Merge branch 'linus' into core/futexes
[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 /**
375  *      ext4_get_branch - read the chain of indirect blocks leading to data
376  *      @inode: inode in question
377  *      @depth: depth of the chain (1 - direct pointer, etc.)
378  *      @offsets: offsets of pointers in inode/indirect blocks
379  *      @chain: place to store the result
380  *      @err: here we store the error value
381  *
382  *      Function fills the array of triples <key, p, bh> and returns %NULL
383  *      if everything went OK or the pointer to the last filled triple
384  *      (incomplete one) otherwise. Upon the return chain[i].key contains
385  *      the number of (i+1)-th block in the chain (as it is stored in memory,
386  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
387  *      number (it points into struct inode for i==0 and into the bh->b_data
388  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
389  *      block for i>0 and NULL for i==0. In other words, it holds the block
390  *      numbers of the chain, addresses they were taken from (and where we can
391  *      verify that chain did not change) and buffer_heads hosting these
392  *      numbers.
393  *
394  *      Function stops when it stumbles upon zero pointer (absent block)
395  *              (pointer to last triple returned, *@err == 0)
396  *      or when it gets an IO error reading an indirect block
397  *              (ditto, *@err == -EIO)
398  *      or when it reads all @depth-1 indirect blocks successfully and finds
399  *      the whole chain, all way to the data (returns %NULL, *err == 0).
400  *
401  *      Need to be called with
402  *      down_read(&EXT4_I(inode)->i_data_sem)
403  */
404 static Indirect *ext4_get_branch(struct inode *inode, int depth,
405                                  ext4_lblk_t  *offsets,
406                                  Indirect chain[4], int *err)
407 {
408         struct super_block *sb = inode->i_sb;
409         Indirect *p = chain;
410         struct buffer_head *bh;
411
412         *err = 0;
413         /* i_data is not going away, no lock needed */
414         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
415         if (!p->key)
416                 goto no_block;
417         while (--depth) {
418                 bh = sb_bread(sb, le32_to_cpu(p->key));
419                 if (!bh)
420                         goto failure;
421                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
422                 /* Reader: end */
423                 if (!p->key)
424                         goto no_block;
425         }
426         return NULL;
427
428 failure:
429         *err = -EIO;
430 no_block:
431         return p;
432 }
433
434 /**
435  *      ext4_find_near - find a place for allocation with sufficient locality
436  *      @inode: owner
437  *      @ind: descriptor of indirect block.
438  *
439  *      This function returns the preferred place for block allocation.
440  *      It is used when heuristic for sequential allocation fails.
441  *      Rules are:
442  *        + if there is a block to the left of our position - allocate near it.
443  *        + if pointer will live in indirect block - allocate near that block.
444  *        + if pointer will live in inode - allocate in the same
445  *          cylinder group.
446  *
447  * In the latter case we colour the starting block by the callers PID to
448  * prevent it from clashing with concurrent allocations for a different inode
449  * in the same block group.   The PID is used here so that functionally related
450  * files will be close-by on-disk.
451  *
452  *      Caller must make sure that @ind is valid and will stay that way.
453  */
454 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
455 {
456         struct ext4_inode_info *ei = EXT4_I(inode);
457         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
458         __le32 *p;
459         ext4_fsblk_t bg_start;
460         ext4_fsblk_t last_block;
461         ext4_grpblk_t colour;
462
463         /* Try to find previous block */
464         for (p = ind->p - 1; p >= start; p--) {
465                 if (*p)
466                         return le32_to_cpu(*p);
467         }
468
469         /* No such thing, so let's try location of indirect block */
470         if (ind->bh)
471                 return ind->bh->b_blocknr;
472
473         /*
474          * It is going to be referred to from the inode itself? OK, just put it
475          * into the same cylinder group then.
476          */
477         bg_start = ext4_group_first_block_no(inode->i_sb, ei->i_block_group);
478         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
479
480         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
481                 colour = (current->pid % 16) *
482                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
483         else
484                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
485         return bg_start + colour;
486 }
487
488 /**
489  *      ext4_find_goal - find a preferred place for allocation.
490  *      @inode: owner
491  *      @block:  block we want
492  *      @partial: pointer to the last triple within a chain
493  *
494  *      Normally this function find the preferred place for block allocation,
495  *      returns it.
496  */
497 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
498                 Indirect *partial)
499 {
500         /*
501          * XXX need to get goal block from mballoc's data structures
502          */
503
504         return ext4_find_near(inode, partial);
505 }
506
507 /**
508  *      ext4_blks_to_allocate: Look up the block map and count the number
509  *      of direct blocks need to be allocated for the given branch.
510  *
511  *      @branch: chain of indirect blocks
512  *      @k: number of blocks need for indirect blocks
513  *      @blks: number of data blocks to be mapped.
514  *      @blocks_to_boundary:  the offset in the indirect block
515  *
516  *      return the total number of blocks to be allocate, including the
517  *      direct and indirect blocks.
518  */
519 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
520                 int blocks_to_boundary)
521 {
522         unsigned int count = 0;
523
524         /*
525          * Simple case, [t,d]Indirect block(s) has not allocated yet
526          * then it's clear blocks on that path have not allocated
527          */
528         if (k > 0) {
529                 /* right now we don't handle cross boundary allocation */
530                 if (blks < blocks_to_boundary + 1)
531                         count += blks;
532                 else
533                         count += blocks_to_boundary + 1;
534                 return count;
535         }
536
537         count++;
538         while (count < blks && count <= blocks_to_boundary &&
539                 le32_to_cpu(*(branch[0].p + count)) == 0) {
540                 count++;
541         }
542         return count;
543 }
544
545 /**
546  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
547  *      @indirect_blks: the number of blocks need to allocate for indirect
548  *                      blocks
549  *
550  *      @new_blocks: on return it will store the new block numbers for
551  *      the indirect blocks(if needed) and the first direct block,
552  *      @blks:  on return it will store the total number of allocated
553  *              direct blocks
554  */
555 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
556                                 ext4_lblk_t iblock, ext4_fsblk_t goal,
557                                 int indirect_blks, int blks,
558                                 ext4_fsblk_t new_blocks[4], int *err)
559 {
560         struct ext4_allocation_request ar;
561         int target, i;
562         unsigned long count = 0, blk_allocated = 0;
563         int index = 0;
564         ext4_fsblk_t current_block = 0;
565         int ret = 0;
566
567         /*
568          * Here we try to allocate the requested multiple blocks at once,
569          * on a best-effort basis.
570          * To build a branch, we should allocate blocks for
571          * the indirect blocks(if not allocated yet), and at least
572          * the first direct block of this branch.  That's the
573          * minimum number of blocks need to allocate(required)
574          */
575         /* first we try to allocate the indirect blocks */
576         target = indirect_blks;
577         while (target > 0) {
578                 count = target;
579                 /* allocating blocks for indirect blocks and direct blocks */
580                 current_block = ext4_new_meta_blocks(handle, inode,
581                                                         goal, &count, err);
582                 if (*err)
583                         goto failed_out;
584
585                 target -= count;
586                 /* allocate blocks for indirect blocks */
587                 while (index < indirect_blks && count) {
588                         new_blocks[index++] = current_block++;
589                         count--;
590                 }
591                 if (count > 0) {
592                         /*
593                          * save the new block number
594                          * for the first direct block
595                          */
596                         new_blocks[index] = current_block;
597                         printk(KERN_INFO "%s returned more blocks than "
598                                                 "requested\n", __func__);
599                         WARN_ON(1);
600                         break;
601                 }
602         }
603
604         target = blks - count ;
605         blk_allocated = count;
606         if (!target)
607                 goto allocated;
608         /* Now allocate data blocks */
609         memset(&ar, 0, sizeof(ar));
610         ar.inode = inode;
611         ar.goal = goal;
612         ar.len = target;
613         ar.logical = iblock;
614         if (S_ISREG(inode->i_mode))
615                 /* enable in-core preallocation only for regular files */
616                 ar.flags = EXT4_MB_HINT_DATA;
617
618         current_block = ext4_mb_new_blocks(handle, &ar, err);
619
620         if (*err && (target == blks)) {
621                 /*
622                  * if the allocation failed and we didn't allocate
623                  * any blocks before
624                  */
625                 goto failed_out;
626         }
627         if (!*err) {
628                 if (target == blks) {
629                 /*
630                  * save the new block number
631                  * for the first direct block
632                  */
633                         new_blocks[index] = current_block;
634                 }
635                 blk_allocated += ar.len;
636         }
637 allocated:
638         /* total number of blocks allocated for direct blocks */
639         ret = blk_allocated;
640         *err = 0;
641         return ret;
642 failed_out:
643         for (i = 0; i < index; i++)
644                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
645         return ret;
646 }
647
648 /**
649  *      ext4_alloc_branch - allocate and set up a chain of blocks.
650  *      @inode: owner
651  *      @indirect_blks: number of allocated indirect blocks
652  *      @blks: number of allocated direct blocks
653  *      @offsets: offsets (in the blocks) to store the pointers to next.
654  *      @branch: place to store the chain in.
655  *
656  *      This function allocates blocks, zeroes out all but the last one,
657  *      links them into chain and (if we are synchronous) writes them to disk.
658  *      In other words, it prepares a branch that can be spliced onto the
659  *      inode. It stores the information about that chain in the branch[], in
660  *      the same format as ext4_get_branch() would do. We are calling it after
661  *      we had read the existing part of chain and partial points to the last
662  *      triple of that (one with zero ->key). Upon the exit we have the same
663  *      picture as after the successful ext4_get_block(), except that in one
664  *      place chain is disconnected - *branch->p is still zero (we did not
665  *      set the last link), but branch->key contains the number that should
666  *      be placed into *branch->p to fill that gap.
667  *
668  *      If allocation fails we free all blocks we've allocated (and forget
669  *      their buffer_heads) and return the error value the from failed
670  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
671  *      as described above and return 0.
672  */
673 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
674                                 ext4_lblk_t iblock, int indirect_blks,
675                                 int *blks, ext4_fsblk_t goal,
676                                 ext4_lblk_t *offsets, Indirect *branch)
677 {
678         int blocksize = inode->i_sb->s_blocksize;
679         int i, n = 0;
680         int err = 0;
681         struct buffer_head *bh;
682         int num;
683         ext4_fsblk_t new_blocks[4];
684         ext4_fsblk_t current_block;
685
686         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
687                                 *blks, new_blocks, &err);
688         if (err)
689                 return err;
690
691         branch[0].key = cpu_to_le32(new_blocks[0]);
692         /*
693          * metadata blocks and data blocks are allocated.
694          */
695         for (n = 1; n <= indirect_blks;  n++) {
696                 /*
697                  * Get buffer_head for parent block, zero it out
698                  * and set the pointer to new one, then send
699                  * parent to disk.
700                  */
701                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
702                 branch[n].bh = bh;
703                 lock_buffer(bh);
704                 BUFFER_TRACE(bh, "call get_create_access");
705                 err = ext4_journal_get_create_access(handle, bh);
706                 if (err) {
707                         unlock_buffer(bh);
708                         brelse(bh);
709                         goto failed;
710                 }
711
712                 memset(bh->b_data, 0, blocksize);
713                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
714                 branch[n].key = cpu_to_le32(new_blocks[n]);
715                 *branch[n].p = branch[n].key;
716                 if (n == indirect_blks) {
717                         current_block = new_blocks[n];
718                         /*
719                          * End of chain, update the last new metablock of
720                          * the chain to point to the new allocated
721                          * data blocks numbers
722                          */
723                         for (i=1; i < num; i++)
724                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
725                 }
726                 BUFFER_TRACE(bh, "marking uptodate");
727                 set_buffer_uptodate(bh);
728                 unlock_buffer(bh);
729
730                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
731                 err = ext4_handle_dirty_metadata(handle, inode, bh);
732                 if (err)
733                         goto failed;
734         }
735         *blks = num;
736         return err;
737 failed:
738         /* Allocation failed, free what we already allocated */
739         for (i = 1; i <= n ; i++) {
740                 BUFFER_TRACE(branch[i].bh, "call jbd2_journal_forget");
741                 ext4_journal_forget(handle, branch[i].bh);
742         }
743         for (i = 0; i < indirect_blks; i++)
744                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
745
746         ext4_free_blocks(handle, inode, new_blocks[i], num, 0);
747
748         return err;
749 }
750
751 /**
752  * ext4_splice_branch - splice the allocated branch onto inode.
753  * @inode: owner
754  * @block: (logical) number of block we are adding
755  * @chain: chain of indirect blocks (with a missing link - see
756  *      ext4_alloc_branch)
757  * @where: location of missing link
758  * @num:   number of indirect blocks we are adding
759  * @blks:  number of direct blocks we are adding
760  *
761  * This function fills the missing link and does all housekeeping needed in
762  * inode (->i_blocks, etc.). In case of success we end up with the full
763  * chain to new block and return 0.
764  */
765 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
766                         ext4_lblk_t block, Indirect *where, int num, int blks)
767 {
768         int i;
769         int err = 0;
770         ext4_fsblk_t current_block;
771
772         /*
773          * If we're splicing into a [td]indirect block (as opposed to the
774          * inode) then we need to get write access to the [td]indirect block
775          * before the splice.
776          */
777         if (where->bh) {
778                 BUFFER_TRACE(where->bh, "get_write_access");
779                 err = ext4_journal_get_write_access(handle, where->bh);
780                 if (err)
781                         goto err_out;
782         }
783         /* That's it */
784
785         *where->p = where->key;
786
787         /*
788          * Update the host buffer_head or inode to point to more just allocated
789          * direct blocks blocks
790          */
791         if (num == 0 && blks > 1) {
792                 current_block = le32_to_cpu(where->key) + 1;
793                 for (i = 1; i < blks; i++)
794                         *(where->p + i) = cpu_to_le32(current_block++);
795         }
796
797         /* We are done with atomic stuff, now do the rest of housekeeping */
798
799         inode->i_ctime = ext4_current_time(inode);
800         ext4_mark_inode_dirty(handle, inode);
801
802         /* had we spliced it onto indirect block? */
803         if (where->bh) {
804                 /*
805                  * If we spliced it onto an indirect block, we haven't
806                  * altered the inode.  Note however that if it is being spliced
807                  * onto an indirect block at the very end of the file (the
808                  * file is growing) then we *will* alter the inode to reflect
809                  * the new i_size.  But that is not done here - it is done in
810                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
811                  */
812                 jbd_debug(5, "splicing indirect only\n");
813                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
814                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
815                 if (err)
816                         goto err_out;
817         } else {
818                 /*
819                  * OK, we spliced it into the inode itself on a direct block.
820                  * Inode was dirtied above.
821                  */
822                 jbd_debug(5, "splicing direct\n");
823         }
824         return err;
825
826 err_out:
827         for (i = 1; i <= num; i++) {
828                 BUFFER_TRACE(where[i].bh, "call jbd2_journal_forget");
829                 ext4_journal_forget(handle, where[i].bh);
830                 ext4_free_blocks(handle, inode,
831                                         le32_to_cpu(where[i-1].key), 1, 0);
832         }
833         ext4_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks, 0);
834
835         return err;
836 }
837
838 /*
839  * Allocation strategy is simple: if we have to allocate something, we will
840  * have to go the whole way to leaf. So let's do it before attaching anything
841  * to tree, set linkage between the newborn blocks, write them if sync is
842  * required, recheck the path, free and repeat if check fails, otherwise
843  * set the last missing link (that will protect us from any truncate-generated
844  * removals - all blocks on the path are immune now) and possibly force the
845  * write on the parent block.
846  * That has a nice additional property: no special recovery from the failed
847  * allocations is needed - we simply release blocks and do not touch anything
848  * reachable from inode.
849  *
850  * `handle' can be NULL if create == 0.
851  *
852  * return > 0, # of blocks mapped or allocated.
853  * return = 0, if plain lookup failed.
854  * return < 0, error case.
855  *
856  *
857  * Need to be called with
858  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
859  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
860  */
861 static int ext4_get_blocks_handle(handle_t *handle, struct inode *inode,
862                                   ext4_lblk_t iblock, unsigned int maxblocks,
863                                   struct buffer_head *bh_result,
864                                   int create, int extend_disksize)
865 {
866         int err = -EIO;
867         ext4_lblk_t offsets[4];
868         Indirect chain[4];
869         Indirect *partial;
870         ext4_fsblk_t goal;
871         int indirect_blks;
872         int blocks_to_boundary = 0;
873         int depth;
874         struct ext4_inode_info *ei = EXT4_I(inode);
875         int count = 0;
876         ext4_fsblk_t first_block = 0;
877         loff_t disksize;
878
879
880         J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
881         J_ASSERT(handle != NULL || create == 0);
882         depth = ext4_block_to_path(inode, iblock, offsets,
883                                         &blocks_to_boundary);
884
885         if (depth == 0)
886                 goto out;
887
888         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
889
890         /* Simplest case - block found, no allocation needed */
891         if (!partial) {
892                 first_block = le32_to_cpu(chain[depth - 1].key);
893                 clear_buffer_new(bh_result);
894                 count++;
895                 /*map more blocks*/
896                 while (count < maxblocks && count <= blocks_to_boundary) {
897                         ext4_fsblk_t blk;
898
899                         blk = le32_to_cpu(*(chain[depth-1].p + count));
900
901                         if (blk == first_block + count)
902                                 count++;
903                         else
904                                 break;
905                 }
906                 goto got_it;
907         }
908
909         /* Next simple case - plain lookup or failed read of indirect block */
910         if (!create || err == -EIO)
911                 goto cleanup;
912
913         /*
914          * Okay, we need to do block allocation.
915         */
916         goal = ext4_find_goal(inode, iblock, partial);
917
918         /* the number of blocks need to allocate for [d,t]indirect blocks */
919         indirect_blks = (chain + depth) - partial - 1;
920
921         /*
922          * Next look up the indirect map to count the totoal number of
923          * direct blocks to allocate for this branch.
924          */
925         count = ext4_blks_to_allocate(partial, indirect_blks,
926                                         maxblocks, blocks_to_boundary);
927         /*
928          * Block out ext4_truncate while we alter the tree
929          */
930         err = ext4_alloc_branch(handle, inode, iblock, indirect_blks,
931                                         &count, goal,
932                                         offsets + (partial - chain), partial);
933
934         /*
935          * The ext4_splice_branch call will free and forget any buffers
936          * on the new chain if there is a failure, but that risks using
937          * up transaction credits, especially for bitmaps where the
938          * credits cannot be returned.  Can we handle this somehow?  We
939          * may need to return -EAGAIN upwards in the worst case.  --sct
940          */
941         if (!err)
942                 err = ext4_splice_branch(handle, inode, iblock,
943                                         partial, indirect_blks, count);
944         /*
945          * i_disksize growing is protected by i_data_sem.  Don't forget to
946          * protect it if you're about to implement concurrent
947          * ext4_get_block() -bzzz
948         */
949         if (!err && extend_disksize) {
950                 disksize = ((loff_t) iblock + count) << inode->i_blkbits;
951                 if (disksize > i_size_read(inode))
952                         disksize = i_size_read(inode);
953                 if (disksize > ei->i_disksize)
954                         ei->i_disksize = disksize;
955         }
956         if (err)
957                 goto cleanup;
958
959         set_buffer_new(bh_result);
960 got_it:
961         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
962         if (count > blocks_to_boundary)
963                 set_buffer_boundary(bh_result);
964         err = count;
965         /* Clean up and exit */
966         partial = chain + depth - 1;    /* the whole chain */
967 cleanup:
968         while (partial > chain) {
969                 BUFFER_TRACE(partial->bh, "call brelse");
970                 brelse(partial->bh);
971                 partial--;
972         }
973         BUFFER_TRACE(bh_result, "returned");
974 out:
975         return err;
976 }
977
978 qsize_t ext4_get_reserved_space(struct inode *inode)
979 {
980         unsigned long long total;
981
982         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
983         total = EXT4_I(inode)->i_reserved_data_blocks +
984                 EXT4_I(inode)->i_reserved_meta_blocks;
985         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
986
987         return total;
988 }
989 /*
990  * Calculate the number of metadata blocks need to reserve
991  * to allocate @blocks for non extent file based file
992  */
993 static int ext4_indirect_calc_metadata_amount(struct inode *inode, int blocks)
994 {
995         int icap = EXT4_ADDR_PER_BLOCK(inode->i_sb);
996         int ind_blks, dind_blks, tind_blks;
997
998         /* number of new indirect blocks needed */
999         ind_blks = (blocks + icap - 1) / icap;
1000
1001         dind_blks = (ind_blks + icap - 1) / icap;
1002
1003         tind_blks = 1;
1004
1005         return ind_blks + dind_blks + tind_blks;
1006 }
1007
1008 /*
1009  * Calculate the number of metadata blocks need to reserve
1010  * to allocate given number of blocks
1011  */
1012 static int ext4_calc_metadata_amount(struct inode *inode, int blocks)
1013 {
1014         if (!blocks)
1015                 return 0;
1016
1017         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
1018                 return ext4_ext_calc_metadata_amount(inode, blocks);
1019
1020         return ext4_indirect_calc_metadata_amount(inode, blocks);
1021 }
1022
1023 static void ext4_da_update_reserve_space(struct inode *inode, int used)
1024 {
1025         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1026         int total, mdb, mdb_free;
1027
1028         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1029         /* recalculate the number of metablocks still need to be reserved */
1030         total = EXT4_I(inode)->i_reserved_data_blocks - used;
1031         mdb = ext4_calc_metadata_amount(inode, total);
1032
1033         /* figure out how many metablocks to release */
1034         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1035         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1036
1037         if (mdb_free) {
1038                 /* Account for allocated meta_blocks */
1039                 mdb_free -= EXT4_I(inode)->i_allocated_meta_blocks;
1040
1041                 /* update fs dirty blocks counter */
1042                 percpu_counter_sub(&sbi->s_dirtyblocks_counter, mdb_free);
1043                 EXT4_I(inode)->i_allocated_meta_blocks = 0;
1044                 EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1045         }
1046
1047         /* update per-inode reservations */
1048         BUG_ON(used  > EXT4_I(inode)->i_reserved_data_blocks);
1049         EXT4_I(inode)->i_reserved_data_blocks -= used;
1050         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1051
1052         /*
1053          * free those over-booking quota for metadata blocks
1054          */
1055
1056         if (mdb_free)
1057                 vfs_dq_release_reservation_block(inode, mdb_free);
1058 }
1059
1060 /*
1061  * The ext4_get_blocks_wrap() function try to look up the requested blocks,
1062  * and returns if the blocks are already mapped.
