Merge branch 'linus' into oprofile-v2
[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
1652         BUG_ON(mpd->next_page <= mpd->first_page);
1653         pagevec_init(&pvec, 0);
1654         index = mpd->first_page;
1655         end = mpd->next_page - 1;
1656
1657         while (index <= end) {
1658                 /* XXX: optimize tail */
1659                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1660                 if (nr_pages == 0)
1661                         break;
1662                 for (i = 0; i < nr_pages; i++) {
1663                         struct page *page = pvec.pages[i];
1664
1665                         index = page->index;
1666                         if (index > end)
1667                                 break;
1668                         index++;
1669
1670                         err = mapping->a_ops->writepage(page, mpd->wbc);
1671                         if (!err)
1672                                 mpd->pages_written++;
1673                         /*
1674                          * In error case, we have to continue because
1675                          * remaining pages are still locked
1676                          * XXX: unlock and re-dirty them?
1677                          */
1678                         if (ret == 0)
1679                                 ret = err;
1680                 }
1681                 pagevec_release(&pvec);
1682         }
1683         return ret;
1684 }
1685
1686 /*
1687  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
1688  *
1689  * @mpd->inode - inode to walk through
1690  * @exbh->b_blocknr - first block on a disk
1691  * @exbh->b_size - amount of space in bytes
1692  * @logical - first logical block to start assignment with
1693  *
1694  * the function goes through all passed space and put actual disk
1695  * block numbers into buffer heads, dropping BH_Delay
1696  */
1697 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
1698                                  struct buffer_head *exbh)
1699 {
1700         struct inode *inode = mpd->inode;
1701         struct address_space *mapping = inode->i_mapping;
1702         int blocks = exbh->b_size >> inode->i_blkbits;
1703         sector_t pblock = exbh->b_blocknr, cur_logical;
1704         struct buffer_head *head, *bh;
1705         pgoff_t index, end;
1706         struct pagevec pvec;
1707         int nr_pages, i;
1708
1709         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1710         end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1711         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1712
1713         pagevec_init(&pvec, 0);
1714
1715         while (index <= end) {
1716                 /* XXX: optimize tail */
1717                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1718                 if (nr_pages == 0)
1719                         break;
1720                 for (i = 0; i < nr_pages; i++) {
1721                         struct page *page = pvec.pages[i];
1722
1723                         index = page->index;
1724                         if (index > end)
1725                                 break;
1726                         index++;
1727
1728                         BUG_ON(!PageLocked(page));
1729                         BUG_ON(PageWriteback(page));
1730                         BUG_ON(!page_has_buffers(page));
1731
1732                         bh = page_buffers(page);
1733                         head = bh;
1734
1735                         /* skip blocks out of the range */
1736                         do {
1737                                 if (cur_logical >= logical)
1738                                         break;
1739                                 cur_logical++;
1740                         } while ((bh = bh->b_this_page) != head);
1741
1742                         do {
1743                                 if (cur_logical >= logical + blocks)
1744                                         break;
1745                                 if (buffer_delay(bh)) {
1746                                         bh->b_blocknr = pblock;
1747                                         clear_buffer_delay(bh);
1748                                         bh->b_bdev = inode->i_sb->s_bdev;
1749                                 } else if (buffer_unwritten(bh)) {
1750                                         bh->b_blocknr = pblock;
1751                                         clear_buffer_unwritten(bh);
1752                                         set_buffer_mapped(bh);
1753                                         set_buffer_new(bh);
1754                                         bh->b_bdev = inode->i_sb->s_bdev;
1755                                 } else if (buffer_mapped(bh))
1756                                         BUG_ON(bh->b_blocknr != pblock);
1757
1758                                 cur_logical++;
1759                                 pblock++;
1760                         } while ((bh = bh->b_this_page) != head);
1761                 }
1762                 pagevec_release(&pvec);
1763         }
1764 }
1765
1766
1767 /*
1768  * __unmap_underlying_blocks - just a helper function to unmap
1769  * set of blocks described by @bh
1770  */
1771 static inline void __unmap_underlying_blocks(struct inode *inode,
1772                                              struct buffer_head *bh)
1773 {
1774         struct block_device *bdev = inode->i_sb->s_bdev;
1775         int blocks, i;
1776
1777         blocks = bh->b_size >> inode->i_blkbits;
1778         for (i = 0; i < blocks; i++)
1779                 unmap_underlying_metadata(bdev, bh->b_blocknr + i);
1780 }
1781
1782 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
1783                                         sector_t logical, long blk_cnt)
1784 {
1785         int nr_pages, i;
1786         pgoff_t index, end;
1787         struct pagevec pvec;
1788         struct inode *inode = mpd->inode;
1789         struct address_space *mapping = inode->i_mapping;
1790
1791         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1792         end   = (logical + blk_cnt - 1) >>
1793                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
1794         while (index <= end) {
1795                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1796                 if (nr_pages == 0)
1797                         break;
1798                 for (i = 0; i < nr_pages; i++) {
1799                         struct page *page = pvec.pages[i];
1800                         index = page->index;
1801                         if (index > end)
1802                                 break;
1803                         index++;
1804
1805                         BUG_ON(!PageLocked(page));
1806                         BUG_ON(PageWriteback(page));
1807                         block_invalidatepage(page, 0);
1808                         ClearPageUptodate(page);
1809                         unlock_page(page);
1810                 }
1811         }
1812         return;
1813 }
1814
1815 static void ext4_print_free_blocks(struct inode *inode)
1816 {
1817         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1818         printk(KERN_EMERG "Total free blocks count %lld\n",
1819                         ext4_count_free_blocks(inode->i_sb));
1820         printk(KERN_EMERG "Free/Dirty block details\n");
1821         printk(KERN_EMERG "free_blocks=%lld\n",
1822                         percpu_counter_sum(&sbi->s_freeblocks_counter));
1823         printk(KERN_EMERG "dirty_blocks=%lld\n",
1824                         percpu_counter_sum(&sbi->s_dirtyblocks_counter));
1825         printk(KERN_EMERG "Block reservation details\n");
1826         printk(KERN_EMERG "i_reserved_data_blocks=%lu\n",
1827                         EXT4_I(inode)->i_reserved_data_blocks);
1828         printk(KERN_EMERG "i_reserved_meta_blocks=%lu\n",
1829                         EXT4_I(inode)->i_reserved_meta_blocks);
1830         return;
1831 }
1832
1833 /*
1834  * mpage_da_map_blocks - go through given space
1835  *
1836  * @mpd->lbh - bh describing space
1837  * @mpd->get_block - the filesystem's block mapper function
1838  *
1839  * The function skips space we know is already mapped to disk blocks.
1840  *
1841  */
1842 static int  mpage_da_map_blocks(struct mpage_da_data *mpd)
1843 {
1844         int err = 0;
1845         struct buffer_head new;
1846         struct buffer_head *lbh = &mpd->lbh;
1847         sector_t next;
1848
1849         /*
1850          * We consider only non-mapped and non-allocated blocks
1851          */
1852         if (buffer_mapped(lbh) && !buffer_delay(lbh))
1853                 return 0;
1854         new.b_state = lbh->b_state;
1855         new.b_blocknr = 0;
1856         new.b_size = lbh->b_size;
1857         next = lbh->b_blocknr;
1858         /*
1859          * If we didn't accumulate anything
1860          * to write simply return
1861          */
1862         if (!new.b_size)
1863                 return 0;
1864         err = mpd->get_block(mpd->inode, next, &new, 1);
1865         if (err) {
1866
1867                 /* If get block returns with error
1868                  * we simply return. Later writepage
1869                  * will redirty the page and writepages
1870                  * will find the dirty page again
1871                  */
1872                 if (err == -EAGAIN)
1873                         return 0;
1874
1875                 if (err == -ENOSPC &&
1876                                 ext4_count_free_blocks(mpd->inode->i_sb)) {
1877                         mpd->retval = err;
1878                         return 0;
1879                 }
1880
1881                 /*
1882                  * get block failure will cause us
1883                  * to loop in writepages. Because
1884                  * a_ops->writepage won't be able to
1885                  * make progress. The page will be redirtied
1886                  * by writepage and writepages will again
1887                  * try to write the same.
