Merge branch 'write_inode2' of git://git.kernel.org/pub/scm/linux/kernel/git/viro...
[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                 if (unlikely(current_block + count > EXT4_MAX_BLOCK_FILE_PHYS)) {
611                         EXT4_ERROR_INODE(inode,
612                                          "current_block %llu + count %lu > %d!",
613                                          current_block, count,
614                                          EXT4_MAX_BLOCK_FILE_PHYS);
615                         *err = -EIO;
616                         goto failed_out;
617                 }
618
619                 target -= count;
620                 /* allocate blocks for indirect blocks */
621                 while (index < indirect_blks && count) {
622                         new_blocks[index++] = current_block++;
623                         count--;
624                 }
625                 if (count > 0) {
626                         /*
627                          * save the new block number
628                          * for the first direct block
629                          */
630                         new_blocks[index] = current_block;
631                         printk(KERN_INFO "%s returned more blocks than "
632                                                 "requested\n", __func__);
633                         WARN_ON(1);
634                         break;
635                 }
636         }
637
638         target = blks - count ;
639         blk_allocated = count;
640         if (!target)
641                 goto allocated;
642         /* Now allocate data blocks */
643         memset(&ar, 0, sizeof(ar));
644         ar.inode = inode;
645         ar.goal = goal;
646         ar.len = target;
647         ar.logical = iblock;
648         if (S_ISREG(inode->i_mode))
649                 /* enable in-core preallocation only for regular files */
650                 ar.flags = EXT4_MB_HINT_DATA;
651
652         current_block = ext4_mb_new_blocks(handle, &ar, err);
653         if (unlikely(current_block + ar.len > EXT4_MAX_BLOCK_FILE_PHYS)) {
654                 EXT4_ERROR_INODE(inode,
655                                  "current_block %llu + ar.len %d > %d!",
656                                  current_block, ar.len,
657                                  EXT4_MAX_BLOCK_FILE_PHYS);
658                 *err = -EIO;
659                 goto failed_out;
660         }
661
662         if (*err && (target == blks)) {
663                 /*
664                  * if the allocation failed and we didn't allocate
665                  * any blocks before
666                  */
667                 goto failed_out;
668         }
669         if (!*err) {
670                 if (target == blks) {
671                         /*
672                          * save the new block number
673                          * for the first direct block
674                          */
675                         new_blocks[index] = current_block;
676                 }
677                 blk_allocated += ar.len;
678         }
679 allocated:
680         /* total number of blocks allocated for direct blocks */
681         ret = blk_allocated;
682         *err = 0;
683         return ret;
684 failed_out:
685         for (i = 0; i < index; i++)
686                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
687         return ret;
688 }
689
690 /**
691  *      ext4_alloc_branch - allocate and set up a chain of blocks.
692  *      @inode: owner
693  *      @indirect_blks: number of allocated indirect blocks
694  *      @blks: number of allocated direct blocks
695  *      @offsets: offsets (in the blocks) to store the pointers to next.
696  *      @branch: place to store the chain in.
697  *
698  *      This function allocates blocks, zeroes out all but the last one,
699  *      links them into chain and (if we are synchronous) writes them to disk.
700  *      In other words, it prepares a branch that can be spliced onto the
701  *      inode. It stores the information about that chain in the branch[], in
702  *      the same format as ext4_get_branch() would do. We are calling it after
703  *      we had read the existing part of chain and partial points to the last
704  *      triple of that (one with zero ->key). Upon the exit we have the same
705  *      picture as after the successful ext4_get_block(), except that in one
706  *      place chain is disconnected - *branch->p is still zero (we did not
707  *      set the last link), but branch->key contains the number that should
708  *      be placed into *branch->p to fill that gap.
709  *
710  *      If allocation fails we free all blocks we've allocated (and forget
711  *      their buffer_heads) and return the error value the from failed
712  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
713  *      as described above and return 0.
714  */
715 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
716                              ext4_lblk_t iblock, int indirect_blks,
717                              int *blks, ext4_fsblk_t goal,
718                              ext4_lblk_t *offsets, Indirect *branch)
719 {
720         int blocksize = inode->i_sb->s_blocksize;
721         int i, n = 0;
722         int err = 0;
723         struct buffer_head *bh;
724         int num;
725         ext4_fsblk_t new_blocks[4];
726         ext4_fsblk_t current_block;
727
728         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
729                                 *blks, new_blocks, &err);
730         if (err)
731                 return err;
732
733         branch[0].key = cpu_to_le32(new_blocks[0]);
734         /*
735          * metadata blocks and data blocks are allocated.
736          */
737         for (n = 1; n <= indirect_blks;  n++) {
738                 /*
739                  * Get buffer_head for parent block, zero it out
740                  * and set the pointer to new one, then send
741                  * parent to disk.
742                  */
743                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
744                 branch[n].bh = bh;
745                 lock_buffer(bh);
746                 BUFFER_TRACE(bh, "call get_create_access");
747                 err = ext4_journal_get_create_access(handle, bh);
748                 if (err) {
749                         /* Don't brelse(bh) here; it's done in
750                          * ext4_journal_forget() below */
751                         unlock_buffer(bh);
752                         goto failed;
753                 }
754
755                 memset(bh->b_data, 0, blocksize);
756                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
757                 branch[n].key = cpu_to_le32(new_blocks[n]);
758                 *branch[n].p = branch[n].key;
759                 if (n == indirect_blks) {
760                         current_block = new_blocks[n];
761                         /*
762                          * End of chain, update the last new metablock of
763                          * the chain to point to the new allocated
764                          * data blocks numbers
765                          */
766                         for (i = 1; i < num; i++)
767                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
768                 }
769                 BUFFER_TRACE(bh, "marking uptodate");
770                 set_buffer_uptodate(bh);
771                 unlock_buffer(bh);
772
773                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
774                 err = ext4_handle_dirty_metadata(handle, inode, bh);
775                 if (err)
776                         goto failed;
777         }
778         *blks = num;
779         return err;
780 failed:
781         /* Allocation failed, free what we already allocated */
782         ext4_free_blocks(handle, inode, 0, new_blocks[0], 1, 0);
783         for (i = 1; i <= n ; i++) {
784                 /* 
785                  * branch[i].bh is newly allocated, so there is no
786                  * need to revoke the block, which is why we don't
787                  * need to set EXT4_FREE_BLOCKS_METADATA.
788                  */
789                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1,
790                                  EXT4_FREE_BLOCKS_FORGET);
791         }
792         for (i = n+1; i < indirect_blks; i++)
793                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
794
795         ext4_free_blocks(handle, inode, 0, new_blocks[i], num, 0);
796
797         return err;
798 }
799
800 /**
801  * ext4_splice_branch - splice the allocated branch onto inode.
802  * @inode: owner
803  * @block: (logical) number of block we are adding
804  * @chain: chain of indirect blocks (with a missing link - see
805  *      ext4_alloc_branch)
806  * @where: location of missing link
807  * @num:   number of indirect blocks we are adding
808  * @blks:  number of direct blocks we are adding
809  *
810  * This function fills the missing link and does all housekeeping needed in
811  * inode (->i_blocks, etc.). In case of success we end up with the full
812  * chain to new block and return 0.
813  */
814 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
815                               ext4_lblk_t block, Indirect *where, int num,
816                               int blks)
817 {
818         int i;
819         int err = 0;
820         ext4_fsblk_t current_block;
821
822         /*
823          * If we're splicing into a [td]indirect block (as opposed to the
824          * inode) then we need to get write access to the [td]indirect block
825          * before the splice.
826          */
827         if (where->bh) {
828                 BUFFER_TRACE(where->bh, "get_write_access");
829                 err = ext4_journal_get_write_access(handle, where->bh);
830                 if (err)
831                         goto err_out;
832         }
833         /* That's it */
834
835         *where->p = where->key;
836
837         /*
838          * Update the host buffer_head or inode to point to more just allocated
839          * direct blocks blocks
840          */
841         if (num == 0 && blks > 1) {
842                 current_block = le32_to_cpu(where->key) + 1;
843                 for (i = 1; i < blks; i++)
844                         *(where->p + i) = cpu_to_le32(current_block++);
845         }
846
847         /* We are done with atomic stuff, now do the rest of housekeeping */
848         /* had we spliced it onto indirect block? */
849         if (where->bh) {
850                 /*
851                  * If we spliced it onto an indirect block, we haven't
852                  * altered the inode.  Note however that if it is being spliced
853                  * onto an indirect block at the very end of the file (the
854                  * file is growing) then we *will* alter the inode to reflect
855                  * the new i_size.  But that is not done here - it is done in
856                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
857                  */
858                 jbd_debug(5, "splicing indirect only\n");
859                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
860                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
861                 if (err)
862                         goto err_out;
863         } else {
864                 /*
865                  * OK, we spliced it into the inode itself on a direct block.
866                  */
867                 ext4_mark_inode_dirty(handle, inode);
868                 jbd_debug(5, "splicing direct\n");
869         }
870         return err;
871
872 err_out:
873         for (i = 1; i <= num; i++) {
874                 /* 
875                  * branch[i].bh is newly allocated, so there is no
876                  * need to revoke the block, which is why we don't
877                  * need to set EXT4_FREE_BLOCKS_METADATA.
878                  */
879                 ext4_free_blocks(handle, inode, where[i].bh, 0, 1,
880                                  EXT4_FREE_BLOCKS_FORGET);
881         }
882         ext4_free_blocks(handle, inode, 0, le32_to_cpu(where[num].key),
883                          blks, 0);
884
885         return err;
886 }
887
888 /*
889  * The ext4_ind_get_blocks() function handles non-extents inodes
890  * (i.e., using the traditional indirect/double-indirect i_blocks
891  * scheme) for ext4_get_blocks().
892  *
893  * Allocation strategy is simple: if we have to allocate something, we will
894  * have to go the whole way to leaf. So let's do it before attaching anything
895  * to tree, set linkage between the newborn blocks, write them if sync is
896  * required, recheck the path, free and repeat if check fails, otherwise
897  * set the last missing link (that will protect us from any truncate-generated
898  * removals - all blocks on the path are immune now) and possibly force the
899  * write on the parent block.
900  * That has a nice additional property: no special recovery from the failed
901  * allocations is needed - we simply release blocks and do not touch anything
902  * reachable from inode.
903  *
904  * `handle' can be NULL if create == 0.
905  *
906  * return > 0, # of blocks mapped or allocated.
907  * return = 0, if plain lookup failed.
908  * return < 0, error case.
909  *
910  * The ext4_ind_get_blocks() function should be called with
911  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
912  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
913  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
914  * blocks.
915  */
916 static int ext4_ind_get_blocks(handle_t *handle, struct inode *inode,
917                                ext4_lblk_t iblock, unsigned int maxblocks,
918                                struct buffer_head *bh_result,
919                                int flags)
920 {
921         int err = -EIO;
922         ext4_lblk_t offsets[4];
923         Indirect chain[4];
924         Indirect *partial;
925         ext4_fsblk_t goal;
926         int indirect_blks;
927         int blocks_to_boundary = 0;
928         int depth;
929         int count = 0;
930         ext4_fsblk_t first_block = 0;
931
932         J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
933         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
934         depth = ext4_block_to_path(inode, iblock, offsets,
935                                    &blocks_to_boundary);
936
937         if (depth == 0)
938                 goto out;
939
940         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
941
942         /* Simplest case - block found, no allocation needed */
943         if (!partial) {
944                 first_block = le32_to_cpu(chain[depth - 1].key);
945                 clear_buffer_new(bh_result);
946                 count++;
947                 /*map more blocks*/
948                 while (count < maxblocks && count <= blocks_to_boundary) {
949                         ext4_fsblk_t blk;
950
951                         blk = le32_to_cpu(*(chain[depth-1].p + count));
952
953                         if (blk == first_block + count)
954                                 count++;
955                         else
956                                 break;
957                 }
958                 goto got_it;
959         }
960
961         /* Next simple case - plain lookup or failed read of indirect block */
962         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
963                 goto cleanup;
964
965         /*
966          * Okay, we need to do block allocation.
967         */
968         goal = ext4_find_goal(inode, iblock, partial);
969
970         /* the number of blocks need to allocate for [d,t]indirect blocks */
971         indirect_blks = (chain + depth) - partial - 1;
972
973         /*
974          * Next look up the indirect map to count the totoal number of
975          * direct blocks to allocate for this branch.
976          */
977         count = ext4_blks_to_allocate(partial, indirect_blks,
978                                         maxblocks, blocks_to_boundary);
979         /*
980          * Block out ext4_truncate while we alter the tree
981          */
982         err = ext4_alloc_branch(handle, inode, iblock, indirect_blks,
983                                 &count, goal,
984                                 offsets + (partial - chain), partial);
985
986         /*
987          * The ext4_splice_branch call will free and forget any buffers
988          * on the new chain if there is a failure, but that risks using
989          * up transaction credits, especially for bitmaps where the
990          * credits cannot be returned.  Can we handle this somehow?  We
991          * may need to return -EAGAIN upwards in the worst case.  --sct
992          */
993         if (!err)
994                 err = ext4_splice_branch(handle, inode, iblock,
995                                          partial, indirect_blks, count);
996         if (err)
997                 goto cleanup;
998
999         set_buffer_new(bh_result);
1000
1001         ext4_update_inode_fsync_trans(handle, inode, 1);
1002 got_it:
1003         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
1004         if (count > blocks_to_boundary)
1005                 set_buffer_boundary(bh_result);
1006         err = count;
1007         /* Clean up and exit */
1008         partial = chain + depth - 1;    /* the whole chain */
1009 cleanup:
1010         while (partial > chain) {
1011                 BUFFER_TRACE(partial->bh, "call brelse");
1012                 brelse(partial->bh);
1013                 partial--;
1014         }
1015         BUFFER_TRACE(bh_result, "returned");
1016 out:
1017         return err;
1018 }
1019
1020 #ifdef CONFIG_QUOTA
1021 qsize_t *ext4_get_reserved_space(struct inode *inode)
1022 {
1023         return &EXT4_I(inode)->i_reserved_quota;
1024 }
1025 #endif
1026
1027 /*
1028  * Calculate the number of metadata blocks need to reserve
1029  * to allocate a new block at @lblocks for non extent file based file
1030  */
1031 static int ext4_indirect_calc_metadata_amount(struct inode *inode,
1032                                               sector_t lblock)
1033 {
1034         struct ext4_inode_info *ei = EXT4_I(inode);
1035         int dind_mask = EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1;
1036         int blk_bits;
1037
1038         if (lblock < EXT4_NDIR_BLOCKS)
1039                 return 0;
1040
1041         lblock -= EXT4_NDIR_BLOCKS;
1042
1043         if (ei->i_da_metadata_calc_len &&
1044             (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
1045                 ei->i_da_metadata_calc_len++;
1046                 return 0;
1047         }
1048         ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
1049         ei->i_da_metadata_calc_len = 1;
1050         blk_bits = roundup_pow_of_two(lblock + 1);
1051         return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
1052 }
1053
1054 /*
1055  * Calculate the number of metadata blocks need to reserve
1056  * to allocate a block located at @lblock
1057  */
1058 static int ext4_calc_metadata_amount(struct inode *inode, sector_t lblock)
1059 {
1060         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
1061                 return ext4_ext_calc_metadata_amount(inode, lblock);
1062
1063         return ext4_indirect_calc_metadata_amount(inode, lblock);
1064 }
1065
1066 /*
1067  * Called with i_data_sem down, which is important since we can call
1068  * ext4_discard_preallocations() from here.
1069  */
1070 void ext4_da_update_reserve_space(struct inode *inode,
1071                                         int used, int quota_claim)
1072 {
1073         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1074         struct ext4_inode_info *ei = EXT4_I(inode);
1075         int mdb_free = 0, allocated_meta_blocks = 0;
1076
1077         spin_lock(&ei->i_block_reservation_lock);
1078         trace_ext4_da_update_reserve_space(inode, used);
1079         if (unlikely(used > ei->i_reserved_data_blocks)) {
1080                 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
1081                          "with only %d reserved data blocks\n",
1082                          __func__, inode->i_ino, used,
1083                          ei->i_reserved_data_blocks);
1084                 WARN_ON(1);
1085                 used = ei->i_reserved_data_blocks;
1086         }
1087
1088         /* Update per-inode reservations */
1089         ei->i_reserved_data_blocks -= used;
1090         used += ei->i_allocated_meta_blocks;
1091         ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
1092         allocated_meta_blocks = ei->i_allocated_meta_blocks;
1093         ei->i_allocated_meta_blocks = 0;
1094         percpu_counter_sub(&sbi->s_dirtyblocks_counter, used);
1095
1096         if (ei->i_reserved_data_blocks == 0) {
1097                 /*
1098                  * We can release all of the reserved metadata blocks
1099                  * only when we have written all of the delayed
1100                  * allocation blocks.
1101                  */
1102                 mdb_free = ei->i_reserved_meta_blocks;
1103                 ei->i_reserved_meta_blocks = 0;
1104                 ei->i_da_metadata_calc_len = 0;
1105                 percpu_counter_sub(&sbi->s_dirtyblocks_counter, mdb_free);
1106         }
1107         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1108
1109         /* Update quota subsystem */
1110         if (quota_claim) {
1111                 vfs_dq_claim_block(inode, used);
1112                 if (mdb_free)
1113                         vfs_dq_release_reservation_block(inode, mdb_free);
1114         } else {
1115                 /*
1116                  * We did fallocate with an offset that is already delayed
1117                  * allocated. So on delayed allocated writeback we should
1118                  * not update the quota for allocated blocks. But then
1119                  * converting an fallocate region to initialized region would
1120                  * have caused a metadata allocation. So claim quota for
1121                  * that
1122                  */
1123                 if (allocated_meta_blocks)
1124                         vfs_dq_claim_block(inode, allocated_meta_blocks);
1125                 vfs_dq_release_reservation_block(inode, mdb_free + used);
1126         }
1127
1128         /*
1129          * If we have done all the pending block allocations and if
1130          * there aren't any writers on the inode, we can discard the
1131          * inode's preallocations.
1132          */
1133         if ((ei->i_reserved_data_blocks == 0) &&
1134             (atomic_read(&inode->i_writecount) == 0))
1135                 ext4_discard_preallocations(inode);
1136 }
1137
1138 static int check_block_validity(struct inode *inode, const char *msg,
1139                                 sector_t logical, sector_t phys, int len)
1140 {
1141         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), phys, len)) {
1142                 __ext4_error(inode->i_sb, msg,
1143                            "inode #%lu logical block %llu mapped to %llu "
1144                            "(size %d)", inode->i_ino,
1145                            (unsigned long long) logical,
1146                            (unsigned long long) phys, len);
1147                 return -EIO;
1148         }
1149         return 0;
1150 }
1151
1152 /*
1153  * Return the number of contiguous dirty pages in a given inode
1154  * starting at page frame idx.
1155  */
1156 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
1157                                     unsigned int max_pages)
1158 {
1159         struct address_space *mapping = inode->i_mapping;
1160         pgoff_t index;
1161         struct pagevec pvec;
1162         pgoff_t num = 0;
1163         int i, nr_pages, done = 0;
1164
1165         if (max_pages == 0)
1166                 return 0;
1167         pagevec_init(&pvec, 0);
1168         while (!done) {
1169                 index = idx;
1170                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
1171                                               PAGECACHE_TAG_DIRTY,
1172                                               (pgoff_t)PAGEVEC_SIZE);
1173                 if (nr_pages == 0)
1174                         break;
1175                 for (i = 0; i < nr_pages; i++) {
1176                         struct page *page = pvec.pages[i];
1177                         struct buffer_head *bh, *head;
1178
1179                         lock_page(page);
1180                         if (unlikely(page->mapping != mapping) ||
1181                             !PageDirty(page) ||
1182                             PageWriteback(page) ||
1183                             page->index != idx) {
1184                                 done = 1;
1185                                 unlock_page(page);
1186                                 break;
1187                         }
1188                         if (page_has_buffers(page)) {
1189                                 bh = head = page_buffers(page);
1190                                 do {
1191                                         if (!buffer_delay(bh) &&
1192                                             !buffer_unwritten(bh))
1193                                                 done = 1;
1194                                         bh = bh->b_this_page;
1195                                 } while (!done && (bh != head));
1196                         }
1197                         unlock_page(page);
1198                         if (done)
1199                                 break;
1200                         idx++;
1201                         num++;
1202                         if (num >= max_pages)
1203                                 break;
1204                 }
1205                 pagevec_release(&pvec);
1206         }
1207         return num;
1208 }
1209
1210 /*
1211  * The ext4_get_blocks() function tries to look up the requested blocks,
1212  * and returns if the blocks are already mapped.
1213  *
1214  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1215  * and store the allocated blocks in the result buffer head and mark it
1216  * mapped.
1217  *
1218  * If file type is extents based, it will call ext4_ext_get_blocks(),
1219  * Otherwise, call with ext4_ind_get_blocks() to handle indirect mapping
1220  * based files
1221  *
1222  * On success, it returns the number of blocks being mapped or allocate.
1223  * if create==0 and the blocks are pre-allocated and uninitialized block,
1224  * the result buffer head is unmapped. If the create ==1, it will make sure
1225  * the buffer head is mapped.
