ext4: use ext4_get_block_write in buffer write
[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
43 #include "ext4_jbd2.h"
44 #include "xattr.h"
45 #include "acl.h"
46 #include "ext4_extents.h"
47
48 #include <trace/events/ext4.h>
49
50 #define MPAGE_DA_EXTENT_TAIL 0x01
51
52 static inline int ext4_begin_ordered_truncate(struct inode *inode,
53                                               loff_t new_size)
54 {
55         return jbd2_journal_begin_ordered_truncate(
56                                         EXT4_SB(inode->i_sb)->s_journal,
57                                         &EXT4_I(inode)->jinode,
58                                         new_size);
59 }
60
61 static void ext4_invalidatepage(struct page *page, unsigned long offset);
62
63 /*
64  * Test whether an inode is a fast symlink.
65  */
66 static int ext4_inode_is_fast_symlink(struct inode *inode)
67 {
68         int ea_blocks = EXT4_I(inode)->i_file_acl ?
69                 (inode->i_sb->s_blocksize >> 9) : 0;
70
71         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
72 }
73
74 /*
75  * Work out how many blocks we need to proceed with the next chunk of a
76  * truncate transaction.
77  */
78 static unsigned long blocks_for_truncate(struct inode *inode)
79 {
80         ext4_lblk_t needed;
81
82         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
83
84         /* Give ourselves just enough room to cope with inodes in which
85          * i_blocks is corrupt: we've seen disk corruptions in the past
86          * which resulted in random data in an inode which looked enough
87          * like a regular file for ext4 to try to delete it.  Things
88          * will go a bit crazy if that happens, but at least we should
89          * try not to panic the whole kernel. */
90         if (needed < 2)
91                 needed = 2;
92
93         /* But we need to bound the transaction so we don't overflow the
94          * journal. */
95         if (needed > EXT4_MAX_TRANS_DATA)
96                 needed = EXT4_MAX_TRANS_DATA;
97
98         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
99 }
100
101 /*
102  * Truncate transactions can be complex and absolutely huge.  So we need to
103  * be able to restart the transaction at a conventient checkpoint to make
104  * sure we don't overflow the journal.
105  *
106  * start_transaction gets us a new handle for a truncate transaction,
107  * and extend_transaction tries to extend the existing one a bit.  If
108  * extend fails, we need to propagate the failure up and restart the
109  * transaction in the top-level truncate loop. --sct
110  */
111 static handle_t *start_transaction(struct inode *inode)
112 {
113         handle_t *result;
114
115         result = ext4_journal_start(inode, blocks_for_truncate(inode));
116         if (!IS_ERR(result))
117                 return result;
118
119         ext4_std_error(inode->i_sb, PTR_ERR(result));
120         return result;
121 }
122
123 /*
124  * Try to extend this transaction for the purposes of truncation.
125  *
126  * Returns 0 if we managed to create more room.  If we can't create more
127  * room, and the transaction must be restarted we return 1.
128  */
129 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
130 {
131         if (!ext4_handle_valid(handle))
132                 return 0;
133         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
134                 return 0;
135         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
136                 return 0;
137         return 1;
138 }
139
140 /*
141  * Restart the transaction associated with *handle.  This does a commit,
142  * so before we call here everything must be consistently dirtied against
143  * this transaction.
144  */
145 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
146                                  int nblocks)
147 {
148         int ret;
149
150         /*
151          * Drop i_data_sem to avoid deadlock with ext4_get_blocks At this
152          * moment, get_block can be called only for blocks inside i_size since
153          * page cache has been already dropped and writes are blocked by
154          * i_mutex. So we can safely drop the i_data_sem here.
155          */
156         BUG_ON(EXT4_JOURNAL(inode) == NULL);
157         jbd_debug(2, "restarting handle %p\n", handle);
158         up_write(&EXT4_I(inode)->i_data_sem);
159         ret = ext4_journal_restart(handle, blocks_for_truncate(inode));
160         down_write(&EXT4_I(inode)->i_data_sem);
161         ext4_discard_preallocations(inode);
162
163         return ret;
164 }
165
166 /*
167  * Called at the last iput() if i_nlink is zero.
168  */
169 void ext4_delete_inode(struct inode *inode)
170 {
171         handle_t *handle;
172         int err;
173
174         if (ext4_should_order_data(inode))
175                 ext4_begin_ordered_truncate(inode, 0);
176         truncate_inode_pages(&inode->i_data, 0);
177
178         if (is_bad_inode(inode))
179                 goto no_delete;
180
181         handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
182         if (IS_ERR(handle)) {
183                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
184                 /*
185                  * If we're going to skip the normal cleanup, we still need to
186                  * make sure that the in-core orphan linked list is properly
187                  * cleaned up.
188                  */
189                 ext4_orphan_del(NULL, inode);
190                 goto no_delete;
191         }
192
193         if (IS_SYNC(inode))
194                 ext4_handle_sync(handle);
195         inode->i_size = 0;
196         err = ext4_mark_inode_dirty(handle, inode);
197         if (err) {
198                 ext4_warning(inode->i_sb,
199                              "couldn't mark inode dirty (err %d)", err);
200                 goto stop_handle;
201         }
202         if (inode->i_blocks)
203                 ext4_truncate(inode);
204
205         /*
206          * ext4_ext_truncate() doesn't reserve any slop when it
207          * restarts journal transactions; therefore there may not be
208          * enough credits left in the handle to remove the inode from
209          * the orphan list and set the dtime field.
210          */
211         if (!ext4_handle_has_enough_credits(handle, 3)) {
212                 err = ext4_journal_extend(handle, 3);
213                 if (err > 0)
214                         err = ext4_journal_restart(handle, 3);
215                 if (err != 0) {
216                         ext4_warning(inode->i_sb,
217                                      "couldn't extend journal (err %d)", err);
218                 stop_handle:
219                         ext4_journal_stop(handle);
220                         goto no_delete;
221                 }
222         }
223
224         /*
225          * Kill off the orphan record which ext4_truncate created.
226          * AKPM: I think this can be inside the above `if'.
227          * Note that ext4_orphan_del() has to be able to cope with the
228          * deletion of a non-existent orphan - this is because we don't
229          * know if ext4_truncate() actually created an orphan record.
230          * (Well, we could do this if we need to, but heck - it works)
231          */
232         ext4_orphan_del(handle, inode);
233         EXT4_I(inode)->i_dtime  = get_seconds();
234
235         /*
236          * One subtle ordering requirement: if anything has gone wrong
237          * (transaction abort, IO errors, whatever), then we can still
238          * do these next steps (the fs will already have been marked as
239          * having errors), but we can't free the inode if the mark_dirty
240          * fails.
241          */
242         if (ext4_mark_inode_dirty(handle, inode))
243                 /* If that failed, just do the required in-core inode clear. */
244                 clear_inode(inode);
245         else
246                 ext4_free_inode(handle, inode);
247         ext4_journal_stop(handle);
248         return;
249 no_delete:
250         clear_inode(inode);     /* We must guarantee clearing of inode... */
251 }
252
253 typedef struct {
254         __le32  *p;
255         __le32  key;
256         struct buffer_head *bh;
257 } Indirect;
258
259 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
260 {
261         p->key = *(p->p = v);
262         p->bh = bh;
263 }
264
265 /**
266  *      ext4_block_to_path - parse the block number into array of offsets
267  *      @inode: inode in question (we are only interested in its superblock)
268  *      @i_block: block number to be parsed
269  *      @offsets: array to store the offsets in
270  *      @boundary: set this non-zero if the referred-to block is likely to be
271  *             followed (on disk) by an indirect block.
272  *
273  *      To store the locations of file's data ext4 uses a data structure common
274  *      for UNIX filesystems - tree of pointers anchored in the inode, with
275  *      data blocks at leaves and indirect blocks in intermediate nodes.
276  *      This function translates the block number into path in that tree -
277  *      return value is the path length and @offsets[n] is the offset of
278  *      pointer to (n+1)th node in the nth one. If @block is out of range
279  *      (negative or too large) warning is printed and zero returned.
280  *
281  *      Note: function doesn't find node addresses, so no IO is needed. All
282  *      we need to know is the capacity of indirect blocks (taken from the
283  *      inode->i_sb).
284  */
285
286 /*
287  * Portability note: the last comparison (check that we fit into triple
288  * indirect block) is spelled differently, because otherwise on an
289  * architecture with 32-bit longs and 8Kb pages we might get into trouble
290  * if our filesystem had 8Kb blocks. We might use long long, but that would
291  * kill us on x86. Oh, well, at least the sign propagation does not matter -
292  * i_block would have to be negative in the very beginning, so we would not
293  * get there at all.
294  */
295
296 static int ext4_block_to_path(struct inode *inode,
297                               ext4_lblk_t i_block,
298                               ext4_lblk_t offsets[4], int *boundary)
299 {
300         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
301         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
302         const long direct_blocks = EXT4_NDIR_BLOCKS,
303                 indirect_blocks = ptrs,
304                 double_blocks = (1 << (ptrs_bits * 2));
305         int n = 0;
306         int final = 0;
307
308         if (i_block < direct_blocks) {
309                 offsets[n++] = i_block;
310                 final = direct_blocks;
311         } else if ((i_block -= direct_blocks) < indirect_blocks) {
312                 offsets[n++] = EXT4_IND_BLOCK;
313                 offsets[n++] = i_block;
314                 final = ptrs;
315         } else if ((i_block -= indirect_blocks) < double_blocks) {
316                 offsets[n++] = EXT4_DIND_BLOCK;
317                 offsets[n++] = i_block >> ptrs_bits;
318                 offsets[n++] = i_block & (ptrs - 1);
319                 final = ptrs;
320         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
321                 offsets[n++] = EXT4_TIND_BLOCK;
322                 offsets[n++] = i_block >> (ptrs_bits * 2);
323                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
324                 offsets[n++] = i_block & (ptrs - 1);
325                 final = ptrs;
326         } else {
327                 ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
328                              i_block + direct_blocks +
329                              indirect_blocks + double_blocks, inode->i_ino);
330         }
331         if (boundary)
332                 *boundary = final - 1 - (i_block & (ptrs - 1));
333         return n;
334 }
335
336 static int __ext4_check_blockref(const char *function, struct inode *inode,
337                                  __le32 *p, unsigned int max)
338 {
339         __le32 *bref = p;
340         unsigned int blk;
341
342         while (bref < p+max) {
343                 blk = le32_to_cpu(*bref++);
344                 if (blk &&
345                     unlikely(!ext4_data_block_valid(EXT4_SB(inode->i_sb),
346                                                     blk, 1))) {
347                         __ext4_error(inode->i_sb, function,
348                                    "invalid block reference %u "
349                                    "in inode #%lu", blk, inode->i_ino);
350                         return -EIO;
351                 }
352         }
353         return 0;
354 }
355
356
357 #define ext4_check_indirect_blockref(inode, bh)                         \
358         __ext4_check_blockref(__func__, inode, (__le32 *)(bh)->b_data,  \
359                               EXT4_ADDR_PER_BLOCK((inode)->i_sb))
360
361 #define ext4_check_inode_blockref(inode)                                \
362         __ext4_check_blockref(__func__, inode, EXT4_I(inode)->i_data,   \
363                               EXT4_NDIR_BLOCKS)
364
365 /**
366  *      ext4_get_branch - read the chain of indirect blocks leading to data
367  *      @inode: inode in question
368  *      @depth: depth of the chain (1 - direct pointer, etc.)
369  *      @offsets: offsets of pointers in inode/indirect blocks
370  *      @chain: place to store the result
371  *      @err: here we store the error value
372  *
373  *      Function fills the array of triples <key, p, bh> and returns %NULL
374  *      if everything went OK or the pointer to the last filled triple
375  *      (incomplete one) otherwise. Upon the return chain[i].key contains
376  *      the number of (i+1)-th block in the chain (as it is stored in memory,
377  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
378  *      number (it points into struct inode for i==0 and into the bh->b_data
379  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
380  *      block for i>0 and NULL for i==0. In other words, it holds the block
381  *      numbers of the chain, addresses they were taken from (and where we can
382  *      verify that chain did not change) and buffer_heads hosting these
383  *      numbers.
384  *
385  *      Function stops when it stumbles upon zero pointer (absent block)
386  *              (pointer to last triple returned, *@err == 0)
387  *      or when it gets an IO error reading an indirect block
388  *              (ditto, *@err == -EIO)
389  *      or when it reads all @depth-1 indirect blocks successfully and finds
390  *      the whole chain, all way to the data (returns %NULL, *err == 0).
391  *
392  *      Need to be called with
393  *      down_read(&EXT4_I(inode)->i_data_sem)
394  */
395 static Indirect *ext4_get_branch(struct inode *inode, int depth,
396                                  ext4_lblk_t  *offsets,
397                                  Indirect chain[4], int *err)
398 {
399         struct super_block *sb = inode->i_sb;
400         Indirect *p = chain;
401         struct buffer_head *bh;
402
403         *err = 0;
404         /* i_data is not going away, no lock needed */
405         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
406         if (!p->key)
407                 goto no_block;
408         while (--depth) {
409                 bh = sb_getblk(sb, le32_to_cpu(p->key));
410                 if (unlikely(!bh))
411                         goto failure;
412
413                 if (!bh_uptodate_or_lock(bh)) {
414                         if (bh_submit_read(bh) < 0) {
415                                 put_bh(bh);
416                                 goto failure;
417                         }
418                         /* validate block references */
419                         if (ext4_check_indirect_blockref(inode, bh)) {
420                                 put_bh(bh);
421                                 goto failure;
422                         }
423                 }
424
425                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
426                 /* Reader: end */
427                 if (!p->key)
428                         goto no_block;
429         }
430         return NULL;
431
432 failure:
433         *err = -EIO;
434 no_block:
435         return p;
436 }
437
438 /**
439  *      ext4_find_near - find a place for allocation with sufficient locality
440  *      @inode: owner
441  *      @ind: descriptor of indirect block.
442  *
443  *      This function returns the preferred place for block allocation.
444  *      It is used when heuristic for sequential allocation fails.
445  *      Rules are:
446  *        + if there is a block to the left of our position - allocate near it.
447  *        + if pointer will live in indirect block - allocate near that block.
448  *        + if pointer will live in inode - allocate in the same
449  *          cylinder group.
450  *
451  * In the latter case we colour the starting block by the callers PID to
452  * prevent it from clashing with concurrent allocations for a different inode
453  * in the same block group.   The PID is used here so that functionally related
454  * files will be close-by on-disk.
455  *
456  *      Caller must make sure that @ind is valid and will stay that way.
457  */
458 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
459 {
460         struct ext4_inode_info *ei = EXT4_I(inode);
461         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
462         __le32 *p;
463         ext4_fsblk_t bg_start;
464         ext4_fsblk_t last_block;
465         ext4_grpblk_t colour;
466         ext4_group_t block_group;
467         int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
468
469         /* Try to find previous block */
470         for (p = ind->p - 1; p >= start; p--) {
471                 if (*p)
472                         return le32_to_cpu(*p);
473         }
474
475         /* No such thing, so let's try location of indirect block */
476         if (ind->bh)
477                 return ind->bh->b_blocknr;
478
479         /*
480          * It is going to be referred to from the inode itself? OK, just put it
481          * into the same cylinder group then.
482          */
483         block_group = ei->i_block_group;
484         if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
485                 block_group &= ~(flex_size-1);
486                 if (S_ISREG(inode->i_mode))
487                         block_group++;
488         }
489         bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
490         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
491
492         /*
493          * If we are doing delayed allocation, we don't need take
494          * colour into account.
495          */
496         if (test_opt(inode->i_sb, DELALLOC))
497                 return bg_start;
498
499         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
500                 colour = (current->pid % 16) *
501                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
502         else
503                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
504         return bg_start + colour;
505 }
506
507 /**
508  *      ext4_find_goal - find a preferred place for allocation.
509  *      @inode: owner
510  *      @block:  block we want
511  *      @partial: pointer to the last triple within a chain
512  *
513  *      Normally this function find the preferred place for block allocation,
514  *      returns it.
515  *      Because this is only used for non-extent files, we limit the block nr
516  *      to 32 bits.
517  */
518 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
519                                    Indirect *partial)
520 {
521         ext4_fsblk_t goal;
522
523         /*
524          * XXX need to get goal block from mballoc's data structures
525          */
526
527         goal = ext4_find_near(inode, partial);
528         goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
529         return goal;
530 }
531
532 /**
533  *      ext4_blks_to_allocate: Look up the block map and count the number
534  *      of direct blocks need to be allocated for the given branch.
535  *
536  *      @branch: chain of indirect blocks
537  *      @k: number of blocks need for indirect blocks
538  *      @blks: number of data blocks to be mapped.
539  *      @blocks_to_boundary:  the offset in the indirect block
540  *
541  *      return the total number of blocks to be allocate, including the
542  *      direct and indirect blocks.
543  */
544 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
545                                  int blocks_to_boundary)
546 {
547         unsigned int count = 0;
548
549         /*
550          * Simple case, [t,d]Indirect block(s) has not allocated yet
551          * then it's clear blocks on that path have not allocated
552          */
553         if (k > 0) {
554                 /* right now we don't handle cross boundary allocation */
555                 if (blks < blocks_to_boundary + 1)
556                         count += blks;
557                 else
558                         count += blocks_to_boundary + 1;
559                 return count;
560         }
561
562         count++;
563         while (count < blks && count <= blocks_to_boundary &&
564                 le32_to_cpu(*(branch[0].p + count)) == 0) {
565                 count++;
566         }
567         return count;
568 }
569
570 /**
571  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
572  *      @indirect_blks: the number of blocks need to allocate for indirect
573  *                      blocks
574  *
575  *      @new_blocks: on return it will store the new block numbers for
576  *      the indirect blocks(if needed) and the first direct block,
577  *      @blks:  on return it will store the total number of allocated
578  *              direct blocks
579  */
580 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
581                              ext4_lblk_t iblock, ext4_fsblk_t goal,
582                              int indirect_blks, int blks,
583                              ext4_fsblk_t new_blocks[4], int *err)
584 {
585         struct ext4_allocation_request ar;
586         int target, i;
587         unsigned long count = 0, blk_allocated = 0;
588         int index = 0;
589         ext4_fsblk_t current_block = 0;
590         int ret = 0;
591
592         /*
593          * Here we try to allocate the requested multiple blocks at once,
594          * on a best-effort basis.
595          * To build a branch, we should allocate blocks for
596          * the indirect blocks(if not allocated yet), and at least
597          * the first direct block of this branch.  That's the
598          * minimum number of blocks need to allocate(required)
599          */
600         /* first we try to allocate the indirect blocks */
601         target = indirect_blks;
602         while (target > 0) {
603                 count = target;
604                 /* allocating blocks for indirect blocks and direct blocks */
605                 current_block = ext4_new_meta_blocks(handle, inode,
606                                                         goal, &count, err);
607                 if (*err)
608                         goto failed_out;
609
610                 BUG_ON(current_block + count > EXT4_MAX_BLOCK_FILE_PHYS);
611
612                 target -= count;
613                 /* allocate blocks for indirect blocks */
614                 while (index < indirect_blks && count) {
615                         new_blocks[index++] = current_block++;
616                         count--;
617                 }
618                 if (count > 0) {
619                         /*
620                          * save the new block number
621                          * for the first direct block
622                          */
623                         new_blocks[index] = current_block;
624                         printk(KERN_INFO "%s returned more blocks than "
625                                                 "requested\n", __func__);
626                         WARN_ON(1);
627                         break;
628                 }
629         }
630
631         target = blks - count ;
632         blk_allocated = count;
633         if (!target)
634                 goto allocated;
635         /* Now allocate data blocks */
636         memset(&ar, 0, sizeof(ar));
637         ar.inode = inode;
638         ar.goal = goal;
639         ar.len = target;
640         ar.logical = iblock;
641         if (S_ISREG(inode->i_mode))
642                 /* enable in-core preallocation only for regular files */
643                 ar.flags = EXT4_MB_HINT_DATA;
644
645         current_block = ext4_mb_new_blocks(handle, &ar, err);
646         BUG_ON(current_block + ar.len > EXT4_MAX_BLOCK_FILE_PHYS);
647
648         if (*err && (target == blks)) {
649                 /*
650                  * if the allocation failed and we didn't allocate
651                  * any blocks before
652                  */
653                 goto failed_out;
654         }
655         if (!*err) {
656                 if (target == blks) {
657                         /*
658                          * save the new block number
659                          * for the first direct block
660                          */
661                         new_blocks[index] = current_block;
662                 }
663                 blk_allocated += ar.len;
664         }
665 allocated:
666         /* total number of blocks allocated for direct blocks */
667         ret = blk_allocated;
668         *err = 0;
669         return ret;
670 failed_out:
671         for (i = 0; i < index; i++)
672                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
673         return ret;
674 }
675
676 /**
677  *      ext4_alloc_branch - allocate and set up a chain of blocks.
678  *      @inode: owner
679  *      @indirect_blks: number of allocated indirect blocks
680  *      @blks: number of allocated direct blocks
681  *      @offsets: offsets (in the blocks) to store the pointers to next.
682  *      @branch: place to store the chain in.
683  *
684  *      This function allocates blocks, zeroes out all but the last one,
685  *      links them into chain and (if we are synchronous) writes them to disk.
686  *      In other words, it prepares a branch that can be spliced onto the
687  *      inode. It stores the information about that chain in the branch[], in
688  *      the same format as ext4_get_branch() would do. We are calling it after
689  *      we had read the existing part of chain and partial points to the last
690  *      triple of that (one with zero ->key). Upon the exit we have the same
691  *      picture as after the successful ext4_get_block(), except that in one
692  *      place chain is disconnected - *branch->p is still zero (we did not
693  *      set the last link), but branch->key contains the number that should
694  *      be placed into *branch->p to fill that gap.
695  *
696  *      If allocation fails we free all blocks we've allocated (and forget
697  *      their buffer_heads) and return the error value the from failed
698  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
699  *      as described above and return 0.
700  */
701 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
702                              ext4_lblk_t iblock, int indirect_blks,
703                              int *blks, ext4_fsblk_t goal,
704                              ext4_lblk_t *offsets, Indirect *branch)
705 {
706         int blocksize = inode->i_sb->s_blocksize;
707         int i, n = 0;
708         int err = 0;
709         struct buffer_head *bh;
710         int num;
711         ext4_fsblk_t new_blocks[4];
712         ext4_fsblk_t current_block;
713
714         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
715                                 *blks, new_blocks, &err);
716         if (err)
717                 return err;
718
719         branch[0].key = cpu_to_le32(new_blocks[0]);
720         /*
721          * metadata blocks and data blocks are allocated.
722          */
723         for (n = 1; n <= indirect_blks;  n++) {
724                 /*
725                  * Get buffer_head for parent block, zero it out
726                  * and set the pointer to new one, then send
727                  * parent to disk.