1888                  */
1889                 printk(KERN_EMERG "%s block allocation failed for inode %lu "
1890                                   "at logical offset %llu with max blocks "
1891                                   "%zd with error %d\n",
1892                                   __func__, mpd->inode->i_ino,
1893                                   (unsigned long long)next,
1894                                   lbh->b_size >> mpd->inode->i_blkbits, err);
1895                 printk(KERN_EMERG "This should not happen.!! "
1896                                         "Data will be lost\n");
1897                 if (err == -ENOSPC) {
1898                         ext4_print_free_blocks(mpd->inode);
1899                 }
1900                 /* invlaidate all the pages */
1901                 ext4_da_block_invalidatepages(mpd, next,
1902                                 lbh->b_size >> mpd->inode->i_blkbits);
1903                 return err;
1904         }
1905         BUG_ON(new.b_size == 0);
1906
1907         if (buffer_new(&new))
1908                 __unmap_underlying_blocks(mpd->inode, &new);
1909
1910         /*
1911          * If blocks are delayed marked, we need to
1912          * put actual blocknr and drop delayed bit
1913          */
1914         if (buffer_delay(lbh) || buffer_unwritten(lbh))
1915                 mpage_put_bnr_to_bhs(mpd, next, &new);
1916
1917         return 0;
1918 }
1919
1920 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
1921                 (1 << BH_Delay) | (1 << BH_Unwritten))
1922
1923 /*
1924  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
1925  *
1926  * @mpd->lbh - extent of blocks
1927  * @logical - logical number of the block in the file
1928  * @bh - bh of the block (used to access block's state)
1929  *
1930  * the function is used to collect contig. blocks in same state
1931  */
1932 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
1933                                    sector_t logical, struct buffer_head *bh)
1934 {
1935         sector_t next;
1936         size_t b_size = bh->b_size;
1937         struct buffer_head *lbh = &mpd->lbh;
1938         int nrblocks = lbh->b_size >> mpd->inode->i_blkbits;
1939
1940         /* check if thereserved journal credits might overflow */
1941         if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
1942                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
1943                         /*
1944                          * With non-extent format we are limited by the journal
1945                          * credit available.  Total credit needed to insert
1946                          * nrblocks contiguous blocks is dependent on the
1947                          * nrblocks.  So limit nrblocks.
1948                          */
1949                         goto flush_it;
1950                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
1951                                 EXT4_MAX_TRANS_DATA) {
1952                         /*
1953                          * Adding the new buffer_head would make it cross the
1954                          * allowed limit for which we have journal credit
1955                          * reserved. So limit the new bh->b_size
1956                          */
1957                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
1958                                                 mpd->inode->i_blkbits;
1959                         /* we will do mpage_da_submit_io in the next loop */
1960                 }
1961         }
1962         /*
1963          * First block in the extent
1964          */
1965         if (lbh->b_size == 0) {
1966                 lbh->b_blocknr = logical;
1967                 lbh->b_size = b_size;
1968                 lbh->b_state = bh->b_state & BH_FLAGS;
1969                 return;
1970         }
1971
1972         next = lbh->b_blocknr + nrblocks;
1973         /*
1974          * Can we merge the block to our big extent?
1975          */
1976         if (logical == next && (bh->b_state & BH_FLAGS) == lbh->b_state) {
1977                 lbh->b_size += b_size;
1978                 return;
1979         }
1980
1981 flush_it:
1982         /*
1983          * We couldn't merge the block to our extent, so we
1984          * need to flush current  extent and start new one
1985          */
1986         if (mpage_da_map_blocks(mpd) == 0)
1987                 mpage_da_submit_io(mpd);
1988         mpd->io_done = 1;
1989         return;
1990 }
1991
1992 /*
1993  * __mpage_da_writepage - finds extent of pages and blocks
1994  *
1995  * @page: page to consider
1996  * @wbc: not used, we just follow rules
1997  * @data: context
1998  *
1999  * The function finds extents of pages and scan them for all blocks.
2000  */
2001 static int __mpage_da_writepage(struct page *page,
2002                                 struct writeback_control *wbc, void *data)
2003 {
2004         struct mpage_da_data *mpd = data;
2005         struct inode *inode = mpd->inode;
2006         struct buffer_head *bh, *head, fake;
2007         sector_t logical;
2008
2009         if (mpd->io_done) {
2010                 /*
2011                  * Rest of the page in the page_vec
2012                  * redirty then and skip then. We will
2013                  * try to to write them again after
2014                  * starting a new transaction
2015                  */
2016                 redirty_page_for_writepage(wbc, page);
2017                 unlock_page(page);
2018                 return MPAGE_DA_EXTENT_TAIL;
2019         }
2020         /*
2021          * Can we merge this page to current extent?
2022          */
2023         if (mpd->next_page != page->index) {
2024                 /*
2025                  * Nope, we can't. So, we map non-allocated blocks
2026                  * and start IO on them using writepage()
2027                  */
2028                 if (mpd->next_page != mpd->first_page) {
2029                         if (mpage_da_map_blocks(mpd) == 0)
2030                                 mpage_da_submit_io(mpd);
2031                         /*
2032                          * skip rest of the page in the page_vec
2033                          */
2034                         mpd->io_done = 1;
2035                         redirty_page_for_writepage(wbc, page);
2036                         unlock_page(page);
2037                         return MPAGE_DA_EXTENT_TAIL;
2038                 }
2039
2040                 /*
2041                  * Start next extent of pages ...
2042                  */
2043                 mpd->first_page = page->index;
2044
2045                 /*
2046                  * ... and blocks
2047                  */
2048                 mpd->lbh.b_size = 0;
2049                 mpd->lbh.b_state = 0;
2050                 mpd->lbh.b_blocknr = 0;
2051         }
2052
2053         mpd->next_page = page->index + 1;
2054         logical = (sector_t) page->index <<
2055                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2056
2057         if (!page_has_buffers(page)) {
2058                 /*
2059                  * There is no attached buffer heads yet (mmap?)
2060                  * we treat the page asfull of dirty blocks
2061                  */
2062                 bh = &fake;
2063                 bh->b_size = PAGE_CACHE_SIZE;
2064                 bh->b_state = 0;
2065                 set_buffer_dirty(bh);
2066                 set_buffer_uptodate(bh);
2067                 mpage_add_bh_to_extent(mpd, logical, bh);
2068                 if (mpd->io_done)
2069                         return MPAGE_DA_EXTENT_TAIL;
2070         } else {
2071                 /*
2072                  * Page with regular buffer heads, just add all dirty ones
2073                  */
2074                 head = page_buffers(page);
2075                 bh = head;
2076                 do {
2077                         BUG_ON(buffer_locked(bh));
2078                         if (buffer_dirty(bh) &&
2079                                 (!buffer_mapped(bh) || buffer_delay(bh))) {
2080                                 mpage_add_bh_to_extent(mpd, logical, bh);
2081                                 if (mpd->io_done)
2082                                         return MPAGE_DA_EXTENT_TAIL;
2083                         }
2084                         logical++;
2085                 } while ((bh = bh->b_this_page) != head);
2086         }
2087
2088         return 0;
2089 }
2090
2091 /*
2092  * mpage_da_writepages - walk the list of dirty pages of the given
2093  * address space, allocates non-allocated blocks, maps newly-allocated
2094  * blocks to existing bhs and issue IO them
2095  *
2096  * @mapping: address space structure to write
2097  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2098  * @get_block: the filesystem's block mapper function.
2099  *
2100  * This is a library function, which implements the writepages()
2101  * address_space_operation.
