Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-2.6
[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/uio.h>
38 #include <linux/bio.h>
39 #include "ext4_jbd2.h"
40 #include "xattr.h"
41 #include "acl.h"
42 #include "ext4_extents.h"
43
44 #define MPAGE_DA_EXTENT_TAIL 0x01
45
46 static inline int ext4_begin_ordered_truncate(struct inode *inode,
47                                               loff_t new_size)
48 {
49         return jbd2_journal_begin_ordered_truncate(&EXT4_I(inode)->jinode,
50                                                    new_size);
51 }
52
53 static void ext4_invalidatepage(struct page *page, unsigned long offset);
54
55 /*
56  * Test whether an inode is a fast symlink.
57  */
58 static int ext4_inode_is_fast_symlink(struct inode *inode)
59 {
60         int ea_blocks = EXT4_I(inode)->i_file_acl ?
61                 (inode->i_sb->s_blocksize >> 9) : 0;
62
63         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
64 }
65
66 /*
67  * The ext4 forget function must perform a revoke if we are freeing data
68  * which has been journaled.  Metadata (eg. indirect blocks) must be
69  * revoked in all cases.
70  *
71  * "bh" may be NULL: a metadata block may have been freed from memory
72  * but there may still be a record of it in the journal, and that record
73  * still needs to be revoked.
74  */
75 int ext4_forget(handle_t *handle, int is_metadata, struct inode *inode,
76                         struct buffer_head *bh, ext4_fsblk_t blocknr)
77 {
78         int err;
79
80         might_sleep();
81
82         BUFFER_TRACE(bh, "enter");
83
84         jbd_debug(4, "forgetting bh %p: is_metadata = %d, mode %o, "
85                   "data mode %lx\n",
86                   bh, is_metadata, inode->i_mode,
87                   test_opt(inode->i_sb, DATA_FLAGS));
88
89         /* Never use the revoke function if we are doing full data
90          * journaling: there is no need to, and a V1 superblock won't
91          * support it.  Otherwise, only skip the revoke on un-journaled
92          * data blocks. */
93
94         if (test_opt(inode->i_sb, DATA_FLAGS) == EXT4_MOUNT_JOURNAL_DATA ||
95             (!is_metadata && !ext4_should_journal_data(inode))) {
96                 if (bh) {
97                         BUFFER_TRACE(bh, "call jbd2_journal_forget");
98                         return ext4_journal_forget(handle, bh);
99                 }
100                 return 0;
101         }
102
103         /*
104          * data!=journal && (is_metadata || should_journal_data(inode))
105          */
106         BUFFER_TRACE(bh, "call ext4_journal_revoke");
107         err = ext4_journal_revoke(handle, blocknr, bh);
108         if (err)
109                 ext4_abort(inode->i_sb, __func__,
110                            "error %d when attempting revoke", err);
111         BUFFER_TRACE(bh, "exit");
112         return err;
113 }
114
115 /*
116  * Work out how many blocks we need to proceed with the next chunk of a
117  * truncate transaction.
118  */
119 static unsigned long blocks_for_truncate(struct inode *inode)
120 {
121         ext4_lblk_t needed;
122
123         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
124
125         /* Give ourselves just enough room to cope with inodes in which
126          * i_blocks is corrupt: we've seen disk corruptions in the past
127          * which resulted in random data in an inode which looked enough
128          * like a regular file for ext4 to try to delete it.  Things
129          * will go a bit crazy if that happens, but at least we should
130          * try not to panic the whole kernel. */
131         if (needed < 2)
132                 needed = 2;
133
134         /* But we need to bound the transaction so we don't overflow the
135          * journal. */
136         if (needed > EXT4_MAX_TRANS_DATA)
137                 needed = EXT4_MAX_TRANS_DATA;
138
139         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
140 }
141
142 /*
143  * Truncate transactions can be complex and absolutely huge.  So we need to
144  * be able to restart the transaction at a conventient checkpoint to make
145  * sure we don't overflow the journal.
146  *
147  * start_transaction gets us a new handle for a truncate transaction,
148  * and extend_transaction tries to extend the existing one a bit.  If
149  * extend fails, we need to propagate the failure up and restart the
150  * transaction in the top-level truncate loop. --sct
151  */
152 static handle_t *start_transaction(struct inode *inode)
153 {
154         handle_t *result;
155
156         result = ext4_journal_start(inode, blocks_for_truncate(inode));
157         if (!IS_ERR(result))
158                 return result;
159
160         ext4_std_error(inode->i_sb, PTR_ERR(result));
161         return result;
162 }
163
164 /*
165  * Try to extend this transaction for the purposes of truncation.
166  *
167  * Returns 0 if we managed to create more room.  If we can't create more
168  * room, and the transaction must be restarted we return 1.
169  */
170 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
171 {
172         if (handle->h_buffer_credits > EXT4_RESERVE_TRANS_BLOCKS)
173                 return 0;
174         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
175                 return 0;
176         return 1;
177 }
178
179 /*
180  * Restart the transaction associated with *handle.  This does a commit,
181  * so before we call here everything must be consistently dirtied against
182  * this transaction.
183  */
184 static int ext4_journal_test_restart(handle_t *handle, struct inode *inode)
185 {
186         jbd_debug(2, "restarting handle %p\n", handle);
187         return ext4_journal_restart(handle, blocks_for_truncate(inode));
188 }
189
190 /*
191  * Called at the last iput() if i_nlink is zero.
192  */
193 void ext4_delete_inode (struct inode * inode)
194 {
195         handle_t *handle;
196         int err;
197
198         if (ext4_should_order_data(inode))
199                 ext4_begin_ordered_truncate(inode, 0);
200         truncate_inode_pages(&inode->i_data, 0);
201
202         if (is_bad_inode(inode))
203                 goto no_delete;
204
205         handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
206         if (IS_ERR(handle)) {
207                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
208                 /*
209                  * If we're going to skip the normal cleanup, we still need to
210                  * make sure that the in-core orphan linked list is properly
211                  * cleaned up.
212                  */
213                 ext4_orphan_del(NULL, inode);
214                 goto no_delete;
215         }
216
217         if (IS_SYNC(inode))
218                 handle->h_sync = 1;
219         inode->i_size = 0;
220         err = ext4_mark_inode_dirty(handle, inode);
221         if (err) {
222                 ext4_warning(inode->i_sb, __func__,
223                              "couldn't mark inode dirty (err %d)", err);
224                 goto stop_handle;
225         }
226         if (inode->i_blocks)
227                 ext4_truncate(inode);
228
229         /*
230          * ext4_ext_truncate() doesn't reserve any slop when it
231          * restarts journal transactions; therefore there may not be
232          * enough credits left in the handle to remove the inode from
233          * the orphan list and set the dtime field.
234          */
235         if (handle->h_buffer_credits < 3) {
236                 err = ext4_journal_extend(handle, 3);
237                 if (err > 0)
238                         err = ext4_journal_restart(handle, 3);
239                 if (err != 0) {
240                         ext4_warning(inode->i_sb, __func__,
241                                      "couldn't extend journal (err %d)", err);
242                 stop_handle:
243                         ext4_journal_stop(handle);
244                         goto no_delete;
245                 }
246         }
247
248         /*
249          * Kill off the orphan record which ext4_truncate created.
250          * AKPM: I think this can be inside the above `if'.
251          * Note that ext4_orphan_del() has to be able to cope with the
252          * deletion of a non-existent orphan - this is because we don't
253          * know if ext4_truncate() actually created an orphan record.
254          * (Well, we could do this if we need to, but heck - it works)
255          */
256         ext4_orphan_del(handle, inode);
257         EXT4_I(inode)->i_dtime  = get_seconds();
258
259         /*
260          * One subtle ordering requirement: if anything has gone wrong
261          * (transaction abort, IO errors, whatever), then we can still
262          * do these next steps (the fs will already have been marked as
263          * having errors), but we can't free the inode if the mark_dirty
264          * fails.
265          */
266         if (ext4_mark_inode_dirty(handle, inode))
267                 /* If that failed, just do the required in-core inode clear. */
268                 clear_inode(inode);
269         else
270                 ext4_free_inode(handle, inode);
271         ext4_journal_stop(handle);
272         return;
273 no_delete:
274         clear_inode(inode);     /* We must guarantee clearing of inode... */
275 }
276
277 typedef struct {
278         __le32  *p;
279         __le32  key;
280         struct buffer_head *bh;
281 } Indirect;
282
283 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
284 {
285         p->key = *(p->p = v);
286         p->bh = bh;
287 }
288
289 /**
290  *      ext4_block_to_path - parse the block number into array of offsets
291  *      @inode: inode in question (we are only interested in its superblock)
292  *      @i_block: block number to be parsed
293  *      @offsets: array to store the offsets in
294  *      @boundary: set this non-zero if the referred-to block is likely to be
295  *             followed (on disk) by an indirect block.
296  *
297  *      To store the locations of file's data ext4 uses a data structure common
298  *      for UNIX filesystems - tree of pointers anchored in the inode, with
299  *      data blocks at leaves and indirect blocks in intermediate nodes.
300  *      This function translates the block number into path in that tree -
301  *      return value is the path length and @offsets[n] is the offset of
302  *      pointer to (n+1)th node in the nth one. If @block is out of range
303  *      (negative or too large) warning is printed and zero returned.
304  *
305  *      Note: function doesn't find node addresses, so no IO is needed. All
306  *      we need to know is the capacity of indirect blocks (taken from the
307  *      inode->i_sb).
308  */
309
310 /*
311  * Portability note: the last comparison (check that we fit into triple
312  * indirect block) is spelled differently, because otherwise on an
313  * architecture with 32-bit longs and 8Kb pages we might get into trouble
314  * if our filesystem had 8Kb blocks. We might use long long, but that would
315  * kill us on x86. Oh, well, at least the sign propagation does not matter -
316  * i_block would have to be negative in the very beginning, so we would not
317  * get there at all.
318  */
319
320 static int ext4_block_to_path(struct inode *inode,
321                         ext4_lblk_t i_block,
322                         ext4_lblk_t offsets[4], int *boundary)
323 {
324         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
325         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
326         const long direct_blocks = EXT4_NDIR_BLOCKS,
327                 indirect_blocks = ptrs,
328                 double_blocks = (1 << (ptrs_bits * 2));
329         int n = 0;
330         int final = 0;
331
332         if (i_block < 0) {
333                 ext4_warning (inode->i_sb, "ext4_block_to_path", "block < 0");
334         } else if (i_block < direct_blocks) {
335                 offsets[n++] = i_block;
336                 final = direct_blocks;
337         } else if ( (i_block -= direct_blocks) < indirect_blocks) {
338                 offsets[n++] = EXT4_IND_BLOCK;
339                 offsets[n++] = i_block;
340                 final = ptrs;
341         } else if ((i_block -= indirect_blocks) < double_blocks) {
342                 offsets[n++] = EXT4_DIND_BLOCK;
343                 offsets[n++] = i_block >> ptrs_bits;
344                 offsets[n++] = i_block & (ptrs - 1);
345                 final = ptrs;
346         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
347                 offsets[n++] = EXT4_TIND_BLOCK;
348                 offsets[n++] = i_block >> (ptrs_bits * 2);
349                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
350                 offsets[n++] = i_block & (ptrs - 1);
351                 final = ptrs;
352         } else {
353                 ext4_warning(inode->i_sb, "ext4_block_to_path",
354                                 "block %lu > max",
355                                 i_block + direct_blocks +
356                                 indirect_blocks + double_blocks);
357         }
358         if (boundary)
359                 *boundary = final - 1 - (i_block & (ptrs - 1));
360         return n;
361 }
362
363 /**
364  *      ext4_get_branch - read the chain of indirect blocks leading to data
365  *      @inode: inode in question
366  *      @depth: depth of the chain (1 - direct pointer, etc.)
367  *      @offsets: offsets of pointers in inode/indirect blocks
368  *      @chain: place to store the result
369  *      @err: here we store the error value
370  *
371  *      Function fills the array of triples <key, p, bh> and returns %NULL
372  *      if everything went OK or the pointer to the last filled triple
373  *      (incomplete one) otherwise. Upon the return chain[i].key contains
374  *      the number of (i+1)-th block in the chain (as it is stored in memory,
375  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
376  *      number (it points into struct inode for i==0 and into the bh->b_data
377  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
378  *      block for i>0 and NULL for i==0. In other words, it holds the block
379  *      numbers of the chain, addresses they were taken from (and where we can
380  *      verify that chain did not change) and buffer_heads hosting these
381  *      numbers.
382  *
383  *      Function stops when it stumbles upon zero pointer (absent block)
384  *              (pointer to last triple returned, *@err == 0)
385  *      or when it gets an IO error reading an indirect block
386  *              (ditto, *@err == -EIO)
387  *      or when it reads all @depth-1 indirect blocks successfully and finds
388  *      the whole chain, all way to the data (returns %NULL, *err == 0).
389  *
390  *      Need to be called with
391  *      down_read(&EXT4_I(inode)->i_data_sem)
392  */
393 static Indirect *ext4_get_branch(struct inode *inode, int depth,
394                                  ext4_lblk_t  *offsets,
395                                  Indirect chain[4], int *err)
396 {
397         struct super_block *sb = inode->i_sb;
398         Indirect *p = chain;
399         struct buffer_head *bh;
400
401         *err = 0;
402         /* i_data is not going away, no lock needed */
403         add_chain (chain, NULL, EXT4_I(inode)->i_data + *offsets);
404         if (!p->key)
405                 goto no_block;
406         while (--depth) {
407                 bh = sb_bread(sb, le32_to_cpu(p->key));
408                 if (!bh)
409                         goto failure;
410                 add_chain(++p, bh, (__le32*)bh->b_data + *++offsets);
411                 /* Reader: end */
412                 if (!p->key)
413                         goto no_block;
414         }
415         return NULL;
416
417 failure:
418         *err = -EIO;
419 no_block:
420         return p;
421 }
422
423 /**
424  *      ext4_find_near - find a place for allocation with sufficient locality
425  *      @inode: owner
426  *      @ind: descriptor of indirect block.
427  *
428  *      This function returns the preferred place for block allocation.
429  *      It is used when heuristic for sequential allocation fails.
430  *      Rules are:
431  *        + if there is a block to the left of our position - allocate near it.
432  *        + if pointer will live in indirect block - allocate near that block.
433  *        + if pointer will live in inode - allocate in the same
434  *          cylinder group.
435  *
436  * In the latter case we colour the starting block by the callers PID to
437  * prevent it from clashing with concurrent allocations for a different inode
438  * in the same block group.   The PID is used here so that functionally related
439  * files will be close-by on-disk.
440  *
441  *      Caller must make sure that @ind is valid and will stay that way.
442  */
443 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
444 {
445         struct ext4_inode_info *ei = EXT4_I(inode);
446         __le32 *start = ind->bh ? (__le32*) ind->bh->b_data : ei->i_data;
447         __le32 *p;
448         ext4_fsblk_t bg_start;
449         ext4_fsblk_t last_block;
450         ext4_grpblk_t colour;
451
452         /* Try to find previous block */
453         for (p = ind->p - 1; p >= start; p--) {
454                 if (*p)
455                         return le32_to_cpu(*p);
456         }
457
458         /* No such thing, so let's try location of indirect block */
459         if (ind->bh)
460                 return ind->bh->b_blocknr;
461
462         /*
463          * It is going to be referred to from the inode itself? OK, just put it
464          * into the same cylinder group then.
465          */
466         bg_start = ext4_group_first_block_no(inode->i_sb, ei->i_block_group);
467         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
468
469         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
470                 colour = (current->pid % 16) *
471                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
472         else
473                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
474         return bg_start + colour;
475 }
476
477 /**
478  *      ext4_find_goal - find a preferred place for allocation.
479  *      @inode: owner
480  *      @block:  block we want
481  *      @partial: pointer to the last triple within a chain
482  *
483  *      Normally this function find the preferred place for block allocation,
484  *      returns it.
485  */
486 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
487                 Indirect *partial)
488 {
489         struct ext4_block_alloc_info *block_i;
490
491         block_i =  EXT4_I(inode)->i_block_alloc_info;
492
493         /*
494          * try the heuristic for sequential allocation,
495          * failing that at least try to get decent locality.
496          */
497         if (block_i && (block == block_i->last_alloc_logical_block + 1)
498                 && (block_i->last_alloc_physical_block != 0)) {
499                 return block_i->last_alloc_physical_block + 1;
500         }
501
502         return ext4_find_near(inode, partial);
503 }
504
505 /**
506  *      ext4_blks_to_allocate: Look up the block map and count the number
507  *      of direct blocks need to be allocated for the given branch.
508  *
509  *      @branch: chain of indirect blocks
510  *      @k: number of blocks need for indirect blocks
511  *      @blks: number of data blocks to be mapped.
512  *      @blocks_to_boundary:  the offset in the indirect block
513  *
514  *      return the total number of blocks to be allocate, including the
515  *      direct and indirect blocks.
516  */
517 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned long blks,
518                 int blocks_to_boundary)
519 {
520         unsigned long count = 0;
521
522         /*
523          * Simple case, [t,d]Indirect block(s) has not allocated yet
524          * then it's clear blocks on that path have not allocated
525          */
526         if (k > 0) {
527                 /* right now we don't handle cross boundary allocation */
528                 if (blks < blocks_to_boundary + 1)
529                         count += blks;
530                 else
531                         count += blocks_to_boundary + 1;
532                 return count;
533         }
534
535         count++;
536         while (count < blks && count <= blocks_to_boundary &&
537                 le32_to_cpu(*(branch[0].p + count)) == 0) {
538                 count++;
539         }
540         return count;
541 }
542
543 /**
544  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
545  *      @indirect_blks: the number of blocks need to allocate for indirect
546  *                      blocks
547  *
548  *      @new_blocks: on return it will store the new block numbers for
549  *      the indirect blocks(if needed) and the first direct block,
550  *      @blks:  on return it will store the total number of allocated
551  *              direct blocks
552  */
553 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
554                                 ext4_lblk_t iblock, ext4_fsblk_t goal,
555                                 int indirect_blks, int blks,
556                                 ext4_fsblk_t new_blocks[4], int *err)
557 {
558         int target, i;
559         unsigned long count = 0, blk_allocated = 0;
560         int index = 0;
561         ext4_fsblk_t current_block = 0;
562         int ret = 0;
563
564         /*
565          * Here we try to allocate the requested multiple blocks at once,
566          * on a best-effort basis.
567          * To build a branch, we should allocate blocks for
568          * the indirect blocks(if not allocated yet), and at least
569          * the first direct block of this branch.  That's the
570          * minimum number of blocks need to allocate(required)
571          */
572         /* first we try to allocate the indirect blocks */
573         target = indirect_blks;
574         while (target > 0) {
575                 count = target;
576                 /* allocating blocks for indirect blocks and direct blocks */
577                 current_block = ext4_new_meta_blocks(handle, inode,
578                                                         goal, &count, err);
579                 if (*err)
580                         goto failed_out;
581
582                 target -= count;
583                 /* allocate blocks for indirect blocks */
584                 while (index < indirect_blks && count) {
585                         new_blocks[index++] = current_block++;
586                         count--;
587                 }
588                 if (count > 0) {
589                         /*
590                          * save the new block number
591                          * for the first direct block
592                          */
593                         new_blocks[index] = current_block;
594                         printk(KERN_INFO "%s returned more blocks than "
595                                                 "requested\n", __func__);
596                         WARN_ON(1);
597                         break;
598                 }
599         }
600
601         target = blks - count ;
602         blk_allocated = count;
603         if (!target)
604                 goto allocated;
605         /* Now allocate data blocks */
606         count = target;
607         /* allocating blocks for data blocks */
608         current_block = ext4_new_blocks(handle, inode, iblock,
609                                                 goal, &count, err);
610         if (*err && (target == blks)) {
611                 /*
612                  * if the allocation failed and we didn't allocate
613                  * any blocks before
614                  */
615                 goto failed_out;
616         }
617         if (!*err) {
618                 if (target == blks) {
619                 /*
620                  * save the new block number
621                  * for the first direct block
622                  */
623                         new_blocks[index] = current_block;
624                 }
625                 blk_allocated += count;
626         }
627 allocated:
628         /* total number of blocks allocated for direct blocks */
629         ret = blk_allocated;
630         *err = 0;
631         return ret;
632 failed_out:
633         for (i = 0; i <index; i++)
634                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
635         return ret;
636 }
637
638 /**
639  *      ext4_alloc_branch - allocate and set up a chain of blocks.
640  *      @inode: owner
641  *      @indirect_blks: number of allocated indirect blocks
642  *      @blks: number of allocated direct blocks
643  *      @offsets: offsets (in the blocks) to store the pointers to next.
644  *      @branch: place to store the chain in.
645  *
646  *      This function allocates blocks, zeroes out all but the last one,
647  *      links them into chain and (if we are synchronous) writes them to disk.
