3223                 if (pos + len > inode->i_size)
3224                         ext4_truncate_failed_write(inode);
3225         }
3226
3227         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3228                 goto retry;
3229 out:
3230         return ret;
3231 }
3232
3233 /*
3234  * Check if we should update i_disksize
3235  * when write to the end of file but not require block allocation
3236  */
3237 static int ext4_da_should_update_i_disksize(struct page *page,
3238                                             unsigned long offset)
3239 {
3240         struct buffer_head *bh;
3241         struct inode *inode = page->mapping->host;
3242         unsigned int idx;
3243         int i;
3244
3245         bh = page_buffers(page);
3246         idx = offset >> inode->i_blkbits;
3247
3248         for (i = 0; i < idx; i++)
3249                 bh = bh->b_this_page;
3250
3251         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3252                 return 0;
3253         return 1;
3254 }
3255
3256 static int ext4_da_write_end(struct file *file,
3257                              struct address_space *mapping,
3258                              loff_t pos, unsigned len, unsigned copied,
3259                              struct page *page, void *fsdata)
3260 {
3261         struct inode *inode = mapping->host;
3262         int ret = 0, ret2;
3263         handle_t *handle = ext4_journal_current_handle();
3264         loff_t new_i_size;
3265         unsigned long start, end;
3266         int write_mode = (int)(unsigned long)fsdata;
3267
3268         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
3269                 if (ext4_should_order_data(inode)) {
3270                         return ext4_ordered_write_end(file, mapping, pos,
3271                                         len, copied, page, fsdata);
3272                 } else if (ext4_should_writeback_data(inode)) {
3273                         return ext4_writeback_write_end(file, mapping, pos,
3274                                         len, copied, page, fsdata);
3275                 } else {
3276                         BUG();
3277                 }
3278         }
3279
3280         trace_ext4_da_write_end(inode, pos, len, copied);
3281         start = pos & (PAGE_CACHE_SIZE - 1);
3282         end = start + copied - 1;
3283
3284         /*
3285          * generic_write_end() will run mark_inode_dirty() if i_size
3286          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3287          * into that.
3288          */
3289
3290         new_i_size = pos + copied;
3291         if (new_i_size > EXT4_I(inode)->i_disksize) {
3292                 if (ext4_da_should_update_i_disksize(page, end)) {
3293                         down_write(&EXT4_I(inode)->i_data_sem);
3294                         if (new_i_size > EXT4_I(inode)->i_disksize) {
3295                                 /*
3296                                  * Updating i_disksize when extending file
3297                                  * without needing block allocation
3298                                  */
3299                                 if (ext4_should_order_data(inode))
3300                                         ret = ext4_jbd2_file_inode(handle,
3301                                                                    inode);
3302
3303                                 EXT4_I(inode)->i_disksize = new_i_size;
3304                         }
3305                         up_write(&EXT4_I(inode)->i_data_sem);
3306                         /* We need to mark inode dirty even if
3307                          * new_i_size is less that inode->i_size
3308                          * bu greater than i_disksize.(hint delalloc)
3309                          */
3310                         ext4_mark_inode_dirty(handle, inode);
3311                 }
3312         }
3313         ret2 = generic_write_end(file, mapping, pos, len, copied,
3314                                                         page, fsdata);
3315         copied = ret2;
3316         if (ret2 < 0)
3317                 ret = ret2;
3318         ret2 = ext4_journal_stop(handle);
3319         if (!ret)
3320                 ret = ret2;
3321
3322         return ret ? ret : copied;
3323 }
3324
3325 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
3326 {
3327         /*
3328          * Drop reserved blocks
3329          */
3330         BUG_ON(!PageLocked(page));
3331         if (!page_has_buffers(page))
3332                 goto out;
3333
3334         ext4_da_page_release_reservation(page, offset);
3335
3336 out:
3337         ext4_invalidatepage(page, offset);
3338
3339         return;
3340 }
3341
3342 /*
3343  * Force all delayed allocation blocks to be allocated for a given inode.
3344  */
3345 int ext4_alloc_da_blocks(struct inode *inode)
3346 {
3347         trace_ext4_alloc_da_blocks(inode);
3348
3349         if (!EXT4_I(inode)->i_reserved_data_blocks &&
3350             !EXT4_I(inode)->i_reserved_meta_blocks)
3351                 return 0;
3352
3353         /*
3354          * We do something simple for now.  The filemap_flush() will
3355          * also start triggering a write of the data blocks, which is
3356          * not strictly speaking necessary (and for users of
3357          * laptop_mode, not even desirable).  However, to do otherwise
3358          * would require replicating code paths in:
3359          *
3360          * ext4_da_writepages() ->
3361          *    write_cache_pages() ---> (via passed in callback function)
3362          *        __mpage_da_writepage() -->
3363          *           mpage_add_bh_to_extent()
3364          *           mpage_da_map_blocks()
3365          *
3366          * The problem is that write_cache_pages(), located in
3367          * mm/page-writeback.c, marks pages clean in preparation for
3368          * doing I/O, which is not desirable if we're not planning on
3369          * doing I/O at all.
3370          *
3371          * We could call write_cache_pages(), and then redirty all of
3372          * the pages by calling redirty_page_for_writeback() but that
3373          * would be ugly in the extreme.  So instead we would need to
3374          * replicate parts of the code in the above functions,
3375          * simplifying them becuase we wouldn't actually intend to
3376          * write out the pages, but rather only collect contiguous
3377          * logical block extents, call the multi-block allocator, and
3378          * then update the buffer heads with the block allocations.
3379          *
3380          * For now, though, we'll cheat by calling filemap_flush(),
3381          * which will map the blocks, and start the I/O, but not
3382          * actually wait for the I/O to complete.
3383          */
3384         return filemap_flush(inode->i_mapping);
3385 }
3386
3387 /*
3388  * bmap() is special.  It gets used by applications such as lilo and by
3389  * the swapper to find the on-disk block of a specific piece of data.
3390  *
3391  * Naturally, this is dangerous if the block concerned is still in the
3392  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3393  * filesystem and enables swap, then they may get a nasty shock when the
3394  * data getting swapped to that swapfile suddenly gets overwritten by
3395  * the original zero's written out previously to the journal and
3396  * awaiting writeback in the kernel's buffer cache.
3397  *
3398  * So, if we see any bmap calls here on a modified, data-journaled file,
3399  * take extra steps to flush any blocks which might be in the cache.
3400  */
3401 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3402 {
3403         struct inode *inode = mapping->host;
3404         journal_t *journal;
3405         int err;
3406
3407         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3408                         test_opt(inode->i_sb, DELALLOC)) {
3409                 /*
3410                  * With delalloc we want to sync the file
3411                  * so that we can make sure we allocate
3412                  * blocks for file
3413                  */
3414                 filemap_write_and_wait(mapping);
3415         }
3416
3417         if (EXT4_JOURNAL(inode) &&
3418             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3419                 /*
3420                  * This is a REALLY heavyweight approach, but the use of
3421                  * bmap on dirty files is expected to be extremely rare:
3422                  * only if we run lilo or swapon on a freshly made file
3423                  * do we expect this to happen.
3424                  *
3425                  * (bmap requires CAP_SYS_RAWIO so this does not
3426                  * represent an unprivileged user DOS attack --- we'd be
3427                  * in trouble if mortal users could trigger this path at
3428                  * will.)
3429                  *
3430                  * NB. EXT4_STATE_JDATA is not set on files other than
3431                  * regular files.  If somebody wants to bmap a directory
3432                  * or symlink and gets confused because the buffer
3433                  * hasn't yet been flushed to disk, they deserve
3434                  * everything they get.
3435                  */
3436
3437                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3438                 journal = EXT4_JOURNAL(inode);
3439                 jbd2_journal_lock_updates(journal);
3440                 err = jbd2_journal_flush(journal);
3441                 jbd2_journal_unlock_updates(journal);
3442
3443                 if (err)
3444                         return 0;
3445         }
3446
3447         return generic_block_bmap(mapping, block, ext4_get_block);
3448 }
3449
3450 static int ext4_readpage(struct file *file, struct page *page)
3451 {
3452         return mpage_readpage(page, ext4_get_block);
3453 }
3454
3455 static int
3456 ext4_readpages(struct file *file, struct address_space *mapping,
3457                 struct list_head *pages, unsigned nr_pages)
3458 {
3459         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3460 }
3461
3462 static void ext4_free_io_end(ext4_io_end_t *io)
3463 {
3464         BUG_ON(!io);
3465         if (io->page)
3466                 put_page(io->page);
3467         iput(io->inode);
3468         kfree(io);
3469 }
3470
3471 static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
3472 {
3473         struct buffer_head *head, *bh;
3474         unsigned int curr_off = 0;
3475
3476         if (!page_has_buffers(page))
3477                 return;
3478         head = bh = page_buffers(page);
3479         do {
3480                 if (offset <= curr_off && test_clear_buffer_uninit(bh)
3481                                         && bh->b_private) {
3482                         ext4_free_io_end(bh->b_private);
3483                         bh->b_private = NULL;
3484                         bh->b_end_io = NULL;
3485                 }
3486                 curr_off = curr_off + bh->b_size;
3487                 bh = bh->b_this_page;
3488         } while (bh != head);
3489 }
3490
3491 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3492 {
3493         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3494
3495         /*
3496          * free any io_end structure allocated for buffers to be discarded
3497          */
3498         if (ext4_should_dioread_nolock(page->mapping->host))
3499                 ext4_invalidatepage_free_endio(page, offset);
3500         /*
3501          * If it's a full truncate we just forget about the pending dirtying
3502          */
3503         if (offset == 0)
3504                 ClearPageChecked(page);
3505
3506         if (journal)
3507                 jbd2_journal_invalidatepage(journal, page, offset);
3508         else
3509                 block_invalidatepage(page, offset);
3510 }
3511
3512 static int ext4_releasepage(struct page *page, gfp_t wait)
3513 {
3514         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3515
3516         WARN_ON(PageChecked(page));
3517         if (!page_has_buffers(page))
3518                 return 0;
3519         if (journal)
3520                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3521         else
3522                 return try_to_free_buffers(page);
3523 }
3524
3525 /*
3526  * O_DIRECT for ext3 (or indirect map) based files
3527  *
3528  * If the O_DIRECT write will extend the file then add this inode to the
3529  * orphan list.  So recovery will truncate it back to the original size
3530  * if the machine crashes during the write.
