Merge branches 'core/debugobjects', 'core/iommu', 'core/locking', 'core/printk',...
[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         /*
490          * XXX need to get goal block from mballoc's data structures
491          */
492
493         return ext4_find_near(inode, partial);
494 }
495
496 /**
497  *      ext4_blks_to_allocate: Look up the block map and count the number
498  *      of direct blocks need to be allocated for the given branch.
499  *
500  *      @branch: chain of indirect blocks
501  *      @k: number of blocks need for indirect blocks
502  *      @blks: number of data blocks to be mapped.
503  *      @blocks_to_boundary:  the offset in the indirect block
504  *
505  *      return the total number of blocks to be allocate, including the
506  *      direct and indirect blocks.
507  */
508 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned long blks,
509                 int blocks_to_boundary)
510 {
511         unsigned long count = 0;
512
513         /*
514          * Simple case, [t,d]Indirect block(s) has not allocated yet
515          * then it's clear blocks on that path have not allocated
516          */
517         if (k > 0) {
518                 /* right now we don't handle cross boundary allocation */
519                 if (blks < blocks_to_boundary + 1)
520                         count += blks;
521                 else
522                         count += blocks_to_boundary + 1;
523                 return count;
524         }
525
526         count++;
527         while (count < blks && count <= blocks_to_boundary &&
528                 le32_to_cpu(*(branch[0].p + count)) == 0) {
529                 count++;
530         }
531         return count;
532 }
533
534 /**
535  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
536  *      @indirect_blks: the number of blocks need to allocate for indirect
537  *                      blocks
538  *
539  *      @new_blocks: on return it will store the new block numbers for
540  *      the indirect blocks(if needed) and the first direct block,
541  *      @blks:  on return it will store the total number of allocated
542  *              direct blocks
543  */
544 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
545                                 ext4_lblk_t iblock, ext4_fsblk_t goal,
546                                 int indirect_blks, int blks,
547                                 ext4_fsblk_t new_blocks[4], int *err)
548 {
549         int target, i;
550         unsigned long count = 0, blk_allocated = 0;
551         int index = 0;
552         ext4_fsblk_t current_block = 0;
553         int ret = 0;
554
555         /*
556          * Here we try to allocate the requested multiple blocks at once,
557          * on a best-effort basis.
558          * To build a branch, we should allocate blocks for
559          * the indirect blocks(if not allocated yet), and at least
560          * the first direct block of this branch.  That's the
561          * minimum number of blocks need to allocate(required)
562          */
563         /* first we try to allocate the indirect blocks */
564         target = indirect_blks;
565         while (target > 0) {
566                 count = target;
567                 /* allocating blocks for indirect blocks and direct blocks */
568                 current_block = ext4_new_meta_blocks(handle, inode,
569                                                         goal, &count, err);
570                 if (*err)
571                         goto failed_out;
572
573                 target -= count;
574                 /* allocate blocks for indirect blocks */
575                 while (index < indirect_blks && count) {
576                         new_blocks[index++] = current_block++;
577                         count--;
578                 }
579                 if (count > 0) {
580                         /*
581                          * save the new block number
582                          * for the first direct block
583                          */
584                         new_blocks[index] = current_block;
585                         printk(KERN_INFO "%s returned more blocks than "
586                                                 "requested\n", __func__);
587                         WARN_ON(1);
588                         break;
589                 }
590         }
591
592         target = blks - count ;
593         blk_allocated = count;
594         if (!target)
595                 goto allocated;
596         /* Now allocate data blocks */
597         count = target;
598         /* allocating blocks for data blocks */
599         current_block = ext4_new_blocks(handle, inode, iblock,
600                                                 goal, &count, err);
601         if (*err && (target == blks)) {
602                 /*
603                  * if the allocation failed and we didn't allocate
604                  * any blocks before
605                  */
606                 goto failed_out;
607         }
608         if (!*err) {
609                 if (target == blks) {
610                 /*
611                  * save the new block number
612                  * for the first direct block
613                  */
614                         new_blocks[index] = current_block;
615                 }
616                 blk_allocated += count;
617         }
618 allocated:
619         /* total number of blocks allocated for direct blocks */
620         ret = blk_allocated;
621         *err = 0;
622         return ret;
623 failed_out:
624         for (i = 0; i < index; i++)
625                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
626         return ret;
627 }
628
629 /**
630  *      ext4_alloc_branch - allocate and set up a chain of blocks.
631  *      @inode: owner
632  *      @indirect_blks: number of allocated indirect blocks
633  *      @blks: number of allocated direct blocks
634  *      @offsets: offsets (in the blocks) to store the pointers to next.
635  *      @branch: place to store the chain in.
636  *
637  *      This function allocates blocks, zeroes out all but the last one,
638  *      links them into chain and (if we are synchronous) writes them to disk.
639  *      In other words, it prepares a branch that can be spliced onto the
640  *      inode. It stores the information about that chain in the branch[], in
641  *      the same format as ext4_get_branch() would do. We are calling it after
642  *      we had read the existing part of chain and partial points to the last
643  *      triple of that (one with zero ->key). Upon the exit we have the same
644  *      picture as after the successful ext4_get_block(), except that in one
645  *      place chain is disconnected - *branch->p is still zero (we did not
646  *      set the last link), but branch->key contains the number that should
647  *      be placed into *branch->p to fill that gap.
648  *
649  *      If allocation fails we free all blocks we've allocated (and forget
650  *      their buffer_heads) and return the error value the from failed
651  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
652  *      as described above and return 0.
653  */
654 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
655                                 ext4_lblk_t iblock, int indirect_blks,
656                                 int *blks, ext4_fsblk_t goal,
657                                 ext4_lblk_t *offsets, Indirect *branch)
658 {
659         int blocksize = inode->i_sb->s_blocksize;
660         int i, n = 0;
661         int err = 0;
662         struct buffer_head *bh;
663         int num;
664         ext4_fsblk_t new_blocks[4];
665         ext4_fsblk_t current_block;
666
667         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
668                                 *blks, new_blocks, &err);
669         if (err)
670                 return err;
671
672         branch[0].key = cpu_to_le32(new_blocks[0]);
673         /*
674          * metadata blocks and data blocks are allocated.
675          */
676         for (n = 1; n <= indirect_blks;  n++) {
677                 /*
678                  * Get buffer_head for parent block, zero it out
679                  * and set the pointer to new one, then send
680                  * parent to disk.
681                  */
682                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
683                 branch[n].bh = bh;
684                 lock_buffer(bh);
685                 BUFFER_TRACE(bh, "call get_create_access");
686                 err = ext4_journal_get_create_access(handle, bh);
687                 if (err) {
688                         unlock_buffer(bh);
689                         brelse(bh);
690                         goto failed;
691                 }
692
693                 memset(bh->b_data, 0, blocksize);
694                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
695                 branch[n].key = cpu_to_le32(new_blocks[n]);
696                 *branch[n].p = branch[n].key;
697                 if (n == indirect_blks) {
698                         current_block = new_blocks[n];
699                         /*
700                          * End of chain, update the last new metablock of
701                          * the chain to point to the new allocated
702                          * data blocks numbers
703                          */
704                         for (i=1; i < num; i++)
705                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
706                 }
707                 BUFFER_TRACE(bh, "marking uptodate");
708                 set_buffer_uptodate(bh);
709                 unlock_buffer(bh);
710
711                 BUFFER_TRACE(bh, "call ext4_journal_dirty_metadata");
712                 err = ext4_journal_dirty_metadata(handle, bh);
713                 if (err)
714                         goto failed;
715         }
716         *blks = num;
717         return err;
718 failed:
719         /* Allocation failed, free what we already allocated */
720         for (i = 1; i <= n ; i++) {
721                 BUFFER_TRACE(branch[i].bh, "call jbd2_journal_forget");
722                 ext4_journal_forget(handle, branch[i].bh);
723         }
724         for (i = 0; i < indirect_blks; i++)
725                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
726
727         ext4_free_blocks(handle, inode, new_blocks[i], num, 0);
728
729         return err;
730 }
731
732 /**
733  * ext4_splice_branch - splice the allocated branch onto inode.
734  * @inode: owner
735  * @block: (logical) number of block we are adding
736  * @chain: chain of indirect blocks (with a missing link - see
737  *      ext4_alloc_branch)
738  * @where: location of missing link
739  * @num:   number of indirect blocks we are adding
740  * @blks:  number of direct blocks we are adding
741  *
742  * This function fills the missing link and does all housekeeping needed in
743  * inode (->i_blocks, etc.). In case of success we end up with the full
744  * chain to new block and return 0.
745  */
746 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
747                         ext4_lblk_t block, Indirect *where, int num, int blks)
748 {
749         int i;
750         int err = 0;
751         ext4_fsblk_t current_block;
752
753         /*
754          * If we're splicing into a [td]indirect block (as opposed to the
755          * inode) then we need to get write access to the [td]indirect block
756          * before the splice.
757          */
758         if (where->bh) {
759                 BUFFER_TRACE(where->bh, "get_write_access");
760                 err = ext4_journal_get_write_access(handle, where->bh);
761                 if (err)
762                         goto err_out;
763         }
764         /* That's it */
765
766         *where->p = where->key;
767
768         /*
769          * Update the host buffer_head or inode to point to more just allocated
770          * direct blocks blocks
771          */
772         if (num == 0 && blks > 1) {
773                 current_block = le32_to_cpu(where->key) + 1;
774                 for (i = 1; i < blks; i++)
775                         *(where->p + i) = cpu_to_le32(current_block++);
776         }
777
778         /* We are done with atomic stuff, now do the rest of housekeeping */
779
780         inode->i_ctime = ext4_current_time(inode);
781         ext4_mark_inode_dirty(handle, inode);
782
783         /* had we spliced it onto indirect block? */
784         if (where->bh) {
785                 /*
786                  * If we spliced it onto an indirect block, we haven't
787                  * altered the inode.  Note however that if it is being spliced
788                  * onto an indirect block at the very end of the file (the
789                  * file is growing) then we *will* alter the inode to reflect
790                  * the new i_size.  But that is not done here - it is done in
791                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
792                  */
793                 jbd_debug(5, "splicing indirect only\n");
794                 BUFFER_TRACE(where->bh, "call ext4_journal_dirty_metadata");
795                 err = ext4_journal_dirty_metadata(handle, where->bh);
796                 if (err)
797                         goto err_out;
798         } else {
799                 /*
800                  * OK, we spliced it into the inode itself on a direct block.
801                  * Inode was dirtied above.
802                  */
803                 jbd_debug(5, "splicing direct\n");
804         }
805         return err;
806
807 err_out:
808         for (i = 1; i <= num; i++) {
809                 BUFFER_TRACE(where[i].bh, "call jbd2_journal_forget");
810                 ext4_journal_forget(handle, where[i].bh);
811                 ext4_free_blocks(handle, inode,
812                                         le32_to_cpu(where[i-1].key), 1, 0);
813         }
814         ext4_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks, 0);
815
816         return err;
817 }
818
819 /*
820  * Allocation strategy is simple: if we have to allocate something, we will
821  * have to go the whole way to leaf. So let's do it before attaching anything
822  * to tree, set linkage between the newborn blocks, write them if sync is
823  * required, recheck the path, free and repeat if check fails, otherwise
824  * set the last missing link (that will protect us from any truncate-generated
825  * removals - all blocks on the path are immune now) and possibly force the
826  * write on the parent block.
827  * That has a nice additional property: no special recovery from the failed
828  * allocations is needed - we simply release blocks and do not touch anything
829  * reachable from inode.
830  *
831  * `handle' can be NULL if create == 0.
832  *
833  * return > 0, # of blocks mapped or allocated.
834  * return = 0, if plain lookup failed.
835  * return < 0, error case.
836  *
837  *
838  * Need to be called with
839  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
840  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
841  */
842 int ext4_get_blocks_handle(handle_t *handle, struct inode *inode,
843                 ext4_lblk_t iblock, unsigned long maxblocks,
844                 struct buffer_head *bh_result,
845                 int create, int extend_disksize)
846 {
847         int err = -EIO;
848         ext4_lblk_t offsets[4];
849         Indirect chain[4];
850         Indirect *partial;
851         ext4_fsblk_t goal;
852         int indirect_blks;
853         int blocks_to_boundary = 0;
854         int depth;
855         struct ext4_inode_info *ei = EXT4_I(inode);
856         int count = 0;
857         ext4_fsblk_t first_block = 0;
858         loff_t disksize;
859
860
861         J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
862         J_ASSERT(handle != NULL || create == 0);
863         depth = ext4_block_to_path(inode, iblock, offsets,
864                                         &blocks_to_boundary);
865
866         if (depth == 0)
867                 goto out;
868
869         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
870
871         /* Simplest case - block found, no allocation needed */
872         if (!partial) {
873                 first_block = le32_to_cpu(chain[depth - 1].key);
874                 clear_buffer_new(bh_result);
875                 count++;
876                 /*map more blocks*/
877                 while (count < maxblocks && count <= blocks_to_boundary) {
878                         ext4_fsblk_t blk;
879
880                         blk = le32_to_cpu(*(chain[depth-1].p + count));
881
882                         if (blk == first_block + count)
883                                 count++;
884                         else
885                                 break;
886                 }
887                 goto got_it;
888         }
889
890         /* Next simple case - plain lookup or failed read of indirect block */
891         if (!create || err == -EIO)
892                 goto cleanup;
893
894         /*
895          * Okay, we need to do block allocation.
896         */
897         goal = ext4_find_goal(inode, iblock, partial);
898
899         /* the number of blocks need to allocate for [d,t]indirect blocks */
900         indirect_blks = (chain + depth) - partial - 1;
901
902         /*
903          * Next look up the indirect map to count the totoal number of
904          * direct blocks to allocate for this branch.
905          */
906         count = ext4_blks_to_allocate(partial, indirect_blks,
907                                         maxblocks, blocks_to_boundary);
908         /*
909          * Block out ext4_truncate while we alter the tree
910          */
911         err = ext4_alloc_branch(handle, inode, iblock, indirect_blks,
912                                         &count, goal,
913                                         offsets + (partial - chain), partial);
914
915         /*
916          * The ext4_splice_branch call will free and forget any buffers
917          * on the new chain if there is a failure, but that risks using
918          * up transaction credits, especially for bitmaps where the
919          * credits cannot be returned.  Can we handle this somehow?  We
920          * may need to return -EAGAIN upwards in the worst case.  --sct
921          */
922         if (!err)
923                 err = ext4_splice_branch(handle, inode, iblock,
924                                         partial, indirect_blks, count);
925         /*
926          * i_disksize growing is protected by i_data_sem.  Don't forget to
927          * protect it if you're about to implement concurrent
928          * ext4_get_block() -bzzz
929         */
930         if (!err && extend_disksize) {
931                 disksize = ((loff_t) iblock + count) << inode->i_blkbits;
932                 if (disksize > i_size_read(inode))
933                         disksize = i_size_read(inode);
934                 if (disksize > ei->i_disksize)
935                         ei->i_disksize = disksize;
936         }
937         if (err)
938                 goto cleanup;
939
940         set_buffer_new(bh_result);
941 got_it:
942         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
943         if (count > blocks_to_boundary)
944                 set_buffer_boundary(bh_result);
945         err = count;
946         /* Clean up and exit */
947         partial = chain + depth - 1;    /* the whole chain */
948 cleanup:
949         while (partial > chain) {
950                 BUFFER_TRACE(partial->bh, "call brelse");
951                 brelse(partial->bh);
952                 partial--;
953         }
954         BUFFER_TRACE(bh_result, "returned");
955 out:
956         return err;
957 }
958
959 /*
960  * Calculate the number of metadata blocks need to reserve
961  * to allocate @blocks for non extent file based file
962  */
963 static int ext4_indirect_calc_metadata_amount(struct inode *inode, int blocks)
964 {
965         int icap = EXT4_ADDR_PER_BLOCK(inode->i_sb);
966         int ind_blks, dind_blks, tind_blks;
967
968         /* number of new indirect blocks needed */
969         ind_blks = (blocks + icap - 1) / icap;
970
971         dind_blks = (ind_blks + icap - 1) / icap;
972
973         tind_blks = 1;
974
975         return ind_blks + dind_blks + tind_blks;
976 }
977
978 /*
979  * Calculate the number of metadata blocks need to reserve
980  * to allocate given number of blocks
981  */
982 static int ext4_calc_metadata_amount(struct inode *inode, int blocks)
983 {
984         if (!blocks)
985                 return 0;
986
987         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
988                 return ext4_ext_calc_metadata_amount(inode, blocks);
989
990         return ext4_indirect_calc_metadata_amount(inode, blocks);
991 }
992
993 static void ext4_da_update_reserve_space(struct inode *inode, int used)
994 {
995         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
996         int total, mdb, mdb_free;
997
998         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
999         /* recalculate the number of metablocks still need to be reserved */
1000         total = EXT4_I(inode)->i_reserved_data_blocks - used;
1001         mdb = ext4_calc_metadata_amount(inode, total);
1002
1003         /* figure out how many metablocks to release */
1004         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1005         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1006
1007         if (mdb_free) {
1008                 /* Account for allocated meta_blocks */
1009                 mdb_free -= EXT4_I(inode)->i_allocated_meta_blocks;
1010
1011                 /* update fs dirty blocks counter */
1012                 percpu_counter_sub(&sbi->s_dirtyblocks_counter, mdb_free);
1013                 EXT4_I(inode)->i_allocated_meta_blocks = 0;
1014                 EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1015         }
1016
1017         /* update per-inode reservations */
1018         BUG_ON(used  > EXT4_I(inode)->i_reserved_data_blocks);
1019         EXT4_I(inode)->i_reserved_data_blocks -= used;
1020
1021         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1022 }
1023
1024 /*
1025  * The ext4_get_blocks_wrap() function try to look up the requested blocks,
1026  * and returns if the blocks are already mapped.
1027  *
1028  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1029  * and store the allocated blocks in the result buffer head and mark it
1030  * mapped.
1031  *
1032  * If file type is extents based, it will call ext4_ext_get_blocks(),
1033  * Otherwise, call with ext4_get_blocks_handle() to handle indirect mapping
1034  * based files
1035  *
1036  * On success, it returns the number of blocks being mapped or allocate.
1037  * if create==0 and the blocks are pre-allocated and uninitialized block,
1038  * the result buffer head is unmapped. If the create ==1, it will make sure
1039  * the buffer head is mapped.
1040  *
1041  * It returns 0 if plain look up failed (blocks have not been allocated), in
1042  * that casem, buffer head is unmapped
1043  *
1044  * It returns the error in case of allocation failure.
1045  */
1046 int ext4_get_blocks_wrap(handle_t *handle, struct inode *inode, sector_t block,
1047                         unsigned long max_blocks, struct buffer_head *bh,
1048                         int create, int extend_disksize, int flag)
1049 {
1050         int retval;
1051
1052         clear_buffer_mapped(bh);
1053
1054         /*
1055          * Try to see if we can get  the block without requesting
1056          * for new file system block.
1057          */
1058         down_read((&EXT4_I(inode)->i_data_sem));
1059         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1060                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1061                                 bh, 0, 0);
1062         } else {
1063                 retval = ext4_get_blocks_handle(handle,
1064                                 inode, block, max_blocks, bh, 0, 0);
1065         }
1066         up_read((&EXT4_I(inode)->i_data_sem));
1067
1068         /* If it is only a block(s) look up */
1069         if (!create)
1070                 return retval;
1071
1072         /*
1073          * Returns if the blocks have already allocated
1074          *
1075          * Note that if blocks have been preallocated
1076          * ext4_ext_get_block() returns th create = 0
1077          * with buffer head unmapped.