1226  *
1227  * It returns 0 if plain look up failed (blocks have not been allocated), in
1228  * that casem, buffer head is unmapped
1229  *
1230  * It returns the error in case of allocation failure.
1231  */
1232 int ext4_get_blocks(handle_t *handle, struct inode *inode, sector_t block,
1233                     unsigned int max_blocks, struct buffer_head *bh,
1234                     int flags)
1235 {
1236         int retval;
1237
1238         clear_buffer_mapped(bh);
1239         clear_buffer_unwritten(bh);
1240
1241         ext_debug("ext4_get_blocks(): inode %lu, flag %d, max_blocks %u,"
1242                   "logical block %lu\n", inode->i_ino, flags, max_blocks,
1243                   (unsigned long)block);
1244         /*
1245          * Try to see if we can get the block without requesting a new
1246          * file system block.
1247          */
1248         down_read((&EXT4_I(inode)->i_data_sem));
1249         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1250                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1251                                 bh, 0);
1252         } else {
1253                 retval = ext4_ind_get_blocks(handle, inode, block, max_blocks,
1254                                              bh, 0);
1255         }
1256         up_read((&EXT4_I(inode)->i_data_sem));
1257
1258         if (retval > 0 && buffer_mapped(bh)) {
1259                 int ret = check_block_validity(inode, "file system corruption",
1260                                                block, bh->b_blocknr, retval);
1261                 if (ret != 0)
1262                         return ret;
1263         }
1264
1265         /* If it is only a block(s) look up */
1266         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1267                 return retval;
1268
1269         /*
1270          * Returns if the blocks have already allocated
1271          *
1272          * Note that if blocks have been preallocated
1273          * ext4_ext_get_block() returns th create = 0
1274          * with buffer head unmapped.
1275          */
1276         if (retval > 0 && buffer_mapped(bh))
1277                 return retval;
1278
1279         /*
1280          * When we call get_blocks without the create flag, the
1281          * BH_Unwritten flag could have gotten set if the blocks
1282          * requested were part of a uninitialized extent.  We need to
1283          * clear this flag now that we are committed to convert all or
1284          * part of the uninitialized extent to be an initialized
1285          * extent.  This is because we need to avoid the combination
1286          * of BH_Unwritten and BH_Mapped flags being simultaneously
1287          * set on the buffer_head.
1288          */
1289         clear_buffer_unwritten(bh);
1290
1291         /*
1292          * New blocks allocate and/or writing to uninitialized extent
1293          * will possibly result in updating i_data, so we take
1294          * the write lock of i_data_sem, and call get_blocks()
1295          * with create == 1 flag.
1296          */
1297         down_write((&EXT4_I(inode)->i_data_sem));
1298
1299         /*
1300          * if the caller is from delayed allocation writeout path
1301          * we have already reserved fs blocks for allocation
1302          * let the underlying get_block() function know to
1303          * avoid double accounting
1304          */
1305         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1306                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1307         /*
1308          * We need to check for EXT4 here because migrate
1309          * could have changed the inode type in between
1310          */
1311         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1312                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1313                                               bh, flags);
1314         } else {
1315                 retval = ext4_ind_get_blocks(handle, inode, block,
1316                                              max_blocks, bh, flags);
1317
1318                 if (retval > 0 && buffer_new(bh)) {
1319                         /*
1320                          * We allocated new blocks which will result in
1321                          * i_data's format changing.  Force the migrate
1322                          * to fail by clearing migrate flags
1323                          */
1324                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
1325                 }
1326
1327                 /*
1328                  * Update reserved blocks/metadata blocks after successful
1329                  * block allocation which had been deferred till now. We don't
1330                  * support fallocate for non extent files. So we can update
1331                  * reserve space here.
1332                  */
1333                 if ((retval > 0) &&
1334                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
1335                         ext4_da_update_reserve_space(inode, retval, 1);
1336         }
1337         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1338                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1339
1340         up_write((&EXT4_I(inode)->i_data_sem));
1341         if (retval > 0 && buffer_mapped(bh)) {
1342                 int ret = check_block_validity(inode, "file system "
1343                                                "corruption after allocation",
1344                                                block, bh->b_blocknr, retval);
1345                 if (ret != 0)
1346                         return ret;
1347         }
1348         return retval;
1349 }
1350
1351 /* Maximum number of blocks we map for direct IO at once. */
1352 #define DIO_MAX_BLOCKS 4096
1353
1354 int ext4_get_block(struct inode *inode, sector_t iblock,
1355                    struct buffer_head *bh_result, int create)
1356 {
1357         handle_t *handle = ext4_journal_current_handle();
1358         int ret = 0, started = 0;
1359         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1360         int dio_credits;
1361
1362         if (create && !handle) {
1363                 /* Direct IO write... */
1364                 if (max_blocks > DIO_MAX_BLOCKS)
1365                         max_blocks = DIO_MAX_BLOCKS;
1366                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
1367                 handle = ext4_journal_start(inode, dio_credits);
1368                 if (IS_ERR(handle)) {
1369                         ret = PTR_ERR(handle);
1370                         goto out;
1371                 }
1372                 started = 1;
1373         }
1374
1375         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
1376                               create ? EXT4_GET_BLOCKS_CREATE : 0);
1377         if (ret > 0) {
1378                 bh_result->b_size = (ret << inode->i_blkbits);
1379                 ret = 0;
1380         }
1381         if (started)
1382                 ext4_journal_stop(handle);
1383 out:
1384         return ret;
1385 }
1386
1387 /*
1388  * `handle' can be NULL if create is zero
1389  */
1390 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1391                                 ext4_lblk_t block, int create, int *errp)
1392 {
1393         struct buffer_head dummy;
1394         int fatal = 0, err;
1395         int flags = 0;
1396
1397         J_ASSERT(handle != NULL || create == 0);
1398
1399         dummy.b_state = 0;
1400         dummy.b_blocknr = -1000;
1401         buffer_trace_init(&dummy.b_history);
1402         if (create)
1403                 flags |= EXT4_GET_BLOCKS_CREATE;
1404         err = ext4_get_blocks(handle, inode, block, 1, &dummy, flags);
1405         /*
1406          * ext4_get_blocks() returns number of blocks mapped. 0 in
1407          * case of a HOLE.
1408          */
1409         if (err > 0) {
1410                 if (err > 1)
1411                         WARN_ON(1);
1412                 err = 0;
1413         }
1414         *errp = err;
1415         if (!err && buffer_mapped(&dummy)) {
1416                 struct buffer_head *bh;
1417                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1418                 if (!bh) {
1419                         *errp = -EIO;
1420                         goto err;
1421                 }
1422                 if (buffer_new(&dummy)) {
1423                         J_ASSERT(create != 0);
1424                         J_ASSERT(handle != NULL);
1425
1426                         /*
1427                          * Now that we do not always journal data, we should
1428                          * keep in mind whether this should always journal the
1429                          * new buffer as metadata.  For now, regular file
1430                          * writes use ext4_get_block instead, so it's not a
1431                          * problem.
1432                          */
1433                         lock_buffer(bh);
1434                         BUFFER_TRACE(bh, "call get_create_access");
1435                         fatal = ext4_journal_get_create_access(handle, bh);
1436                         if (!fatal && !buffer_uptodate(bh)) {
1437                                 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1438                                 set_buffer_uptodate(bh);
1439                         }
1440                         unlock_buffer(bh);
1441                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1442                         err = ext4_handle_dirty_metadata(handle, inode, bh);
1443                         if (!fatal)
1444                                 fatal = err;
1445                 } else {
1446                         BUFFER_TRACE(bh, "not a new buffer");
1447                 }
1448                 if (fatal) {
1449                         *errp = fatal;
1450                         brelse(bh);
1451                         bh = NULL;
1452                 }
1453                 return bh;
1454         }
1455 err:
1456         return NULL;
1457 }
1458
1459 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1460                                ext4_lblk_t block, int create, int *err)
1461 {
1462         struct buffer_head *bh;
1463
1464         bh = ext4_getblk(handle, inode, block, create, err);
1465         if (!bh)
1466                 return bh;
1467         if (buffer_uptodate(bh))
1468                 return bh;
1469         ll_rw_block(READ_META, 1, &bh);
1470         wait_on_buffer(bh);
1471         if (buffer_uptodate(bh))
1472                 return bh;
1473         put_bh(bh);
1474         *err = -EIO;
1475         return NULL;
1476 }
1477
1478 static int walk_page_buffers(handle_t *handle,
1479                              struct buffer_head *head,
1480                              unsigned from,
1481                              unsigned to,
1482                              int *partial,
1483                              int (*fn)(handle_t *handle,
1484                                        struct buffer_head *bh))
1485 {
1486         struct buffer_head *bh;
1487         unsigned block_start, block_end;
1488         unsigned blocksize = head->b_size;
1489         int err, ret = 0;
1490         struct buffer_head *next;
1491
1492         for (bh = head, block_start = 0;
1493              ret == 0 && (bh != head || !block_start);
1494              block_start = block_end, bh = next) {
1495                 next = bh->b_this_page;
1496                 block_end = block_start + blocksize;
1497                 if (block_end <= from || block_start >= to) {
1498                         if (partial && !buffer_uptodate(bh))
1499                                 *partial = 1;
1500                         continue;
1501                 }
1502                 err = (*fn)(handle, bh);
1503                 if (!ret)
1504                         ret = err;
1505         }
1506         return ret;
1507 }
1508
1509 /*
1510  * To preserve ordering, it is essential that the hole instantiation and
1511  * the data write be encapsulated in a single transaction.  We cannot
1512  * close off a transaction and start a new one between the ext4_get_block()
1513  * and the commit_write().  So doing the jbd2_journal_start at the start of
1514  * prepare_write() is the right place.
1515  *
1516  * Also, this function can nest inside ext4_writepage() ->
1517  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1518  * has generated enough buffer credits to do the whole page.  So we won't
1519  * block on the journal in that case, which is good, because the caller may
1520  * be PF_MEMALLOC.
1521  *
1522  * By accident, ext4 can be reentered when a transaction is open via
1523  * quota file writes.  If we were to commit the transaction while thus
1524  * reentered, there can be a deadlock - we would be holding a quota
1525  * lock, and the commit would never complete if another thread had a
1526  * transaction open and was blocking on the quota lock - a ranking
1527  * violation.
1528  *
1529  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1530  * will _not_ run commit under these circumstances because handle->h_ref
1531  * is elevated.  We'll still have enough credits for the tiny quotafile
1532  * write.
1533  */
1534 static int do_journal_get_write_access(handle_t *handle,
1535                                        struct buffer_head *bh)
1536 {
1537         if (!buffer_mapped(bh) || buffer_freed(bh))
1538                 return 0;
1539         return ext4_journal_get_write_access(handle, bh);
1540 }
1541
1542 /*
1543  * Truncate blocks that were not used by write. We have to truncate the
1544  * pagecache as well so that corresponding buffers get properly unmapped.
1545  */
1546 static void ext4_truncate_failed_write(struct inode *inode)
1547 {
1548         truncate_inode_pages(inode->i_mapping, inode->i_size);
1549         ext4_truncate(inode);
1550 }
1551
1552 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
1553                    struct buffer_head *bh_result, int create);
1554 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1555                             loff_t pos, unsigned len, unsigned flags,
1556                             struct page **pagep, void **fsdata)
1557 {
1558         struct inode *inode = mapping->host;
1559         int ret, needed_blocks;
1560         handle_t *handle;
1561         int retries = 0;
1562         struct page *page;
1563         pgoff_t index;
1564         unsigned from, to;
1565
1566         trace_ext4_write_begin(inode, pos, len, flags);
1567         /*
1568          * Reserve one block more for addition to orphan list in case
1569          * we allocate blocks but write fails for some reason
1570          */
1571         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1572         index = pos >> PAGE_CACHE_SHIFT;
1573         from = pos & (PAGE_CACHE_SIZE - 1);
1574         to = from + len;
1575
1576 retry:
1577         handle = ext4_journal_start(inode, needed_blocks);
1578         if (IS_ERR(handle)) {
1579                 ret = PTR_ERR(handle);
1580                 goto out;
1581         }
1582
1583         /* We cannot recurse into the filesystem as the transaction is already
1584          * started */
1585         flags |= AOP_FLAG_NOFS;
1586
1587         page = grab_cache_page_write_begin(mapping, index, flags);
1588         if (!page) {
1589                 ext4_journal_stop(handle);
1590                 ret = -ENOMEM;
1591                 goto out;
1592         }
1593         *pagep = page;
1594
1595         if (ext4_should_dioread_nolock(inode))
1596                 ret = block_write_begin(file, mapping, pos, len, flags, pagep,
1597                                 fsdata, ext4_get_block_write);
1598         else
1599                 ret = block_write_begin(file, mapping, pos, len, flags, pagep,
1600                                 fsdata, ext4_get_block);
1601
1602         if (!ret && ext4_should_journal_data(inode)) {
1603                 ret = walk_page_buffers(handle, page_buffers(page),
1604                                 from, to, NULL, do_journal_get_write_access);
1605         }
1606
1607         if (ret) {
1608                 unlock_page(page);
1609                 page_cache_release(page);
1610                 /*
1611                  * block_write_begin may have instantiated a few blocks
1612                  * outside i_size.  Trim these off again. Don't need
1613                  * i_size_read because we hold i_mutex.
1614                  *
1615                  * Add inode to orphan list in case we crash before
1616                  * truncate finishes
1617                  */
1618                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1619                         ext4_orphan_add(handle, inode);
1620
1621                 ext4_journal_stop(handle);
1622                 if (pos + len > inode->i_size) {
1623                         ext4_truncate_failed_write(inode);
1624                         /*
1625                          * If truncate failed early the inode might
1626                          * still be on the orphan list; we need to
1627                          * make sure the inode is removed from the
1628                          * orphan list in that case.
1629                          */
1630                         if (inode->i_nlink)
1631                                 ext4_orphan_del(NULL, inode);
1632                 }
1633         }
1634
1635         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1636                 goto retry;
1637 out:
1638         return ret;
1639 }
1640
1641 /* For write_end() in data=journal mode */
1642 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1643 {
1644         if (!buffer_mapped(bh) || buffer_freed(bh))
1645                 return 0;
1646         set_buffer_uptodate(bh);
1647         return ext4_handle_dirty_metadata(handle, NULL, bh);
1648 }
1649
1650 static int ext4_generic_write_end(struct file *file,
1651                                   struct address_space *mapping,
1652                                   loff_t pos, unsigned len, unsigned copied,
1653                                   struct page *page, void *fsdata)
1654 {
1655         int i_size_changed = 0;
1656         struct inode *inode = mapping->host;
1657         handle_t *handle = ext4_journal_current_handle();
1658
1659         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1660
1661         /*
1662          * No need to use i_size_read() here, the i_size
1663          * cannot change under us because we hold i_mutex.
1664          *
1665          * But it's important to update i_size while still holding page lock:
1666          * page writeout could otherwise come in and zero beyond i_size.
1667          */
1668         if (pos + copied > inode->i_size) {
1669                 i_size_write(inode, pos + copied);
1670                 i_size_changed = 1;
1671         }
1672
1673         if (pos + copied >  EXT4_I(inode)->i_disksize) {
1674                 /* We need to mark inode dirty even if
1675                  * new_i_size is less that inode->i_size
1676                  * bu greater than i_disksize.(hint delalloc)
1677                  */
1678                 ext4_update_i_disksize(inode, (pos + copied));
1679                 i_size_changed = 1;
1680         }
1681         unlock_page(page);
1682         page_cache_release(page);
1683
1684         /*
1685          * Don't mark the inode dirty under page lock. First, it unnecessarily
1686          * makes the holding time of page lock longer. Second, it forces lock
1687          * ordering of page lock and transaction start for journaling
1688          * filesystems.
1689          */
1690         if (i_size_changed)
1691                 ext4_mark_inode_dirty(handle, inode);
1692
1693         return copied;
1694 }
1695
1696 /*
1697  * We need to pick up the new inode size which generic_commit_write gave us
1698  * `file' can be NULL - eg, when called from page_symlink().
1699  *
1700  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1701  * buffers are managed internally.
1702  */
1703 static int ext4_ordered_write_end(struct file *file,
1704                                   struct address_space *mapping,
1705                                   loff_t pos, unsigned len, unsigned copied,
1706                                   struct page *page, void *fsdata)
1707 {
1708         handle_t *handle = ext4_journal_current_handle();
1709         struct inode *inode = mapping->host;
1710         int ret = 0, ret2;
1711
1712         trace_ext4_ordered_write_end(inode, pos, len, copied);
1713         ret = ext4_jbd2_file_inode(handle, inode);
1714
1715         if (ret == 0) {
1716                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1717                                                         page, fsdata);
1718                 copied = ret2;
1719                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1720                         /* if we have allocated more blocks and copied
1721                          * less. We will have blocks allocated outside
1722                          * inode->i_size. So truncate them
1723                          */
1724                         ext4_orphan_add(handle, inode);
1725                 if (ret2 < 0)
1726                         ret = ret2;
1727         }
1728         ret2 = ext4_journal_stop(handle);
1729         if (!ret)
1730                 ret = ret2;
1731
1732         if (pos + len > inode->i_size) {
1733                 ext4_truncate_failed_write(inode);
1734                 /*
1735                  * If truncate failed early the inode might still be
1736                  * on the orphan list; we need to make sure the inode
1737                  * is removed from the orphan list in that case.
1738                  */
1739                 if (inode->i_nlink)
1740                         ext4_orphan_del(NULL, inode);
1741         }
1742
1743
1744         return ret ? ret : copied;
1745 }
1746
1747 static int ext4_writeback_write_end(struct file *file,
1748                                     struct address_space *mapping,
1749                                     loff_t pos, unsigned len, unsigned copied,
1750                                     struct page *page, void *fsdata)
1751 {
1752         handle_t *handle = ext4_journal_current_handle();
1753         struct inode *inode = mapping->host;
1754         int ret = 0, ret2;
1755
1756         trace_ext4_writeback_write_end(inode, pos, len, copied);
1757         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1758                                                         page, fsdata);
1759         copied = ret2;
1760         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1761                 /* if we have allocated more blocks and copied
1762                  * less. We will have blocks allocated outside
1763                  * inode->i_size. So truncate them
1764                  */
1765                 ext4_orphan_add(handle, inode);
1766
1767         if (ret2 < 0)
1768                 ret = ret2;
1769
1770         ret2 = ext4_journal_stop(handle);
1771         if (!ret)
1772                 ret = ret2;
1773
1774         if (pos + len > inode->i_size) {
1775                 ext4_truncate_failed_write(inode);
1776                 /*
1777                  * If truncate failed early the inode might still be
1778                  * on the orphan list; we need to make sure the inode
1779                  * is removed from the orphan list in that case.
1780                  */
1781                 if (inode->i_nlink)
1782                         ext4_orphan_del(NULL, inode);
1783         }
1784
1785         return ret ? ret : copied;
1786 }
1787
1788 static int ext4_journalled_write_end(struct file *file,
1789                                      struct address_space *mapping,
1790                                      loff_t pos, unsigned len, unsigned copied,
1791                                      struct page *page, void *fsdata)
1792 {
1793         handle_t *handle = ext4_journal_current_handle();
1794         struct inode *inode = mapping->host;
1795         int ret = 0, ret2;
1796         int partial = 0;
1797         unsigned from, to;
1798         loff_t new_i_size;
1799
1800         trace_ext4_journalled_write_end(inode, pos, len, copied);
1801         from = pos & (PAGE_CACHE_SIZE - 1);
1802         to = from + len;
1803
1804         if (copied < len) {
1805                 if (!PageUptodate(page))
1806                         copied = 0;
1807                 page_zero_new_buffers(page, from+copied, to);
1808         }
1809
1810         ret = walk_page_buffers(handle, page_buffers(page), from,
1811                                 to, &partial, write_end_fn);
1812         if (!partial)
1813                 SetPageUptodate(page);
1814         new_i_size = pos + copied;
1815         if (new_i_size > inode->i_size)
1816                 i_size_write(inode, pos+copied);
1817         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1818         if (new_i_size > EXT4_I(inode)->i_disksize) {
1819                 ext4_update_i_disksize(inode, new_i_size);
1820                 ret2 = ext4_mark_inode_dirty(handle, inode);
1821                 if (!ret)
1822                         ret = ret2;
1823         }
1824
1825         unlock_page(page);
1826         page_cache_release(page);
1827         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1828                 /* if we have allocated more blocks and copied
1829                  * less. We will have blocks allocated outside
1830                  * inode->i_size. So truncate them
1831                  */
1832                 ext4_orphan_add(handle, inode);
1833
1834         ret2 = ext4_journal_stop(handle);
1835         if (!ret)
1836                 ret = ret2;
1837         if (pos + len > inode->i_size) {
1838                 ext4_truncate_failed_write(inode);
1839                 /*
1840                  * If truncate failed early the inode might still be
1841                  * on the orphan list; we need to make sure the inode
1842                  * is removed from the orphan list in that case.