648  *      In other words, it prepares a branch that can be spliced onto the
649  *      inode. It stores the information about that chain in the branch[], in
650  *      the same format as ext4_get_branch() would do. We are calling it after
651  *      we had read the existing part of chain and partial points to the last
652  *      triple of that (one with zero ->key). Upon the exit we have the same
653  *      picture as after the successful ext4_get_block(), except that in one
654  *      place chain is disconnected - *branch->p is still zero (we did not
655  *      set the last link), but branch->key contains the number that should
656  *      be placed into *branch->p to fill that gap.
657  *
658  *      If allocation fails we free all blocks we've allocated (and forget
659  *      their buffer_heads) and return the error value the from failed
660  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
661  *      as described above and return 0.
662  */
663 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
664                                 ext4_lblk_t iblock, int indirect_blks,
665                                 int *blks, ext4_fsblk_t goal,
666                                 ext4_lblk_t *offsets, Indirect *branch)
667 {
668         int blocksize = inode->i_sb->s_blocksize;
669         int i, n = 0;
670         int err = 0;
671         struct buffer_head *bh;
672         int num;
673         ext4_fsblk_t new_blocks[4];
674         ext4_fsblk_t current_block;
675
676         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
677                                 *blks, new_blocks, &err);
678         if (err)
679                 return err;
680
681         branch[0].key = cpu_to_le32(new_blocks[0]);
682         /*
683          * metadata blocks and data blocks are allocated.
684          */
685         for (n = 1; n <= indirect_blks;  n++) {
686                 /*
687                  * Get buffer_head for parent block, zero it out
688                  * and set the pointer to new one, then send
689                  * parent to disk.
690                  */
691                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
692                 branch[n].bh = bh;
693                 lock_buffer(bh);
694                 BUFFER_TRACE(bh, "call get_create_access");
695                 err = ext4_journal_get_create_access(handle, bh);
696                 if (err) {
697                         unlock_buffer(bh);
698                         brelse(bh);
699                         goto failed;
700                 }
701
702                 memset(bh->b_data, 0, blocksize);
703                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
704                 branch[n].key = cpu_to_le32(new_blocks[n]);
705                 *branch[n].p = branch[n].key;
706                 if ( n == indirect_blks) {
707                         current_block = new_blocks[n];
708                         /*
709                          * End of chain, update the last new metablock of
710                          * the chain to point to the new allocated
711                          * data blocks numbers
712                          */
713                         for (i=1; i < num; i++)
714                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
715                 }
716                 BUFFER_TRACE(bh, "marking uptodate");
717                 set_buffer_uptodate(bh);
718                 unlock_buffer(bh);
719
720                 BUFFER_TRACE(bh, "call ext4_journal_dirty_metadata");
721                 err = ext4_journal_dirty_metadata(handle, bh);
722                 if (err)
723                         goto failed;
724         }
725         *blks = num;
726         return err;
727 failed:
728         /* Allocation failed, free what we already allocated */
729         for (i = 1; i <= n ; i++) {
730                 BUFFER_TRACE(branch[i].bh, "call jbd2_journal_forget");
731                 ext4_journal_forget(handle, branch[i].bh);
732         }
733         for (i = 0; i <indirect_blks; i++)
734                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
735
736         ext4_free_blocks(handle, inode, new_blocks[i], num, 0);
737
738         return err;
739 }
740
741 /**
742  * ext4_splice_branch - splice the allocated branch onto inode.
743  * @inode: owner
744  * @block: (logical) number of block we are adding
745  * @chain: chain of indirect blocks (with a missing link - see
746  *      ext4_alloc_branch)
747  * @where: location of missing link
748  * @num:   number of indirect blocks we are adding
749  * @blks:  number of direct blocks we are adding
750  *
751  * This function fills the missing link and does all housekeeping needed in
752  * inode (->i_blocks, etc.). In case of success we end up with the full
753  * chain to new block and return 0.
754  */
755 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
756                         ext4_lblk_t block, Indirect *where, int num, int blks)
757 {
758         int i;
759         int err = 0;
760         struct ext4_block_alloc_info *block_i;
761         ext4_fsblk_t current_block;
762
763         block_i = EXT4_I(inode)->i_block_alloc_info;
764         /*
765          * If we're splicing into a [td]indirect block (as opposed to the
766          * inode) then we need to get write access to the [td]indirect block
767          * before the splice.
768          */
769         if (where->bh) {
770                 BUFFER_TRACE(where->bh, "get_write_access");
771                 err = ext4_journal_get_write_access(handle, where->bh);
772                 if (err)
773                         goto err_out;
774         }
775         /* That's it */
776
777         *where->p = where->key;
778
779         /*
780          * Update the host buffer_head or inode to point to more just allocated
781          * direct blocks blocks
782          */
783         if (num == 0 && blks > 1) {
784                 current_block = le32_to_cpu(where->key) + 1;
785                 for (i = 1; i < blks; i++)
786                         *(where->p + i ) = cpu_to_le32(current_block++);
787         }
788
789         /*
790          * update the most recently allocated logical & physical block
791          * in i_block_alloc_info, to assist find the proper goal block for next
792          * allocation
793          */
794         if (block_i) {
795                 block_i->last_alloc_logical_block = block + blks - 1;
796                 block_i->last_alloc_physical_block =
797                                 le32_to_cpu(where[num].key) + blks - 1;
798         }
799
800         /* We are done with atomic stuff, now do the rest of housekeeping */
801
802         inode->i_ctime = ext4_current_time(inode);
803         ext4_mark_inode_dirty(handle, inode);
804
805         /* had we spliced it onto indirect block? */
806         if (where->bh) {
807                 /*
808                  * If we spliced it onto an indirect block, we haven't
809                  * altered the inode.  Note however that if it is being spliced
810                  * onto an indirect block at the very end of the file (the
811                  * file is growing) then we *will* alter the inode to reflect
812                  * the new i_size.  But that is not done here - it is done in
813                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
814                  */
815                 jbd_debug(5, "splicing indirect only\n");
816                 BUFFER_TRACE(where->bh, "call ext4_journal_dirty_metadata");
817                 err = ext4_journal_dirty_metadata(handle, where->bh);
818                 if (err)
819                         goto err_out;
820         } else {
821                 /*
822                  * OK, we spliced it into the inode itself on a direct block.
823                  * Inode was dirtied above.
824                  */
825                 jbd_debug(5, "splicing direct\n");
826         }
827         return err;
828
829 err_out:
830         for (i = 1; i <= num; i++) {
831                 BUFFER_TRACE(where[i].bh, "call jbd2_journal_forget");
832                 ext4_journal_forget(handle, where[i].bh);
833                 ext4_free_blocks(handle, inode,
834                                         le32_to_cpu(where[i-1].key), 1, 0);
835         }
836         ext4_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks, 0);
837
838         return err;
839 }
840
841 /*
842  * Allocation strategy is simple: if we have to allocate something, we will
843  * have to go the whole way to leaf. So let's do it before attaching anything
844  * to tree, set linkage between the newborn blocks, write them if sync is
845  * required, recheck the path, free and repeat if check fails, otherwise
846  * set the last missing link (that will protect us from any truncate-generated
847  * removals - all blocks on the path are immune now) and possibly force the
848  * write on the parent block.
849  * That has a nice additional property: no special recovery from the failed
850  * allocations is needed - we simply release blocks and do not touch anything
851  * reachable from inode.
852  *
853  * `handle' can be NULL if create == 0.
854  *
855  * return > 0, # of blocks mapped or allocated.
856  * return = 0, if plain lookup failed.
857  * return < 0, error case.
858  *
859  *
860  * Need to be called with
861  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
862  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
863  */
864 int ext4_get_blocks_handle(handle_t *handle, struct inode *inode,
865                 ext4_lblk_t iblock, unsigned long maxblocks,
866                 struct buffer_head *bh_result,
867                 int create, int extend_disksize)
868 {
869         int err = -EIO;
870         ext4_lblk_t offsets[4];
871         Indirect chain[4];
872         Indirect *partial;
873         ext4_fsblk_t goal;
874         int indirect_blks;
875         int blocks_to_boundary = 0;
876         int depth;
877         struct ext4_inode_info *ei = EXT4_I(inode);
878         int count = 0;
879         ext4_fsblk_t first_block = 0;
880         loff_t disksize;
881
882
883         J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
884         J_ASSERT(handle != NULL || create == 0);
885         depth = ext4_block_to_path(inode, iblock, offsets,
886                                         &blocks_to_boundary);
887
888         if (depth == 0)
889                 goto out;
890
891         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
892
893         /* Simplest case - block found, no allocation needed */
894         if (!partial) {
895                 first_block = le32_to_cpu(chain[depth - 1].key);
896                 clear_buffer_new(bh_result);
897                 count++;
898                 /*map more blocks*/
899                 while (count < maxblocks && count <= blocks_to_boundary) {
900                         ext4_fsblk_t blk;
901
902                         blk = le32_to_cpu(*(chain[depth-1].p + count));
903
904                         if (blk == first_block + count)
905                                 count++;
906                         else
907                                 break;
908                 }
909                 goto got_it;
910         }
911
912         /* Next simple case - plain lookup or failed read of indirect block */
913         if (!create || err == -EIO)
914                 goto cleanup;
915
916         /*
917          * Okay, we need to do block allocation.  Lazily initialize the block
918          * allocation info here if necessary
919         */
920         if (S_ISREG(inode->i_mode) && (!ei->i_block_alloc_info))
921                 ext4_init_block_alloc_info(inode);
922
923         goal = ext4_find_goal(inode, iblock, partial);
924
925         /* the number of blocks need to allocate for [d,t]indirect blocks */
926         indirect_blks = (chain + depth) - partial - 1;
927
928         /*
929          * Next look up the indirect map to count the totoal number of
930          * direct blocks to allocate for this branch.
931          */
932         count = ext4_blks_to_allocate(partial, indirect_blks,
933                                         maxblocks, blocks_to_boundary);
934         /*
935          * Block out ext4_truncate while we alter the tree
936          */
937         err = ext4_alloc_branch(handle, inode, iblock, indirect_blks,
938                                         &count, goal,
939                                         offsets + (partial - chain), partial);
940
941         /*
942          * The ext4_splice_branch call will free and forget any buffers
943          * on the new chain if there is a failure, but that risks using
944          * up transaction credits, especially for bitmaps where the
945          * credits cannot be returned.  Can we handle this somehow?  We
946          * may need to return -EAGAIN upwards in the worst case.  --sct
947          */
948         if (!err)
949                 err = ext4_splice_branch(handle, inode, iblock,
950                                         partial, indirect_blks, count);
951         /*
952          * i_disksize growing is protected by i_data_sem.  Don't forget to
953          * protect it if you're about to implement concurrent
954          * ext4_get_block() -bzzz
955         */
956         if (!err && extend_disksize) {
957                 disksize = ((loff_t) iblock + count) << inode->i_blkbits;
958                 if (disksize > i_size_read(inode))
959                         disksize = i_size_read(inode);
960                 if (disksize > ei->i_disksize)
961                         ei->i_disksize = disksize;
962         }
963         if (err)
964                 goto cleanup;
965
966         set_buffer_new(bh_result);
967 got_it:
968         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
969         if (count > blocks_to_boundary)
970                 set_buffer_boundary(bh_result);
971         err = count;
972         /* Clean up and exit */
973         partial = chain + depth - 1;    /* the whole chain */
974 cleanup:
975         while (partial > chain) {
976                 BUFFER_TRACE(partial->bh, "call brelse");
977                 brelse(partial->bh);
978                 partial--;
979         }
980         BUFFER_TRACE(bh_result, "returned");
981 out:
982         return err;
983 }
984
985 /*
986  * Calculate the number of metadata blocks need to reserve
987  * to allocate @blocks for non extent file based file
988  */
989 static int ext4_indirect_calc_metadata_amount(struct inode *inode, int blocks)
990 {
991         int icap = EXT4_ADDR_PER_BLOCK(inode->i_sb);
992         int ind_blks, dind_blks, tind_blks;
993
994         /* number of new indirect blocks needed */
995         ind_blks = (blocks + icap - 1) / icap;
996
997         dind_blks = (ind_blks + icap - 1) / icap;
998
999         tind_blks = 1;
1000
1001         return ind_blks + dind_blks + tind_blks;
1002 }
1003
1004 /*
1005  * Calculate the number of metadata blocks need to reserve
1006  * to allocate given number of blocks
1007  */
1008 static int ext4_calc_metadata_amount(struct inode *inode, int blocks)
1009 {
1010         if (!blocks)
1011                 return 0;
1012
1013         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
1014                 return ext4_ext_calc_metadata_amount(inode, blocks);
1015
1016         return ext4_indirect_calc_metadata_amount(inode, blocks);
1017 }
1018
1019 static void ext4_da_update_reserve_space(struct inode *inode, int used)
1020 {
1021         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1022         int total, mdb, mdb_free;
1023
1024         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1025         /* recalculate the number of metablocks still need to be reserved */
1026         total = EXT4_I(inode)->i_reserved_data_blocks - used;
1027         mdb = ext4_calc_metadata_amount(inode, total);
1028
1029         /* figure out how many metablocks to release */
1030         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1031         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1032
1033         /* Account for allocated meta_blocks */
1034         mdb_free -= EXT4_I(inode)->i_allocated_meta_blocks;
1035
1036         /* update fs free blocks counter for truncate case */
1037         percpu_counter_add(&sbi->s_freeblocks_counter, mdb_free);
1038
1039         /* update per-inode reservations */
1040         BUG_ON(used  > EXT4_I(inode)->i_reserved_data_blocks);
1041         EXT4_I(inode)->i_reserved_data_blocks -= used;
1042
1043         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1044         EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1045         EXT4_I(inode)->i_allocated_meta_blocks = 0;
1046         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1047 }
1048
1049 /*
1050  * The ext4_get_blocks_wrap() function try to look up the requested blocks,
1051  * and returns if the blocks are already mapped.
1052  *
1053  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1054  * and store the allocated blocks in the result buffer head and mark it
1055  * mapped.
1056  *
1057  * If file type is extents based, it will call ext4_ext_get_blocks(),
1058  * Otherwise, call with ext4_get_blocks_handle() to handle indirect mapping
1059  * based files
1060  *
1061  * On success, it returns the number of blocks being mapped or allocate.
1062  * if create==0 and the blocks are pre-allocated and uninitialized block,
1063  * the result buffer head is unmapped. If the create ==1, it will make sure
1064  * the buffer head is mapped.
1065  *
1066  * It returns 0 if plain look up failed (blocks have not been allocated), in
1067  * that casem, buffer head is unmapped
1068  *
1069  * It returns the error in case of allocation failure.
1070  */
1071 int ext4_get_blocks_wrap(handle_t *handle, struct inode *inode, sector_t block,
1072                         unsigned long max_blocks, struct buffer_head *bh,
1073                         int create, int extend_disksize, int flag)
1074 {
1075         int retval;
1076
1077         clear_buffer_mapped(bh);
1078
1079         /*
1080          * Try to see if we can get  the block without requesting
1081          * for new file system block.
1082          */
1083         down_read((&EXT4_I(inode)->i_data_sem));
1084         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1085                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1086                                 bh, 0, 0);
1087         } else {
1088                 retval = ext4_get_blocks_handle(handle,
1089                                 inode, block, max_blocks, bh, 0, 0);
1090         }
1091         up_read((&EXT4_I(inode)->i_data_sem));
1092
1093         /* If it is only a block(s) look up */
1094         if (!create)
1095                 return retval;
1096
1097         /*
1098          * Returns if the blocks have already allocated
1099          *
1100          * Note that if blocks have been preallocated
1101          * ext4_ext_get_block() returns th create = 0
1102          * with buffer head unmapped.
1103          */
1104         if (retval > 0 && buffer_mapped(bh))
1105                 return retval;
1106
1107         /*
1108          * New blocks allocate and/or writing to uninitialized extent
1109          * will possibly result in updating i_data, so we take
1110          * the write lock of i_data_sem, and call get_blocks()
1111          * with create == 1 flag.
1112          */
1113         down_write((&EXT4_I(inode)->i_data_sem));
1114
1115         /*
1116          * if the caller is from delayed allocation writeout path
1117          * we have already reserved fs blocks for allocation
1118          * let the underlying get_block() function know to
1119          * avoid double accounting
1120          */
1121         if (flag)
1122                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1123         /*
1124          * We need to check for EXT4 here because migrate
1125          * could have changed the inode type in between
1126          */
1127         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1128                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1129                                 bh, create, extend_disksize);
1130         } else {
1131                 retval = ext4_get_blocks_handle(handle, inode, block,
1132                                 max_blocks, bh, create, extend_disksize);
1133
1134                 if (retval > 0 && buffer_new(bh)) {
1135                         /*
1136                          * We allocated new blocks which will result in
1137                          * i_data's format changing.  Force the migrate
1138                          * to fail by clearing migrate flags
1139                          */
1140                         EXT4_I(inode)->i_flags = EXT4_I(inode)->i_flags &
1141                                                         ~EXT4_EXT_MIGRATE;
1142                 }
1143         }
1144
1145         if (flag) {
1146                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1147                 /*
1148                  * Update reserved blocks/metadata blocks
1149                  * after successful block allocation
1150                  * which were deferred till now
1151                  */
1152                 if ((retval > 0) && buffer_delay(bh))
1153                         ext4_da_update_reserve_space(inode, retval);
1154         }
1155
1156         up_write((&EXT4_I(inode)->i_data_sem));
1157         return retval;
1158 }
1159
1160 /* Maximum number of blocks we map for direct IO at once. */
1161 #define DIO_MAX_BLOCKS 4096
1162
1163 static int ext4_get_block(struct inode *inode, sector_t iblock,
1164                         struct buffer_head *bh_result, int create)
1165 {
1166         handle_t *handle = ext4_journal_current_handle();
1167         int ret = 0, started = 0;
1168         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1169         int dio_credits;
1170
1171         if (create && !handle) {
1172                 /* Direct IO write... */
1173                 if (max_blocks > DIO_MAX_BLOCKS)
1174                         max_blocks = DIO_MAX_BLOCKS;
1175                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
1176                 handle = ext4_journal_start(inode, dio_credits);
1177                 if (IS_ERR(handle)) {
1178                         ret = PTR_ERR(handle);
1179                         goto out;
1180                 }
1181                 started = 1;
1182         }
1183
1184         ret = ext4_get_blocks_wrap(handle, inode, iblock,
1185                                         max_blocks, bh_result, create, 0, 0);
1186         if (ret > 0) {
1187                 bh_result->b_size = (ret << inode->i_blkbits);
1188                 ret = 0;
1189         }
1190         if (started)
1191                 ext4_journal_stop(handle);
1192 out:
1193         return ret;
1194 }
1195
1196 /*
1197  * `handle' can be NULL if create is zero
1198  */
1199 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1200                                 ext4_lblk_t block, int create, int *errp)
1201 {
1202         struct buffer_head dummy;
1203         int fatal = 0, err;
1204
1205         J_ASSERT(handle != NULL || create == 0);
1206
1207         dummy.b_state = 0;
1208         dummy.b_blocknr = -1000;
1209         buffer_trace_init(&dummy.b_history);
1210         err = ext4_get_blocks_wrap(handle, inode, block, 1,
1211                                         &dummy, create, 1, 0);
1212         /*
1213          * ext4_get_blocks_handle() returns number of blocks
1214          * mapped. 0 in case of a HOLE.
1215          */
1216         if (err > 0) {
1217                 if (err > 1)
1218                         WARN_ON(1);
1219                 err = 0;
1220         }
1221         *errp = err;
1222         if (!err && buffer_mapped(&dummy)) {
1223                 struct buffer_head *bh;
1224                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1225                 if (!bh) {
1226                         *errp = -EIO;
1227                         goto err;
1228                 }
1229                 if (buffer_new(&dummy)) {
1230                         J_ASSERT(create != 0);
1231                         J_ASSERT(handle != NULL);
1232
1233                         /*
1234                          * Now that we do not always journal data, we should
1235                          * keep in mind whether this should always journal the
1236                          * new buffer as metadata.  For now, regular file
1237                          * writes use ext4_get_block instead, so it's not a
1238                          * problem.