728                  */
729                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
730                 branch[n].bh = bh;
731                 lock_buffer(bh);
732                 BUFFER_TRACE(bh, "call get_create_access");
733                 err = ext4_journal_get_create_access(handle, bh);
734                 if (err) {
735                         /* Don't brelse(bh) here; it's done in
736                          * ext4_journal_forget() below */
737                         unlock_buffer(bh);
738                         goto failed;
739                 }
740
741                 memset(bh->b_data, 0, blocksize);
742                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
743                 branch[n].key = cpu_to_le32(new_blocks[n]);
744                 *branch[n].p = branch[n].key;
745                 if (n == indirect_blks) {
746                         current_block = new_blocks[n];
747                         /*
748                          * End of chain, update the last new metablock of
749                          * the chain to point to the new allocated
750                          * data blocks numbers
751                          */
752                         for (i = 1; i < num; i++)
753                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
754                 }
755                 BUFFER_TRACE(bh, "marking uptodate");
756                 set_buffer_uptodate(bh);
757                 unlock_buffer(bh);
758
759                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
760                 err = ext4_handle_dirty_metadata(handle, inode, bh);
761                 if (err)
762                         goto failed;
763         }
764         *blks = num;
765         return err;
766 failed:
767         /* Allocation failed, free what we already allocated */
768         ext4_free_blocks(handle, inode, 0, new_blocks[0], 1, 0);
769         for (i = 1; i <= n ; i++) {
770                 /* 
771                  * branch[i].bh is newly allocated, so there is no
772                  * need to revoke the block, which is why we don't
773                  * need to set EXT4_FREE_BLOCKS_METADATA.
774                  */
775                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1,
776                                  EXT4_FREE_BLOCKS_FORGET);
777         }
778         for (i = n+1; i < indirect_blks; i++)
779                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
780
781         ext4_free_blocks(handle, inode, 0, new_blocks[i], num, 0);
782
783         return err;
784 }
785
786 /**
787  * ext4_splice_branch - splice the allocated branch onto inode.
788  * @inode: owner
789  * @block: (logical) number of block we are adding
790  * @chain: chain of indirect blocks (with a missing link - see
791  *      ext4_alloc_branch)
792  * @where: location of missing link
793  * @num:   number of indirect blocks we are adding
794  * @blks:  number of direct blocks we are adding
795  *
796  * This function fills the missing link and does all housekeeping needed in
797  * inode (->i_blocks, etc.). In case of success we end up with the full
798  * chain to new block and return 0.
799  */
800 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
801                               ext4_lblk_t block, Indirect *where, int num,
802                               int blks)
803 {
804         int i;
805         int err = 0;
806         ext4_fsblk_t current_block;
807
808         /*
809          * If we're splicing into a [td]indirect block (as opposed to the
810          * inode) then we need to get write access to the [td]indirect block
811          * before the splice.
812          */
813         if (where->bh) {
814                 BUFFER_TRACE(where->bh, "get_write_access");
815                 err = ext4_journal_get_write_access(handle, where->bh);
816                 if (err)
817                         goto err_out;
818         }
819         /* That's it */
820
821         *where->p = where->key;
822
823         /*
824          * Update the host buffer_head or inode to point to more just allocated
825          * direct blocks blocks
826          */
827         if (num == 0 && blks > 1) {
828                 current_block = le32_to_cpu(where->key) + 1;
829                 for (i = 1; i < blks; i++)
830                         *(where->p + i) = cpu_to_le32(current_block++);
831         }
832
833         /* We are done with atomic stuff, now do the rest of housekeeping */
834         /* had we spliced it onto indirect block? */
835         if (where->bh) {
836                 /*
837                  * If we spliced it onto an indirect block, we haven't
838                  * altered the inode.  Note however that if it is being spliced
839                  * onto an indirect block at the very end of the file (the
840                  * file is growing) then we *will* alter the inode to reflect
841                  * the new i_size.  But that is not done here - it is done in
842                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
843                  */
844                 jbd_debug(5, "splicing indirect only\n");
845                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
846                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
847                 if (err)
848                         goto err_out;
849         } else {
850                 /*
851                  * OK, we spliced it into the inode itself on a direct block.
852                  */
853                 ext4_mark_inode_dirty(handle, inode);
854                 jbd_debug(5, "splicing direct\n");
855         }
856         return err;
857
858 err_out:
859         for (i = 1; i <= num; i++) {
860                 /* 
861                  * branch[i].bh is newly allocated, so there is no
862                  * need to revoke the block, which is why we don't
863                  * need to set EXT4_FREE_BLOCKS_METADATA.
864                  */
865                 ext4_free_blocks(handle, inode, where[i].bh, 0, 1,
866                                  EXT4_FREE_BLOCKS_FORGET);
867         }
868         ext4_free_blocks(handle, inode, 0, le32_to_cpu(where[num].key),
869                          blks, 0);
870
871         return err;
872 }
873
874 /*
875  * The ext4_ind_get_blocks() function handles non-extents inodes
876  * (i.e., using the traditional indirect/double-indirect i_blocks
877  * scheme) for ext4_get_blocks().
878  *
879  * Allocation strategy is simple: if we have to allocate something, we will
880  * have to go the whole way to leaf. So let's do it before attaching anything
881  * to tree, set linkage between the newborn blocks, write them if sync is
882  * required, recheck the path, free and repeat if check fails, otherwise
883  * set the last missing link (that will protect us from any truncate-generated
884  * removals - all blocks on the path are immune now) and possibly force the
885  * write on the parent block.
886  * That has a nice additional property: no special recovery from the failed
887  * allocations is needed - we simply release blocks and do not touch anything
888  * reachable from inode.
889  *
890  * `handle' can be NULL if create == 0.
891  *
892  * return > 0, # of blocks mapped or allocated.
893  * return = 0, if plain lookup failed.
894  * return < 0, error case.
895  *
896  * The ext4_ind_get_blocks() function should be called with
897  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
898  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
899  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
900  * blocks.
901  */
902 static int ext4_ind_get_blocks(handle_t *handle, struct inode *inode,
903                                ext4_lblk_t iblock, unsigned int maxblocks,
904                                struct buffer_head *bh_result,
905                                int flags)
906 {
907         int err = -EIO;
908         ext4_lblk_t offsets[4];
909         Indirect chain[4];
910         Indirect *partial;
911         ext4_fsblk_t goal;
912         int indirect_blks;
913         int blocks_to_boundary = 0;
914         int depth;
915         int count = 0;
916         ext4_fsblk_t first_block = 0;
917
918         J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
919         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
920         depth = ext4_block_to_path(inode, iblock, offsets,
921                                    &blocks_to_boundary);
922
923         if (depth == 0)
924                 goto out;
925
926         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
927
928         /* Simplest case - block found, no allocation needed */
929         if (!partial) {
930                 first_block = le32_to_cpu(chain[depth - 1].key);
931                 clear_buffer_new(bh_result);
932                 count++;
933                 /*map more blocks*/
934                 while (count < maxblocks && count <= blocks_to_boundary) {
935                         ext4_fsblk_t blk;
936
937                         blk = le32_to_cpu(*(chain[depth-1].p + count));
938
939                         if (blk == first_block + count)
940                                 count++;
941                         else
942                                 break;
943                 }
944                 goto got_it;
945         }
946
947         /* Next simple case - plain lookup or failed read of indirect block */
948         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
949                 goto cleanup;
950
951         /*
952          * Okay, we need to do block allocation.
953         */
954         goal = ext4_find_goal(inode, iblock, partial);
955
956         /* the number of blocks need to allocate for [d,t]indirect blocks */
957         indirect_blks = (chain + depth) - partial - 1;
958
959         /*
960          * Next look up the indirect map to count the totoal number of
961          * direct blocks to allocate for this branch.
962          */
963         count = ext4_blks_to_allocate(partial, indirect_blks,
964                                         maxblocks, blocks_to_boundary);
965         /*
966          * Block out ext4_truncate while we alter the tree
967          */
968         err = ext4_alloc_branch(handle, inode, iblock, indirect_blks,
969                                 &count, goal,
970                                 offsets + (partial - chain), partial);
971
972         /*
973          * The ext4_splice_branch call will free and forget any buffers
974          * on the new chain if there is a failure, but that risks using
975          * up transaction credits, especially for bitmaps where the
976          * credits cannot be returned.  Can we handle this somehow?  We
977          * may need to return -EAGAIN upwards in the worst case.  --sct
978          */
979         if (!err)
980                 err = ext4_splice_branch(handle, inode, iblock,
981                                          partial, indirect_blks, count);
982         if (err)
983                 goto cleanup;
984
985         set_buffer_new(bh_result);
986
987         ext4_update_inode_fsync_trans(handle, inode, 1);
988 got_it:
989         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
990         if (count > blocks_to_boundary)
991                 set_buffer_boundary(bh_result);
992         err = count;
993         /* Clean up and exit */
994         partial = chain + depth - 1;    /* the whole chain */
995 cleanup:
996         while (partial > chain) {
997                 BUFFER_TRACE(partial->bh, "call brelse");
998                 brelse(partial->bh);
999                 partial--;
1000         }
1001         BUFFER_TRACE(bh_result, "returned");
1002 out:
1003         return err;
1004 }
1005
1006 #ifdef CONFIG_QUOTA
1007 qsize_t *ext4_get_reserved_space(struct inode *inode)
1008 {
1009         return &EXT4_I(inode)->i_reserved_quota;
1010 }
1011 #endif
1012
1013 /*
1014  * Calculate the number of metadata blocks need to reserve
1015  * to allocate a new block at @lblocks for non extent file based file
1016  */
1017 static int ext4_indirect_calc_metadata_amount(struct inode *inode,
1018                                               sector_t lblock)
1019 {
1020         struct ext4_inode_info *ei = EXT4_I(inode);
1021         int dind_mask = EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1;
1022         int blk_bits;
1023
1024         if (lblock < EXT4_NDIR_BLOCKS)
1025                 return 0;
1026
1027         lblock -= EXT4_NDIR_BLOCKS;
1028
1029         if (ei->i_da_metadata_calc_len &&
1030             (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
1031                 ei->i_da_metadata_calc_len++;
1032                 return 0;
1033         }
1034         ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
1035         ei->i_da_metadata_calc_len = 1;
1036         blk_bits = roundup_pow_of_two(lblock + 1);
1037         return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
1038 }
1039
1040 /*
1041  * Calculate the number of metadata blocks need to reserve
1042  * to allocate a block located at @lblock
1043  */
1044 static int ext4_calc_metadata_amount(struct inode *inode, sector_t lblock)
1045 {
1046         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
1047                 return ext4_ext_calc_metadata_amount(inode, lblock);
1048
1049         return ext4_indirect_calc_metadata_amount(inode, lblock);
1050 }
1051
1052 /*
1053  * Called with i_data_sem down, which is important since we can call
1054  * ext4_discard_preallocations() from here.
1055  */
1056 void ext4_da_update_reserve_space(struct inode *inode,
1057                                         int used, int quota_claim)
1058 {
1059         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1060         struct ext4_inode_info *ei = EXT4_I(inode);
1061         int mdb_free = 0, allocated_meta_blocks = 0;
1062
1063         spin_lock(&ei->i_block_reservation_lock);
1064         trace_ext4_da_update_reserve_space(inode, used);
1065         if (unlikely(used > ei->i_reserved_data_blocks)) {
1066                 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
1067                          "with only %d reserved data blocks\n",
1068                          __func__, inode->i_ino, used,
1069                          ei->i_reserved_data_blocks);
1070                 WARN_ON(1);
1071                 used = ei->i_reserved_data_blocks;
1072         }
1073
1074         /* Update per-inode reservations */
1075         ei->i_reserved_data_blocks -= used;
1076         used += ei->i_allocated_meta_blocks;
1077         ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
1078         allocated_meta_blocks = ei->i_allocated_meta_blocks;
1079         ei->i_allocated_meta_blocks = 0;
1080         percpu_counter_sub(&sbi->s_dirtyblocks_counter, used);
1081
1082         if (ei->i_reserved_data_blocks == 0) {
1083                 /*
1084                  * We can release all of the reserved metadata blocks
1085                  * only when we have written all of the delayed
1086                  * allocation blocks.
1087                  */
1088                 mdb_free = ei->i_reserved_meta_blocks;
1089                 ei->i_reserved_meta_blocks = 0;
1090                 ei->i_da_metadata_calc_len = 0;
1091                 percpu_counter_sub(&sbi->s_dirtyblocks_counter, mdb_free);
1092         }
1093         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1094
1095         /* Update quota subsystem */
1096         if (quota_claim) {
1097                 vfs_dq_claim_block(inode, used);
1098                 if (mdb_free)
1099                         vfs_dq_release_reservation_block(inode, mdb_free);
1100         } else {
1101                 /*
1102                  * We did fallocate with an offset that is already delayed
1103                  * allocated. So on delayed allocated writeback we should
1104                  * not update the quota for allocated blocks. But then
1105                  * converting an fallocate region to initialized region would
1106                  * have caused a metadata allocation. So claim quota for
1107                  * that
1108                  */
1109                 if (allocated_meta_blocks)
1110                         vfs_dq_claim_block(inode, allocated_meta_blocks);
1111                 vfs_dq_release_reservation_block(inode, mdb_free + used);
1112         }
1113
1114         /*
1115          * If we have done all the pending block allocations and if
1116          * there aren't any writers on the inode, we can discard the
1117          * inode's preallocations.
1118          */
1119         if ((ei->i_reserved_data_blocks == 0) &&
1120             (atomic_read(&inode->i_writecount) == 0))
1121                 ext4_discard_preallocations(inode);
1122 }
1123
1124 static int check_block_validity(struct inode *inode, const char *msg,
1125                                 sector_t logical, sector_t phys, int len)
1126 {
1127         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), phys, len)) {
1128                 __ext4_error(inode->i_sb, msg,
1129                            "inode #%lu logical block %llu mapped to %llu "
1130                            "(size %d)", inode->i_ino,
1131                            (unsigned long long) logical,
1132                            (unsigned long long) phys, len);
1133                 return -EIO;
1134         }
1135         return 0;
1136 }
1137
1138 /*
1139  * Return the number of contiguous dirty pages in a given inode
1140  * starting at page frame idx.
1141  */
1142 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
1143                                     unsigned int max_pages)
1144 {
1145         struct address_space *mapping = inode->i_mapping;
1146         pgoff_t index;
1147         struct pagevec pvec;
1148         pgoff_t num = 0;
1149         int i, nr_pages, done = 0;
1150
1151         if (max_pages == 0)
1152                 return 0;
1153         pagevec_init(&pvec, 0);
1154         while (!done) {
1155                 index = idx;
1156                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
1157                                               PAGECACHE_TAG_DIRTY,
1158                                               (pgoff_t)PAGEVEC_SIZE);
1159                 if (nr_pages == 0)
1160                         break;
1161                 for (i = 0; i < nr_pages; i++) {
1162                         struct page *page = pvec.pages[i];
1163                         struct buffer_head *bh, *head;
1164
1165                         lock_page(page);
1166                         if (unlikely(page->mapping != mapping) ||
1167                             !PageDirty(page) ||
1168                             PageWriteback(page) ||
1169                             page->index != idx) {
1170                                 done = 1;
1171                                 unlock_page(page);
1172                                 break;
1173                         }
1174                         if (page_has_buffers(page)) {
1175                                 bh = head = page_buffers(page);
1176                                 do {
1177                                         if (!buffer_delay(bh) &&
1178                                             !buffer_unwritten(bh))
1179                                                 done = 1;
1180                                         bh = bh->b_this_page;
1181                                 } while (!done && (bh != head));
1182                         }
1183                         unlock_page(page);
1184                         if (done)
1185                                 break;
1186                         idx++;
1187                         num++;
1188                         if (num >= max_pages)
1189                                 break;
1190                 }
1191                 pagevec_release(&pvec);
1192         }
1193         return num;
1194 }
1195
1196 /*
1197  * The ext4_get_blocks() function tries to look up the requested blocks,
1198  * and returns if the blocks are already mapped.
1199  *
1200  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1201  * and store the allocated blocks in the result buffer head and mark it
1202  * mapped.
1203  *
1204  * If file type is extents based, it will call ext4_ext_get_blocks(),
1205  * Otherwise, call with ext4_ind_get_blocks() to handle indirect mapping
1206  * based files
1207  *
1208  * On success, it returns the number of blocks being mapped or allocate.
1209  * if create==0 and the blocks are pre-allocated and uninitialized block,
1210  * the result buffer head is unmapped. If the create ==1, it will make sure
1211  * the buffer head is mapped.
1212  *
1213  * It returns 0 if plain look up failed (blocks have not been allocated), in
1214  * that casem, buffer head is unmapped
1215  *
1216  * It returns the error in case of allocation failure.
1217  */
1218 int ext4_get_blocks(handle_t *handle, struct inode *inode, sector_t block,
1219                     unsigned int max_blocks, struct buffer_head *bh,
1220                     int flags)
1221 {
1222         int retval;
1223
1224         clear_buffer_mapped(bh);
1225         clear_buffer_unwritten(bh);
1226
1227         ext_debug("ext4_get_blocks(): inode %lu, flag %d, max_blocks %u,"
1228                   "logical block %lu\n", inode->i_ino, flags, max_blocks,
1229                   (unsigned long)block);
1230         /*
1231          * Try to see if we can get the block without requesting a new
1232          * file system block.
1233          */
1234         down_read((&EXT4_I(inode)->i_data_sem));
1235         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1236                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1237                                 bh, 0);
1238         } else {
1239                 retval = ext4_ind_get_blocks(handle, inode, block, max_blocks,
1240                                              bh, 0);
1241         }
1242         up_read((&EXT4_I(inode)->i_data_sem));
1243
1244         if (retval > 0 && buffer_mapped(bh)) {
1245                 int ret = check_block_validity(inode, "file system corruption",
1246                                                block, bh->b_blocknr, retval);
1247                 if (ret != 0)
1248                         return ret;
1249         }
1250
1251         /* If it is only a block(s) look up */
1252         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1253                 return retval;
1254
1255         /*
1256          * Returns if the blocks have already allocated
1257          *
1258          * Note that if blocks have been preallocated
1259          * ext4_ext_get_block() returns th create = 0
1260          * with buffer head unmapped.
1261          */
1262         if (retval > 0 && buffer_mapped(bh))
1263                 return retval;
1264
1265         /*
1266          * When we call get_blocks without the create flag, the
1267          * BH_Unwritten flag could have gotten set if the blocks
1268          * requested were part of a uninitialized extent.  We need to
1269          * clear this flag now that we are committed to convert all or
1270          * part of the uninitialized extent to be an initialized
1271          * extent.  This is because we need to avoid the combination
1272          * of BH_Unwritten and BH_Mapped flags being simultaneously
1273          * set on the buffer_head.
1274          */
1275         clear_buffer_unwritten(bh);
1276
1277         /*
1278          * New blocks allocate and/or writing to uninitialized extent
1279          * will possibly result in updating i_data, so we take
1280          * the write lock of i_data_sem, and call get_blocks()
1281          * with create == 1 flag.
1282          */
1283         down_write((&EXT4_I(inode)->i_data_sem));
1284
1285         /*
1286          * if the caller is from delayed allocation writeout path
1287          * we have already reserved fs blocks for allocation
1288          * let the underlying get_block() function know to
1289          * avoid double accounting
1290          */
1291         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1292                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1293         /*
1294          * We need to check for EXT4 here because migrate
1295          * could have changed the inode type in between
1296          */
1297         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1298                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1299                                               bh, flags);
1300         } else {
1301                 retval = ext4_ind_get_blocks(handle, inode, block,
1302                                              max_blocks, bh, flags);
1303
1304                 if (retval > 0 && buffer_new(bh)) {
1305                         /*
1306                          * We allocated new blocks which will result in
1307                          * i_data's format changing.  Force the migrate
1308                          * to fail by clearing migrate flags
1309                          */
1310                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
1311                 }
1312
1313                 /*
1314                  * Update reserved blocks/metadata blocks after successful
1315                  * block allocation which had been deferred till now. We don't
1316                  * support fallocate for non extent files. So we can update
1317                  * reserve space here.
1318                  */
1319                 if ((retval > 0) &&
1320                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
1321                         ext4_da_update_reserve_space(inode, retval, 1);
1322         }
1323         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1324                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1325
1326         up_write((&EXT4_I(inode)->i_data_sem));
1327         if (retval > 0 && buffer_mapped(bh)) {
1328                 int ret = check_block_validity(inode, "file system "
1329                                                "corruption after allocation",
1330                                                block, bh->b_blocknr, retval);
1331                 if (ret != 0)
1332                         return ret;
1333         }
1334         return retval;
1335 }
1336
1337 /* Maximum number of blocks we map for direct IO at once. */
1338 #define DIO_MAX_BLOCKS 4096
1339
1340 int ext4_get_block(struct inode *inode, sector_t iblock,
1341                    struct buffer_head *bh_result, int create)
1342 {
1343         handle_t *handle = ext4_journal_current_handle();
1344         int ret = 0, started = 0;
1345         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1346         int dio_credits;
1347
1348         if (create && !handle) {
1349                 /* Direct IO write... */
1350                 if (max_blocks > DIO_MAX_BLOCKS)
1351                         max_blocks = DIO_MAX_BLOCKS;
1352                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
1353                 handle = ext4_journal_start(inode, dio_credits);
1354                 if (IS_ERR(handle)) {
1355                         ret = PTR_ERR(handle);
1356                         goto out;
1357                 }
1358                 started = 1;
1359         }
1360
1361         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
1362                               create ? EXT4_GET_BLOCKS_CREATE : 0);
1363         if (ret > 0) {
1364                 bh_result->b_size = (ret << inode->i_blkbits);
1365                 ret = 0;
1366         }
1367         if (started)
1368                 ext4_journal_stop(handle);
1369 out:
1370         return ret;
1371 }
1372
1373 /*
1374  * `handle' can be NULL if create is zero
1375  */
1376 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1377                                 ext4_lblk_t block, int create, int *errp)
1378 {
1379         struct buffer_head dummy;
1380         int fatal = 0, err;
1381         int flags = 0;
1382
1383         J_ASSERT(handle != NULL || create == 0);
1384
1385         dummy.b_state = 0;
1386         dummy.b_blocknr = -1000;
1387         buffer_trace_init(&dummy.b_history);
1388         if (create)
1389                 flags |= EXT4_GET_BLOCKS_CREATE;
1390         err = ext4_get_blocks(handle, inode, block, 1, &dummy, flags);
1391         /*
1392          * ext4_get_blocks() returns number of blocks mapped. 0 in
1393          * case of a HOLE.
1394          */
1395         if (err > 0) {
1396                 if (err > 1)
1397                         WARN_ON(1);
1398                 err = 0;
1399         }
1400         *errp = err;
1401         if (!err && buffer_mapped(&dummy)) {
1402                 struct buffer_head *bh;
1403                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1404                 if (!bh) {
1405                         *errp = -EIO;
1406                         goto err;
1407                 }
1408                 if (buffer_new(&dummy)) {
1409                         J_ASSERT(create != 0);
1410                         J_ASSERT(handle != NULL);
1411
1412                         /*
1413                          * Now that we do not always journal data, we should
1414                          * keep in mind whether this should always journal the
1415                          * new buffer as metadata.  For now, regular file
1416                          * writes use ext4_get_block instead, so it's not a
1417                          * problem.