1078          */
1079         if (retval > 0 && buffer_mapped(bh))
1080                 return retval;
1081
1082         /*
1083          * New blocks allocate and/or writing to uninitialized extent
1084          * will possibly result in updating i_data, so we take
1085          * the write lock of i_data_sem, and call get_blocks()
1086          * with create == 1 flag.
1087          */
1088         down_write((&EXT4_I(inode)->i_data_sem));
1089
1090         /*
1091          * if the caller is from delayed allocation writeout path
1092          * we have already reserved fs blocks for allocation
1093          * let the underlying get_block() function know to
1094          * avoid double accounting
1095          */
1096         if (flag)
1097                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1098         /*
1099          * We need to check for EXT4 here because migrate
1100          * could have changed the inode type in between
1101          */
1102         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1103                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1104                                 bh, create, extend_disksize);
1105         } else {
1106                 retval = ext4_get_blocks_handle(handle, inode, block,
1107                                 max_blocks, bh, create, extend_disksize);
1108
1109                 if (retval > 0 && buffer_new(bh)) {
1110                         /*
1111                          * We allocated new blocks which will result in
1112                          * i_data's format changing.  Force the migrate
1113                          * to fail by clearing migrate flags
1114                          */
1115                         EXT4_I(inode)->i_flags = EXT4_I(inode)->i_flags &
1116                                                         ~EXT4_EXT_MIGRATE;
1117                 }
1118         }
1119
1120         if (flag) {
1121                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1122                 /*
1123                  * Update reserved blocks/metadata blocks
1124                  * after successful block allocation
1125                  * which were deferred till now
1126                  */
1127                 if ((retval > 0) && buffer_delay(bh))
1128                         ext4_da_update_reserve_space(inode, retval);
1129         }
1130
1131         up_write((&EXT4_I(inode)->i_data_sem));
1132         return retval;
1133 }
1134
1135 /* Maximum number of blocks we map for direct IO at once. */
1136 #define DIO_MAX_BLOCKS 4096
1137
1138 int ext4_get_block(struct inode *inode, sector_t iblock,
1139                    struct buffer_head *bh_result, int create)
1140 {
1141         handle_t *handle = ext4_journal_current_handle();
1142         int ret = 0, started = 0;
1143         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1144         int dio_credits;
1145
1146         if (create && !handle) {
1147                 /* Direct IO write... */
1148                 if (max_blocks > DIO_MAX_BLOCKS)
1149                         max_blocks = DIO_MAX_BLOCKS;
1150                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
1151                 handle = ext4_journal_start(inode, dio_credits);
1152                 if (IS_ERR(handle)) {
1153                         ret = PTR_ERR(handle);
1154                         goto out;
1155                 }
1156                 started = 1;
1157         }
1158
1159         ret = ext4_get_blocks_wrap(handle, inode, iblock,
1160                                         max_blocks, bh_result, create, 0, 0);
1161         if (ret > 0) {
1162                 bh_result->b_size = (ret << inode->i_blkbits);
1163                 ret = 0;
1164         }
1165         if (started)
1166                 ext4_journal_stop(handle);
1167 out:
1168         return ret;
1169 }
1170
1171 /*
1172  * `handle' can be NULL if create is zero
1173  */
1174 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1175                                 ext4_lblk_t block, int create, int *errp)
1176 {
1177         struct buffer_head dummy;
1178         int fatal = 0, err;
1179
1180         J_ASSERT(handle != NULL || create == 0);
1181
1182         dummy.b_state = 0;
1183         dummy.b_blocknr = -1000;
1184         buffer_trace_init(&dummy.b_history);
1185         err = ext4_get_blocks_wrap(handle, inode, block, 1,
1186                                         &dummy, create, 1, 0);
1187         /*
1188          * ext4_get_blocks_handle() returns number of blocks
1189          * mapped. 0 in case of a HOLE.
1190          */
1191         if (err > 0) {
1192                 if (err > 1)
1193                         WARN_ON(1);
1194                 err = 0;
1195         }
1196         *errp = err;
1197         if (!err && buffer_mapped(&dummy)) {
1198                 struct buffer_head *bh;
1199                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1200                 if (!bh) {
1201                         *errp = -EIO;
1202                         goto err;
1203                 }
1204                 if (buffer_new(&dummy)) {
1205                         J_ASSERT(create != 0);
1206                         J_ASSERT(handle != NULL);
1207
1208                         /*
1209                          * Now that we do not always journal data, we should
1210                          * keep in mind whether this should always journal the
1211                          * new buffer as metadata.  For now, regular file
1212                          * writes use ext4_get_block instead, so it's not a
1213                          * problem.
1214                          */
1215                         lock_buffer(bh);
1216                         BUFFER_TRACE(bh, "call get_create_access");
1217                         fatal = ext4_journal_get_create_access(handle, bh);
1218                         if (!fatal && !buffer_uptodate(bh)) {
1219                                 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1220                                 set_buffer_uptodate(bh);
1221                         }
1222                         unlock_buffer(bh);
1223                         BUFFER_TRACE(bh, "call ext4_journal_dirty_metadata");
1224                         err = ext4_journal_dirty_metadata(handle, bh);
1225                         if (!fatal)
1226                                 fatal = err;
1227                 } else {
1228                         BUFFER_TRACE(bh, "not a new buffer");
1229                 }
1230                 if (fatal) {
1231                         *errp = fatal;
1232                         brelse(bh);
1233                         bh = NULL;
1234                 }
1235                 return bh;
1236         }
1237 err:
1238         return NULL;
1239 }
1240
1241 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1242                                ext4_lblk_t block, int create, int *err)
1243 {
1244         struct buffer_head *bh;
1245
1246         bh = ext4_getblk(handle, inode, block, create, err);
1247         if (!bh)
1248                 return bh;
1249         if (buffer_uptodate(bh))
1250                 return bh;
1251         ll_rw_block(READ_META, 1, &bh);
1252         wait_on_buffer(bh);
1253         if (buffer_uptodate(bh))
1254                 return bh;
1255         put_bh(bh);
1256         *err = -EIO;
1257         return NULL;
1258 }
1259
1260 static int walk_page_buffers(handle_t *handle,
1261                              struct buffer_head *head,
1262                              unsigned from,
1263                              unsigned to,
1264                              int *partial,
1265                              int (*fn)(handle_t *handle,
1266                                        struct buffer_head *bh))
1267 {
1268         struct buffer_head *bh;
1269         unsigned block_start, block_end;
1270         unsigned blocksize = head->b_size;
1271         int err, ret = 0;
1272         struct buffer_head *next;
1273
1274         for (bh = head, block_start = 0;
1275              ret == 0 && (bh != head || !block_start);
1276              block_start = block_end, bh = next)
1277         {
1278                 next = bh->b_this_page;
1279                 block_end = block_start + blocksize;
1280                 if (block_end <= from || block_start >= to) {
1281                         if (partial && !buffer_uptodate(bh))
1282                                 *partial = 1;
1283                         continue;
1284                 }
1285                 err = (*fn)(handle, bh);
1286                 if (!ret)
1287                         ret = err;
1288         }
1289         return ret;
1290 }
1291
1292 /*
1293  * To preserve ordering, it is essential that the hole instantiation and
1294  * the data write be encapsulated in a single transaction.  We cannot
1295  * close off a transaction and start a new one between the ext4_get_block()
1296  * and the commit_write().  So doing the jbd2_journal_start at the start of
1297  * prepare_write() is the right place.
1298  *
1299  * Also, this function can nest inside ext4_writepage() ->
1300  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1301  * has generated enough buffer credits to do the whole page.  So we won't
1302  * block on the journal in that case, which is good, because the caller may
1303  * be PF_MEMALLOC.
1304  *
1305  * By accident, ext4 can be reentered when a transaction is open via
1306  * quota file writes.  If we were to commit the transaction while thus
1307  * reentered, there can be a deadlock - we would be holding a quota
1308  * lock, and the commit would never complete if another thread had a
1309  * transaction open and was blocking on the quota lock - a ranking
1310  * violation.
1311  *
1312  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1313  * will _not_ run commit under these circumstances because handle->h_ref
1314  * is elevated.  We'll still have enough credits for the tiny quotafile
1315  * write.
1316  */
1317 static int do_journal_get_write_access(handle_t *handle,
1318                                         struct buffer_head *bh)
1319 {
1320         if (!buffer_mapped(bh) || buffer_freed(bh))
1321                 return 0;
1322         return ext4_journal_get_write_access(handle, bh);
1323 }
1324
1325 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1326                                 loff_t pos, unsigned len, unsigned flags,
1327                                 struct page **pagep, void **fsdata)
1328 {
1329         struct inode *inode = mapping->host;
1330         int ret, needed_blocks = ext4_writepage_trans_blocks(inode);
1331         handle_t *handle;
1332         int retries = 0;
1333         struct page *page;
1334         pgoff_t index;
1335         unsigned from, to;
1336
1337         index = pos >> PAGE_CACHE_SHIFT;
1338         from = pos & (PAGE_CACHE_SIZE - 1);
1339         to = from + len;
1340
1341 retry:
1342         handle = ext4_journal_start(inode, needed_blocks);
1343         if (IS_ERR(handle)) {
1344                 ret = PTR_ERR(handle);
1345                 goto out;
1346         }
1347
1348         page = __grab_cache_page(mapping, index);
1349         if (!page) {
1350                 ext4_journal_stop(handle);
1351                 ret = -ENOMEM;
1352                 goto out;
1353         }
1354         *pagep = page;
1355
1356         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1357                                                         ext4_get_block);
1358
1359         if (!ret && ext4_should_journal_data(inode)) {
1360                 ret = walk_page_buffers(handle, page_buffers(page),
1361                                 from, to, NULL, do_journal_get_write_access);
1362         }
1363
1364         if (ret) {
1365                 unlock_page(page);
1366                 ext4_journal_stop(handle);
1367                 page_cache_release(page);
1368                 /*
1369                  * block_write_begin may have instantiated a few blocks
1370                  * outside i_size.  Trim these off again. Don't need
1371                  * i_size_read because we hold i_mutex.
1372                  */
1373                 if (pos + len > inode->i_size)
1374                         vmtruncate(inode, inode->i_size);
1375         }
1376
1377         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1378                 goto retry;
1379 out:
1380         return ret;
1381 }
1382
1383 /* For write_end() in data=journal mode */
1384 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1385 {
1386         if (!buffer_mapped(bh) || buffer_freed(bh))
1387                 return 0;
1388         set_buffer_uptodate(bh);
1389         return ext4_journal_dirty_metadata(handle, bh);
1390 }
1391
1392 /*
1393  * We need to pick up the new inode size which generic_commit_write gave us
1394  * `file' can be NULL - eg, when called from page_symlink().
1395  *
1396  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1397  * buffers are managed internally.
1398  */
1399 static int ext4_ordered_write_end(struct file *file,
1400                                 struct address_space *mapping,
1401                                 loff_t pos, unsigned len, unsigned copied,
1402                                 struct page *page, void *fsdata)
1403 {
1404         handle_t *handle = ext4_journal_current_handle();
1405         struct inode *inode = mapping->host;
1406         int ret = 0, ret2;
1407
1408         ret = ext4_jbd2_file_inode(handle, inode);
1409
1410         if (ret == 0) {
1411                 loff_t new_i_size;
1412
1413                 new_i_size = pos + copied;
1414                 if (new_i_size > EXT4_I(inode)->i_disksize) {
1415                         ext4_update_i_disksize(inode, new_i_size);
1416                         /* We need to mark inode dirty even if
1417                          * new_i_size is less that inode->i_size
1418                          * bu greater than i_disksize.(hint delalloc)
1419                          */
1420                         ext4_mark_inode_dirty(handle, inode);
1421                 }
1422
1423                 ret2 = generic_write_end(file, mapping, pos, len, copied,
1424                                                         page, fsdata);
1425                 copied = ret2;
1426                 if (ret2 < 0)
1427                         ret = ret2;
1428         }
1429         ret2 = ext4_journal_stop(handle);
1430         if (!ret)
1431                 ret = ret2;
1432
1433         return ret ? ret : copied;
1434 }
1435
1436 static int ext4_writeback_write_end(struct file *file,
1437                                 struct address_space *mapping,
1438                                 loff_t pos, unsigned len, unsigned copied,
1439                                 struct page *page, void *fsdata)
1440 {
1441         handle_t *handle = ext4_journal_current_handle();
1442         struct inode *inode = mapping->host;
1443         int ret = 0, ret2;
1444         loff_t new_i_size;
1445
1446         new_i_size = pos + copied;
1447         if (new_i_size > EXT4_I(inode)->i_disksize) {
1448                 ext4_update_i_disksize(inode, new_i_size);
1449                 /* We need to mark inode dirty even if
1450                  * new_i_size is less that inode->i_size
1451                  * bu greater than i_disksize.(hint delalloc)
1452                  */
1453                 ext4_mark_inode_dirty(handle, inode);
1454         }
1455
1456         ret2 = generic_write_end(file, mapping, pos, len, copied,
1457                                                         page, fsdata);
1458         copied = ret2;
1459         if (ret2 < 0)
1460                 ret = ret2;
1461
1462         ret2 = ext4_journal_stop(handle);
1463         if (!ret)
1464                 ret = ret2;
1465
1466         return ret ? ret : copied;
1467 }
1468
1469 static int ext4_journalled_write_end(struct file *file,
1470                                 struct address_space *mapping,
1471                                 loff_t pos, unsigned len, unsigned copied,
1472                                 struct page *page, void *fsdata)
1473 {
1474         handle_t *handle = ext4_journal_current_handle();
1475         struct inode *inode = mapping->host;
1476         int ret = 0, ret2;
1477         int partial = 0;
1478         unsigned from, to;
1479         loff_t new_i_size;
1480
1481         from = pos & (PAGE_CACHE_SIZE - 1);
1482         to = from + len;
1483
1484         if (copied < len) {
1485                 if (!PageUptodate(page))
1486                         copied = 0;
1487                 page_zero_new_buffers(page, from+copied, to);
1488         }
1489
1490         ret = walk_page_buffers(handle, page_buffers(page), from,
1491                                 to, &partial, write_end_fn);
1492         if (!partial)
1493                 SetPageUptodate(page);
1494         new_i_size = pos + copied;
1495         if (new_i_size > inode->i_size)
1496                 i_size_write(inode, pos+copied);
1497         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
1498         if (new_i_size > EXT4_I(inode)->i_disksize) {
1499                 ext4_update_i_disksize(inode, new_i_size);
1500                 ret2 = ext4_mark_inode_dirty(handle, inode);
1501                 if (!ret)
1502                         ret = ret2;
1503         }
1504
1505         unlock_page(page);
1506         ret2 = ext4_journal_stop(handle);
1507         if (!ret)
1508                 ret = ret2;
1509         page_cache_release(page);
1510
1511         return ret ? ret : copied;
1512 }
1513
1514 static int ext4_da_reserve_space(struct inode *inode, int nrblocks)
1515 {
1516         int retries = 0;
1517        struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1518        unsigned long md_needed, mdblocks, total = 0;
1519
1520         /*
1521          * recalculate the amount of metadata blocks to reserve
1522          * in order to allocate nrblocks
1523          * worse case is one extent per block
1524          */
1525 repeat:
1526         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1527         total = EXT4_I(inode)->i_reserved_data_blocks + nrblocks;
1528         mdblocks = ext4_calc_metadata_amount(inode, total);
1529         BUG_ON(mdblocks < EXT4_I(inode)->i_reserved_meta_blocks);
1530
1531         md_needed = mdblocks - EXT4_I(inode)->i_reserved_meta_blocks;
1532         total = md_needed + nrblocks;
1533
1534         if (ext4_claim_free_blocks(sbi, total)) {
1535                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1536                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1537                         yield();
1538                         goto repeat;
1539                 }
1540                 return -ENOSPC;
1541         }
1542         EXT4_I(inode)->i_reserved_data_blocks += nrblocks;
1543         EXT4_I(inode)->i_reserved_meta_blocks = mdblocks;
1544
1545         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1546         return 0;       /* success */
1547 }
1548
1549 static void ext4_da_release_space(struct inode *inode, int to_free)
1550 {
1551         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1552         int total, mdb, mdb_free, release;
1553
1554         if (!to_free)
1555                 return;         /* Nothing to release, exit */
1556
1557         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1558
1559         if (!EXT4_I(inode)->i_reserved_data_blocks) {
1560                 /*
1561                  * if there is no reserved blocks, but we try to free some
1562                  * then the counter is messed up somewhere.
1563                  * but since this function is called from invalidate
1564                  * page, it's harmless to return without any action
1565                  */
1566                 printk(KERN_INFO "ext4 delalloc try to release %d reserved "
1567                             "blocks for inode %lu, but there is no reserved "
1568                             "data blocks\n", to_free, inode->i_ino);
1569                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1570                 return;
1571         }
1572
1573         /* recalculate the number of metablocks still need to be reserved */
1574         total = EXT4_I(inode)->i_reserved_data_blocks - to_free;
1575         mdb = ext4_calc_metadata_amount(inode, total);
1576
1577         /* figure out how many metablocks to release */
1578         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1579         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1580
1581         release = to_free + mdb_free;
1582
1583         /* update fs dirty blocks counter for truncate case */
1584         percpu_counter_sub(&sbi->s_dirtyblocks_counter, release);
1585
1586         /* update per-inode reservations */
1587         BUG_ON(to_free > EXT4_I(inode)->i_reserved_data_blocks);
1588         EXT4_I(inode)->i_reserved_data_blocks -= to_free;
1589
1590         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1591         EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1592         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1593 }
1594
1595 static void ext4_da_page_release_reservation(struct page *page,
1596                                                 unsigned long offset)
1597 {
1598         int to_release = 0;
1599         struct buffer_head *head, *bh;
1600         unsigned int curr_off = 0;
1601
1602         head = page_buffers(page);
1603         bh = head;
1604         do {
1605                 unsigned int next_off = curr_off + bh->b_size;
1606
1607                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1608                         to_release++;
1609                         clear_buffer_delay(bh);
1610                 }
1611                 curr_off = next_off;
1612         } while ((bh = bh->b_this_page) != head);
1613         ext4_da_release_space(page->mapping->host, to_release);
1614 }
1615
1616 /*
1617  * Delayed allocation stuff
1618  */
1619
1620 struct mpage_da_data {
1621         struct inode *inode;
1622         struct buffer_head lbh;                 /* extent of blocks */
1623         unsigned long first_page, next_page;    /* extent of pages */
1624         get_block_t *get_block;
1625         struct writeback_control *wbc;
1626         int io_done;
1627         long pages_written;
1628         int retval;
1629 };
1630
1631 /*
1632  * mpage_da_submit_io - walks through extent of pages and try to write
1633  * them with writepage() call back
1634  *
1635  * @mpd->inode: inode
1636  * @mpd->first_page: first page of the extent
1637  * @mpd->next_page: page after the last page of the extent
1638  * @mpd->get_block: the filesystem's block mapper function
1639  *
1640  * By the time mpage_da_submit_io() is called we expect all blocks
1641  * to be allocated. this may be wrong if allocation failed.
1642  *
1643  * As pages are already locked by write_cache_pages(), we can't use it
1644  */
1645 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1646 {
1647         struct address_space *mapping = mpd->inode->i_mapping;
1648         int ret = 0, err, nr_pages, i;
1649         unsigned long index, end;
1650         struct pagevec pvec;
1651         long pages_skipped;
1652
1653         BUG_ON(mpd->next_page <= mpd->first_page);
1654         pagevec_init(&pvec, 0);
1655         index = mpd->first_page;
1656         end = mpd->next_page - 1;
1657
1658         while (index <= end) {
1659                 /*
1660                  * We can use PAGECACHE_TAG_DIRTY lookup here because
1661                  * even though we have cleared the dirty flag on the page
1662                  * We still keep the page in the radix tree with tag
1663                  * PAGECACHE_TAG_DIRTY. See clear_page_dirty_for_io.