1239                          */
1240                         lock_buffer(bh);
1241                         BUFFER_TRACE(bh, "call get_create_access");
1242                         fatal = ext4_journal_get_create_access(handle, bh);
1243                         if (!fatal && !buffer_uptodate(bh)) {
1244                                 memset(bh->b_data,0,inode->i_sb->s_blocksize);
1245                                 set_buffer_uptodate(bh);
1246                         }
1247                         unlock_buffer(bh);
1248                         BUFFER_TRACE(bh, "call ext4_journal_dirty_metadata");
1249                         err = ext4_journal_dirty_metadata(handle, bh);
1250                         if (!fatal)
1251                                 fatal = err;
1252                 } else {
1253                         BUFFER_TRACE(bh, "not a new buffer");
1254                 }
1255                 if (fatal) {
1256                         *errp = fatal;
1257                         brelse(bh);
1258                         bh = NULL;
1259                 }
1260                 return bh;
1261         }
1262 err:
1263         return NULL;
1264 }
1265
1266 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1267                                ext4_lblk_t block, int create, int *err)
1268 {
1269         struct buffer_head * bh;
1270
1271         bh = ext4_getblk(handle, inode, block, create, err);
1272         if (!bh)
1273                 return bh;
1274         if (buffer_uptodate(bh))
1275                 return bh;
1276         ll_rw_block(READ_META, 1, &bh);
1277         wait_on_buffer(bh);
1278         if (buffer_uptodate(bh))
1279                 return bh;
1280         put_bh(bh);
1281         *err = -EIO;
1282         return NULL;
1283 }
1284
1285 static int walk_page_buffers(   handle_t *handle,
1286                                 struct buffer_head *head,
1287                                 unsigned from,
1288                                 unsigned to,
1289                                 int *partial,
1290                                 int (*fn)(      handle_t *handle,
1291                                                 struct buffer_head *bh))
1292 {
1293         struct buffer_head *bh;
1294         unsigned block_start, block_end;
1295         unsigned blocksize = head->b_size;
1296         int err, ret = 0;
1297         struct buffer_head *next;
1298
1299         for (   bh = head, block_start = 0;
1300                 ret == 0 && (bh != head || !block_start);
1301                 block_start = block_end, bh = next)
1302         {
1303                 next = bh->b_this_page;
1304                 block_end = block_start + blocksize;
1305                 if (block_end <= from || block_start >= to) {
1306                         if (partial && !buffer_uptodate(bh))
1307                                 *partial = 1;
1308                         continue;
1309                 }
1310                 err = (*fn)(handle, bh);
1311                 if (!ret)
1312                         ret = err;
1313         }
1314         return ret;
1315 }
1316
1317 /*
1318  * To preserve ordering, it is essential that the hole instantiation and
1319  * the data write be encapsulated in a single transaction.  We cannot
1320  * close off a transaction and start a new one between the ext4_get_block()
1321  * and the commit_write().  So doing the jbd2_journal_start at the start of
1322  * prepare_write() is the right place.
1323  *
1324  * Also, this function can nest inside ext4_writepage() ->
1325  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1326  * has generated enough buffer credits to do the whole page.  So we won't
1327  * block on the journal in that case, which is good, because the caller may
1328  * be PF_MEMALLOC.
1329  *
1330  * By accident, ext4 can be reentered when a transaction is open via
1331  * quota file writes.  If we were to commit the transaction while thus
1332  * reentered, there can be a deadlock - we would be holding a quota
1333  * lock, and the commit would never complete if another thread had a
1334  * transaction open and was blocking on the quota lock - a ranking
1335  * violation.
1336  *
1337  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1338  * will _not_ run commit under these circumstances because handle->h_ref
1339  * is elevated.  We'll still have enough credits for the tiny quotafile
1340  * write.
1341  */
1342 static int do_journal_get_write_access(handle_t *handle,
1343                                         struct buffer_head *bh)
1344 {
1345         if (!buffer_mapped(bh) || buffer_freed(bh))
1346                 return 0;
1347         return ext4_journal_get_write_access(handle, bh);
1348 }
1349
1350 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1351                                 loff_t pos, unsigned len, unsigned flags,
1352                                 struct page **pagep, void **fsdata)
1353 {
1354         struct inode *inode = mapping->host;
1355         int ret, needed_blocks = ext4_writepage_trans_blocks(inode);
1356         handle_t *handle;
1357         int retries = 0;
1358         struct page *page;
1359         pgoff_t index;
1360         unsigned from, to;
1361
1362         index = pos >> PAGE_CACHE_SHIFT;
1363         from = pos & (PAGE_CACHE_SIZE - 1);
1364         to = from + len;
1365
1366 retry:
1367         handle = ext4_journal_start(inode, needed_blocks);
1368         if (IS_ERR(handle)) {
1369                 ret = PTR_ERR(handle);
1370                 goto out;
1371         }
1372
1373         page = __grab_cache_page(mapping, index);
1374         if (!page) {
1375                 ext4_journal_stop(handle);
1376                 ret = -ENOMEM;
1377                 goto out;
1378         }
1379         *pagep = page;
1380
1381         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1382                                                         ext4_get_block);
1383
1384         if (!ret && ext4_should_journal_data(inode)) {
1385                 ret = walk_page_buffers(handle, page_buffers(page),
1386                                 from, to, NULL, do_journal_get_write_access);
1387         }
1388
1389         if (ret) {
1390                 unlock_page(page);
1391                 ext4_journal_stop(handle);
1392                 page_cache_release(page);
1393         }
1394
1395         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1396                 goto retry;
1397 out:
1398         return ret;
1399 }
1400
1401 /* For write_end() in data=journal mode */
1402 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1403 {
1404         if (!buffer_mapped(bh) || buffer_freed(bh))
1405                 return 0;
1406         set_buffer_uptodate(bh);
1407         return ext4_journal_dirty_metadata(handle, bh);
1408 }
1409
1410 /*
1411  * We need to pick up the new inode size which generic_commit_write gave us
1412  * `file' can be NULL - eg, when called from page_symlink().
1413  *
1414  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1415  * buffers are managed internally.
1416  */
1417 static int ext4_ordered_write_end(struct file *file,
1418                                 struct address_space *mapping,
1419                                 loff_t pos, unsigned len, unsigned copied,
1420                                 struct page *page, void *fsdata)
1421 {
1422         handle_t *handle = ext4_journal_current_handle();
1423         struct inode *inode = mapping->host;
1424         int ret = 0, ret2;
1425
1426         ret = ext4_jbd2_file_inode(handle, inode);
1427
1428         if (ret == 0) {
1429                 /*
1430                  * generic_write_end() will run mark_inode_dirty() if i_size
1431                  * changes.  So let's piggyback the i_disksize mark_inode_dirty
1432                  * into that.
1433                  */
1434                 loff_t new_i_size;
1435
1436                 new_i_size = pos + copied;
1437                 if (new_i_size > EXT4_I(inode)->i_disksize)
1438                         EXT4_I(inode)->i_disksize = new_i_size;
1439                 ret2 = generic_write_end(file, mapping, pos, len, copied,
1440                                                         page, fsdata);
1441                 copied = ret2;
1442                 if (ret2 < 0)
1443                         ret = ret2;
1444         }
1445         ret2 = ext4_journal_stop(handle);
1446         if (!ret)
1447                 ret = ret2;
1448
1449         return ret ? ret : copied;
1450 }
1451
1452 static int ext4_writeback_write_end(struct file *file,
1453                                 struct address_space *mapping,
1454                                 loff_t pos, unsigned len, unsigned copied,
1455                                 struct page *page, void *fsdata)
1456 {
1457         handle_t *handle = ext4_journal_current_handle();
1458         struct inode *inode = mapping->host;
1459         int ret = 0, ret2;
1460         loff_t new_i_size;
1461
1462         new_i_size = pos + copied;
1463         if (new_i_size > EXT4_I(inode)->i_disksize)
1464                 EXT4_I(inode)->i_disksize = new_i_size;
1465
1466         ret2 = generic_write_end(file, mapping, pos, len, copied,
1467                                                         page, fsdata);
1468         copied = ret2;
1469         if (ret2 < 0)
1470                 ret = ret2;
1471
1472         ret2 = ext4_journal_stop(handle);
1473         if (!ret)
1474                 ret = ret2;
1475
1476         return ret ? ret : copied;
1477 }
1478
1479 static int ext4_journalled_write_end(struct file *file,
1480                                 struct address_space *mapping,
1481                                 loff_t pos, unsigned len, unsigned copied,
1482                                 struct page *page, void *fsdata)
1483 {
1484         handle_t *handle = ext4_journal_current_handle();
1485         struct inode *inode = mapping->host;
1486         int ret = 0, ret2;
1487         int partial = 0;
1488         unsigned from, to;
1489
1490         from = pos & (PAGE_CACHE_SIZE - 1);
1491         to = from + len;
1492
1493         if (copied < len) {
1494                 if (!PageUptodate(page))
1495                         copied = 0;
1496                 page_zero_new_buffers(page, from+copied, to);
1497         }
1498
1499         ret = walk_page_buffers(handle, page_buffers(page), from,
1500                                 to, &partial, write_end_fn);
1501         if (!partial)
1502                 SetPageUptodate(page);
1503         if (pos+copied > inode->i_size)
1504                 i_size_write(inode, pos+copied);
1505         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
1506         if (inode->i_size > EXT4_I(inode)->i_disksize) {
1507                 EXT4_I(inode)->i_disksize = inode->i_size;
1508                 ret2 = ext4_mark_inode_dirty(handle, inode);
1509                 if (!ret)
1510                         ret = ret2;
1511         }
1512
1513         unlock_page(page);
1514         ret2 = ext4_journal_stop(handle);
1515         if (!ret)
1516                 ret = ret2;
1517         page_cache_release(page);
1518
1519         return ret ? ret : copied;
1520 }
1521
1522 static int ext4_da_reserve_space(struct inode *inode, int nrblocks)
1523 {
1524        struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1525        unsigned long md_needed, mdblocks, total = 0;
1526
1527         /*
1528          * recalculate the amount of metadata blocks to reserve
1529          * in order to allocate nrblocks
1530          * worse case is one extent per block
1531          */
1532         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1533         total = EXT4_I(inode)->i_reserved_data_blocks + nrblocks;
1534         mdblocks = ext4_calc_metadata_amount(inode, total);
1535         BUG_ON(mdblocks < EXT4_I(inode)->i_reserved_meta_blocks);
1536
1537         md_needed = mdblocks - EXT4_I(inode)->i_reserved_meta_blocks;
1538         total = md_needed + nrblocks;
1539
1540         if (ext4_has_free_blocks(sbi, total) < total) {
1541                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1542                 return -ENOSPC;
1543         }
1544         /* reduce fs free blocks counter */
1545         percpu_counter_sub(&sbi->s_freeblocks_counter, total);
1546
1547         EXT4_I(inode)->i_reserved_data_blocks += nrblocks;
1548         EXT4_I(inode)->i_reserved_meta_blocks = mdblocks;
1549
1550         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1551         return 0;       /* success */
1552 }
1553
1554 static void ext4_da_release_space(struct inode *inode, int to_free)
1555 {
1556         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1557         int total, mdb, mdb_free, release;
1558
1559         if (!to_free)
1560                 return;         /* Nothing to release, exit */
1561
1562         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1563
1564         if (!EXT4_I(inode)->i_reserved_data_blocks) {
1565                 /*
1566                  * if there is no reserved blocks, but we try to free some
1567                  * then the counter is messed up somewhere.
1568                  * but since this function is called from invalidate
1569                  * page, it's harmless to return without any action
1570                  */
1571                 printk(KERN_INFO "ext4 delalloc try to release %d reserved "
1572                             "blocks for inode %lu, but there is no reserved "
1573                             "data blocks\n", to_free, inode->i_ino);
1574                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1575                 return;
1576         }
1577
1578         /* recalculate the number of metablocks still need to be reserved */
1579         total = EXT4_I(inode)->i_reserved_data_blocks - to_free;
1580         mdb = ext4_calc_metadata_amount(inode, total);
1581
1582         /* figure out how many metablocks to release */
1583         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1584         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1585
1586         release = to_free + mdb_free;
1587
1588         /* update fs free blocks counter for truncate case */
1589         percpu_counter_add(&sbi->s_freeblocks_counter, release);
1590
1591         /* update per-inode reservations */
1592         BUG_ON(to_free > EXT4_I(inode)->i_reserved_data_blocks);
1593         EXT4_I(inode)->i_reserved_data_blocks -= to_free;
1594
1595         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1596         EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1597         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1598 }
1599
1600 static void ext4_da_page_release_reservation(struct page *page,
1601                                                 unsigned long offset)
1602 {
1603         int to_release = 0;
1604         struct buffer_head *head, *bh;
1605         unsigned int curr_off = 0;
1606
1607         head = page_buffers(page);
1608         bh = head;
1609         do {
1610                 unsigned int next_off = curr_off + bh->b_size;
1611
1612                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1613                         to_release++;
1614                         clear_buffer_delay(bh);
1615                 }
1616                 curr_off = next_off;
1617         } while ((bh = bh->b_this_page) != head);
1618         ext4_da_release_space(page->mapping->host, to_release);
1619 }
1620
1621 /*
1622  * Delayed allocation stuff
1623  */
1624
1625 struct mpage_da_data {
1626         struct inode *inode;
1627         struct buffer_head lbh;                 /* extent of blocks */
1628         unsigned long first_page, next_page;    /* extent of pages */
1629         get_block_t *get_block;
1630         struct writeback_control *wbc;
1631         int io_done;
1632         long pages_written;
1633 };
1634
1635 /*
1636  * mpage_da_submit_io - walks through extent of pages and try to write
1637  * them with writepage() call back
1638  *
1639  * @mpd->inode: inode
1640  * @mpd->first_page: first page of the extent
1641  * @mpd->next_page: page after the last page of the extent
1642  * @mpd->get_block: the filesystem's block mapper function
1643  *
1644  * By the time mpage_da_submit_io() is called we expect all blocks
1645  * to be allocated. this may be wrong if allocation failed.
1646  *
1647  * As pages are already locked by write_cache_pages(), we can't use it
1648  */
1649 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1650 {
1651         struct address_space *mapping = mpd->inode->i_mapping;
1652         int ret = 0, err, nr_pages, i;
1653         unsigned long index, end;
1654         struct pagevec pvec;
1655
1656         BUG_ON(mpd->next_page <= mpd->first_page);
1657         pagevec_init(&pvec, 0);
1658         index = mpd->first_page;
1659         end = mpd->next_page - 1;
1660
1661         while (index <= end) {
1662                 /* XXX: optimize tail */
1663                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1664                 if (nr_pages == 0)
1665                         break;
1666                 for (i = 0; i < nr_pages; i++) {
1667                         struct page *page = pvec.pages[i];
1668
1669                         index = page->index;
1670                         if (index > end)
1671                                 break;
1672                         index++;
1673
1674                         err = mapping->a_ops->writepage(page, mpd->wbc);
1675                         if (!err)
1676                                 mpd->pages_written++;
1677                         /*
1678                          * In error case, we have to continue because
1679                          * remaining pages are still locked
1680                          * XXX: unlock and re-dirty them?
1681                          */
1682                         if (ret == 0)
1683                                 ret = err;
1684                 }
1685                 pagevec_release(&pvec);
1686         }
1687         return ret;
1688 }
1689
1690 /*
1691  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
1692  *
1693  * @mpd->inode - inode to walk through
1694  * @exbh->b_blocknr - first block on a disk
1695  * @exbh->b_size - amount of space in bytes
1696  * @logical - first logical block to start assignment with
1697  *
1698  * the function goes through all passed space and put actual disk
1699  * block numbers into buffer heads, dropping BH_Delay
1700  */
1701 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
1702                                  struct buffer_head *exbh)
1703 {
1704         struct inode *inode = mpd->inode;
1705         struct address_space *mapping = inode->i_mapping;
1706         int blocks = exbh->b_size >> inode->i_blkbits;
1707         sector_t pblock = exbh->b_blocknr, cur_logical;
1708         struct buffer_head *head, *bh;
1709         pgoff_t index, end;
1710         struct pagevec pvec;
1711         int nr_pages, i;
1712
1713         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1714         end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1715         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1716
1717         pagevec_init(&pvec, 0);
1718
1719         while (index <= end) {
1720                 /* XXX: optimize tail */
1721                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1722                 if (nr_pages == 0)
1723                         break;
1724                 for (i = 0; i < nr_pages; i++) {
1725                         struct page *page = pvec.pages[i];
1726
1727                         index = page->index;
1728                         if (index > end)
1729                                 break;
1730                         index++;
1731
1732                         BUG_ON(!PageLocked(page));
1733                         BUG_ON(PageWriteback(page));
1734                         BUG_ON(!page_has_buffers(page));
1735
1736                         bh = page_buffers(page);
1737                         head = bh;
1738
1739                         /* skip blocks out of the range */
1740                         do {
1741                                 if (cur_logical >= logical)
1742                                         break;
1743                                 cur_logical++;
1744                         } while ((bh = bh->b_this_page) != head);
1745
1746                         do {
1747                                 if (cur_logical >= logical + blocks)
1748                                         break;
1749                                 if (buffer_delay(bh)) {
1750                                         bh->b_blocknr = pblock;
1751                                         clear_buffer_delay(bh);
1752                                         bh->b_bdev = inode->i_sb->s_bdev;
1753                                 } else if (buffer_unwritten(bh)) {
1754                                         bh->b_blocknr = pblock;
1755                                         clear_buffer_unwritten(bh);
1756                                         set_buffer_mapped(bh);
1757                                         set_buffer_new(bh);
1758                                         bh->b_bdev = inode->i_sb->s_bdev;
1759                                 } else if (buffer_mapped(bh))
1760                                         BUG_ON(bh->b_blocknr != pblock);
1761
1762                                 cur_logical++;
1763                                 pblock++;
1764                         } while ((bh = bh->b_this_page) != head);
1765                 }
1766                 pagevec_release(&pvec);
1767         }
1768 }
1769
1770
1771 /*
1772  * __unmap_underlying_blocks - just a helper function to unmap
1773  * set of blocks described by @bh
1774  */
1775 static inline void __unmap_underlying_blocks(struct inode *inode,
1776                                              struct buffer_head *bh)
1777 {
1778         struct block_device *bdev = inode->i_sb->s_bdev;
1779         int blocks, i;
1780
1781         blocks = bh->b_size >> inode->i_blkbits;
1782         for (i = 0; i < blocks; i++)
1783                 unmap_underlying_metadata(bdev, bh->b_blocknr + i);
1784 }
1785
1786 /*
1787  * mpage_da_map_blocks - go through given space
1788  *
1789  * @mpd->lbh - bh describing space
1790  * @mpd->get_block - the filesystem's block mapper function
1791  *
1792  * The function skips space we know is already mapped to disk blocks.
1793  *
1794  */
1795 static void mpage_da_map_blocks(struct mpage_da_data *mpd)
1796 {
1797         int err = 0;
1798         struct buffer_head *lbh = &mpd->lbh;
1799         sector_t next = lbh->b_blocknr;
1800         struct buffer_head new;
1801
1802         /*
1803          * We consider only non-mapped and non-allocated blocks
1804          */
1805         if (buffer_mapped(lbh) && !buffer_delay(lbh))
1806                 return;
1807
1808         new.b_state = lbh->b_state;
1809         new.b_blocknr = 0;
1810         new.b_size = lbh->b_size;
1811
1812         /*
1813          * If we didn't accumulate anything
1814          * to write simply return
1815          */
1816         if (!new.b_size)
1817                 return;
1818         err = mpd->get_block(mpd->inode, next, &new, 1);
1819         if (err)
1820                 return;
1821         BUG_ON(new.b_size == 0);
1822
1823         if (buffer_new(&new))
1824                 __unmap_underlying_blocks(mpd->inode, &new);
1825
1826         /*
1827          * If blocks are delayed marked, we need to
1828          * put actual blocknr and drop delayed bit
1829          */
1830         if (buffer_delay(lbh) || buffer_unwritten(lbh))
1831                 mpage_put_bnr_to_bhs(mpd, next, &new);
1832
1833         return;
1834 }
1835
1836 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
1837                 (1 << BH_Delay) | (1 << BH_Unwritten))
1838
1839 /*
1840  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
1841  *
1842  * @mpd->lbh - extent of blocks
1843  * @logical - logical number of the block in the file
1844  * @bh - bh of the block (used to access block's state)
1845  *
1846  * the function is used to collect contig. blocks in same state
1847  */
1848 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
1849                                    sector_t logical, struct buffer_head *bh)
1850 {
1851         sector_t next;
1852         size_t b_size = bh->b_size;
1853         struct buffer_head *lbh = &mpd->lbh;
1854         int nrblocks = lbh->b_size >> mpd->inode->i_blkbits;
1855
1856         /* check if thereserved journal credits might overflow */
1857         if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
1858                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
1859                         /*
1860                          * With non-extent format we are limited by the journal
1861                          * credit available.  Total credit needed to insert
1862                          * nrblocks contiguous blocks is dependent on the
1863                          * nrblocks.  So limit nrblocks.
1864                          */
1865                         goto flush_it;
1866                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
1867                                 EXT4_MAX_TRANS_DATA) {
1868                         /*
1869                          * Adding the new buffer_head would make it cross the
1870                          * allowed limit for which we have journal credit
1871                          * reserved. So limit the new bh->b_size
1872                          */
1873                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
1874                                                 mpd->inode->i_blkbits;
1875                         /* we will do mpage_da_submit_io in the next loop */
1876                 }
1877         }
1878         /*
1879          * First block in the extent
1880          */
1881         if (lbh->b_size == 0) {
1882                 lbh->b_blocknr = logical;
1883                 lbh->b_size = b_size;
1884                 lbh->b_state = bh->b_state & BH_FLAGS;
1885                 return;
1886         }
1887
1888         next = lbh->b_blocknr + nrblocks;
1889         /*
1890          * Can we merge the block to our big extent?