1418                          */
1419                         lock_buffer(bh);
1420                         BUFFER_TRACE(bh, "call get_create_access");
1421                         fatal = ext4_journal_get_create_access(handle, bh);
1422                         if (!fatal && !buffer_uptodate(bh)) {
1423                                 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1424                                 set_buffer_uptodate(bh);
1425                         }
1426                         unlock_buffer(bh);
1427                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1428                         err = ext4_handle_dirty_metadata(handle, inode, bh);
1429                         if (!fatal)
1430                                 fatal = err;
1431                 } else {
1432                         BUFFER_TRACE(bh, "not a new buffer");
1433                 }
1434                 if (fatal) {
1435                         *errp = fatal;
1436                         brelse(bh);
1437                         bh = NULL;
1438                 }
1439                 return bh;
1440         }
1441 err:
1442         return NULL;
1443 }
1444
1445 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1446                                ext4_lblk_t block, int create, int *err)
1447 {
1448         struct buffer_head *bh;
1449
1450         bh = ext4_getblk(handle, inode, block, create, err);
1451         if (!bh)
1452                 return bh;
1453         if (buffer_uptodate(bh))
1454                 return bh;
1455         ll_rw_block(READ_META, 1, &bh);
1456         wait_on_buffer(bh);
1457         if (buffer_uptodate(bh))
1458                 return bh;
1459         put_bh(bh);
1460         *err = -EIO;
1461         return NULL;
1462 }
1463
1464 static int walk_page_buffers(handle_t *handle,
1465                              struct buffer_head *head,
1466                              unsigned from,
1467                              unsigned to,
1468                              int *partial,
1469                              int (*fn)(handle_t *handle,
1470                                        struct buffer_head *bh))
1471 {
1472         struct buffer_head *bh;
1473         unsigned block_start, block_end;
1474         unsigned blocksize = head->b_size;
1475         int err, ret = 0;
1476         struct buffer_head *next;
1477
1478         for (bh = head, block_start = 0;
1479              ret == 0 && (bh != head || !block_start);
1480              block_start = block_end, bh = next) {
1481                 next = bh->b_this_page;
1482                 block_end = block_start + blocksize;
1483                 if (block_end <= from || block_start >= to) {
1484                         if (partial && !buffer_uptodate(bh))
1485                                 *partial = 1;
1486                         continue;
1487                 }
1488                 err = (*fn)(handle, bh);
1489                 if (!ret)
1490                         ret = err;
1491         }
1492         return ret;
1493 }
1494
1495 /*
1496  * To preserve ordering, it is essential that the hole instantiation and
1497  * the data write be encapsulated in a single transaction.  We cannot
1498  * close off a transaction and start a new one between the ext4_get_block()
1499  * and the commit_write().  So doing the jbd2_journal_start at the start of
1500  * prepare_write() is the right place.
1501  *
1502  * Also, this function can nest inside ext4_writepage() ->
1503  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1504  * has generated enough buffer credits to do the whole page.  So we won't
1505  * block on the journal in that case, which is good, because the caller may
1506  * be PF_MEMALLOC.
1507  *
1508  * By accident, ext4 can be reentered when a transaction is open via
1509  * quota file writes.  If we were to commit the transaction while thus
1510  * reentered, there can be a deadlock - we would be holding a quota
1511  * lock, and the commit would never complete if another thread had a
1512  * transaction open and was blocking on the quota lock - a ranking
1513  * violation.
1514  *
1515  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1516  * will _not_ run commit under these circumstances because handle->h_ref
1517  * is elevated.  We'll still have enough credits for the tiny quotafile
1518  * write.
1519  */
1520 static int do_journal_get_write_access(handle_t *handle,
1521                                        struct buffer_head *bh)
1522 {
1523         if (!buffer_mapped(bh) || buffer_freed(bh))
1524                 return 0;
1525         return ext4_journal_get_write_access(handle, bh);
1526 }
1527
1528 /*
1529  * Truncate blocks that were not used by write. We have to truncate the
1530  * pagecache as well so that corresponding buffers get properly unmapped.
1531  */
1532 static void ext4_truncate_failed_write(struct inode *inode)
1533 {
1534         truncate_inode_pages(inode->i_mapping, inode->i_size);
1535         ext4_truncate(inode);
1536 }
1537
1538 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
1539                    struct buffer_head *bh_result, int create);
1540 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1541                             loff_t pos, unsigned len, unsigned flags,
1542                             struct page **pagep, void **fsdata)
1543 {
1544         struct inode *inode = mapping->host;
1545         int ret, needed_blocks;
1546         handle_t *handle;
1547         int retries = 0;
1548         struct page *page;
1549         pgoff_t index;
1550         unsigned from, to;
1551
1552         trace_ext4_write_begin(inode, pos, len, flags);
1553         /*
1554          * Reserve one block more for addition to orphan list in case
1555          * we allocate blocks but write fails for some reason
1556          */
1557         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1558         index = pos >> PAGE_CACHE_SHIFT;
1559         from = pos & (PAGE_CACHE_SIZE - 1);
1560         to = from + len;
1561
1562 retry:
1563         handle = ext4_journal_start(inode, needed_blocks);
1564         if (IS_ERR(handle)) {
1565                 ret = PTR_ERR(handle);
1566                 goto out;
1567         }
1568
1569         /* We cannot recurse into the filesystem as the transaction is already
1570          * started */
1571         flags |= AOP_FLAG_NOFS;
1572
1573         page = grab_cache_page_write_begin(mapping, index, flags);
1574         if (!page) {
1575                 ext4_journal_stop(handle);
1576                 ret = -ENOMEM;
1577                 goto out;
1578         }
1579         *pagep = page;
1580
1581         if (ext4_should_dioread_nolock(inode))
1582                 ret = block_write_begin(file, mapping, pos, len, flags, pagep,
1583                                 fsdata, ext4_get_block_write);
1584         else
1585                 ret = block_write_begin(file, mapping, pos, len, flags, pagep,
1586                                 fsdata, ext4_get_block);
1587
1588         if (!ret && ext4_should_journal_data(inode)) {
1589                 ret = walk_page_buffers(handle, page_buffers(page),
1590                                 from, to, NULL, do_journal_get_write_access);
1591         }
1592
1593         if (ret) {
1594                 unlock_page(page);
1595                 page_cache_release(page);
1596                 /*
1597                  * block_write_begin may have instantiated a few blocks
1598                  * outside i_size.  Trim these off again. Don't need
1599                  * i_size_read because we hold i_mutex.
1600                  *
1601                  * Add inode to orphan list in case we crash before
1602                  * truncate finishes
1603                  */
1604                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1605                         ext4_orphan_add(handle, inode);
1606
1607                 ext4_journal_stop(handle);
1608                 if (pos + len > inode->i_size) {
1609                         ext4_truncate_failed_write(inode);
1610                         /*
1611                          * If truncate failed early the inode might
1612                          * still be on the orphan list; we need to
1613                          * make sure the inode is removed from the
1614                          * orphan list in that case.
1615                          */
1616                         if (inode->i_nlink)
1617                                 ext4_orphan_del(NULL, inode);
1618                 }
1619         }
1620
1621         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1622                 goto retry;
1623 out:
1624         return ret;
1625 }
1626
1627 /* For write_end() in data=journal mode */
1628 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1629 {
1630         if (!buffer_mapped(bh) || buffer_freed(bh))
1631                 return 0;
1632         set_buffer_uptodate(bh);
1633         return ext4_handle_dirty_metadata(handle, NULL, bh);
1634 }
1635
1636 static int ext4_generic_write_end(struct file *file,
1637                                   struct address_space *mapping,
1638                                   loff_t pos, unsigned len, unsigned copied,
1639                                   struct page *page, void *fsdata)
1640 {
1641         int i_size_changed = 0;
1642         struct inode *inode = mapping->host;
1643         handle_t *handle = ext4_journal_current_handle();
1644
1645         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1646
1647         /*
1648          * No need to use i_size_read() here, the i_size
1649          * cannot change under us because we hold i_mutex.
1650          *
1651          * But it's important to update i_size while still holding page lock:
1652          * page writeout could otherwise come in and zero beyond i_size.
1653          */
1654         if (pos + copied > inode->i_size) {
1655                 i_size_write(inode, pos + copied);
1656                 i_size_changed = 1;
1657         }
1658
1659         if (pos + copied >  EXT4_I(inode)->i_disksize) {
1660                 /* We need to mark inode dirty even if
1661                  * new_i_size is less that inode->i_size
1662                  * bu greater than i_disksize.(hint delalloc)
1663                  */
1664                 ext4_update_i_disksize(inode, (pos + copied));
1665                 i_size_changed = 1;
1666         }
1667         unlock_page(page);
1668         page_cache_release(page);
1669
1670         /*
1671          * Don't mark the inode dirty under page lock. First, it unnecessarily
1672          * makes the holding time of page lock longer. Second, it forces lock
1673          * ordering of page lock and transaction start for journaling
1674          * filesystems.
1675          */
1676         if (i_size_changed)
1677                 ext4_mark_inode_dirty(handle, inode);
1678
1679         return copied;
1680 }
1681
1682 /*
1683  * We need to pick up the new inode size which generic_commit_write gave us
1684  * `file' can be NULL - eg, when called from page_symlink().
1685  *
1686  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1687  * buffers are managed internally.
1688  */
1689 static int ext4_ordered_write_end(struct file *file,
1690                                   struct address_space *mapping,
1691                                   loff_t pos, unsigned len, unsigned copied,
1692                                   struct page *page, void *fsdata)
1693 {
1694         handle_t *handle = ext4_journal_current_handle();
1695         struct inode *inode = mapping->host;
1696         int ret = 0, ret2;
1697
1698         trace_ext4_ordered_write_end(inode, pos, len, copied);
1699         ret = ext4_jbd2_file_inode(handle, inode);
1700
1701         if (ret == 0) {
1702                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1703                                                         page, fsdata);
1704                 copied = ret2;
1705                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1706                         /* if we have allocated more blocks and copied
1707                          * less. We will have blocks allocated outside
1708                          * inode->i_size. So truncate them
1709                          */
1710                         ext4_orphan_add(handle, inode);
1711                 if (ret2 < 0)
1712                         ret = ret2;
1713         }
1714         ret2 = ext4_journal_stop(handle);
1715         if (!ret)
1716                 ret = ret2;
1717
1718         if (pos + len > inode->i_size) {
1719                 ext4_truncate_failed_write(inode);
1720                 /*
1721                  * If truncate failed early the inode might still be
1722                  * on the orphan list; we need to make sure the inode
1723                  * is removed from the orphan list in that case.
1724                  */
1725                 if (inode->i_nlink)
1726                         ext4_orphan_del(NULL, inode);
1727         }
1728
1729
1730         return ret ? ret : copied;
1731 }
1732
1733 static int ext4_writeback_write_end(struct file *file,
1734                                     struct address_space *mapping,
1735                                     loff_t pos, unsigned len, unsigned copied,
1736                                     struct page *page, void *fsdata)
1737 {
1738         handle_t *handle = ext4_journal_current_handle();
1739         struct inode *inode = mapping->host;
1740         int ret = 0, ret2;
1741
1742         trace_ext4_writeback_write_end(inode, pos, len, copied);
1743         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1744                                                         page, fsdata);
1745         copied = ret2;
1746         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1747                 /* if we have allocated more blocks and copied
1748                  * less. We will have blocks allocated outside
1749                  * inode->i_size. So truncate them
1750                  */
1751                 ext4_orphan_add(handle, inode);
1752
1753         if (ret2 < 0)
1754                 ret = ret2;
1755
1756         ret2 = ext4_journal_stop(handle);
1757         if (!ret)
1758                 ret = ret2;
1759
1760         if (pos + len > inode->i_size) {
1761                 ext4_truncate_failed_write(inode);
1762                 /*
1763                  * If truncate failed early the inode might still be
1764                  * on the orphan list; we need to make sure the inode
1765                  * is removed from the orphan list in that case.
1766                  */
1767                 if (inode->i_nlink)
1768                         ext4_orphan_del(NULL, inode);
1769         }
1770
1771         return ret ? ret : copied;
1772 }
1773
1774 static int ext4_journalled_write_end(struct file *file,
1775                                      struct address_space *mapping,
1776                                      loff_t pos, unsigned len, unsigned copied,
1777                                      struct page *page, void *fsdata)
1778 {
1779         handle_t *handle = ext4_journal_current_handle();
1780         struct inode *inode = mapping->host;
1781         int ret = 0, ret2;
1782         int partial = 0;
1783         unsigned from, to;
1784         loff_t new_i_size;
1785
1786         trace_ext4_journalled_write_end(inode, pos, len, copied);
1787         from = pos & (PAGE_CACHE_SIZE - 1);
1788         to = from + len;
1789
1790         if (copied < len) {
1791                 if (!PageUptodate(page))
1792                         copied = 0;
1793                 page_zero_new_buffers(page, from+copied, to);
1794         }
1795
1796         ret = walk_page_buffers(handle, page_buffers(page), from,
1797                                 to, &partial, write_end_fn);
1798         if (!partial)
1799                 SetPageUptodate(page);
1800         new_i_size = pos + copied;
1801         if (new_i_size > inode->i_size)
1802                 i_size_write(inode, pos+copied);
1803         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1804         if (new_i_size > EXT4_I(inode)->i_disksize) {
1805                 ext4_update_i_disksize(inode, new_i_size);
1806                 ret2 = ext4_mark_inode_dirty(handle, inode);
1807                 if (!ret)
1808                         ret = ret2;
1809         }
1810
1811         unlock_page(page);
1812         page_cache_release(page);
1813         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1814                 /* if we have allocated more blocks and copied
1815                  * less. We will have blocks allocated outside
1816                  * inode->i_size. So truncate them
1817                  */
1818                 ext4_orphan_add(handle, inode);
1819
1820         ret2 = ext4_journal_stop(handle);
1821         if (!ret)
1822                 ret = ret2;
1823         if (pos + len > inode->i_size) {
1824                 ext4_truncate_failed_write(inode);
1825                 /*
1826                  * If truncate failed early the inode might still be
1827                  * on the orphan list; we need to make sure the inode
1828                  * is removed from the orphan list in that case.
1829                  */
1830                 if (inode->i_nlink)
1831                         ext4_orphan_del(NULL, inode);
1832         }
1833
1834         return ret ? ret : copied;
1835 }
1836
1837 /*
1838  * Reserve a single block located at lblock
1839  */
1840 static int ext4_da_reserve_space(struct inode *inode, sector_t lblock)
1841 {
1842         int retries = 0;
1843         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1844         struct ext4_inode_info *ei = EXT4_I(inode);
1845         unsigned long md_needed, md_reserved;
1846
1847         /*
1848          * recalculate the amount of metadata blocks to reserve
1849          * in order to allocate nrblocks
1850          * worse case is one extent per block
1851          */
1852 repeat:
1853         spin_lock(&ei->i_block_reservation_lock);
1854         md_reserved = ei->i_reserved_meta_blocks;
1855         md_needed = ext4_calc_metadata_amount(inode, lblock);
1856         trace_ext4_da_reserve_space(inode, md_needed);
1857         spin_unlock(&ei->i_block_reservation_lock);
1858
1859         /*
1860          * Make quota reservation here to prevent quota overflow
1861          * later. Real quota accounting is done at pages writeout
1862          * time.
1863          */
1864         if (vfs_dq_reserve_block(inode, md_needed + 1))
1865                 return -EDQUOT;
1866
1867         if (ext4_claim_free_blocks(sbi, md_needed + 1)) {
1868                 vfs_dq_release_reservation_block(inode, md_needed + 1);
1869                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1870                         yield();
1871                         goto repeat;
1872                 }
1873                 return -ENOSPC;
1874         }
1875         spin_lock(&ei->i_block_reservation_lock);
1876         ei->i_reserved_data_blocks++;
1877         ei->i_reserved_meta_blocks += md_needed;
1878         spin_unlock(&ei->i_block_reservation_lock);
1879
1880         return 0;       /* success */
1881 }
1882
1883 static void ext4_da_release_space(struct inode *inode, int to_free)
1884 {
1885         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1886         struct ext4_inode_info *ei = EXT4_I(inode);
1887
1888         if (!to_free)
1889                 return;         /* Nothing to release, exit */
1890
1891         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1892
1893         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1894                 /*
1895                  * if there aren't enough reserved blocks, then the
1896                  * counter is messed up somewhere.  Since this
1897                  * function is called from invalidate page, it's
1898                  * harmless to return without any action.
1899                  */
1900                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1901                          "ino %lu, to_free %d with only %d reserved "
1902                          "data blocks\n", inode->i_ino, to_free,
1903                          ei->i_reserved_data_blocks);
1904                 WARN_ON(1);
1905                 to_free = ei->i_reserved_data_blocks;
1906         }
1907         ei->i_reserved_data_blocks -= to_free;
1908
1909         if (ei->i_reserved_data_blocks == 0) {
1910                 /*
1911                  * We can release all of the reserved metadata blocks
1912                  * only when we have written all of the delayed
1913                  * allocation blocks.
1914                  */
1915                 to_free += ei->i_reserved_meta_blocks;
1916                 ei->i_reserved_meta_blocks = 0;
1917                 ei->i_da_metadata_calc_len = 0;
1918         }
1919
1920         /* update fs dirty blocks counter */
1921         percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1922
1923         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1924
1925         vfs_dq_release_reservation_block(inode, to_free);
1926 }
1927
1928 static void ext4_da_page_release_reservation(struct page *page,
1929                                              unsigned long offset)
1930 {
1931         int to_release = 0;
1932         struct buffer_head *head, *bh;
1933         unsigned int curr_off = 0;
1934
1935         head = page_buffers(page);
1936         bh = head;
1937         do {
1938                 unsigned int next_off = curr_off + bh->b_size;
1939
1940                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1941                         to_release++;
1942                         clear_buffer_delay(bh);
1943                 }
1944                 curr_off = next_off;
1945         } while ((bh = bh->b_this_page) != head);
1946         ext4_da_release_space(page->mapping->host, to_release);
1947 }
1948
1949 /*
1950  * Delayed allocation stuff
1951  */
1952
1953 /*
1954  * mpage_da_submit_io - walks through extent of pages and try to write
1955  * them with writepage() call back
1956  *
1957  * @mpd->inode: inode
1958  * @mpd->first_page: first page of the extent
1959  * @mpd->next_page: page after the last page of the extent
1960  *
1961  * By the time mpage_da_submit_io() is called we expect all blocks
1962  * to be allocated. this may be wrong if allocation failed.
1963  *
1964  * As pages are already locked by write_cache_pages(), we can't use it
1965  */
1966 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1967 {
1968         long pages_skipped;
1969         struct pagevec pvec;
1970         unsigned long index, end;
1971         int ret = 0, err, nr_pages, i;
1972         struct inode *inode = mpd->inode;
1973         struct address_space *mapping = inode->i_mapping;
1974
1975         BUG_ON(mpd->next_page <= mpd->first_page);
1976         /*
1977          * We need to start from the first_page to the next_page - 1
1978          * to make sure we also write the mapped dirty buffer_heads.
1979          * If we look at mpd->b_blocknr we would only be looking
1980          * at the currently mapped buffer_heads.
1981          */
1982         index = mpd->first_page;
1983         end = mpd->next_page - 1;
1984
1985         pagevec_init(&pvec, 0);
1986         while (index <= end) {
1987                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1988                 if (nr_pages == 0)
1989                         break;
1990                 for (i = 0; i < nr_pages; i++) {
1991                         struct page *page = pvec.pages[i];
1992
1993                         index = page->index;
1994                         if (index > end)
1995                                 break;
1996                         index++;
1997
1998                         BUG_ON(!PageLocked(page));
1999                         BUG_ON(PageWriteback(page));
2000
2001                         pages_skipped = mpd->wbc->pages_skipped;
2002                         err = mapping->a_ops->writepage(page, mpd->wbc);
2003                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
2004                                 /*
2005                                  * have successfully written the page
2006                                  * without skipping the same
2007                                  */
2008                                 mpd->pages_written++;
2009                         /*
2010                          * In error case, we have to continue because
2011                          * remaining pages are still locked
2012                          * XXX: unlock and re-dirty them?
2013                          */
2014                         if (ret == 0)
2015                                 ret = err;
2016                 }
2017                 pagevec_release(&pvec);
2018         }
2019         return ret;
2020 }
2021
2022 /*
2023  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
2024  *
2025  * @mpd->inode - inode to walk through
2026  * @exbh->b_blocknr - first block on a disk
2027  * @exbh->b_size - amount of space in bytes
2028  * @logical - first logical block to start assignment with
2029  *
2030  * the function goes through all passed space and put actual disk
2031  * block numbers into buffer heads, dropping BH_Delay and BH_Unwritten
2032  */
2033 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
2034                                  struct buffer_head *exbh)
2035 {
2036         struct inode *inode = mpd->inode;
2037         struct address_space *mapping = inode->i_mapping;
2038         int blocks = exbh->b_size >> inode->i_blkbits;
2039         sector_t pblock = exbh->b_blocknr, cur_logical;
2040         struct buffer_head *head, *bh;
2041         pgoff_t index, end;
2042         struct pagevec pvec;
2043         int nr_pages, i;
2044
2045         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2046         end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2047         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2048
2049         pagevec_init(&pvec, 0);
2050
2051         while (index <= end) {
2052                 /* XXX: optimize tail */
2053                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2054                 if (nr_pages == 0)
2055                         break;
2056                 for (i = 0; i < nr_pages; i++) {
2057                         struct page *page = pvec.pages[i];
2058
2059                         index = page->index;
2060                         if (index > end)
2061                                 break;
2062                         index++;
2063
2064                         BUG_ON(!PageLocked(page));
2065                         BUG_ON(PageWriteback(page));
2066                         BUG_ON(!page_has_buffers(page));
2067
2068                         bh = page_buffers(page);
2069                         head = bh;
2070
2071                         /* skip blocks out of the range */
2072                         do {
2073                                 if (cur_logical >= logical)
2074                                         break;
2075                                 cur_logical++;
2076                         } while ((bh = bh->b_this_page) != head);
2077
2078                         do {
2079                                 if (cur_logical >= logical + blocks)
2080                                         break;
2081
2082                                 if (buffer_delay(bh) ||
2083                                                 buffer_unwritten(bh)) {
2084
2085                                         BUG_ON(bh->b_bdev != inode->i_sb->s_bdev);
2086
2087                                         if (buffer_delay(bh)) {
2088                                                 clear_buffer_delay(bh);
2089                                                 bh->b_blocknr = pblock;
2090                                         } else {
2091                                                 /*
2092                                                  * unwritten already should have
2093                                                  * blocknr assigned. Verify that
2094                                                  */
2095                                                 clear_buffer_unwritten(bh);
2096                                                 BUG_ON(bh->b_blocknr != pblock);
2097                                         }
2098
2099                                 } else if (buffer_mapped(bh))
2100                                         BUG_ON(bh->b_blocknr != pblock);
2101
2102                                 if (buffer_uninit(exbh))
2103                                         set_buffer_uninit(bh);
2104                                 cur_logical++;
2105                                 pblock++;
2106                         } while ((bh = bh->b_this_page) != head);
2107                 }
2108                 pagevec_release(&pvec);
2109         }
2110 }
2111
2112
2113 /*
2114  * __unmap_underlying_blocks - just a helper function to unmap
2115  * set of blocks described by @bh
2116  */
2117 static inline void __unmap_underlying_blocks(struct inode *inode,
2118                                              struct buffer_head *bh)
2119 {
2120         struct block_device *bdev = inode->i_sb->s_bdev;
2121         int blocks, i;
2122
2123         blocks = bh->b_size >> inode->i_blkbits;
2124         for (i = 0; i < blocks; i++)
2125                 unmap_underlying_metadata(bdev, bh->b_blocknr + i);
2126 }
2127
2128 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2129                                         sector_t logical, long blk_cnt)
2130 {
2131         int nr_pages, i;
2132         pgoff_t index, end;
2133         struct pagevec pvec;
2134         struct inode *inode = mpd->inode;
2135         struct address_space *mapping = inode->i_mapping;
2136
2137         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2138         end   = (logical + blk_cnt - 1) >>
2139                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2140         while (index <= end) {
2141                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2142                 if (nr_pages == 0)
2143                         break;
2144                 for (i = 0; i < nr_pages; i++) {
2145                         struct page *page = pvec.pages[i];
2146                         index = page->index;
2147                         if (index > end)
2148                                 break;
2149                         index++;
2150
2151                         BUG_ON(!PageLocked(page));
2152                         BUG_ON(PageWriteback(page));
2153                         block_invalidatepage(page, 0);
2154                         ClearPageUptodate(page);
2155                         unlock_page(page);
2156                 }
2157         }
2158         return;
2159 }
2160
2161 static void ext4_print_free_blocks(struct inode *inode)
2162 {
2163         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2164         printk(KERN_CRIT "Total free blocks count %lld\n",
2165                ext4_count_free_blocks(inode->i_sb));
2166         printk(KERN_CRIT "Free/Dirty block details\n");
2167         printk(KERN_CRIT "free_blocks=%lld\n",
2168                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2169         printk(KERN_CRIT "dirty_blocks=%lld\n",
2170                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2171         printk(KERN_CRIT "Block reservation details\n");
2172         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2173                EXT4_I(inode)->i_reserved_data_blocks);
2174         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2175                EXT4_I(inode)->i_reserved_meta_blocks);
2176         return;
2177 }
2178
2179 /*
2180  * mpage_da_map_blocks - go through given space
2181  *
2182  * @mpd - bh describing space
2183  *
2184  * The function skips space we know is already mapped to disk blocks.