1063  *
1064  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1065  * and store the allocated blocks in the result buffer head and mark it
1066  * mapped.
1067  *
1068  * If file type is extents based, it will call ext4_ext_get_blocks(),
1069  * Otherwise, call with ext4_get_blocks_handle() to handle indirect mapping
1070  * based files
1071  *
1072  * On success, it returns the number of blocks being mapped or allocate.
1073  * if create==0 and the blocks are pre-allocated and uninitialized block,
1074  * the result buffer head is unmapped. If the create ==1, it will make sure
1075  * the buffer head is mapped.
1076  *
1077  * It returns 0 if plain look up failed (blocks have not been allocated), in
1078  * that casem, buffer head is unmapped
1079  *
1080  * It returns the error in case of allocation failure.
1081  */
1082 int ext4_get_blocks_wrap(handle_t *handle, struct inode *inode, sector_t block,
1083                         unsigned int max_blocks, struct buffer_head *bh,
1084                         int create, int extend_disksize, int flag)
1085 {
1086         int retval;
1087
1088         clear_buffer_mapped(bh);
1089
1090         /*
1091          * Try to see if we can get  the block without requesting
1092          * for new file system block.
1093          */
1094         down_read((&EXT4_I(inode)->i_data_sem));
1095         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1096                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1097                                 bh, 0, 0);
1098         } else {
1099                 retval = ext4_get_blocks_handle(handle,
1100                                 inode, block, max_blocks, bh, 0, 0);
1101         }
1102         up_read((&EXT4_I(inode)->i_data_sem));
1103
1104         /* If it is only a block(s) look up */
1105         if (!create)
1106                 return retval;
1107
1108         /*
1109          * Returns if the blocks have already allocated
1110          *
1111          * Note that if blocks have been preallocated
1112          * ext4_ext_get_block() returns th create = 0
1113          * with buffer head unmapped.
1114          */
1115         if (retval > 0 && buffer_mapped(bh))
1116                 return retval;
1117
1118         /*
1119          * New blocks allocate and/or writing to uninitialized extent
1120          * will possibly result in updating i_data, so we take
1121          * the write lock of i_data_sem, and call get_blocks()
1122          * with create == 1 flag.
1123          */
1124         down_write((&EXT4_I(inode)->i_data_sem));
1125
1126         /*
1127          * if the caller is from delayed allocation writeout path
1128          * we have already reserved fs blocks for allocation
1129          * let the underlying get_block() function know to
1130          * avoid double accounting
1131          */
1132         if (flag)
1133                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1134         /*
1135          * We need to check for EXT4 here because migrate
1136          * could have changed the inode type in between
1137          */
1138         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1139                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1140                                 bh, create, extend_disksize);
1141         } else {
1142                 retval = ext4_get_blocks_handle(handle, inode, block,
1143                                 max_blocks, bh, create, extend_disksize);
1144
1145                 if (retval > 0 && buffer_new(bh)) {
1146                         /*
1147                          * We allocated new blocks which will result in
1148                          * i_data's format changing.  Force the migrate
1149                          * to fail by clearing migrate flags
1150                          */
1151                         EXT4_I(inode)->i_flags = EXT4_I(inode)->i_flags &
1152                                                         ~EXT4_EXT_MIGRATE;
1153                 }
1154         }
1155
1156         if (flag) {
1157                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1158                 /*
1159                  * Update reserved blocks/metadata blocks
1160                  * after successful block allocation
1161                  * which were deferred till now
1162                  */
1163                 if ((retval > 0) && buffer_delay(bh))
1164                         ext4_da_update_reserve_space(inode, retval);
1165         }
1166
1167         up_write((&EXT4_I(inode)->i_data_sem));
1168         return retval;
1169 }
1170
1171 /* Maximum number of blocks we map for direct IO at once. */
1172 #define DIO_MAX_BLOCKS 4096
1173
1174 int ext4_get_block(struct inode *inode, sector_t iblock,
1175                    struct buffer_head *bh_result, int create)
1176 {
1177         handle_t *handle = ext4_journal_current_handle();
1178         int ret = 0, started = 0;
1179         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1180         int dio_credits;
1181
1182         if (create && !handle) {
1183                 /* Direct IO write... */
1184                 if (max_blocks > DIO_MAX_BLOCKS)
1185                         max_blocks = DIO_MAX_BLOCKS;
1186                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
1187                 handle = ext4_journal_start(inode, dio_credits);
1188                 if (IS_ERR(handle)) {
1189                         ret = PTR_ERR(handle);
1190                         goto out;
1191                 }
1192                 started = 1;
1193         }
1194
1195         ret = ext4_get_blocks_wrap(handle, inode, iblock,
1196                                         max_blocks, bh_result, create, 0, 0);
1197         if (ret > 0) {
1198                 bh_result->b_size = (ret << inode->i_blkbits);
1199                 ret = 0;
1200         }
1201         if (started)
1202                 ext4_journal_stop(handle);
1203 out:
1204         return ret;
1205 }
1206
1207 /*
1208  * `handle' can be NULL if create is zero
1209  */
1210 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1211                                 ext4_lblk_t block, int create, int *errp)
1212 {
1213         struct buffer_head dummy;
1214         int fatal = 0, err;
1215
1216         J_ASSERT(handle != NULL || create == 0);
1217
1218         dummy.b_state = 0;
1219         dummy.b_blocknr = -1000;
1220         buffer_trace_init(&dummy.b_history);
1221         err = ext4_get_blocks_wrap(handle, inode, block, 1,
1222                                         &dummy, create, 1, 0);
1223         /*
1224          * ext4_get_blocks_handle() returns number of blocks
1225          * mapped. 0 in case of a HOLE.
1226          */
1227         if (err > 0) {
1228                 if (err > 1)
1229                         WARN_ON(1);
1230                 err = 0;
1231         }
1232         *errp = err;
1233         if (!err && buffer_mapped(&dummy)) {
1234                 struct buffer_head *bh;
1235                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1236                 if (!bh) {
1237                         *errp = -EIO;
1238                         goto err;
1239                 }
1240                 if (buffer_new(&dummy)) {
1241                         J_ASSERT(create != 0);
1242                         J_ASSERT(handle != NULL);
1243
1244                         /*
1245                          * Now that we do not always journal data, we should
1246                          * keep in mind whether this should always journal the
1247                          * new buffer as metadata.  For now, regular file
1248                          * writes use ext4_get_block instead, so it's not a
1249                          * problem.
1250                          */
1251                         lock_buffer(bh);
1252                         BUFFER_TRACE(bh, "call get_create_access");
1253                         fatal = ext4_journal_get_create_access(handle, bh);
1254                         if (!fatal && !buffer_uptodate(bh)) {
1255                                 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1256                                 set_buffer_uptodate(bh);
1257                         }
1258                         unlock_buffer(bh);
1259                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1260                         err = ext4_handle_dirty_metadata(handle, inode, bh);
1261                         if (!fatal)
1262                                 fatal = err;
1263                 } else {
1264                         BUFFER_TRACE(bh, "not a new buffer");
1265                 }
1266                 if (fatal) {
1267                         *errp = fatal;
1268                         brelse(bh);
1269                         bh = NULL;
1270                 }
1271                 return bh;
1272         }
1273 err:
1274         return NULL;
1275 }
1276
1277 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1278                                ext4_lblk_t block, int create, int *err)
1279 {
1280         struct buffer_head *bh;
1281
1282         bh = ext4_getblk(handle, inode, block, create, err);
1283         if (!bh)
1284                 return bh;
1285         if (buffer_uptodate(bh))
1286                 return bh;
1287         ll_rw_block(READ_META, 1, &bh);
1288         wait_on_buffer(bh);
1289         if (buffer_uptodate(bh))
1290                 return bh;
1291         put_bh(bh);
1292         *err = -EIO;
1293         return NULL;
1294 }
1295
1296 static int walk_page_buffers(handle_t *handle,
1297                              struct buffer_head *head,
1298                              unsigned from,
1299                              unsigned to,
1300                              int *partial,
1301                              int (*fn)(handle_t *handle,
1302                                        struct buffer_head *bh))
1303 {
1304         struct buffer_head *bh;
1305         unsigned block_start, block_end;
1306         unsigned blocksize = head->b_size;
1307         int err, ret = 0;
1308         struct buffer_head *next;
1309
1310         for (bh = head, block_start = 0;
1311              ret == 0 && (bh != head || !block_start);
1312              block_start = block_end, bh = next)
1313         {
1314                 next = bh->b_this_page;
1315                 block_end = block_start + blocksize;
1316                 if (block_end <= from || block_start >= to) {
1317                         if (partial && !buffer_uptodate(bh))
1318                                 *partial = 1;
1319                         continue;
1320                 }
1321                 err = (*fn)(handle, bh);
1322                 if (!ret)
1323                         ret = err;
1324         }
1325         return ret;
1326 }
1327
1328 /*
1329  * To preserve ordering, it is essential that the hole instantiation and
1330  * the data write be encapsulated in a single transaction.  We cannot
1331  * close off a transaction and start a new one between the ext4_get_block()
1332  * and the commit_write().  So doing the jbd2_journal_start at the start of
1333  * prepare_write() is the right place.
1334  *
1335  * Also, this function can nest inside ext4_writepage() ->
1336  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1337  * has generated enough buffer credits to do the whole page.  So we won't
1338  * block on the journal in that case, which is good, because the caller may
1339  * be PF_MEMALLOC.
1340  *
1341  * By accident, ext4 can be reentered when a transaction is open via
1342  * quota file writes.  If we were to commit the transaction while thus
1343  * reentered, there can be a deadlock - we would be holding a quota
1344  * lock, and the commit would never complete if another thread had a
1345  * transaction open and was blocking on the quota lock - a ranking
1346  * violation.
1347  *
1348  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1349  * will _not_ run commit under these circumstances because handle->h_ref
1350  * is elevated.  We'll still have enough credits for the tiny quotafile
1351  * write.
1352  */
1353 static int do_journal_get_write_access(handle_t *handle,
1354                                         struct buffer_head *bh)
1355 {
1356         if (!buffer_mapped(bh) || buffer_freed(bh))
1357                 return 0;
1358         return ext4_journal_get_write_access(handle, bh);
1359 }
1360
1361 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1362                                 loff_t pos, unsigned len, unsigned flags,
1363                                 struct page **pagep, void **fsdata)
1364 {
1365         struct inode *inode = mapping->host;
1366         int ret, needed_blocks = ext4_writepage_trans_blocks(inode);
1367         handle_t *handle;
1368         int retries = 0;
1369         struct page *page;
1370         pgoff_t index;
1371         unsigned from, to;
1372
1373         trace_mark(ext4_write_begin,
1374                    "dev %s ino %lu pos %llu len %u flags %u",
1375                    inode->i_sb->s_id, inode->i_ino,
1376                    (unsigned long long) pos, len, flags);
1377         index = pos >> PAGE_CACHE_SHIFT;
1378         from = pos & (PAGE_CACHE_SIZE - 1);
1379         to = from + len;
1380
1381 retry:
1382         handle = ext4_journal_start(inode, needed_blocks);
1383         if (IS_ERR(handle)) {
1384                 ret = PTR_ERR(handle);
1385                 goto out;
1386         }
1387
1388         /* We cannot recurse into the filesystem as the transaction is already
1389          * started */
1390         flags |= AOP_FLAG_NOFS;
1391
1392         page = grab_cache_page_write_begin(mapping, index, flags);
1393         if (!page) {
1394                 ext4_journal_stop(handle);
1395                 ret = -ENOMEM;
1396                 goto out;
1397         }
1398         *pagep = page;
1399
1400         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1401                                 ext4_get_block);
1402
1403         if (!ret && ext4_should_journal_data(inode)) {
1404                 ret = walk_page_buffers(handle, page_buffers(page),
1405                                 from, to, NULL, do_journal_get_write_access);
1406         }
1407
1408         if (ret) {
1409                 unlock_page(page);
1410                 ext4_journal_stop(handle);
1411                 page_cache_release(page);
1412                 /*
1413                  * block_write_begin may have instantiated a few blocks
1414                  * outside i_size.  Trim these off again. Don't need
1415                  * i_size_read because we hold i_mutex.
1416                  */
1417                 if (pos + len > inode->i_size)
1418                         vmtruncate(inode, inode->i_size);
1419         }
1420
1421         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1422                 goto retry;
1423 out:
1424         return ret;
1425 }
1426
1427 /* For write_end() in data=journal mode */
1428 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1429 {
1430         if (!buffer_mapped(bh) || buffer_freed(bh))
1431                 return 0;
1432         set_buffer_uptodate(bh);
1433         return ext4_handle_dirty_metadata(handle, NULL, bh);
1434 }
1435
1436 /*
1437  * We need to pick up the new inode size which generic_commit_write gave us
1438  * `file' can be NULL - eg, when called from page_symlink().
1439  *
1440  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1441  * buffers are managed internally.
1442  */
1443 static int ext4_ordered_write_end(struct file *file,
1444                                 struct address_space *mapping,
1445                                 loff_t pos, unsigned len, unsigned copied,
1446                                 struct page *page, void *fsdata)
1447 {
1448         handle_t *handle = ext4_journal_current_handle();
1449         struct inode *inode = mapping->host;
1450         int ret = 0, ret2;
1451
1452         trace_mark(ext4_ordered_write_end,
1453                    "dev %s ino %lu pos %llu len %u copied %u",
1454                    inode->i_sb->s_id, inode->i_ino,
1455                    (unsigned long long) pos, len, copied);
1456         ret = ext4_jbd2_file_inode(handle, inode);
1457
1458         if (ret == 0) {
1459                 loff_t new_i_size;
1460
1461                 new_i_size = pos + copied;
1462                 if (new_i_size > EXT4_I(inode)->i_disksize) {
1463                         ext4_update_i_disksize(inode, new_i_size);
1464                         /* We need to mark inode dirty even if
1465                          * new_i_size is less that inode->i_size
1466                          * bu greater than i_disksize.(hint delalloc)
1467                          */
1468                         ext4_mark_inode_dirty(handle, inode);
1469                 }
1470
1471                 ret2 = generic_write_end(file, mapping, pos, len, copied,
1472                                                         page, fsdata);
1473                 copied = ret2;
1474                 if (ret2 < 0)
1475                         ret = ret2;
1476         }
1477         ret2 = ext4_journal_stop(handle);
1478         if (!ret)
1479                 ret = ret2;
1480
1481         return ret ? ret : copied;
1482 }
1483
1484 static int ext4_writeback_write_end(struct file *file,
1485                                 struct address_space *mapping,
1486                                 loff_t pos, unsigned len, unsigned copied,
1487                                 struct page *page, void *fsdata)
1488 {
1489         handle_t *handle = ext4_journal_current_handle();
1490         struct inode *inode = mapping->host;
1491         int ret = 0, ret2;
1492         loff_t new_i_size;
1493
1494         trace_mark(ext4_writeback_write_end,
1495                    "dev %s ino %lu pos %llu len %u copied %u",
1496                    inode->i_sb->s_id, inode->i_ino,
1497                    (unsigned long long) pos, len, copied);
1498         new_i_size = pos + copied;
1499         if (new_i_size > EXT4_I(inode)->i_disksize) {
1500                 ext4_update_i_disksize(inode, new_i_size);
1501                 /* We need to mark inode dirty even if
1502                  * new_i_size is less that inode->i_size
1503                  * bu greater than i_disksize.(hint delalloc)
1504                  */
1505                 ext4_mark_inode_dirty(handle, inode);
1506         }
1507
1508         ret2 = generic_write_end(file, mapping, pos, len, copied,
1509                                                         page, fsdata);
1510         copied = ret2;
1511         if (ret2 < 0)
1512                 ret = ret2;
1513
1514         ret2 = ext4_journal_stop(handle);
1515         if (!ret)
1516                 ret = ret2;
1517
1518         return ret ? ret : copied;
1519 }
1520
1521 static int ext4_journalled_write_end(struct file *file,
1522                                 struct address_space *mapping,
1523                                 loff_t pos, unsigned len, unsigned copied,
1524                                 struct page *page, void *fsdata)
1525 {
1526         handle_t *handle = ext4_journal_current_handle();
1527         struct inode *inode = mapping->host;
1528         int ret = 0, ret2;
1529         int partial = 0;
1530         unsigned from, to;
1531         loff_t new_i_size;
1532
1533         trace_mark(ext4_journalled_write_end,
1534                    "dev %s ino %lu pos %llu len %u copied %u",
1535                    inode->i_sb->s_id, inode->i_ino,
1536                    (unsigned long long) pos, len, copied);
1537         from = pos & (PAGE_CACHE_SIZE - 1);
1538         to = from + len;
1539
1540         if (copied < len) {
1541                 if (!PageUptodate(page))
1542                         copied = 0;
1543                 page_zero_new_buffers(page, from+copied, to);
1544         }
1545
1546         ret = walk_page_buffers(handle, page_buffers(page), from,
1547                                 to, &partial, write_end_fn);
1548         if (!partial)
1549                 SetPageUptodate(page);
1550         new_i_size = pos + copied;
1551         if (new_i_size > inode->i_size)
1552                 i_size_write(inode, pos+copied);
1553         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
1554         if (new_i_size > EXT4_I(inode)->i_disksize) {
1555                 ext4_update_i_disksize(inode, new_i_size);
1556                 ret2 = ext4_mark_inode_dirty(handle, inode);
1557                 if (!ret)
1558                         ret = ret2;
1559         }
1560
1561         unlock_page(page);
1562         ret2 = ext4_journal_stop(handle);
1563         if (!ret)
1564                 ret = ret2;
1565         page_cache_release(page);
1566
1567         return ret ? ret : copied;
1568 }
1569
1570 static int ext4_da_reserve_space(struct inode *inode, int nrblocks)
1571 {
1572         int retries = 0;
1573         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1574         unsigned long md_needed, mdblocks, total = 0;
1575
1576         /*
1577          * recalculate the amount of metadata blocks to reserve
1578          * in order to allocate nrblocks
1579          * worse case is one extent per block
1580          */
1581 repeat:
1582         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1583         total = EXT4_I(inode)->i_reserved_data_blocks + nrblocks;
1584         mdblocks = ext4_calc_metadata_amount(inode, total);
1585         BUG_ON(mdblocks < EXT4_I(inode)->i_reserved_meta_blocks);
1586
1587         md_needed = mdblocks - EXT4_I(inode)->i_reserved_meta_blocks;
1588         total = md_needed + nrblocks;
1589
1590         /*
1591          * Make quota reservation here to prevent quota overflow
1592          * later. Real quota accounting is done at pages writeout
1593          * time.
1594          */
1595         if (vfs_dq_reserve_block(inode, total)) {
1596                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1597                 return -EDQUOT;
1598         }
1599
1600         if (ext4_claim_free_blocks(sbi, total)) {
1601                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1602                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1603                         yield();
1604                         goto repeat;
1605                 }
1606                 vfs_dq_release_reservation_block(inode, total);
1607                 return -ENOSPC;
1608         }
1609         EXT4_I(inode)->i_reserved_data_blocks += nrblocks;
1610         EXT4_I(inode)->i_reserved_meta_blocks = mdblocks;
1611
1612         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1613         return 0;       /* success */
1614 }
1615
1616 static void ext4_da_release_space(struct inode *inode, int to_free)
1617 {
1618         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1619         int total, mdb, mdb_free, release;
1620
1621         if (!to_free)
1622                 return;         /* Nothing to release, exit */
1623
1624         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1625
1626         if (!EXT4_I(inode)->i_reserved_data_blocks) {
1627                 /*
1628                  * if there is no reserved blocks, but we try to free some
1629                  * then the counter is messed up somewhere.