1891          */
1892         if (logical == next && (bh->b_state & BH_FLAGS) == lbh->b_state) {
1893                 lbh->b_size += b_size;
1894                 return;
1895         }
1896
1897 flush_it:
1898         /*
1899          * We couldn't merge the block to our extent, so we
1900          * need to flush current  extent and start new one
1901          */
1902         mpage_da_map_blocks(mpd);
1903         mpage_da_submit_io(mpd);
1904         mpd->io_done = 1;
1905         return;
1906 }
1907
1908 /*
1909  * __mpage_da_writepage - finds extent of pages and blocks
1910  *
1911  * @page: page to consider
1912  * @wbc: not used, we just follow rules
1913  * @data: context
1914  *
1915  * The function finds extents of pages and scan them for all blocks.
1916  */
1917 static int __mpage_da_writepage(struct page *page,
1918                                 struct writeback_control *wbc, void *data)
1919 {
1920         struct mpage_da_data *mpd = data;
1921         struct inode *inode = mpd->inode;
1922         struct buffer_head *bh, *head, fake;
1923         sector_t logical;
1924
1925         if (mpd->io_done) {
1926                 /*
1927                  * Rest of the page in the page_vec
1928                  * redirty then and skip then. We will
1929                  * try to to write them again after
1930                  * starting a new transaction
1931                  */
1932                 redirty_page_for_writepage(wbc, page);
1933                 unlock_page(page);
1934                 return MPAGE_DA_EXTENT_TAIL;
1935         }
1936         /*
1937          * Can we merge this page to current extent?
1938          */
1939         if (mpd->next_page != page->index) {
1940                 /*
1941                  * Nope, we can't. So, we map non-allocated blocks
1942                  * and start IO on them using writepage()
1943                  */
1944                 if (mpd->next_page != mpd->first_page) {
1945                         mpage_da_map_blocks(mpd);
1946                         mpage_da_submit_io(mpd);
1947                         /*
1948                          * skip rest of the page in the page_vec
1949                          */
1950                         mpd->io_done = 1;
1951                         redirty_page_for_writepage(wbc, page);
1952                         unlock_page(page);
1953                         return MPAGE_DA_EXTENT_TAIL;
1954                 }
1955
1956                 /*
1957                  * Start next extent of pages ...
1958                  */
1959                 mpd->first_page = page->index;
1960
1961                 /*
1962                  * ... and blocks
1963                  */
1964                 mpd->lbh.b_size = 0;
1965                 mpd->lbh.b_state = 0;
1966                 mpd->lbh.b_blocknr = 0;
1967         }
1968
1969         mpd->next_page = page->index + 1;
1970         logical = (sector_t) page->index <<
1971                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
1972
1973         if (!page_has_buffers(page)) {
1974                 /*
1975                  * There is no attached buffer heads yet (mmap?)
1976                  * we treat the page asfull of dirty blocks
1977                  */
1978                 bh = &fake;
1979                 bh->b_size = PAGE_CACHE_SIZE;
1980                 bh->b_state = 0;
1981                 set_buffer_dirty(bh);
1982                 set_buffer_uptodate(bh);
1983                 mpage_add_bh_to_extent(mpd, logical, bh);
1984                 if (mpd->io_done)
1985                         return MPAGE_DA_EXTENT_TAIL;
1986         } else {
1987                 /*
1988                  * Page with regular buffer heads, just add all dirty ones
1989                  */
1990                 head = page_buffers(page);
1991                 bh = head;
1992                 do {
1993                         BUG_ON(buffer_locked(bh));
1994                         if (buffer_dirty(bh) &&
1995                                 (!buffer_mapped(bh) || buffer_delay(bh))) {
1996                                 mpage_add_bh_to_extent(mpd, logical, bh);
1997                                 if (mpd->io_done)
1998                                         return MPAGE_DA_EXTENT_TAIL;
1999                         }
2000                         logical++;
2001                 } while ((bh = bh->b_this_page) != head);
2002         }
2003
2004         return 0;
2005 }
2006
2007 /*
2008  * mpage_da_writepages - walk the list of dirty pages of the given
2009  * address space, allocates non-allocated blocks, maps newly-allocated
2010  * blocks to existing bhs and issue IO them
2011  *
2012  * @mapping: address space structure to write
2013  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2014  * @get_block: the filesystem's block mapper function.
2015  *
2016  * This is a library function, which implements the writepages()
2017  * address_space_operation.
2018  */
2019 static int mpage_da_writepages(struct address_space *mapping,
2020                                struct writeback_control *wbc,
2021                                get_block_t get_block)
2022 {
2023         struct mpage_da_data mpd;
2024         long to_write;
2025         int ret;
2026
2027         if (!get_block)
2028                 return generic_writepages(mapping, wbc);
2029
2030         mpd.wbc = wbc;
2031         mpd.inode = mapping->host;
2032         mpd.lbh.b_size = 0;
2033         mpd.lbh.b_state = 0;
2034         mpd.lbh.b_blocknr = 0;
2035         mpd.first_page = 0;
2036         mpd.next_page = 0;
2037         mpd.get_block = get_block;
2038         mpd.io_done = 0;
2039         mpd.pages_written = 0;
2040
2041         to_write = wbc->nr_to_write;
2042
2043         ret = write_cache_pages(mapping, wbc, __mpage_da_writepage, &mpd);
2044
2045         /*
2046          * Handle last extent of pages
2047          */
2048         if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2049                 mpage_da_map_blocks(&mpd);
2050                 mpage_da_submit_io(&mpd);
2051         }
2052
2053         wbc->nr_to_write = to_write - mpd.pages_written;
2054         return ret;
2055 }
2056
2057 /*
2058  * this is a special callback for ->write_begin() only
2059  * it's intention is to return mapped block or reserve space
2060  */
2061 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2062                                   struct buffer_head *bh_result, int create)
2063 {
2064         int ret = 0;
2065
2066         BUG_ON(create == 0);
2067         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2068
2069         /*
2070          * first, we need to know whether the block is allocated already
2071          * preallocated blocks are unmapped but should treated
2072          * the same as allocated blocks.
2073          */
2074         ret = ext4_get_blocks_wrap(NULL, inode, iblock, 1,  bh_result, 0, 0, 0);
2075         if ((ret == 0) && !buffer_delay(bh_result)) {
2076                 /* the block isn't (pre)allocated yet, let's reserve space */
2077                 /*
2078                  * XXX: __block_prepare_write() unmaps passed block,
2079                  * is it OK?
2080                  */
2081                 ret = ext4_da_reserve_space(inode, 1);
2082                 if (ret)
2083                         /* not enough space to reserve */
2084                         return ret;
2085
2086                 map_bh(bh_result, inode->i_sb, 0);
2087                 set_buffer_new(bh_result);
2088                 set_buffer_delay(bh_result);
2089         } else if (ret > 0) {
2090                 bh_result->b_size = (ret << inode->i_blkbits);
2091                 ret = 0;
2092         }
2093
2094         return ret;
2095 }
2096 #define         EXT4_DELALLOC_RSVED     1
2097 static int ext4_da_get_block_write(struct inode *inode, sector_t iblock,
2098                                    struct buffer_head *bh_result, int create)
2099 {
2100         int ret;
2101         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2102         loff_t disksize = EXT4_I(inode)->i_disksize;
2103         handle_t *handle = NULL;
2104
2105         handle = ext4_journal_current_handle();
2106         if (!handle) {
2107                 ret = ext4_get_blocks_wrap(handle, inode, iblock, max_blocks,
2108                                    bh_result, 0, 0, 0);
2109                 BUG_ON(!ret);
2110         } else {
2111                 ret = ext4_get_blocks_wrap(handle, inode, iblock, max_blocks,
2112                                    bh_result, create, 0, EXT4_DELALLOC_RSVED);
2113         }
2114
2115         if (ret > 0) {
2116                 bh_result->b_size = (ret << inode->i_blkbits);
2117
2118                 /*
2119                  * Update on-disk size along with block allocation
2120                  * we don't use 'extend_disksize' as size may change
2121                  * within already allocated block -bzzz
2122                  */
2123                 disksize = ((loff_t) iblock + ret) << inode->i_blkbits;
2124                 if (disksize > i_size_read(inode))
2125                         disksize = i_size_read(inode);
2126                 if (disksize > EXT4_I(inode)->i_disksize) {
2127                         /*
2128                          * XXX: replace with spinlock if seen contended -bzzz
2129                          */
2130                         down_write(&EXT4_I(inode)->i_data_sem);
2131                         if (disksize > EXT4_I(inode)->i_disksize)
2132                                 EXT4_I(inode)->i_disksize = disksize;
2133                         up_write(&EXT4_I(inode)->i_data_sem);
2134
2135                         if (EXT4_I(inode)->i_disksize == disksize) {
2136                                 ret = ext4_mark_inode_dirty(handle, inode);
2137                                 return ret;
2138                         }
2139                 }
2140                 ret = 0;
2141         }
2142         return ret;
2143 }
2144
2145 static int ext4_bh_unmapped_or_delay(handle_t *handle, struct buffer_head *bh)
2146 {
2147         /*
2148          * unmapped buffer is possible for holes.
2149          * delay buffer is possible with delayed allocation
2150          */
2151         return ((!buffer_mapped(bh) || buffer_delay(bh)) && buffer_dirty(bh));
2152 }
2153
2154 static int ext4_normal_get_block_write(struct inode *inode, sector_t iblock,
2155                                    struct buffer_head *bh_result, int create)
2156 {
2157         int ret = 0;
2158         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2159
2160         /*
2161          * we don't want to do block allocation in writepage
2162          * so call get_block_wrap with create = 0
2163          */
2164         ret = ext4_get_blocks_wrap(NULL, inode, iblock, max_blocks,
2165                                    bh_result, 0, 0, 0);
2166         if (ret > 0) {
2167                 bh_result->b_size = (ret << inode->i_blkbits);
2168                 ret = 0;
2169         }
2170         return ret;
2171 }
2172
2173 /*
2174  * get called vi ext4_da_writepages after taking page lock (have journal handle)
2175  * get called via journal_submit_inode_data_buffers (no journal handle)
2176  * get called via shrink_page_list via pdflush (no journal handle)
2177  * or grab_page_cache when doing write_begin (have journal handle)
2178  */
2179 static int ext4_da_writepage(struct page *page,
2180                                 struct writeback_control *wbc)
2181 {
2182         int ret = 0;
2183         loff_t size;
2184         unsigned long len;
2185         struct buffer_head *page_bufs;
2186         struct inode *inode = page->mapping->host;
2187
2188         size = i_size_read(inode);
2189         if (page->index == size >> PAGE_CACHE_SHIFT)
2190                 len = size & ~PAGE_CACHE_MASK;
2191         else
2192                 len = PAGE_CACHE_SIZE;
2193
2194         if (page_has_buffers(page)) {
2195                 page_bufs = page_buffers(page);
2196                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2197                                         ext4_bh_unmapped_or_delay)) {
2198                         /*
2199                          * We don't want to do  block allocation
2200                          * So redirty the page and return
2201                          * We may reach here when we do a journal commit
2202                          * via journal_submit_inode_data_buffers.
2203                          * If we don't have mapping block we just ignore
2204                          * them. We can also reach here via shrink_page_list
2205                          */
2206                         redirty_page_for_writepage(wbc, page);
2207                         unlock_page(page);
2208                         return 0;
2209                 }
2210         } else {
2211                 /*
2212                  * The test for page_has_buffers() is subtle:
2213                  * We know the page is dirty but it lost buffers. That means
2214                  * that at some moment in time after write_begin()/write_end()
2215                  * has been called all buffers have been clean and thus they
2216                  * must have been written at least once. So they are all
2217                  * mapped and we can happily proceed with mapping them
2218                  * and writing the page.
2219                  *
2220                  * Try to initialize the buffer_heads and check whether
2221                  * all are mapped and non delay. We don't want to
2222                  * do block allocation here.
2223                  */
2224                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
2225                                                 ext4_normal_get_block_write);
2226                 if (!ret) {
2227                         page_bufs = page_buffers(page);
2228                         /* check whether all are mapped and non delay */
2229                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2230                                                 ext4_bh_unmapped_or_delay)) {
2231                                 redirty_page_for_writepage(wbc, page);
2232                                 unlock_page(page);
2233                                 return 0;
2234                         }
2235                 } else {
2236                         /*
2237                          * We can't do block allocation here
2238                          * so just redity the page and unlock
2239                          * and return
2240                          */
2241                         redirty_page_for_writepage(wbc, page);
2242                         unlock_page(page);
2243                         return 0;
2244                 }
2245         }
2246
2247         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2248                 ret = nobh_writepage(page, ext4_normal_get_block_write, wbc);
2249         else
2250                 ret = block_write_full_page(page,
2251                                                 ext4_normal_get_block_write,
2252                                                 wbc);
2253
2254         return ret;
2255 }
2256
2257 /*
2258  * This is called via ext4_da_writepages() to
2259  * calulate the total number of credits to reserve to fit
2260  * a single extent allocation into a single transaction,
2261  * ext4_da_writpeages() will loop calling this before
2262  * the block allocation.
2263  */
2264
2265 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2266 {
2267         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2268
2269         /*
2270          * With non-extent format the journal credit needed to
2271          * insert nrblocks contiguous block is dependent on
2272          * number of contiguous block. So we will limit
2273          * number of contiguous block to a sane value
2274          */
2275         if (!(inode->i_flags & EXT4_EXTENTS_FL) &&
2276             (max_blocks > EXT4_MAX_TRANS_DATA))
2277                 max_blocks = EXT4_MAX_TRANS_DATA;
2278
2279         return ext4_chunk_trans_blocks(inode, max_blocks);
2280 }
2281
2282 static int ext4_da_writepages(struct address_space *mapping,
2283                               struct writeback_control *wbc)
2284 {
2285         handle_t *handle = NULL;
2286         loff_t range_start = 0;
2287         struct inode *inode = mapping->host;
2288         int needed_blocks, ret = 0, nr_to_writebump = 0;
2289         long to_write, pages_skipped = 0;
2290         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2291
2292         /*
2293          * No pages to write? This is mainly a kludge to avoid starting
2294          * a transaction for special inodes like journal inode on last iput()
2295          * because that could violate lock ordering on umount
2296          */
2297         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2298                 return 0;
2299         /*
2300          * Make sure nr_to_write is >= sbi->s_mb_stream_request
2301          * This make sure small files blocks are allocated in
2302          * single attempt. This ensure that small files
2303          * get less fragmented.
2304          */
2305         if (wbc->nr_to_write < sbi->s_mb_stream_request) {
2306                 nr_to_writebump = sbi->s_mb_stream_request - wbc->nr_to_write;
2307                 wbc->nr_to_write = sbi->s_mb_stream_request;
2308         }
2309
2310         if (!wbc->range_cyclic)
2311                 /*
2312                  * If range_cyclic is not set force range_cont
2313                  * and save the old writeback_index
2314                  */
2315                 wbc->range_cont = 1;
2316
2317         range_start =  wbc->range_start;
2318         pages_skipped = wbc->pages_skipped;
2319
2320 restart_loop:
2321         to_write = wbc->nr_to_write;
2322         while (!ret && to_write > 0) {
2323
2324                 /*
2325                  * we  insert one extent at a time. So we need
2326                  * credit needed for single extent allocation.
2327                  * journalled mode is currently not supported
2328                  * by delalloc
2329                  */
2330                 BUG_ON(ext4_should_journal_data(inode));
2331                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2332
2333                 /* start a new transaction*/
2334                 handle = ext4_journal_start(inode, needed_blocks);
2335                 if (IS_ERR(handle)) {
2336                         ret = PTR_ERR(handle);
2337                         printk(KERN_EMERG "%s: jbd2_start: "
2338                                "%ld pages, ino %lu; err %d\n", __func__,
2339                                 wbc->nr_to_write, inode->i_ino, ret);
2340                         dump_stack();
2341                         goto out_writepages;
2342                 }
2343                 if (ext4_should_order_data(inode)) {
2344                         /*
2345                          * With ordered mode we need to add
2346                          * the inode to the journal handl
2347                          * when we do block allocation.
2348                          */
2349                         ret = ext4_jbd2_file_inode(handle, inode);
2350                         if (ret) {
2351                                 ext4_journal_stop(handle);
2352                                 goto out_writepages;
2353                         }
2354                 }
2355
2356                 to_write -= wbc->nr_to_write;
2357                 ret = mpage_da_writepages(mapping, wbc,
2358                                           ext4_da_get_block_write);
2359                 ext4_journal_stop(handle);
2360                 if (ret == MPAGE_DA_EXTENT_TAIL) {
2361                         /*
2362                          * got one extent now try with
2363                          * rest of the pages
2364                          */
2365                         to_write += wbc->nr_to_write;
2366                         ret = 0;
2367                 } else if (wbc->nr_to_write) {
2368                         /*
2369                          * There is no more writeout needed
2370                          * or we requested for a noblocking writeout
2371                          * and we found the device congested
2372                          */
2373                         to_write += wbc->nr_to_write;
2374                         break;
2375                 }
2376                 wbc->nr_to_write = to_write;
2377         }
2378
2379         if (wbc->range_cont && (pages_skipped != wbc->pages_skipped)) {
2380                 /* We skipped pages in this loop */
2381                 wbc->range_start = range_start;
2382                 wbc->nr_to_write = to_write +
2383                                 wbc->pages_skipped - pages_skipped;
2384                 wbc->pages_skipped = pages_skipped;
2385                 goto restart_loop;
2386         }
2387
2388 out_writepages:
2389         wbc->nr_to_write = to_write - nr_to_writebump;
2390         wbc->range_start = range_start;
2391         return ret;
2392 }
2393
2394 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2395                                 loff_t pos, unsigned len, unsigned flags,
2396                                 struct page **pagep, void **fsdata)
2397 {
2398         int ret, retries = 0;
2399         struct page *page;
2400         pgoff_t index;
2401         unsigned from, to;
2402         struct inode *inode = mapping->host;
2403         handle_t *handle;
2404
2405         index = pos >> PAGE_CACHE_SHIFT;
2406         from = pos & (PAGE_CACHE_SIZE - 1);
2407         to = from + len;
2408
2409 retry:
2410         /*
2411          * With delayed allocation, we don't log the i_disksize update
2412          * if there is delayed block allocation. But we still need
2413          * to journalling the i_disksize update if writes to the end
2414          * of file which has an already mapped buffer.
2415          */
2416         handle = ext4_journal_start(inode, 1);
2417         if (IS_ERR(handle)) {
2418                 ret = PTR_ERR(handle);
2419                 goto out;
2420         }
2421
2422         page = __grab_cache_page(mapping, index);
2423         if (!page) {
2424                 ext4_journal_stop(handle);
2425                 ret = -ENOMEM;
2426                 goto out;
2427         }
2428         *pagep = page;
2429
2430         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
2431                                                         ext4_da_get_block_prep);
2432         if (ret < 0) {
2433                 unlock_page(page);
2434                 ext4_journal_stop(handle);
2435                 page_cache_release(page);
2436         }
2437
2438         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
2439                 goto retry;
2440 out:
2441         return ret;
2442 }
2443
2444 /*
2445  * Check if we should update i_disksize
2446  * when write to the end of file but not require block allocation
2447  */
2448 static int ext4_da_should_update_i_disksize(struct page *page,
2449                                          unsigned long offset)
2450 {
2451         struct buffer_head *bh;
2452         struct inode *inode = page->mapping->host;
2453         unsigned int idx;
2454         int i;
2455
2456         bh = page_buffers(page);
2457         idx = offset >> inode->i_blkbits;
2458
2459         for (i=0; i < idx; i++)
2460                 bh = bh->b_this_page;
2461
2462         if (!buffer_mapped(bh) || (buffer_delay(bh)))
2463                 return 0;
2464         return 1;
2465 }
2466
2467 static int ext4_da_write_end(struct file *file,
2468                                 struct address_space *mapping,
2469                                 loff_t pos, unsigned len, unsigned copied,
2470                                 struct page *page, void *fsdata)
2471 {
2472         struct inode *inode = mapping->host;
2473         int ret = 0, ret2;
2474         handle_t *handle = ext4_journal_current_handle();
2475         loff_t new_i_size;
2476         unsigned long start, end;
2477
2478         start = pos & (PAGE_CACHE_SIZE - 1);
2479         end = start + copied -1;
2480
2481         /*
2482          * generic_write_end() will run mark_inode_dirty() if i_size
2483          * changes.  So let's piggyback the i_disksize mark_inode_dirty
2484          * into that.
2485          */
2486
2487         new_i_size = pos + copied;
2488         if (new_i_size > EXT4_I(inode)->i_disksize) {
2489                 if (ext4_da_should_update_i_disksize(page, end)) {
2490                         down_write(&EXT4_I(inode)->i_data_sem);
2491                         if (new_i_size > EXT4_I(inode)->i_disksize) {
2492                                 /*
2493                                  * Updating i_disksize when extending file
2494                                  * without needing block allocation
2495                                  */
2496                                 if (ext4_should_order_data(inode))
2497                                         ret = ext4_jbd2_file_inode(handle,
2498                                                                    inode);
2499
2500                                 EXT4_I(inode)->i_disksize = new_i_size;
2501                         }
2502                         up_write(&EXT4_I(inode)->i_data_sem);
2503                 }
2504         }
2505         ret2 = generic_write_end(file, mapping, pos, len, copied,
2506                                                         page, fsdata);
2507         copied = ret2;
2508         if (ret2 < 0)
2509                 ret = ret2;
2510         ret2 = ext4_journal_stop(handle);
2511         if (!ret)
2512                 ret = ret2;
2513
2514         return ret ? ret : copied;
2515 }
2516
2517 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
2518 {
2519         /*
2520          * Drop reserved blocks
2521          */
2522         BUG_ON(!PageLocked(page));
2523         if (!page_has_buffers(page))
2524                 goto out;
2525
2526         ext4_da_page_release_reservation(page, offset);
2527
2528 out:
2529         ext4_invalidatepage(page, offset);
2530
2531         return;
2532 }
2533
2534
2535 /*
2536  * bmap() is special.  It gets used by applications such as lilo and by
2537  * the swapper to find the on-disk block of a specific piece of data.
