ext4: Avoid leaking blocks after a block allocation failure
[linux-2.6] / 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 "ext4_jbd2.h"
41 #include "xattr.h"
42 #include "acl.h"
43 #include "ext4_extents.h"
44
45 #define MPAGE_DA_EXTENT_TAIL 0x01
46
47 static inline int ext4_begin_ordered_truncate(struct inode *inode,
48                                               loff_t new_size)
49 {
50         return jbd2_journal_begin_ordered_truncate(
51                                         EXT4_SB(inode->i_sb)->s_journal,
52                                         &EXT4_I(inode)->jinode,
53                                         new_size);
54 }
55
56 static void ext4_invalidatepage(struct page *page, unsigned long offset);
57
58 /*
59  * Test whether an inode is a fast symlink.
60  */
61 static int ext4_inode_is_fast_symlink(struct inode *inode)
62 {
63         int ea_blocks = EXT4_I(inode)->i_file_acl ?
64                 (inode->i_sb->s_blocksize >> 9) : 0;
65
66         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
67 }
68
69 /*
70  * The ext4 forget function must perform a revoke if we are freeing data
71  * which has been journaled.  Metadata (eg. indirect blocks) must be
72  * revoked in all cases.
73  *
74  * "bh" may be NULL: a metadata block may have been freed from memory
75  * but there may still be a record of it in the journal, and that record
76  * still needs to be revoked.
77  *
78  * If the handle isn't valid we're not journaling so there's nothing to do.
79  */
80 int ext4_forget(handle_t *handle, int is_metadata, struct inode *inode,
81                         struct buffer_head *bh, ext4_fsblk_t blocknr)
82 {
83         int err;
84
85         if (!ext4_handle_valid(handle))
86                 return 0;
87
88         might_sleep();
89
90         BUFFER_TRACE(bh, "enter");
91
92         jbd_debug(4, "forgetting bh %p: is_metadata = %d, mode %o, "
93                   "data mode %lx\n",
94                   bh, is_metadata, inode->i_mode,
95                   test_opt(inode->i_sb, DATA_FLAGS));
96
97         /* Never use the revoke function if we are doing full data
98          * journaling: there is no need to, and a V1 superblock won't
99          * support it.  Otherwise, only skip the revoke on un-journaled
100          * data blocks. */
101
102         if (test_opt(inode->i_sb, DATA_FLAGS) == EXT4_MOUNT_JOURNAL_DATA ||
103             (!is_metadata && !ext4_should_journal_data(inode))) {
104                 if (bh) {
105                         BUFFER_TRACE(bh, "call jbd2_journal_forget");
106                         return ext4_journal_forget(handle, bh);
107                 }
108                 return 0;
109         }
110
111         /*
112          * data!=journal && (is_metadata || should_journal_data(inode))
113          */
114         BUFFER_TRACE(bh, "call ext4_journal_revoke");
115         err = ext4_journal_revoke(handle, blocknr, bh);
116         if (err)
117                 ext4_abort(inode->i_sb, __func__,
118                            "error %d when attempting revoke", err);
119         BUFFER_TRACE(bh, "exit");
120         return err;
121 }
122
123 /*
124  * Work out how many blocks we need to proceed with the next chunk of a
125  * truncate transaction.
126  */
127 static unsigned long blocks_for_truncate(struct inode *inode)
128 {
129         ext4_lblk_t needed;
130
131         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
132
133         /* Give ourselves just enough room to cope with inodes in which
134          * i_blocks is corrupt: we've seen disk corruptions in the past
135          * which resulted in random data in an inode which looked enough
136          * like a regular file for ext4 to try to delete it.  Things
137          * will go a bit crazy if that happens, but at least we should
138          * try not to panic the whole kernel. */
139         if (needed < 2)
140                 needed = 2;
141
142         /* But we need to bound the transaction so we don't overflow the
143          * journal. */
144         if (needed > EXT4_MAX_TRANS_DATA)
145                 needed = EXT4_MAX_TRANS_DATA;
146
147         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
148 }
149
150 /*
151  * Truncate transactions can be complex and absolutely huge.  So we need to
152  * be able to restart the transaction at a conventient checkpoint to make
153  * sure we don't overflow the journal.
154  *
155  * start_transaction gets us a new handle for a truncate transaction,
156  * and extend_transaction tries to extend the existing one a bit.  If
157  * extend fails, we need to propagate the failure up and restart the
158  * transaction in the top-level truncate loop. --sct
159  */
160 static handle_t *start_transaction(struct inode *inode)
161 {
162         handle_t *result;
163
164         result = ext4_journal_start(inode, blocks_for_truncate(inode));
165         if (!IS_ERR(result))
166                 return result;
167
168         ext4_std_error(inode->i_sb, PTR_ERR(result));
169         return result;
170 }
171
172 /*
173  * Try to extend this transaction for the purposes of truncation.
174  *
175  * Returns 0 if we managed to create more room.  If we can't create more
176  * room, and the transaction must be restarted we return 1.
177  */
178 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
179 {
180         if (!ext4_handle_valid(handle))
181                 return 0;
182         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
183                 return 0;
184         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
185                 return 0;
186         return 1;
187 }
188
189 /*
190  * Restart the transaction associated with *handle.  This does a commit,
191  * so before we call here everything must be consistently dirtied against
192  * this transaction.
193  */
194 static int ext4_journal_test_restart(handle_t *handle, struct inode *inode)
195 {
196         BUG_ON(EXT4_JOURNAL(inode) == NULL);
197         jbd_debug(2, "restarting handle %p\n", handle);
198         return ext4_journal_restart(handle, blocks_for_truncate(inode));
199 }
200
201 /*
202  * Called at the last iput() if i_nlink is zero.
203  */
204 void ext4_delete_inode(struct inode *inode)
205 {
206         handle_t *handle;
207         int err;
208
209         if (ext4_should_order_data(inode))
210                 ext4_begin_ordered_truncate(inode, 0);
211         truncate_inode_pages(&inode->i_data, 0);
212
213         if (is_bad_inode(inode))
214                 goto no_delete;
215
216         handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
217         if (IS_ERR(handle)) {
218                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
219                 /*
220                  * If we're going to skip the normal cleanup, we still need to
221                  * make sure that the in-core orphan linked list is properly
222                  * cleaned up.
223                  */
224                 ext4_orphan_del(NULL, inode);
225                 goto no_delete;
226         }
227
228         if (IS_SYNC(inode))
229                 ext4_handle_sync(handle);
230         inode->i_size = 0;
231         err = ext4_mark_inode_dirty(handle, inode);
232         if (err) {
233                 ext4_warning(inode->i_sb, __func__,
234                              "couldn't mark inode dirty (err %d)", err);
235                 goto stop_handle;
236         }
237         if (inode->i_blocks)
238                 ext4_truncate(inode);
239
240         /*
241          * ext4_ext_truncate() doesn't reserve any slop when it
242          * restarts journal transactions; therefore there may not be
243          * enough credits left in the handle to remove the inode from
244          * the orphan list and set the dtime field.
245          */
246         if (!ext4_handle_has_enough_credits(handle, 3)) {
247                 err = ext4_journal_extend(handle, 3);
248                 if (err > 0)
249                         err = ext4_journal_restart(handle, 3);
250                 if (err != 0) {
251                         ext4_warning(inode->i_sb, __func__,
252                                      "couldn't extend journal (err %d)", err);
253                 stop_handle:
254                         ext4_journal_stop(handle);
255                         goto no_delete;
256                 }
257         }
258
259         /*
260          * Kill off the orphan record which ext4_truncate created.
261          * AKPM: I think this can be inside the above `if'.
262          * Note that ext4_orphan_del() has to be able to cope with the
263          * deletion of a non-existent orphan - this is because we don't
264          * know if ext4_truncate() actually created an orphan record.
265          * (Well, we could do this if we need to, but heck - it works)
266          */
267         ext4_orphan_del(handle, inode);
268         EXT4_I(inode)->i_dtime  = get_seconds();
269
270         /*
271          * One subtle ordering requirement: if anything has gone wrong
272          * (transaction abort, IO errors, whatever), then we can still
273          * do these next steps (the fs will already have been marked as
274          * having errors), but we can't free the inode if the mark_dirty
275          * fails.
276          */
277         if (ext4_mark_inode_dirty(handle, inode))
278                 /* If that failed, just do the required in-core inode clear. */
279                 clear_inode(inode);
280         else
281                 ext4_free_inode(handle, inode);
282         ext4_journal_stop(handle);
283         return;
284 no_delete:
285         clear_inode(inode);     /* We must guarantee clearing of inode... */
286 }
287
288 typedef struct {
289         __le32  *p;
290         __le32  key;
291         struct buffer_head *bh;
292 } Indirect;
293
294 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
295 {
296         p->key = *(p->p = v);
297         p->bh = bh;
298 }
299
300 /**
301  *      ext4_block_to_path - parse the block number into array of offsets
302  *      @inode: inode in question (we are only interested in its superblock)
303  *      @i_block: block number to be parsed
304  *      @offsets: array to store the offsets in
305  *      @boundary: set this non-zero if the referred-to block is likely to be
306  *             followed (on disk) by an indirect block.
307  *
308  *      To store the locations of file's data ext4 uses a data structure common
309  *      for UNIX filesystems - tree of pointers anchored in the inode, with
310  *      data blocks at leaves and indirect blocks in intermediate nodes.
311  *      This function translates the block number into path in that tree -
312  *      return value is the path length and @offsets[n] is the offset of
313  *      pointer to (n+1)th node in the nth one. If @block is out of range
314  *      (negative or too large) warning is printed and zero returned.
315  *
316  *      Note: function doesn't find node addresses, so no IO is needed. All
317  *      we need to know is the capacity of indirect blocks (taken from the
318  *      inode->i_sb).
319  */
320
321 /*
322  * Portability note: the last comparison (check that we fit into triple
323  * indirect block) is spelled differently, because otherwise on an
324  * architecture with 32-bit longs and 8Kb pages we might get into trouble
325  * if our filesystem had 8Kb blocks. We might use long long, but that would
326  * kill us on x86. Oh, well, at least the sign propagation does not matter -
327  * i_block would have to be negative in the very beginning, so we would not
328  * get there at all.
329  */
330
331 static int ext4_block_to_path(struct inode *inode,
332                         ext4_lblk_t i_block,
333                         ext4_lblk_t offsets[4], int *boundary)
334 {
335         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
336         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
337         const long direct_blocks = EXT4_NDIR_BLOCKS,
338                 indirect_blocks = ptrs,
339                 double_blocks = (1 << (ptrs_bits * 2));
340         int n = 0;
341         int final = 0;
342
343         if (i_block < 0) {
344                 ext4_warning(inode->i_sb, "ext4_block_to_path", "block < 0");
345         } else if (i_block < direct_blocks) {
346                 offsets[n++] = i_block;
347                 final = direct_blocks;
348         } else if ((i_block -= direct_blocks) < indirect_blocks) {
349                 offsets[n++] = EXT4_IND_BLOCK;
350                 offsets[n++] = i_block;
351                 final = ptrs;
352         } else if ((i_block -= indirect_blocks) < double_blocks) {
353                 offsets[n++] = EXT4_DIND_BLOCK;
354                 offsets[n++] = i_block >> ptrs_bits;
355                 offsets[n++] = i_block & (ptrs - 1);
356                 final = ptrs;
357         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
358                 offsets[n++] = EXT4_TIND_BLOCK;
359                 offsets[n++] = i_block >> (ptrs_bits * 2);
360                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
361                 offsets[n++] = i_block & (ptrs - 1);
362                 final = ptrs;
363         } else {
364                 ext4_warning(inode->i_sb, "ext4_block_to_path",
365                                 "block %lu > max in inode %lu",
366                                 i_block + direct_blocks +
367                                 indirect_blocks + double_blocks, inode->i_ino);
368         }
369         if (boundary)
370                 *boundary = final - 1 - (i_block & (ptrs - 1));
371         return n;
372 }
373
374 static int __ext4_check_blockref(const char *function, struct inode *inode,
375                                  __le32 *p, unsigned int max)
376 {
377         __le32 *bref = p;
378         unsigned int blk;
379
380         while (bref < p+max) {
381                 blk = le32_to_cpu(*bref++);
382                 if (blk && 
383                     unlikely(!ext4_data_block_valid(EXT4_SB(inode->i_sb), 
384                                                     blk, 1))) {
385                         ext4_error(inode->i_sb, function,
386                                    "invalid block reference %u "
387                                    "in inode #%lu", blk, inode->i_ino);
388                         return -EIO;
389                 }
390         }
391         return 0;
392 }
393
394
395 #define ext4_check_indirect_blockref(inode, bh)                         \
396         __ext4_check_blockref(__func__, inode, (__le32 *)(bh)->b_data,  \
397                               EXT4_ADDR_PER_BLOCK((inode)->i_sb))
398
399 #define ext4_check_inode_blockref(inode)                                \
400         __ext4_check_blockref(__func__, inode, EXT4_I(inode)->i_data,   \
401                               EXT4_NDIR_BLOCKS)
402
403 /**
404  *      ext4_get_branch - read the chain of indirect blocks leading to data
405  *      @inode: inode in question
406  *      @depth: depth of the chain (1 - direct pointer, etc.)
407  *      @offsets: offsets of pointers in inode/indirect blocks
408  *      @chain: place to store the result
409  *      @err: here we store the error value
410  *
411  *      Function fills the array of triples <key, p, bh> and returns %NULL
412  *      if everything went OK or the pointer to the last filled triple
413  *      (incomplete one) otherwise. Upon the return chain[i].key contains
414  *      the number of (i+1)-th block in the chain (as it is stored in memory,
415  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
416  *      number (it points into struct inode for i==0 and into the bh->b_data
417  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
418  *      block for i>0 and NULL for i==0. In other words, it holds the block
419  *      numbers of the chain, addresses they were taken from (and where we can
420  *      verify that chain did not change) and buffer_heads hosting these
421  *      numbers.
422  *
423  *      Function stops when it stumbles upon zero pointer (absent block)
424  *              (pointer to last triple returned, *@err == 0)
425  *      or when it gets an IO error reading an indirect block
426  *              (ditto, *@err == -EIO)
427  *      or when it reads all @depth-1 indirect blocks successfully and finds
428  *      the whole chain, all way to the data (returns %NULL, *err == 0).
429  *
430  *      Need to be called with
431  *      down_read(&EXT4_I(inode)->i_data_sem)
432  */
433 static Indirect *ext4_get_branch(struct inode *inode, int depth,
434                                  ext4_lblk_t  *offsets,
435                                  Indirect chain[4], int *err)
436 {
437         struct super_block *sb = inode->i_sb;
438         Indirect *p = chain;
439         struct buffer_head *bh;
440
441         *err = 0;
442         /* i_data is not going away, no lock needed */
443         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
444         if (!p->key)
445                 goto no_block;
446         while (--depth) {
447                 bh = sb_getblk(sb, le32_to_cpu(p->key));
448                 if (unlikely(!bh))
449                         goto failure;
450                   
451                 if (!bh_uptodate_or_lock(bh)) {
452                         if (bh_submit_read(bh) < 0) {
453                                 put_bh(bh);
454                                 goto failure;
455                         }
456                         /* validate block references */
457                         if (ext4_check_indirect_blockref(inode, bh)) {
458                                 put_bh(bh);
459                                 goto failure;
460                         }
461                 }
462                 
463                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
464                 /* Reader: end */
465                 if (!p->key)
466                         goto no_block;
467         }
468         return NULL;
469
470 failure:
471         *err = -EIO;
472 no_block:
473         return p;
474 }
475
476 /**
477  *      ext4_find_near - find a place for allocation with sufficient locality
478  *      @inode: owner
479  *      @ind: descriptor of indirect block.
480  *
481  *      This function returns the preferred place for block allocation.
482  *      It is used when heuristic for sequential allocation fails.
483  *      Rules are:
484  *        + if there is a block to the left of our position - allocate near it.
485  *        + if pointer will live in indirect block - allocate near that block.
486  *        + if pointer will live in inode - allocate in the same
487  *          cylinder group.
488  *
489  * In the latter case we colour the starting block by the callers PID to
490  * prevent it from clashing with concurrent allocations for a different inode
491  * in the same block group.   The PID is used here so that functionally related
492  * files will be close-by on-disk.
493  *
494  *      Caller must make sure that @ind is valid and will stay that way.
495  */
496 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
497 {
498         struct ext4_inode_info *ei = EXT4_I(inode);
499         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
500         __le32 *p;
501         ext4_fsblk_t bg_start;
502         ext4_fsblk_t last_block;
503         ext4_grpblk_t colour;
504         ext4_group_t block_group;
505         int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
506
507         /* Try to find previous block */
508         for (p = ind->p - 1; p >= start; p--) {
509                 if (*p)
510                         return le32_to_cpu(*p);
511         }
512
513         /* No such thing, so let's try location of indirect block */
514         if (ind->bh)
515                 return ind->bh->b_blocknr;
516
517         /*
518          * It is going to be referred to from the inode itself? OK, just put it
519          * into the same cylinder group then.
520          */
521         block_group = ei->i_block_group;
522         if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
523                 block_group &= ~(flex_size-1);
524                 if (S_ISREG(inode->i_mode))
525                         block_group++;
526         }
527         bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
528         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
529
530         /*
531          * If we are doing delayed allocation, we don't need take
532          * colour into account.
533          */
534         if (test_opt(inode->i_sb, DELALLOC))
535                 return bg_start;
536
537         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
538                 colour = (current->pid % 16) *
539                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
540         else
541                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
542         return bg_start + colour;
543 }
544
545 /**
546  *      ext4_find_goal - find a preferred place for allocation.
547  *      @inode: owner
548  *      @block:  block we want
549  *      @partial: pointer to the last triple within a chain
550  *
551  *      Normally this function find the preferred place for block allocation,
552  *      returns it.
553  */
554 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
555                 Indirect *partial)
556 {
557         /*
558          * XXX need to get goal block from mballoc's data structures
559          */
560
561         return ext4_find_near(inode, partial);
562 }
563
564 /**
565  *      ext4_blks_to_allocate: Look up the block map and count the number
566  *      of direct blocks need to be allocated for the given branch.
567  *
568  *      @branch: chain of indirect blocks
569  *      @k: number of blocks need for indirect blocks
570  *      @blks: number of data blocks to be mapped.
571  *      @blocks_to_boundary:  the offset in the indirect block
572  *
573  *      return the total number of blocks to be allocate, including the
574  *      direct and indirect blocks.
575  */
576 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
577                 int blocks_to_boundary)
578 {
579         unsigned int count = 0;
580
581         /*
582          * Simple case, [t,d]Indirect block(s) has not allocated yet
583          * then it's clear blocks on that path have not allocated
584          */
585         if (k > 0) {
586                 /* right now we don't handle cross boundary allocation */
587                 if (blks < blocks_to_boundary + 1)
588                         count += blks;
589                 else
590                         count += blocks_to_boundary + 1;
591                 return count;
592         }
593
594         count++;
595         while (count < blks && count <= blocks_to_boundary &&
596                 le32_to_cpu(*(branch[0].p + count)) == 0) {
597                 count++;
598         }
599         return count;
600 }
601
602 /**
603  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
604  *      @indirect_blks: the number of blocks need to allocate for indirect
605  *                      blocks
606  *
607  *      @new_blocks: on return it will store the new block numbers for
608  *      the indirect blocks(if needed) and the first direct block,
609  *      @blks:  on return it will store the total number of allocated
610  *              direct blocks
611  */
612 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
613                                 ext4_lblk_t iblock, ext4_fsblk_t goal,
614                                 int indirect_blks, int blks,
615                                 ext4_fsblk_t new_blocks[4], int *err)
616 {
617         struct ext4_allocation_request ar;
618         int target, i;
619         unsigned long count = 0, blk_allocated = 0;
620         int index = 0;
621         ext4_fsblk_t current_block = 0;
622         int ret = 0;
623
624         /*
625          * Here we try to allocate the requested multiple blocks at once,
626          * on a best-effort basis.
627          * To build a branch, we should allocate blocks for
628          * the indirect blocks(if not allocated yet), and at least
629          * the first direct block of this branch.  That's the
630          * minimum number of blocks need to allocate(required)
631          */
632         /* first we try to allocate the indirect blocks */
633         target = indirect_blks;
634         while (target > 0) {
635                 count = target;
636                 /* allocating blocks for indirect blocks and direct blocks */
637                 current_block = ext4_new_meta_blocks(handle, inode,
638                                                         goal, &count, err);
639                 if (*err)
640                         goto failed_out;
641
642                 target -= count;
643                 /* allocate blocks for indirect blocks */
644                 while (index < indirect_blks && count) {
645                         new_blocks[index++] = current_block++;
646                         count--;
647                 }
648                 if (count > 0) {
649                         /*
650                          * save the new block number
651                          * for the first direct block
652                          */
653                         new_blocks[index] = current_block;
654                         printk(KERN_INFO "%s returned more blocks than "
655                                                 "requested\n", __func__);
656                         WARN_ON(1);
657                         break;
658                 }
659         }
660
661         target = blks - count ;
662         blk_allocated = count;
663         if (!target)
664                 goto allocated;
665         /* Now allocate data blocks */
666         memset(&ar, 0, sizeof(ar));
667         ar.inode = inode;
668         ar.goal = goal;
669         ar.len = target;
670         ar.logical = iblock;
671         if (S_ISREG(inode->i_mode))
672                 /* enable in-core preallocation only for regular files */
673                 ar.flags = EXT4_MB_HINT_DATA;
674
675         current_block = ext4_mb_new_blocks(handle, &ar, err);
676
677         if (*err && (target == blks)) {
678                 /*
679                  * if the allocation failed and we didn't allocate
680                  * any blocks before
681                  */
682                 goto failed_out;
683         }
684         if (!*err) {
685                 if (target == blks) {
686                 /*
687                  * save the new block number
688                  * for the first direct block
689                  */
690                         new_blocks[index] = current_block;
691                 }
692                 blk_allocated += ar.len;
693         }
694 allocated:
695         /* total number of blocks allocated for direct blocks */
696         ret = blk_allocated;
697         *err = 0;
698         return ret;
699 failed_out:
700         for (i = 0; i < index; i++)
701                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
702         return ret;
703 }
704
705 /**
706  *      ext4_alloc_branch - allocate and set up a chain of blocks.
707  *      @inode: owner
708  *      @indirect_blks: number of allocated indirect blocks
709  *      @blks: number of allocated direct blocks
710  *      @offsets: offsets (in the blocks) to store the pointers to next.
711  *      @branch: place to store the chain in.
712  *
713  *      This function allocates blocks, zeroes out all but the last one,
714  *      links them into chain and (if we are synchronous) writes them to disk.
715  *      In other words, it prepares a branch that can be spliced onto the
716  *      inode. It stores the information about that chain in the branch[], in
717  *      the same format as ext4_get_branch() would do. We are calling it after
718  *      we had read the existing part of chain and partial points to the last
719  *      triple of that (one with zero ->key). Upon the exit we have the same
720  *      picture as after the successful ext4_get_block(), except that in one
721  *      place chain is disconnected - *branch->p is still zero (we did not
722  *      set the last link), but branch->key contains the number that should
723  *      be placed into *branch->p to fill that gap.
