ext4: fix ext4_get_blocks references
[linux-2.6.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/jbd2.h>
29 #include <linux/highuid.h>
30 #include <linux/pagemap.h>
31 #include <linux/quotaops.h>
32 #include <linux/string.h>
33 #include <linux/buffer_head.h>
34 #include <linux/writeback.h>
35 #include <linux/pagevec.h>
36 #include <linux/mpage.h>
37 #include <linux/namei.h>
38 #include <linux/uio.h>
39 #include <linux/bio.h>
40 #include <linux/workqueue.h>
41 #include <linux/kernel.h>
42 #include <linux/slab.h>
43
44 #include "ext4_jbd2.h"
45 #include "xattr.h"
46 #include "acl.h"
47 #include "ext4_extents.h"
48
49 #include <trace/events/ext4.h>
50
51 #define MPAGE_DA_EXTENT_TAIL 0x01
52
53 static inline int ext4_begin_ordered_truncate(struct inode *inode,
54                                               loff_t new_size)
55 {
56         return jbd2_journal_begin_ordered_truncate(
57                                         EXT4_SB(inode->i_sb)->s_journal,
58                                         &EXT4_I(inode)->jinode,
59                                         new_size);
60 }
61
62 static void ext4_invalidatepage(struct page *page, unsigned long offset);
63
64 /*
65  * Test whether an inode is a fast symlink.
66  */
67 static int ext4_inode_is_fast_symlink(struct inode *inode)
68 {
69         int ea_blocks = EXT4_I(inode)->i_file_acl ?
70                 (inode->i_sb->s_blocksize >> 9) : 0;
71
72         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
73 }
74
75 /*
76  * Work out how many blocks we need to proceed with the next chunk of a
77  * truncate transaction.
78  */
79 static unsigned long blocks_for_truncate(struct inode *inode)
80 {
81         ext4_lblk_t needed;
82
83         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
84
85         /* Give ourselves just enough room to cope with inodes in which
86          * i_blocks is corrupt: we've seen disk corruptions in the past
87          * which resulted in random data in an inode which looked enough
88          * like a regular file for ext4 to try to delete it.  Things
89          * will go a bit crazy if that happens, but at least we should
90          * try not to panic the whole kernel. */
91         if (needed < 2)
92                 needed = 2;
93
94         /* But we need to bound the transaction so we don't overflow the
95          * journal. */
96         if (needed > EXT4_MAX_TRANS_DATA)
97                 needed = EXT4_MAX_TRANS_DATA;
98
99         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
100 }
101
102 /*
103  * Truncate transactions can be complex and absolutely huge.  So we need to
104  * be able to restart the transaction at a conventient checkpoint to make
105  * sure we don't overflow the journal.
106  *
107  * start_transaction gets us a new handle for a truncate transaction,
108  * and extend_transaction tries to extend the existing one a bit.  If
109  * extend fails, we need to propagate the failure up and restart the
110  * transaction in the top-level truncate loop. --sct
111  */
112 static handle_t *start_transaction(struct inode *inode)
113 {
114         handle_t *result;
115
116         result = ext4_journal_start(inode, blocks_for_truncate(inode));
117         if (!IS_ERR(result))
118                 return result;
119
120         ext4_std_error(inode->i_sb, PTR_ERR(result));
121         return result;
122 }
123
124 /*
125  * Try to extend this transaction for the purposes of truncation.
126  *
127  * Returns 0 if we managed to create more room.  If we can't create more
128  * room, and the transaction must be restarted we return 1.
129  */
130 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
131 {
132         if (!ext4_handle_valid(handle))
133                 return 0;
134         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
135                 return 0;
136         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
137                 return 0;
138         return 1;
139 }
140
141 /*
142  * Restart the transaction associated with *handle.  This does a commit,
143  * so before we call here everything must be consistently dirtied against
144  * this transaction.
145  */
146 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
147                                  int nblocks)
148 {
149         int ret;
150
151         /*
152          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
153          * moment, get_block can be called only for blocks inside i_size since
154          * page cache has been already dropped and writes are blocked by
155          * i_mutex. So we can safely drop the i_data_sem here.
156          */
157         BUG_ON(EXT4_JOURNAL(inode) == NULL);
158         jbd_debug(2, "restarting handle %p\n", handle);
159         up_write(&EXT4_I(inode)->i_data_sem);
160         ret = ext4_journal_restart(handle, blocks_for_truncate(inode));
161         down_write(&EXT4_I(inode)->i_data_sem);
162         ext4_discard_preallocations(inode);
163
164         return ret;
165 }
166
167 /*
168  * Called at the last iput() if i_nlink is zero.
169  */
170 void ext4_delete_inode(struct inode *inode)
171 {
172         handle_t *handle;
173         int err;
174
175         if (!is_bad_inode(inode))
176                 dquot_initialize(inode);
177
178         if (ext4_should_order_data(inode))
179                 ext4_begin_ordered_truncate(inode, 0);
180         truncate_inode_pages(&inode->i_data, 0);
181
182         if (is_bad_inode(inode))
183                 goto no_delete;
184
185         handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
186         if (IS_ERR(handle)) {
187                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
188                 /*
189                  * If we're going to skip the normal cleanup, we still need to
190                  * make sure that the in-core orphan linked list is properly
191                  * cleaned up.
192                  */
193                 ext4_orphan_del(NULL, inode);
194                 goto no_delete;
195         }
196
197         if (IS_SYNC(inode))
198                 ext4_handle_sync(handle);
199         inode->i_size = 0;
200         err = ext4_mark_inode_dirty(handle, inode);
201         if (err) {
202                 ext4_warning(inode->i_sb,
203                              "couldn't mark inode dirty (err %d)", err);
204                 goto stop_handle;
205         }
206         if (inode->i_blocks)
207                 ext4_truncate(inode);
208
209         /*
210          * ext4_ext_truncate() doesn't reserve any slop when it
211          * restarts journal transactions; therefore there may not be
212          * enough credits left in the handle to remove the inode from
213          * the orphan list and set the dtime field.
214          */
215         if (!ext4_handle_has_enough_credits(handle, 3)) {
216                 err = ext4_journal_extend(handle, 3);
217                 if (err > 0)
218                         err = ext4_journal_restart(handle, 3);
219                 if (err != 0) {
220                         ext4_warning(inode->i_sb,
221                                      "couldn't extend journal (err %d)", err);
222                 stop_handle:
223                         ext4_journal_stop(handle);
224                         goto no_delete;
225                 }
226         }
227
228         /*
229          * Kill off the orphan record which ext4_truncate created.
230          * AKPM: I think this can be inside the above `if'.
231          * Note that ext4_orphan_del() has to be able to cope with the
232          * deletion of a non-existent orphan - this is because we don't
233          * know if ext4_truncate() actually created an orphan record.
234          * (Well, we could do this if we need to, but heck - it works)
235          */
236         ext4_orphan_del(handle, inode);
237         EXT4_I(inode)->i_dtime  = get_seconds();
238
239         /*
240          * One subtle ordering requirement: if anything has gone wrong
241          * (transaction abort, IO errors, whatever), then we can still
242          * do these next steps (the fs will already have been marked as
243          * having errors), but we can't free the inode if the mark_dirty
244          * fails.
245          */
246         if (ext4_mark_inode_dirty(handle, inode))
247                 /* If that failed, just do the required in-core inode clear. */
248                 clear_inode(inode);
249         else
250                 ext4_free_inode(handle, inode);
251         ext4_journal_stop(handle);
252         return;
253 no_delete:
254         clear_inode(inode);     /* We must guarantee clearing of inode... */
255 }
256
257 typedef struct {
258         __le32  *p;
259         __le32  key;
260         struct buffer_head *bh;
261 } Indirect;
262
263 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
264 {
265         p->key = *(p->p = v);
266         p->bh = bh;
267 }
268
269 /**
270  *      ext4_block_to_path - parse the block number into array of offsets
271  *      @inode: inode in question (we are only interested in its superblock)
272  *      @i_block: block number to be parsed
273  *      @offsets: array to store the offsets in
274  *      @boundary: set this non-zero if the referred-to block is likely to be
275  *             followed (on disk) by an indirect block.
276  *
277  *      To store the locations of file's data ext4 uses a data structure common
278  *      for UNIX filesystems - tree of pointers anchored in the inode, with
279  *      data blocks at leaves and indirect blocks in intermediate nodes.
280  *      This function translates the block number into path in that tree -
281  *      return value is the path length and @offsets[n] is the offset of
282  *      pointer to (n+1)th node in the nth one. If @block is out of range
283  *      (negative or too large) warning is printed and zero returned.
284  *
285  *      Note: function doesn't find node addresses, so no IO is needed. All
286  *      we need to know is the capacity of indirect blocks (taken from the
287  *      inode->i_sb).
288  */
289
290 /*
291  * Portability note: the last comparison (check that we fit into triple
292  * indirect block) is spelled differently, because otherwise on an
293  * architecture with 32-bit longs and 8Kb pages we might get into trouble
294  * if our filesystem had 8Kb blocks. We might use long long, but that would
295  * kill us on x86. Oh, well, at least the sign propagation does not matter -
296  * i_block would have to be negative in the very beginning, so we would not
297  * get there at all.
298  */
299
300 static int ext4_block_to_path(struct inode *inode,
301                               ext4_lblk_t i_block,
302                               ext4_lblk_t offsets[4], int *boundary)
303 {
304         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
305         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
306         const long direct_blocks = EXT4_NDIR_BLOCKS,
307                 indirect_blocks = ptrs,
308                 double_blocks = (1 << (ptrs_bits * 2));
309         int n = 0;
310         int final = 0;
311
312         if (i_block < direct_blocks) {
313                 offsets[n++] = i_block;
314                 final = direct_blocks;
315         } else if ((i_block -= direct_blocks) < indirect_blocks) {
316                 offsets[n++] = EXT4_IND_BLOCK;
317                 offsets[n++] = i_block;
318                 final = ptrs;
319         } else if ((i_block -= indirect_blocks) < double_blocks) {
320                 offsets[n++] = EXT4_DIND_BLOCK;
321                 offsets[n++] = i_block >> ptrs_bits;
322                 offsets[n++] = i_block & (ptrs - 1);
323                 final = ptrs;
324         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
325                 offsets[n++] = EXT4_TIND_BLOCK;
326                 offsets[n++] = i_block >> (ptrs_bits * 2);
327                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
328                 offsets[n++] = i_block & (ptrs - 1);
329                 final = ptrs;
330         } else {
331                 ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
332                              i_block + direct_blocks +
333                              indirect_blocks + double_blocks, inode->i_ino);
334         }
335         if (boundary)
336                 *boundary = final - 1 - (i_block & (ptrs - 1));
337         return n;
338 }
339
340 static int __ext4_check_blockref(const char *function, unsigned int line,
341                                  struct inode *inode,
342                                  __le32 *p, unsigned int max)
343 {
344         struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
345         __le32 *bref = p;
346         unsigned int blk;
347
348         while (bref < p+max) {
349                 blk = le32_to_cpu(*bref++);
350                 if (blk &&
351                     unlikely(!ext4_data_block_valid(EXT4_SB(inode->i_sb),
352                                                     blk, 1))) {
353                         es->s_last_error_block = cpu_to_le64(blk);
354                         ext4_error_inode(inode, function, line, blk,
355                                          "invalid block");
356                         return -EIO;
357                 }
358         }
359         return 0;
360 }
361
362
363 #define ext4_check_indirect_blockref(inode, bh)                         \
364         __ext4_check_blockref(__func__, __LINE__, inode,                \
365                               (__le32 *)(bh)->b_data,                   \
366                               EXT4_ADDR_PER_BLOCK((inode)->i_sb))
367
368 #define ext4_check_inode_blockref(inode)                                \
369         __ext4_check_blockref(__func__, __LINE__, inode,                \
370                               EXT4_I(inode)->i_data,                    \
371                               EXT4_NDIR_BLOCKS)
372
373 /**
374  *      ext4_get_branch - read the chain of indirect blocks leading to data
375  *      @inode: inode in question
376  *      @depth: depth of the chain (1 - direct pointer, etc.)
377  *      @offsets: offsets of pointers in inode/indirect blocks
378  *      @chain: place to store the result
379  *      @err: here we store the error value
380  *
381  *      Function fills the array of triples <key, p, bh> and returns %NULL
382  *      if everything went OK or the pointer to the last filled triple
383  *      (incomplete one) otherwise. Upon the return chain[i].key contains
384  *      the number of (i+1)-th block in the chain (as it is stored in memory,
385  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
386  *      number (it points into struct inode for i==0 and into the bh->b_data
387  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
388  *      block for i>0 and NULL for i==0. In other words, it holds the block
389  *      numbers of the chain, addresses they were taken from (and where we can
390  *      verify that chain did not change) and buffer_heads hosting these
391  *      numbers.
392  *
393  *      Function stops when it stumbles upon zero pointer (absent block)
394  *              (pointer to last triple returned, *@err == 0)
395  *      or when it gets an IO error reading an indirect block
396  *              (ditto, *@err == -EIO)
397  *      or when it reads all @depth-1 indirect blocks successfully and finds
398  *      the whole chain, all way to the data (returns %NULL, *err == 0).
399  *
400  *      Need to be called with
401  *      down_read(&EXT4_I(inode)->i_data_sem)
402  */
403 static Indirect *ext4_get_branch(struct inode *inode, int depth,
404                                  ext4_lblk_t  *offsets,
405                                  Indirect chain[4], int *err)
406 {
407         struct super_block *sb = inode->i_sb;
408         Indirect *p = chain;
409         struct buffer_head *bh;
410
411         *err = 0;
412         /* i_data is not going away, no lock needed */
413         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
414         if (!p->key)
415                 goto no_block;
416         while (--depth) {
417                 bh = sb_getblk(sb, le32_to_cpu(p->key));
418                 if (unlikely(!bh))
419                         goto failure;
420
421                 if (!bh_uptodate_or_lock(bh)) {
422                         if (bh_submit_read(bh) < 0) {
423                                 put_bh(bh);
424                                 goto failure;
425                         }
426                         /* validate block references */
427                         if (ext4_check_indirect_blockref(inode, bh)) {
428                                 put_bh(bh);
429                                 goto failure;
430                         }
431                 }
432
433                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
434                 /* Reader: end */
435                 if (!p->key)
436                         goto no_block;
437         }
438         return NULL;
439
440 failure:
441         *err = -EIO;
442 no_block:
443         return p;
444 }
445
446 /**
447  *      ext4_find_near - find a place for allocation with sufficient locality
448  *      @inode: owner
449  *      @ind: descriptor of indirect block.
450  *
451  *      This function returns the preferred place for block allocation.
452  *      It is used when heuristic for sequential allocation fails.
453  *      Rules are:
454  *        + if there is a block to the left of our position - allocate near it.
455  *        + if pointer will live in indirect block - allocate near that block.
456  *        + if pointer will live in inode - allocate in the same
457  *          cylinder group.
458  *
459  * In the latter case we colour the starting block by the callers PID to
460  * prevent it from clashing with concurrent allocations for a different inode
461  * in the same block group.   The PID is used here so that functionally related
462  * files will be close-by on-disk.
463  *
464  *      Caller must make sure that @ind is valid and will stay that way.
465  */
466 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
467 {
468         struct ext4_inode_info *ei = EXT4_I(inode);
469         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
470         __le32 *p;
471         ext4_fsblk_t bg_start;
472         ext4_fsblk_t last_block;
473         ext4_grpblk_t colour;
474         ext4_group_t block_group;
475         int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
476
477         /* Try to find previous block */
478         for (p = ind->p - 1; p >= start; p--) {
479                 if (*p)
480                         return le32_to_cpu(*p);
481         }
482
483         /* No such thing, so let's try location of indirect block */
484         if (ind->bh)
485                 return ind->bh->b_blocknr;
486
487         /*
488          * It is going to be referred to from the inode itself? OK, just put it
489          * into the same cylinder group then.
490          */
491         block_group = ei->i_block_group;
492         if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
493                 block_group &= ~(flex_size-1);
494                 if (S_ISREG(inode->i_mode))
495                         block_group++;
496         }
497         bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
498         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
499
500         /*
501          * If we are doing delayed allocation, we don't need take
502          * colour into account.
503          */
504         if (test_opt(inode->i_sb, DELALLOC))
505                 return bg_start;
506
507         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
508                 colour = (current->pid % 16) *
509                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
510         else
511                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
512         return bg_start + colour;
513 }
514
515 /**
516  *      ext4_find_goal - find a preferred place for allocation.
517  *      @inode: owner
518  *      @block:  block we want
519  *      @partial: pointer to the last triple within a chain
520  *
521  *      Normally this function find the preferred place for block allocation,
522  *      returns it.
523  *      Because this is only used for non-extent files, we limit the block nr
524  *      to 32 bits.
525  */
526 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
527                                    Indirect *partial)
528 {
529         ext4_fsblk_t goal;
530
531         /*
532          * XXX need to get goal block from mballoc's data structures
533          */
534
535         goal = ext4_find_near(inode, partial);
536         goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
537         return goal;
538 }
539
540 /**
541  *      ext4_blks_to_allocate: Look up the block map and count the number
542  *      of direct blocks need to be allocated for the given branch.
543  *
544  *      @branch: chain of indirect blocks
545  *      @k: number of blocks need for indirect blocks
546  *      @blks: number of data blocks to be mapped.
547  *      @blocks_to_boundary:  the offset in the indirect block
548  *
549  *      return the total number of blocks to be allocate, including the
550  *      direct and indirect blocks.
551  */
552 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
553                                  int blocks_to_boundary)
554 {
555         unsigned int count = 0;
556
557         /*
558          * Simple case, [t,d]Indirect block(s) has not allocated yet
559          * then it's clear blocks on that path have not allocated
560          */
561         if (k > 0) {
562                 /* right now we don't handle cross boundary allocation */
563                 if (blks < blocks_to_boundary + 1)
564                         count += blks;
565                 else
566                         count += blocks_to_boundary + 1;
567                 return count;
568         }
569
570         count++;
571         while (count < blks && count <= blocks_to_boundary &&
572                 le32_to_cpu(*(branch[0].p + count)) == 0) {
573                 count++;
574         }
575         return count;
576 }
577
578 /**
579  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
580  *      @indirect_blks: the number of blocks need to allocate for indirect
581  *                      blocks
582  *
583  *      @new_blocks: on return it will store the new block numbers for
584  *      the indirect blocks(if needed) and the first direct block,
585  *      @blks:  on return it will store the total number of allocated
586  *              direct blocks
587  */
588 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
589                              ext4_lblk_t iblock, ext4_fsblk_t goal,
590                              int indirect_blks, int blks,
591                              ext4_fsblk_t new_blocks[4], int *err)
592 {
593         struct ext4_allocation_request ar;
594         int target, i;
595         unsigned long count = 0, blk_allocated = 0;
596         int index = 0;
597         ext4_fsblk_t current_block = 0;
598         int ret = 0;
599
600         /*
601          * Here we try to allocate the requested multiple blocks at once,
602          * on a best-effort basis.
603          * To build a branch, we should allocate blocks for
604          * the indirect blocks(if not allocated yet), and at least
605          * the first direct block of this branch.  That's the
606          * minimum number of blocks need to allocate(required)
607          */
608         /* first we try to allocate the indirect blocks */
609         target = indirect_blks;
610         while (target > 0) {
611                 count = target;
612                 /* allocating blocks for indirect blocks and direct blocks */
613                 current_block = ext4_new_meta_blocks(handle, inode,
614                                                         goal, &count, err);
615                 if (*err)
616                         goto failed_out;
617
618                 if (unlikely(current_block + count > EXT4_MAX_BLOCK_FILE_PHYS)) {
619                         EXT4_ERROR_INODE(inode,
620                                          "current_block %llu + count %lu > %d!",
621                                          current_block, count,
622                                          EXT4_MAX_BLOCK_FILE_PHYS);
623                         *err = -EIO;
624                         goto failed_out;
625                 }
626
627                 target -= count;
628                 /* allocate blocks for indirect blocks */
629                 while (index < indirect_blks && count) {
630                         new_blocks[index++] = current_block++;
631                         count--;
632                 }
633                 if (count > 0) {
634                         /*
635                          * save the new block number
636                          * for the first direct block
637                          */
638                         new_blocks[index] = current_block;
639                         printk(KERN_INFO "%s returned more blocks than "
640                                                 "requested\n", __func__);
641                         WARN_ON(1);
642                         break;
643                 }
644         }
645
646         target = blks - count ;
647         blk_allocated = count;
648         if (!target)
649                 goto allocated;
650         /* Now allocate data blocks */
651         memset(&ar, 0, sizeof(ar));
652         ar.inode = inode;
653         ar.goal = goal;
654         ar.len = target;
655         ar.logical = iblock;
656         if (S_ISREG(inode->i_mode))
657                 /* enable in-core preallocation only for regular files */
658                 ar.flags = EXT4_MB_HINT_DATA;
659
660         current_block = ext4_mb_new_blocks(handle, &ar, err);
661         if (unlikely(current_block + ar.len > EXT4_MAX_BLOCK_FILE_PHYS)) {
662                 EXT4_ERROR_INODE(inode,
663                                  "current_block %llu + ar.len %d > %d!",
664                                  current_block, ar.len,
665                                  EXT4_MAX_BLOCK_FILE_PHYS);
666                 *err = -EIO;
667                 goto failed_out;
668         }
669
670         if (*err && (target == blks)) {
671                 /*
672                  * if the allocation failed and we didn't allocate
673                  * any blocks before
674                  */
675                 goto failed_out;
676         }
677         if (!*err) {
678                 if (target == blks) {
679                         /*
680                          * save the new block number
681                          * for the first direct block
682                          */
683                         new_blocks[index] = current_block;
684                 }
685                 blk_allocated += ar.len;
686         }
687 allocated:
688         /* total number of blocks allocated for direct blocks */
689         ret = blk_allocated;
690         *err = 0;
691         return ret;
692 failed_out:
693         for (i = 0; i < index; i++)
694                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
695         return ret;
696 }
697
698 /**
699  *      ext4_alloc_branch - allocate and set up a chain of blocks.
700  *      @inode: owner
701  *      @indirect_blks: number of allocated indirect blocks
702  *      @blks: number of allocated direct blocks
703  *      @offsets: offsets (in the blocks) to store the pointers to next.
704  *      @branch: place to store the chain in.
705  *
706  *      This function allocates blocks, zeroes out all but the last one,
707  *      links them into chain and (if we are synchronous) writes them to disk.
708  *      In other words, it prepares a branch that can be spliced onto the
709  *      inode. It stores the information about that chain in the branch[], in
710  *      the same format as ext4_get_branch() would do. We are calling it after
711  *      we had read the existing part of chain and partial points to the last
712  *      triple of that (one with zero ->key). Upon the exit we have the same
713  *      picture as after the successful ext4_get_block(), except that in one
714  *      place chain is disconnected - *branch->p is still zero (we did not
715  *      set the last link), but branch->key contains the number that should
716  *      be placed into *branch->p to fill that gap.
717  *
718  *      If allocation fails we free all blocks we've allocated (and forget
719  *      their buffer_heads) and return the error value the from failed
720  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
721  *      as described above and return 0.