3531  *
3532  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3533  * crashes then stale disk data _may_ be exposed inside the file. But current
3534  * VFS code falls back into buffered path in that case so we are safe.
3535  */
3536 static ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
3537                               const struct iovec *iov, loff_t offset,
3538                               unsigned long nr_segs)
3539 {
3540         struct file *file = iocb->ki_filp;
3541         struct inode *inode = file->f_mapping->host;
3542         struct ext4_inode_info *ei = EXT4_I(inode);
3543         handle_t *handle;
3544         ssize_t ret;
3545         int orphan = 0;
3546         size_t count = iov_length(iov, nr_segs);
3547         int retries = 0;
3548
3549         if (rw == WRITE) {
3550                 loff_t final_size = offset + count;
3551
3552                 if (final_size > inode->i_size) {
3553                         /* Credits for sb + inode write */
3554                         handle = ext4_journal_start(inode, 2);
3555                         if (IS_ERR(handle)) {
3556                                 ret = PTR_ERR(handle);
3557                                 goto out;
3558                         }
3559                         ret = ext4_orphan_add(handle, inode);
3560                         if (ret) {
3561                                 ext4_journal_stop(handle);
3562                                 goto out;
3563                         }
3564                         orphan = 1;
3565                         ei->i_disksize = inode->i_size;
3566                         ext4_journal_stop(handle);
3567                 }
3568         }
3569
3570 retry:
3571         if (rw == READ && ext4_should_dioread_nolock(inode))
3572                 ret = __blockdev_direct_IO(rw, iocb, inode,
3573                                  inode->i_sb->s_bdev, iov,
3574                                  offset, nr_segs,
3575                                  ext4_get_block, NULL, NULL, 0);
3576         else {
3577                 ret = blockdev_direct_IO(rw, iocb, inode,
3578                                  inode->i_sb->s_bdev, iov,
3579                                  offset, nr_segs,
3580                                  ext4_get_block, NULL);
3581
3582                 if (unlikely((rw & WRITE) && ret < 0)) {
3583                         loff_t isize = i_size_read(inode);
3584                         loff_t end = offset + iov_length(iov, nr_segs);
3585
3586                         if (end > isize)
3587                                 vmtruncate(inode, isize);
3588                 }
3589         }
3590         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3591                 goto retry;
3592
3593         if (orphan) {
3594                 int err;
3595
3596                 /* Credits for sb + inode write */
3597                 handle = ext4_journal_start(inode, 2);
3598                 if (IS_ERR(handle)) {
3599                         /* This is really bad luck. We've written the data
3600                          * but cannot extend i_size. Bail out and pretend
3601                          * the write failed... */
3602                         ret = PTR_ERR(handle);
3603                         if (inode->i_nlink)
3604                                 ext4_orphan_del(NULL, inode);
3605
3606                         goto out;
3607                 }
3608                 if (inode->i_nlink)
3609                         ext4_orphan_del(handle, inode);
3610                 if (ret > 0) {
3611                         loff_t end = offset + ret;
3612                         if (end > inode->i_size) {
3613                                 ei->i_disksize = end;
3614                                 i_size_write(inode, end);
3615                                 /*
3616                                  * We're going to return a positive `ret'
3617                                  * here due to non-zero-length I/O, so there's
3618                                  * no way of reporting error returns from
3619                                  * ext4_mark_inode_dirty() to userspace.  So
3620                                  * ignore it.
3621                                  */
3622                                 ext4_mark_inode_dirty(handle, inode);
3623                         }
3624                 }
3625                 err = ext4_journal_stop(handle);
3626                 if (ret == 0)
3627                         ret = err;
3628         }
3629 out:
3630         return ret;
3631 }
3632
3633 /*
3634  * ext4_get_block used when preparing for a DIO write or buffer write.
3635  * We allocate an uinitialized extent if blocks haven't been allocated.
3636  * The extent will be converted to initialized after the IO is complete.
3637  */
3638 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
3639                    struct buffer_head *bh_result, int create)
3640 {
3641         ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
3642                    inode->i_ino, create);
3643         return _ext4_get_block(inode, iblock, bh_result,
3644                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
3645 }
3646
3647 static void dump_completed_IO(struct inode * inode)
3648 {
3649 #ifdef  EXT4_DEBUG
3650         struct list_head *cur, *before, *after;
3651         ext4_io_end_t *io, *io0, *io1;
3652         unsigned long flags;
3653
3654         if (list_empty(&EXT4_I(inode)->i_completed_io_list)){
3655                 ext4_debug("inode %lu completed_io list is empty\n", inode->i_ino);
3656                 return;
3657         }
3658
3659         ext4_debug("Dump inode %lu completed_io list \n", inode->i_ino);
3660         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3661         list_for_each_entry(io, &EXT4_I(inode)->i_completed_io_list, list){
3662                 cur = &io->list;
3663                 before = cur->prev;
3664                 io0 = container_of(before, ext4_io_end_t, list);
3665                 after = cur->next;
3666                 io1 = container_of(after, ext4_io_end_t, list);
3667
3668                 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
3669                             io, inode->i_ino, io0, io1);
3670         }
3671         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3672 #endif
3673 }
3674
3675 /*
3676  * check a range of space and convert unwritten extents to written.
3677  */
3678 static int ext4_end_io_nolock(ext4_io_end_t *io)
3679 {
3680         struct inode *inode = io->inode;
3681         loff_t offset = io->offset;
3682         ssize_t size = io->size;
3683         int ret = 0;
3684
3685         ext4_debug("ext4_end_io_nolock: io 0x%p from inode %lu,list->next 0x%p,"
3686                    "list->prev 0x%p\n",
3687                    io, inode->i_ino, io->list.next, io->list.prev);
3688
3689         if (list_empty(&io->list))
3690                 return ret;
3691
3692         if (io->flag != EXT4_IO_UNWRITTEN)
3693                 return ret;
3694
3695         ret = ext4_convert_unwritten_extents(inode, offset, size);
3696         if (ret < 0) {
3697                 printk(KERN_EMERG "%s: failed to convert unwritten"
3698                         "extents to written extents, error is %d"
3699                         " io is still on inode %lu aio dio list\n",
3700                        __func__, ret, inode->i_ino);
3701                 return ret;
3702         }
3703
3704         if (io->iocb)
3705                 aio_complete(io->iocb, io->result, 0);
3706         /* clear the DIO AIO unwritten flag */
3707         io->flag = 0;
3708         return ret;
3709 }
3710
3711 /*
3712  * work on completed aio dio IO, to convert unwritten extents to extents
3713  */
3714 static void ext4_end_io_work(struct work_struct *work)
3715 {
3716         ext4_io_end_t           *io = container_of(work, ext4_io_end_t, work);
3717         struct inode            *inode = io->inode;
3718         struct ext4_inode_info  *ei = EXT4_I(inode);
3719         unsigned long           flags;
3720         int                     ret;
3721
3722         mutex_lock(&inode->i_mutex);
3723         ret = ext4_end_io_nolock(io);
3724         if (ret < 0) {
3725                 mutex_unlock(&inode->i_mutex);
3726                 return;
3727         }
3728
3729         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3730         if (!list_empty(&io->list))
3731                 list_del_init(&io->list);
3732         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3733         mutex_unlock(&inode->i_mutex);
3734         ext4_free_io_end(io);
3735 }
3736
3737 /*
3738  * This function is called from ext4_sync_file().
3739  *
3740  * When IO is completed, the work to convert unwritten extents to
3741  * written is queued on workqueue but may not get immediately
3742  * scheduled. When fsync is called, we need to ensure the
3743  * conversion is complete before fsync returns.
3744  * The inode keeps track of a list of pending/completed IO that
3745  * might needs to do the conversion. This function walks through
3746  * the list and convert the related unwritten extents for completed IO
3747  * to written.
3748  * The function return the number of pending IOs on success.
3749  */
3750 int flush_completed_IO(struct inode *inode)
3751 {
3752         ext4_io_end_t *io;
3753         struct ext4_inode_info *ei = EXT4_I(inode);
3754         unsigned long flags;
3755         int ret = 0;
3756         int ret2 = 0;
3757
3758         if (list_empty(&ei->i_completed_io_list))
3759                 return ret;
3760
3761         dump_completed_IO(inode);
3762         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3763         while (!list_empty(&ei->i_completed_io_list)){
3764                 io = list_entry(ei->i_completed_io_list.next,
3765                                 ext4_io_end_t, list);
3766                 /*
3767                  * Calling ext4_end_io_nolock() to convert completed
3768                  * IO to written.
3769                  *
3770                  * When ext4_sync_file() is called, run_queue() may already
3771                  * about to flush the work corresponding to this io structure.
3772                  * It will be upset if it founds the io structure related
3773                  * to the work-to-be schedule is freed.
3774                  *
3775                  * Thus we need to keep the io structure still valid here after
3776                  * convertion finished. The io structure has a flag to
3777                  * avoid double converting from both fsync and background work
3778                  * queue work.