724  *
725  *      If allocation fails we free all blocks we've allocated (and forget
726  *      their buffer_heads) and return the error value the from failed
727  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
728  *      as described above and return 0.
729  */
730 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
731                                 ext4_lblk_t iblock, int indirect_blks,
732                                 int *blks, ext4_fsblk_t goal,
733                                 ext4_lblk_t *offsets, Indirect *branch)
734 {
735         int blocksize = inode->i_sb->s_blocksize;
736         int i, n = 0;
737         int err = 0;
738         struct buffer_head *bh;
739         int num;
740         ext4_fsblk_t new_blocks[4];
741         ext4_fsblk_t current_block;
742
743         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
744                                 *blks, new_blocks, &err);
745         if (err)
746                 return err;
747
748         branch[0].key = cpu_to_le32(new_blocks[0]);
749         /*
750          * metadata blocks and data blocks are allocated.
751          */
752         for (n = 1; n <= indirect_blks;  n++) {
753                 /*
754                  * Get buffer_head for parent block, zero it out
755                  * and set the pointer to new one, then send
756                  * parent to disk.
757                  */
758                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
759                 branch[n].bh = bh;
760                 lock_buffer(bh);
761                 BUFFER_TRACE(bh, "call get_create_access");
762                 err = ext4_journal_get_create_access(handle, bh);
763                 if (err) {
764                         unlock_buffer(bh);
765                         brelse(bh);
766                         goto failed;
767                 }
768
769                 memset(bh->b_data, 0, blocksize);
770                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
771                 branch[n].key = cpu_to_le32(new_blocks[n]);
772                 *branch[n].p = branch[n].key;
773                 if (n == indirect_blks) {
774                         current_block = new_blocks[n];
775                         /*
776                          * End of chain, update the last new metablock of
777                          * the chain to point to the new allocated
778                          * data blocks numbers
779                          */
780                         for (i=1; i < num; i++)
781                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
782                 }
783                 BUFFER_TRACE(bh, "marking uptodate");
784                 set_buffer_uptodate(bh);
785                 unlock_buffer(bh);
786
787                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
788                 err = ext4_handle_dirty_metadata(handle, inode, bh);
789                 if (err)
790                         goto failed;
791         }
792         *blks = num;
793         return err;
794 failed:
795         /* Allocation failed, free what we already allocated */
796         for (i = 1; i <= n ; i++) {
797                 BUFFER_TRACE(branch[i].bh, "call jbd2_journal_forget");
798                 ext4_journal_forget(handle, branch[i].bh);
799         }
800         for (i = 0; i < indirect_blks; i++)
801                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
802
803         ext4_free_blocks(handle, inode, new_blocks[i], num, 0);
804
805         return err;
806 }
807
808 /**
809  * ext4_splice_branch - splice the allocated branch onto inode.
810  * @inode: owner
811  * @block: (logical) number of block we are adding
812  * @chain: chain of indirect blocks (with a missing link - see
813  *      ext4_alloc_branch)
814  * @where: location of missing link
815  * @num:   number of indirect blocks we are adding
816  * @blks:  number of direct blocks we are adding
817  *
818  * This function fills the missing link and does all housekeeping needed in
819  * inode (->i_blocks, etc.). In case of success we end up with the full
820  * chain to new block and return 0.
821  */
822 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
823                         ext4_lblk_t block, Indirect *where, int num, int blks)
824 {
825         int i;
826         int err = 0;
827         ext4_fsblk_t current_block;
828
829         /*
830          * If we're splicing into a [td]indirect block (as opposed to the
831          * inode) then we need to get write access to the [td]indirect block
832          * before the splice.
833          */
834         if (where->bh) {
835                 BUFFER_TRACE(where->bh, "get_write_access");
836                 err = ext4_journal_get_write_access(handle, where->bh);
837                 if (err)
838                         goto err_out;
839         }
840         /* That's it */
841
842         *where->p = where->key;
843
844         /*
845          * Update the host buffer_head or inode to point to more just allocated
846          * direct blocks blocks
847          */
848         if (num == 0 && blks > 1) {
849                 current_block = le32_to_cpu(where->key) + 1;
850                 for (i = 1; i < blks; i++)
851                         *(where->p + i) = cpu_to_le32(current_block++);
852         }
853
854         /* We are done with atomic stuff, now do the rest of housekeeping */
855
856         inode->i_ctime = ext4_current_time(inode);
857         ext4_mark_inode_dirty(handle, inode);
858
859         /* had we spliced it onto indirect block? */
860         if (where->bh) {
861                 /*
862                  * If we spliced it onto an indirect block, we haven't
863                  * altered the inode.  Note however that if it is being spliced
864                  * onto an indirect block at the very end of the file (the
865                  * file is growing) then we *will* alter the inode to reflect
866                  * the new i_size.  But that is not done here - it is done in
867                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
868                  */
869                 jbd_debug(5, "splicing indirect only\n");
870                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
871                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
872                 if (err)
873                         goto err_out;
874         } else {
875                 /*
876                  * OK, we spliced it into the inode itself on a direct block.
877                  * Inode was dirtied above.
878                  */
879                 jbd_debug(5, "splicing direct\n");
880         }
881         return err;
882
883 err_out:
884         for (i = 1; i <= num; i++) {
885                 BUFFER_TRACE(where[i].bh, "call jbd2_journal_forget");
886                 ext4_journal_forget(handle, where[i].bh);
887                 ext4_free_blocks(handle, inode,
888                                         le32_to_cpu(where[i-1].key), 1, 0);
889         }
890         ext4_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks, 0);
891
892         return err;
893 }
894
895 /*
896  * The ext4_ind_get_blocks() function handles non-extents inodes
897  * (i.e., using the traditional indirect/double-indirect i_blocks
898  * scheme) for ext4_get_blocks().
899  *
900  * Allocation strategy is simple: if we have to allocate something, we will
901  * have to go the whole way to leaf. So let's do it before attaching anything
902  * to tree, set linkage between the newborn blocks, write them if sync is
903  * required, recheck the path, free and repeat if check fails, otherwise
904  * set the last missing link (that will protect us from any truncate-generated
905  * removals - all blocks on the path are immune now) and possibly force the
906  * write on the parent block.
907  * That has a nice additional property: no special recovery from the failed
908  * allocations is needed - we simply release blocks and do not touch anything
909  * reachable from inode.
910  *
911  * `handle' can be NULL if create == 0.
912  *
913  * return > 0, # of blocks mapped or allocated.
914  * return = 0, if plain lookup failed.
915  * return < 0, error case.
916  *
917  * The ext4_ind_get_blocks() function should be called with
918  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
919  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
920  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
921  * blocks.
922  */
923 static int ext4_ind_get_blocks(handle_t *handle, struct inode *inode,
924                                   ext4_lblk_t iblock, unsigned int maxblocks,
925                                   struct buffer_head *bh_result,
926                                   int flags)
927 {
928         int err = -EIO;
929         ext4_lblk_t offsets[4];
930         Indirect chain[4];
931         Indirect *partial;
932         ext4_fsblk_t goal;
933         int indirect_blks;
934         int blocks_to_boundary = 0;
935         int depth;
936         int count = 0;
937         ext4_fsblk_t first_block = 0;
938
939         J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
940         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
941         depth = ext4_block_to_path(inode, iblock, offsets,
942                                         &blocks_to_boundary);
943
944         if (depth == 0)
945                 goto out;
946
947         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
948
949         /* Simplest case - block found, no allocation needed */
950         if (!partial) {
951                 first_block = le32_to_cpu(chain[depth - 1].key);
952                 clear_buffer_new(bh_result);
953                 count++;
954                 /*map more blocks*/
955                 while (count < maxblocks && count <= blocks_to_boundary) {
956                         ext4_fsblk_t blk;
957
958                         blk = le32_to_cpu(*(chain[depth-1].p + count));
959
960                         if (blk == first_block + count)
961                                 count++;
962                         else
963                                 break;
964                 }
965                 goto got_it;
966         }
967
968         /* Next simple case - plain lookup or failed read of indirect block */
969         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
970                 goto cleanup;
971
972         /*
973          * Okay, we need to do block allocation.
974         */
975         goal = ext4_find_goal(inode, iblock, partial);
976
977         /* the number of blocks need to allocate for [d,t]indirect blocks */
978         indirect_blks = (chain + depth) - partial - 1;
979
980         /*
981          * Next look up the indirect map to count the totoal number of
982          * direct blocks to allocate for this branch.
983          */
984         count = ext4_blks_to_allocate(partial, indirect_blks,
985                                         maxblocks, blocks_to_boundary);
986         /*
987          * Block out ext4_truncate while we alter the tree
988          */
989         err = ext4_alloc_branch(handle, inode, iblock, indirect_blks,
990                                         &count, goal,
991                                         offsets + (partial - chain), partial);
992
993         /*
994          * The ext4_splice_branch call will free and forget any buffers
995          * on the new chain if there is a failure, but that risks using
996          * up transaction credits, especially for bitmaps where the
997          * credits cannot be returned.  Can we handle this somehow?  We
998          * may need to return -EAGAIN upwards in the worst case.  --sct
999          */
1000         if (!err)
1001                 err = ext4_splice_branch(handle, inode, iblock,
1002                                         partial, indirect_blks, count);
1003         else 
1004                 goto cleanup;
1005
1006         set_buffer_new(bh_result);
1007 got_it:
1008         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
1009         if (count > blocks_to_boundary)
1010                 set_buffer_boundary(bh_result);
1011         err = count;
1012         /* Clean up and exit */
1013         partial = chain + depth - 1;    /* the whole chain */
1014 cleanup:
1015         while (partial > chain) {
1016                 BUFFER_TRACE(partial->bh, "call brelse");
1017                 brelse(partial->bh);
1018                 partial--;
1019         }
1020         BUFFER_TRACE(bh_result, "returned");
1021 out:
1022         return err;
1023 }
1024
1025 qsize_t ext4_get_reserved_space(struct inode *inode)
1026 {
1027         unsigned long long total;
1028
1029         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1030         total = EXT4_I(inode)->i_reserved_data_blocks +
1031                 EXT4_I(inode)->i_reserved_meta_blocks;
1032         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1033
1034         return total;
1035 }
1036 /*
1037  * Calculate the number of metadata blocks need to reserve
1038  * to allocate @blocks for non extent file based file
1039  */
1040 static int ext4_indirect_calc_metadata_amount(struct inode *inode, int blocks)
1041 {
1042         int icap = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1043         int ind_blks, dind_blks, tind_blks;
1044
1045         /* number of new indirect blocks needed */
1046         ind_blks = (blocks + icap - 1) / icap;
1047
1048         dind_blks = (ind_blks + icap - 1) / icap;
1049
1050         tind_blks = 1;
1051
1052         return ind_blks + dind_blks + tind_blks;
1053 }
1054
1055 /*
1056  * Calculate the number of metadata blocks need to reserve
1057  * to allocate given number of blocks
1058  */
1059 static int ext4_calc_metadata_amount(struct inode *inode, int blocks)
1060 {
1061         if (!blocks)
1062                 return 0;
1063
1064         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
1065                 return ext4_ext_calc_metadata_amount(inode, blocks);
1066
1067         return ext4_indirect_calc_metadata_amount(inode, blocks);
1068 }
1069
1070 static void ext4_da_update_reserve_space(struct inode *inode, int used)
1071 {
1072         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1073         int total, mdb, mdb_free;
1074
1075         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1076         /* recalculate the number of metablocks still need to be reserved */
1077         total = EXT4_I(inode)->i_reserved_data_blocks - used;
1078         mdb = ext4_calc_metadata_amount(inode, total);
1079
1080         /* figure out how many metablocks to release */
1081         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1082         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1083
1084         if (mdb_free) {
1085                 /* Account for allocated meta_blocks */
1086                 mdb_free -= EXT4_I(inode)->i_allocated_meta_blocks;
1087
1088                 /* update fs dirty blocks counter */
1089                 percpu_counter_sub(&sbi->s_dirtyblocks_counter, mdb_free);
1090                 EXT4_I(inode)->i_allocated_meta_blocks = 0;
1091                 EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1092         }
1093
1094         /* update per-inode reservations */
1095         BUG_ON(used  > EXT4_I(inode)->i_reserved_data_blocks);
1096         EXT4_I(inode)->i_reserved_data_blocks -= used;
1097         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1098
1099         /*
1100          * free those over-booking quota for metadata blocks
1101          */
1102         if (mdb_free)
1103                 vfs_dq_release_reservation_block(inode, mdb_free);
1104
1105         /*
1106          * If we have done all the pending block allocations and if
1107          * there aren't any writers on the inode, we can discard the
1108          * inode's preallocations.
1109          */
1110         if (!total && (atomic_read(&inode->i_writecount) == 0))
1111                 ext4_discard_preallocations(inode);
1112 }
1113
1114 static int check_block_validity(struct inode *inode, sector_t logical,
1115                                 sector_t phys, int len)
1116 {
1117         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), phys, len)) {
1118                 ext4_error(inode->i_sb, "check_block_validity",
1119                            "inode #%lu logical block %llu mapped to %llu "
1120                            "(size %d)", inode->i_ino,
1121                            (unsigned long long) logical,
1122                            (unsigned long long) phys, len);
1123                 WARN_ON(1);
1124                 return -EIO;
1125         }
1126         return 0;
1127 }
1128
1129 /*
1130  * The ext4_get_blocks() function tries to look up the requested blocks,
1131  * and returns if the blocks are already mapped.
1132  *
1133  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1134  * and store the allocated blocks in the result buffer head and mark it
1135  * mapped.
1136  *
1137  * If file type is extents based, it will call ext4_ext_get_blocks(),
1138  * Otherwise, call with ext4_ind_get_blocks() to handle indirect mapping
1139  * based files
1140  *
1141  * On success, it returns the number of blocks being mapped or allocate.
1142  * if create==0 and the blocks are pre-allocated and uninitialized block,
1143  * the result buffer head is unmapped. If the create ==1, it will make sure
1144  * the buffer head is mapped.
1145  *
1146  * It returns 0 if plain look up failed (blocks have not been allocated), in
1147  * that casem, buffer head is unmapped
1148  *
1149  * It returns the error in case of allocation failure.
1150  */
1151 int ext4_get_blocks(handle_t *handle, struct inode *inode, sector_t block,
1152                     unsigned int max_blocks, struct buffer_head *bh,
1153                     int flags)
1154 {
1155         int retval;
1156
1157         clear_buffer_mapped(bh);
1158         clear_buffer_unwritten(bh);
1159
1160         /*
1161          * Try to see if we can get the block without requesting a new
1162          * file system block.
1163          */
1164         down_read((&EXT4_I(inode)->i_data_sem));
1165         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1166                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1167                                 bh, 0);
1168         } else {
1169                 retval = ext4_ind_get_blocks(handle, inode, block, max_blocks,
1170                                              bh, 0);
1171         }
1172         up_read((&EXT4_I(inode)->i_data_sem));
1173
1174         if (retval > 0 && buffer_mapped(bh)) {
1175                 int ret = check_block_validity(inode, block, 
1176                                                bh->b_blocknr, retval);
1177                 if (ret != 0)
1178                         return ret;
1179         }
1180
1181         /* If it is only a block(s) look up */
1182         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1183                 return retval;
1184
1185         /*
1186          * Returns if the blocks have already allocated
1187          *
1188          * Note that if blocks have been preallocated
1189          * ext4_ext_get_block() returns th create = 0
1190          * with buffer head unmapped.
1191          */
1192         if (retval > 0 && buffer_mapped(bh))
1193                 return retval;
1194
1195         /*
1196          * When we call get_blocks without the create flag, the
1197          * BH_Unwritten flag could have gotten set if the blocks
1198          * requested were part of a uninitialized extent.  We need to
1199          * clear this flag now that we are committed to convert all or
1200          * part of the uninitialized extent to be an initialized
1201          * extent.  This is because we need to avoid the combination
1202          * of BH_Unwritten and BH_Mapped flags being simultaneously
1203          * set on the buffer_head.
1204          */
1205         clear_buffer_unwritten(bh);
1206
1207         /*
1208          * New blocks allocate and/or writing to uninitialized extent
1209          * will possibly result in updating i_data, so we take
1210          * the write lock of i_data_sem, and call get_blocks()
1211          * with create == 1 flag.
1212          */
1213         down_write((&EXT4_I(inode)->i_data_sem));
1214
1215         /*
1216          * if the caller is from delayed allocation writeout path
1217          * we have already reserved fs blocks for allocation
1218          * let the underlying get_block() function know to
1219          * avoid double accounting
1220          */
1221         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1222                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1223         /*
1224          * We need to check for EXT4 here because migrate
1225          * could have changed the inode type in between
1226          */
1227         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1228                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1229                                               bh, flags);
1230         } else {
1231                 retval = ext4_ind_get_blocks(handle, inode, block,
1232                                              max_blocks, bh, flags);
1233
1234                 if (retval > 0 && buffer_new(bh)) {
1235                         /*
1236                          * We allocated new blocks which will result in
1237                          * i_data's format changing.  Force the migrate
1238                          * to fail by clearing migrate flags
1239                          */
1240                         EXT4_I(inode)->i_flags = EXT4_I(inode)->i_flags &
1241                                                         ~EXT4_EXT_MIGRATE;
1242                 }
1243         }
1244
1245         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1246                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1247
1248         /*
1249          * Update reserved blocks/metadata blocks after successful
1250          * block allocation which had been deferred till now.
1251          */
1252         if ((retval > 0) && (flags & EXT4_GET_BLOCKS_UPDATE_RESERVE_SPACE))
1253                 ext4_da_update_reserve_space(inode, retval);
1254
1255         up_write((&EXT4_I(inode)->i_data_sem));
1256         if (retval > 0 && buffer_mapped(bh)) {
1257                 int ret = check_block_validity(inode, block, 
1258                                                bh->b_blocknr, retval);
1259                 if (ret != 0)
1260                         return ret;
1261         }
1262         return retval;
1263 }
1264
1265 /* Maximum number of blocks we map for direct IO at once. */
1266 #define DIO_MAX_BLOCKS 4096
1267
1268 int ext4_get_block(struct inode *inode, sector_t iblock,
1269                    struct buffer_head *bh_result, int create)
1270 {
1271         handle_t *handle = ext4_journal_current_handle();
1272         int ret = 0, started = 0;
1273         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1274         int dio_credits;
1275
1276         if (create && !handle) {
1277                 /* Direct IO write... */
1278                 if (max_blocks > DIO_MAX_BLOCKS)
1279                         max_blocks = DIO_MAX_BLOCKS;
1280                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
1281                 handle = ext4_journal_start(inode, dio_credits);
1282                 if (IS_ERR(handle)) {
1283                         ret = PTR_ERR(handle);
1284                         goto out;
1285                 }
1286                 started = 1;
1287         }
1288
1289         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
1290                               create ? EXT4_GET_BLOCKS_CREATE : 0);
1291         if (ret > 0) {
1292                 bh_result->b_size = (ret << inode->i_blkbits);
1293                 ret = 0;
1294         }
1295         if (started)
1296                 ext4_journal_stop(handle);
1297 out:
1298         return ret;
1299 }
1300
1301 /*
1302  * `handle' can be NULL if create is zero
1303  */
1304 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1305                                 ext4_lblk_t block, int create, int *errp)
1306 {
1307         struct buffer_head dummy;
1308         int fatal = 0, err;
1309         int flags = 0;
1310
1311         J_ASSERT(handle != NULL || create == 0);
1312
1313         dummy.b_state = 0;
1314         dummy.b_blocknr = -1000;
1315         buffer_trace_init(&dummy.b_history);
1316         if (create)
1317                 flags |= EXT4_GET_BLOCKS_CREATE;
1318         err = ext4_get_blocks(handle, inode, block, 1, &dummy, flags);
1319         /*
1320          * ext4_get_blocks() returns number of blocks mapped. 0 in
1321          * case of a HOLE.
1322          */
1323         if (err > 0) {
1324                 if (err > 1)
1325                         WARN_ON(1);
1326                 err = 0;
1327         }
1328         *errp = err;
1329         if (!err && buffer_mapped(&dummy)) {
1330                 struct buffer_head *bh;
1331                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1332                 if (!bh) {
1333                         *errp = -EIO;
1334                         goto err;
1335                 }
1336                 if (buffer_new(&dummy)) {
1337                         J_ASSERT(create != 0);
1338                         J_ASSERT(handle != NULL);
1339
1340                         /*
1341                          * Now that we do not always journal data, we should
1342                          * keep in mind whether this should always journal the
1343                          * new buffer as metadata.  For now, regular file
1344                          * writes use ext4_get_block instead, so it's not a
1345                          * problem.
1346                          */
1347                         lock_buffer(bh);
1348                         BUFFER_TRACE(bh, "call get_create_access");
1349                         fatal = ext4_journal_get_create_access(handle, bh);
1350                         if (!fatal && !buffer_uptodate(bh)) {
1351                                 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1352                                 set_buffer_uptodate(bh);
1353                         }
1354                         unlock_buffer(bh);
1355                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1356                         err = ext4_handle_dirty_metadata(handle, inode, bh);
1357                         if (!fatal)
1358                                 fatal = err;
1359                 } else {
1360                         BUFFER_TRACE(bh, "not a new buffer");
1361                 }
1362                 if (fatal) {
1363                         *errp = fatal;
1364                         brelse(bh);
1365                         bh = NULL;
1366                 }
1367                 return bh;
1368         }
1369 err:
1370         return NULL;
1371 }
1372
1373 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1374                                ext4_lblk_t block, int create, int *err)
1375 {
1376         struct buffer_head *bh;
1377
1378         bh = ext4_getblk(handle, inode, block, create, err);
1379         if (!bh)
1380                 return bh;
1381         if (buffer_uptodate(bh))
1382                 return bh;
1383         ll_rw_block(READ_META, 1, &bh);
1384         wait_on_buffer(bh);
1385         if (buffer_uptodate(bh))
1386                 return bh;
1387         put_bh(bh);
1388         *err = -EIO;
1389         return NULL;
1390 }
1391
1392 static int walk_page_buffers(handle_t *handle,
1393                              struct buffer_head *head,
1394                              unsigned from,
1395                              unsigned to,
1396                              int *partial,
1397                              int (*fn)(handle_t *handle,
1398                                        struct buffer_head *bh))
1399 {
1400         struct buffer_head *bh;
1401         unsigned block_start, block_end;
1402         unsigned blocksize = head->b_size;
1403         int err, ret = 0;
1404         struct buffer_head *next;
1405
1406         for (bh = head, block_start = 0;
1407              ret == 0 && (bh != head || !block_start);
1408              block_start = block_end, bh = next)
1409         {
1410                 next = bh->b_this_page;
1411                 block_end = block_start + blocksize;
1412                 if (block_end <= from || block_start >= to) {
1413                         if (partial && !buffer_uptodate(bh))
1414                                 *partial = 1;
1415                         continue;
1416                 }
1417                 err = (*fn)(handle, bh);
1418                 if (!ret)
1419                         ret = err;
1420         }
1421         return ret;
1422 }
1423
1424 /*
1425  * To preserve ordering, it is essential that the hole instantiation and
1426  * the data write be encapsulated in a single transaction.  We cannot
1427  * close off a transaction and start a new one between the ext4_get_block()
1428  * and the commit_write().  So doing the jbd2_journal_start at the start of
1429  * prepare_write() is the right place.