722  */
723 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
724                              ext4_lblk_t iblock, int indirect_blks,
725                              int *blks, ext4_fsblk_t goal,
726                              ext4_lblk_t *offsets, Indirect *branch)
727 {
728         int blocksize = inode->i_sb->s_blocksize;
729         int i, n = 0;
730         int err = 0;
731         struct buffer_head *bh;
732         int num;
733         ext4_fsblk_t new_blocks[4];
734         ext4_fsblk_t current_block;
735
736         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
737                                 *blks, new_blocks, &err);
738         if (err)
739                 return err;
740
741         branch[0].key = cpu_to_le32(new_blocks[0]);
742         /*
743          * metadata blocks and data blocks are allocated.
744          */
745         for (n = 1; n <= indirect_blks;  n++) {
746                 /*
747                  * Get buffer_head for parent block, zero it out
748                  * and set the pointer to new one, then send
749                  * parent to disk.
750                  */
751                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
752                 branch[n].bh = bh;
753                 lock_buffer(bh);
754                 BUFFER_TRACE(bh, "call get_create_access");
755                 err = ext4_journal_get_create_access(handle, bh);
756                 if (err) {
757                         /* Don't brelse(bh) here; it's done in
758                          * ext4_journal_forget() below */
759                         unlock_buffer(bh);
760                         goto failed;
761                 }
762
763                 memset(bh->b_data, 0, blocksize);
764                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
765                 branch[n].key = cpu_to_le32(new_blocks[n]);
766                 *branch[n].p = branch[n].key;
767                 if (n == indirect_blks) {
768                         current_block = new_blocks[n];
769                         /*
770                          * End of chain, update the last new metablock of
771                          * the chain to point to the new allocated
772                          * data blocks numbers
773                          */
774                         for (i = 1; i < num; i++)
775                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
776                 }
777                 BUFFER_TRACE(bh, "marking uptodate");
778                 set_buffer_uptodate(bh);
779                 unlock_buffer(bh);
780
781                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
782                 err = ext4_handle_dirty_metadata(handle, inode, bh);
783                 if (err)
784                         goto failed;
785         }
786         *blks = num;
787         return err;
788 failed:
789         /* Allocation failed, free what we already allocated */
790         ext4_free_blocks(handle, inode, 0, new_blocks[0], 1, 0);
791         for (i = 1; i <= n ; i++) {
792                 /*
793                  * branch[i].bh is newly allocated, so there is no
794                  * need to revoke the block, which is why we don't
795                  * need to set EXT4_FREE_BLOCKS_METADATA.
796                  */
797                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1,
798                                  EXT4_FREE_BLOCKS_FORGET);
799         }
800         for (i = n+1; i < indirect_blks; i++)
801                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
802
803         ext4_free_blocks(handle, inode, 0, new_blocks[i], num, 0);
804
805         return err;
806 }
807
808 /**
809  * ext4_splice_branch - splice the allocated branch onto inode.
810  * @inode: owner
811  * @block: (logical) number of block we are adding
812  * @chain: chain of indirect blocks (with a missing link - see
813  *      ext4_alloc_branch)
814  * @where: location of missing link
815  * @num:   number of indirect blocks we are adding
816  * @blks:  number of direct blocks we are adding
817  *
818  * This function fills the missing link and does all housekeeping needed in
819  * inode (->i_blocks, etc.). In case of success we end up with the full
820  * chain to new block and return 0.
821  */
822 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
823                               ext4_lblk_t block, Indirect *where, int num,
824                               int blks)
825 {
826         int i;
827         int err = 0;
828         ext4_fsblk_t current_block;
829
830         /*
831          * If we're splicing into a [td]indirect block (as opposed to the
832          * inode) then we need to get write access to the [td]indirect block
833          * before the splice.
834          */
835         if (where->bh) {
836                 BUFFER_TRACE(where->bh, "get_write_access");
837                 err = ext4_journal_get_write_access(handle, where->bh);
838                 if (err)
839                         goto err_out;
840         }
841         /* That's it */
842
843         *where->p = where->key;
844
845         /*
846          * Update the host buffer_head or inode to point to more just allocated
847          * direct blocks blocks
848          */
849         if (num == 0 && blks > 1) {
850                 current_block = le32_to_cpu(where->key) + 1;
851                 for (i = 1; i < blks; i++)
852                         *(where->p + i) = cpu_to_le32(current_block++);
853         }
854
855         /* We are done with atomic stuff, now do the rest of housekeeping */
856         /* had we spliced it onto indirect block? */
857         if (where->bh) {
858                 /*
859                  * If we spliced it onto an indirect block, we haven't
860                  * altered the inode.  Note however that if it is being spliced
861                  * onto an indirect block at the very end of the file (the
862                  * file is growing) then we *will* alter the inode to reflect
863                  * the new i_size.  But that is not done here - it is done in
864                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
865                  */
866                 jbd_debug(5, "splicing indirect only\n");
867                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
868                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
869                 if (err)
870                         goto err_out;
871         } else {
872                 /*
873                  * OK, we spliced it into the inode itself on a direct block.
874                  */
875                 ext4_mark_inode_dirty(handle, inode);
876                 jbd_debug(5, "splicing direct\n");
877         }
878         return err;
879
880 err_out:
881         for (i = 1; i <= num; i++) {
882                 /*
883                  * branch[i].bh is newly allocated, so there is no
884                  * need to revoke the block, which is why we don't
885                  * need to set EXT4_FREE_BLOCKS_METADATA.
886                  */
887                 ext4_free_blocks(handle, inode, where[i].bh, 0, 1,
888                                  EXT4_FREE_BLOCKS_FORGET);
889         }
890         ext4_free_blocks(handle, inode, 0, le32_to_cpu(where[num].key),
891                          blks, 0);
892
893         return err;
894 }
895
896 /*
897  * The ext4_ind_map_blocks() function handles non-extents inodes
898  * (i.e., using the traditional indirect/double-indirect i_blocks
899  * scheme) for ext4_map_blocks().
900  *
901  * Allocation strategy is simple: if we have to allocate something, we will
902  * have to go the whole way to leaf. So let's do it before attaching anything
903  * to tree, set linkage between the newborn blocks, write them if sync is
904  * required, recheck the path, free and repeat if check fails, otherwise
905  * set the last missing link (that will protect us from any truncate-generated
906  * removals - all blocks on the path are immune now) and possibly force the
907  * write on the parent block.
908  * That has a nice additional property: no special recovery from the failed
909  * allocations is needed - we simply release blocks and do not touch anything
910  * reachable from inode.
911  *
912  * `handle' can be NULL if create == 0.
913  *
914  * return > 0, # of blocks mapped or allocated.
915  * return = 0, if plain lookup failed.
916  * return < 0, error case.
917  *
918  * The ext4_ind_get_blocks() function should be called with
919  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
920  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
921  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
922  * blocks.
923  */
924 static int ext4_ind_map_blocks(handle_t *handle, struct inode *inode,
925                                struct ext4_map_blocks *map,
926                                int flags)
927 {
928         int err = -EIO;
929         ext4_lblk_t offsets[4];
930         Indirect chain[4];
931         Indirect *partial;
932         ext4_fsblk_t goal;
933         int indirect_blks;
934         int blocks_to_boundary = 0;
935         int depth;
936         int count = 0;
937         ext4_fsblk_t first_block = 0;
938
939         J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
940         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
941         depth = ext4_block_to_path(inode, map->m_lblk, offsets,
942                                    &blocks_to_boundary);
943
944         if (depth == 0)
945                 goto out;
946
947         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
948
949         /* Simplest case - block found, no allocation needed */
950         if (!partial) {
951                 first_block = le32_to_cpu(chain[depth - 1].key);
952                 count++;
953                 /*map more blocks*/
954                 while (count < map->m_len && count <= blocks_to_boundary) {
955                         ext4_fsblk_t blk;
956
957                         blk = le32_to_cpu(*(chain[depth-1].p + count));
958
959                         if (blk == first_block + count)
960                                 count++;
961                         else
962                                 break;
963                 }
964                 goto got_it;
965         }
966
967         /* Next simple case - plain lookup or failed read of indirect block */
968         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
969                 goto cleanup;
970
971         /*
972          * Okay, we need to do block allocation.
973         */
974         goal = ext4_find_goal(inode, map->m_lblk, partial);
975
976         /* the number of blocks need to allocate for [d,t]indirect blocks */
977         indirect_blks = (chain + depth) - partial - 1;
978
979         /*
980          * Next look up the indirect map to count the totoal number of
981          * direct blocks to allocate for this branch.
982          */
983         count = ext4_blks_to_allocate(partial, indirect_blks,
984                                       map->m_len, blocks_to_boundary);
985         /*
986          * Block out ext4_truncate while we alter the tree
987          */
988         err = ext4_alloc_branch(handle, inode, map->m_lblk, indirect_blks,
989                                 &count, goal,
990                                 offsets + (partial - chain), partial);
991
992         /*
993          * The ext4_splice_branch call will free and forget any buffers
994          * on the new chain if there is a failure, but that risks using
995          * up transaction credits, especially for bitmaps where the
996          * credits cannot be returned.  Can we handle this somehow?  We
997          * may need to return -EAGAIN upwards in the worst case.  --sct
998          */
999         if (!err)
1000                 err = ext4_splice_branch(handle, inode, map->m_lblk,
1001                                          partial, indirect_blks, count);
1002         if (err)
1003                 goto cleanup;
1004
1005         map->m_flags |= EXT4_MAP_NEW;
1006
1007         ext4_update_inode_fsync_trans(handle, inode, 1);
1008 got_it:
1009         map->m_flags |= EXT4_MAP_MAPPED;
1010         map->m_pblk = le32_to_cpu(chain[depth-1].key);
1011         map->m_len = count;
1012         if (count > blocks_to_boundary)
1013                 map->m_flags |= EXT4_MAP_BOUNDARY;
1014         err = count;
1015         /* Clean up and exit */
1016         partial = chain + depth - 1;    /* the whole chain */
1017 cleanup:
1018         while (partial > chain) {
1019                 BUFFER_TRACE(partial->bh, "call brelse");
1020                 brelse(partial->bh);
1021                 partial--;
1022         }
1023 out:
1024         return err;
1025 }
1026
1027 #ifdef CONFIG_QUOTA
1028 qsize_t *ext4_get_reserved_space(struct inode *inode)
1029 {
1030         return &EXT4_I(inode)->i_reserved_quota;
1031 }
1032 #endif
1033
1034 /*
1035  * Calculate the number of metadata blocks need to reserve
1036  * to allocate a new block at @lblocks for non extent file based file
1037  */
1038 static int ext4_indirect_calc_metadata_amount(struct inode *inode,
1039                                               sector_t lblock)
1040 {
1041         struct ext4_inode_info *ei = EXT4_I(inode);
1042         sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
1043         int blk_bits;
1044
1045         if (lblock < EXT4_NDIR_BLOCKS)
1046                 return 0;
1047
1048         lblock -= EXT4_NDIR_BLOCKS;
1049
1050         if (ei->i_da_metadata_calc_len &&
1051             (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
1052                 ei->i_da_metadata_calc_len++;
1053                 return 0;
1054         }
1055         ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
1056         ei->i_da_metadata_calc_len = 1;
1057         blk_bits = order_base_2(lblock);
1058         return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
1059 }
1060
1061 /*
1062  * Calculate the number of metadata blocks need to reserve
1063  * to allocate a block located at @lblock
1064  */
1065 static int ext4_calc_metadata_amount(struct inode *inode, sector_t lblock)
1066 {
1067         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1068                 return ext4_ext_calc_metadata_amount(inode, lblock);
1069
1070         return ext4_indirect_calc_metadata_amount(inode, lblock);
1071 }
1072
1073 /*
1074  * Called with i_data_sem down, which is important since we can call
1075  * ext4_discard_preallocations() from here.
1076  */
1077 void ext4_da_update_reserve_space(struct inode *inode,
1078                                         int used, int quota_claim)
1079 {
1080         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1081         struct ext4_inode_info *ei = EXT4_I(inode);
1082
1083         spin_lock(&ei->i_block_reservation_lock);
1084         trace_ext4_da_update_reserve_space(inode, used);
1085         if (unlikely(used > ei->i_reserved_data_blocks)) {
1086                 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
1087                          "with only %d reserved data blocks\n",
1088                          __func__, inode->i_ino, used,
1089                          ei->i_reserved_data_blocks);
1090                 WARN_ON(1);
1091                 used = ei->i_reserved_data_blocks;
1092         }
1093
1094         /* Update per-inode reservations */
1095         ei->i_reserved_data_blocks -= used;
1096         ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
1097         percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1098                            used + ei->i_allocated_meta_blocks);
1099         ei->i_allocated_meta_blocks = 0;
1100
1101         if (ei->i_reserved_data_blocks == 0) {
1102                 /*
1103                  * We can release all of the reserved metadata blocks
1104                  * only when we have written all of the delayed
1105                  * allocation blocks.
1106                  */
1107                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1108                                    ei->i_reserved_meta_blocks);
1109                 ei->i_reserved_meta_blocks = 0;
1110                 ei->i_da_metadata_calc_len = 0;
1111         }
1112         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1113
1114         /* Update quota subsystem for data blocks */
1115         if (quota_claim)
1116                 dquot_claim_block(inode, used);
1117         else {
1118                 /*
1119                  * We did fallocate with an offset that is already delayed
1120                  * allocated. So on delayed allocated writeback we should
1121                  * not re-claim the quota for fallocated blocks.
1122                  */
1123                 dquot_release_reservation_block(inode, used);
1124         }
1125
1126         /*
1127          * If we have done all the pending block allocations and if
1128          * there aren't any writers on the inode, we can discard the
1129          * inode's preallocations.
1130          */
1131         if ((ei->i_reserved_data_blocks == 0) &&
1132             (atomic_read(&inode->i_writecount) == 0))
1133                 ext4_discard_preallocations(inode);
1134 }
1135
1136 static int __check_block_validity(struct inode *inode, const char *func,
1137                                 unsigned int line,
1138                                 struct ext4_map_blocks *map)
1139 {
1140         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
1141                                    map->m_len)) {
1142                 ext4_error_inode(inode, func, line, map->m_pblk,
1143                                  "lblock %lu mapped to illegal pblock "
1144                                  "(length %d)", (unsigned long) map->m_lblk,
1145                                  map->m_len);
1146                 return -EIO;
1147         }
1148         return 0;
1149 }
1150
1151 #define check_block_validity(inode, map)        \
1152         __check_block_validity((inode), __func__, __LINE__, (map))
1153
1154 /*
1155  * Return the number of contiguous dirty pages in a given inode
1156  * starting at page frame idx.
1157  */
1158 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
1159                                     unsigned int max_pages)
1160 {
1161         struct address_space *mapping = inode->i_mapping;
1162         pgoff_t index;
1163         struct pagevec pvec;
1164         pgoff_t num = 0;
1165         int i, nr_pages, done = 0;
1166
1167         if (max_pages == 0)
1168                 return 0;
1169         pagevec_init(&pvec, 0);
1170         while (!done) {
1171                 index = idx;
1172                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
1173                                               PAGECACHE_TAG_DIRTY,
1174                                               (pgoff_t)PAGEVEC_SIZE);
1175                 if (nr_pages == 0)
1176                         break;
1177                 for (i = 0; i < nr_pages; i++) {
1178                         struct page *page = pvec.pages[i];
1179                         struct buffer_head *bh, *head;
1180
1181                         lock_page(page);
1182                         if (unlikely(page->mapping != mapping) ||
1183                             !PageDirty(page) ||
1184                             PageWriteback(page) ||
1185                             page->index != idx) {
1186                                 done = 1;
1187                                 unlock_page(page);
1188                                 break;
1189                         }
1190                         if (page_has_buffers(page)) {
1191                                 bh = head = page_buffers(page);
1192                                 do {
1193                                         if (!buffer_delay(bh) &&
1194                                             !buffer_unwritten(bh))
1195                                                 done = 1;
1196                                         bh = bh->b_this_page;
1197                                 } while (!done && (bh != head));
1198                         }
1199                         unlock_page(page);
1200                         if (done)
1201                                 break;
1202                         idx++;
1203                         num++;
1204                         if (num >= max_pages)
1205                                 break;
1206                 }
1207                 pagevec_release(&pvec);
1208         }
1209         return num;
1210 }
1211
1212 /*
1213  * The ext4_map_blocks() function tries to look up the requested blocks,
1214  * and returns if the blocks are already mapped.
1215  *
1216  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1217  * and store the allocated blocks in the result buffer head and mark it
1218  * mapped.
1219  *
1220  * If file type is extents based, it will call ext4_ext_map_blocks(),
1221  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
1222  * based files
1223  *
1224  * On success, it returns the number of blocks being mapped or allocate.
1225  * if create==0 and the blocks are pre-allocated and uninitialized block,
1226  * the result buffer head is unmapped. If the create ==1, it will make sure
1227  * the buffer head is mapped.
1228  *
1229  * It returns 0 if plain look up failed (blocks have not been allocated), in
1230  * that casem, buffer head is unmapped
1231  *
1232  * It returns the error in case of allocation failure.
1233  */
1234 int ext4_map_blocks(handle_t *handle, struct inode *inode,
1235                     struct ext4_map_blocks *map, int flags)
1236 {
1237         int retval;
1238
1239         map->m_flags = 0;
1240         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
1241                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
1242                   (unsigned long) map->m_lblk);
1243         /*
1244          * Try to see if we can get the block without requesting a new
1245          * file system block.
1246          */
1247         down_read((&EXT4_I(inode)->i_data_sem));
1248         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
1249                 retval = ext4_ext_map_blocks(handle, inode, map, 0);
1250         } else {
1251                 retval = ext4_ind_map_blocks(handle, inode, map, 0);
1252         }
1253         up_read((&EXT4_I(inode)->i_data_sem));
1254
1255         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
1256                 int ret = check_block_validity(inode, map);
1257                 if (ret != 0)
1258                         return ret;
1259         }
1260
1261         /* If it is only a block(s) look up */
1262         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1263                 return retval;
1264
1265         /*
1266          * Returns if the blocks have already allocated
1267          *
1268          * Note that if blocks have been preallocated
1269          * ext4_ext_get_block() returns th create = 0
1270          * with buffer head unmapped.
1271          */
1272         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
1273                 return retval;
1274
1275         /*
1276          * When we call get_blocks without the create flag, the
1277          * BH_Unwritten flag could have gotten set if the blocks
1278          * requested were part of a uninitialized extent.  We need to
1279          * clear this flag now that we are committed to convert all or
1280          * part of the uninitialized extent to be an initialized
1281          * extent.  This is because we need to avoid the combination
1282          * of BH_Unwritten and BH_Mapped flags being simultaneously
1283          * set on the buffer_head.
1284          */
1285         map->m_flags &= ~EXT4_MAP_UNWRITTEN;
1286
1287         /*
1288          * New blocks allocate and/or writing to uninitialized extent
1289          * will possibly result in updating i_data, so we take
1290          * the write lock of i_data_sem, and call get_blocks()
1291          * with create == 1 flag.
1292          */
1293         down_write((&EXT4_I(inode)->i_data_sem));
1294
1295         /*
1296          * if the caller is from delayed allocation writeout path
1297          * we have already reserved fs blocks for allocation
1298          * let the underlying get_block() function know to
1299          * avoid double accounting
1300          */
1301         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1302                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1303         /*
1304          * We need to check for EXT4 here because migrate
1305          * could have changed the inode type in between
1306          */
1307         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
1308                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
1309         } else {
1310                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
1311
1312                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
1313                         /*
1314                          * We allocated new blocks which will result in
1315                          * i_data's format changing.  Force the migrate
1316                          * to fail by clearing migrate flags
1317                          */
1318                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
1319                 }
1320
1321                 /*
1322                  * Update reserved blocks/metadata blocks after successful
1323                  * block allocation which had been deferred till now. We don't
1324                  * support fallocate for non extent files. So we can update
1325                  * reserve space here.
1326                  */
1327                 if ((retval > 0) &&
1328                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
1329                         ext4_da_update_reserve_space(inode, retval, 1);
1330         }
1331         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1332                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1333
1334         up_write((&EXT4_I(inode)->i_data_sem));
1335         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
1336                 int ret = check_block_validity(inode, map);
1337                 if (ret != 0)
1338                         return ret;
1339         }
1340         return retval;
1341 }
1342
1343 /* Maximum number of blocks we map for direct IO at once. */
1344 #define DIO_MAX_BLOCKS 4096
1345
1346 static int _ext4_get_block(struct inode *inode, sector_t iblock,
1347                            struct buffer_head *bh, int flags)
1348 {
1349         handle_t *handle = ext4_journal_current_handle();
1350         struct ext4_map_blocks map;
1351         int ret = 0, started = 0;
1352         int dio_credits;
1353
1354         map.m_lblk = iblock;
1355         map.m_len = bh->b_size >> inode->i_blkbits;
1356
1357         if (flags && !handle) {
1358                 /* Direct IO write... */
1359                 if (map.m_len > DIO_MAX_BLOCKS)
1360                         map.m_len = DIO_MAX_BLOCKS;
1361                 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
1362                 handle = ext4_journal_start(inode, dio_credits);
1363                 if (IS_ERR(handle)) {
1364                         ret = PTR_ERR(handle);
1365                         return ret;
1366                 }
1367                 started = 1;
1368         }
1369
1370         ret = ext4_map_blocks(handle, inode, &map, flags);
1371         if (ret > 0) {
1372                 map_bh(bh, inode->i_sb, map.m_pblk);
1373                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
1374                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
1375                 ret = 0;
1376         }
1377         if (started)
1378                 ext4_journal_stop(handle);
1379         return ret;
1380 }
1381
1382 int ext4_get_block(struct inode *inode, sector_t iblock,
1383                    struct buffer_head *bh, int create)
1384 {
1385         return _ext4_get_block(inode, iblock, bh,
1386                                create ? EXT4_GET_BLOCKS_CREATE : 0);
1387 }
1388
1389 /*
1390  * `handle' can be NULL if create is zero
1391  */
1392 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1393                                 ext4_lblk_t block, int create, int *errp)
1394 {
1395         struct ext4_map_blocks map;
1396         struct buffer_head *bh;
1397         int fatal = 0, err;
1398
1399         J_ASSERT(handle != NULL || create == 0);
1400
1401         map.m_lblk = block;
1402         map.m_len = 1;
1403         err = ext4_map_blocks(handle, inode, &map,
1404                               create ? EXT4_GET_BLOCKS_CREATE : 0);
1405
1406         if (err < 0)
1407                 *errp = err;
1408         if (err <= 0)
1409                 return NULL;
1410         *errp = 0;
1411
1412         bh = sb_getblk(inode->i_sb, map.m_pblk);
1413         if (!bh) {
1414                 *errp = -EIO;
1415                 return NULL;
1416         }
1417         if (map.m_flags & EXT4_MAP_NEW) {
1418                 J_ASSERT(create != 0);
1419                 J_ASSERT(handle != NULL);
1420
1421                 /*
1422                  * Now that we do not always journal data, we should
1423                  * keep in mind whether this should always journal the
1424                  * new buffer as metadata.  For now, regular file
1425                  * writes use ext4_get_block instead, so it's not a
1426                  * problem.