3779                  */
3780                 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3781                 ret = ext4_end_io_nolock(io);
3782                 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3783                 if (ret < 0)
3784                         ret2 = ret;
3785                 else
3786                         list_del_init(&io->list);
3787         }
3788         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3789         return (ret2 < 0) ? ret2 : 0;
3790 }
3791
3792 static ext4_io_end_t *ext4_init_io_end (struct inode *inode, gfp_t flags)
3793 {
3794         ext4_io_end_t *io = NULL;
3795
3796         io = kmalloc(sizeof(*io), flags);
3797
3798         if (io) {
3799                 igrab(inode);
3800                 io->inode = inode;
3801                 io->flag = 0;
3802                 io->offset = 0;
3803                 io->size = 0;
3804                 io->page = NULL;
3805                 io->iocb = NULL;
3806                 io->result = 0;
3807                 INIT_WORK(&io->work, ext4_end_io_work);
3808                 INIT_LIST_HEAD(&io->list);
3809         }
3810
3811         return io;
3812 }
3813
3814 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3815                             ssize_t size, void *private, int ret,
3816                             bool is_async)
3817 {
3818         ext4_io_end_t *io_end = iocb->private;
3819         struct workqueue_struct *wq;
3820         unsigned long flags;
3821         struct ext4_inode_info *ei;
3822
3823         /* if not async direct IO or dio with 0 bytes write, just return */
3824         if (!io_end || !size)
3825                 goto out;
3826
3827         ext_debug("ext4_end_io_dio(): io_end 0x%p"
3828                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
3829                   iocb->private, io_end->inode->i_ino, iocb, offset,
3830                   size);
3831
3832         /* if not aio dio with unwritten extents, just free io and return */
3833         if (io_end->flag != EXT4_IO_UNWRITTEN){
3834                 ext4_free_io_end(io_end);
3835                 iocb->private = NULL;
3836 out:
3837                 if (is_async)
3838                         aio_complete(iocb, ret, 0);
3839                 return;
3840         }
3841
3842         io_end->offset = offset;
3843         io_end->size = size;
3844         if (is_async) {
3845                 io_end->iocb = iocb;
3846                 io_end->result = ret;
3847         }
3848         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
3849
3850         /* queue the work to convert unwritten extents to written */
3851         queue_work(wq, &io_end->work);
3852
3853         /* Add the io_end to per-inode completed aio dio list*/
3854         ei = EXT4_I(io_end->inode);
3855         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3856         list_add_tail(&io_end->list, &ei->i_completed_io_list);
3857         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3858         iocb->private = NULL;
3859 }
3860
3861 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
3862 {
3863         ext4_io_end_t *io_end = bh->b_private;
3864         struct workqueue_struct *wq;
3865         struct inode *inode;
3866         unsigned long flags;
3867
3868         if (!test_clear_buffer_uninit(bh) || !io_end)
3869                 goto out;
3870
3871         if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
3872                 printk("sb umounted, discard end_io request for inode %lu\n",
3873                         io_end->inode->i_ino);
3874                 ext4_free_io_end(io_end);
3875                 goto out;
3876         }
3877
3878         io_end->flag = EXT4_IO_UNWRITTEN;
3879         inode = io_end->inode;
3880
3881         /* Add the io_end to per-inode completed io list*/
3882         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3883         list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
3884         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3885
3886         wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
3887         /* queue the work to convert unwritten extents to written */
3888         queue_work(wq, &io_end->work);
3889 out:
3890         bh->b_private = NULL;
3891         bh->b_end_io = NULL;
3892         clear_buffer_uninit(bh);
3893         end_buffer_async_write(bh, uptodate);
3894 }
3895
3896 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
3897 {
3898         ext4_io_end_t *io_end;
3899         struct page *page = bh->b_page;
3900         loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
3901         size_t size = bh->b_size;
3902
3903 retry:
3904         io_end = ext4_init_io_end(inode, GFP_ATOMIC);
3905         if (!io_end) {
3906                 if (printk_ratelimit())
3907                         printk(KERN_WARNING "%s: allocation fail\n", __func__);
3908                 schedule();
3909                 goto retry;
3910         }
3911         io_end->offset = offset;
3912         io_end->size = size;
3913         /*
3914          * We need to hold a reference to the page to make sure it
3915          * doesn't get evicted before ext4_end_io_work() has a chance
3916          * to convert the extent from written to unwritten.
3917          */
3918         io_end->page = page;
3919         get_page(io_end->page);
3920
3921         bh->b_private = io_end;
3922         bh->b_end_io = ext4_end_io_buffer_write;
3923         return 0;
3924 }
3925
3926 /*
3927  * For ext4 extent files, ext4 will do direct-io write to holes,
3928  * preallocated extents, and those write extend the file, no need to
3929  * fall back to buffered IO.
3930  *
3931  * For holes, we fallocate those blocks, mark them as unintialized
3932  * If those blocks were preallocated, we mark sure they are splited, but
3933  * still keep the range to write as unintialized.
3934  *
3935  * The unwrritten extents will be converted to written when DIO is completed.
3936  * For async direct IO, since the IO may still pending when return, we
3937  * set up an end_io call back function, which will do the convertion
3938  * when async direct IO completed.
3939  *
3940  * If the O_DIRECT write will extend the file then add this inode to the
3941  * orphan list.  So recovery will truncate it back to the original size
3942  * if the machine crashes during the write.
3943  *
3944  */
3945 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
3946                               const struct iovec *iov, loff_t offset,
3947                               unsigned long nr_segs)
3948 {
3949         struct file *file = iocb->ki_filp;
3950         struct inode *inode = file->f_mapping->host;
3951         ssize_t ret;
3952         size_t count = iov_length(iov, nr_segs);
3953
3954         loff_t final_size = offset + count;
3955         if (rw == WRITE && final_size <= inode->i_size) {
3956                 /*
3957                  * We could direct write to holes and fallocate.
3958                  *
3959                  * Allocated blocks to fill the hole are marked as uninitialized
3960                  * to prevent paralel buffered read to expose the stale data
3961                  * before DIO complete the data IO.
3962                  *
3963                  * As to previously fallocated extents, ext4 get_block
3964                  * will just simply mark the buffer mapped but still
3965                  * keep the extents uninitialized.
3966                  *
3967                  * for non AIO case, we will convert those unwritten extents
3968                  * to written after return back from blockdev_direct_IO.
3969                  *
3970                  * for async DIO, the conversion needs to be defered when
3971                  * the IO is completed. The ext4 end_io callback function
3972                  * will be called to take care of the conversion work.
3973                  * Here for async case, we allocate an io_end structure to
3974                  * hook to the iocb.
3975                  */
3976                 iocb->private = NULL;
3977                 EXT4_I(inode)->cur_aio_dio = NULL;
3978                 if (!is_sync_kiocb(iocb)) {
3979                         iocb->private = ext4_init_io_end(inode, GFP_NOFS);
3980                         if (!iocb->private)
3981                                 return -ENOMEM;
3982                         /*
3983                          * we save the io structure for current async
3984                          * direct IO, so that later ext4_map_blocks()
3985                          * could flag the io structure whether there
3986                          * is a unwritten extents needs to be converted
3987                          * when IO is completed.
3988                          */
3989                         EXT4_I(inode)->cur_aio_dio = iocb->private;
3990                 }
3991
3992                 ret = blockdev_direct_IO(rw, iocb, inode,
3993                                          inode->i_sb->s_bdev, iov,
3994                                          offset, nr_segs,
3995                                          ext4_get_block_write,
3996                                          ext4_end_io_dio);
3997                 if (iocb->private)
3998                         EXT4_I(inode)->cur_aio_dio = NULL;
3999                 /*
4000                  * The io_end structure takes a reference to the inode,
4001                  * that structure needs to be destroyed and the
4002                  * reference to the inode need to be dropped, when IO is
4003                  * complete, even with 0 byte write, or failed.
4004                  *
4005                  * In the successful AIO DIO case, the io_end structure will be
4006                  * desctroyed and the reference to the inode will be dropped
4007                  * after the end_io call back function is called.
4008                  *
4009                  * In the case there is 0 byte write, or error case, since
4010                  * VFS direct IO won't invoke the end_io call back function,
4011                  * we need to free the end_io structure here.
4012                  */
4013                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
4014                         ext4_free_io_end(iocb->private);
4015                         iocb->private = NULL;
4016                 } else if (ret > 0 && ext4_test_inode_state(inode,
4017                                                 EXT4_STATE_DIO_UNWRITTEN)) {
4018                         int err;
4019                         /*
4020                          * for non AIO case, since the IO is already
4021                          * completed, we could do the convertion right here
4022                          */
4023                         err = ext4_convert_unwritten_extents(inode,
4024                                                              offset, ret);
4025                         if (err < 0)
4026                                 ret = err;
4027                         ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
4028                 }
4029                 return ret;
4030         }
4031
4032         /* for write the the end of file case, we fall back to old way */
4033         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
4034 }
4035
4036 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
4037                               const struct iovec *iov, loff_t offset,
4038                               unsigned long nr_segs)
4039 {
4040         struct file *file = iocb->ki_filp;
4041         struct inode *inode = file->f_mapping->host;
4042
4043         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4044                 return ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
4045
4046         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
4047 }
4048
4049 /*
4050  * Pages can be marked dirty completely asynchronously from ext4's journalling
4051  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
4052  * much here because ->set_page_dirty is called under VFS locks.  The page is
4053  * not necessarily locked.
4054  *
4055  * We cannot just dirty the page and leave attached buffers clean, because the
4056  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
4057  * or jbddirty because all the journalling code will explode.
4058  *
4059  * So what we do is to mark the page "pending dirty" and next time writepage
4060  * is called, propagate that into the buffers appropriately.