1843                  */
1844                 if (inode->i_nlink)
1845                         ext4_orphan_del(NULL, inode);
1846         }
1847
1848         return ret ? ret : copied;
1849 }
1850
1851 /*
1852  * Reserve a single block located at lblock
1853  */
1854 static int ext4_da_reserve_space(struct inode *inode, sector_t lblock)
1855 {
1856         int retries = 0;
1857         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1858         struct ext4_inode_info *ei = EXT4_I(inode);
1859         unsigned long md_needed, md_reserved;
1860
1861         /*
1862          * recalculate the amount of metadata blocks to reserve
1863          * in order to allocate nrblocks
1864          * worse case is one extent per block
1865          */
1866 repeat:
1867         spin_lock(&ei->i_block_reservation_lock);
1868         md_reserved = ei->i_reserved_meta_blocks;
1869         md_needed = ext4_calc_metadata_amount(inode, lblock);
1870         trace_ext4_da_reserve_space(inode, md_needed);
1871         spin_unlock(&ei->i_block_reservation_lock);
1872
1873         /*
1874          * Make quota reservation here to prevent quota overflow
1875          * later. Real quota accounting is done at pages writeout
1876          * time.
1877          */
1878         if (vfs_dq_reserve_block(inode, md_needed + 1))
1879                 return -EDQUOT;
1880
1881         if (ext4_claim_free_blocks(sbi, md_needed + 1)) {
1882                 vfs_dq_release_reservation_block(inode, md_needed + 1);
1883                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1884                         yield();
1885                         goto repeat;
1886                 }
1887                 return -ENOSPC;
1888         }
1889         spin_lock(&ei->i_block_reservation_lock);
1890         ei->i_reserved_data_blocks++;
1891         ei->i_reserved_meta_blocks += md_needed;
1892         spin_unlock(&ei->i_block_reservation_lock);
1893
1894         return 0;       /* success */
1895 }
1896
1897 static void ext4_da_release_space(struct inode *inode, int to_free)
1898 {
1899         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1900         struct ext4_inode_info *ei = EXT4_I(inode);
1901
1902         if (!to_free)
1903                 return;         /* Nothing to release, exit */
1904
1905         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1906
1907         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1908                 /*
1909                  * if there aren't enough reserved blocks, then the
1910                  * counter is messed up somewhere.  Since this
1911                  * function is called from invalidate page, it's
1912                  * harmless to return without any action.
1913                  */
1914                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1915                          "ino %lu, to_free %d with only %d reserved "
1916                          "data blocks\n", inode->i_ino, to_free,
1917                          ei->i_reserved_data_blocks);
1918                 WARN_ON(1);
1919                 to_free = ei->i_reserved_data_blocks;
1920         }
1921         ei->i_reserved_data_blocks -= to_free;
1922
1923         if (ei->i_reserved_data_blocks == 0) {
1924                 /*
1925                  * We can release all of the reserved metadata blocks
1926                  * only when we have written all of the delayed
1927                  * allocation blocks.
1928                  */
1929                 to_free += ei->i_reserved_meta_blocks;
1930                 ei->i_reserved_meta_blocks = 0;
1931                 ei->i_da_metadata_calc_len = 0;
1932         }
1933
1934         /* update fs dirty blocks counter */
1935         percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1936
1937         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1938
1939         vfs_dq_release_reservation_block(inode, to_free);
1940 }
1941
1942 static void ext4_da_page_release_reservation(struct page *page,
1943                                              unsigned long offset)
1944 {
1945         int to_release = 0;
1946         struct buffer_head *head, *bh;
1947         unsigned int curr_off = 0;
1948
1949         head = page_buffers(page);
1950         bh = head;
1951         do {
1952                 unsigned int next_off = curr_off + bh->b_size;
1953
1954                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1955                         to_release++;
1956                         clear_buffer_delay(bh);
1957                 }
1958                 curr_off = next_off;
1959         } while ((bh = bh->b_this_page) != head);
1960         ext4_da_release_space(page->mapping->host, to_release);
1961 }
1962
1963 /*
1964  * Delayed allocation stuff
1965  */
1966
1967 /*
1968  * mpage_da_submit_io - walks through extent of pages and try to write
1969  * them with writepage() call back
1970  *
1971  * @mpd->inode: inode
1972  * @mpd->first_page: first page of the extent
1973  * @mpd->next_page: page after the last page of the extent
1974  *
1975  * By the time mpage_da_submit_io() is called we expect all blocks
1976  * to be allocated. this may be wrong if allocation failed.
1977  *
1978  * As pages are already locked by write_cache_pages(), we can't use it
1979  */
1980 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1981 {
1982         long pages_skipped;
1983         struct pagevec pvec;
1984         unsigned long index, end;
1985         int ret = 0, err, nr_pages, i;
1986         struct inode *inode = mpd->inode;
1987         struct address_space *mapping = inode->i_mapping;
1988
1989         BUG_ON(mpd->next_page <= mpd->first_page);
1990         /*
1991          * We need to start from the first_page to the next_page - 1
1992          * to make sure we also write the mapped dirty buffer_heads.
1993          * If we look at mpd->b_blocknr we would only be looking
1994          * at the currently mapped buffer_heads.
1995          */
1996         index = mpd->first_page;
1997         end = mpd->next_page - 1;
1998
1999         pagevec_init(&pvec, 0);
2000         while (index <= end) {
2001                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2002                 if (nr_pages == 0)
2003                         break;
2004                 for (i = 0; i < nr_pages; i++) {
2005                         struct page *page = pvec.pages[i];
2006
2007                         index = page->index;
2008                         if (index > end)
2009                                 break;
2010                         index++;
2011
2012                         BUG_ON(!PageLocked(page));
2013                         BUG_ON(PageWriteback(page));
2014
2015                         pages_skipped = mpd->wbc->pages_skipped;
2016                         err = mapping->a_ops->writepage(page, mpd->wbc);
2017                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
2018                                 /*
2019                                  * have successfully written the page
2020                                  * without skipping the same
2021                                  */
2022                                 mpd->pages_written++;
2023                         /*
2024                          * In error case, we have to continue because
2025                          * remaining pages are still locked
2026                          * XXX: unlock and re-dirty them?
2027                          */
2028                         if (ret == 0)
2029                                 ret = err;
2030                 }
2031                 pagevec_release(&pvec);
2032         }
2033         return ret;
2034 }
2035
2036 /*
2037  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
2038  *
2039  * @mpd->inode - inode to walk through
2040  * @exbh->b_blocknr - first block on a disk
2041  * @exbh->b_size - amount of space in bytes
2042  * @logical - first logical block to start assignment with
2043  *
2044  * the function goes through all passed space and put actual disk
2045  * block numbers into buffer heads, dropping BH_Delay and BH_Unwritten
2046  */
2047 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
2048                                  struct buffer_head *exbh)
2049 {
2050         struct inode *inode = mpd->inode;
2051         struct address_space *mapping = inode->i_mapping;
2052         int blocks = exbh->b_size >> inode->i_blkbits;
2053         sector_t pblock = exbh->b_blocknr, cur_logical;
2054         struct buffer_head *head, *bh;
2055         pgoff_t index, end;
2056         struct pagevec pvec;
2057         int nr_pages, i;
2058
2059         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2060         end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2061         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2062
2063         pagevec_init(&pvec, 0);
2064
2065         while (index <= end) {
2066                 /* XXX: optimize tail */
2067                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2068                 if (nr_pages == 0)
2069                         break;
2070                 for (i = 0; i < nr_pages; i++) {
2071                         struct page *page = pvec.pages[i];
2072
2073                         index = page->index;
2074                         if (index > end)
2075                                 break;
2076                         index++;
2077
2078                         BUG_ON(!PageLocked(page));
2079                         BUG_ON(PageWriteback(page));
2080                         BUG_ON(!page_has_buffers(page));
2081
2082                         bh = page_buffers(page);
2083                         head = bh;
2084
2085                         /* skip blocks out of the range */
2086                         do {
2087                                 if (cur_logical >= logical)
2088                                         break;
2089                                 cur_logical++;
2090                         } while ((bh = bh->b_this_page) != head);
2091
2092                         do {
2093                                 if (cur_logical >= logical + blocks)
2094                                         break;
2095
2096                                 if (buffer_delay(bh) ||
2097                                                 buffer_unwritten(bh)) {
2098
2099                                         BUG_ON(bh->b_bdev != inode->i_sb->s_bdev);
2100
2101                                         if (buffer_delay(bh)) {
2102                                                 clear_buffer_delay(bh);
2103                                                 bh->b_blocknr = pblock;
2104                                         } else {
2105                                                 /*
2106                                                  * unwritten already should have
2107                                                  * blocknr assigned. Verify that
2108                                                  */
2109                                                 clear_buffer_unwritten(bh);
2110                                                 BUG_ON(bh->b_blocknr != pblock);
2111                                         }
2112
2113                                 } else if (buffer_mapped(bh))
2114                                         BUG_ON(bh->b_blocknr != pblock);
2115
2116                                 if (buffer_uninit(exbh))
2117                                         set_buffer_uninit(bh);
2118                                 cur_logical++;
2119                                 pblock++;
2120                         } while ((bh = bh->b_this_page) != head);
2121                 }
2122                 pagevec_release(&pvec);
2123         }
2124 }
2125
2126
2127 /*
2128  * __unmap_underlying_blocks - just a helper function to unmap
2129  * set of blocks described by @bh
2130  */
2131 static inline void __unmap_underlying_blocks(struct inode *inode,
2132                                              struct buffer_head *bh)
2133 {
2134         struct block_device *bdev = inode->i_sb->s_bdev;
2135         int blocks, i;
2136
2137         blocks = bh->b_size >> inode->i_blkbits;
2138         for (i = 0; i < blocks; i++)
2139                 unmap_underlying_metadata(bdev, bh->b_blocknr + i);
2140 }
2141
2142 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2143                                         sector_t logical, long blk_cnt)
2144 {
2145         int nr_pages, i;
2146         pgoff_t index, end;
2147         struct pagevec pvec;
2148         struct inode *inode = mpd->inode;
2149         struct address_space *mapping = inode->i_mapping;
2150
2151         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2152         end   = (logical + blk_cnt - 1) >>
2153                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2154         while (index <= end) {
2155                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2156                 if (nr_pages == 0)
2157                         break;
2158                 for (i = 0; i < nr_pages; i++) {
2159                         struct page *page = pvec.pages[i];
2160                         if (page->index > end)
2161                                 break;
2162                         BUG_ON(!PageLocked(page));
2163                         BUG_ON(PageWriteback(page));
2164                         block_invalidatepage(page, 0);
2165                         ClearPageUptodate(page);
2166                         unlock_page(page);
2167                 }
2168                 index = pvec.pages[nr_pages - 1]->index + 1;
2169                 pagevec_release(&pvec);
2170         }
2171         return;
2172 }
2173
2174 static void ext4_print_free_blocks(struct inode *inode)
2175 {
2176         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2177         printk(KERN_CRIT "Total free blocks count %lld\n",
2178                ext4_count_free_blocks(inode->i_sb));
2179         printk(KERN_CRIT "Free/Dirty block details\n");
2180         printk(KERN_CRIT "free_blocks=%lld\n",
2181                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2182         printk(KERN_CRIT "dirty_blocks=%lld\n",
2183                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2184         printk(KERN_CRIT "Block reservation details\n");
2185         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2186                EXT4_I(inode)->i_reserved_data_blocks);
2187         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2188                EXT4_I(inode)->i_reserved_meta_blocks);
2189         return;
2190 }
2191
2192 /*
2193  * mpage_da_map_blocks - go through given space
2194  *
2195  * @mpd - bh describing space
2196  *
2197  * The function skips space we know is already mapped to disk blocks.
2198  *
2199  */
2200 static int mpage_da_map_blocks(struct mpage_da_data *mpd)
2201 {
2202         int err, blks, get_blocks_flags;
2203         struct buffer_head new;
2204         sector_t next = mpd->b_blocknr;
2205         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2206         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2207         handle_t *handle = NULL;
2208
2209         /*
2210          * We consider only non-mapped and non-allocated blocks
2211          */
2212         if ((mpd->b_state  & (1 << BH_Mapped)) &&
2213                 !(mpd->b_state & (1 << BH_Delay)) &&
2214                 !(mpd->b_state & (1 << BH_Unwritten)))
2215                 return 0;
2216
2217         /*
2218          * If we didn't accumulate anything to write simply return
2219          */
2220         if (!mpd->b_size)
2221                 return 0;
2222
2223         handle = ext4_journal_current_handle();
2224         BUG_ON(!handle);
2225
2226         /*
2227          * Call ext4_get_blocks() to allocate any delayed allocation
2228          * blocks, or to convert an uninitialized extent to be
2229          * initialized (in the case where we have written into
2230          * one or more preallocated blocks).
2231          *
2232          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2233          * indicate that we are on the delayed allocation path.  This
2234          * affects functions in many different parts of the allocation
2235          * call path.  This flag exists primarily because we don't
2236          * want to change *many* call functions, so ext4_get_blocks()
2237          * will set the magic i_delalloc_reserved_flag once the
2238          * inode's allocation semaphore is taken.
2239          *
2240          * If the blocks in questions were delalloc blocks, set
2241          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2242          * variables are updated after the blocks have been allocated.
2243          */
2244         new.b_state = 0;
2245         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
2246         if (ext4_should_dioread_nolock(mpd->inode))
2247                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2248         if (mpd->b_state & (1 << BH_Delay))
2249                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2250
2251         blks = ext4_get_blocks(handle, mpd->inode, next, max_blocks,
2252                                &new, get_blocks_flags);
2253         if (blks < 0) {
2254                 err = blks;
2255                 /*
2256                  * If get block returns with error we simply
2257                  * return. Later writepage will redirty the page and
2258                  * writepages will find the dirty page again
2259                  */
2260                 if (err == -EAGAIN)
2261                         return 0;
2262
2263                 if (err == -ENOSPC &&
2264                     ext4_count_free_blocks(mpd->inode->i_sb)) {
2265                         mpd->retval = err;
2266                         return 0;
2267                 }
2268
2269                 /*
2270                  * get block failure will cause us to loop in
2271                  * writepages, because a_ops->writepage won't be able
2272                  * to make progress. The page will be redirtied by
2273                  * writepage and writepages will again try to write
2274                  * the same.
2275                  */
2276                 ext4_msg(mpd->inode->i_sb, KERN_CRIT,
2277                          "delayed block allocation failed for inode %lu at "
2278                          "logical offset %llu with max blocks %zd with "
2279                          "error %d\n", mpd->inode->i_ino,
2280                          (unsigned long long) next,
2281                          mpd->b_size >> mpd->inode->i_blkbits, err);
2282                 printk(KERN_CRIT "This should not happen!!  "
2283                        "Data will be lost\n");
2284                 if (err == -ENOSPC) {
2285                         ext4_print_free_blocks(mpd->inode);
2286                 }
2287                 /* invalidate all the pages */
2288                 ext4_da_block_invalidatepages(mpd, next,
2289                                 mpd->b_size >> mpd->inode->i_blkbits);
2290                 return err;
2291         }
2292         BUG_ON(blks == 0);
2293
2294         new.b_size = (blks << mpd->inode->i_blkbits);
2295
2296         if (buffer_new(&new))
2297                 __unmap_underlying_blocks(mpd->inode, &new);
2298
2299         /*
2300          * If blocks are delayed marked, we need to
2301          * put actual blocknr and drop delayed bit
2302          */
2303         if ((mpd->b_state & (1 << BH_Delay)) ||
2304             (mpd->b_state & (1 << BH_Unwritten)))
2305                 mpage_put_bnr_to_bhs(mpd, next, &new);
2306
2307         if (ext4_should_order_data(mpd->inode)) {
2308                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2309                 if (err)
2310                         return err;
2311         }
2312
2313         /*
2314          * Update on-disk size along with block allocation.
2315          */
2316         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2317         if (disksize > i_size_read(mpd->inode))
2318                 disksize = i_size_read(mpd->inode);
2319         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2320                 ext4_update_i_disksize(mpd->inode, disksize);
2321                 return ext4_mark_inode_dirty(handle, mpd->inode);
2322         }
2323
2324         return 0;
2325 }
2326
2327 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2328                 (1 << BH_Delay) | (1 << BH_Unwritten))
2329
2330 /*
2331  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2332  *
2333  * @mpd->lbh - extent of blocks
2334  * @logical - logical number of the block in the file
2335  * @bh - bh of the block (used to access block's state)
2336  *
2337  * the function is used to collect contig. blocks in same state
2338  */
2339 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2340                                    sector_t logical, size_t b_size,
2341                                    unsigned long b_state)
2342 {
2343         sector_t next;
2344         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2345
2346         /* check if thereserved journal credits might overflow */
2347         if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
2348                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2349                         /*
2350                          * With non-extent format we are limited by the journal
2351                          * credit available.  Total credit needed to insert
2352                          * nrblocks contiguous blocks is dependent on the
2353                          * nrblocks.  So limit nrblocks.
2354                          */
2355                         goto flush_it;
2356                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2357                                 EXT4_MAX_TRANS_DATA) {
2358                         /*
2359                          * Adding the new buffer_head would make it cross the
2360                          * allowed limit for which we have journal credit
2361                          * reserved. So limit the new bh->b_size
2362                          */
2363                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2364                                                 mpd->inode->i_blkbits;
2365                         /* we will do mpage_da_submit_io in the next loop */
2366                 }
2367         }
2368         /*
2369          * First block in the extent
2370          */
2371         if (mpd->b_size == 0) {
2372                 mpd->b_blocknr = logical;
2373                 mpd->b_size = b_size;
2374                 mpd->b_state = b_state & BH_FLAGS;
2375                 return;
2376         }
2377
2378         next = mpd->b_blocknr + nrblocks;
2379         /*
2380          * Can we merge the block to our big extent?
2381          */
2382         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2383                 mpd->b_size += b_size;
2384                 return;
2385         }
2386
2387 flush_it:
2388         /*
2389          * We couldn't merge the block to our extent, so we
2390          * need to flush current  extent and start new one
2391          */
2392         if (mpage_da_map_blocks(mpd) == 0)
2393                 mpage_da_submit_io(mpd);
2394         mpd->io_done = 1;
2395         return;
2396 }
2397
2398 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
2399 {
2400         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
2401 }
2402
2403 /*
2404  * __mpage_da_writepage - finds extent of pages and blocks
2405  *
2406  * @page: page to consider
2407  * @wbc: not used, we just follow rules
2408  * @data: context
2409  *
2410  * The function finds extents of pages and scan them for all blocks.
2411  */
2412 static int __mpage_da_writepage(struct page *page,
2413                                 struct writeback_control *wbc, void *data)
2414 {
2415         struct mpage_da_data *mpd = data;
2416         struct inode *inode = mpd->inode;
2417         struct buffer_head *bh, *head;
2418         sector_t logical;
2419
2420         if (mpd->io_done) {
2421                 /*
2422                  * Rest of the page in the page_vec
2423                  * redirty then and skip then. We will
2424                  * try to write them again after
2425                  * starting a new transaction
2426                  */
2427                 redirty_page_for_writepage(wbc, page);
2428                 unlock_page(page);
2429                 return MPAGE_DA_EXTENT_TAIL;
2430         }
2431         /*
2432          * Can we merge this page to current extent?
2433          */
2434         if (mpd->next_page != page->index) {
2435                 /*
2436                  * Nope, we can't. So, we map non-allocated blocks
2437                  * and start IO on them using writepage()
2438                  */
2439                 if (mpd->next_page != mpd->first_page) {
2440                         if (mpage_da_map_blocks(mpd) == 0)
2441                                 mpage_da_submit_io(mpd);
2442                         /*
2443                          * skip rest of the page in the page_vec
2444                          */
2445                         mpd->io_done = 1;
2446                         redirty_page_for_writepage(wbc, page);
2447                         unlock_page(page);
2448                         return MPAGE_DA_EXTENT_TAIL;
2449                 }
2450
2451                 /*
2452                  * Start next extent of pages ...
2453                  */
2454                 mpd->first_page = page->index;
2455
2456                 /*
2457                  * ... and blocks
2458                  */
2459                 mpd->b_size = 0;
2460                 mpd->b_state = 0;
2461                 mpd->b_blocknr = 0;
2462         }
2463
2464         mpd->next_page = page->index + 1;
2465         logical = (sector_t) page->index <<
2466                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2467
2468         if (!page_has_buffers(page)) {
2469                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2470                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2471                 if (mpd->io_done)
2472                         return MPAGE_DA_EXTENT_TAIL;
2473         } else {
2474                 /*
2475                  * Page with regular buffer heads, just add all dirty ones
2476                  */
2477                 head = page_buffers(page);
2478                 bh = head;
2479                 do {
2480                         BUG_ON(buffer_locked(bh));
2481                         /*
2482                          * We need to try to allocate
2483                          * unmapped blocks in the same page.
2484                          * Otherwise we won't make progress
2485                          * with the page in ext4_writepage
2486                          */
2487                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2488                                 mpage_add_bh_to_extent(mpd, logical,
2489                                                        bh->b_size,
2490                                                        bh->b_state);
2491                                 if (mpd->io_done)
2492                                         return MPAGE_DA_EXTENT_TAIL;
2493                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2494                                 /*
2495                                  * mapped dirty buffer. We need to update
2496                                  * the b_state because we look at
2497                                  * b_state in mpage_da_map_blocks. We don't
2498                                  * update b_size because if we find an
2499                                  * unmapped buffer_head later we need to
2500                                  * use the b_state flag of that buffer_head.