1427                  */
1428                 lock_buffer(bh);
1429                 BUFFER_TRACE(bh, "call get_create_access");
1430                 fatal = ext4_journal_get_create_access(handle, bh);
1431                 if (!fatal && !buffer_uptodate(bh)) {
1432                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1433                         set_buffer_uptodate(bh);
1434                 }
1435                 unlock_buffer(bh);
1436                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1437                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1438                 if (!fatal)
1439                         fatal = err;
1440         } else {
1441                 BUFFER_TRACE(bh, "not a new buffer");
1442         }
1443         if (fatal) {
1444                 *errp = fatal;
1445                 brelse(bh);
1446                 bh = NULL;
1447         }
1448         return bh;
1449 }
1450
1451 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1452                                ext4_lblk_t block, int create, int *err)
1453 {
1454         struct buffer_head *bh;
1455
1456         bh = ext4_getblk(handle, inode, block, create, err);
1457         if (!bh)
1458                 return bh;
1459         if (buffer_uptodate(bh))
1460                 return bh;
1461         ll_rw_block(READ_META, 1, &bh);
1462         wait_on_buffer(bh);
1463         if (buffer_uptodate(bh))
1464                 return bh;
1465         put_bh(bh);
1466         *err = -EIO;
1467         return NULL;
1468 }
1469
1470 static int walk_page_buffers(handle_t *handle,
1471                              struct buffer_head *head,
1472                              unsigned from,
1473                              unsigned to,
1474                              int *partial,
1475                              int (*fn)(handle_t *handle,
1476                                        struct buffer_head *bh))
1477 {
1478         struct buffer_head *bh;
1479         unsigned block_start, block_end;
1480         unsigned blocksize = head->b_size;
1481         int err, ret = 0;
1482         struct buffer_head *next;
1483
1484         for (bh = head, block_start = 0;
1485              ret == 0 && (bh != head || !block_start);
1486              block_start = block_end, bh = next) {
1487                 next = bh->b_this_page;
1488                 block_end = block_start + blocksize;
1489                 if (block_end <= from || block_start >= to) {
1490                         if (partial && !buffer_uptodate(bh))
1491                                 *partial = 1;
1492                         continue;
1493                 }
1494                 err = (*fn)(handle, bh);
1495                 if (!ret)
1496                         ret = err;
1497         }
1498         return ret;
1499 }
1500
1501 /*
1502  * To preserve ordering, it is essential that the hole instantiation and
1503  * the data write be encapsulated in a single transaction.  We cannot
1504  * close off a transaction and start a new one between the ext4_get_block()
1505  * and the commit_write().  So doing the jbd2_journal_start at the start of
1506  * prepare_write() is the right place.
1507  *
1508  * Also, this function can nest inside ext4_writepage() ->
1509  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1510  * has generated enough buffer credits to do the whole page.  So we won't
1511  * block on the journal in that case, which is good, because the caller may
1512  * be PF_MEMALLOC.
1513  *
1514  * By accident, ext4 can be reentered when a transaction is open via
1515  * quota file writes.  If we were to commit the transaction while thus
1516  * reentered, there can be a deadlock - we would be holding a quota
1517  * lock, and the commit would never complete if another thread had a
1518  * transaction open and was blocking on the quota lock - a ranking
1519  * violation.
1520  *
1521  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1522  * will _not_ run commit under these circumstances because handle->h_ref
1523  * is elevated.  We'll still have enough credits for the tiny quotafile
1524  * write.
1525  */
1526 static int do_journal_get_write_access(handle_t *handle,
1527                                        struct buffer_head *bh)
1528 {
1529         if (!buffer_mapped(bh) || buffer_freed(bh))
1530                 return 0;
1531         return ext4_journal_get_write_access(handle, bh);
1532 }
1533
1534 /*
1535  * Truncate blocks that were not used by write. We have to truncate the
1536  * pagecache as well so that corresponding buffers get properly unmapped.
1537  */
1538 static void ext4_truncate_failed_write(struct inode *inode)
1539 {
1540         truncate_inode_pages(inode->i_mapping, inode->i_size);
1541         ext4_truncate(inode);
1542 }
1543
1544 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
1545                    struct buffer_head *bh_result, int create);
1546 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1547                             loff_t pos, unsigned len, unsigned flags,
1548                             struct page **pagep, void **fsdata)
1549 {
1550         struct inode *inode = mapping->host;
1551         int ret, needed_blocks;
1552         handle_t *handle;
1553         int retries = 0;
1554         struct page *page;
1555         pgoff_t index;
1556         unsigned from, to;
1557
1558         trace_ext4_write_begin(inode, pos, len, flags);
1559         /*
1560          * Reserve one block more for addition to orphan list in case
1561          * we allocate blocks but write fails for some reason
1562          */
1563         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1564         index = pos >> PAGE_CACHE_SHIFT;
1565         from = pos & (PAGE_CACHE_SIZE - 1);
1566         to = from + len;
1567
1568 retry:
1569         handle = ext4_journal_start(inode, needed_blocks);
1570         if (IS_ERR(handle)) {
1571                 ret = PTR_ERR(handle);
1572                 goto out;
1573         }
1574
1575         /* We cannot recurse into the filesystem as the transaction is already
1576          * started */
1577         flags |= AOP_FLAG_NOFS;
1578
1579         page = grab_cache_page_write_begin(mapping, index, flags);
1580         if (!page) {
1581                 ext4_journal_stop(handle);
1582                 ret = -ENOMEM;
1583                 goto out;
1584         }
1585         *pagep = page;
1586
1587         if (ext4_should_dioread_nolock(inode))
1588                 ret = block_write_begin(file, mapping, pos, len, flags, pagep,
1589                                 fsdata, ext4_get_block_write);
1590         else
1591                 ret = block_write_begin(file, mapping, pos, len, flags, pagep,
1592                                 fsdata, ext4_get_block);
1593
1594         if (!ret && ext4_should_journal_data(inode)) {
1595                 ret = walk_page_buffers(handle, page_buffers(page),
1596                                 from, to, NULL, do_journal_get_write_access);
1597         }
1598
1599         if (ret) {
1600                 unlock_page(page);
1601                 page_cache_release(page);
1602                 /*
1603                  * block_write_begin may have instantiated a few blocks
1604                  * outside i_size.  Trim these off again. Don't need
1605                  * i_size_read because we hold i_mutex.
1606                  *
1607                  * Add inode to orphan list in case we crash before
1608                  * truncate finishes
1609                  */
1610                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1611                         ext4_orphan_add(handle, inode);
1612
1613                 ext4_journal_stop(handle);
1614                 if (pos + len > inode->i_size) {
1615                         ext4_truncate_failed_write(inode);
1616                         /*
1617                          * If truncate failed early the inode might
1618                          * still be on the orphan list; we need to
1619                          * make sure the inode is removed from the
1620                          * orphan list in that case.
1621                          */
1622                         if (inode->i_nlink)
1623                                 ext4_orphan_del(NULL, inode);
1624                 }
1625         }
1626
1627         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1628                 goto retry;
1629 out:
1630         return ret;
1631 }
1632
1633 /* For write_end() in data=journal mode */
1634 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1635 {
1636         if (!buffer_mapped(bh) || buffer_freed(bh))
1637                 return 0;
1638         set_buffer_uptodate(bh);
1639         return ext4_handle_dirty_metadata(handle, NULL, bh);
1640 }
1641
1642 static int ext4_generic_write_end(struct file *file,
1643                                   struct address_space *mapping,
1644                                   loff_t pos, unsigned len, unsigned copied,
1645                                   struct page *page, void *fsdata)
1646 {
1647         int i_size_changed = 0;
1648         struct inode *inode = mapping->host;
1649         handle_t *handle = ext4_journal_current_handle();
1650
1651         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1652
1653         /*
1654          * No need to use i_size_read() here, the i_size
1655          * cannot change under us because we hold i_mutex.
1656          *
1657          * But it's important to update i_size while still holding page lock:
1658          * page writeout could otherwise come in and zero beyond i_size.
1659          */
1660         if (pos + copied > inode->i_size) {
1661                 i_size_write(inode, pos + copied);
1662                 i_size_changed = 1;
1663         }
1664
1665         if (pos + copied >  EXT4_I(inode)->i_disksize) {
1666                 /* We need to mark inode dirty even if
1667                  * new_i_size is less that inode->i_size
1668                  * bu greater than i_disksize.(hint delalloc)
1669                  */
1670                 ext4_update_i_disksize(inode, (pos + copied));
1671                 i_size_changed = 1;
1672         }
1673         unlock_page(page);
1674         page_cache_release(page);
1675
1676         /*
1677          * Don't mark the inode dirty under page lock. First, it unnecessarily
1678          * makes the holding time of page lock longer. Second, it forces lock
1679          * ordering of page lock and transaction start for journaling
1680          * filesystems.
1681          */
1682         if (i_size_changed)
1683                 ext4_mark_inode_dirty(handle, inode);
1684
1685         return copied;
1686 }
1687
1688 /*
1689  * We need to pick up the new inode size which generic_commit_write gave us
1690  * `file' can be NULL - eg, when called from page_symlink().
1691  *
1692  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1693  * buffers are managed internally.
1694  */
1695 static int ext4_ordered_write_end(struct file *file,
1696                                   struct address_space *mapping,
1697                                   loff_t pos, unsigned len, unsigned copied,
1698                                   struct page *page, void *fsdata)
1699 {
1700         handle_t *handle = ext4_journal_current_handle();
1701         struct inode *inode = mapping->host;
1702         int ret = 0, ret2;
1703
1704         trace_ext4_ordered_write_end(inode, pos, len, copied);
1705         ret = ext4_jbd2_file_inode(handle, inode);
1706
1707         if (ret == 0) {
1708                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1709                                                         page, fsdata);
1710                 copied = ret2;
1711                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1712                         /* if we have allocated more blocks and copied
1713                          * less. We will have blocks allocated outside
1714                          * inode->i_size. So truncate them
1715                          */
1716                         ext4_orphan_add(handle, inode);
1717                 if (ret2 < 0)
1718                         ret = ret2;
1719         }
1720         ret2 = ext4_journal_stop(handle);
1721         if (!ret)
1722                 ret = ret2;
1723
1724         if (pos + len > inode->i_size) {
1725                 ext4_truncate_failed_write(inode);
1726                 /*
1727                  * If truncate failed early the inode might still be
1728                  * on the orphan list; we need to make sure the inode
1729                  * is removed from the orphan list in that case.
1730                  */
1731                 if (inode->i_nlink)
1732                         ext4_orphan_del(NULL, inode);
1733         }
1734
1735
1736         return ret ? ret : copied;
1737 }
1738
1739 static int ext4_writeback_write_end(struct file *file,
1740                                     struct address_space *mapping,
1741                                     loff_t pos, unsigned len, unsigned copied,
1742                                     struct page *page, void *fsdata)
1743 {
1744         handle_t *handle = ext4_journal_current_handle();
1745         struct inode *inode = mapping->host;
1746         int ret = 0, ret2;
1747
1748         trace_ext4_writeback_write_end(inode, pos, len, copied);
1749         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1750                                                         page, fsdata);
1751         copied = ret2;
1752         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1753                 /* if we have allocated more blocks and copied
1754                  * less. We will have blocks allocated outside
1755                  * inode->i_size. So truncate them
1756                  */
1757                 ext4_orphan_add(handle, inode);
1758
1759         if (ret2 < 0)
1760                 ret = ret2;
1761
1762         ret2 = ext4_journal_stop(handle);
1763         if (!ret)
1764                 ret = ret2;
1765
1766         if (pos + len > inode->i_size) {
1767                 ext4_truncate_failed_write(inode);
1768                 /*
1769                  * If truncate failed early the inode might still be
1770                  * on the orphan list; we need to make sure the inode
1771                  * is removed from the orphan list in that case.
1772                  */
1773                 if (inode->i_nlink)
1774                         ext4_orphan_del(NULL, inode);
1775         }
1776
1777         return ret ? ret : copied;
1778 }
1779
1780 static int ext4_journalled_write_end(struct file *file,
1781                                      struct address_space *mapping,
1782                                      loff_t pos, unsigned len, unsigned copied,
1783                                      struct page *page, void *fsdata)
1784 {
1785         handle_t *handle = ext4_journal_current_handle();
1786         struct inode *inode = mapping->host;
1787         int ret = 0, ret2;
1788         int partial = 0;
1789         unsigned from, to;
1790         loff_t new_i_size;
1791
1792         trace_ext4_journalled_write_end(inode, pos, len, copied);
1793         from = pos & (PAGE_CACHE_SIZE - 1);
1794         to = from + len;
1795
1796         if (copied < len) {
1797                 if (!PageUptodate(page))
1798                         copied = 0;
1799                 page_zero_new_buffers(page, from+copied, to);
1800         }
1801
1802         ret = walk_page_buffers(handle, page_buffers(page), from,
1803                                 to, &partial, write_end_fn);
1804         if (!partial)
1805                 SetPageUptodate(page);
1806         new_i_size = pos + copied;
1807         if (new_i_size > inode->i_size)
1808                 i_size_write(inode, pos+copied);
1809         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1810         if (new_i_size > EXT4_I(inode)->i_disksize) {
1811                 ext4_update_i_disksize(inode, new_i_size);
1812                 ret2 = ext4_mark_inode_dirty(handle, inode);
1813                 if (!ret)
1814                         ret = ret2;
1815         }
1816
1817         unlock_page(page);
1818         page_cache_release(page);
1819         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1820                 /* if we have allocated more blocks and copied
1821                  * less. We will have blocks allocated outside
1822                  * inode->i_size. So truncate them
1823                  */
1824                 ext4_orphan_add(handle, inode);
1825
1826         ret2 = ext4_journal_stop(handle);
1827         if (!ret)
1828                 ret = ret2;
1829         if (pos + len > inode->i_size) {
1830                 ext4_truncate_failed_write(inode);
1831                 /*
1832                  * If truncate failed early the inode might still be
1833                  * on the orphan list; we need to make sure the inode
1834                  * is removed from the orphan list in that case.
1835                  */
1836                 if (inode->i_nlink)
1837                         ext4_orphan_del(NULL, inode);
1838         }
1839
1840         return ret ? ret : copied;
1841 }
1842
1843 /*
1844  * Reserve a single block located at lblock
1845  */
1846 static int ext4_da_reserve_space(struct inode *inode, sector_t lblock)
1847 {
1848         int retries = 0;
1849         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1850         struct ext4_inode_info *ei = EXT4_I(inode);
1851         unsigned long md_needed;
1852         int ret;
1853
1854         /*
1855          * recalculate the amount of metadata blocks to reserve
1856          * in order to allocate nrblocks
1857          * worse case is one extent per block
1858          */
1859 repeat:
1860         spin_lock(&ei->i_block_reservation_lock);
1861         md_needed = ext4_calc_metadata_amount(inode, lblock);
1862         trace_ext4_da_reserve_space(inode, md_needed);
1863         spin_unlock(&ei->i_block_reservation_lock);
1864
1865         /*
1866          * We will charge metadata quota at writeout time; this saves
1867          * us from metadata over-estimation, though we may go over by
1868          * a small amount in the end.  Here we just reserve for data.
1869          */
1870         ret = dquot_reserve_block(inode, 1);
1871         if (ret)
1872                 return ret;
1873         /*
1874          * We do still charge estimated metadata to the sb though;
1875          * we cannot afford to run out of free blocks.
1876          */
1877         if (ext4_claim_free_blocks(sbi, md_needed + 1)) {
1878                 dquot_release_reservation_block(inode, 1);
1879                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1880                         yield();
1881                         goto repeat;
1882                 }
1883                 return -ENOSPC;
1884         }
1885         spin_lock(&ei->i_block_reservation_lock);
1886         ei->i_reserved_data_blocks++;
1887         ei->i_reserved_meta_blocks += md_needed;
1888         spin_unlock(&ei->i_block_reservation_lock);
1889
1890         return 0;       /* success */
1891 }
1892
1893 static void ext4_da_release_space(struct inode *inode, int to_free)
1894 {
1895         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1896         struct ext4_inode_info *ei = EXT4_I(inode);
1897
1898         if (!to_free)
1899                 return;         /* Nothing to release, exit */
1900
1901         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1902
1903         trace_ext4_da_release_space(inode, to_free);
1904         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1905                 /*
1906                  * if there aren't enough reserved blocks, then the
1907                  * counter is messed up somewhere.  Since this
1908                  * function is called from invalidate page, it's
1909                  * harmless to return without any action.
1910                  */
1911                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1912                          "ino %lu, to_free %d with only %d reserved "
1913                          "data blocks\n", inode->i_ino, to_free,
1914                          ei->i_reserved_data_blocks);
1915                 WARN_ON(1);
1916                 to_free = ei->i_reserved_data_blocks;
1917         }
1918         ei->i_reserved_data_blocks -= to_free;
1919
1920         if (ei->i_reserved_data_blocks == 0) {
1921                 /*
1922                  * We can release all of the reserved metadata blocks
1923                  * only when we have written all of the delayed
1924                  * allocation blocks.
1925                  */
1926                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1927                                    ei->i_reserved_meta_blocks);
1928                 ei->i_reserved_meta_blocks = 0;
1929                 ei->i_da_metadata_calc_len = 0;
1930         }
1931
1932         /* update fs dirty data blocks counter */
1933         percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1934
1935         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1936
1937         dquot_release_reservation_block(inode, to_free);
1938 }
1939
1940 static void ext4_da_page_release_reservation(struct page *page,
1941                                              unsigned long offset)
1942 {
1943         int to_release = 0;
1944         struct buffer_head *head, *bh;
1945         unsigned int curr_off = 0;
1946
1947         head = page_buffers(page);
1948         bh = head;
1949         do {
1950                 unsigned int next_off = curr_off + bh->b_size;
1951
1952                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1953                         to_release++;
1954                         clear_buffer_delay(bh);
1955                 }
1956                 curr_off = next_off;
1957         } while ((bh = bh->b_this_page) != head);
1958         ext4_da_release_space(page->mapping->host, to_release);
1959 }
1960
1961 /*
1962  * Delayed allocation stuff
1963  */
1964
1965 /*
1966  * mpage_da_submit_io - walks through extent of pages and try to write
1967  * them with writepage() call back
1968  *
1969  * @mpd->inode: inode
1970  * @mpd->first_page: first page of the extent
1971  * @mpd->next_page: page after the last page of the extent
1972  *
1973  * By the time mpage_da_submit_io() is called we expect all blocks
1974  * to be allocated. this may be wrong if allocation failed.
1975  *
1976  * As pages are already locked by write_cache_pages(), we can't use it
1977  */
1978 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1979 {
1980         long pages_skipped;
1981         struct pagevec pvec;
1982         unsigned long index, end;
1983         int ret = 0, err, nr_pages, i;
1984         struct inode *inode = mpd->inode;
1985         struct address_space *mapping = inode->i_mapping;
1986
1987         BUG_ON(mpd->next_page <= mpd->first_page);
1988         /*
1989          * We need to start from the first_page to the next_page - 1
1990          * to make sure we also write the mapped dirty buffer_heads.
1991          * If we look at mpd->b_blocknr we would only be looking
1992          * at the currently mapped buffer_heads.
1993          */
1994         index = mpd->first_page;
1995         end = mpd->next_page - 1;
1996
1997         pagevec_init(&pvec, 0);
1998         while (index <= end) {
1999                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2000                 if (nr_pages == 0)
2001                         break;
2002                 for (i = 0; i < nr_pages; i++) {
2003                         struct page *page = pvec.pages[i];
2004
2005                         index = page->index;
2006                         if (index > end)
2007                                 break;
2008                         index++;
2009
2010                         BUG_ON(!PageLocked(page));
2011                         BUG_ON(PageWriteback(page));
2012
2013                         pages_skipped = mpd->wbc->pages_skipped;
2014                         err = mapping->a_ops->writepage(page, mpd->wbc);
2015                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
2016                                 /*
2017                                  * have successfully written the page
2018                                  * without skipping the same
2019                                  */
2020                                 mpd->pages_written++;
2021                         /*
2022                          * In error case, we have to continue because
2023                          * remaining pages are still locked
2024                          * XXX: unlock and re-dirty them?
2025                          */
2026                         if (ret == 0)
2027                                 ret = err;
2028                 }
2029                 pagevec_release(&pvec);
2030         }
2031         return ret;
2032 }
2033
2034 /*
2035  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
2036  *
2037  * the function goes through all passed space and put actual disk
2038  * block numbers into buffer heads, dropping BH_Delay and BH_Unwritten
2039  */
2040 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd,
2041                                  struct ext4_map_blocks *map)
2042 {
2043         struct inode *inode = mpd->inode;
2044         struct address_space *mapping = inode->i_mapping;
2045         int blocks = map->m_len;
2046         sector_t pblock = map->m_pblk, cur_logical;
2047         struct buffer_head *head, *bh;
2048         pgoff_t index, end;
2049         struct pagevec pvec;
2050         int nr_pages, i;
2051
2052         index = map->m_lblk >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2053         end = (map->m_lblk + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2054         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2055
2056         pagevec_init(&pvec, 0);
2057
2058         while (index <= end) {
2059                 /* XXX: optimize tail */
2060                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2061                 if (nr_pages == 0)
2062                         break;
2063                 for (i = 0; i < nr_pages; i++) {
2064                         struct page *page = pvec.pages[i];
2065
2066                         index = page->index;
2067                         if (index > end)
2068                                 break;
2069                         index++;
2070
2071                         BUG_ON(!PageLocked(page));
2072                         BUG_ON(PageWriteback(page));
2073                         BUG_ON(!page_has_buffers(page));
2074
2075                         bh = page_buffers(page);
2076                         head = bh;
2077
2078                         /* skip blocks out of the range */
2079                         do {
2080                                 if (cur_logical >= map->m_lblk)
2081                                         break;
2082                                 cur_logical++;
2083                         } while ((bh = bh->b_this_page) != head);
2084
2085                         do {
2086                                 if (cur_logical >= map->m_lblk + blocks)
2087                                         break;
2088
2089                                 if (buffer_delay(bh) || buffer_unwritten(bh)) {
2090
2091                                         BUG_ON(bh->b_bdev != inode->i_sb->s_bdev);
2092
2093                                         if (buffer_delay(bh)) {
2094                                                 clear_buffer_delay(bh);
2095                                                 bh->b_blocknr = pblock;
2096                                         } else {
2097                                                 /*
2098                                                  * unwritten already should have
2099                                                  * blocknr assigned. Verify that
2100                                                  */
2101                                                 clear_buffer_unwritten(bh);
2102                                                 BUG_ON(bh->b_blocknr != pblock);
2103                                         }
2104
2105                                 } else if (buffer_mapped(bh))
2106                                         BUG_ON(bh->b_blocknr != pblock);
2107
2108                                 if (map->m_flags & EXT4_MAP_UNINIT)
2109                                         set_buffer_uninit(bh);
2110                                 cur_logical++;
2111                                 pblock++;
2112                         } while ((bh = bh->b_this_page) != head);
2113                 }
2114                 pagevec_release(&pvec);
2115         }
2116 }
2117
2118
2119 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2120                                         sector_t logical, long blk_cnt)
2121 {
2122         int nr_pages, i;
2123         pgoff_t index, end;
2124         struct pagevec pvec;
2125         struct inode *inode = mpd->inode;
2126         struct address_space *mapping = inode->i_mapping;
2127
2128         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2129         end   = (logical + blk_cnt - 1) >>
2130                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2131         while (index <= end) {
2132                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2133                 if (nr_pages == 0)
2134                         break;
2135                 for (i = 0; i < nr_pages; i++) {
2136                         struct page *page = pvec.pages[i];
2137                         if (page->index > end)
2138                                 break;
2139                         BUG_ON(!PageLocked(page));
2140                         BUG_ON(PageWriteback(page));
2141                         block_invalidatepage(page, 0);
2142                         ClearPageUptodate(page);
2143                         unlock_page(page);
2144                 }
2145                 index = pvec.pages[nr_pages - 1]->index + 1;
2146                 pagevec_release(&pvec);
2147         }
2148         return;
2149 }
2150
2151 static void ext4_print_free_blocks(struct inode *inode)
2152 {
2153         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2154         printk(KERN_CRIT "Total free blocks count %lld\n",
2155                ext4_count_free_blocks(inode->i_sb));
2156         printk(KERN_CRIT "Free/Dirty block details\n");
2157         printk(KERN_CRIT "free_blocks=%lld\n",
2158                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2159         printk(KERN_CRIT "dirty_blocks=%lld\n",
2160                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2161         printk(KERN_CRIT "Block reservation details\n");
2162         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2163                EXT4_I(inode)->i_reserved_data_blocks);
2164         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2165                EXT4_I(inode)->i_reserved_meta_blocks);
2166         return;
2167 }
2168
2169 /*
2170  * mpage_da_map_blocks - go through given space
2171  *
2172  * @mpd - bh describing space
2173  *
2174  * The function skips space we know is already mapped to disk blocks.