2538  *
2539  * Naturally, this is dangerous if the block concerned is still in the
2540  * journal.  If somebody makes a swapfile on an ext4 data-journaling
2541  * filesystem and enables swap, then they may get a nasty shock when the
2542  * data getting swapped to that swapfile suddenly gets overwritten by
2543  * the original zero's written out previously to the journal and
2544  * awaiting writeback in the kernel's buffer cache.
2545  *
2546  * So, if we see any bmap calls here on a modified, data-journaled file,
2547  * take extra steps to flush any blocks which might be in the cache.
2548  */
2549 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
2550 {
2551         struct inode *inode = mapping->host;
2552         journal_t *journal;
2553         int err;
2554
2555         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
2556                         test_opt(inode->i_sb, DELALLOC)) {
2557                 /*
2558                  * With delalloc we want to sync the file
2559                  * so that we can make sure we allocate
2560                  * blocks for file
2561                  */
2562                 filemap_write_and_wait(mapping);
2563         }
2564
2565         if (EXT4_I(inode)->i_state & EXT4_STATE_JDATA) {
2566                 /*
2567                  * This is a REALLY heavyweight approach, but the use of
2568                  * bmap on dirty files is expected to be extremely rare:
2569                  * only if we run lilo or swapon on a freshly made file
2570                  * do we expect this to happen.
2571                  *
2572                  * (bmap requires CAP_SYS_RAWIO so this does not
2573                  * represent an unprivileged user DOS attack --- we'd be
2574                  * in trouble if mortal users could trigger this path at
2575                  * will.)
2576                  *
2577                  * NB. EXT4_STATE_JDATA is not set on files other than
2578                  * regular files.  If somebody wants to bmap a directory
2579                  * or symlink and gets confused because the buffer
2580                  * hasn't yet been flushed to disk, they deserve
2581                  * everything they get.
2582                  */
2583
2584                 EXT4_I(inode)->i_state &= ~EXT4_STATE_JDATA;
2585                 journal = EXT4_JOURNAL(inode);
2586                 jbd2_journal_lock_updates(journal);
2587                 err = jbd2_journal_flush(journal);
2588                 jbd2_journal_unlock_updates(journal);
2589
2590                 if (err)
2591                         return 0;
2592         }
2593
2594         return generic_block_bmap(mapping,block,ext4_get_block);
2595 }
2596
2597 static int bget_one(handle_t *handle, struct buffer_head *bh)
2598 {
2599         get_bh(bh);
2600         return 0;
2601 }
2602
2603 static int bput_one(handle_t *handle, struct buffer_head *bh)
2604 {
2605         put_bh(bh);
2606         return 0;
2607 }
2608
2609 /*
2610  * Note that we don't need to start a transaction unless we're journaling data
2611  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2612  * need to file the inode to the transaction's list in ordered mode because if
2613  * we are writing back data added by write(), the inode is already there and if
2614  * we are writing back data modified via mmap(), noone guarantees in which
2615  * transaction the data will hit the disk. In case we are journaling data, we
2616  * cannot start transaction directly because transaction start ranks above page
2617  * lock so we have to do some magic.
2618  *
2619  * In all journaling modes block_write_full_page() will start the I/O.
2620  *
2621  * Problem:
2622  *
2623  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2624  *              ext4_writepage()
2625  *
2626  * Similar for:
2627  *
2628  *      ext4_file_write() -> generic_file_write() -> __alloc_pages() -> ...
2629  *
2630  * Same applies to ext4_get_block().  We will deadlock on various things like
2631  * lock_journal and i_data_sem
2632  *
2633  * Setting PF_MEMALLOC here doesn't work - too many internal memory
2634  * allocations fail.
2635  *
2636  * 16May01: If we're reentered then journal_current_handle() will be
2637  *          non-zero. We simply *return*.
2638  *
2639  * 1 July 2001: @@@ FIXME:
2640  *   In journalled data mode, a data buffer may be metadata against the
2641  *   current transaction.  But the same file is part of a shared mapping
2642  *   and someone does a writepage() on it.
2643  *
2644  *   We will move the buffer onto the async_data list, but *after* it has
2645  *   been dirtied. So there's a small window where we have dirty data on
2646  *   BJ_Metadata.
2647  *
2648  *   Note that this only applies to the last partial page in the file.  The
2649  *   bit which block_write_full_page() uses prepare/commit for.  (That's
2650  *   broken code anyway: it's wrong for msync()).
2651  *
2652  *   It's a rare case: affects the final partial page, for journalled data
2653  *   where the file is subject to bith write() and writepage() in the same
2654  *   transction.  To fix it we'll need a custom block_write_full_page().
2655  *   We'll probably need that anyway for journalling writepage() output.
2656  *
2657  * We don't honour synchronous mounts for writepage().  That would be
2658  * disastrous.  Any write() or metadata operation will sync the fs for
2659  * us.
2660  *
2661  */
2662 static int __ext4_normal_writepage(struct page *page,
2663                                 struct writeback_control *wbc)
2664 {
2665         struct inode *inode = page->mapping->host;
2666
2667         if (test_opt(inode->i_sb, NOBH))
2668                 return nobh_writepage(page,
2669                                         ext4_normal_get_block_write, wbc);
2670         else
2671                 return block_write_full_page(page,
2672                                                 ext4_normal_get_block_write,
2673                                                 wbc);
2674 }
2675
2676 static int ext4_normal_writepage(struct page *page,
2677                                 struct writeback_control *wbc)
2678 {
2679         struct inode *inode = page->mapping->host;
2680         loff_t size = i_size_read(inode);
2681         loff_t len;
2682
2683         J_ASSERT(PageLocked(page));
2684         if (page->index == size >> PAGE_CACHE_SHIFT)
2685                 len = size & ~PAGE_CACHE_MASK;
2686         else
2687                 len = PAGE_CACHE_SIZE;
2688
2689         if (page_has_buffers(page)) {
2690                 /* if page has buffers it should all be mapped
2691                  * and allocated. If there are not buffers attached
2692                  * to the page we know the page is dirty but it lost
2693                  * buffers. That means that at some moment in time
2694                  * after write_begin() / write_end() has been called
2695                  * all buffers have been clean and thus they must have been
2696                  * written at least once. So they are all mapped and we can
2697                  * happily proceed with mapping them and writing the page.
2698                  */
2699                 BUG_ON(walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
2700                                         ext4_bh_unmapped_or_delay));
2701         }
2702
2703         if (!ext4_journal_current_handle())
2704                 return __ext4_normal_writepage(page, wbc);
2705
2706         redirty_page_for_writepage(wbc, page);
2707         unlock_page(page);
2708         return 0;
2709 }
2710
2711 static int __ext4_journalled_writepage(struct page *page,
2712                                 struct writeback_control *wbc)
2713 {
2714         struct address_space *mapping = page->mapping;
2715         struct inode *inode = mapping->host;
2716         struct buffer_head *page_bufs;
2717         handle_t *handle = NULL;
2718         int ret = 0;
2719         int err;
2720
2721         ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
2722                                         ext4_normal_get_block_write);
2723         if (ret != 0)
2724                 goto out_unlock;
2725
2726         page_bufs = page_buffers(page);
2727         walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE, NULL,
2728                                                                 bget_one);
2729         /* As soon as we unlock the page, it can go away, but we have
2730          * references to buffers so we are safe */
2731         unlock_page(page);
2732
2733         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2734         if (IS_ERR(handle)) {
2735                 ret = PTR_ERR(handle);
2736                 goto out;
2737         }
2738
2739         ret = walk_page_buffers(handle, page_bufs, 0,
2740                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
2741
2742         err = walk_page_buffers(handle, page_bufs, 0,
2743                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
2744         if (ret == 0)
2745                 ret = err;
2746         err = ext4_journal_stop(handle);
2747         if (!ret)
2748                 ret = err;
2749
2750         walk_page_buffers(handle, page_bufs, 0,
2751                                 PAGE_CACHE_SIZE, NULL, bput_one);
2752         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
2753         goto out;
2754
2755 out_unlock:
2756         unlock_page(page);
2757 out:
2758         return ret;
2759 }
2760
2761 static int ext4_journalled_writepage(struct page *page,
2762                                 struct writeback_control *wbc)
2763 {
2764         struct inode *inode = page->mapping->host;
2765         loff_t size = i_size_read(inode);
2766         loff_t len;
2767
2768         J_ASSERT(PageLocked(page));
2769         if (page->index == size >> PAGE_CACHE_SHIFT)
2770                 len = size & ~PAGE_CACHE_MASK;
2771         else
2772                 len = PAGE_CACHE_SIZE;
2773
2774         if (page_has_buffers(page)) {
2775                 /* if page has buffers it should all be mapped
2776                  * and allocated. If there are not buffers attached
2777                  * to the page we know the page is dirty but it lost
2778                  * buffers. That means that at some moment in time
2779                  * after write_begin() / write_end() has been called
2780                  * all buffers have been clean and thus they must have been
2781                  * written at least once. So they are all mapped and we can
2782                  * happily proceed with mapping them and writing the page.
2783                  */
2784                 BUG_ON(walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
2785                                         ext4_bh_unmapped_or_delay));
2786         }
2787
2788         if (ext4_journal_current_handle())
2789                 goto no_write;
2790
2791         if (PageChecked(page)) {
2792                 /*
2793                  * It's mmapped pagecache.  Add buffers and journal it.  There
2794                  * doesn't seem much point in redirtying the page here.
2795                  */
2796                 ClearPageChecked(page);
2797                 return __ext4_journalled_writepage(page, wbc);
2798         } else {
2799                 /*
2800                  * It may be a page full of checkpoint-mode buffers.  We don't
2801                  * really know unless we go poke around in the buffer_heads.
2802                  * But block_write_full_page will do the right thing.
2803                  */
2804                 return block_write_full_page(page,
2805                                                 ext4_normal_get_block_write,
2806                                                 wbc);
2807         }
2808 no_write:
2809         redirty_page_for_writepage(wbc, page);
2810         unlock_page(page);
2811         return 0;
2812 }
2813
2814 static int ext4_readpage(struct file *file, struct page *page)
2815 {
2816         return mpage_readpage(page, ext4_get_block);
2817 }
2818
2819 static int
2820 ext4_readpages(struct file *file, struct address_space *mapping,
2821                 struct list_head *pages, unsigned nr_pages)
2822 {
2823         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
2824 }
2825
2826 static void ext4_invalidatepage(struct page *page, unsigned long offset)
2827 {
2828         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2829
2830         /*
2831          * If it's a full truncate we just forget about the pending dirtying
2832          */
2833         if (offset == 0)
2834                 ClearPageChecked(page);
2835
2836         jbd2_journal_invalidatepage(journal, page, offset);
2837 }
2838
2839 static int ext4_releasepage(struct page *page, gfp_t wait)
2840 {
2841         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2842
2843         WARN_ON(PageChecked(page));
2844         if (!page_has_buffers(page))
2845                 return 0;
2846         return jbd2_journal_try_to_free_buffers(journal, page, wait);
2847 }
2848
2849 /*
2850  * If the O_DIRECT write will extend the file then add this inode to the
2851  * orphan list.  So recovery will truncate it back to the original size
2852  * if the machine crashes during the write.
2853  *
2854  * If the O_DIRECT write is intantiating holes inside i_size and the machine
2855  * crashes then stale disk data _may_ be exposed inside the file. But current
2856  * VFS code falls back into buffered path in that case so we are safe.
2857  */
2858 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
2859                         const struct iovec *iov, loff_t offset,
2860                         unsigned long nr_segs)
2861 {
2862         struct file *file = iocb->ki_filp;
2863         struct inode *inode = file->f_mapping->host;
2864         struct ext4_inode_info *ei = EXT4_I(inode);
2865         handle_t *handle;
2866         ssize_t ret;
2867         int orphan = 0;
2868         size_t count = iov_length(iov, nr_segs);
2869
2870         if (rw == WRITE) {
2871                 loff_t final_size = offset + count;
2872
2873                 if (final_size > inode->i_size) {
2874                         /* Credits for sb + inode write */
2875                         handle = ext4_journal_start(inode, 2);
2876                         if (IS_ERR(handle)) {
2877                                 ret = PTR_ERR(handle);
2878                                 goto out;
2879                         }
2880                         ret = ext4_orphan_add(handle, inode);
2881                         if (ret) {
2882                                 ext4_journal_stop(handle);
2883                                 goto out;
2884                         }
2885                         orphan = 1;
2886                         ei->i_disksize = inode->i_size;
2887                         ext4_journal_stop(handle);
2888                 }
2889         }
2890
2891         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
2892                                  offset, nr_segs,
2893                                  ext4_get_block, NULL);
2894
2895         if (orphan) {
2896                 int err;
2897
2898                 /* Credits for sb + inode write */
2899                 handle = ext4_journal_start(inode, 2);
2900                 if (IS_ERR(handle)) {
2901                         /* This is really bad luck. We've written the data
2902                          * but cannot extend i_size. Bail out and pretend
2903                          * the write failed... */
2904                         ret = PTR_ERR(handle);
2905                         goto out;
2906                 }
2907                 if (inode->i_nlink)
2908                         ext4_orphan_del(handle, inode);
2909                 if (ret > 0) {
2910                         loff_t end = offset + ret;
2911                         if (end > inode->i_size) {
2912                                 ei->i_disksize = end;
2913                                 i_size_write(inode, end);
2914                                 /*
2915                                  * We're going to return a positive `ret'
2916                                  * here due to non-zero-length I/O, so there's
2917                                  * no way of reporting error returns from
2918                                  * ext4_mark_inode_dirty() to userspace.  So
2919                                  * ignore it.
2920                                  */
2921                                 ext4_mark_inode_dirty(handle, inode);
2922                         }
2923                 }
2924                 err = ext4_journal_stop(handle);
2925                 if (ret == 0)
2926                         ret = err;
2927         }
2928 out:
2929         return ret;
2930 }
2931
2932 /*
2933  * Pages can be marked dirty completely asynchronously from ext4's journalling
2934  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
2935  * much here because ->set_page_dirty is called under VFS locks.  The page is
2936  * not necessarily locked.
2937  *
2938  * We cannot just dirty the page and leave attached buffers clean, because the
2939  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
2940  * or jbddirty because all the journalling code will explode.
2941  *
2942  * So what we do is to mark the page "pending dirty" and next time writepage
2943  * is called, propagate that into the buffers appropriately.
2944  */
2945 static int ext4_journalled_set_page_dirty(struct page *page)
2946 {
2947         SetPageChecked(page);
2948         return __set_page_dirty_nobuffers(page);
2949 }
2950
2951 static const struct address_space_operations ext4_ordered_aops = {
2952         .readpage               = ext4_readpage,
2953         .readpages              = ext4_readpages,
2954         .writepage              = ext4_normal_writepage,
2955         .sync_page              = block_sync_page,
2956         .write_begin            = ext4_write_begin,
2957         .write_end              = ext4_ordered_write_end,
2958         .bmap                   = ext4_bmap,
2959         .invalidatepage         = ext4_invalidatepage,
2960         .releasepage            = ext4_releasepage,
2961         .direct_IO              = ext4_direct_IO,
2962         .migratepage            = buffer_migrate_page,
2963         .is_partially_uptodate  = block_is_partially_uptodate,
2964 };
2965
2966 static const struct address_space_operations ext4_writeback_aops = {
2967         .readpage               = ext4_readpage,
2968         .readpages              = ext4_readpages,
2969         .writepage              = ext4_normal_writepage,
2970         .sync_page              = block_sync_page,
2971         .write_begin            = ext4_write_begin,
2972         .write_end              = ext4_writeback_write_end,
2973         .bmap                   = ext4_bmap,
2974         .invalidatepage         = ext4_invalidatepage,
2975         .releasepage            = ext4_releasepage,
2976         .direct_IO              = ext4_direct_IO,
2977         .migratepage            = buffer_migrate_page,
2978         .is_partially_uptodate  = block_is_partially_uptodate,
2979 };
2980
2981 static const struct address_space_operations ext4_journalled_aops = {
2982         .readpage               = ext4_readpage,
2983         .readpages              = ext4_readpages,
2984         .writepage              = ext4_journalled_writepage,
2985         .sync_page              = block_sync_page,
2986         .write_begin            = ext4_write_begin,
2987         .write_end              = ext4_journalled_write_end,
2988         .set_page_dirty         = ext4_journalled_set_page_dirty,
2989         .bmap                   = ext4_bmap,
2990         .invalidatepage         = ext4_invalidatepage,
2991         .releasepage            = ext4_releasepage,
2992         .is_partially_uptodate  = block_is_partially_uptodate,
2993 };
2994
2995 static const struct address_space_operations ext4_da_aops = {
2996         .readpage               = ext4_readpage,
2997         .readpages              = ext4_readpages,
2998         .writepage              = ext4_da_writepage,
2999         .writepages             = ext4_da_writepages,
3000         .sync_page              = block_sync_page,
3001         .write_begin            = ext4_da_write_begin,
3002         .write_end              = ext4_da_write_end,
3003         .bmap                   = ext4_bmap,
3004         .invalidatepage         = ext4_da_invalidatepage,
3005         .releasepage            = ext4_releasepage,
3006         .direct_IO              = ext4_direct_IO,
3007         .migratepage            = buffer_migrate_page,
3008         .is_partially_uptodate  = block_is_partially_uptodate,
3009 };
3010
3011 void ext4_set_aops(struct inode *inode)
3012 {
3013         if (ext4_should_order_data(inode) &&
3014                 test_opt(inode->i_sb, DELALLOC))
3015                 inode->i_mapping->a_ops = &ext4_da_aops;
3016         else if (ext4_should_order_data(inode))
3017                 inode->i_mapping->a_ops = &ext4_ordered_aops;
3018         else if (ext4_should_writeback_data(inode) &&
3019                  test_opt(inode->i_sb, DELALLOC))
3020                 inode->i_mapping->a_ops = &ext4_da_aops;
3021         else if (ext4_should_writeback_data(inode))
3022                 inode->i_mapping->a_ops = &ext4_writeback_aops;
3023         else
3024                 inode->i_mapping->a_ops = &ext4_journalled_aops;
3025 }
3026
3027 /*
3028  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3029  * up to the end of the block which corresponds to `from'.
3030  * This required during truncate. We need to physically zero the tail end
3031  * of that block so it doesn't yield old data if the file is later grown.
3032  */
3033 int ext4_block_truncate_page(handle_t *handle,
3034                 struct address_space *mapping, loff_t from)
3035 {
3036         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3037         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3038         unsigned blocksize, length, pos;
3039         ext4_lblk_t iblock;
3040         struct inode *inode = mapping->host;
3041         struct buffer_head *bh;
3042         struct page *page;
3043         int err = 0;
3044
3045         page = grab_cache_page(mapping, from >> PAGE_CACHE_SHIFT);
3046         if (!page)
3047                 return -EINVAL;
3048
3049         blocksize = inode->i_sb->s_blocksize;
3050         length = blocksize - (offset & (blocksize - 1));
3051         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3052
3053         /*
3054          * For "nobh" option,  we can only work if we don't need to
3055          * read-in the page - otherwise we create buffers to do the IO.