2501                                  */
2502                                 if (mpd->b_size == 0)
2503                                         mpd->b_state = bh->b_state & BH_FLAGS;
2504                         }
2505                         logical++;
2506                 } while ((bh = bh->b_this_page) != head);
2507         }
2508
2509         return 0;
2510 }
2511
2512 /*
2513  * This is a special get_blocks_t callback which is used by
2514  * ext4_da_write_begin().  It will either return mapped block or
2515  * reserve space for a single block.
2516  *
2517  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2518  * We also have b_blocknr = -1 and b_bdev initialized properly
2519  *
2520  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2521  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2522  * initialized properly.
2523  */
2524 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2525                                   struct buffer_head *bh_result, int create)
2526 {
2527         int ret = 0;
2528         sector_t invalid_block = ~((sector_t) 0xffff);
2529
2530         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2531                 invalid_block = ~0;
2532
2533         BUG_ON(create == 0);
2534         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2535
2536         /*
2537          * first, we need to know whether the block is allocated already
2538          * preallocated blocks are unmapped but should treated
2539          * the same as allocated blocks.
2540          */
2541         ret = ext4_get_blocks(NULL, inode, iblock, 1,  bh_result, 0);
2542         if ((ret == 0) && !buffer_delay(bh_result)) {
2543                 /* the block isn't (pre)allocated yet, let's reserve space */
2544                 /*
2545                  * XXX: __block_prepare_write() unmaps passed block,
2546                  * is it OK?
2547                  */
2548                 ret = ext4_da_reserve_space(inode, iblock);
2549                 if (ret)
2550                         /* not enough space to reserve */
2551                         return ret;
2552
2553                 map_bh(bh_result, inode->i_sb, invalid_block);
2554                 set_buffer_new(bh_result);
2555                 set_buffer_delay(bh_result);
2556         } else if (ret > 0) {
2557                 bh_result->b_size = (ret << inode->i_blkbits);
2558                 if (buffer_unwritten(bh_result)) {
2559                         /* A delayed write to unwritten bh should
2560                          * be marked new and mapped.  Mapped ensures
2561                          * that we don't do get_block multiple times
2562                          * when we write to the same offset and new
2563                          * ensures that we do proper zero out for
2564                          * partial write.
2565                          */
2566                         set_buffer_new(bh_result);
2567                         set_buffer_mapped(bh_result);
2568                 }
2569                 ret = 0;
2570         }
2571
2572         return ret;
2573 }
2574
2575 /*
2576  * This function is used as a standard get_block_t calback function
2577  * when there is no desire to allocate any blocks.  It is used as a
2578  * callback function for block_prepare_write(), nobh_writepage(), and
2579  * block_write_full_page().  These functions should only try to map a
2580  * single block at a time.
2581  *
2582  * Since this function doesn't do block allocations even if the caller
2583  * requests it by passing in create=1, it is critically important that
2584  * any caller checks to make sure that any buffer heads are returned
2585  * by this function are either all already mapped or marked for
2586  * delayed allocation before calling nobh_writepage() or
2587  * block_write_full_page().  Otherwise, b_blocknr could be left
2588  * unitialized, and the page write functions will be taken by
2589  * surprise.
2590  */
2591 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2592                                    struct buffer_head *bh_result, int create)
2593 {
2594         int ret = 0;
2595         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2596
2597         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2598
2599         /*
2600          * we don't want to do block allocation in writepage
2601          * so call get_block_wrap with create = 0
2602          */
2603         ret = ext4_get_blocks(NULL, inode, iblock, max_blocks, bh_result, 0);
2604         if (ret > 0) {
2605                 bh_result->b_size = (ret << inode->i_blkbits);
2606                 ret = 0;
2607         }
2608         return ret;
2609 }
2610
2611 static int bget_one(handle_t *handle, struct buffer_head *bh)
2612 {
2613         get_bh(bh);
2614         return 0;
2615 }
2616
2617 static int bput_one(handle_t *handle, struct buffer_head *bh)
2618 {
2619         put_bh(bh);
2620         return 0;
2621 }
2622
2623 static int __ext4_journalled_writepage(struct page *page,
2624                                        unsigned int len)
2625 {
2626         struct address_space *mapping = page->mapping;
2627         struct inode *inode = mapping->host;
2628         struct buffer_head *page_bufs;
2629         handle_t *handle = NULL;
2630         int ret = 0;
2631         int err;
2632
2633         page_bufs = page_buffers(page);
2634         BUG_ON(!page_bufs);
2635         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2636         /* As soon as we unlock the page, it can go away, but we have
2637          * references to buffers so we are safe */
2638         unlock_page(page);
2639
2640         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2641         if (IS_ERR(handle)) {
2642                 ret = PTR_ERR(handle);
2643                 goto out;
2644         }
2645
2646         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2647                                 do_journal_get_write_access);
2648
2649         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2650                                 write_end_fn);
2651         if (ret == 0)
2652                 ret = err;
2653         err = ext4_journal_stop(handle);
2654         if (!ret)
2655                 ret = err;
2656
2657         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2658         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2659 out:
2660         return ret;
2661 }
2662
2663 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
2664 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
2665
2666 /*
2667  * Note that we don't need to start a transaction unless we're journaling data
2668  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2669  * need to file the inode to the transaction's list in ordered mode because if
2670  * we are writing back data added by write(), the inode is already there and if
2671  * we are writing back data modified via mmap(), noone guarantees in which
2672  * transaction the data will hit the disk. In case we are journaling data, we
2673  * cannot start transaction directly because transaction start ranks above page
2674  * lock so we have to do some magic.
2675  *
2676  * This function can get called via...
2677  *   - ext4_da_writepages after taking page lock (have journal handle)
2678  *   - journal_submit_inode_data_buffers (no journal handle)
2679  *   - shrink_page_list via pdflush (no journal handle)
2680  *   - grab_page_cache when doing write_begin (have journal handle)
2681  *
2682  * We don't do any block allocation in this function. If we have page with
2683  * multiple blocks we need to write those buffer_heads that are mapped. This
2684  * is important for mmaped based write. So if we do with blocksize 1K
2685  * truncate(f, 1024);
2686  * a = mmap(f, 0, 4096);
2687  * a[0] = 'a';
2688  * truncate(f, 4096);
2689  * we have in the page first buffer_head mapped via page_mkwrite call back
2690  * but other bufer_heads would be unmapped but dirty(dirty done via the
2691  * do_wp_page). So writepage should write the first block. If we modify
2692  * the mmap area beyond 1024 we will again get a page_fault and the
2693  * page_mkwrite callback will do the block allocation and mark the
2694  * buffer_heads mapped.
2695  *
2696  * We redirty the page if we have any buffer_heads that is either delay or
2697  * unwritten in the page.
2698  *
2699  * We can get recursively called as show below.
2700  *
2701  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2702  *              ext4_writepage()
2703  *
2704  * But since we don't do any block allocation we should not deadlock.
2705  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2706  */
2707 static int ext4_writepage(struct page *page,
2708                           struct writeback_control *wbc)
2709 {
2710         int ret = 0;
2711         loff_t size;
2712         unsigned int len;
2713         struct buffer_head *page_bufs = NULL;
2714         struct inode *inode = page->mapping->host;
2715
2716         trace_ext4_writepage(inode, page);
2717         size = i_size_read(inode);
2718         if (page->index == size >> PAGE_CACHE_SHIFT)
2719                 len = size & ~PAGE_CACHE_MASK;
2720         else
2721                 len = PAGE_CACHE_SIZE;
2722
2723         if (page_has_buffers(page)) {
2724                 page_bufs = page_buffers(page);
2725                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2726                                         ext4_bh_delay_or_unwritten)) {
2727                         /*
2728                          * We don't want to do  block allocation
2729                          * So redirty the page and return
2730                          * We may reach here when we do a journal commit
2731                          * via journal_submit_inode_data_buffers.
2732                          * If we don't have mapping block we just ignore
2733                          * them. We can also reach here via shrink_page_list
2734                          */
2735                         redirty_page_for_writepage(wbc, page);
2736                         unlock_page(page);
2737                         return 0;
2738                 }
2739         } else {
2740                 /*
2741                  * The test for page_has_buffers() is subtle:
2742                  * We know the page is dirty but it lost buffers. That means
2743                  * that at some moment in time after write_begin()/write_end()
2744                  * has been called all buffers have been clean and thus they
2745                  * must have been written at least once. So they are all
2746                  * mapped and we can happily proceed with mapping them
2747                  * and writing the page.
2748                  *
2749                  * Try to initialize the buffer_heads and check whether
2750                  * all are mapped and non delay. We don't want to
2751                  * do block allocation here.
2752                  */
2753                 ret = block_prepare_write(page, 0, len,
2754                                           noalloc_get_block_write);
2755                 if (!ret) {
2756                         page_bufs = page_buffers(page);
2757                         /* check whether all are mapped and non delay */
2758                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2759                                                 ext4_bh_delay_or_unwritten)) {
2760                                 redirty_page_for_writepage(wbc, page);
2761                                 unlock_page(page);
2762                                 return 0;
2763                         }
2764                 } else {
2765                         /*
2766                          * We can't do block allocation here
2767                          * so just redity the page and unlock
2768                          * and return
2769                          */
2770                         redirty_page_for_writepage(wbc, page);
2771                         unlock_page(page);
2772                         return 0;
2773                 }
2774                 /* now mark the buffer_heads as dirty and uptodate */
2775                 block_commit_write(page, 0, len);
2776         }
2777
2778         if (PageChecked(page) && ext4_should_journal_data(inode)) {
2779                 /*
2780                  * It's mmapped pagecache.  Add buffers and journal it.  There
2781                  * doesn't seem much point in redirtying the page here.
2782                  */
2783                 ClearPageChecked(page);
2784                 return __ext4_journalled_writepage(page, len);
2785         }
2786
2787         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2788                 ret = nobh_writepage(page, noalloc_get_block_write, wbc);
2789         else if (page_bufs && buffer_uninit(page_bufs)) {
2790                 ext4_set_bh_endio(page_bufs, inode);
2791                 ret = block_write_full_page_endio(page, noalloc_get_block_write,
2792                                             wbc, ext4_end_io_buffer_write);
2793         } else
2794                 ret = block_write_full_page(page, noalloc_get_block_write,
2795                                             wbc);
2796
2797         return ret;
2798 }
2799
2800 /*
2801  * This is called via ext4_da_writepages() to
2802  * calulate the total number of credits to reserve to fit
2803  * a single extent allocation into a single transaction,
2804  * ext4_da_writpeages() will loop calling this before
2805  * the block allocation.
2806  */
2807
2808 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2809 {
2810         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2811
2812         /*
2813          * With non-extent format the journal credit needed to
2814          * insert nrblocks contiguous block is dependent on
2815          * number of contiguous block. So we will limit
2816          * number of contiguous block to a sane value
2817          */
2818         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) &&
2819             (max_blocks > EXT4_MAX_TRANS_DATA))
2820                 max_blocks = EXT4_MAX_TRANS_DATA;
2821
2822         return ext4_chunk_trans_blocks(inode, max_blocks);
2823 }
2824
2825 static int ext4_da_writepages(struct address_space *mapping,
2826                               struct writeback_control *wbc)
2827 {
2828         pgoff_t index;
2829         int range_whole = 0;
2830         handle_t *handle = NULL;
2831         struct mpage_da_data mpd;
2832         struct inode *inode = mapping->host;
2833         int no_nrwrite_index_update;
2834         int pages_written = 0;
2835         long pages_skipped;
2836         unsigned int max_pages;
2837         int range_cyclic, cycled = 1, io_done = 0;
2838         int needed_blocks, ret = 0;
2839         long desired_nr_to_write, nr_to_writebump = 0;
2840         loff_t range_start = wbc->range_start;
2841         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2842
2843         trace_ext4_da_writepages(inode, wbc);
2844
2845         /*
2846          * No pages to write? This is mainly a kludge to avoid starting
2847          * a transaction for special inodes like journal inode on last iput()
2848          * because that could violate lock ordering on umount
2849          */
2850         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2851                 return 0;
2852
2853         /*
2854          * If the filesystem has aborted, it is read-only, so return
2855          * right away instead of dumping stack traces later on that
2856          * will obscure the real source of the problem.  We test
2857          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2858          * the latter could be true if the filesystem is mounted
2859          * read-only, and in that case, ext4_da_writepages should
2860          * *never* be called, so if that ever happens, we would want
2861          * the stack trace.
2862          */
2863         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2864                 return -EROFS;
2865
2866         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2867                 range_whole = 1;
2868
2869         range_cyclic = wbc->range_cyclic;
2870         if (wbc->range_cyclic) {
2871                 index = mapping->writeback_index;
2872                 if (index)
2873                         cycled = 0;
2874                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2875                 wbc->range_end  = LLONG_MAX;
2876                 wbc->range_cyclic = 0;
2877         } else
2878                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2879
2880         /*
2881          * This works around two forms of stupidity.  The first is in
2882          * the writeback code, which caps the maximum number of pages
2883          * written to be 1024 pages.  This is wrong on multiple
2884          * levels; different architectues have a different page size,
2885          * which changes the maximum amount of data which gets
2886          * written.  Secondly, 4 megabytes is way too small.  XFS
2887          * forces this value to be 16 megabytes by multiplying
2888          * nr_to_write parameter by four, and then relies on its
2889          * allocator to allocate larger extents to make them
2890          * contiguous.  Unfortunately this brings us to the second
2891          * stupidity, which is that ext4's mballoc code only allocates
2892          * at most 2048 blocks.  So we force contiguous writes up to
2893          * the number of dirty blocks in the inode, or
2894          * sbi->max_writeback_mb_bump whichever is smaller.
2895          */
2896         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2897         if (!range_cyclic && range_whole)
2898                 desired_nr_to_write = wbc->nr_to_write * 8;
2899         else
2900                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2901                                                            max_pages);
2902         if (desired_nr_to_write > max_pages)
2903                 desired_nr_to_write = max_pages;
2904
2905         if (wbc->nr_to_write < desired_nr_to_write) {
2906                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2907                 wbc->nr_to_write = desired_nr_to_write;
2908         }
2909
2910         mpd.wbc = wbc;
2911         mpd.inode = mapping->host;
2912
2913         /*
2914          * we don't want write_cache_pages to update
2915          * nr_to_write and writeback_index
2916          */
2917         no_nrwrite_index_update = wbc->no_nrwrite_index_update;
2918         wbc->no_nrwrite_index_update = 1;
2919         pages_skipped = wbc->pages_skipped;
2920
2921 retry:
2922         while (!ret && wbc->nr_to_write > 0) {
2923
2924                 /*
2925                  * we  insert one extent at a time. So we need
2926                  * credit needed for single extent allocation.
2927                  * journalled mode is currently not supported
2928                  * by delalloc
2929                  */
2930                 BUG_ON(ext4_should_journal_data(inode));
2931                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2932
2933                 /* start a new transaction*/
2934                 handle = ext4_journal_start(inode, needed_blocks);
2935                 if (IS_ERR(handle)) {
2936                         ret = PTR_ERR(handle);
2937                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2938                                "%ld pages, ino %lu; err %d\n", __func__,
2939                                 wbc->nr_to_write, inode->i_ino, ret);
2940                         goto out_writepages;
2941                 }
2942
2943                 /*
2944                  * Now call __mpage_da_writepage to find the next
2945                  * contiguous region of logical blocks that need
2946                  * blocks to be allocated by ext4.  We don't actually
2947                  * submit the blocks for I/O here, even though
2948                  * write_cache_pages thinks it will, and will set the
2949                  * pages as clean for write before calling
2950                  * __mpage_da_writepage().
2951                  */
2952                 mpd.b_size = 0;
2953                 mpd.b_state = 0;
2954                 mpd.b_blocknr = 0;
2955                 mpd.first_page = 0;
2956                 mpd.next_page = 0;
2957                 mpd.io_done = 0;
2958                 mpd.pages_written = 0;
2959                 mpd.retval = 0;
2960                 ret = write_cache_pages(mapping, wbc, __mpage_da_writepage,
2961                                         &mpd);
2962                 /*
2963                  * If we have a contiguous extent of pages and we
2964                  * haven't done the I/O yet, map the blocks and submit
2965                  * them for I/O.
2966                  */
2967                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2968                         if (mpage_da_map_blocks(&mpd) == 0)
2969                                 mpage_da_submit_io(&mpd);
2970                         mpd.io_done = 1;
2971                         ret = MPAGE_DA_EXTENT_TAIL;
2972                 }
2973                 trace_ext4_da_write_pages(inode, &mpd);
2974                 wbc->nr_to_write -= mpd.pages_written;
2975
2976                 ext4_journal_stop(handle);
2977
2978                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2979                         /* commit the transaction which would
2980                          * free blocks released in the transaction
2981                          * and try again
2982                          */
2983                         jbd2_journal_force_commit_nested(sbi->s_journal);
2984                         wbc->pages_skipped = pages_skipped;
2985                         ret = 0;
2986                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2987                         /*
2988                          * got one extent now try with
2989                          * rest of the pages
2990                          */
2991                         pages_written += mpd.pages_written;
2992                         wbc->pages_skipped = pages_skipped;
2993                         ret = 0;
2994                         io_done = 1;
2995                 } else if (wbc->nr_to_write)
2996                         /*
2997                          * There is no more writeout needed
2998                          * or we requested for a noblocking writeout
2999                          * and we found the device congested
3000                          */
3001                         break;
3002         }
3003         if (!io_done && !cycled) {
3004                 cycled = 1;
3005                 index = 0;
3006                 wbc->range_start = index << PAGE_CACHE_SHIFT;
3007                 wbc->range_end  = mapping->writeback_index - 1;
3008                 goto retry;
3009         }
3010         if (pages_skipped != wbc->pages_skipped)
3011                 ext4_msg(inode->i_sb, KERN_CRIT,
3012                          "This should not happen leaving %s "
3013                          "with nr_to_write = %ld ret = %d\n",
3014                          __func__, wbc->nr_to_write, ret);
3015
3016         /* Update index */
3017         index += pages_written;
3018         wbc->range_cyclic = range_cyclic;
3019         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
3020                 /*
3021                  * set the writeback_index so that range_cyclic
3022                  * mode will write it back later
3023                  */
3024                 mapping->writeback_index = index;
3025
3026 out_writepages:
3027         if (!no_nrwrite_index_update)
3028                 wbc->no_nrwrite_index_update = 0;
3029         wbc->nr_to_write -= nr_to_writebump;
3030         wbc->range_start = range_start;
3031         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
3032         return ret;
3033 }
3034
3035 #define FALL_BACK_TO_NONDELALLOC 1
3036 static int ext4_nonda_switch(struct super_block *sb)
3037 {
3038         s64 free_blocks, dirty_blocks;
3039         struct ext4_sb_info *sbi = EXT4_SB(sb);
3040
3041         /*
3042          * switch to non delalloc mode if we are running low
3043          * on free block. The free block accounting via percpu
3044          * counters can get slightly wrong with percpu_counter_batch getting
3045          * accumulated on each CPU without updating global counters
3046          * Delalloc need an accurate free block accounting. So switch
3047          * to non delalloc when we are near to error range.
3048          */
3049         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
3050         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
3051         if (2 * free_blocks < 3 * dirty_blocks ||
3052                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
3053                 /*
3054                  * free block count is less than 150% of dirty blocks
3055                  * or free blocks is less than watermark
3056                  */
3057                 return 1;
3058         }
3059         /*
3060          * Even if we don't switch but are nearing capacity,
3061          * start pushing delalloc when 1/2 of free blocks are dirty.
3062          */
3063         if (free_blocks < 2 * dirty_blocks)
3064                 writeback_inodes_sb_if_idle(sb);
3065
3066         return 0;
3067 }
3068
3069 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3070                                loff_t pos, unsigned len, unsigned flags,
3071                                struct page **pagep, void **fsdata)
3072 {
3073         int ret, retries = 0, quota_retries = 0;
3074         struct page *page;
3075         pgoff_t index;
3076         unsigned from, to;
3077         struct inode *inode = mapping->host;
3078         handle_t *handle;
3079
3080         index = pos >> PAGE_CACHE_SHIFT;
3081         from = pos & (PAGE_CACHE_SIZE - 1);
3082         to = from + len;
3083
3084         if (ext4_nonda_switch(inode->i_sb)) {
3085                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3086                 return ext4_write_begin(file, mapping, pos,
3087                                         len, flags, pagep, fsdata);
3088         }
3089         *fsdata = (void *)0;
3090         trace_ext4_da_write_begin(inode, pos, len, flags);
3091 retry:
3092         /*
3093          * With delayed allocation, we don't log the i_disksize update
3094          * if there is delayed block allocation. But we still need
3095          * to journalling the i_disksize update if writes to the end
3096          * of file which has an already mapped buffer.
3097          */
3098         handle = ext4_journal_start(inode, 1);
3099         if (IS_ERR(handle)) {
3100                 ret = PTR_ERR(handle);
3101                 goto out;
3102         }
3103         /* We cannot recurse into the filesystem as the transaction is already
3104          * started */
3105         flags |= AOP_FLAG_NOFS;
3106
3107         page = grab_cache_page_write_begin(mapping, index, flags);
3108         if (!page) {
3109                 ext4_journal_stop(handle);
3110                 ret = -ENOMEM;
3111                 goto out;
3112         }
3113         *pagep = page;
3114
3115         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
3116                                 ext4_da_get_block_prep);
3117         if (ret < 0) {
3118                 unlock_page(page);
3119                 ext4_journal_stop(handle);
3120                 page_cache_release(page);
3121                 /*
3122                  * block_write_begin may have instantiated a few blocks
3123                  * outside i_size.  Trim these off again. Don't need
3124                  * i_size_read because we hold i_mutex.