2175  *
2176  */
2177 static int mpage_da_map_blocks(struct mpage_da_data *mpd)
2178 {
2179         int err, blks, get_blocks_flags;
2180         struct ext4_map_blocks map;
2181         sector_t next = mpd->b_blocknr;
2182         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2183         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2184         handle_t *handle = NULL;
2185
2186         /*
2187          * We consider only non-mapped and non-allocated blocks
2188          */
2189         if ((mpd->b_state  & (1 << BH_Mapped)) &&
2190                 !(mpd->b_state & (1 << BH_Delay)) &&
2191                 !(mpd->b_state & (1 << BH_Unwritten)))
2192                 return 0;
2193
2194         /*
2195          * If we didn't accumulate anything to write simply return
2196          */
2197         if (!mpd->b_size)
2198                 return 0;
2199
2200         handle = ext4_journal_current_handle();
2201         BUG_ON(!handle);
2202
2203         /*
2204          * Call ext4_map_blocks() to allocate any delayed allocation
2205          * blocks, or to convert an uninitialized extent to be
2206          * initialized (in the case where we have written into
2207          * one or more preallocated blocks).
2208          *
2209          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2210          * indicate that we are on the delayed allocation path.  This
2211          * affects functions in many different parts of the allocation
2212          * call path.  This flag exists primarily because we don't
2213          * want to change *many* call functions, so ext4_map_blocks()
2214          * will set the magic i_delalloc_reserved_flag once the
2215          * inode's allocation semaphore is taken.
2216          *
2217          * If the blocks in questions were delalloc blocks, set
2218          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2219          * variables are updated after the blocks have been allocated.
2220          */
2221         map.m_lblk = next;
2222         map.m_len = max_blocks;
2223         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
2224         if (ext4_should_dioread_nolock(mpd->inode))
2225                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2226         if (mpd->b_state & (1 << BH_Delay))
2227                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2228
2229         blks = ext4_map_blocks(handle, mpd->inode, &map, get_blocks_flags);
2230         if (blks < 0) {
2231                 err = blks;
2232                 /*
2233                  * If get block returns with error we simply
2234                  * return. Later writepage will redirty the page and
2235                  * writepages will find the dirty page again
2236                  */
2237                 if (err == -EAGAIN)
2238                         return 0;
2239
2240                 if (err == -ENOSPC &&
2241                     ext4_count_free_blocks(mpd->inode->i_sb)) {
2242                         mpd->retval = err;
2243                         return 0;
2244                 }
2245
2246                 /*
2247                  * get block failure will cause us to loop in
2248                  * writepages, because a_ops->writepage won't be able
2249                  * to make progress. The page will be redirtied by
2250                  * writepage and writepages will again try to write
2251                  * the same.
2252                  */
2253                 ext4_msg(mpd->inode->i_sb, KERN_CRIT,
2254                          "delayed block allocation failed for inode %lu at "
2255                          "logical offset %llu with max blocks %zd with "
2256                          "error %d", mpd->inode->i_ino,
2257                          (unsigned long long) next,
2258                          mpd->b_size >> mpd->inode->i_blkbits, err);
2259                 printk(KERN_CRIT "This should not happen!!  "
2260                        "Data will be lost\n");
2261                 if (err == -ENOSPC) {
2262                         ext4_print_free_blocks(mpd->inode);
2263                 }
2264                 /* invalidate all the pages */
2265                 ext4_da_block_invalidatepages(mpd, next,
2266                                 mpd->b_size >> mpd->inode->i_blkbits);
2267                 return err;
2268         }
2269         BUG_ON(blks == 0);
2270
2271         if (map.m_flags & EXT4_MAP_NEW) {
2272                 struct block_device *bdev = mpd->inode->i_sb->s_bdev;
2273                 int i;
2274
2275                 for (i = 0; i < map.m_len; i++)
2276                         unmap_underlying_metadata(bdev, map.m_pblk + i);
2277         }
2278
2279         /*
2280          * If blocks are delayed marked, we need to
2281          * put actual blocknr and drop delayed bit
2282          */
2283         if ((mpd->b_state & (1 << BH_Delay)) ||
2284             (mpd->b_state & (1 << BH_Unwritten)))
2285                 mpage_put_bnr_to_bhs(mpd, &map);
2286
2287         if (ext4_should_order_data(mpd->inode)) {
2288                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2289                 if (err)
2290                         return err;
2291         }
2292
2293         /*
2294          * Update on-disk size along with block allocation.
2295          */
2296         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2297         if (disksize > i_size_read(mpd->inode))
2298                 disksize = i_size_read(mpd->inode);
2299         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2300                 ext4_update_i_disksize(mpd->inode, disksize);
2301                 return ext4_mark_inode_dirty(handle, mpd->inode);
2302         }
2303
2304         return 0;
2305 }
2306
2307 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2308                 (1 << BH_Delay) | (1 << BH_Unwritten))
2309
2310 /*
2311  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2312  *
2313  * @mpd->lbh - extent of blocks
2314  * @logical - logical number of the block in the file
2315  * @bh - bh of the block (used to access block's state)
2316  *
2317  * the function is used to collect contig. blocks in same state
2318  */
2319 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2320                                    sector_t logical, size_t b_size,
2321                                    unsigned long b_state)
2322 {
2323         sector_t next;
2324         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2325
2326         /*
2327          * XXX Don't go larger than mballoc is willing to allocate
2328          * This is a stopgap solution.  We eventually need to fold
2329          * mpage_da_submit_io() into this function and then call
2330          * ext4_map_blocks() multiple times in a loop
2331          */
2332         if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
2333                 goto flush_it;
2334
2335         /* check if thereserved journal credits might overflow */
2336         if (!(ext4_test_inode_flag(mpd->inode, EXT4_INODE_EXTENTS))) {
2337                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2338                         /*
2339                          * With non-extent format we are limited by the journal
2340                          * credit available.  Total credit needed to insert
2341                          * nrblocks contiguous blocks is dependent on the
2342                          * nrblocks.  So limit nrblocks.
2343                          */
2344                         goto flush_it;
2345                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2346                                 EXT4_MAX_TRANS_DATA) {
2347                         /*
2348                          * Adding the new buffer_head would make it cross the
2349                          * allowed limit for which we have journal credit
2350                          * reserved. So limit the new bh->b_size
2351                          */
2352                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2353                                                 mpd->inode->i_blkbits;
2354                         /* we will do mpage_da_submit_io in the next loop */
2355                 }
2356         }
2357         /*
2358          * First block in the extent
2359          */
2360         if (mpd->b_size == 0) {
2361                 mpd->b_blocknr = logical;
2362                 mpd->b_size = b_size;
2363                 mpd->b_state = b_state & BH_FLAGS;
2364                 return;
2365         }
2366
2367         next = mpd->b_blocknr + nrblocks;
2368         /*
2369          * Can we merge the block to our big extent?
2370          */
2371         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2372                 mpd->b_size += b_size;
2373                 return;
2374         }
2375
2376 flush_it:
2377         /*
2378          * We couldn't merge the block to our extent, so we
2379          * need to flush current  extent and start new one
2380          */
2381         if (mpage_da_map_blocks(mpd) == 0)
2382                 mpage_da_submit_io(mpd);
2383         mpd->io_done = 1;
2384         return;
2385 }
2386
2387 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
2388 {
2389         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
2390 }
2391
2392 /*
2393  * __mpage_da_writepage - finds extent of pages and blocks
2394  *
2395  * @page: page to consider
2396  * @wbc: not used, we just follow rules
2397  * @data: context
2398  *
2399  * The function finds extents of pages and scan them for all blocks.
2400  */
2401 static int __mpage_da_writepage(struct page *page,
2402                                 struct writeback_control *wbc, void *data)
2403 {
2404         struct mpage_da_data *mpd = data;
2405         struct inode *inode = mpd->inode;
2406         struct buffer_head *bh, *head;
2407         sector_t logical;
2408
2409         /*
2410          * Can we merge this page to current extent?
2411          */
2412         if (mpd->next_page != page->index) {
2413                 /*
2414                  * Nope, we can't. So, we map non-allocated blocks
2415                  * and start IO on them using writepage()
2416                  */
2417                 if (mpd->next_page != mpd->first_page) {
2418                         if (mpage_da_map_blocks(mpd) == 0)
2419                                 mpage_da_submit_io(mpd);
2420                         /*
2421                          * skip rest of the page in the page_vec
2422                          */
2423                         mpd->io_done = 1;
2424                         redirty_page_for_writepage(wbc, page);
2425                         unlock_page(page);
2426                         return MPAGE_DA_EXTENT_TAIL;
2427                 }
2428
2429                 /*
2430                  * Start next extent of pages ...
2431                  */
2432                 mpd->first_page = page->index;
2433
2434                 /*
2435                  * ... and blocks
2436                  */
2437                 mpd->b_size = 0;
2438                 mpd->b_state = 0;
2439                 mpd->b_blocknr = 0;
2440         }
2441
2442         mpd->next_page = page->index + 1;
2443         logical = (sector_t) page->index <<
2444                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2445
2446         if (!page_has_buffers(page)) {
2447                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2448                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2449                 if (mpd->io_done)
2450                         return MPAGE_DA_EXTENT_TAIL;
2451         } else {
2452                 /*
2453                  * Page with regular buffer heads, just add all dirty ones
2454                  */
2455                 head = page_buffers(page);
2456                 bh = head;
2457                 do {
2458                         BUG_ON(buffer_locked(bh));
2459                         /*
2460                          * We need to try to allocate
2461                          * unmapped blocks in the same page.
2462                          * Otherwise we won't make progress
2463                          * with the page in ext4_writepage
2464                          */
2465                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2466                                 mpage_add_bh_to_extent(mpd, logical,
2467                                                        bh->b_size,
2468                                                        bh->b_state);
2469                                 if (mpd->io_done)
2470                                         return MPAGE_DA_EXTENT_TAIL;
2471                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2472                                 /*
2473                                  * mapped dirty buffer. We need to update
2474                                  * the b_state because we look at
2475                                  * b_state in mpage_da_map_blocks. We don't
2476                                  * update b_size because if we find an
2477                                  * unmapped buffer_head later we need to
2478                                  * use the b_state flag of that buffer_head.
2479                                  */
2480                                 if (mpd->b_size == 0)
2481                                         mpd->b_state = bh->b_state & BH_FLAGS;
2482                         }
2483                         logical++;
2484                 } while ((bh = bh->b_this_page) != head);
2485         }
2486
2487         return 0;
2488 }
2489
2490 /*
2491  * This is a special get_blocks_t callback which is used by
2492  * ext4_da_write_begin().  It will either return mapped block or
2493  * reserve space for a single block.
2494  *
2495  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2496  * We also have b_blocknr = -1 and b_bdev initialized properly
2497  *
2498  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2499  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2500  * initialized properly.
2501  */
2502 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2503                                   struct buffer_head *bh, int create)
2504 {
2505         struct ext4_map_blocks map;
2506         int ret = 0;
2507         sector_t invalid_block = ~((sector_t) 0xffff);
2508
2509         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2510                 invalid_block = ~0;
2511
2512         BUG_ON(create == 0);
2513         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
2514
2515         map.m_lblk = iblock;
2516         map.m_len = 1;
2517
2518         /*
2519          * first, we need to know whether the block is allocated already
2520          * preallocated blocks are unmapped but should treated
2521          * the same as allocated blocks.
2522          */
2523         ret = ext4_map_blocks(NULL, inode, &map, 0);
2524         if (ret < 0)
2525                 return ret;
2526         if (ret == 0) {
2527                 if (buffer_delay(bh))
2528                         return 0; /* Not sure this could or should happen */
2529                 /*
2530                  * XXX: __block_prepare_write() unmaps passed block,
2531                  * is it OK?
2532                  */
2533                 ret = ext4_da_reserve_space(inode, iblock);
2534                 if (ret)
2535                         /* not enough space to reserve */
2536                         return ret;
2537
2538                 map_bh(bh, inode->i_sb, invalid_block);
2539                 set_buffer_new(bh);
2540                 set_buffer_delay(bh);
2541                 return 0;
2542         }
2543
2544         map_bh(bh, inode->i_sb, map.m_pblk);
2545         bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
2546
2547         if (buffer_unwritten(bh)) {
2548                 /* A delayed write to unwritten bh should be marked
2549                  * new and mapped.  Mapped ensures that we don't do
2550                  * get_block multiple times when we write to the same
2551                  * offset and new ensures that we do proper zero out
2552                  * for partial write.
2553                  */
2554                 set_buffer_new(bh);
2555                 set_buffer_mapped(bh);
2556         }
2557         return 0;
2558 }
2559
2560 /*
2561  * This function is used as a standard get_block_t calback function
2562  * when there is no desire to allocate any blocks.  It is used as a
2563  * callback function for block_prepare_write() and block_write_full_page().
2564  * These functions should only try to map a single block at a time.
2565  *
2566  * Since this function doesn't do block allocations even if the caller
2567  * requests it by passing in create=1, it is critically important that
2568  * any caller checks to make sure that any buffer heads are returned
2569  * by this function are either all already mapped or marked for
2570  * delayed allocation before calling  block_write_full_page().  Otherwise,
2571  * b_blocknr could be left unitialized, and the page write functions will
2572  * be taken by surprise.
2573  */
2574 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2575                                    struct buffer_head *bh_result, int create)
2576 {
2577         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2578         return _ext4_get_block(inode, iblock, bh_result, 0);
2579 }
2580
2581 static int bget_one(handle_t *handle, struct buffer_head *bh)
2582 {
2583         get_bh(bh);
2584         return 0;
2585 }
2586
2587 static int bput_one(handle_t *handle, struct buffer_head *bh)
2588 {
2589         put_bh(bh);
2590         return 0;
2591 }
2592
2593 static int __ext4_journalled_writepage(struct page *page,
2594                                        unsigned int len)
2595 {
2596         struct address_space *mapping = page->mapping;
2597         struct inode *inode = mapping->host;
2598         struct buffer_head *page_bufs;
2599         handle_t *handle = NULL;
2600         int ret = 0;
2601         int err;
2602
2603         page_bufs = page_buffers(page);
2604         BUG_ON(!page_bufs);
2605         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2606         /* As soon as we unlock the page, it can go away, but we have
2607          * references to buffers so we are safe */
2608         unlock_page(page);
2609
2610         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2611         if (IS_ERR(handle)) {
2612                 ret = PTR_ERR(handle);
2613                 goto out;
2614         }
2615
2616         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2617                                 do_journal_get_write_access);
2618
2619         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2620                                 write_end_fn);
2621         if (ret == 0)
2622                 ret = err;
2623         err = ext4_journal_stop(handle);
2624         if (!ret)
2625                 ret = err;
2626
2627         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2628         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2629 out:
2630         return ret;
2631 }
2632
2633 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
2634 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
2635
2636 /*
2637  * Note that we don't need to start a transaction unless we're journaling data
2638  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2639  * need to file the inode to the transaction's list in ordered mode because if
2640  * we are writing back data added by write(), the inode is already there and if
2641  * we are writing back data modified via mmap(), noone guarantees in which
2642  * transaction the data will hit the disk. In case we are journaling data, we
2643  * cannot start transaction directly because transaction start ranks above page
2644  * lock so we have to do some magic.
2645  *
2646  * This function can get called via...
2647  *   - ext4_da_writepages after taking page lock (have journal handle)
2648  *   - journal_submit_inode_data_buffers (no journal handle)
2649  *   - shrink_page_list via pdflush (no journal handle)
2650  *   - grab_page_cache when doing write_begin (have journal handle)
2651  *
2652  * We don't do any block allocation in this function. If we have page with
2653  * multiple blocks we need to write those buffer_heads that are mapped. This
2654  * is important for mmaped based write. So if we do with blocksize 1K
2655  * truncate(f, 1024);
2656  * a = mmap(f, 0, 4096);
2657  * a[0] = 'a';
2658  * truncate(f, 4096);
2659  * we have in the page first buffer_head mapped via page_mkwrite call back
2660  * but other bufer_heads would be unmapped but dirty(dirty done via the
2661  * do_wp_page). So writepage should write the first block. If we modify
2662  * the mmap area beyond 1024 we will again get a page_fault and the
2663  * page_mkwrite callback will do the block allocation and mark the
2664  * buffer_heads mapped.
2665  *
2666  * We redirty the page if we have any buffer_heads that is either delay or
2667  * unwritten in the page.
2668  *
2669  * We can get recursively called as show below.
2670  *
2671  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2672  *              ext4_writepage()
2673  *
2674  * But since we don't do any block allocation we should not deadlock.
2675  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2676  */
2677 static int ext4_writepage(struct page *page,
2678                           struct writeback_control *wbc)
2679 {
2680         int ret = 0;
2681         loff_t size;
2682         unsigned int len;
2683         struct buffer_head *page_bufs = NULL;
2684         struct inode *inode = page->mapping->host;
2685
2686         trace_ext4_writepage(inode, page);
2687         size = i_size_read(inode);
2688         if (page->index == size >> PAGE_CACHE_SHIFT)
2689                 len = size & ~PAGE_CACHE_MASK;
2690         else
2691                 len = PAGE_CACHE_SIZE;
2692
2693         if (page_has_buffers(page)) {
2694                 page_bufs = page_buffers(page);
2695                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2696                                         ext4_bh_delay_or_unwritten)) {
2697                         /*
2698                          * We don't want to do  block allocation
2699                          * So redirty the page and return
2700                          * We may reach here when we do a journal commit
2701                          * via journal_submit_inode_data_buffers.
2702                          * If we don't have mapping block we just ignore
2703                          * them. We can also reach here via shrink_page_list
2704                          */
2705                         redirty_page_for_writepage(wbc, page);
2706                         unlock_page(page);
2707                         return 0;
2708                 }
2709         } else {
2710                 /*
2711                  * The test for page_has_buffers() is subtle:
2712                  * We know the page is dirty but it lost buffers. That means
2713                  * that at some moment in time after write_begin()/write_end()
2714                  * has been called all buffers have been clean and thus they
2715                  * must have been written at least once. So they are all
2716                  * mapped and we can happily proceed with mapping them
2717                  * and writing the page.
2718                  *
2719                  * Try to initialize the buffer_heads and check whether
2720                  * all are mapped and non delay. We don't want to
2721                  * do block allocation here.
2722                  */
2723                 ret = block_prepare_write(page, 0, len,
2724                                           noalloc_get_block_write);
2725                 if (!ret) {
2726                         page_bufs = page_buffers(page);
2727                         /* check whether all are mapped and non delay */
2728                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2729                                                 ext4_bh_delay_or_unwritten)) {
2730                                 redirty_page_for_writepage(wbc, page);
2731                                 unlock_page(page);
2732                                 return 0;
2733                         }
2734                 } else {
2735                         /*
2736                          * We can't do block allocation here
2737                          * so just redity the page and unlock
2738                          * and return
2739                          */
2740                         redirty_page_for_writepage(wbc, page);
2741                         unlock_page(page);
2742                         return 0;
2743                 }
2744                 /* now mark the buffer_heads as dirty and uptodate */
2745                 block_commit_write(page, 0, len);
2746         }
2747
2748         if (PageChecked(page) && ext4_should_journal_data(inode)) {
2749                 /*
2750                  * It's mmapped pagecache.  Add buffers and journal it.  There
2751                  * doesn't seem much point in redirtying the page here.
2752                  */
2753                 ClearPageChecked(page);
2754                 return __ext4_journalled_writepage(page, len);
2755         }
2756
2757         if (page_bufs && buffer_uninit(page_bufs)) {
2758                 ext4_set_bh_endio(page_bufs, inode);
2759                 ret = block_write_full_page_endio(page, noalloc_get_block_write,
2760                                             wbc, ext4_end_io_buffer_write);
2761         } else
2762                 ret = block_write_full_page(page, noalloc_get_block_write,
2763                                             wbc);
2764
2765         return ret;
2766 }
2767
2768 /*
2769  * This is called via ext4_da_writepages() to
2770  * calulate the total number of credits to reserve to fit
2771  * a single extent allocation into a single transaction,
2772  * ext4_da_writpeages() will loop calling this before
2773  * the block allocation.
2774  */
2775
2776 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2777 {
2778         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2779
2780         /*
2781          * With non-extent format the journal credit needed to
2782          * insert nrblocks contiguous block is dependent on
2783          * number of contiguous block. So we will limit
2784          * number of contiguous block to a sane value
2785          */
2786         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) &&
2787             (max_blocks > EXT4_MAX_TRANS_DATA))
2788                 max_blocks = EXT4_MAX_TRANS_DATA;
2789
2790         return ext4_chunk_trans_blocks(inode, max_blocks);
2791 }
2792
2793 /*
2794  * write_cache_pages_da - walk the list of dirty pages of the given
2795  * address space and call the callback function (which usually writes
2796  * the pages).
2797  *
2798  * This is a forked version of write_cache_pages().  Differences:
2799  *      Range cyclic is ignored.
2800  *      no_nrwrite_index_update is always presumed true
2801  */
2802 static int write_cache_pages_da(struct address_space *mapping,
2803                                 struct writeback_control *wbc,
2804                                 struct mpage_da_data *mpd)
2805 {
2806         int ret = 0;
2807         int done = 0;
2808         struct pagevec pvec;
2809         int nr_pages;
2810         pgoff_t index;
2811         pgoff_t end;            /* Inclusive */
2812         long nr_to_write = wbc->nr_to_write;
2813
2814         pagevec_init(&pvec, 0);
2815         index = wbc->range_start >> PAGE_CACHE_SHIFT;
2816         end = wbc->range_end >> PAGE_CACHE_SHIFT;
2817
2818         while (!done && (index <= end)) {
2819                 int i;
2820
2821                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2822                               PAGECACHE_TAG_DIRTY,
2823                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2824                 if (nr_pages == 0)
2825                         break;
2826
2827                 for (i = 0; i < nr_pages; i++) {
2828                         struct page *page = pvec.pages[i];
2829
2830                         /*
2831                          * At this point, the page may be truncated or
2832                          * invalidated (changing page->mapping to NULL), or
2833                          * even swizzled back from swapper_space to tmpfs file
2834                          * mapping. However, page->index will not change
2835                          * because we have a reference on the page.