1430  *
1431  * Also, this function can nest inside ext4_writepage() ->
1432  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1433  * has generated enough buffer credits to do the whole page.  So we won't
1434  * block on the journal in that case, which is good, because the caller may
1435  * be PF_MEMALLOC.
1436  *
1437  * By accident, ext4 can be reentered when a transaction is open via
1438  * quota file writes.  If we were to commit the transaction while thus
1439  * reentered, there can be a deadlock - we would be holding a quota
1440  * lock, and the commit would never complete if another thread had a
1441  * transaction open and was blocking on the quota lock - a ranking
1442  * violation.
1443  *
1444  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1445  * will _not_ run commit under these circumstances because handle->h_ref
1446  * is elevated.  We'll still have enough credits for the tiny quotafile
1447  * write.
1448  */
1449 static int do_journal_get_write_access(handle_t *handle,
1450                                         struct buffer_head *bh)
1451 {
1452         if (!buffer_mapped(bh) || buffer_freed(bh))
1453                 return 0;
1454         return ext4_journal_get_write_access(handle, bh);
1455 }
1456
1457 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1458                                 loff_t pos, unsigned len, unsigned flags,
1459                                 struct page **pagep, void **fsdata)
1460 {
1461         struct inode *inode = mapping->host;
1462         int ret, needed_blocks;
1463         handle_t *handle;
1464         int retries = 0;
1465         struct page *page;
1466         pgoff_t index;
1467         unsigned from, to;
1468
1469         trace_mark(ext4_write_begin,
1470                    "dev %s ino %lu pos %llu len %u flags %u",
1471                    inode->i_sb->s_id, inode->i_ino,
1472                    (unsigned long long) pos, len, flags);
1473         /*
1474          * Reserve one block more for addition to orphan list in case
1475          * we allocate blocks but write fails for some reason
1476          */
1477         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1478         index = pos >> PAGE_CACHE_SHIFT;
1479         from = pos & (PAGE_CACHE_SIZE - 1);
1480         to = from + len;
1481
1482 retry:
1483         handle = ext4_journal_start(inode, needed_blocks);
1484         if (IS_ERR(handle)) {
1485                 ret = PTR_ERR(handle);
1486                 goto out;
1487         }
1488
1489         /* We cannot recurse into the filesystem as the transaction is already
1490          * started */
1491         flags |= AOP_FLAG_NOFS;
1492
1493         page = grab_cache_page_write_begin(mapping, index, flags);
1494         if (!page) {
1495                 ext4_journal_stop(handle);
1496                 ret = -ENOMEM;
1497                 goto out;
1498         }
1499         *pagep = page;
1500
1501         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1502                                 ext4_get_block);
1503
1504         if (!ret && ext4_should_journal_data(inode)) {
1505                 ret = walk_page_buffers(handle, page_buffers(page),
1506                                 from, to, NULL, do_journal_get_write_access);
1507         }
1508
1509         if (ret) {
1510                 unlock_page(page);
1511                 page_cache_release(page);
1512                 /*
1513                  * block_write_begin may have instantiated a few blocks
1514                  * outside i_size.  Trim these off again. Don't need
1515                  * i_size_read because we hold i_mutex.
1516                  *
1517                  * Add inode to orphan list in case we crash before
1518                  * truncate finishes
1519                  */
1520                 if (pos + len > inode->i_size)
1521                         ext4_orphan_add(handle, inode);
1522
1523                 ext4_journal_stop(handle);
1524                 if (pos + len > inode->i_size) {
1525                         vmtruncate(inode, inode->i_size);
1526                         /* 
1527                          * If vmtruncate failed early the inode might
1528                          * still be on the orphan list; we need to
1529                          * make sure the inode is removed from the
1530                          * orphan list in that case.
1531                          */
1532                         if (inode->i_nlink)
1533                                 ext4_orphan_del(NULL, inode);
1534                 }
1535         }
1536
1537         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1538                 goto retry;
1539 out:
1540         return ret;
1541 }
1542
1543 /* For write_end() in data=journal mode */
1544 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1545 {
1546         if (!buffer_mapped(bh) || buffer_freed(bh))
1547                 return 0;
1548         set_buffer_uptodate(bh);
1549         return ext4_handle_dirty_metadata(handle, NULL, bh);
1550 }
1551
1552 /*
1553  * We need to pick up the new inode size which generic_commit_write gave us
1554  * `file' can be NULL - eg, when called from page_symlink().
1555  *
1556  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1557  * buffers are managed internally.
1558  */
1559 static int ext4_ordered_write_end(struct file *file,
1560                                 struct address_space *mapping,
1561                                 loff_t pos, unsigned len, unsigned copied,
1562                                 struct page *page, void *fsdata)
1563 {
1564         handle_t *handle = ext4_journal_current_handle();
1565         struct inode *inode = mapping->host;
1566         int ret = 0, ret2;
1567
1568         trace_mark(ext4_ordered_write_end,
1569                    "dev %s ino %lu pos %llu len %u copied %u",
1570                    inode->i_sb->s_id, inode->i_ino,
1571                    (unsigned long long) pos, len, copied);
1572         ret = ext4_jbd2_file_inode(handle, inode);
1573
1574         if (ret == 0) {
1575                 loff_t new_i_size;
1576
1577                 new_i_size = pos + copied;
1578                 if (new_i_size > EXT4_I(inode)->i_disksize) {
1579                         ext4_update_i_disksize(inode, new_i_size);
1580                         /* We need to mark inode dirty even if
1581                          * new_i_size is less that inode->i_size
1582                          * bu greater than i_disksize.(hint delalloc)
1583                          */
1584                         ext4_mark_inode_dirty(handle, inode);
1585                 }
1586
1587                 ret2 = generic_write_end(file, mapping, pos, len, copied,
1588                                                         page, fsdata);
1589                 copied = ret2;
1590                 if (ret2 < 0)
1591                         ret = ret2;
1592         }
1593         ret2 = ext4_journal_stop(handle);
1594         if (!ret)
1595                 ret = ret2;
1596
1597         return ret ? ret : copied;
1598 }
1599
1600 static int ext4_writeback_write_end(struct file *file,
1601                                 struct address_space *mapping,
1602                                 loff_t pos, unsigned len, unsigned copied,
1603                                 struct page *page, void *fsdata)
1604 {
1605         handle_t *handle = ext4_journal_current_handle();
1606         struct inode *inode = mapping->host;
1607         int ret = 0, ret2;
1608         loff_t new_i_size;
1609
1610         trace_mark(ext4_writeback_write_end,
1611                    "dev %s ino %lu pos %llu len %u copied %u",
1612                    inode->i_sb->s_id, inode->i_ino,
1613                    (unsigned long long) pos, len, copied);
1614         new_i_size = pos + copied;
1615         if (new_i_size > EXT4_I(inode)->i_disksize) {
1616                 ext4_update_i_disksize(inode, new_i_size);
1617                 /* We need to mark inode dirty even if
1618                  * new_i_size is less that inode->i_size
1619                  * bu greater than i_disksize.(hint delalloc)
1620                  */
1621                 ext4_mark_inode_dirty(handle, inode);
1622         }
1623
1624         ret2 = generic_write_end(file, mapping, pos, len, copied,
1625                                                         page, fsdata);
1626         copied = ret2;
1627         if (ret2 < 0)
1628                 ret = ret2;
1629
1630         ret2 = ext4_journal_stop(handle);
1631         if (!ret)
1632                 ret = ret2;
1633
1634         return ret ? ret : copied;
1635 }
1636
1637 static int ext4_journalled_write_end(struct file *file,
1638                                 struct address_space *mapping,
1639                                 loff_t pos, unsigned len, unsigned copied,
1640                                 struct page *page, void *fsdata)
1641 {
1642         handle_t *handle = ext4_journal_current_handle();
1643         struct inode *inode = mapping->host;
1644         int ret = 0, ret2;
1645         int partial = 0;
1646         unsigned from, to;
1647         loff_t new_i_size;
1648
1649         trace_mark(ext4_journalled_write_end,
1650                    "dev %s ino %lu pos %llu len %u copied %u",
1651                    inode->i_sb->s_id, inode->i_ino,
1652                    (unsigned long long) pos, len, copied);
1653         from = pos & (PAGE_CACHE_SIZE - 1);
1654         to = from + len;
1655
1656         if (copied < len) {
1657                 if (!PageUptodate(page))
1658                         copied = 0;
1659                 page_zero_new_buffers(page, from+copied, to);
1660         }
1661
1662         ret = walk_page_buffers(handle, page_buffers(page), from,
1663                                 to, &partial, write_end_fn);
1664         if (!partial)
1665                 SetPageUptodate(page);
1666         new_i_size = pos + copied;
1667         if (new_i_size > inode->i_size)
1668                 i_size_write(inode, pos+copied);
1669         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
1670         if (new_i_size > EXT4_I(inode)->i_disksize) {
1671                 ext4_update_i_disksize(inode, new_i_size);
1672                 ret2 = ext4_mark_inode_dirty(handle, inode);
1673                 if (!ret)
1674                         ret = ret2;
1675         }
1676
1677         unlock_page(page);
1678         ret2 = ext4_journal_stop(handle);
1679         if (!ret)
1680                 ret = ret2;
1681         page_cache_release(page);
1682
1683         return ret ? ret : copied;
1684 }
1685
1686 static int ext4_da_reserve_space(struct inode *inode, int nrblocks)
1687 {
1688         int retries = 0;
1689         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1690         unsigned long md_needed, mdblocks, total = 0;
1691
1692         /*
1693          * recalculate the amount of metadata blocks to reserve
1694          * in order to allocate nrblocks
1695          * worse case is one extent per block
1696          */
1697 repeat:
1698         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1699         total = EXT4_I(inode)->i_reserved_data_blocks + nrblocks;
1700         mdblocks = ext4_calc_metadata_amount(inode, total);
1701         BUG_ON(mdblocks < EXT4_I(inode)->i_reserved_meta_blocks);
1702
1703         md_needed = mdblocks - EXT4_I(inode)->i_reserved_meta_blocks;
1704         total = md_needed + nrblocks;
1705
1706         /*
1707          * Make quota reservation here to prevent quota overflow
1708          * later. Real quota accounting is done at pages writeout
1709          * time.
1710          */
1711         if (vfs_dq_reserve_block(inode, total)) {
1712                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1713                 return -EDQUOT;
1714         }
1715
1716         if (ext4_claim_free_blocks(sbi, total)) {
1717                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1718                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1719                         yield();
1720                         goto repeat;
1721                 }
1722                 vfs_dq_release_reservation_block(inode, total);
1723                 return -ENOSPC;
1724         }
1725         EXT4_I(inode)->i_reserved_data_blocks += nrblocks;
1726         EXT4_I(inode)->i_reserved_meta_blocks = mdblocks;
1727
1728         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1729         return 0;       /* success */
1730 }
1731
1732 static void ext4_da_release_space(struct inode *inode, int to_free)
1733 {
1734         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1735         int total, mdb, mdb_free, release;
1736
1737         if (!to_free)
1738                 return;         /* Nothing to release, exit */
1739
1740         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1741
1742         if (!EXT4_I(inode)->i_reserved_data_blocks) {
1743                 /*
1744                  * if there is no reserved blocks, but we try to free some
1745                  * then the counter is messed up somewhere.
1746                  * but since this function is called from invalidate
1747                  * page, it's harmless to return without any action
1748                  */
1749                 printk(KERN_INFO "ext4 delalloc try to release %d reserved "
1750                             "blocks for inode %lu, but there is no reserved "
1751                             "data blocks\n", to_free, inode->i_ino);
1752                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1753                 return;
1754         }
1755
1756         /* recalculate the number of metablocks still need to be reserved */
1757         total = EXT4_I(inode)->i_reserved_data_blocks - to_free;
1758         mdb = ext4_calc_metadata_amount(inode, total);
1759
1760         /* figure out how many metablocks to release */
1761         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1762         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1763
1764         release = to_free + mdb_free;
1765
1766         /* update fs dirty blocks counter for truncate case */
1767         percpu_counter_sub(&sbi->s_dirtyblocks_counter, release);
1768
1769         /* update per-inode reservations */
1770         BUG_ON(to_free > EXT4_I(inode)->i_reserved_data_blocks);
1771         EXT4_I(inode)->i_reserved_data_blocks -= to_free;
1772
1773         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1774         EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1775         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1776
1777         vfs_dq_release_reservation_block(inode, release);
1778 }
1779
1780 static void ext4_da_page_release_reservation(struct page *page,
1781                                                 unsigned long offset)
1782 {
1783         int to_release = 0;
1784         struct buffer_head *head, *bh;
1785         unsigned int curr_off = 0;
1786
1787         head = page_buffers(page);
1788         bh = head;
1789         do {
1790                 unsigned int next_off = curr_off + bh->b_size;
1791
1792                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1793                         to_release++;
1794                         clear_buffer_delay(bh);
1795                 }
1796                 curr_off = next_off;
1797         } while ((bh = bh->b_this_page) != head);
1798         ext4_da_release_space(page->mapping->host, to_release);
1799 }
1800
1801 /*
1802  * Delayed allocation stuff
1803  */
1804
1805 struct mpage_da_data {
1806         struct inode *inode;
1807         sector_t b_blocknr;             /* start block number of extent */
1808         size_t b_size;                  /* size of extent */
1809         unsigned long b_state;          /* state of the extent */
1810         unsigned long first_page, next_page;    /* extent of pages */
1811         struct writeback_control *wbc;
1812         int io_done;
1813         int pages_written;
1814         int retval;
1815 };
1816
1817 /*
1818  * mpage_da_submit_io - walks through extent of pages and try to write
1819  * them with writepage() call back
1820  *
1821  * @mpd->inode: inode
1822  * @mpd->first_page: first page of the extent
1823  * @mpd->next_page: page after the last page of the extent
1824  *
1825  * By the time mpage_da_submit_io() is called we expect all blocks
1826  * to be allocated. this may be wrong if allocation failed.
1827  *
1828  * As pages are already locked by write_cache_pages(), we can't use it
1829  */
1830 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1831 {
1832         long pages_skipped;
1833         struct pagevec pvec;
1834         unsigned long index, end;
1835         int ret = 0, err, nr_pages, i;
1836         struct inode *inode = mpd->inode;
1837         struct address_space *mapping = inode->i_mapping;
1838
1839         BUG_ON(mpd->next_page <= mpd->first_page);
1840         /*
1841          * We need to start from the first_page to the next_page - 1
1842          * to make sure we also write the mapped dirty buffer_heads.
1843          * If we look at mpd->b_blocknr we would only be looking
1844          * at the currently mapped buffer_heads.
1845          */
1846         index = mpd->first_page;
1847         end = mpd->next_page - 1;
1848
1849         pagevec_init(&pvec, 0);
1850         while (index <= end) {
1851                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1852                 if (nr_pages == 0)
1853                         break;
1854                 for (i = 0; i < nr_pages; i++) {
1855                         struct page *page = pvec.pages[i];
1856
1857                         index = page->index;
1858                         if (index > end)
1859                                 break;
1860                         index++;
1861
1862                         BUG_ON(!PageLocked(page));
1863                         BUG_ON(PageWriteback(page));
1864
1865                         pages_skipped = mpd->wbc->pages_skipped;
1866                         err = mapping->a_ops->writepage(page, mpd->wbc);
1867                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
1868                                 /*
1869                                  * have successfully written the page
1870                                  * without skipping the same
1871                                  */
1872                                 mpd->pages_written++;
1873                         /*
1874                          * In error case, we have to continue because
1875                          * remaining pages are still locked
1876                          * XXX: unlock and re-dirty them?
1877                          */
1878                         if (ret == 0)
1879                                 ret = err;
1880                 }
1881                 pagevec_release(&pvec);
1882         }
1883         return ret;
1884 }
1885
1886 /*
1887  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
1888  *
1889  * @mpd->inode - inode to walk through
1890  * @exbh->b_blocknr - first block on a disk
1891  * @exbh->b_size - amount of space in bytes
1892  * @logical - first logical block to start assignment with
1893  *
1894  * the function goes through all passed space and put actual disk
1895  * block numbers into buffer heads, dropping BH_Delay and BH_Unwritten
1896  */
1897 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
1898                                  struct buffer_head *exbh)
1899 {
1900         struct inode *inode = mpd->inode;
1901         struct address_space *mapping = inode->i_mapping;
1902         int blocks = exbh->b_size >> inode->i_blkbits;
1903         sector_t pblock = exbh->b_blocknr, cur_logical;
1904         struct buffer_head *head, *bh;
1905         pgoff_t index, end;
1906         struct pagevec pvec;
1907         int nr_pages, i;
1908
1909         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1910         end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1911         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1912
1913         pagevec_init(&pvec, 0);
1914
1915         while (index <= end) {
1916                 /* XXX: optimize tail */
1917                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1918                 if (nr_pages == 0)
1919                         break;
1920                 for (i = 0; i < nr_pages; i++) {
1921                         struct page *page = pvec.pages[i];
1922
1923                         index = page->index;
1924                         if (index > end)
1925                                 break;
1926                         index++;
1927
1928                         BUG_ON(!PageLocked(page));
1929                         BUG_ON(PageWriteback(page));
1930                         BUG_ON(!page_has_buffers(page));
1931
1932                         bh = page_buffers(page);
1933                         head = bh;
1934
1935                         /* skip blocks out of the range */
1936                         do {
1937                                 if (cur_logical >= logical)
1938                                         break;
1939                                 cur_logical++;
1940                         } while ((bh = bh->b_this_page) != head);
1941
1942                         do {
1943                                 if (cur_logical >= logical + blocks)
1944                                         break;
1945
1946                                 if (buffer_delay(bh) ||
1947                                                 buffer_unwritten(bh)) {
1948
1949                                         BUG_ON(bh->b_bdev != inode->i_sb->s_bdev);
1950
1951                                         if (buffer_delay(bh)) {
1952                                                 clear_buffer_delay(bh);
1953                                                 bh->b_blocknr = pblock;
1954                                         } else {
1955                                                 /*
1956                                                  * unwritten already should have
1957                                                  * blocknr assigned. Verify that
1958                                                  */
1959                                                 clear_buffer_unwritten(bh);
1960                                                 BUG_ON(bh->b_blocknr != pblock);
1961                                         }
1962
1963                                 } else if (buffer_mapped(bh))
1964                                         BUG_ON(bh->b_blocknr != pblock);
1965
1966                                 cur_logical++;
1967                                 pblock++;
1968                         } while ((bh = bh->b_this_page) != head);
1969                 }
1970                 pagevec_release(&pvec);
1971         }
1972 }
1973
1974
1975 /*
1976  * __unmap_underlying_blocks - just a helper function to unmap
1977  * set of blocks described by @bh
1978  */
1979 static inline void __unmap_underlying_blocks(struct inode *inode,
1980                                              struct buffer_head *bh)
1981 {
1982         struct block_device *bdev = inode->i_sb->s_bdev;
1983         int blocks, i;
1984
1985         blocks = bh->b_size >> inode->i_blkbits;
1986         for (i = 0; i < blocks; i++)
1987                 unmap_underlying_metadata(bdev, bh->b_blocknr + i);
1988 }
1989
1990 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
1991                                         sector_t logical, long blk_cnt)
1992 {
1993         int nr_pages, i;
1994         pgoff_t index, end;
1995         struct pagevec pvec;
1996         struct inode *inode = mpd->inode;
1997         struct address_space *mapping = inode->i_mapping;
1998
1999         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2000         end   = (logical + blk_cnt - 1) >>
2001                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2002         while (index <= end) {
2003                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2004                 if (nr_pages == 0)
2005                         break;
2006                 for (i = 0; i < nr_pages; i++) {
2007                         struct page *page = pvec.pages[i];
2008                         index = page->index;
2009                         if (index > end)
2010                                 break;
2011                         index++;
2012
2013                         BUG_ON(!PageLocked(page));
2014                         BUG_ON(PageWriteback(page));
2015                         block_invalidatepage(page, 0);
2016                         ClearPageUptodate(page);
2017                         unlock_page(page);
2018                 }
2019         }
2020         return;
2021 }
2022
2023 static void ext4_print_free_blocks(struct inode *inode)
2024 {
2025         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2026         printk(KERN_EMERG "Total free blocks count %lld\n",
2027                         ext4_count_free_blocks(inode->i_sb));
2028         printk(KERN_EMERG "Free/Dirty block details\n");
2029         printk(KERN_EMERG "free_blocks=%lld\n",
2030                         (long long)percpu_counter_sum(&sbi->s_freeblocks_counter));
2031         printk(KERN_EMERG "dirty_blocks=%lld\n",
2032                         (long long)percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2033         printk(KERN_EMERG "Block reservation details\n");
2034         printk(KERN_EMERG "i_reserved_data_blocks=%u\n",
2035                         EXT4_I(inode)->i_reserved_data_blocks);
2036         printk(KERN_EMERG "i_reserved_meta_blocks=%u\n",
2037                         EXT4_I(inode)->i_reserved_meta_blocks);
2038         return;
2039 }
2040
2041 /*
2042  * mpage_da_map_blocks - go through given space
2043  *
2044  * @mpd - bh describing space
2045  *
2046  * The function skips space we know is already mapped to disk blocks.
2047  *
2048  */
2049 static int mpage_da_map_blocks(struct mpage_da_data *mpd)
2050 {
2051         int err, blks, get_blocks_flags;
2052         struct buffer_head new;
2053         sector_t next = mpd->b_blocknr;
2054         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2055         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2056         handle_t *handle = NULL;
2057
2058         /*
2059          * We consider only non-mapped and non-allocated blocks
2060          */
2061         if ((mpd->b_state  & (1 << BH_Mapped)) &&
2062                 !(mpd->b_state & (1 << BH_Delay)) &&
2063                 !(mpd->b_state & (1 << BH_Unwritten)))
2064                 return 0;
2065
2066         /*
2067          * If we didn't accumulate anything to write simply return
2068          */
2069         if (!mpd->b_size)
2070                 return 0;
2071
2072         handle = ext4_journal_current_handle();
2073         BUG_ON(!handle);
2074
2075         /*
2076          * Call ext4_get_blocks() to allocate any delayed allocation
2077          * blocks, or to convert an uninitialized extent to be
2078          * initialized (in the case where we have written into
2079          * one or more preallocated blocks).
2080          *
2081          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2082          * indicate that we are on the delayed allocation path.  This
2083          * affects functions in many different parts of the allocation
2084          * call path.  This flag exists primarily because we don't
2085          * want to change *many* call functions, so ext4_get_blocks()
2086          * will set the magic i_delalloc_reserved_flag once the
2087          * inode's allocation semaphore is taken.
2088          *
2089          * If the blocks in questions were delalloc blocks, set
2090          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2091          * variables are updated after the blocks have been allocated.