1630                  * but since this function is called from invalidate
1631                  * page, it's harmless to return without any action
1632                  */
1633                 printk(KERN_INFO "ext4 delalloc try to release %d reserved "
1634                             "blocks for inode %lu, but there is no reserved "
1635                             "data blocks\n", to_free, inode->i_ino);
1636                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1637                 return;
1638         }
1639
1640         /* recalculate the number of metablocks still need to be reserved */
1641         total = EXT4_I(inode)->i_reserved_data_blocks - to_free;
1642         mdb = ext4_calc_metadata_amount(inode, total);
1643
1644         /* figure out how many metablocks to release */
1645         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1646         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1647
1648         release = to_free + mdb_free;
1649
1650         /* update fs dirty blocks counter for truncate case */
1651         percpu_counter_sub(&sbi->s_dirtyblocks_counter, release);
1652
1653         /* update per-inode reservations */
1654         BUG_ON(to_free > EXT4_I(inode)->i_reserved_data_blocks);
1655         EXT4_I(inode)->i_reserved_data_blocks -= to_free;
1656
1657         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1658         EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1659         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1660
1661         vfs_dq_release_reservation_block(inode, release);
1662 }
1663
1664 static void ext4_da_page_release_reservation(struct page *page,
1665                                                 unsigned long offset)
1666 {
1667         int to_release = 0;
1668         struct buffer_head *head, *bh;
1669         unsigned int curr_off = 0;
1670
1671         head = page_buffers(page);
1672         bh = head;
1673         do {
1674                 unsigned int next_off = curr_off + bh->b_size;
1675
1676                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1677                         to_release++;
1678                         clear_buffer_delay(bh);
1679                 }
1680                 curr_off = next_off;
1681         } while ((bh = bh->b_this_page) != head);
1682         ext4_da_release_space(page->mapping->host, to_release);
1683 }
1684
1685 /*
1686  * Delayed allocation stuff
1687  */
1688
1689 struct mpage_da_data {
1690         struct inode *inode;
1691         struct buffer_head lbh;                 /* extent of blocks */
1692         unsigned long first_page, next_page;    /* extent of pages */
1693         get_block_t *get_block;
1694         struct writeback_control *wbc;
1695         int io_done;
1696         int pages_written;
1697         int retval;
1698 };
1699
1700 /*
1701  * mpage_da_submit_io - walks through extent of pages and try to write
1702  * them with writepage() call back
1703  *
1704  * @mpd->inode: inode
1705  * @mpd->first_page: first page of the extent
1706  * @mpd->next_page: page after the last page of the extent
1707  * @mpd->get_block: the filesystem's block mapper function
1708  *
1709  * By the time mpage_da_submit_io() is called we expect all blocks
1710  * to be allocated. this may be wrong if allocation failed.
1711  *
1712  * As pages are already locked by write_cache_pages(), we can't use it
1713  */
1714 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1715 {
1716         long pages_skipped;
1717         struct pagevec pvec;
1718         unsigned long index, end;
1719         int ret = 0, err, nr_pages, i;
1720         struct inode *inode = mpd->inode;
1721         struct address_space *mapping = inode->i_mapping;
1722
1723         BUG_ON(mpd->next_page <= mpd->first_page);
1724         /*
1725          * We need to start from the first_page to the next_page - 1
1726          * to make sure we also write the mapped dirty buffer_heads.
1727          * If we look at mpd->lbh.b_blocknr we would only be looking
1728          * at the currently mapped buffer_heads.
1729          */
1730         index = mpd->first_page;
1731         end = mpd->next_page - 1;
1732
1733         pagevec_init(&pvec, 0);
1734         while (index <= end) {
1735                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1736                 if (nr_pages == 0)
1737                         break;
1738                 for (i = 0; i < nr_pages; i++) {
1739                         struct page *page = pvec.pages[i];
1740
1741                         index = page->index;
1742                         if (index > end)
1743                                 break;
1744                         index++;
1745
1746                         BUG_ON(!PageLocked(page));
1747                         BUG_ON(PageWriteback(page));
1748
1749                         pages_skipped = mpd->wbc->pages_skipped;
1750                         err = mapping->a_ops->writepage(page, mpd->wbc);
1751                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
1752                                 /*
1753                                  * have successfully written the page
1754                                  * without skipping the same
1755                                  */
1756                                 mpd->pages_written++;
1757                         /*
1758                          * In error case, we have to continue because
1759                          * remaining pages are still locked
1760                          * XXX: unlock and re-dirty them?
1761                          */
1762                         if (ret == 0)
1763                                 ret = err;
1764                 }
1765                 pagevec_release(&pvec);
1766         }
1767         return ret;
1768 }
1769
1770 /*
1771  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
1772  *
1773  * @mpd->inode - inode to walk through
1774  * @exbh->b_blocknr - first block on a disk
1775  * @exbh->b_size - amount of space in bytes
1776  * @logical - first logical block to start assignment with
1777  *
1778  * the function goes through all passed space and put actual disk
1779  * block numbers into buffer heads, dropping BH_Delay
1780  */
1781 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
1782                                  struct buffer_head *exbh)
1783 {
1784         struct inode *inode = mpd->inode;
1785         struct address_space *mapping = inode->i_mapping;
1786         int blocks = exbh->b_size >> inode->i_blkbits;
1787         sector_t pblock = exbh->b_blocknr, cur_logical;
1788         struct buffer_head *head, *bh;
1789         pgoff_t index, end;
1790         struct pagevec pvec;
1791         int nr_pages, i;
1792
1793         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1794         end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1795         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1796
1797         pagevec_init(&pvec, 0);
1798
1799         while (index <= end) {
1800                 /* XXX: optimize tail */
1801                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1802                 if (nr_pages == 0)
1803                         break;
1804                 for (i = 0; i < nr_pages; i++) {
1805                         struct page *page = pvec.pages[i];
1806
1807                         index = page->index;
1808                         if (index > end)
1809                                 break;
1810                         index++;
1811
1812                         BUG_ON(!PageLocked(page));
1813                         BUG_ON(PageWriteback(page));
1814                         BUG_ON(!page_has_buffers(page));
1815
1816                         bh = page_buffers(page);
1817                         head = bh;
1818
1819                         /* skip blocks out of the range */
1820                         do {
1821                                 if (cur_logical >= logical)
1822                                         break;
1823                                 cur_logical++;
1824                         } while ((bh = bh->b_this_page) != head);
1825
1826                         do {
1827                                 if (cur_logical >= logical + blocks)
1828                                         break;
1829                                 if (buffer_delay(bh)) {
1830                                         bh->b_blocknr = pblock;
1831                                         clear_buffer_delay(bh);
1832                                         bh->b_bdev = inode->i_sb->s_bdev;
1833                                 } else if (buffer_unwritten(bh)) {
1834                                         bh->b_blocknr = pblock;
1835                                         clear_buffer_unwritten(bh);
1836                                         set_buffer_mapped(bh);
1837                                         set_buffer_new(bh);
1838                                         bh->b_bdev = inode->i_sb->s_bdev;
1839                                 } else if (buffer_mapped(bh))
1840                                         BUG_ON(bh->b_blocknr != pblock);
1841
1842                                 cur_logical++;
1843                                 pblock++;
1844                         } while ((bh = bh->b_this_page) != head);
1845                 }
1846                 pagevec_release(&pvec);
1847         }
1848 }
1849
1850
1851 /*
1852  * __unmap_underlying_blocks - just a helper function to unmap
1853  * set of blocks described by @bh
1854  */
1855 static inline void __unmap_underlying_blocks(struct inode *inode,
1856                                              struct buffer_head *bh)
1857 {
1858         struct block_device *bdev = inode->i_sb->s_bdev;
1859         int blocks, i;
1860
1861         blocks = bh->b_size >> inode->i_blkbits;
1862         for (i = 0; i < blocks; i++)
1863                 unmap_underlying_metadata(bdev, bh->b_blocknr + i);
1864 }
1865
1866 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
1867                                         sector_t logical, long blk_cnt)
1868 {
1869         int nr_pages, i;
1870         pgoff_t index, end;
1871         struct pagevec pvec;
1872         struct inode *inode = mpd->inode;
1873         struct address_space *mapping = inode->i_mapping;
1874
1875         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1876         end   = (logical + blk_cnt - 1) >>
1877                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
1878         while (index <= end) {
1879                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1880                 if (nr_pages == 0)
1881                         break;
1882                 for (i = 0; i < nr_pages; i++) {
1883                         struct page *page = pvec.pages[i];
1884                         index = page->index;
1885                         if (index > end)
1886                                 break;
1887                         index++;
1888
1889                         BUG_ON(!PageLocked(page));
1890                         BUG_ON(PageWriteback(page));
1891                         block_invalidatepage(page, 0);
1892                         ClearPageUptodate(page);
1893                         unlock_page(page);
1894                 }
1895         }
1896         return;
1897 }
1898
1899 static void ext4_print_free_blocks(struct inode *inode)
1900 {
1901         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1902         printk(KERN_EMERG "Total free blocks count %lld\n",
1903                         ext4_count_free_blocks(inode->i_sb));
1904         printk(KERN_EMERG "Free/Dirty block details\n");
1905         printk(KERN_EMERG "free_blocks=%lld\n",
1906                         (long long)percpu_counter_sum(&sbi->s_freeblocks_counter));
1907         printk(KERN_EMERG "dirty_blocks=%lld\n",
1908                         (long long)percpu_counter_sum(&sbi->s_dirtyblocks_counter));
1909         printk(KERN_EMERG "Block reservation details\n");
1910         printk(KERN_EMERG "i_reserved_data_blocks=%u\n",
1911                         EXT4_I(inode)->i_reserved_data_blocks);
1912         printk(KERN_EMERG "i_reserved_meta_blocks=%u\n",
1913                         EXT4_I(inode)->i_reserved_meta_blocks);
1914         return;
1915 }
1916
1917 /*
1918  * mpage_da_map_blocks - go through given space
1919  *
1920  * @mpd->lbh - bh describing space
1921  * @mpd->get_block - the filesystem's block mapper function
1922  *
1923  * The function skips space we know is already mapped to disk blocks.
1924  *
1925  */
1926 static int  mpage_da_map_blocks(struct mpage_da_data *mpd)
1927 {
1928         int err = 0;
1929         struct buffer_head new;
1930         struct buffer_head *lbh = &mpd->lbh;
1931         sector_t next;
1932
1933         /*
1934          * We consider only non-mapped and non-allocated blocks
1935          */
1936         if (buffer_mapped(lbh) && !buffer_delay(lbh))
1937                 return 0;
1938         new.b_state = lbh->b_state;
1939         new.b_blocknr = 0;
1940         new.b_size = lbh->b_size;
1941         next = lbh->b_blocknr;
1942         /*
1943          * If we didn't accumulate anything
1944          * to write simply return
1945          */
1946         if (!new.b_size)
1947                 return 0;
1948         err = mpd->get_block(mpd->inode, next, &new, 1);
1949         if (err) {
1950
1951                 /* If get block returns with error
1952                  * we simply return. Later writepage
1953                  * will redirty the page and writepages
1954                  * will find the dirty page again
1955                  */
1956                 if (err == -EAGAIN)
1957                         return 0;
1958
1959                 if (err == -ENOSPC &&
1960                                 ext4_count_free_blocks(mpd->inode->i_sb)) {
1961                         mpd->retval = err;
1962                         return 0;
1963                 }
1964
1965                 /*
1966                  * get block failure will cause us
1967                  * to loop in writepages. Because
1968                  * a_ops->writepage won't be able to
1969                  * make progress. The page will be redirtied
1970                  * by writepage and writepages will again
1971                  * try to write the same.
1972                  */
1973                 printk(KERN_EMERG "%s block allocation failed for inode %lu "
1974                                   "at logical offset %llu with max blocks "
1975                                   "%zd with error %d\n",
1976                                   __func__, mpd->inode->i_ino,
1977                                   (unsigned long long)next,
1978                                   lbh->b_size >> mpd->inode->i_blkbits, err);
1979                 printk(KERN_EMERG "This should not happen.!! "
1980                                         "Data will be lost\n");
1981                 if (err == -ENOSPC) {
1982                         ext4_print_free_blocks(mpd->inode);
1983                 }
1984                 /* invlaidate all the pages */
1985                 ext4_da_block_invalidatepages(mpd, next,
1986                                 lbh->b_size >> mpd->inode->i_blkbits);
1987                 return err;
1988         }
1989         BUG_ON(new.b_size == 0);
1990
1991         if (buffer_new(&new))
1992                 __unmap_underlying_blocks(mpd->inode, &new);
1993
1994         /*
1995          * If blocks are delayed marked, we need to
1996          * put actual blocknr and drop delayed bit
1997          */
1998         if (buffer_delay(lbh) || buffer_unwritten(lbh))
1999                 mpage_put_bnr_to_bhs(mpd, next, &new);
2000
2001         return 0;
2002 }
2003
2004 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2005                 (1 << BH_Delay) | (1 << BH_Unwritten))
2006
2007 /*
2008  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2009  *
2010  * @mpd->lbh - extent of blocks
2011  * @logical - logical number of the block in the file
2012  * @bh - bh of the block (used to access block's state)
2013  *
2014  * the function is used to collect contig. blocks in same state
2015  */
2016 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2017                                    sector_t logical, struct buffer_head *bh)
2018 {
2019         sector_t next;
2020         size_t b_size = bh->b_size;
2021         struct buffer_head *lbh = &mpd->lbh;
2022         int nrblocks = lbh->b_size >> mpd->inode->i_blkbits;
2023
2024         /* check if thereserved journal credits might overflow */
2025         if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
2026                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2027                         /*
2028                          * With non-extent format we are limited by the journal
2029                          * credit available.  Total credit needed to insert
2030                          * nrblocks contiguous blocks is dependent on the
2031                          * nrblocks.  So limit nrblocks.
2032                          */
2033                         goto flush_it;
2034                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2035                                 EXT4_MAX_TRANS_DATA) {
2036                         /*
2037                          * Adding the new buffer_head would make it cross the
2038                          * allowed limit for which we have journal credit
2039                          * reserved. So limit the new bh->b_size
2040                          */
2041                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2042                                                 mpd->inode->i_blkbits;
2043                         /* we will do mpage_da_submit_io in the next loop */
2044                 }
2045         }
2046         /*
2047          * First block in the extent
2048          */
2049         if (lbh->b_size == 0) {
2050                 lbh->b_blocknr = logical;
2051                 lbh->b_size = b_size;
2052                 lbh->b_state = bh->b_state & BH_FLAGS;
2053                 return;
2054         }
2055
2056         next = lbh->b_blocknr + nrblocks;
2057         /*
2058          * Can we merge the block to our big extent?
2059          */
2060         if (logical == next && (bh->b_state & BH_FLAGS) == lbh->b_state) {
2061                 lbh->b_size += b_size;
2062                 return;
2063         }
2064
2065 flush_it:
2066         /*
2067          * We couldn't merge the block to our extent, so we
2068          * need to flush current  extent and start new one
2069          */
2070         if (mpage_da_map_blocks(mpd) == 0)
2071                 mpage_da_submit_io(mpd);
2072         mpd->io_done = 1;
2073         return;
2074 }
2075
2076 /*
2077  * __mpage_da_writepage - finds extent of pages and blocks
2078  *
2079  * @page: page to consider
2080  * @wbc: not used, we just follow rules
2081  * @data: context
2082  *
2083  * The function finds extents of pages and scan them for all blocks.
2084  */
2085 static int __mpage_da_writepage(struct page *page,
2086                                 struct writeback_control *wbc, void *data)
2087 {
2088         struct mpage_da_data *mpd = data;
2089         struct inode *inode = mpd->inode;
2090         struct buffer_head *bh, *head, fake;
2091         sector_t logical;
2092
2093         if (mpd->io_done) {
2094                 /*
2095                  * Rest of the page in the page_vec
2096                  * redirty then and skip then. We will
2097                  * try to to write them again after
2098                  * starting a new transaction
2099                  */
2100                 redirty_page_for_writepage(wbc, page);
2101                 unlock_page(page);
2102                 return MPAGE_DA_EXTENT_TAIL;
2103         }
2104         /*
2105          * Can we merge this page to current extent?
2106          */
2107         if (mpd->next_page != page->index) {
2108                 /*
2109                  * Nope, we can't. So, we map non-allocated blocks
2110                  * and start IO on them using writepage()
2111                  */
2112                 if (mpd->next_page != mpd->first_page) {
2113                         if (mpage_da_map_blocks(mpd) == 0)
2114                                 mpage_da_submit_io(mpd);
2115                         /*
2116                          * skip rest of the page in the page_vec
2117                          */
2118                         mpd->io_done = 1;
2119                         redirty_page_for_writepage(wbc, page);
2120                         unlock_page(page);
2121                         return MPAGE_DA_EXTENT_TAIL;
2122                 }
2123
2124                 /*
2125                  * Start next extent of pages ...
2126                  */
2127                 mpd->first_page = page->index;
2128
2129                 /*
2130                  * ... and blocks
2131                  */
2132                 mpd->lbh.b_size = 0;
2133                 mpd->lbh.b_state = 0;
2134                 mpd->lbh.b_blocknr = 0;
2135         }
2136
2137         mpd->next_page = page->index + 1;
2138         logical = (sector_t) page->index <<
2139                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2140
2141         if (!page_has_buffers(page)) {
2142                 /*
2143                  * There is no attached buffer heads yet (mmap?)
2144                  * we treat the page asfull of dirty blocks
2145                  */
2146                 bh = &fake;
2147                 bh->b_size = PAGE_CACHE_SIZE;
2148                 bh->b_state = 0;
2149                 set_buffer_dirty(bh);
2150                 set_buffer_uptodate(bh);
2151                 mpage_add_bh_to_extent(mpd, logical, bh);
2152                 if (mpd->io_done)
2153                         return MPAGE_DA_EXTENT_TAIL;
2154         } else {
2155                 /*
2156                  * Page with regular buffer heads, just add all dirty ones
2157                  */
2158                 head = page_buffers(page);
2159                 bh = head;
2160                 do {
2161                         BUG_ON(buffer_locked(bh));
2162                         /*
2163                          * We need to try to allocate
2164                          * unmapped blocks in the same page.
2165                          * Otherwise we won't make progress
2166                          * with the page in ext4_da_writepage
2167                          */
2168                         if (buffer_dirty(bh) &&
2169                                 (!buffer_mapped(bh) || buffer_delay(bh))) {
2170                                 mpage_add_bh_to_extent(mpd, logical, bh);
2171                                 if (mpd->io_done)
2172                                         return MPAGE_DA_EXTENT_TAIL;
2173                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2174                                 /*
2175                                  * mapped dirty buffer. We need to update
2176                                  * the b_state because we look at
2177                                  * b_state in mpage_da_map_blocks. We don't
2178                                  * update b_size because if we find an
2179                                  * unmapped buffer_head later we need to
2180                                  * use the b_state flag of that buffer_head.
2181                                  */
2182                                 if (mpd->lbh.b_size == 0)
2183                                         mpd->lbh.b_state =
2184                                                 bh->b_state & BH_FLAGS;
2185                         }
2186                         logical++;
2187                 } while ((bh = bh->b_this_page) != head);
2188         }
2189
2190         return 0;
2191 }
2192
2193 /*
2194  * mpage_da_writepages - walk the list of dirty pages of the given
2195  * address space, allocates non-allocated blocks, maps newly-allocated
2196  * blocks to existing bhs and issue IO them
2197  *
2198  * @mapping: address space structure to write
2199  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2200  * @get_block: the filesystem's block mapper function.
2201  *
2202  * This is a library function, which implements the writepages()
2203  * address_space_operation.
2204  */
2205 static int mpage_da_writepages(struct address_space *mapping,
2206                                struct writeback_control *wbc,
2207                                struct mpage_da_data *mpd)
2208 {
2209         int ret;
2210
2211         if (!mpd->get_block)
2212                 return generic_writepages(mapping, wbc);
2213
2214         mpd->lbh.b_size = 0;
2215         mpd->lbh.b_state = 0;
2216         mpd->lbh.b_blocknr = 0;
2217         mpd->first_page = 0;
2218         mpd->next_page = 0;
2219         mpd->io_done = 0;
2220         mpd->pages_written = 0;
2221         mpd->retval = 0;
2222
2223         ret = write_cache_pages(mapping, wbc, __mpage_da_writepage, mpd);
2224         /*
2225          * Handle last extent of pages
2226          */
2227         if (!mpd->io_done && mpd->next_page != mpd->first_page) {
2228                 if (mpage_da_map_blocks(mpd) == 0)
2229                         mpage_da_submit_io(mpd);
2230
2231                 mpd->io_done = 1;
2232                 ret = MPAGE_DA_EXTENT_TAIL;
2233         }
2234         wbc->nr_to_write -= mpd->pages_written;
2235         return ret;
2236 }
2237
2238 /*
2239  * this is a special callback for ->write_begin() only
2240  * it's intention is to return mapped block or reserve space
2241  */
2242 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2243                                   struct buffer_head *bh_result, int create)
2244 {
2245         int ret = 0;
2246
2247         BUG_ON(create == 0);
2248         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2249
2250         /*
2251          * first, we need to know whether the block is allocated already
2252          * preallocated blocks are unmapped but should treated
2253          * the same as allocated blocks.