2102  */
2103 static int mpage_da_writepages(struct address_space *mapping,
2104                                struct writeback_control *wbc,
2105                                struct mpage_da_data *mpd)
2106 {
2107         long to_write;
2108         int ret;
2109
2110         if (!mpd->get_block)
2111                 return generic_writepages(mapping, wbc);
2112
2113         mpd->lbh.b_size = 0;
2114         mpd->lbh.b_state = 0;
2115         mpd->lbh.b_blocknr = 0;
2116         mpd->first_page = 0;
2117         mpd->next_page = 0;
2118         mpd->io_done = 0;
2119         mpd->pages_written = 0;
2120         mpd->retval = 0;
2121
2122         to_write = wbc->nr_to_write;
2123
2124         ret = write_cache_pages(mapping, wbc, __mpage_da_writepage, mpd);
2125
2126         /*
2127          * Handle last extent of pages
2128          */
2129         if (!mpd->io_done && mpd->next_page != mpd->first_page) {
2130                 if (mpage_da_map_blocks(mpd) == 0)
2131                         mpage_da_submit_io(mpd);
2132         }
2133
2134         wbc->nr_to_write = to_write - mpd->pages_written;
2135         return ret;
2136 }
2137
2138 /*
2139  * this is a special callback for ->write_begin() only
2140  * it's intention is to return mapped block or reserve space
2141  */
2142 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2143                                   struct buffer_head *bh_result, int create)
2144 {
2145         int ret = 0;
2146
2147         BUG_ON(create == 0);
2148         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2149
2150         /*
2151          * first, we need to know whether the block is allocated already
2152          * preallocated blocks are unmapped but should treated
2153          * the same as allocated blocks.
2154          */
2155         ret = ext4_get_blocks_wrap(NULL, inode, iblock, 1,  bh_result, 0, 0, 0);
2156         if ((ret == 0) && !buffer_delay(bh_result)) {
2157                 /* the block isn't (pre)allocated yet, let's reserve space */
2158                 /*
2159                  * XXX: __block_prepare_write() unmaps passed block,
2160                  * is it OK?
2161                  */
2162                 ret = ext4_da_reserve_space(inode, 1);
2163                 if (ret)
2164                         /* not enough space to reserve */
2165                         return ret;
2166
2167                 map_bh(bh_result, inode->i_sb, 0);
2168                 set_buffer_new(bh_result);
2169                 set_buffer_delay(bh_result);
2170         } else if (ret > 0) {
2171                 bh_result->b_size = (ret << inode->i_blkbits);
2172                 ret = 0;
2173         }
2174
2175         return ret;
2176 }
2177 #define         EXT4_DELALLOC_RSVED     1
2178 static int ext4_da_get_block_write(struct inode *inode, sector_t iblock,
2179                                    struct buffer_head *bh_result, int create)
2180 {
2181         int ret;
2182         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2183         loff_t disksize = EXT4_I(inode)->i_disksize;
2184         handle_t *handle = NULL;
2185
2186         handle = ext4_journal_current_handle();
2187         BUG_ON(!handle);
2188         ret = ext4_get_blocks_wrap(handle, inode, iblock, max_blocks,
2189                         bh_result, create, 0, EXT4_DELALLOC_RSVED);
2190         if (ret > 0) {
2191
2192                 bh_result->b_size = (ret << inode->i_blkbits);
2193
2194                 if (ext4_should_order_data(inode)) {
2195                         int retval;
2196                         retval = ext4_jbd2_file_inode(handle, inode);
2197                         if (retval)
2198                                 /*
2199                                  * Failed to add inode for ordered
2200                                  * mode. Don't update file size
2201                                  */
2202                                 return retval;
2203                 }
2204
2205                 /*
2206                  * Update on-disk size along with block allocation
2207                  * we don't use 'extend_disksize' as size may change
2208                  * within already allocated block -bzzz
2209                  */
2210                 disksize = ((loff_t) iblock + ret) << inode->i_blkbits;
2211                 if (disksize > i_size_read(inode))
2212                         disksize = i_size_read(inode);
2213                 if (disksize > EXT4_I(inode)->i_disksize) {
2214                         ext4_update_i_disksize(inode, disksize);
2215                         ret = ext4_mark_inode_dirty(handle, inode);
2216                         return ret;
2217                 }
2218                 ret = 0;
2219         }
2220         return ret;
2221 }
2222
2223 static int ext4_bh_unmapped_or_delay(handle_t *handle, struct buffer_head *bh)
2224 {
2225         /*
2226          * unmapped buffer is possible for holes.
2227          * delay buffer is possible with delayed allocation
2228          */
2229         return ((!buffer_mapped(bh) || buffer_delay(bh)) && buffer_dirty(bh));
2230 }
2231
2232 static int ext4_normal_get_block_write(struct inode *inode, sector_t iblock,
2233                                    struct buffer_head *bh_result, int create)
2234 {
2235         int ret = 0;
2236         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2237
2238         /*
2239          * we don't want to do block allocation in writepage
2240          * so call get_block_wrap with create = 0
2241          */
2242         ret = ext4_get_blocks_wrap(NULL, inode, iblock, max_blocks,
2243                                    bh_result, 0, 0, 0);
2244         if (ret > 0) {
2245                 bh_result->b_size = (ret << inode->i_blkbits);
2246                 ret = 0;
2247         }
2248         return ret;
2249 }
2250
2251 /*
2252  * get called vi ext4_da_writepages after taking page lock (have journal handle)
2253  * get called via journal_submit_inode_data_buffers (no journal handle)
2254  * get called via shrink_page_list via pdflush (no journal handle)
2255  * or grab_page_cache when doing write_begin (have journal handle)
2256  */
2257 static int ext4_da_writepage(struct page *page,
2258                                 struct writeback_control *wbc)
2259 {
2260         int ret = 0;
2261         loff_t size;
2262         unsigned long len;
2263         struct buffer_head *page_bufs;
2264         struct inode *inode = page->mapping->host;
2265
2266         size = i_size_read(inode);
2267         if (page->index == size >> PAGE_CACHE_SHIFT)
2268                 len = size & ~PAGE_CACHE_MASK;
2269         else
2270                 len = PAGE_CACHE_SIZE;
2271
2272         if (page_has_buffers(page)) {
2273                 page_bufs = page_buffers(page);
2274                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2275                                         ext4_bh_unmapped_or_delay)) {
2276                         /*
2277                          * We don't want to do  block allocation
2278                          * So redirty the page and return
2279                          * We may reach here when we do a journal commit
2280                          * via journal_submit_inode_data_buffers.
2281                          * If we don't have mapping block we just ignore
2282                          * them. We can also reach here via shrink_page_list
2283                          */
2284                         redirty_page_for_writepage(wbc, page);
2285                         unlock_page(page);
2286                         return 0;
2287                 }
2288         } else {
2289                 /*
2290                  * The test for page_has_buffers() is subtle:
2291                  * We know the page is dirty but it lost buffers. That means
2292                  * that at some moment in time after write_begin()/write_end()
2293                  * has been called all buffers have been clean and thus they
2294                  * must have been written at least once. So they are all
2295                  * mapped and we can happily proceed with mapping them
2296                  * and writing the page.
2297                  *
2298                  * Try to initialize the buffer_heads and check whether
2299                  * all are mapped and non delay. We don't want to
2300                  * do block allocation here.
2301                  */
2302                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
2303                                                 ext4_normal_get_block_write);
2304                 if (!ret) {
2305                         page_bufs = page_buffers(page);
2306                         /* check whether all are mapped and non delay */
2307                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2308                                                 ext4_bh_unmapped_or_delay)) {
2309                                 redirty_page_for_writepage(wbc, page);
2310                                 unlock_page(page);
2311                                 return 0;
2312                         }
2313                 } else {
2314                         /*
2315                          * We can't do block allocation here
2316                          * so just redity the page and unlock
2317                          * and return
2318                          */
2319                         redirty_page_for_writepage(wbc, page);
2320                         unlock_page(page);
2321                         return 0;
2322                 }
2323         }
2324
2325         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2326                 ret = nobh_writepage(page, ext4_normal_get_block_write, wbc);
2327         else
2328                 ret = block_write_full_page(page,
2329                                                 ext4_normal_get_block_write,
2330                                                 wbc);
2331
2332         return ret;
2333 }
2334
2335 /*
2336  * This is called via ext4_da_writepages() to
2337  * calulate the total number of credits to reserve to fit
2338  * a single extent allocation into a single transaction,
2339  * ext4_da_writpeages() will loop calling this before
2340  * the block allocation.