2092          */
2093         new.b_state = 0;
2094         get_blocks_flags = (EXT4_GET_BLOCKS_CREATE |
2095                             EXT4_GET_BLOCKS_DELALLOC_RESERVE);
2096         if (mpd->b_state & (1 << BH_Delay))
2097                 get_blocks_flags |= EXT4_GET_BLOCKS_UPDATE_RESERVE_SPACE;
2098         blks = ext4_get_blocks(handle, mpd->inode, next, max_blocks,
2099                                &new, get_blocks_flags);
2100         if (blks < 0) {
2101                 err = blks;
2102                 /*
2103                  * If get block returns with error we simply
2104                  * return. Later writepage will redirty the page and
2105                  * writepages will find the dirty page again
2106                  */
2107                 if (err == -EAGAIN)
2108                         return 0;
2109
2110                 if (err == -ENOSPC &&
2111                     ext4_count_free_blocks(mpd->inode->i_sb)) {
2112                         mpd->retval = err;
2113                         return 0;
2114                 }
2115
2116                 /*
2117                  * get block failure will cause us to loop in
2118                  * writepages, because a_ops->writepage won't be able
2119                  * to make progress. The page will be redirtied by
2120                  * writepage and writepages will again try to write
2121                  * the same.
2122                  */
2123                 printk(KERN_EMERG "%s block allocation failed for inode %lu "
2124                                   "at logical offset %llu with max blocks "
2125                                   "%zd with error %d\n",
2126                                   __func__, mpd->inode->i_ino,
2127                                   (unsigned long long)next,
2128                                   mpd->b_size >> mpd->inode->i_blkbits, err);
2129                 printk(KERN_EMERG "This should not happen.!! "
2130                                         "Data will be lost\n");
2131                 if (err == -ENOSPC) {
2132                         ext4_print_free_blocks(mpd->inode);
2133                 }
2134                 /* invalidate all the pages */
2135                 ext4_da_block_invalidatepages(mpd, next,
2136                                 mpd->b_size >> mpd->inode->i_blkbits);
2137                 return err;
2138         }
2139         BUG_ON(blks == 0);
2140
2141         new.b_size = (blks << mpd->inode->i_blkbits);
2142
2143         if (buffer_new(&new))
2144                 __unmap_underlying_blocks(mpd->inode, &new);
2145
2146         /*
2147          * If blocks are delayed marked, we need to
2148          * put actual blocknr and drop delayed bit
2149          */
2150         if ((mpd->b_state & (1 << BH_Delay)) ||
2151             (mpd->b_state & (1 << BH_Unwritten)))
2152                 mpage_put_bnr_to_bhs(mpd, next, &new);
2153
2154         if (ext4_should_order_data(mpd->inode)) {
2155                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2156                 if (err)
2157                         return err;
2158         }
2159
2160         /*
2161          * Update on-disk size along with block allocation.
2162          */
2163         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2164         if (disksize > i_size_read(mpd->inode))
2165                 disksize = i_size_read(mpd->inode);
2166         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2167                 ext4_update_i_disksize(mpd->inode, disksize);
2168                 return ext4_mark_inode_dirty(handle, mpd->inode);
2169         }
2170
2171         return 0;
2172 }
2173
2174 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2175                 (1 << BH_Delay) | (1 << BH_Unwritten))
2176
2177 /*
2178  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2179  *
2180  * @mpd->lbh - extent of blocks
2181  * @logical - logical number of the block in the file
2182  * @bh - bh of the block (used to access block's state)
2183  *
2184  * the function is used to collect contig. blocks in same state
2185  */
2186 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2187                                    sector_t logical, size_t b_size,
2188                                    unsigned long b_state)
2189 {
2190         sector_t next;
2191         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2192
2193         /* check if thereserved journal credits might overflow */
2194         if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
2195                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2196                         /*
2197                          * With non-extent format we are limited by the journal
2198                          * credit available.  Total credit needed to insert
2199                          * nrblocks contiguous blocks is dependent on the
2200                          * nrblocks.  So limit nrblocks.
2201                          */
2202                         goto flush_it;
2203                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2204                                 EXT4_MAX_TRANS_DATA) {
2205                         /*
2206                          * Adding the new buffer_head would make it cross the
2207                          * allowed limit for which we have journal credit
2208                          * reserved. So limit the new bh->b_size
2209                          */
2210                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2211                                                 mpd->inode->i_blkbits;
2212                         /* we will do mpage_da_submit_io in the next loop */
2213                 }
2214         }
2215         /*
2216          * First block in the extent
2217          */
2218         if (mpd->b_size == 0) {
2219                 mpd->b_blocknr = logical;
2220                 mpd->b_size = b_size;
2221                 mpd->b_state = b_state & BH_FLAGS;
2222                 return;
2223         }
2224
2225         next = mpd->b_blocknr + nrblocks;
2226         /*
2227          * Can we merge the block to our big extent?
2228          */
2229         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2230                 mpd->b_size += b_size;
2231                 return;
2232         }
2233
2234 flush_it:
2235         /*
2236          * We couldn't merge the block to our extent, so we
2237          * need to flush current  extent and start new one
2238          */
2239         if (mpage_da_map_blocks(mpd) == 0)
2240                 mpage_da_submit_io(mpd);
2241         mpd->io_done = 1;
2242         return;
2243 }
2244
2245 static int ext4_bh_unmapped_or_delay(handle_t *handle, struct buffer_head *bh)
2246 {
2247         /*
2248          * unmapped buffer is possible for holes.
2249          * delay buffer is possible with delayed allocation.
2250          * We also need to consider unwritten buffer as unmapped.
2251          */
2252         return (!buffer_mapped(bh) || buffer_delay(bh) ||
2253                                 buffer_unwritten(bh)) && buffer_dirty(bh);
2254 }
2255
2256 /*
2257  * __mpage_da_writepage - finds extent of pages and blocks
2258  *
2259  * @page: page to consider
2260  * @wbc: not used, we just follow rules
2261  * @data: context
2262  *
2263  * The function finds extents of pages and scan them for all blocks.
2264  */
2265 static int __mpage_da_writepage(struct page *page,
2266                                 struct writeback_control *wbc, void *data)
2267 {
2268         struct mpage_da_data *mpd = data;
2269         struct inode *inode = mpd->inode;
2270         struct buffer_head *bh, *head;
2271         sector_t logical;
2272
2273         if (mpd->io_done) {
2274                 /*
2275                  * Rest of the page in the page_vec
2276                  * redirty then and skip then. We will
2277                  * try to to write them again after
2278                  * starting a new transaction
2279                  */
2280                 redirty_page_for_writepage(wbc, page);
2281                 unlock_page(page);
2282                 return MPAGE_DA_EXTENT_TAIL;
2283         }
2284         /*
2285          * Can we merge this page to current extent?
2286          */
2287         if (mpd->next_page != page->index) {
2288                 /*
2289                  * Nope, we can't. So, we map non-allocated blocks
2290                  * and start IO on them using writepage()
2291                  */
2292                 if (mpd->next_page != mpd->first_page) {
2293                         if (mpage_da_map_blocks(mpd) == 0)
2294                                 mpage_da_submit_io(mpd);
2295                         /*
2296                          * skip rest of the page in the page_vec
2297                          */
2298                         mpd->io_done = 1;
2299                         redirty_page_for_writepage(wbc, page);
2300                         unlock_page(page);
2301                         return MPAGE_DA_EXTENT_TAIL;
2302                 }
2303
2304                 /*
2305                  * Start next extent of pages ...
2306                  */
2307                 mpd->first_page = page->index;
2308
2309                 /*
2310                  * ... and blocks
2311                  */
2312                 mpd->b_size = 0;
2313                 mpd->b_state = 0;
2314                 mpd->b_blocknr = 0;
2315         }
2316
2317         mpd->next_page = page->index + 1;
2318         logical = (sector_t) page->index <<
2319                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2320
2321         if (!page_has_buffers(page)) {
2322                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2323                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2324                 if (mpd->io_done)
2325                         return MPAGE_DA_EXTENT_TAIL;
2326         } else {
2327                 /*
2328                  * Page with regular buffer heads, just add all dirty ones
2329                  */
2330                 head = page_buffers(page);
2331                 bh = head;
2332                 do {
2333                         BUG_ON(buffer_locked(bh));
2334                         /*
2335                          * We need to try to allocate
2336                          * unmapped blocks in the same page.
2337                          * Otherwise we won't make progress
2338                          * with the page in ext4_da_writepage
2339                          */
2340                         if (ext4_bh_unmapped_or_delay(NULL, bh)) {
2341                                 mpage_add_bh_to_extent(mpd, logical,
2342                                                        bh->b_size,
2343                                                        bh->b_state);
2344                                 if (mpd->io_done)
2345                                         return MPAGE_DA_EXTENT_TAIL;
2346                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2347                                 /*
2348                                  * mapped dirty buffer. We need to update
2349                                  * the b_state because we look at
2350                                  * b_state in mpage_da_map_blocks. We don't
2351                                  * update b_size because if we find an
2352                                  * unmapped buffer_head later we need to
2353                                  * use the b_state flag of that buffer_head.
2354                                  */
2355                                 if (mpd->b_size == 0)
2356                                         mpd->b_state = bh->b_state & BH_FLAGS;
2357                         }
2358                         logical++;
2359                 } while ((bh = bh->b_this_page) != head);
2360         }
2361
2362         return 0;
2363 }
2364
2365 /*
2366  * This is a special get_blocks_t callback which is used by
2367  * ext4_da_write_begin().  It will either return mapped block or
2368  * reserve space for a single block.
2369  *
2370  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2371  * We also have b_blocknr = -1 and b_bdev initialized properly
2372  *
2373  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2374  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2375  * initialized properly.
2376  */
2377 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2378                                   struct buffer_head *bh_result, int create)
2379 {
2380         int ret = 0;
2381         sector_t invalid_block = ~((sector_t) 0xffff);
2382
2383         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2384                 invalid_block = ~0;
2385
2386         BUG_ON(create == 0);
2387         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2388
2389         /*
2390          * first, we need to know whether the block is allocated already
2391          * preallocated blocks are unmapped but should treated
2392          * the same as allocated blocks.
2393          */
2394         ret = ext4_get_blocks(NULL, inode, iblock, 1,  bh_result, 0);
2395         if ((ret == 0) && !buffer_delay(bh_result)) {
2396                 /* the block isn't (pre)allocated yet, let's reserve space */
2397                 /*
2398                  * XXX: __block_prepare_write() unmaps passed block,
2399                  * is it OK?
2400                  */
2401                 ret = ext4_da_reserve_space(inode, 1);
2402                 if (ret)
2403                         /* not enough space to reserve */
2404                         return ret;
2405
2406                 map_bh(bh_result, inode->i_sb, invalid_block);
2407                 set_buffer_new(bh_result);
2408                 set_buffer_delay(bh_result);
2409         } else if (ret > 0) {
2410                 bh_result->b_size = (ret << inode->i_blkbits);
2411                 if (buffer_unwritten(bh_result)) {
2412                         /* A delayed write to unwritten bh should
2413                          * be marked new and mapped.  Mapped ensures
2414                          * that we don't do get_block multiple times
2415                          * when we write to the same offset and new
2416                          * ensures that we do proper zero out for
2417                          * partial write.
2418                          */
2419                         set_buffer_new(bh_result);
2420                         set_buffer_mapped(bh_result);
2421                 }
2422                 ret = 0;
2423         }
2424
2425         return ret;
2426 }
2427
2428 /*
2429  * This function is used as a standard get_block_t calback function
2430  * when there is no desire to allocate any blocks.  It is used as a
2431  * callback function for block_prepare_write(), nobh_writepage(), and
2432  * block_write_full_page().  These functions should only try to map a
2433  * single block at a time.
2434  *
2435  * Since this function doesn't do block allocations even if the caller
2436  * requests it by passing in create=1, it is critically important that
2437  * any caller checks to make sure that any buffer heads are returned
2438  * by this function are either all already mapped or marked for
2439  * delayed allocation before calling nobh_writepage() or
2440  * block_write_full_page().  Otherwise, b_blocknr could be left
2441  * unitialized, and the page write functions will be taken by
2442  * surprise.
2443  */
2444 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2445                                    struct buffer_head *bh_result, int create)
2446 {
2447         int ret = 0;
2448         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2449
2450         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2451
2452         /*
2453          * we don't want to do block allocation in writepage
2454          * so call get_block_wrap with create = 0
2455          */
2456         ret = ext4_get_blocks(NULL, inode, iblock, max_blocks, bh_result, 0);
2457         BUG_ON(create && ret == 0);
2458         if (ret > 0) {
2459                 bh_result->b_size = (ret << inode->i_blkbits);
2460                 ret = 0;
2461         }
2462         return ret;
2463 }
2464
2465 /*
2466  * This function can get called via...
2467  *   - ext4_da_writepages after taking page lock (have journal handle)
2468  *   - journal_submit_inode_data_buffers (no journal handle)
2469  *   - shrink_page_list via pdflush (no journal handle)
2470  *   - grab_page_cache when doing write_begin (have journal handle)
2471  */
2472 static int ext4_da_writepage(struct page *page,
2473                                 struct writeback_control *wbc)
2474 {
2475         int ret = 0;
2476         loff_t size;
2477         unsigned int len;
2478         struct buffer_head *page_bufs;
2479         struct inode *inode = page->mapping->host;
2480
2481         trace_mark(ext4_da_writepage,
2482                    "dev %s ino %lu page_index %lu",
2483                    inode->i_sb->s_id, inode->i_ino, page->index);
2484         size = i_size_read(inode);
2485         if (page->index == size >> PAGE_CACHE_SHIFT)
2486                 len = size & ~PAGE_CACHE_MASK;
2487         else
2488                 len = PAGE_CACHE_SIZE;
2489
2490         if (page_has_buffers(page)) {
2491                 page_bufs = page_buffers(page);
2492                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2493                                         ext4_bh_unmapped_or_delay)) {
2494                         /*
2495                          * We don't want to do  block allocation
2496                          * So redirty the page and return
2497                          * We may reach here when we do a journal commit
2498                          * via journal_submit_inode_data_buffers.
2499                          * If we don't have mapping block we just ignore
2500                          * them. We can also reach here via shrink_page_list
2501                          */
2502                         redirty_page_for_writepage(wbc, page);
2503                         unlock_page(page);
2504                         return 0;
2505                 }
2506         } else {
2507                 /*
2508                  * The test for page_has_buffers() is subtle:
2509                  * We know the page is dirty but it lost buffers. That means
2510                  * that at some moment in time after write_begin()/write_end()
2511                  * has been called all buffers have been clean and thus they
2512                  * must have been written at least once. So they are all
2513                  * mapped and we can happily proceed with mapping them
2514                  * and writing the page.
2515                  *
2516                  * Try to initialize the buffer_heads and check whether
2517                  * all are mapped and non delay. We don't want to
2518                  * do block allocation here.
2519                  */
2520                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
2521                                           noalloc_get_block_write);
2522                 if (!ret) {
2523                         page_bufs = page_buffers(page);
2524                         /* check whether all are mapped and non delay */
2525                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2526                                                 ext4_bh_unmapped_or_delay)) {
2527                                 redirty_page_for_writepage(wbc, page);
2528                                 unlock_page(page);
2529                                 return 0;
2530                         }
2531                 } else {
2532                         /*
2533                          * We can't do block allocation here
2534                          * so just redity the page and unlock
2535                          * and return
2536                          */
2537                         redirty_page_for_writepage(wbc, page);
2538                         unlock_page(page);
2539                         return 0;
2540                 }
2541                 /* now mark the buffer_heads as dirty and uptodate */
2542                 block_commit_write(page, 0, PAGE_CACHE_SIZE);
2543         }
2544
2545         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2546                 ret = nobh_writepage(page, noalloc_get_block_write, wbc);
2547         else
2548                 ret = block_write_full_page(page, noalloc_get_block_write,
2549                                             wbc);
2550
2551         return ret;
2552 }
2553
2554 /*
2555  * This is called via ext4_da_writepages() to
2556  * calulate the total number of credits to reserve to fit
2557  * a single extent allocation into a single transaction,
2558  * ext4_da_writpeages() will loop calling this before
2559  * the block allocation.
2560  */
2561
2562 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2563 {
2564         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2565
2566         /*
2567          * With non-extent format the journal credit needed to
2568          * insert nrblocks contiguous block is dependent on
2569          * number of contiguous block. So we will limit
2570          * number of contiguous block to a sane value
2571          */
2572         if (!(inode->i_flags & EXT4_EXTENTS_FL) &&
2573             (max_blocks > EXT4_MAX_TRANS_DATA))
2574                 max_blocks = EXT4_MAX_TRANS_DATA;
2575
2576         return ext4_chunk_trans_blocks(inode, max_blocks);
2577 }
2578
2579 static int ext4_da_writepages(struct address_space *mapping,
2580                               struct writeback_control *wbc)
2581 {
2582         pgoff_t index;
2583         int range_whole = 0;
2584         handle_t *handle = NULL;
2585         struct mpage_da_data mpd;
2586         struct inode *inode = mapping->host;
2587         int no_nrwrite_index_update;
2588         int pages_written = 0;
2589         long pages_skipped;
2590         int range_cyclic, cycled = 1, io_done = 0;
2591         int needed_blocks, ret = 0, nr_to_writebump = 0;
2592         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2593
2594         trace_mark(ext4_da_writepages,
2595                    "dev %s ino %lu nr_t_write %ld "
2596                    "pages_skipped %ld range_start %llu "
2597                    "range_end %llu nonblocking %d "
2598                    "for_kupdate %d for_reclaim %d "
2599                    "for_writepages %d range_cyclic %d",
2600                    inode->i_sb->s_id, inode->i_ino,
2601                    wbc->nr_to_write, wbc->pages_skipped,
2602                    (unsigned long long) wbc->range_start,
2603                    (unsigned long long) wbc->range_end,
2604                    wbc->nonblocking, wbc->for_kupdate,
2605                    wbc->for_reclaim, wbc->for_writepages,
2606                    wbc->range_cyclic);
2607
2608         /*
2609          * No pages to write? This is mainly a kludge to avoid starting
2610          * a transaction for special inodes like journal inode on last iput()
2611          * because that could violate lock ordering on umount
2612          */
2613         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2614                 return 0;
2615
2616         /*
2617          * If the filesystem has aborted, it is read-only, so return
2618          * right away instead of dumping stack traces later on that
2619          * will obscure the real source of the problem.  We test
2620          * EXT4_MOUNT_ABORT instead of sb->s_flag's MS_RDONLY because
2621          * the latter could be true if the filesystem is mounted
2622          * read-only, and in that case, ext4_da_writepages should
2623          * *never* be called, so if that ever happens, we would want
2624          * the stack trace.
2625          */
2626         if (unlikely(sbi->s_mount_opt & EXT4_MOUNT_ABORT))
2627                 return -EROFS;
2628
2629         /*
2630          * Make sure nr_to_write is >= sbi->s_mb_stream_request
2631          * This make sure small files blocks are allocated in
2632          * single attempt. This ensure that small files
2633          * get less fragmented.
2634          */
2635         if (wbc->nr_to_write < sbi->s_mb_stream_request) {
2636                 nr_to_writebump = sbi->s_mb_stream_request - wbc->nr_to_write;
2637                 wbc->nr_to_write = sbi->s_mb_stream_request;
2638         }
2639         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2640                 range_whole = 1;
2641
2642         range_cyclic = wbc->range_cyclic;
2643         if (wbc->range_cyclic) {
2644                 index = mapping->writeback_index;
2645                 if (index)
2646                         cycled = 0;
2647                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2648                 wbc->range_end  = LLONG_MAX;
2649                 wbc->range_cyclic = 0;
2650         } else
2651                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2652
2653         mpd.wbc = wbc;
2654         mpd.inode = mapping->host;
2655
2656         /*
2657          * we don't want write_cache_pages to update
2658          * nr_to_write and writeback_index
2659          */
2660         no_nrwrite_index_update = wbc->no_nrwrite_index_update;
2661         wbc->no_nrwrite_index_update = 1;
2662         pages_skipped = wbc->pages_skipped;
2663
2664 retry:
2665         while (!ret && wbc->nr_to_write > 0) {
2666
2667                 /*
2668                  * we  insert one extent at a time. So we need
2669                  * credit needed for single extent allocation.
2670                  * journalled mode is currently not supported
2671                  * by delalloc
2672                  */
2673                 BUG_ON(ext4_should_journal_data(inode));
2674                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2675
2676                 /* start a new transaction*/
2677                 handle = ext4_journal_start(inode, needed_blocks);
2678                 if (IS_ERR(handle)) {
2679                         ret = PTR_ERR(handle);
2680                         printk(KERN_CRIT "%s: jbd2_start: "
2681                                "%ld pages, ino %lu; err %d\n", __func__,
2682                                 wbc->nr_to_write, inode->i_ino, ret);
2683                         dump_stack();
2684                         goto out_writepages;
2685                 }
2686
2687                 /*
2688                  * Now call __mpage_da_writepage to find the next
2689                  * contiguous region of logical blocks that need
2690                  * blocks to be allocated by ext4.  We don't actually
2691                  * submit the blocks for I/O here, even though
2692                  * write_cache_pages thinks it will, and will set the
2693                  * pages as clean for write before calling
2694                  * __mpage_da_writepage().
2695                  */
2696                 mpd.b_size = 0;
2697                 mpd.b_state = 0;
2698                 mpd.b_blocknr = 0;
2699                 mpd.first_page = 0;
2700                 mpd.next_page = 0;
2701                 mpd.io_done = 0;
2702                 mpd.pages_written = 0;
2703                 mpd.retval = 0;
2704                 ret = write_cache_pages(mapping, wbc, __mpage_da_writepage,
2705                                         &mpd);
2706                 /*
2707                  * If we have a contigous extent of pages and we
2708                  * haven't done the I/O yet, map the blocks and submit
2709                  * them for I/O.
2710                  */
2711                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2712                         if (mpage_da_map_blocks(&mpd) == 0)
2713                                 mpage_da_submit_io(&mpd);
2714                         mpd.io_done = 1;
2715                         ret = MPAGE_DA_EXTENT_TAIL;
2716                 }
2717                 wbc->nr_to_write -= mpd.pages_written;
2718
2719                 ext4_journal_stop(handle);
2720
2721                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2722                         /* commit the transaction which would
2723                          * free blocks released in the transaction
2724                          * and try again
2725                          */
2726                         jbd2_journal_force_commit_nested(sbi->s_journal);
2727                         wbc->pages_skipped = pages_skipped;
2728                         ret = 0;
2729                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2730                         /*
2731                          * got one extent now try with
2732                          * rest of the pages
2733                          */
2734                         pages_written += mpd.pages_written;
2735                         wbc->pages_skipped = pages_skipped;
2736                         ret = 0;
2737                         io_done = 1;
2738                 } else if (wbc->nr_to_write)
2739                         /*
2740                          * There is no more writeout needed
2741                          * or we requested for a noblocking writeout
2742                          * and we found the device congested
2743                          */
2744                         break;
2745         }
2746         if (!io_done && !cycled) {
2747                 cycled = 1;
2748                 index = 0;
2749                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2750                 wbc->range_end  = mapping->writeback_index - 1;
2751                 goto retry;
2752         }
2753         if (pages_skipped != wbc->pages_skipped)
2754                 printk(KERN_EMERG "This should not happen leaving %s "
2755                                 "with nr_to_write = %ld ret = %d\n",
2756                                 __func__, wbc->nr_to_write, ret);
2757
2758         /* Update index */
2759         index += pages_written;
2760         wbc->range_cyclic = range_cyclic;
2761         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2762                 /*
2763                  * set the writeback_index so that range_cyclic
2764                  * mode will write it back later
2765                  */
2766                 mapping->writeback_index = index;
2767
2768 out_writepages:
2769         if (!no_nrwrite_index_update)
2770                 wbc->no_nrwrite_index_update = 0;
2771         wbc->nr_to_write -= nr_to_writebump;
2772         trace_mark(ext4_da_writepage_result,
2773                    "dev %s ino %lu ret %d pages_written %d "
2774                    "pages_skipped %ld congestion %d "
2775                    "more_io %d no_nrwrite_index_update %d",
2776                    inode->i_sb->s_id, inode->i_ino, ret,
2777                    pages_written, wbc->pages_skipped,
2778                    wbc->encountered_congestion, wbc->more_io,
2779                    wbc->no_nrwrite_index_update);
2780         return ret;
2781 }
2782
2783 #define FALL_BACK_TO_NONDELALLOC 1
2784 static int ext4_nonda_switch(struct super_block *sb)
2785 {
2786         s64 free_blocks, dirty_blocks;
2787         struct ext4_sb_info *sbi = EXT4_SB(sb);
2788
2789         /*
2790          * switch to non delalloc mode if we are running low
2791          * on free block. The free block accounting via percpu
2792          * counters can get slightly wrong with percpu_counter_batch getting
2793          * accumulated on each CPU without updating global counters
2794          * Delalloc need an accurate free block accounting. So switch
2795          * to non delalloc when we are near to error range.