2836                          */
2837                         if (page->index > end) {
2838                                 done = 1;
2839                                 break;
2840                         }
2841
2842                         lock_page(page);
2843
2844                         /*
2845                          * Page truncated or invalidated. We can freely skip it
2846                          * then, even for data integrity operations: the page
2847                          * has disappeared concurrently, so there could be no
2848                          * real expectation of this data interity operation
2849                          * even if there is now a new, dirty page at the same
2850                          * pagecache address.
2851                          */
2852                         if (unlikely(page->mapping != mapping)) {
2853 continue_unlock:
2854                                 unlock_page(page);
2855                                 continue;
2856                         }
2857
2858                         if (!PageDirty(page)) {
2859                                 /* someone wrote it for us */
2860                                 goto continue_unlock;
2861                         }
2862
2863                         if (PageWriteback(page)) {
2864                                 if (wbc->sync_mode != WB_SYNC_NONE)
2865                                         wait_on_page_writeback(page);
2866                                 else
2867                                         goto continue_unlock;
2868                         }
2869
2870                         BUG_ON(PageWriteback(page));
2871                         if (!clear_page_dirty_for_io(page))
2872                                 goto continue_unlock;
2873
2874                         ret = __mpage_da_writepage(page, wbc, mpd);
2875                         if (unlikely(ret)) {
2876                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2877                                         unlock_page(page);
2878                                         ret = 0;
2879                                 } else {
2880                                         done = 1;
2881                                         break;
2882                                 }
2883                         }
2884
2885                         if (nr_to_write > 0) {
2886                                 nr_to_write--;
2887                                 if (nr_to_write == 0 &&
2888                                     wbc->sync_mode == WB_SYNC_NONE) {
2889                                         /*
2890                                          * We stop writing back only if we are
2891                                          * not doing integrity sync. In case of
2892                                          * integrity sync we have to keep going
2893                                          * because someone may be concurrently
2894                                          * dirtying pages, and we might have
2895                                          * synced a lot of newly appeared dirty
2896                                          * pages, but have not synced all of the
2897                                          * old dirty pages.
2898                                          */
2899                                         done = 1;
2900                                         break;
2901                                 }
2902                         }
2903                 }
2904                 pagevec_release(&pvec);
2905                 cond_resched();
2906         }
2907         return ret;
2908 }
2909
2910
2911 static int ext4_da_writepages(struct address_space *mapping,
2912                               struct writeback_control *wbc)
2913 {
2914         pgoff_t index;
2915         int range_whole = 0;
2916         handle_t *handle = NULL;
2917         struct mpage_da_data mpd;
2918         struct inode *inode = mapping->host;
2919         int pages_written = 0;
2920         long pages_skipped;
2921         unsigned int max_pages;
2922         int range_cyclic, cycled = 1, io_done = 0;
2923         int needed_blocks, ret = 0;
2924         long desired_nr_to_write, nr_to_writebump = 0;
2925         loff_t range_start = wbc->range_start;
2926         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2927
2928         trace_ext4_da_writepages(inode, wbc);
2929
2930         /*
2931          * No pages to write? This is mainly a kludge to avoid starting
2932          * a transaction for special inodes like journal inode on last iput()
2933          * because that could violate lock ordering on umount
2934          */
2935         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2936                 return 0;
2937
2938         /*
2939          * If the filesystem has aborted, it is read-only, so return
2940          * right away instead of dumping stack traces later on that
2941          * will obscure the real source of the problem.  We test
2942          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2943          * the latter could be true if the filesystem is mounted
2944          * read-only, and in that case, ext4_da_writepages should
2945          * *never* be called, so if that ever happens, we would want
2946          * the stack trace.
2947          */
2948         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2949                 return -EROFS;
2950
2951         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2952                 range_whole = 1;
2953
2954         range_cyclic = wbc->range_cyclic;
2955         if (wbc->range_cyclic) {
2956                 index = mapping->writeback_index;
2957                 if (index)
2958                         cycled = 0;
2959                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2960                 wbc->range_end  = LLONG_MAX;
2961                 wbc->range_cyclic = 0;
2962         } else
2963                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2964
2965         /*
2966          * This works around two forms of stupidity.  The first is in
2967          * the writeback code, which caps the maximum number of pages
2968          * written to be 1024 pages.  This is wrong on multiple
2969          * levels; different architectues have a different page size,
2970          * which changes the maximum amount of data which gets
2971          * written.  Secondly, 4 megabytes is way too small.  XFS
2972          * forces this value to be 16 megabytes by multiplying
2973          * nr_to_write parameter by four, and then relies on its
2974          * allocator to allocate larger extents to make them
2975          * contiguous.  Unfortunately this brings us to the second
2976          * stupidity, which is that ext4's mballoc code only allocates
2977          * at most 2048 blocks.  So we force contiguous writes up to
2978          * the number of dirty blocks in the inode, or
2979          * sbi->max_writeback_mb_bump whichever is smaller.
2980          */
2981         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2982         if (!range_cyclic && range_whole)
2983                 desired_nr_to_write = wbc->nr_to_write * 8;
2984         else
2985                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2986                                                            max_pages);
2987         if (desired_nr_to_write > max_pages)
2988                 desired_nr_to_write = max_pages;
2989
2990         if (wbc->nr_to_write < desired_nr_to_write) {
2991                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2992                 wbc->nr_to_write = desired_nr_to_write;
2993         }
2994
2995         mpd.wbc = wbc;
2996         mpd.inode = mapping->host;
2997
2998         pages_skipped = wbc->pages_skipped;
2999
3000 retry:
3001         while (!ret && wbc->nr_to_write > 0) {
3002
3003                 /*
3004                  * we  insert one extent at a time. So we need
3005                  * credit needed for single extent allocation.
3006                  * journalled mode is currently not supported
3007                  * by delalloc
3008                  */
3009                 BUG_ON(ext4_should_journal_data(inode));
3010                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
3011
3012                 /* start a new transaction*/
3013                 handle = ext4_journal_start(inode, needed_blocks);
3014                 if (IS_ERR(handle)) {
3015                         ret = PTR_ERR(handle);
3016                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
3017                                "%ld pages, ino %lu; err %d", __func__,
3018                                 wbc->nr_to_write, inode->i_ino, ret);
3019                         goto out_writepages;
3020                 }
3021
3022                 /*
3023                  * Now call __mpage_da_writepage to find the next
3024                  * contiguous region of logical blocks that need
3025                  * blocks to be allocated by ext4.  We don't actually
3026                  * submit the blocks for I/O here, even though
3027                  * write_cache_pages thinks it will, and will set the
3028                  * pages as clean for write before calling
3029                  * __mpage_da_writepage().
3030                  */
3031                 mpd.b_size = 0;
3032                 mpd.b_state = 0;
3033                 mpd.b_blocknr = 0;
3034                 mpd.first_page = 0;
3035                 mpd.next_page = 0;
3036                 mpd.io_done = 0;
3037                 mpd.pages_written = 0;
3038                 mpd.retval = 0;
3039                 ret = write_cache_pages_da(mapping, wbc, &mpd);
3040                 /*
3041                  * If we have a contiguous extent of pages and we
3042                  * haven't done the I/O yet, map the blocks and submit
3043                  * them for I/O.
3044                  */
3045                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
3046                         if (mpage_da_map_blocks(&mpd) == 0)
3047                                 mpage_da_submit_io(&mpd);
3048                         mpd.io_done = 1;
3049                         ret = MPAGE_DA_EXTENT_TAIL;
3050                 }
3051                 trace_ext4_da_write_pages(inode, &mpd);
3052                 wbc->nr_to_write -= mpd.pages_written;
3053
3054                 ext4_journal_stop(handle);
3055
3056                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
3057                         /* commit the transaction which would
3058                          * free blocks released in the transaction
3059                          * and try again
3060                          */
3061                         jbd2_journal_force_commit_nested(sbi->s_journal);
3062                         wbc->pages_skipped = pages_skipped;
3063                         ret = 0;
3064                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
3065                         /*
3066                          * got one extent now try with
3067                          * rest of the pages
3068                          */
3069                         pages_written += mpd.pages_written;
3070                         wbc->pages_skipped = pages_skipped;
3071                         ret = 0;
3072                         io_done = 1;
3073                 } else if (wbc->nr_to_write)
3074                         /*
3075                          * There is no more writeout needed
3076                          * or we requested for a noblocking writeout
3077                          * and we found the device congested
3078                          */
3079                         break;
3080         }
3081         if (!io_done && !cycled) {
3082                 cycled = 1;
3083                 index = 0;
3084                 wbc->range_start = index << PAGE_CACHE_SHIFT;
3085                 wbc->range_end  = mapping->writeback_index - 1;
3086                 goto retry;
3087         }
3088         if (pages_skipped != wbc->pages_skipped)
3089                 ext4_msg(inode->i_sb, KERN_CRIT,
3090                          "This should not happen leaving %s "
3091                          "with nr_to_write = %ld ret = %d",
3092                          __func__, wbc->nr_to_write, ret);
3093
3094         /* Update index */
3095         index += pages_written;
3096         wbc->range_cyclic = range_cyclic;
3097         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
3098                 /*
3099                  * set the writeback_index so that range_cyclic
3100                  * mode will write it back later
3101                  */
3102                 mapping->writeback_index = index;
3103
3104 out_writepages:
3105         wbc->nr_to_write -= nr_to_writebump;
3106         wbc->range_start = range_start;
3107         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
3108         return ret;
3109 }
3110
3111 #define FALL_BACK_TO_NONDELALLOC 1
3112 static int ext4_nonda_switch(struct super_block *sb)
3113 {
3114         s64 free_blocks, dirty_blocks;
3115         struct ext4_sb_info *sbi = EXT4_SB(sb);
3116
3117         /*
3118          * switch to non delalloc mode if we are running low
3119          * on free block. The free block accounting via percpu
3120          * counters can get slightly wrong with percpu_counter_batch getting
3121          * accumulated on each CPU without updating global counters
3122          * Delalloc need an accurate free block accounting. So switch
3123          * to non delalloc when we are near to error range.
3124          */
3125         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
3126         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
3127         if (2 * free_blocks < 3 * dirty_blocks ||
3128                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
3129                 /*
3130                  * free block count is less than 150% of dirty blocks
3131                  * or free blocks is less than watermark
3132                  */
3133                 return 1;
3134         }
3135         /*
3136          * Even if we don't switch but are nearing capacity,
3137          * start pushing delalloc when 1/2 of free blocks are dirty.
3138          */
3139         if (free_blocks < 2 * dirty_blocks)
3140                 writeback_inodes_sb_if_idle(sb);
3141
3142         return 0;
3143 }
3144
3145 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3146                                loff_t pos, unsigned len, unsigned flags,
3147                                struct page **pagep, void **fsdata)
3148 {
3149         int ret, retries = 0;
3150         struct page *page;
3151         pgoff_t index;
3152         struct inode *inode = mapping->host;
3153         handle_t *handle;
3154
3155         index = pos >> PAGE_CACHE_SHIFT;
3156
3157         if (ext4_nonda_switch(inode->i_sb)) {
3158                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3159                 return ext4_write_begin(file, mapping, pos,
3160                                         len, flags, pagep, fsdata);
3161         }
3162         *fsdata = (void *)0;
3163         trace_ext4_da_write_begin(inode, pos, len, flags);
3164 retry:
3165         /*
3166          * With delayed allocation, we don't log the i_disksize update
3167          * if there is delayed block allocation. But we still need
3168          * to journalling the i_disksize update if writes to the end
3169          * of file which has an already mapped buffer.
3170          */
3171         handle = ext4_journal_start(inode, 1);
3172         if (IS_ERR(handle)) {
3173                 ret = PTR_ERR(handle);
3174                 goto out;
3175         }
3176         /* We cannot recurse into the filesystem as the transaction is already
3177          * started */
3178         flags |= AOP_FLAG_NOFS;
3179
3180         page = grab_cache_page_write_begin(mapping, index, flags);
3181         if (!page) {
3182                 ext4_journal_stop(handle);
3183                 ret = -ENOMEM;
3184                 goto out;
3185         }
3186         *pagep = page;
3187
3188         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
3189                                 ext4_da_get_block_prep);
3190         if (ret < 0) {
3191                 unlock_page(page);
3192                 ext4_journal_stop(handle);
3193                 page_cache_release(page);
3194                 /*
3195                  * block_write_begin may have instantiated a few blocks
3196                  * outside i_size.  Trim these off again. Don't need
3197                  * i_size_read because we hold i_mutex.
3198                  */
3199                 if (pos + len > inode->i_size)
3200                         ext4_truncate_failed_write(inode);
3201         }
3202
3203         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3204                 goto retry;
3205 out:
3206         return ret;
3207 }
3208
3209 /*
3210  * Check if we should update i_disksize
3211  * when write to the end of file but not require block allocation
3212  */
3213 static int ext4_da_should_update_i_disksize(struct page *page,
3214                                             unsigned long offset)
3215 {
3216         struct buffer_head *bh;
3217         struct inode *inode = page->mapping->host;
3218         unsigned int idx;
3219         int i;
3220
3221         bh = page_buffers(page);
3222         idx = offset >> inode->i_blkbits;
3223
3224         for (i = 0; i < idx; i++)
3225                 bh = bh->b_this_page;
3226
3227         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3228                 return 0;
3229         return 1;
3230 }
3231
3232 static int ext4_da_write_end(struct file *file,
3233                              struct address_space *mapping,
3234                              loff_t pos, unsigned len, unsigned copied,
3235                              struct page *page, void *fsdata)
3236 {
3237         struct inode *inode = mapping->host;
3238         int ret = 0, ret2;
3239         handle_t *handle = ext4_journal_current_handle();
3240         loff_t new_i_size;
3241         unsigned long start, end;
3242         int write_mode = (int)(unsigned long)fsdata;
3243
3244         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
3245                 if (ext4_should_order_data(inode)) {
3246                         return ext4_ordered_write_end(file, mapping, pos,
3247                                         len, copied, page, fsdata);
3248                 } else if (ext4_should_writeback_data(inode)) {
3249                         return ext4_writeback_write_end(file, mapping, pos,
3250                                         len, copied, page, fsdata);
3251                 } else {
3252                         BUG();
3253                 }
3254         }
3255
3256         trace_ext4_da_write_end(inode, pos, len, copied);
3257         start = pos & (PAGE_CACHE_SIZE - 1);
3258         end = start + copied - 1;
3259
3260         /*
3261          * generic_write_end() will run mark_inode_dirty() if i_size
3262          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3263          * into that.
3264          */
3265
3266         new_i_size = pos + copied;
3267         if (new_i_size > EXT4_I(inode)->i_disksize) {
3268                 if (ext4_da_should_update_i_disksize(page, end)) {
3269                         down_write(&EXT4_I(inode)->i_data_sem);
3270                         if (new_i_size > EXT4_I(inode)->i_disksize) {
3271                                 /*
3272                                  * Updating i_disksize when extending file
3273                                  * without needing block allocation
3274                                  */
3275                                 if (ext4_should_order_data(inode))
3276                                         ret = ext4_jbd2_file_inode(handle,
3277                                                                    inode);
3278
3279                                 EXT4_I(inode)->i_disksize = new_i_size;
3280                         }
3281                         up_write(&EXT4_I(inode)->i_data_sem);
3282                         /* We need to mark inode dirty even if
3283                          * new_i_size is less that inode->i_size
3284                          * bu greater than i_disksize.(hint delalloc)
3285                          */
3286                         ext4_mark_inode_dirty(handle, inode);
3287                 }
3288         }
3289         ret2 = generic_write_end(file, mapping, pos, len, copied,
3290                                                         page, fsdata);
3291         copied = ret2;
3292         if (ret2 < 0)
3293                 ret = ret2;
3294         ret2 = ext4_journal_stop(handle);
3295         if (!ret)
3296                 ret = ret2;
3297
3298         return ret ? ret : copied;
3299 }
3300
3301 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
3302 {
3303         /*
3304          * Drop reserved blocks
3305          */
3306         BUG_ON(!PageLocked(page));
3307         if (!page_has_buffers(page))
3308                 goto out;
3309
3310         ext4_da_page_release_reservation(page, offset);
3311
3312 out:
3313         ext4_invalidatepage(page, offset);
3314
3315         return;
3316 }
3317
3318 /*
3319  * Force all delayed allocation blocks to be allocated for a given inode.
3320  */
3321 int ext4_alloc_da_blocks(struct inode *inode)
3322 {
3323         trace_ext4_alloc_da_blocks(inode);
3324
3325         if (!EXT4_I(inode)->i_reserved_data_blocks &&
3326             !EXT4_I(inode)->i_reserved_meta_blocks)
3327                 return 0;
3328
3329         /*
3330          * We do something simple for now.  The filemap_flush() will
3331          * also start triggering a write of the data blocks, which is
3332          * not strictly speaking necessary (and for users of
3333          * laptop_mode, not even desirable).  However, to do otherwise
3334          * would require replicating code paths in:
3335          *
3336          * ext4_da_writepages() ->
3337          *    write_cache_pages() ---> (via passed in callback function)
3338          *        __mpage_da_writepage() -->
3339          *           mpage_add_bh_to_extent()
3340          *           mpage_da_map_blocks()
3341          *
3342          * The problem is that write_cache_pages(), located in
3343          * mm/page-writeback.c, marks pages clean in preparation for
3344          * doing I/O, which is not desirable if we're not planning on
3345          * doing I/O at all.
3346          *
3347          * We could call write_cache_pages(), and then redirty all of
3348          * the pages by calling redirty_page_for_writeback() but that
3349          * would be ugly in the extreme.  So instead we would need to
3350          * replicate parts of the code in the above functions,
3351          * simplifying them becuase we wouldn't actually intend to
3352          * write out the pages, but rather only collect contiguous
3353          * logical block extents, call the multi-block allocator, and
3354          * then update the buffer heads with the block allocations.
3355          *
3356          * For now, though, we'll cheat by calling filemap_flush(),
3357          * which will map the blocks, and start the I/O, but not
3358          * actually wait for the I/O to complete.
3359          */
3360         return filemap_flush(inode->i_mapping);
3361 }
3362
3363 /*
3364  * bmap() is special.  It gets used by applications such as lilo and by
3365  * the swapper to find the on-disk block of a specific piece of data.
3366  *
3367  * Naturally, this is dangerous if the block concerned is still in the
3368  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3369  * filesystem and enables swap, then they may get a nasty shock when the
3370  * data getting swapped to that swapfile suddenly gets overwritten by
3371  * the original zero's written out previously to the journal and
3372  * awaiting writeback in the kernel's buffer cache.
3373  *
3374  * So, if we see any bmap calls here on a modified, data-journaled file,
3375  * take extra steps to flush any blocks which might be in the cache.
3376  */
3377 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3378 {
3379         struct inode *inode = mapping->host;
3380         journal_t *journal;
3381         int err;
3382
3383         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3384                         test_opt(inode->i_sb, DELALLOC)) {
3385                 /*
3386                  * With delalloc we want to sync the file
3387                  * so that we can make sure we allocate
3388                  * blocks for file
3389                  */
3390                 filemap_write_and_wait(mapping);
3391         }
3392
3393         if (EXT4_JOURNAL(inode) &&
3394             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3395                 /*
3396                  * This is a REALLY heavyweight approach, but the use of
3397                  * bmap on dirty files is expected to be extremely rare:
3398                  * only if we run lilo or swapon on a freshly made file
3399                  * do we expect this to happen.
3400                  *
3401                  * (bmap requires CAP_SYS_RAWIO so this does not
3402                  * represent an unprivileged user DOS attack --- we'd be
3403                  * in trouble if mortal users could trigger this path at
3404                  * will.)
3405                  *
3406                  * NB. EXT4_STATE_JDATA is not set on files other than
3407                  * regular files.  If somebody wants to bmap a directory
3408                  * or symlink and gets confused because the buffer
3409                  * hasn't yet been flushed to disk, they deserve
3410                  * everything they get.
3411                  */
3412
3413                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3414                 journal = EXT4_JOURNAL(inode);
3415                 jbd2_journal_lock_updates(journal);
3416                 err = jbd2_journal_flush(journal);
3417                 jbd2_journal_unlock_updates(journal);
3418
3419                 if (err)
3420                         return 0;
3421         }
3422
3423         return generic_block_bmap(mapping, block, ext4_get_block);
3424 }
3425
3426 static int ext4_readpage(struct file *file, struct page *page)
3427 {
3428         return mpage_readpage(page, ext4_get_block);
3429 }
3430
3431 static int
3432 ext4_readpages(struct file *file, struct address_space *mapping,
3433                 struct list_head *pages, unsigned nr_pages)
3434 {
3435         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3436 }
3437
3438 static void ext4_free_io_end(ext4_io_end_t *io)
3439 {
3440         BUG_ON(!io);
3441         if (io->page)
3442                 put_page(io->page);
3443         iput(io->inode);
3444         kfree(io);
3445 }
3446
3447 static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
3448 {
3449         struct buffer_head *head, *bh;
3450         unsigned int curr_off = 0;
3451
3452         if (!page_has_buffers(page))
3453                 return;
3454         head = bh = page_buffers(page);
3455         do {
3456                 if (offset <= curr_off && test_clear_buffer_uninit(bh)
3457                                         && bh->b_private) {
3458                         ext4_free_io_end(bh->b_private);
3459                         bh->b_private = NULL;
3460                         bh->b_end_io = NULL;
3461                 }
3462                 curr_off = curr_off + bh->b_size;
3463                 bh = bh->b_this_page;
3464         } while (bh != head);
3465 }
3466
3467 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3468 {
3469         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3470
3471         /*
3472          * free any io_end structure allocated for buffers to be discarded
3473          */
3474         if (ext4_should_dioread_nolock(page->mapping->host))
3475                 ext4_invalidatepage_free_endio(page, offset);
3476         /*
3477          * If it's a full truncate we just forget about the pending dirtying
3478          */
3479         if (offset == 0)
3480                 ClearPageChecked(page);
3481
3482         if (journal)
3483                 jbd2_journal_invalidatepage(journal, page, offset);
3484         else
3485                 block_invalidatepage(page, offset);
3486 }
3487
3488 static int ext4_releasepage(struct page *page, gfp_t wait)
3489 {
3490         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3491
3492         WARN_ON(PageChecked(page));
3493         if (!page_has_buffers(page))
3494                 return 0;
3495         if (journal)
3496                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3497         else
3498                 return try_to_free_buffers(page);
3499 }
3500
3501 /*
3502  * O_DIRECT for ext3 (or indirect map) based files
3503  *
3504  * If the O_DIRECT write will extend the file then add this inode to the
3505  * orphan list.  So recovery will truncate it back to the original size
3506  * if the machine crashes during the write.
3507  *
3508  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3509  * crashes then stale disk data _may_ be exposed inside the file. But current
3510  * VFS code falls back into buffered path in that case so we are safe.