2341  */
2342
2343 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2344 {
2345         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2346
2347         /*
2348          * With non-extent format the journal credit needed to
2349          * insert nrblocks contiguous block is dependent on
2350          * number of contiguous block. So we will limit
2351          * number of contiguous block to a sane value
2352          */
2353         if (!(inode->i_flags & EXT4_EXTENTS_FL) &&
2354             (max_blocks > EXT4_MAX_TRANS_DATA))
2355                 max_blocks = EXT4_MAX_TRANS_DATA;
2356
2357         return ext4_chunk_trans_blocks(inode, max_blocks);
2358 }
2359
2360 static int ext4_da_writepages(struct address_space *mapping,
2361                               struct writeback_control *wbc)
2362 {
2363         handle_t *handle = NULL;
2364         loff_t range_start = 0;
2365         struct mpage_da_data mpd;
2366         struct inode *inode = mapping->host;
2367         int needed_blocks, ret = 0, nr_to_writebump = 0;
2368         long to_write, pages_skipped = 0;
2369         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2370
2371         /*
2372          * No pages to write? This is mainly a kludge to avoid starting
2373          * a transaction for special inodes like journal inode on last iput()
2374          * because that could violate lock ordering on umount
2375          */
2376         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2377                 return 0;
2378         /*
2379          * Make sure nr_to_write is >= sbi->s_mb_stream_request
2380          * This make sure small files blocks are allocated in
2381          * single attempt. This ensure that small files
2382          * get less fragmented.
2383          */
2384         if (wbc->nr_to_write < sbi->s_mb_stream_request) {
2385                 nr_to_writebump = sbi->s_mb_stream_request - wbc->nr_to_write;
2386                 wbc->nr_to_write = sbi->s_mb_stream_request;
2387         }
2388
2389         if (!wbc->range_cyclic)
2390                 /*
2391                  * If range_cyclic is not set force range_cont
2392                  * and save the old writeback_index
2393                  */
2394                 wbc->range_cont = 1;
2395
2396         range_start =  wbc->range_start;
2397         pages_skipped = wbc->pages_skipped;
2398
2399         mpd.wbc = wbc;
2400         mpd.inode = mapping->host;
2401
2402 restart_loop:
2403         to_write = wbc->nr_to_write;
2404         while (!ret && to_write > 0) {
2405
2406                 /*
2407                  * we  insert one extent at a time. So we need
2408                  * credit needed for single extent allocation.
2409                  * journalled mode is currently not supported
2410                  * by delalloc
2411                  */
2412                 BUG_ON(ext4_should_journal_data(inode));
2413                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2414
2415                 /* start a new transaction*/
2416                 handle = ext4_journal_start(inode, needed_blocks);
2417                 if (IS_ERR(handle)) {
2418                         ret = PTR_ERR(handle);
2419                         printk(KERN_EMERG "%s: jbd2_start: "
2420                                "%ld pages, ino %lu; err %d\n", __func__,
2421                                 wbc->nr_to_write, inode->i_ino, ret);
2422                         dump_stack();
2423                         goto out_writepages;
2424                 }
2425                 to_write -= wbc->nr_to_write;
2426
2427                 mpd.get_block = ext4_da_get_block_write;
2428                 ret = mpage_da_writepages(mapping, wbc, &mpd);
2429
2430                 ext4_journal_stop(handle);
2431
2432                 if (mpd.retval == -ENOSPC)
2433                         jbd2_journal_force_commit_nested(sbi->s_journal);
2434
2435                 /* reset the retry count */
2436                 if (ret == MPAGE_DA_EXTENT_TAIL) {
2437                         /*
2438                          * got one extent now try with
2439                          * rest of the pages
2440                          */
2441                         to_write += wbc->nr_to_write;
2442                         ret = 0;
2443                 } else if (wbc->nr_to_write) {
2444                         /*
2445                          * There is no more writeout needed
2446                          * or we requested for a noblocking writeout
2447                          * and we found the device congested
2448                          */
2449                         to_write += wbc->nr_to_write;
2450                         break;
2451                 }
2452                 wbc->nr_to_write = to_write;
2453         }
2454
2455         if (wbc->range_cont && (pages_skipped != wbc->pages_skipped)) {
2456                 /* We skipped pages in this loop */
2457                 wbc->range_start = range_start;
2458                 wbc->nr_to_write = to_write +
2459                                 wbc->pages_skipped - pages_skipped;
2460                 wbc->pages_skipped = pages_skipped;
2461                 goto restart_loop;
2462         }
2463
2464 out_writepages:
2465         wbc->nr_to_write = to_write - nr_to_writebump;
2466         wbc->range_start = range_start;
2467         return ret;
2468 }
2469
2470 #define FALL_BACK_TO_NONDELALLOC 1
2471 static int ext4_nonda_switch(struct super_block *sb)
2472 {
2473         s64 free_blocks, dirty_blocks;
2474         struct ext4_sb_info *sbi = EXT4_SB(sb);
2475
2476         /*
2477          * switch to non delalloc mode if we are running low
2478          * on free block. The free block accounting via percpu
2479          * counters can get slightly wrong with FBC_BATCH getting
2480          * accumulated on each CPU without updating global counters
2481          * Delalloc need an accurate free block accounting. So switch
2482          * to non delalloc when we are near to error range.
2483          */
2484         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
2485         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
2486         if (2 * free_blocks < 3 * dirty_blocks ||
2487                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
2488                 /*
2489                  * free block count is less that 150% of dirty blocks
2490                  * or free blocks is less that watermark
2491                  */
2492                 return 1;
2493         }
2494         return 0;
2495 }
2496
2497 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2498                                 loff_t pos, unsigned len, unsigned flags,
2499                                 struct page **pagep, void **fsdata)
2500 {
2501         int ret, retries = 0;
2502         struct page *page;
2503         pgoff_t index;
2504         unsigned from, to;
2505         struct inode *inode = mapping->host;
2506         handle_t *handle;
2507
2508         index = pos >> PAGE_CACHE_SHIFT;
2509         from = pos & (PAGE_CACHE_SIZE - 1);
2510         to = from + len;
2511
2512         if (ext4_nonda_switch(inode->i_sb)) {
2513                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2514                 return ext4_write_begin(file, mapping, pos,
2515                                         len, flags, pagep, fsdata);
2516         }
2517         *fsdata = (void *)0;
2518 retry:
2519         /*
2520          * With delayed allocation, we don't log the i_disksize update
2521          * if there is delayed block allocation. But we still need
2522          * to journalling the i_disksize update if writes to the end
2523          * of file which has an already mapped buffer.
2524          */
2525         handle = ext4_journal_start(inode, 1);
2526         if (IS_ERR(handle)) {
2527                 ret = PTR_ERR(handle);
2528                 goto out;
2529         }
2530
2531         page = __grab_cache_page(mapping, index);
2532         if (!page) {
2533                 ext4_journal_stop(handle);
2534                 ret = -ENOMEM;
2535                 goto out;
2536         }
2537         *pagep = page;
2538
2539         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
2540                                                         ext4_da_get_block_prep);
2541         if (ret < 0) {
2542                 unlock_page(page);
2543                 ext4_journal_stop(handle);
2544                 page_cache_release(page);
2545                 /*
2546                  * block_write_begin may have instantiated a few blocks
2547                  * outside i_size.  Trim these off again. Don't need
2548                  * i_size_read because we hold i_mutex.