3125                  */
3126                 if (pos + len > inode->i_size)
3127                         ext4_truncate_failed_write(inode);
3128         }
3129
3130         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3131                 goto retry;
3132
3133         if ((ret == -EDQUOT) &&
3134             EXT4_I(inode)->i_reserved_meta_blocks &&
3135             (quota_retries++ < 3)) {
3136                 /*
3137                  * Since we often over-estimate the number of meta
3138                  * data blocks required, we may sometimes get a
3139                  * spurios out of quota error even though there would
3140                  * be enough space once we write the data blocks and
3141                  * find out how many meta data blocks were _really_
3142                  * required.  So try forcing the inode write to see if
3143                  * that helps.
3144                  */
3145                 write_inode_now(inode, (quota_retries == 3));
3146                 goto retry;
3147         }
3148 out:
3149         return ret;
3150 }
3151
3152 /*
3153  * Check if we should update i_disksize
3154  * when write to the end of file but not require block allocation
3155  */
3156 static int ext4_da_should_update_i_disksize(struct page *page,
3157                                             unsigned long offset)
3158 {
3159         struct buffer_head *bh;
3160         struct inode *inode = page->mapping->host;
3161         unsigned int idx;
3162         int i;
3163
3164         bh = page_buffers(page);
3165         idx = offset >> inode->i_blkbits;
3166
3167         for (i = 0; i < idx; i++)
3168                 bh = bh->b_this_page;
3169
3170         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3171                 return 0;
3172         return 1;
3173 }
3174
3175 static int ext4_da_write_end(struct file *file,
3176                              struct address_space *mapping,
3177                              loff_t pos, unsigned len, unsigned copied,
3178                              struct page *page, void *fsdata)
3179 {
3180         struct inode *inode = mapping->host;
3181         int ret = 0, ret2;
3182         handle_t *handle = ext4_journal_current_handle();
3183         loff_t new_i_size;
3184         unsigned long start, end;
3185         int write_mode = (int)(unsigned long)fsdata;
3186
3187         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
3188                 if (ext4_should_order_data(inode)) {
3189                         return ext4_ordered_write_end(file, mapping, pos,
3190                                         len, copied, page, fsdata);
3191                 } else if (ext4_should_writeback_data(inode)) {
3192                         return ext4_writeback_write_end(file, mapping, pos,
3193                                         len, copied, page, fsdata);
3194                 } else {
3195                         BUG();
3196                 }
3197         }
3198
3199         trace_ext4_da_write_end(inode, pos, len, copied);
3200         start = pos & (PAGE_CACHE_SIZE - 1);
3201         end = start + copied - 1;
3202
3203         /*
3204          * generic_write_end() will run mark_inode_dirty() if i_size
3205          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3206          * into that.
3207          */
3208
3209         new_i_size = pos + copied;
3210         if (new_i_size > EXT4_I(inode)->i_disksize) {
3211                 if (ext4_da_should_update_i_disksize(page, end)) {
3212                         down_write(&EXT4_I(inode)->i_data_sem);
3213                         if (new_i_size > EXT4_I(inode)->i_disksize) {
3214                                 /*
3215                                  * Updating i_disksize when extending file
3216                                  * without needing block allocation
3217                                  */
3218                                 if (ext4_should_order_data(inode))
3219                                         ret = ext4_jbd2_file_inode(handle,
3220                                                                    inode);
3221
3222                                 EXT4_I(inode)->i_disksize = new_i_size;
3223                         }
3224                         up_write(&EXT4_I(inode)->i_data_sem);
3225                         /* We need to mark inode dirty even if
3226                          * new_i_size is less that inode->i_size
3227                          * bu greater than i_disksize.(hint delalloc)
3228                          */
3229                         ext4_mark_inode_dirty(handle, inode);
3230                 }
3231         }
3232         ret2 = generic_write_end(file, mapping, pos, len, copied,
3233                                                         page, fsdata);
3234         copied = ret2;
3235         if (ret2 < 0)
3236                 ret = ret2;
3237         ret2 = ext4_journal_stop(handle);
3238         if (!ret)
3239                 ret = ret2;
3240
3241         return ret ? ret : copied;
3242 }
3243
3244 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
3245 {
3246         /*
3247          * Drop reserved blocks
3248          */
3249         BUG_ON(!PageLocked(page));
3250         if (!page_has_buffers(page))
3251                 goto out;
3252
3253         ext4_da_page_release_reservation(page, offset);
3254
3255 out:
3256         ext4_invalidatepage(page, offset);
3257
3258         return;
3259 }
3260
3261 /*
3262  * Force all delayed allocation blocks to be allocated for a given inode.
3263  */
3264 int ext4_alloc_da_blocks(struct inode *inode)
3265 {
3266         trace_ext4_alloc_da_blocks(inode);
3267
3268         if (!EXT4_I(inode)->i_reserved_data_blocks &&
3269             !EXT4_I(inode)->i_reserved_meta_blocks)
3270                 return 0;
3271
3272         /*
3273          * We do something simple for now.  The filemap_flush() will
3274          * also start triggering a write of the data blocks, which is
3275          * not strictly speaking necessary (and for users of
3276          * laptop_mode, not even desirable).  However, to do otherwise
3277          * would require replicating code paths in:
3278          *
3279          * ext4_da_writepages() ->
3280          *    write_cache_pages() ---> (via passed in callback function)
3281          *        __mpage_da_writepage() -->
3282          *           mpage_add_bh_to_extent()
3283          *           mpage_da_map_blocks()
3284          *
3285          * The problem is that write_cache_pages(), located in
3286          * mm/page-writeback.c, marks pages clean in preparation for
3287          * doing I/O, which is not desirable if we're not planning on
3288          * doing I/O at all.
3289          *
3290          * We could call write_cache_pages(), and then redirty all of
3291          * the pages by calling redirty_page_for_writeback() but that
3292          * would be ugly in the extreme.  So instead we would need to
3293          * replicate parts of the code in the above functions,
3294          * simplifying them becuase we wouldn't actually intend to
3295          * write out the pages, but rather only collect contiguous
3296          * logical block extents, call the multi-block allocator, and
3297          * then update the buffer heads with the block allocations.
3298          *
3299          * For now, though, we'll cheat by calling filemap_flush(),
3300          * which will map the blocks, and start the I/O, but not
3301          * actually wait for the I/O to complete.
3302          */
3303         return filemap_flush(inode->i_mapping);
3304 }
3305
3306 /*
3307  * bmap() is special.  It gets used by applications such as lilo and by
3308  * the swapper to find the on-disk block of a specific piece of data.
3309  *
3310  * Naturally, this is dangerous if the block concerned is still in the
3311  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3312  * filesystem and enables swap, then they may get a nasty shock when the
3313  * data getting swapped to that swapfile suddenly gets overwritten by
3314  * the original zero's written out previously to the journal and
3315  * awaiting writeback in the kernel's buffer cache.
3316  *
3317  * So, if we see any bmap calls here on a modified, data-journaled file,
3318  * take extra steps to flush any blocks which might be in the cache.
3319  */
3320 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3321 {
3322         struct inode *inode = mapping->host;
3323         journal_t *journal;
3324         int err;
3325
3326         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3327                         test_opt(inode->i_sb, DELALLOC)) {
3328                 /*
3329                  * With delalloc we want to sync the file
3330                  * so that we can make sure we allocate
3331                  * blocks for file
3332                  */
3333                 filemap_write_and_wait(mapping);
3334         }
3335
3336         if (EXT4_JOURNAL(inode) &&
3337             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3338                 /*
3339                  * This is a REALLY heavyweight approach, but the use of
3340                  * bmap on dirty files is expected to be extremely rare:
3341                  * only if we run lilo or swapon on a freshly made file
3342                  * do we expect this to happen.
3343                  *
3344                  * (bmap requires CAP_SYS_RAWIO so this does not
3345                  * represent an unprivileged user DOS attack --- we'd be
3346                  * in trouble if mortal users could trigger this path at
3347                  * will.)
3348                  *
3349                  * NB. EXT4_STATE_JDATA is not set on files other than
3350                  * regular files.  If somebody wants to bmap a directory
3351                  * or symlink and gets confused because the buffer
3352                  * hasn't yet been flushed to disk, they deserve
3353                  * everything they get.
3354                  */
3355
3356                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3357                 journal = EXT4_JOURNAL(inode);
3358                 jbd2_journal_lock_updates(journal);
3359                 err = jbd2_journal_flush(journal);
3360                 jbd2_journal_unlock_updates(journal);
3361
3362                 if (err)
3363                         return 0;
3364         }
3365
3366         return generic_block_bmap(mapping, block, ext4_get_block);
3367 }
3368
3369 static int ext4_readpage(struct file *file, struct page *page)
3370 {
3371         return mpage_readpage(page, ext4_get_block);
3372 }
3373
3374 static int
3375 ext4_readpages(struct file *file, struct address_space *mapping,
3376                 struct list_head *pages, unsigned nr_pages)
3377 {
3378         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3379 }
3380
3381 static void ext4_free_io_end(ext4_io_end_t *io)
3382 {
3383         BUG_ON(!io);
3384         if (io->page)
3385                 put_page(io->page);
3386         iput(io->inode);
3387         kfree(io);
3388 }
3389
3390 static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
3391 {
3392         struct buffer_head *head, *bh;
3393         unsigned int curr_off = 0;
3394
3395         if (!page_has_buffers(page))
3396                 return;
3397         head = bh = page_buffers(page);
3398         do {
3399                 if (offset <= curr_off && test_clear_buffer_uninit(bh)
3400                                         && bh->b_private) {
3401                         ext4_free_io_end(bh->b_private);
3402                         bh->b_private = NULL;
3403                         bh->b_end_io = NULL;
3404                 }
3405                 curr_off = curr_off + bh->b_size;
3406                 bh = bh->b_this_page;
3407         } while (bh != head);
3408 }
3409
3410 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3411 {
3412         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3413
3414         /*
3415          * free any io_end structure allocated for buffers to be discarded
3416          */
3417         if (ext4_should_dioread_nolock(page->mapping->host))
3418                 ext4_invalidatepage_free_endio(page, offset);
3419         /*
3420          * If it's a full truncate we just forget about the pending dirtying
3421          */
3422         if (offset == 0)
3423                 ClearPageChecked(page);
3424
3425         if (journal)
3426                 jbd2_journal_invalidatepage(journal, page, offset);
3427         else
3428                 block_invalidatepage(page, offset);
3429 }
3430
3431 static int ext4_releasepage(struct page *page, gfp_t wait)
3432 {
3433         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3434
3435         WARN_ON(PageChecked(page));
3436         if (!page_has_buffers(page))
3437                 return 0;
3438         if (journal)
3439                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3440         else
3441                 return try_to_free_buffers(page);
3442 }
3443
3444 /*
3445  * O_DIRECT for ext3 (or indirect map) based files
3446  *
3447  * If the O_DIRECT write will extend the file then add this inode to the
3448  * orphan list.  So recovery will truncate it back to the original size
3449  * if the machine crashes during the write.
3450  *
3451  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3452  * crashes then stale disk data _may_ be exposed inside the file. But current
3453  * VFS code falls back into buffered path in that case so we are safe.
3454  */
3455 static ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
3456                               const struct iovec *iov, loff_t offset,
3457                               unsigned long nr_segs)
3458 {
3459         struct file *file = iocb->ki_filp;
3460         struct inode *inode = file->f_mapping->host;
3461         struct ext4_inode_info *ei = EXT4_I(inode);
3462         handle_t *handle;
3463         ssize_t ret;
3464         int orphan = 0;
3465         size_t count = iov_length(iov, nr_segs);
3466         int retries = 0;
3467
3468         if (rw == WRITE) {
3469                 loff_t final_size = offset + count;
3470
3471                 if (final_size > inode->i_size) {
3472                         /* Credits for sb + inode write */
3473                         handle = ext4_journal_start(inode, 2);
3474                         if (IS_ERR(handle)) {
3475                                 ret = PTR_ERR(handle);
3476                                 goto out;
3477                         }
3478                         ret = ext4_orphan_add(handle, inode);
3479                         if (ret) {
3480                                 ext4_journal_stop(handle);
3481                                 goto out;
3482                         }
3483                         orphan = 1;
3484                         ei->i_disksize = inode->i_size;
3485                         ext4_journal_stop(handle);
3486                 }
3487         }
3488
3489 retry:
3490         if (rw == READ && ext4_should_dioread_nolock(inode))
3491                 ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
3492                                  inode->i_sb->s_bdev, iov,
3493                                  offset, nr_segs,
3494                                  ext4_get_block, NULL);
3495         else
3496                 ret = blockdev_direct_IO(rw, iocb, inode,
3497                                  inode->i_sb->s_bdev, iov,
3498                                  offset, nr_segs,
3499                                  ext4_get_block, NULL);
3500         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3501                 goto retry;
3502
3503         if (orphan) {
3504                 int err;
3505
3506                 /* Credits for sb + inode write */
3507                 handle = ext4_journal_start(inode, 2);
3508                 if (IS_ERR(handle)) {
3509                         /* This is really bad luck. We've written the data
3510                          * but cannot extend i_size. Bail out and pretend
3511                          * the write failed... */
3512                         ret = PTR_ERR(handle);
3513                         if (inode->i_nlink)
3514                                 ext4_orphan_del(NULL, inode);
3515
3516                         goto out;
3517                 }
3518                 if (inode->i_nlink)
3519                         ext4_orphan_del(handle, inode);
3520                 if (ret > 0) {
3521                         loff_t end = offset + ret;
3522                         if (end > inode->i_size) {
3523                                 ei->i_disksize = end;
3524                                 i_size_write(inode, end);
3525                                 /*
3526                                  * We're going to return a positive `ret'
3527                                  * here due to non-zero-length I/O, so there's
3528                                  * no way of reporting error returns from
3529                                  * ext4_mark_inode_dirty() to userspace.  So
3530                                  * ignore it.
3531                                  */
3532                                 ext4_mark_inode_dirty(handle, inode);
3533                         }
3534                 }
3535                 err = ext4_journal_stop(handle);
3536                 if (ret == 0)
3537                         ret = err;
3538         }
3539 out:
3540         return ret;
3541 }
3542
3543 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
3544                    struct buffer_head *bh_result, int create)
3545 {
3546         handle_t *handle = ext4_journal_current_handle();
3547         int ret = 0;
3548         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
3549         int dio_credits;
3550         int started = 0;
3551
3552         ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
3553                    inode->i_ino, create);
3554         /*
3555          * ext4_get_block in prepare for a DIO write or buffer write.
3556          * We allocate an uinitialized extent if blocks haven't been allocated.
3557          * The extent will be converted to initialized after IO complete.
3558          */
3559         create = EXT4_GET_BLOCKS_IO_CREATE_EXT;
3560
3561         if (!handle) {
3562                 if (max_blocks > DIO_MAX_BLOCKS)
3563                         max_blocks = DIO_MAX_BLOCKS;
3564                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
3565                 handle = ext4_journal_start(inode, dio_credits);
3566                 if (IS_ERR(handle)) {
3567                         ret = PTR_ERR(handle);
3568                         goto out;
3569                 }
3570                 started = 1;
3571         }
3572
3573         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
3574                               create);
3575         if (ret > 0) {
3576                 bh_result->b_size = (ret << inode->i_blkbits);
3577                 ret = 0;
3578         }
3579         if (started)
3580                 ext4_journal_stop(handle);
3581 out:
3582         return ret;
3583 }
3584
3585 static void dump_completed_IO(struct inode * inode)
3586 {
3587 #ifdef  EXT4_DEBUG
3588         struct list_head *cur, *before, *after;
3589         ext4_io_end_t *io, *io0, *io1;
3590         unsigned long flags;
3591
3592         if (list_empty(&EXT4_I(inode)->i_completed_io_list)){
3593                 ext4_debug("inode %lu completed_io list is empty\n", inode->i_ino);
3594                 return;
3595         }
3596
3597         ext4_debug("Dump inode %lu completed_io list \n", inode->i_ino);
3598         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3599         list_for_each_entry(io, &EXT4_I(inode)->i_completed_io_list, list){
3600                 cur = &io->list;
3601                 before = cur->prev;
3602                 io0 = container_of(before, ext4_io_end_t, list);
3603                 after = cur->next;
3604                 io1 = container_of(after, ext4_io_end_t, list);
3605
3606                 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
3607                             io, inode->i_ino, io0, io1);
3608         }
3609         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3610 #endif
3611 }
3612
3613 /*
3614  * check a range of space and convert unwritten extents to written.
3615  */
3616 static int ext4_end_io_nolock(ext4_io_end_t *io)
3617 {
3618         struct inode *inode = io->inode;
3619         loff_t offset = io->offset;
3620         ssize_t size = io->size;
3621         int ret = 0;
3622
3623         ext4_debug("ext4_end_io_nolock: io 0x%p from inode %lu,list->next 0x%p,"
3624                    "list->prev 0x%p\n",
3625                    io, inode->i_ino, io->list.next, io->list.prev);
3626
3627         if (list_empty(&io->list))
3628                 return ret;
3629
3630         if (io->flag != EXT4_IO_UNWRITTEN)
3631                 return ret;
3632
3633         ret = ext4_convert_unwritten_extents(inode, offset, size);
3634         if (ret < 0) {
3635                 printk(KERN_EMERG "%s: failed to convert unwritten"
3636                         "extents to written extents, error is %d"
3637                         " io is still on inode %lu aio dio list\n",
3638                        __func__, ret, inode->i_ino);
3639                 return ret;
3640         }
3641
3642         /* clear the DIO AIO unwritten flag */
3643         io->flag = 0;
3644         return ret;
3645 }
3646
3647 /*
3648  * work on completed aio dio IO, to convert unwritten extents to extents
3649  */
3650 static void ext4_end_io_work(struct work_struct *work)
3651 {
3652         ext4_io_end_t           *io = container_of(work, ext4_io_end_t, work);
3653         struct inode            *inode = io->inode;
3654         struct ext4_inode_info  *ei = EXT4_I(inode);
3655         unsigned long           flags;
3656         int                     ret;
3657
3658         mutex_lock(&inode->i_mutex);
3659         ret = ext4_end_io_nolock(io);
3660         if (ret < 0) {
3661                 mutex_unlock(&inode->i_mutex);
3662                 return;
3663         }
3664
3665         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3666         if (!list_empty(&io->list))
3667                 list_del_init(&io->list);
3668         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3669         mutex_unlock(&inode->i_mutex);
3670         ext4_free_io_end(io);
3671 }
3672
3673 /*
3674  * This function is called from ext4_sync_file().
3675  *
3676  * When IO is completed, the work to convert unwritten extents to
3677  * written is queued on workqueue but may not get immediately
3678  * scheduled. When fsync is called, we need to ensure the
3679  * conversion is complete before fsync returns.
3680  * The inode keeps track of a list of pending/completed IO that
3681  * might needs to do the conversion. This function walks through
3682  * the list and convert the related unwritten extents for completed IO
3683  * to written.
3684  * The function return the number of pending IOs on success.
3685  */
3686 int flush_completed_IO(struct inode *inode)
3687 {
3688         ext4_io_end_t *io;
3689         struct ext4_inode_info *ei = EXT4_I(inode);
3690         unsigned long flags;
3691         int ret = 0;
3692         int ret2 = 0;
3693
3694         if (list_empty(&ei->i_completed_io_list))
3695                 return ret;
3696
3697         dump_completed_IO(inode);
3698         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3699         while (!list_empty(&ei->i_completed_io_list)){
3700                 io = list_entry(ei->i_completed_io_list.next,
3701                                 ext4_io_end_t, list);
3702                 /*
3703                  * Calling ext4_end_io_nolock() to convert completed
3704                  * IO to written.
3705                  *
3706                  * When ext4_sync_file() is called, run_queue() may already
3707                  * about to flush the work corresponding to this io structure.
3708                  * It will be upset if it founds the io structure related
3709                  * to the work-to-be schedule is freed.
3710                  *
3711                  * Thus we need to keep the io structure still valid here after
3712                  * convertion finished. The io structure has a flag to
3713                  * avoid double converting from both fsync and background work
3714                  * queue work.