1664                  * The PAGECACHE_TAG_DIRTY is cleared in set_page_writeback
1665                  * which is called via the below writepage callback.
1666                  */
1667                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
1668                                         PAGECACHE_TAG_DIRTY,
1669                                         min(end - index,
1670                                         (pgoff_t)PAGEVEC_SIZE-1) + 1);
1671                 if (nr_pages == 0)
1672                         break;
1673                 for (i = 0; i < nr_pages; i++) {
1674                         struct page *page = pvec.pages[i];
1675
1676                         pages_skipped = mpd->wbc->pages_skipped;
1677                         err = mapping->a_ops->writepage(page, mpd->wbc);
1678                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
1679                                 /*
1680                                  * have successfully written the page
1681                                  * without skipping the same
1682                                  */
1683                                 mpd->pages_written++;
1684                         /*
1685                          * In error case, we have to continue because
1686                          * remaining pages are still locked
1687                          * XXX: unlock and re-dirty them?
1688                          */
1689                         if (ret == 0)
1690                                 ret = err;
1691                 }
1692                 pagevec_release(&pvec);
1693         }
1694         return ret;
1695 }
1696
1697 /*
1698  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
1699  *
1700  * @mpd->inode - inode to walk through
1701  * @exbh->b_blocknr - first block on a disk
1702  * @exbh->b_size - amount of space in bytes
1703  * @logical - first logical block to start assignment with
1704  *
1705  * the function goes through all passed space and put actual disk
1706  * block numbers into buffer heads, dropping BH_Delay
1707  */
1708 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
1709                                  struct buffer_head *exbh)
1710 {
1711         struct inode *inode = mpd->inode;
1712         struct address_space *mapping = inode->i_mapping;
1713         int blocks = exbh->b_size >> inode->i_blkbits;
1714         sector_t pblock = exbh->b_blocknr, cur_logical;
1715         struct buffer_head *head, *bh;
1716         pgoff_t index, end;
1717         struct pagevec pvec;
1718         int nr_pages, i;
1719
1720         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1721         end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1722         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1723
1724         pagevec_init(&pvec, 0);
1725
1726         while (index <= end) {
1727                 /* XXX: optimize tail */
1728                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1729                 if (nr_pages == 0)
1730                         break;
1731                 for (i = 0; i < nr_pages; i++) {
1732                         struct page *page = pvec.pages[i];
1733
1734                         index = page->index;
1735                         if (index > end)
1736                                 break;
1737                         index++;
1738
1739                         BUG_ON(!PageLocked(page));
1740                         BUG_ON(PageWriteback(page));
1741                         BUG_ON(!page_has_buffers(page));
1742
1743                         bh = page_buffers(page);
1744                         head = bh;
1745
1746                         /* skip blocks out of the range */
1747                         do {
1748                                 if (cur_logical >= logical)
1749                                         break;
1750                                 cur_logical++;
1751                         } while ((bh = bh->b_this_page) != head);
1752
1753                         do {
1754                                 if (cur_logical >= logical + blocks)
1755                                         break;
1756                                 if (buffer_delay(bh)) {
1757                                         bh->b_blocknr = pblock;
1758                                         clear_buffer_delay(bh);
1759                                         bh->b_bdev = inode->i_sb->s_bdev;
1760                                 } else if (buffer_unwritten(bh)) {
1761                                         bh->b_blocknr = pblock;
1762                                         clear_buffer_unwritten(bh);
1763                                         set_buffer_mapped(bh);
1764                                         set_buffer_new(bh);
1765                                         bh->b_bdev = inode->i_sb->s_bdev;
1766                                 } else if (buffer_mapped(bh))
1767                                         BUG_ON(bh->b_blocknr != pblock);
1768
1769                                 cur_logical++;
1770                                 pblock++;
1771                         } while ((bh = bh->b_this_page) != head);
1772                 }
1773                 pagevec_release(&pvec);
1774         }
1775 }
1776
1777
1778 /*
1779  * __unmap_underlying_blocks - just a helper function to unmap
1780  * set of blocks described by @bh
1781  */
1782 static inline void __unmap_underlying_blocks(struct inode *inode,
1783                                              struct buffer_head *bh)
1784 {
1785         struct block_device *bdev = inode->i_sb->s_bdev;
1786         int blocks, i;
1787
1788         blocks = bh->b_size >> inode->i_blkbits;
1789         for (i = 0; i < blocks; i++)
1790                 unmap_underlying_metadata(bdev, bh->b_blocknr + i);
1791 }
1792
1793 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
1794                                         sector_t logical, long blk_cnt)
1795 {
1796         int nr_pages, i;
1797         pgoff_t index, end;
1798         struct pagevec pvec;
1799         struct inode *inode = mpd->inode;
1800         struct address_space *mapping = inode->i_mapping;
1801
1802         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1803         end   = (logical + blk_cnt - 1) >>
1804                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
1805         while (index <= end) {
1806                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1807                 if (nr_pages == 0)
1808                         break;
1809                 for (i = 0; i < nr_pages; i++) {
1810                         struct page *page = pvec.pages[i];
1811                         index = page->index;
1812                         if (index > end)
1813                                 break;
1814                         index++;
1815
1816                         BUG_ON(!PageLocked(page));
1817                         BUG_ON(PageWriteback(page));
1818                         block_invalidatepage(page, 0);
1819                         ClearPageUptodate(page);
1820                         unlock_page(page);
1821                 }
1822         }
1823         return;
1824 }
1825
1826 static void ext4_print_free_blocks(struct inode *inode)
1827 {
1828         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1829         printk(KERN_EMERG "Total free blocks count %lld\n",
1830                         ext4_count_free_blocks(inode->i_sb));
1831         printk(KERN_EMERG "Free/Dirty block details\n");
1832         printk(KERN_EMERG "free_blocks=%lld\n",
1833                         percpu_counter_sum(&sbi->s_freeblocks_counter));
1834         printk(KERN_EMERG "dirty_blocks=%lld\n",
1835                         percpu_counter_sum(&sbi->s_dirtyblocks_counter));
1836         printk(KERN_EMERG "Block reservation details\n");
1837         printk(KERN_EMERG "i_reserved_data_blocks=%lu\n",
1838                         EXT4_I(inode)->i_reserved_data_blocks);
1839         printk(KERN_EMERG "i_reserved_meta_blocks=%lu\n",
1840                         EXT4_I(inode)->i_reserved_meta_blocks);
1841         return;
1842 }
1843
1844 /*
1845  * mpage_da_map_blocks - go through given space
1846  *
1847  * @mpd->lbh - bh describing space
1848  * @mpd->get_block - the filesystem's block mapper function
1849  *
1850  * The function skips space we know is already mapped to disk blocks.
1851  *
1852  */
1853 static int  mpage_da_map_blocks(struct mpage_da_data *mpd)
1854 {
1855         int err = 0;
1856         struct buffer_head new;
1857         struct buffer_head *lbh = &mpd->lbh;
1858         sector_t next;
1859
1860         /*
1861          * We consider only non-mapped and non-allocated blocks
1862          */
1863         if (buffer_mapped(lbh) && !buffer_delay(lbh))
1864                 return 0;
1865         new.b_state = lbh->b_state;
1866         new.b_blocknr = 0;
1867         new.b_size = lbh->b_size;
1868         next = lbh->b_blocknr;
1869         /*
1870          * If we didn't accumulate anything
1871          * to write simply return
1872          */
1873         if (!new.b_size)
1874                 return 0;
1875         err = mpd->get_block(mpd->inode, next, &new, 1);
1876         if (err) {
1877
1878                 /* If get block returns with error
1879                  * we simply return. Later writepage
1880                  * will redirty the page and writepages
1881                  * will find the dirty page again
1882                  */
1883                 if (err == -EAGAIN)
1884                         return 0;
1885
1886                 if (err == -ENOSPC &&
1887                                 ext4_count_free_blocks(mpd->inode->i_sb)) {
1888                         mpd->retval = err;
1889                         return 0;
1890                 }
1891
1892                 /*
1893                  * get block failure will cause us
1894                  * to loop in writepages. Because
1895                  * a_ops->writepage won't be able to
1896                  * make progress. The page will be redirtied
1897                  * by writepage and writepages will again
1898                  * try to write the same.
1899                  */
1900                 printk(KERN_EMERG "%s block allocation failed for inode %lu "
1901                                   "at logical offset %llu with max blocks "
1902                                   "%zd with error %d\n",
1903                                   __func__, mpd->inode->i_ino,
1904                                   (unsigned long long)next,
1905                                   lbh->b_size >> mpd->inode->i_blkbits, err);
1906                 printk(KERN_EMERG "This should not happen.!! "
1907                                         "Data will be lost\n");
1908                 if (err == -ENOSPC) {
1909                         ext4_print_free_blocks(mpd->inode);
1910                 }
1911                 /* invlaidate all the pages */
1912                 ext4_da_block_invalidatepages(mpd, next,
1913                                 lbh->b_size >> mpd->inode->i_blkbits);
1914                 return err;
1915         }
1916         BUG_ON(new.b_size == 0);
1917
1918         if (buffer_new(&new))
1919                 __unmap_underlying_blocks(mpd->inode, &new);
1920
1921         /*
1922          * If blocks are delayed marked, we need to
1923          * put actual blocknr and drop delayed bit
1924          */
1925         if (buffer_delay(lbh) || buffer_unwritten(lbh))
1926                 mpage_put_bnr_to_bhs(mpd, next, &new);
1927
1928         return 0;
1929 }
1930
1931 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
1932                 (1 << BH_Delay) | (1 << BH_Unwritten))
1933
1934 /*
1935  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
1936  *
1937  * @mpd->lbh - extent of blocks
1938  * @logical - logical number of the block in the file
1939  * @bh - bh of the block (used to access block's state)
1940  *
1941  * the function is used to collect contig. blocks in same state
1942  */
1943 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
1944                                    sector_t logical, struct buffer_head *bh)
1945 {
1946         sector_t next;
1947         size_t b_size = bh->b_size;
1948         struct buffer_head *lbh = &mpd->lbh;
1949         int nrblocks = lbh->b_size >> mpd->inode->i_blkbits;
1950
1951         /* check if thereserved journal credits might overflow */
1952         if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
1953                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
1954                         /*
1955                          * With non-extent format we are limited by the journal
1956                          * credit available.  Total credit needed to insert
1957                          * nrblocks contiguous blocks is dependent on the
1958                          * nrblocks.  So limit nrblocks.
1959                          */
1960                         goto flush_it;
1961                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
1962                                 EXT4_MAX_TRANS_DATA) {
1963                         /*
1964                          * Adding the new buffer_head would make it cross the
1965                          * allowed limit for which we have journal credit
1966                          * reserved. So limit the new bh->b_size
1967                          */
1968                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
1969                                                 mpd->inode->i_blkbits;
1970                         /* we will do mpage_da_submit_io in the next loop */
1971                 }
1972         }
1973         /*
1974          * First block in the extent
1975          */
1976         if (lbh->b_size == 0) {
1977                 lbh->b_blocknr = logical;
1978                 lbh->b_size = b_size;
1979                 lbh->b_state = bh->b_state & BH_FLAGS;
1980                 return;
1981         }
1982
1983         next = lbh->b_blocknr + nrblocks;
1984         /*
1985          * Can we merge the block to our big extent?
1986          */
1987         if (logical == next && (bh->b_state & BH_FLAGS) == lbh->b_state) {
1988                 lbh->b_size += b_size;
1989                 return;
1990         }
1991
1992 flush_it:
1993         /*
1994          * We couldn't merge the block to our extent, so we
1995          * need to flush current  extent and start new one
1996          */
1997         if (mpage_da_map_blocks(mpd) == 0)
1998                 mpage_da_submit_io(mpd);
1999         mpd->io_done = 1;
2000         return;
2001 }
2002
2003 /*
2004  * __mpage_da_writepage - finds extent of pages and blocks
2005  *
2006  * @page: page to consider
2007  * @wbc: not used, we just follow rules
2008  * @data: context
2009  *
2010  * The function finds extents of pages and scan them for all blocks.
2011  */
2012 static int __mpage_da_writepage(struct page *page,
2013                                 struct writeback_control *wbc, void *data)
2014 {
2015         struct mpage_da_data *mpd = data;
2016         struct inode *inode = mpd->inode;
2017         struct buffer_head *bh, *head, fake;
2018         sector_t logical;
2019
2020         if (mpd->io_done) {
2021                 /*
2022                  * Rest of the page in the page_vec
2023                  * redirty then and skip then. We will
2024                  * try to to write them again after
2025                  * starting a new transaction
2026                  */
2027                 redirty_page_for_writepage(wbc, page);
2028                 unlock_page(page);
2029                 return MPAGE_DA_EXTENT_TAIL;
2030         }
2031         /*
2032          * Can we merge this page to current extent?
2033          */
2034         if (mpd->next_page != page->index) {
2035                 /*
2036                  * Nope, we can't. So, we map non-allocated blocks
2037                  * and start IO on them using writepage()
2038                  */
2039                 if (mpd->next_page != mpd->first_page) {
2040                         if (mpage_da_map_blocks(mpd) == 0)
2041                                 mpage_da_submit_io(mpd);
2042                         /*
2043                          * skip rest of the page in the page_vec
2044                          */
2045                         mpd->io_done = 1;
2046                         redirty_page_for_writepage(wbc, page);
2047                         unlock_page(page);
2048                         return MPAGE_DA_EXTENT_TAIL;
2049                 }
2050
2051                 /*
2052                  * Start next extent of pages ...
2053                  */
2054                 mpd->first_page = page->index;
2055
2056                 /*
2057                  * ... and blocks
2058                  */
2059                 mpd->lbh.b_size = 0;
2060                 mpd->lbh.b_state = 0;
2061                 mpd->lbh.b_blocknr = 0;
2062         }
2063
2064         mpd->next_page = page->index + 1;
2065         logical = (sector_t) page->index <<
2066                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2067
2068         if (!page_has_buffers(page)) {
2069                 /*
2070                  * There is no attached buffer heads yet (mmap?)
2071                  * we treat the page asfull of dirty blocks
2072                  */
2073                 bh = &fake;
2074                 bh->b_size = PAGE_CACHE_SIZE;
2075                 bh->b_state = 0;
2076                 set_buffer_dirty(bh);
2077                 set_buffer_uptodate(bh);
2078                 mpage_add_bh_to_extent(mpd, logical, bh);
2079                 if (mpd->io_done)
2080                         return MPAGE_DA_EXTENT_TAIL;
2081         } else {
2082                 /*
2083                  * Page with regular buffer heads, just add all dirty ones
2084                  */
2085                 head = page_buffers(page);
2086                 bh = head;
2087                 do {
2088                         BUG_ON(buffer_locked(bh));
2089                         if (buffer_dirty(bh) &&
2090                                 (!buffer_mapped(bh) || buffer_delay(bh))) {
2091                                 mpage_add_bh_to_extent(mpd, logical, bh);
2092                                 if (mpd->io_done)
2093                                         return MPAGE_DA_EXTENT_TAIL;
2094                         }
2095                         logical++;
2096                 } while ((bh = bh->b_this_page) != head);
2097         }
2098
2099         return 0;
2100 }
2101
2102 /*
2103  * mpage_da_writepages - walk the list of dirty pages of the given
2104  * address space, allocates non-allocated blocks, maps newly-allocated
2105  * blocks to existing bhs and issue IO them
2106  *
2107  * @mapping: address space structure to write
2108  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2109  * @get_block: the filesystem's block mapper function.
2110  *
2111  * This is a library function, which implements the writepages()
2112  * address_space_operation.
2113  */
2114 static int mpage_da_writepages(struct address_space *mapping,
2115                                struct writeback_control *wbc,
2116                                struct mpage_da_data *mpd)
2117 {
2118         int ret;
2119
2120         if (!mpd->get_block)
2121                 return generic_writepages(mapping, wbc);
2122
2123         mpd->lbh.b_size = 0;
2124         mpd->lbh.b_state = 0;
2125         mpd->lbh.b_blocknr = 0;
2126         mpd->first_page = 0;
2127         mpd->next_page = 0;
2128         mpd->io_done = 0;
2129         mpd->pages_written = 0;
2130         mpd->retval = 0;
2131
2132         ret = write_cache_pages(mapping, wbc, __mpage_da_writepage, mpd);
2133         /*
2134          * Handle last extent of pages
2135          */
2136         if (!mpd->io_done && mpd->next_page != mpd->first_page) {
2137                 if (mpage_da_map_blocks(mpd) == 0)
2138                         mpage_da_submit_io(mpd);
2139
2140                 mpd->io_done = 1;
2141                 ret = MPAGE_DA_EXTENT_TAIL;
2142         }
2143         wbc->nr_to_write -= mpd->pages_written;
2144         return ret;
2145 }
2146
2147 /*
2148  * this is a special callback for ->write_begin() only
2149  * it's intention is to return mapped block or reserve space
2150  */
2151 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2152                                   struct buffer_head *bh_result, int create)
2153 {
2154         int ret = 0;
2155
2156         BUG_ON(create == 0);
2157         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2158
2159         /*
2160          * first, we need to know whether the block is allocated already
2161          * preallocated blocks are unmapped but should treated
2162          * the same as allocated blocks.
2163          */
2164         ret = ext4_get_blocks_wrap(NULL, inode, iblock, 1,  bh_result, 0, 0, 0);
2165         if ((ret == 0) && !buffer_delay(bh_result)) {
2166                 /* the block isn't (pre)allocated yet, let's reserve space */
2167                 /*
2168                  * XXX: __block_prepare_write() unmaps passed block,
2169                  * is it OK?
2170                  */
2171                 ret = ext4_da_reserve_space(inode, 1);
2172                 if (ret)
2173                         /* not enough space to reserve */
2174                         return ret;
2175
2176                 map_bh(bh_result, inode->i_sb, 0);
2177                 set_buffer_new(bh_result);
2178                 set_buffer_delay(bh_result);
2179         } else if (ret > 0) {
2180                 bh_result->b_size = (ret << inode->i_blkbits);
2181                 ret = 0;
2182         }
2183
2184         return ret;
2185 }
2186 #define         EXT4_DELALLOC_RSVED     1
2187 static int ext4_da_get_block_write(struct inode *inode, sector_t iblock,
2188                                    struct buffer_head *bh_result, int create)
2189 {
2190         int ret;
2191         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2192         loff_t disksize = EXT4_I(inode)->i_disksize;
2193         handle_t *handle = NULL;
2194
2195         handle = ext4_journal_current_handle();
2196         BUG_ON(!handle);
2197         ret = ext4_get_blocks_wrap(handle, inode, iblock, max_blocks,
2198                         bh_result, create, 0, EXT4_DELALLOC_RSVED);
2199         if (ret > 0) {
2200
2201                 bh_result->b_size = (ret << inode->i_blkbits);
2202
2203                 if (ext4_should_order_data(inode)) {
2204                         int retval;
2205                         retval = ext4_jbd2_file_inode(handle, inode);
2206                         if (retval)
2207                                 /*
2208                                  * Failed to add inode for ordered
2209                                  * mode. Don't update file size
2210                                  */
2211                                 return retval;
2212                 }
2213
2214                 /*
2215                  * Update on-disk size along with block allocation
2216                  * we don't use 'extend_disksize' as size may change
2217                  * within already allocated block -bzzz
2218                  */
2219                 disksize = ((loff_t) iblock + ret) << inode->i_blkbits;
2220                 if (disksize > i_size_read(inode))
2221                         disksize = i_size_read(inode);
2222                 if (disksize > EXT4_I(inode)->i_disksize) {
2223                         ext4_update_i_disksize(inode, disksize);
2224                         ret = ext4_mark_inode_dirty(handle, inode);
2225                         return ret;
2226                 }
2227                 ret = 0;
2228         }
2229         return ret;
2230 }
2231
2232 static int ext4_bh_unmapped_or_delay(handle_t *handle, struct buffer_head *bh)
2233 {
2234         /*
2235          * unmapped buffer is possible for holes.