2254          */
2255         ret = ext4_get_blocks_wrap(NULL, inode, iblock, 1,  bh_result, 0, 0, 0);
2256         if ((ret == 0) && !buffer_delay(bh_result)) {
2257                 /* the block isn't (pre)allocated yet, let's reserve space */
2258                 /*
2259                  * XXX: __block_prepare_write() unmaps passed block,
2260                  * is it OK?
2261                  */
2262                 ret = ext4_da_reserve_space(inode, 1);
2263                 if (ret)
2264                         /* not enough space to reserve */
2265                         return ret;
2266
2267                 map_bh(bh_result, inode->i_sb, 0);
2268                 set_buffer_new(bh_result);
2269                 set_buffer_delay(bh_result);
2270         } else if (ret > 0) {
2271                 bh_result->b_size = (ret << inode->i_blkbits);
2272                 ret = 0;
2273         }
2274
2275         return ret;
2276 }
2277 #define         EXT4_DELALLOC_RSVED     1
2278 static int ext4_da_get_block_write(struct inode *inode, sector_t iblock,
2279                                    struct buffer_head *bh_result, int create)
2280 {
2281         int ret;
2282         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2283         loff_t disksize = EXT4_I(inode)->i_disksize;
2284         handle_t *handle = NULL;
2285
2286         handle = ext4_journal_current_handle();
2287         BUG_ON(!handle);
2288         ret = ext4_get_blocks_wrap(handle, inode, iblock, max_blocks,
2289                         bh_result, create, 0, EXT4_DELALLOC_RSVED);
2290         if (ret > 0) {
2291
2292                 bh_result->b_size = (ret << inode->i_blkbits);
2293
2294                 if (ext4_should_order_data(inode)) {
2295                         int retval;
2296                         retval = ext4_jbd2_file_inode(handle, inode);
2297                         if (retval)
2298                                 /*
2299                                  * Failed to add inode for ordered
2300                                  * mode. Don't update file size
2301                                  */
2302                                 return retval;
2303                 }
2304
2305                 /*
2306                  * Update on-disk size along with block allocation
2307                  * we don't use 'extend_disksize' as size may change
2308                  * within already allocated block -bzzz
2309                  */
2310                 disksize = ((loff_t) iblock + ret) << inode->i_blkbits;
2311                 if (disksize > i_size_read(inode))
2312                         disksize = i_size_read(inode);
2313                 if (disksize > EXT4_I(inode)->i_disksize) {
2314                         ext4_update_i_disksize(inode, disksize);
2315                         ret = ext4_mark_inode_dirty(handle, inode);
2316                         return ret;
2317                 }
2318                 ret = 0;
2319         }
2320         return ret;
2321 }
2322
2323 static int ext4_bh_unmapped_or_delay(handle_t *handle, struct buffer_head *bh)
2324 {
2325         /*
2326          * unmapped buffer is possible for holes.
2327          * delay buffer is possible with delayed allocation
2328          */
2329         return ((!buffer_mapped(bh) || buffer_delay(bh)) && buffer_dirty(bh));
2330 }
2331
2332 static int ext4_normal_get_block_write(struct inode *inode, sector_t iblock,
2333                                    struct buffer_head *bh_result, int create)
2334 {
2335         int ret = 0;
2336         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2337
2338         /*
2339          * we don't want to do block allocation in writepage
2340          * so call get_block_wrap with create = 0
2341          */
2342         ret = ext4_get_blocks_wrap(NULL, inode, iblock, max_blocks,
2343                                    bh_result, 0, 0, 0);
2344         if (ret > 0) {
2345                 bh_result->b_size = (ret << inode->i_blkbits);
2346                 ret = 0;
2347         }
2348         return ret;
2349 }
2350
2351 /*
2352  * get called vi ext4_da_writepages after taking page lock (have journal handle)
2353  * get called via journal_submit_inode_data_buffers (no journal handle)
2354  * get called via shrink_page_list via pdflush (no journal handle)
2355  * or grab_page_cache when doing write_begin (have journal handle)
2356  */
2357 static int ext4_da_writepage(struct page *page,
2358                                 struct writeback_control *wbc)
2359 {
2360         int ret = 0;
2361         loff_t size;
2362         unsigned int len;
2363         struct buffer_head *page_bufs;
2364         struct inode *inode = page->mapping->host;
2365
2366         trace_mark(ext4_da_writepage,
2367                    "dev %s ino %lu page_index %lu",
2368                    inode->i_sb->s_id, inode->i_ino, page->index);
2369         size = i_size_read(inode);
2370         if (page->index == size >> PAGE_CACHE_SHIFT)
2371                 len = size & ~PAGE_CACHE_MASK;
2372         else
2373                 len = PAGE_CACHE_SIZE;
2374
2375         if (page_has_buffers(page)) {
2376                 page_bufs = page_buffers(page);
2377                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2378                                         ext4_bh_unmapped_or_delay)) {
2379                         /*
2380                          * We don't want to do  block allocation
2381                          * So redirty the page and return
2382                          * We may reach here when we do a journal commit
2383                          * via journal_submit_inode_data_buffers.
2384                          * If we don't have mapping block we just ignore
2385                          * them. We can also reach here via shrink_page_list
2386                          */
2387                         redirty_page_for_writepage(wbc, page);
2388                         unlock_page(page);
2389                         return 0;
2390                 }
2391         } else {
2392                 /*
2393                  * The test for page_has_buffers() is subtle:
2394                  * We know the page is dirty but it lost buffers. That means
2395                  * that at some moment in time after write_begin()/write_end()
2396                  * has been called all buffers have been clean and thus they
2397                  * must have been written at least once. So they are all
2398                  * mapped and we can happily proceed with mapping them
2399                  * and writing the page.
2400                  *
2401                  * Try to initialize the buffer_heads and check whether
2402                  * all are mapped and non delay. We don't want to
2403                  * do block allocation here.
2404                  */
2405                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
2406                                                 ext4_normal_get_block_write);
2407                 if (!ret) {
2408                         page_bufs = page_buffers(page);
2409                         /* check whether all are mapped and non delay */
2410                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2411                                                 ext4_bh_unmapped_or_delay)) {
2412                                 redirty_page_for_writepage(wbc, page);
2413                                 unlock_page(page);
2414                                 return 0;
2415                         }
2416                 } else {
2417                         /*
2418                          * We can't do block allocation here
2419                          * so just redity the page and unlock
2420                          * and return
2421                          */
2422                         redirty_page_for_writepage(wbc, page);
2423                         unlock_page(page);
2424                         return 0;
2425                 }
2426                 /* now mark the buffer_heads as dirty and uptodate */
2427                 block_commit_write(page, 0, PAGE_CACHE_SIZE);
2428         }
2429
2430         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2431                 ret = nobh_writepage(page, ext4_normal_get_block_write, wbc);
2432         else
2433                 ret = block_write_full_page(page,
2434                                                 ext4_normal_get_block_write,
2435                                                 wbc);
2436
2437         return ret;
2438 }
2439
2440 /*
2441  * This is called via ext4_da_writepages() to
2442  * calulate the total number of credits to reserve to fit
2443  * a single extent allocation into a single transaction,
2444  * ext4_da_writpeages() will loop calling this before
2445  * the block allocation.
2446  */
2447
2448 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2449 {
2450         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2451
2452         /*
2453          * With non-extent format the journal credit needed to
2454          * insert nrblocks contiguous block is dependent on
2455          * number of contiguous block. So we will limit
2456          * number of contiguous block to a sane value
2457          */
2458         if (!(inode->i_flags & EXT4_EXTENTS_FL) &&
2459             (max_blocks > EXT4_MAX_TRANS_DATA))
2460                 max_blocks = EXT4_MAX_TRANS_DATA;
2461
2462         return ext4_chunk_trans_blocks(inode, max_blocks);
2463 }
2464
2465 static int ext4_da_writepages(struct address_space *mapping,
2466                               struct writeback_control *wbc)
2467 {
2468         pgoff_t index;
2469         int range_whole = 0;
2470         handle_t *handle = NULL;
2471         struct mpage_da_data mpd;
2472         struct inode *inode = mapping->host;
2473         int no_nrwrite_index_update;
2474         int pages_written = 0;
2475         long pages_skipped;
2476         int range_cyclic, cycled = 1, io_done = 0;
2477         int needed_blocks, ret = 0, nr_to_writebump = 0;
2478         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2479
2480         trace_mark(ext4_da_writepages,
2481                    "dev %s ino %lu nr_t_write %ld "
2482                    "pages_skipped %ld range_start %llu "
2483                    "range_end %llu nonblocking %d "
2484                    "for_kupdate %d for_reclaim %d "
2485                    "for_writepages %d range_cyclic %d",
2486                    inode->i_sb->s_id, inode->i_ino,
2487                    wbc->nr_to_write, wbc->pages_skipped,
2488                    (unsigned long long) wbc->range_start,
2489                    (unsigned long long) wbc->range_end,
2490                    wbc->nonblocking, wbc->for_kupdate,
2491                    wbc->for_reclaim, wbc->for_writepages,
2492                    wbc->range_cyclic);
2493
2494         /*
2495          * No pages to write? This is mainly a kludge to avoid starting
2496          * a transaction for special inodes like journal inode on last iput()
2497          * because that could violate lock ordering on umount
2498          */
2499         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2500                 return 0;
2501
2502         /*
2503          * If the filesystem has aborted, it is read-only, so return
2504          * right away instead of dumping stack traces later on that
2505          * will obscure the real source of the problem.  We test
2506          * EXT4_MOUNT_ABORT instead of sb->s_flag's MS_RDONLY because
2507          * the latter could be true if the filesystem is mounted
2508          * read-only, and in that case, ext4_da_writepages should
2509          * *never* be called, so if that ever happens, we would want
2510          * the stack trace.
2511          */
2512         if (unlikely(sbi->s_mount_opt & EXT4_MOUNT_ABORT))
2513                 return -EROFS;
2514
2515         /*
2516          * Make sure nr_to_write is >= sbi->s_mb_stream_request
2517          * This make sure small files blocks are allocated in
2518          * single attempt. This ensure that small files
2519          * get less fragmented.
2520          */
2521         if (wbc->nr_to_write < sbi->s_mb_stream_request) {
2522                 nr_to_writebump = sbi->s_mb_stream_request - wbc->nr_to_write;
2523                 wbc->nr_to_write = sbi->s_mb_stream_request;
2524         }
2525         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2526                 range_whole = 1;
2527
2528         range_cyclic = wbc->range_cyclic;
2529         if (wbc->range_cyclic) {
2530                 index = mapping->writeback_index;
2531                 if (index)
2532                         cycled = 0;
2533                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2534                 wbc->range_end  = LLONG_MAX;
2535                 wbc->range_cyclic = 0;
2536         } else
2537                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2538
2539         mpd.wbc = wbc;
2540         mpd.inode = mapping->host;
2541
2542         /*
2543          * we don't want write_cache_pages to update
2544          * nr_to_write and writeback_index
2545          */
2546         no_nrwrite_index_update = wbc->no_nrwrite_index_update;
2547         wbc->no_nrwrite_index_update = 1;
2548         pages_skipped = wbc->pages_skipped;
2549
2550 retry:
2551         while (!ret && wbc->nr_to_write > 0) {
2552
2553                 /*
2554                  * we  insert one extent at a time. So we need
2555                  * credit needed for single extent allocation.
2556                  * journalled mode is currently not supported
2557                  * by delalloc
2558                  */
2559                 BUG_ON(ext4_should_journal_data(inode));
2560                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2561
2562                 /* start a new transaction*/
2563                 handle = ext4_journal_start(inode, needed_blocks);
2564                 if (IS_ERR(handle)) {
2565                         ret = PTR_ERR(handle);
2566                         printk(KERN_CRIT "%s: jbd2_start: "
2567                                "%ld pages, ino %lu; err %d\n", __func__,
2568                                 wbc->nr_to_write, inode->i_ino, ret);
2569                         dump_stack();
2570                         goto out_writepages;
2571                 }
2572                 mpd.get_block = ext4_da_get_block_write;
2573                 ret = mpage_da_writepages(mapping, wbc, &mpd);
2574
2575                 ext4_journal_stop(handle);
2576
2577                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2578                         /* commit the transaction which would
2579                          * free blocks released in the transaction
2580                          * and try again
2581                          */
2582                         jbd2_journal_force_commit_nested(sbi->s_journal);
2583                         wbc->pages_skipped = pages_skipped;
2584                         ret = 0;
2585                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2586                         /*
2587                          * got one extent now try with
2588                          * rest of the pages
2589                          */
2590                         pages_written += mpd.pages_written;
2591                         wbc->pages_skipped = pages_skipped;
2592                         ret = 0;
2593                         io_done = 1;
2594                 } else if (wbc->nr_to_write)
2595                         /*
2596                          * There is no more writeout needed
2597                          * or we requested for a noblocking writeout
2598                          * and we found the device congested
2599                          */
2600                         break;
2601         }
2602         if (!io_done && !cycled) {
2603                 cycled = 1;
2604                 index = 0;
2605                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2606                 wbc->range_end  = mapping->writeback_index - 1;
2607                 goto retry;
2608         }
2609         if (pages_skipped != wbc->pages_skipped)
2610                 printk(KERN_EMERG "This should not happen leaving %s "
2611                                 "with nr_to_write = %ld ret = %d\n",
2612                                 __func__, wbc->nr_to_write, ret);
2613
2614         /* Update index */
2615         index += pages_written;
2616         wbc->range_cyclic = range_cyclic;
2617         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2618                 /*
2619                  * set the writeback_index so that range_cyclic
2620                  * mode will write it back later
2621                  */
2622                 mapping->writeback_index = index;
2623
2624 out_writepages:
2625         if (!no_nrwrite_index_update)
2626                 wbc->no_nrwrite_index_update = 0;
2627         wbc->nr_to_write -= nr_to_writebump;
2628         trace_mark(ext4_da_writepage_result,
2629                    "dev %s ino %lu ret %d pages_written %d "
2630                    "pages_skipped %ld congestion %d "
2631                    "more_io %d no_nrwrite_index_update %d",
2632                    inode->i_sb->s_id, inode->i_ino, ret,
2633                    pages_written, wbc->pages_skipped,
2634                    wbc->encountered_congestion, wbc->more_io,
2635                    wbc->no_nrwrite_index_update);
2636         return ret;
2637 }
2638
2639 #define FALL_BACK_TO_NONDELALLOC 1
2640 static int ext4_nonda_switch(struct super_block *sb)
2641 {
2642         s64 free_blocks, dirty_blocks;
2643         struct ext4_sb_info *sbi = EXT4_SB(sb);
2644
2645         /*
2646          * switch to non delalloc mode if we are running low
2647          * on free block. The free block accounting via percpu
2648          * counters can get slightly wrong with percpu_counter_batch getting
2649          * accumulated on each CPU without updating global counters
2650          * Delalloc need an accurate free block accounting. So switch
2651          * to non delalloc when we are near to error range.
2652          */
2653         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
2654         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
2655         if (2 * free_blocks < 3 * dirty_blocks ||
2656                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
2657                 /*
2658                  * free block count is less that 150% of dirty blocks
2659                  * or free blocks is less that watermark
2660                  */
2661                 return 1;
2662         }
2663         return 0;
2664 }
2665
2666 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2667                                 loff_t pos, unsigned len, unsigned flags,
2668                                 struct page **pagep, void **fsdata)
2669 {
2670         int ret, retries = 0;
2671         struct page *page;
2672         pgoff_t index;
2673         unsigned from, to;
2674         struct inode *inode = mapping->host;
2675         handle_t *handle;
2676
2677         index = pos >> PAGE_CACHE_SHIFT;
2678         from = pos & (PAGE_CACHE_SIZE - 1);
2679         to = from + len;
2680
2681         if (ext4_nonda_switch(inode->i_sb)) {
2682                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2683                 return ext4_write_begin(file, mapping, pos,
2684                                         len, flags, pagep, fsdata);
2685         }
2686         *fsdata = (void *)0;
2687
2688         trace_mark(ext4_da_write_begin,
2689                    "dev %s ino %lu pos %llu len %u flags %u",
2690                    inode->i_sb->s_id, inode->i_ino,
2691                    (unsigned long long) pos, len, flags);
2692 retry:
2693         /*
2694          * With delayed allocation, we don't log the i_disksize update
2695          * if there is delayed block allocation. But we still need
2696          * to journalling the i_disksize update if writes to the end
2697          * of file which has an already mapped buffer.
2698          */
2699         handle = ext4_journal_start(inode, 1);
2700         if (IS_ERR(handle)) {
2701                 ret = PTR_ERR(handle);
2702                 goto out;
2703         }
2704         /* We cannot recurse into the filesystem as the transaction is already
2705          * started */
2706         flags |= AOP_FLAG_NOFS;
2707
2708         page = grab_cache_page_write_begin(mapping, index, flags);
2709         if (!page) {
2710                 ext4_journal_stop(handle);
2711                 ret = -ENOMEM;
2712                 goto out;
2713         }
2714         *pagep = page;
2715
2716         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
2717                                                         ext4_da_get_block_prep);
2718         if (ret < 0) {
2719                 unlock_page(page);
2720                 ext4_journal_stop(handle);
2721                 page_cache_release(page);
2722                 /*
2723                  * block_write_begin may have instantiated a few blocks
2724                  * outside i_size.  Trim these off again. Don't need
2725                  * i_size_read because we hold i_mutex.
2726                  */
2727                 if (pos + len > inode->i_size)
2728                         vmtruncate(inode, inode->i_size);
2729         }
2730
2731         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
2732                 goto retry;
2733 out:
2734         return ret;
2735 }
2736
2737 /*
2738  * Check if we should update i_disksize
2739  * when write to the end of file but not require block allocation
2740  */
2741 static int ext4_da_should_update_i_disksize(struct page *page,
2742                                          unsigned long offset)
2743 {
2744         struct buffer_head *bh;
2745         struct inode *inode = page->mapping->host;
2746         unsigned int idx;
2747         int i;
2748
2749         bh = page_buffers(page);
2750         idx = offset >> inode->i_blkbits;
2751
2752         for (i = 0; i < idx; i++)
2753                 bh = bh->b_this_page;
2754
2755         if (!buffer_mapped(bh) || (buffer_delay(bh)))
2756                 return 0;
2757         return 1;
2758 }
2759
2760 static int ext4_da_write_end(struct file *file,
2761                                 struct address_space *mapping,
2762                                 loff_t pos, unsigned len, unsigned copied,
2763                                 struct page *page, void *fsdata)
2764 {
2765         struct inode *inode = mapping->host;
2766         int ret = 0, ret2;
2767         handle_t *handle = ext4_journal_current_handle();
2768         loff_t new_i_size;
2769         unsigned long start, end;
2770         int write_mode = (int)(unsigned long)fsdata;
2771
2772         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
2773                 if (ext4_should_order_data(inode)) {
2774                         return ext4_ordered_write_end(file, mapping, pos,
2775                                         len, copied, page, fsdata);
2776                 } else if (ext4_should_writeback_data(inode)) {
2777                         return ext4_writeback_write_end(file, mapping, pos,
2778                                         len, copied, page, fsdata);
2779                 } else {
2780                         BUG();
2781                 }
2782         }
2783
2784         trace_mark(ext4_da_write_end,
2785                    "dev %s ino %lu pos %llu len %u copied %u",
2786                    inode->i_sb->s_id, inode->i_ino,
2787                    (unsigned long long) pos, len, copied);
2788         start = pos & (PAGE_CACHE_SIZE - 1);
2789         end = start + copied - 1;
2790
2791         /*
2792          * generic_write_end() will run mark_inode_dirty() if i_size
2793          * changes.  So let's piggyback the i_disksize mark_inode_dirty
2794          * into that.
2795          */
2796
2797         new_i_size = pos + copied;
2798         if (new_i_size > EXT4_I(inode)->i_disksize) {
2799                 if (ext4_da_should_update_i_disksize(page, end)) {
2800                         down_write(&EXT4_I(inode)->i_data_sem);
2801                         if (new_i_size > EXT4_I(inode)->i_disksize) {
2802                                 /*
2803                                  * Updating i_disksize when extending file
2804                                  * without needing block allocation
2805                                  */
2806                                 if (ext4_should_order_data(inode))
2807                                         ret = ext4_jbd2_file_inode(handle,
2808                                                                    inode);
2809
2810                                 EXT4_I(inode)->i_disksize = new_i_size;
2811                         }
2812                         up_write(&EXT4_I(inode)->i_data_sem);
2813                         /* We need to mark inode dirty even if
2814                          * new_i_size is less that inode->i_size
2815                          * bu greater than i_disksize.(hint delalloc)
2816                          */
2817                         ext4_mark_inode_dirty(handle, inode);
2818                 }
2819         }
2820         ret2 = generic_write_end(file, mapping, pos, len, copied,
2821                                                         page, fsdata);
2822         copied = ret2;
2823         if (ret2 < 0)
2824                 ret = ret2;
2825         ret2 = ext4_journal_stop(handle);
2826         if (!ret)
2827                 ret = ret2;
2828
2829         return ret ? ret : copied;
2830 }
2831
2832 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
2833 {
2834         /*
2835          * Drop reserved blocks
2836          */
2837         BUG_ON(!PageLocked(page));
2838         if (!page_has_buffers(page))
2839                 goto out;
2840
2841         ext4_da_page_release_reservation(page, offset);
2842
2843 out:
2844         ext4_invalidatepage(page, offset);
2845
2846         return;
2847 }
2848
2849
2850 /*
2851  * bmap() is special.  It gets used by applications such as lilo and by
2852  * the swapper to find the on-disk block of a specific piece of data.