4061  */
4062 static int ext4_journalled_set_page_dirty(struct page *page)
4063 {
4064         SetPageChecked(page);
4065         return __set_page_dirty_nobuffers(page);
4066 }
4067
4068 static const struct address_space_operations ext4_ordered_aops = {
4069         .readpage               = ext4_readpage,
4070         .readpages              = ext4_readpages,
4071         .writepage              = ext4_writepage,
4072         .sync_page              = block_sync_page,
4073         .write_begin            = ext4_write_begin,
4074         .write_end              = ext4_ordered_write_end,
4075         .bmap                   = ext4_bmap,
4076         .invalidatepage         = ext4_invalidatepage,
4077         .releasepage            = ext4_releasepage,
4078         .direct_IO              = ext4_direct_IO,
4079         .migratepage            = buffer_migrate_page,
4080         .is_partially_uptodate  = block_is_partially_uptodate,
4081         .error_remove_page      = generic_error_remove_page,
4082 };
4083
4084 static const struct address_space_operations ext4_writeback_aops = {
4085         .readpage               = ext4_readpage,
4086         .readpages              = ext4_readpages,
4087         .writepage              = ext4_writepage,
4088         .sync_page              = block_sync_page,
4089         .write_begin            = ext4_write_begin,
4090         .write_end              = ext4_writeback_write_end,
4091         .bmap                   = ext4_bmap,
4092         .invalidatepage         = ext4_invalidatepage,
4093         .releasepage            = ext4_releasepage,
4094         .direct_IO              = ext4_direct_IO,
4095         .migratepage            = buffer_migrate_page,
4096         .is_partially_uptodate  = block_is_partially_uptodate,
4097         .error_remove_page      = generic_error_remove_page,
4098 };
4099
4100 static const struct address_space_operations ext4_journalled_aops = {
4101         .readpage               = ext4_readpage,
4102         .readpages              = ext4_readpages,
4103         .writepage              = ext4_writepage,
4104         .sync_page              = block_sync_page,
4105         .write_begin            = ext4_write_begin,
4106         .write_end              = ext4_journalled_write_end,
4107         .set_page_dirty         = ext4_journalled_set_page_dirty,
4108         .bmap                   = ext4_bmap,
4109         .invalidatepage         = ext4_invalidatepage,
4110         .releasepage            = ext4_releasepage,
4111         .is_partially_uptodate  = block_is_partially_uptodate,
4112         .error_remove_page      = generic_error_remove_page,
4113 };
4114
4115 static const struct address_space_operations ext4_da_aops = {
4116         .readpage               = ext4_readpage,
4117         .readpages              = ext4_readpages,
4118         .writepage              = ext4_writepage,
4119         .writepages             = ext4_da_writepages,
4120         .sync_page              = block_sync_page,
4121         .write_begin            = ext4_da_write_begin,
4122         .write_end              = ext4_da_write_end,
4123         .bmap                   = ext4_bmap,
4124         .invalidatepage         = ext4_da_invalidatepage,
4125         .releasepage            = ext4_releasepage,
4126         .direct_IO              = ext4_direct_IO,
4127         .migratepage            = buffer_migrate_page,
4128         .is_partially_uptodate  = block_is_partially_uptodate,
4129         .error_remove_page      = generic_error_remove_page,
4130 };
4131
4132 void ext4_set_aops(struct inode *inode)
4133 {
4134         if (ext4_should_order_data(inode) &&
4135                 test_opt(inode->i_sb, DELALLOC))
4136                 inode->i_mapping->a_ops = &ext4_da_aops;
4137         else if (ext4_should_order_data(inode))
4138                 inode->i_mapping->a_ops = &ext4_ordered_aops;
4139         else if (ext4_should_writeback_data(inode) &&
4140                  test_opt(inode->i_sb, DELALLOC))
4141                 inode->i_mapping->a_ops = &ext4_da_aops;
4142         else if (ext4_should_writeback_data(inode))
4143                 inode->i_mapping->a_ops = &ext4_writeback_aops;
4144         else
4145                 inode->i_mapping->a_ops = &ext4_journalled_aops;
4146 }
4147
4148 /*
4149  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
4150  * up to the end of the block which corresponds to `from'.
4151  * This required during truncate. We need to physically zero the tail end
4152  * of that block so it doesn't yield old data if the file is later grown.
4153  */
4154 int ext4_block_truncate_page(handle_t *handle,
4155                 struct address_space *mapping, loff_t from)
4156 {
4157         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
4158         unsigned offset = from & (PAGE_CACHE_SIZE-1);
4159         unsigned blocksize, length, pos;
4160         ext4_lblk_t iblock;
4161         struct inode *inode = mapping->host;
4162         struct buffer_head *bh;
4163         struct page *page;
4164         int err = 0;
4165
4166         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
4167                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
4168         if (!page)
4169                 return -EINVAL;
4170
4171         blocksize = inode->i_sb->s_blocksize;
4172         length = blocksize - (offset & (blocksize - 1));
4173         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
4174
4175         if (!page_has_buffers(page))
4176                 create_empty_buffers(page, blocksize, 0);
4177
4178         /* Find the buffer that contains "offset" */
4179         bh = page_buffers(page);
4180         pos = blocksize;
4181         while (offset >= pos) {
4182                 bh = bh->b_this_page;
4183                 iblock++;
4184                 pos += blocksize;
4185         }
4186
4187         err = 0;
4188         if (buffer_freed(bh)) {
4189                 BUFFER_TRACE(bh, "freed: skip");
4190                 goto unlock;
4191         }
4192
4193         if (!buffer_mapped(bh)) {
4194                 BUFFER_TRACE(bh, "unmapped");
4195                 ext4_get_block(inode, iblock, bh, 0);
4196                 /* unmapped? It's a hole - nothing to do */
4197                 if (!buffer_mapped(bh)) {
4198                         BUFFER_TRACE(bh, "still unmapped");
4199                         goto unlock;
4200                 }
4201         }
4202
4203         /* Ok, it's mapped. Make sure it's up-to-date */
4204         if (PageUptodate(page))
4205                 set_buffer_uptodate(bh);
4206
4207         if (!buffer_uptodate(bh)) {
4208                 err = -EIO;
4209                 ll_rw_block(READ, 1, &bh);
4210                 wait_on_buffer(bh);
4211                 /* Uhhuh. Read error. Complain and punt. */
4212                 if (!buffer_uptodate(bh))
4213                         goto unlock;
4214         }
4215
4216         if (ext4_should_journal_data(inode)) {
4217                 BUFFER_TRACE(bh, "get write access");
4218                 err = ext4_journal_get_write_access(handle, bh);
4219                 if (err)
4220                         goto unlock;
4221         }
4222
4223         zero_user(page, offset, length);
4224
4225         BUFFER_TRACE(bh, "zeroed end of block");
4226
4227         err = 0;
4228         if (ext4_should_journal_data(inode)) {
4229                 err = ext4_handle_dirty_metadata(handle, inode, bh);
4230         } else {
4231                 if (ext4_should_order_data(inode))
4232                         err = ext4_jbd2_file_inode(handle, inode);
4233                 mark_buffer_dirty(bh);
4234         }
4235
4236 unlock:
4237         unlock_page(page);
4238         page_cache_release(page);
4239         return err;
4240 }
4241
4242 /*
4243  * Probably it should be a library function... search for first non-zero word
4244  * or memcmp with zero_page, whatever is better for particular architecture.
4245  * Linus?
4246  */
4247 static inline int all_zeroes(__le32 *p, __le32 *q)
4248 {
4249         while (p < q)
4250                 if (*p++)
4251                         return 0;
4252         return 1;
4253 }
4254
4255 /**
4256  *      ext4_find_shared - find the indirect blocks for partial truncation.
4257  *      @inode:   inode in question
4258  *      @depth:   depth of the affected branch
4259  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
4260  *      @chain:   place to store the pointers to partial indirect blocks
4261  *      @top:     place to the (detached) top of branch
4262  *
4263  *      This is a helper function used by ext4_truncate().
4264  *
4265  *      When we do truncate() we may have to clean the ends of several
4266  *      indirect blocks but leave the blocks themselves alive. Block is
4267  *      partially truncated if some data below the new i_size is refered
4268  *      from it (and it is on the path to the first completely truncated
4269  *      data block, indeed).  We have to free the top of that path along
4270  *      with everything to the right of the path. Since no allocation
4271  *      past the truncation point is possible until ext4_truncate()
4272  *      finishes, we may safely do the latter, but top of branch may
4273  *      require special attention - pageout below the truncation point
4274  *      might try to populate it.
4275  *
4276  *      We atomically detach the top of branch from the tree, store the
4277  *      block number of its root in *@top, pointers to buffer_heads of
4278  *      partially truncated blocks - in @chain[].bh and pointers to
4279  *      their last elements that should not be removed - in
4280  *      @chain[].p. Return value is the pointer to last filled element
4281  *      of @chain.
4282  *
4283  *      The work left to caller to do the actual freeing of subtrees:
4284  *              a) free the subtree starting from *@top
4285  *              b) free the subtrees whose roots are stored in
4286  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
4287  *              c) free the subtrees growing from the inode past the @chain[0].
4288  *                      (no partially truncated stuff there).  */
4289
4290 static Indirect *ext4_find_shared(struct inode *inode, int depth,
4291                                   ext4_lblk_t offsets[4], Indirect chain[4],
4292                                   __le32 *top)
4293 {
4294         Indirect *partial, *p;
4295         int k, err;
4296
4297         *top = 0;
4298         /* Make k index the deepest non-null offset + 1 */
4299         for (k = depth; k > 1 && !offsets[k-1]; k--)
4300                 ;
4301         partial = ext4_get_branch(inode, k, offsets, chain, &err);
4302         /* Writer: pointers */
4303         if (!partial)
4304                 partial = chain + k-1;
4305         /*
4306          * If the branch acquired continuation since we've looked at it -
4307          * fine, it should all survive and (new) top doesn't belong to us.
4308          */
4309         if (!partial->key && *partial->p)
4310                 /* Writer: end */
4311                 goto no_top;
4312         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
4313                 ;
4314         /*
4315          * OK, we've found the last block that must survive. The rest of our
4316          * branch should be detached before unlocking. However, if that rest
4317          * of branch is all ours and does not grow immediately from the inode
4318          * it's easier to cheat and just decrement partial->p.
4319          */
4320         if (p == chain + k - 1 && p > chain) {
4321                 p->p--;
4322         } else {
4323                 *top = *p->p;
4324                 /* Nope, don't do this in ext4.  Must leave the tree intact */
4325 #if 0
4326                 *p->p = 0;
4327 #endif
4328         }
4329         /* Writer: end */
4330
4331         while (partial > p) {
4332                 brelse(partial->bh);
4333                 partial--;
4334         }
4335 no_top:
4336         return partial;
4337 }
4338
4339 /*
4340  * Zero a number of block pointers in either an inode or an indirect block.
4341  * If we restart the transaction we must again get write access to the
4342  * indirect block for further modification.
4343  *
4344  * We release `count' blocks on disk, but (last - first) may be greater
4345  * than `count' because there can be holes in there.