2236          * delay buffer is possible with delayed allocation
2237          */
2238         return ((!buffer_mapped(bh) || buffer_delay(bh)) && buffer_dirty(bh));
2239 }
2240
2241 static int ext4_normal_get_block_write(struct inode *inode, sector_t iblock,
2242                                    struct buffer_head *bh_result, int create)
2243 {
2244         int ret = 0;
2245         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2246
2247         /*
2248          * we don't want to do block allocation in writepage
2249          * so call get_block_wrap with create = 0
2250          */
2251         ret = ext4_get_blocks_wrap(NULL, inode, iblock, max_blocks,
2252                                    bh_result, 0, 0, 0);
2253         if (ret > 0) {
2254                 bh_result->b_size = (ret << inode->i_blkbits);
2255                 ret = 0;
2256         }
2257         return ret;
2258 }
2259
2260 /*
2261  * get called vi ext4_da_writepages after taking page lock (have journal handle)
2262  * get called via journal_submit_inode_data_buffers (no journal handle)
2263  * get called via shrink_page_list via pdflush (no journal handle)
2264  * or grab_page_cache when doing write_begin (have journal handle)
2265  */
2266 static int ext4_da_writepage(struct page *page,
2267                                 struct writeback_control *wbc)
2268 {
2269         int ret = 0;
2270         loff_t size;
2271         unsigned long len;
2272         struct buffer_head *page_bufs;
2273         struct inode *inode = page->mapping->host;
2274
2275         size = i_size_read(inode);
2276         if (page->index == size >> PAGE_CACHE_SHIFT)
2277                 len = size & ~PAGE_CACHE_MASK;
2278         else
2279                 len = PAGE_CACHE_SIZE;
2280
2281         if (page_has_buffers(page)) {
2282                 page_bufs = page_buffers(page);
2283                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2284                                         ext4_bh_unmapped_or_delay)) {
2285                         /*
2286                          * We don't want to do  block allocation
2287                          * So redirty the page and return
2288                          * We may reach here when we do a journal commit
2289                          * via journal_submit_inode_data_buffers.
2290                          * If we don't have mapping block we just ignore
2291                          * them. We can also reach here via shrink_page_list
2292                          */
2293                         redirty_page_for_writepage(wbc, page);
2294                         unlock_page(page);
2295                         return 0;
2296                 }
2297         } else {
2298                 /*
2299                  * The test for page_has_buffers() is subtle:
2300                  * We know the page is dirty but it lost buffers. That means
2301                  * that at some moment in time after write_begin()/write_end()
2302                  * has been called all buffers have been clean and thus they
2303                  * must have been written at least once. So they are all
2304                  * mapped and we can happily proceed with mapping them
2305                  * and writing the page.
2306                  *
2307                  * Try to initialize the buffer_heads and check whether
2308                  * all are mapped and non delay. We don't want to
2309                  * do block allocation here.
2310                  */
2311                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
2312                                                 ext4_normal_get_block_write);
2313                 if (!ret) {
2314                         page_bufs = page_buffers(page);
2315                         /* check whether all are mapped and non delay */
2316                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2317                                                 ext4_bh_unmapped_or_delay)) {
2318                                 redirty_page_for_writepage(wbc, page);
2319                                 unlock_page(page);
2320                                 return 0;
2321                         }
2322                 } else {
2323                         /*
2324                          * We can't do block allocation here
2325                          * so just redity the page and unlock
2326                          * and return
2327                          */
2328                         redirty_page_for_writepage(wbc, page);
2329                         unlock_page(page);
2330                         return 0;
2331                 }
2332                 /* now mark the buffer_heads as dirty and uptodate */
2333                 block_commit_write(page, 0, PAGE_CACHE_SIZE);
2334         }
2335
2336         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2337                 ret = nobh_writepage(page, ext4_normal_get_block_write, wbc);
2338         else
2339                 ret = block_write_full_page(page,
2340                                                 ext4_normal_get_block_write,
2341                                                 wbc);
2342
2343         return ret;
2344 }
2345
2346 /*
2347  * This is called via ext4_da_writepages() to
2348  * calulate the total number of credits to reserve to fit
2349  * a single extent allocation into a single transaction,
2350  * ext4_da_writpeages() will loop calling this before
2351  * the block allocation.
2352  */
2353
2354 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2355 {
2356         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2357
2358         /*
2359          * With non-extent format the journal credit needed to
2360          * insert nrblocks contiguous block is dependent on
2361          * number of contiguous block. So we will limit
2362          * number of contiguous block to a sane value
2363          */
2364         if (!(inode->i_flags & EXT4_EXTENTS_FL) &&
2365             (max_blocks > EXT4_MAX_TRANS_DATA))
2366                 max_blocks = EXT4_MAX_TRANS_DATA;
2367
2368         return ext4_chunk_trans_blocks(inode, max_blocks);
2369 }
2370
2371 static int ext4_da_writepages(struct address_space *mapping,
2372                               struct writeback_control *wbc)
2373 {
2374         pgoff_t index;
2375         int range_whole = 0;
2376         handle_t *handle = NULL;
2377         struct mpage_da_data mpd;
2378         struct inode *inode = mapping->host;
2379         int no_nrwrite_index_update;
2380         long pages_written = 0, pages_skipped;
2381         int needed_blocks, ret = 0, nr_to_writebump = 0;
2382         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2383
2384         /*
2385          * No pages to write? This is mainly a kludge to avoid starting
2386          * a transaction for special inodes like journal inode on last iput()
2387          * because that could violate lock ordering on umount
2388          */
2389         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2390                 return 0;
2391         /*
2392          * Make sure nr_to_write is >= sbi->s_mb_stream_request
2393          * This make sure small files blocks are allocated in
2394          * single attempt. This ensure that small files
2395          * get less fragmented.
2396          */
2397         if (wbc->nr_to_write < sbi->s_mb_stream_request) {
2398                 nr_to_writebump = sbi->s_mb_stream_request - wbc->nr_to_write;
2399                 wbc->nr_to_write = sbi->s_mb_stream_request;
2400         }
2401         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2402                 range_whole = 1;
2403
2404         if (wbc->range_cyclic)
2405                 index = mapping->writeback_index;
2406         else
2407                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2408
2409         mpd.wbc = wbc;
2410         mpd.inode = mapping->host;
2411
2412         /*
2413          * we don't want write_cache_pages to update
2414          * nr_to_write and writeback_index
2415          */
2416         no_nrwrite_index_update = wbc->no_nrwrite_index_update;
2417         wbc->no_nrwrite_index_update = 1;
2418         pages_skipped = wbc->pages_skipped;
2419
2420         while (!ret && wbc->nr_to_write > 0) {
2421
2422                 /*
2423                  * we  insert one extent at a time. So we need
2424                  * credit needed for single extent allocation.
2425                  * journalled mode is currently not supported
2426                  * by delalloc
2427                  */
2428                 BUG_ON(ext4_should_journal_data(inode));
2429                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2430
2431                 /* start a new transaction*/
2432                 handle = ext4_journal_start(inode, needed_blocks);
2433                 if (IS_ERR(handle)) {
2434                         ret = PTR_ERR(handle);
2435                         printk(KERN_EMERG "%s: jbd2_start: "
2436                                "%ld pages, ino %lu; err %d\n", __func__,
2437                                 wbc->nr_to_write, inode->i_ino, ret);
2438                         dump_stack();
2439                         goto out_writepages;
2440                 }
2441                 mpd.get_block = ext4_da_get_block_write;
2442                 ret = mpage_da_writepages(mapping, wbc, &mpd);
2443
2444                 ext4_journal_stop(handle);
2445
2446                 if (mpd.retval == -ENOSPC) {
2447                         /* commit the transaction which would
2448                          * free blocks released in the transaction
2449                          * and try again
2450                          */
2451                         jbd2_journal_force_commit_nested(sbi->s_journal);
2452                         wbc->pages_skipped = pages_skipped;
2453                         ret = 0;
2454                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2455                         /*
2456                          * got one extent now try with
2457                          * rest of the pages
2458                          */
2459                         pages_written += mpd.pages_written;
2460                         wbc->pages_skipped = pages_skipped;
2461                         ret = 0;
2462                 } else if (wbc->nr_to_write)
2463                         /*
2464                          * There is no more writeout needed
2465                          * or we requested for a noblocking writeout
2466                          * and we found the device congested
2467                          */
2468                         break;
2469         }
2470         if (pages_skipped != wbc->pages_skipped)
2471                 printk(KERN_EMERG "This should not happen leaving %s "
2472                                 "with nr_to_write = %ld ret = %d\n",
2473                                 __func__, wbc->nr_to_write, ret);
2474
2475         /* Update index */
2476         index += pages_written;
2477         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2478                 /*
2479                  * set the writeback_index so that range_cyclic
2480                  * mode will write it back later
2481                  */
2482                 mapping->writeback_index = index;
2483
2484 out_writepages:
2485         if (!no_nrwrite_index_update)
2486                 wbc->no_nrwrite_index_update = 0;
2487         wbc->nr_to_write -= nr_to_writebump;
2488         return ret;
2489 }
2490
2491 #define FALL_BACK_TO_NONDELALLOC 1
2492 static int ext4_nonda_switch(struct super_block *sb)
2493 {
2494         s64 free_blocks, dirty_blocks;
2495         struct ext4_sb_info *sbi = EXT4_SB(sb);
2496
2497         /*
2498          * switch to non delalloc mode if we are running low
2499          * on free block. The free block accounting via percpu
2500          * counters can get slightly wrong with FBC_BATCH getting
2501          * accumulated on each CPU without updating global counters
2502          * Delalloc need an accurate free block accounting. So switch
2503          * to non delalloc when we are near to error range.
2504          */
2505         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
2506         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
2507         if (2 * free_blocks < 3 * dirty_blocks ||
2508                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
2509                 /*
2510                  * free block count is less that 150% of dirty blocks
2511                  * or free blocks is less that watermark
2512                  */
2513                 return 1;
2514         }
2515         return 0;
2516 }
2517
2518 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2519                                 loff_t pos, unsigned len, unsigned flags,
2520                                 struct page **pagep, void **fsdata)
2521 {
2522         int ret, retries = 0;
2523         struct page *page;
2524         pgoff_t index;
2525         unsigned from, to;
2526         struct inode *inode = mapping->host;
2527         handle_t *handle;
2528
2529         index = pos >> PAGE_CACHE_SHIFT;
2530         from = pos & (PAGE_CACHE_SIZE - 1);
2531         to = from + len;
2532
2533         if (ext4_nonda_switch(inode->i_sb)) {
2534                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2535                 return ext4_write_begin(file, mapping, pos,
2536                                         len, flags, pagep, fsdata);
2537         }
2538         *fsdata = (void *)0;
2539 retry:
2540         /*
2541          * With delayed allocation, we don't log the i_disksize update
2542          * if there is delayed block allocation. But we still need
2543          * to journalling the i_disksize update if writes to the end
2544          * of file which has an already mapped buffer.
2545          */
2546         handle = ext4_journal_start(inode, 1);
2547         if (IS_ERR(handle)) {
2548                 ret = PTR_ERR(handle);
2549                 goto out;
2550         }
2551
2552         page = __grab_cache_page(mapping, index);
2553         if (!page) {
2554                 ext4_journal_stop(handle);
2555                 ret = -ENOMEM;
2556                 goto out;
2557         }
2558         *pagep = page;
2559
2560         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
2561                                                         ext4_da_get_block_prep);
2562         if (ret < 0) {
2563                 unlock_page(page);
2564                 ext4_journal_stop(handle);
2565                 page_cache_release(page);
2566                 /*
2567                  * block_write_begin may have instantiated a few blocks
2568                  * outside i_size.  Trim these off again. Don't need
2569                  * i_size_read because we hold i_mutex.
2570                  */
2571                 if (pos + len > inode->i_size)
2572                         vmtruncate(inode, inode->i_size);
2573         }
2574
2575         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
2576                 goto retry;
2577 out:
2578         return ret;
2579 }
2580
2581 /*
2582  * Check if we should update i_disksize
2583  * when write to the end of file but not require block allocation
2584  */
2585 static int ext4_da_should_update_i_disksize(struct page *page,
2586                                          unsigned long offset)
2587 {
2588         struct buffer_head *bh;
2589         struct inode *inode = page->mapping->host;
2590         unsigned int idx;
2591         int i;
2592
2593         bh = page_buffers(page);
2594         idx = offset >> inode->i_blkbits;
2595
2596         for (i = 0; i < idx; i++)
2597                 bh = bh->b_this_page;
2598
2599         if (!buffer_mapped(bh) || (buffer_delay(bh)))
2600                 return 0;
2601         return 1;
2602 }
2603
2604 static int ext4_da_write_end(struct file *file,
2605                                 struct address_space *mapping,
2606                                 loff_t pos, unsigned len, unsigned copied,
2607                                 struct page *page, void *fsdata)
2608 {
2609         struct inode *inode = mapping->host;
2610         int ret = 0, ret2;
2611         handle_t *handle = ext4_journal_current_handle();
2612         loff_t new_i_size;
2613         unsigned long start, end;
2614         int write_mode = (int)(unsigned long)fsdata;
2615
2616         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
2617                 if (ext4_should_order_data(inode)) {
2618                         return ext4_ordered_write_end(file, mapping, pos,
2619                                         len, copied, page, fsdata);
2620                 } else if (ext4_should_writeback_data(inode)) {
2621                         return ext4_writeback_write_end(file, mapping, pos,
2622                                         len, copied, page, fsdata);
2623                 } else {
2624                         BUG();
2625                 }
2626         }
2627
2628         start = pos & (PAGE_CACHE_SIZE - 1);
2629         end = start + copied - 1;
2630
2631         /*
2632          * generic_write_end() will run mark_inode_dirty() if i_size
2633          * changes.  So let's piggyback the i_disksize mark_inode_dirty
2634          * into that.
2635          */
2636
2637         new_i_size = pos + copied;
2638         if (new_i_size > EXT4_I(inode)->i_disksize) {
2639                 if (ext4_da_should_update_i_disksize(page, end)) {
2640                         down_write(&EXT4_I(inode)->i_data_sem);
2641                         if (new_i_size > EXT4_I(inode)->i_disksize) {
2642                                 /*
2643                                  * Updating i_disksize when extending file
2644                                  * without needing block allocation
2645                                  */
2646                                 if (ext4_should_order_data(inode))
2647                                         ret = ext4_jbd2_file_inode(handle,
2648                                                                    inode);
2649
2650                                 EXT4_I(inode)->i_disksize = new_i_size;
2651                         }
2652                         up_write(&EXT4_I(inode)->i_data_sem);
2653                         /* We need to mark inode dirty even if
2654                          * new_i_size is less that inode->i_size
2655                          * bu greater than i_disksize.(hint delalloc)
2656                          */
2657                         ext4_mark_inode_dirty(handle, inode);
2658                 }
2659         }
2660         ret2 = generic_write_end(file, mapping, pos, len, copied,
2661                                                         page, fsdata);
2662         copied = ret2;
2663         if (ret2 < 0)
2664                 ret = ret2;
2665         ret2 = ext4_journal_stop(handle);
2666         if (!ret)
2667                 ret = ret2;
2668
2669         return ret ? ret : copied;
2670 }
2671
2672 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
2673 {
2674         /*
2675          * Drop reserved blocks
2676          */
2677         BUG_ON(!PageLocked(page));
2678         if (!page_has_buffers(page))
2679                 goto out;
2680
2681         ext4_da_page_release_reservation(page, offset);
2682
2683 out:
2684         ext4_invalidatepage(page, offset);
2685
2686         return;
2687 }
2688
2689
2690 /*
2691  * bmap() is special.  It gets used by applications such as lilo and by
2692  * the swapper to find the on-disk block of a specific piece of data.
2693  *
2694  * Naturally, this is dangerous if the block concerned is still in the
2695  * journal.  If somebody makes a swapfile on an ext4 data-journaling
2696  * filesystem and enables swap, then they may get a nasty shock when the
2697  * data getting swapped to that swapfile suddenly gets overwritten by
2698  * the original zero's written out previously to the journal and
2699  * awaiting writeback in the kernel's buffer cache.
2700  *
2701  * So, if we see any bmap calls here on a modified, data-journaled file,
2702  * take extra steps to flush any blocks which might be in the cache.
2703  */
2704 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
2705 {
2706         struct inode *inode = mapping->host;
2707         journal_t *journal;
2708         int err;
2709
2710         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
2711                         test_opt(inode->i_sb, DELALLOC)) {
2712                 /*
2713                  * With delalloc we want to sync the file
2714                  * so that we can make sure we allocate
2715                  * blocks for file
2716                  */
2717                 filemap_write_and_wait(mapping);
2718         }
2719
2720         if (EXT4_I(inode)->i_state & EXT4_STATE_JDATA) {
2721                 /*
2722                  * This is a REALLY heavyweight approach, but the use of
2723                  * bmap on dirty files is expected to be extremely rare:
2724                  * only if we run lilo or swapon on a freshly made file
2725                  * do we expect this to happen.
2726                  *
2727                  * (bmap requires CAP_SYS_RAWIO so this does not
2728                  * represent an unprivileged user DOS attack --- we'd be
2729                  * in trouble if mortal users could trigger this path at
2730                  * will.)
2731                  *
2732                  * NB. EXT4_STATE_JDATA is not set on files other than
2733                  * regular files.  If somebody wants to bmap a directory
2734                  * or symlink and gets confused because the buffer
2735                  * hasn't yet been flushed to disk, they deserve
2736                  * everything they get.
2737                  */
2738
2739                 EXT4_I(inode)->i_state &= ~EXT4_STATE_JDATA;
2740                 journal = EXT4_JOURNAL(inode);
2741                 jbd2_journal_lock_updates(journal);
2742                 err = jbd2_journal_flush(journal);
2743                 jbd2_journal_unlock_updates(journal);
2744
2745                 if (err)
2746                         return 0;
2747         }
2748
2749         return generic_block_bmap(mapping, block, ext4_get_block);
2750 }
2751
2752 static int bget_one(handle_t *handle, struct buffer_head *bh)
2753 {
2754         get_bh(bh);
2755         return 0;
2756 }
2757
2758 static int bput_one(handle_t *handle, struct buffer_head *bh)
2759 {
2760         put_bh(bh);
2761         return 0;
2762 }
2763
2764 /*
2765  * Note that we don't need to start a transaction unless we're journaling data
2766  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2767  * need to file the inode to the transaction's list in ordered mode because if
2768  * we are writing back data added by write(), the inode is already there and if
2769  * we are writing back data modified via mmap(), noone guarantees in which
2770  * transaction the data will hit the disk. In case we are journaling data, we
2771  * cannot start transaction directly because transaction start ranks above page
2772  * lock so we have to do some magic.
2773  *
2774  * In all journaling modes block_write_full_page() will start the I/O.
2775  *
2776  * Problem:
2777  *
2778  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2779  *              ext4_writepage()
2780  *
2781  * Similar for:
2782  *
2783  *      ext4_file_write() -> generic_file_write() -> __alloc_pages() -> ...