2796          */
2797         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
2798         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
2799         if (2 * free_blocks < 3 * dirty_blocks ||
2800                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
2801                 /*
2802                  * free block count is less that 150% of dirty blocks
2803                  * or free blocks is less that watermark
2804                  */
2805                 return 1;
2806         }
2807         return 0;
2808 }
2809
2810 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2811                                 loff_t pos, unsigned len, unsigned flags,
2812                                 struct page **pagep, void **fsdata)
2813 {
2814         int ret, retries = 0;
2815         struct page *page;
2816         pgoff_t index;
2817         unsigned from, to;
2818         struct inode *inode = mapping->host;
2819         handle_t *handle;
2820
2821         index = pos >> PAGE_CACHE_SHIFT;
2822         from = pos & (PAGE_CACHE_SIZE - 1);
2823         to = from + len;
2824
2825         if (ext4_nonda_switch(inode->i_sb)) {
2826                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2827                 return ext4_write_begin(file, mapping, pos,
2828                                         len, flags, pagep, fsdata);
2829         }
2830         *fsdata = (void *)0;
2831
2832         trace_mark(ext4_da_write_begin,
2833                    "dev %s ino %lu pos %llu len %u flags %u",
2834                    inode->i_sb->s_id, inode->i_ino,
2835                    (unsigned long long) pos, len, flags);
2836 retry:
2837         /*
2838          * With delayed allocation, we don't log the i_disksize update
2839          * if there is delayed block allocation. But we still need
2840          * to journalling the i_disksize update if writes to the end
2841          * of file which has an already mapped buffer.
2842          */
2843         handle = ext4_journal_start(inode, 1);
2844         if (IS_ERR(handle)) {
2845                 ret = PTR_ERR(handle);
2846                 goto out;
2847         }
2848         /* We cannot recurse into the filesystem as the transaction is already
2849          * started */
2850         flags |= AOP_FLAG_NOFS;
2851
2852         page = grab_cache_page_write_begin(mapping, index, flags);
2853         if (!page) {
2854                 ext4_journal_stop(handle);
2855                 ret = -ENOMEM;
2856                 goto out;
2857         }
2858         *pagep = page;
2859
2860         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
2861                                 ext4_da_get_block_prep);
2862         if (ret < 0) {
2863                 unlock_page(page);
2864                 ext4_journal_stop(handle);
2865                 page_cache_release(page);
2866                 /*
2867                  * block_write_begin may have instantiated a few blocks
2868                  * outside i_size.  Trim these off again. Don't need
2869                  * i_size_read because we hold i_mutex.
2870                  */
2871                 if (pos + len > inode->i_size)
2872                         vmtruncate(inode, inode->i_size);
2873         }
2874
2875         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
2876                 goto retry;
2877 out:
2878         return ret;
2879 }
2880
2881 /*
2882  * Check if we should update i_disksize
2883  * when write to the end of file but not require block allocation
2884  */
2885 static int ext4_da_should_update_i_disksize(struct page *page,
2886                                          unsigned long offset)
2887 {
2888         struct buffer_head *bh;
2889         struct inode *inode = page->mapping->host;
2890         unsigned int idx;
2891         int i;
2892
2893         bh = page_buffers(page);
2894         idx = offset >> inode->i_blkbits;
2895
2896         for (i = 0; i < idx; i++)
2897                 bh = bh->b_this_page;
2898
2899         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
2900                 return 0;
2901         return 1;
2902 }
2903
2904 static int ext4_da_write_end(struct file *file,
2905                                 struct address_space *mapping,
2906                                 loff_t pos, unsigned len, unsigned copied,
2907                                 struct page *page, void *fsdata)
2908 {
2909         struct inode *inode = mapping->host;
2910         int ret = 0, ret2;
2911         handle_t *handle = ext4_journal_current_handle();
2912         loff_t new_i_size;
2913         unsigned long start, end;
2914         int write_mode = (int)(unsigned long)fsdata;
2915
2916         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
2917                 if (ext4_should_order_data(inode)) {
2918                         return ext4_ordered_write_end(file, mapping, pos,
2919                                         len, copied, page, fsdata);
2920                 } else if (ext4_should_writeback_data(inode)) {
2921                         return ext4_writeback_write_end(file, mapping, pos,
2922                                         len, copied, page, fsdata);
2923                 } else {
2924                         BUG();
2925                 }
2926         }
2927
2928         trace_mark(ext4_da_write_end,
2929                    "dev %s ino %lu pos %llu len %u copied %u",
2930                    inode->i_sb->s_id, inode->i_ino,
2931                    (unsigned long long) pos, len, copied);
2932         start = pos & (PAGE_CACHE_SIZE - 1);
2933         end = start + copied - 1;
2934
2935         /*
2936          * generic_write_end() will run mark_inode_dirty() if i_size
2937          * changes.  So let's piggyback the i_disksize mark_inode_dirty
2938          * into that.
2939          */
2940
2941         new_i_size = pos + copied;
2942         if (new_i_size > EXT4_I(inode)->i_disksize) {
2943                 if (ext4_da_should_update_i_disksize(page, end)) {
2944                         down_write(&EXT4_I(inode)->i_data_sem);
2945                         if (new_i_size > EXT4_I(inode)->i_disksize) {
2946                                 /*
2947                                  * Updating i_disksize when extending file
2948                                  * without needing block allocation
2949                                  */
2950                                 if (ext4_should_order_data(inode))
2951                                         ret = ext4_jbd2_file_inode(handle,
2952                                                                    inode);
2953
2954                                 EXT4_I(inode)->i_disksize = new_i_size;
2955                         }
2956                         up_write(&EXT4_I(inode)->i_data_sem);
2957                         /* We need to mark inode dirty even if
2958                          * new_i_size is less that inode->i_size
2959                          * bu greater than i_disksize.(hint delalloc)
2960                          */
2961                         ext4_mark_inode_dirty(handle, inode);
2962                 }
2963         }
2964         ret2 = generic_write_end(file, mapping, pos, len, copied,
2965                                                         page, fsdata);
2966         copied = ret2;
2967         if (ret2 < 0)
2968                 ret = ret2;
2969         ret2 = ext4_journal_stop(handle);
2970         if (!ret)
2971                 ret = ret2;
2972
2973         return ret ? ret : copied;
2974 }
2975
2976 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
2977 {
2978         /*
2979          * Drop reserved blocks
2980          */
2981         BUG_ON(!PageLocked(page));
2982         if (!page_has_buffers(page))
2983                 goto out;
2984
2985         ext4_da_page_release_reservation(page, offset);
2986
2987 out:
2988         ext4_invalidatepage(page, offset);
2989
2990         return;
2991 }
2992
2993 /*
2994  * Force all delayed allocation blocks to be allocated for a given inode.
2995  */
2996 int ext4_alloc_da_blocks(struct inode *inode)
2997 {
2998         if (!EXT4_I(inode)->i_reserved_data_blocks &&
2999             !EXT4_I(inode)->i_reserved_meta_blocks)
3000                 return 0;
3001
3002         /*
3003          * We do something simple for now.  The filemap_flush() will
3004          * also start triggering a write of the data blocks, which is
3005          * not strictly speaking necessary (and for users of
3006          * laptop_mode, not even desirable).  However, to do otherwise
3007          * would require replicating code paths in:
3008          * 
3009          * ext4_da_writepages() ->
3010          *    write_cache_pages() ---> (via passed in callback function)
3011          *        __mpage_da_writepage() -->
3012          *           mpage_add_bh_to_extent()
3013          *           mpage_da_map_blocks()
3014          *
3015          * The problem is that write_cache_pages(), located in
3016          * mm/page-writeback.c, marks pages clean in preparation for
3017          * doing I/O, which is not desirable if we're not planning on
3018          * doing I/O at all.
3019          *
3020          * We could call write_cache_pages(), and then redirty all of
3021          * the pages by calling redirty_page_for_writeback() but that
3022          * would be ugly in the extreme.  So instead we would need to
3023          * replicate parts of the code in the above functions,
3024          * simplifying them becuase we wouldn't actually intend to
3025          * write out the pages, but rather only collect contiguous
3026          * logical block extents, call the multi-block allocator, and
3027          * then update the buffer heads with the block allocations.
3028          * 
3029          * For now, though, we'll cheat by calling filemap_flush(),
3030          * which will map the blocks, and start the I/O, but not
3031          * actually wait for the I/O to complete.
3032          */
3033         return filemap_flush(inode->i_mapping);
3034 }
3035
3036 /*
3037  * bmap() is special.  It gets used by applications such as lilo and by
3038  * the swapper to find the on-disk block of a specific piece of data.
3039  *
3040  * Naturally, this is dangerous if the block concerned is still in the
3041  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3042  * filesystem and enables swap, then they may get a nasty shock when the
3043  * data getting swapped to that swapfile suddenly gets overwritten by
3044  * the original zero's written out previously to the journal and
3045  * awaiting writeback in the kernel's buffer cache.
3046  *
3047  * So, if we see any bmap calls here on a modified, data-journaled file,
3048  * take extra steps to flush any blocks which might be in the cache.
3049  */
3050 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3051 {
3052         struct inode *inode = mapping->host;
3053         journal_t *journal;
3054         int err;
3055
3056         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3057                         test_opt(inode->i_sb, DELALLOC)) {
3058                 /*
3059                  * With delalloc we want to sync the file
3060                  * so that we can make sure we allocate
3061                  * blocks for file
3062                  */
3063                 filemap_write_and_wait(mapping);
3064         }
3065
3066         if (EXT4_JOURNAL(inode) && EXT4_I(inode)->i_state & EXT4_STATE_JDATA) {
3067                 /*
3068                  * This is a REALLY heavyweight approach, but the use of
3069                  * bmap on dirty files is expected to be extremely rare:
3070                  * only if we run lilo or swapon on a freshly made file
3071                  * do we expect this to happen.
3072                  *
3073                  * (bmap requires CAP_SYS_RAWIO so this does not
3074                  * represent an unprivileged user DOS attack --- we'd be
3075                  * in trouble if mortal users could trigger this path at
3076                  * will.)
3077                  *
3078                  * NB. EXT4_STATE_JDATA is not set on files other than
3079                  * regular files.  If somebody wants to bmap a directory
3080                  * or symlink and gets confused because the buffer
3081                  * hasn't yet been flushed to disk, they deserve
3082                  * everything they get.
3083                  */
3084
3085                 EXT4_I(inode)->i_state &= ~EXT4_STATE_JDATA;
3086                 journal = EXT4_JOURNAL(inode);
3087                 jbd2_journal_lock_updates(journal);
3088                 err = jbd2_journal_flush(journal);
3089                 jbd2_journal_unlock_updates(journal);
3090
3091                 if (err)
3092                         return 0;
3093         }
3094
3095         return generic_block_bmap(mapping, block, ext4_get_block);
3096 }
3097
3098 static int bget_one(handle_t *handle, struct buffer_head *bh)
3099 {
3100         get_bh(bh);
3101         return 0;
3102 }
3103
3104 static int bput_one(handle_t *handle, struct buffer_head *bh)
3105 {
3106         put_bh(bh);
3107         return 0;
3108 }
3109
3110 /*
3111  * Note that we don't need to start a transaction unless we're journaling data
3112  * because we should have holes filled from ext4_page_mkwrite(). We even don't
3113  * need to file the inode to the transaction's list in ordered mode because if
3114  * we are writing back data added by write(), the inode is already there and if
3115  * we are writing back data modified via mmap(), noone guarantees in which
3116  * transaction the data will hit the disk. In case we are journaling data, we
3117  * cannot start transaction directly because transaction start ranks above page
3118  * lock so we have to do some magic.
3119  *
3120  * In all journaling modes block_write_full_page() will start the I/O.
3121  *
3122  * Problem:
3123  *
3124  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
3125  *              ext4_writepage()
3126  *
3127  * Similar for:
3128  *
3129  *      ext4_file_write() -> generic_file_write() -> __alloc_pages() -> ...
3130  *
3131  * Same applies to ext4_get_block().  We will deadlock on various things like
3132  * lock_journal and i_data_sem
3133  *
3134  * Setting PF_MEMALLOC here doesn't work - too many internal memory
3135  * allocations fail.
3136  *
3137  * 16May01: If we're reentered then journal_current_handle() will be
3138  *          non-zero. We simply *return*.
3139  *
3140  * 1 July 2001: @@@ FIXME:
3141  *   In journalled data mode, a data buffer may be metadata against the
3142  *   current transaction.  But the same file is part of a shared mapping
3143  *   and someone does a writepage() on it.
3144  *
3145  *   We will move the buffer onto the async_data list, but *after* it has
3146  *   been dirtied. So there's a small window where we have dirty data on
3147  *   BJ_Metadata.
3148  *
3149  *   Note that this only applies to the last partial page in the file.  The
3150  *   bit which block_write_full_page() uses prepare/commit for.  (That's
3151  *   broken code anyway: it's wrong for msync()).
3152  *
3153  *   It's a rare case: affects the final partial page, for journalled data
3154  *   where the file is subject to bith write() and writepage() in the same
3155  *   transction.  To fix it we'll need a custom block_write_full_page().
3156  *   We'll probably need that anyway for journalling writepage() output.
3157  *
3158  * We don't honour synchronous mounts for writepage().  That would be
3159  * disastrous.  Any write() or metadata operation will sync the fs for
3160  * us.
3161  *
3162  */
3163 static int __ext4_normal_writepage(struct page *page,
3164                                 struct writeback_control *wbc)
3165 {
3166         struct inode *inode = page->mapping->host;
3167
3168         if (test_opt(inode->i_sb, NOBH))
3169                 return nobh_writepage(page, noalloc_get_block_write, wbc);
3170         else
3171                 return block_write_full_page(page, noalloc_get_block_write,
3172                                              wbc);
3173 }
3174
3175 static int ext4_normal_writepage(struct page *page,
3176                                 struct writeback_control *wbc)
3177 {
3178         struct inode *inode = page->mapping->host;
3179         loff_t size = i_size_read(inode);
3180         loff_t len;
3181
3182         trace_mark(ext4_normal_writepage,
3183                    "dev %s ino %lu page_index %lu",
3184                    inode->i_sb->s_id, inode->i_ino, page->index);
3185         J_ASSERT(PageLocked(page));
3186         if (page->index == size >> PAGE_CACHE_SHIFT)
3187                 len = size & ~PAGE_CACHE_MASK;
3188         else
3189                 len = PAGE_CACHE_SIZE;
3190
3191         if (page_has_buffers(page)) {
3192                 /* if page has buffers it should all be mapped
3193                  * and allocated. If there are not buffers attached
3194                  * to the page we know the page is dirty but it lost
3195                  * buffers. That means that at some moment in time
3196                  * after write_begin() / write_end() has been called
3197                  * all buffers have been clean and thus they must have been
3198                  * written at least once. So they are all mapped and we can
3199                  * happily proceed with mapping them and writing the page.
3200                  */
3201                 BUG_ON(walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
3202                                         ext4_bh_unmapped_or_delay));
3203         }
3204
3205         if (!ext4_journal_current_handle())
3206                 return __ext4_normal_writepage(page, wbc);
3207
3208         redirty_page_for_writepage(wbc, page);
3209         unlock_page(page);
3210         return 0;
3211 }
3212
3213 static int __ext4_journalled_writepage(struct page *page,
3214                                 struct writeback_control *wbc)
3215 {
3216         struct address_space *mapping = page->mapping;
3217         struct inode *inode = mapping->host;
3218         struct buffer_head *page_bufs;
3219         handle_t *handle = NULL;
3220         int ret = 0;
3221         int err;
3222
3223         ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
3224                                   noalloc_get_block_write);
3225         if (ret != 0)
3226                 goto out_unlock;
3227
3228         page_bufs = page_buffers(page);
3229         walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE, NULL,
3230                                                                 bget_one);
3231         /* As soon as we unlock the page, it can go away, but we have
3232          * references to buffers so we are safe */
3233         unlock_page(page);
3234
3235         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
3236         if (IS_ERR(handle)) {
3237                 ret = PTR_ERR(handle);
3238                 goto out;
3239         }
3240
3241         ret = walk_page_buffers(handle, page_bufs, 0,
3242                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
3243
3244         err = walk_page_buffers(handle, page_bufs, 0,
3245                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
3246         if (ret == 0)
3247                 ret = err;
3248         err = ext4_journal_stop(handle);
3249         if (!ret)
3250                 ret = err;
3251
3252         walk_page_buffers(handle, page_bufs, 0,
3253                                 PAGE_CACHE_SIZE, NULL, bput_one);
3254         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
3255         goto out;
3256
3257 out_unlock:
3258         unlock_page(page);
3259 out:
3260         return ret;
3261 }
3262
3263 static int ext4_journalled_writepage(struct page *page,
3264                                 struct writeback_control *wbc)
3265 {
3266         struct inode *inode = page->mapping->host;
3267         loff_t size = i_size_read(inode);
3268         loff_t len;
3269
3270         trace_mark(ext4_journalled_writepage,
3271                    "dev %s ino %lu page_index %lu",
3272                    inode->i_sb->s_id, inode->i_ino, page->index);
3273         J_ASSERT(PageLocked(page));
3274         if (page->index == size >> PAGE_CACHE_SHIFT)
3275                 len = size & ~PAGE_CACHE_MASK;
3276         else
3277                 len = PAGE_CACHE_SIZE;
3278
3279         if (page_has_buffers(page)) {
3280                 /* if page has buffers it should all be mapped
3281                  * and allocated. If there are not buffers attached
3282                  * to the page we know the page is dirty but it lost
3283                  * buffers. That means that at some moment in time
3284                  * after write_begin() / write_end() has been called
3285                  * all buffers have been clean and thus they must have been
3286                  * written at least once. So they are all mapped and we can
3287                  * happily proceed with mapping them and writing the page.
3288                  */
3289                 BUG_ON(walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
3290                                         ext4_bh_unmapped_or_delay));
3291         }
3292
3293         if (ext4_journal_current_handle())
3294                 goto no_write;
3295
3296         if (PageChecked(page)) {
3297                 /*
3298                  * It's mmapped pagecache.  Add buffers and journal it.  There
3299                  * doesn't seem much point in redirtying the page here.
3300                  */
3301                 ClearPageChecked(page);
3302                 return __ext4_journalled_writepage(page, wbc);
3303         } else {
3304                 /*
3305                  * It may be a page full of checkpoint-mode buffers.  We don't
3306                  * really know unless we go poke around in the buffer_heads.
3307                  * But block_write_full_page will do the right thing.
3308                  */
3309                 return block_write_full_page(page, noalloc_get_block_write,
3310                                              wbc);
3311         }
3312 no_write:
3313         redirty_page_for_writepage(wbc, page);
3314         unlock_page(page);
3315         return 0;
3316 }
3317
3318 static int ext4_readpage(struct file *file, struct page *page)
3319 {
3320         return mpage_readpage(page, ext4_get_block);
3321 }
3322
3323 static int
3324 ext4_readpages(struct file *file, struct address_space *mapping,
3325                 struct list_head *pages, unsigned nr_pages)
3326 {
3327         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3328 }
3329
3330 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3331 {
3332         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3333
3334         /*
3335          * If it's a full truncate we just forget about the pending dirtying
3336          */
3337         if (offset == 0)
3338                 ClearPageChecked(page);
3339
3340         if (journal)
3341                 jbd2_journal_invalidatepage(journal, page, offset);
3342         else
3343                 block_invalidatepage(page, offset);
3344 }
3345
3346 static int ext4_releasepage(struct page *page, gfp_t wait)
3347 {
3348         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3349
3350         WARN_ON(PageChecked(page));
3351         if (!page_has_buffers(page))
3352                 return 0;
3353         if (journal)
3354                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3355         else
3356                 return try_to_free_buffers(page);
3357 }
3358
3359 /*
3360  * If the O_DIRECT write will extend the file then add this inode to the
3361  * orphan list.  So recovery will truncate it back to the original size
3362  * if the machine crashes during the write.
3363  *
3364  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3365  * crashes then stale disk data _may_ be exposed inside the file. But current
3366  * VFS code falls back into buffered path in that case so we are safe.
3367  */
3368 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3369                         const struct iovec *iov, loff_t offset,
3370                         unsigned long nr_segs)
3371 {
3372         struct file *file = iocb->ki_filp;
3373         struct inode *inode = file->f_mapping->host;
3374         struct ext4_inode_info *ei = EXT4_I(inode);
3375         handle_t *handle;
3376         ssize_t ret;
3377         int orphan = 0;
3378         size_t count = iov_length(iov, nr_segs);
3379
3380         if (rw == WRITE) {
3381                 loff_t final_size = offset + count;
3382
3383                 if (final_size > inode->i_size) {
3384                         /* Credits for sb + inode write */
3385                         handle = ext4_journal_start(inode, 2);
3386                         if (IS_ERR(handle)) {
3387                                 ret = PTR_ERR(handle);
3388                                 goto out;
3389                         }
3390                         ret = ext4_orphan_add(handle, inode);
3391                         if (ret) {
3392                                 ext4_journal_stop(handle);
3393                                 goto out;
3394                         }
3395                         orphan = 1;
3396                         ei->i_disksize = inode->i_size;
3397                         ext4_journal_stop(handle);
3398                 }
3399         }
3400
3401         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
3402                                  offset, nr_segs,
3403                                  ext4_get_block, NULL);
3404
3405         if (orphan) {
3406                 int err;
3407
3408                 /* Credits for sb + inode write */
3409                 handle = ext4_journal_start(inode, 2);
3410                 if (IS_ERR(handle)) {
3411                         /* This is really bad luck. We've written the data
3412                          * but cannot extend i_size. Bail out and pretend
3413                          * the write failed... */
3414                         ret = PTR_ERR(handle);
3415                         goto out;
3416                 }
3417                 if (inode->i_nlink)
3418                         ext4_orphan_del(handle, inode);
3419                 if (ret > 0) {
3420                         loff_t end = offset + ret;
3421                         if (end > inode->i_size) {
3422                                 ei->i_disksize = end;
3423                                 i_size_write(inode, end);
3424                                 /*
3425                                  * We're going to return a positive `ret'
3426                                  * here due to non-zero-length I/O, so there's
3427                                  * no way of reporting error returns from
3428                                  * ext4_mark_inode_dirty() to userspace.  So
3429                                  * ignore it.