2185  *
2186  */
2187 static int mpage_da_map_blocks(struct mpage_da_data *mpd)
2188 {
2189         int err, blks, get_blocks_flags;
2190         struct buffer_head new;
2191         sector_t next = mpd->b_blocknr;
2192         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2193         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2194         handle_t *handle = NULL;
2195
2196         /*
2197          * We consider only non-mapped and non-allocated blocks
2198          */
2199         if ((mpd->b_state  & (1 << BH_Mapped)) &&
2200                 !(mpd->b_state & (1 << BH_Delay)) &&
2201                 !(mpd->b_state & (1 << BH_Unwritten)))
2202                 return 0;
2203
2204         /*
2205          * If we didn't accumulate anything to write simply return
2206          */
2207         if (!mpd->b_size)
2208                 return 0;
2209
2210         handle = ext4_journal_current_handle();
2211         BUG_ON(!handle);
2212
2213         /*
2214          * Call ext4_get_blocks() to allocate any delayed allocation
2215          * blocks, or to convert an uninitialized extent to be
2216          * initialized (in the case where we have written into
2217          * one or more preallocated blocks).
2218          *
2219          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2220          * indicate that we are on the delayed allocation path.  This
2221          * affects functions in many different parts of the allocation
2222          * call path.  This flag exists primarily because we don't
2223          * want to change *many* call functions, so ext4_get_blocks()
2224          * will set the magic i_delalloc_reserved_flag once the
2225          * inode's allocation semaphore is taken.
2226          *
2227          * If the blocks in questions were delalloc blocks, set
2228          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2229          * variables are updated after the blocks have been allocated.
2230          */
2231         new.b_state = 0;
2232         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
2233         if (ext4_should_dioread_nolock(mpd->inode))
2234                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2235         if (mpd->b_state & (1 << BH_Delay))
2236                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2237
2238         blks = ext4_get_blocks(handle, mpd->inode, next, max_blocks,
2239                                &new, get_blocks_flags);
2240         if (blks < 0) {
2241                 err = blks;
2242                 /*
2243                  * If get block returns with error we simply
2244                  * return. Later writepage will redirty the page and
2245                  * writepages will find the dirty page again
2246                  */
2247                 if (err == -EAGAIN)
2248                         return 0;
2249
2250                 if (err == -ENOSPC &&
2251                     ext4_count_free_blocks(mpd->inode->i_sb)) {
2252                         mpd->retval = err;
2253                         return 0;
2254                 }
2255
2256                 /*
2257                  * get block failure will cause us to loop in
2258                  * writepages, because a_ops->writepage won't be able
2259                  * to make progress. The page will be redirtied by
2260                  * writepage and writepages will again try to write
2261                  * the same.
2262                  */
2263                 ext4_msg(mpd->inode->i_sb, KERN_CRIT,
2264                          "delayed block allocation failed for inode %lu at "
2265                          "logical offset %llu with max blocks %zd with "
2266                          "error %d\n", mpd->inode->i_ino,
2267                          (unsigned long long) next,
2268                          mpd->b_size >> mpd->inode->i_blkbits, err);
2269                 printk(KERN_CRIT "This should not happen!!  "
2270                        "Data will be lost\n");
2271                 if (err == -ENOSPC) {
2272                         ext4_print_free_blocks(mpd->inode);
2273                 }
2274                 /* invalidate all the pages */
2275                 ext4_da_block_invalidatepages(mpd, next,
2276                                 mpd->b_size >> mpd->inode->i_blkbits);
2277                 return err;
2278         }
2279         BUG_ON(blks == 0);
2280
2281         new.b_size = (blks << mpd->inode->i_blkbits);
2282
2283         if (buffer_new(&new))
2284                 __unmap_underlying_blocks(mpd->inode, &new);
2285
2286         /*
2287          * If blocks are delayed marked, we need to
2288          * put actual blocknr and drop delayed bit
2289          */
2290         if ((mpd->b_state & (1 << BH_Delay)) ||
2291             (mpd->b_state & (1 << BH_Unwritten)))
2292                 mpage_put_bnr_to_bhs(mpd, next, &new);
2293
2294         if (ext4_should_order_data(mpd->inode)) {
2295                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2296                 if (err)
2297                         return err;
2298         }
2299
2300         /*
2301          * Update on-disk size along with block allocation.
2302          */
2303         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2304         if (disksize > i_size_read(mpd->inode))
2305                 disksize = i_size_read(mpd->inode);
2306         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2307                 ext4_update_i_disksize(mpd->inode, disksize);
2308                 return ext4_mark_inode_dirty(handle, mpd->inode);
2309         }
2310
2311         return 0;
2312 }
2313
2314 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2315                 (1 << BH_Delay) | (1 << BH_Unwritten))
2316
2317 /*
2318  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2319  *
2320  * @mpd->lbh - extent of blocks
2321  * @logical - logical number of the block in the file
2322  * @bh - bh of the block (used to access block's state)
2323  *
2324  * the function is used to collect contig. blocks in same state
2325  */
2326 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2327                                    sector_t logical, size_t b_size,
2328                                    unsigned long b_state)
2329 {
2330         sector_t next;
2331         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2332
2333         /* check if thereserved journal credits might overflow */
2334         if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
2335                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2336                         /*
2337                          * With non-extent format we are limited by the journal
2338                          * credit available.  Total credit needed to insert
2339                          * nrblocks contiguous blocks is dependent on the
2340                          * nrblocks.  So limit nrblocks.
2341                          */
2342                         goto flush_it;
2343                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2344                                 EXT4_MAX_TRANS_DATA) {
2345                         /*
2346                          * Adding the new buffer_head would make it cross the
2347                          * allowed limit for which we have journal credit
2348                          * reserved. So limit the new bh->b_size
2349                          */
2350                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2351                                                 mpd->inode->i_blkbits;
2352                         /* we will do mpage_da_submit_io in the next loop */
2353                 }
2354         }
2355         /*
2356          * First block in the extent
2357          */
2358         if (mpd->b_size == 0) {
2359                 mpd->b_blocknr = logical;
2360                 mpd->b_size = b_size;
2361                 mpd->b_state = b_state & BH_FLAGS;
2362                 return;
2363         }
2364
2365         next = mpd->b_blocknr + nrblocks;
2366         /*
2367          * Can we merge the block to our big extent?
2368          */
2369         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2370                 mpd->b_size += b_size;
2371                 return;
2372         }
2373
2374 flush_it:
2375         /*
2376          * We couldn't merge the block to our extent, so we
2377          * need to flush current  extent and start new one
2378          */
2379         if (mpage_da_map_blocks(mpd) == 0)
2380                 mpage_da_submit_io(mpd);
2381         mpd->io_done = 1;
2382         return;
2383 }
2384
2385 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
2386 {
2387         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
2388 }
2389
2390 /*
2391  * __mpage_da_writepage - finds extent of pages and blocks
2392  *
2393  * @page: page to consider
2394  * @wbc: not used, we just follow rules
2395  * @data: context
2396  *
2397  * The function finds extents of pages and scan them for all blocks.
2398  */
2399 static int __mpage_da_writepage(struct page *page,
2400                                 struct writeback_control *wbc, void *data)
2401 {
2402         struct mpage_da_data *mpd = data;
2403         struct inode *inode = mpd->inode;
2404         struct buffer_head *bh, *head;
2405         sector_t logical;
2406
2407         if (mpd->io_done) {
2408                 /*
2409                  * Rest of the page in the page_vec
2410                  * redirty then and skip then. We will
2411                  * try to write them again after
2412                  * starting a new transaction
2413                  */
2414                 redirty_page_for_writepage(wbc, page);
2415                 unlock_page(page);
2416                 return MPAGE_DA_EXTENT_TAIL;
2417         }
2418         /*
2419          * Can we merge this page to current extent?
2420          */
2421         if (mpd->next_page != page->index) {
2422                 /*
2423                  * Nope, we can't. So, we map non-allocated blocks
2424                  * and start IO on them using writepage()
2425                  */
2426                 if (mpd->next_page != mpd->first_page) {
2427                         if (mpage_da_map_blocks(mpd) == 0)
2428                                 mpage_da_submit_io(mpd);
2429                         /*
2430                          * skip rest of the page in the page_vec
2431                          */
2432                         mpd->io_done = 1;
2433                         redirty_page_for_writepage(wbc, page);
2434                         unlock_page(page);
2435                         return MPAGE_DA_EXTENT_TAIL;
2436                 }
2437
2438                 /*
2439                  * Start next extent of pages ...
2440                  */
2441                 mpd->first_page = page->index;
2442
2443                 /*
2444                  * ... and blocks
2445                  */
2446                 mpd->b_size = 0;
2447                 mpd->b_state = 0;
2448                 mpd->b_blocknr = 0;
2449         }
2450
2451         mpd->next_page = page->index + 1;
2452         logical = (sector_t) page->index <<
2453                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2454
2455         if (!page_has_buffers(page)) {
2456                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2457                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2458                 if (mpd->io_done)
2459                         return MPAGE_DA_EXTENT_TAIL;
2460         } else {
2461                 /*
2462                  * Page with regular buffer heads, just add all dirty ones
2463                  */
2464                 head = page_buffers(page);
2465                 bh = head;
2466                 do {
2467                         BUG_ON(buffer_locked(bh));
2468                         /*
2469                          * We need to try to allocate
2470                          * unmapped blocks in the same page.
2471                          * Otherwise we won't make progress
2472                          * with the page in ext4_writepage
2473                          */
2474                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2475                                 mpage_add_bh_to_extent(mpd, logical,
2476                                                        bh->b_size,
2477                                                        bh->b_state);
2478                                 if (mpd->io_done)
2479                                         return MPAGE_DA_EXTENT_TAIL;
2480                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2481                                 /*
2482                                  * mapped dirty buffer. We need to update
2483                                  * the b_state because we look at
2484                                  * b_state in mpage_da_map_blocks. We don't
2485                                  * update b_size because if we find an
2486                                  * unmapped buffer_head later we need to
2487                                  * use the b_state flag of that buffer_head.
2488                                  */
2489                                 if (mpd->b_size == 0)
2490                                         mpd->b_state = bh->b_state & BH_FLAGS;
2491                         }
2492                         logical++;
2493                 } while ((bh = bh->b_this_page) != head);
2494         }
2495
2496         return 0;
2497 }
2498
2499 /*
2500  * This is a special get_blocks_t callback which is used by
2501  * ext4_da_write_begin().  It will either return mapped block or
2502  * reserve space for a single block.
2503  *
2504  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2505  * We also have b_blocknr = -1 and b_bdev initialized properly
2506  *
2507  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2508  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2509  * initialized properly.
2510  */
2511 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2512                                   struct buffer_head *bh_result, int create)
2513 {
2514         int ret = 0;
2515         sector_t invalid_block = ~((sector_t) 0xffff);
2516
2517         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2518                 invalid_block = ~0;
2519
2520         BUG_ON(create == 0);
2521         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2522
2523         /*
2524          * first, we need to know whether the block is allocated already
2525          * preallocated blocks are unmapped but should treated
2526          * the same as allocated blocks.
2527          */
2528         ret = ext4_get_blocks(NULL, inode, iblock, 1,  bh_result, 0);
2529         if ((ret == 0) && !buffer_delay(bh_result)) {
2530                 /* the block isn't (pre)allocated yet, let's reserve space */
2531                 /*
2532                  * XXX: __block_prepare_write() unmaps passed block,
2533                  * is it OK?
2534                  */
2535                 ret = ext4_da_reserve_space(inode, iblock);
2536                 if (ret)
2537                         /* not enough space to reserve */
2538                         return ret;
2539
2540                 map_bh(bh_result, inode->i_sb, invalid_block);
2541                 set_buffer_new(bh_result);
2542                 set_buffer_delay(bh_result);
2543         } else if (ret > 0) {
2544                 bh_result->b_size = (ret << inode->i_blkbits);
2545                 if (buffer_unwritten(bh_result)) {
2546                         /* A delayed write to unwritten bh should
2547                          * be marked new and mapped.  Mapped ensures
2548                          * that we don't do get_block multiple times
2549                          * when we write to the same offset and new
2550                          * ensures that we do proper zero out for
2551                          * partial write.
2552                          */
2553                         set_buffer_new(bh_result);
2554                         set_buffer_mapped(bh_result);
2555                 }
2556                 ret = 0;
2557         }
2558
2559         return ret;
2560 }
2561
2562 /*
2563  * This function is used as a standard get_block_t calback function
2564  * when there is no desire to allocate any blocks.  It is used as a
2565  * callback function for block_prepare_write(), nobh_writepage(), and
2566  * block_write_full_page().  These functions should only try to map a
2567  * single block at a time.
2568  *
2569  * Since this function doesn't do block allocations even if the caller
2570  * requests it by passing in create=1, it is critically important that
2571  * any caller checks to make sure that any buffer heads are returned
2572  * by this function are either all already mapped or marked for
2573  * delayed allocation before calling nobh_writepage() or
2574  * block_write_full_page().  Otherwise, b_blocknr could be left
2575  * unitialized, and the page write functions will be taken by
2576  * surprise.
2577  */
2578 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2579                                    struct buffer_head *bh_result, int create)
2580 {
2581         int ret = 0;
2582         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2583
2584         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2585
2586         /*
2587          * we don't want to do block allocation in writepage
2588          * so call get_block_wrap with create = 0
2589          */
2590         ret = ext4_get_blocks(NULL, inode, iblock, max_blocks, bh_result, 0);
2591         if (ret > 0) {
2592                 bh_result->b_size = (ret << inode->i_blkbits);
2593                 ret = 0;
2594         }
2595         return ret;
2596 }
2597
2598 static int bget_one(handle_t *handle, struct buffer_head *bh)
2599 {
2600         get_bh(bh);
2601         return 0;
2602 }
2603
2604 static int bput_one(handle_t *handle, struct buffer_head *bh)
2605 {
2606         put_bh(bh);
2607         return 0;
2608 }
2609
2610 static int __ext4_journalled_writepage(struct page *page,
2611                                        unsigned int len)
2612 {
2613         struct address_space *mapping = page->mapping;
2614         struct inode *inode = mapping->host;
2615         struct buffer_head *page_bufs;
2616         handle_t *handle = NULL;
2617         int ret = 0;
2618         int err;
2619
2620         page_bufs = page_buffers(page);
2621         BUG_ON(!page_bufs);
2622         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2623         /* As soon as we unlock the page, it can go away, but we have
2624          * references to buffers so we are safe */
2625         unlock_page(page);
2626
2627         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2628         if (IS_ERR(handle)) {
2629                 ret = PTR_ERR(handle);
2630                 goto out;
2631         }
2632
2633         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2634                                 do_journal_get_write_access);
2635
2636         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2637                                 write_end_fn);
2638         if (ret == 0)
2639                 ret = err;
2640         err = ext4_journal_stop(handle);
2641         if (!ret)
2642                 ret = err;
2643
2644         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2645         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2646 out:
2647         return ret;
2648 }
2649
2650 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
2651 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
2652
2653 /*
2654  * Note that we don't need to start a transaction unless we're journaling data
2655  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2656  * need to file the inode to the transaction's list in ordered mode because if
2657  * we are writing back data added by write(), the inode is already there and if
2658  * we are writing back data modified via mmap(), noone guarantees in which
2659  * transaction the data will hit the disk. In case we are journaling data, we
2660  * cannot start transaction directly because transaction start ranks above page
2661  * lock so we have to do some magic.
2662  *
2663  * This function can get called via...
2664  *   - ext4_da_writepages after taking page lock (have journal handle)
2665  *   - journal_submit_inode_data_buffers (no journal handle)
2666  *   - shrink_page_list via pdflush (no journal handle)
2667  *   - grab_page_cache when doing write_begin (have journal handle)
2668  *
2669  * We don't do any block allocation in this function. If we have page with
2670  * multiple blocks we need to write those buffer_heads that are mapped. This
2671  * is important for mmaped based write. So if we do with blocksize 1K
2672  * truncate(f, 1024);
2673  * a = mmap(f, 0, 4096);
2674  * a[0] = 'a';
2675  * truncate(f, 4096);
2676  * we have in the page first buffer_head mapped via page_mkwrite call back
2677  * but other bufer_heads would be unmapped but dirty(dirty done via the
2678  * do_wp_page). So writepage should write the first block. If we modify
2679  * the mmap area beyond 1024 we will again get a page_fault and the
2680  * page_mkwrite callback will do the block allocation and mark the
2681  * buffer_heads mapped.
2682  *
2683  * We redirty the page if we have any buffer_heads that is either delay or
2684  * unwritten in the page.
2685  *
2686  * We can get recursively called as show below.
2687  *
2688  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2689  *              ext4_writepage()
2690  *
2691  * But since we don't do any block allocation we should not deadlock.
2692  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2693  */
2694 static int ext4_writepage(struct page *page,
2695                           struct writeback_control *wbc)
2696 {
2697         int ret = 0;
2698         loff_t size;
2699         unsigned int len;
2700         struct buffer_head *page_bufs = NULL;
2701         struct inode *inode = page->mapping->host;
2702
2703         trace_ext4_writepage(inode, page);
2704         size = i_size_read(inode);
2705         if (page->index == size >> PAGE_CACHE_SHIFT)
2706                 len = size & ~PAGE_CACHE_MASK;
2707         else
2708                 len = PAGE_CACHE_SIZE;
2709
2710         if (page_has_buffers(page)) {
2711                 page_bufs = page_buffers(page);
2712                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2713                                         ext4_bh_delay_or_unwritten)) {
2714                         /*
2715                          * We don't want to do  block allocation
2716                          * So redirty the page and return
2717                          * We may reach here when we do a journal commit
2718                          * via journal_submit_inode_data_buffers.
2719                          * If we don't have mapping block we just ignore
2720                          * them. We can also reach here via shrink_page_list
2721                          */
2722                         redirty_page_for_writepage(wbc, page);
2723                         unlock_page(page);
2724                         return 0;
2725                 }
2726         } else {
2727                 /*
2728                  * The test for page_has_buffers() is subtle:
2729                  * We know the page is dirty but it lost buffers. That means
2730                  * that at some moment in time after write_begin()/write_end()
2731                  * has been called all buffers have been clean and thus they
2732                  * must have been written at least once. So they are all
2733                  * mapped and we can happily proceed with mapping them
2734                  * and writing the page.
2735                  *
2736                  * Try to initialize the buffer_heads and check whether
2737                  * all are mapped and non delay. We don't want to
2738                  * do block allocation here.
2739                  */
2740                 ret = block_prepare_write(page, 0, len,
2741                                           noalloc_get_block_write);
2742                 if (!ret) {
2743                         page_bufs = page_buffers(page);
2744                         /* check whether all are mapped and non delay */
2745                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2746                                                 ext4_bh_delay_or_unwritten)) {
2747                                 redirty_page_for_writepage(wbc, page);
2748                                 unlock_page(page);
2749                                 return 0;
2750                         }
2751                 } else {
2752                         /*
2753                          * We can't do block allocation here
2754                          * so just redity the page and unlock
2755                          * and return
2756                          */
2757                         redirty_page_for_writepage(wbc, page);
2758                         unlock_page(page);
2759                         return 0;
2760                 }
2761                 /* now mark the buffer_heads as dirty and uptodate */
2762                 block_commit_write(page, 0, len);
2763         }
2764
2765         if (PageChecked(page) && ext4_should_journal_data(inode)) {
2766                 /*
2767                  * It's mmapped pagecache.  Add buffers and journal it.  There
2768                  * doesn't seem much point in redirtying the page here.
2769                  */
2770                 ClearPageChecked(page);
2771                 return __ext4_journalled_writepage(page, len);
2772         }
2773
2774         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2775                 ret = nobh_writepage(page, noalloc_get_block_write, wbc);
2776         else if (page_bufs && buffer_uninit(page_bufs)) {
2777                 ext4_set_bh_endio(page_bufs, inode);
2778                 ret = block_write_full_page_endio(page, noalloc_get_block_write,
2779                                             wbc, ext4_end_io_buffer_write);
2780         } else
2781                 ret = block_write_full_page(page, noalloc_get_block_write,
2782                                             wbc);
2783
2784         return ret;
2785 }
2786
2787 /*
2788  * This is called via ext4_da_writepages() to
2789  * calulate the total number of credits to reserve to fit
2790  * a single extent allocation into a single transaction,
2791  * ext4_da_writpeages() will loop calling this before
2792  * the block allocation.
2793  */
2794
2795 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2796 {
2797         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2798
2799         /*
2800          * With non-extent format the journal credit needed to
2801          * insert nrblocks contiguous block is dependent on
2802          * number of contiguous block. So we will limit
2803          * number of contiguous block to a sane value
2804          */
2805         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) &&
2806             (max_blocks > EXT4_MAX_TRANS_DATA))
2807                 max_blocks = EXT4_MAX_TRANS_DATA;
2808
2809         return ext4_chunk_trans_blocks(inode, max_blocks);
2810 }
2811
2812 static int ext4_da_writepages(struct address_space *mapping,
2813                               struct writeback_control *wbc)
2814 {
2815         pgoff_t index;
2816         int range_whole = 0;
2817         handle_t *handle = NULL;
2818         struct mpage_da_data mpd;
2819         struct inode *inode = mapping->host;
2820         int no_nrwrite_index_update;
2821         int pages_written = 0;
2822         long pages_skipped;
2823         unsigned int max_pages;
2824         int range_cyclic, cycled = 1, io_done = 0;
2825         int needed_blocks, ret = 0;
2826         long desired_nr_to_write, nr_to_writebump = 0;
2827         loff_t range_start = wbc->range_start;
2828         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2829
2830         trace_ext4_da_writepages(inode, wbc);
2831
2832         /*
2833          * No pages to write? This is mainly a kludge to avoid starting
2834          * a transaction for special inodes like journal inode on last iput()
2835          * because that could violate lock ordering on umount
2836          */
2837         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2838                 return 0;
2839
2840         /*
2841          * If the filesystem has aborted, it is read-only, so return
2842          * right away instead of dumping stack traces later on that
2843          * will obscure the real source of the problem.  We test
2844          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2845          * the latter could be true if the filesystem is mounted
2846          * read-only, and in that case, ext4_da_writepages should
2847          * *never* be called, so if that ever happens, we would want
2848          * the stack trace.