3511  */
3512 static ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
3513                               const struct iovec *iov, loff_t offset,
3514                               unsigned long nr_segs)
3515 {
3516         struct file *file = iocb->ki_filp;
3517         struct inode *inode = file->f_mapping->host;
3518         struct ext4_inode_info *ei = EXT4_I(inode);
3519         handle_t *handle;
3520         ssize_t ret;
3521         int orphan = 0;
3522         size_t count = iov_length(iov, nr_segs);
3523         int retries = 0;
3524
3525         if (rw == WRITE) {
3526                 loff_t final_size = offset + count;
3527
3528                 if (final_size > inode->i_size) {
3529                         /* Credits for sb + inode write */
3530                         handle = ext4_journal_start(inode, 2);
3531                         if (IS_ERR(handle)) {
3532                                 ret = PTR_ERR(handle);
3533                                 goto out;
3534                         }
3535                         ret = ext4_orphan_add(handle, inode);
3536                         if (ret) {
3537                                 ext4_journal_stop(handle);
3538                                 goto out;
3539                         }
3540                         orphan = 1;
3541                         ei->i_disksize = inode->i_size;
3542                         ext4_journal_stop(handle);
3543                 }
3544         }
3545
3546 retry:
3547         if (rw == READ && ext4_should_dioread_nolock(inode))
3548                 ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
3549                                  inode->i_sb->s_bdev, iov,
3550                                  offset, nr_segs,
3551                                  ext4_get_block, NULL);
3552         else
3553                 ret = blockdev_direct_IO(rw, iocb, inode,
3554                                  inode->i_sb->s_bdev, iov,
3555                                  offset, nr_segs,
3556                                  ext4_get_block, NULL);
3557         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3558                 goto retry;
3559
3560         if (orphan) {
3561                 int err;
3562
3563                 /* Credits for sb + inode write */
3564                 handle = ext4_journal_start(inode, 2);
3565                 if (IS_ERR(handle)) {
3566                         /* This is really bad luck. We've written the data
3567                          * but cannot extend i_size. Bail out and pretend
3568                          * the write failed... */
3569                         ret = PTR_ERR(handle);
3570                         if (inode->i_nlink)
3571                                 ext4_orphan_del(NULL, inode);
3572
3573                         goto out;
3574                 }
3575                 if (inode->i_nlink)
3576                         ext4_orphan_del(handle, inode);
3577                 if (ret > 0) {
3578                         loff_t end = offset + ret;
3579                         if (end > inode->i_size) {
3580                                 ei->i_disksize = end;
3581                                 i_size_write(inode, end);
3582                                 /*
3583                                  * We're going to return a positive `ret'
3584                                  * here due to non-zero-length I/O, so there's
3585                                  * no way of reporting error returns from
3586                                  * ext4_mark_inode_dirty() to userspace.  So
3587                                  * ignore it.
3588                                  */
3589                                 ext4_mark_inode_dirty(handle, inode);
3590                         }
3591                 }
3592                 err = ext4_journal_stop(handle);
3593                 if (ret == 0)
3594                         ret = err;
3595         }
3596 out:
3597         return ret;
3598 }
3599
3600 /*
3601  * ext4_get_block used when preparing for a DIO write or buffer write.
3602  * We allocate an uinitialized extent if blocks haven't been allocated.
3603  * The extent will be converted to initialized after the IO is complete.
3604  */
3605 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
3606                    struct buffer_head *bh_result, int create)
3607 {
3608         ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
3609                    inode->i_ino, create);
3610         return _ext4_get_block(inode, iblock, bh_result,
3611                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
3612 }
3613
3614 static void dump_completed_IO(struct inode * inode)
3615 {
3616 #ifdef  EXT4_DEBUG
3617         struct list_head *cur, *before, *after;
3618         ext4_io_end_t *io, *io0, *io1;
3619         unsigned long flags;
3620
3621         if (list_empty(&EXT4_I(inode)->i_completed_io_list)){
3622                 ext4_debug("inode %lu completed_io list is empty\n", inode->i_ino);
3623                 return;
3624         }
3625
3626         ext4_debug("Dump inode %lu completed_io list \n", inode->i_ino);
3627         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3628         list_for_each_entry(io, &EXT4_I(inode)->i_completed_io_list, list){
3629                 cur = &io->list;
3630                 before = cur->prev;
3631                 io0 = container_of(before, ext4_io_end_t, list);
3632                 after = cur->next;
3633                 io1 = container_of(after, ext4_io_end_t, list);
3634
3635                 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
3636                             io, inode->i_ino, io0, io1);
3637         }
3638         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3639 #endif
3640 }
3641
3642 /*
3643  * check a range of space and convert unwritten extents to written.
3644  */
3645 static int ext4_end_io_nolock(ext4_io_end_t *io)
3646 {
3647         struct inode *inode = io->inode;
3648         loff_t offset = io->offset;
3649         ssize_t size = io->size;
3650         int ret = 0;
3651
3652         ext4_debug("ext4_end_io_nolock: io 0x%p from inode %lu,list->next 0x%p,"
3653                    "list->prev 0x%p\n",
3654                    io, inode->i_ino, io->list.next, io->list.prev);
3655
3656         if (list_empty(&io->list))
3657                 return ret;
3658
3659         if (io->flag != EXT4_IO_UNWRITTEN)
3660                 return ret;
3661
3662         ret = ext4_convert_unwritten_extents(inode, offset, size);
3663         if (ret < 0) {
3664                 printk(KERN_EMERG "%s: failed to convert unwritten"
3665                         "extents to written extents, error is %d"
3666                         " io is still on inode %lu aio dio list\n",
3667                        __func__, ret, inode->i_ino);
3668                 return ret;
3669         }
3670
3671         if (io->iocb)
3672                 aio_complete(io->iocb, io->result, 0);
3673         /* clear the DIO AIO unwritten flag */
3674         io->flag = 0;
3675         return ret;
3676 }
3677
3678 /*
3679  * work on completed aio dio IO, to convert unwritten extents to extents
3680  */
3681 static void ext4_end_io_work(struct work_struct *work)
3682 {
3683         ext4_io_end_t           *io = container_of(work, ext4_io_end_t, work);
3684         struct inode            *inode = io->inode;
3685         struct ext4_inode_info  *ei = EXT4_I(inode);
3686         unsigned long           flags;
3687         int                     ret;
3688
3689         mutex_lock(&inode->i_mutex);
3690         ret = ext4_end_io_nolock(io);
3691         if (ret < 0) {
3692                 mutex_unlock(&inode->i_mutex);
3693                 return;
3694         }
3695
3696         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3697         if (!list_empty(&io->list))
3698                 list_del_init(&io->list);
3699         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3700         mutex_unlock(&inode->i_mutex);
3701         ext4_free_io_end(io);
3702 }
3703
3704 /*
3705  * This function is called from ext4_sync_file().
3706  *
3707  * When IO is completed, the work to convert unwritten extents to
3708  * written is queued on workqueue but may not get immediately
3709  * scheduled. When fsync is called, we need to ensure the
3710  * conversion is complete before fsync returns.
3711  * The inode keeps track of a list of pending/completed IO that
3712  * might needs to do the conversion. This function walks through
3713  * the list and convert the related unwritten extents for completed IO
3714  * to written.
3715  * The function return the number of pending IOs on success.
3716  */
3717 int flush_completed_IO(struct inode *inode)
3718 {
3719         ext4_io_end_t *io;
3720         struct ext4_inode_info *ei = EXT4_I(inode);
3721         unsigned long flags;
3722         int ret = 0;
3723         int ret2 = 0;
3724
3725         if (list_empty(&ei->i_completed_io_list))
3726                 return ret;
3727
3728         dump_completed_IO(inode);
3729         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3730         while (!list_empty(&ei->i_completed_io_list)){
3731                 io = list_entry(ei->i_completed_io_list.next,
3732                                 ext4_io_end_t, list);
3733                 /*
3734                  * Calling ext4_end_io_nolock() to convert completed
3735                  * IO to written.
3736                  *
3737                  * When ext4_sync_file() is called, run_queue() may already
3738                  * about to flush the work corresponding to this io structure.
3739                  * It will be upset if it founds the io structure related
3740                  * to the work-to-be schedule is freed.
3741                  *
3742                  * Thus we need to keep the io structure still valid here after
3743                  * convertion finished. The io structure has a flag to
3744                  * avoid double converting from both fsync and background work
3745                  * queue work.
3746                  */
3747                 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3748                 ret = ext4_end_io_nolock(io);
3749                 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3750                 if (ret < 0)
3751                         ret2 = ret;
3752                 else
3753                         list_del_init(&io->list);
3754         }
3755         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3756         return (ret2 < 0) ? ret2 : 0;
3757 }
3758
3759 static ext4_io_end_t *ext4_init_io_end (struct inode *inode, gfp_t flags)
3760 {
3761         ext4_io_end_t *io = NULL;
3762
3763         io = kmalloc(sizeof(*io), flags);
3764
3765         if (io) {
3766                 igrab(inode);
3767                 io->inode = inode;
3768                 io->flag = 0;
3769                 io->offset = 0;
3770                 io->size = 0;
3771                 io->page = NULL;
3772                 io->iocb = NULL;
3773                 io->result = 0;
3774                 INIT_WORK(&io->work, ext4_end_io_work);
3775                 INIT_LIST_HEAD(&io->list);
3776         }
3777
3778         return io;
3779 }
3780
3781 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3782                             ssize_t size, void *private, int ret,
3783                             bool is_async)
3784 {
3785         ext4_io_end_t *io_end = iocb->private;
3786         struct workqueue_struct *wq;
3787         unsigned long flags;
3788         struct ext4_inode_info *ei;
3789
3790         /* if not async direct IO or dio with 0 bytes write, just return */
3791         if (!io_end || !size)
3792                 goto out;
3793
3794         ext_debug("ext4_end_io_dio(): io_end 0x%p"
3795                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
3796                   iocb->private, io_end->inode->i_ino, iocb, offset,
3797                   size);
3798
3799         /* if not aio dio with unwritten extents, just free io and return */
3800         if (io_end->flag != EXT4_IO_UNWRITTEN){
3801                 ext4_free_io_end(io_end);
3802                 iocb->private = NULL;
3803 out:
3804                 if (is_async)
3805                         aio_complete(iocb, ret, 0);
3806                 return;
3807         }
3808
3809         io_end->offset = offset;
3810         io_end->size = size;
3811         if (is_async) {
3812                 io_end->iocb = iocb;
3813                 io_end->result = ret;
3814         }
3815         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
3816
3817         /* queue the work to convert unwritten extents to written */
3818         queue_work(wq, &io_end->work);
3819
3820         /* Add the io_end to per-inode completed aio dio list*/
3821         ei = EXT4_I(io_end->inode);
3822         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3823         list_add_tail(&io_end->list, &ei->i_completed_io_list);
3824         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3825         iocb->private = NULL;
3826 }
3827
3828 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
3829 {
3830         ext4_io_end_t *io_end = bh->b_private;
3831         struct workqueue_struct *wq;
3832         struct inode *inode;
3833         unsigned long flags;
3834
3835         if (!test_clear_buffer_uninit(bh) || !io_end)
3836                 goto out;
3837
3838         if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
3839                 printk("sb umounted, discard end_io request for inode %lu\n",
3840                         io_end->inode->i_ino);
3841                 ext4_free_io_end(io_end);
3842                 goto out;
3843         }
3844
3845         io_end->flag = EXT4_IO_UNWRITTEN;
3846         inode = io_end->inode;
3847
3848         /* Add the io_end to per-inode completed io list*/
3849         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3850         list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
3851         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3852
3853         wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
3854         /* queue the work to convert unwritten extents to written */
3855         queue_work(wq, &io_end->work);
3856 out:
3857         bh->b_private = NULL;
3858         bh->b_end_io = NULL;
3859         clear_buffer_uninit(bh);
3860         end_buffer_async_write(bh, uptodate);
3861 }
3862
3863 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
3864 {
3865         ext4_io_end_t *io_end;
3866         struct page *page = bh->b_page;
3867         loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
3868         size_t size = bh->b_size;
3869
3870 retry:
3871         io_end = ext4_init_io_end(inode, GFP_ATOMIC);
3872         if (!io_end) {
3873                 if (printk_ratelimit())
3874                         printk(KERN_WARNING "%s: allocation fail\n", __func__);
3875                 schedule();
3876                 goto retry;
3877         }
3878         io_end->offset = offset;
3879         io_end->size = size;
3880         /*
3881          * We need to hold a reference to the page to make sure it
3882          * doesn't get evicted before ext4_end_io_work() has a chance
3883          * to convert the extent from written to unwritten.
3884          */
3885         io_end->page = page;
3886         get_page(io_end->page);
3887
3888         bh->b_private = io_end;
3889         bh->b_end_io = ext4_end_io_buffer_write;
3890         return 0;
3891 }
3892
3893 /*
3894  * For ext4 extent files, ext4 will do direct-io write to holes,
3895  * preallocated extents, and those write extend the file, no need to
3896  * fall back to buffered IO.
3897  *
3898  * For holes, we fallocate those blocks, mark them as unintialized
3899  * If those blocks were preallocated, we mark sure they are splited, but
3900  * still keep the range to write as unintialized.
3901  *
3902  * The unwrritten extents will be converted to written when DIO is completed.
3903  * For async direct IO, since the IO may still pending when return, we
3904  * set up an end_io call back function, which will do the convertion
3905  * when async direct IO completed.
3906  *
3907  * If the O_DIRECT write will extend the file then add this inode to the
3908  * orphan list.  So recovery will truncate it back to the original size
3909  * if the machine crashes during the write.
3910  *
3911  */
3912 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
3913                               const struct iovec *iov, loff_t offset,
3914                               unsigned long nr_segs)
3915 {
3916         struct file *file = iocb->ki_filp;
3917         struct inode *inode = file->f_mapping->host;
3918         ssize_t ret;
3919         size_t count = iov_length(iov, nr_segs);
3920
3921         loff_t final_size = offset + count;
3922         if (rw == WRITE && final_size <= inode->i_size) {
3923                 /*
3924                  * We could direct write to holes and fallocate.
3925                  *
3926                  * Allocated blocks to fill the hole are marked as uninitialized
3927                  * to prevent paralel buffered read to expose the stale data
3928                  * before DIO complete the data IO.
3929                  *
3930                  * As to previously fallocated extents, ext4 get_block
3931                  * will just simply mark the buffer mapped but still
3932                  * keep the extents uninitialized.
3933                  *
3934                  * for non AIO case, we will convert those unwritten extents
3935                  * to written after return back from blockdev_direct_IO.
3936                  *
3937                  * for async DIO, the conversion needs to be defered when
3938                  * the IO is completed. The ext4 end_io callback function
3939                  * will be called to take care of the conversion work.
3940                  * Here for async case, we allocate an io_end structure to
3941                  * hook to the iocb.
3942                  */
3943                 iocb->private = NULL;
3944                 EXT4_I(inode)->cur_aio_dio = NULL;
3945                 if (!is_sync_kiocb(iocb)) {
3946                         iocb->private = ext4_init_io_end(inode, GFP_NOFS);
3947                         if (!iocb->private)
3948                                 return -ENOMEM;
3949                         /*
3950                          * we save the io structure for current async
3951                          * direct IO, so that later ext4_map_blocks()
3952                          * could flag the io structure whether there
3953                          * is a unwritten extents needs to be converted
3954                          * when IO is completed.
3955                          */
3956                         EXT4_I(inode)->cur_aio_dio = iocb->private;
3957                 }
3958
3959                 ret = blockdev_direct_IO(rw, iocb, inode,
3960                                          inode->i_sb->s_bdev, iov,
3961                                          offset, nr_segs,
3962                                          ext4_get_block_write,
3963                                          ext4_end_io_dio);
3964                 if (iocb->private)
3965                         EXT4_I(inode)->cur_aio_dio = NULL;
3966                 /*
3967                  * The io_end structure takes a reference to the inode,
3968                  * that structure needs to be destroyed and the
3969                  * reference to the inode need to be dropped, when IO is
3970                  * complete, even with 0 byte write, or failed.
3971                  *
3972                  * In the successful AIO DIO case, the io_end structure will be
3973                  * desctroyed and the reference to the inode will be dropped
3974                  * after the end_io call back function is called.
3975                  *
3976                  * In the case there is 0 byte write, or error case, since
3977                  * VFS direct IO won't invoke the end_io call back function,
3978                  * we need to free the end_io structure here.
3979                  */
3980                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
3981                         ext4_free_io_end(iocb->private);
3982                         iocb->private = NULL;
3983                 } else if (ret > 0 && ext4_test_inode_state(inode,
3984                                                 EXT4_STATE_DIO_UNWRITTEN)) {
3985                         int err;
3986                         /*
3987                          * for non AIO case, since the IO is already
3988                          * completed, we could do the convertion right here
3989                          */
3990                         err = ext4_convert_unwritten_extents(inode,
3991                                                              offset, ret);
3992                         if (err < 0)
3993                                 ret = err;
3994                         ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3995                 }
3996                 return ret;
3997         }
3998
3999         /* for write the the end of file case, we fall back to old way */
4000         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
4001 }
4002
4003 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
4004                               const struct iovec *iov, loff_t offset,
4005                               unsigned long nr_segs)
4006 {
4007         struct file *file = iocb->ki_filp;
4008         struct inode *inode = file->f_mapping->host;
4009
4010         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
4011                 return ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
4012
4013         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
4014 }
4015
4016 /*
4017  * Pages can be marked dirty completely asynchronously from ext4's journalling
4018  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
4019  * much here because ->set_page_dirty is called under VFS locks.  The page is
4020  * not necessarily locked.
4021  *
4022  * We cannot just dirty the page and leave attached buffers clean, because the
4023  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
4024  * or jbddirty because all the journalling code will explode.
4025  *
4026  * So what we do is to mark the page "pending dirty" and next time writepage
4027  * is called, propagate that into the buffers appropriately.
4028  */
4029 static int ext4_journalled_set_page_dirty(struct page *page)
4030 {
4031         SetPageChecked(page);
4032         return __set_page_dirty_nobuffers(page);
4033 }
4034
4035 static const struct address_space_operations ext4_ordered_aops = {
4036         .readpage               = ext4_readpage,
4037         .readpages              = ext4_readpages,
4038         .writepage              = ext4_writepage,
4039         .sync_page              = block_sync_page,
4040         .write_begin            = ext4_write_begin,
4041         .write_end              = ext4_ordered_write_end,
4042         .bmap                   = ext4_bmap,
4043         .invalidatepage         = ext4_invalidatepage,
4044         .releasepage            = ext4_releasepage,
4045         .direct_IO              = ext4_direct_IO,
4046         .migratepage            = buffer_migrate_page,
4047         .is_partially_uptodate  = block_is_partially_uptodate,
4048         .error_remove_page      = generic_error_remove_page,
4049 };
4050
4051 static const struct address_space_operations ext4_writeback_aops = {
4052         .readpage               = ext4_readpage,
4053         .readpages              = ext4_readpages,
4054         .writepage              = ext4_writepage,
4055         .sync_page              = block_sync_page,
4056         .write_begin            = ext4_write_begin,
4057         .write_end              = ext4_writeback_write_end,
4058         .bmap                   = ext4_bmap,
4059         .invalidatepage         = ext4_invalidatepage,
4060         .releasepage            = ext4_releasepage,
4061         .direct_IO              = ext4_direct_IO,
4062         .migratepage            = buffer_migrate_page,
4063         .is_partially_uptodate  = block_is_partially_uptodate,
4064         .error_remove_page      = generic_error_remove_page,
4065 };
4066
4067 static const struct address_space_operations ext4_journalled_aops = {
4068         .readpage               = ext4_readpage,
4069         .readpages              = ext4_readpages,
4070         .writepage              = ext4_writepage,
4071         .sync_page              = block_sync_page,
4072         .write_begin            = ext4_write_begin,
4073         .write_end              = ext4_journalled_write_end,
4074         .set_page_dirty         = ext4_journalled_set_page_dirty,
4075         .bmap                   = ext4_bmap,
4076         .invalidatepage         = ext4_invalidatepage,
4077         .releasepage            = ext4_releasepage,
4078         .is_partially_uptodate  = block_is_partially_uptodate,
4079         .error_remove_page      = generic_error_remove_page,
4080 };
4081
4082 static const struct address_space_operations ext4_da_aops = {
4083         .readpage               = ext4_readpage,
4084         .readpages              = ext4_readpages,
4085         .writepage              = ext4_writepage,
4086         .writepages             = ext4_da_writepages,
4087         .sync_page              = block_sync_page,
4088         .write_begin            = ext4_da_write_begin,
4089         .write_end              = ext4_da_write_end,
4090         .bmap                   = ext4_bmap,
4091         .invalidatepage         = ext4_da_invalidatepage,
4092         .releasepage            = ext4_releasepage,
4093         .direct_IO              = ext4_direct_IO,
4094         .migratepage            = buffer_migrate_page,
4095         .is_partially_uptodate  = block_is_partially_uptodate,
4096         .error_remove_page      = generic_error_remove_page,
4097 };
4098
4099 void ext4_set_aops(struct inode *inode)
4100 {
4101         if (ext4_should_order_data(inode) &&
4102                 test_opt(inode->i_sb, DELALLOC))
4103                 inode->i_mapping->a_ops = &ext4_da_aops;
4104         else if (ext4_should_order_data(inode))
4105                 inode->i_mapping->a_ops = &ext4_ordered_aops;
4106         else if (ext4_should_writeback_data(inode) &&
4107                  test_opt(inode->i_sb, DELALLOC))
4108                 inode->i_mapping->a_ops = &ext4_da_aops;
4109         else if (ext4_should_writeback_data(inode))
4110                 inode->i_mapping->a_ops = &ext4_writeback_aops;
4111         else
4112                 inode->i_mapping->a_ops = &ext4_journalled_aops;
4113 }
4114
4115 /*
4116  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
4117  * up to the end of the block which corresponds to `from'.
4118  * This required during truncate. We need to physically zero the tail end
4119  * of that block so it doesn't yield old data if the file is later grown.
4120  */
4121 int ext4_block_truncate_page(handle_t *handle,
4122                 struct address_space *mapping, loff_t from)
4123 {
4124         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
4125         unsigned offset = from & (PAGE_CACHE_SIZE-1);
4126         unsigned blocksize, length, pos;
4127         ext4_lblk_t iblock;
4128         struct inode *inode = mapping->host;
4129         struct buffer_head *bh;
4130         struct page *page;
4131         int err = 0;
4132
4133         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
4134                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
4135         if (!page)
4136                 return -EINVAL;
4137
4138         blocksize = inode->i_sb->s_blocksize;
4139         length = blocksize - (offset & (blocksize - 1));
4140         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
4141
4142         if (!page_has_buffers(page))
4143                 create_empty_buffers(page, blocksize, 0);
4144
4145         /* Find the buffer that contains "offset" */
4146         bh = page_buffers(page);
4147         pos = blocksize;
4148         while (offset >= pos) {
4149                 bh = bh->b_this_page;
4150                 iblock++;
4151                 pos += blocksize;
4152         }
4153
4154         err = 0;
4155         if (buffer_freed(bh)) {
4156                 BUFFER_TRACE(bh, "freed: skip");
4157                 goto unlock;
4158         }
4159
4160         if (!buffer_mapped(bh)) {
4161                 BUFFER_TRACE(bh, "unmapped");
4162                 ext4_get_block(inode, iblock, bh, 0);
4163                 /* unmapped? It's a hole - nothing to do */
4164                 if (!buffer_mapped(bh)) {
4165                         BUFFER_TRACE(bh, "still unmapped");
4166                         goto unlock;
4167                 }
4168         }
4169
4170         /* Ok, it's mapped. Make sure it's up-to-date */
4171         if (PageUptodate(page))
4172                 set_buffer_uptodate(bh);
4173
4174         if (!buffer_uptodate(bh)) {
4175                 err = -EIO;
4176                 ll_rw_block(READ, 1, &bh);
4177                 wait_on_buffer(bh);
4178                 /* Uhhuh. Read error. Complain and punt. */
4179                 if (!buffer_uptodate(bh))
4180                         goto unlock;
4181         }
4182
4183         if (ext4_should_journal_data(inode)) {
4184                 BUFFER_TRACE(bh, "get write access");
4185                 err = ext4_journal_get_write_access(handle, bh);
4186                 if (err)
4187                         goto unlock;
4188         }
4189
4190         zero_user(page, offset, length);
4191
4192         BUFFER_TRACE(bh, "zeroed end of block");
4193
4194         err = 0;
4195         if (ext4_should_journal_data(inode)) {
4196                 err = ext4_handle_dirty_metadata(handle, inode, bh);
4197         } else {
4198                 if (ext4_should_order_data(inode))
4199                         err = ext4_jbd2_file_inode(handle, inode);
4200                 mark_buffer_dirty(bh);
4201         }
4202
4203 unlock:
4204         unlock_page(page);
4205         page_cache_release(page);
4206         return err;
4207 }
4208
4209 /*
4210  * Probably it should be a library function... search for first non-zero word
4211  * or memcmp with zero_page, whatever is better for particular architecture.