3430                                  */
3431                                 ext4_mark_inode_dirty(handle, inode);
3432                         }
3433                 }
3434                 err = ext4_journal_stop(handle);
3435                 if (ret == 0)
3436                         ret = err;
3437         }
3438 out:
3439         return ret;
3440 }
3441
3442 /*
3443  * Pages can be marked dirty completely asynchronously from ext4's journalling
3444  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3445  * much here because ->set_page_dirty is called under VFS locks.  The page is
3446  * not necessarily locked.
3447  *
3448  * We cannot just dirty the page and leave attached buffers clean, because the
3449  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3450  * or jbddirty because all the journalling code will explode.
3451  *
3452  * So what we do is to mark the page "pending dirty" and next time writepage
3453  * is called, propagate that into the buffers appropriately.
3454  */
3455 static int ext4_journalled_set_page_dirty(struct page *page)
3456 {
3457         SetPageChecked(page);
3458         return __set_page_dirty_nobuffers(page);
3459 }
3460
3461 static const struct address_space_operations ext4_ordered_aops = {
3462         .readpage               = ext4_readpage,
3463         .readpages              = ext4_readpages,
3464         .writepage              = ext4_normal_writepage,
3465         .sync_page              = block_sync_page,
3466         .write_begin            = ext4_write_begin,
3467         .write_end              = ext4_ordered_write_end,
3468         .bmap                   = ext4_bmap,
3469         .invalidatepage         = ext4_invalidatepage,
3470         .releasepage            = ext4_releasepage,
3471         .direct_IO              = ext4_direct_IO,
3472         .migratepage            = buffer_migrate_page,
3473         .is_partially_uptodate  = block_is_partially_uptodate,
3474 };
3475
3476 static const struct address_space_operations ext4_writeback_aops = {
3477         .readpage               = ext4_readpage,
3478         .readpages              = ext4_readpages,
3479         .writepage              = ext4_normal_writepage,
3480         .sync_page              = block_sync_page,
3481         .write_begin            = ext4_write_begin,
3482         .write_end              = ext4_writeback_write_end,
3483         .bmap                   = ext4_bmap,
3484         .invalidatepage         = ext4_invalidatepage,
3485         .releasepage            = ext4_releasepage,
3486         .direct_IO              = ext4_direct_IO,
3487         .migratepage            = buffer_migrate_page,
3488         .is_partially_uptodate  = block_is_partially_uptodate,
3489 };
3490
3491 static const struct address_space_operations ext4_journalled_aops = {
3492         .readpage               = ext4_readpage,
3493         .readpages              = ext4_readpages,
3494         .writepage              = ext4_journalled_writepage,
3495         .sync_page              = block_sync_page,
3496         .write_begin            = ext4_write_begin,
3497         .write_end              = ext4_journalled_write_end,
3498         .set_page_dirty         = ext4_journalled_set_page_dirty,
3499         .bmap                   = ext4_bmap,
3500         .invalidatepage         = ext4_invalidatepage,
3501         .releasepage            = ext4_releasepage,
3502         .is_partially_uptodate  = block_is_partially_uptodate,
3503 };
3504
3505 static const struct address_space_operations ext4_da_aops = {
3506         .readpage               = ext4_readpage,
3507         .readpages              = ext4_readpages,
3508         .writepage              = ext4_da_writepage,
3509         .writepages             = ext4_da_writepages,
3510         .sync_page              = block_sync_page,
3511         .write_begin            = ext4_da_write_begin,
3512         .write_end              = ext4_da_write_end,
3513         .bmap                   = ext4_bmap,
3514         .invalidatepage         = ext4_da_invalidatepage,
3515         .releasepage            = ext4_releasepage,
3516         .direct_IO              = ext4_direct_IO,
3517         .migratepage            = buffer_migrate_page,
3518         .is_partially_uptodate  = block_is_partially_uptodate,
3519 };
3520
3521 void ext4_set_aops(struct inode *inode)
3522 {
3523         if (ext4_should_order_data(inode) &&
3524                 test_opt(inode->i_sb, DELALLOC))
3525                 inode->i_mapping->a_ops = &ext4_da_aops;
3526         else if (ext4_should_order_data(inode))
3527                 inode->i_mapping->a_ops = &ext4_ordered_aops;
3528         else if (ext4_should_writeback_data(inode) &&
3529                  test_opt(inode->i_sb, DELALLOC))
3530                 inode->i_mapping->a_ops = &ext4_da_aops;
3531         else if (ext4_should_writeback_data(inode))
3532                 inode->i_mapping->a_ops = &ext4_writeback_aops;
3533         else
3534                 inode->i_mapping->a_ops = &ext4_journalled_aops;
3535 }
3536
3537 /*
3538  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3539  * up to the end of the block which corresponds to `from'.
3540  * This required during truncate. We need to physically zero the tail end
3541  * of that block so it doesn't yield old data if the file is later grown.
3542  */
3543 int ext4_block_truncate_page(handle_t *handle,
3544                 struct address_space *mapping, loff_t from)
3545 {
3546         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3547         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3548         unsigned blocksize, length, pos;
3549         ext4_lblk_t iblock;
3550         struct inode *inode = mapping->host;
3551         struct buffer_head *bh;
3552         struct page *page;
3553         int err = 0;
3554
3555         page = grab_cache_page(mapping, from >> PAGE_CACHE_SHIFT);
3556         if (!page)
3557                 return -EINVAL;
3558
3559         blocksize = inode->i_sb->s_blocksize;
3560         length = blocksize - (offset & (blocksize - 1));
3561         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3562
3563         /*
3564          * For "nobh" option,  we can only work if we don't need to
3565          * read-in the page - otherwise we create buffers to do the IO.
3566          */
3567         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
3568              ext4_should_writeback_data(inode) && PageUptodate(page)) {
3569                 zero_user(page, offset, length);
3570                 set_page_dirty(page);
3571                 goto unlock;
3572         }
3573
3574         if (!page_has_buffers(page))
3575                 create_empty_buffers(page, blocksize, 0);
3576
3577         /* Find the buffer that contains "offset" */
3578         bh = page_buffers(page);
3579         pos = blocksize;
3580         while (offset >= pos) {
3581                 bh = bh->b_this_page;
3582                 iblock++;
3583                 pos += blocksize;
3584         }
3585
3586         err = 0;
3587         if (buffer_freed(bh)) {
3588                 BUFFER_TRACE(bh, "freed: skip");
3589                 goto unlock;
3590         }
3591
3592         if (!buffer_mapped(bh)) {
3593                 BUFFER_TRACE(bh, "unmapped");
3594                 ext4_get_block(inode, iblock, bh, 0);
3595                 /* unmapped? It's a hole - nothing to do */
3596                 if (!buffer_mapped(bh)) {
3597                         BUFFER_TRACE(bh, "still unmapped");
3598                         goto unlock;
3599                 }
3600         }
3601
3602         /* Ok, it's mapped. Make sure it's up-to-date */
3603         if (PageUptodate(page))
3604                 set_buffer_uptodate(bh);
3605
3606         if (!buffer_uptodate(bh)) {
3607                 err = -EIO;
3608                 ll_rw_block(READ, 1, &bh);
3609                 wait_on_buffer(bh);
3610                 /* Uhhuh. Read error. Complain and punt. */
3611                 if (!buffer_uptodate(bh))
3612                         goto unlock;
3613         }
3614
3615         if (ext4_should_journal_data(inode)) {
3616                 BUFFER_TRACE(bh, "get write access");
3617                 err = ext4_journal_get_write_access(handle, bh);
3618                 if (err)
3619                         goto unlock;
3620         }
3621
3622         zero_user(page, offset, length);
3623
3624         BUFFER_TRACE(bh, "zeroed end of block");
3625
3626         err = 0;
3627         if (ext4_should_journal_data(inode)) {
3628                 err = ext4_handle_dirty_metadata(handle, inode, bh);
3629         } else {
3630                 if (ext4_should_order_data(inode))
3631                         err = ext4_jbd2_file_inode(handle, inode);
3632                 mark_buffer_dirty(bh);
3633         }
3634
3635 unlock:
3636         unlock_page(page);
3637         page_cache_release(page);
3638         return err;
3639 }
3640
3641 /*
3642  * Probably it should be a library function... search for first non-zero word
3643  * or memcmp with zero_page, whatever is better for particular architecture.
3644  * Linus?
3645  */
3646 static inline int all_zeroes(__le32 *p, __le32 *q)
3647 {
3648         while (p < q)
3649                 if (*p++)
3650                         return 0;
3651         return 1;
3652 }
3653
3654 /**
3655  *      ext4_find_shared - find the indirect blocks for partial truncation.
3656  *      @inode:   inode in question
3657  *      @depth:   depth of the affected branch
3658  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
3659  *      @chain:   place to store the pointers to partial indirect blocks
3660  *      @top:     place to the (detached) top of branch
3661  *
3662  *      This is a helper function used by ext4_truncate().
3663  *
3664  *      When we do truncate() we may have to clean the ends of several
3665  *      indirect blocks but leave the blocks themselves alive. Block is
3666  *      partially truncated if some data below the new i_size is refered
3667  *      from it (and it is on the path to the first completely truncated
3668  *      data block, indeed).  We have to free the top of that path along
3669  *      with everything to the right of the path. Since no allocation
3670  *      past the truncation point is possible until ext4_truncate()
3671  *      finishes, we may safely do the latter, but top of branch may
3672  *      require special attention - pageout below the truncation point
3673  *      might try to populate it.
3674  *
3675  *      We atomically detach the top of branch from the tree, store the
3676  *      block number of its root in *@top, pointers to buffer_heads of
3677  *      partially truncated blocks - in @chain[].bh and pointers to
3678  *      their last elements that should not be removed - in
3679  *      @chain[].p. Return value is the pointer to last filled element
3680  *      of @chain.
3681  *
3682  *      The work left to caller to do the actual freeing of subtrees:
3683  *              a) free the subtree starting from *@top
3684  *              b) free the subtrees whose roots are stored in
3685  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
3686  *              c) free the subtrees growing from the inode past the @chain[0].
3687  *                      (no partially truncated stuff there).  */
3688
3689 static Indirect *ext4_find_shared(struct inode *inode, int depth,
3690                         ext4_lblk_t offsets[4], Indirect chain[4], __le32 *top)
3691 {
3692         Indirect *partial, *p;
3693         int k, err;
3694
3695         *top = 0;
3696         /* Make k index the deepest non-null offest + 1 */
3697         for (k = depth; k > 1 && !offsets[k-1]; k--)
3698                 ;
3699         partial = ext4_get_branch(inode, k, offsets, chain, &err);
3700         /* Writer: pointers */
3701         if (!partial)
3702                 partial = chain + k-1;
3703         /*
3704          * If the branch acquired continuation since we've looked at it -
3705          * fine, it should all survive and (new) top doesn't belong to us.
3706          */
3707         if (!partial->key && *partial->p)
3708                 /* Writer: end */
3709                 goto no_top;
3710         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
3711                 ;
3712         /*
3713          * OK, we've found the last block that must survive. The rest of our
3714          * branch should be detached before unlocking. However, if that rest
3715          * of branch is all ours and does not grow immediately from the inode
3716          * it's easier to cheat and just decrement partial->p.
3717          */
3718         if (p == chain + k - 1 && p > chain) {
3719                 p->p--;
3720         } else {
3721                 *top = *p->p;
3722                 /* Nope, don't do this in ext4.  Must leave the tree intact */
3723 #if 0
3724                 *p->p = 0;
3725 #endif
3726         }
3727         /* Writer: end */
3728
3729         while (partial > p) {
3730                 brelse(partial->bh);
3731                 partial--;
3732         }
3733 no_top:
3734         return partial;
3735 }
3736
3737 /*
3738  * Zero a number of block pointers in either an inode or an indirect block.
3739  * If we restart the transaction we must again get write access to the
3740  * indirect block for further modification.
3741  *
3742  * We release `count' blocks on disk, but (last - first) may be greater
3743  * than `count' because there can be holes in there.
3744  */
3745 static void ext4_clear_blocks(handle_t *handle, struct inode *inode,
3746                 struct buffer_head *bh, ext4_fsblk_t block_to_free,
3747                 unsigned long count, __le32 *first, __le32 *last)
3748 {
3749         __le32 *p;
3750         if (try_to_extend_transaction(handle, inode)) {
3751                 if (bh) {
3752                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
3753                         ext4_handle_dirty_metadata(handle, inode, bh);
3754                 }
3755                 ext4_mark_inode_dirty(handle, inode);
3756                 ext4_journal_test_restart(handle, inode);
3757                 if (bh) {
3758                         BUFFER_TRACE(bh, "retaking write access");
3759                         ext4_journal_get_write_access(handle, bh);
3760                 }
3761         }
3762
3763         /*
3764          * Any buffers which are on the journal will be in memory. We find
3765          * them on the hash table so jbd2_journal_revoke() will run jbd2_journal_forget()
3766          * on them.  We've already detached each block from the file, so
3767          * bforget() in jbd2_journal_forget() should be safe.
3768          *
3769          * AKPM: turn on bforget in jbd2_journal_forget()!!!
3770          */
3771         for (p = first; p < last; p++) {
3772                 u32 nr = le32_to_cpu(*p);
3773                 if (nr) {
3774                         struct buffer_head *tbh;
3775
3776                         *p = 0;
3777                         tbh = sb_find_get_block(inode->i_sb, nr);
3778                         ext4_forget(handle, 0, inode, tbh, nr);
3779                 }
3780         }
3781
3782         ext4_free_blocks(handle, inode, block_to_free, count, 0);
3783 }
3784
3785 /**
3786  * ext4_free_data - free a list of data blocks
3787  * @handle:     handle for this transaction
3788  * @inode:      inode we are dealing with
3789  * @this_bh:    indirect buffer_head which contains *@first and *@last
3790  * @first:      array of block numbers
3791  * @last:       points immediately past the end of array
3792  *
3793  * We are freeing all blocks refered from that array (numbers are stored as
3794  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
3795  *
3796  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
3797  * blocks are contiguous then releasing them at one time will only affect one
3798  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
3799  * actually use a lot of journal space.
3800  *
3801  * @this_bh will be %NULL if @first and @last point into the inode's direct
3802  * block pointers.
3803  */
3804 static void ext4_free_data(handle_t *handle, struct inode *inode,
3805                            struct buffer_head *this_bh,
3806                            __le32 *first, __le32 *last)
3807 {
3808         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
3809         unsigned long count = 0;            /* Number of blocks in the run */
3810         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
3811                                                corresponding to
3812                                                block_to_free */
3813         ext4_fsblk_t nr;                    /* Current block # */
3814         __le32 *p;                          /* Pointer into inode/ind
3815                                                for current block */
3816         int err;
3817
3818         if (this_bh) {                          /* For indirect block */
3819                 BUFFER_TRACE(this_bh, "get_write_access");
3820                 err = ext4_journal_get_write_access(handle, this_bh);
3821                 /* Important: if we can't update the indirect pointers
3822                  * to the blocks, we can't free them. */
3823                 if (err)
3824                         return;
3825         }
3826
3827         for (p = first; p < last; p++) {
3828                 nr = le32_to_cpu(*p);
3829                 if (nr) {
3830                         /* accumulate blocks to free if they're contiguous */
3831                         if (count == 0) {
3832                                 block_to_free = nr;
3833                                 block_to_free_p = p;
3834                                 count = 1;
3835                         } else if (nr == block_to_free + count) {
3836                                 count++;
3837                         } else {
3838                                 ext4_clear_blocks(handle, inode, this_bh,
3839                                                   block_to_free,
3840                                                   count, block_to_free_p, p);
3841                                 block_to_free = nr;
3842                                 block_to_free_p = p;
3843                                 count = 1;
3844                         }
3845                 }
3846         }
3847
3848         if (count > 0)
3849                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
3850                                   count, block_to_free_p, p);
3851
3852         if (this_bh) {
3853                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
3854
3855                 /*
3856                  * The buffer head should have an attached journal head at this
3857                  * point. However, if the data is corrupted and an indirect
3858                  * block pointed to itself, it would have been detached when
3859                  * the block was cleared. Check for this instead of OOPSing.
3860                  */
3861                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
3862                         ext4_handle_dirty_metadata(handle, inode, this_bh);
3863                 else
3864                         ext4_error(inode->i_sb, __func__,
3865                                    "circular indirect block detected, "
3866                                    "inode=%lu, block=%llu",
3867                                    inode->i_ino,
3868                                    (unsigned long long) this_bh->b_blocknr);
3869         }
3870 }
3871
3872 /**
3873  *      ext4_free_branches - free an array of branches
3874  *      @handle: JBD handle for this transaction
3875  *      @inode: inode we are dealing with
3876  *      @parent_bh: the buffer_head which contains *@first and *@last
3877  *      @first: array of block numbers
3878  *      @last:  pointer immediately past the end of array
3879  *      @depth: depth of the branches to free
3880  *
3881  *      We are freeing all blocks refered from these branches (numbers are
3882  *      stored as little-endian 32-bit) and updating @inode->i_blocks
3883  *      appropriately.
3884  */
3885 static void ext4_free_branches(handle_t *handle, struct inode *inode,
3886                                struct buffer_head *parent_bh,
3887                                __le32 *first, __le32 *last, int depth)
3888 {
3889         ext4_fsblk_t nr;
3890         __le32 *p;
3891
3892         if (ext4_handle_is_aborted(handle))
3893                 return;
3894
3895         if (depth--) {
3896                 struct buffer_head *bh;
3897                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
3898                 p = last;
3899                 while (--p >= first) {
3900                         nr = le32_to_cpu(*p);
3901                         if (!nr)
3902                                 continue;               /* A hole */
3903
3904                         /* Go read the buffer for the next level down */
3905                         bh = sb_bread(inode->i_sb, nr);
3906
3907                         /*
3908                          * A read failure? Report error and clear slot
3909                          * (should be rare).
3910                          */
3911                         if (!bh) {
3912                                 ext4_error(inode->i_sb, "ext4_free_branches",
3913                                            "Read failure, inode=%lu, block=%llu",
3914                                            inode->i_ino, nr);
3915                                 continue;
3916                         }
3917
3918                         /* This zaps the entire block.  Bottom up. */
3919                         BUFFER_TRACE(bh, "free child branches");
3920                         ext4_free_branches(handle, inode, bh,
3921                                         (__le32 *) bh->b_data,
3922                                         (__le32 *) bh->b_data + addr_per_block,
3923                                         depth);
3924
3925                         /*
3926                          * We've probably journalled the indirect block several
3927                          * times during the truncate.  But it's no longer
3928                          * needed and we now drop it from the transaction via
3929                          * jbd2_journal_revoke().
3930                          *
3931                          * That's easy if it's exclusively part of this
3932                          * transaction.  But if it's part of the committing
3933                          * transaction then jbd2_journal_forget() will simply
3934                          * brelse() it.  That means that if the underlying
3935                          * block is reallocated in ext4_get_block(),
3936                          * unmap_underlying_metadata() will find this block
3937                          * and will try to get rid of it.  damn, damn.
3938                          *
3939                          * If this block has already been committed to the
3940                          * journal, a revoke record will be written.  And
3941                          * revoke records must be emitted *before* clearing
3942                          * this block's bit in the bitmaps.
3943                          */
3944                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
3945
3946                         /*
3947                          * Everything below this this pointer has been
3948                          * released.  Now let this top-of-subtree go.
3949                          *
3950                          * We want the freeing of this indirect block to be
3951                          * atomic in the journal with the updating of the
3952                          * bitmap block which owns it.  So make some room in
3953                          * the journal.
3954                          *
3955                          * We zero the parent pointer *after* freeing its
3956                          * pointee in the bitmaps, so if extend_transaction()
3957                          * for some reason fails to put the bitmap changes and
3958                          * the release into the same transaction, recovery
3959                          * will merely complain about releasing a free block,
3960                          * rather than leaking blocks.
3961                          */
3962                         if (ext4_handle_is_aborted(handle))
3963                                 return;
3964                         if (try_to_extend_transaction(handle, inode)) {
3965                                 ext4_mark_inode_dirty(handle, inode);
3966                                 ext4_journal_test_restart(handle, inode);
3967                         }
3968
3969                         ext4_free_blocks(handle, inode, nr, 1, 1);
3970
3971                         if (parent_bh) {
3972                                 /*
3973                                  * The block which we have just freed is
3974                                  * pointed to by an indirect block: journal it
3975                                  */
3976                                 BUFFER_TRACE(parent_bh, "get_write_access");
3977                                 if (!ext4_journal_get_write_access(handle,
3978                                                                    parent_bh)){
3979                                         *p = 0;
3980                                         BUFFER_TRACE(parent_bh,
3981                                         "call ext4_handle_dirty_metadata");
3982                                         ext4_handle_dirty_metadata(handle,
3983                                                                    inode,
3984                                                                    parent_bh);
3985                                 }
3986                         }
3987                 }
3988         } else {
3989                 /* We have reached the bottom of the tree. */
3990                 BUFFER_TRACE(parent_bh, "free data blocks");
3991                 ext4_free_data(handle, inode, parent_bh, first, last);
3992         }
3993 }
3994
3995 int ext4_can_truncate(struct inode *inode)
3996 {
3997         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
3998                 return 0;
3999         if (S_ISREG(inode->i_mode))
4000                 return 1;
4001         if (S_ISDIR(inode->i_mode))
4002                 return 1;
4003         if (S_ISLNK(inode->i_mode))
4004                 return !ext4_inode_is_fast_symlink(inode);
4005         return 0;
4006 }
4007
4008 /*
4009  * ext4_truncate()
4010  *
4011  * We block out ext4_get_block() block instantiations across the entire
4012  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4013  * simultaneously on behalf of the same inode.
4014  *
4015  * As we work through the truncate and commmit bits of it to the journal there
4016  * is one core, guiding principle: the file's tree must always be consistent on
4017  * disk.  We must be able to restart the truncate after a crash.
4018  *
4019  * The file's tree may be transiently inconsistent in memory (although it
4020  * probably isn't), but whenever we close off and commit a journal transaction,
4021  * the contents of (the filesystem + the journal) must be consistent and
4022  * restartable.  It's pretty simple, really: bottom up, right to left (although
4023  * left-to-right works OK too).
4024  *
4025  * Note that at recovery time, journal replay occurs *before* the restart of
4026  * truncate against the orphan inode list.
4027  *
4028  * The committed inode has the new, desired i_size (which is the same as
4029  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4030  * that this inode's truncate did not complete and it will again call
4031  * ext4_truncate() to have another go.  So there will be instantiated blocks
4032  * to the right of the truncation point in a crashed ext4 filesystem.  But
4033  * that's fine - as long as they are linked from the inode, the post-crash
4034  * ext4_truncate() run will find them and release them.