4346  */
4347 static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
4348                              struct buffer_head *bh,
4349                              ext4_fsblk_t block_to_free,
4350                              unsigned long count, __le32 *first,
4351                              __le32 *last)
4352 {
4353         __le32 *p;
4354         int     flags = EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_VALIDATED;
4355
4356         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
4357                 flags |= EXT4_FREE_BLOCKS_METADATA;
4358
4359         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
4360                                    count)) {
4361                 EXT4_ERROR_INODE(inode, "attempt to clear invalid "
4362                                  "blocks %llu len %lu",
4363                                  (unsigned long long) block_to_free, count);
4364                 return 1;
4365         }
4366
4367         if (try_to_extend_transaction(handle, inode)) {
4368                 if (bh) {
4369                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4370                         ext4_handle_dirty_metadata(handle, inode, bh);
4371                 }
4372                 ext4_mark_inode_dirty(handle, inode);
4373                 ext4_truncate_restart_trans(handle, inode,
4374                                             blocks_for_truncate(inode));
4375                 if (bh) {
4376                         BUFFER_TRACE(bh, "retaking write access");
4377                         ext4_journal_get_write_access(handle, bh);
4378                 }
4379         }
4380
4381         for (p = first; p < last; p++)
4382                 *p = 0;
4383
4384         ext4_free_blocks(handle, inode, 0, block_to_free, count, flags);
4385         return 0;
4386 }
4387
4388 /**
4389  * ext4_free_data - free a list of data blocks
4390  * @handle:     handle for this transaction
4391  * @inode:      inode we are dealing with
4392  * @this_bh:    indirect buffer_head which contains *@first and *@last
4393  * @first:      array of block numbers
4394  * @last:       points immediately past the end of array
4395  *
4396  * We are freeing all blocks refered from that array (numbers are stored as
4397  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
4398  *
4399  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
4400  * blocks are contiguous then releasing them at one time will only affect one
4401  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
4402  * actually use a lot of journal space.
4403  *
4404  * @this_bh will be %NULL if @first and @last point into the inode's direct
4405  * block pointers.
4406  */
4407 static void ext4_free_data(handle_t *handle, struct inode *inode,
4408                            struct buffer_head *this_bh,
4409                            __le32 *first, __le32 *last)
4410 {
4411         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
4412         unsigned long count = 0;            /* Number of blocks in the run */
4413         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
4414                                                corresponding to
4415                                                block_to_free */
4416         ext4_fsblk_t nr;                    /* Current block # */
4417         __le32 *p;                          /* Pointer into inode/ind
4418                                                for current block */
4419         int err;
4420
4421         if (this_bh) {                          /* For indirect block */
4422                 BUFFER_TRACE(this_bh, "get_write_access");
4423                 err = ext4_journal_get_write_access(handle, this_bh);
4424                 /* Important: if we can't update the indirect pointers
4425                  * to the blocks, we can't free them. */
4426                 if (err)
4427                         return;
4428         }
4429
4430         for (p = first; p < last; p++) {
4431                 nr = le32_to_cpu(*p);
4432                 if (nr) {
4433                         /* accumulate blocks to free if they're contiguous */
4434                         if (count == 0) {
4435                                 block_to_free = nr;
4436                                 block_to_free_p = p;
4437                                 count = 1;
4438                         } else if (nr == block_to_free + count) {
4439                                 count++;
4440                         } else {
4441                                 if (ext4_clear_blocks(handle, inode, this_bh,
4442                                                       block_to_free, count,
4443                                                       block_to_free_p, p))
4444                                         break;
4445                                 block_to_free = nr;
4446                                 block_to_free_p = p;
4447                                 count = 1;
4448                         }
4449                 }
4450         }
4451
4452         if (count > 0)
4453                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
4454                                   count, block_to_free_p, p);
4455
4456         if (this_bh) {
4457                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
4458
4459                 /*
4460                  * The buffer head should have an attached journal head at this
4461                  * point. However, if the data is corrupted and an indirect
4462                  * block pointed to itself, it would have been detached when
4463                  * the block was cleared. Check for this instead of OOPSing.
4464                  */
4465                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
4466                         ext4_handle_dirty_metadata(handle, inode, this_bh);
4467                 else
4468                         EXT4_ERROR_INODE(inode,
4469                                          "circular indirect block detected at "
4470                                          "block %llu",
4471                                 (unsigned long long) this_bh->b_blocknr);
4472         }
4473 }
4474
4475 /**
4476  *      ext4_free_branches - free an array of branches
4477  *      @handle: JBD handle for this transaction
4478  *      @inode: inode we are dealing with
4479  *      @parent_bh: the buffer_head which contains *@first and *@last
4480  *      @first: array of block numbers
4481  *      @last:  pointer immediately past the end of array
4482  *      @depth: depth of the branches to free
4483  *
4484  *      We are freeing all blocks refered from these branches (numbers are
4485  *      stored as little-endian 32-bit) and updating @inode->i_blocks
4486  *      appropriately.
4487  */
4488 static void ext4_free_branches(handle_t *handle, struct inode *inode,
4489                                struct buffer_head *parent_bh,
4490                                __le32 *first, __le32 *last, int depth)
4491 {
4492         ext4_fsblk_t nr;
4493         __le32 *p;
4494
4495         if (ext4_handle_is_aborted(handle))
4496                 return;
4497
4498         if (depth--) {
4499                 struct buffer_head *bh;
4500                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4501                 p = last;
4502                 while (--p >= first) {
4503                         nr = le32_to_cpu(*p);
4504                         if (!nr)
4505                                 continue;               /* A hole */
4506
4507                         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
4508                                                    nr, 1)) {
4509                                 EXT4_ERROR_INODE(inode,
4510                                                  "invalid indirect mapped "
4511                                                  "block %lu (level %d)",
4512                                                  (unsigned long) nr, depth);
4513                                 break;
4514                         }
4515
4516                         /* Go read the buffer for the next level down */
4517                         bh = sb_bread(inode->i_sb, nr);
4518
4519                         /*
4520                          * A read failure? Report error and clear slot
4521                          * (should be rare).
4522                          */
4523                         if (!bh) {
4524                                 EXT4_ERROR_INODE_BLOCK(inode, nr,
4525                                                        "Read failure");
4526                                 continue;
4527                         }
4528
4529                         /* This zaps the entire block.  Bottom up. */
4530                         BUFFER_TRACE(bh, "free child branches");
4531                         ext4_free_branches(handle, inode, bh,
4532                                         (__le32 *) bh->b_data,
4533                                         (__le32 *) bh->b_data + addr_per_block,
4534                                         depth);
4535
4536                         /*
4537                          * Everything below this this pointer has been
4538                          * released.  Now let this top-of-subtree go.
4539                          *
4540                          * We want the freeing of this indirect block to be
4541                          * atomic in the journal with the updating of the
4542                          * bitmap block which owns it.  So make some room in
4543                          * the journal.
4544                          *
4545                          * We zero the parent pointer *after* freeing its
4546                          * pointee in the bitmaps, so if extend_transaction()
4547                          * for some reason fails to put the bitmap changes and
4548                          * the release into the same transaction, recovery
4549                          * will merely complain about releasing a free block,
4550                          * rather than leaking blocks.
4551                          */
4552                         if (ext4_handle_is_aborted(handle))
4553                                 return;
4554                         if (try_to_extend_transaction(handle, inode)) {
4555                                 ext4_mark_inode_dirty(handle, inode);
4556                                 ext4_truncate_restart_trans(handle, inode,
4557                                             blocks_for_truncate(inode));
4558                         }
4559
4560                         /*
4561                          * The forget flag here is critical because if
4562                          * we are journaling (and not doing data
4563                          * journaling), we have to make sure a revoke
4564                          * record is written to prevent the journal
4565                          * replay from overwriting the (former)
4566                          * indirect block if it gets reallocated as a
4567                          * data block.  This must happen in the same
4568                          * transaction where the data blocks are
4569                          * actually freed.
4570                          */
4571                         ext4_free_blocks(handle, inode, 0, nr, 1,
4572                                          EXT4_FREE_BLOCKS_METADATA|
4573                                          EXT4_FREE_BLOCKS_FORGET);
4574
4575                         if (parent_bh) {
4576                                 /*
4577                                  * The block which we have just freed is
4578                                  * pointed to by an indirect block: journal it
4579                                  */
4580                                 BUFFER_TRACE(parent_bh, "get_write_access");
4581                                 if (!ext4_journal_get_write_access(handle,
4582                                                                    parent_bh)){
4583                                         *p = 0;
4584                                         BUFFER_TRACE(parent_bh,
4585                                         "call ext4_handle_dirty_metadata");
4586                                         ext4_handle_dirty_metadata(handle,
4587                                                                    inode,
4588                                                                    parent_bh);
4589                                 }
4590                         }
4591                 }
4592         } else {
4593                 /* We have reached the bottom of the tree. */
4594                 BUFFER_TRACE(parent_bh, "free data blocks");
4595                 ext4_free_data(handle, inode, parent_bh, first, last);
4596         }
4597 }
4598
4599 int ext4_can_truncate(struct inode *inode)
4600 {
4601         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4602                 return 0;
4603         if (S_ISREG(inode->i_mode))
4604                 return 1;
4605         if (S_ISDIR(inode->i_mode))
4606                 return 1;
4607         if (S_ISLNK(inode->i_mode))
4608                 return !ext4_inode_is_fast_symlink(inode);
4609         return 0;
4610 }
4611
4612 /*
4613  * ext4_truncate()
4614  *
4615  * We block out ext4_get_block() block instantiations across the entire
4616  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4617  * simultaneously on behalf of the same inode.
4618  *
4619  * As we work through the truncate and commmit bits of it to the journal there
4620  * is one core, guiding principle: the file's tree must always be consistent on
4621  * disk.  We must be able to restart the truncate after a crash.
4622  *
4623  * The file's tree may be transiently inconsistent in memory (although it
4624  * probably isn't), but whenever we close off and commit a journal transaction,
4625  * the contents of (the filesystem + the journal) must be consistent and
4626  * restartable.  It's pretty simple, really: bottom up, right to left (although
4627  * left-to-right works OK too).
4628  *
4629  * Note that at recovery time, journal replay occurs *before* the restart of
4630  * truncate against the orphan inode list.
4631  *
4632  * The committed inode has the new, desired i_size (which is the same as
4633  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4634  * that this inode's truncate did not complete and it will again call
4635  * ext4_truncate() to have another go.  So there will be instantiated blocks
4636  * to the right of the truncation point in a crashed ext4 filesystem.  But
4637  * that's fine - as long as they are linked from the inode, the post-crash
4638  * ext4_truncate() run will find them and release them.