2784  *
2785  * Same applies to ext4_get_block().  We will deadlock on various things like
2786  * lock_journal and i_data_sem
2787  *
2788  * Setting PF_MEMALLOC here doesn't work - too many internal memory
2789  * allocations fail.
2790  *
2791  * 16May01: If we're reentered then journal_current_handle() will be
2792  *          non-zero. We simply *return*.
2793  *
2794  * 1 July 2001: @@@ FIXME:
2795  *   In journalled data mode, a data buffer may be metadata against the
2796  *   current transaction.  But the same file is part of a shared mapping
2797  *   and someone does a writepage() on it.
2798  *
2799  *   We will move the buffer onto the async_data list, but *after* it has
2800  *   been dirtied. So there's a small window where we have dirty data on
2801  *   BJ_Metadata.
2802  *
2803  *   Note that this only applies to the last partial page in the file.  The
2804  *   bit which block_write_full_page() uses prepare/commit for.  (That's
2805  *   broken code anyway: it's wrong for msync()).
2806  *
2807  *   It's a rare case: affects the final partial page, for journalled data
2808  *   where the file is subject to bith write() and writepage() in the same
2809  *   transction.  To fix it we'll need a custom block_write_full_page().
2810  *   We'll probably need that anyway for journalling writepage() output.
2811  *
2812  * We don't honour synchronous mounts for writepage().  That would be
2813  * disastrous.  Any write() or metadata operation will sync the fs for
2814  * us.
2815  *
2816  */
2817 static int __ext4_normal_writepage(struct page *page,
2818                                 struct writeback_control *wbc)
2819 {
2820         struct inode *inode = page->mapping->host;
2821
2822         if (test_opt(inode->i_sb, NOBH))
2823                 return nobh_writepage(page,
2824                                         ext4_normal_get_block_write, wbc);
2825         else
2826                 return block_write_full_page(page,
2827                                                 ext4_normal_get_block_write,
2828                                                 wbc);
2829 }
2830
2831 static int ext4_normal_writepage(struct page *page,
2832                                 struct writeback_control *wbc)
2833 {
2834         struct inode *inode = page->mapping->host;
2835         loff_t size = i_size_read(inode);
2836         loff_t len;
2837
2838         J_ASSERT(PageLocked(page));
2839         if (page->index == size >> PAGE_CACHE_SHIFT)
2840                 len = size & ~PAGE_CACHE_MASK;
2841         else
2842                 len = PAGE_CACHE_SIZE;
2843
2844         if (page_has_buffers(page)) {
2845                 /* if page has buffers it should all be mapped
2846                  * and allocated. If there are not buffers attached
2847                  * to the page we know the page is dirty but it lost
2848                  * buffers. That means that at some moment in time
2849                  * after write_begin() / write_end() has been called
2850                  * all buffers have been clean and thus they must have been
2851                  * written at least once. So they are all mapped and we can
2852                  * happily proceed with mapping them and writing the page.
2853                  */
2854                 BUG_ON(walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
2855                                         ext4_bh_unmapped_or_delay));
2856         }
2857
2858         if (!ext4_journal_current_handle())
2859                 return __ext4_normal_writepage(page, wbc);
2860
2861         redirty_page_for_writepage(wbc, page);
2862         unlock_page(page);
2863         return 0;
2864 }
2865
2866 static int __ext4_journalled_writepage(struct page *page,
2867                                 struct writeback_control *wbc)
2868 {
2869         struct address_space *mapping = page->mapping;
2870         struct inode *inode = mapping->host;
2871         struct buffer_head *page_bufs;
2872         handle_t *handle = NULL;
2873         int ret = 0;
2874         int err;
2875
2876         ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
2877                                         ext4_normal_get_block_write);
2878         if (ret != 0)
2879                 goto out_unlock;
2880
2881         page_bufs = page_buffers(page);
2882         walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE, NULL,
2883                                                                 bget_one);
2884         /* As soon as we unlock the page, it can go away, but we have
2885          * references to buffers so we are safe */
2886         unlock_page(page);
2887
2888         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2889         if (IS_ERR(handle)) {
2890                 ret = PTR_ERR(handle);
2891                 goto out;
2892         }
2893
2894         ret = walk_page_buffers(handle, page_bufs, 0,
2895                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
2896
2897         err = walk_page_buffers(handle, page_bufs, 0,
2898                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
2899         if (ret == 0)
2900                 ret = err;
2901         err = ext4_journal_stop(handle);
2902         if (!ret)
2903                 ret = err;
2904
2905         walk_page_buffers(handle, page_bufs, 0,
2906                                 PAGE_CACHE_SIZE, NULL, bput_one);
2907         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
2908         goto out;
2909
2910 out_unlock:
2911         unlock_page(page);
2912 out:
2913         return ret;
2914 }
2915
2916 static int ext4_journalled_writepage(struct page *page,
2917                                 struct writeback_control *wbc)
2918 {
2919         struct inode *inode = page->mapping->host;
2920         loff_t size = i_size_read(inode);
2921         loff_t len;
2922
2923         J_ASSERT(PageLocked(page));
2924         if (page->index == size >> PAGE_CACHE_SHIFT)
2925                 len = size & ~PAGE_CACHE_MASK;
2926         else
2927                 len = PAGE_CACHE_SIZE;
2928
2929         if (page_has_buffers(page)) {
2930                 /* if page has buffers it should all be mapped
2931                  * and allocated. If there are not buffers attached
2932                  * to the page we know the page is dirty but it lost
2933                  * buffers. That means that at some moment in time
2934                  * after write_begin() / write_end() has been called
2935                  * all buffers have been clean and thus they must have been
2936                  * written at least once. So they are all mapped and we can
2937                  * happily proceed with mapping them and writing the page.
2938                  */
2939                 BUG_ON(walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
2940                                         ext4_bh_unmapped_or_delay));
2941         }
2942
2943         if (ext4_journal_current_handle())
2944                 goto no_write;
2945
2946         if (PageChecked(page)) {
2947                 /*
2948                  * It's mmapped pagecache.  Add buffers and journal it.  There
2949                  * doesn't seem much point in redirtying the page here.
2950                  */
2951                 ClearPageChecked(page);
2952                 return __ext4_journalled_writepage(page, wbc);
2953         } else {
2954                 /*
2955                  * It may be a page full of checkpoint-mode buffers.  We don't
2956                  * really know unless we go poke around in the buffer_heads.
2957                  * But block_write_full_page will do the right thing.
2958                  */
2959                 return block_write_full_page(page,
2960                                                 ext4_normal_get_block_write,
2961                                                 wbc);
2962         }
2963 no_write:
2964         redirty_page_for_writepage(wbc, page);
2965         unlock_page(page);
2966         return 0;
2967 }
2968
2969 static int ext4_readpage(struct file *file, struct page *page)
2970 {
2971         return mpage_readpage(page, ext4_get_block);
2972 }
2973
2974 static int
2975 ext4_readpages(struct file *file, struct address_space *mapping,
2976                 struct list_head *pages, unsigned nr_pages)
2977 {
2978         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
2979 }
2980
2981 static void ext4_invalidatepage(struct page *page, unsigned long offset)
2982 {
2983         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2984
2985         /*
2986          * If it's a full truncate we just forget about the pending dirtying
2987          */
2988         if (offset == 0)
2989                 ClearPageChecked(page);
2990
2991         jbd2_journal_invalidatepage(journal, page, offset);
2992 }
2993
2994 static int ext4_releasepage(struct page *page, gfp_t wait)
2995 {
2996         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2997
2998         WARN_ON(PageChecked(page));
2999         if (!page_has_buffers(page))
3000                 return 0;
3001         return jbd2_journal_try_to_free_buffers(journal, page, wait);
3002 }
3003
3004 /*
3005  * If the O_DIRECT write will extend the file then add this inode to the
3006  * orphan list.  So recovery will truncate it back to the original size
3007  * if the machine crashes during the write.
3008  *
3009  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3010  * crashes then stale disk data _may_ be exposed inside the file. But current
3011  * VFS code falls back into buffered path in that case so we are safe.
3012  */
3013 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3014                         const struct iovec *iov, loff_t offset,
3015                         unsigned long nr_segs)
3016 {
3017         struct file *file = iocb->ki_filp;
3018         struct inode *inode = file->f_mapping->host;
3019         struct ext4_inode_info *ei = EXT4_I(inode);
3020         handle_t *handle;
3021         ssize_t ret;
3022         int orphan = 0;
3023         size_t count = iov_length(iov, nr_segs);
3024
3025         if (rw == WRITE) {
3026                 loff_t final_size = offset + count;
3027
3028                 if (final_size > inode->i_size) {
3029                         /* Credits for sb + inode write */
3030                         handle = ext4_journal_start(inode, 2);
3031                         if (IS_ERR(handle)) {
3032                                 ret = PTR_ERR(handle);
3033                                 goto out;
3034                         }
3035                         ret = ext4_orphan_add(handle, inode);
3036                         if (ret) {
3037                                 ext4_journal_stop(handle);
3038                                 goto out;
3039                         }
3040                         orphan = 1;
3041                         ei->i_disksize = inode->i_size;
3042                         ext4_journal_stop(handle);
3043                 }
3044         }
3045
3046         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
3047                                  offset, nr_segs,
3048                                  ext4_get_block, NULL);
3049
3050         if (orphan) {
3051                 int err;
3052
3053                 /* Credits for sb + inode write */
3054                 handle = ext4_journal_start(inode, 2);
3055                 if (IS_ERR(handle)) {
3056                         /* This is really bad luck. We've written the data
3057                          * but cannot extend i_size. Bail out and pretend
3058                          * the write failed... */
3059                         ret = PTR_ERR(handle);
3060                         goto out;
3061                 }
3062                 if (inode->i_nlink)
3063                         ext4_orphan_del(handle, inode);
3064                 if (ret > 0) {
3065                         loff_t end = offset + ret;
3066                         if (end > inode->i_size) {
3067                                 ei->i_disksize = end;
3068                                 i_size_write(inode, end);
3069                                 /*
3070                                  * We're going to return a positive `ret'
3071                                  * here due to non-zero-length I/O, so there's
3072                                  * no way of reporting error returns from
3073                                  * ext4_mark_inode_dirty() to userspace.  So
3074                                  * ignore it.
3075                                  */
3076                                 ext4_mark_inode_dirty(handle, inode);
3077                         }
3078                 }
3079                 err = ext4_journal_stop(handle);
3080                 if (ret == 0)
3081                         ret = err;
3082         }
3083 out:
3084         return ret;
3085 }
3086
3087 /*
3088  * Pages can be marked dirty completely asynchronously from ext4's journalling
3089  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3090  * much here because ->set_page_dirty is called under VFS locks.  The page is
3091  * not necessarily locked.
3092  *
3093  * We cannot just dirty the page and leave attached buffers clean, because the
3094  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3095  * or jbddirty because all the journalling code will explode.
3096  *
3097  * So what we do is to mark the page "pending dirty" and next time writepage
3098  * is called, propagate that into the buffers appropriately.
3099  */
3100 static int ext4_journalled_set_page_dirty(struct page *page)
3101 {
3102         SetPageChecked(page);
3103         return __set_page_dirty_nobuffers(page);
3104 }
3105
3106 static const struct address_space_operations ext4_ordered_aops = {
3107         .readpage               = ext4_readpage,
3108         .readpages              = ext4_readpages,
3109         .writepage              = ext4_normal_writepage,
3110         .sync_page              = block_sync_page,
3111         .write_begin            = ext4_write_begin,
3112         .write_end              = ext4_ordered_write_end,
3113         .bmap                   = ext4_bmap,
3114         .invalidatepage         = ext4_invalidatepage,
3115         .releasepage            = ext4_releasepage,
3116         .direct_IO              = ext4_direct_IO,
3117         .migratepage            = buffer_migrate_page,
3118         .is_partially_uptodate  = block_is_partially_uptodate,
3119 };
3120
3121 static const struct address_space_operations ext4_writeback_aops = {
3122         .readpage               = ext4_readpage,
3123         .readpages              = ext4_readpages,
3124         .writepage              = ext4_normal_writepage,
3125         .sync_page              = block_sync_page,
3126         .write_begin            = ext4_write_begin,
3127         .write_end              = ext4_writeback_write_end,
3128         .bmap                   = ext4_bmap,
3129         .invalidatepage         = ext4_invalidatepage,
3130         .releasepage            = ext4_releasepage,
3131         .direct_IO              = ext4_direct_IO,
3132         .migratepage            = buffer_migrate_page,
3133         .is_partially_uptodate  = block_is_partially_uptodate,
3134 };
3135
3136 static const struct address_space_operations ext4_journalled_aops = {
3137         .readpage               = ext4_readpage,
3138         .readpages              = ext4_readpages,
3139         .writepage              = ext4_journalled_writepage,
3140         .sync_page              = block_sync_page,
3141         .write_begin            = ext4_write_begin,
3142         .write_end              = ext4_journalled_write_end,
3143         .set_page_dirty         = ext4_journalled_set_page_dirty,
3144         .bmap                   = ext4_bmap,
3145         .invalidatepage         = ext4_invalidatepage,
3146         .releasepage            = ext4_releasepage,
3147         .is_partially_uptodate  = block_is_partially_uptodate,
3148 };
3149
3150 static const struct address_space_operations ext4_da_aops = {
3151         .readpage               = ext4_readpage,
3152         .readpages              = ext4_readpages,
3153         .writepage              = ext4_da_writepage,
3154         .writepages             = ext4_da_writepages,
3155         .sync_page              = block_sync_page,
3156         .write_begin            = ext4_da_write_begin,
3157         .write_end              = ext4_da_write_end,
3158         .bmap                   = ext4_bmap,
3159         .invalidatepage         = ext4_da_invalidatepage,
3160         .releasepage            = ext4_releasepage,
3161         .direct_IO              = ext4_direct_IO,
3162         .migratepage            = buffer_migrate_page,
3163         .is_partially_uptodate  = block_is_partially_uptodate,
3164 };
3165
3166 void ext4_set_aops(struct inode *inode)
3167 {
3168         if (ext4_should_order_data(inode) &&
3169                 test_opt(inode->i_sb, DELALLOC))
3170                 inode->i_mapping->a_ops = &ext4_da_aops;
3171         else if (ext4_should_order_data(inode))
3172                 inode->i_mapping->a_ops = &ext4_ordered_aops;
3173         else if (ext4_should_writeback_data(inode) &&
3174                  test_opt(inode->i_sb, DELALLOC))
3175                 inode->i_mapping->a_ops = &ext4_da_aops;
3176         else if (ext4_should_writeback_data(inode))
3177                 inode->i_mapping->a_ops = &ext4_writeback_aops;
3178         else
3179                 inode->i_mapping->a_ops = &ext4_journalled_aops;
3180 }
3181
3182 /*
3183  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3184  * up to the end of the block which corresponds to `from'.
3185  * This required during truncate. We need to physically zero the tail end
3186  * of that block so it doesn't yield old data if the file is later grown.
3187  */
3188 int ext4_block_truncate_page(handle_t *handle,
3189                 struct address_space *mapping, loff_t from)
3190 {
3191         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3192         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3193         unsigned blocksize, length, pos;
3194         ext4_lblk_t iblock;
3195         struct inode *inode = mapping->host;
3196         struct buffer_head *bh;
3197         struct page *page;
3198         int err = 0;
3199
3200         page = grab_cache_page(mapping, from >> PAGE_CACHE_SHIFT);
3201         if (!page)
3202                 return -EINVAL;
3203
3204         blocksize = inode->i_sb->s_blocksize;
3205         length = blocksize - (offset & (blocksize - 1));
3206         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3207
3208         /*
3209          * For "nobh" option,  we can only work if we don't need to
3210          * read-in the page - otherwise we create buffers to do the IO.
3211          */
3212         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
3213              ext4_should_writeback_data(inode) && PageUptodate(page)) {
3214                 zero_user(page, offset, length);
3215                 set_page_dirty(page);
3216                 goto unlock;
3217         }
3218
3219         if (!page_has_buffers(page))
3220                 create_empty_buffers(page, blocksize, 0);
3221
3222         /* Find the buffer that contains "offset" */
3223         bh = page_buffers(page);
3224         pos = blocksize;
3225         while (offset >= pos) {
3226                 bh = bh->b_this_page;
3227                 iblock++;
3228                 pos += blocksize;
3229         }
3230
3231         err = 0;
3232         if (buffer_freed(bh)) {
3233                 BUFFER_TRACE(bh, "freed: skip");
3234                 goto unlock;
3235         }
3236
3237         if (!buffer_mapped(bh)) {
3238                 BUFFER_TRACE(bh, "unmapped");
3239                 ext4_get_block(inode, iblock, bh, 0);
3240                 /* unmapped? It's a hole - nothing to do */
3241                 if (!buffer_mapped(bh)) {
3242                         BUFFER_TRACE(bh, "still unmapped");
3243                         goto unlock;
3244                 }
3245         }
3246
3247         /* Ok, it's mapped. Make sure it's up-to-date */
3248         if (PageUptodate(page))
3249                 set_buffer_uptodate(bh);
3250
3251         if (!buffer_uptodate(bh)) {
3252                 err = -EIO;
3253                 ll_rw_block(READ, 1, &bh);
3254                 wait_on_buffer(bh);
3255                 /* Uhhuh. Read error. Complain and punt. */
3256                 if (!buffer_uptodate(bh))
3257                         goto unlock;
3258         }
3259
3260         if (ext4_should_journal_data(inode)) {
3261                 BUFFER_TRACE(bh, "get write access");
3262                 err = ext4_journal_get_write_access(handle, bh);
3263                 if (err)
3264                         goto unlock;
3265         }
3266
3267         zero_user(page, offset, length);
3268
3269         BUFFER_TRACE(bh, "zeroed end of block");
3270
3271         err = 0;
3272         if (ext4_should_journal_data(inode)) {
3273                 err = ext4_journal_dirty_metadata(handle, bh);
3274         } else {
3275                 if (ext4_should_order_data(inode))
3276                         err = ext4_jbd2_file_inode(handle, inode);
3277                 mark_buffer_dirty(bh);
3278         }
3279
3280 unlock:
3281         unlock_page(page);
3282         page_cache_release(page);
3283         return err;
3284 }
3285
3286 /*
3287  * Probably it should be a library function... search for first non-zero word
3288  * or memcmp with zero_page, whatever is better for particular architecture.
3289  * Linus?
3290  */
3291 static inline int all_zeroes(__le32 *p, __le32 *q)
3292 {
3293         while (p < q)
3294                 if (*p++)
3295                         return 0;
3296         return 1;
3297 }
3298
3299 /**
3300  *      ext4_find_shared - find the indirect blocks for partial truncation.
3301  *      @inode:   inode in question
3302  *      @depth:   depth of the affected branch
3303  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
3304  *      @chain:   place to store the pointers to partial indirect blocks
3305  *      @top:     place to the (detached) top of branch
3306  *
3307  *      This is a helper function used by ext4_truncate().
3308  *
3309  *      When we do truncate() we may have to clean the ends of several
3310  *      indirect blocks but leave the blocks themselves alive. Block is
3311  *      partially truncated if some data below the new i_size is refered
3312  *      from it (and it is on the path to the first completely truncated
3313  *      data block, indeed).  We have to free the top of that path along
3314  *      with everything to the right of the path. Since no allocation
3315  *      past the truncation point is possible until ext4_truncate()
3316  *      finishes, we may safely do the latter, but top of branch may
3317  *      require special attention - pageout below the truncation point
3318  *      might try to populate it.
3319  *
3320  *      We atomically detach the top of branch from the tree, store the
3321  *      block number of its root in *@top, pointers to buffer_heads of
3322  *      partially truncated blocks - in @chain[].bh and pointers to
3323  *      their last elements that should not be removed - in
3324  *      @chain[].p. Return value is the pointer to last filled element
3325  *      of @chain.