3056          */
3057         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
3058              ext4_should_writeback_data(inode) && PageUptodate(page)) {
3059                 zero_user(page, offset, length);
3060                 set_page_dirty(page);
3061                 goto unlock;
3062         }
3063
3064         if (!page_has_buffers(page))
3065                 create_empty_buffers(page, blocksize, 0);
3066
3067         /* Find the buffer that contains "offset" */
3068         bh = page_buffers(page);
3069         pos = blocksize;
3070         while (offset >= pos) {
3071                 bh = bh->b_this_page;
3072                 iblock++;
3073                 pos += blocksize;
3074         }
3075
3076         err = 0;
3077         if (buffer_freed(bh)) {
3078                 BUFFER_TRACE(bh, "freed: skip");
3079                 goto unlock;
3080         }
3081
3082         if (!buffer_mapped(bh)) {
3083                 BUFFER_TRACE(bh, "unmapped");
3084                 ext4_get_block(inode, iblock, bh, 0);
3085                 /* unmapped? It's a hole - nothing to do */
3086                 if (!buffer_mapped(bh)) {
3087                         BUFFER_TRACE(bh, "still unmapped");
3088                         goto unlock;
3089                 }
3090         }
3091
3092         /* Ok, it's mapped. Make sure it's up-to-date */
3093         if (PageUptodate(page))
3094                 set_buffer_uptodate(bh);
3095
3096         if (!buffer_uptodate(bh)) {
3097                 err = -EIO;
3098                 ll_rw_block(READ, 1, &bh);
3099                 wait_on_buffer(bh);
3100                 /* Uhhuh. Read error. Complain and punt. */
3101                 if (!buffer_uptodate(bh))
3102                         goto unlock;
3103         }
3104
3105         if (ext4_should_journal_data(inode)) {
3106                 BUFFER_TRACE(bh, "get write access");
3107                 err = ext4_journal_get_write_access(handle, bh);
3108                 if (err)
3109                         goto unlock;
3110         }
3111
3112         zero_user(page, offset, length);
3113
3114         BUFFER_TRACE(bh, "zeroed end of block");
3115
3116         err = 0;
3117         if (ext4_should_journal_data(inode)) {
3118                 err = ext4_journal_dirty_metadata(handle, bh);
3119         } else {
3120                 if (ext4_should_order_data(inode))
3121                         err = ext4_jbd2_file_inode(handle, inode);
3122                 mark_buffer_dirty(bh);
3123         }
3124
3125 unlock:
3126         unlock_page(page);
3127         page_cache_release(page);
3128         return err;
3129 }
3130
3131 /*
3132  * Probably it should be a library function... search for first non-zero word
3133  * or memcmp with zero_page, whatever is better for particular architecture.
3134  * Linus?
3135  */
3136 static inline int all_zeroes(__le32 *p, __le32 *q)
3137 {
3138         while (p < q)
3139                 if (*p++)
3140                         return 0;
3141         return 1;
3142 }
3143
3144 /**
3145  *      ext4_find_shared - find the indirect blocks for partial truncation.
3146  *      @inode:   inode in question
3147  *      @depth:   depth of the affected branch
3148  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
3149  *      @chain:   place to store the pointers to partial indirect blocks
3150  *      @top:     place to the (detached) top of branch
3151  *
3152  *      This is a helper function used by ext4_truncate().
3153  *
3154  *      When we do truncate() we may have to clean the ends of several
3155  *      indirect blocks but leave the blocks themselves alive. Block is
3156  *      partially truncated if some data below the new i_size is refered
3157  *      from it (and it is on the path to the first completely truncated
3158  *      data block, indeed).  We have to free the top of that path along
3159  *      with everything to the right of the path. Since no allocation
3160  *      past the truncation point is possible until ext4_truncate()
3161  *      finishes, we may safely do the latter, but top of branch may
3162  *      require special attention - pageout below the truncation point
3163  *      might try to populate it.
3164  *
3165  *      We atomically detach the top of branch from the tree, store the
3166  *      block number of its root in *@top, pointers to buffer_heads of
3167  *      partially truncated blocks - in @chain[].bh and pointers to
3168  *      their last elements that should not be removed - in
3169  *      @chain[].p. Return value is the pointer to last filled element
3170  *      of @chain.
3171  *
3172  *      The work left to caller to do the actual freeing of subtrees:
3173  *              a) free the subtree starting from *@top
3174  *              b) free the subtrees whose roots are stored in
3175  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
3176  *              c) free the subtrees growing from the inode past the @chain[0].
3177  *                      (no partially truncated stuff there).  */
3178
3179 static Indirect *ext4_find_shared(struct inode *inode, int depth,
3180                         ext4_lblk_t offsets[4], Indirect chain[4], __le32 *top)
3181 {
3182         Indirect *partial, *p;
3183         int k, err;
3184
3185         *top = 0;
3186         /* Make k index the deepest non-null offest + 1 */
3187         for (k = depth; k > 1 && !offsets[k-1]; k--)
3188                 ;
3189         partial = ext4_get_branch(inode, k, offsets, chain, &err);
3190         /* Writer: pointers */
3191         if (!partial)
3192                 partial = chain + k-1;
3193         /*
3194          * If the branch acquired continuation since we've looked at it -
3195          * fine, it should all survive and (new) top doesn't belong to us.
3196          */
3197         if (!partial->key && *partial->p)
3198                 /* Writer: end */
3199                 goto no_top;
3200         for (p=partial; p>chain && all_zeroes((__le32*)p->bh->b_data,p->p); p--)
3201                 ;
3202         /*
3203          * OK, we've found the last block that must survive. The rest of our
3204          * branch should be detached before unlocking. However, if that rest
3205          * of branch is all ours and does not grow immediately from the inode
3206          * it's easier to cheat and just decrement partial->p.
3207          */
3208         if (p == chain + k - 1 && p > chain) {
3209                 p->p--;
3210         } else {
3211                 *top = *p->p;
3212                 /* Nope, don't do this in ext4.  Must leave the tree intact */
3213 #if 0
3214                 *p->p = 0;
3215 #endif
3216         }
3217         /* Writer: end */
3218
3219         while(partial > p) {
3220                 brelse(partial->bh);
3221                 partial--;
3222         }
3223 no_top:
3224         return partial;
3225 }
3226
3227 /*
3228  * Zero a number of block pointers in either an inode or an indirect block.
3229  * If we restart the transaction we must again get write access to the
3230  * indirect block for further modification.
3231  *
3232  * We release `count' blocks on disk, but (last - first) may be greater
3233  * than `count' because there can be holes in there.
3234  */
3235 static void ext4_clear_blocks(handle_t *handle, struct inode *inode,
3236                 struct buffer_head *bh, ext4_fsblk_t block_to_free,
3237                 unsigned long count, __le32 *first, __le32 *last)
3238 {
3239         __le32 *p;
3240         if (try_to_extend_transaction(handle, inode)) {
3241                 if (bh) {
3242                         BUFFER_TRACE(bh, "call ext4_journal_dirty_metadata");
3243                         ext4_journal_dirty_metadata(handle, bh);
3244                 }
3245                 ext4_mark_inode_dirty(handle, inode);
3246                 ext4_journal_test_restart(handle, inode);
3247                 if (bh) {
3248                         BUFFER_TRACE(bh, "retaking write access");
3249                         ext4_journal_get_write_access(handle, bh);
3250                 }
3251         }
3252
3253         /*
3254          * Any buffers which are on the journal will be in memory. We find
3255          * them on the hash table so jbd2_journal_revoke() will run jbd2_journal_forget()
3256          * on them.  We've already detached each block from the file, so
3257          * bforget() in jbd2_journal_forget() should be safe.
3258          *
3259          * AKPM: turn on bforget in jbd2_journal_forget()!!!
3260          */
3261         for (p = first; p < last; p++) {
3262                 u32 nr = le32_to_cpu(*p);
3263                 if (nr) {
3264                         struct buffer_head *tbh;
3265
3266                         *p = 0;
3267                         tbh = sb_find_get_block(inode->i_sb, nr);
3268                         ext4_forget(handle, 0, inode, tbh, nr);
3269                 }
3270         }
3271
3272         ext4_free_blocks(handle, inode, block_to_free, count, 0);
3273 }
3274
3275 /**
3276  * ext4_free_data - free a list of data blocks
3277  * @handle:     handle for this transaction
3278  * @inode:      inode we are dealing with
3279  * @this_bh:    indirect buffer_head which contains *@first and *@last
3280  * @first:      array of block numbers
3281  * @last:       points immediately past the end of array
3282  *
3283  * We are freeing all blocks refered from that array (numbers are stored as
3284  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
3285  *
3286  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
3287  * blocks are contiguous then releasing them at one time will only affect one
3288  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
3289  * actually use a lot of journal space.
3290  *
3291  * @this_bh will be %NULL if @first and @last point into the inode's direct
3292  * block pointers.
3293  */
3294 static void ext4_free_data(handle_t *handle, struct inode *inode,
3295                            struct buffer_head *this_bh,
3296                            __le32 *first, __le32 *last)
3297 {
3298         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
3299         unsigned long count = 0;            /* Number of blocks in the run */
3300         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
3301                                                corresponding to
3302                                                block_to_free */
3303         ext4_fsblk_t nr;                    /* Current block # */
3304         __le32 *p;                          /* Pointer into inode/ind
3305                                                for current block */
3306         int err;
3307
3308         if (this_bh) {                          /* For indirect block */
3309                 BUFFER_TRACE(this_bh, "get_write_access");
3310                 err = ext4_journal_get_write_access(handle, this_bh);
3311                 /* Important: if we can't update the indirect pointers
3312                  * to the blocks, we can't free them. */
3313                 if (err)
3314                         return;
3315         }
3316
3317         for (p = first; p < last; p++) {
3318                 nr = le32_to_cpu(*p);
3319                 if (nr) {
3320                         /* accumulate blocks to free if they're contiguous */
3321                         if (count == 0) {
3322                                 block_to_free = nr;
3323                                 block_to_free_p = p;
3324                                 count = 1;
3325                         } else if (nr == block_to_free + count) {
3326                                 count++;
3327                         } else {
3328                                 ext4_clear_blocks(handle, inode, this_bh,
3329                                                   block_to_free,
3330                                                   count, block_to_free_p, p);
3331                                 block_to_free = nr;
3332                                 block_to_free_p = p;
3333                                 count = 1;
3334                         }
3335                 }
3336         }
3337
3338         if (count > 0)
3339                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
3340                                   count, block_to_free_p, p);
3341
3342         if (this_bh) {
3343                 BUFFER_TRACE(this_bh, "call ext4_journal_dirty_metadata");
3344
3345                 /*
3346                  * The buffer head should have an attached journal head at this
3347                  * point. However, if the data is corrupted and an indirect
3348                  * block pointed to itself, it would have been detached when
3349                  * the block was cleared. Check for this instead of OOPSing.
3350                  */
3351                 if (bh2jh(this_bh))
3352                         ext4_journal_dirty_metadata(handle, this_bh);
3353                 else
3354                         ext4_error(inode->i_sb, __func__,
3355                                    "circular indirect block detected, "
3356                                    "inode=%lu, block=%llu",
3357                                    inode->i_ino,
3358                                    (unsigned long long) this_bh->b_blocknr);
3359         }
3360 }
3361
3362 /**
3363  *      ext4_free_branches - free an array of branches
3364  *      @handle: JBD handle for this transaction
3365  *      @inode: inode we are dealing with
3366  *      @parent_bh: the buffer_head which contains *@first and *@last
3367  *      @first: array of block numbers
3368  *      @last:  pointer immediately past the end of array
3369  *      @depth: depth of the branches to free
3370  *
3371  *      We are freeing all blocks refered from these branches (numbers are
3372  *      stored as little-endian 32-bit) and updating @inode->i_blocks
3373  *      appropriately.
3374  */
3375 static void ext4_free_branches(handle_t *handle, struct inode *inode,
3376                                struct buffer_head *parent_bh,
3377                                __le32 *first, __le32 *last, int depth)
3378 {
3379         ext4_fsblk_t nr;
3380         __le32 *p;
3381
3382         if (is_handle_aborted(handle))
3383                 return;
3384
3385         if (depth--) {
3386                 struct buffer_head *bh;
3387                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
3388                 p = last;
3389                 while (--p >= first) {
3390                         nr = le32_to_cpu(*p);
3391                         if (!nr)
3392                                 continue;               /* A hole */
3393
3394                         /* Go read the buffer for the next level down */
3395                         bh = sb_bread(inode->i_sb, nr);
3396
3397                         /*
3398                          * A read failure? Report error and clear slot
3399                          * (should be rare).
3400                          */
3401                         if (!bh) {
3402                                 ext4_error(inode->i_sb, "ext4_free_branches",
3403                                            "Read failure, inode=%lu, block=%llu",
3404                                            inode->i_ino, nr);
3405                                 continue;
3406                         }
3407
3408                         /* This zaps the entire block.  Bottom up. */
3409                         BUFFER_TRACE(bh, "free child branches");
3410                         ext4_free_branches(handle, inode, bh,
3411                                            (__le32*)bh->b_data,
3412                                            (__le32*)bh->b_data + addr_per_block,
3413                                            depth);
3414
3415                         /*
3416                          * We've probably journalled the indirect block several
3417                          * times during the truncate.  But it's no longer
3418                          * needed and we now drop it from the transaction via
3419                          * jbd2_journal_revoke().
3420                          *
3421                          * That's easy if it's exclusively part of this
3422                          * transaction.  But if it's part of the committing
3423                          * transaction then jbd2_journal_forget() will simply
3424                          * brelse() it.  That means that if the underlying
3425                          * block is reallocated in ext4_get_block(),
3426                          * unmap_underlying_metadata() will find this block
3427                          * and will try to get rid of it.  damn, damn.
3428                          *
3429                          * If this block has already been committed to the
3430                          * journal, a revoke record will be written.  And
3431                          * revoke records must be emitted *before* clearing
3432                          * this block's bit in the bitmaps.
3433                          */
3434                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
3435
3436                         /*
3437                          * Everything below this this pointer has been
3438                          * released.  Now let this top-of-subtree go.
3439                          *
3440                          * We want the freeing of this indirect block to be
3441                          * atomic in the journal with the updating of the
3442                          * bitmap block which owns it.  So make some room in
3443                          * the journal.
3444                          *
3445                          * We zero the parent pointer *after* freeing its
3446                          * pointee in the bitmaps, so if extend_transaction()
3447                          * for some reason fails to put the bitmap changes and
3448                          * the release into the same transaction, recovery
3449                          * will merely complain about releasing a free block,
3450                          * rather than leaking blocks.
3451                          */
3452                         if (is_handle_aborted(handle))
3453                                 return;
3454                         if (try_to_extend_transaction(handle, inode)) {
3455                                 ext4_mark_inode_dirty(handle, inode);
3456                                 ext4_journal_test_restart(handle, inode);
3457                         }
3458
3459                         ext4_free_blocks(handle, inode, nr, 1, 1);
3460
3461                         if (parent_bh) {
3462                                 /*
3463                                  * The block which we have just freed is
3464                                  * pointed to by an indirect block: journal it
3465                                  */
3466                                 BUFFER_TRACE(parent_bh, "get_write_access");
3467                                 if (!ext4_journal_get_write_access(handle,
3468                                                                    parent_bh)){
3469                                         *p = 0;
3470                                         BUFFER_TRACE(parent_bh,
3471                                         "call ext4_journal_dirty_metadata");
3472                                         ext4_journal_dirty_metadata(handle,
3473                                                                     parent_bh);
3474                                 }
3475                         }
3476                 }
3477         } else {
3478                 /* We have reached the bottom of the tree. */
3479                 BUFFER_TRACE(parent_bh, "free data blocks");
3480                 ext4_free_data(handle, inode, parent_bh, first, last);
3481         }
3482 }
3483
3484 int ext4_can_truncate(struct inode *inode)
3485 {
3486         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
3487                 return 0;
3488         if (S_ISREG(inode->i_mode))
3489                 return 1;
3490         if (S_ISDIR(inode->i_mode))
3491                 return 1;
3492         if (S_ISLNK(inode->i_mode))
3493                 return !ext4_inode_is_fast_symlink(inode);
3494         return 0;
3495 }
3496
3497 /*
3498  * ext4_truncate()
3499  *
3500  * We block out ext4_get_block() block instantiations across the entire
3501  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
3502  * simultaneously on behalf of the same inode.
3503  *
3504  * As we work through the truncate and commmit bits of it to the journal there
3505  * is one core, guiding principle: the file's tree must always be consistent on
3506  * disk.  We must be able to restart the truncate after a crash.
3507  *
3508  * The file's tree may be transiently inconsistent in memory (although it
3509  * probably isn't), but whenever we close off and commit a journal transaction,
3510  * the contents of (the filesystem + the journal) must be consistent and
3511  * restartable.  It's pretty simple, really: bottom up, right to left (although
3512  * left-to-right works OK too).
3513  *
3514  * Note that at recovery time, journal replay occurs *before* the restart of
3515  * truncate against the orphan inode list.
3516  *
3517  * The committed inode has the new, desired i_size (which is the same as
3518  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
3519  * that this inode's truncate did not complete and it will again call
3520  * ext4_truncate() to have another go.  So there will be instantiated blocks
3521  * to the right of the truncation point in a crashed ext4 filesystem.  But
3522  * that's fine - as long as they are linked from the inode, the post-crash
3523  * ext4_truncate() run will find them and release them.
3524  */
3525 void ext4_truncate(struct inode *inode)
3526 {
3527         handle_t *handle;
3528         struct ext4_inode_info *ei = EXT4_I(inode);
3529         __le32 *i_data = ei->i_data;
3530         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
3531         struct address_space *mapping = inode->i_mapping;
3532         ext4_lblk_t offsets[4];
3533         Indirect chain[4];
3534         Indirect *partial;
3535         __le32 nr = 0;
3536         int n;
3537         ext4_lblk_t last_block;
3538         unsigned blocksize = inode->i_sb->s_blocksize;
3539
3540         if (!ext4_can_truncate(inode))
3541                 return;
3542
3543         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
3544                 ext4_ext_truncate(inode);
3545                 return;
3546         }
3547
3548         handle = start_transaction(inode);
3549         if (IS_ERR(handle))
3550                 return;         /* AKPM: return what? */
3551
3552         last_block = (inode->i_size + blocksize-1)
3553                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
3554
3555         if (inode->i_size & (blocksize - 1))
3556                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
3557                         goto out_stop;
3558
3559         n = ext4_block_to_path(inode, last_block, offsets, NULL);
3560         if (n == 0)
3561                 goto out_stop;  /* error */
3562
3563         /*
3564          * OK.  This truncate is going to happen.  We add the inode to the
3565          * orphan list, so that if this truncate spans multiple transactions,
3566          * and we crash, we will resume the truncate when the filesystem
3567          * recovers.  It also marks the inode dirty, to catch the new size.
3568          *
3569          * Implication: the file must always be in a sane, consistent
3570          * truncatable state while each transaction commits.
3571          */
3572         if (ext4_orphan_add(handle, inode))
3573                 goto out_stop;
3574
3575         /*
3576          * From here we block out all ext4_get_block() callers who want to
3577          * modify the block allocation tree.
3578          */
3579         down_write(&ei->i_data_sem);
3580
3581         ext4_discard_reservation(inode);
3582
3583         /*
3584          * The orphan list entry will now protect us from any crash which
3585          * occurs before the truncate completes, so it is now safe to propagate
3586          * the new, shorter inode size (held for now in i_size) into the
3587          * on-disk inode. We do this via i_disksize, which is the value which
3588          * ext4 *really* writes onto the disk inode.
3589          */
3590         ei->i_disksize = inode->i_size;
3591
3592         if (n == 1) {           /* direct blocks */
3593                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
3594                                i_data + EXT4_NDIR_BLOCKS);
3595                 goto do_indirects;
3596         }
3597
3598         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
3599         /* Kill the top of shared branch (not detached) */
3600         if (nr) {
3601                 if (partial == chain) {
3602                         /* Shared branch grows from the inode */
3603                         ext4_free_branches(handle, inode, NULL,
3604                                            &nr, &nr+1, (chain+n-1) - partial);
3605                         *partial->p = 0;
3606                         /*
3607                          * We mark the inode dirty prior to restart,
3608                          * and prior to stop.  No need for it here.
3609                          */
3610                 } else {
3611                         /* Shared branch grows from an indirect block */
3612                         BUFFER_TRACE(partial->bh, "get_write_access");
3613                         ext4_free_branches(handle, inode, partial->bh,
3614                                         partial->p,
3615                                         partial->p+1, (chain+n-1) - partial);
3616                 }
3617         }
3618         /* Clear the ends of indirect blocks on the shared branch */
3619         while (partial > chain) {
3620                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
3621                                    (__le32*)partial->bh->b_data+addr_per_block,
3622                                    (chain+n-1) - partial);
3623                 BUFFER_TRACE(partial->bh, "call brelse");
3624                 brelse (partial->bh);
3625                 partial--;
3626         }
3627 do_indirects:
3628         /* Kill the remaining (whole) subtrees */
3629         switch (offsets[0]) {
3630         default:
3631                 nr = i_data[EXT4_IND_BLOCK];
3632                 if (nr) {
3633                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
3634                         i_data[EXT4_IND_BLOCK] = 0;
3635                 }
3636         case EXT4_IND_BLOCK:
3637                 nr = i_data[EXT4_DIND_BLOCK];
3638                 if (nr) {
3639                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
3640                         i_data[EXT4_DIND_BLOCK] = 0;
3641                 }
3642         case EXT4_DIND_BLOCK:
3643                 nr = i_data[EXT4_TIND_BLOCK];
3644                 if (nr) {
3645                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
3646                         i_data[EXT4_TIND_BLOCK] = 0;
3647                 }
3648         case EXT4_TIND_BLOCK:
3649                 ;
3650         }
3651
3652         up_write(&ei->i_data_sem);
3653         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
3654         ext4_mark_inode_dirty(handle, inode);
3655
3656         /*
3657          * In a multi-transaction truncate, we only make the final transaction
3658          * synchronous
3659          */
3660         if (IS_SYNC(inode))
3661                 handle->h_sync = 1;
3662 out_stop:
3663         /*
3664          * If this was a simple ftruncate(), and the file will remain alive
3665          * then we need to clear up the orphan record which we created above.