4639  */
4640 void ext4_truncate(struct inode *inode)
4641 {
4642         handle_t *handle;
4643         struct ext4_inode_info *ei = EXT4_I(inode);
4644         __le32 *i_data = ei->i_data;
4645         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4646         struct address_space *mapping = inode->i_mapping;
4647         ext4_lblk_t offsets[4];
4648         Indirect chain[4];
4649         Indirect *partial;
4650         __le32 nr = 0;
4651         int n;
4652         ext4_lblk_t last_block;
4653         unsigned blocksize = inode->i_sb->s_blocksize;
4654
4655         if (!ext4_can_truncate(inode))
4656                 return;
4657
4658         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4659
4660         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4661                 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4662
4663         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
4664                 ext4_ext_truncate(inode);
4665                 return;
4666         }
4667
4668         handle = start_transaction(inode);
4669         if (IS_ERR(handle))
4670                 return;         /* AKPM: return what? */
4671
4672         last_block = (inode->i_size + blocksize-1)
4673                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4674
4675         if (inode->i_size & (blocksize - 1))
4676                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4677                         goto out_stop;
4678
4679         n = ext4_block_to_path(inode, last_block, offsets, NULL);
4680         if (n == 0)
4681                 goto out_stop;  /* error */
4682
4683         /*
4684          * OK.  This truncate is going to happen.  We add the inode to the
4685          * orphan list, so that if this truncate spans multiple transactions,
4686          * and we crash, we will resume the truncate when the filesystem
4687          * recovers.  It also marks the inode dirty, to catch the new size.
4688          *
4689          * Implication: the file must always be in a sane, consistent
4690          * truncatable state while each transaction commits.
4691          */
4692         if (ext4_orphan_add(handle, inode))
4693                 goto out_stop;
4694
4695         /*
4696          * From here we block out all ext4_get_block() callers who want to
4697          * modify the block allocation tree.
4698          */
4699         down_write(&ei->i_data_sem);
4700
4701         ext4_discard_preallocations(inode);
4702
4703         /*
4704          * The orphan list entry will now protect us from any crash which
4705          * occurs before the truncate completes, so it is now safe to propagate
4706          * the new, shorter inode size (held for now in i_size) into the
4707          * on-disk inode. We do this via i_disksize, which is the value which
4708          * ext4 *really* writes onto the disk inode.
4709          */
4710         ei->i_disksize = inode->i_size;
4711
4712         if (n == 1) {           /* direct blocks */
4713                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4714                                i_data + EXT4_NDIR_BLOCKS);
4715                 goto do_indirects;
4716         }
4717
4718         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4719         /* Kill the top of shared branch (not detached) */
4720         if (nr) {
4721                 if (partial == chain) {
4722                         /* Shared branch grows from the inode */
4723                         ext4_free_branches(handle, inode, NULL,
4724                                            &nr, &nr+1, (chain+n-1) - partial);
4725                         *partial->p = 0;
4726                         /*
4727                          * We mark the inode dirty prior to restart,
4728                          * and prior to stop.  No need for it here.
4729                          */
4730                 } else {
4731                         /* Shared branch grows from an indirect block */
4732                         BUFFER_TRACE(partial->bh, "get_write_access");
4733                         ext4_free_branches(handle, inode, partial->bh,
4734                                         partial->p,
4735                                         partial->p+1, (chain+n-1) - partial);
4736                 }
4737         }
4738         /* Clear the ends of indirect blocks on the shared branch */
4739         while (partial > chain) {
4740                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4741                                    (__le32*)partial->bh->b_data+addr_per_block,
4742                                    (chain+n-1) - partial);
4743                 BUFFER_TRACE(partial->bh, "call brelse");
4744                 brelse(partial->bh);
4745                 partial--;
4746         }
4747 do_indirects:
4748         /* Kill the remaining (whole) subtrees */
4749         switch (offsets[0]) {
4750         default:
4751                 nr = i_data[EXT4_IND_BLOCK];
4752                 if (nr) {
4753                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4754                         i_data[EXT4_IND_BLOCK] = 0;
4755                 }
4756         case EXT4_IND_BLOCK:
4757                 nr = i_data[EXT4_DIND_BLOCK];
4758                 if (nr) {
4759                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4760                         i_data[EXT4_DIND_BLOCK] = 0;
4761                 }
4762         case EXT4_DIND_BLOCK:
4763                 nr = i_data[EXT4_TIND_BLOCK];
4764                 if (nr) {
4765                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4766                         i_data[EXT4_TIND_BLOCK] = 0;
4767                 }
4768         case EXT4_TIND_BLOCK:
4769                 ;
4770         }
4771
4772         up_write(&ei->i_data_sem);
4773         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4774         ext4_mark_inode_dirty(handle, inode);
4775
4776         /*
4777          * In a multi-transaction truncate, we only make the final transaction
4778          * synchronous
4779          */
4780         if (IS_SYNC(inode))
4781                 ext4_handle_sync(handle);
4782 out_stop:
4783         /*
4784          * If this was a simple ftruncate(), and the file will remain alive
4785          * then we need to clear up the orphan record which we created above.
4786          * However, if this was a real unlink then we were called by
4787          * ext4_delete_inode(), and we allow that function to clean up the
4788          * orphan info for us.
4789          */
4790         if (inode->i_nlink)
4791                 ext4_orphan_del(handle, inode);
4792
4793         ext4_journal_stop(handle);
4794 }
4795
4796 /*
4797  * ext4_get_inode_loc returns with an extra refcount against the inode's
4798  * underlying buffer_head on success. If 'in_mem' is true, we have all
4799  * data in memory that is needed to recreate the on-disk version of this
4800  * inode.
4801  */
4802 static int __ext4_get_inode_loc(struct inode *inode,
4803                                 struct ext4_iloc *iloc, int in_mem)
4804 {
4805         struct ext4_group_desc  *gdp;
4806         struct buffer_head      *bh;
4807         struct super_block      *sb = inode->i_sb;
4808         ext4_fsblk_t            block;
4809         int                     inodes_per_block, inode_offset;
4810
4811         iloc->bh = NULL;
4812         if (!ext4_valid_inum(sb, inode->i_ino))
4813                 return -EIO;
4814
4815         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4816         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4817         if (!gdp)
4818                 return -EIO;
4819
4820         /*
4821          * Figure out the offset within the block group inode table
4822          */
4823         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4824         inode_offset = ((inode->i_ino - 1) %
4825                         EXT4_INODES_PER_GROUP(sb));
4826         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4827         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4828
4829         bh = sb_getblk(sb, block);
4830         if (!bh) {
4831                 EXT4_ERROR_INODE_BLOCK(inode, block,
4832                                        "unable to read itable block");
4833                 return -EIO;
4834         }
4835         if (!buffer_uptodate(bh)) {
4836                 lock_buffer(bh);
4837
4838                 /*
4839                  * If the buffer has the write error flag, we have failed
4840                  * to write out another inode in the same block.  In this
4841                  * case, we don't have to read the block because we may
4842                  * read the old inode data successfully.
4843                  */
4844                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4845                         set_buffer_uptodate(bh);
4846
4847                 if (buffer_uptodate(bh)) {
4848                         /* someone brought it uptodate while we waited */
4849                         unlock_buffer(bh);
4850                         goto has_buffer;
4851                 }
4852
4853                 /*
4854                  * If we have all information of the inode in memory and this
4855                  * is the only valid inode in the block, we need not read the
4856                  * block.
4857                  */
4858                 if (in_mem) {
4859                         struct buffer_head *bitmap_bh;
4860                         int i, start;
4861
4862                         start = inode_offset & ~(inodes_per_block - 1);
4863
4864                         /* Is the inode bitmap in cache? */
4865                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4866                         if (!bitmap_bh)
4867                                 goto make_io;
4868
4869                         /*
4870                          * If the inode bitmap isn't in cache then the
4871                          * optimisation may end up performing two reads instead
4872                          * of one, so skip it.
4873                          */
4874                         if (!buffer_uptodate(bitmap_bh)) {
4875                                 brelse(bitmap_bh);
4876                                 goto make_io;
4877                         }
4878                         for (i = start; i < start + inodes_per_block; i++) {
4879                                 if (i == inode_offset)
4880                                         continue;
4881                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4882                                         break;
4883                         }
4884                         brelse(bitmap_bh);
4885                         if (i == start + inodes_per_block) {
4886                                 /* all other inodes are free, so skip I/O */
4887                                 memset(bh->b_data, 0, bh->b_size);
4888                                 set_buffer_uptodate(bh);
4889                                 unlock_buffer(bh);
4890                                 goto has_buffer;
4891                         }
4892                 }
4893
4894 make_io:
4895                 /*
4896                  * If we need to do any I/O, try to pre-readahead extra
4897                  * blocks from the inode table.
4898                  */
4899                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4900                         ext4_fsblk_t b, end, table;
4901                         unsigned num;
4902
4903                         table = ext4_inode_table(sb, gdp);
4904                         /* s_inode_readahead_blks is always a power of 2 */
4905                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4906                         if (table > b)
4907                                 b = table;
4908                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4909                         num = EXT4_INODES_PER_GROUP(sb);
4910                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4911                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4912                                 num -= ext4_itable_unused_count(sb, gdp);
4913                         table += num / inodes_per_block;
4914                         if (end > table)
4915                                 end = table;
4916                         while (b <= end)
4917                                 sb_breadahead(sb, b++);
4918                 }
4919
4920                 /*
4921                  * There are other valid inodes in the buffer, this inode
4922                  * has in-inode xattrs, or we don't have this inode in memory.
4923                  * Read the block from disk.