2849          */
2850         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2851                 return -EROFS;
2852
2853         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2854                 range_whole = 1;
2855
2856         range_cyclic = wbc->range_cyclic;
2857         if (wbc->range_cyclic) {
2858                 index = mapping->writeback_index;
2859                 if (index)
2860                         cycled = 0;
2861                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2862                 wbc->range_end  = LLONG_MAX;
2863                 wbc->range_cyclic = 0;
2864         } else
2865                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2866
2867         /*
2868          * This works around two forms of stupidity.  The first is in
2869          * the writeback code, which caps the maximum number of pages
2870          * written to be 1024 pages.  This is wrong on multiple
2871          * levels; different architectues have a different page size,
2872          * which changes the maximum amount of data which gets
2873          * written.  Secondly, 4 megabytes is way too small.  XFS
2874          * forces this value to be 16 megabytes by multiplying
2875          * nr_to_write parameter by four, and then relies on its
2876          * allocator to allocate larger extents to make them
2877          * contiguous.  Unfortunately this brings us to the second
2878          * stupidity, which is that ext4's mballoc code only allocates
2879          * at most 2048 blocks.  So we force contiguous writes up to
2880          * the number of dirty blocks in the inode, or
2881          * sbi->max_writeback_mb_bump whichever is smaller.
2882          */
2883         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2884         if (!range_cyclic && range_whole)
2885                 desired_nr_to_write = wbc->nr_to_write * 8;
2886         else
2887                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2888                                                            max_pages);
2889         if (desired_nr_to_write > max_pages)
2890                 desired_nr_to_write = max_pages;
2891
2892         if (wbc->nr_to_write < desired_nr_to_write) {
2893                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2894                 wbc->nr_to_write = desired_nr_to_write;
2895         }
2896
2897         mpd.wbc = wbc;
2898         mpd.inode = mapping->host;
2899
2900         /*
2901          * we don't want write_cache_pages to update
2902          * nr_to_write and writeback_index
2903          */
2904         no_nrwrite_index_update = wbc->no_nrwrite_index_update;
2905         wbc->no_nrwrite_index_update = 1;
2906         pages_skipped = wbc->pages_skipped;
2907
2908 retry:
2909         while (!ret && wbc->nr_to_write > 0) {
2910
2911                 /*
2912                  * we  insert one extent at a time. So we need
2913                  * credit needed for single extent allocation.
2914                  * journalled mode is currently not supported
2915                  * by delalloc
2916                  */
2917                 BUG_ON(ext4_should_journal_data(inode));
2918                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2919
2920                 /* start a new transaction*/
2921                 handle = ext4_journal_start(inode, needed_blocks);
2922                 if (IS_ERR(handle)) {
2923                         ret = PTR_ERR(handle);
2924                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2925                                "%ld pages, ino %lu; err %d\n", __func__,
2926                                 wbc->nr_to_write, inode->i_ino, ret);
2927                         goto out_writepages;
2928                 }
2929
2930                 /*
2931                  * Now call __mpage_da_writepage to find the next
2932                  * contiguous region of logical blocks that need
2933                  * blocks to be allocated by ext4.  We don't actually
2934                  * submit the blocks for I/O here, even though
2935                  * write_cache_pages thinks it will, and will set the
2936                  * pages as clean for write before calling
2937                  * __mpage_da_writepage().
2938                  */
2939                 mpd.b_size = 0;
2940                 mpd.b_state = 0;
2941                 mpd.b_blocknr = 0;
2942                 mpd.first_page = 0;
2943                 mpd.next_page = 0;
2944                 mpd.io_done = 0;
2945                 mpd.pages_written = 0;
2946                 mpd.retval = 0;
2947                 ret = write_cache_pages(mapping, wbc, __mpage_da_writepage,
2948                                         &mpd);
2949                 /*
2950                  * If we have a contiguous extent of pages and we
2951                  * haven't done the I/O yet, map the blocks and submit
2952                  * them for I/O.
2953                  */
2954                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2955                         if (mpage_da_map_blocks(&mpd) == 0)
2956                                 mpage_da_submit_io(&mpd);
2957                         mpd.io_done = 1;
2958                         ret = MPAGE_DA_EXTENT_TAIL;
2959                 }
2960                 trace_ext4_da_write_pages(inode, &mpd);
2961                 wbc->nr_to_write -= mpd.pages_written;
2962
2963                 ext4_journal_stop(handle);
2964
2965                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2966                         /* commit the transaction which would
2967                          * free blocks released in the transaction
2968                          * and try again
2969                          */
2970                         jbd2_journal_force_commit_nested(sbi->s_journal);
2971                         wbc->pages_skipped = pages_skipped;
2972                         ret = 0;
2973                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2974                         /*
2975                          * got one extent now try with
2976                          * rest of the pages
2977                          */
2978                         pages_written += mpd.pages_written;
2979                         wbc->pages_skipped = pages_skipped;
2980                         ret = 0;
2981                         io_done = 1;
2982                 } else if (wbc->nr_to_write)
2983                         /*
2984                          * There is no more writeout needed
2985                          * or we requested for a noblocking writeout
2986                          * and we found the device congested
2987                          */
2988                         break;
2989         }
2990         if (!io_done && !cycled) {
2991                 cycled = 1;
2992                 index = 0;
2993                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2994                 wbc->range_end  = mapping->writeback_index - 1;
2995                 goto retry;
2996         }
2997         if (pages_skipped != wbc->pages_skipped)
2998                 ext4_msg(inode->i_sb, KERN_CRIT,
2999                          "This should not happen leaving %s "
3000                          "with nr_to_write = %ld ret = %d\n",
3001                          __func__, wbc->nr_to_write, ret);
3002
3003         /* Update index */
3004         index += pages_written;
3005         wbc->range_cyclic = range_cyclic;
3006         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
3007                 /*
3008                  * set the writeback_index so that range_cyclic
3009                  * mode will write it back later
3010                  */
3011                 mapping->writeback_index = index;
3012
3013 out_writepages:
3014         if (!no_nrwrite_index_update)
3015                 wbc->no_nrwrite_index_update = 0;
3016         wbc->nr_to_write -= nr_to_writebump;
3017         wbc->range_start = range_start;
3018         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
3019         return ret;
3020 }
3021
3022 #define FALL_BACK_TO_NONDELALLOC 1
3023 static int ext4_nonda_switch(struct super_block *sb)
3024 {
3025         s64 free_blocks, dirty_blocks;
3026         struct ext4_sb_info *sbi = EXT4_SB(sb);
3027
3028         /*
3029          * switch to non delalloc mode if we are running low
3030          * on free block. The free block accounting via percpu
3031          * counters can get slightly wrong with percpu_counter_batch getting
3032          * accumulated on each CPU without updating global counters
3033          * Delalloc need an accurate free block accounting. So switch
3034          * to non delalloc when we are near to error range.
3035          */
3036         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
3037         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
3038         if (2 * free_blocks < 3 * dirty_blocks ||
3039                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
3040                 /*
3041                  * free block count is less than 150% of dirty blocks
3042                  * or free blocks is less than watermark
3043                  */
3044                 return 1;
3045         }
3046         /*
3047          * Even if we don't switch but are nearing capacity,
3048          * start pushing delalloc when 1/2 of free blocks are dirty.
3049          */
3050         if (free_blocks < 2 * dirty_blocks)
3051                 writeback_inodes_sb_if_idle(sb);
3052
3053         return 0;
3054 }
3055
3056 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3057                                loff_t pos, unsigned len, unsigned flags,
3058                                struct page **pagep, void **fsdata)
3059 {
3060         int ret, retries = 0, quota_retries = 0;
3061         struct page *page;
3062         pgoff_t index;
3063         unsigned from, to;
3064         struct inode *inode = mapping->host;
3065         handle_t *handle;
3066
3067         index = pos >> PAGE_CACHE_SHIFT;
3068         from = pos & (PAGE_CACHE_SIZE - 1);
3069         to = from + len;
3070
3071         if (ext4_nonda_switch(inode->i_sb)) {
3072                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3073                 return ext4_write_begin(file, mapping, pos,
3074                                         len, flags, pagep, fsdata);
3075         }
3076         *fsdata = (void *)0;
3077         trace_ext4_da_write_begin(inode, pos, len, flags);
3078 retry:
3079         /*
3080          * With delayed allocation, we don't log the i_disksize update
3081          * if there is delayed block allocation. But we still need
3082          * to journalling the i_disksize update if writes to the end
3083          * of file which has an already mapped buffer.
3084          */
3085         handle = ext4_journal_start(inode, 1);
3086         if (IS_ERR(handle)) {
3087                 ret = PTR_ERR(handle);
3088                 goto out;
3089         }
3090         /* We cannot recurse into the filesystem as the transaction is already
3091          * started */
3092         flags |= AOP_FLAG_NOFS;
3093
3094         page = grab_cache_page_write_begin(mapping, index, flags);
3095         if (!page) {
3096                 ext4_journal_stop(handle);
3097                 ret = -ENOMEM;
3098                 goto out;
3099         }
3100         *pagep = page;
3101
3102         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
3103                                 ext4_da_get_block_prep);
3104         if (ret < 0) {
3105                 unlock_page(page);
3106                 ext4_journal_stop(handle);
3107                 page_cache_release(page);
3108                 /*
3109                  * block_write_begin may have instantiated a few blocks
3110                  * outside i_size.  Trim these off again. Don't need
3111                  * i_size_read because we hold i_mutex.
3112                  */
3113                 if (pos + len > inode->i_size)
3114                         ext4_truncate_failed_write(inode);
3115         }
3116
3117         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3118                 goto retry;
3119
3120         if ((ret == -EDQUOT) &&
3121             EXT4_I(inode)->i_reserved_meta_blocks &&
3122             (quota_retries++ < 3)) {
3123                 /*
3124                  * Since we often over-estimate the number of meta
3125                  * data blocks required, we may sometimes get a
3126                  * spurios out of quota error even though there would
3127                  * be enough space once we write the data blocks and
3128                  * find out how many meta data blocks were _really_
3129                  * required.  So try forcing the inode write to see if
3130                  * that helps.
3131                  */
3132                 write_inode_now(inode, (quota_retries == 3));
3133                 goto retry;
3134         }
3135 out:
3136         return ret;
3137 }
3138
3139 /*
3140  * Check if we should update i_disksize
3141  * when write to the end of file but not require block allocation
3142  */
3143 static int ext4_da_should_update_i_disksize(struct page *page,
3144                                             unsigned long offset)
3145 {
3146         struct buffer_head *bh;
3147         struct inode *inode = page->mapping->host;
3148         unsigned int idx;
3149         int i;
3150
3151         bh = page_buffers(page);
3152         idx = offset >> inode->i_blkbits;
3153
3154         for (i = 0; i < idx; i++)
3155                 bh = bh->b_this_page;
3156
3157         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3158                 return 0;
3159         return 1;
3160 }
3161
3162 static int ext4_da_write_end(struct file *file,
3163                              struct address_space *mapping,
3164                              loff_t pos, unsigned len, unsigned copied,
3165                              struct page *page, void *fsdata)
3166 {
3167         struct inode *inode = mapping->host;
3168         int ret = 0, ret2;
3169         handle_t *handle = ext4_journal_current_handle();
3170         loff_t new_i_size;
3171         unsigned long start, end;
3172         int write_mode = (int)(unsigned long)fsdata;
3173
3174         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
3175                 if (ext4_should_order_data(inode)) {
3176                         return ext4_ordered_write_end(file, mapping, pos,
3177                                         len, copied, page, fsdata);
3178                 } else if (ext4_should_writeback_data(inode)) {
3179                         return ext4_writeback_write_end(file, mapping, pos,
3180                                         len, copied, page, fsdata);
3181                 } else {
3182                         BUG();
3183                 }
3184         }
3185
3186         trace_ext4_da_write_end(inode, pos, len, copied);
3187         start = pos & (PAGE_CACHE_SIZE - 1);
3188         end = start + copied - 1;
3189
3190         /*
3191          * generic_write_end() will run mark_inode_dirty() if i_size
3192          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3193          * into that.
3194          */
3195
3196         new_i_size = pos + copied;
3197         if (new_i_size > EXT4_I(inode)->i_disksize) {
3198                 if (ext4_da_should_update_i_disksize(page, end)) {
3199                         down_write(&EXT4_I(inode)->i_data_sem);
3200                         if (new_i_size > EXT4_I(inode)->i_disksize) {
3201                                 /*
3202                                  * Updating i_disksize when extending file
3203                                  * without needing block allocation
3204                                  */
3205                                 if (ext4_should_order_data(inode))
3206                                         ret = ext4_jbd2_file_inode(handle,
3207                                                                    inode);
3208
3209                                 EXT4_I(inode)->i_disksize = new_i_size;
3210                         }
3211                         up_write(&EXT4_I(inode)->i_data_sem);
3212                         /* We need to mark inode dirty even if
3213                          * new_i_size is less that inode->i_size
3214                          * bu greater than i_disksize.(hint delalloc)
3215                          */
3216                         ext4_mark_inode_dirty(handle, inode);
3217                 }
3218         }
3219         ret2 = generic_write_end(file, mapping, pos, len, copied,
3220                                                         page, fsdata);
3221         copied = ret2;
3222         if (ret2 < 0)
3223                 ret = ret2;
3224         ret2 = ext4_journal_stop(handle);
3225         if (!ret)
3226                 ret = ret2;
3227
3228         return ret ? ret : copied;
3229 }
3230
3231 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
3232 {
3233         /*
3234          * Drop reserved blocks
3235          */
3236         BUG_ON(!PageLocked(page));
3237         if (!page_has_buffers(page))
3238                 goto out;
3239
3240         ext4_da_page_release_reservation(page, offset);
3241
3242 out:
3243         ext4_invalidatepage(page, offset);
3244
3245         return;
3246 }
3247
3248 /*
3249  * Force all delayed allocation blocks to be allocated for a given inode.
3250  */
3251 int ext4_alloc_da_blocks(struct inode *inode)
3252 {
3253         trace_ext4_alloc_da_blocks(inode);
3254
3255         if (!EXT4_I(inode)->i_reserved_data_blocks &&
3256             !EXT4_I(inode)->i_reserved_meta_blocks)
3257                 return 0;
3258
3259         /*
3260          * We do something simple for now.  The filemap_flush() will
3261          * also start triggering a write of the data blocks, which is
3262          * not strictly speaking necessary (and for users of
3263          * laptop_mode, not even desirable).  However, to do otherwise
3264          * would require replicating code paths in:
3265          *
3266          * ext4_da_writepages() ->
3267          *    write_cache_pages() ---> (via passed in callback function)
3268          *        __mpage_da_writepage() -->
3269          *           mpage_add_bh_to_extent()
3270          *           mpage_da_map_blocks()
3271          *
3272          * The problem is that write_cache_pages(), located in
3273          * mm/page-writeback.c, marks pages clean in preparation for
3274          * doing I/O, which is not desirable if we're not planning on
3275          * doing I/O at all.
3276          *
3277          * We could call write_cache_pages(), and then redirty all of
3278          * the pages by calling redirty_page_for_writeback() but that
3279          * would be ugly in the extreme.  So instead we would need to
3280          * replicate parts of the code in the above functions,
3281          * simplifying them becuase we wouldn't actually intend to
3282          * write out the pages, but rather only collect contiguous
3283          * logical block extents, call the multi-block allocator, and
3284          * then update the buffer heads with the block allocations.
3285          *
3286          * For now, though, we'll cheat by calling filemap_flush(),
3287          * which will map the blocks, and start the I/O, but not
3288          * actually wait for the I/O to complete.
3289          */
3290         return filemap_flush(inode->i_mapping);
3291 }
3292
3293 /*
3294  * bmap() is special.  It gets used by applications such as lilo and by
3295  * the swapper to find the on-disk block of a specific piece of data.
3296  *
3297  * Naturally, this is dangerous if the block concerned is still in the
3298  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3299  * filesystem and enables swap, then they may get a nasty shock when the
3300  * data getting swapped to that swapfile suddenly gets overwritten by
3301  * the original zero's written out previously to the journal and
3302  * awaiting writeback in the kernel's buffer cache.
3303  *
3304  * So, if we see any bmap calls here on a modified, data-journaled file,
3305  * take extra steps to flush any blocks which might be in the cache.
3306  */
3307 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3308 {
3309         struct inode *inode = mapping->host;
3310         journal_t *journal;
3311         int err;
3312
3313         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3314                         test_opt(inode->i_sb, DELALLOC)) {
3315                 /*
3316                  * With delalloc we want to sync the file
3317                  * so that we can make sure we allocate
3318                  * blocks for file
3319                  */
3320                 filemap_write_and_wait(mapping);
3321         }
3322
3323         if (EXT4_JOURNAL(inode) &&
3324             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3325                 /*
3326                  * This is a REALLY heavyweight approach, but the use of
3327                  * bmap on dirty files is expected to be extremely rare:
3328                  * only if we run lilo or swapon on a freshly made file
3329                  * do we expect this to happen.
3330                  *
3331                  * (bmap requires CAP_SYS_RAWIO so this does not
3332                  * represent an unprivileged user DOS attack --- we'd be
3333                  * in trouble if mortal users could trigger this path at
3334                  * will.)
3335                  *
3336                  * NB. EXT4_STATE_JDATA is not set on files other than
3337                  * regular files.  If somebody wants to bmap a directory
3338                  * or symlink and gets confused because the buffer
3339                  * hasn't yet been flushed to disk, they deserve
3340                  * everything they get.
3341                  */
3342
3343                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3344                 journal = EXT4_JOURNAL(inode);
3345                 jbd2_journal_lock_updates(journal);
3346                 err = jbd2_journal_flush(journal);
3347                 jbd2_journal_unlock_updates(journal);
3348
3349                 if (err)
3350                         return 0;
3351         }
3352
3353         return generic_block_bmap(mapping, block, ext4_get_block);
3354 }
3355
3356 static int ext4_readpage(struct file *file, struct page *page)
3357 {
3358         return mpage_readpage(page, ext4_get_block);
3359 }
3360
3361 static int
3362 ext4_readpages(struct file *file, struct address_space *mapping,
3363                 struct list_head *pages, unsigned nr_pages)
3364 {
3365         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3366 }
3367
3368 static void ext4_free_io_end(ext4_io_end_t *io)
3369 {
3370         BUG_ON(!io);
3371         if (io->page)
3372                 put_page(io->page);
3373         iput(io->inode);
3374         kfree(io);
3375 }
3376
3377 static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
3378 {
3379         struct buffer_head *head, *bh;
3380         unsigned int curr_off = 0;
3381
3382         if (!page_has_buffers(page))
3383                 return;
3384         head = bh = page_buffers(page);
3385         do {
3386                 if (offset <= curr_off && test_clear_buffer_uninit(bh)
3387                                         && bh->b_private) {
3388                         ext4_free_io_end(bh->b_private);
3389                         bh->b_private = NULL;
3390                         bh->b_end_io = NULL;
3391                 }
3392                 curr_off = curr_off + bh->b_size;
3393                 bh = bh->b_this_page;
3394         } while (bh != head);
3395 }
3396
3397 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3398 {
3399         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3400
3401         /*
3402          * free any io_end structure allocated for buffers to be discarded
3403          */
3404         if (ext4_should_dioread_nolock(page->mapping->host))
3405                 ext4_invalidatepage_free_endio(page, offset);
3406         /*
3407          * If it's a full truncate we just forget about the pending dirtying
3408          */
3409         if (offset == 0)
3410                 ClearPageChecked(page);
3411
3412         if (journal)
3413                 jbd2_journal_invalidatepage(journal, page, offset);
3414         else
3415                 block_invalidatepage(page, offset);
3416 }
3417
3418 static int ext4_releasepage(struct page *page, gfp_t wait)
3419 {
3420         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3421
3422         WARN_ON(PageChecked(page));
3423         if (!page_has_buffers(page))
3424                 return 0;
3425         if (journal)
3426                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3427         else
3428                 return try_to_free_buffers(page);
3429 }
3430
3431 /*
3432  * O_DIRECT for ext3 (or indirect map) based files
3433  *
3434  * If the O_DIRECT write will extend the file then add this inode to the
3435  * orphan list.  So recovery will truncate it back to the original size
3436  * if the machine crashes during the write.
3437  *
3438  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3439  * crashes then stale disk data _may_ be exposed inside the file. But current
3440  * VFS code falls back into buffered path in that case so we are safe.
3441  */
3442 static ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
3443                               const struct iovec *iov, loff_t offset,
3444                               unsigned long nr_segs)
3445 {
3446         struct file *file = iocb->ki_filp;
3447         struct inode *inode = file->f_mapping->host;
3448         struct ext4_inode_info *ei = EXT4_I(inode);
3449         handle_t *handle;
3450         ssize_t ret;
3451         int orphan = 0;
3452         size_t count = iov_length(iov, nr_segs);
3453         int retries = 0;
3454
3455         if (rw == WRITE) {
3456                 loff_t final_size = offset + count;
3457
3458                 if (final_size > inode->i_size) {
3459                         /* Credits for sb + inode write */
3460                         handle = ext4_journal_start(inode, 2);
3461                         if (IS_ERR(handle)) {
3462                                 ret = PTR_ERR(handle);
3463                                 goto out;
3464                         }
3465                         ret = ext4_orphan_add(handle, inode);
3466                         if (ret) {
3467                                 ext4_journal_stop(handle);
3468                                 goto out;
3469                         }
3470                         orphan = 1;
3471                         ei->i_disksize = inode->i_size;
3472                         ext4_journal_stop(handle);
3473                 }
3474         }
3475
3476 retry:
3477         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
3478                                  offset, nr_segs,
3479                                  ext4_get_block, NULL);
3480         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3481                 goto retry;
3482
3483         if (orphan) {
3484                 int err;
3485
3486                 /* Credits for sb + inode write */
3487                 handle = ext4_journal_start(inode, 2);
3488                 if (IS_ERR(handle)) {
3489                         /* This is really bad luck. We've written the data
3490                          * but cannot extend i_size. Bail out and pretend
3491                          * the write failed... */
3492                         ret = PTR_ERR(handle);
3493                         if (inode->i_nlink)
3494                                 ext4_orphan_del(NULL, inode);
3495
3496                         goto out;
3497                 }
3498                 if (inode->i_nlink)
3499                         ext4_orphan_del(handle, inode);
3500                 if (ret > 0) {
3501                         loff_t end = offset + ret;
3502                         if (end > inode->i_size) {
3503                                 ei->i_disksize = end;
3504                                 i_size_write(inode, end);
3505                                 /*
3506                                  * We're going to return a positive `ret'
3507                                  * here due to non-zero-length I/O, so there's
3508                                  * no way of reporting error returns from
3509                                  * ext4_mark_inode_dirty() to userspace.  So
3510                                  * ignore it.
3511                                  */
3512                                 ext4_mark_inode_dirty(handle, inode);
3513                         }
3514                 }
3515                 err = ext4_journal_stop(handle);
3516                 if (ret == 0)
3517                         ret = err;
3518         }
3519 out:
3520         return ret;
3521 }
3522
3523 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
3524                    struct buffer_head *bh_result, int create)
3525 {
3526         handle_t *handle = ext4_journal_current_handle();
3527         int ret = 0;
3528         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
3529         int dio_credits;
3530         int started = 0;
3531
3532         ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
3533                    inode->i_ino, create);
3534         /*
3535          * ext4_get_block in prepare for a DIO write or buffer write.
3536          * We allocate an uinitialized extent if blocks haven't been allocated.
3537          * The extent will be converted to initialized after IO complete.