2853  *
2854  * Naturally, this is dangerous if the block concerned is still in the
2855  * journal.  If somebody makes a swapfile on an ext4 data-journaling
2856  * filesystem and enables swap, then they may get a nasty shock when the
2857  * data getting swapped to that swapfile suddenly gets overwritten by
2858  * the original zero's written out previously to the journal and
2859  * awaiting writeback in the kernel's buffer cache.
2860  *
2861  * So, if we see any bmap calls here on a modified, data-journaled file,
2862  * take extra steps to flush any blocks which might be in the cache.
2863  */
2864 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
2865 {
2866         struct inode *inode = mapping->host;
2867         journal_t *journal;
2868         int err;
2869
2870         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
2871                         test_opt(inode->i_sb, DELALLOC)) {
2872                 /*
2873                  * With delalloc we want to sync the file
2874                  * so that we can make sure we allocate
2875                  * blocks for file
2876                  */
2877                 filemap_write_and_wait(mapping);
2878         }
2879
2880         if (EXT4_JOURNAL(inode) && EXT4_I(inode)->i_state & EXT4_STATE_JDATA) {
2881                 /*
2882                  * This is a REALLY heavyweight approach, but the use of
2883                  * bmap on dirty files is expected to be extremely rare:
2884                  * only if we run lilo or swapon on a freshly made file
2885                  * do we expect this to happen.
2886                  *
2887                  * (bmap requires CAP_SYS_RAWIO so this does not
2888                  * represent an unprivileged user DOS attack --- we'd be
2889                  * in trouble if mortal users could trigger this path at
2890                  * will.)
2891                  *
2892                  * NB. EXT4_STATE_JDATA is not set on files other than
2893                  * regular files.  If somebody wants to bmap a directory
2894                  * or symlink and gets confused because the buffer
2895                  * hasn't yet been flushed to disk, they deserve
2896                  * everything they get.
2897                  */
2898
2899                 EXT4_I(inode)->i_state &= ~EXT4_STATE_JDATA;
2900                 journal = EXT4_JOURNAL(inode);
2901                 jbd2_journal_lock_updates(journal);
2902                 err = jbd2_journal_flush(journal);
2903                 jbd2_journal_unlock_updates(journal);
2904
2905                 if (err)
2906                         return 0;
2907         }
2908
2909         return generic_block_bmap(mapping, block, ext4_get_block);
2910 }
2911
2912 static int bget_one(handle_t *handle, struct buffer_head *bh)
2913 {
2914         get_bh(bh);
2915         return 0;
2916 }
2917
2918 static int bput_one(handle_t *handle, struct buffer_head *bh)
2919 {
2920         put_bh(bh);
2921         return 0;
2922 }
2923
2924 /*
2925  * Note that we don't need to start a transaction unless we're journaling data
2926  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2927  * need to file the inode to the transaction's list in ordered mode because if
2928  * we are writing back data added by write(), the inode is already there and if
2929  * we are writing back data modified via mmap(), noone guarantees in which
2930  * transaction the data will hit the disk. In case we are journaling data, we
2931  * cannot start transaction directly because transaction start ranks above page
2932  * lock so we have to do some magic.
2933  *
2934  * In all journaling modes block_write_full_page() will start the I/O.
2935  *
2936  * Problem:
2937  *
2938  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2939  *              ext4_writepage()
2940  *
2941  * Similar for:
2942  *
2943  *      ext4_file_write() -> generic_file_write() -> __alloc_pages() -> ...
2944  *
2945  * Same applies to ext4_get_block().  We will deadlock on various things like
2946  * lock_journal and i_data_sem
2947  *
2948  * Setting PF_MEMALLOC here doesn't work - too many internal memory
2949  * allocations fail.
2950  *
2951  * 16May01: If we're reentered then journal_current_handle() will be
2952  *          non-zero. We simply *return*.
2953  *
2954  * 1 July 2001: @@@ FIXME:
2955  *   In journalled data mode, a data buffer may be metadata against the
2956  *   current transaction.  But the same file is part of a shared mapping
2957  *   and someone does a writepage() on it.
2958  *
2959  *   We will move the buffer onto the async_data list, but *after* it has
2960  *   been dirtied. So there's a small window where we have dirty data on
2961  *   BJ_Metadata.
2962  *
2963  *   Note that this only applies to the last partial page in the file.  The
2964  *   bit which block_write_full_page() uses prepare/commit for.  (That's
2965  *   broken code anyway: it's wrong for msync()).
2966  *
2967  *   It's a rare case: affects the final partial page, for journalled data
2968  *   where the file is subject to bith write() and writepage() in the same
2969  *   transction.  To fix it we'll need a custom block_write_full_page().
2970  *   We'll probably need that anyway for journalling writepage() output.
2971  *
2972  * We don't honour synchronous mounts for writepage().  That would be
2973  * disastrous.  Any write() or metadata operation will sync the fs for
2974  * us.
2975  *
2976  */
2977 static int __ext4_normal_writepage(struct page *page,
2978                                 struct writeback_control *wbc)
2979 {
2980         struct inode *inode = page->mapping->host;
2981
2982         if (test_opt(inode->i_sb, NOBH))
2983                 return nobh_writepage(page,
2984                                         ext4_normal_get_block_write, wbc);
2985         else
2986                 return block_write_full_page(page,
2987                                                 ext4_normal_get_block_write,
2988                                                 wbc);
2989 }
2990
2991 static int ext4_normal_writepage(struct page *page,
2992                                 struct writeback_control *wbc)
2993 {
2994         struct inode *inode = page->mapping->host;
2995         loff_t size = i_size_read(inode);
2996         loff_t len;
2997
2998         trace_mark(ext4_normal_writepage,
2999                    "dev %s ino %lu page_index %lu",
3000                    inode->i_sb->s_id, inode->i_ino, page->index);
3001         J_ASSERT(PageLocked(page));
3002         if (page->index == size >> PAGE_CACHE_SHIFT)
3003                 len = size & ~PAGE_CACHE_MASK;
3004         else
3005                 len = PAGE_CACHE_SIZE;
3006
3007         if (page_has_buffers(page)) {
3008                 /* if page has buffers it should all be mapped
3009                  * and allocated. If there are not buffers attached
3010                  * to the page we know the page is dirty but it lost
3011                  * buffers. That means that at some moment in time
3012                  * after write_begin() / write_end() has been called
3013                  * all buffers have been clean and thus they must have been
3014                  * written at least once. So they are all mapped and we can
3015                  * happily proceed with mapping them and writing the page.
3016                  */
3017                 BUG_ON(walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
3018                                         ext4_bh_unmapped_or_delay));
3019         }
3020
3021         if (!ext4_journal_current_handle())
3022                 return __ext4_normal_writepage(page, wbc);
3023
3024         redirty_page_for_writepage(wbc, page);
3025         unlock_page(page);
3026         return 0;
3027 }
3028
3029 static int __ext4_journalled_writepage(struct page *page,
3030                                 struct writeback_control *wbc)
3031 {
3032         struct address_space *mapping = page->mapping;
3033         struct inode *inode = mapping->host;
3034         struct buffer_head *page_bufs;
3035         handle_t *handle = NULL;
3036         int ret = 0;
3037         int err;
3038
3039         ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
3040                                         ext4_normal_get_block_write);
3041         if (ret != 0)
3042                 goto out_unlock;
3043
3044         page_bufs = page_buffers(page);
3045         walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE, NULL,
3046                                                                 bget_one);
3047         /* As soon as we unlock the page, it can go away, but we have
3048          * references to buffers so we are safe */
3049         unlock_page(page);
3050
3051         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
3052         if (IS_ERR(handle)) {
3053                 ret = PTR_ERR(handle);
3054                 goto out;
3055         }
3056
3057         ret = walk_page_buffers(handle, page_bufs, 0,
3058                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
3059
3060         err = walk_page_buffers(handle, page_bufs, 0,
3061                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
3062         if (ret == 0)
3063                 ret = err;
3064         err = ext4_journal_stop(handle);
3065         if (!ret)
3066                 ret = err;
3067
3068         walk_page_buffers(handle, page_bufs, 0,
3069                                 PAGE_CACHE_SIZE, NULL, bput_one);
3070         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
3071         goto out;
3072
3073 out_unlock:
3074         unlock_page(page);
3075 out:
3076         return ret;
3077 }
3078
3079 static int ext4_journalled_writepage(struct page *page,
3080                                 struct writeback_control *wbc)
3081 {
3082         struct inode *inode = page->mapping->host;
3083         loff_t size = i_size_read(inode);
3084         loff_t len;
3085
3086         trace_mark(ext4_journalled_writepage,
3087                    "dev %s ino %lu page_index %lu",
3088                    inode->i_sb->s_id, inode->i_ino, page->index);
3089         J_ASSERT(PageLocked(page));
3090         if (page->index == size >> PAGE_CACHE_SHIFT)
3091                 len = size & ~PAGE_CACHE_MASK;
3092         else
3093                 len = PAGE_CACHE_SIZE;
3094
3095         if (page_has_buffers(page)) {
3096                 /* if page has buffers it should all be mapped
3097                  * and allocated. If there are not buffers attached
3098                  * to the page we know the page is dirty but it lost
3099                  * buffers. That means that at some moment in time
3100                  * after write_begin() / write_end() has been called
3101                  * all buffers have been clean and thus they must have been
3102                  * written at least once. So they are all mapped and we can
3103                  * happily proceed with mapping them and writing the page.
3104                  */
3105                 BUG_ON(walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
3106                                         ext4_bh_unmapped_or_delay));
3107         }
3108
3109         if (ext4_journal_current_handle())
3110                 goto no_write;
3111
3112         if (PageChecked(page)) {
3113                 /*
3114                  * It's mmapped pagecache.  Add buffers and journal it.  There
3115                  * doesn't seem much point in redirtying the page here.
3116                  */
3117                 ClearPageChecked(page);
3118                 return __ext4_journalled_writepage(page, wbc);
3119         } else {
3120                 /*
3121                  * It may be a page full of checkpoint-mode buffers.  We don't
3122                  * really know unless we go poke around in the buffer_heads.
3123                  * But block_write_full_page will do the right thing.
3124                  */
3125                 return block_write_full_page(page,
3126                                                 ext4_normal_get_block_write,
3127                                                 wbc);
3128         }
3129 no_write:
3130         redirty_page_for_writepage(wbc, page);
3131         unlock_page(page);
3132         return 0;
3133 }
3134
3135 static int ext4_readpage(struct file *file, struct page *page)
3136 {
3137         return mpage_readpage(page, ext4_get_block);
3138 }
3139
3140 static int
3141 ext4_readpages(struct file *file, struct address_space *mapping,
3142                 struct list_head *pages, unsigned nr_pages)
3143 {
3144         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3145 }
3146
3147 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3148 {
3149         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3150
3151         /*
3152          * If it's a full truncate we just forget about the pending dirtying
3153          */
3154         if (offset == 0)
3155                 ClearPageChecked(page);
3156
3157         if (journal)
3158                 jbd2_journal_invalidatepage(journal, page, offset);
3159         else
3160                 block_invalidatepage(page, offset);
3161 }
3162
3163 static int ext4_releasepage(struct page *page, gfp_t wait)
3164 {
3165         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3166
3167         WARN_ON(PageChecked(page));
3168         if (!page_has_buffers(page))
3169                 return 0;
3170         if (journal)
3171                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3172         else
3173                 return try_to_free_buffers(page);
3174 }
3175
3176 /*
3177  * If the O_DIRECT write will extend the file then add this inode to the
3178  * orphan list.  So recovery will truncate it back to the original size
3179  * if the machine crashes during the write.
3180  *
3181  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3182  * crashes then stale disk data _may_ be exposed inside the file. But current
3183  * VFS code falls back into buffered path in that case so we are safe.
3184  */
3185 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3186                         const struct iovec *iov, loff_t offset,
3187                         unsigned long nr_segs)
3188 {
3189         struct file *file = iocb->ki_filp;
3190         struct inode *inode = file->f_mapping->host;
3191         struct ext4_inode_info *ei = EXT4_I(inode);
3192         handle_t *handle;
3193         ssize_t ret;
3194         int orphan = 0;
3195         size_t count = iov_length(iov, nr_segs);
3196
3197         if (rw == WRITE) {
3198                 loff_t final_size = offset + count;
3199
3200                 if (final_size > inode->i_size) {
3201                         /* Credits for sb + inode write */
3202                         handle = ext4_journal_start(inode, 2);
3203                         if (IS_ERR(handle)) {
3204                                 ret = PTR_ERR(handle);
3205                                 goto out;
3206                         }
3207                         ret = ext4_orphan_add(handle, inode);
3208                         if (ret) {
3209                                 ext4_journal_stop(handle);
3210                                 goto out;
3211                         }
3212                         orphan = 1;
3213                         ei->i_disksize = inode->i_size;
3214                         ext4_journal_stop(handle);
3215                 }
3216         }
3217
3218         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
3219                                  offset, nr_segs,
3220                                  ext4_get_block, NULL);
3221
3222         if (orphan) {
3223                 int err;
3224
3225                 /* Credits for sb + inode write */
3226                 handle = ext4_journal_start(inode, 2);
3227                 if (IS_ERR(handle)) {
3228                         /* This is really bad luck. We've written the data
3229                          * but cannot extend i_size. Bail out and pretend
3230                          * the write failed... */
3231                         ret = PTR_ERR(handle);
3232                         goto out;
3233                 }
3234                 if (inode->i_nlink)
3235                         ext4_orphan_del(handle, inode);
3236                 if (ret > 0) {
3237                         loff_t end = offset + ret;
3238                         if (end > inode->i_size) {
3239                                 ei->i_disksize = end;
3240                                 i_size_write(inode, end);
3241                                 /*
3242                                  * We're going to return a positive `ret'
3243                                  * here due to non-zero-length I/O, so there's
3244                                  * no way of reporting error returns from
3245                                  * ext4_mark_inode_dirty() to userspace.  So
3246                                  * ignore it.
3247                                  */
3248                                 ext4_mark_inode_dirty(handle, inode);
3249                         }
3250                 }
3251                 err = ext4_journal_stop(handle);
3252                 if (ret == 0)
3253                         ret = err;
3254         }
3255 out:
3256         return ret;
3257 }
3258
3259 /*
3260  * Pages can be marked dirty completely asynchronously from ext4's journalling
3261  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3262  * much here because ->set_page_dirty is called under VFS locks.  The page is
3263  * not necessarily locked.
3264  *
3265  * We cannot just dirty the page and leave attached buffers clean, because the
3266  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3267  * or jbddirty because all the journalling code will explode.
3268  *
3269  * So what we do is to mark the page "pending dirty" and next time writepage
3270  * is called, propagate that into the buffers appropriately.
3271  */
3272 static int ext4_journalled_set_page_dirty(struct page *page)
3273 {
3274         SetPageChecked(page);
3275         return __set_page_dirty_nobuffers(page);
3276 }
3277
3278 static const struct address_space_operations ext4_ordered_aops = {
3279         .readpage               = ext4_readpage,
3280         .readpages              = ext4_readpages,
3281         .writepage              = ext4_normal_writepage,
3282         .sync_page              = block_sync_page,
3283         .write_begin            = ext4_write_begin,
3284         .write_end              = ext4_ordered_write_end,
3285         .bmap                   = ext4_bmap,
3286         .invalidatepage         = ext4_invalidatepage,
3287         .releasepage            = ext4_releasepage,
3288         .direct_IO              = ext4_direct_IO,
3289         .migratepage            = buffer_migrate_page,
3290         .is_partially_uptodate  = block_is_partially_uptodate,
3291 };
3292
3293 static const struct address_space_operations ext4_writeback_aops = {
3294         .readpage               = ext4_readpage,
3295         .readpages              = ext4_readpages,
3296         .writepage              = ext4_normal_writepage,
3297         .sync_page              = block_sync_page,
3298         .write_begin            = ext4_write_begin,
3299         .write_end              = ext4_writeback_write_end,
3300         .bmap                   = ext4_bmap,
3301         .invalidatepage         = ext4_invalidatepage,
3302         .releasepage            = ext4_releasepage,
3303         .direct_IO              = ext4_direct_IO,
3304         .migratepage            = buffer_migrate_page,
3305         .is_partially_uptodate  = block_is_partially_uptodate,
3306 };
3307
3308 static const struct address_space_operations ext4_journalled_aops = {
3309         .readpage               = ext4_readpage,
3310         .readpages              = ext4_readpages,
3311         .writepage              = ext4_journalled_writepage,
3312         .sync_page              = block_sync_page,
3313         .write_begin            = ext4_write_begin,
3314         .write_end              = ext4_journalled_write_end,
3315         .set_page_dirty         = ext4_journalled_set_page_dirty,
3316         .bmap                   = ext4_bmap,
3317         .invalidatepage         = ext4_invalidatepage,
3318         .releasepage            = ext4_releasepage,
3319         .is_partially_uptodate  = block_is_partially_uptodate,
3320 };
3321
3322 static const struct address_space_operations ext4_da_aops = {
3323         .readpage               = ext4_readpage,
3324         .readpages              = ext4_readpages,
3325         .writepage              = ext4_da_writepage,
3326         .writepages             = ext4_da_writepages,
3327         .sync_page              = block_sync_page,
3328         .write_begin            = ext4_da_write_begin,
3329         .write_end              = ext4_da_write_end,
3330         .bmap                   = ext4_bmap,
3331         .invalidatepage         = ext4_da_invalidatepage,
3332         .releasepage            = ext4_releasepage,
3333         .direct_IO              = ext4_direct_IO,
3334         .migratepage            = buffer_migrate_page,
3335         .is_partially_uptodate  = block_is_partially_uptodate,
3336 };
3337
3338 void ext4_set_aops(struct inode *inode)
3339 {
3340         if (ext4_should_order_data(inode) &&
3341                 test_opt(inode->i_sb, DELALLOC))
3342                 inode->i_mapping->a_ops = &ext4_da_aops;
3343         else if (ext4_should_order_data(inode))
3344                 inode->i_mapping->a_ops = &ext4_ordered_aops;
3345         else if (ext4_should_writeback_data(inode) &&
3346                  test_opt(inode->i_sb, DELALLOC))
3347                 inode->i_mapping->a_ops = &ext4_da_aops;
3348         else if (ext4_should_writeback_data(inode))
3349                 inode->i_mapping->a_ops = &ext4_writeback_aops;
3350         else
3351                 inode->i_mapping->a_ops = &ext4_journalled_aops;
3352 }
3353
3354 /*
3355  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3356  * up to the end of the block which corresponds to `from'.
3357  * This required during truncate. We need to physically zero the tail end
3358  * of that block so it doesn't yield old data if the file is later grown.
3359  */
3360 int ext4_block_truncate_page(handle_t *handle,
3361                 struct address_space *mapping, loff_t from)
3362 {
3363         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3364         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3365         unsigned blocksize, length, pos;
3366         ext4_lblk_t iblock;
3367         struct inode *inode = mapping->host;
3368         struct buffer_head *bh;
3369         struct page *page;
3370         int err = 0;
3371
3372         page = grab_cache_page(mapping, from >> PAGE_CACHE_SHIFT);
3373         if (!page)
3374                 return -EINVAL;
3375
3376         blocksize = inode->i_sb->s_blocksize;
3377         length = blocksize - (offset & (blocksize - 1));
3378         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3379
3380         /*
3381          * For "nobh" option,  we can only work if we don't need to
3382          * read-in the page - otherwise we create buffers to do the IO.