4924                  */
4925                 get_bh(bh);
4926                 bh->b_end_io = end_buffer_read_sync;
4927                 submit_bh(READ_META, bh);
4928                 wait_on_buffer(bh);
4929                 if (!buffer_uptodate(bh)) {
4930                         EXT4_ERROR_INODE_BLOCK(inode, block,
4931                                                "unable to read itable block");
4932                         brelse(bh);
4933                         return -EIO;
4934                 }
4935         }
4936 has_buffer:
4937         iloc->bh = bh;
4938         return 0;
4939 }
4940
4941 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4942 {
4943         /* We have all inode data except xattrs in memory here. */
4944         return __ext4_get_inode_loc(inode, iloc,
4945                 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
4946 }
4947
4948 void ext4_set_inode_flags(struct inode *inode)
4949 {
4950         unsigned int flags = EXT4_I(inode)->i_flags;
4951
4952         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4953         if (flags & EXT4_SYNC_FL)
4954                 inode->i_flags |= S_SYNC;
4955         if (flags & EXT4_APPEND_FL)
4956                 inode->i_flags |= S_APPEND;
4957         if (flags & EXT4_IMMUTABLE_FL)
4958                 inode->i_flags |= S_IMMUTABLE;
4959         if (flags & EXT4_NOATIME_FL)
4960                 inode->i_flags |= S_NOATIME;
4961         if (flags & EXT4_DIRSYNC_FL)
4962                 inode->i_flags |= S_DIRSYNC;
4963 }
4964
4965 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4966 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4967 {
4968         unsigned int vfs_fl;
4969         unsigned long old_fl, new_fl;
4970
4971         do {
4972                 vfs_fl = ei->vfs_inode.i_flags;
4973                 old_fl = ei->i_flags;
4974                 new_fl = old_fl & ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4975                                 EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|
4976                                 EXT4_DIRSYNC_FL);
4977                 if (vfs_fl & S_SYNC)
4978                         new_fl |= EXT4_SYNC_FL;
4979                 if (vfs_fl & S_APPEND)
4980                         new_fl |= EXT4_APPEND_FL;
4981                 if (vfs_fl & S_IMMUTABLE)
4982                         new_fl |= EXT4_IMMUTABLE_FL;
4983                 if (vfs_fl & S_NOATIME)
4984                         new_fl |= EXT4_NOATIME_FL;
4985                 if (vfs_fl & S_DIRSYNC)
4986                         new_fl |= EXT4_DIRSYNC_FL;
4987         } while (cmpxchg(&ei->i_flags, old_fl, new_fl) != old_fl);
4988 }
4989
4990 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4991                                   struct ext4_inode_info *ei)
4992 {
4993         blkcnt_t i_blocks ;
4994         struct inode *inode = &(ei->vfs_inode);
4995         struct super_block *sb = inode->i_sb;
4996
4997         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4998                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4999                 /* we are using combined 48 bit field */
5000                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
5001                                         le32_to_cpu(raw_inode->i_blocks_lo);
5002                 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
5003                         /* i_blocks represent file system block size */
5004                         return i_blocks  << (inode->i_blkbits - 9);
5005                 } else {
5006                         return i_blocks;
5007                 }
5008         } else {
5009                 return le32_to_cpu(raw_inode->i_blocks_lo);
5010         }
5011 }
5012
5013 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
5014 {
5015         struct ext4_iloc iloc;
5016         struct ext4_inode *raw_inode;
5017         struct ext4_inode_info *ei;
5018         struct inode *inode;
5019         journal_t *journal = EXT4_SB(sb)->s_journal;
5020         long ret;
5021         int block;
5022
5023         inode = iget_locked(sb, ino);
5024         if (!inode)
5025                 return ERR_PTR(-ENOMEM);
5026         if (!(inode->i_state & I_NEW))
5027                 return inode;
5028
5029         ei = EXT4_I(inode);
5030         iloc.bh = 0;
5031
5032         ret = __ext4_get_inode_loc(inode, &iloc, 0);
5033         if (ret < 0)
5034                 goto bad_inode;
5035         raw_inode = ext4_raw_inode(&iloc);
5036         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
5037         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
5038         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
5039         if (!(test_opt(inode->i_sb, NO_UID32))) {
5040                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
5041                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
5042         }
5043         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
5044
5045         ei->i_state_flags = 0;
5046         ei->i_dir_start_lookup = 0;
5047         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
5048         /* We now have enough fields to check if the inode was active or not.
5049          * This is needed because nfsd might try to access dead inodes
5050          * the test is that same one that e2fsck uses
5051          * NeilBrown 1999oct15
5052          */
5053         if (inode->i_nlink == 0) {
5054                 if (inode->i_mode == 0 ||
5055                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
5056                         /* this inode is deleted */
5057                         ret = -ESTALE;
5058                         goto bad_inode;
5059                 }
5060                 /* The only unlinked inodes we let through here have
5061                  * valid i_mode and are being read by the orphan
5062                  * recovery code: that's fine, we're about to complete
5063                  * the process of deleting those. */
5064         }
5065         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
5066         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
5067         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
5068         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
5069                 ei->i_file_acl |=
5070                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
5071         inode->i_size = ext4_isize(raw_inode);
5072         ei->i_disksize = inode->i_size;
5073 #ifdef CONFIG_QUOTA
5074         ei->i_reserved_quota = 0;
5075 #endif
5076         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
5077         ei->i_block_group = iloc.block_group;
5078         ei->i_last_alloc_group = ~0;
5079         /*
5080          * NOTE! The in-memory inode i_data array is in little-endian order
5081          * even on big-endian machines: we do NOT byteswap the block numbers!
5082          */
5083         for (block = 0; block < EXT4_N_BLOCKS; block++)
5084                 ei->i_data[block] = raw_inode->i_block[block];
5085         INIT_LIST_HEAD(&ei->i_orphan);
5086
5087         /*
5088          * Set transaction id's of transactions that have to be committed
5089          * to finish f[data]sync. We set them to currently running transaction
5090          * as we cannot be sure that the inode or some of its metadata isn't
5091          * part of the transaction - the inode could have been reclaimed and
5092          * now it is reread from disk.
5093          */
5094         if (journal) {
5095                 transaction_t *transaction;
5096                 tid_t tid;
5097
5098                 read_lock(&journal->j_state_lock);
5099                 if (journal->j_running_transaction)
5100                         transaction = journal->j_running_transaction;
5101                 else
5102                         transaction = journal->j_committing_transaction;
5103                 if (transaction)
5104                         tid = transaction->t_tid;
5105                 else
5106                         tid = journal->j_commit_sequence;
5107                 read_unlock(&journal->j_state_lock);
5108                 ei->i_sync_tid = tid;
5109                 ei->i_datasync_tid = tid;
5110         }
5111
5112         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5113                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
5114                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
5115                     EXT4_INODE_SIZE(inode->i_sb)) {
5116                         ret = -EIO;
5117                         goto bad_inode;
5118                 }
5119                 if (ei->i_extra_isize == 0) {
5120                         /* The extra space is currently unused. Use it. */
5121                         ei->i_extra_isize = sizeof(struct ext4_inode) -
5122                                             EXT4_GOOD_OLD_INODE_SIZE;
5123                 } else {
5124                         __le32 *magic = (void *)raw_inode +
5125                                         EXT4_GOOD_OLD_INODE_SIZE +
5126                                         ei->i_extra_isize;
5127                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
5128                                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
5129                 }
5130         } else
5131                 ei->i_extra_isize = 0;
5132
5133         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
5134         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
5135         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
5136         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
5137
5138         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
5139         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5140                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5141                         inode->i_version |=
5142                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
5143         }
5144
5145         ret = 0;
5146         if (ei->i_file_acl &&
5147             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
5148                 EXT4_ERROR_INODE(inode, "bad extended attribute block %llu",
5149                                  ei->i_file_acl);
5150                 ret = -EIO;
5151                 goto bad_inode;
5152         } else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5153                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5154                     (S_ISLNK(inode->i_mode) &&
5155                      !ext4_inode_is_fast_symlink(inode)))
5156                         /* Validate extent which is part of inode */
5157                         ret = ext4_ext_check_inode(inode);
5158         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5159                    (S_ISLNK(inode->i_mode) &&
5160                     !ext4_inode_is_fast_symlink(inode))) {
5161                 /* Validate block references which are part of inode */
5162                 ret = ext4_check_inode_blockref(inode);
5163         }
5164         if (ret)
5165                 goto bad_inode;
5166
5167         if (S_ISREG(inode->i_mode)) {
5168                 inode->i_op = &ext4_file_inode_operations;
5169                 inode->i_fop = &ext4_file_operations;
5170                 ext4_set_aops(inode);
5171         } else if (S_ISDIR(inode->i_mode)) {
5172                 inode->i_op = &ext4_dir_inode_operations;
5173                 inode->i_fop = &ext4_dir_operations;
5174         } else if (S_ISLNK(inode->i_mode)) {
5175                 if (ext4_inode_is_fast_symlink(inode)) {
5176                         inode->i_op = &ext4_fast_symlink_inode_operations;
5177                         nd_terminate_link(ei->i_data, inode->i_size,
5178                                 sizeof(ei->i_data) - 1);
5179                 } else {
5180                         inode->i_op = &ext4_symlink_inode_operations;
5181                         ext4_set_aops(inode);
5182                 }
5183         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
5184               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
5185                 inode->i_op = &ext4_special_inode_operations;
5186                 if (raw_inode->i_block[0])
5187                         init_special_inode(inode, inode->i_mode,
5188                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
5189                 else
5190                         init_special_inode(inode, inode->i_mode,
5191                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
5192         } else {
5193                 ret = -EIO;
5194                 EXT4_ERROR_INODE(inode, "bogus i_mode (%o)", inode->i_mode);
5195                 goto bad_inode;
5196         }
5197         brelse(iloc.bh);
5198         ext4_set_inode_flags(inode);
5199         unlock_new_inode(inode);
5200         return inode;
5201
5202 bad_inode:
5203         brelse(iloc.bh);
5204         iget_failed(inode);
5205         return ERR_PTR(ret);
5206 }
5207
5208 static int ext4_inode_blocks_set(handle_t *handle,
5209                                 struct ext4_inode *raw_inode,
5210                                 struct ext4_inode_info *ei)
5211 {
5212         struct inode *inode = &(ei->vfs_inode);
5213         u64 i_blocks = inode->i_blocks;
5214         struct super_block *sb = inode->i_sb;
5215
5216         if (i_blocks <= ~0U) {
5217                 /*
5218                  * i_blocks can be represnted in a 32 bit variable
5219                  * as multiple of 512 bytes
5220                  */
5221                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5222                 raw_inode->i_blocks_high = 0;
5223                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5224                 return 0;
5225         }
5226         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
5227                 return -EFBIG;
5228
5229         if (i_blocks <= 0xffffffffffffULL) {
5230                 /*
5231                  * i_blocks can be represented in a 48 bit variable
5232                  * as multiple of 512 bytes
5233                  */
5234                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5235                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5236                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5237         } else {
5238                 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5239                 /* i_block is stored in file system block size */
5240                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
5241                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5242                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5243         }
5244         return 0;
5245 }
5246
5247 /*
5248  * Post the struct inode info into an on-disk inode location in the
5249  * buffer-cache.  This gobbles the caller's reference to the
5250  * buffer_head in the inode location struct.