3326  *
3327  *      The work left to caller to do the actual freeing of subtrees:
3328  *              a) free the subtree starting from *@top
3329  *              b) free the subtrees whose roots are stored in
3330  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
3331  *              c) free the subtrees growing from the inode past the @chain[0].
3332  *                      (no partially truncated stuff there).  */
3333
3334 static Indirect *ext4_find_shared(struct inode *inode, int depth,
3335                         ext4_lblk_t offsets[4], Indirect chain[4], __le32 *top)
3336 {
3337         Indirect *partial, *p;
3338         int k, err;
3339
3340         *top = 0;
3341         /* Make k index the deepest non-null offest + 1 */
3342         for (k = depth; k > 1 && !offsets[k-1]; k--)
3343                 ;
3344         partial = ext4_get_branch(inode, k, offsets, chain, &err);
3345         /* Writer: pointers */
3346         if (!partial)
3347                 partial = chain + k-1;
3348         /*
3349          * If the branch acquired continuation since we've looked at it -
3350          * fine, it should all survive and (new) top doesn't belong to us.
3351          */
3352         if (!partial->key && *partial->p)
3353                 /* Writer: end */
3354                 goto no_top;
3355         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
3356                 ;
3357         /*
3358          * OK, we've found the last block that must survive. The rest of our
3359          * branch should be detached before unlocking. However, if that rest
3360          * of branch is all ours and does not grow immediately from the inode
3361          * it's easier to cheat and just decrement partial->p.
3362          */
3363         if (p == chain + k - 1 && p > chain) {
3364                 p->p--;
3365         } else {
3366                 *top = *p->p;
3367                 /* Nope, don't do this in ext4.  Must leave the tree intact */
3368 #if 0
3369                 *p->p = 0;
3370 #endif
3371         }
3372         /* Writer: end */
3373
3374         while (partial > p) {
3375                 brelse(partial->bh);
3376                 partial--;
3377         }
3378 no_top:
3379         return partial;
3380 }
3381
3382 /*
3383  * Zero a number of block pointers in either an inode or an indirect block.
3384  * If we restart the transaction we must again get write access to the
3385  * indirect block for further modification.
3386  *
3387  * We release `count' blocks on disk, but (last - first) may be greater
3388  * than `count' because there can be holes in there.
3389  */
3390 static void ext4_clear_blocks(handle_t *handle, struct inode *inode,
3391                 struct buffer_head *bh, ext4_fsblk_t block_to_free,
3392                 unsigned long count, __le32 *first, __le32 *last)
3393 {
3394         __le32 *p;
3395         if (try_to_extend_transaction(handle, inode)) {
3396                 if (bh) {
3397                         BUFFER_TRACE(bh, "call ext4_journal_dirty_metadata");
3398                         ext4_journal_dirty_metadata(handle, bh);
3399                 }
3400                 ext4_mark_inode_dirty(handle, inode);
3401                 ext4_journal_test_restart(handle, inode);
3402                 if (bh) {
3403                         BUFFER_TRACE(bh, "retaking write access");
3404                         ext4_journal_get_write_access(handle, bh);
3405                 }
3406         }
3407
3408         /*
3409          * Any buffers which are on the journal will be in memory. We find
3410          * them on the hash table so jbd2_journal_revoke() will run jbd2_journal_forget()
3411          * on them.  We've already detached each block from the file, so
3412          * bforget() in jbd2_journal_forget() should be safe.
3413          *
3414          * AKPM: turn on bforget in jbd2_journal_forget()!!!
3415          */
3416         for (p = first; p < last; p++) {
3417                 u32 nr = le32_to_cpu(*p);
3418                 if (nr) {
3419                         struct buffer_head *tbh;
3420
3421                         *p = 0;
3422                         tbh = sb_find_get_block(inode->i_sb, nr);
3423                         ext4_forget(handle, 0, inode, tbh, nr);
3424                 }
3425         }
3426
3427         ext4_free_blocks(handle, inode, block_to_free, count, 0);
3428 }
3429
3430 /**
3431  * ext4_free_data - free a list of data blocks
3432  * @handle:     handle for this transaction
3433  * @inode:      inode we are dealing with
3434  * @this_bh:    indirect buffer_head which contains *@first and *@last
3435  * @first:      array of block numbers
3436  * @last:       points immediately past the end of array
3437  *
3438  * We are freeing all blocks refered from that array (numbers are stored as
3439  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
3440  *
3441  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
3442  * blocks are contiguous then releasing them at one time will only affect one
3443  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
3444  * actually use a lot of journal space.
3445  *
3446  * @this_bh will be %NULL if @first and @last point into the inode's direct
3447  * block pointers.
3448  */
3449 static void ext4_free_data(handle_t *handle, struct inode *inode,
3450                            struct buffer_head *this_bh,
3451                            __le32 *first, __le32 *last)
3452 {
3453         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
3454         unsigned long count = 0;            /* Number of blocks in the run */
3455         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
3456                                                corresponding to
3457                                                block_to_free */
3458         ext4_fsblk_t nr;                    /* Current block # */
3459         __le32 *p;                          /* Pointer into inode/ind
3460                                                for current block */
3461         int err;
3462
3463         if (this_bh) {                          /* For indirect block */
3464                 BUFFER_TRACE(this_bh, "get_write_access");
3465                 err = ext4_journal_get_write_access(handle, this_bh);
3466                 /* Important: if we can't update the indirect pointers
3467                  * to the blocks, we can't free them. */
3468                 if (err)
3469                         return;
3470         }
3471
3472         for (p = first; p < last; p++) {
3473                 nr = le32_to_cpu(*p);
3474                 if (nr) {
3475                         /* accumulate blocks to free if they're contiguous */
3476                         if (count == 0) {
3477                                 block_to_free = nr;
3478                                 block_to_free_p = p;
3479                                 count = 1;
3480                         } else if (nr == block_to_free + count) {
3481                                 count++;
3482                         } else {
3483                                 ext4_clear_blocks(handle, inode, this_bh,
3484                                                   block_to_free,
3485                                                   count, block_to_free_p, p);
3486                                 block_to_free = nr;
3487                                 block_to_free_p = p;
3488                                 count = 1;
3489                         }
3490                 }
3491         }
3492
3493         if (count > 0)
3494                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
3495                                   count, block_to_free_p, p);
3496
3497         if (this_bh) {
3498                 BUFFER_TRACE(this_bh, "call ext4_journal_dirty_metadata");
3499
3500                 /*
3501                  * The buffer head should have an attached journal head at this
3502                  * point. However, if the data is corrupted and an indirect
3503                  * block pointed to itself, it would have been detached when
3504                  * the block was cleared. Check for this instead of OOPSing.
3505                  */
3506                 if (bh2jh(this_bh))
3507                         ext4_journal_dirty_metadata(handle, this_bh);
3508                 else
3509                         ext4_error(inode->i_sb, __func__,
3510                                    "circular indirect block detected, "
3511                                    "inode=%lu, block=%llu",
3512                                    inode->i_ino,
3513                                    (unsigned long long) this_bh->b_blocknr);
3514         }
3515 }
3516
3517 /**
3518  *      ext4_free_branches - free an array of branches
3519  *      @handle: JBD handle for this transaction
3520  *      @inode: inode we are dealing with
3521  *      @parent_bh: the buffer_head which contains *@first and *@last
3522  *      @first: array of block numbers
3523  *      @last:  pointer immediately past the end of array
3524  *      @depth: depth of the branches to free
3525  *
3526  *      We are freeing all blocks refered from these branches (numbers are
3527  *      stored as little-endian 32-bit) and updating @inode->i_blocks
3528  *      appropriately.
3529  */
3530 static void ext4_free_branches(handle_t *handle, struct inode *inode,
3531                                struct buffer_head *parent_bh,
3532                                __le32 *first, __le32 *last, int depth)
3533 {
3534         ext4_fsblk_t nr;
3535         __le32 *p;
3536
3537         if (is_handle_aborted(handle))
3538                 return;
3539
3540         if (depth--) {
3541                 struct buffer_head *bh;
3542                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
3543                 p = last;
3544                 while (--p >= first) {
3545                         nr = le32_to_cpu(*p);
3546                         if (!nr)
3547                                 continue;               /* A hole */
3548
3549                         /* Go read the buffer for the next level down */
3550                         bh = sb_bread(inode->i_sb, nr);
3551
3552                         /*
3553                          * A read failure? Report error and clear slot
3554                          * (should be rare).
3555                          */
3556                         if (!bh) {
3557                                 ext4_error(inode->i_sb, "ext4_free_branches",
3558                                            "Read failure, inode=%lu, block=%llu",
3559                                            inode->i_ino, nr);
3560                                 continue;
3561                         }
3562
3563                         /* This zaps the entire block.  Bottom up. */
3564                         BUFFER_TRACE(bh, "free child branches");
3565                         ext4_free_branches(handle, inode, bh,
3566                                         (__le32 *) bh->b_data,
3567                                         (__le32 *) bh->b_data + addr_per_block,
3568                                         depth);
3569
3570                         /*
3571                          * We've probably journalled the indirect block several
3572                          * times during the truncate.  But it's no longer
3573                          * needed and we now drop it from the transaction via
3574                          * jbd2_journal_revoke().
3575                          *
3576                          * That's easy if it's exclusively part of this
3577                          * transaction.  But if it's part of the committing
3578                          * transaction then jbd2_journal_forget() will simply
3579                          * brelse() it.  That means that if the underlying
3580                          * block is reallocated in ext4_get_block(),
3581                          * unmap_underlying_metadata() will find this block
3582                          * and will try to get rid of it.  damn, damn.
3583                          *
3584                          * If this block has already been committed to the
3585                          * journal, a revoke record will be written.  And
3586                          * revoke records must be emitted *before* clearing
3587                          * this block's bit in the bitmaps.
3588                          */
3589                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
3590
3591                         /*
3592                          * Everything below this this pointer has been
3593                          * released.  Now let this top-of-subtree go.
3594                          *
3595                          * We want the freeing of this indirect block to be
3596                          * atomic in the journal with the updating of the
3597                          * bitmap block which owns it.  So make some room in
3598                          * the journal.
3599                          *
3600                          * We zero the parent pointer *after* freeing its
3601                          * pointee in the bitmaps, so if extend_transaction()
3602                          * for some reason fails to put the bitmap changes and
3603                          * the release into the same transaction, recovery
3604                          * will merely complain about releasing a free block,
3605                          * rather than leaking blocks.
3606                          */
3607                         if (is_handle_aborted(handle))
3608                                 return;
3609                         if (try_to_extend_transaction(handle, inode)) {
3610                                 ext4_mark_inode_dirty(handle, inode);
3611                                 ext4_journal_test_restart(handle, inode);
3612                         }
3613
3614                         ext4_free_blocks(handle, inode, nr, 1, 1);
3615
3616                         if (parent_bh) {
3617                                 /*
3618                                  * The block which we have just freed is
3619                                  * pointed to by an indirect block: journal it
3620                                  */
3621                                 BUFFER_TRACE(parent_bh, "get_write_access");
3622                                 if (!ext4_journal_get_write_access(handle,
3623                                                                    parent_bh)){
3624                                         *p = 0;
3625                                         BUFFER_TRACE(parent_bh,
3626                                         "call ext4_journal_dirty_metadata");
3627                                         ext4_journal_dirty_metadata(handle,
3628                                                                     parent_bh);
3629                                 }
3630                         }
3631                 }
3632         } else {
3633                 /* We have reached the bottom of the tree. */
3634                 BUFFER_TRACE(parent_bh, "free data blocks");
3635                 ext4_free_data(handle, inode, parent_bh, first, last);
3636         }
3637 }
3638
3639 int ext4_can_truncate(struct inode *inode)
3640 {
3641         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
3642                 return 0;
3643         if (S_ISREG(inode->i_mode))
3644                 return 1;
3645         if (S_ISDIR(inode->i_mode))
3646                 return 1;
3647         if (S_ISLNK(inode->i_mode))
3648                 return !ext4_inode_is_fast_symlink(inode);
3649         return 0;
3650 }
3651
3652 /*
3653  * ext4_truncate()
3654  *
3655  * We block out ext4_get_block() block instantiations across the entire
3656  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
3657  * simultaneously on behalf of the same inode.
3658  *
3659  * As we work through the truncate and commmit bits of it to the journal there
3660  * is one core, guiding principle: the file's tree must always be consistent on
3661  * disk.  We must be able to restart the truncate after a crash.
3662  *
3663  * The file's tree may be transiently inconsistent in memory (although it
3664  * probably isn't), but whenever we close off and commit a journal transaction,
3665  * the contents of (the filesystem + the journal) must be consistent and
3666  * restartable.  It's pretty simple, really: bottom up, right to left (although
3667  * left-to-right works OK too).
3668  *
3669  * Note that at recovery time, journal replay occurs *before* the restart of
3670  * truncate against the orphan inode list.
3671  *
3672  * The committed inode has the new, desired i_size (which is the same as
3673  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
3674  * that this inode's truncate did not complete and it will again call
3675  * ext4_truncate() to have another go.  So there will be instantiated blocks
3676  * to the right of the truncation point in a crashed ext4 filesystem.  But
3677  * that's fine - as long as they are linked from the inode, the post-crash
3678  * ext4_truncate() run will find them and release them.
3679  */
3680 void ext4_truncate(struct inode *inode)
3681 {
3682         handle_t *handle;
3683         struct ext4_inode_info *ei = EXT4_I(inode);
3684         __le32 *i_data = ei->i_data;
3685         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
3686         struct address_space *mapping = inode->i_mapping;
3687         ext4_lblk_t offsets[4];
3688         Indirect chain[4];
3689         Indirect *partial;
3690         __le32 nr = 0;
3691         int n;
3692         ext4_lblk_t last_block;
3693         unsigned blocksize = inode->i_sb->s_blocksize;
3694
3695         if (!ext4_can_truncate(inode))
3696                 return;
3697
3698         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
3699                 ext4_ext_truncate(inode);
3700                 return;
3701         }
3702
3703         handle = start_transaction(inode);
3704         if (IS_ERR(handle))
3705                 return;         /* AKPM: return what? */
3706
3707         last_block = (inode->i_size + blocksize-1)
3708                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
3709
3710         if (inode->i_size & (blocksize - 1))
3711                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
3712                         goto out_stop;
3713
3714         n = ext4_block_to_path(inode, last_block, offsets, NULL);
3715         if (n == 0)
3716                 goto out_stop;  /* error */
3717
3718         /*
3719          * OK.  This truncate is going to happen.  We add the inode to the
3720          * orphan list, so that if this truncate spans multiple transactions,
3721          * and we crash, we will resume the truncate when the filesystem
3722          * recovers.  It also marks the inode dirty, to catch the new size.
3723          *
3724          * Implication: the file must always be in a sane, consistent
3725          * truncatable state while each transaction commits.
3726          */
3727         if (ext4_orphan_add(handle, inode))
3728                 goto out_stop;
3729
3730         /*
3731          * From here we block out all ext4_get_block() callers who want to
3732          * modify the block allocation tree.
3733          */
3734         down_write(&ei->i_data_sem);
3735
3736         ext4_discard_preallocations(inode);
3737
3738         /*
3739          * The orphan list entry will now protect us from any crash which
3740          * occurs before the truncate completes, so it is now safe to propagate
3741          * the new, shorter inode size (held for now in i_size) into the
3742          * on-disk inode. We do this via i_disksize, which is the value which
3743          * ext4 *really* writes onto the disk inode.
3744          */
3745         ei->i_disksize = inode->i_size;
3746
3747         if (n == 1) {           /* direct blocks */
3748                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
3749                                i_data + EXT4_NDIR_BLOCKS);
3750                 goto do_indirects;
3751         }
3752
3753         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
3754         /* Kill the top of shared branch (not detached) */
3755         if (nr) {
3756                 if (partial == chain) {
3757                         /* Shared branch grows from the inode */
3758                         ext4_free_branches(handle, inode, NULL,
3759                                            &nr, &nr+1, (chain+n-1) - partial);
3760                         *partial->p = 0;
3761                         /*
3762                          * We mark the inode dirty prior to restart,
3763                          * and prior to stop.  No need for it here.
3764                          */
3765                 } else {
3766                         /* Shared branch grows from an indirect block */
3767                         BUFFER_TRACE(partial->bh, "get_write_access");
3768                         ext4_free_branches(handle, inode, partial->bh,
3769                                         partial->p,
3770                                         partial->p+1, (chain+n-1) - partial);
3771                 }
3772         }
3773         /* Clear the ends of indirect blocks on the shared branch */
3774         while (partial > chain) {
3775                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
3776                                    (__le32*)partial->bh->b_data+addr_per_block,
3777                                    (chain+n-1) - partial);
3778                 BUFFER_TRACE(partial->bh, "call brelse");
3779                 brelse (partial->bh);
3780                 partial--;
3781         }
3782 do_indirects:
3783         /* Kill the remaining (whole) subtrees */
3784         switch (offsets[0]) {
3785         default:
3786                 nr = i_data[EXT4_IND_BLOCK];
3787                 if (nr) {
3788                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
3789                         i_data[EXT4_IND_BLOCK] = 0;
3790                 }
3791         case EXT4_IND_BLOCK:
3792                 nr = i_data[EXT4_DIND_BLOCK];
3793                 if (nr) {
3794                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
3795                         i_data[EXT4_DIND_BLOCK] = 0;
3796                 }
3797         case EXT4_DIND_BLOCK:
3798                 nr = i_data[EXT4_TIND_BLOCK];
3799                 if (nr) {
3800                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
3801                         i_data[EXT4_TIND_BLOCK] = 0;
3802                 }
3803         case EXT4_TIND_BLOCK:
3804                 ;
3805         }
3806
3807         up_write(&ei->i_data_sem);
3808         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
3809         ext4_mark_inode_dirty(handle, inode);
3810
3811         /*
3812          * In a multi-transaction truncate, we only make the final transaction
3813          * synchronous
3814          */
3815         if (IS_SYNC(inode))
3816                 handle->h_sync = 1;
3817 out_stop:
3818         /*
3819          * If this was a simple ftruncate(), and the file will remain alive
3820          * then we need to clear up the orphan record which we created above.
3821          * However, if this was a real unlink then we were called by
3822          * ext4_delete_inode(), and we allow that function to clean up the
3823          * orphan info for us.
3824          */
3825         if (inode->i_nlink)
3826                 ext4_orphan_del(handle, inode);
3827
3828         ext4_journal_stop(handle);
3829 }
3830
3831 /*
3832  * ext4_get_inode_loc returns with an extra refcount against the inode's
3833  * underlying buffer_head on success. If 'in_mem' is true, we have all
3834  * data in memory that is needed to recreate the on-disk version of this
3835  * inode.
3836  */
3837 static int __ext4_get_inode_loc(struct inode *inode,
3838                                 struct ext4_iloc *iloc, int in_mem)
3839 {
3840         struct ext4_group_desc  *gdp;
3841         struct buffer_head      *bh;
3842         struct super_block      *sb = inode->i_sb;
3843         ext4_fsblk_t            block;
3844         int                     inodes_per_block, inode_offset;
3845
3846         iloc->bh = 0;
3847         if (!ext4_valid_inum(sb, inode->i_ino))
3848                 return -EIO;
3849
3850         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
3851         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
3852         if (!gdp)
3853                 return -EIO;
3854
3855         /*
3856          * Figure out the offset within the block group inode table
3857          */
3858         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
3859         inode_offset = ((inode->i_ino - 1) %
3860                         EXT4_INODES_PER_GROUP(sb));
3861         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
3862         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
3863
3864         bh = sb_getblk(sb, block);
3865         if (!bh) {
3866                 ext4_error(sb, "ext4_get_inode_loc", "unable to read "
3867                            "inode block - inode=%lu, block=%llu",
3868                            inode->i_ino, block);
3869                 return -EIO;
3870         }
3871         if (!buffer_uptodate(bh)) {
3872                 lock_buffer(bh);
3873
3874                 /*
3875                  * If the buffer has the write error flag, we have failed
3876                  * to write out another inode in the same block.  In this
3877                  * case, we don't have to read the block because we may
3878                  * read the old inode data successfully.