3715                  */
3716                 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3717                 ret = ext4_end_io_nolock(io);
3718                 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3719                 if (ret < 0)
3720                         ret2 = ret;
3721                 else
3722                         list_del_init(&io->list);
3723         }
3724         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3725         return (ret2 < 0) ? ret2 : 0;
3726 }
3727
3728 static ext4_io_end_t *ext4_init_io_end (struct inode *inode, gfp_t flags)
3729 {
3730         ext4_io_end_t *io = NULL;
3731
3732         io = kmalloc(sizeof(*io), flags);
3733
3734         if (io) {
3735                 igrab(inode);
3736                 io->inode = inode;
3737                 io->flag = 0;
3738                 io->offset = 0;
3739                 io->size = 0;
3740                 io->page = NULL;
3741                 INIT_WORK(&io->work, ext4_end_io_work);
3742                 INIT_LIST_HEAD(&io->list);
3743         }
3744
3745         return io;
3746 }
3747
3748 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3749                             ssize_t size, void *private)
3750 {
3751         ext4_io_end_t *io_end = iocb->private;
3752         struct workqueue_struct *wq;
3753         unsigned long flags;
3754         struct ext4_inode_info *ei;
3755
3756         /* if not async direct IO or dio with 0 bytes write, just return */
3757         if (!io_end || !size)
3758                 return;
3759
3760         ext_debug("ext4_end_io_dio(): io_end 0x%p"
3761                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
3762                   iocb->private, io_end->inode->i_ino, iocb, offset,
3763                   size);
3764
3765         /* if not aio dio with unwritten extents, just free io and return */
3766         if (io_end->flag != EXT4_IO_UNWRITTEN){
3767                 ext4_free_io_end(io_end);
3768                 iocb->private = NULL;
3769                 return;
3770         }
3771
3772         io_end->offset = offset;
3773         io_end->size = size;
3774         io_end->flag = EXT4_IO_UNWRITTEN;
3775         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
3776
3777         /* queue the work to convert unwritten extents to written */
3778         queue_work(wq, &io_end->work);
3779
3780         /* Add the io_end to per-inode completed aio dio list*/
3781         ei = EXT4_I(io_end->inode);
3782         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3783         list_add_tail(&io_end->list, &ei->i_completed_io_list);
3784         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3785         iocb->private = NULL;
3786 }
3787
3788 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
3789 {
3790         ext4_io_end_t *io_end = bh->b_private;
3791         struct workqueue_struct *wq;
3792         struct inode *inode;
3793         unsigned long flags;
3794
3795         if (!test_clear_buffer_uninit(bh) || !io_end)
3796                 goto out;
3797
3798         if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
3799                 printk("sb umounted, discard end_io request for inode %lu\n",
3800                         io_end->inode->i_ino);
3801                 ext4_free_io_end(io_end);
3802                 goto out;
3803         }
3804
3805         io_end->flag = EXT4_IO_UNWRITTEN;
3806         inode = io_end->inode;
3807
3808         /* Add the io_end to per-inode completed io list*/
3809         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3810         list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
3811         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3812
3813         wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
3814         /* queue the work to convert unwritten extents to written */
3815         queue_work(wq, &io_end->work);
3816 out:
3817         bh->b_private = NULL;
3818         bh->b_end_io = NULL;
3819         clear_buffer_uninit(bh);
3820         end_buffer_async_write(bh, uptodate);
3821 }
3822
3823 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
3824 {
3825         ext4_io_end_t *io_end;
3826         struct page *page = bh->b_page;
3827         loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
3828         size_t size = bh->b_size;
3829
3830 retry:
3831         io_end = ext4_init_io_end(inode, GFP_ATOMIC);
3832         if (!io_end) {
3833                 if (printk_ratelimit())
3834                         printk(KERN_WARNING "%s: allocation fail\n", __func__);
3835                 schedule();
3836                 goto retry;
3837         }
3838         io_end->offset = offset;
3839         io_end->size = size;
3840         /*
3841          * We need to hold a reference to the page to make sure it
3842          * doesn't get evicted before ext4_end_io_work() has a chance
3843          * to convert the extent from written to unwritten.
3844          */
3845         io_end->page = page;
3846         get_page(io_end->page);
3847
3848         bh->b_private = io_end;
3849         bh->b_end_io = ext4_end_io_buffer_write;
3850         return 0;
3851 }
3852
3853 /*
3854  * For ext4 extent files, ext4 will do direct-io write to holes,
3855  * preallocated extents, and those write extend the file, no need to
3856  * fall back to buffered IO.
3857  *
3858  * For holes, we fallocate those blocks, mark them as unintialized
3859  * If those blocks were preallocated, we mark sure they are splited, but
3860  * still keep the range to write as unintialized.
3861  *
3862  * The unwrritten extents will be converted to written when DIO is completed.
3863  * For async direct IO, since the IO may still pending when return, we
3864  * set up an end_io call back function, which will do the convertion
3865  * when async direct IO completed.
3866  *
3867  * If the O_DIRECT write will extend the file then add this inode to the
3868  * orphan list.  So recovery will truncate it back to the original size
3869  * if the machine crashes during the write.
3870  *
3871  */
3872 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
3873                               const struct iovec *iov, loff_t offset,
3874                               unsigned long nr_segs)
3875 {
3876         struct file *file = iocb->ki_filp;
3877         struct inode *inode = file->f_mapping->host;
3878         ssize_t ret;
3879         size_t count = iov_length(iov, nr_segs);
3880
3881         loff_t final_size = offset + count;
3882         if (rw == WRITE && final_size <= inode->i_size) {
3883                 /*
3884                  * We could direct write to holes and fallocate.
3885                  *
3886                  * Allocated blocks to fill the hole are marked as uninitialized
3887                  * to prevent paralel buffered read to expose the stale data
3888                  * before DIO complete the data IO.
3889                  *
3890                  * As to previously fallocated extents, ext4 get_block
3891                  * will just simply mark the buffer mapped but still
3892                  * keep the extents uninitialized.
3893                  *
3894                  * for non AIO case, we will convert those unwritten extents
3895                  * to written after return back from blockdev_direct_IO.
3896                  *
3897                  * for async DIO, the conversion needs to be defered when
3898                  * the IO is completed. The ext4 end_io callback function
3899                  * will be called to take care of the conversion work.
3900                  * Here for async case, we allocate an io_end structure to
3901                  * hook to the iocb.
3902                  */
3903                 iocb->private = NULL;
3904                 EXT4_I(inode)->cur_aio_dio = NULL;
3905                 if (!is_sync_kiocb(iocb)) {
3906                         iocb->private = ext4_init_io_end(inode, GFP_NOFS);
3907                         if (!iocb->private)
3908                                 return -ENOMEM;
3909                         /*
3910                          * we save the io structure for current async
3911                          * direct IO, so that later ext4_get_blocks()
3912                          * could flag the io structure whether there
3913                          * is a unwritten extents needs to be converted
3914                          * when IO is completed.
3915                          */
3916                         EXT4_I(inode)->cur_aio_dio = iocb->private;
3917                 }
3918
3919                 ret = blockdev_direct_IO(rw, iocb, inode,
3920                                          inode->i_sb->s_bdev, iov,
3921                                          offset, nr_segs,
3922                                          ext4_get_block_write,
3923                                          ext4_end_io_dio);
3924                 if (iocb->private)
3925                         EXT4_I(inode)->cur_aio_dio = NULL;
3926                 /*
3927                  * The io_end structure takes a reference to the inode,
3928                  * that structure needs to be destroyed and the
3929                  * reference to the inode need to be dropped, when IO is
3930                  * complete, even with 0 byte write, or failed.
3931                  *
3932                  * In the successful AIO DIO case, the io_end structure will be
3933                  * desctroyed and the reference to the inode will be dropped
3934                  * after the end_io call back function is called.
3935                  *
3936                  * In the case there is 0 byte write, or error case, since
3937                  * VFS direct IO won't invoke the end_io call back function,
3938                  * we need to free the end_io structure here.
3939                  */
3940                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
3941                         ext4_free_io_end(iocb->private);
3942                         iocb->private = NULL;
3943                 } else if (ret > 0 && ext4_test_inode_state(inode,
3944                                                 EXT4_STATE_DIO_UNWRITTEN)) {
3945                         int err;
3946                         /*
3947                          * for non AIO case, since the IO is already
3948                          * completed, we could do the convertion right here
3949                          */
3950                         err = ext4_convert_unwritten_extents(inode,
3951                                                              offset, ret);
3952                         if (err < 0)
3953                                 ret = err;
3954                         ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3955                 }
3956                 return ret;
3957         }
3958
3959         /* for write the the end of file case, we fall back to old way */
3960         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3961 }
3962
3963 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3964                               const struct iovec *iov, loff_t offset,
3965                               unsigned long nr_segs)
3966 {
3967         struct file *file = iocb->ki_filp;
3968         struct inode *inode = file->f_mapping->host;
3969
3970         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
3971                 return ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
3972
3973         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3974 }
3975
3976 /*
3977  * Pages can be marked dirty completely asynchronously from ext4's journalling
3978  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3979  * much here because ->set_page_dirty is called under VFS locks.  The page is
3980  * not necessarily locked.
3981  *
3982  * We cannot just dirty the page and leave attached buffers clean, because the
3983  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3984  * or jbddirty because all the journalling code will explode.
3985  *
3986  * So what we do is to mark the page "pending dirty" and next time writepage
3987  * is called, propagate that into the buffers appropriately.
3988  */
3989 static int ext4_journalled_set_page_dirty(struct page *page)
3990 {
3991         SetPageChecked(page);
3992         return __set_page_dirty_nobuffers(page);
3993 }
3994
3995 static const struct address_space_operations ext4_ordered_aops = {
3996         .readpage               = ext4_readpage,
3997         .readpages              = ext4_readpages,
3998         .writepage              = ext4_writepage,
3999         .sync_page              = block_sync_page,
4000         .write_begin            = ext4_write_begin,
4001         .write_end              = ext4_ordered_write_end,
4002         .bmap                   = ext4_bmap,
4003         .invalidatepage         = ext4_invalidatepage,
4004         .releasepage            = ext4_releasepage,
4005         .direct_IO              = ext4_direct_IO,
4006         .migratepage            = buffer_migrate_page,
4007         .is_partially_uptodate  = block_is_partially_uptodate,
4008         .error_remove_page      = generic_error_remove_page,
4009 };
4010
4011 static const struct address_space_operations ext4_writeback_aops = {
4012         .readpage               = ext4_readpage,
4013         .readpages              = ext4_readpages,
4014         .writepage              = ext4_writepage,
4015         .sync_page              = block_sync_page,
4016         .write_begin            = ext4_write_begin,
4017         .write_end              = ext4_writeback_write_end,
4018         .bmap                   = ext4_bmap,
4019         .invalidatepage         = ext4_invalidatepage,
4020         .releasepage            = ext4_releasepage,
4021         .direct_IO              = ext4_direct_IO,
4022         .migratepage            = buffer_migrate_page,
4023         .is_partially_uptodate  = block_is_partially_uptodate,
4024         .error_remove_page      = generic_error_remove_page,
4025 };
4026
4027 static const struct address_space_operations ext4_journalled_aops = {
4028         .readpage               = ext4_readpage,
4029         .readpages              = ext4_readpages,
4030         .writepage              = ext4_writepage,
4031         .sync_page              = block_sync_page,
4032         .write_begin            = ext4_write_begin,
4033         .write_end              = ext4_journalled_write_end,
4034         .set_page_dirty         = ext4_journalled_set_page_dirty,
4035         .bmap                   = ext4_bmap,
4036         .invalidatepage         = ext4_invalidatepage,
4037         .releasepage            = ext4_releasepage,
4038         .is_partially_uptodate  = block_is_partially_uptodate,
4039         .error_remove_page      = generic_error_remove_page,
4040 };
4041
4042 static const struct address_space_operations ext4_da_aops = {
4043         .readpage               = ext4_readpage,
4044         .readpages              = ext4_readpages,
4045         .writepage              = ext4_writepage,
4046         .writepages             = ext4_da_writepages,
4047         .sync_page              = block_sync_page,
4048         .write_begin            = ext4_da_write_begin,
4049         .write_end              = ext4_da_write_end,
4050         .bmap                   = ext4_bmap,
4051         .invalidatepage         = ext4_da_invalidatepage,
4052         .releasepage            = ext4_releasepage,
4053         .direct_IO              = ext4_direct_IO,
4054         .migratepage            = buffer_migrate_page,
4055         .is_partially_uptodate  = block_is_partially_uptodate,
4056         .error_remove_page      = generic_error_remove_page,
4057 };
4058
4059 void ext4_set_aops(struct inode *inode)
4060 {
4061         if (ext4_should_order_data(inode) &&
4062                 test_opt(inode->i_sb, DELALLOC))
4063                 inode->i_mapping->a_ops = &ext4_da_aops;
4064         else if (ext4_should_order_data(inode))
4065                 inode->i_mapping->a_ops = &ext4_ordered_aops;
4066         else if (ext4_should_writeback_data(inode) &&
4067                  test_opt(inode->i_sb, DELALLOC))
4068                 inode->i_mapping->a_ops = &ext4_da_aops;
4069         else if (ext4_should_writeback_data(inode))
4070                 inode->i_mapping->a_ops = &ext4_writeback_aops;
4071         else
4072                 inode->i_mapping->a_ops = &ext4_journalled_aops;
4073 }
4074
4075 /*
4076  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
4077  * up to the end of the block which corresponds to `from'.
4078  * This required during truncate. We need to physically zero the tail end
4079  * of that block so it doesn't yield old data if the file is later grown.
4080  */
4081 int ext4_block_truncate_page(handle_t *handle,
4082                 struct address_space *mapping, loff_t from)
4083 {
4084         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
4085         unsigned offset = from & (PAGE_CACHE_SIZE-1);
4086         unsigned blocksize, length, pos;
4087         ext4_lblk_t iblock;
4088         struct inode *inode = mapping->host;
4089         struct buffer_head *bh;
4090         struct page *page;
4091         int err = 0;
4092
4093         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
4094                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
4095         if (!page)
4096                 return -EINVAL;
4097
4098         blocksize = inode->i_sb->s_blocksize;
4099         length = blocksize - (offset & (blocksize - 1));
4100         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
4101
4102         /*
4103          * For "nobh" option,  we can only work if we don't need to
4104          * read-in the page - otherwise we create buffers to do the IO.
4105          */
4106         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
4107              ext4_should_writeback_data(inode) && PageUptodate(page)) {
4108                 zero_user(page, offset, length);
4109                 set_page_dirty(page);
4110                 goto unlock;
4111         }
4112
4113         if (!page_has_buffers(page))
4114                 create_empty_buffers(page, blocksize, 0);
4115
4116         /* Find the buffer that contains "offset" */
4117         bh = page_buffers(page);
4118         pos = blocksize;
4119         while (offset >= pos) {
4120                 bh = bh->b_this_page;
4121                 iblock++;
4122                 pos += blocksize;
4123         }
4124
4125         err = 0;
4126         if (buffer_freed(bh)) {
4127                 BUFFER_TRACE(bh, "freed: skip");
4128                 goto unlock;
4129         }
4130
4131         if (!buffer_mapped(bh)) {
4132                 BUFFER_TRACE(bh, "unmapped");
4133                 ext4_get_block(inode, iblock, bh, 0);
4134                 /* unmapped? It's a hole - nothing to do */
4135                 if (!buffer_mapped(bh)) {
4136                         BUFFER_TRACE(bh, "still unmapped");
4137                         goto unlock;
4138                 }
4139         }
4140
4141         /* Ok, it's mapped. Make sure it's up-to-date */
4142         if (PageUptodate(page))
4143                 set_buffer_uptodate(bh);
4144
4145         if (!buffer_uptodate(bh)) {
4146                 err = -EIO;
4147                 ll_rw_block(READ, 1, &bh);
4148                 wait_on_buffer(bh);
4149                 /* Uhhuh. Read error. Complain and punt. */
4150                 if (!buffer_uptodate(bh))
4151                         goto unlock;
4152         }
4153
4154         if (ext4_should_journal_data(inode)) {
4155                 BUFFER_TRACE(bh, "get write access");
4156                 err = ext4_journal_get_write_access(handle, bh);
4157                 if (err)
4158                         goto unlock;
4159         }
4160
4161         zero_user(page, offset, length);
4162
4163         BUFFER_TRACE(bh, "zeroed end of block");
4164
4165         err = 0;
4166         if (ext4_should_journal_data(inode)) {
4167                 err = ext4_handle_dirty_metadata(handle, inode, bh);
4168         } else {
4169                 if (ext4_should_order_data(inode))
4170                         err = ext4_jbd2_file_inode(handle, inode);
4171                 mark_buffer_dirty(bh);
4172         }
4173
4174 unlock:
4175         unlock_page(page);
4176         page_cache_release(page);
4177         return err;
4178 }
4179
4180 /*
4181  * Probably it should be a library function... search for first non-zero word
4182  * or memcmp with zero_page, whatever is better for particular architecture.
4183  * Linus?
4184  */
4185 static inline int all_zeroes(__le32 *p, __le32 *q)
4186 {
4187         while (p < q)
4188                 if (*p++)
4189                         return 0;
4190         return 1;
4191 }
4192
4193 /**
4194  *      ext4_find_shared - find the indirect blocks for partial truncation.
4195  *      @inode:   inode in question
4196  *      @depth:   depth of the affected branch
4197  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
4198  *      @chain:   place to store the pointers to partial indirect blocks
4199  *      @top:     place to the (detached) top of branch
4200  *
4201  *      This is a helper function used by ext4_truncate().
4202  *
4203  *      When we do truncate() we may have to clean the ends of several
4204  *      indirect blocks but leave the blocks themselves alive. Block is
4205  *      partially truncated if some data below the new i_size is refered
4206  *      from it (and it is on the path to the first completely truncated
4207  *      data block, indeed).  We have to free the top of that path along
4208  *      with everything to the right of the path. Since no allocation
4209  *      past the truncation point is possible until ext4_truncate()
4210  *      finishes, we may safely do the latter, but top of branch may
4211  *      require special attention - pageout below the truncation point
4212  *      might try to populate it.
4213  *
4214  *      We atomically detach the top of branch from the tree, store the
4215  *      block number of its root in *@top, pointers to buffer_heads of
4216  *      partially truncated blocks - in @chain[].bh and pointers to
4217  *      their last elements that should not be removed - in
4218  *      @chain[].p. Return value is the pointer to last filled element
4219  *      of @chain.
4220  *
4221  *      The work left to caller to do the actual freeing of subtrees:
4222  *              a) free the subtree starting from *@top
4223  *              b) free the subtrees whose roots are stored in
4224  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
4225  *              c) free the subtrees growing from the inode past the @chain[0].
4226  *                      (no partially truncated stuff there).  */
4227
4228 static Indirect *ext4_find_shared(struct inode *inode, int depth,
4229                                   ext4_lblk_t offsets[4], Indirect chain[4],
4230                                   __le32 *top)
4231 {
4232         Indirect *partial, *p;
4233         int k, err;
4234
4235         *top = 0;
4236         /* Make k index the deepest non-null offset + 1 */
4237         for (k = depth; k > 1 && !offsets[k-1]; k--)
4238                 ;
4239         partial = ext4_get_branch(inode, k, offsets, chain, &err);
4240         /* Writer: pointers */
4241         if (!partial)
4242                 partial = chain + k-1;
4243         /*
4244          * If the branch acquired continuation since we've looked at it -
4245          * fine, it should all survive and (new) top doesn't belong to us.
4246          */
4247         if (!partial->key && *partial->p)
4248                 /* Writer: end */
4249                 goto no_top;
4250         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
4251                 ;
4252         /*
4253          * OK, we've found the last block that must survive. The rest of our
4254          * branch should be detached before unlocking. However, if that rest
4255          * of branch is all ours and does not grow immediately from the inode
4256          * it's easier to cheat and just decrement partial->p.
4257          */
4258         if (p == chain + k - 1 && p > chain) {
4259                 p->p--;
4260         } else {
4261                 *top = *p->p;
4262                 /* Nope, don't do this in ext4.  Must leave the tree intact */
4263 #if 0
4264                 *p->p = 0;
4265 #endif
4266         }
4267         /* Writer: end */
4268
4269         while (partial > p) {
4270                 brelse(partial->bh);
4271                 partial--;
4272         }
4273 no_top:
4274         return partial;
4275 }
4276
4277 /*
4278  * Zero a number of block pointers in either an inode or an indirect block.
4279  * If we restart the transaction we must again get write access to the
4280  * indirect block for further modification.
4281  *
4282  * We release `count' blocks on disk, but (last - first) may be greater
4283  * than `count' because there can be holes in there.
4284  */
4285 static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
4286                              struct buffer_head *bh,
4287                              ext4_fsblk_t block_to_free,
4288                              unsigned long count, __le32 *first,
4289                              __le32 *last)
4290 {
4291         __le32 *p;
4292         int     flags = EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_VALIDATED;
4293
4294         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
4295                 flags |= EXT4_FREE_BLOCKS_METADATA;
4296
4297         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
4298                                    count)) {
4299                 ext4_error(inode->i_sb, "inode #%lu: "
4300                            "attempt to clear blocks %llu len %lu, invalid",
4301                            inode->i_ino, (unsigned long long) block_to_free,
4302                            count);
4303                 return 1;
4304         }
4305
4306         if (try_to_extend_transaction(handle, inode)) {
4307                 if (bh) {
4308                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4309                         ext4_handle_dirty_metadata(handle, inode, bh);
4310                 }
4311                 ext4_mark_inode_dirty(handle, inode);
4312                 ext4_truncate_restart_trans(handle, inode,
4313                                             blocks_for_truncate(inode));
4314                 if (bh) {
4315                         BUFFER_TRACE(bh, "retaking write access");
4316                         ext4_journal_get_write_access(handle, bh);
4317                 }
4318         }
4319
4320         for (p = first; p < last; p++)
4321                 *p = 0;
4322
4323         ext4_free_blocks(handle, inode, 0, block_to_free, count, flags);
4324         return 0;
4325 }
4326
4327 /**
4328  * ext4_free_data - free a list of data blocks
4329  * @handle:     handle for this transaction
4330  * @inode:      inode we are dealing with
4331  * @this_bh:    indirect buffer_head which contains *@first and *@last
4332  * @first:      array of block numbers
4333  * @last:       points immediately past the end of array
4334  *
4335  * We are freeing all blocks refered from that array (numbers are stored as
4336  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
4337  *
4338  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
4339  * blocks are contiguous then releasing them at one time will only affect one
4340  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
4341  * actually use a lot of journal space.