3538          */
3539         create = EXT4_GET_BLOCKS_IO_CREATE_EXT;
3540
3541         if (!handle) {
3542                 if (max_blocks > DIO_MAX_BLOCKS)
3543                         max_blocks = DIO_MAX_BLOCKS;
3544                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
3545                 handle = ext4_journal_start(inode, dio_credits);
3546                 if (IS_ERR(handle)) {
3547                         ret = PTR_ERR(handle);
3548                         goto out;
3549                 }
3550                 started = 1;
3551         }
3552
3553         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
3554                               create);
3555         if (ret > 0) {
3556                 bh_result->b_size = (ret << inode->i_blkbits);
3557                 ret = 0;
3558         }
3559         if (started)
3560                 ext4_journal_stop(handle);
3561 out:
3562         return ret;
3563 }
3564
3565 static void dump_completed_IO(struct inode * inode)
3566 {
3567 #ifdef  EXT4_DEBUG
3568         struct list_head *cur, *before, *after;
3569         ext4_io_end_t *io, *io0, *io1;
3570         unsigned long flags;
3571
3572         if (list_empty(&EXT4_I(inode)->i_completed_io_list)){
3573                 ext4_debug("inode %lu completed_io list is empty\n", inode->i_ino);
3574                 return;
3575         }
3576
3577         ext4_debug("Dump inode %lu completed_io list \n", inode->i_ino);
3578         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3579         list_for_each_entry(io, &EXT4_I(inode)->i_completed_io_list, list){
3580                 cur = &io->list;
3581                 before = cur->prev;
3582                 io0 = container_of(before, ext4_io_end_t, list);
3583                 after = cur->next;
3584                 io1 = container_of(after, ext4_io_end_t, list);
3585
3586                 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
3587                             io, inode->i_ino, io0, io1);
3588         }
3589         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3590 #endif
3591 }
3592
3593 /*
3594  * check a range of space and convert unwritten extents to written.
3595  */
3596 static int ext4_end_io_nolock(ext4_io_end_t *io)
3597 {
3598         struct inode *inode = io->inode;
3599         loff_t offset = io->offset;
3600         ssize_t size = io->size;
3601         int ret = 0;
3602
3603         ext4_debug("ext4_end_io_nolock: io 0x%p from inode %lu,list->next 0x%p,"
3604                    "list->prev 0x%p\n",
3605                    io, inode->i_ino, io->list.next, io->list.prev);
3606
3607         if (list_empty(&io->list))
3608                 return ret;
3609
3610         if (io->flag != EXT4_IO_UNWRITTEN)
3611                 return ret;
3612
3613         ret = ext4_convert_unwritten_extents(inode, offset, size);
3614         if (ret < 0) {
3615                 printk(KERN_EMERG "%s: failed to convert unwritten"
3616                         "extents to written extents, error is %d"
3617                         " io is still on inode %lu aio dio list\n",
3618                        __func__, ret, inode->i_ino);
3619                 return ret;
3620         }
3621
3622         /* clear the DIO AIO unwritten flag */
3623         io->flag = 0;
3624         return ret;
3625 }
3626
3627 /*
3628  * work on completed aio dio IO, to convert unwritten extents to extents
3629  */
3630 static void ext4_end_io_work(struct work_struct *work)
3631 {
3632         ext4_io_end_t           *io = container_of(work, ext4_io_end_t, work);
3633         struct inode            *inode = io->inode;
3634         struct ext4_inode_info  *ei = EXT4_I(inode);
3635         unsigned long           flags;
3636         int                     ret;
3637
3638         mutex_lock(&inode->i_mutex);
3639         ret = ext4_end_io_nolock(io);
3640         if (ret < 0) {
3641                 mutex_unlock(&inode->i_mutex);
3642                 return;
3643         }
3644
3645         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3646         if (!list_empty(&io->list))
3647                 list_del_init(&io->list);
3648         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3649         mutex_unlock(&inode->i_mutex);
3650         ext4_free_io_end(io);
3651 }
3652
3653 /*
3654  * This function is called from ext4_sync_file().
3655  *
3656  * When IO is completed, the work to convert unwritten extents to
3657  * written is queued on workqueue but may not get immediately
3658  * scheduled. When fsync is called, we need to ensure the
3659  * conversion is complete before fsync returns.
3660  * The inode keeps track of a list of pending/completed IO that
3661  * might needs to do the conversion. This function walks through
3662  * the list and convert the related unwritten extents for completed IO
3663  * to written.
3664  * The function return the number of pending IOs on success.
3665  */
3666 int flush_completed_IO(struct inode *inode)
3667 {
3668         ext4_io_end_t *io;
3669         struct ext4_inode_info *ei = EXT4_I(inode);
3670         unsigned long flags;
3671         int ret = 0;
3672         int ret2 = 0;
3673
3674         if (list_empty(&ei->i_completed_io_list))
3675                 return ret;
3676
3677         dump_completed_IO(inode);
3678         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3679         while (!list_empty(&ei->i_completed_io_list)){
3680                 io = list_entry(ei->i_completed_io_list.next,
3681                                 ext4_io_end_t, list);
3682                 /*
3683                  * Calling ext4_end_io_nolock() to convert completed
3684                  * IO to written.
3685                  *
3686                  * When ext4_sync_file() is called, run_queue() may already
3687                  * about to flush the work corresponding to this io structure.
3688                  * It will be upset if it founds the io structure related
3689                  * to the work-to-be schedule is freed.
3690                  *
3691                  * Thus we need to keep the io structure still valid here after
3692                  * convertion finished. The io structure has a flag to
3693                  * avoid double converting from both fsync and background work
3694                  * queue work.
3695                  */
3696                 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3697                 ret = ext4_end_io_nolock(io);
3698                 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3699                 if (ret < 0)
3700                         ret2 = ret;
3701                 else
3702                         list_del_init(&io->list);
3703         }
3704         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3705         return (ret2 < 0) ? ret2 : 0;
3706 }
3707
3708 static ext4_io_end_t *ext4_init_io_end (struct inode *inode, gfp_t flags)
3709 {
3710         ext4_io_end_t *io = NULL;
3711
3712         io = kmalloc(sizeof(*io), flags);
3713
3714         if (io) {
3715                 igrab(inode);
3716                 io->inode = inode;
3717                 io->flag = 0;
3718                 io->offset = 0;
3719                 io->size = 0;
3720                 io->page = NULL;
3721                 INIT_WORK(&io->work, ext4_end_io_work);
3722                 INIT_LIST_HEAD(&io->list);
3723         }
3724
3725         return io;
3726 }
3727
3728 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3729                             ssize_t size, void *private)
3730 {
3731         ext4_io_end_t *io_end = iocb->private;
3732         struct workqueue_struct *wq;
3733         unsigned long flags;
3734         struct ext4_inode_info *ei;
3735
3736         /* if not async direct IO or dio with 0 bytes write, just return */
3737         if (!io_end || !size)
3738                 return;
3739
3740         ext_debug("ext4_end_io_dio(): io_end 0x%p"
3741                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
3742                   iocb->private, io_end->inode->i_ino, iocb, offset,
3743                   size);
3744
3745         /* if not aio dio with unwritten extents, just free io and return */
3746         if (io_end->flag != EXT4_IO_UNWRITTEN){
3747                 ext4_free_io_end(io_end);
3748                 iocb->private = NULL;
3749                 return;
3750         }
3751
3752         io_end->offset = offset;
3753         io_end->size = size;
3754         io_end->flag = EXT4_IO_UNWRITTEN;
3755         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
3756
3757         /* queue the work to convert unwritten extents to written */
3758         queue_work(wq, &io_end->work);
3759
3760         /* Add the io_end to per-inode completed aio dio list*/
3761         ei = EXT4_I(io_end->inode);
3762         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3763         list_add_tail(&io_end->list, &ei->i_completed_io_list);
3764         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3765         iocb->private = NULL;
3766 }
3767
3768 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
3769 {
3770         ext4_io_end_t *io_end = bh->b_private;
3771         struct workqueue_struct *wq;
3772         struct inode *inode;
3773         unsigned long flags;
3774
3775         if (!test_clear_buffer_uninit(bh) || !io_end)
3776                 goto out;
3777
3778         if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
3779                 printk("sb umounted, discard end_io request for inode %lu\n",
3780                         io_end->inode->i_ino);
3781                 ext4_free_io_end(io_end);
3782                 goto out;
3783         }
3784
3785         io_end->flag = EXT4_IO_UNWRITTEN;
3786         inode = io_end->inode;
3787
3788         /* Add the io_end to per-inode completed io list*/
3789         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3790         list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
3791         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3792
3793         wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
3794         /* queue the work to convert unwritten extents to written */
3795         queue_work(wq, &io_end->work);
3796 out:
3797         bh->b_private = NULL;
3798         bh->b_end_io = NULL;
3799         clear_buffer_uninit(bh);
3800         end_buffer_async_write(bh, uptodate);
3801 }
3802
3803 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
3804 {
3805         ext4_io_end_t *io_end;
3806         struct page *page = bh->b_page;
3807         loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
3808         size_t size = bh->b_size;
3809
3810 retry:
3811         io_end = ext4_init_io_end(inode, GFP_ATOMIC);
3812         if (!io_end) {
3813                 if (printk_ratelimit())
3814                         printk(KERN_WARNING "%s: allocation fail\n", __func__);
3815                 schedule();
3816                 goto retry;
3817         }
3818         io_end->offset = offset;
3819         io_end->size = size;
3820         /*
3821          * We need to hold a reference to the page to make sure it
3822          * doesn't get evicted before ext4_end_io_work() has a chance
3823          * to convert the extent from written to unwritten.
3824          */
3825         io_end->page = page;
3826         get_page(io_end->page);
3827
3828         bh->b_private = io_end;
3829         bh->b_end_io = ext4_end_io_buffer_write;
3830         return 0;
3831 }
3832
3833 /*
3834  * For ext4 extent files, ext4 will do direct-io write to holes,
3835  * preallocated extents, and those write extend the file, no need to
3836  * fall back to buffered IO.
3837  *
3838  * For holes, we fallocate those blocks, mark them as unintialized
3839  * If those blocks were preallocated, we mark sure they are splited, but
3840  * still keep the range to write as unintialized.
3841  *
3842  * The unwrritten extents will be converted to written when DIO is completed.
3843  * For async direct IO, since the IO may still pending when return, we
3844  * set up an end_io call back function, which will do the convertion
3845  * when async direct IO completed.
3846  *
3847  * If the O_DIRECT write will extend the file then add this inode to the
3848  * orphan list.  So recovery will truncate it back to the original size
3849  * if the machine crashes during the write.
3850  *
3851  */
3852 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
3853                               const struct iovec *iov, loff_t offset,
3854                               unsigned long nr_segs)
3855 {
3856         struct file *file = iocb->ki_filp;
3857         struct inode *inode = file->f_mapping->host;
3858         ssize_t ret;
3859         size_t count = iov_length(iov, nr_segs);
3860
3861         loff_t final_size = offset + count;
3862         if (rw == WRITE && final_size <= inode->i_size) {
3863                 /*
3864                  * We could direct write to holes and fallocate.
3865                  *
3866                  * Allocated blocks to fill the hole are marked as uninitialized
3867                  * to prevent paralel buffered read to expose the stale data
3868                  * before DIO complete the data IO.
3869                  *
3870                  * As to previously fallocated extents, ext4 get_block
3871                  * will just simply mark the buffer mapped but still
3872                  * keep the extents uninitialized.
3873                  *
3874                  * for non AIO case, we will convert those unwritten extents
3875                  * to written after return back from blockdev_direct_IO.
3876                  *
3877                  * for async DIO, the conversion needs to be defered when
3878                  * the IO is completed. The ext4 end_io callback function
3879                  * will be called to take care of the conversion work.
3880                  * Here for async case, we allocate an io_end structure to
3881                  * hook to the iocb.
3882                  */
3883                 iocb->private = NULL;
3884                 EXT4_I(inode)->cur_aio_dio = NULL;
3885                 if (!is_sync_kiocb(iocb)) {
3886                         iocb->private = ext4_init_io_end(inode, GFP_NOFS);
3887                         if (!iocb->private)
3888                                 return -ENOMEM;
3889                         /*
3890                          * we save the io structure for current async
3891                          * direct IO, so that later ext4_get_blocks()
3892                          * could flag the io structure whether there
3893                          * is a unwritten extents needs to be converted
3894                          * when IO is completed.
3895                          */
3896                         EXT4_I(inode)->cur_aio_dio = iocb->private;
3897                 }
3898
3899                 ret = blockdev_direct_IO(rw, iocb, inode,
3900                                          inode->i_sb->s_bdev, iov,
3901                                          offset, nr_segs,
3902                                          ext4_get_block_write,
3903                                          ext4_end_io_dio);
3904                 if (iocb->private)
3905                         EXT4_I(inode)->cur_aio_dio = NULL;
3906                 /*
3907                  * The io_end structure takes a reference to the inode,
3908                  * that structure needs to be destroyed and the
3909                  * reference to the inode need to be dropped, when IO is
3910                  * complete, even with 0 byte write, or failed.
3911                  *
3912                  * In the successful AIO DIO case, the io_end structure will be
3913                  * desctroyed and the reference to the inode will be dropped
3914                  * after the end_io call back function is called.
3915                  *
3916                  * In the case there is 0 byte write, or error case, since
3917                  * VFS direct IO won't invoke the end_io call back function,
3918                  * we need to free the end_io structure here.
3919                  */
3920                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
3921                         ext4_free_io_end(iocb->private);
3922                         iocb->private = NULL;
3923                 } else if (ret > 0 && ext4_test_inode_state(inode,
3924                                                 EXT4_STATE_DIO_UNWRITTEN)) {
3925                         int err;
3926                         /*
3927                          * for non AIO case, since the IO is already
3928                          * completed, we could do the convertion right here
3929                          */
3930                         err = ext4_convert_unwritten_extents(inode,
3931                                                              offset, ret);
3932                         if (err < 0)
3933                                 ret = err;
3934                         ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3935                 }
3936                 return ret;
3937         }
3938
3939         /* for write the the end of file case, we fall back to old way */
3940         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3941 }
3942
3943 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3944                               const struct iovec *iov, loff_t offset,
3945                               unsigned long nr_segs)
3946 {
3947         struct file *file = iocb->ki_filp;
3948         struct inode *inode = file->f_mapping->host;
3949
3950         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
3951                 return ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
3952
3953         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3954 }
3955
3956 /*
3957  * Pages can be marked dirty completely asynchronously from ext4's journalling
3958  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3959  * much here because ->set_page_dirty is called under VFS locks.  The page is
3960  * not necessarily locked.
3961  *
3962  * We cannot just dirty the page and leave attached buffers clean, because the
3963  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3964  * or jbddirty because all the journalling code will explode.
3965  *
3966  * So what we do is to mark the page "pending dirty" and next time writepage
3967  * is called, propagate that into the buffers appropriately.
3968  */
3969 static int ext4_journalled_set_page_dirty(struct page *page)
3970 {
3971         SetPageChecked(page);
3972         return __set_page_dirty_nobuffers(page);
3973 }
3974
3975 static const struct address_space_operations ext4_ordered_aops = {
3976         .readpage               = ext4_readpage,
3977         .readpages              = ext4_readpages,
3978         .writepage              = ext4_writepage,
3979         .sync_page              = block_sync_page,
3980         .write_begin            = ext4_write_begin,
3981         .write_end              = ext4_ordered_write_end,
3982         .bmap                   = ext4_bmap,
3983         .invalidatepage         = ext4_invalidatepage,
3984         .releasepage            = ext4_releasepage,
3985         .direct_IO              = ext4_direct_IO,
3986         .migratepage            = buffer_migrate_page,
3987         .is_partially_uptodate  = block_is_partially_uptodate,
3988         .error_remove_page      = generic_error_remove_page,
3989 };
3990
3991 static const struct address_space_operations ext4_writeback_aops = {
3992         .readpage               = ext4_readpage,
3993         .readpages              = ext4_readpages,
3994         .writepage              = ext4_writepage,
3995         .sync_page              = block_sync_page,
3996         .write_begin            = ext4_write_begin,
3997         .write_end              = ext4_writeback_write_end,
3998         .bmap                   = ext4_bmap,
3999         .invalidatepage         = ext4_invalidatepage,
4000         .releasepage            = ext4_releasepage,
4001         .direct_IO              = ext4_direct_IO,
4002         .migratepage            = buffer_migrate_page,
4003         .is_partially_uptodate  = block_is_partially_uptodate,
4004         .error_remove_page      = generic_error_remove_page,
4005 };
4006
4007 static const struct address_space_operations ext4_journalled_aops = {
4008         .readpage               = ext4_readpage,
4009         .readpages              = ext4_readpages,
4010         .writepage              = ext4_writepage,
4011         .sync_page              = block_sync_page,
4012         .write_begin            = ext4_write_begin,
4013         .write_end              = ext4_journalled_write_end,
4014         .set_page_dirty         = ext4_journalled_set_page_dirty,
4015         .bmap                   = ext4_bmap,
4016         .invalidatepage         = ext4_invalidatepage,
4017         .releasepage            = ext4_releasepage,
4018         .is_partially_uptodate  = block_is_partially_uptodate,
4019         .error_remove_page      = generic_error_remove_page,
4020 };
4021
4022 static const struct address_space_operations ext4_da_aops = {
4023         .readpage               = ext4_readpage,
4024         .readpages              = ext4_readpages,
4025         .writepage              = ext4_writepage,
4026         .writepages             = ext4_da_writepages,
4027         .sync_page              = block_sync_page,
4028         .write_begin            = ext4_da_write_begin,
4029         .write_end              = ext4_da_write_end,
4030         .bmap                   = ext4_bmap,
4031         .invalidatepage         = ext4_da_invalidatepage,
4032         .releasepage            = ext4_releasepage,
4033         .direct_IO              = ext4_direct_IO,
4034         .migratepage            = buffer_migrate_page,
4035         .is_partially_uptodate  = block_is_partially_uptodate,
4036         .error_remove_page      = generic_error_remove_page,
4037 };
4038
4039 void ext4_set_aops(struct inode *inode)
4040 {
4041         if (ext4_should_order_data(inode) &&
4042                 test_opt(inode->i_sb, DELALLOC))
4043                 inode->i_mapping->a_ops = &ext4_da_aops;
4044         else if (ext4_should_order_data(inode))
4045                 inode->i_mapping->a_ops = &ext4_ordered_aops;
4046         else if (ext4_should_writeback_data(inode) &&
4047                  test_opt(inode->i_sb, DELALLOC))
4048                 inode->i_mapping->a_ops = &ext4_da_aops;
4049         else if (ext4_should_writeback_data(inode))
4050                 inode->i_mapping->a_ops = &ext4_writeback_aops;
4051         else
4052                 inode->i_mapping->a_ops = &ext4_journalled_aops;
4053 }
4054
4055 /*
4056  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
4057  * up to the end of the block which corresponds to `from'.
4058  * This required during truncate. We need to physically zero the tail end
4059  * of that block so it doesn't yield old data if the file is later grown.
4060  */
4061 int ext4_block_truncate_page(handle_t *handle,
4062                 struct address_space *mapping, loff_t from)
4063 {
4064         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
4065         unsigned offset = from & (PAGE_CACHE_SIZE-1);
4066         unsigned blocksize, length, pos;
4067         ext4_lblk_t iblock;
4068         struct inode *inode = mapping->host;
4069         struct buffer_head *bh;
4070         struct page *page;
4071         int err = 0;
4072
4073         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
4074                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
4075         if (!page)
4076                 return -EINVAL;
4077
4078         blocksize = inode->i_sb->s_blocksize;
4079         length = blocksize - (offset & (blocksize - 1));
4080         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
4081
4082         /*
4083          * For "nobh" option,  we can only work if we don't need to
4084          * read-in the page - otherwise we create buffers to do the IO.
4085          */
4086         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
4087              ext4_should_writeback_data(inode) && PageUptodate(page)) {
4088                 zero_user(page, offset, length);
4089                 set_page_dirty(page);
4090                 goto unlock;
4091         }
4092
4093         if (!page_has_buffers(page))
4094                 create_empty_buffers(page, blocksize, 0);
4095
4096         /* Find the buffer that contains "offset" */
4097         bh = page_buffers(page);
4098         pos = blocksize;
4099         while (offset >= pos) {
4100                 bh = bh->b_this_page;
4101                 iblock++;
4102                 pos += blocksize;
4103         }
4104
4105         err = 0;
4106         if (buffer_freed(bh)) {
4107                 BUFFER_TRACE(bh, "freed: skip");
4108                 goto unlock;
4109         }
4110
4111         if (!buffer_mapped(bh)) {
4112                 BUFFER_TRACE(bh, "unmapped");
4113                 ext4_get_block(inode, iblock, bh, 0);
4114                 /* unmapped? It's a hole - nothing to do */
4115                 if (!buffer_mapped(bh)) {
4116                         BUFFER_TRACE(bh, "still unmapped");
4117                         goto unlock;
4118                 }
4119         }
4120
4121         /* Ok, it's mapped. Make sure it's up-to-date */
4122         if (PageUptodate(page))
4123                 set_buffer_uptodate(bh);
4124
4125         if (!buffer_uptodate(bh)) {
4126                 err = -EIO;
4127                 ll_rw_block(READ, 1, &bh);
4128                 wait_on_buffer(bh);
4129                 /* Uhhuh. Read error. Complain and punt. */
4130                 if (!buffer_uptodate(bh))
4131                         goto unlock;
4132         }
4133
4134         if (ext4_should_journal_data(inode)) {
4135                 BUFFER_TRACE(bh, "get write access");
4136                 err = ext4_journal_get_write_access(handle, bh);
4137                 if (err)
4138                         goto unlock;
4139         }
4140
4141         zero_user(page, offset, length);
4142
4143         BUFFER_TRACE(bh, "zeroed end of block");
4144
4145         err = 0;
4146         if (ext4_should_journal_data(inode)) {
4147                 err = ext4_handle_dirty_metadata(handle, inode, bh);
4148         } else {
4149                 if (ext4_should_order_data(inode))
4150                         err = ext4_jbd2_file_inode(handle, inode);
4151                 mark_buffer_dirty(bh);
4152         }
4153
4154 unlock:
4155         unlock_page(page);
4156         page_cache_release(page);
4157         return err;
4158 }
4159
4160 /*
4161  * Probably it should be a library function... search for first non-zero word
4162  * or memcmp with zero_page, whatever is better for particular architecture.
4163  * Linus?
4164  */
4165 static inline int all_zeroes(__le32 *p, __le32 *q)
4166 {
4167         while (p < q)
4168                 if (*p++)
4169                         return 0;
4170         return 1;
4171 }
4172
4173 /**
4174  *      ext4_find_shared - find the indirect blocks for partial truncation.
4175  *      @inode:   inode in question
4176  *      @depth:   depth of the affected branch
4177  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
4178  *      @chain:   place to store the pointers to partial indirect blocks
4179  *      @top:     place to the (detached) top of branch
4180  *
4181  *      This is a helper function used by ext4_truncate().
4182  *
4183  *      When we do truncate() we may have to clean the ends of several
4184  *      indirect blocks but leave the blocks themselves alive. Block is
4185  *      partially truncated if some data below the new i_size is refered
4186  *      from it (and it is on the path to the first completely truncated
4187  *      data block, indeed).  We have to free the top of that path along
4188  *      with everything to the right of the path. Since no allocation
4189  *      past the truncation point is possible until ext4_truncate()
4190  *      finishes, we may safely do the latter, but top of branch may
4191  *      require special attention - pageout below the truncation point
4192  *      might try to populate it.
4193  *
4194  *      We atomically detach the top of branch from the tree, store the
4195  *      block number of its root in *@top, pointers to buffer_heads of
4196  *      partially truncated blocks - in @chain[].bh and pointers to
4197  *      their last elements that should not be removed - in
4198  *      @chain[].p. Return value is the pointer to last filled element
4199  *      of @chain.