3383          */
3384         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
3385              ext4_should_writeback_data(inode) && PageUptodate(page)) {
3386                 zero_user(page, offset, length);
3387                 set_page_dirty(page);
3388                 goto unlock;
3389         }
3390
3391         if (!page_has_buffers(page))
3392                 create_empty_buffers(page, blocksize, 0);
3393
3394         /* Find the buffer that contains "offset" */
3395         bh = page_buffers(page);
3396         pos = blocksize;
3397         while (offset >= pos) {
3398                 bh = bh->b_this_page;
3399                 iblock++;
3400                 pos += blocksize;
3401         }
3402
3403         err = 0;
3404         if (buffer_freed(bh)) {
3405                 BUFFER_TRACE(bh, "freed: skip");
3406                 goto unlock;
3407         }
3408
3409         if (!buffer_mapped(bh)) {
3410                 BUFFER_TRACE(bh, "unmapped");
3411                 ext4_get_block(inode, iblock, bh, 0);
3412                 /* unmapped? It's a hole - nothing to do */
3413                 if (!buffer_mapped(bh)) {
3414                         BUFFER_TRACE(bh, "still unmapped");
3415                         goto unlock;
3416                 }
3417         }
3418
3419         /* Ok, it's mapped. Make sure it's up-to-date */
3420         if (PageUptodate(page))
3421                 set_buffer_uptodate(bh);
3422
3423         if (!buffer_uptodate(bh)) {
3424                 err = -EIO;
3425                 ll_rw_block(READ, 1, &bh);
3426                 wait_on_buffer(bh);
3427                 /* Uhhuh. Read error. Complain and punt. */
3428                 if (!buffer_uptodate(bh))
3429                         goto unlock;
3430         }
3431
3432         if (ext4_should_journal_data(inode)) {
3433                 BUFFER_TRACE(bh, "get write access");
3434                 err = ext4_journal_get_write_access(handle, bh);
3435                 if (err)
3436                         goto unlock;
3437         }
3438
3439         zero_user(page, offset, length);
3440
3441         BUFFER_TRACE(bh, "zeroed end of block");
3442
3443         err = 0;
3444         if (ext4_should_journal_data(inode)) {
3445                 err = ext4_handle_dirty_metadata(handle, inode, bh);
3446         } else {
3447                 if (ext4_should_order_data(inode))
3448                         err = ext4_jbd2_file_inode(handle, inode);
3449                 mark_buffer_dirty(bh);
3450         }
3451
3452 unlock:
3453         unlock_page(page);
3454         page_cache_release(page);
3455         return err;
3456 }
3457
3458 /*
3459  * Probably it should be a library function... search for first non-zero word
3460  * or memcmp with zero_page, whatever is better for particular architecture.
3461  * Linus?
3462  */
3463 static inline int all_zeroes(__le32 *p, __le32 *q)
3464 {
3465         while (p < q)
3466                 if (*p++)
3467                         return 0;
3468         return 1;
3469 }
3470
3471 /**
3472  *      ext4_find_shared - find the indirect blocks for partial truncation.
3473  *      @inode:   inode in question
3474  *      @depth:   depth of the affected branch
3475  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
3476  *      @chain:   place to store the pointers to partial indirect blocks
3477  *      @top:     place to the (detached) top of branch
3478  *
3479  *      This is a helper function used by ext4_truncate().
3480  *
3481  *      When we do truncate() we may have to clean the ends of several
3482  *      indirect blocks but leave the blocks themselves alive. Block is
3483  *      partially truncated if some data below the new i_size is refered
3484  *      from it (and it is on the path to the first completely truncated
3485  *      data block, indeed).  We have to free the top of that path along
3486  *      with everything to the right of the path. Since no allocation
3487  *      past the truncation point is possible until ext4_truncate()
3488  *      finishes, we may safely do the latter, but top of branch may
3489  *      require special attention - pageout below the truncation point
3490  *      might try to populate it.
3491  *
3492  *      We atomically detach the top of branch from the tree, store the
3493  *      block number of its root in *@top, pointers to buffer_heads of
3494  *      partially truncated blocks - in @chain[].bh and pointers to
3495  *      their last elements that should not be removed - in
3496  *      @chain[].p. Return value is the pointer to last filled element
3497  *      of @chain.
3498  *
3499  *      The work left to caller to do the actual freeing of subtrees:
3500  *              a) free the subtree starting from *@top
3501  *              b) free the subtrees whose roots are stored in
3502  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
3503  *              c) free the subtrees growing from the inode past the @chain[0].
3504  *                      (no partially truncated stuff there).  */
3505
3506 static Indirect *ext4_find_shared(struct inode *inode, int depth,
3507                         ext4_lblk_t offsets[4], Indirect chain[4], __le32 *top)
3508 {
3509         Indirect *partial, *p;
3510         int k, err;
3511
3512         *top = 0;
3513         /* Make k index the deepest non-null offest + 1 */
3514         for (k = depth; k > 1 && !offsets[k-1]; k--)
3515                 ;
3516         partial = ext4_get_branch(inode, k, offsets, chain, &err);
3517         /* Writer: pointers */
3518         if (!partial)
3519                 partial = chain + k-1;
3520         /*
3521          * If the branch acquired continuation since we've looked at it -
3522          * fine, it should all survive and (new) top doesn't belong to us.
3523          */
3524         if (!partial->key && *partial->p)
3525                 /* Writer: end */
3526                 goto no_top;
3527         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
3528                 ;
3529         /*
3530          * OK, we've found the last block that must survive. The rest of our
3531          * branch should be detached before unlocking. However, if that rest
3532          * of branch is all ours and does not grow immediately from the inode
3533          * it's easier to cheat and just decrement partial->p.
3534          */
3535         if (p == chain + k - 1 && p > chain) {
3536                 p->p--;
3537         } else {
3538                 *top = *p->p;
3539                 /* Nope, don't do this in ext4.  Must leave the tree intact */
3540 #if 0
3541                 *p->p = 0;
3542 #endif
3543         }
3544         /* Writer: end */
3545
3546         while (partial > p) {
3547                 brelse(partial->bh);
3548                 partial--;
3549         }
3550 no_top:
3551         return partial;
3552 }
3553
3554 /*
3555  * Zero a number of block pointers in either an inode or an indirect block.
3556  * If we restart the transaction we must again get write access to the
3557  * indirect block for further modification.
3558  *
3559  * We release `count' blocks on disk, but (last - first) may be greater
3560  * than `count' because there can be holes in there.
3561  */
3562 static void ext4_clear_blocks(handle_t *handle, struct inode *inode,
3563                 struct buffer_head *bh, ext4_fsblk_t block_to_free,
3564                 unsigned long count, __le32 *first, __le32 *last)
3565 {
3566         __le32 *p;
3567         if (try_to_extend_transaction(handle, inode)) {
3568                 if (bh) {
3569                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
3570                         ext4_handle_dirty_metadata(handle, inode, bh);
3571                 }
3572                 ext4_mark_inode_dirty(handle, inode);
3573                 ext4_journal_test_restart(handle, inode);
3574                 if (bh) {
3575                         BUFFER_TRACE(bh, "retaking write access");
3576                         ext4_journal_get_write_access(handle, bh);
3577                 }
3578         }
3579
3580         /*
3581          * Any buffers which are on the journal will be in memory. We find
3582          * them on the hash table so jbd2_journal_revoke() will run jbd2_journal_forget()
3583          * on them.  We've already detached each block from the file, so
3584          * bforget() in jbd2_journal_forget() should be safe.
3585          *
3586          * AKPM: turn on bforget in jbd2_journal_forget()!!!
3587          */
3588         for (p = first; p < last; p++) {
3589                 u32 nr = le32_to_cpu(*p);
3590                 if (nr) {
3591                         struct buffer_head *tbh;
3592
3593                         *p = 0;
3594                         tbh = sb_find_get_block(inode->i_sb, nr);
3595                         ext4_forget(handle, 0, inode, tbh, nr);
3596                 }
3597         }
3598
3599         ext4_free_blocks(handle, inode, block_to_free, count, 0);
3600 }
3601
3602 /**
3603  * ext4_free_data - free a list of data blocks
3604  * @handle:     handle for this transaction
3605  * @inode:      inode we are dealing with
3606  * @this_bh:    indirect buffer_head which contains *@first and *@last
3607  * @first:      array of block numbers
3608  * @last:       points immediately past the end of array
3609  *
3610  * We are freeing all blocks refered from that array (numbers are stored as
3611  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
3612  *
3613  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
3614  * blocks are contiguous then releasing them at one time will only affect one
3615  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
3616  * actually use a lot of journal space.
3617  *
3618  * @this_bh will be %NULL if @first and @last point into the inode's direct
3619  * block pointers.
3620  */
3621 static void ext4_free_data(handle_t *handle, struct inode *inode,
3622                            struct buffer_head *this_bh,
3623                            __le32 *first, __le32 *last)
3624 {
3625         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
3626         unsigned long count = 0;            /* Number of blocks in the run */
3627         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
3628                                                corresponding to
3629                                                block_to_free */
3630         ext4_fsblk_t nr;                    /* Current block # */
3631         __le32 *p;                          /* Pointer into inode/ind
3632                                                for current block */
3633         int err;
3634
3635         if (this_bh) {                          /* For indirect block */
3636                 BUFFER_TRACE(this_bh, "get_write_access");
3637                 err = ext4_journal_get_write_access(handle, this_bh);
3638                 /* Important: if we can't update the indirect pointers
3639                  * to the blocks, we can't free them. */
3640                 if (err)
3641                         return;
3642         }
3643
3644         for (p = first; p < last; p++) {
3645                 nr = le32_to_cpu(*p);
3646                 if (nr) {
3647                         /* accumulate blocks to free if they're contiguous */
3648                         if (count == 0) {
3649                                 block_to_free = nr;
3650                                 block_to_free_p = p;
3651                                 count = 1;
3652                         } else if (nr == block_to_free + count) {
3653                                 count++;
3654                         } else {
3655                                 ext4_clear_blocks(handle, inode, this_bh,
3656                                                   block_to_free,
3657                                                   count, block_to_free_p, p);
3658                                 block_to_free = nr;
3659                                 block_to_free_p = p;
3660                                 count = 1;
3661                         }
3662                 }
3663         }
3664
3665         if (count > 0)
3666                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
3667                                   count, block_to_free_p, p);
3668
3669         if (this_bh) {
3670                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
3671
3672                 /*
3673                  * The buffer head should have an attached journal head at this
3674                  * point. However, if the data is corrupted and an indirect
3675                  * block pointed to itself, it would have been detached when
3676                  * the block was cleared. Check for this instead of OOPSing.
3677                  */
3678                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
3679                         ext4_handle_dirty_metadata(handle, inode, this_bh);
3680                 else
3681                         ext4_error(inode->i_sb, __func__,
3682                                    "circular indirect block detected, "
3683                                    "inode=%lu, block=%llu",
3684                                    inode->i_ino,
3685                                    (unsigned long long) this_bh->b_blocknr);
3686         }
3687 }
3688
3689 /**
3690  *      ext4_free_branches - free an array of branches
3691  *      @handle: JBD handle for this transaction
3692  *      @inode: inode we are dealing with
3693  *      @parent_bh: the buffer_head which contains *@first and *@last
3694  *      @first: array of block numbers
3695  *      @last:  pointer immediately past the end of array
3696  *      @depth: depth of the branches to free
3697  *
3698  *      We are freeing all blocks refered from these branches (numbers are
3699  *      stored as little-endian 32-bit) and updating @inode->i_blocks
3700  *      appropriately.
3701  */
3702 static void ext4_free_branches(handle_t *handle, struct inode *inode,
3703                                struct buffer_head *parent_bh,
3704                                __le32 *first, __le32 *last, int depth)
3705 {
3706         ext4_fsblk_t nr;
3707         __le32 *p;
3708
3709         if (ext4_handle_is_aborted(handle))
3710                 return;
3711
3712         if (depth--) {
3713                 struct buffer_head *bh;
3714                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
3715                 p = last;
3716                 while (--p >= first) {
3717                         nr = le32_to_cpu(*p);
3718                         if (!nr)
3719                                 continue;               /* A hole */
3720
3721                         /* Go read the buffer for the next level down */
3722                         bh = sb_bread(inode->i_sb, nr);
3723
3724                         /*
3725                          * A read failure? Report error and clear slot
3726                          * (should be rare).
3727                          */
3728                         if (!bh) {
3729                                 ext4_error(inode->i_sb, "ext4_free_branches",
3730                                            "Read failure, inode=%lu, block=%llu",
3731                                            inode->i_ino, nr);
3732                                 continue;
3733                         }
3734
3735                         /* This zaps the entire block.  Bottom up. */
3736                         BUFFER_TRACE(bh, "free child branches");
3737                         ext4_free_branches(handle, inode, bh,
3738                                         (__le32 *) bh->b_data,
3739                                         (__le32 *) bh->b_data + addr_per_block,
3740                                         depth);
3741
3742                         /*
3743                          * We've probably journalled the indirect block several
3744                          * times during the truncate.  But it's no longer
3745                          * needed and we now drop it from the transaction via
3746                          * jbd2_journal_revoke().
3747                          *
3748                          * That's easy if it's exclusively part of this
3749                          * transaction.  But if it's part of the committing
3750                          * transaction then jbd2_journal_forget() will simply
3751                          * brelse() it.  That means that if the underlying
3752                          * block is reallocated in ext4_get_block(),
3753                          * unmap_underlying_metadata() will find this block
3754                          * and will try to get rid of it.  damn, damn.
3755                          *
3756                          * If this block has already been committed to the
3757                          * journal, a revoke record will be written.  And
3758                          * revoke records must be emitted *before* clearing
3759                          * this block's bit in the bitmaps.
3760                          */
3761                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
3762
3763                         /*
3764                          * Everything below this this pointer has been
3765                          * released.  Now let this top-of-subtree go.
3766                          *
3767                          * We want the freeing of this indirect block to be
3768                          * atomic in the journal with the updating of the
3769                          * bitmap block which owns it.  So make some room in
3770                          * the journal.
3771                          *
3772                          * We zero the parent pointer *after* freeing its
3773                          * pointee in the bitmaps, so if extend_transaction()
3774                          * for some reason fails to put the bitmap changes and
3775                          * the release into the same transaction, recovery
3776                          * will merely complain about releasing a free block,
3777                          * rather than leaking blocks.
3778                          */
3779                         if (ext4_handle_is_aborted(handle))
3780                                 return;
3781                         if (try_to_extend_transaction(handle, inode)) {
3782                                 ext4_mark_inode_dirty(handle, inode);
3783                                 ext4_journal_test_restart(handle, inode);
3784                         }
3785
3786                         ext4_free_blocks(handle, inode, nr, 1, 1);
3787
3788                         if (parent_bh) {
3789                                 /*
3790                                  * The block which we have just freed is
3791                                  * pointed to by an indirect block: journal it
3792                                  */
3793                                 BUFFER_TRACE(parent_bh, "get_write_access");
3794                                 if (!ext4_journal_get_write_access(handle,
3795                                                                    parent_bh)){
3796                                         *p = 0;
3797                                         BUFFER_TRACE(parent_bh,
3798                                         "call ext4_handle_dirty_metadata");
3799                                         ext4_handle_dirty_metadata(handle,
3800                                                                    inode,
3801                                                                    parent_bh);
3802                                 }
3803                         }
3804                 }
3805         } else {
3806                 /* We have reached the bottom of the tree. */
3807                 BUFFER_TRACE(parent_bh, "free data blocks");
3808                 ext4_free_data(handle, inode, parent_bh, first, last);
3809         }
3810 }
3811
3812 int ext4_can_truncate(struct inode *inode)
3813 {
3814         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
3815                 return 0;
3816         if (S_ISREG(inode->i_mode))
3817                 return 1;
3818         if (S_ISDIR(inode->i_mode))
3819                 return 1;
3820         if (S_ISLNK(inode->i_mode))
3821                 return !ext4_inode_is_fast_symlink(inode);
3822         return 0;
3823 }
3824
3825 /*
3826  * ext4_truncate()
3827  *
3828  * We block out ext4_get_block() block instantiations across the entire
3829  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
3830  * simultaneously on behalf of the same inode.
3831  *
3832  * As we work through the truncate and commmit bits of it to the journal there
3833  * is one core, guiding principle: the file's tree must always be consistent on
3834  * disk.  We must be able to restart the truncate after a crash.
3835  *
3836  * The file's tree may be transiently inconsistent in memory (although it
3837  * probably isn't), but whenever we close off and commit a journal transaction,
3838  * the contents of (the filesystem + the journal) must be consistent and
3839  * restartable.  It's pretty simple, really: bottom up, right to left (although
3840  * left-to-right works OK too).
3841  *
3842  * Note that at recovery time, journal replay occurs *before* the restart of
3843  * truncate against the orphan inode list.
3844  *
3845  * The committed inode has the new, desired i_size (which is the same as
3846  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
3847  * that this inode's truncate did not complete and it will again call
3848  * ext4_truncate() to have another go.  So there will be instantiated blocks
3849  * to the right of the truncation point in a crashed ext4 filesystem.  But
3850  * that's fine - as long as they are linked from the inode, the post-crash
3851  * ext4_truncate() run will find them and release them.
3852  */
3853 void ext4_truncate(struct inode *inode)
3854 {
3855         handle_t *handle;
3856         struct ext4_inode_info *ei = EXT4_I(inode);
3857         __le32 *i_data = ei->i_data;
3858         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
3859         struct address_space *mapping = inode->i_mapping;
3860         ext4_lblk_t offsets[4];
3861         Indirect chain[4];
3862         Indirect *partial;
3863         __le32 nr = 0;
3864         int n;
3865         ext4_lblk_t last_block;
3866         unsigned blocksize = inode->i_sb->s_blocksize;
3867
3868         if (!ext4_can_truncate(inode))
3869                 return;
3870
3871         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
3872                 ext4_ext_truncate(inode);
3873                 return;
3874         }
3875
3876         handle = start_transaction(inode);
3877         if (IS_ERR(handle))
3878                 return;         /* AKPM: return what? */
3879
3880         last_block = (inode->i_size + blocksize-1)
3881                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
3882
3883         if (inode->i_size & (blocksize - 1))
3884                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
3885                         goto out_stop;
3886
3887         n = ext4_block_to_path(inode, last_block, offsets, NULL);
3888         if (n == 0)
3889                 goto out_stop;  /* error */
3890
3891         /*
3892          * OK.  This truncate is going to happen.  We add the inode to the
3893          * orphan list, so that if this truncate spans multiple transactions,
3894          * and we crash, we will resume the truncate when the filesystem
3895          * recovers.  It also marks the inode dirty, to catch the new size.
3896          *
3897          * Implication: the file must always be in a sane, consistent
3898          * truncatable state while each transaction commits.
3899          */
3900         if (ext4_orphan_add(handle, inode))
3901                 goto out_stop;
3902
3903         /*
3904          * From here we block out all ext4_get_block() callers who want to
3905          * modify the block allocation tree.
3906          */
3907         down_write(&ei->i_data_sem);
3908
3909         ext4_discard_preallocations(inode);
3910
3911         /*
3912          * The orphan list entry will now protect us from any crash which
3913          * occurs before the truncate completes, so it is now safe to propagate
3914          * the new, shorter inode size (held for now in i_size) into the
3915          * on-disk inode. We do this via i_disksize, which is the value which
3916          * ext4 *really* writes onto the disk inode.
3917          */
3918         ei->i_disksize = inode->i_size;
3919
3920         if (n == 1) {           /* direct blocks */
3921                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
3922                                i_data + EXT4_NDIR_BLOCKS);
3923                 goto do_indirects;
3924         }
3925
3926         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
3927         /* Kill the top of shared branch (not detached) */
3928         if (nr) {
3929                 if (partial == chain) {
3930                         /* Shared branch grows from the inode */
3931                         ext4_free_branches(handle, inode, NULL,
3932                                            &nr, &nr+1, (chain+n-1) - partial);
3933                         *partial->p = 0;
3934                         /*
3935                          * We mark the inode dirty prior to restart,
3936                          * and prior to stop.  No need for it here.
3937                          */
3938                 } else {
3939                         /* Shared branch grows from an indirect block */
3940                         BUFFER_TRACE(partial->bh, "get_write_access");
3941                         ext4_free_branches(handle, inode, partial->bh,
3942                                         partial->p,
3943                                         partial->p+1, (chain+n-1) - partial);
3944                 }
3945         }
3946         /* Clear the ends of indirect blocks on the shared branch */
3947         while (partial > chain) {
3948                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
3949                                    (__le32*)partial->bh->b_data+addr_per_block,
3950                                    (chain+n-1) - partial);
3951                 BUFFER_TRACE(partial->bh, "call brelse");
3952                 brelse (partial->bh);
3953                 partial--;
3954         }
3955 do_indirects:
3956         /* Kill the remaining (whole) subtrees */
3957         switch (offsets[0]) {
3958         default:
3959                 nr = i_data[EXT4_IND_BLOCK];
3960                 if (nr) {
3961                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
3962                         i_data[EXT4_IND_BLOCK] = 0;
3963                 }
3964         case EXT4_IND_BLOCK:
3965                 nr = i_data[EXT4_DIND_BLOCK];
3966                 if (nr) {
3967                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
3968                         i_data[EXT4_DIND_BLOCK] = 0;
3969                 }
3970         case EXT4_DIND_BLOCK:
3971                 nr = i_data[EXT4_TIND_BLOCK];
3972                 if (nr) {
3973                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
3974                         i_data[EXT4_TIND_BLOCK] = 0;
3975                 }
3976         case EXT4_TIND_BLOCK:
3977                 ;
3978         }
3979
3980         up_write(&ei->i_data_sem);
3981         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
3982         ext4_mark_inode_dirty(handle, inode);
3983
3984         /*
3985          * In a multi-transaction truncate, we only make the final transaction
3986          * synchronous
3987          */
3988         if (IS_SYNC(inode))
3989                 ext4_handle_sync(handle);
3990 out_stop:
3991         /*
3992          * If this was a simple ftruncate(), and the file will remain alive
3993          * then we need to clear up the orphan record which we created above.