2549                  */
2550                 if (pos + len > inode->i_size)
2551                         vmtruncate(inode, inode->i_size);
2552         }
2553
2554         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
2555                 goto retry;
2556 out:
2557         return ret;
2558 }
2559
2560 /*
2561  * Check if we should update i_disksize
2562  * when write to the end of file but not require block allocation
2563  */
2564 static int ext4_da_should_update_i_disksize(struct page *page,
2565                                          unsigned long offset)
2566 {
2567         struct buffer_head *bh;
2568         struct inode *inode = page->mapping->host;
2569         unsigned int idx;
2570         int i;
2571
2572         bh = page_buffers(page);
2573         idx = offset >> inode->i_blkbits;
2574
2575         for (i = 0; i < idx; i++)
2576                 bh = bh->b_this_page;
2577
2578         if (!buffer_mapped(bh) || (buffer_delay(bh)))
2579                 return 0;
2580         return 1;
2581 }
2582
2583 static int ext4_da_write_end(struct file *file,
2584                                 struct address_space *mapping,
2585                                 loff_t pos, unsigned len, unsigned copied,
2586                                 struct page *page, void *fsdata)
2587 {
2588         struct inode *inode = mapping->host;
2589         int ret = 0, ret2;
2590         handle_t *handle = ext4_journal_current_handle();
2591         loff_t new_i_size;
2592         unsigned long start, end;
2593         int write_mode = (int)(unsigned long)fsdata;
2594
2595         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
2596                 if (ext4_should_order_data(inode)) {
2597                         return ext4_ordered_write_end(file, mapping, pos,
2598                                         len, copied, page, fsdata);
2599                 } else if (ext4_should_writeback_data(inode)) {
2600                         return ext4_writeback_write_end(file, mapping, pos,
2601                                         len, copied, page, fsdata);
2602                 } else {
2603                         BUG();
2604                 }
2605         }
2606
2607         start = pos & (PAGE_CACHE_SIZE - 1);
2608         end = start + copied - 1;
2609
2610         /*
2611          * generic_write_end() will run mark_inode_dirty() if i_size
2612          * changes.  So let's piggyback the i_disksize mark_inode_dirty
2613          * into that.
2614          */
2615
2616         new_i_size = pos + copied;
2617         if (new_i_size > EXT4_I(inode)->i_disksize) {
2618                 if (ext4_da_should_update_i_disksize(page, end)) {
2619                         down_write(&EXT4_I(inode)->i_data_sem);
2620                         if (new_i_size > EXT4_I(inode)->i_disksize) {
2621                                 /*
2622                                  * Updating i_disksize when extending file
2623                                  * without needing block allocation
2624                                  */
2625                                 if (ext4_should_order_data(inode))
2626                                         ret = ext4_jbd2_file_inode(handle,
2627                                                                    inode);
2628
2629                                 EXT4_I(inode)->i_disksize = new_i_size;
2630                         }
2631                         up_write(&EXT4_I(inode)->i_data_sem);
2632                         /* We need to mark inode dirty even if
2633                          * new_i_size is less that inode->i_size
2634                          * bu greater than i_disksize.(hint delalloc)
2635                          */
2636                         ext4_mark_inode_dirty(handle, inode);
2637                 }
2638         }
2639         ret2 = generic_write_end(file, mapping, pos, len, copied,
2640                                                         page, fsdata);
2641         copied = ret2;
2642         if (ret2 < 0)
2643                 ret = ret2;
2644         ret2 = ext4_journal_stop(handle);
2645         if (!ret)
2646                 ret = ret2;
2647
2648         return ret ? ret : copied;
2649 }
2650
2651 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
2652 {
2653         /*
2654          * Drop reserved blocks
2655          */
2656         BUG_ON(!PageLocked(page));
2657         if (!page_has_buffers(page))
2658                 goto out;
2659
2660         ext4_da_page_release_reservation(page, offset);
2661
2662 out:
2663         ext4_invalidatepage(page, offset);
2664
2665         return;
2666 }
2667
2668
2669 /*
2670  * bmap() is special.  It gets used by applications such as lilo and by
2671  * the swapper to find the on-disk block of a specific piece of data.
2672  *
2673  * Naturally, this is dangerous if the block concerned is still in the
2674  * journal.  If somebody makes a swapfile on an ext4 data-journaling
2675  * filesystem and enables swap, then they may get a nasty shock when the
2676  * data getting swapped to that swapfile suddenly gets overwritten by
2677  * the original zero's written out previously to the journal and
2678  * awaiting writeback in the kernel's buffer cache.
2679  *
2680  * So, if we see any bmap calls here on a modified, data-journaled file,
2681  * take extra steps to flush any blocks which might be in the cache.
2682  */
2683 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
2684 {
2685         struct inode *inode = mapping->host;
2686         journal_t *journal;
2687         int err;
2688
2689         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
2690                         test_opt(inode->i_sb, DELALLOC)) {
2691                 /*
2692                  * With delalloc we want to sync the file
2693                  * so that we can make sure we allocate
2694                  * blocks for file
2695                  */
2696                 filemap_write_and_wait(mapping);
2697         }
2698
2699         if (EXT4_I(inode)->i_state & EXT4_STATE_JDATA) {
2700                 /*
2701                  * This is a REALLY heavyweight approach, but the use of
2702                  * bmap on dirty files is expected to be extremely rare:
2703                  * only if we run lilo or swapon on a freshly made file
2704                  * do we expect this to happen.
2705                  *
2706                  * (bmap requires CAP_SYS_RAWIO so this does not
2707                  * represent an unprivileged user DOS attack --- we'd be
2708                  * in trouble if mortal users could trigger this path at
2709                  * will.)
2710                  *
2711                  * NB. EXT4_STATE_JDATA is not set on files other than
2712                  * regular files.  If somebody wants to bmap a directory
2713                  * or symlink and gets confused because the buffer
2714                  * hasn't yet been flushed to disk, they deserve
2715                  * everything they get.
2716                  */
2717
2718                 EXT4_I(inode)->i_state &= ~EXT4_STATE_JDATA;
2719                 journal = EXT4_JOURNAL(inode);
2720                 jbd2_journal_lock_updates(journal);
2721                 err = jbd2_journal_flush(journal);
2722                 jbd2_journal_unlock_updates(journal);
2723
2724                 if (err)
2725                         return 0;
2726         }
2727
2728         return generic_block_bmap(mapping, block, ext4_get_block);
2729 }
2730
2731 static int bget_one(handle_t *handle, struct buffer_head *bh)
2732 {
2733         get_bh(bh);
2734         return 0;
2735 }
2736
2737 static int bput_one(handle_t *handle, struct buffer_head *bh)
2738 {
2739         put_bh(bh);
2740         return 0;
2741 }
2742
2743 /*
2744  * Note that we don't need to start a transaction unless we're journaling data
2745  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2746  * need to file the inode to the transaction's list in ordered mode because if
2747  * we are writing back data added by write(), the inode is already there and if
2748  * we are writing back data modified via mmap(), noone guarantees in which
2749  * transaction the data will hit the disk. In case we are journaling data, we
2750  * cannot start transaction directly because transaction start ranks above page
2751  * lock so we have to do some magic.
2752  *
2753  * In all journaling modes block_write_full_page() will start the I/O.
2754  *
2755  * Problem:
2756  *
2757  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2758  *              ext4_writepage()
2759  *
2760  * Similar for:
2761  *
2762  *      ext4_file_write() -> generic_file_write() -> __alloc_pages() -> ...
2763  *
2764  * Same applies to ext4_get_block().  We will deadlock on various things like
2765  * lock_journal and i_data_sem
2766  *
2767  * Setting PF_MEMALLOC here doesn't work - too many internal memory
2768  * allocations fail.
2769  *
2770  * 16May01: If we're reentered then journal_current_handle() will be
2771  *          non-zero. We simply *return*.
2772  *
2773  * 1 July 2001: @@@ FIXME:
2774  *   In journalled data mode, a data buffer may be metadata against the
2775  *   current transaction.  But the same file is part of a shared mapping
2776  *   and someone does a writepage() on it.
2777  *
2778  *   We will move the buffer onto the async_data list, but *after* it has
2779  *   been dirtied. So there's a small window where we have dirty data on
2780  *   BJ_Metadata.
2781  *
2782  *   Note that this only applies to the last partial page in the file.  The
2783  *   bit which block_write_full_page() uses prepare/commit for.  (That's
2784  *   broken code anyway: it's wrong for msync()).