4200  *
4201  *      The work left to caller to do the actual freeing of subtrees:
4202  *              a) free the subtree starting from *@top
4203  *              b) free the subtrees whose roots are stored in
4204  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
4205  *              c) free the subtrees growing from the inode past the @chain[0].
4206  *                      (no partially truncated stuff there).  */
4207
4208 static Indirect *ext4_find_shared(struct inode *inode, int depth,
4209                                   ext4_lblk_t offsets[4], Indirect chain[4],
4210                                   __le32 *top)
4211 {
4212         Indirect *partial, *p;
4213         int k, err;
4214
4215         *top = 0;
4216         /* Make k index the deepest non-null offset + 1 */
4217         for (k = depth; k > 1 && !offsets[k-1]; k--)
4218                 ;
4219         partial = ext4_get_branch(inode, k, offsets, chain, &err);
4220         /* Writer: pointers */
4221         if (!partial)
4222                 partial = chain + k-1;
4223         /*
4224          * If the branch acquired continuation since we've looked at it -
4225          * fine, it should all survive and (new) top doesn't belong to us.
4226          */
4227         if (!partial->key && *partial->p)
4228                 /* Writer: end */
4229                 goto no_top;
4230         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
4231                 ;
4232         /*
4233          * OK, we've found the last block that must survive. The rest of our
4234          * branch should be detached before unlocking. However, if that rest
4235          * of branch is all ours and does not grow immediately from the inode
4236          * it's easier to cheat and just decrement partial->p.
4237          */
4238         if (p == chain + k - 1 && p > chain) {
4239                 p->p--;
4240         } else {
4241                 *top = *p->p;
4242                 /* Nope, don't do this in ext4.  Must leave the tree intact */
4243 #if 0
4244                 *p->p = 0;
4245 #endif
4246         }
4247         /* Writer: end */
4248
4249         while (partial > p) {
4250                 brelse(partial->bh);
4251                 partial--;
4252         }
4253 no_top:
4254         return partial;
4255 }
4256
4257 /*
4258  * Zero a number of block pointers in either an inode or an indirect block.
4259  * If we restart the transaction we must again get write access to the
4260  * indirect block for further modification.
4261  *
4262  * We release `count' blocks on disk, but (last - first) may be greater
4263  * than `count' because there can be holes in there.
4264  */
4265 static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
4266                              struct buffer_head *bh,
4267                              ext4_fsblk_t block_to_free,
4268                              unsigned long count, __le32 *first,
4269                              __le32 *last)
4270 {
4271         __le32 *p;
4272         int     flags = EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_VALIDATED;
4273
4274         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
4275                 flags |= EXT4_FREE_BLOCKS_METADATA;
4276
4277         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
4278                                    count)) {
4279                 ext4_error(inode->i_sb, "inode #%lu: "
4280                            "attempt to clear blocks %llu len %lu, invalid",
4281                            inode->i_ino, (unsigned long long) block_to_free,
4282                            count);
4283                 return 1;
4284         }
4285
4286         if (try_to_extend_transaction(handle, inode)) {
4287                 if (bh) {
4288                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4289                         ext4_handle_dirty_metadata(handle, inode, bh);
4290                 }
4291                 ext4_mark_inode_dirty(handle, inode);
4292                 ext4_truncate_restart_trans(handle, inode,
4293                                             blocks_for_truncate(inode));
4294                 if (bh) {
4295                         BUFFER_TRACE(bh, "retaking write access");
4296                         ext4_journal_get_write_access(handle, bh);
4297                 }
4298         }
4299
4300         for (p = first; p < last; p++)
4301                 *p = 0;
4302
4303         ext4_free_blocks(handle, inode, 0, block_to_free, count, flags);
4304         return 0;
4305 }
4306
4307 /**
4308  * ext4_free_data - free a list of data blocks
4309  * @handle:     handle for this transaction
4310  * @inode:      inode we are dealing with
4311  * @this_bh:    indirect buffer_head which contains *@first and *@last
4312  * @first:      array of block numbers
4313  * @last:       points immediately past the end of array
4314  *
4315  * We are freeing all blocks refered from that array (numbers are stored as
4316  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
4317  *
4318  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
4319  * blocks are contiguous then releasing them at one time will only affect one
4320  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
4321  * actually use a lot of journal space.
4322  *
4323  * @this_bh will be %NULL if @first and @last point into the inode's direct
4324  * block pointers.
4325  */
4326 static void ext4_free_data(handle_t *handle, struct inode *inode,
4327                            struct buffer_head *this_bh,
4328                            __le32 *first, __le32 *last)
4329 {
4330         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
4331         unsigned long count = 0;            /* Number of blocks in the run */
4332         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
4333                                                corresponding to
4334                                                block_to_free */
4335         ext4_fsblk_t nr;                    /* Current block # */
4336         __le32 *p;                          /* Pointer into inode/ind
4337                                                for current block */
4338         int err;
4339
4340         if (this_bh) {                          /* For indirect block */
4341                 BUFFER_TRACE(this_bh, "get_write_access");
4342                 err = ext4_journal_get_write_access(handle, this_bh);
4343                 /* Important: if we can't update the indirect pointers
4344                  * to the blocks, we can't free them. */
4345                 if (err)
4346                         return;
4347         }
4348
4349         for (p = first; p < last; p++) {
4350                 nr = le32_to_cpu(*p);
4351                 if (nr) {
4352                         /* accumulate blocks to free if they're contiguous */
4353                         if (count == 0) {
4354                                 block_to_free = nr;
4355                                 block_to_free_p = p;
4356                                 count = 1;
4357                         } else if (nr == block_to_free + count) {
4358                                 count++;
4359                         } else {
4360                                 if (ext4_clear_blocks(handle, inode, this_bh,
4361                                                       block_to_free, count,
4362                                                       block_to_free_p, p))
4363                                         break;
4364                                 block_to_free = nr;
4365                                 block_to_free_p = p;
4366                                 count = 1;
4367                         }
4368                 }
4369         }
4370
4371         if (count > 0)
4372                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
4373                                   count, block_to_free_p, p);
4374
4375         if (this_bh) {
4376                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
4377
4378                 /*
4379                  * The buffer head should have an attached journal head at this
4380                  * point. However, if the data is corrupted and an indirect
4381                  * block pointed to itself, it would have been detached when
4382                  * the block was cleared. Check for this instead of OOPSing.
4383                  */
4384                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
4385                         ext4_handle_dirty_metadata(handle, inode, this_bh);
4386                 else
4387                         ext4_error(inode->i_sb,
4388                                    "circular indirect block detected, "
4389                                    "inode=%lu, block=%llu",
4390                                    inode->i_ino,
4391                                    (unsigned long long) this_bh->b_blocknr);
4392         }
4393 }
4394
4395 /**
4396  *      ext4_free_branches - free an array of branches
4397  *      @handle: JBD handle for this transaction
4398  *      @inode: inode we are dealing with
4399  *      @parent_bh: the buffer_head which contains *@first and *@last
4400  *      @first: array of block numbers
4401  *      @last:  pointer immediately past the end of array
4402  *      @depth: depth of the branches to free
4403  *
4404  *      We are freeing all blocks refered from these branches (numbers are
4405  *      stored as little-endian 32-bit) and updating @inode->i_blocks
4406  *      appropriately.
4407  */
4408 static void ext4_free_branches(handle_t *handle, struct inode *inode,
4409                                struct buffer_head *parent_bh,
4410                                __le32 *first, __le32 *last, int depth)
4411 {
4412         ext4_fsblk_t nr;
4413         __le32 *p;
4414
4415         if (ext4_handle_is_aborted(handle))
4416                 return;
4417
4418         if (depth--) {
4419                 struct buffer_head *bh;
4420                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4421                 p = last;
4422                 while (--p >= first) {
4423                         nr = le32_to_cpu(*p);
4424                         if (!nr)
4425                                 continue;               /* A hole */
4426
4427                         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
4428                                                    nr, 1)) {
4429                                 ext4_error(inode->i_sb,
4430                                            "indirect mapped block in inode "
4431                                            "#%lu invalid (level %d, blk #%lu)",
4432                                            inode->i_ino, depth,
4433                                            (unsigned long) nr);
4434                                 break;
4435                         }
4436
4437                         /* Go read the buffer for the next level down */
4438                         bh = sb_bread(inode->i_sb, nr);
4439
4440                         /*
4441                          * A read failure? Report error and clear slot
4442                          * (should be rare).
4443                          */
4444                         if (!bh) {
4445                                 ext4_error(inode->i_sb,
4446                                            "Read failure, inode=%lu, block=%llu",
4447                                            inode->i_ino, nr);
4448                                 continue;
4449                         }
4450
4451                         /* This zaps the entire block.  Bottom up. */
4452                         BUFFER_TRACE(bh, "free child branches");
4453                         ext4_free_branches(handle, inode, bh,
4454                                         (__le32 *) bh->b_data,
4455                                         (__le32 *) bh->b_data + addr_per_block,
4456                                         depth);
4457
4458                         /*
4459                          * We've probably journalled the indirect block several
4460                          * times during the truncate.  But it's no longer
4461                          * needed and we now drop it from the transaction via
4462                          * jbd2_journal_revoke().
4463                          *
4464                          * That's easy if it's exclusively part of this
4465                          * transaction.  But if it's part of the committing
4466                          * transaction then jbd2_journal_forget() will simply
4467                          * brelse() it.  That means that if the underlying
4468                          * block is reallocated in ext4_get_block(),
4469                          * unmap_underlying_metadata() will find this block
4470                          * and will try to get rid of it.  damn, damn.
4471                          *
4472                          * If this block has already been committed to the
4473                          * journal, a revoke record will be written.  And
4474                          * revoke records must be emitted *before* clearing
4475                          * this block's bit in the bitmaps.
4476                          */
4477                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
4478
4479                         /*
4480                          * Everything below this this pointer has been
4481                          * released.  Now let this top-of-subtree go.
4482                          *
4483                          * We want the freeing of this indirect block to be
4484                          * atomic in the journal with the updating of the
4485                          * bitmap block which owns it.  So make some room in
4486                          * the journal.
4487                          *
4488                          * We zero the parent pointer *after* freeing its
4489                          * pointee in the bitmaps, so if extend_transaction()
4490                          * for some reason fails to put the bitmap changes and
4491                          * the release into the same transaction, recovery
4492                          * will merely complain about releasing a free block,
4493                          * rather than leaking blocks.
4494                          */
4495                         if (ext4_handle_is_aborted(handle))
4496                                 return;
4497                         if (try_to_extend_transaction(handle, inode)) {
4498                                 ext4_mark_inode_dirty(handle, inode);
4499                                 ext4_truncate_restart_trans(handle, inode,
4500                                             blocks_for_truncate(inode));
4501                         }
4502
4503                         ext4_free_blocks(handle, inode, 0, nr, 1,
4504                                          EXT4_FREE_BLOCKS_METADATA);
4505
4506                         if (parent_bh) {
4507                                 /*
4508                                  * The block which we have just freed is
4509                                  * pointed to by an indirect block: journal it
4510                                  */
4511                                 BUFFER_TRACE(parent_bh, "get_write_access");
4512                                 if (!ext4_journal_get_write_access(handle,
4513                                                                    parent_bh)){
4514                                         *p = 0;
4515                                         BUFFER_TRACE(parent_bh,
4516                                         "call ext4_handle_dirty_metadata");
4517                                         ext4_handle_dirty_metadata(handle,
4518                                                                    inode,
4519                                                                    parent_bh);
4520                                 }
4521                         }
4522                 }
4523         } else {
4524                 /* We have reached the bottom of the tree. */
4525                 BUFFER_TRACE(parent_bh, "free data blocks");
4526                 ext4_free_data(handle, inode, parent_bh, first, last);
4527         }
4528 }
4529
4530 int ext4_can_truncate(struct inode *inode)
4531 {
4532         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4533                 return 0;
4534         if (S_ISREG(inode->i_mode))
4535                 return 1;
4536         if (S_ISDIR(inode->i_mode))
4537                 return 1;
4538         if (S_ISLNK(inode->i_mode))
4539                 return !ext4_inode_is_fast_symlink(inode);
4540         return 0;
4541 }
4542
4543 /*
4544  * ext4_truncate()
4545  *
4546  * We block out ext4_get_block() block instantiations across the entire
4547  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4548  * simultaneously on behalf of the same inode.
4549  *
4550  * As we work through the truncate and commmit bits of it to the journal there
4551  * is one core, guiding principle: the file's tree must always be consistent on
4552  * disk.  We must be able to restart the truncate after a crash.
4553  *
4554  * The file's tree may be transiently inconsistent in memory (although it
4555  * probably isn't), but whenever we close off and commit a journal transaction,
4556  * the contents of (the filesystem + the journal) must be consistent and
4557  * restartable.  It's pretty simple, really: bottom up, right to left (although
4558  * left-to-right works OK too).
4559  *
4560  * Note that at recovery time, journal replay occurs *before* the restart of
4561  * truncate against the orphan inode list.
4562  *
4563  * The committed inode has the new, desired i_size (which is the same as
4564  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4565  * that this inode's truncate did not complete and it will again call
4566  * ext4_truncate() to have another go.  So there will be instantiated blocks
4567  * to the right of the truncation point in a crashed ext4 filesystem.  But
4568  * that's fine - as long as they are linked from the inode, the post-crash
4569  * ext4_truncate() run will find them and release them.
4570  */
4571 void ext4_truncate(struct inode *inode)
4572 {
4573         handle_t *handle;
4574         struct ext4_inode_info *ei = EXT4_I(inode);
4575         __le32 *i_data = ei->i_data;
4576         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4577         struct address_space *mapping = inode->i_mapping;
4578         ext4_lblk_t offsets[4];
4579         Indirect chain[4];
4580         Indirect *partial;
4581         __le32 nr = 0;
4582         int n;
4583         ext4_lblk_t last_block;
4584         unsigned blocksize = inode->i_sb->s_blocksize;
4585
4586         if (!ext4_can_truncate(inode))
4587                 return;
4588
4589         EXT4_I(inode)->i_flags &= ~EXT4_EOFBLOCKS_FL;
4590
4591         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4592                 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4593
4594         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
4595                 ext4_ext_truncate(inode);
4596                 return;
4597         }
4598
4599         handle = start_transaction(inode);
4600         if (IS_ERR(handle))
4601                 return;         /* AKPM: return what? */
4602
4603         last_block = (inode->i_size + blocksize-1)
4604                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4605
4606         if (inode->i_size & (blocksize - 1))
4607                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4608                         goto out_stop;
4609
4610         n = ext4_block_to_path(inode, last_block, offsets, NULL);
4611         if (n == 0)
4612                 goto out_stop;  /* error */
4613
4614         /*
4615          * OK.  This truncate is going to happen.  We add the inode to the
4616          * orphan list, so that if this truncate spans multiple transactions,
4617          * and we crash, we will resume the truncate when the filesystem
4618          * recovers.  It also marks the inode dirty, to catch the new size.
4619          *
4620          * Implication: the file must always be in a sane, consistent
4621          * truncatable state while each transaction commits.
4622          */
4623         if (ext4_orphan_add(handle, inode))
4624                 goto out_stop;
4625
4626         /*
4627          * From here we block out all ext4_get_block() callers who want to
4628          * modify the block allocation tree.
4629          */
4630         down_write(&ei->i_data_sem);
4631
4632         ext4_discard_preallocations(inode);
4633
4634         /*
4635          * The orphan list entry will now protect us from any crash which
4636          * occurs before the truncate completes, so it is now safe to propagate
4637          * the new, shorter inode size (held for now in i_size) into the
4638          * on-disk inode. We do this via i_disksize, which is the value which
4639          * ext4 *really* writes onto the disk inode.
4640          */
4641         ei->i_disksize = inode->i_size;
4642
4643         if (n == 1) {           /* direct blocks */
4644                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4645                                i_data + EXT4_NDIR_BLOCKS);
4646                 goto do_indirects;
4647         }
4648
4649         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4650         /* Kill the top of shared branch (not detached) */
4651         if (nr) {
4652                 if (partial == chain) {
4653                         /* Shared branch grows from the inode */
4654                         ext4_free_branches(handle, inode, NULL,
4655                                            &nr, &nr+1, (chain+n-1) - partial);
4656                         *partial->p = 0;
4657                         /*
4658                          * We mark the inode dirty prior to restart,
4659                          * and prior to stop.  No need for it here.
4660                          */
4661                 } else {
4662                         /* Shared branch grows from an indirect block */
4663                         BUFFER_TRACE(partial->bh, "get_write_access");
4664                         ext4_free_branches(handle, inode, partial->bh,
4665                                         partial->p,
4666                                         partial->p+1, (chain+n-1) - partial);
4667                 }
4668         }
4669         /* Clear the ends of indirect blocks on the shared branch */
4670         while (partial > chain) {
4671                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4672                                    (__le32*)partial->bh->b_data+addr_per_block,
4673                                    (chain+n-1) - partial);
4674                 BUFFER_TRACE(partial->bh, "call brelse");
4675                 brelse(partial->bh);
4676                 partial--;
4677         }
4678 do_indirects:
4679         /* Kill the remaining (whole) subtrees */
4680         switch (offsets[0]) {
4681         default:
4682                 nr = i_data[EXT4_IND_BLOCK];
4683                 if (nr) {
4684                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4685                         i_data[EXT4_IND_BLOCK] = 0;
4686                 }
4687         case EXT4_IND_BLOCK:
4688                 nr = i_data[EXT4_DIND_BLOCK];
4689                 if (nr) {
4690                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4691                         i_data[EXT4_DIND_BLOCK] = 0;
4692                 }
4693         case EXT4_DIND_BLOCK:
4694                 nr = i_data[EXT4_TIND_BLOCK];
4695                 if (nr) {
4696                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4697                         i_data[EXT4_TIND_BLOCK] = 0;
4698                 }
4699         case EXT4_TIND_BLOCK:
4700                 ;
4701         }
4702
4703         up_write(&ei->i_data_sem);
4704         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4705         ext4_mark_inode_dirty(handle, inode);
4706
4707         /*
4708          * In a multi-transaction truncate, we only make the final transaction
4709          * synchronous
4710          */
4711         if (IS_SYNC(inode))
4712                 ext4_handle_sync(handle);
4713 out_stop:
4714         /*
4715          * If this was a simple ftruncate(), and the file will remain alive
4716          * then we need to clear up the orphan record which we created above.
4717          * However, if this was a real unlink then we were called by
4718          * ext4_delete_inode(), and we allow that function to clean up the
4719          * orphan info for us.
4720          */
4721         if (inode->i_nlink)
4722                 ext4_orphan_del(handle, inode);
4723
4724         ext4_journal_stop(handle);
4725 }
4726
4727 /*
4728  * ext4_get_inode_loc returns with an extra refcount against the inode's
4729  * underlying buffer_head on success. If 'in_mem' is true, we have all
4730  * data in memory that is needed to recreate the on-disk version of this
4731  * inode.
4732  */
4733 static int __ext4_get_inode_loc(struct inode *inode,
4734                                 struct ext4_iloc *iloc, int in_mem)
4735 {
4736         struct ext4_group_desc  *gdp;
4737         struct buffer_head      *bh;
4738         struct super_block      *sb = inode->i_sb;
4739         ext4_fsblk_t            block;
4740         int                     inodes_per_block, inode_offset;
4741
4742         iloc->bh = NULL;
4743         if (!ext4_valid_inum(sb, inode->i_ino))
4744                 return -EIO;
4745
4746         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4747         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4748         if (!gdp)
4749                 return -EIO;
4750
4751         /*
4752          * Figure out the offset within the block group inode table
4753          */
4754         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4755         inode_offset = ((inode->i_ino - 1) %
4756                         EXT4_INODES_PER_GROUP(sb));
4757         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4758         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4759
4760         bh = sb_getblk(sb, block);
4761         if (!bh) {
4762                 ext4_error(sb, "unable to read inode block - "
4763                            "inode=%lu, block=%llu", inode->i_ino, block);
4764                 return -EIO;
4765         }
4766         if (!buffer_uptodate(bh)) {
4767                 lock_buffer(bh);
4768
4769                 /*
4770                  * If the buffer has the write error flag, we have failed
4771                  * to write out another inode in the same block.  In this
4772                  * case, we don't have to read the block because we may
4773                  * read the old inode data successfully.
4774                  */
4775                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4776                         set_buffer_uptodate(bh);
4777
4778                 if (buffer_uptodate(bh)) {
4779                         /* someone brought it uptodate while we waited */
4780                         unlock_buffer(bh);
4781                         goto has_buffer;
4782                 }
4783
4784                 /*
4785                  * If we have all information of the inode in memory and this
4786                  * is the only valid inode in the block, we need not read the
4787                  * block.
4788                  */
4789                 if (in_mem) {
4790                         struct buffer_head *bitmap_bh;
4791                         int i, start;
4792
4793                         start = inode_offset & ~(inodes_per_block - 1);
4794
4795                         /* Is the inode bitmap in cache? */
4796                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4797                         if (!bitmap_bh)
4798                                 goto make_io;
4799
4800                         /*
4801                          * If the inode bitmap isn't in cache then the
4802                          * optimisation may end up performing two reads instead
4803                          * of one, so skip it.
4804                          */
4805                         if (!buffer_uptodate(bitmap_bh)) {
4806                                 brelse(bitmap_bh);
4807                                 goto make_io;
4808                         }
4809                         for (i = start; i < start + inodes_per_block; i++) {
4810                                 if (i == inode_offset)
4811                                         continue;
4812                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4813                                         break;
4814                         }
4815                         brelse(bitmap_bh);
4816                         if (i == start + inodes_per_block) {
4817                                 /* all other inodes are free, so skip I/O */
4818                                 memset(bh->b_data, 0, bh->b_size);
4819                                 set_buffer_uptodate(bh);
4820                                 unlock_buffer(bh);
4821                                 goto has_buffer;
4822                         }
4823                 }
4824
4825 make_io:
4826                 /*
4827                  * If we need to do any I/O, try to pre-readahead extra
4828                  * blocks from the inode table.
4829                  */
4830                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4831                         ext4_fsblk_t b, end, table;
4832                         unsigned num;
4833
4834                         table = ext4_inode_table(sb, gdp);
4835                         /* s_inode_readahead_blks is always a power of 2 */
4836                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4837                         if (table > b)
4838                                 b = table;
4839                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4840                         num = EXT4_INODES_PER_GROUP(sb);
4841                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4842                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4843                                 num -= ext4_itable_unused_count(sb, gdp);
4844                         table += num / inodes_per_block;
4845                         if (end > table)
4846                                 end = table;
4847                         while (b <= end)
4848                                 sb_breadahead(sb, b++);
4849                 }
4850
4851                 /*
4852                  * There are other valid inodes in the buffer, this inode
4853                  * has in-inode xattrs, or we don't have this inode in memory.
4854                  * Read the block from disk.