3994          * However, if this was a real unlink then we were called by
3995          * ext4_delete_inode(), and we allow that function to clean up the
3996          * orphan info for us.
3997          */
3998         if (inode->i_nlink)
3999                 ext4_orphan_del(handle, inode);
4000
4001         ext4_journal_stop(handle);
4002 }
4003
4004 /*
4005  * ext4_get_inode_loc returns with an extra refcount against the inode's
4006  * underlying buffer_head on success. If 'in_mem' is true, we have all
4007  * data in memory that is needed to recreate the on-disk version of this
4008  * inode.
4009  */
4010 static int __ext4_get_inode_loc(struct inode *inode,
4011                                 struct ext4_iloc *iloc, int in_mem)
4012 {
4013         struct ext4_group_desc  *gdp;
4014         struct buffer_head      *bh;
4015         struct super_block      *sb = inode->i_sb;
4016         ext4_fsblk_t            block;
4017         int                     inodes_per_block, inode_offset;
4018
4019         iloc->bh = NULL;
4020         if (!ext4_valid_inum(sb, inode->i_ino))
4021                 return -EIO;
4022
4023         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4024         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4025         if (!gdp)
4026                 return -EIO;
4027
4028         /*
4029          * Figure out the offset within the block group inode table
4030          */
4031         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4032         inode_offset = ((inode->i_ino - 1) %
4033                         EXT4_INODES_PER_GROUP(sb));
4034         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4035         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4036
4037         bh = sb_getblk(sb, block);
4038         if (!bh) {
4039                 ext4_error(sb, "ext4_get_inode_loc", "unable to read "
4040                            "inode block - inode=%lu, block=%llu",
4041                            inode->i_ino, block);
4042                 return -EIO;
4043         }
4044         if (!buffer_uptodate(bh)) {
4045                 lock_buffer(bh);
4046
4047                 /*
4048                  * If the buffer has the write error flag, we have failed
4049                  * to write out another inode in the same block.  In this
4050                  * case, we don't have to read the block because we may
4051                  * read the old inode data successfully.
4052                  */
4053                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4054                         set_buffer_uptodate(bh);
4055
4056                 if (buffer_uptodate(bh)) {
4057                         /* someone brought it uptodate while we waited */
4058                         unlock_buffer(bh);
4059                         goto has_buffer;
4060                 }
4061
4062                 /*
4063                  * If we have all information of the inode in memory and this
4064                  * is the only valid inode in the block, we need not read the
4065                  * block.
4066                  */
4067                 if (in_mem) {
4068                         struct buffer_head *bitmap_bh;
4069                         int i, start;
4070
4071                         start = inode_offset & ~(inodes_per_block - 1);
4072
4073                         /* Is the inode bitmap in cache? */
4074                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4075                         if (!bitmap_bh)
4076                                 goto make_io;
4077
4078                         /*
4079                          * If the inode bitmap isn't in cache then the
4080                          * optimisation may end up performing two reads instead
4081                          * of one, so skip it.
4082                          */
4083                         if (!buffer_uptodate(bitmap_bh)) {
4084                                 brelse(bitmap_bh);
4085                                 goto make_io;
4086                         }
4087                         for (i = start; i < start + inodes_per_block; i++) {
4088                                 if (i == inode_offset)
4089                                         continue;
4090                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4091                                         break;
4092                         }
4093                         brelse(bitmap_bh);
4094                         if (i == start + inodes_per_block) {
4095                                 /* all other inodes are free, so skip I/O */
4096                                 memset(bh->b_data, 0, bh->b_size);
4097                                 set_buffer_uptodate(bh);
4098                                 unlock_buffer(bh);
4099                                 goto has_buffer;
4100                         }
4101                 }
4102
4103 make_io:
4104                 /*
4105                  * If we need to do any I/O, try to pre-readahead extra
4106                  * blocks from the inode table.
4107                  */
4108                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4109                         ext4_fsblk_t b, end, table;
4110                         unsigned num;
4111
4112                         table = ext4_inode_table(sb, gdp);
4113                         /* Make sure s_inode_readahead_blks is a power of 2 */
4114                         while (EXT4_SB(sb)->s_inode_readahead_blks &
4115                                (EXT4_SB(sb)->s_inode_readahead_blks-1))
4116                                 EXT4_SB(sb)->s_inode_readahead_blks = 
4117                                    (EXT4_SB(sb)->s_inode_readahead_blks &
4118                                     (EXT4_SB(sb)->s_inode_readahead_blks-1));
4119                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4120                         if (table > b)
4121                                 b = table;
4122                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4123                         num = EXT4_INODES_PER_GROUP(sb);
4124                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4125                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4126                                 num -= ext4_itable_unused_count(sb, gdp);
4127                         table += num / inodes_per_block;
4128                         if (end > table)
4129                                 end = table;
4130                         while (b <= end)
4131                                 sb_breadahead(sb, b++);
4132                 }
4133
4134                 /*
4135                  * There are other valid inodes in the buffer, this inode
4136                  * has in-inode xattrs, or we don't have this inode in memory.
4137                  * Read the block from disk.
4138                  */
4139                 get_bh(bh);
4140                 bh->b_end_io = end_buffer_read_sync;
4141                 submit_bh(READ_META, bh);
4142                 wait_on_buffer(bh);
4143                 if (!buffer_uptodate(bh)) {
4144                         ext4_error(sb, __func__,
4145                                    "unable to read inode block - inode=%lu, "
4146                                    "block=%llu", inode->i_ino, block);
4147                         brelse(bh);
4148                         return -EIO;
4149                 }
4150         }
4151 has_buffer:
4152         iloc->bh = bh;
4153         return 0;
4154 }
4155
4156 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4157 {
4158         /* We have all inode data except xattrs in memory here. */
4159         return __ext4_get_inode_loc(inode, iloc,
4160                 !(EXT4_I(inode)->i_state & EXT4_STATE_XATTR));
4161 }
4162
4163 void ext4_set_inode_flags(struct inode *inode)
4164 {
4165         unsigned int flags = EXT4_I(inode)->i_flags;
4166
4167         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4168         if (flags & EXT4_SYNC_FL)
4169                 inode->i_flags |= S_SYNC;
4170         if (flags & EXT4_APPEND_FL)
4171                 inode->i_flags |= S_APPEND;
4172         if (flags & EXT4_IMMUTABLE_FL)
4173                 inode->i_flags |= S_IMMUTABLE;
4174         if (flags & EXT4_NOATIME_FL)
4175                 inode->i_flags |= S_NOATIME;
4176         if (flags & EXT4_DIRSYNC_FL)
4177                 inode->i_flags |= S_DIRSYNC;
4178 }
4179
4180 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4181 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4182 {
4183         unsigned int flags = ei->vfs_inode.i_flags;
4184
4185         ei->i_flags &= ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4186                         EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|EXT4_DIRSYNC_FL);
4187         if (flags & S_SYNC)
4188                 ei->i_flags |= EXT4_SYNC_FL;
4189         if (flags & S_APPEND)
4190                 ei->i_flags |= EXT4_APPEND_FL;
4191         if (flags & S_IMMUTABLE)
4192                 ei->i_flags |= EXT4_IMMUTABLE_FL;
4193         if (flags & S_NOATIME)
4194                 ei->i_flags |= EXT4_NOATIME_FL;
4195         if (flags & S_DIRSYNC)
4196                 ei->i_flags |= EXT4_DIRSYNC_FL;
4197 }
4198 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4199                                         struct ext4_inode_info *ei)
4200 {
4201         blkcnt_t i_blocks ;
4202         struct inode *inode = &(ei->vfs_inode);
4203         struct super_block *sb = inode->i_sb;
4204
4205         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4206                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4207                 /* we are using combined 48 bit field */
4208                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4209                                         le32_to_cpu(raw_inode->i_blocks_lo);
4210                 if (ei->i_flags & EXT4_HUGE_FILE_FL) {
4211                         /* i_blocks represent file system block size */
4212                         return i_blocks  << (inode->i_blkbits - 9);
4213                 } else {
4214                         return i_blocks;
4215                 }
4216         } else {
4217                 return le32_to_cpu(raw_inode->i_blocks_lo);
4218         }
4219 }
4220
4221 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4222 {
4223         struct ext4_iloc iloc;
4224         struct ext4_inode *raw_inode;
4225         struct ext4_inode_info *ei;
4226         struct buffer_head *bh;
4227         struct inode *inode;
4228         long ret;
4229         int block;
4230
4231         inode = iget_locked(sb, ino);
4232         if (!inode)
4233                 return ERR_PTR(-ENOMEM);
4234         if (!(inode->i_state & I_NEW))
4235                 return inode;
4236
4237         ei = EXT4_I(inode);
4238 #ifdef CONFIG_EXT4_FS_POSIX_ACL
4239         ei->i_acl = EXT4_ACL_NOT_CACHED;
4240         ei->i_default_acl = EXT4_ACL_NOT_CACHED;
4241 #endif
4242
4243         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4244         if (ret < 0)
4245                 goto bad_inode;
4246         bh = iloc.bh;
4247         raw_inode = ext4_raw_inode(&iloc);
4248         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4249         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4250         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4251         if (!(test_opt(inode->i_sb, NO_UID32))) {
4252                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4253                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4254         }
4255         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
4256
4257         ei->i_state = 0;
4258         ei->i_dir_start_lookup = 0;
4259         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4260         /* We now have enough fields to check if the inode was active or not.
4261          * This is needed because nfsd might try to access dead inodes
4262          * the test is that same one that e2fsck uses
4263          * NeilBrown 1999oct15
4264          */
4265         if (inode->i_nlink == 0) {
4266                 if (inode->i_mode == 0 ||
4267                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
4268                         /* this inode is deleted */
4269                         brelse(bh);
4270                         ret = -ESTALE;
4271                         goto bad_inode;
4272                 }
4273                 /* The only unlinked inodes we let through here have
4274                  * valid i_mode and are being read by the orphan
4275                  * recovery code: that's fine, we're about to complete
4276                  * the process of deleting those. */
4277         }
4278         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4279         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4280         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4281         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
4282             cpu_to_le32(EXT4_OS_HURD)) {
4283                 ei->i_file_acl |=
4284                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4285         }
4286         inode->i_size = ext4_isize(raw_inode);
4287         ei->i_disksize = inode->i_size;
4288         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4289         ei->i_block_group = iloc.block_group;
4290         /*
4291          * NOTE! The in-memory inode i_data array is in little-endian order
4292          * even on big-endian machines: we do NOT byteswap the block numbers!
4293          */
4294         for (block = 0; block < EXT4_N_BLOCKS; block++)
4295                 ei->i_data[block] = raw_inode->i_block[block];
4296         INIT_LIST_HEAD(&ei->i_orphan);
4297
4298         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4299                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4300                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4301                     EXT4_INODE_SIZE(inode->i_sb)) {
4302                         brelse(bh);
4303                         ret = -EIO;
4304                         goto bad_inode;
4305                 }
4306                 if (ei->i_extra_isize == 0) {
4307                         /* The extra space is currently unused. Use it. */
4308                         ei->i_extra_isize = sizeof(struct ext4_inode) -
4309                                             EXT4_GOOD_OLD_INODE_SIZE;
4310                 } else {
4311                         __le32 *magic = (void *)raw_inode +
4312                                         EXT4_GOOD_OLD_INODE_SIZE +
4313                                         ei->i_extra_isize;
4314                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
4315                                  ei->i_state |= EXT4_STATE_XATTR;
4316                 }
4317         } else
4318                 ei->i_extra_isize = 0;
4319
4320         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4321         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4322         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4323         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4324
4325         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4326         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4327                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4328                         inode->i_version |=
4329                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4330         }
4331
4332         if (S_ISREG(inode->i_mode)) {
4333                 inode->i_op = &ext4_file_inode_operations;
4334                 inode->i_fop = &ext4_file_operations;
4335                 ext4_set_aops(inode);
4336         } else if (S_ISDIR(inode->i_mode)) {
4337                 inode->i_op = &ext4_dir_inode_operations;
4338                 inode->i_fop = &ext4_dir_operations;
4339         } else if (S_ISLNK(inode->i_mode)) {
4340                 if (ext4_inode_is_fast_symlink(inode)) {
4341                         inode->i_op = &ext4_fast_symlink_inode_operations;
4342                         nd_terminate_link(ei->i_data, inode->i_size,
4343                                 sizeof(ei->i_data) - 1);
4344                 } else {
4345                         inode->i_op = &ext4_symlink_inode_operations;
4346                         ext4_set_aops(inode);
4347                 }
4348         } else {
4349                 inode->i_op = &ext4_special_inode_operations;
4350                 if (raw_inode->i_block[0])
4351                         init_special_inode(inode, inode->i_mode,
4352                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4353                 else
4354                         init_special_inode(inode, inode->i_mode,
4355                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4356         }
4357         brelse(iloc.bh);
4358         ext4_set_inode_flags(inode);
4359         unlock_new_inode(inode);
4360         return inode;
4361
4362 bad_inode:
4363         iget_failed(inode);
4364         return ERR_PTR(ret);
4365 }
4366
4367 static int ext4_inode_blocks_set(handle_t *handle,
4368                                 struct ext4_inode *raw_inode,
4369                                 struct ext4_inode_info *ei)
4370 {
4371         struct inode *inode = &(ei->vfs_inode);
4372         u64 i_blocks = inode->i_blocks;
4373         struct super_block *sb = inode->i_sb;
4374
4375         if (i_blocks <= ~0U) {
4376                 /*
4377                  * i_blocks can be represnted in a 32 bit variable
4378                  * as multiple of 512 bytes
4379                  */
4380                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4381                 raw_inode->i_blocks_high = 0;
4382                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4383                 return 0;
4384         }
4385         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
4386                 return -EFBIG;
4387
4388         if (i_blocks <= 0xffffffffffffULL) {
4389                 /*
4390                  * i_blocks can be represented in a 48 bit variable
4391                  * as multiple of 512 bytes
4392                  */
4393                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4394                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4395                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4396         } else {
4397                 ei->i_flags |= EXT4_HUGE_FILE_FL;
4398                 /* i_block is stored in file system block size */
4399                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4400                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4401                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4402         }
4403         return 0;
4404 }
4405
4406 /*
4407  * Post the struct inode info into an on-disk inode location in the
4408  * buffer-cache.  This gobbles the caller's reference to the
4409  * buffer_head in the inode location struct.
4410  *
4411  * The caller must have write access to iloc->bh.
4412  */
4413 static int ext4_do_update_inode(handle_t *handle,
4414                                 struct inode *inode,
4415                                 struct ext4_iloc *iloc)
4416 {
4417         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
4418         struct ext4_inode_info *ei = EXT4_I(inode);
4419         struct buffer_head *bh = iloc->bh;
4420         int err = 0, rc, block;
4421
4422         /* For fields not not tracking in the in-memory inode,
4423          * initialise them to zero for new inodes. */
4424         if (ei->i_state & EXT4_STATE_NEW)
4425                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
4426
4427         ext4_get_inode_flags(ei);
4428         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
4429         if (!(test_opt(inode->i_sb, NO_UID32))) {
4430                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
4431                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
4432 /*
4433  * Fix up interoperability with old kernels. Otherwise, old inodes get
4434  * re-used with the upper 16 bits of the uid/gid intact
4435  */
4436                 if (!ei->i_dtime) {
4437                         raw_inode->i_uid_high =
4438                                 cpu_to_le16(high_16_bits(inode->i_uid));
4439                         raw_inode->i_gid_high =
4440                                 cpu_to_le16(high_16_bits(inode->i_gid));
4441                 } else {
4442                         raw_inode->i_uid_high = 0;
4443                         raw_inode->i_gid_high = 0;
4444                 }
4445         } else {
4446                 raw_inode->i_uid_low =
4447                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
4448                 raw_inode->i_gid_low =
4449                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
4450                 raw_inode->i_uid_high = 0;
4451                 raw_inode->i_gid_high = 0;
4452         }
4453         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
4454
4455         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
4456         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
4457         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
4458         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
4459
4460         if (ext4_inode_blocks_set(handle, raw_inode, ei))
4461                 goto out_brelse;
4462         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
4463         /* clear the migrate flag in the raw_inode */
4464         raw_inode->i_flags = cpu_to_le32(ei->i_flags & ~EXT4_EXT_MIGRATE);
4465         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
4466             cpu_to_le32(EXT4_OS_HURD))
4467                 raw_inode->i_file_acl_high =
4468                         cpu_to_le16(ei->i_file_acl >> 32);
4469         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
4470         ext4_isize_set(raw_inode, ei->i_disksize);
4471         if (ei->i_disksize > 0x7fffffffULL) {
4472                 struct super_block *sb = inode->i_sb;
4473                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
4474                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
4475                                 EXT4_SB(sb)->s_es->s_rev_level ==
4476                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
4477                         /* If this is the first large file
4478                          * created, add a flag to the superblock.
4479                          */
4480                         err = ext4_journal_get_write_access(handle,
4481                                         EXT4_SB(sb)->s_sbh);
4482                         if (err)
4483                                 goto out_brelse;
4484                         ext4_update_dynamic_rev(sb);
4485                         EXT4_SET_RO_COMPAT_FEATURE(sb,
4486                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
4487                         sb->s_dirt = 1;
4488                         ext4_handle_sync(handle);
4489                         err = ext4_handle_dirty_metadata(handle, inode,
4490                                         EXT4_SB(sb)->s_sbh);
4491                 }
4492         }
4493         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
4494         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
4495                 if (old_valid_dev(inode->i_rdev)) {
4496                         raw_inode->i_block[0] =
4497                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
4498                         raw_inode->i_block[1] = 0;
4499                 } else {
4500                         raw_inode->i_block[0] = 0;
4501                         raw_inode->i_block[1] =
4502                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
4503                         raw_inode->i_block[2] = 0;
4504                 }
4505         } else for (block = 0; block < EXT4_N_BLOCKS; block++)
4506                 raw_inode->i_block[block] = ei->i_data[block];
4507
4508         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
4509         if (ei->i_extra_isize) {
4510                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4511                         raw_inode->i_version_hi =
4512                         cpu_to_le32(inode->i_version >> 32);
4513                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
4514         }
4515
4516         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4517         rc = ext4_handle_dirty_metadata(handle, inode, bh);
4518         if (!err)
4519                 err = rc;
4520         ei->i_state &= ~EXT4_STATE_NEW;
4521
4522 out_brelse:
4523         brelse(bh);
4524         ext4_std_error(inode->i_sb, err);
4525         return err;
4526 }
4527
4528 /*
4529  * ext4_write_inode()
4530  *
4531  * We are called from a few places:
4532  *
4533  * - Within generic_file_write() for O_SYNC files.
4534  *   Here, there will be no transaction running. We wait for any running
4535  *   trasnaction to commit.
4536  *
4537  * - Within sys_sync(), kupdate and such.
4538  *   We wait on commit, if tol to.
4539  *
4540  * - Within prune_icache() (PF_MEMALLOC == true)
4541  *   Here we simply return.  We can't afford to block kswapd on the
4542  *   journal commit.
4543  *
4544  * In all cases it is actually safe for us to return without doing anything,
4545  * because the inode has been copied into a raw inode buffer in
4546  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
4547  * knfsd.
4548  *
4549  * Note that we are absolutely dependent upon all inode dirtiers doing the
4550  * right thing: they *must* call mark_inode_dirty() after dirtying info in
4551  * which we are interested.
4552  *
4553  * It would be a bug for them to not do this.  The code:
4554  *
4555  *      mark_inode_dirty(inode)
4556  *      stuff();
4557  *      inode->i_size = expr;
4558  *
4559  * is in error because a kswapd-driven write_inode() could occur while
4560  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
4561  * will no longer be on the superblock's dirty inode list.
4562  */
4563 int ext4_write_inode(struct inode *inode, int wait)
4564 {
4565         if (current->flags & PF_MEMALLOC)
4566                 return 0;
4567
4568         if (ext4_journal_current_handle()) {
4569                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
4570                 dump_stack();
4571                 return -EIO;
4572         }
4573
4574         if (!wait)
4575                 return 0;
4576
4577         return ext4_force_commit(inode->i_sb);
4578 }
4579
4580 int __ext4_write_dirty_metadata(struct inode *inode, struct buffer_head *bh)
4581 {
4582         int err = 0;
4583
4584         mark_buffer_dirty(bh);
4585         if (inode && inode_needs_sync(inode)) {
4586                 sync_dirty_buffer(bh);
4587                 if (buffer_req(bh) && !buffer_uptodate(bh)) {
4588                         ext4_error(inode->i_sb, __func__,
4589                                    "IO error syncing inode, "
4590                                    "inode=%lu, block=%llu",
4591                                    inode->i_ino,
4592                                    (unsigned long long)bh->b_blocknr);
4593                         err = -EIO;
4594                 }
4595         }
4596         return err;
4597 }
4598
4599 /*
4600  * ext4_setattr()
4601  *
4602  * Called from notify_change.