2785  *
2786  *   It's a rare case: affects the final partial page, for journalled data
2787  *   where the file is subject to bith write() and writepage() in the same
2788  *   transction.  To fix it we'll need a custom block_write_full_page().
2789  *   We'll probably need that anyway for journalling writepage() output.
2790  *
2791  * We don't honour synchronous mounts for writepage().  That would be
2792  * disastrous.  Any write() or metadata operation will sync the fs for
2793  * us.
2794  *
2795  */
2796 static int __ext4_normal_writepage(struct page *page,
2797                                 struct writeback_control *wbc)
2798 {
2799         struct inode *inode = page->mapping->host;
2800
2801         if (test_opt(inode->i_sb, NOBH))
2802                 return nobh_writepage(page,
2803                                         ext4_normal_get_block_write, wbc);
2804         else
2805                 return block_write_full_page(page,
2806                                                 ext4_normal_get_block_write,
2807                                                 wbc);
2808 }
2809
2810 static int ext4_normal_writepage(struct page *page,
2811                                 struct writeback_control *wbc)
2812 {
2813         struct inode *inode = page->mapping->host;
2814         loff_t size = i_size_read(inode);
2815         loff_t len;
2816
2817         J_ASSERT(PageLocked(page));
2818         if (page->index == size >> PAGE_CACHE_SHIFT)
2819                 len = size & ~PAGE_CACHE_MASK;
2820         else
2821                 len = PAGE_CACHE_SIZE;
2822
2823         if (page_has_buffers(page)) {
2824                 /* if page has buffers it should all be mapped
2825                  * and allocated. If there are not buffers attached
2826                  * to the page we know the page is dirty but it lost
2827                  * buffers. That means that at some moment in time
2828                  * after write_begin() / write_end() has been called
2829                  * all buffers have been clean and thus they must have been
2830                  * written at least once. So they are all mapped and we can
2831                  * happily proceed with mapping them and writing the page.
2832                  */
2833                 BUG_ON(walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
2834                                         ext4_bh_unmapped_or_delay));
2835         }
2836
2837         if (!ext4_journal_current_handle())
2838                 return __ext4_normal_writepage(page, wbc);
2839
2840         redirty_page_for_writepage(wbc, page);
2841         unlock_page(page);
2842         return 0;
2843 }
2844
2845 static int __ext4_journalled_writepage(struct page *page,
2846                                 struct writeback_control *wbc)
2847 {
2848         struct address_space *mapping = page->mapping;
2849         struct inode *inode = mapping->host;
2850         struct buffer_head *page_bufs;
2851         handle_t *handle = NULL;
2852         int ret = 0;
2853         int err;
2854
2855         ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
2856                                         ext4_normal_get_block_write);
2857         if (ret != 0)
2858                 goto out_unlock;
2859
2860         page_bufs = page_buffers(page);
2861         walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE, NULL,
2862                                                                 bget_one);
2863         /* As soon as we unlock the page, it can go away, but we have
2864          * references to buffers so we are safe */
2865         unlock_page(page);
2866
2867         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2868         if (IS_ERR(handle)) {
2869                 ret = PTR_ERR(handle);
2870                 goto out;
2871         }
2872
2873         ret = walk_page_buffers(handle, page_bufs, 0,
2874                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
2875
2876         err = walk_page_buffers(handle, page_bufs, 0,
2877                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
2878         if (ret == 0)
2879                 ret = err;
2880         err = ext4_journal_stop(handle);
2881         if (!ret)
2882                 ret = err;
2883
2884         walk_page_buffers(handle, page_bufs, 0,
2885                                 PAGE_CACHE_SIZE, NULL, bput_one);
2886         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
2887         goto out;
2888
2889 out_unlock:
2890         unlock_page(page);
2891 out:
2892         return ret;
2893 }
2894
2895 static int ext4_journalled_writepage(struct page *page,
2896                                 struct writeback_control *wbc)
2897 {
2898         struct inode *inode = page->mapping->host;
2899         loff_t size = i_size_read(inode);
2900         loff_t len;
2901
2902         J_ASSERT(PageLocked(page));
2903         if (page->index == size >> PAGE_CACHE_SHIFT)
2904                 len = size & ~PAGE_CACHE_MASK;
2905         else
2906                 len = PAGE_CACHE_SIZE;
2907
2908         if (page_has_buffers(page)) {
2909                 /* if page has buffers it should all be mapped
2910                  * and allocated. If there are not buffers attached
2911                  * to the page we know the page is dirty but it lost
2912                  * buffers. That means that at some moment in time
2913                  * after write_begin() / write_end() has been called
2914                  * all buffers have been clean and thus they must have been
2915                  * written at least once. So they are all mapped and we can
2916                  * happily proceed with mapping them and writing the page.
2917                  */
2918                 BUG_ON(walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
2919                                         ext4_bh_unmapped_or_delay));
2920         }
2921
2922         if (ext4_journal_current_handle())
2923                 goto no_write;
2924
2925         if (PageChecked(page)) {
2926                 /*
2927                  * It's mmapped pagecache.  Add buffers and journal it.  There
2928                  * doesn't seem much point in redirtying the page here.
2929                  */
2930                 ClearPageChecked(page);
2931                 return __ext4_journalled_writepage(page, wbc);
2932         } else {
2933                 /*
2934                  * It may be a page full of checkpoint-mode buffers.  We don't
2935                  * really know unless we go poke around in the buffer_heads.
2936                  * But block_write_full_page will do the right thing.
2937                  */
2938                 return block_write_full_page(page,
2939                                                 ext4_normal_get_block_write,
2940                                                 wbc);
2941         }
2942 no_write:
2943         redirty_page_for_writepage(wbc, page);
2944         unlock_page(page);
2945         return 0;
2946 }
2947
2948 static int ext4_readpage(struct file *file, struct page *page)
2949 {
2950         return mpage_readpage(page, ext4_get_block);
2951 }
2952
2953 static int
2954 ext4_readpages(struct file *file, struct address_space *mapping,
2955                 struct list_head *pages, unsigned nr_pages)
2956 {
2957         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
2958 }
2959
2960 static void ext4_invalidatepage(struct page *page, unsigned long offset)
2961 {
2962         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2963
2964         /*
2965          * If it's a full truncate we just forget about the pending dirtying
2966          */
2967         if (offset == 0)
2968                 ClearPageChecked(page);
2969
2970         jbd2_journal_invalidatepage(journal, page, offset);
2971 }
2972
2973 static int ext4_releasepage(struct page *page, gfp_t wait)
2974 {
2975         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2976
2977         WARN_ON(PageChecked(page));
2978         if (!page_has_buffers(page))
2979                 return 0;
2980         return jbd2_journal_try_to_free_buffers(journal, page, wait);
2981 }
2982
2983 /*
2984  * If the O_DIRECT write will extend the file then add this inode to the
2985  * orphan list.  So recovery will truncate it back to the original size
2986  * if the machine crashes during the write.
2987  *
2988  * If the O_DIRECT write is intantiating holes inside i_size and the machine
2989  * crashes then stale disk data _may_ be exposed inside the file. But current
2990  * VFS code falls back into buffered path in that case so we are safe.