4035  */
4036 void ext4_truncate(struct inode *inode)
4037 {
4038         handle_t *handle;
4039         struct ext4_inode_info *ei = EXT4_I(inode);
4040         __le32 *i_data = ei->i_data;
4041         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4042         struct address_space *mapping = inode->i_mapping;
4043         ext4_lblk_t offsets[4];
4044         Indirect chain[4];
4045         Indirect *partial;
4046         __le32 nr = 0;
4047         int n;
4048         ext4_lblk_t last_block;
4049         unsigned blocksize = inode->i_sb->s_blocksize;
4050
4051         if (!ext4_can_truncate(inode))
4052                 return;
4053
4054         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4055                 ei->i_state |= EXT4_STATE_DA_ALLOC_CLOSE;
4056
4057         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
4058                 ext4_ext_truncate(inode);
4059                 return;
4060         }
4061
4062         handle = start_transaction(inode);
4063         if (IS_ERR(handle))
4064                 return;         /* AKPM: return what? */
4065
4066         last_block = (inode->i_size + blocksize-1)
4067                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4068
4069         if (inode->i_size & (blocksize - 1))
4070                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4071                         goto out_stop;
4072
4073         n = ext4_block_to_path(inode, last_block, offsets, NULL);
4074         if (n == 0)
4075                 goto out_stop;  /* error */
4076
4077         /*
4078          * OK.  This truncate is going to happen.  We add the inode to the
4079          * orphan list, so that if this truncate spans multiple transactions,
4080          * and we crash, we will resume the truncate when the filesystem
4081          * recovers.  It also marks the inode dirty, to catch the new size.
4082          *
4083          * Implication: the file must always be in a sane, consistent
4084          * truncatable state while each transaction commits.
4085          */
4086         if (ext4_orphan_add(handle, inode))
4087                 goto out_stop;
4088
4089         /*
4090          * From here we block out all ext4_get_block() callers who want to
4091          * modify the block allocation tree.
4092          */
4093         down_write(&ei->i_data_sem);
4094
4095         ext4_discard_preallocations(inode);
4096
4097         /*
4098          * The orphan list entry will now protect us from any crash which
4099          * occurs before the truncate completes, so it is now safe to propagate
4100          * the new, shorter inode size (held for now in i_size) into the
4101          * on-disk inode. We do this via i_disksize, which is the value which
4102          * ext4 *really* writes onto the disk inode.
4103          */
4104         ei->i_disksize = inode->i_size;
4105
4106         if (n == 1) {           /* direct blocks */
4107                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4108                                i_data + EXT4_NDIR_BLOCKS);
4109                 goto do_indirects;
4110         }
4111
4112         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4113         /* Kill the top of shared branch (not detached) */
4114         if (nr) {
4115                 if (partial == chain) {
4116                         /* Shared branch grows from the inode */
4117                         ext4_free_branches(handle, inode, NULL,
4118                                            &nr, &nr+1, (chain+n-1) - partial);
4119                         *partial->p = 0;
4120                         /*
4121                          * We mark the inode dirty prior to restart,
4122                          * and prior to stop.  No need for it here.
4123                          */
4124                 } else {
4125                         /* Shared branch grows from an indirect block */
4126                         BUFFER_TRACE(partial->bh, "get_write_access");
4127                         ext4_free_branches(handle, inode, partial->bh,
4128                                         partial->p,
4129                                         partial->p+1, (chain+n-1) - partial);
4130                 }
4131         }
4132         /* Clear the ends of indirect blocks on the shared branch */
4133         while (partial > chain) {
4134                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4135                                    (__le32*)partial->bh->b_data+addr_per_block,
4136                                    (chain+n-1) - partial);
4137                 BUFFER_TRACE(partial->bh, "call brelse");
4138                 brelse (partial->bh);
4139                 partial--;
4140         }
4141 do_indirects:
4142         /* Kill the remaining (whole) subtrees */
4143         switch (offsets[0]) {
4144         default:
4145                 nr = i_data[EXT4_IND_BLOCK];
4146                 if (nr) {
4147                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4148                         i_data[EXT4_IND_BLOCK] = 0;
4149                 }
4150         case EXT4_IND_BLOCK:
4151                 nr = i_data[EXT4_DIND_BLOCK];
4152                 if (nr) {
4153                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4154                         i_data[EXT4_DIND_BLOCK] = 0;
4155                 }
4156         case EXT4_DIND_BLOCK:
4157                 nr = i_data[EXT4_TIND_BLOCK];
4158                 if (nr) {
4159                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4160                         i_data[EXT4_TIND_BLOCK] = 0;
4161                 }
4162         case EXT4_TIND_BLOCK:
4163                 ;
4164         }
4165
4166         up_write(&ei->i_data_sem);
4167         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4168         ext4_mark_inode_dirty(handle, inode);
4169
4170         /*
4171          * In a multi-transaction truncate, we only make the final transaction
4172          * synchronous
4173          */
4174         if (IS_SYNC(inode))
4175                 ext4_handle_sync(handle);
4176 out_stop:
4177         /*
4178          * If this was a simple ftruncate(), and the file will remain alive
4179          * then we need to clear up the orphan record which we created above.
4180          * However, if this was a real unlink then we were called by
4181          * ext4_delete_inode(), and we allow that function to clean up the
4182          * orphan info for us.
4183          */
4184         if (inode->i_nlink)
4185                 ext4_orphan_del(handle, inode);
4186
4187         ext4_journal_stop(handle);
4188 }
4189
4190 /*
4191  * ext4_get_inode_loc returns with an extra refcount against the inode's
4192  * underlying buffer_head on success. If 'in_mem' is true, we have all
4193  * data in memory that is needed to recreate the on-disk version of this
4194  * inode.
4195  */
4196 static int __ext4_get_inode_loc(struct inode *inode,
4197                                 struct ext4_iloc *iloc, int in_mem)
4198 {
4199         struct ext4_group_desc  *gdp;
4200         struct buffer_head      *bh;
4201         struct super_block      *sb = inode->i_sb;
4202         ext4_fsblk_t            block;
4203         int                     inodes_per_block, inode_offset;
4204
4205         iloc->bh = NULL;
4206         if (!ext4_valid_inum(sb, inode->i_ino))
4207                 return -EIO;
4208
4209         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4210         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4211         if (!gdp)
4212                 return -EIO;
4213
4214         /*
4215          * Figure out the offset within the block group inode table
4216          */
4217         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4218         inode_offset = ((inode->i_ino - 1) %
4219                         EXT4_INODES_PER_GROUP(sb));
4220         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4221         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4222
4223         bh = sb_getblk(sb, block);
4224         if (!bh) {
4225                 ext4_error(sb, "ext4_get_inode_loc", "unable to read "
4226                            "inode block - inode=%lu, block=%llu",
4227                            inode->i_ino, block);
4228                 return -EIO;
4229         }
4230         if (!buffer_uptodate(bh)) {
4231                 lock_buffer(bh);
4232
4233                 /*
4234                  * If the buffer has the write error flag, we have failed
4235                  * to write out another inode in the same block.  In this
4236                  * case, we don't have to read the block because we may
4237                  * read the old inode data successfully.
4238                  */
4239                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4240                         set_buffer_uptodate(bh);
4241
4242                 if (buffer_uptodate(bh)) {
4243                         /* someone brought it uptodate while we waited */
4244                         unlock_buffer(bh);
4245                         goto has_buffer;
4246                 }
4247
4248                 /*
4249                  * If we have all information of the inode in memory and this
4250                  * is the only valid inode in the block, we need not read the
4251                  * block.
4252                  */
4253                 if (in_mem) {
4254                         struct buffer_head *bitmap_bh;
4255                         int i, start;
4256
4257                         start = inode_offset & ~(inodes_per_block - 1);
4258
4259                         /* Is the inode bitmap in cache? */
4260                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4261                         if (!bitmap_bh)
4262                                 goto make_io;
4263
4264                         /*
4265                          * If the inode bitmap isn't in cache then the
4266                          * optimisation may end up performing two reads instead
4267                          * of one, so skip it.
4268                          */
4269                         if (!buffer_uptodate(bitmap_bh)) {
4270                                 brelse(bitmap_bh);
4271                                 goto make_io;
4272                         }
4273                         for (i = start; i < start + inodes_per_block; i++) {
4274                                 if (i == inode_offset)
4275                                         continue;
4276                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4277                                         break;
4278                         }
4279                         brelse(bitmap_bh);
4280                         if (i == start + inodes_per_block) {
4281                                 /* all other inodes are free, so skip I/O */
4282                                 memset(bh->b_data, 0, bh->b_size);
4283                                 set_buffer_uptodate(bh);
4284                                 unlock_buffer(bh);
4285                                 goto has_buffer;
4286                         }
4287                 }
4288
4289 make_io:
4290                 /*
4291                  * If we need to do any I/O, try to pre-readahead extra
4292                  * blocks from the inode table.
4293                  */
4294                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4295                         ext4_fsblk_t b, end, table;
4296                         unsigned num;
4297
4298                         table = ext4_inode_table(sb, gdp);
4299                         /* s_inode_readahead_blks is always a power of 2 */
4300                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4301                         if (table > b)
4302                                 b = table;
4303                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4304                         num = EXT4_INODES_PER_GROUP(sb);
4305                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4306                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4307                                 num -= ext4_itable_unused_count(sb, gdp);
4308                         table += num / inodes_per_block;
4309                         if (end > table)
4310                                 end = table;
4311                         while (b <= end)
4312                                 sb_breadahead(sb, b++);
4313                 }
4314
4315                 /*
4316                  * There are other valid inodes in the buffer, this inode
4317                  * has in-inode xattrs, or we don't have this inode in memory.
4318                  * Read the block from disk.
4319                  */
4320                 get_bh(bh);
4321                 bh->b_end_io = end_buffer_read_sync;
4322                 submit_bh(READ_META, bh);
4323                 wait_on_buffer(bh);
4324                 if (!buffer_uptodate(bh)) {
4325                         ext4_error(sb, __func__,
4326                                    "unable to read inode block - inode=%lu, "
4327                                    "block=%llu", inode->i_ino, block);
4328                         brelse(bh);
4329                         return -EIO;
4330                 }
4331         }
4332 has_buffer:
4333         iloc->bh = bh;
4334         return 0;
4335 }
4336
4337 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4338 {
4339         /* We have all inode data except xattrs in memory here. */
4340         return __ext4_get_inode_loc(inode, iloc,
4341                 !(EXT4_I(inode)->i_state & EXT4_STATE_XATTR));
4342 }
4343
4344 void ext4_set_inode_flags(struct inode *inode)
4345 {
4346         unsigned int flags = EXT4_I(inode)->i_flags;
4347
4348         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4349         if (flags & EXT4_SYNC_FL)
4350                 inode->i_flags |= S_SYNC;
4351         if (flags & EXT4_APPEND_FL)
4352                 inode->i_flags |= S_APPEND;
4353         if (flags & EXT4_IMMUTABLE_FL)
4354                 inode->i_flags |= S_IMMUTABLE;
4355         if (flags & EXT4_NOATIME_FL)
4356                 inode->i_flags |= S_NOATIME;
4357         if (flags & EXT4_DIRSYNC_FL)
4358                 inode->i_flags |= S_DIRSYNC;
4359 }
4360
4361 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4362 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4363 {
4364         unsigned int flags = ei->vfs_inode.i_flags;
4365
4366         ei->i_flags &= ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4367                         EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|EXT4_DIRSYNC_FL);
4368         if (flags & S_SYNC)
4369                 ei->i_flags |= EXT4_SYNC_FL;
4370         if (flags & S_APPEND)
4371                 ei->i_flags |= EXT4_APPEND_FL;
4372         if (flags & S_IMMUTABLE)
4373                 ei->i_flags |= EXT4_IMMUTABLE_FL;
4374         if (flags & S_NOATIME)
4375                 ei->i_flags |= EXT4_NOATIME_FL;
4376         if (flags & S_DIRSYNC)
4377                 ei->i_flags |= EXT4_DIRSYNC_FL;
4378 }
4379 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4380                                         struct ext4_inode_info *ei)
4381 {
4382         blkcnt_t i_blocks ;
4383         struct inode *inode = &(ei->vfs_inode);
4384         struct super_block *sb = inode->i_sb;
4385
4386         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4387                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4388                 /* we are using combined 48 bit field */
4389                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4390                                         le32_to_cpu(raw_inode->i_blocks_lo);
4391                 if (ei->i_flags & EXT4_HUGE_FILE_FL) {
4392                         /* i_blocks represent file system block size */
4393                         return i_blocks  << (inode->i_blkbits - 9);
4394                 } else {
4395                         return i_blocks;
4396                 }
4397         } else {
4398                 return le32_to_cpu(raw_inode->i_blocks_lo);
4399         }
4400 }
4401
4402 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4403 {
4404         struct ext4_iloc iloc;
4405         struct ext4_inode *raw_inode;
4406         struct ext4_inode_info *ei;
4407         struct buffer_head *bh;
4408         struct inode *inode;
4409         long ret;
4410         int block;
4411
4412         inode = iget_locked(sb, ino);
4413         if (!inode)
4414                 return ERR_PTR(-ENOMEM);
4415         if (!(inode->i_state & I_NEW))
4416                 return inode;
4417
4418         ei = EXT4_I(inode);
4419 #ifdef CONFIG_EXT4_FS_POSIX_ACL
4420         ei->i_acl = EXT4_ACL_NOT_CACHED;
4421         ei->i_default_acl = EXT4_ACL_NOT_CACHED;
4422 #endif
4423
4424         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4425         if (ret < 0)
4426                 goto bad_inode;
4427         bh = iloc.bh;
4428         raw_inode = ext4_raw_inode(&iloc);
4429         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4430         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4431         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4432         if (!(test_opt(inode->i_sb, NO_UID32))) {
4433                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4434                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4435         }
4436         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
4437
4438         ei->i_state = 0;
4439         ei->i_dir_start_lookup = 0;
4440         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4441         /* We now have enough fields to check if the inode was active or not.
4442          * This is needed because nfsd might try to access dead inodes
4443          * the test is that same one that e2fsck uses
4444          * NeilBrown 1999oct15
4445          */
4446         if (inode->i_nlink == 0) {
4447                 if (inode->i_mode == 0 ||
4448                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
4449                         /* this inode is deleted */
4450                         brelse(bh);
4451                         ret = -ESTALE;
4452                         goto bad_inode;
4453                 }
4454                 /* The only unlinked inodes we let through here have
4455                  * valid i_mode and are being read by the orphan
4456                  * recovery code: that's fine, we're about to complete
4457                  * the process of deleting those. */
4458         }
4459         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4460         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4461         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4462         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
4463                 ei->i_file_acl |=
4464                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4465         inode->i_size = ext4_isize(raw_inode);
4466         ei->i_disksize = inode->i_size;
4467         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4468         ei->i_block_group = iloc.block_group;
4469         ei->i_last_alloc_group = ~0;
4470         /*
4471          * NOTE! The in-memory inode i_data array is in little-endian order
4472          * even on big-endian machines: we do NOT byteswap the block numbers!
4473          */
4474         for (block = 0; block < EXT4_N_BLOCKS; block++)
4475                 ei->i_data[block] = raw_inode->i_block[block];
4476         INIT_LIST_HEAD(&ei->i_orphan);
4477
4478         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4479                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4480                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4481                     EXT4_INODE_SIZE(inode->i_sb)) {
4482                         brelse(bh);
4483                         ret = -EIO;
4484                         goto bad_inode;
4485                 }
4486                 if (ei->i_extra_isize == 0) {
4487                         /* The extra space is currently unused. Use it. */
4488                         ei->i_extra_isize = sizeof(struct ext4_inode) -
4489                                             EXT4_GOOD_OLD_INODE_SIZE;
4490                 } else {
4491                         __le32 *magic = (void *)raw_inode +
4492                                         EXT4_GOOD_OLD_INODE_SIZE +
4493                                         ei->i_extra_isize;
4494                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
4495                                  ei->i_state |= EXT4_STATE_XATTR;
4496                 }
4497         } else
4498                 ei->i_extra_isize = 0;
4499
4500         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4501         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4502         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4503         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4504
4505         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4506         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4507                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4508                         inode->i_version |=
4509                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4510         }
4511
4512         ret = 0;
4513         if (ei->i_file_acl &&
4514             ((ei->i_file_acl < 
4515               (le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block) +
4516                EXT4_SB(sb)->s_gdb_count)) ||
4517              (ei->i_file_acl >= ext4_blocks_count(EXT4_SB(sb)->s_es)))) {
4518                 ext4_error(sb, __func__,
4519                            "bad extended attribute block %llu in inode #%lu",
4520                            ei->i_file_acl, inode->i_ino);
4521                 ret = -EIO;
4522                 goto bad_inode;
4523         } else if (ei->i_flags & EXT4_EXTENTS_FL) {
4524                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4525                     (S_ISLNK(inode->i_mode) &&
4526                      !ext4_inode_is_fast_symlink(inode)))
4527                         /* Validate extent which is part of inode */
4528                         ret = ext4_ext_check_inode(inode);
4529         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4530                    (S_ISLNK(inode->i_mode) &&
4531                     !ext4_inode_is_fast_symlink(inode))) {
4532                 /* Validate block references which are part of inode */
4533                 ret = ext4_check_inode_blockref(inode);
4534         }
4535         if (ret) {
4536                 brelse(bh);
4537                 goto bad_inode;
4538         }
4539
4540         if (S_ISREG(inode->i_mode)) {
4541                 inode->i_op = &ext4_file_inode_operations;
4542                 inode->i_fop = &ext4_file_operations;
4543                 ext4_set_aops(inode);
4544         } else if (S_ISDIR(inode->i_mode)) {
4545                 inode->i_op = &ext4_dir_inode_operations;
4546                 inode->i_fop = &ext4_dir_operations;
4547         } else if (S_ISLNK(inode->i_mode)) {
4548                 if (ext4_inode_is_fast_symlink(inode)) {
4549                         inode->i_op = &ext4_fast_symlink_inode_operations;
4550                         nd_terminate_link(ei->i_data, inode->i_size,
4551                                 sizeof(ei->i_data) - 1);
4552                 } else {
4553                         inode->i_op = &ext4_symlink_inode_operations;
4554                         ext4_set_aops(inode);
4555                 }
4556         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
4557               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
4558                 inode->i_op = &ext4_special_inode_operations;
4559                 if (raw_inode->i_block[0])
4560                         init_special_inode(inode, inode->i_mode,
4561                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4562                 else
4563                         init_special_inode(inode, inode->i_mode,
4564                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4565         } else {
4566                 brelse(bh);
4567                 ret = -EIO;
4568                 ext4_error(inode->i_sb, __func__, 
4569                            "bogus i_mode (%o) for inode=%lu",
4570                            inode->i_mode, inode->i_ino);
4571                 goto bad_inode;
4572         }
4573         brelse(iloc.bh);
4574         ext4_set_inode_flags(inode);
4575         unlock_new_inode(inode);
4576         return inode;
4577
4578 bad_inode:
4579         iget_failed(inode);
4580         return ERR_PTR(ret);
4581 }
4582
4583 static int ext4_inode_blocks_set(handle_t *handle,
4584                                 struct ext4_inode *raw_inode,
4585                                 struct ext4_inode_info *ei)
4586 {
4587         struct inode *inode = &(ei->vfs_inode);
4588         u64 i_blocks = inode->i_blocks;
4589         struct super_block *sb = inode->i_sb;
4590
4591         if (i_blocks <= ~0U) {
4592                 /*
4593                  * i_blocks can be represnted in a 32 bit variable
4594                  * as multiple of 512 bytes
4595                  */
4596                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4597                 raw_inode->i_blocks_high = 0;
4598                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4599                 return 0;
4600         }
4601         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
4602                 return -EFBIG;
4603
4604         if (i_blocks <= 0xffffffffffffULL) {
4605                 /*
4606                  * i_blocks can be represented in a 48 bit variable
4607                  * as multiple of 512 bytes
4608                  */
4609                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4610                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4611                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4612         } else {
4613                 ei->i_flags |= EXT4_HUGE_FILE_FL;
4614                 /* i_block is stored in file system block size */
4615                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4616                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4617                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4618         }
4619         return 0;
4620 }
4621
4622 /*
4623  * Post the struct inode info into an on-disk inode location in the
4624  * buffer-cache.  This gobbles the caller's reference to the
4625  * buffer_head in the inode location struct.
4626  *
4627  * The caller must have write access to iloc->bh.
4628  */
4629 static int ext4_do_update_inode(handle_t *handle,
4630                                 struct inode *inode,
4631                                 struct ext4_iloc *iloc)
4632 {
4633         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
4634         struct ext4_inode_info *ei = EXT4_I(inode);
4635         struct buffer_head *bh = iloc->bh;
4636         int err = 0, rc, block;
4637
4638         /* For fields not not tracking in the in-memory inode,
4639          * initialise them to zero for new inodes. */
4640         if (ei->i_state & EXT4_STATE_NEW)
4641                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
4642
4643         ext4_get_inode_flags(ei);
4644         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
4645         if (!(test_opt(inode->i_sb, NO_UID32))) {
4646                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
4647                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
4648 /*
4649  * Fix up interoperability with old kernels. Otherwise, old inodes get
4650  * re-used with the upper 16 bits of the uid/gid intact
4651  */
4652                 if (!ei->i_dtime) {
4653                         raw_inode->i_uid_high =
4654                                 cpu_to_le16(high_16_bits(inode->i_uid));
4655                         raw_inode->i_gid_high =
4656                                 cpu_to_le16(high_16_bits(inode->i_gid));
4657                 } else {
4658                         raw_inode->i_uid_high = 0;
4659                         raw_inode->i_gid_high = 0;
4660                 }
4661         } else {
4662                 raw_inode->i_uid_low =
4663                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
4664                 raw_inode->i_gid_low =
4665                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
4666                 raw_inode->i_uid_high = 0;
4667                 raw_inode->i_gid_high = 0;
4668         }
4669         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
4670
4671         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
4672         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
4673         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
4674         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
4675
4676         if (ext4_inode_blocks_set(handle, raw_inode, ei))
4677                 goto out_brelse;
4678         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
4679         /* clear the migrate flag in the raw_inode */
4680         raw_inode->i_flags = cpu_to_le32(ei->i_flags & ~EXT4_EXT_MIGRATE);
4681         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
4682             cpu_to_le32(EXT4_OS_HURD))
4683                 raw_inode->i_file_acl_high =
4684                         cpu_to_le16(ei->i_file_acl >> 32);
4685         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
4686         ext4_isize_set(raw_inode, ei->i_disksize);
4687         if (ei->i_disksize > 0x7fffffffULL) {
4688                 struct super_block *sb = inode->i_sb;
4689                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
4690                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
4691                                 EXT4_SB(sb)->s_es->s_rev_level ==
4692                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
4693                         /* If this is the first large file
4694                          * created, add a flag to the superblock.