3666          * However, if this was a real unlink then we were called by
3667          * ext4_delete_inode(), and we allow that function to clean up the
3668          * orphan info for us.
3669          */
3670         if (inode->i_nlink)
3671                 ext4_orphan_del(handle, inode);
3672
3673         ext4_journal_stop(handle);
3674 }
3675
3676 static ext4_fsblk_t ext4_get_inode_block(struct super_block *sb,
3677                 unsigned long ino, struct ext4_iloc *iloc)
3678 {
3679         ext4_group_t block_group;
3680         unsigned long offset;
3681         ext4_fsblk_t block;
3682         struct ext4_group_desc *gdp;
3683
3684         if (!ext4_valid_inum(sb, ino)) {
3685                 /*
3686                  * This error is already checked for in namei.c unless we are
3687                  * looking at an NFS filehandle, in which case no error
3688                  * report is needed
3689                  */
3690                 return 0;
3691         }
3692
3693         block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb);
3694         gdp = ext4_get_group_desc(sb, block_group, NULL);
3695         if (!gdp)
3696                 return 0;
3697
3698         /*
3699          * Figure out the offset within the block group inode table
3700          */
3701         offset = ((ino - 1) % EXT4_INODES_PER_GROUP(sb)) *
3702                 EXT4_INODE_SIZE(sb);
3703         block = ext4_inode_table(sb, gdp) +
3704                 (offset >> EXT4_BLOCK_SIZE_BITS(sb));
3705
3706         iloc->block_group = block_group;
3707         iloc->offset = offset & (EXT4_BLOCK_SIZE(sb) - 1);
3708         return block;
3709 }
3710
3711 /*
3712  * ext4_get_inode_loc returns with an extra refcount against the inode's
3713  * underlying buffer_head on success. If 'in_mem' is true, we have all
3714  * data in memory that is needed to recreate the on-disk version of this
3715  * inode.
3716  */
3717 static int __ext4_get_inode_loc(struct inode *inode,
3718                                 struct ext4_iloc *iloc, int in_mem)
3719 {
3720         ext4_fsblk_t block;
3721         struct buffer_head *bh;
3722
3723         block = ext4_get_inode_block(inode->i_sb, inode->i_ino, iloc);
3724         if (!block)
3725                 return -EIO;
3726
3727         bh = sb_getblk(inode->i_sb, block);
3728         if (!bh) {
3729                 ext4_error (inode->i_sb, "ext4_get_inode_loc",
3730                                 "unable to read inode block - "
3731                                 "inode=%lu, block=%llu",
3732                                  inode->i_ino, block);
3733                 return -EIO;
3734         }
3735         if (!buffer_uptodate(bh)) {
3736                 lock_buffer(bh);
3737
3738                 /*
3739                  * If the buffer has the write error flag, we have failed
3740                  * to write out another inode in the same block.  In this
3741                  * case, we don't have to read the block because we may
3742                  * read the old inode data successfully.
3743                  */
3744                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
3745                         set_buffer_uptodate(bh);
3746
3747                 if (buffer_uptodate(bh)) {
3748                         /* someone brought it uptodate while we waited */
3749                         unlock_buffer(bh);
3750                         goto has_buffer;
3751                 }
3752
3753                 /*
3754                  * If we have all information of the inode in memory and this
3755                  * is the only valid inode in the block, we need not read the
3756                  * block.
3757                  */
3758                 if (in_mem) {
3759                         struct buffer_head *bitmap_bh;
3760                         struct ext4_group_desc *desc;
3761                         int inodes_per_buffer;
3762                         int inode_offset, i;
3763                         ext4_group_t block_group;
3764                         int start;
3765
3766                         block_group = (inode->i_ino - 1) /
3767                                         EXT4_INODES_PER_GROUP(inode->i_sb);
3768                         inodes_per_buffer = bh->b_size /
3769                                 EXT4_INODE_SIZE(inode->i_sb);
3770                         inode_offset = ((inode->i_ino - 1) %
3771                                         EXT4_INODES_PER_GROUP(inode->i_sb));
3772                         start = inode_offset & ~(inodes_per_buffer - 1);
3773
3774                         /* Is the inode bitmap in cache? */
3775                         desc = ext4_get_group_desc(inode->i_sb,
3776                                                 block_group, NULL);
3777                         if (!desc)
3778                                 goto make_io;
3779
3780                         bitmap_bh = sb_getblk(inode->i_sb,
3781                                 ext4_inode_bitmap(inode->i_sb, desc));
3782                         if (!bitmap_bh)
3783                                 goto make_io;
3784
3785                         /*
3786                          * If the inode bitmap isn't in cache then the
3787                          * optimisation may end up performing two reads instead
3788                          * of one, so skip it.
3789                          */
3790                         if (!buffer_uptodate(bitmap_bh)) {
3791                                 brelse(bitmap_bh);
3792                                 goto make_io;
3793                         }
3794                         for (i = start; i < start + inodes_per_buffer; i++) {
3795                                 if (i == inode_offset)
3796                                         continue;
3797                                 if (ext4_test_bit(i, bitmap_bh->b_data))
3798                                         break;
3799                         }
3800                         brelse(bitmap_bh);
3801                         if (i == start + inodes_per_buffer) {
3802                                 /* all other inodes are free, so skip I/O */
3803                                 memset(bh->b_data, 0, bh->b_size);
3804                                 set_buffer_uptodate(bh);
3805                                 unlock_buffer(bh);
3806                                 goto has_buffer;
3807                         }
3808                 }
3809
3810 make_io:
3811                 /*
3812                  * There are other valid inodes in the buffer, this inode
3813                  * has in-inode xattrs, or we don't have this inode in memory.
3814                  * Read the block from disk.
3815                  */
3816                 get_bh(bh);
3817                 bh->b_end_io = end_buffer_read_sync;
3818                 submit_bh(READ_META, bh);
3819                 wait_on_buffer(bh);
3820                 if (!buffer_uptodate(bh)) {
3821                         ext4_error(inode->i_sb, "ext4_get_inode_loc",
3822                                         "unable to read inode block - "
3823                                         "inode=%lu, block=%llu",
3824                                         inode->i_ino, block);
3825                         brelse(bh);
3826                         return -EIO;
3827                 }
3828         }
3829 has_buffer:
3830         iloc->bh = bh;
3831         return 0;
3832 }
3833
3834 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
3835 {
3836         /* We have all inode data except xattrs in memory here. */
3837         return __ext4_get_inode_loc(inode, iloc,
3838                 !(EXT4_I(inode)->i_state & EXT4_STATE_XATTR));
3839 }
3840
3841 void ext4_set_inode_flags(struct inode *inode)
3842 {
3843         unsigned int flags = EXT4_I(inode)->i_flags;
3844
3845         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
3846         if (flags & EXT4_SYNC_FL)
3847                 inode->i_flags |= S_SYNC;
3848         if (flags & EXT4_APPEND_FL)
3849                 inode->i_flags |= S_APPEND;
3850         if (flags & EXT4_IMMUTABLE_FL)
3851                 inode->i_flags |= S_IMMUTABLE;
3852         if (flags & EXT4_NOATIME_FL)
3853                 inode->i_flags |= S_NOATIME;
3854         if (flags & EXT4_DIRSYNC_FL)
3855                 inode->i_flags |= S_DIRSYNC;
3856 }
3857
3858 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
3859 void ext4_get_inode_flags(struct ext4_inode_info *ei)
3860 {
3861         unsigned int flags = ei->vfs_inode.i_flags;
3862
3863         ei->i_flags &= ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
3864                         EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|EXT4_DIRSYNC_FL);
3865         if (flags & S_SYNC)
3866                 ei->i_flags |= EXT4_SYNC_FL;
3867         if (flags & S_APPEND)
3868                 ei->i_flags |= EXT4_APPEND_FL;
3869         if (flags & S_IMMUTABLE)
3870                 ei->i_flags |= EXT4_IMMUTABLE_FL;
3871         if (flags & S_NOATIME)
3872                 ei->i_flags |= EXT4_NOATIME_FL;
3873         if (flags & S_DIRSYNC)
3874                 ei->i_flags |= EXT4_DIRSYNC_FL;
3875 }
3876 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
3877                                         struct ext4_inode_info *ei)
3878 {
3879         blkcnt_t i_blocks ;
3880         struct inode *inode = &(ei->vfs_inode);
3881         struct super_block *sb = inode->i_sb;
3882
3883         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
3884                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
3885                 /* we are using combined 48 bit field */
3886                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
3887                                         le32_to_cpu(raw_inode->i_blocks_lo);
3888                 if (ei->i_flags & EXT4_HUGE_FILE_FL) {
3889                         /* i_blocks represent file system block size */
3890                         return i_blocks  << (inode->i_blkbits - 9);
3891                 } else {
3892                         return i_blocks;
3893                 }
3894         } else {
3895                 return le32_to_cpu(raw_inode->i_blocks_lo);
3896         }
3897 }
3898
3899 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
3900 {
3901         struct ext4_iloc iloc;
3902         struct ext4_inode *raw_inode;
3903         struct ext4_inode_info *ei;
3904         struct buffer_head *bh;
3905         struct inode *inode;
3906         long ret;
3907         int block;
3908
3909         inode = iget_locked(sb, ino);
3910         if (!inode)
3911                 return ERR_PTR(-ENOMEM);
3912         if (!(inode->i_state & I_NEW))
3913                 return inode;
3914
3915         ei = EXT4_I(inode);
3916 #ifdef CONFIG_EXT4DEV_FS_POSIX_ACL
3917         ei->i_acl = EXT4_ACL_NOT_CACHED;
3918         ei->i_default_acl = EXT4_ACL_NOT_CACHED;
3919 #endif
3920         ei->i_block_alloc_info = NULL;
3921
3922         ret = __ext4_get_inode_loc(inode, &iloc, 0);
3923         if (ret < 0)
3924                 goto bad_inode;
3925         bh = iloc.bh;
3926         raw_inode = ext4_raw_inode(&iloc);
3927         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
3928         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
3929         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
3930         if(!(test_opt (inode->i_sb, NO_UID32))) {
3931                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
3932                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
3933         }
3934         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
3935
3936         ei->i_state = 0;
3937         ei->i_dir_start_lookup = 0;
3938         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
3939         /* We now have enough fields to check if the inode was active or not.
3940          * This is needed because nfsd might try to access dead inodes
3941          * the test is that same one that e2fsck uses
3942          * NeilBrown 1999oct15
3943          */
3944         if (inode->i_nlink == 0) {
3945                 if (inode->i_mode == 0 ||
3946                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
3947                         /* this inode is deleted */
3948                         brelse (bh);
3949                         ret = -ESTALE;
3950                         goto bad_inode;
3951                 }
3952                 /* The only unlinked inodes we let through here have
3953                  * valid i_mode and are being read by the orphan
3954                  * recovery code: that's fine, we're about to complete
3955                  * the process of deleting those. */
3956         }
3957         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
3958         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
3959         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
3960         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
3961             cpu_to_le32(EXT4_OS_HURD)) {
3962                 ei->i_file_acl |=
3963                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
3964         }
3965         inode->i_size = ext4_isize(raw_inode);
3966         ei->i_disksize = inode->i_size;
3967         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
3968         ei->i_block_group = iloc.block_group;
3969         /*
3970          * NOTE! The in-memory inode i_data array is in little-endian order
3971          * even on big-endian machines: we do NOT byteswap the block numbers!
3972          */
3973         for (block = 0; block < EXT4_N_BLOCKS; block++)
3974                 ei->i_data[block] = raw_inode->i_block[block];
3975         INIT_LIST_HEAD(&ei->i_orphan);
3976
3977         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
3978                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
3979                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
3980                     EXT4_INODE_SIZE(inode->i_sb)) {
3981                         brelse (bh);
3982                         ret = -EIO;
3983                         goto bad_inode;
3984                 }
3985                 if (ei->i_extra_isize == 0) {
3986                         /* The extra space is currently unused. Use it. */
3987                         ei->i_extra_isize = sizeof(struct ext4_inode) -
3988                                             EXT4_GOOD_OLD_INODE_SIZE;
3989                 } else {
3990                         __le32 *magic = (void *)raw_inode +
3991                                         EXT4_GOOD_OLD_INODE_SIZE +
3992                                         ei->i_extra_isize;
3993                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
3994                                  ei->i_state |= EXT4_STATE_XATTR;
3995                 }
3996         } else
3997                 ei->i_extra_isize = 0;
3998
3999         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4000         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4001         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4002         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4003
4004         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4005         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4006                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4007                         inode->i_version |=
4008                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4009         }
4010
4011         if (S_ISREG(inode->i_mode)) {
4012                 inode->i_op = &ext4_file_inode_operations;
4013                 inode->i_fop = &ext4_file_operations;
4014                 ext4_set_aops(inode);
4015         } else if (S_ISDIR(inode->i_mode)) {
4016                 inode->i_op = &ext4_dir_inode_operations;
4017                 inode->i_fop = &ext4_dir_operations;
4018         } else if (S_ISLNK(inode->i_mode)) {
4019                 if (ext4_inode_is_fast_symlink(inode))
4020                         inode->i_op = &ext4_fast_symlink_inode_operations;
4021                 else {
4022                         inode->i_op = &ext4_symlink_inode_operations;
4023                         ext4_set_aops(inode);
4024                 }
4025         } else {
4026                 inode->i_op = &ext4_special_inode_operations;
4027                 if (raw_inode->i_block[0])
4028                         init_special_inode(inode, inode->i_mode,
4029                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4030                 else
4031                         init_special_inode(inode, inode->i_mode,
4032                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4033         }
4034         brelse (iloc.bh);
4035         ext4_set_inode_flags(inode);
4036         unlock_new_inode(inode);
4037         return inode;
4038
4039 bad_inode:
4040         iget_failed(inode);
4041         return ERR_PTR(ret);
4042 }
4043
4044 static int ext4_inode_blocks_set(handle_t *handle,
4045                                 struct ext4_inode *raw_inode,
4046                                 struct ext4_inode_info *ei)
4047 {
4048         struct inode *inode = &(ei->vfs_inode);
4049         u64 i_blocks = inode->i_blocks;
4050         struct super_block *sb = inode->i_sb;
4051         int err = 0;
4052
4053         if (i_blocks <= ~0U) {
4054                 /*
4055                  * i_blocks can be represnted in a 32 bit variable
4056                  * as multiple of 512 bytes
4057                  */
4058                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4059                 raw_inode->i_blocks_high = 0;
4060                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4061         } else if (i_blocks <= 0xffffffffffffULL) {
4062                 /*
4063                  * i_blocks can be represented in a 48 bit variable
4064                  * as multiple of 512 bytes
4065                  */
4066                 err = ext4_update_rocompat_feature(handle, sb,
4067                                             EXT4_FEATURE_RO_COMPAT_HUGE_FILE);
4068                 if (err)
4069                         goto  err_out;
4070                 /* i_block is stored in the split  48 bit fields */
4071                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4072                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4073                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4074         } else {
4075                 /*
4076                  * i_blocks should be represented in a 48 bit variable
4077                  * as multiple of  file system block size
4078                  */
4079                 err = ext4_update_rocompat_feature(handle, sb,
4080                                             EXT4_FEATURE_RO_COMPAT_HUGE_FILE);
4081                 if (err)
4082                         goto  err_out;
4083                 ei->i_flags |= EXT4_HUGE_FILE_FL;
4084                 /* i_block is stored in file system block size */
4085                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4086                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4087                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4088         }
4089 err_out:
4090         return err;
4091 }
4092
4093 /*
4094  * Post the struct inode info into an on-disk inode location in the
4095  * buffer-cache.  This gobbles the caller's reference to the
4096  * buffer_head in the inode location struct.
4097  *
4098  * The caller must have write access to iloc->bh.
4099  */
4100 static int ext4_do_update_inode(handle_t *handle,
4101                                 struct inode *inode,
4102                                 struct ext4_iloc *iloc)
4103 {
4104         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
4105         struct ext4_inode_info *ei = EXT4_I(inode);
4106         struct buffer_head *bh = iloc->bh;
4107         int err = 0, rc, block;
4108
4109         /* For fields not not tracking in the in-memory inode,
4110          * initialise them to zero for new inodes. */
4111         if (ei->i_state & EXT4_STATE_NEW)
4112                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
4113
4114         ext4_get_inode_flags(ei);
4115         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
4116         if(!(test_opt(inode->i_sb, NO_UID32))) {
4117                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
4118                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
4119 /*
4120  * Fix up interoperability with old kernels. Otherwise, old inodes get
4121  * re-used with the upper 16 bits of the uid/gid intact
4122  */
4123                 if(!ei->i_dtime) {
4124                         raw_inode->i_uid_high =
4125                                 cpu_to_le16(high_16_bits(inode->i_uid));
4126                         raw_inode->i_gid_high =
4127                                 cpu_to_le16(high_16_bits(inode->i_gid));
4128                 } else {
4129                         raw_inode->i_uid_high = 0;
4130                         raw_inode->i_gid_high = 0;
4131                 }
4132         } else {
4133                 raw_inode->i_uid_low =
4134                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
4135                 raw_inode->i_gid_low =
4136                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
4137                 raw_inode->i_uid_high = 0;
4138                 raw_inode->i_gid_high = 0;
4139         }
4140         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
4141
4142         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
4143         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
4144         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
4145         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
4146
4147         if (ext4_inode_blocks_set(handle, raw_inode, ei))
4148                 goto out_brelse;
4149         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
4150         /* clear the migrate flag in the raw_inode */
4151         raw_inode->i_flags = cpu_to_le32(ei->i_flags & ~EXT4_EXT_MIGRATE);
4152         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
4153             cpu_to_le32(EXT4_OS_HURD))
4154                 raw_inode->i_file_acl_high =
4155                         cpu_to_le16(ei->i_file_acl >> 32);
4156         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
4157         ext4_isize_set(raw_inode, ei->i_disksize);
4158         if (ei->i_disksize > 0x7fffffffULL) {
4159                 struct super_block *sb = inode->i_sb;
4160                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
4161                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
4162                                 EXT4_SB(sb)->s_es->s_rev_level ==
4163                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
4164                         /* If this is the first large file
4165                          * created, add a flag to the superblock.
4166                          */
4167                         err = ext4_journal_get_write_access(handle,
4168                                         EXT4_SB(sb)->s_sbh);
4169                         if (err)
4170                                 goto out_brelse;
4171                         ext4_update_dynamic_rev(sb);
4172                         EXT4_SET_RO_COMPAT_FEATURE(sb,
4173                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
4174                         sb->s_dirt = 1;
4175                         handle->h_sync = 1;
4176                         err = ext4_journal_dirty_metadata(handle,
4177                                         EXT4_SB(sb)->s_sbh);
4178                 }
4179         }
4180         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
4181         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
4182                 if (old_valid_dev(inode->i_rdev)) {
4183                         raw_inode->i_block[0] =
4184                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
4185                         raw_inode->i_block[1] = 0;
4186                 } else {
4187                         raw_inode->i_block[0] = 0;
4188                         raw_inode->i_block[1] =
4189                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
4190                         raw_inode->i_block[2] = 0;
4191                 }
4192         } else for (block = 0; block < EXT4_N_BLOCKS; block++)
4193                 raw_inode->i_block[block] = ei->i_data[block];
4194
4195         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
4196         if (ei->i_extra_isize) {
4197                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4198                         raw_inode->i_version_hi =
4199                         cpu_to_le32(inode->i_version >> 32);
4200                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
4201         }
4202
4203
4204         BUFFER_TRACE(bh, "call ext4_journal_dirty_metadata");
4205         rc = ext4_journal_dirty_metadata(handle, bh);
4206         if (!err)
4207                 err = rc;
4208         ei->i_state &= ~EXT4_STATE_NEW;
4209
4210 out_brelse:
4211         brelse (bh);
4212         ext4_std_error(inode->i_sb, err);
4213         return err;
4214 }
4215
4216 /*
4217  * ext4_write_inode()
4218  *
4219  * We are called from a few places:
4220  *
4221  * - Within generic_file_write() for O_SYNC files.
4222  *   Here, there will be no transaction running. We wait for any running
4223  *   trasnaction to commit.
4224  *
4225  * - Within sys_sync(), kupdate and such.
4226  *   We wait on commit, if tol to.
4227  *
4228  * - Within prune_icache() (PF_MEMALLOC == true)
4229  *   Here we simply return.  We can't afford to block kswapd on the
4230  *   journal commit.
4231  *
4232  * In all cases it is actually safe for us to return without doing anything,
4233  * because the inode has been copied into a raw inode buffer in
4234  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
4235  * knfsd.
4236  *
4237  * Note that we are absolutely dependent upon all inode dirtiers doing the
4238  * right thing: they *must* call mark_inode_dirty() after dirtying info in
4239  * which we are interested.
4240  *
4241  * It would be a bug for them to not do this.  The code:
4242  *
4243  *      mark_inode_dirty(inode)
4244  *      stuff();
4245  *      inode->i_size = expr;
4246  *
4247  * is in error because a kswapd-driven write_inode() could occur while
4248  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
4249  * will no longer be on the superblock's dirty inode list.