2991  */
2992 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
2993                         const struct iovec *iov, loff_t offset,
2994                         unsigned long nr_segs)
2995 {
2996         struct file *file = iocb->ki_filp;
2997         struct inode *inode = file->f_mapping->host;
2998         struct ext4_inode_info *ei = EXT4_I(inode);
2999         handle_t *handle;
3000         ssize_t ret;
3001         int orphan = 0;
3002         size_t count = iov_length(iov, nr_segs);
3003
3004         if (rw == WRITE) {
3005                 loff_t final_size = offset + count;
3006
3007                 if (final_size > inode->i_size) {
3008                         /* Credits for sb + inode write */
3009                         handle = ext4_journal_start(inode, 2);
3010                         if (IS_ERR(handle)) {
3011                                 ret = PTR_ERR(handle);
3012                                 goto out;
3013                         }
3014                         ret = ext4_orphan_add(handle, inode);
3015                         if (ret) {
3016                                 ext4_journal_stop(handle);
3017                                 goto out;
3018                         }
3019                         orphan = 1;
3020                         ei->i_disksize = inode->i_size;
3021                         ext4_journal_stop(handle);
3022                 }
3023         }
3024
3025         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
3026                                  offset, nr_segs,
3027                                  ext4_get_block, NULL);
3028
3029         if (orphan) {
3030                 int err;
3031
3032                 /* Credits for sb + inode write */
3033                 handle = ext4_journal_start(inode, 2);
3034                 if (IS_ERR(handle)) {
3035                         /* This is really bad luck. We've written the data
3036                          * but cannot extend i_size. Bail out and pretend
3037                          * the write failed... */
3038                         ret = PTR_ERR(handle);
3039                         goto out;
3040                 }
3041                 if (inode->i_nlink)
3042                         ext4_orphan_del(handle, inode);
3043                 if (ret > 0) {
3044                         loff_t end = offset + ret;
3045                         if (end > inode->i_size) {
3046                                 ei->i_disksize = end;
3047                                 i_size_write(inode, end);
3048                                 /*
3049                                  * We're going to return a positive `ret'
3050                                  * here due to non-zero-length I/O, so there's
3051                                  * no way of reporting error returns from
3052                                  * ext4_mark_inode_dirty() to userspace.  So
3053                                  * ignore it.
3054                                  */
3055                                 ext4_mark_inode_dirty(handle, inode);
3056                         }
3057                 }
3058                 err = ext4_journal_stop(handle);
3059                 if (ret == 0)
3060                         ret = err;
3061         }
3062 out:
3063         return ret;
3064 }
3065
3066 /*
3067  * Pages can be marked dirty completely asynchronously from ext4's journalling
3068  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3069  * much here because ->set_page_dirty is called under VFS locks.  The page is
3070  * not necessarily locked.
3071  *
3072  * We cannot just dirty the page and leave attached buffers clean, because the
3073  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3074  * or jbddirty because all the journalling code will explode.
3075  *
3076  * So what we do is to mark the page "pending dirty" and next time writepage
3077  * is called, propagate that into the buffers appropriately.
3078  */
3079 static int ext4_journalled_set_page_dirty(struct page *page)
3080 {
3081         SetPageChecked(page);
3082         return __set_page_dirty_nobuffers(page);
3083 }
3084
3085 static const struct address_space_operations ext4_ordered_aops = {
3086         .readpage               = ext4_readpage,
3087         .readpages              = ext4_readpages,
3088         .writepage              = ext4_normal_writepage,
3089         .sync_page              = block_sync_page,
3090         .write_begin            = ext4_write_begin,
3091         .write_end              = ext4_ordered_write_end,
3092         .bmap                   = ext4_bmap,
3093         .invalidatepage         = ext4_invalidatepage,
3094         .releasepage            = ext4_releasepage,
3095         .direct_IO              = ext4_direct_IO,
3096         .migratepage            = buffer_migrate_page,
3097         .is_partially_uptodate  = block_is_partially_uptodate,
3098 };
3099
3100 static const struct address_space_operations ext4_writeback_aops = {
3101         .readpage               = ext4_readpage,
3102         .readpages              = ext4_readpages,
3103         .writepage              = ext4_normal_writepage,
3104         .sync_page              = block_sync_page,
3105         .write_begin            = ext4_write_begin,
3106         .write_end              = ext4_writeback_write_end,
3107         .bmap                   = ext4_bmap,
3108         .invalidatepage         = ext4_invalidatepage,
3109         .releasepage            = ext4_releasepage,
3110         .direct_IO              = ext4_direct_IO,
3111         .migratepage            = buffer_migrate_page,
3112         .is_partially_uptodate  = block_is_partially_uptodate,
3113 };
3114
3115 static const struct address_space_operations ext4_journalled_aops = {
3116         .readpage               = ext4_readpage,
3117         .readpages              = ext4_readpages,
3118         .writepage              = ext4_journalled_writepage,
3119         .sync_page              = block_sync_page,
3120         .write_begin            = ext4_write_begin,
3121         .write_end              = ext4_journalled_write_end,
3122         .set_page_dirty         = ext4_journalled_set_page_dirty,
3123         .bmap                   = ext4_bmap,
3124         .invalidatepage         = ext4_invalidatepage,
3125         .releasepage            = ext4_releasepage,
3126         .is_partially_uptodate  = block_is_partially_uptodate,
3127 };
3128
3129 static const struct address_space_operations ext4_da_aops = {
3130         .readpage               = ext4_readpage,
3131         .readpages              = ext4_readpages,
3132         .writepage              = ext4_da_writepage,
3133         .writepages             = ext4_da_writepages,
3134         .sync_page              = block_sync_page,
3135         .write_begin            = ext4_da_write_begin,
3136         .write_end              = ext4_da_write_end,
3137         .bmap                   = ext4_bmap,
3138         .invalidatepage         = ext4_da_invalidatepage,
3139         .releasepage            = ext4_releasepage,
3140         .direct_IO              = ext4_direct_IO,
3141         .migratepage            = buffer_migrate_page,
3142         .is_partially_uptodate  = block_is_partially_uptodate,
3143 };
3144
3145 void ext4_set_aops(struct inode *inode)
3146 {
3147         if (ext4_should_order_data(inode) &&
3148                 test_opt(inode->i_sb, DELALLOC))
3149                 inode->i_mapping->a_ops = &ext4_da_aops;
3150         else if (ext4_should_order_data(inode))
3151                 inode->i_mapping->a_ops = &ext4_ordered_aops;
3152         else if (ext4_should_writeback_data(inode) &&
3153                  test_opt(inode->i_sb, DELALLOC))
3154                 inode->i_mapping->a_ops = &ext4_da_aops;
3155         else if (ext4_should_writeback_data(inode))
3156                 inode->i_mapping->a_ops = &ext4_writeback_aops;
3157         else
3158                 inode->i_mapping->a_ops = &ext4_journalled_aops;
3159 }
3160
3161 /*
3162  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3163  * up to the end of the block which corresponds to `from'.
3164  * This required during truncate. We need to physically zero the tail end
3165  * of that block so it doesn't yield old data if the file is later grown.
3166  */
3167 int ext4_block_truncate_page(handle_t *handle,
3168                 struct address_space *mapping, loff_t from)
3169 {
3170         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3171         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3172         unsigned blocksize, length, pos;
3173         ext4_lblk_t iblock;
3174         struct inode *inode = mapping->host;
3175         struct buffer_head *bh;
3176         struct page *page;
3177         int err = 0;
3178
3179         page = grab_cache_page(mapping, from >> PAGE_CACHE_SHIFT);
3180         if (!page)
3181                 return -EINVAL;
3182
3183         blocksize = inode->i_sb->s_blocksize;
3184         length = blocksize - (offset & (blocksize - 1));
3185         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3186
3187         /*
3188          * For "nobh" option,  we can only work if we don't need to
3189          * read-in the page - otherwise we create buffers to do the IO.