4342  *
4343  * @this_bh will be %NULL if @first and @last point into the inode's direct
4344  * block pointers.
4345  */
4346 static void ext4_free_data(handle_t *handle, struct inode *inode,
4347                            struct buffer_head *this_bh,
4348                            __le32 *first, __le32 *last)
4349 {
4350         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
4351         unsigned long count = 0;            /* Number of blocks in the run */
4352         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
4353                                                corresponding to
4354                                                block_to_free */
4355         ext4_fsblk_t nr;                    /* Current block # */
4356         __le32 *p;                          /* Pointer into inode/ind
4357                                                for current block */
4358         int err;
4359
4360         if (this_bh) {                          /* For indirect block */
4361                 BUFFER_TRACE(this_bh, "get_write_access");
4362                 err = ext4_journal_get_write_access(handle, this_bh);
4363                 /* Important: if we can't update the indirect pointers
4364                  * to the blocks, we can't free them. */
4365                 if (err)
4366                         return;
4367         }
4368
4369         for (p = first; p < last; p++) {
4370                 nr = le32_to_cpu(*p);
4371                 if (nr) {
4372                         /* accumulate blocks to free if they're contiguous */
4373                         if (count == 0) {
4374                                 block_to_free = nr;
4375                                 block_to_free_p = p;
4376                                 count = 1;
4377                         } else if (nr == block_to_free + count) {
4378                                 count++;
4379                         } else {
4380                                 if (ext4_clear_blocks(handle, inode, this_bh,
4381                                                       block_to_free, count,
4382                                                       block_to_free_p, p))
4383                                         break;
4384                                 block_to_free = nr;
4385                                 block_to_free_p = p;
4386                                 count = 1;
4387                         }
4388                 }
4389         }
4390
4391         if (count > 0)
4392                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
4393                                   count, block_to_free_p, p);
4394
4395         if (this_bh) {
4396                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
4397
4398                 /*
4399                  * The buffer head should have an attached journal head at this
4400                  * point. However, if the data is corrupted and an indirect
4401                  * block pointed to itself, it would have been detached when
4402                  * the block was cleared. Check for this instead of OOPSing.
4403                  */
4404                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
4405                         ext4_handle_dirty_metadata(handle, inode, this_bh);
4406                 else
4407                         ext4_error(inode->i_sb,
4408                                    "circular indirect block detected, "
4409                                    "inode=%lu, block=%llu",
4410                                    inode->i_ino,
4411                                    (unsigned long long) this_bh->b_blocknr);
4412         }
4413 }
4414
4415 /**
4416  *      ext4_free_branches - free an array of branches
4417  *      @handle: JBD handle for this transaction
4418  *      @inode: inode we are dealing with
4419  *      @parent_bh: the buffer_head which contains *@first and *@last
4420  *      @first: array of block numbers
4421  *      @last:  pointer immediately past the end of array
4422  *      @depth: depth of the branches to free
4423  *
4424  *      We are freeing all blocks refered from these branches (numbers are
4425  *      stored as little-endian 32-bit) and updating @inode->i_blocks
4426  *      appropriately.
4427  */
4428 static void ext4_free_branches(handle_t *handle, struct inode *inode,
4429                                struct buffer_head *parent_bh,
4430                                __le32 *first, __le32 *last, int depth)
4431 {
4432         ext4_fsblk_t nr;
4433         __le32 *p;
4434
4435         if (ext4_handle_is_aborted(handle))
4436                 return;
4437
4438         if (depth--) {
4439                 struct buffer_head *bh;
4440                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4441                 p = last;
4442                 while (--p >= first) {
4443                         nr = le32_to_cpu(*p);
4444                         if (!nr)
4445                                 continue;               /* A hole */
4446
4447                         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
4448                                                    nr, 1)) {
4449                                 ext4_error(inode->i_sb,
4450                                            "indirect mapped block in inode "
4451                                            "#%lu invalid (level %d, blk #%lu)",
4452                                            inode->i_ino, depth,
4453                                            (unsigned long) nr);
4454                                 break;
4455                         }
4456
4457                         /* Go read the buffer for the next level down */
4458                         bh = sb_bread(inode->i_sb, nr);
4459
4460                         /*
4461                          * A read failure? Report error and clear slot
4462                          * (should be rare).
4463                          */
4464                         if (!bh) {
4465                                 ext4_error(inode->i_sb,
4466                                            "Read failure, inode=%lu, block=%llu",
4467                                            inode->i_ino, nr);
4468                                 continue;
4469                         }
4470
4471                         /* This zaps the entire block.  Bottom up. */
4472                         BUFFER_TRACE(bh, "free child branches");
4473                         ext4_free_branches(handle, inode, bh,
4474                                         (__le32 *) bh->b_data,
4475                                         (__le32 *) bh->b_data + addr_per_block,
4476                                         depth);
4477
4478                         /*
4479                          * We've probably journalled the indirect block several
4480                          * times during the truncate.  But it's no longer
4481                          * needed and we now drop it from the transaction via
4482                          * jbd2_journal_revoke().
4483                          *
4484                          * That's easy if it's exclusively part of this
4485                          * transaction.  But if it's part of the committing
4486                          * transaction then jbd2_journal_forget() will simply
4487                          * brelse() it.  That means that if the underlying
4488                          * block is reallocated in ext4_get_block(),
4489                          * unmap_underlying_metadata() will find this block
4490                          * and will try to get rid of it.  damn, damn.
4491                          *
4492                          * If this block has already been committed to the
4493                          * journal, a revoke record will be written.  And
4494                          * revoke records must be emitted *before* clearing
4495                          * this block's bit in the bitmaps.
4496                          */
4497                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
4498
4499                         /*
4500                          * Everything below this this pointer has been
4501                          * released.  Now let this top-of-subtree go.
4502                          *
4503                          * We want the freeing of this indirect block to be
4504                          * atomic in the journal with the updating of the
4505                          * bitmap block which owns it.  So make some room in
4506                          * the journal.
4507                          *
4508                          * We zero the parent pointer *after* freeing its
4509                          * pointee in the bitmaps, so if extend_transaction()
4510                          * for some reason fails to put the bitmap changes and
4511                          * the release into the same transaction, recovery
4512                          * will merely complain about releasing a free block,
4513                          * rather than leaking blocks.
4514                          */
4515                         if (ext4_handle_is_aborted(handle))
4516                                 return;
4517                         if (try_to_extend_transaction(handle, inode)) {
4518                                 ext4_mark_inode_dirty(handle, inode);
4519                                 ext4_truncate_restart_trans(handle, inode,
4520                                             blocks_for_truncate(inode));
4521                         }
4522
4523                         ext4_free_blocks(handle, inode, 0, nr, 1,
4524                                          EXT4_FREE_BLOCKS_METADATA);
4525
4526                         if (parent_bh) {
4527                                 /*
4528                                  * The block which we have just freed is
4529                                  * pointed to by an indirect block: journal it
4530                                  */
4531                                 BUFFER_TRACE(parent_bh, "get_write_access");
4532                                 if (!ext4_journal_get_write_access(handle,
4533                                                                    parent_bh)){
4534                                         *p = 0;
4535                                         BUFFER_TRACE(parent_bh,
4536                                         "call ext4_handle_dirty_metadata");
4537                                         ext4_handle_dirty_metadata(handle,
4538                                                                    inode,
4539                                                                    parent_bh);
4540                                 }
4541                         }
4542                 }
4543         } else {
4544                 /* We have reached the bottom of the tree. */
4545                 BUFFER_TRACE(parent_bh, "free data blocks");
4546                 ext4_free_data(handle, inode, parent_bh, first, last);
4547         }
4548 }
4549
4550 int ext4_can_truncate(struct inode *inode)
4551 {
4552         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4553                 return 0;
4554         if (S_ISREG(inode->i_mode))
4555                 return 1;
4556         if (S_ISDIR(inode->i_mode))
4557                 return 1;
4558         if (S_ISLNK(inode->i_mode))
4559                 return !ext4_inode_is_fast_symlink(inode);
4560         return 0;
4561 }
4562
4563 /*
4564  * ext4_truncate()
4565  *
4566  * We block out ext4_get_block() block instantiations across the entire
4567  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4568  * simultaneously on behalf of the same inode.
4569  *
4570  * As we work through the truncate and commmit bits of it to the journal there
4571  * is one core, guiding principle: the file's tree must always be consistent on
4572  * disk.  We must be able to restart the truncate after a crash.
4573  *
4574  * The file's tree may be transiently inconsistent in memory (although it
4575  * probably isn't), but whenever we close off and commit a journal transaction,
4576  * the contents of (the filesystem + the journal) must be consistent and
4577  * restartable.  It's pretty simple, really: bottom up, right to left (although
4578  * left-to-right works OK too).
4579  *
4580  * Note that at recovery time, journal replay occurs *before* the restart of
4581  * truncate against the orphan inode list.
4582  *
4583  * The committed inode has the new, desired i_size (which is the same as
4584  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4585  * that this inode's truncate did not complete and it will again call
4586  * ext4_truncate() to have another go.  So there will be instantiated blocks
4587  * to the right of the truncation point in a crashed ext4 filesystem.  But
4588  * that's fine - as long as they are linked from the inode, the post-crash
4589  * ext4_truncate() run will find them and release them.
4590  */
4591 void ext4_truncate(struct inode *inode)
4592 {
4593         handle_t *handle;
4594         struct ext4_inode_info *ei = EXT4_I(inode);
4595         __le32 *i_data = ei->i_data;
4596         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4597         struct address_space *mapping = inode->i_mapping;
4598         ext4_lblk_t offsets[4];
4599         Indirect chain[4];
4600         Indirect *partial;
4601         __le32 nr = 0;
4602         int n;
4603         ext4_lblk_t last_block;
4604         unsigned blocksize = inode->i_sb->s_blocksize;
4605
4606         if (!ext4_can_truncate(inode))
4607                 return;
4608
4609         EXT4_I(inode)->i_flags &= ~EXT4_EOFBLOCKS_FL;
4610
4611         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4612                 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4613
4614         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
4615                 ext4_ext_truncate(inode);
4616                 return;
4617         }
4618
4619         handle = start_transaction(inode);
4620         if (IS_ERR(handle))
4621                 return;         /* AKPM: return what? */
4622
4623         last_block = (inode->i_size + blocksize-1)
4624                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4625
4626         if (inode->i_size & (blocksize - 1))
4627                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4628                         goto out_stop;
4629
4630         n = ext4_block_to_path(inode, last_block, offsets, NULL);
4631         if (n == 0)
4632                 goto out_stop;  /* error */
4633
4634         /*
4635          * OK.  This truncate is going to happen.  We add the inode to the
4636          * orphan list, so that if this truncate spans multiple transactions,
4637          * and we crash, we will resume the truncate when the filesystem
4638          * recovers.  It also marks the inode dirty, to catch the new size.
4639          *
4640          * Implication: the file must always be in a sane, consistent
4641          * truncatable state while each transaction commits.
4642          */
4643         if (ext4_orphan_add(handle, inode))
4644                 goto out_stop;
4645
4646         /*
4647          * From here we block out all ext4_get_block() callers who want to
4648          * modify the block allocation tree.
4649          */
4650         down_write(&ei->i_data_sem);
4651
4652         ext4_discard_preallocations(inode);
4653
4654         /*
4655          * The orphan list entry will now protect us from any crash which
4656          * occurs before the truncate completes, so it is now safe to propagate
4657          * the new, shorter inode size (held for now in i_size) into the
4658          * on-disk inode. We do this via i_disksize, which is the value which
4659          * ext4 *really* writes onto the disk inode.
4660          */
4661         ei->i_disksize = inode->i_size;
4662
4663         if (n == 1) {           /* direct blocks */
4664                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4665                                i_data + EXT4_NDIR_BLOCKS);
4666                 goto do_indirects;
4667         }
4668
4669         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4670         /* Kill the top of shared branch (not detached) */
4671         if (nr) {
4672                 if (partial == chain) {
4673                         /* Shared branch grows from the inode */
4674                         ext4_free_branches(handle, inode, NULL,
4675                                            &nr, &nr+1, (chain+n-1) - partial);
4676                         *partial->p = 0;
4677                         /*
4678                          * We mark the inode dirty prior to restart,
4679                          * and prior to stop.  No need for it here.
4680                          */
4681                 } else {
4682                         /* Shared branch grows from an indirect block */
4683                         BUFFER_TRACE(partial->bh, "get_write_access");
4684                         ext4_free_branches(handle, inode, partial->bh,
4685                                         partial->p,
4686                                         partial->p+1, (chain+n-1) - partial);
4687                 }
4688         }
4689         /* Clear the ends of indirect blocks on the shared branch */
4690         while (partial > chain) {
4691                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4692                                    (__le32*)partial->bh->b_data+addr_per_block,
4693                                    (chain+n-1) - partial);
4694                 BUFFER_TRACE(partial->bh, "call brelse");
4695                 brelse(partial->bh);
4696                 partial--;
4697         }
4698 do_indirects:
4699         /* Kill the remaining (whole) subtrees */
4700         switch (offsets[0]) {
4701         default:
4702                 nr = i_data[EXT4_IND_BLOCK];
4703                 if (nr) {
4704                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4705                         i_data[EXT4_IND_BLOCK] = 0;
4706                 }
4707         case EXT4_IND_BLOCK:
4708                 nr = i_data[EXT4_DIND_BLOCK];
4709                 if (nr) {
4710                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4711                         i_data[EXT4_DIND_BLOCK] = 0;
4712                 }
4713         case EXT4_DIND_BLOCK:
4714                 nr = i_data[EXT4_TIND_BLOCK];
4715                 if (nr) {
4716                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4717                         i_data[EXT4_TIND_BLOCK] = 0;
4718                 }
4719         case EXT4_TIND_BLOCK:
4720                 ;
4721         }
4722
4723         up_write(&ei->i_data_sem);
4724         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4725         ext4_mark_inode_dirty(handle, inode);
4726
4727         /*
4728          * In a multi-transaction truncate, we only make the final transaction
4729          * synchronous
4730          */
4731         if (IS_SYNC(inode))
4732                 ext4_handle_sync(handle);
4733 out_stop:
4734         /*
4735          * If this was a simple ftruncate(), and the file will remain alive
4736          * then we need to clear up the orphan record which we created above.
4737          * However, if this was a real unlink then we were called by
4738          * ext4_delete_inode(), and we allow that function to clean up the
4739          * orphan info for us.
4740          */
4741         if (inode->i_nlink)
4742                 ext4_orphan_del(handle, inode);
4743
4744         ext4_journal_stop(handle);
4745 }
4746
4747 /*
4748  * ext4_get_inode_loc returns with an extra refcount against the inode's
4749  * underlying buffer_head on success. If 'in_mem' is true, we have all
4750  * data in memory that is needed to recreate the on-disk version of this
4751  * inode.
4752  */
4753 static int __ext4_get_inode_loc(struct inode *inode,
4754                                 struct ext4_iloc *iloc, int in_mem)
4755 {
4756         struct ext4_group_desc  *gdp;
4757         struct buffer_head      *bh;
4758         struct super_block      *sb = inode->i_sb;
4759         ext4_fsblk_t            block;
4760         int                     inodes_per_block, inode_offset;
4761
4762         iloc->bh = NULL;
4763         if (!ext4_valid_inum(sb, inode->i_ino))
4764                 return -EIO;
4765
4766         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4767         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4768         if (!gdp)
4769                 return -EIO;
4770
4771         /*
4772          * Figure out the offset within the block group inode table
4773          */
4774         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4775         inode_offset = ((inode->i_ino - 1) %
4776                         EXT4_INODES_PER_GROUP(sb));
4777         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4778         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4779
4780         bh = sb_getblk(sb, block);
4781         if (!bh) {
4782                 ext4_error(sb, "unable to read inode block - "
4783                            "inode=%lu, block=%llu", inode->i_ino, block);
4784                 return -EIO;
4785         }
4786         if (!buffer_uptodate(bh)) {
4787                 lock_buffer(bh);
4788
4789                 /*
4790                  * If the buffer has the write error flag, we have failed
4791                  * to write out another inode in the same block.  In this
4792                  * case, we don't have to read the block because we may
4793                  * read the old inode data successfully.
4794                  */
4795                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4796                         set_buffer_uptodate(bh);
4797
4798                 if (buffer_uptodate(bh)) {
4799                         /* someone brought it uptodate while we waited */
4800                         unlock_buffer(bh);
4801                         goto has_buffer;
4802                 }
4803
4804                 /*
4805                  * If we have all information of the inode in memory and this
4806                  * is the only valid inode in the block, we need not read the
4807                  * block.
4808                  */
4809                 if (in_mem) {
4810                         struct buffer_head *bitmap_bh;
4811                         int i, start;
4812
4813                         start = inode_offset & ~(inodes_per_block - 1);
4814
4815                         /* Is the inode bitmap in cache? */
4816                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4817                         if (!bitmap_bh)
4818                                 goto make_io;
4819
4820                         /*
4821                          * If the inode bitmap isn't in cache then the
4822                          * optimisation may end up performing two reads instead
4823                          * of one, so skip it.
4824                          */
4825                         if (!buffer_uptodate(bitmap_bh)) {
4826                                 brelse(bitmap_bh);
4827                                 goto make_io;
4828                         }
4829                         for (i = start; i < start + inodes_per_block; i++) {
4830                                 if (i == inode_offset)
4831                                         continue;
4832                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4833                                         break;
4834                         }
4835                         brelse(bitmap_bh);
4836                         if (i == start + inodes_per_block) {
4837                                 /* all other inodes are free, so skip I/O */
4838                                 memset(bh->b_data, 0, bh->b_size);
4839                                 set_buffer_uptodate(bh);
4840                                 unlock_buffer(bh);
4841                                 goto has_buffer;
4842                         }
4843                 }
4844
4845 make_io:
4846                 /*
4847                  * If we need to do any I/O, try to pre-readahead extra
4848                  * blocks from the inode table.
4849                  */
4850                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4851                         ext4_fsblk_t b, end, table;
4852                         unsigned num;
4853
4854                         table = ext4_inode_table(sb, gdp);
4855                         /* s_inode_readahead_blks is always a power of 2 */
4856                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4857                         if (table > b)
4858                                 b = table;
4859                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4860                         num = EXT4_INODES_PER_GROUP(sb);
4861                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4862                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4863                                 num -= ext4_itable_unused_count(sb, gdp);
4864                         table += num / inodes_per_block;
4865                         if (end > table)
4866                                 end = table;
4867                         while (b <= end)
4868                                 sb_breadahead(sb, b++);
4869                 }
4870
4871                 /*
4872                  * There are other valid inodes in the buffer, this inode
4873                  * has in-inode xattrs, or we don't have this inode in memory.
4874                  * Read the block from disk.