3879                  */
3880                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
3881                         set_buffer_uptodate(bh);
3882
3883                 if (buffer_uptodate(bh)) {
3884                         /* someone brought it uptodate while we waited */
3885                         unlock_buffer(bh);
3886                         goto has_buffer;
3887                 }
3888
3889                 /*
3890                  * If we have all information of the inode in memory and this
3891                  * is the only valid inode in the block, we need not read the
3892                  * block.
3893                  */
3894                 if (in_mem) {
3895                         struct buffer_head *bitmap_bh;
3896                         int i, start;
3897
3898                         start = inode_offset & ~(inodes_per_block - 1);
3899
3900                         /* Is the inode bitmap in cache? */
3901                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
3902                         if (!bitmap_bh)
3903                                 goto make_io;
3904
3905                         /*
3906                          * If the inode bitmap isn't in cache then the
3907                          * optimisation may end up performing two reads instead
3908                          * of one, so skip it.
3909                          */
3910                         if (!buffer_uptodate(bitmap_bh)) {
3911                                 brelse(bitmap_bh);
3912                                 goto make_io;
3913                         }
3914                         for (i = start; i < start + inodes_per_block; i++) {
3915                                 if (i == inode_offset)
3916                                         continue;
3917                                 if (ext4_test_bit(i, bitmap_bh->b_data))
3918                                         break;
3919                         }
3920                         brelse(bitmap_bh);
3921                         if (i == start + inodes_per_block) {
3922                                 /* all other inodes are free, so skip I/O */
3923                                 memset(bh->b_data, 0, bh->b_size);
3924                                 set_buffer_uptodate(bh);
3925                                 unlock_buffer(bh);
3926                                 goto has_buffer;
3927                         }
3928                 }
3929
3930 make_io:
3931                 /*
3932                  * If we need to do any I/O, try to pre-readahead extra
3933                  * blocks from the inode table.
3934                  */
3935                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
3936                         ext4_fsblk_t b, end, table;
3937                         unsigned num;
3938
3939                         table = ext4_inode_table(sb, gdp);
3940                         /* Make sure s_inode_readahead_blks is a power of 2 */
3941                         while (EXT4_SB(sb)->s_inode_readahead_blks &
3942                                (EXT4_SB(sb)->s_inode_readahead_blks-1))
3943                                 EXT4_SB(sb)->s_inode_readahead_blks = 
3944                                    (EXT4_SB(sb)->s_inode_readahead_blks &
3945                                     (EXT4_SB(sb)->s_inode_readahead_blks-1));
3946                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
3947                         if (table > b)
3948                                 b = table;
3949                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
3950                         num = EXT4_INODES_PER_GROUP(sb);
3951                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
3952                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
3953                                 num -= le16_to_cpu(gdp->bg_itable_unused);
3954                         table += num / inodes_per_block;
3955                         if (end > table)
3956                                 end = table;
3957                         while (b <= end)
3958                                 sb_breadahead(sb, b++);
3959                 }
3960
3961                 /*
3962                  * There are other valid inodes in the buffer, this inode
3963                  * has in-inode xattrs, or we don't have this inode in memory.
3964                  * Read the block from disk.
3965                  */
3966                 get_bh(bh);
3967                 bh->b_end_io = end_buffer_read_sync;
3968                 submit_bh(READ_META, bh);
3969                 wait_on_buffer(bh);
3970                 if (!buffer_uptodate(bh)) {
3971                         ext4_error(sb, __func__,
3972                                    "unable to read inode block - inode=%lu, "
3973                                    "block=%llu", inode->i_ino, block);
3974                         brelse(bh);
3975                         return -EIO;
3976                 }
3977         }
3978 has_buffer:
3979         iloc->bh = bh;
3980         return 0;
3981 }
3982
3983 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
3984 {
3985         /* We have all inode data except xattrs in memory here. */
3986         return __ext4_get_inode_loc(inode, iloc,
3987                 !(EXT4_I(inode)->i_state & EXT4_STATE_XATTR));
3988 }
3989
3990 void ext4_set_inode_flags(struct inode *inode)
3991 {
3992         unsigned int flags = EXT4_I(inode)->i_flags;
3993
3994         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
3995         if (flags & EXT4_SYNC_FL)
3996                 inode->i_flags |= S_SYNC;
3997         if (flags & EXT4_APPEND_FL)
3998                 inode->i_flags |= S_APPEND;
3999         if (flags & EXT4_IMMUTABLE_FL)
4000                 inode->i_flags |= S_IMMUTABLE;
4001         if (flags & EXT4_NOATIME_FL)
4002                 inode->i_flags |= S_NOATIME;
4003         if (flags & EXT4_DIRSYNC_FL)
4004                 inode->i_flags |= S_DIRSYNC;
4005 }
4006
4007 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4008 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4009 {
4010         unsigned int flags = ei->vfs_inode.i_flags;
4011
4012         ei->i_flags &= ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4013                         EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|EXT4_DIRSYNC_FL);
4014         if (flags & S_SYNC)
4015                 ei->i_flags |= EXT4_SYNC_FL;
4016         if (flags & S_APPEND)
4017                 ei->i_flags |= EXT4_APPEND_FL;
4018         if (flags & S_IMMUTABLE)
4019                 ei->i_flags |= EXT4_IMMUTABLE_FL;
4020         if (flags & S_NOATIME)
4021                 ei->i_flags |= EXT4_NOATIME_FL;
4022         if (flags & S_DIRSYNC)
4023                 ei->i_flags |= EXT4_DIRSYNC_FL;
4024 }
4025 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4026                                         struct ext4_inode_info *ei)
4027 {
4028         blkcnt_t i_blocks ;
4029         struct inode *inode = &(ei->vfs_inode);
4030         struct super_block *sb = inode->i_sb;
4031
4032         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4033                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4034                 /* we are using combined 48 bit field */
4035                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4036                                         le32_to_cpu(raw_inode->i_blocks_lo);
4037                 if (ei->i_flags & EXT4_HUGE_FILE_FL) {
4038                         /* i_blocks represent file system block size */
4039                         return i_blocks  << (inode->i_blkbits - 9);
4040                 } else {
4041                         return i_blocks;
4042                 }
4043         } else {
4044                 return le32_to_cpu(raw_inode->i_blocks_lo);
4045         }
4046 }
4047
4048 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4049 {
4050         struct ext4_iloc iloc;
4051         struct ext4_inode *raw_inode;
4052         struct ext4_inode_info *ei;
4053         struct buffer_head *bh;
4054         struct inode *inode;
4055         long ret;
4056         int block;
4057
4058         inode = iget_locked(sb, ino);
4059         if (!inode)
4060                 return ERR_PTR(-ENOMEM);
4061         if (!(inode->i_state & I_NEW))
4062                 return inode;
4063
4064         ei = EXT4_I(inode);
4065 #ifdef CONFIG_EXT4_FS_POSIX_ACL
4066         ei->i_acl = EXT4_ACL_NOT_CACHED;
4067         ei->i_default_acl = EXT4_ACL_NOT_CACHED;
4068 #endif
4069
4070         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4071         if (ret < 0)
4072                 goto bad_inode;
4073         bh = iloc.bh;
4074         raw_inode = ext4_raw_inode(&iloc);
4075         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4076         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4077         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4078         if (!(test_opt(inode->i_sb, NO_UID32))) {
4079                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4080                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4081         }
4082         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
4083
4084         ei->i_state = 0;
4085         ei->i_dir_start_lookup = 0;
4086         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4087         /* We now have enough fields to check if the inode was active or not.
4088          * This is needed because nfsd might try to access dead inodes
4089          * the test is that same one that e2fsck uses
4090          * NeilBrown 1999oct15
4091          */
4092         if (inode->i_nlink == 0) {
4093                 if (inode->i_mode == 0 ||
4094                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
4095                         /* this inode is deleted */
4096                         brelse(bh);
4097                         ret = -ESTALE;
4098                         goto bad_inode;
4099                 }
4100                 /* The only unlinked inodes we let through here have
4101                  * valid i_mode and are being read by the orphan
4102                  * recovery code: that's fine, we're about to complete
4103                  * the process of deleting those. */
4104         }
4105         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4106         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4107         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4108         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
4109             cpu_to_le32(EXT4_OS_HURD)) {
4110                 ei->i_file_acl |=
4111                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4112         }
4113         inode->i_size = ext4_isize(raw_inode);
4114         ei->i_disksize = inode->i_size;
4115         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4116         ei->i_block_group = iloc.block_group;
4117         /*
4118          * NOTE! The in-memory inode i_data array is in little-endian order
4119          * even on big-endian machines: we do NOT byteswap the block numbers!
4120          */
4121         for (block = 0; block < EXT4_N_BLOCKS; block++)
4122                 ei->i_data[block] = raw_inode->i_block[block];
4123         INIT_LIST_HEAD(&ei->i_orphan);
4124
4125         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4126                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4127                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4128                     EXT4_INODE_SIZE(inode->i_sb)) {
4129                         brelse(bh);
4130                         ret = -EIO;
4131                         goto bad_inode;
4132                 }
4133                 if (ei->i_extra_isize == 0) {
4134                         /* The extra space is currently unused. Use it. */
4135                         ei->i_extra_isize = sizeof(struct ext4_inode) -
4136                                             EXT4_GOOD_OLD_INODE_SIZE;
4137                 } else {
4138                         __le32 *magic = (void *)raw_inode +
4139                                         EXT4_GOOD_OLD_INODE_SIZE +
4140                                         ei->i_extra_isize;
4141                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
4142                                  ei->i_state |= EXT4_STATE_XATTR;
4143                 }
4144         } else
4145                 ei->i_extra_isize = 0;
4146
4147         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4148         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4149         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4150         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4151
4152         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4153         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4154                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4155                         inode->i_version |=
4156                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4157         }
4158
4159         if (S_ISREG(inode->i_mode)) {
4160                 inode->i_op = &ext4_file_inode_operations;
4161                 inode->i_fop = &ext4_file_operations;
4162                 ext4_set_aops(inode);
4163         } else if (S_ISDIR(inode->i_mode)) {
4164                 inode->i_op = &ext4_dir_inode_operations;
4165                 inode->i_fop = &ext4_dir_operations;
4166         } else if (S_ISLNK(inode->i_mode)) {
4167                 if (ext4_inode_is_fast_symlink(inode))
4168                         inode->i_op = &ext4_fast_symlink_inode_operations;
4169                 else {
4170                         inode->i_op = &ext4_symlink_inode_operations;
4171                         ext4_set_aops(inode);
4172                 }
4173         } else {
4174                 inode->i_op = &ext4_special_inode_operations;
4175                 if (raw_inode->i_block[0])
4176                         init_special_inode(inode, inode->i_mode,
4177                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4178                 else
4179                         init_special_inode(inode, inode->i_mode,
4180                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4181         }
4182         brelse(iloc.bh);
4183         ext4_set_inode_flags(inode);
4184         unlock_new_inode(inode);
4185         return inode;
4186
4187 bad_inode:
4188         iget_failed(inode);
4189         return ERR_PTR(ret);
4190 }
4191
4192 static int ext4_inode_blocks_set(handle_t *handle,
4193                                 struct ext4_inode *raw_inode,
4194                                 struct ext4_inode_info *ei)
4195 {
4196         struct inode *inode = &(ei->vfs_inode);
4197         u64 i_blocks = inode->i_blocks;
4198         struct super_block *sb = inode->i_sb;
4199
4200         if (i_blocks <= ~0U) {
4201                 /*
4202                  * i_blocks can be represnted in a 32 bit variable
4203                  * as multiple of 512 bytes
4204                  */
4205                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4206                 raw_inode->i_blocks_high = 0;
4207                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4208                 return 0;
4209         }
4210         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
4211                 return -EFBIG;
4212
4213         if (i_blocks <= 0xffffffffffffULL) {
4214                 /*
4215                  * i_blocks can be represented in a 48 bit variable
4216                  * as multiple of 512 bytes
4217                  */
4218                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4219                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4220                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4221         } else {
4222                 ei->i_flags |= EXT4_HUGE_FILE_FL;
4223                 /* i_block is stored in file system block size */
4224                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4225                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4226                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4227         }
4228         return 0;
4229 }
4230
4231 /*
4232  * Post the struct inode info into an on-disk inode location in the
4233  * buffer-cache.  This gobbles the caller's reference to the
4234  * buffer_head in the inode location struct.
4235  *
4236  * The caller must have write access to iloc->bh.
4237  */
4238 static int ext4_do_update_inode(handle_t *handle,
4239                                 struct inode *inode,
4240                                 struct ext4_iloc *iloc)
4241 {
4242         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
4243         struct ext4_inode_info *ei = EXT4_I(inode);
4244         struct buffer_head *bh = iloc->bh;
4245         int err = 0, rc, block;
4246
4247         /* For fields not not tracking in the in-memory inode,
4248          * initialise them to zero for new inodes. */
4249         if (ei->i_state & EXT4_STATE_NEW)
4250                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
4251
4252         ext4_get_inode_flags(ei);
4253         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
4254         if (!(test_opt(inode->i_sb, NO_UID32))) {
4255                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
4256                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
4257 /*
4258  * Fix up interoperability with old kernels. Otherwise, old inodes get
4259  * re-used with the upper 16 bits of the uid/gid intact
4260  */
4261                 if (!ei->i_dtime) {
4262                         raw_inode->i_uid_high =
4263                                 cpu_to_le16(high_16_bits(inode->i_uid));
4264                         raw_inode->i_gid_high =
4265                                 cpu_to_le16(high_16_bits(inode->i_gid));
4266                 } else {
4267                         raw_inode->i_uid_high = 0;
4268                         raw_inode->i_gid_high = 0;
4269                 }
4270         } else {
4271                 raw_inode->i_uid_low =
4272                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
4273                 raw_inode->i_gid_low =
4274                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
4275                 raw_inode->i_uid_high = 0;
4276                 raw_inode->i_gid_high = 0;
4277         }
4278         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
4279
4280         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
4281         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
4282         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
4283         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
4284
4285         if (ext4_inode_blocks_set(handle, raw_inode, ei))
4286                 goto out_brelse;
4287         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
4288         /* clear the migrate flag in the raw_inode */
4289         raw_inode->i_flags = cpu_to_le32(ei->i_flags & ~EXT4_EXT_MIGRATE);
4290         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
4291             cpu_to_le32(EXT4_OS_HURD))
4292                 raw_inode->i_file_acl_high =
4293                         cpu_to_le16(ei->i_file_acl >> 32);
4294         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
4295         ext4_isize_set(raw_inode, ei->i_disksize);
4296         if (ei->i_disksize > 0x7fffffffULL) {
4297                 struct super_block *sb = inode->i_sb;
4298                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
4299                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
4300                                 EXT4_SB(sb)->s_es->s_rev_level ==
4301                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
4302                         /* If this is the first large file
4303                          * created, add a flag to the superblock.
4304                          */
4305                         err = ext4_journal_get_write_access(handle,
4306                                         EXT4_SB(sb)->s_sbh);
4307                         if (err)
4308                                 goto out_brelse;
4309                         ext4_update_dynamic_rev(sb);
4310                         EXT4_SET_RO_COMPAT_FEATURE(sb,
4311                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
4312                         sb->s_dirt = 1;
4313                         handle->h_sync = 1;
4314                         err = ext4_journal_dirty_metadata(handle,
4315                                         EXT4_SB(sb)->s_sbh);
4316                 }
4317         }
4318         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
4319         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
4320                 if (old_valid_dev(inode->i_rdev)) {
4321                         raw_inode->i_block[0] =
4322                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
4323                         raw_inode->i_block[1] = 0;
4324                 } else {
4325                         raw_inode->i_block[0] = 0;
4326                         raw_inode->i_block[1] =
4327                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
4328                         raw_inode->i_block[2] = 0;
4329                 }
4330         } else for (block = 0; block < EXT4_N_BLOCKS; block++)
4331                 raw_inode->i_block[block] = ei->i_data[block];
4332
4333         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
4334         if (ei->i_extra_isize) {
4335                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4336                         raw_inode->i_version_hi =
4337                         cpu_to_le32(inode->i_version >> 32);
4338                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
4339         }
4340
4341
4342         BUFFER_TRACE(bh, "call ext4_journal_dirty_metadata");
4343         rc = ext4_journal_dirty_metadata(handle, bh);
4344         if (!err)
4345                 err = rc;
4346         ei->i_state &= ~EXT4_STATE_NEW;
4347
4348 out_brelse:
4349         brelse(bh);
4350         ext4_std_error(inode->i_sb, err);
4351         return err;
4352 }
4353
4354 /*
4355  * ext4_write_inode()
4356  *
4357  * We are called from a few places:
4358  *
4359  * - Within generic_file_write() for O_SYNC files.
4360  *   Here, there will be no transaction running. We wait for any running
4361  *   trasnaction to commit.
4362  *
4363  * - Within sys_sync(), kupdate and such.
4364  *   We wait on commit, if tol to.
4365  *
4366  * - Within prune_icache() (PF_MEMALLOC == true)
4367  *   Here we simply return.  We can't afford to block kswapd on the
4368  *   journal commit.
4369  *
4370  * In all cases it is actually safe for us to return without doing anything,
4371  * because the inode has been copied into a raw inode buffer in
4372  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
4373  * knfsd.
4374  *
4375  * Note that we are absolutely dependent upon all inode dirtiers doing the
4376  * right thing: they *must* call mark_inode_dirty() after dirtying info in
4377  * which we are interested.
4378  *
4379  * It would be a bug for them to not do this.  The code:
4380  *
4381  *      mark_inode_dirty(inode)
4382  *      stuff();
4383  *      inode->i_size = expr;
4384  *
4385  * is in error because a kswapd-driven write_inode() could occur while
4386  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
4387  * will no longer be on the superblock's dirty inode list.
4388  */
4389 int ext4_write_inode(struct inode *inode, int wait)
4390 {
4391         if (current->flags & PF_MEMALLOC)
4392                 return 0;
4393
4394         if (ext4_journal_current_handle()) {
4395                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
4396                 dump_stack();
4397                 return -EIO;
4398         }
4399
4400         if (!wait)
4401                 return 0;
4402
4403         return ext4_force_commit(inode->i_sb);
4404 }
4405
4406 /*
4407  * ext4_setattr()
4408  *
4409  * Called from notify_change.
4410  *
4411  * We want to trap VFS attempts to truncate the file as soon as
4412  * possible.  In particular, we want to make sure that when the VFS
4413  * shrinks i_size, we put the inode on the orphan list and modify
4414  * i_disksize immediately, so that during the subsequent flushing of
4415  * dirty pages and freeing of disk blocks, we can guarantee that any
4416  * commit will leave the blocks being flushed in an unused state on
4417  * disk.  (On recovery, the inode will get truncated and the blocks will
4418  * be freed, so we have a strong guarantee that no future commit will
4419  * leave these blocks visible to the user.)
4420  *
4421  * Another thing we have to assure is that if we are in ordered mode
4422  * and inode is still attached to the committing transaction, we must
4423  * we start writeout of all the dirty pages which are being truncated.
4424  * This way we are sure that all the data written in the previous
4425  * transaction are already on disk (truncate waits for pages under
4426  * writeback).