4603  *
4604  * We want to trap VFS attempts to truncate the file as soon as
4605  * possible.  In particular, we want to make sure that when the VFS
4606  * shrinks i_size, we put the inode on the orphan list and modify
4607  * i_disksize immediately, so that during the subsequent flushing of
4608  * dirty pages and freeing of disk blocks, we can guarantee that any
4609  * commit will leave the blocks being flushed in an unused state on
4610  * disk.  (On recovery, the inode will get truncated and the blocks will
4611  * be freed, so we have a strong guarantee that no future commit will
4612  * leave these blocks visible to the user.)
4613  *
4614  * Another thing we have to assure is that if we are in ordered mode
4615  * and inode is still attached to the committing transaction, we must
4616  * we start writeout of all the dirty pages which are being truncated.
4617  * This way we are sure that all the data written in the previous
4618  * transaction are already on disk (truncate waits for pages under
4619  * writeback).
4620  *
4621  * Called with inode->i_mutex down.
4622  */
4623 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
4624 {
4625         struct inode *inode = dentry->d_inode;
4626         int error, rc = 0;
4627         const unsigned int ia_valid = attr->ia_valid;
4628
4629         error = inode_change_ok(inode, attr);
4630         if (error)
4631                 return error;
4632
4633         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
4634                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
4635                 handle_t *handle;
4636
4637                 /* (user+group)*(old+new) structure, inode write (sb,
4638                  * inode block, ? - but truncate inode update has it) */
4639                 handle = ext4_journal_start(inode, 2*(EXT4_QUOTA_INIT_BLOCKS(inode->i_sb)+
4640                                         EXT4_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
4641                 if (IS_ERR(handle)) {
4642                         error = PTR_ERR(handle);
4643                         goto err_out;
4644                 }
4645                 error = vfs_dq_transfer(inode, attr) ? -EDQUOT : 0;
4646                 if (error) {
4647                         ext4_journal_stop(handle);
4648                         return error;
4649                 }
4650                 /* Update corresponding info in inode so that everything is in
4651                  * one transaction */
4652                 if (attr->ia_valid & ATTR_UID)
4653                         inode->i_uid = attr->ia_uid;
4654                 if (attr->ia_valid & ATTR_GID)
4655                         inode->i_gid = attr->ia_gid;
4656                 error = ext4_mark_inode_dirty(handle, inode);
4657                 ext4_journal_stop(handle);
4658         }
4659
4660         if (attr->ia_valid & ATTR_SIZE) {
4661                 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)) {
4662                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4663
4664                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
4665                                 error = -EFBIG;
4666                                 goto err_out;
4667                         }
4668                 }
4669         }
4670
4671         if (S_ISREG(inode->i_mode) &&
4672             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
4673                 handle_t *handle;
4674
4675                 handle = ext4_journal_start(inode, 3);
4676                 if (IS_ERR(handle)) {
4677                         error = PTR_ERR(handle);
4678                         goto err_out;
4679                 }
4680
4681                 error = ext4_orphan_add(handle, inode);
4682                 EXT4_I(inode)->i_disksize = attr->ia_size;
4683                 rc = ext4_mark_inode_dirty(handle, inode);
4684                 if (!error)
4685                         error = rc;
4686                 ext4_journal_stop(handle);
4687
4688                 if (ext4_should_order_data(inode)) {
4689                         error = ext4_begin_ordered_truncate(inode,
4690                                                             attr->ia_size);
4691                         if (error) {
4692                                 /* Do as much error cleanup as possible */
4693                                 handle = ext4_journal_start(inode, 3);
4694                                 if (IS_ERR(handle)) {
4695                                         ext4_orphan_del(NULL, inode);
4696                                         goto err_out;
4697                                 }
4698                                 ext4_orphan_del(handle, inode);
4699                                 ext4_journal_stop(handle);
4700                                 goto err_out;
4701                         }
4702                 }
4703         }
4704
4705         rc = inode_setattr(inode, attr);
4706
4707         /* If inode_setattr's call to ext4_truncate failed to get a
4708          * transaction handle at all, we need to clean up the in-core
4709          * orphan list manually. */
4710         if (inode->i_nlink)
4711                 ext4_orphan_del(NULL, inode);
4712
4713         if (!rc && (ia_valid & ATTR_MODE))
4714                 rc = ext4_acl_chmod(inode);
4715
4716 err_out:
4717         ext4_std_error(inode->i_sb, error);
4718         if (!error)
4719                 error = rc;
4720         return error;
4721 }
4722
4723 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
4724                  struct kstat *stat)
4725 {
4726         struct inode *inode;
4727         unsigned long delalloc_blocks;
4728
4729         inode = dentry->d_inode;
4730         generic_fillattr(inode, stat);
4731
4732         /*
4733          * We can't update i_blocks if the block allocation is delayed
4734          * otherwise in the case of system crash before the real block
4735          * allocation is done, we will have i_blocks inconsistent with
4736          * on-disk file blocks.
4737          * We always keep i_blocks updated together with real
4738          * allocation. But to not confuse with user, stat
4739          * will return the blocks that include the delayed allocation
4740          * blocks for this file.
4741          */
4742         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
4743         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
4744         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
4745
4746         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
4747         return 0;
4748 }
4749
4750 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
4751                                       int chunk)
4752 {
4753         int indirects;
4754
4755         /* if nrblocks are contiguous */
4756         if (chunk) {
4757                 /*
4758                  * With N contiguous data blocks, it need at most
4759                  * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) indirect blocks
4760                  * 2 dindirect blocks
4761                  * 1 tindirect block
4762                  */
4763                 indirects = nrblocks / EXT4_ADDR_PER_BLOCK(inode->i_sb);
4764                 return indirects + 3;
4765         }
4766         /*
4767          * if nrblocks are not contiguous, worse case, each block touch
4768          * a indirect block, and each indirect block touch a double indirect
4769          * block, plus a triple indirect block
4770          */
4771         indirects = nrblocks * 2 + 1;
4772         return indirects;
4773 }
4774
4775 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4776 {
4777         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL))
4778                 return ext4_indirect_trans_blocks(inode, nrblocks, chunk);
4779         return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
4780 }
4781
4782 /*
4783  * Account for index blocks, block groups bitmaps and block group
4784  * descriptor blocks if modify datablocks and index blocks
4785  * worse case, the indexs blocks spread over different block groups
4786  *
4787  * If datablocks are discontiguous, they are possible to spread over
4788  * different block groups too. If they are contiugous, with flexbg,
4789  * they could still across block group boundary.
4790  *
4791  * Also account for superblock, inode, quota and xattr blocks
4792  */
4793 int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4794 {
4795         int groups, gdpblocks;
4796         int idxblocks;
4797         int ret = 0;
4798
4799         /*
4800          * How many index blocks need to touch to modify nrblocks?
4801          * The "Chunk" flag indicating whether the nrblocks is
4802          * physically contiguous on disk
4803          *
4804          * For Direct IO and fallocate, they calls get_block to allocate
4805          * one single extent at a time, so they could set the "Chunk" flag
4806          */
4807         idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
4808
4809         ret = idxblocks;
4810
4811         /*
4812          * Now let's see how many group bitmaps and group descriptors need
4813          * to account
4814          */
4815         groups = idxblocks;
4816         if (chunk)
4817                 groups += 1;
4818         else
4819                 groups += nrblocks;
4820
4821         gdpblocks = groups;
4822         if (groups > EXT4_SB(inode->i_sb)->s_groups_count)
4823                 groups = EXT4_SB(inode->i_sb)->s_groups_count;
4824         if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
4825                 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
4826
4827         /* bitmaps and block group descriptor blocks */
4828         ret += groups + gdpblocks;
4829
4830         /* Blocks for super block, inode, quota and xattr blocks */
4831         ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
4832
4833         return ret;
4834 }
4835
4836 /*
4837  * Calulate the total number of credits to reserve to fit
4838  * the modification of a single pages into a single transaction,
4839  * which may include multiple chunks of block allocations.
4840  *
4841  * This could be called via ext4_write_begin()
4842  *
4843  * We need to consider the worse case, when
4844  * one new block per extent.
4845  */
4846 int ext4_writepage_trans_blocks(struct inode *inode)
4847 {
4848         int bpp = ext4_journal_blocks_per_page(inode);
4849         int ret;
4850
4851         ret = ext4_meta_trans_blocks(inode, bpp, 0);
4852
4853         /* Account for data blocks for journalled mode */
4854         if (ext4_should_journal_data(inode))
4855                 ret += bpp;
4856         return ret;
4857 }
4858
4859 /*
4860  * Calculate the journal credits for a chunk of data modification.
4861  *
4862  * This is called from DIO, fallocate or whoever calling
4863  * ext4_get_blocks_wrap() to map/allocate a chunk of contigous disk blocks.
4864  *
4865  * journal buffers for data blocks are not included here, as DIO
4866  * and fallocate do no need to journal data buffers.
4867  */
4868 int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
4869 {
4870         return ext4_meta_trans_blocks(inode, nrblocks, 1);
4871 }
4872
4873 /*
4874  * The caller must have previously called ext4_reserve_inode_write().
4875  * Give this, we know that the caller already has write access to iloc->bh.
4876  */
4877 int ext4_mark_iloc_dirty(handle_t *handle,
4878                 struct inode *inode, struct ext4_iloc *iloc)
4879 {
4880         int err = 0;
4881
4882         if (test_opt(inode->i_sb, I_VERSION))
4883                 inode_inc_iversion(inode);
4884
4885         /* the do_update_inode consumes one bh->b_count */
4886         get_bh(iloc->bh);
4887
4888         /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
4889         err = ext4_do_update_inode(handle, inode, iloc);
4890         put_bh(iloc->bh);
4891         return err;
4892 }
4893
4894 /*
4895  * On success, We end up with an outstanding reference count against
4896  * iloc->bh.  This _must_ be cleaned up later.
4897  */
4898
4899 int
4900 ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
4901                          struct ext4_iloc *iloc)
4902 {
4903         int err;
4904
4905         err = ext4_get_inode_loc(inode, iloc);
4906         if (!err) {
4907                 BUFFER_TRACE(iloc->bh, "get_write_access");
4908                 err = ext4_journal_get_write_access(handle, iloc->bh);
4909                 if (err) {
4910                         brelse(iloc->bh);
4911                         iloc->bh = NULL;
4912                 }
4913         }
4914         ext4_std_error(inode->i_sb, err);
4915         return err;
4916 }
4917
4918 /*
4919  * Expand an inode by new_extra_isize bytes.
4920  * Returns 0 on success or negative error number on failure.
4921  */
4922 static int ext4_expand_extra_isize(struct inode *inode,
4923                                    unsigned int new_extra_isize,
4924                                    struct ext4_iloc iloc,
4925                                    handle_t *handle)
4926 {
4927         struct ext4_inode *raw_inode;
4928         struct ext4_xattr_ibody_header *header;
4929         struct ext4_xattr_entry *entry;
4930
4931         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
4932                 return 0;
4933
4934         raw_inode = ext4_raw_inode(&iloc);
4935
4936         header = IHDR(inode, raw_inode);
4937         entry = IFIRST(header);
4938
4939         /* No extended attributes present */
4940         if (!(EXT4_I(inode)->i_state & EXT4_STATE_XATTR) ||
4941                 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
4942                 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
4943                         new_extra_isize);
4944                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
4945                 return 0;
4946         }
4947
4948         /* try to expand with EAs present */
4949         return ext4_expand_extra_isize_ea(inode, new_extra_isize,
4950                                           raw_inode, handle);
4951 }
4952
4953 /*
4954  * What we do here is to mark the in-core inode as clean with respect to inode
4955  * dirtiness (it may still be data-dirty).
4956  * This means that the in-core inode may be reaped by prune_icache
4957  * without having to perform any I/O.  This is a very good thing,
4958  * because *any* task may call prune_icache - even ones which
4959  * have a transaction open against a different journal.
4960  *
4961  * Is this cheating?  Not really.  Sure, we haven't written the
4962  * inode out, but prune_icache isn't a user-visible syncing function.
4963  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
4964  * we start and wait on commits.
4965  *
4966  * Is this efficient/effective?  Well, we're being nice to the system
4967  * by cleaning up our inodes proactively so they can be reaped
4968  * without I/O.  But we are potentially leaving up to five seconds'
4969  * worth of inodes floating about which prune_icache wants us to
4970  * write out.  One way to fix that would be to get prune_icache()
4971  * to do a write_super() to free up some memory.  It has the desired
4972  * effect.
4973  */
4974 int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
4975 {
4976         struct ext4_iloc iloc;
4977         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4978         static unsigned int mnt_count;
4979         int err, ret;
4980
4981         might_sleep();
4982         err = ext4_reserve_inode_write(handle, inode, &iloc);
4983         if (ext4_handle_valid(handle) &&
4984             EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
4985             !(EXT4_I(inode)->i_state & EXT4_STATE_NO_EXPAND)) {
4986                 /*
4987                  * We need extra buffer credits since we may write into EA block
4988                  * with this same handle. If journal_extend fails, then it will
4989                  * only result in a minor loss of functionality for that inode.
4990                  * If this is felt to be critical, then e2fsck should be run to
4991                  * force a large enough s_min_extra_isize.
4992                  */
4993                 if ((jbd2_journal_extend(handle,
4994                              EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
4995                         ret = ext4_expand_extra_isize(inode,
4996                                                       sbi->s_want_extra_isize,
4997                                                       iloc, handle);
4998                         if (ret) {
4999                                 EXT4_I(inode)->i_state |= EXT4_STATE_NO_EXPAND;
5000                                 if (mnt_count !=
5001                                         le16_to_cpu(sbi->s_es->s_mnt_count)) {
5002                                         ext4_warning(inode->i_sb, __func__,
5003                                         "Unable to expand inode %lu. Delete"
5004                                         " some EAs or run e2fsck.",
5005                                         inode->i_ino);
5006                                         mnt_count =
5007                                           le16_to_cpu(sbi->s_es->s_mnt_count);
5008                                 }
5009                         }
5010                 }
5011         }
5012         if (!err)
5013                 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
5014         return err;
5015 }
5016
5017 /*
5018  * ext4_dirty_inode() is called from __mark_inode_dirty()
5019  *
5020  * We're really interested in the case where a file is being extended.
5021  * i_size has been changed by generic_commit_write() and we thus need
5022  * to include the updated inode in the current transaction.
5023  *
5024  * Also, vfs_dq_alloc_block() will always dirty the inode when blocks
5025  * are allocated to the file.
5026  *
5027  * If the inode is marked synchronous, we don't honour that here - doing
5028  * so would cause a commit on atime updates, which we don't bother doing.
5029  * We handle synchronous inodes at the highest possible level.
5030  */
5031 void ext4_dirty_inode(struct inode *inode)
5032 {
5033         handle_t *current_handle = ext4_journal_current_handle();
5034         handle_t *handle;
5035
5036         if (!ext4_handle_valid(current_handle)) {
5037                 ext4_mark_inode_dirty(current_handle, inode);
5038                 return;
5039         }
5040
5041         handle = ext4_journal_start(inode, 2);
5042         if (IS_ERR(handle))
5043                 goto out;
5044         if (current_handle &&
5045                 current_handle->h_transaction != handle->h_transaction) {
5046                 /* This task has a transaction open against a different fs */
5047                 printk(KERN_EMERG "%s: transactions do not match!\n",
5048                        __func__);
5049         } else {
5050                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
5051                                 current_handle);
5052                 ext4_mark_inode_dirty(handle, inode);
5053         }
5054         ext4_journal_stop(handle);
5055 out:
5056         return;
5057 }
5058
5059 #if 0
5060 /*
5061  * Bind an inode's backing buffer_head into this transaction, to prevent
5062  * it from being flushed to disk early.  Unlike
5063  * ext4_reserve_inode_write, this leaves behind no bh reference and
5064  * returns no iloc structure, so the caller needs to repeat the iloc
5065  * lookup to mark the inode dirty later.
5066  */
5067 static int ext4_pin_inode(handle_t *handle, struct inode *inode)
5068 {
5069         struct ext4_iloc iloc;
5070
5071         int err = 0;
5072         if (handle) {
5073                 err = ext4_get_inode_loc(inode, &iloc);
5074                 if (!err) {
5075                         BUFFER_TRACE(iloc.bh, "get_write_access");
5076                         err = jbd2_journal_get_write_access(handle, iloc.bh);
5077                         if (!err)
5078                                 err = ext4_handle_dirty_metadata(handle,
5079                                                                  inode,
5080                                                                  iloc.bh);
5081                         brelse(iloc.bh);
5082                 }
5083         }
5084         ext4_std_error(inode->i_sb, err);
5085         return err;
5086 }
5087 #endif
5088
5089 int ext4_change_inode_journal_flag(struct inode *inode, int val)
5090 {
5091         journal_t *journal;
5092         handle_t *handle;
5093         int err;
5094
5095         /*
5096          * We have to be very careful here: changing a data block's
5097          * journaling status dynamically is dangerous.  If we write a
5098          * data block to the journal, change the status and then delete
5099          * that block, we risk forgetting to revoke the old log record
5100          * from the journal and so a subsequent replay can corrupt data.
5101          * So, first we make sure that the journal is empty and that
5102          * nobody is changing anything.
5103          */
5104
5105         journal = EXT4_JOURNAL(inode);
5106         if (!journal)
5107                 return 0;
5108         if (is_journal_aborted(journal))
5109                 return -EROFS;
5110
5111         jbd2_journal_lock_updates(journal);
5112         jbd2_journal_flush(journal);
5113
5114         /*
5115          * OK, there are no updates running now, and all cached data is
5116          * synced to disk.  We are now in a completely consistent state
5117          * which doesn't have anything in the journal, and we know that
5118          * no filesystem updates are running, so it is safe to modify
5119          * the inode's in-core data-journaling state flag now.
5120          */
5121
5122         if (val)
5123                 EXT4_I(inode)->i_flags |= EXT4_JOURNAL_DATA_FL;
5124         else
5125                 EXT4_I(inode)->i_flags &= ~EXT4_JOURNAL_DATA_FL;
5126         ext4_set_aops(inode);
5127
5128         jbd2_journal_unlock_updates(journal);
5129
5130         /* Finally we can mark the inode as dirty. */
5131
5132         handle = ext4_journal_start(inode, 1);
5133         if (IS_ERR(handle))
5134                 return PTR_ERR(handle);
5135
5136         err = ext4_mark_inode_dirty(handle, inode);
5137         ext4_handle_sync(handle);
5138         ext4_journal_stop(handle);
5139         ext4_std_error(inode->i_sb, err);
5140
5141         return err;
5142 }
5143
5144 static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
5145 {
5146         return !buffer_mapped(bh);
5147 }
5148
5149 int ext4_page_mkwrite(struct vm_area_struct *vma, struct page *page)
5150 {
5151         loff_t size;
5152         unsigned long len;
5153         int ret = -EINVAL;
5154         void *fsdata;
5155         struct file *file = vma->vm_file;
5156         struct inode *inode = file->f_path.dentry->d_inode;
5157         struct address_space *mapping = inode->i_mapping;
5158
5159         /*
5160          * Get i_alloc_sem to stop truncates messing with the inode. We cannot
5161          * get i_mutex because we are already holding mmap_sem.
5162          */
5163         down_read(&inode->i_alloc_sem);
5164         size = i_size_read(inode);
5165         if (page->mapping != mapping || size <= page_offset(page)
5166             || !PageUptodate(page)) {
5167                 /* page got truncated from under us? */
5168                 goto out_unlock;
5169         }
5170         ret = 0;
5171         if (PageMappedToDisk(page))
5172                 goto out_unlock;
5173
5174         if (page->index == size >> PAGE_CACHE_SHIFT)
5175                 len = size & ~PAGE_CACHE_MASK;
5176         else
5177                 len = PAGE_CACHE_SIZE;
5178
5179         if (page_has_buffers(page)) {
5180                 /* return if we have all the buffers mapped */
5181                 if (!walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
5182                                        ext4_bh_unmapped))
5183                         goto out_unlock;
5184         }
5185         /*
5186          * OK, we need to fill the hole... Do write_begin write_end
5187          * to do block allocation/reservation.We are not holding
5188          * inode.i__mutex here. That allow * parallel write_begin,
5189          * write_end call. lock_page prevent this from happening
5190          * on the same page though
5191          */
5192         ret = mapping->a_ops->write_begin(file, mapping, page_offset(page),
5193                         len, AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
5194         if (ret < 0)
5195                 goto out_unlock;
5196         ret = mapping->a_ops->write_end(file, mapping, page_offset(page),
5197                         len, len, page, fsdata);
5198         if (ret < 0)
5199                 goto out_unlock;
5200         ret = 0;
5201 out_unlock:
5202         up_read(&inode->i_alloc_sem);
5203         return ret;
5204 }