5251  *
5252  * The caller must have write access to iloc->bh.
5253  */
5254 static int ext4_do_update_inode(handle_t *handle,
5255                                 struct inode *inode,
5256                                 struct ext4_iloc *iloc)
5257 {
5258         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5259         struct ext4_inode_info *ei = EXT4_I(inode);
5260         struct buffer_head *bh = iloc->bh;
5261         int err = 0, rc, block;
5262
5263         /* For fields not not tracking in the in-memory inode,
5264          * initialise them to zero for new inodes. */
5265         if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5266                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5267
5268         ext4_get_inode_flags(ei);
5269         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5270         if (!(test_opt(inode->i_sb, NO_UID32))) {
5271                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
5272                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
5273 /*
5274  * Fix up interoperability with old kernels. Otherwise, old inodes get
5275  * re-used with the upper 16 bits of the uid/gid intact
5276  */
5277                 if (!ei->i_dtime) {
5278                         raw_inode->i_uid_high =
5279                                 cpu_to_le16(high_16_bits(inode->i_uid));
5280                         raw_inode->i_gid_high =
5281                                 cpu_to_le16(high_16_bits(inode->i_gid));
5282                 } else {
5283                         raw_inode->i_uid_high = 0;
5284                         raw_inode->i_gid_high = 0;
5285                 }
5286         } else {
5287                 raw_inode->i_uid_low =
5288                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
5289                 raw_inode->i_gid_low =
5290                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
5291                 raw_inode->i_uid_high = 0;
5292                 raw_inode->i_gid_high = 0;
5293         }
5294         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5295
5296         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5297         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5298         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5299         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5300
5301         if (ext4_inode_blocks_set(handle, raw_inode, ei))
5302                 goto out_brelse;
5303         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5304         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
5305         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
5306             cpu_to_le32(EXT4_OS_HURD))
5307                 raw_inode->i_file_acl_high =
5308                         cpu_to_le16(ei->i_file_acl >> 32);
5309         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5310         ext4_isize_set(raw_inode, ei->i_disksize);
5311         if (ei->i_disksize > 0x7fffffffULL) {
5312                 struct super_block *sb = inode->i_sb;
5313                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
5314                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
5315                                 EXT4_SB(sb)->s_es->s_rev_level ==
5316                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
5317                         /* If this is the first large file
5318                          * created, add a flag to the superblock.
5319                          */
5320                         err = ext4_journal_get_write_access(handle,
5321                                         EXT4_SB(sb)->s_sbh);
5322                         if (err)
5323                                 goto out_brelse;
5324                         ext4_update_dynamic_rev(sb);
5325                         EXT4_SET_RO_COMPAT_FEATURE(sb,
5326                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
5327                         sb->s_dirt = 1;
5328                         ext4_handle_sync(handle);
5329                         err = ext4_handle_dirty_metadata(handle, NULL,
5330                                         EXT4_SB(sb)->s_sbh);
5331                 }
5332         }
5333         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5334         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5335                 if (old_valid_dev(inode->i_rdev)) {
5336                         raw_inode->i_block[0] =
5337                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
5338                         raw_inode->i_block[1] = 0;
5339                 } else {
5340                         raw_inode->i_block[0] = 0;
5341                         raw_inode->i_block[1] =
5342                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
5343                         raw_inode->i_block[2] = 0;
5344                 }
5345         } else
5346                 for (block = 0; block < EXT4_N_BLOCKS; block++)
5347                         raw_inode->i_block[block] = ei->i_data[block];
5348
5349         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
5350         if (ei->i_extra_isize) {
5351                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5352                         raw_inode->i_version_hi =
5353                         cpu_to_le32(inode->i_version >> 32);
5354                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
5355         }
5356
5357         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5358         rc = ext4_handle_dirty_metadata(handle, NULL, bh);
5359         if (!err)
5360                 err = rc;
5361         ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5362
5363         ext4_update_inode_fsync_trans(handle, inode, 0);
5364 out_brelse:
5365         brelse(bh);
5366         ext4_std_error(inode->i_sb, err);
5367         return err;
5368 }
5369
5370 /*
5371  * ext4_write_inode()
5372  *
5373  * We are called from a few places:
5374  *
5375  * - Within generic_file_write() for O_SYNC files.
5376  *   Here, there will be no transaction running. We wait for any running
5377  *   trasnaction to commit.
5378  *
5379  * - Within sys_sync(), kupdate and such.
5380  *   We wait on commit, if tol to.
5381  *
5382  * - Within prune_icache() (PF_MEMALLOC == true)
5383  *   Here we simply return.  We can't afford to block kswapd on the
5384  *   journal commit.
5385  *
5386  * In all cases it is actually safe for us to return without doing anything,
5387  * because the inode has been copied into a raw inode buffer in
5388  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
5389  * knfsd.
5390  *
5391  * Note that we are absolutely dependent upon all inode dirtiers doing the
5392  * right thing: they *must* call mark_inode_dirty() after dirtying info in
5393  * which we are interested.
5394  *
5395  * It would be a bug for them to not do this.  The code:
5396  *
5397  *      mark_inode_dirty(inode)
5398  *      stuff();
5399  *      inode->i_size = expr;
5400  *
5401  * is in error because a kswapd-driven write_inode() could occur while
5402  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
5403  * will no longer be on the superblock's dirty inode list.
5404  */
5405 int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5406 {
5407         int err;
5408
5409         if (current->flags & PF_MEMALLOC)
5410                 return 0;
5411
5412         if (EXT4_SB(inode->i_sb)->s_journal) {
5413                 if (ext4_journal_current_handle()) {
5414                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5415                         dump_stack();
5416                         return -EIO;
5417                 }
5418
5419                 if (wbc->sync_mode != WB_SYNC_ALL)
5420                         return 0;
5421
5422                 err = ext4_force_commit(inode->i_sb);
5423         } else {
5424                 struct ext4_iloc iloc;
5425
5426                 err = __ext4_get_inode_loc(inode, &iloc, 0);
5427                 if (err)
5428                         return err;
5429                 if (wbc->sync_mode == WB_SYNC_ALL)
5430                         sync_dirty_buffer(iloc.bh);
5431                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5432                         EXT4_ERROR_INODE_BLOCK(inode, iloc.bh->b_blocknr,
5433                                          "IO error syncing inode");
5434                         err = -EIO;
5435                 }
5436                 brelse(iloc.bh);
5437         }
5438         return err;
5439 }
5440
5441 /*
5442  * ext4_setattr()
5443  *
5444  * Called from notify_change.
5445  *
5446  * We want to trap VFS attempts to truncate the file as soon as
5447  * possible.  In particular, we want to make sure that when the VFS
5448  * shrinks i_size, we put the inode on the orphan list and modify
5449  * i_disksize immediately, so that during the subsequent flushing of
5450  * dirty pages and freeing of disk blocks, we can guarantee that any
5451  * commit will leave the blocks being flushed in an unused state on
5452  * disk.  (On recovery, the inode will get truncated and the blocks will
5453  * be freed, so we have a strong guarantee that no future commit will
5454  * leave these blocks visible to the user.)
5455  *
5456  * Another thing we have to assure is that if we are in ordered mode
5457  * and inode is still attached to the committing transaction, we must
5458  * we start writeout of all the dirty pages which are being truncated.
5459  * This way we are sure that all the data written in the previous
5460  * transaction are already on disk (truncate waits for pages under
5461  * writeback).
5462  *
5463  * Called with inode->i_mutex down.
5464  */
5465 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5466 {
5467         struct inode *inode = dentry->d_inode;
5468         int error, rc = 0;
5469         const unsigned int ia_valid = attr->ia_valid;
5470
5471         error = inode_change_ok(inode, attr);
5472         if (error)
5473                 return error;
5474
5475         if (is_quota_modification(inode, attr))
5476                 dquot_initialize(inode);
5477         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
5478                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
5479                 handle_t *handle;
5480
5481                 /* (user+group)*(old+new) structure, inode write (sb,
5482                  * inode block, ? - but truncate inode update has it) */
5483                 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
5484                                         EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
5485                 if (IS_ERR(handle)) {
5486                         error = PTR_ERR(handle);
5487                         goto err_out;
5488                 }
5489                 error = dquot_transfer(inode, attr);
5490                 if (error) {
5491                         ext4_journal_stop(handle);
5492                         return error;
5493                 }
5494                 /* Update corresponding info in inode so that everything is in
5495                  * one transaction */
5496                 if (attr->ia_valid & ATTR_UID)
5497                         inode->i_uid = attr->ia_uid;
5498                 if (attr->ia_valid & ATTR_GID)
5499                         inode->i_gid = attr->ia_gid;
5500                 error = ext4_mark_inode_dirty(handle, inode);
5501                 ext4_journal_stop(handle);
5502         }
5503
5504         if (attr->ia_valid & ATTR_SIZE) {
5505                 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
5506                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5507
5508                         if (attr->ia_size > sbi->s_bitmap_maxbytes)
5509                                 return -EFBIG;
5510                 }
5511         }
5512
5513         if (S_ISREG(inode->i_mode) &&
5514             attr->ia_valid & ATTR_SIZE &&
5515             (attr->ia_size < inode->i_size ||
5516              (ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS)))) {
5517                 handle_t *handle;
5518
5519                 handle = ext4_journal_start(inode, 3);
5520                 if (IS_ERR(handle)) {
5521                         error = PTR_ERR(handle);
5522                         goto err_out;
5523                 }
5524
5525                 error = ext4_orphan_add(handle, inode);
5526                 EXT4_I(inode)->i_disksize = attr->ia_size;
5527                 rc = ext4_mark_inode_dirty(handle, inode);
5528                 if (!error)
5529                         error = rc;
5530                 ext4_journal_stop(handle);
5531
5532                 if (ext4_should_order_data(inode)) {
5533                         error = ext4_begin_ordered_truncate(inode,
5534                                                             attr->ia_size);
5535                         if (error) {
5536                                 /* Do as much error cleanup as possible */
5537                                 handle = ext4_journal_start(inode, 3);
5538                                 if (IS_ERR(handle)) {
5539      &nbs