4855                  */
4856                 get_bh(bh);
4857                 bh->b_end_io = end_buffer_read_sync;
4858                 submit_bh(READ_META, bh);
4859                 wait_on_buffer(bh);
4860                 if (!buffer_uptodate(bh)) {
4861                         ext4_error(sb, "unable to read inode block - inode=%lu,"
4862                                    " block=%llu", inode->i_ino, block);
4863                         brelse(bh);
4864                         return -EIO;
4865                 }
4866         }
4867 has_buffer:
4868         iloc->bh = bh;
4869         return 0;
4870 }
4871
4872 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4873 {
4874         /* We have all inode data except xattrs in memory here. */
4875         return __ext4_get_inode_loc(inode, iloc,
4876                 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
4877 }
4878
4879 void ext4_set_inode_flags(struct inode *inode)
4880 {
4881         unsigned int flags = EXT4_I(inode)->i_flags;
4882
4883         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4884         if (flags & EXT4_SYNC_FL)
4885                 inode->i_flags |= S_SYNC;
4886         if (flags & EXT4_APPEND_FL)
4887                 inode->i_flags |= S_APPEND;
4888         if (flags & EXT4_IMMUTABLE_FL)
4889                 inode->i_flags |= S_IMMUTABLE;
4890         if (flags & EXT4_NOATIME_FL)
4891                 inode->i_flags |= S_NOATIME;
4892         if (flags & EXT4_DIRSYNC_FL)
4893                 inode->i_flags |= S_DIRSYNC;
4894 }
4895
4896 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4897 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4898 {
4899         unsigned int flags = ei->vfs_inode.i_flags;
4900
4901         ei->i_flags &= ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4902                         EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|EXT4_DIRSYNC_FL);
4903         if (flags & S_SYNC)
4904                 ei->i_flags |= EXT4_SYNC_FL;
4905         if (flags & S_APPEND)
4906                 ei->i_flags |= EXT4_APPEND_FL;
4907         if (flags & S_IMMUTABLE)
4908                 ei->i_flags |= EXT4_IMMUTABLE_FL;
4909         if (flags & S_NOATIME)
4910                 ei->i_flags |= EXT4_NOATIME_FL;
4911         if (flags & S_DIRSYNC)
4912                 ei->i_flags |= EXT4_DIRSYNC_FL;
4913 }
4914
4915 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4916                                   struct ext4_inode_info *ei)
4917 {
4918         blkcnt_t i_blocks ;
4919         struct inode *inode = &(ei->vfs_inode);
4920         struct super_block *sb = inode->i_sb;
4921
4922         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4923                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4924                 /* we are using combined 48 bit field */
4925                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4926                                         le32_to_cpu(raw_inode->i_blocks_lo);
4927                 if (ei->i_flags & EXT4_HUGE_FILE_FL) {
4928                         /* i_blocks represent file system block size */
4929                         return i_blocks  << (inode->i_blkbits - 9);
4930                 } else {
4931                         return i_blocks;
4932                 }
4933         } else {
4934                 return le32_to_cpu(raw_inode->i_blocks_lo);
4935         }
4936 }
4937
4938 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4939 {
4940         struct ext4_iloc iloc;
4941         struct ext4_inode *raw_inode;
4942         struct ext4_inode_info *ei;
4943         struct inode *inode;
4944         journal_t *journal = EXT4_SB(sb)->s_journal;
4945         long ret;
4946         int block;
4947
4948         inode = iget_locked(sb, ino);
4949         if (!inode)
4950                 return ERR_PTR(-ENOMEM);
4951         if (!(inode->i_state & I_NEW))
4952                 return inode;
4953
4954         ei = EXT4_I(inode);
4955         iloc.bh = 0;
4956
4957         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4958         if (ret < 0)
4959                 goto bad_inode;
4960         raw_inode = ext4_raw_inode(&iloc);
4961         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4962         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4963         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4964         if (!(test_opt(inode->i_sb, NO_UID32))) {
4965                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4966                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4967         }
4968         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
4969
4970         ei->i_state_flags = 0;
4971         ei->i_dir_start_lookup = 0;
4972         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4973         /* We now have enough fields to check if the inode was active or not.
4974          * This is needed because nfsd might try to access dead inodes
4975          * the test is that same one that e2fsck uses
4976          * NeilBrown 1999oct15
4977          */
4978         if (inode->i_nlink == 0) {
4979                 if (inode->i_mode == 0 ||
4980                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
4981                         /* this inode is deleted */
4982                         ret = -ESTALE;
4983                         goto bad_inode;
4984                 }
4985                 /* The only unlinked inodes we let through here have
4986                  * valid i_mode and are being read by the orphan
4987                  * recovery code: that's fine, we're about to complete
4988                  * the process of deleting those. */
4989         }
4990         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4991         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4992         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4993         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
4994                 ei->i_file_acl |=
4995                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4996         inode->i_size = ext4_isize(raw_inode);
4997         ei->i_disksize = inode->i_size;
4998 #ifdef CONFIG_QUOTA
4999         ei->i_reserved_quota = 0;
5000 #endif
5001         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
5002         ei->i_block_group = iloc.block_group;
5003         ei->i_last_alloc_group = ~0;
5004         /*
5005          * NOTE! The in-memory inode i_data array is in little-endian order
5006          * even on big-endian machines: we do NOT byteswap the block numbers!
5007          */
5008         for (block = 0; block < EXT4_N_BLOCKS; block++)
5009                 ei->i_data[block] = raw_inode->i_block[block];
5010         INIT_LIST_HEAD(&ei->i_orphan);
5011
5012         /*
5013          * Set transaction id's of transactions that have to be committed
5014          * to finish f[data]sync. We set them to currently running transaction
5015          * as we cannot be sure that the inode or some of its metadata isn't
5016          * part of the transaction - the inode could have been reclaimed and
5017          * now it is reread from disk.
5018          */
5019         if (journal) {
5020                 transaction_t *transaction;
5021                 tid_t tid;
5022
5023                 spin_lock(&journal->j_state_lock);
5024                 if (journal->j_running_transaction)
5025                         transaction = journal->j_running_transaction;
5026                 else
5027                         transaction = journal->j_committing_transaction;
5028                 if (transaction)
5029                         tid = transaction->t_tid;
5030                 else
5031                         tid = journal->j_commit_sequence;
5032                 spin_unlock(&journal->j_state_lock);
5033                 ei->i_sync_tid = tid;
5034                 ei->i_datasync_tid = tid;
5035         }
5036
5037         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5038                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
5039                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
5040                     EXT4_INODE_SIZE(inode->i_sb)) {
5041                         ret = -EIO;
5042                         goto bad_inode;
5043                 }
5044                 if (ei->i_extra_isize == 0) {
5045                         /* The extra space is currently unused. Use it. */
5046                         ei->i_extra_isize = sizeof(struct ext4_inode) -
5047                                             EXT4_GOOD_OLD_INODE_SIZE;
5048                 } else {
5049                         __le32 *magic = (void *)raw_inode +
5050                                         EXT4_GOOD_OLD_INODE_SIZE +
5051                                         ei->i_extra_isize;
5052                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
5053                                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
5054                 }
5055         } else
5056                 ei->i_extra_isize = 0;
5057
5058         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
5059         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
5060         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
5061         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
5062
5063         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
5064         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5065                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5066                         inode->i_version |=
5067                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
5068         }
5069
5070         ret = 0;
5071         if (ei->i_file_acl &&
5072             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
5073                 ext4_error(sb, "bad extended attribute block %llu inode #%lu",
5074                            ei->i_file_acl, inode->i_ino);
5075                 ret = -EIO;
5076                 goto bad_inode;
5077         } else if (ei->i_flags & EXT4_EXTENTS_FL) {
5078                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5079                     (S_ISLNK(inode->i_mode) &&
5080                      !ext4_inode_is_fast_symlink(inode)))
5081                         /* Validate extent which is part of inode */
5082                         ret = ext4_ext_check_inode(inode);
5083         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5084                    (S_ISLNK(inode->i_mode) &&
5085                     !ext4_inode_is_fast_symlink(inode))) {
5086                 /* Validate block references which are part of inode */
5087                 ret = ext4_check_inode_blockref(inode);
5088         }
5089         if (ret)
5090                 goto bad_inode;
5091
5092         if (S_ISREG(inode->i_mode)) {
5093                 inode->i_op = &ext4_file_inode_operations;
5094                 inode->i_fop = &ext4_file_operations;
5095                 ext4_set_aops(inode);
5096         } else if (S_ISDIR(inode->i_mode)) {
5097                 inode->i_op = &ext4_dir_inode_operations;
5098                 inode->i_fop = &ext4_dir_operations;
5099         } else if (S_ISLNK(inode->i_mode)) {
5100                 if (ext4_inode_is_fast_symlink(inode)) {
5101                         inode->i_op = &ext4_fast_symlink_inode_operations;
5102                         nd_terminate_link(ei->i_data, inode->i_size,
5103                                 sizeof(ei->i_data) - 1);
5104                 } else {
5105                         inode->i_op = &ext4_symlink_inode_operations;
5106                         ext4_set_aops(inode);
5107                 }
5108         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
5109               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
5110                 inode->i_op = &ext4_special_inode_operations;
5111                 if (raw_inode->i_block[0])
5112                         init_special_inode(inode, inode->i_mode,
5113                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
5114                 else
5115                         init_special_inode(inode, inode->i_mode,
5116                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
5117         } else {
5118                 ret = -EIO;
5119                 ext4_error(inode->i_sb, "bogus i_mode (%o) for inode=%lu",
5120                            inode->i_mode, inode->i_ino);
5121                 goto bad_inode;
5122         }
5123         brelse(iloc.bh);
5124         ext4_set_inode_flags(inode);
5125         unlock_new_inode(inode);
5126         return inode;
5127
5128 bad_inode:
5129         brelse(iloc.bh);
5130         iget_failed(inode);
5131         return ERR_PTR(ret);
5132 }
5133
5134 static int ext4_inode_blocks_set(handle_t *handle,
5135                                 struct ext4_inode *raw_inode,
5136                                 struct ext4_inode_info *ei)
5137 {
5138         struct inode *inode = &(ei->vfs_inode);
5139         u64 i_blocks = inode->i_blocks;
5140         struct super_block *sb = inode->i_sb;
5141
5142         if (i_blocks <= ~0U) {
5143                 /*
5144                  * i_blocks can be represnted in a 32 bit variable
5145                  * as multiple of 512 bytes
5146                  */
5147                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5148                 raw_inode->i_blocks_high = 0;
5149                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
5150                 return 0;
5151         }
5152         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
5153                 return -EFBIG;
5154
5155         if (i_blocks <= 0xffffffffffffULL) {
5156                 /*
5157                  * i_blocks can be represented in a 48 bit variable
5158                  * as multiple of 512 bytes
5159                  */
5160                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5161                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5162                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
5163         } else {
5164                 ei->i_flags |= EXT4_HUGE_FILE_FL;
5165                 /* i_block is stored in file system block size */
5166                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
5167                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5168                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5169         }
5170         return 0;
5171 }
5172
5173 /*
5174  * Post the struct inode info into an on-disk inode location in the
5175  * buffer-cache.  This gobbles the caller's reference to the
5176  * buffer_head in the inode location struct.
5177  *
5178  * The caller must have write access to iloc->bh.
5179  */
5180 static int ext4_do_update_inode(handle_t *handle,
5181                                 struct inode *inode,
5182                                 struct ext4_iloc *iloc)
5183 {
5184         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5185         struct ext4_inode_info *ei = EXT4_I(inode);
5186         struct buffer_head *bh = iloc->bh;
5187         int err = 0, rc, block;
5188
5189         /* For fields not not tracking in the in-memory inode,
5190          * initialise them to zero for new inodes. */
5191         if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5192                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5193
5194         ext4_get_inode_flags(ei);
5195         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5196         if (!(test_opt(inode->i_sb, NO_UID32))) {
5197                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
5198                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
5199 /*
5200  * Fix up interoperability with old kernels. Otherwise, old inodes get
5201  * re-used with the upper 16 bits of the uid/gid intact
5202  */
5203                 if (!ei->i_dtime) {
5204                         raw_inode->i_uid_high =
5205                                 cpu_to_le16(high_16_bits(inode->i_uid));
5206                         raw_inode->i_gid_high =
5207                                 cpu_to_le16(high_16_bits(inode->i_gid));
5208                 } else {
5209                         raw_inode->i_uid_high = 0;
5210                         raw_inode->i_gid_high = 0;
5211                 }
5212         } else {
5213                 raw_inode->i_uid_low =
5214                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
5215                 raw_inode->i_gid_low =
5216                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
5217                 raw_inode->i_uid_high = 0;
5218                 raw_inode->i_gid_high = 0;
5219         }
5220         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5221
5222         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5223         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5224         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5225         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5226
5227         if (ext4_inode_blocks_set(handle, raw_inode, ei))
5228                 goto out_brelse;
5229         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5230         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
5231         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
5232             cpu_to_le32(EXT4_OS_HURD))
5233                 raw_inode->i_file_acl_high =
5234                         cpu_to_le16(ei->i_file_acl >> 32);
5235         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5236         ext4_isize_set(raw_inode, ei->i_disksize);
5237         if (ei->i_disksize > 0x7fffffffULL) {
5238                 struct super_block *sb = inode->i_sb;
5239                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
5240                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
5241                                 EXT4_SB(sb)->s_es->s_rev_level ==
5242                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
5243                         /* If this is the first large file
5244                          * created, add a flag to the superblock.
5245                          */
5246                         err = ext4_journal_get_write_access(handle,
5247                                         EXT4_SB(sb)->s_sbh);
5248                         if (err)
5249                                 goto out_brelse;
5250                         ext4_update_dynamic_rev(sb);
5251                         EXT4_SET_RO_COMPAT_FEATURE(sb,
5252                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
5253                         sb->s_dirt = 1;
5254                         ext4_handle_sync(handle);
5255                         err = ext4_handle_dirty_metadata(handle, NULL,
5256                                         EXT4_SB(sb)->s_sbh);
5257                 }
5258         }
5259         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5260         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5261                 if (old_valid_dev(inode->i_rdev)) {
5262                         raw_inode->i_block[0] =
5263                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
5264                         raw_inode->i_block[1] = 0;
5265                 } else {
5266                         raw_inode->i_block[0] = 0;
5267                         raw_inode->i_block[1] =
5268                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
5269                         raw_inode->i_block[2] = 0;
5270                 }
5271         } else
5272                 for (block = 0; block < EXT4_N_BLOCKS; block++)
5273                         raw_inode->i_block[block] = ei->i_data[block];
5274
5275         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
5276         if (ei->i_extra_isize) {
5277                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5278                         raw_inode->i_version_hi =
5279                         cpu_to_le32(inode->i_version >> 32);
5280                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
5281         }
5282
5283         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5284         rc = ext4_handle_dirty_metadata(handle, NULL, bh);
5285         if (!err)
5286                 err = rc;
5287         ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5288
5289         ext4_update_inode_fsync_trans(handle, inode, 0);
5290 out_brelse:
5291         brelse(bh);
5292         ext4_std_error(inode->i_sb, err);
5293         return err;
5294 }
5295
5296 /*
5297  * ext4_write_inode()
5298  *
5299  * We are called from a few places:
5300  *
5301  * - Within generic_file_write() for O_SYNC files.
5302  *   Here, there will be no transaction running. We wait for any running
5303  *   trasnaction to commit.
5304  *
5305  * - Within sys_sync(), kupdate and such.
5306  *   We wait on commit, if tol to.
5307  *
5308  * - Within prune_icache() (PF_MEMALLOC == true)
5309  *   Here we simply return.  We can't afford to block kswapd on the
5310  *   journal commit.
5311  *
5312  * In all cases it is actually safe for us to return without doing anything,
5313  * because the inode has been copied into a raw inode buffer in
5314  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
5315  * knfsd.
5316  *
5317  * Note that we are absolutely dependent upon all inode dirtiers doing the
5318  * right thing: they *must* call mark_inode_dirty() after dirtying info in
5319  * which we are interested.
5320  *
5321  * It would be a bug for them to not do this.  The code:
5322  *
5323  *      mark_inode_dirty(inode)
5324  *      stuff();
5325  *      inode->i_size = expr;
5326  *
5327  * is in error because a kswapd-driven write_inode() could occur while
5328  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
5329  * will no longer be on the superblock's dirty inode list.
5330  */
5331 int ext4_write_inode(struct inode *inode, int wait)
5332 {
5333         int err;
5334
5335         if (current->flags & PF_MEMALLOC)
5336                 return 0;
5337
5338         if (EXT4_SB(inode->i_sb)->s_journal) {
5339                 if (ext4_journal_current_handle()) {
5340                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5341                         dump_stack();
5342                         return -EIO;
5343                 }
5344
5345                 if (!wait)
5346                         return 0;
5347
5348                 err = ext4_force_commit(inode->i_sb);
5349         } else {
5350                 struct ext4_iloc iloc;
5351
5352                 err = ext4_get_inode_loc(inode, &iloc);
5353                 if (err)
5354                         return err;
5355                 if (wait)
5356                         sync_dirty_buffer(iloc.bh);
5357                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5358                         ext4_error(inode->i_sb, "IO error syncing inode, "
5359                                    "inode=%lu, block=%llu", inode->i_ino,
5360                                    (unsigned long long)iloc.bh->b_blocknr);
5361                         err = -EIO;
5362                 }
5363         }
5364         return err;
5365 }
5366
5367 /*
5368  * ext4_setattr()
5369  *
5370  * Called from notify_change.
5371  *
5372  * We want to trap VFS attempts to truncate the file as soon as
5373  * possible.  In particular, we want to make sure that when the VFS
5374  * shrinks i_size, we put the inode on the orphan list and modify
5375  * i_disksize immediately, so that during the subsequent flushing of
5376  * dirty pages and freeing of disk blocks, we can guarantee that any
5377  * commit will leave the blocks being flushed in an unused state on
5378  * disk.  (On recovery, the inode will get truncated and the blocks will
5379  * be freed, so we have a strong guarantee that no future commit will
5380  * leave these blocks visible to the user.)
5381  *
5382  * Another thing we have to assure is that if we are in ordered mode
5383  * and inode is still attached to the committing transaction, we must
5384  * we start writeout of all the dirty pages which are being truncated.
5385  * This way we are sure that all the data written in the previous
5386  * transaction are already on disk (truncate waits for pages under
5387  * writeback).
5388  *
5389  * Called with inode->i_mutex down.
5390  */
5391 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5392 {
5393         struct inode *inode = dentry->d_inode;
5394         int error, rc = 0;
5395         const unsigned int ia_valid = attr->ia_valid;
5396
5397         error = inode_change_ok(inode, attr);
5398         if (error)
5399                 return error;
5400
5401         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
5402                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
5403                 handle_t *handle;
5404
5405                 /* (user+group)*(old+new) structure, inode write (sb,
5406                  * inode block, ? - but truncate inode update has it) */
5407                 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
5408                                         EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
5409                 if (IS_ERR(handle)) {
5410                         error = PTR_ERR(handle);
5411                         goto err_out;
5412                 }
5413                 error = vfs_dq_transfer(inode, attr) ? -EDQUOT : 0;
5414                 if (error) {
5415                         ext4_journal_stop(handle);
5416                         return error;
5417                 }
5418                 /* Update corresponding info in inode so that everything is in
5419                  * one transaction */
5420                 if (attr->ia_valid & ATTR_UID)
5421                         inode->i_uid = attr->ia_uid;
5422                 if (attr->ia_valid & ATTR_GID)
5423                         inode->i_gid = attr->ia_gid;
5424                 error = ext4_mark_inode_dirty(handle, inode);
5425                 ext4_journal_stop(handle);
5426         }
5427
5428         if (attr->ia_valid & ATTR_SIZE) {
5429                 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)) {
5430                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5431
5432                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
5433                                 error = -EFBIG;
5434                                 goto err_out;
5435                         }
5436                 }
5437         }
5438
5439         if (S_ISREG(inode->i_mode) &&
5440             attr->ia_valid & ATTR_SIZE &&
5441             (attr->ia_size < inode->i_size ||
5442              (EXT4_I(inode)->i_flags & EXT4_EOFBLOCKS_FL))) {
5443                 handle_t *handle;
5444
5445                 handle = ext4_journal_start(inode, 3);
5446                 if (IS_ERR(handle)) {
5447                         error = PTR_ERR(handle);
5448                         goto err_out;
5449                 }
5450
5451                 error = ext4_orphan_add(handle, inode);
5452                 EXT4_I(inode)->i_disksize = attr->ia_size;
5453                 rc = ext4_mark_inode_dirty(handle, inode);
5454                 if (!error)
5455                         error = rc;
5456                 ext4_journal_stop(handle);
5457
5458                 if (ext4_should_order_data(inode)) {
5459                         error = ext4_begin_ordered_truncate(inode,
5460                                                             attr->ia_size);
5461                         if (error) {
5462                                 /* Do as much error cleanup as possible */
5463                                 handle = ext4_journal_start(inode, 3);
5464                                 if (IS_ERR(handle)) {
5465                                         ext4_orphan_del(NULL, inode);
5466                                         goto err_out;
5467                                 }
5468                                 ext4_orphan_del(handle, inode);
5469                                 ext4_journal_stop(handle);
5470                                 goto err_out;
5471                         }
5472                 }
5473                 /* ext4_truncate will clear the flag */
5474                 if ((EXT4_I(inode)->i_flags & EXT4_EOFBLOCKS_FL))
5475                         ext4_truncate(inode);
5476         }
5477
5478         rc = inode_setattr(inode, attr);
5479
5480         /* If inode_setattr's call to ext4_truncate failed to get a
5481          * transaction handle at all, we need to clean up the in-core
5482          * orphan list manually. */
5483         if (inode->i_nlink)
5484                 ext4_orphan_del(NULL, inode);
5485
5486         if (!rc && (ia_valid & ATTR_MODE))
5487                 rc = ext4_acl_chmod(inode);
5488
5489 err_out:
5490         ext4_std_error(inode->i_sb, error);
5491         if (!error)
5492                 error = rc;
5493         return error;
5494 }
5495
5496 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
5497                  struct kstat *stat)
5498 {
5499         struct inode *inode;
5500         unsigned long delalloc_blocks;
5501
5502         inode = dentry->d_inode;
5503         generic_fillattr(inode, stat);
5504
5505         /*
5506          * We can't update i_blocks if the block allocation is delayed
5507          * otherwise in the case of system crash before the real block
5508          * allocation is done, we will have i_blocks inconsistent with
5509          * on-disk file blocks.
5510          * We always keep i_blocks updated together with real
5511          * allocation. But to not confuse with user, stat
5512          * will return the blocks that include the delayed allocation
5513          * blocks for this file.
5514          */
5515         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
5516         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
5517         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
5518
5519         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
5520         return 0;
5521 }
5522
5523 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
5524                                       int chunk)
5525 {
5526         int indirects;
5527
5528         /* if nrblocks are contiguous */
5529         if (chunk) {
5530                 /*
5531                  * With N contiguous data blocks, it need at most
5532                  * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) indirect blocks
5533                  * 2 dindirect blocks
5534                  * 1 tindirect block
5535                  */
5536                 indirects = nrblocks / EXT4_ADDR_PER_BLOCK(inode->i_sb);
5537                 return indirects + 3;
5538         }
5539         /*
5540          * if nrblocks are not contiguous, worse case, each block touch
5541          * a indirect block, and each indirect block touch a double indirect
5542          * block, plus a triple indirect block
5543          */
5544         indirects = nrblocks * 2 + 1;
5545         return indirects;
5546 }
5547
5548 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5549 {
5550         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL))
5551                 return ext4_indirect_trans_blocks(inode, nrblocks, chunk);
5552         return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
5553 }
5554
5555 /*
5556  * Account for index blocks, block groups bitmaps and block group
5557  * descriptor blocks if modify datablocks and index blocks
5558  * worse case, the indexs blocks spread over different block groups
5559  *
5560  * If datablocks are discontiguous, they are possible to spread over
5561  * different block groups too. If they are contiuguous, with flexbg,
5562  * they could still across block group boundary.
5563  *
5564  * Also account for superblock, inode, quota and xattr blocks
5565  */