4875                  */
4876                 get_bh(bh);
4877                 bh->b_end_io = end_buffer_read_sync;
4878                 submit_bh(READ_META, bh);
4879                 wait_on_buffer(bh);
4880                 if (!buffer_uptodate(bh)) {
4881                         ext4_error(sb, "unable to read inode block - inode=%lu,"
4882                                    " block=%llu", inode->i_ino, block);
4883                         brelse(bh);
4884                         return -EIO;
4885                 }
4886         }
4887 has_buffer:
4888         iloc->bh = bh;
4889         return 0;
4890 }
4891
4892 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4893 {
4894         /* We have all inode data except xattrs in memory here. */
4895         return __ext4_get_inode_loc(inode, iloc,
4896                 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
4897 }
4898
4899 void ext4_set_inode_flags(struct inode *inode)
4900 {
4901         unsigned int flags = EXT4_I(inode)->i_flags;
4902
4903         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4904         if (flags & EXT4_SYNC_FL)
4905                 inode->i_flags |= S_SYNC;
4906         if (flags & EXT4_APPEND_FL)
4907                 inode->i_flags |= S_APPEND;
4908         if (flags & EXT4_IMMUTABLE_FL)
4909                 inode->i_flags |= S_IMMUTABLE;
4910         if (flags & EXT4_NOATIME_FL)
4911                 inode->i_flags |= S_NOATIME;
4912         if (flags & EXT4_DIRSYNC_FL)
4913                 inode->i_flags |= S_DIRSYNC;
4914 }
4915
4916 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4917 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4918 {
4919         unsigned int flags = ei->vfs_inode.i_flags;
4920
4921         ei->i_flags &= ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4922                         EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|EXT4_DIRSYNC_FL);
4923         if (flags & S_SYNC)
4924                 ei->i_flags |= EXT4_SYNC_FL;
4925         if (flags & S_APPEND)
4926                 ei->i_flags |= EXT4_APPEND_FL;
4927         if (flags & S_IMMUTABLE)
4928                 ei->i_flags |= EXT4_IMMUTABLE_FL;
4929         if (flags & S_NOATIME)
4930                 ei->i_flags |= EXT4_NOATIME_FL;
4931         if (flags & S_DIRSYNC)
4932                 ei->i_flags |= EXT4_DIRSYNC_FL;
4933 }
4934
4935 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4936                                   struct ext4_inode_info *ei)
4937 {
4938         blkcnt_t i_blocks ;
4939         struct inode *inode = &(ei->vfs_inode);
4940         struct super_block *sb = inode->i_sb;
4941
4942         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4943                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4944                 /* we are using combined 48 bit field */
4945                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4946                                         le32_to_cpu(raw_inode->i_blocks_lo);
4947                 if (ei->i_flags & EXT4_HUGE_FILE_FL) {
4948                         /* i_blocks represent file system block size */
4949                         return i_blocks  << (inode->i_blkbits - 9);
4950                 } else {
4951                         return i_blocks;
4952                 }
4953         } else {
4954                 return le32_to_cpu(raw_inode->i_blocks_lo);
4955         }
4956 }
4957
4958 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4959 {
4960         struct ext4_iloc iloc;
4961         struct ext4_inode *raw_inode;
4962         struct ext4_inode_info *ei;
4963         struct inode *inode;
4964         journal_t *journal = EXT4_SB(sb)->s_journal;
4965         long ret;
4966         int block;
4967
4968         inode = iget_locked(sb, ino);
4969         if (!inode)
4970                 return ERR_PTR(-ENOMEM);
4971         if (!(inode->i_state & I_NEW))
4972                 return inode;
4973
4974         ei = EXT4_I(inode);
4975         iloc.bh = 0;
4976
4977         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4978         if (ret < 0)
4979                 goto bad_inode;
4980         raw_inode = ext4_raw_inode(&iloc);
4981         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4982         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4983         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4984         if (!(test_opt(inode->i_sb, NO_UID32))) {
4985                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4986                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4987         }
4988         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
4989
4990         ei->i_state_flags = 0;
4991         ei->i_dir_start_lookup = 0;
4992         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4993         /* We now have enough fields to check if the inode was active or not.
4994          * This is needed because nfsd might try to access dead inodes
4995          * the test is that same one that e2fsck uses
4996          * NeilBrown 1999oct15
4997          */
4998         if (inode->i_nlink == 0) {
4999                 if (inode->i_mode == 0 ||
5000                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
5001                         /* this inode is deleted */
5002                         ret = -ESTALE;
5003                         goto bad_inode;
5004                 }
5005                 /* The only unlinked inodes we let through here have
5006                  * valid i_mode and are being read by the orphan
5007                  * recovery code: that's fine, we're about to complete
5008                  * the process of deleting those. */
5009         }
5010         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
5011         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
5012         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
5013         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
5014                 ei->i_file_acl |=
5015                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
5016         inode->i_size = ext4_isize(raw_inode);
5017         ei->i_disksize = inode->i_size;
5018 #ifdef CONFIG_QUOTA
5019         ei->i_reserved_quota = 0;
5020 #endif
5021         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
5022         ei->i_block_group = iloc.block_group;
5023         ei->i_last_alloc_group = ~0;
5024         /*
5025          * NOTE! The in-memory inode i_data array is in little-endian order
5026          * even on big-endian machines: we do NOT byteswap the block numbers!
5027          */
5028         for (block = 0; block < EXT4_N_BLOCKS; block++)
5029                 ei->i_data[block] = raw_inode->i_block[block];
5030         INIT_LIST_HEAD(&ei->i_orphan);
5031
5032         /*
5033          * Set transaction id's of transactions that have to be committed
5034          * to finish f[data]sync. We set them to currently running transaction
5035          * as we cannot be sure that the inode or some of its metadata isn't
5036          * part of the transaction - the inode could have been reclaimed and
5037          * now it is reread from disk.
5038          */
5039         if (journal) {
5040                 transaction_t *transaction;
5041                 tid_t tid;
5042
5043                 spin_lock(&journal->j_state_lock);
5044                 if (journal->j_running_transaction)
5045                         transaction = journal->j_running_transaction;
5046                 else
5047                         transaction = journal->j_committing_transaction;
5048                 if (transaction)
5049                         tid = transaction->t_tid;
5050                 else
5051                         tid = journal->j_commit_sequence;
5052                 spin_unlock(&journal->j_state_lock);
5053                 ei->i_sync_tid = tid;
5054                 ei->i_datasync_tid = tid;
5055         }
5056
5057         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5058                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
5059                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
5060                     EXT4_INODE_SIZE(inode->i_sb)) {
5061                         ret = -EIO;
5062                         goto bad_inode;
5063                 }
5064                 if (ei->i_extra_isize == 0) {
5065                         /* The extra space is currently unused. Use it. */
5066                         ei->i_extra_isize = sizeof(struct ext4_inode) -
5067                                             EXT4_GOOD_OLD_INODE_SIZE;
5068                 } else {
5069                         __le32 *magic = (void *)raw_inode +
5070                                         EXT4_GOOD_OLD_INODE_SIZE +
5071                                         ei->i_extra_isize;
5072                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
5073                                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
5074                 }
5075         } else
5076                 ei->i_extra_isize = 0;
5077
5078         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
5079         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
5080         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
5081         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
5082
5083         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
5084         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
5085                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5086                         inode->i_version |=
5087                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
5088         }
5089
5090         ret = 0;
5091         if (ei->i_file_acl &&
5092             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
5093                 ext4_error(sb, "bad extended attribute block %llu inode #%lu",
5094                            ei->i_file_acl, inode->i_ino);
5095                 ret = -EIO;
5096                 goto bad_inode;
5097         } else if (ei->i_flags & EXT4_EXTENTS_FL) {
5098                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5099                     (S_ISLNK(inode->i_mode) &&
5100                      !ext4_inode_is_fast_symlink(inode)))
5101                         /* Validate extent which is part of inode */
5102                         ret = ext4_ext_check_inode(inode);
5103         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5104                    (S_ISLNK(inode->i_mode) &&
5105                     !ext4_inode_is_fast_symlink(inode))) {
5106                 /* Validate block references which are part of inode */
5107                 ret = ext4_check_inode_blockref(inode);
5108         }
5109         if (ret)
5110                 goto bad_inode;
5111
5112         if (S_ISREG(inode->i_mode)) {
5113                 inode->i_op = &ext4_file_inode_operations;
5114                 inode->i_fop = &ext4_file_operations;
5115                 ext4_set_aops(inode);
5116         } else if (S_ISDIR(inode->i_mode)) {
5117                 inode->i_op = &ext4_dir_inode_operations;
5118                 inode->i_fop = &ext4_dir_operations;
5119         } else if (S_ISLNK(inode->i_mode)) {
5120                 if (ext4_inode_is_fast_symlink(inode)) {
5121                         inode->i_op = &ext4_fast_symlink_inode_operations;
5122                         nd_terminate_link(ei->i_data, inode->i_size,
5123                                 sizeof(ei->i_data) - 1);
5124                 } else {
5125                         inode->i_op = &ext4_symlink_inode_operations;
5126                         ext4_set_aops(inode);
5127                 }
5128         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
5129               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
5130                 inode->i_op = &ext4_special_inode_operations;
5131                 if (raw_inode->i_block[0])
5132                         init_special_inode(inode, inode->i_mode,
5133                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
5134                 else
5135                         init_special_inode(inode, inode->i_mode,
5136                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
5137         } else {
5138                 ret = -EIO;
5139                 ext4_error(inode->i_sb, "bogus i_mode (%o) for inode=%lu",
5140                            inode->i_mode, inode->i_ino);
5141                 goto bad_inode;
5142         }
5143         brelse(iloc.bh);
5144         ext4_set_inode_flags(inode);
5145         unlock_new_inode(inode);
5146         return inode;
5147
5148 bad_inode:
5149         brelse(iloc.bh);
5150         iget_failed(inode);
5151         return ERR_PTR(ret);
5152 }
5153
5154 static int ext4_inode_blocks_set(handle_t *handle,
5155                                 struct ext4_inode *raw_inode,
5156                                 struct ext4_inode_info *ei)
5157 {
5158         struct inode *inode = &(ei->vfs_inode);
5159         u64 i_blocks = inode->i_blocks;
5160         struct super_block *sb = inode->i_sb;
5161
5162         if (i_blocks <= ~0U) {
5163                 /*
5164                  * i_blocks can be represnted in a 32 bit variable
5165                  * as multiple of 512 bytes
5166                  */
5167                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5168                 raw_inode->i_blocks_high = 0;
5169                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
5170                 return 0;
5171         }
5172         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
5173                 return -EFBIG;
5174
5175         if (i_blocks <= 0xffffffffffffULL) {
5176                 /*
5177                  * i_blocks can be represented in a 48 bit variable
5178                  * as multiple of 512 bytes
5179                  */
5180                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5181                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5182                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
5183         } else {
5184                 ei->i_flags |= EXT4_HUGE_FILE_FL;
5185                 /* i_block is stored in file system block size */
5186                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
5187                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5188                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5189         }
5190         return 0;
5191 }
5192
5193 /*
5194  * Post the struct inode info into an on-disk inode location in the
5195  * buffer-cache.  This gobbles the caller's reference to the
5196  * buffer_head in the inode location struct.
5197  *
5198  * The caller must have write access to iloc->bh.
5199  */
5200 static int ext4_do_update_inode(handle_t *handle,
5201                                 struct inode *inode,
5202                                 struct ext4_iloc *iloc)
5203 {
5204         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5205         struct ext4_inode_info *ei = EXT4_I(inode);
5206         struct buffer_head *bh = iloc->bh;
5207         int err = 0, rc, block;
5208
5209         /* For fields not not tracking in the in-memory inode,
5210          * initialise them to zero for new inodes. */
5211         if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5212                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5213
5214         ext4_get_inode_flags(ei);
5215         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5216         if (!(test_opt(inode->i_sb, NO_UID32))) {
5217                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
5218                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
5219 /*
5220  * Fix up interoperability with old kernels. Otherwise, old inodes get
5221  * re-used with the upper 16 bits of the uid/gid intact
5222  */
5223                 if (!ei->i_dtime) {
5224                         raw_inode->i_uid_high =
5225                                 cpu_to_le16(high_16_bits(inode->i_uid));
5226                         raw_inode->i_gid_high =
5227                                 cpu_to_le16(high_16_bits(inode->i_gid));
5228                 } else {
5229                         raw_inode->i_uid_high = 0;
5230                         raw_inode->i_gid_high = 0;
5231                 }
5232         } else {
5233                 raw_inode->i_uid_low =
5234                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
5235                 raw_inode->i_gid_low =
5236                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
5237                 raw_inode->i_uid_high = 0;
5238                 raw_inode->i_gid_high = 0;
5239         }
5240         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5241
5242         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5243         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5244         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5245         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5246
5247         if (ext4_inode_blocks_set(handle, raw_inode, ei))
5248                 goto out_brelse;
5249         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5250         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
5251         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
5252             cpu_to_le32(EXT4_OS_HURD))
5253                 raw_inode->i_file_acl_high =
5254                         cpu_to_le16(ei->i_file_acl >> 32);
5255         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5256         ext4_isize_set(raw_inode, ei->i_disksize);
5257         if (ei->i_disksize > 0x7fffffffULL) {
5258                 struct super_block *sb = inode->i_sb;
5259                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
5260                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
5261                                 EXT4_SB(sb)->s_es->s_rev_level ==
5262                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
5263                         /* If this is the first large file
5264                          * created, add a flag to the superblock.
5265                          */
5266                         err = ext4_journal_get_write_access(handle,
5267                                         EXT4_SB(sb)->s_sbh);
5268                         if (err)
5269                                 goto out_brelse;
5270                         ext4_update_dynamic_rev(sb);
5271                         EXT4_SET_RO_COMPAT_FEATURE(sb,
5272                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
5273                         sb->s_dirt = 1;
5274                         ext4_handle_sync(handle);
5275                         err = ext4_handle_dirty_metadata(handle, NULL,
5276                                         EXT4_SB(sb)->s_sbh);
5277                 }
5278         }
5279         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5280         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5281                 if (old_valid_dev(inode->i_rdev)) {
5282                         raw_inode->i_block[0] =
5283                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
5284                         raw_inode->i_block[1] = 0;
5285                 } else {
5286                         raw_inode->i_block[0] = 0;
5287                         raw_inode->i_block[1] =
5288                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
5289                         raw_inode->i_block[2] = 0;
5290                 }
5291         } else
5292                 for (block = 0; block < EXT4_N_BLOCKS; block++)
5293                         raw_inode->i_block[block] = ei->i_data[block];
5294
5295         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
5296         if (ei->i_extra_isize) {
5297                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5298                         raw_inode->i_version_hi =
5299                         cpu_to_le32(inode->i_version >> 32);
5300                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
5301         }
5302
5303         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5304         rc = ext4_handle_dirty_metadata(handle, NULL, bh);
5305         if (!err)
5306                 err = rc;
5307         ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5308
5309         ext4_update_inode_fsync_trans(handle, inode, 0);
5310 out_brelse:
5311         brelse(bh);
5312         ext4_std_error(inode->i_sb, err);
5313         return err;
5314 }
5315
5316 /*
5317  * ext4_write_inode()
5318  *
5319  * We are called from a few places:
5320  *
5321  * - Within generic_file_write() for O_SYNC files.
5322  *   Here, there will be no transaction running. We wait for any running
5323  *   trasnaction to commit.
5324  *
5325  * - Within sys_sync(), kupdate and such.
5326  *   We wait on commit, if tol to.
5327  *
5328  * - Within prune_icache() (PF_MEMALLOC == true)
5329  *   Here we simply return.  We can't afford to block kswapd on the
5330  *   journal commit.
5331  *
5332  * In all cases it is actually safe for us to return without doing anything,
5333  * because the inode has been copied into a raw inode buffer in
5334  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
5335  * knfsd.
5336  *
5337  * Note that we are absolutely dependent upon all inode dirtiers doing the
5338  * right thing: they *must* call mark_inode_dirty() after dirtying info in
5339  * which we are interested.
5340  *
5341  * It would be a bug for them to not do this.  The code:
5342  *
5343  *      mark_inode_dirty(inode)
5344  *      stuff();
5345  *      inode->i_size = expr;
5346  *
5347  * is in error because a kswapd-driven write_inode() could occur while
5348  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
5349  * will no longer be on the superblock's dirty inode list.
5350  */
5351 int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5352 {
5353         int err;
5354
5355         if (current->flags & PF_MEMALLOC)
5356                 return 0;
5357
5358         if (EXT4_SB(inode->i_sb)->s_journal) {
5359                 if (ext4_journal_current_handle()) {
5360                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5361                         dump_stack();
5362                         return -EIO;
5363                 }
5364
5365                 if (wbc->sync_mode != WB_SYNC_ALL)
5366                         return 0;
5367
5368                 err = ext4_force_commit(inode->i_sb);
5369         } else {
5370                 struct ext4_iloc iloc;
5371
5372                 err = ext4_get_inode_loc(inode, &iloc);
5373                 if (err)
5374                         return err;
5375                 if (wbc->sync_mode == WB_SYNC_ALL)
5376                         sync_dirty_buffer(iloc.bh);
5377                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5378                         ext4_error(inode->i_sb, "IO error syncing inode, "
5379                                    "inode=%lu, block=%llu", inode->i_ino,
5380                                    (unsigned long long)iloc.bh->b_blocknr);
5381                         err = -EIO;
5382                 }
5383         }
5384         return err;
5385 }
5386
5387 /*
5388  * ext4_setattr()
5389  *
5390  * Called from notify_change.
5391  *
5392  * We want to trap VFS attempts to truncate the file as soon as
5393  * possible.  In particular, we want to make sure that when the VFS
5394  * shrinks i_size, we put the inode on the orphan list and modify
5395  * i_disksize immediately, so that during the subsequent flushing of
5396  * dirty pages and freeing of disk blocks, we can guarantee that any
5397  * commit will leave the blocks being flushed in an unused state on
5398  * disk.  (On recovery, the inode will get truncated and the blocks will
5399  * be freed, so we have a strong guarantee that no future commit will
5400  * leave these blocks visible to the user.)
5401  *
5402  * Another thing we have to assure is that if we are in ordered mode
5403  * and inode is still attached to the committing transaction, we must
5404  * we start writeout of all the dirty pages which are being truncated.
5405  * This way we are sure that all the data written in the previous
5406  * transaction are already on disk (truncate waits for pages under
5407  * writeback).
5408  *
5409  * Called with inode->i_mutex down.
5410  */
5411 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5412 {
5413         struct inode *inode = dentry->d_inode;
5414         int error, rc = 0;
5415         const unsigned int ia_valid = attr->ia_valid;
5416
5417         error = inode_change_ok(inode, attr);
5418         if (error)
5419                 return error;
5420
5421         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
5422                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
5423                 handle_t *handle;
5424
5425                 /* (user+group)*(old+new) structure, inode write (sb,
5426                  * inode block, ? - but truncate inode update has it) */
5427                 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
5428                                         EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
5429                 if (IS_ERR(handle)) {
5430                         error = PTR_ERR(handle);
5431                         goto err_out;
5432                 }
5433                 error = vfs_dq_transfer(inode, attr) ? -EDQUOT : 0;
5434                 if (error) {
5435                         ext4_journal_stop(handle);
5436                         return error;
5437                 }
5438                 /* Update corresponding info in inode so that everything is in
5439                  * one transaction */
5440                 if (attr->ia_valid & ATTR_UID)
5441                         inode->i_uid = attr->ia_uid;
5442                 if (attr->ia_valid & ATTR_GID)
5443                         inode->i_gid = attr->ia_gid;
5444                 error = ext4_mark_inode_dirty(handle, inode);
5445                 ext4_journal_stop(handle);
5446         }
5447
5448         if (attr->ia_valid & ATTR_SIZE) {
5449                 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)) {
5450                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5451
5452                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
5453                                 error = -EFBIG;
5454                                 goto err_out;
5455                         }
5456                 }
5457         }
5458
5459         if (S_ISREG(inode->i_mode) &&
5460             attr->ia_valid & ATTR_SIZE &&
5461             (attr->ia_size < inode->i_size ||
5462              (EXT4_I(inode)->i_flags & EXT4_EOFBLOCKS_FL))) {
5463                 handle_t *handle;
5464
5465                 handle = ext4_journal_start(inode, 3);
5466                 if (IS_ERR(handle)) {
5467                         error = PTR_ERR(handle);
5468                         goto err_out;
5469                 }
5470
5471                 error = ext4_orphan_add(handle, inode);
5472                 EXT4_I(inode)->i_disksize = attr->ia_size;
5473                 rc = ext4_mark_inode_dirty(handle, inode);
5474                 if (!error)
5475                         error = rc;
5476                 ext4_journal_stop(handle);
5477
5478                 if (ext4_should_order_data(inode)) {
5479                         error = ext4_begin_ordered_truncate(inode,
5480                                                             attr->ia_size);
5481                         if (error) {
5482                                 /* Do as much error cleanup as possible */
5483                                 handle = ext4_journal_start(inode, 3);
5484                                 if (IS_ERR(handle)) {
5485                                         ext4_orphan_del(NULL, inode);
5486                                         goto err_out;
5487                                 }
5488                                 ext4_orphan_del(handle, inode);
5489                                 ext4_journal_stop(handle);
5490                                 goto err_out;
5491                         }
5492                 }
5493                 /* ext4_truncate will clear the flag */
5494                 if ((EXT4_I(inode)->i_flags & EXT4_EOFBLOCKS_FL))
5495                         ext4_truncate(inode);
5496         }
5497
5498         rc = inode_setattr(inode, attr);
5499
5500         /* If inode_setattr's call to ext4_truncate failed to get a
5501          * transaction handle at all, we need to clean up the in-core
5502          * orphan list manually. */
5503         if (inode->i_nlink)
5504                 ext4_orphan_del(NULL, inode);
5505
5506         if (!rc && (ia_valid & ATTR_MODE))
5507                 rc = ext4_acl_chmod(inode);
5508
5509 err_out:
5510         ext4_std_error(inode->i_sb, error);
5511         if (!error)
5512                 error = rc;
5513         return error;
5514 }
5515
5516 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
5517                  struct kstat *stat)
5518 {
5519         struct inode *inode;
5520         unsigned long delalloc_blocks;
5521
5522         inode = dentry->d_inode;
5523         generic_fillattr(inode, stat);
5524
5525         /*
5526          * We can't update i_blocks if the block allocation is delayed
5527          * otherwise in the case of system crash before the real block
5528          * allocation is done, we will have i_blocks inconsistent with
5529          * on-disk file blocks.
5530          * We always keep i_blocks updated together with real
5531          * allocation. But to not confuse with user, stat
5532          * will return the blocks that include the delayed allocation
5533          * blocks for this file.
5534          */
5535         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
5536         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
5537         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
5538
5539         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
5540         return 0;
5541 }
5542
5543 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
5544                                       int chunk)
5545 {
5546         int indirects;
5547
5548         /* if nrblocks are contiguous */
5549         if (chunk) {
5550                 /*
5551                  * With N contiguous data blocks, it need at most
5552    &nbs