4212  * Linus?
4213  */
4214 static inline int all_zeroes(__le32 *p, __le32 *q)
4215 {
4216         while (p < q)
4217                 if (*p++)
4218                         return 0;
4219         return 1;
4220 }
4221
4222 /**
4223  *      ext4_find_shared - find the indirect blocks for partial truncation.
4224  *      @inode:   inode in question
4225  *      @depth:   depth of the affected branch
4226  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
4227  *      @chain:   place to store the pointers to partial indirect blocks
4228  *      @top:     place to the (detached) top of branch
4229  *
4230  *      This is a helper function used by ext4_truncate().
4231  *
4232  *      When we do truncate() we may have to clean the ends of several
4233  *      indirect blocks but leave the blocks themselves alive. Block is
4234  *      partially truncated if some data below the new i_size is refered
4235  *      from it (and it is on the path to the first completely truncated
4236  *      data block, indeed).  We have to free the top of that path along
4237  *      with everything to the right of the path. Since no allocation
4238  *      past the truncation point is possible until ext4_truncate()
4239  *      finishes, we may safely do the latter, but top of branch may
4240  *      require special attention - pageout below the truncation point
4241  *      might try to populate it.
4242  *
4243  *      We atomically detach the top of branch from the tree, store the
4244  *      block number of its root in *@top, pointers to buffer_heads of
4245  *      partially truncated blocks - in @chain[].bh and pointers to
4246  *      their last elements that should not be removed - in
4247  *      @chain[].p. Return value is the pointer to last filled element
4248  *      of @chain.
4249  *
4250  *      The work left to caller to do the actual freeing of subtrees:
4251  *              a) free the subtree starting from *@top
4252  *              b) free the subtrees whose roots are stored in
4253  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
4254  *              c) free the subtrees growing from the inode past the @chain[0].
4255  *                      (no partially truncated stuff there).  */
4256
4257 static Indirect *ext4_find_shared(struct inode *inode, int depth,
4258                                   ext4_lblk_t offsets[4], Indirect chain[4],
4259                                   __le32 *top)
4260 {
4261         Indirect *partial, *p;
4262         int k, err;
4263
4264         *top = 0;
4265         /* Make k index the deepest non-null offset + 1 */
4266         for (k = depth; k > 1 && !offsets[k-1]; k--)
4267                 ;
4268         partial = ext4_get_branch(inode, k, offsets, chain, &err);
4269         /* Writer: pointers */
4270         if (!partial)
4271                 partial = chain + k-1;
4272         /*
4273          * If the branch acquired continuation since we've looked at it -
4274          * fine, it should all survive and (new) top doesn't belong to us.
4275          */
4276         if (!partial->key && *partial->p)
4277                 /* Writer: end */
4278                 goto no_top;
4279         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
4280                 ;
4281         /*
4282          * OK, we've found the last block that must survive. The rest of our
4283          * branch should be detached before unlocking. However, if that rest
4284          * of branch is all ours and does not grow immediately from the inode
4285          * it's easier to cheat and just decrement partial->p.
4286          */
4287         if (p == chain + k - 1 && p > chain) {
4288                 p->p--;
4289         } else {
4290                 *top = *p->p;
4291                 /* Nope, don't do this in ext4.  Must leave the tree intact */
4292 #if 0
4293                 *p->p = 0;
4294 #endif
4295         }
4296         /* Writer: end */
4297
4298         while (partial > p) {
4299                 brelse(partial->bh);
4300                 partial--;
4301         }
4302 no_top:
4303         return partial;
4304 }
4305
4306 /*
4307  * Zero a number of block pointers in either an inode or an indirect block.
4308  * If we restart the transaction we must again get write access to the
4309  * indirect block for further modification.
4310  *
4311  * We release `count' blocks on disk, but (last - first) may be greater
4312  * than `count' because there can be holes in there.
4313  */
4314 static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
4315                              struct buffer_head *bh,
4316                              ext4_fsblk_t block_to_free,
4317                              unsigned long count, __le32 *first,
4318                              __le32 *last)
4319 {
4320         __le32 *p;
4321         int     flags = EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_VALIDATED;
4322
4323         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
4324                 flags |= EXT4_FREE_BLOCKS_METADATA;
4325
4326         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
4327                                    count)) {
4328                 EXT4_ERROR_INODE(inode, "attempt to clear invalid "
4329                                  "blocks %llu len %lu",
4330                                  (unsigned long long) block_to_free, count);
4331                 return 1;
4332         }
4333
4334         if (try_to_extend_transaction(handle, inode)) {
4335                 if (bh) {
4336                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4337                         ext4_handle_dirty_metadata(handle, inode, bh);
4338                 }
4339                 ext4_mark_inode_dirty(handle, inode);
4340                 ext4_truncate_restart_trans(handle, inode,
4341                                             blocks_for_truncate(inode));
4342                 if (bh) {
4343                         BUFFER_TRACE(bh, "retaking write access");
4344                         ext4_journal_get_write_access(handle, bh);
4345                 }
4346         }
4347
4348         for (p = first; p < last; p++)
4349                 *p = 0;
4350
4351         ext4_free_blocks(handle, inode, 0, block_to_free, count, flags);
4352         return 0;
4353 }
4354
4355 /**
4356  * ext4_free_data - free a list of data blocks
4357  * @handle:     handle for this transaction
4358  * @inode:      inode we are dealing with
4359  * @this_bh:    indirect buffer_head which contains *@first and *@last
4360  * @first:      array of block numbers
4361  * @last:       points immediately past the end of array
4362  *
4363  * We are freeing all blocks refered from that array (numbers are stored as
4364  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
4365  *
4366  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
4367  * blocks are contiguous then releasing them at one time will only affect one
4368  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
4369  * actually use a lot of journal space.
4370  *
4371  * @this_bh will be %NULL if @first and @last point into the inode's direct
4372  * block pointers.
4373  */
4374 static void ext4_free_data(handle_t *handle, struct inode *inode,
4375                            struct buffer_head *this_bh,
4376                            __le32 *first, __le32 *last)
4377 {
4378         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
4379         unsigned long count = 0;            /* Number of blocks in the run */
4380         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
4381                                                corresponding to
4382                                                block_to_free */
4383         ext4_fsblk_t nr;                    /* Current block # */
4384         __le32 *p;                          /* Pointer into inode/ind
4385                                                for current block */
4386         int err;
4387
4388         if (this_bh) {                          /* For indirect block */
4389                 BUFFER_TRACE(this_bh, "get_write_access");
4390                 err = ext4_journal_get_write_access(handle, this_bh);
4391                 /* Important: if we can't update the indirect pointers
4392                  * to the blocks, we can't free them. */
4393                 if (err)
4394                         return;
4395         }
4396
4397         for (p = first; p < last; p++) {
4398                 nr = le32_to_cpu(*p);
4399                 if (nr) {
4400                         /* accumulate blocks to free if they're contiguous */
4401                         if (count == 0) {
4402                                 block_to_free = nr;
4403                                 block_to_free_p = p;
4404                                 count = 1;
4405                         } else if (nr == block_to_free + count) {
4406                                 count++;
4407                         } else {
4408                                 if (ext4_clear_blocks(handle, inode, this_bh,
4409                                                       block_to_free, count,
4410                                                       block_to_free_p, p))
4411                                         break;
4412                                 block_to_free = nr;
4413                                 block_to_free_p = p;
4414                                 count = 1;
4415                         }
4416                 }
4417         }
4418
4419         if (count > 0)
4420                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
4421                                   count, block_to_free_p, p);
4422
4423         if (this_bh) {
4424                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
4425
4426                 /*
4427                  * The buffer head should have an attached journal head at this
4428                  * point. However, if the data is corrupted and an indirect
4429                  * block pointed to itself, it would have been detached when
4430                  * the block was cleared. Check for this instead of OOPSing.
4431                  */
4432                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
4433                         ext4_handle_dirty_metadata(handle, inode, this_bh);
4434                 else
4435                         EXT4_ERROR_INODE(inode,
4436                                          "circular indirect block detected at "
4437                                          "block %llu",
4438                                 (unsigned long long) this_bh->b_blocknr);
4439         }
4440 }
4441
4442 /**
4443  *      ext4_free_branches - free an array of branches
4444  *      @handle: JBD handle for this transaction
4445  *      @inode: inode we are dealing with
4446  *      @parent_bh: the buffer_head which contains *@first and *@last
4447  *      @first: array of block numbers
4448  *      @last:  pointer immediately past the end of array
4449  *      @depth: depth of the branches to free
4450  *
4451  *      We are freeing all blocks refered from these branches (numbers are
4452  *      stored as little-endian 32-bit) and updating @inode->i_blocks
4453  *      appropriately.
4454  */
4455 static void ext4_free_branches(handle_t *handle, struct inode *inode,
4456                                struct buffer_head *parent_bh,
4457                                __le32 *first, __le32 *last, int depth)
4458 {
4459         ext4_fsblk_t nr;
4460         __le32 *p;
4461
4462         if (ext4_handle_is_aborted(handle))
4463                 return;
4464
4465         if (depth--) {
4466                 struct buffer_head *bh;
4467                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4468                 p = last;
4469                 while (--p >= first) {
4470                         nr = le32_to_cpu(*p);
4471                         if (!nr)
4472                                 continue;               /* A hole */
4473
4474                         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
4475                                                    nr, 1)) {
4476                                 EXT4_ERROR_INODE(inode,
4477                                                  "invalid indirect mapped "
4478                                                  "block %lu (level %d)",
4479                                                  (unsigned long) nr, depth);
4480                                 break;
4481                         }
4482
4483                         /* Go read the buffer for the next level down */
4484                         bh = sb_bread(inode->i_sb, nr);
4485
4486                         /*
4487                          * A read failure? Report error and clear slot
4488                          * (should be rare).
4489                          */
4490                         if (!bh) {
4491                                 EXT4_ERROR_INODE_BLOCK(inode, nr,
4492                                                        "Read failure");
4493                                 continue;
4494                         }
4495
4496                         /* This zaps the entire block.  Bottom up. */
4497                         BUFFER_TRACE(bh, "free child branches");
4498                         ext4_free_branches(handle, inode, bh,
4499                                         (__le32 *) bh->b_data,
4500                                         (__le32 *) bh->b_data + addr_per_block,
4501                                         depth);
4502
4503                         /*
4504                          * Everything below this this pointer has been
4505                          * released.  Now let this top-of-subtree go.
4506                          *
4507                          * We want the freeing of this indirect block to be
4508                          * atomic in the journal with the updating of the
4509                          * bitmap block which owns it.  So make some room in
4510                          * the journal.
4511                          *
4512                          * We zero the parent pointer *after* freeing its
4513                          * pointee in the bitmaps, so if extend_transaction()
4514                          * for some reason fails to put the bitmap changes and
4515                          * the release into the same transaction, recovery
4516                          * will merely complain about releasing a free block,
4517                          * rather than leaking blocks.
4518                          */
4519                         if (ext4_handle_is_aborted(handle))
4520                                 return;
4521                         if (try_to_extend_transaction(handle, inode)) {
4522                                 ext4_mark_inode_dirty(handle, inode);
4523                                 ext4_truncate_restart_trans(handle, inode,
4524                                             blocks_for_truncate(inode));
4525                         }
4526
4527                         /*
4528                          * The forget flag here is critical because if
4529                          * we are journaling (and not doing data
4530                          * journaling), we have to make sure a revoke
4531                          * record is written to prevent the journal
4532                          * replay from overwriting the (former)
4533                          * indirect block if it gets reallocated as a
4534                          * data block.  This must happen in the same
4535                          * transaction where the data blocks are
4536                          * actually freed.
4537                          */
4538                         ext4_free_blocks(handle, inode, 0, nr, 1,
4539                                          EXT4_FREE_BLOCKS_METADATA|
4540                                          EXT4_FREE_BLOCKS_FORGET);
4541
4542                         if (parent_bh) {
4543                                 /*
4544                                  * The block which we have just freed is
4545                                  * pointed to by an indirect block: journal it
4546                                  */
4547                                 BUFFER_TRACE(parent_bh, "get_write_access");
4548                                 if (!ext4_journal_get_write_access(handle,
4549                                                                    parent_bh)){
4550                                         *p = 0;
4551                                         BUFFER_TRACE(parent_bh,
4552                                         "call ext4_handle_dirty_metadata");
4553                                         ext4_handle_dirty_metadata(handle,
4554                                                                    inode,
4555                                                                    parent_bh);
4556                                 }
4557                         }
4558                 }
4559         } else {
4560                 /* We have reached the bottom of the tree. */
4561                 BUFFER_TRACE(parent_bh, "free data blocks");
4562                 ext4_free_data(handle, inode, parent_bh, first, last);
4563         }
4564 }
4565
4566 int ext4_can_truncate(struct inode *inode)
4567 {
4568         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4569                 return 0;
4570         if (S_ISREG(inode->i_mode))
4571                 return 1;
4572         if (S_ISDIR(inode->i_mode))
4573                 return 1;
4574         if (S_ISLNK(inode->i_mode))
4575                 return !ext4_inode_is_fast_symlink(inode);
4576         return 0;
4577 }
4578
4579 /*
4580  * ext4_truncate()
4581  *
4582  * We block out ext4_get_block() block instantiations across the entire
4583  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4584  * simultaneously on behalf of the same inode.
4585  *
4586  * As we work through the truncate and commmit bits of it to the journal there
4587  * is one core, guiding principle: the file's tree must always be consistent on
4588  * disk.  We must be able to restart the truncate after a crash.
4589  *
4590  * The file's tree may be transiently inconsistent in memory (although it
4591  * probably isn't), but whenever we close off and commit a journal transaction,
4592  * the contents of (the filesystem + the journal) must be consistent and
4593  * restartable.  It's pretty simple, really: bottom up, right to left (although
4594  * left-to-right works OK too).
4595  *
4596  * Note that at recovery time, journal replay occurs *before* the restart of
4597  * truncate against the orphan inode list.
4598  *
4599  * The committed inode has the new, desired i_size (which is the same as
4600  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4601  * that this inode's truncate did not complete and it will again call
4602  * ext4_truncate() to have another go.  So there will be instantiated blocks
4603  * to the right of the truncation point in a crashed ext4 filesystem.  But
4604  * that's fine - as long as they are linked from the inode, the post-crash
4605  * ext4_truncate() run will find them and release them.
4606  */
4607 void ext4_truncate(struct inode *inode)
4608 {
4609         handle_t *handle;
4610         struct ext4_inode_info *ei = EXT4_I(inode);
4611         __le32 *i_data = ei->i_data;
4612         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4613         struct address_space *mapping = inode->i_mapping;
4614         ext4_lblk_t offsets[4];
4615         Indirect chain[4];
4616         Indirect *partial;
4617         __le32 nr = 0;
4618         int n;
4619         ext4_lblk_t last_block;
4620         unsigned blocksize = inode->i_sb->s_blocksize;
4621
4622         if (!ext4_can_truncate(inode))
4623                 return;
4624
4625         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4626
4627         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4628                 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4629
4630         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
4631                 ext4_ext_truncate(inode);
4632                 return;
4633         }
4634
4635         handle = start_transaction(inode);
4636         if (IS_ERR(handle))
4637                 return;         /* AKPM: return what? */
4638
4639         last_block = (inode->i_size + blocksize-1)
4640                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4641
4642         if (inode->i_size & (blocksize - 1))
4643                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4644                         goto out_stop;
4645
4646         n = ext4_block_to_path(inode, last_block, offsets, NULL);
4647         if (n == 0)
4648                 goto out_stop;  /* error */
4649
4650         /*
4651          * OK.  This truncate is going to happen.  We add the inode to the
4652          * orphan list, so that if this truncate spans multiple transactions,
4653          * and we crash, we will resume the truncate when the filesystem
4654          * recovers.  It also marks the inode dirty, to catch the new size.
4655          *
4656          * Implication: the file must always be in a sane, consistent
4657          * truncatable state while each transaction commits.
4658          */
4659         if (ext4_orphan_add(handle, inode))
4660                 goto out_stop;
4661
4662         /*
4663          * From here we block out all ext4_get_block() callers who want to
4664          * modify the block allocation tree.
4665          */
4666         down_write(&ei->i_data_sem);
4667
4668         ext4_discard_preallocations(inode);
4669
4670         /*
4671          * The orphan list entry will now protect us from any crash which
4672          * occurs before the truncate completes, so it is now safe to propagate
4673          * the new, shorter inode size (held for now in i_size) into the
4674          * on-disk inode. We do this via i_disksize, which is the value which
4675          * ext4 *really* writes onto the disk inode.
4676          */
4677         ei->i_disksize = inode->i_size;
4678
4679         if (n == 1) {           /* direct blocks */
4680                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4681                                i_data + EXT4_NDIR_BLOCKS);
4682                 goto do_indirects;
4683         }
4684
4685         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4686         /* Kill the top of shared branch (not detached) */
4687         if (nr) {
4688                 if (partial == chain) {
4689                         /* Shared branch grows from the inode */
4690                         ext4_free_branches(handle, inode, NULL,
4691                                            &nr, &nr+1, (chain+n-1) - partial);
4692                         *partial->p = 0;
4693                         /*
4694                          * We mark the inode dirty prior to restart,
4695                          * and prior to stop.  No need for it here.
4696                          */
4697                 } else {
4698                         /* Shared branch grows from an indirect block */
4699                         BUFFER_TRACE(partial->bh, "get_write_access");
4700                         ext4_free_branches(handle, inode, partial->bh,
4701                                         partial->p,
4702                                         partial->p+1, (chain+n-1) - partial);
4703                 }
4704         }
4705         /* Clear the ends of indirect blocks on the shared branch */
4706         while (partial > chain) {
4707                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4708                                    (__le32*)partial->bh->b_data+addr_per_block,
4709                                    (chain+n-1) - partial);
4710                 BUFFER_TRACE(partial->bh, "call brelse");
4711                 brelse(partial->bh);
4712                 partial--;
4713         }
4714 do_indirects:
4715         /* Kill the remaining (whole) subtrees */
4716         switch (offsets[0]) {
4717         default:
4718                 nr = i_data[EXT4_IND_BLOCK];
4719                 if (nr) {
4720                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4721                         i_data[EXT4_IND_BLOCK] = 0;
4722                 }
4723         case EXT4_IND_BLOCK:
4724                 nr = i_data[EXT4_DIND_BLOCK];
4725                 if (nr) {
4726                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4727                         i_data[EXT4_DIND_BLOCK] = 0;
4728                 }
4729         case EXT4_DIND_BLOCK:
4730                 nr = i_data[EXT4_TIND_BLOCK];
4731                 if (nr) {
4732                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4733                         i_data[EXT4_TIND_BLOCK] = 0;
4734                 }
4735         case EXT4_TIND_BLOCK:
4736                 ;
4737         }
4738
4739         up_write(&ei->i_data_sem);
4740         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4741         ext4_mark_inode_dirty(handle, inode);
4742
4743         /*
4744          * In a multi-transaction truncate, we only make the final transaction
4745          * synchronous
4746          */
4747         if (IS_SYNC(inode))
4748                 ext4_handle_sync(handle);
4749 out_stop:
4750         /*
4751          * If this was a simple ftruncate(), and the file will remain alive
4752          * then we need to clear up the orphan record which we created above.
4753          * However, if this was a real unlink then we were called by
4754          * ext4_delete_inode(), and we allow that function to clean up the
4755          * orphan info for us.
4756          */
4757         if (inode->i_nlink)
4758                 ext4_orphan_del(handle, inode);
4759
4760         ext4_journal_stop(handle);
4761 }
4762
4763 /*
4764  * ext4_get_inode_loc returns with an extra refcount against the inode's
4765  * underlying buffer_head on success. If 'in_mem' is true, we have all
4766  * data in memory that is needed to recreate the on-disk version of this
4767  * inode.
4768  */
4769 static int __ext4_get_inode_loc(struct inode *inode,
4770                                 struct ext4_iloc *iloc, int in_mem)
4771 {
4772         struct ext4_group_desc  *gdp;
4773         struct buffer_head      *bh;
4774         struct super_block      *sb = inode->i_sb;
4775         ext4_fsblk_t            block;
4776         int                     inodes_per_block, inode_offset;
4777
4778         iloc->bh = NULL;
4779         if (!ext4_valid_inum(sb, inode->i_ino))
4780                 return -EIO;
4781
4782         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4783         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4784         if (!gdp)
4785                 return -EIO;
4786
4787         /*
4788          * Figure out the offset within the block group inode table
4789          */
4790         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4791         inode_offset = ((inode->i_ino - 1) %
4792                         EXT4_INODES_PER_GROUP(sb));
4793         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4794         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4795
4796         bh = sb_getblk(sb, block);
4797         if (!bh) {
4798                 EXT4_ERROR_INODE_BLOCK(inode, block,
4799                                        "unable to read itable block");
4800                 return -EIO;
4801         }
4802         if (!buffer_uptodate(bh)) {
4803                 lock_buffer(bh);
4804
4805                 /*
4806                  * If the buffer has the write error flag, we have failed
4807                  * to write out another inode in the same block.  In this
4808                  * case, we don't have to read the block because we may
4809                  * read the old inode data successfully.
4810                  */
4811                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4812                         set_buffer_uptodate(bh);
4813
4814                 if (buffer_uptodate(bh)) {
4815                         /* someone brought it uptodate while we waited */
4816                         unlock_buffer(bh);
4817                         goto has_buffer;
4818                 }
4819
4820                 /*
4821                  * If we have all information of the inode in memory and this
4822                  * is the only valid inode in the block, we need not read the
4823                  * block.
4824                  */
4825                 if (in_mem) {
4826                         struct buffer_head *bitmap_bh;
4827                         int i, start;
4828
4829                         start = inode_offset & ~(inodes_per_block - 1);
4830
4831                         /* Is the inode bitmap in cache? */
4832                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4833                         if (!bitmap_bh)
4834                                 goto make_io;
4835
4836                         /*
4837                          * If the inode bitmap isn't in cache then the
4838                          * optimisation may end up performing two reads instead
4839                          * of one, so skip it.
4840                          */
4841                         if (!buffer_uptodate(bitmap_bh)) {
4842                                 brelse(bitmap_bh);
4843                                 goto make_io;
4844                         }
4845                         for (i = start; i < start + inodes_per_block; i++) {
4846                                 if (i == inode_offset)
4847                                         continue;
4848                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4849                                         break;
4850                         }
4851                         brelse(bitmap_bh);
4852                         if (i == start + inodes_per_block) {
4853                                 /* all other inodes are free, so skip I/O */
4854                                 memset(bh->b_data, 0, bh->b_size);
4855                                 set_buffer_uptodate(bh);
4856                                 unlock_buffer(bh);
4857                                 goto has_buffer;
4858                         }
4859                 }
4860
4861 make_io:
4862                 /*
4863                  * If we need to do any I/O, try to pre-readahead extra
4864                  * blocks from the inode table.