4427  *
4428  * Called with inode->i_mutex down.
4429  */
4430 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
4431 {
4432         struct inode *inode = dentry->d_inode;
4433         int error, rc = 0;
4434         const unsigned int ia_valid = attr->ia_valid;
4435
4436         error = inode_change_ok(inode, attr);
4437         if (error)
4438                 return error;
4439
4440         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
4441                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
4442                 handle_t *handle;
4443
4444                 /* (user+group)*(old+new) structure, inode write (sb,
4445                  * inode block, ? - but truncate inode update has it) */
4446                 handle = ext4_journal_start(inode, 2*(EXT4_QUOTA_INIT_BLOCKS(inode->i_sb)+
4447                                         EXT4_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
4448                 if (IS_ERR(handle)) {
4449                         error = PTR_ERR(handle);
4450                         goto err_out;
4451                 }
4452                 error = DQUOT_TRANSFER(inode, attr) ? -EDQUOT : 0;
4453                 if (error) {
4454                         ext4_journal_stop(handle);
4455                         return error;
4456                 }
4457                 /* Update corresponding info in inode so that everything is in
4458                  * one transaction */
4459                 if (attr->ia_valid & ATTR_UID)
4460                         inode->i_uid = attr->ia_uid;
4461                 if (attr->ia_valid & ATTR_GID)
4462                         inode->i_gid = attr->ia_gid;
4463                 error = ext4_mark_inode_dirty(handle, inode);
4464                 ext4_journal_stop(handle);
4465         }
4466
4467         if (attr->ia_valid & ATTR_SIZE) {
4468                 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)) {
4469                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4470
4471                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
4472                                 error = -EFBIG;
4473                                 goto err_out;
4474                         }
4475                 }
4476         }
4477
4478         if (S_ISREG(inode->i_mode) &&
4479             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
4480                 handle_t *handle;
4481
4482                 handle = ext4_journal_start(inode, 3);
4483                 if (IS_ERR(handle)) {
4484                         error = PTR_ERR(handle);
4485                         goto err_out;
4486                 }
4487
4488                 error = ext4_orphan_add(handle, inode);
4489                 EXT4_I(inode)->i_disksize = attr->ia_size;
4490                 rc = ext4_mark_inode_dirty(handle, inode);
4491                 if (!error)
4492                         error = rc;
4493                 ext4_journal_stop(handle);
4494
4495                 if (ext4_should_order_data(inode)) {
4496                         error = ext4_begin_ordered_truncate(inode,
4497                                                             attr->ia_size);
4498                         if (error) {
4499                                 /* Do as much error cleanup as possible */
4500                                 handle = ext4_journal_start(inode, 3);
4501                                 if (IS_ERR(handle)) {
4502                                         ext4_orphan_del(NULL, inode);
4503                                         goto err_out;
4504                                 }
4505                                 ext4_orphan_del(handle, inode);
4506                                 ext4_journal_stop(handle);
4507                                 goto err_out;
4508                         }
4509                 }
4510         }
4511
4512         rc = inode_setattr(inode, attr);
4513
4514         /* If inode_setattr's call to ext4_truncate failed to get a
4515          * transaction handle at all, we need to clean up the in-core
4516          * orphan list manually. */
4517         if (inode->i_nlink)
4518                 ext4_orphan_del(NULL, inode);
4519
4520         if (!rc && (ia_valid & ATTR_MODE))
4521                 rc = ext4_acl_chmod(inode);
4522
4523 err_out:
4524         ext4_std_error(inode->i_sb, error);
4525         if (!error)
4526                 error = rc;
4527         return error;
4528 }
4529
4530 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
4531                  struct kstat *stat)
4532 {
4533         struct inode *inode;
4534         unsigned long delalloc_blocks;
4535
4536         inode = dentry->d_inode;
4537         generic_fillattr(inode, stat);
4538
4539         /*
4540          * We can't update i_blocks if the block allocation is delayed
4541          * otherwise in the case of system crash before the real block
4542          * allocation is done, we will have i_blocks inconsistent with
4543          * on-disk file blocks.
4544          * We always keep i_blocks updated together with real
4545          * allocation. But to not confuse with user, stat
4546          * will return the blocks that include the delayed allocation
4547          * blocks for this file.
4548          */
4549         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
4550         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
4551         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
4552
4553         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
4554         return 0;
4555 }
4556
4557 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
4558                                       int chunk)
4559 {
4560         int indirects;
4561
4562         /* if nrblocks are contiguous */
4563         if (chunk) {
4564                 /*
4565                  * With N contiguous data blocks, it need at most
4566                  * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) indirect blocks
4567                  * 2 dindirect blocks
4568                  * 1 tindirect block
4569                  */
4570                 indirects = nrblocks / EXT4_ADDR_PER_BLOCK(inode->i_sb);
4571                 return indirects + 3;
4572         }
4573         /*
4574          * if nrblocks are not contiguous, worse case, each block touch
4575          * a indirect block, and each indirect block touch a double indirect
4576          * block, plus a triple indirect block
4577          */
4578         indirects = nrblocks * 2 + 1;
4579         return indirects;
4580 }
4581
4582 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4583 {
4584         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL))
4585                 return ext4_indirect_trans_blocks(inode, nrblocks, chunk);
4586         return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
4587 }
4588
4589 /*
4590  * Account for index blocks, block groups bitmaps and block group
4591  * descriptor blocks if modify datablocks and index blocks
4592  * worse case, the indexs blocks spread over different block groups
4593  *
4594  * If datablocks are discontiguous, they are possible to spread over
4595  * different block groups too. If they are contiugous, with flexbg,
4596  * they could still across block group boundary.
4597  *
4598  * Also account for superblock, inode, quota and xattr blocks
4599  */
4600 int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4601 {
4602         int groups, gdpblocks;
4603         int idxblocks;
4604         int ret = 0;
4605
4606         /*
4607          * How many index blocks need to touch to modify nrblocks?
4608          * The "Chunk" flag indicating whether the nrblocks is
4609          * physically contiguous on disk
4610          *
4611          * For Direct IO and fallocate, they calls get_block to allocate
4612          * one single extent at a time, so they could set the "Chunk" flag
4613          */
4614         idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
4615
4616         ret = idxblocks;
4617
4618         /*
4619          * Now let's see how many group bitmaps and group descriptors need
4620          * to account
4621          */
4622         groups = idxblocks;
4623         if (chunk)
4624                 groups += 1;
4625         else
4626                 groups += nrblocks;
4627
4628         gdpblocks = groups;
4629         if (groups > EXT4_SB(inode->i_sb)->s_groups_count)
4630                 groups = EXT4_SB(inode->i_sb)->s_groups_count;
4631         if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
4632                 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
4633
4634         /* bitmaps and block group descriptor blocks */
4635         ret += groups + gdpblocks;
4636
4637         /* Blocks for super block, inode, quota and xattr blocks */
4638         ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
4639
4640         return ret;
4641 }
4642
4643 /*
4644  * Calulate the total number of credits to reserve to fit
4645  * the modification of a single pages into a single transaction,
4646  * which may include multiple chunks of block allocations.
4647  *
4648  * This could be called via ext4_write_begin()
4649  *
4650  * We need to consider the worse case, when
4651  * one new block per extent.
4652  */
4653 int ext4_writepage_trans_blocks(struct inode *inode)
4654 {
4655         int bpp = ext4_journal_blocks_per_page(inode);
4656         int ret;
4657
4658         ret = ext4_meta_trans_blocks(inode, bpp, 0);
4659
4660         /* Account for data blocks for journalled mode */
4661         if (ext4_should_journal_data(inode))
4662                 ret += bpp;
4663         return ret;
4664 }
4665
4666 /*
4667  * Calculate the journal credits for a chunk of data modification.
4668  *
4669  * This is called from DIO, fallocate or whoever calling
4670  * ext4_get_blocks_wrap() to map/allocate a chunk of contigous disk blocks.
4671  *
4672  * journal buffers for data blocks are not included here, as DIO
4673  * and fallocate do no need to journal data buffers.
4674  */
4675 int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
4676 {
4677         return ext4_meta_trans_blocks(inode, nrblocks, 1);
4678 }
4679
4680 /*
4681  * The caller must have previously called ext4_reserve_inode_write().
4682  * Give this, we know that the caller already has write access to iloc->bh.
4683  */
4684 int ext4_mark_iloc_dirty(handle_t *handle,
4685                 struct inode *inode, struct ext4_iloc *iloc)
4686 {
4687         int err = 0;
4688
4689         if (test_opt(inode->i_sb, I_VERSION))
4690                 inode_inc_iversion(inode);
4691
4692         /* the do_update_inode consumes one bh->b_count */
4693         get_bh(iloc->bh);
4694
4695         /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
4696         err = ext4_do_update_inode(handle, inode, iloc);
4697         put_bh(iloc->bh);
4698         return err;
4699 }
4700
4701 /*
4702  * On success, We end up with an outstanding reference count against
4703  * iloc->bh.  This _must_ be cleaned up later.
4704  */
4705
4706 int
4707 ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
4708                          struct ext4_iloc *iloc)
4709 {
4710         int err = 0;
4711         if (handle) {
4712                 err = ext4_get_inode_loc(inode, iloc);
4713                 if (!err) {
4714                         BUFFER_TRACE(iloc->bh, "get_write_access");
4715                         err = ext4_journal_get_write_access(handle, iloc->bh);
4716                         if (err) {
4717                                 brelse(iloc->bh);
4718                                 iloc->bh = NULL;
4719                         }
4720                 }
4721         }
4722         ext4_std_error(inode->i_sb, err);
4723         return err;
4724 }
4725
4726 /*
4727  * Expand an inode by new_extra_isize bytes.
4728  * Returns 0 on success or negative error number on failure.
4729  */
4730 static int ext4_expand_extra_isize(struct inode *inode,
4731                                    unsigned int new_extra_isize,
4732                                    struct ext4_iloc iloc,
4733                                    handle_t *handle)
4734 {
4735         struct ext4_inode *raw_inode;
4736         struct ext4_xattr_ibody_header *header;
4737         struct ext4_xattr_entry *entry;
4738
4739         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
4740                 return 0;
4741
4742         raw_inode = ext4_raw_inode(&iloc);
4743
4744         header = IHDR(inode, raw_inode);
4745         entry = IFIRST(header);
4746
4747         /* No extended attributes present */
4748         if (!(EXT4_I(inode)->i_state & EXT4_STATE_XATTR) ||
4749                 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
4750                 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
4751                         new_extra_isize);
4752                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
4753                 return 0;
4754         }
4755
4756         /* try to expand with EAs present */
4757         return ext4_expand_extra_isize_ea(inode, new_extra_isize,
4758                                           raw_inode, handle);
4759 }
4760
4761 /*
4762  * What we do here is to mark the in-core inode as clean with respect to inode
4763  * dirtiness (it may still be data-dirty).
4764  * This means that the in-core inode may be reaped by prune_icache
4765  * without having to perform any I/O.  This is a very good thing,
4766  * because *any* task may call prune_icache - even ones which
4767  * have a transaction open against a different journal.
4768  *
4769  * Is this cheating?  Not really.  Sure, we haven't written the
4770  * inode out, but prune_icache isn't a user-visible syncing function.
4771  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
4772  * we start and wait on commits.
4773  *
4774  * Is this efficient/effective?  Well, we're being nice to the system
4775  * by cleaning up our inodes proactively so they can be reaped
4776  * without I/O.  But we are potentially leaving up to five seconds'
4777  * worth of inodes floating about which prune_icache wants us to
4778  * write out.  One way to fix that would be to get prune_icache()
4779  * to do a write_super() to free up some memory.  It has the desired
4780  * effect.
4781  */
4782 int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
4783 {
4784         struct ext4_iloc iloc;
4785         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4786         static unsigned int mnt_count;
4787         int err, ret;
4788
4789         might_sleep();
4790         err = ext4_reserve_inode_write(handle, inode, &iloc);
4791         if (EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
4792             !(EXT4_I(inode)->i_state & EXT4_STATE_NO_EXPAND)) {
4793                 /*
4794                  * We need extra buffer credits since we may write into EA block
4795                  * with this same handle. If journal_extend fails, then it will
4796                  * only result in a minor loss of functionality for that inode.
4797                  * If this is felt to be critical, then e2fsck should be run to
4798                  * force a large enough s_min_extra_isize.
4799                  */
4800                 if ((jbd2_journal_extend(handle,
4801                              EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
4802                         ret = ext4_expand_extra_isize(inode,
4803                                                       sbi->s_want_extra_isize,
4804                                                       iloc, handle);
4805                         if (ret) {
4806                                 EXT4_I(inode)->i_state |= EXT4_STATE_NO_EXPAND;
4807                                 if (mnt_count !=
4808                                         le16_to_cpu(sbi->s_es->s_mnt_count)) {
4809                                         ext4_warning(inode->i_sb, __func__,
4810                                         "Unable to expand inode %lu. Delete"
4811                                         " some EAs or run e2fsck.",
4812                                         inode->i_ino);
4813                                         mnt_count =
4814                                           le16_to_cpu(sbi->s_es->s_mnt_count);
4815                                 }
4816                         }
4817                 }
4818         }
4819         if (!err)
4820                 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
4821         return err;
4822 }
4823
4824 /*
4825  * ext4_dirty_inode() is called from __mark_inode_dirty()
4826  *
4827  * We're really interested in the case where a file is being extended.
4828  * i_size has been changed by generic_commit_write() and we thus need
4829  * to include the updated inode in the current transaction.
4830  *
4831  * Also, DQUOT_ALLOC_SPACE() will always dirty the inode when blocks
4832  * are allocated to the file.
4833  *
4834  * If the inode is marked synchronous, we don't honour that here - doing
4835  * so would cause a commit on atime updates, which we don't bother doing.
4836  * We handle synchronous inodes at the highest possible level.
4837  */
4838 void ext4_dirty_inode(struct inode *inode)
4839 {
4840         handle_t *current_handle = ext4_journal_current_handle();
4841         handle_t *handle;
4842
4843         handle = ext4_journal_start(inode, 2);
4844         if (IS_ERR(handle))
4845                 goto out;
4846         if (current_handle &&
4847                 current_handle->h_transaction != handle->h_transaction) {
4848                 /* This task has a transaction open against a different fs */
4849                 printk(KERN_EMERG "%s: transactions do not match!\n",
4850                        __func__);
4851         } else {
4852                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
4853                                 current_handle);
4854                 ext4_mark_inode_dirty(handle, inode);
4855         }
4856         ext4_journal_stop(handle);
4857 out:
4858         return;
4859 }
4860
4861 #if 0
4862 /*
4863  * Bind an inode's backing buffer_head into this transaction, to prevent
4864  * it from being flushed to disk early.  Unlike
4865  * ext4_reserve_inode_write, this leaves behind no bh reference and
4866  * returns no iloc structure, so the caller needs to repeat the iloc
4867  * lookup to mark the inode dirty later.
4868  */
4869 static int ext4_pin_inode(handle_t *handle, struct inode *inode)
4870 {
4871         struct ext4_iloc iloc;
4872
4873         int err = 0;
4874         if (handle) {
4875                 err = ext4_get_inode_loc(inode, &iloc);
4876                 if (!err) {
4877                         BUFFER_TRACE(iloc.bh, "get_write_access");
4878                         err = jbd2_journal_get_write_access(handle, iloc.bh);
4879                         if (!err)
4880                                 err = ext4_journal_dirty_metadata(handle,
4881                                                                   iloc.bh);
4882                         brelse(iloc.bh);
4883                 }
4884         }
4885         ext4_std_error(inode->i_sb, err);
4886         return err;
4887 }
4888 #endif
4889
4890 int ext4_change_inode_journal_flag(struct inode *inode, int val)
4891 {
4892         journal_t *journal;
4893         handle_t *handle;
4894         int err;
4895
4896         /*
4897          * We have to be very careful here: changing a data block's
4898          * journaling status dynamically is dangerous.  If we write a
4899          * data block to the journal, change the status and then delete
4900          * that block, we risk forgetting to revoke the old log record
4901          * from the journal and so a subsequent replay can corrupt data.
4902          * So, first we make sure that the journal is empty and that
4903          * nobody is changing anything.
4904          */
4905
4906         journal = EXT4_JOURNAL(inode);
4907         if (is_journal_aborted(journal))
4908                 return -EROFS;
4909
4910         jbd2_journal_lock_updates(journal);
4911         jbd2_journal_flush(journal);
4912
4913         /*
4914          * OK, there are no updates running now, and all cached data is
4915          * synced to disk.  We are now in a completely consistent state
4916          * which doesn't have anything in the journal, and we know that
4917          * no filesystem updates are running, so it is safe to modify
4918          * the inode's in-core data-journaling state flag now.
4919          */
4920
4921         if (val)
4922                 EXT4_I(inode)->i_flags |= EXT4_JOURNAL_DATA_FL;
4923         else
4924                 EXT4_I(inode)->i_flags &= ~EXT4_JOURNAL_DATA_FL;
4925         ext4_set_aops(inode);
4926
4927         jbd2_journal_unlock_updates(journal);
4928
4929         /* Finally we can mark the inode as dirty. */
4930
4931         handle = ext4_journal_start(inode, 1);
4932         if (IS_ERR(handle))
4933                 return PTR_ERR(handle);
4934
4935         err = ext4_mark_inode_dirty(handle, inode);
4936         handle->h_sync = 1;
4937         ext4_journal_stop(handle);
4938         ext4_std_error(inode->i_sb, err);
4939
4940         return err;
4941 }
4942
4943 static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
4944 {
4945         return !buffer_mapped(bh);
4946 }
4947
4948 int ext4_page_mkwrite(struct vm_area_struct *vma, struct page *page)
4949 {
4950         loff_t size;
4951         unsigned long len;
4952         int ret = -EINVAL;
4953         void *fsdata;
4954         struct file *file = vma->vm_file;
4955         struct inode *inode = file->f_path.dentry->d_inode;
4956         struct address_space *mapping = inode->i_mapping;
4957
4958         /*
4959          * Get i_alloc_sem to stop truncates messing with the inode. We cannot
4960          * get i_mutex because we are already holding mmap_sem.
4961          */
4962         down_read(&inode->i_alloc_sem);
4963         size = i_size_read(inode);
4964         if (page->mapping != mapping || size <= page_offset(page)
4965             || !PageUptodate(page)) {
4966                 /* page got truncated from under us? */
4967                 goto out_unlock;
4968         }
4969         ret = 0;
4970         if (PageMappedToDisk(page))
4971                 goto out_unlock;
4972
4973         if (page->index == size >> PAGE_CACHE_SHIFT)
4974                 len = size & ~PAGE_CACHE_MASK;
4975         else
4976                 len = PAGE_CACHE_SIZE;
4977
4978         if (page_has_buffers(page)) {
4979                 /* return if we have all the buffers mapped */
4980                 if (!walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
4981                                        ext4_bh_unmapped))
4982                         goto out_unlock;
4983         }
4984         /*
4985          * OK, we need to fill the hole... Do write_begin write_end
4986          * to do block allocation/reservation.We are not holding
4987          * inode.i__mutex here. That allow * parallel write_begin,
4988          * write_end call. lock_page prevent this from happening
4989          * on the same page though
4990          */
4991         ret = mapping->a_ops->write_begin(file, mapping, page_offset(page),
4992                         len, AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
4993         if (ret < 0)
4994                 goto out_unlock;
4995         ret = mapping->a_ops->write_end(file, mapping, page_offset(page),
4996                         len, len, page, fsdata);
4997         if (ret < 0)
4998                 goto out_unlock;
4999         ret = 0;
5000 out_unlock:
5001         up_read(&inode->i_alloc_sem);
5002         return ret;
5003 }