4695                          */
4696                         err = ext4_journal_get_write_access(handle,
4697                                         EXT4_SB(sb)->s_sbh);
4698                         if (err)
4699                                 goto out_brelse;
4700                         ext4_update_dynamic_rev(sb);
4701                         EXT4_SET_RO_COMPAT_FEATURE(sb,
4702                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
4703                         sb->s_dirt = 1;
4704                         ext4_handle_sync(handle);
4705                         err = ext4_handle_dirty_metadata(handle, inode,
4706                                         EXT4_SB(sb)->s_sbh);
4707                 }
4708         }
4709         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
4710         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
4711                 if (old_valid_dev(inode->i_rdev)) {
4712                         raw_inode->i_block[0] =
4713                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
4714                         raw_inode->i_block[1] = 0;
4715                 } else {
4716                         raw_inode->i_block[0] = 0;
4717                         raw_inode->i_block[1] =
4718                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
4719                         raw_inode->i_block[2] = 0;
4720                 }
4721         } else for (block = 0; block < EXT4_N_BLOCKS; block++)
4722                 raw_inode->i_block[block] = ei->i_data[block];
4723
4724         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
4725         if (ei->i_extra_isize) {
4726                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4727                         raw_inode->i_version_hi =
4728                         cpu_to_le32(inode->i_version >> 32);
4729                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
4730         }
4731
4732         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4733         rc = ext4_handle_dirty_metadata(handle, inode, bh);
4734         if (!err)
4735                 err = rc;
4736         ei->i_state &= ~EXT4_STATE_NEW;
4737
4738 out_brelse:
4739         brelse(bh);
4740         ext4_std_error(inode->i_sb, err);
4741         return err;
4742 }
4743
4744 /*
4745  * ext4_write_inode()
4746  *
4747  * We are called from a few places:
4748  *
4749  * - Within generic_file_write() for O_SYNC files.
4750  *   Here, there will be no transaction running. We wait for any running
4751  *   trasnaction to commit.
4752  *
4753  * - Within sys_sync(), kupdate and such.
4754  *   We wait on commit, if tol to.
4755  *
4756  * - Within prune_icache() (PF_MEMALLOC == true)
4757  *   Here we simply return.  We can't afford to block kswapd on the
4758  *   journal commit.
4759  *
4760  * In all cases it is actually safe for us to return without doing anything,
4761  * because the inode has been copied into a raw inode buffer in
4762  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
4763  * knfsd.
4764  *
4765  * Note that we are absolutely dependent upon all inode dirtiers doing the
4766  * right thing: they *must* call mark_inode_dirty() after dirtying info in
4767  * which we are interested.
4768  *
4769  * It would be a bug for them to not do this.  The code:
4770  *
4771  *      mark_inode_dirty(inode)
4772  *      stuff();
4773  *      inode->i_size = expr;
4774  *
4775  * is in error because a kswapd-driven write_inode() could occur while
4776  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
4777  * will no longer be on the superblock's dirty inode list.
4778  */
4779 int ext4_write_inode(struct inode *inode, int wait)
4780 {
4781         if (current->flags & PF_MEMALLOC)
4782                 return 0;
4783
4784         if (ext4_journal_current_handle()) {
4785                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
4786                 dump_stack();
4787                 return -EIO;
4788         }
4789
4790         if (!wait)
4791                 return 0;
4792
4793         return ext4_force_commit(inode->i_sb);
4794 }
4795
4796 /*
4797  * ext4_setattr()
4798  *
4799  * Called from notify_change.
4800  *
4801  * We want to trap VFS attempts to truncate the file as soon as
4802  * possible.  In particular, we want to make sure that when the VFS
4803  * shrinks i_size, we put the inode on the orphan list and modify
4804  * i_disksize immediately, so that during the subsequent flushing of
4805  * dirty pages and freeing of disk blocks, we can guarantee that any
4806  * commit will leave the blocks being flushed in an unused state on
4807  * disk.  (On recovery, the inode will get truncated and the blocks will
4808  * be freed, so we have a strong guarantee that no future commit will
4809  * leave these blocks visible to the user.)
4810  *
4811  * Another thing we have to assure is that if we are in ordered mode
4812  * and inode is still attached to the committing transaction, we must
4813  * we start writeout of all the dirty pages which are being truncated.
4814  * This way we are sure that all the data written in the previous
4815  * transaction are already on disk (truncate waits for pages under
4816  * writeback).
4817  *
4818  * Called with inode->i_mutex down.
4819  */
4820 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
4821 {
4822         struct inode *inode = dentry->d_inode;
4823         int error, rc = 0;
4824         const unsigned int ia_valid = attr->ia_valid;
4825
4826         error = inode_change_ok(inode, attr);
4827         if (error)
4828                 return error;
4829
4830         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
4831                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
4832                 handle_t *handle;
4833
4834                 /* (user+group)*(old+new) structure, inode write (sb,
4835                  * inode block, ? - but truncate inode update has it) */
4836                 handle = ext4_journal_start(inode, 2*(EXT4_QUOTA_INIT_BLOCKS(inode->i_sb)+
4837                                         EXT4_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
4838                 if (IS_ERR(handle)) {
4839                         error = PTR_ERR(handle);
4840                         goto err_out;
4841                 }
4842                 error = vfs_dq_transfer(inode, attr) ? -EDQUOT : 0;
4843                 if (error) {
4844                         ext4_journal_stop(handle);
4845                         return error;
4846                 }
4847                 /* Update corresponding info in inode so that everything is in
4848                  * one transaction */
4849                 if (attr->ia_valid & ATTR_UID)
4850                         inode->i_uid = attr->ia_uid;
4851                 if (attr->ia_valid & ATTR_GID)
4852                         inode->i_gid = attr->ia_gid;
4853                 error = ext4_mark_inode_dirty(handle, inode);
4854                 ext4_journal_stop(handle);
4855         }
4856
4857         if (attr->ia_valid & ATTR_SIZE) {
4858                 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)) {
4859                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4860
4861                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
4862                                 error = -EFBIG;
4863                                 goto err_out;
4864                         }
4865                 }
4866         }
4867
4868         if (S_ISREG(inode->i_mode) &&
4869             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
4870                 handle_t *handle;
4871
4872                 handle = ext4_journal_start(inode, 3);
4873                 if (IS_ERR(handle)) {
4874                         error = PTR_ERR(handle);
4875                         goto err_out;
4876                 }
4877
4878                 error = ext4_orphan_add(handle, inode);
4879                 EXT4_I(inode)->i_disksize = attr->ia_size;
4880                 rc = ext4_mark_inode_dirty(handle, inode);
4881                 if (!error)
4882                         error = rc;
4883                 ext4_journal_stop(handle);
4884
4885                 if (ext4_should_order_data(inode)) {
4886                         error = ext4_begin_ordered_truncate(inode,
4887                                                             attr->ia_size);
4888                         if (error) {
4889                                 /* Do as much error cleanup as possible */
4890                                 handle = ext4_journal_start(inode, 3);
4891                                 if (IS_ERR(handle)) {
4892                                         ext4_orphan_del(NULL, inode);
4893                                         goto err_out;
4894                                 }
4895                                 ext4_orphan_del(handle, inode);
4896                                 ext4_journal_stop(handle);
4897                                 goto err_out;
4898                         }
4899                 }
4900         }
4901
4902         rc = inode_setattr(inode, attr);
4903
4904         /* If inode_setattr's call to ext4_truncate failed to get a
4905          * transaction handle at all, we need to clean up the in-core
4906          * orphan list manually. */
4907         if (inode->i_nlink)
4908                 ext4_orphan_del(NULL, inode);
4909
4910         if (!rc && (ia_valid & ATTR_MODE))
4911                 rc = ext4_acl_chmod(inode);
4912
4913 err_out:
4914         ext4_std_error(inode->i_sb, error);
4915         if (!error)
4916                 error = rc;
4917         return error;
4918 }
4919
4920 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
4921                  struct kstat *stat)
4922 {
4923         struct inode *inode;
4924         unsigned long delalloc_blocks;
4925
4926         inode = dentry->d_inode;
4927         generic_fillattr(inode, stat);
4928
4929         /*
4930          * We can't update i_blocks if the block allocation is delayed
4931          * otherwise in the case of system crash before the real block
4932          * allocation is done, we will have i_blocks inconsistent with
4933          * on-disk file blocks.
4934          * We always keep i_blocks updated together with real
4935          * allocation. But to not confuse with user, stat
4936          * will return the blocks that include the delayed allocation
4937          * blocks for this file.
4938          */
4939         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
4940         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
4941         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
4942
4943         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
4944         return 0;
4945 }
4946
4947 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
4948                                       int chunk)
4949 {
4950         int indirects;
4951
4952         /* if nrblocks are contiguous */
4953         if (chunk) {
4954                 /*
4955                  * With N contiguous data blocks, it need at most
4956                  * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) indirect blocks
4957                  * 2 dindirect blocks
4958                  * 1 tindirect block
4959                  */
4960                 indirects = nrblocks / EXT4_ADDR_PER_BLOCK(inode->i_sb);
4961                 return indirects + 3;
4962         }
4963         /*
4964          * if nrblocks are not contiguous, worse case, each block touch
4965          * a indirect block, and each indirect block touch a double indirect
4966          * block, plus a triple indirect block
4967          */
4968         indirects = nrblocks * 2 + 1;
4969         return indirects;
4970 }
4971
4972 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4973 {
4974         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL))
4975                 return ext4_indirect_trans_blocks(inode, nrblocks, chunk);
4976         return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
4977 }
4978
4979 /*
4980  * Account for index blocks, block groups bitmaps and block group
4981  * descriptor blocks if modify datablocks and index blocks
4982  * worse case, the indexs blocks spread over different block groups
4983  *
4984  * If datablocks are discontiguous, they are possible to spread over
4985  * different block groups too. If they are contiugous, with flexbg,
4986  * they could still across block group boundary.
4987  *
4988  * Also account for superblock, inode, quota and xattr blocks
4989  */
4990 int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4991 {
4992         ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
4993         int gdpblocks;
4994         int idxblocks;
4995         int ret = 0;
4996
4997         /*
4998          * How many index blocks need to touch to modify nrblocks?
4999          * The "Chunk" flag indicating whether the nrblocks is
5000          * physically contiguous on disk
5001          *
5002          * For Direct IO and fallocate, they calls get_block to allocate
5003          * one single extent at a time, so they could set the "Chunk" flag
5004          */
5005         idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
5006
5007         ret = idxblocks;
5008
5009         /*
5010          * Now let's see how many group bitmaps and group descriptors need
5011          * to account
5012          */
5013         groups = idxblocks;
5014         if (chunk)
5015                 groups += 1;
5016         else
5017                 groups += nrblocks;
5018
5019         gdpblocks = groups;
5020         if (groups > ngroups)
5021                 groups = ngroups;
5022         if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
5023                 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
5024
5025         /* bitmaps and block group descriptor blocks */
5026         ret += groups + gdpblocks;
5027
5028         /* Blocks for super block, inode, quota and xattr blocks */
5029         ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
5030
5031         return ret;
5032 }
5033
5034 /*
5035  * Calulate the total number of credits to reserve to fit
5036  * the modification of a single pages into a single transaction,
5037  * which may include multiple chunks of block allocations.
5038  *
5039  * This could be called via ext4_write_begin()
5040  *
5041  * We need to consider the worse case, when
5042  * one new block per extent.
5043  */
5044 int ext4_writepage_trans_blocks(struct inode *inode)
5045 {
5046         int bpp = ext4_journal_blocks_per_page(inode);
5047         int ret;
5048
5049         ret = ext4_meta_trans_blocks(inode, bpp, 0);
5050
5051         /* Account for data blocks for journalled mode */
5052         if (ext4_should_journal_data(inode))
5053                 ret += bpp;
5054         return ret;
5055 }
5056
5057 /*
5058  * Calculate the journal credits for a chunk of data modification.
5059  *
5060  * This is called from DIO, fallocate or whoever calling
5061  * ext4_get_blocks() to map/allocate a chunk of contigous disk blocks.
5062  *
5063  * journal buffers for data blocks are not included here, as DIO
5064  * and fallocate do no need to journal data buffers.
5065  */
5066 int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
5067 {
5068         return ext4_meta_trans_blocks(inode, nrblocks, 1);
5069 }
5070
5071 /*
5072  * The caller must have previously called ext4_reserve_inode_write().
5073  * Give this, we know that the caller already has write access to iloc->bh.
5074  */
5075 int ext4_mark_iloc_dirty(handle_t *handle,
5076                 struct inode *inode, struct ext4_iloc *iloc)
5077 {
5078         int err = 0;
5079
5080         if (test_opt(inode->i_sb, I_VERSION))
5081                 inode_inc_iversion(inode);
5082
5083         /* the do_update_inode consumes one bh->b_count */
5084         get_bh(iloc->bh);
5085
5086         /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
5087         err = ext4_do_update_inode(handle, inode, iloc);
5088         put_bh(iloc->bh);
5089         return err;
5090 }
5091
5092 /*
5093  * On success, We end up with an outstanding reference count against
5094  * iloc->bh.  This _must_ be cleaned up later.
5095  */
5096
5097 int
5098 ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
5099                          struct ext4_iloc *iloc)
5100 {
5101         int err;
5102
5103         err = ext4_get_inode_loc(inode, iloc);
5104         if (!err) {
5105                 BUFFER_TRACE(iloc->bh, "get_write_access");
5106                 err = ext4_journal_get_write_access(handle, iloc->bh);
5107                 if (err) {
5108                         brelse(iloc->bh);
5109                         iloc->bh = NULL;
5110                 }
5111         }
5112         ext4_std_error(inode->i_sb, err);
5113         return err;
5114 }
5115
5116 /*
5117  * Expand an inode by new_extra_isize bytes.
5118  * Returns 0 on success or negative error number on failure.
5119  */
5120 static int ext4_expand_extra_isize(struct inode *inode,
5121                                    unsigned int new_extra_isize,
5122                                    struct ext4_iloc iloc,
5123                                    handle_t *handle)
5124 {
5125         struct ext4_inode *raw_inode;
5126         struct ext4_xattr_ibody_header *header;
5127         struct ext4_xattr_entry *entry;
5128
5129         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
5130                 return 0;
5131
5132         raw_inode = ext4_raw_inode(&iloc);
5133
5134         header = IHDR(inode, raw_inode);
5135         entry = IFIRST(header);
5136
5137         /* No extended attributes present */
5138         if (!(EXT4_I(inode)->i_state & EXT4_STATE_XATTR) ||
5139                 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
5140                 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
5141                         new_extra_isize);
5142                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
5143                 return 0;
5144         }
5145
5146         /* try to expand with EAs present */
5147         return ext4_expand_extra_isize_ea(inode, new_extra_isize,
5148                                           raw_inode, handle);
5149 }
5150
5151 /*
5152  * What we do here is to mark the in-core inode as clean with respect to inode
5153  * dirtiness (it may still be data-dirty).
5154  * This means that the in-core inode may be reaped by prune_icache
5155  * without having to perform any I/O.  This is a very good thing,
5156  * because *any* task may call prune_icache - even ones which
5157  * have a transaction open against a different journal.
5158  *
5159  * Is this cheating?  Not really.  Sure, we haven't written the
5160  * inode out, but prune_icache isn't a user-visible syncing function.
5161  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
5162  * we start and wait on commits.
5163  *
5164  * Is this efficient/effective?  Well, we're being nice to the system
5165  * by cleaning up our inodes proactively so they can be reaped
5166  * without I/O.  But we are potentially leaving up to five seconds'
5167  * worth of inodes floating about which prune_icache wants us to
5168  * write out.  One way to fix that would be to get prune_icache()
5169  * to do a write_super() to free up some memory.  It has the desired
5170  * effect.
5171  */
5172 int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
5173 {
5174         struct ext4_iloc iloc;
5175         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5176         static unsigned int mnt_count;
5177         int err, ret;
5178
5179         might_sleep();
5180         err = ext4_reserve_inode_write(handle, inode, &iloc);
5181         if (ext4_handle_valid(handle) &&
5182             EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
5183             !(EXT4_I(inode)->i_state & EXT4_STATE_NO_EXPAND)) {
5184                 /*
5185                  * We need extra buffer credits since we may write into EA block
5186                  * with this same handle. If journal_extend fails, then it will
5187                  * only result in a minor loss of functionality for that inode.
5188                  * If this is felt to be critical, then e2fsck should be run to
5189                  * force a large enough s_min_extra_isize.
5190                  */
5191                 if ((jbd2_journal_extend(handle,
5192                              EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
5193                         ret = ext4_expand_extra_isize(inode,
5194                                                       sbi->s_want_extra_isize,
5195                                                       iloc, handle);
5196                         if (ret) {
5197                                 EXT4_I(inode)->i_state |= EXT4_STATE_NO_EXPAND;
5198                                 if (mnt_count !=
5199                                         le16_to_cpu(sbi->s_es->s_mnt_count)) {
5200                                         ext4_warning(inode->i_sb, __func__,
5201                                         "Unable to expand inode %lu. Delete"
5202                                         " some EAs or run e2fsck.",
5203                                         inode->i_ino);
5204                                         mnt_count =
5205                                           le16_to_cpu(sbi->s_es->s_mnt_count);
5206                                 }
5207                         }
5208                 }
5209         }
5210         if (!err)
5211                 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
5212         return err;
5213 }
5214
5215 /*
5216  * ext4_dirty_inode() is called from __mark_inode_dirty()
5217  *
5218  * We're really interested in the case where a file is being extended.
5219  * i_size has been changed by generic_commit_write() and we thus need
5220  * to include the updated inode in the current transaction.
5221  *
5222  * Also, vfs_dq_alloc_block() will always dirty the inode when blocks
5223  * are allocated to the file.
5224  *
5225  * If the inode is marked synchronous, we don't honour that here - doing
5226  * so would cause a commit on atime updates, which we don't bother doing.
5227  * We handle synchronous inodes at the highest possible level.
5228  */
5229 void ext4_dirty_inode(struct inode *inode)
5230 {
5231         handle_t *current_handle = ext4_journal_current_handle();
5232         handle_t *handle;
5233
5234         if (!ext4_handle_valid(current_handle)) {
5235                 ext4_mark_inode_dirty(current_handle, inode);
5236                 return;
5237         }
5238
5239         handle = ext4_journal_start(inode, 2);
5240         if (IS_ERR(handle))
5241                 goto out;
5242         if (current_handle &&
5243                 current_handle->h_transaction != handle->h_transaction) {
5244                 /* This task has a transaction open against a different fs */
5245                 printk(KERN_EMERG "%s: transactions do not match!\n",
5246                        __func__);
5247         } else {
5248                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
5249                                 current_handle);
5250                 ext4_mark_inode_dirty(handle, inode);
5251         }
5252         ext4_journal_stop(handle);
5253 out:
5254         return;
5255 }
5256
5257 #if 0
5258 /*
5259  * Bind an inode's backing buffer_head into this transaction, to prevent
5260  * it from being flushed to disk early.  Unlike
5261  * ext4_reserve_inode_write, this leaves behind no bh reference and
5262  * returns no iloc structure, so the caller needs to repeat the iloc
5263  * lookup to mark the inode dirty later.
5264  */
5265 static int ext4_pin_inode(handle_t *handle, struct inode *inode)
5266 {
5267         struct ext4_iloc iloc;
5268
5269         int err = 0;
5270         if (handle) {
5271                 err = ext4_get_inode_loc(inode, &iloc);
5272                 if (!err) {
5273                         BUFFER_TRACE(iloc.bh, "get_write_access");
5274                         err = jbd2_journal_get_write_access(handle, iloc.bh);
5275                         if (!err)
5276                                 err = ext4_handle_dirty_metadata(handle,
5277                                                                  inode,
5278                                                                  iloc.bh);
5279                         brelse(iloc.bh);
5280                 }
5281         }
5282         ext4_std_error(inode->i_sb, err);
5283         return err;
5284 }
5285 #endif
5286
5287 int ext4_change_inode_journal_flag(struct inode *inode, int val)
5288 {
5289         journal_t *journal;
5290         handle_t *handle;
5291         int err;
5292
5293         /*
5294          * We have to be very careful here: changing a data block's
5295          * journaling status dynamically is dangerous.  If we write a
5296          * data block to the journal, change the status and then delete
5297          * that block, we risk forgetting to revoke the old log record
5298          * from the journal and so a subsequent replay can corrupt data.
5299          * So, first we make sure that the journal is empty and that
5300          * nobody is changing anything.
5301          */
5302
5303         journal = EXT4_JOURNAL(inode);
5304         if (!journal)
5305                 return 0;
5306         if (is_journal_aborted(journal))
5307                 return -EROFS;
5308
5309         jbd2_journal_lock_updates(journal);
5310         jbd2_journal_flush(journal);
5311
5312         /*
5313          * OK, there are no updates running now, and all cached data is
5314          * synced to disk.  We are now in a completely consistent state
5315          * which doesn't have anything in the journal, and we know that
5316          * no filesystem updates are running, so it is safe to modify
5317          * the inode's in-core data-journaling state flag now.
5318          */
5319
5320         if (val)
5321                 EXT4_I(inode)->i_flags |= EXT4_JOURNAL_DATA_FL;
5322         else
5323                 EXT4_I(inode)->i_flags &= ~EXT4_JOURNAL_DATA_FL;
5324         ext4_set_aops(inode);
5325
5326         jbd2_journal_unlock_updates(journal);
5327
5328         /* Finally we can mark the inode as dirty. */
5329
5330         handle = ext4_journal_start(inode, 1);
5331         if (IS_ERR(handle))
5332                 return PTR_ERR(handle);
5333
5334         err = ext4_mark_inode_dirty(handle, inode);
5335         ext4_handle_sync(handle);
5336         ext4_journal_stop(handle);
5337         ext4_std_error(inode->i_sb, err);
5338
5339         return err;
5340 }
5341
5342 static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
5343 {
5344         return !buffer_mapped(bh);
5345 }
5346
5347 int ext4_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
5348 {
5349         struct page *page = vmf->page;
5350         loff_t size;
5351         unsigned long len;
5352         int ret = -EINVAL;
5353         void *fsdata;
5354         struct file *file = vma->vm_file;
5355         struct inode *inode = file->f_path.dentry->d_inode;
5356         struct address_space *mapping = inode->i_mapping;
5357
5358         /*
5359          * Get i_alloc_sem to stop truncates messing with the inode. We cannot
5360          * get i_mutex because we are already holding mmap_sem.
5361          */
5362         down_read(&inode->i_alloc_sem);
5363         size = i_size_read(inode);
5364         if (page->mapping != mapping || size <= page_offset(page)
5365             || !PageUptodate(page)) {
5366                 /* page got truncated from under us? */
5367                 goto out_unlock;
5368         }
5369         ret = 0;
5370         if (PageMappedToDisk(page))
5371                 goto out_unlock;
5372
5373         if (page->index == size >> PAGE_CACHE_SHIFT)
5374                 len = size & ~PAGE_CACHE_MASK;
5375         else
5376                 len = PAGE_CACHE_SIZE;
5377
5378         if (page_has_buffers(page)) {
5379                 /* return if we have all the buffers mapped */
5380                 if (!walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
5381                                        ext4_bh_unmapped))
5382                         goto out_unlock;
5383         }
5384         /*
5385          * OK, we need to fill the hole... Do write_begin write_end
5386          * to do block allocation/reservation.We are not holding
5387          * inode.i__mutex here. That allow * parallel write_begin,
5388          * write_end call. lock_page prevent this from happening
5389          * on the same page though
5390          */
5391         ret = mapping->a_ops->write_begin(file, mapping, page_offset(page),
5392                         len, AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
5393         if (ret < 0)
5394                 goto out_unlock;
5395         ret = mapping->a_ops->write_end(file, mapping, page_offset(page),
5396                         len, len, page, fsdata);
5397         if (ret < 0)
5398                 goto out_unlock;
5399         ret = 0;
5400 out_unlock:
5401         if (ret)
5402                 ret = VM_FAULT_SIGBUS;
5403         up_read(&inode->i_alloc_sem);
5404         return ret;
5405 }