4250  */
4251 int ext4_write_inode(struct inode *inode, int wait)
4252 {
4253         if (current->flags & PF_MEMALLOC)
4254                 return 0;
4255
4256         if (ext4_journal_current_handle()) {
4257                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
4258                 dump_stack();
4259                 return -EIO;
4260         }
4261
4262         if (!wait)
4263                 return 0;
4264
4265         return ext4_force_commit(inode->i_sb);
4266 }
4267
4268 /*
4269  * ext4_setattr()
4270  *
4271  * Called from notify_change.
4272  *
4273  * We want to trap VFS attempts to truncate the file as soon as
4274  * possible.  In particular, we want to make sure that when the VFS
4275  * shrinks i_size, we put the inode on the orphan list and modify
4276  * i_disksize immediately, so that during the subsequent flushing of
4277  * dirty pages and freeing of disk blocks, we can guarantee that any
4278  * commit will leave the blocks being flushed in an unused state on
4279  * disk.  (On recovery, the inode will get truncated and the blocks will
4280  * be freed, so we have a strong guarantee that no future commit will
4281  * leave these blocks visible to the user.)
4282  *
4283  * Another thing we have to assure is that if we are in ordered mode
4284  * and inode is still attached to the committing transaction, we must
4285  * we start writeout of all the dirty pages which are being truncated.
4286  * This way we are sure that all the data written in the previous
4287  * transaction are already on disk (truncate waits for pages under
4288  * writeback).
4289  *
4290  * Called with inode->i_mutex down.
4291  */
4292 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
4293 {
4294         struct inode *inode = dentry->d_inode;
4295         int error, rc = 0;
4296         const unsigned int ia_valid = attr->ia_valid;
4297
4298         error = inode_change_ok(inode, attr);
4299         if (error)
4300                 return error;
4301
4302         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
4303                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
4304                 handle_t *handle;
4305
4306                 /* (user+group)*(old+new) structure, inode write (sb,
4307                  * inode block, ? - but truncate inode update has it) */
4308                 handle = ext4_journal_start(inode, 2*(EXT4_QUOTA_INIT_BLOCKS(inode->i_sb)+
4309                                         EXT4_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
4310                 if (IS_ERR(handle)) {
4311                         error = PTR_ERR(handle);
4312                         goto err_out;
4313                 }
4314                 error = DQUOT_TRANSFER(inode, attr) ? -EDQUOT : 0;
4315                 if (error) {
4316                         ext4_journal_stop(handle);
4317                         return error;
4318                 }
4319                 /* Update corresponding info in inode so that everything is in
4320                  * one transaction */
4321                 if (attr->ia_valid & ATTR_UID)
4322                         inode->i_uid = attr->ia_uid;
4323                 if (attr->ia_valid & ATTR_GID)
4324                         inode->i_gid = attr->ia_gid;
4325                 error = ext4_mark_inode_dirty(handle, inode);
4326                 ext4_journal_stop(handle);
4327         }
4328
4329         if (attr->ia_valid & ATTR_SIZE) {
4330                 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)) {
4331                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4332
4333                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
4334                                 error = -EFBIG;
4335                                 goto err_out;
4336                         }
4337                 }
4338         }
4339
4340         if (S_ISREG(inode->i_mode) &&
4341             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
4342                 handle_t *handle;
4343
4344                 handle = ext4_journal_start(inode, 3);
4345                 if (IS_ERR(handle)) {
4346                         error = PTR_ERR(handle);
4347                         goto err_out;
4348                 }
4349
4350                 error = ext4_orphan_add(handle, inode);
4351                 EXT4_I(inode)->i_disksize = attr->ia_size;
4352                 rc = ext4_mark_inode_dirty(handle, inode);
4353                 if (!error)
4354                         error = rc;
4355                 ext4_journal_stop(handle);
4356
4357                 if (ext4_should_order_data(inode)) {
4358                         error = ext4_begin_ordered_truncate(inode,
4359                                                             attr->ia_size);
4360                         if (error) {
4361                                 /* Do as much error cleanup as possible */
4362                                 handle = ext4_journal_start(inode, 3);
4363                                 if (IS_ERR(handle)) {
4364                                         ext4_orphan_del(NULL, inode);
4365                                         goto err_out;
4366                                 }
4367                                 ext4_orphan_del(handle, inode);
4368                                 ext4_journal_stop(handle);
4369                                 goto err_out;
4370                         }
4371                 }
4372         }
4373
4374         rc = inode_setattr(inode, attr);
4375
4376         /* If inode_setattr's call to ext4_truncate failed to get a
4377          * transaction handle at all, we need to clean up the in-core
4378          * orphan list manually. */
4379         if (inode->i_nlink)
4380                 ext4_orphan_del(NULL, inode);
4381
4382         if (!rc && (ia_valid & ATTR_MODE))
4383                 rc = ext4_acl_chmod(inode);
4384
4385 err_out:
4386         ext4_std_error(inode->i_sb, error);
4387         if (!error)
4388                 error = rc;
4389         return error;
4390 }
4391
4392 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
4393                  struct kstat *stat)
4394 {
4395         struct inode *inode;
4396         unsigned long delalloc_blocks;
4397
4398         inode = dentry->d_inode;
4399         generic_fillattr(inode, stat);
4400
4401         /*
4402          * We can't update i_blocks if the block allocation is delayed
4403          * otherwise in the case of system crash before the real block
4404          * allocation is done, we will have i_blocks inconsistent with
4405          * on-disk file blocks.
4406          * We always keep i_blocks updated together with real
4407          * allocation. But to not confuse with user, stat
4408          * will return the blocks that include the delayed allocation
4409          * blocks for this file.
4410          */
4411         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
4412         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
4413         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
4414
4415         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
4416         return 0;
4417 }
4418
4419 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
4420                                       int chunk)
4421 {
4422         int indirects;
4423
4424         /* if nrblocks are contiguous */
4425         if (chunk) {
4426                 /*
4427                  * With N contiguous data blocks, it need at most
4428                  * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) indirect blocks
4429                  * 2 dindirect blocks
4430                  * 1 tindirect block
4431                  */
4432                 indirects = nrblocks / EXT4_ADDR_PER_BLOCK(inode->i_sb);
4433                 return indirects + 3;
4434         }
4435         /*
4436          * if nrblocks are not contiguous, worse case, each block touch
4437          * a indirect block, and each indirect block touch a double indirect
4438          * block, plus a triple indirect block
4439          */
4440         indirects = nrblocks * 2 + 1;
4441         return indirects;
4442 }
4443
4444 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4445 {
4446         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL))
4447                 return ext4_indirect_trans_blocks(inode, nrblocks, 0);
4448         return ext4_ext_index_trans_blocks(inode, nrblocks, 0);
4449 }
4450 /*
4451  * Account for index blocks, block groups bitmaps and block group
4452  * descriptor blocks if modify datablocks and index blocks
4453  * worse case, the indexs blocks spread over different block groups
4454  *
4455  * If datablocks are discontiguous, they are possible to spread over
4456  * different block groups too. If they are contiugous, with flexbg,
4457  * they could still across block group boundary.
4458  *
4459  * Also account for superblock, inode, quota and xattr blocks
4460  */
4461 int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4462 {
4463         int groups, gdpblocks;
4464         int idxblocks;
4465         int ret = 0;
4466
4467         /*
4468          * How many index blocks need to touch to modify nrblocks?
4469          * The "Chunk" flag indicating whether the nrblocks is
4470          * physically contiguous on disk
4471          *
4472          * For Direct IO and fallocate, they calls get_block to allocate
4473          * one single extent at a time, so they could set the "Chunk" flag
4474          */
4475         idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
4476
4477         ret = idxblocks;
4478
4479         /*
4480          * Now let's see how many group bitmaps and group descriptors need
4481          * to account
4482          */
4483         groups = idxblocks;
4484         if (chunk)
4485                 groups += 1;
4486         else
4487                 groups += nrblocks;
4488
4489         gdpblocks = groups;
4490         if (groups > EXT4_SB(inode->i_sb)->s_groups_count)
4491                 groups = EXT4_SB(inode->i_sb)->s_groups_count;
4492         if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
4493                 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
4494
4495         /* bitmaps and block group descriptor blocks */
4496         ret += groups + gdpblocks;
4497
4498         /* Blocks for super block, inode, quota and xattr blocks */
4499         ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
4500
4501         return ret;
4502 }
4503
4504 /*
4505  * Calulate the total number of credits to reserve to fit
4506  * the modification of a single pages into a single transaction,
4507  * which may include multiple chunks of block allocations.
4508  *
4509  * This could be called via ext4_write_begin()
4510  *
4511  * We need to consider the worse case, when
4512  * one new block per extent.
4513  */
4514 int ext4_writepage_trans_blocks(struct inode *inode)
4515 {
4516         int bpp = ext4_journal_blocks_per_page(inode);
4517         int ret;
4518
4519         ret = ext4_meta_trans_blocks(inode, bpp, 0);
4520
4521         /* Account for data blocks for journalled mode */
4522         if (ext4_should_journal_data(inode))
4523                 ret += bpp;
4524         return ret;
4525 }
4526
4527 /*
4528  * Calculate the journal credits for a chunk of data modification.
4529  *
4530  * This is called from DIO, fallocate or whoever calling
4531  * ext4_get_blocks_wrap() to map/allocate a chunk of contigous disk blocks.
4532  *
4533  * journal buffers for data blocks are not included here, as DIO
4534  * and fallocate do no need to journal data buffers.
4535  */
4536 int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
4537 {
4538         return ext4_meta_trans_blocks(inode, nrblocks, 1);
4539 }
4540
4541 /*
4542  * The caller must have previously called ext4_reserve_inode_write().
4543  * Give this, we know that the caller already has write access to iloc->bh.
4544  */
4545 int ext4_mark_iloc_dirty(handle_t *handle,
4546                 struct inode *inode, struct ext4_iloc *iloc)
4547 {
4548         int err = 0;
4549
4550         if (test_opt(inode->i_sb, I_VERSION))
4551                 inode_inc_iversion(inode);
4552
4553         /* the do_update_inode consumes one bh->b_count */
4554         get_bh(iloc->bh);
4555
4556         /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
4557         err = ext4_do_update_inode(handle, inode, iloc);
4558         put_bh(iloc->bh);
4559         return err;
4560 }
4561
4562 /*
4563  * On success, We end up with an outstanding reference count against
4564  * iloc->bh.  This _must_ be cleaned up later.
4565  */
4566
4567 int
4568 ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
4569                          struct ext4_iloc *iloc)
4570 {
4571         int err = 0;
4572         if (handle) {
4573                 err = ext4_get_inode_loc(inode, iloc);
4574                 if (!err) {
4575                         BUFFER_TRACE(iloc->bh, "get_write_access");
4576                         err = ext4_journal_get_write_access(handle, iloc->bh);
4577                         if (err) {
4578                                 brelse(iloc->bh);
4579                                 iloc->bh = NULL;
4580                         }
4581                 }
4582         }
4583         ext4_std_error(inode->i_sb, err);
4584         return err;
4585 }
4586
4587 /*
4588  * Expand an inode by new_extra_isize bytes.
4589  * Returns 0 on success or negative error number on failure.
4590  */
4591 static int ext4_expand_extra_isize(struct inode *inode,
4592                                    unsigned int new_extra_isize,
4593                                    struct ext4_iloc iloc,
4594                                    handle_t *handle)
4595 {
4596         struct ext4_inode *raw_inode;
4597         struct ext4_xattr_ibody_header *header;
4598         struct ext4_xattr_entry *entry;
4599
4600         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
4601                 return 0;
4602
4603         raw_inode = ext4_raw_inode(&iloc);
4604
4605         header = IHDR(inode, raw_inode);
4606         entry = IFIRST(header);
4607
4608         /* No extended attributes present */
4609         if (!(EXT4_I(inode)->i_state & EXT4_STATE_XATTR) ||
4610                 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
4611                 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
4612                         new_extra_isize);
4613                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
4614                 return 0;
4615         }
4616
4617         /* try to expand with EAs present */
4618         return ext4_expand_extra_isize_ea(inode, new_extra_isize,
4619                                           raw_inode, handle);
4620 }
4621
4622 /*
4623  * What we do here is to mark the in-core inode as clean with respect to inode
4624  * dirtiness (it may still be data-dirty).
4625  * This means that the in-core inode may be reaped by prune_icache
4626  * without having to perform any I/O.  This is a very good thing,
4627  * because *any* task may call prune_icache - even ones which
4628  * have a transaction open against a different journal.
4629  *
4630  * Is this cheating?  Not really.  Sure, we haven't written the
4631  * inode out, but prune_icache isn't a user-visible syncing function.
4632  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
4633  * we start and wait on commits.
4634  *
4635  * Is this efficient/effective?  Well, we're being nice to the system
4636  * by cleaning up our inodes proactively so they can be reaped
4637  * without I/O.  But we are potentially leaving up to five seconds'
4638  * worth of inodes floating about which prune_icache wants us to
4639  * write out.  One way to fix that would be to get prune_icache()
4640  * to do a write_super() to free up some memory.  It has the desired
4641  * effect.
4642  */
4643 int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
4644 {
4645         struct ext4_iloc iloc;
4646         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4647         static unsigned int mnt_count;
4648         int err, ret;
4649
4650         might_sleep();
4651         err = ext4_reserve_inode_write(handle, inode, &iloc);
4652         if (EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
4653             !(EXT4_I(inode)->i_state & EXT4_STATE_NO_EXPAND)) {
4654                 /*
4655                  * We need extra buffer credits since we may write into EA block
4656                  * with this same handle. If journal_extend fails, then it will
4657                  * only result in a minor loss of functionality for that inode.
4658                  * If this is felt to be critical, then e2fsck should be run to
4659                  * force a large enough s_min_extra_isize.
4660                  */
4661                 if ((jbd2_journal_extend(handle,
4662                              EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
4663                         ret = ext4_expand_extra_isize(inode,
4664                                                       sbi->s_want_extra_isize,
4665                                                       iloc, handle);
4666                         if (ret) {
4667                                 EXT4_I(inode)->i_state |= EXT4_STATE_NO_EXPAND;
4668                                 if (mnt_count !=
4669                                         le16_to_cpu(sbi->s_es->s_mnt_count)) {
4670                                         ext4_warning(inode->i_sb, __func__,
4671                                         "Unable to expand inode %lu. Delete"
4672                                         " some EAs or run e2fsck.",
4673                                         inode->i_ino);
4674                                         mnt_count =
4675                                           le16_to_cpu(sbi->s_es->s_mnt_count);
4676                                 }
4677                         }
4678                 }
4679         }
4680         if (!err)
4681                 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
4682         return err;
4683 }
4684
4685 /*
4686  * ext4_dirty_inode() is called from __mark_inode_dirty()
4687  *
4688  * We're really interested in the case where a file is being extended.
4689  * i_size has been changed by generic_commit_write() and we thus need
4690  * to include the updated inode in the current transaction.
4691  *
4692  * Also, DQUOT_ALLOC_SPACE() will always dirty the inode when blocks
4693  * are allocated to the file.
4694  *
4695  * If the inode is marked synchronous, we don't honour that here - doing
4696  * so would cause a commit on atime updates, which we don't bother doing.
4697  * We handle synchronous inodes at the highest possible level.
4698  */
4699 void ext4_dirty_inode(struct inode *inode)
4700 {
4701         handle_t *current_handle = ext4_journal_current_handle();
4702         handle_t *handle;
4703
4704         handle = ext4_journal_start(inode, 2);
4705         if (IS_ERR(handle))
4706                 goto out;
4707         if (current_handle &&
4708                 current_handle->h_transaction != handle->h_transaction) {
4709                 /* This task has a transaction open against a different fs */
4710                 printk(KERN_EMERG "%s: transactions do not match!\n",
4711                        __func__);
4712         } else {
4713                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
4714                                 current_handle);
4715                 ext4_mark_inode_dirty(handle, inode);
4716         }
4717         ext4_journal_stop(handle);
4718 out:
4719         return;
4720 }
4721
4722 #if 0
4723 /*
4724  * Bind an inode's backing buffer_head into this transaction, to prevent
4725  * it from being flushed to disk early.  Unlike
4726  * ext4_reserve_inode_write, this leaves behind no bh reference and
4727  * returns no iloc structure, so the caller needs to repeat the iloc
4728  * lookup to mark the inode dirty later.
4729  */
4730 static int ext4_pin_inode(handle_t *handle, struct inode *inode)
4731 {
4732         struct ext4_iloc iloc;
4733
4734         int err = 0;
4735         if (handle) {
4736                 err = ext4_get_inode_loc(inode, &iloc);
4737                 if (!err) {
4738                         BUFFER_TRACE(iloc.bh, "get_write_access");
4739                         err = jbd2_journal_get_write_access(handle, iloc.bh);
4740                         if (!err)
4741                                 err = ext4_journal_dirty_metadata(handle,
4742                                                                   iloc.bh);
4743                         brelse(iloc.bh);
4744                 }
4745         }
4746         ext4_std_error(inode->i_sb, err);
4747         return err;
4748 }
4749 #endif
4750
4751 int ext4_change_inode_journal_flag(struct inode *inode, int val)
4752 {
4753         journal_t *journal;
4754         handle_t *handle;
4755         int err;
4756
4757         /*
4758          * We have to be very careful here: changing a data block's
4759          * journaling status dynamically is dangerous.  If we write a
4760          * data block to the journal, change the status and then delete
4761          * that block, we risk forgetting to revoke the old log record
4762          * from the journal and so a subsequent replay can corrupt data.
4763          * So, first we make sure that the journal is empty and that
4764          * nobody is changing anything.
4765          */
4766
4767         journal = EXT4_JOURNAL(inode);
4768         if (is_journal_aborted(journal))
4769                 return -EROFS;
4770
4771         jbd2_journal_lock_updates(journal);
4772         jbd2_journal_flush(journal);
4773
4774         /*
4775          * OK, there are no updates running now, and all cached data is
4776          * synced to disk.  We are now in a completely consistent state
4777          * which doesn't have anything in the journal, and we know that
4778          * no filesystem updates are running, so it is safe to modify
4779          * the inode's in-core data-journaling state flag now.
4780          */
4781
4782         if (val)
4783                 EXT4_I(inode)->i_flags |= EXT4_JOURNAL_DATA_FL;
4784         else
4785                 EXT4_I(inode)->i_flags &= ~EXT4_JOURNAL_DATA_FL;
4786         ext4_set_aops(inode);
4787
4788         jbd2_journal_unlock_updates(journal);
4789
4790         /* Finally we can mark the inode as dirty. */
4791
4792         handle = ext4_journal_start(inode, 1);
4793         if (IS_ERR(handle))
4794                 return PTR_ERR(handle);
4795
4796         err = ext4_mark_inode_dirty(handle, inode);
4797         handle->h_sync = 1;
4798         ext4_journal_stop(handle);
4799         ext4_std_error(inode->i_sb, err);
4800
4801         return err;
4802 }
4803
4804 static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
4805 {
4806         return !buffer_mapped(bh);
4807 }
4808
4809 int ext4_page_mkwrite(struct vm_area_struct *vma, struct page *page)
4810 {
4811         loff_t size;
4812         unsigned long len;
4813         int ret = -EINVAL;
4814         struct file *file = vma->vm_file;
4815         struct inode *inode = file->f_path.dentry->d_inode;
4816         struct address_space *mapping = inode->i_mapping;
4817
4818         /*
4819          * Get i_alloc_sem to stop truncates messing with the inode. We cannot
4820          * get i_mutex because we are already holding mmap_sem.
4821          */
4822         down_read(&inode->i_alloc_sem);
4823         size = i_size_read(inode);
4824         if (page->mapping != mapping || size <= page_offset(page)
4825             || !PageUptodate(page)) {
4826                 /* page got truncated from under us? */
4827                 goto out_unlock;
4828         }
4829         ret = 0;
4830         if (PageMappedToDisk(page))
4831                 goto out_unlock;
4832
4833         if (page->index == size >> PAGE_CACHE_SHIFT)
4834                 len = size & ~PAGE_CACHE_MASK;
4835         else
4836                 len = PAGE_CACHE_SIZE;
4837
4838         if (page_has_buffers(page)) {
4839                 /* return if we have all the buffers mapped */
4840                 if (!walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
4841                                        ext4_bh_unmapped))
4842                         goto out_unlock;
4843         }
4844         /*
4845          * OK, we need to fill the hole... Do write_begin write_end
4846          * to do block allocation/reservation.We are not holding
4847          * inode.i__mutex here. That allow * parallel write_begin,
4848          * write_end call. lock_page prevent this from happening
4849          * on the same page though
4850          */
4851         ret = mapping->a_ops->write_begin(file, mapping, page_offset(page),
4852                         len, AOP_FLAG_UNINTERRUPTIBLE, &page, NULL);
4853         if (ret < 0)
4854                 goto out_unlock;
4855         ret = mapping->a_ops->write_end(file, mapping, page_offset(page),
4856                         len, len, page, NULL);
4857         if (ret < 0)
4858                 goto out_unlock;
4859         ret = 0;
4860 out_unlock:
4861         up_read(&inode->i_alloc_sem);
4862         return ret;
4863 }