4865                  */
4866                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4867                         ext4_fsblk_t b, end, table;
4868                         unsigned num;
4869
4870                         table = ext4_inode_table(sb, gdp);
4871                         /* s_inode_readahead_blks is always a power of 2 */
4872                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4873                         if (table > b)
4874                                 b = table;
4875                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4876                         num = EXT4_INODES_PER_GROUP(sb);
4877                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4878                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4879                                 num -= ext4_itable_unused_count(sb, gdp);
4880                         table += num / inodes_per_block;
4881                         if (end > table)
4882                                 end = table;
4883                         while (b <= end)
4884                                 sb_breadahead(sb, b++);
4885                 }
4886
4887                 /*
4888                  * There are other valid inodes in the buffer, this inode
4889                  * has in-inode xattrs, or we don't have this inode in memory.
4890                  * Read the block from disk.
4891                  */
4892                 get_bh(bh);
4893                 bh->b_end_io = end_buffer_read_sync;
4894                 submit_bh(READ_META, bh);
4895                 wait_on_buffer(bh);
4896                 if (!buffer_uptodate(bh)) {
4897                         EXT4_ERROR_INODE_BLOCK(inode, block,
4898                                                "unable to read itable block");
4899                         brelse(bh);
4900                         return -EIO;
4901                 }
4902         }
4903 has_buffer:
4904         iloc->bh = bh;
4905         return 0;
4906 }
4907
4908 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4909 {
4910         /* We have all inode data except xattrs in memory here. */
4911         return __ext4_get_inode_loc(inode, iloc,
4912                 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
4913 }
4914
4915 void ext4_set_inode_flags(struct inode *inode)
4916 {
4917         unsigned int flags = EXT4_I(inode)->i_flags;
4918
4919         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4920         if (flags & EXT4_SYNC_FL)
4921                 inode->i_flags |= S_SYNC;
4922         if (flags & EXT4_APPEND_FL)
4923                 inode->i_flags |= S_APPEND;
4924         if (flags & EXT4_IMMUTABLE_FL)
4925                 inode->i_flags |= S_IMMUTABLE;
4926         if (flags & EXT4_NOATIME_FL)
4927                 inode->i_flags |= S_NOATIME;
4928         if (flags & EXT4_DIRSYNC_FL)
4929                 inode->i_flags |= S_DIRSYNC;
4930 }
4931
4932 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4933 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4934 {
4935         unsigned int vfs_fl;
4936         unsigned long old_fl, new_fl;
4937
4938         do {
4939                 vfs_fl = ei->vfs_inode.i_flags;
4940                 old_fl = ei->i_flags;
4941                 new_fl = old_fl & ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4942                                 EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|
4943                                 EXT4_DIRSYNC_FL);
4944                 if (vfs_fl & S_SYNC)
4945                         new_fl |= EXT4_SYNC_FL;
4946                 if (vfs_fl & S_APPEND)
4947                         new_fl |= EXT4_APPEND_FL;
4948                 if (vfs_fl & S_IMMUTABLE)
4949                         new_fl |= EXT4_IMMUTABLE_FL;
4950                 if (vfs_fl & S_NOATIME)
4951                         new_fl |= EXT4_NOATIME_FL;
4952                 if (vfs_fl & S_DIRSYNC)
4953                         new_fl |= EXT4_DIRSYNC_FL;
4954         } while (cmpxchg(&ei->i_flags, old_fl, new_fl) != old_fl);
4955 }
4956
4957 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4958                                   struct ext4_inode_info *ei)
4959 {
4960         blkcnt_t i_blocks ;
4961         struct inode *inode = &(ei->vfs_inode);
4962         struct super_block *sb = inode->i_sb;
4963
4964         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4965                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4966                 /* we are using combined 48 bit field */
4967                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4968                                         le32_to_cpu(raw_inode->i_blocks_lo);
4969                 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
4970                         /* i_blocks represent file system block size */
4971                         return i_blocks  << (inode->i_blkbits - 9);
4972                 } else {
4973                         return i_blocks;
4974                 }
4975         } else {
4976                 return le32_to_cpu(raw_inode->i_blocks_lo);
4977         }
4978 }
4979
4980 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4981 {
4982         struct ext4_iloc iloc;
4983         struct ext4_inode *raw_inode;
4984         struct ext4_inode_info *ei;
4985         struct inode *inode;
4986         journal_t *journal = EXT4_SB(sb)->s_journal;
4987         long ret;
4988         int block;
4989
4990         inode = iget_locked(sb, ino);
4991         if (!inode)
4992                 return ERR_PTR(-ENOMEM);
4993         if (!(inode->i_state & I_NEW))
4994                 return inode;
4995
4996         ei = EXT4_I(inode);
4997         iloc.bh = 0;
4998
4999         ret = __ext4_get_inode_loc(inode, &iloc, 0);
5000         if (ret < 0)
5001                 goto bad_inode;
5002         raw_inode = ext4_raw_inode(&iloc);
5003         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
5004         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
5005         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
5006         if (!(test_opt(inode->i_sb, NO_UID32))) {
5007                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
5008                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
5009         }
5010         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
5011
5012         ei->i_state_flags = 0;
5013         ei->i_dir_start_lookup = 0;
5014         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
5015         /* We now have enough fields to check if the inode was active or not.
5016          * This is needed because nfsd might try to access dead inodes
5017          * the test is that same one that e2fsck uses
5018          * NeilBrown 1999oct15
5019          */
5020         if (inode->i_nlink == 0) {
5021                 if (inode->i_mode == 0 ||
5022                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
5023                         /* this inode is deleted */
5024                         ret = -ESTALE;
5025                         goto bad_inode;
5026                 }
5027                 /* The only unlinked inodes we let through here have
5028                  * valid i_mode and are being read by the orphan
5029                  * recovery code: that's fine, we're about to complete
5030                  * the process of deleting those. */
5031         }
5032         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
5033         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
5034         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
5035         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
5036                 ei->i_file_acl |=
5037                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
5038         inode->i_size = ext4_isize(raw_inode);
5039         ei->i_disksize = inode->i_size;
5040 #ifdef CONFIG_QUOTA
5041         ei->i_reserved_quota = 0;
5042 #endif
5043         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
5044         ei->i_block_group = iloc.block_group;
5045         ei->i_last_alloc_group = ~0;
5046         /*
5047          * NOTE! The in-memory inode i_data array is in little-endian order
5048          * even on big-endian machines: we do NOT byteswap the block numbers!
5049          */
5050         for (block = 0; block < EXT4_N_BLOCKS; block++)
5051                 ei->i_data[block] = raw_inode->i_block[block];
5052         INIT_LIST_HEAD(&ei->i_orphan);
5053
5054         /*
5055          * Set transaction id's of transactions that have to be committed
5056          * to finish f[data]sync. We set them to currently running transaction
5057          * as we cannot be sure that the inode or some of its metadata isn't
5058          * part of the transaction - the inode could have been reclaimed and
5059          * now it is reread from disk.
5060          */
5061         if (journal) {
5062                 transaction_t *transaction;
5063                 tid_t tid;
5064
5065                 spin_lock(&journal->j_state_lock);
5066                 if (journal->j_running_transaction)
5067                         transaction = journal->j_running_transaction;
5068                 else
5069                         transaction = journal->j_committing_transaction;
5070                 if (transaction)
5071                         tid = transaction->t_tid;
5072                 else
5073                         tid = journal->j_commit_sequence;
5074                 spin_unlock(&journal->j_state_lock);
5075                 ei->i_sync_tid = tid;
5076                 ei->i_datasync_tid = tid;
5077         }
5078
5079         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5080                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
5081                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
5082                     EXT4_INODE_SIZE(inode->i_sb)) {
5083                         ret = -EIO;
5084                         goto bad_inode;
5085                 }
5086                 if (ei->i_extra_isize == 0) {
5087                         /* The extra space is currently unused. Use it. */
5088                         ei->i_extra_isize = sizeof(struct ext4_inode) -
5089                                             EXT4_GOOD_OLD_INODE_SIZE;
5090                 } else {
5091                         __le32 *magic = (void *)raw_inode +
5092                                         EXT4_GOOD_OLD_INODE_SIZE +
5093                                         ei->i_extra_isize;
5094                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
5095                                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
5096                 }
5097         } else
5098                 ei->i_extra_isize = 0;
5099
5100         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
5101         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
5102         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
5103         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
5104
5105         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
5106         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5107                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5108                         inode->i_version |=
5109                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
5110         }
5111
5112         ret = 0;
5113         if (ei->i_file_acl &&
5114             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
5115                 EXT4_ERROR_INODE(inode, "bad extended attribute block %llu",
5116                                  ei->i_file_acl);
5117                 ret = -EIO;
5118                 goto bad_inode;
5119         } else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5120                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5121                     (S_ISLNK(inode->i_mode) &&
5122                      !ext4_inode_is_fast_symlink(inode)))
5123                         /* Validate extent which is part of inode */
5124                         ret = ext4_ext_check_inode(inode);
5125         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5126                    (S_ISLNK(inode->i_mode) &&
5127                     !ext4_inode_is_fast_symlink(inode))) {
5128                 /* Validate block references which are part of inode */
5129                 ret = ext4_check_inode_blockref(inode);
5130         }
5131         if (ret)
5132                 goto bad_inode;
5133
5134         if (S_ISREG(inode->i_mode)) {
5135                 inode->i_op = &ext4_file_inode_operations;
5136                 inode->i_fop = &ext4_file_operations;
5137                 ext4_set_aops(inode);
5138         } else if (S_ISDIR(inode->i_mode)) {
5139                 inode->i_op = &ext4_dir_inode_operations;
5140                 inode->i_fop = &ext4_dir_operations;
5141         } else if (S_ISLNK(inode->i_mode)) {
5142                 if (ext4_inode_is_fast_symlink(inode)) {
5143                         inode->i_op = &ext4_fast_symlink_inode_operations;
5144                         nd_terminate_link(ei->i_data, inode->i_size,
5145                                 sizeof(ei->i_data) - 1);
5146                 } else {
5147                         inode->i_op = &ext4_symlink_inode_operations;
5148                         ext4_set_aops(inode);
5149                 }
5150         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
5151               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
5152                 inode->i_op = &ext4_special_inode_operations;
5153                 if (raw_inode->i_block[0])
5154                         init_special_inode(inode, inode->i_mode,
5155                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
5156                 else
5157                         init_special_inode(inode, inode->i_mode,
5158                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
5159         } else {
5160                 ret = -EIO;
5161                 EXT4_ERROR_INODE(inode, "bogus i_mode (%o)", inode->i_mode);
5162                 goto bad_inode;
5163         }
5164         brelse(iloc.bh);
5165         ext4_set_inode_flags(inode);
5166         unlock_new_inode(inode);
5167         return inode;
5168
5169 bad_inode:
5170         brelse(iloc.bh);
5171         iget_failed(inode);
5172         return ERR_PTR(ret);
5173 }
5174
5175 static int ext4_inode_blocks_set(handle_t *handle,
5176                                 struct ext4_inode *raw_inode,
5177                                 struct ext4_inode_info *ei)
5178 {
5179         struct inode *inode = &(ei->vfs_inode);
5180         u64 i_blocks = inode->i_blocks;
5181         struct super_block *sb = inode->i_sb;
5182
5183         if (i_blocks <= ~0U) {
5184                 /*
5185                  * i_blocks can be represnted in a 32 bit variable
5186                  * as multiple of 512 bytes
5187                  */
5188                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5189                 raw_inode->i_blocks_high = 0;
5190                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5191                 return 0;
5192         }
5193         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
5194                 return -EFBIG;
5195
5196         if (i_blocks <= 0xffffffffffffULL) {
5197                 /*
5198                  * i_blocks can be represented in a 48 bit variable
5199                  * as multiple of 512 bytes
5200                  */
5201                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5202                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5203                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5204         } else {
5205                 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5206                 /* i_block is stored in file system block size */
5207                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
5208                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5209                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5210         }
5211         return 0;
5212 }
5213
5214 /*
5215  * Post the struct inode info into an on-disk inode location in the
5216  * buffer-cache.  This gobbles the caller's reference to the
5217  * buffer_head in the inode location struct.
5218  *
5219  * The caller must have write access to iloc->bh.
5220  */
5221 static int ext4_do_update_inode(handle_t *handle,
5222                                 struct inode *inode,
5223                                 struct ext4_iloc *iloc)
5224 {
5225         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5226         struct ext4_inode_info *ei = EXT4_I(inode);
5227         struct buffer_head *bh = iloc->bh;
5228         int err = 0, rc, block;
5229
5230         /* For fields not not tracking in the in-memory inode,
5231          * initialise them to zero for new inodes. */
5232         if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5233                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5234
5235         ext4_get_inode_flags(ei);
5236         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5237         if (!(test_opt(inode->i_sb, NO_UID32))) {
5238                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
5239                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
5240 /*
5241  * Fix up interoperability with old kernels. Otherwise, old inodes get
5242  * re-used with the upper 16 bits of the uid/gid intact
5243  */
5244                 if (!ei->i_dtime) {
5245                         raw_inode->i_uid_high =
5246                                 cpu_to_le16(high_16_bits(inode->i_uid));
5247                         raw_inode->i_gid_high =
5248                                 cpu_to_le16(high_16_bits(inode->i_gid));
5249                 } else {
5250                         raw_inode->i_uid_high = 0;
5251                         raw_inode->i_gid_high = 0;
5252                 }
5253         } else {
5254                 raw_inode->i_uid_low =
5255                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
5256                 raw_inode->i_gid_low =
5257                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
5258                 raw_inode->i_uid_high = 0;
5259                 raw_inode->i_gid_high = 0;
5260         }
5261         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5262
5263         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5264         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5265         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5266         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5267
5268         if (ext4_inode_blocks_set(handle, raw_inode, ei))
5269                 goto out_brelse;
5270         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5271         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
5272         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
5273             cpu_to_le32(EXT4_OS_HURD))
5274                 raw_inode->i_file_acl_high =
5275                         cpu_to_le16(ei->i_file_acl >> 32);
5276         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5277         ext4_isize_set(raw_inode, ei->i_disksize);
5278         if (ei->i_disksize > 0x7fffffffULL) {
5279                 struct super_block *sb = inode->i_sb;
5280                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
5281                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
5282                                 EXT4_SB(sb)->s_es->s_rev_level ==
5283                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
5284                         /* If this is the first large file
5285                          * created, add a flag to the superblock.
5286                          */
5287                         err = ext4_journal_get_write_access(handle,
5288                                         EXT4_SB(sb)->s_sbh);
5289                         if (err)
5290                                 goto out_brelse;
5291                         ext4_update_dynamic_rev(sb);
5292                         EXT4_SET_RO_COMPAT_FEATURE(sb,
5293                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
5294                         sb->s_dirt = 1;
5295                         ext4_handle_sync(handle);
5296                         err = ext4_handle_dirty_metadata(handle, NULL,
5297                                         EXT4_SB(sb)->s_sbh);
5298                 }
5299         }
5300         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5301         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5302                 if (old_valid_dev(inode->i_rdev)) {
5303                         raw_inode->i_block[0] =
5304                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
5305                         raw_inode->i_block[1] = 0;
5306                 } else {
5307                         raw_inode->i_block[0] = 0;
5308                         raw_inode->i_block[1] =
5309                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
5310                         raw_inode->i_block[2] = 0;
5311                 }
5312         } else
5313                 for (block = 0; block < EXT4_N_BLOCKS; block++)
5314                         raw_inode->i_block[block] = ei->i_data[block];
5315
5316         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
5317         if (ei->i_extra_isize) {
5318                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5319                         raw_inode->i_version_hi =
5320                         cpu_to_le32(inode->i_version >> 32);
5321                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
5322         }
5323
5324         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5325         rc = ext4_handle_dirty_metadata(handle, NULL, bh);
5326         if (!err)
5327                 err = rc;
5328         ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5329
5330         ext4_update_inode_fsync_trans(handle, inode, 0);
5331 out_brelse:
5332         brelse(bh);
5333         ext4_std_error(inode->i_sb, err);
5334         return err;
5335 }
5336
5337 /*
5338  * ext4_write_inode()
5339  *
5340  * We are called from a few places:
5341  *
5342  * - Within generic_file_write() for O_SYNC files.
5343  *   Here, there will be no transaction running. We wait for any running
5344  *   trasnaction to commit.
5345  *
5346  * - Within sys_sync(), kupdate and such.
5347  *   We wait on commit, if tol to.
5348  *
5349  * - Within prune_icache() (PF_MEMALLOC == true)
5350  *   Here we simply return.  We can't afford to block kswapd on the
5351  *   journal commit.
5352  *
5353  * In all cases it is actually safe for us to return without doing anything,
5354  * because the inode has been copied into a raw inode buffer in
5355  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
5356  * knfsd.
5357  *
5358  * Note that we are absolutely dependent upon all inode dirtiers doing the
5359  * right thing: they *must* call mark_inode_dirty() after dirtying info in
5360  * which we are interested.
5361  *
5362  * It would be a bug for them to not do this.  The code:
5363  *
5364  *      mark_inode_dirty(inode)
5365  *      stuff();
5366  *      inode->i_size = expr;
5367  *
5368  * is in error because a kswapd-driven write_inode() could occur while
5369  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
5370  * will no longer be on the superblock's dirty inode list.
5371  */
5372 int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5373 {
5374         int err;
5375
5376         if (current->flags & PF_MEMALLOC)
5377                 return 0;
5378
5379         if (EXT4_SB(inode->i_sb)->s_journal) {
5380                 if (ext4_journal_current_handle()) {
5381                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5382                         dump_stack();
5383                         return -EIO;
5384                 }
5385
5386                 if (wbc->sync_mode != WB_SYNC_ALL)
5387                         return 0;
5388
5389                 err = ext4_force_commit(inode->i_sb);
5390         } else {
5391                 struct ext4_iloc iloc;
5392
5393                 err = __ext4_get_inode_loc(inode, &iloc, 0);
5394                 if (err)
5395                         return err;
5396                 if (wbc->sync_mode == WB_SYNC_ALL)
5397                         sync_dirty_buffer(iloc.bh);
5398                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5399                         EXT4_ERROR_INODE_BLOCK(inode, iloc.bh->b_blocknr,
5400                                          "IO error syncing inode");
5401                         err = -EIO;
5402                 }
5403                 brelse(iloc.bh);
5404         }
5405         return err;
5406 }
5407
5408 /*
5409  * ext4_setattr()
5410  *
5411  * Called from notify_change.
5412  *
5413  * We want to trap VFS attempts to truncate the file as soon as
5414  * possible.  In particular, we want to make sure that when the VFS
5415  * shrinks i_size, we put the inode on the orphan list and modify
5416  * i_disksize immediately, so that during the subsequent flushing of
5417  * dirty pages and freeing of disk blocks, we can guarantee that any
5418  * commit will leave the blocks being flushed in an unused state on
5419  * disk.  (On recovery, the inode will get truncated and the blocks will
5420  * be freed, so we have a strong guarantee that no future commit will
5421  * leave these blocks visible to the user.)
5422  *
5423  * Another thing we have to assure is that if we are in ordered mode
5424  * and inode is still attached to the committing transaction, we must
5425  * we start writeout of all the dirty pages which are being truncated.
5426  * This way we are sure that all the data written in the previous
5427  * transaction are already on disk (truncate waits for pages under
5428  * writeback).
5429  *
5430  * Called with inode->i_mutex down.
5431  */
5432 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5433 {
5434         struct inode *inode = dentry->d_inode;
5435         int error, rc = 0;
5436         const unsigned int ia_valid = attr->ia_valid;
5437
5438         error = inode_change_ok(inode, attr);
5439         if (error)
5440                 return error;
5441
5442         if (is_quota_modification(inode, attr))
5443                 dquot_initialize(inode);
5444         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
5445                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
5446                 handle_t *handle;
5447
5448                 /* (user+group)*(old+new) structure, inode write (sb,
5449                  * inode block, ? - but truncate inode update has it) */
5450                 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
5451                                         EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
5452                 if (IS_ERR(handle)) {
5453                         error = PTR_ERR(handle);
5454                         goto err_out;
5455                 }
5456                 error = dquot_transfer(inode, attr);
5457                 if (error) {
5458                         ext4_journal_stop(handle);
5459                         return error;
5460                 }
5461                 /* Update corresponding info in inode so that everything is in
5462                  * one transaction */
5463                 if (attr->ia_valid & ATTR_UID)
5464                         inode->i_uid = attr->ia_uid;
5465                 if (attr->ia_valid & ATTR_GID)
5466                         inode->i_gid = attr->ia_gid;
5467                 error = ext4_mark_inode_dirty(handle, inode);
5468                 ext4_journal_stop(handle);
5469         }
5470
5471         if (attr->ia_valid & ATTR_SIZE) {
5472                 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
5473                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5474
5475                         if (attr->ia_size > sbi->s_bitmap_maxbytes)
5476                                 return -EFBIG;
5477                 }
5478         }
5479
5480         if (S_ISREG(inode->i_mode) &&
5481             attr->ia_valid & ATTR_SIZE &&
5482             (attr->ia_size < inode->i_size ||
5483              (ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS)))) {
5484                 handle_t *handle;
5485
5486                 handle = ext4_journal_start(inode, 3);
5487                 if (IS_ERR(handle)) {
5488                         error = PTR_ERR(handle);
5489                         goto err_out;
5490                 }
5491
5492                 error = ext4_orphan_add(handle, inode);
5493                 EXT4_I(inode)->i_disksize = attr->ia_size;
5494                 rc = ext4_mark_inode_dirty(handle, inode);
5495                 if (!error)
5496                         error = rc;
5497                 ext4_journal_stop(handle);
5498
5499                 if (ext4_should_order_data(inode)) {
5500                         error = ext4_begin_ordered_truncate(inode,
5501                                                             attr->ia_size);
5502                         if (error) {
5503                                 /* Do as much error cleanup as possible */
5504                                 handle = ext4_journal_start(inode, 3);
5505                                 if (IS_ERR(handle)) {
5506                                         ext4_orphan_del(NULL, inode);
5507                                         goto err_out;
5508                                 }
5509                                 ext4_orphan_del(handle, inode);
5510                                 ext4_journal_stop(handle);
5511                                 goto err_out;
5512                         }
5513                 }
5514                 /* ext4_truncate will clear the flag */
5515                 if ((ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS)))
5516                         ext4_truncate(inode);
5517         }
5518
5519         rc = inode_setattr(inode, attr);
5520
5521         /* If inode_setattr's call to ext4_truncate failed to get a
5522          * transaction handle at all, we need to clean up the in-core
5523          * orphan list manually. */
5524         if (inode->i_nlink)
5525                 ext4_orphan_del(NULL, inode);
5526
5527         if (!rc && (ia_valid & ATTR_MODE))
5528                 rc = ext4_acl_chmod(inode);
5529
5530 err_out:
5531         ext4_std_error(inode->i_sb, error);
5532         if (!error)
5533                 error = rc;
5534         return error;
5535 }
5536
5537 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
5538                  struct kstat *stat)
5539 {
5540         struct inode *inode;
5541