3190          */
3191         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
3192              ext4_should_writeback_data(inode) && PageUptodate(page)) {
3193                 zero_user(page, offset, length);
3194                 set_page_dirty(page);
3195                 goto unlock;
3196         }
3197
3198         if (!page_has_buffers(page))
3199                 create_empty_buffers(page, blocksize, 0);
3200
3201         /* Find the buffer that contains "offset" */
3202         bh = page_buffers(page);
3203         pos = blocksize;
3204         while (offset >= pos) {
3205                 bh = bh->b_this_page;
3206                 iblock++;
3207                 pos += blocksize;
3208         }
3209
3210         err = 0;
3211         if (buffer_freed(bh)) {
3212                 BUFFER_TRACE(bh, "freed: skip");
3213                 goto unlock;
3214         }
3215
3216         if (!buffer_mapped(bh)) {
3217                 BUFFER_TRACE(bh, "unmapped");
3218                 ext4_get_block(inode, iblock, bh, 0);
3219                 /* unmapped? It's a hole - nothing to do */
3220                 if (!buffer_mapped(bh)) {
3221                         BUFFER_TRACE(bh, "still unmapped");
3222                         goto unlock;
3223                 }
3224         }
3225
3226         /* Ok, it's mapped. Make sure it's up-to-date */
3227         if (PageUptodate(page))
3228                 set_buffer_uptodate(bh);
3229
3230         if (!buffer_uptodate(bh)) {
3231                 err = -EIO;
3232                 ll_rw_block(READ, 1, &bh);
3233                 wait_on_buffer(bh);
3234                 /* Uhhuh. Read error. Complain and punt. */
3235                 if (!buffer_uptodate(bh))
3236                         goto unlock;
3237         }
3238
3239         if (ext4_should_journal_data(inode)) {
3240                 BUFFER_TRACE(bh, "get write access");
3241                 err = ext4_journal_get_write_access(handle, bh);
3242                 if (err)
3243                         goto unlock;
3244         }
3245
3246         zero_user(page, offset, length);
3247
3248         BUFFER_TRACE(bh, "zeroed end of block");
3249
3250         err = 0;
3251         if (ext4_should_journal_data(inode)) {
3252                 err = ext4_journal_dirty_metadata(handle, bh);
3253         } else {
3254                 if (ext4_should_order_data(inode))
3255                         err = ext4_jbd2_file_inode(handle, inode);
3256                 mark_buffer_dirty(bh);
3257         }
3258
3259 unlock:
3260         unlock_page(page);
3261         page_cache_release(page);
3262         return err;
3263 }
3264
3265 /*
3266  * Probably it should be a library function... search for first non-zero word
3267  * or memcmp with zero_page, whatever is better for particular architecture.
3268  * Linus?
3269  */
3270 static inline int all_zeroes(__le32 *p, __le32 *q)
3271 {
3272         while (p < q)
3273                 if (*p++)
3274                         return 0;
3275         return 1;
3276 }
3277
3278 /**
3279  *      ext4_find_shared - find the indirect blocks for partial truncation.
3280  *      @inode:   inode in question
3281  *      @depth:   depth of the affected branch
3282  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
3283  *      @chain:   place to store the pointers to partial indirect blocks
3284  *      @top:     place to the (detached) top of branch
3285  *
3286  *      This is a helper function used by ext4_truncate().
3287  *
3288  *      When we do truncate() we may have to clean the ends of several
3289  *      indirect blocks but leave the blocks themselves alive. Block is
3290  *      partially truncated if some data below the new i_size is refered
3291  *      from it (and it is on the path to the first completely truncated
3292  *      data block, indeed).  We have to free the top of that path along
3293  *      with everything to the right of the path. Since no allocation
3294  *      past the truncation point is possible until ext4_truncate()
3295  *      finishes, we may safely do the latter, but top of branch may
3296  *      require special attention - pageout below the truncation point
3297  *      might try to populate it.
3298  *
3299  *      We atomically detach the top of branch from the tree, store the
3300  *      block number of its root in *@top, pointers to buffer_heads of
3301  *      partially truncated blocks - in @chain[].bh and pointers to
3302  *      their last elements that should not be removed - in
3303  *      @chain[].p. Return value is the pointer to last filled element
3304  *      of @chain.
3305  *
3306  *      The work left to caller to do the actual freeing of subtrees:
3307  *              a) free the subtree starting from *@top
3308  *              b) free the subtrees whose roots are stored in
3309  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
3310  *              c) free the subtrees growing from the inode past the @chain[0].
3311  *                      (no partially truncated stuff there).  */
3312
3313 static Indirect *ext4_find_shared(struct inode *inode, int depth,
3314                         ext4_lblk_t offsets[4], Indirect chain[4], __le32 *top)
3315 {
3316         Indirect *partial, *p;
3317         int k, err;
3318
3319         *top = 0;
3320         /* Make k index the deepest non-null offest + 1 */
3321         for (k = depth; k > 1 && !offsets[k-1]; k--)
3322                 ;
3323         partial = ext4_get_branch(inode, k, offsets, chain, &err);
3324         /* Writer: pointers */
3325         if (!partial)
3326                 partial = chain + k-1;
3327         /*
3328          * If the branch acquired continuation since we've looked at it -
3329          * fine, it should all survive and (new) top doesn't belong to us.
3330          */
3331         if (!partial->key && *partial->p)
3332                 /* Writer: end */
3333                 goto no_top;
3334         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
3335                 ;
3336         /*
3337          * OK, we've found the last block that must survive. The rest of our
3338          * branch should be detached before unlocking. However, if that rest
3339          * of branch is all ours and does not grow immediately from the inode
3340          * it's easier to cheat and just decrement partial->p.
3341          */
3342         if (p == chain + k - 1 && p > chain) {
3343                 p->p--;
3344         } else {
3345                 *top = *p->p;
3346                 /* Nope, don't do this in ext4.  Must leave the tree intact */
3347 #if 0
3348                 *p->p = 0;
3349 #endif
3350         }
3351         /* Writer: end */
3352
3353         while (partial > p) {
3354                 brelse(partial->bh);
3355                 partial--;
3356         }
3357 no_top:
3358         return partial;
3359 }
3360
3361 /*
3362  * Zero a number of block pointers in either an inode or an indirect block.
3363  * If we restart the transaction we must again get write access to the
3364  * indirect block for further modification.
3365  *
3366  * We release `count' blocks on disk, but (last - first) may be greater
3367  * than `count' because there can be holes in there.
3368  */
3369 static void ext4_clear_blocks(handle_t *handle, struct inode *inode,
3370                 struct buffer_head *bh, ext4_fsblk_t block_to_free,
3371                 unsigned long count, __le32 *first, __le32 *last)
3372 {
3373         __le32 *p;
3374         if (try_to_extend_transaction(handle, inode)) {
3375                 if (bh) {
3376                         BUFFER_TRACE(bh, "call ext4_journal_dirty_metadata");
3377                         ext4_journal_dirty_metadata(handle, bh);
3378                 }
3379                 ext4_mark_inode_dirty(handle, inode);
3380                 ext4_journal_test_restart(handle, inode);
3381                 if (bh) {
3382                         BUFFER_TRACE(bh, "retaking write access");
3383                         ext4_journal_get_write_access(handle, bh);
3384                 }
3385         }
3386
3387         /*
3388          * Any buffers which are on the journal will be in memory. We find
3389          * them on the hash table so jbd2_journal_revoke() will run jbd2_journal_forget()
3390          * on them.  We've already detached each block from the file, so
3391          * bforget() in jbd2_journal_forget() should be safe.
3392          *
3393          * AKPM: turn on bforget in jbd2_journal_forget()!!!
3394          */
3395         for (p = first; p < last; p++) {
3396                 u32 nr = le32_to_cpu(*p);
3397                 if (nr) {
3398                         struct buffer_head *tbh;
3399
3400                         *p = 0;
3401                         tbh = sb_find_get_block(inode->i_sb, nr);
3402                         ext4_forget(handle, 0, inode, tbh, nr);
3403                 }
3404         }
3405
3406         ext4_free_blocks(handle, inode, block_to_free, count, 0);
3407 }
3408
3409 /**
3410  * ext4_free_data - free a list of data blocks
3411  * @handle:     handle for this transaction
3412  * @inode:      inode we are dealing with
3413  * @this_bh:    indirect buffer_head which contains *@first and *@last
3414  * @first:      array of block numbers
3415  * @last:       points immediately past the end of array
3416  *
3417  * We are freeing all blocks refered from that array (numbers are stored as
3418  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
3419  *
3420  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
3421  * blocks are contiguous then releasing them at one time will only affect one
3422  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
3423  * actually use a lot of journal space.
3424  *
3425  * @this_bh will be %NULL if @first and @last point into the inode's direct
3426  * block pointers.
3427  */
3428 static void ext4_free_data(handle_t *handle, struct inode *inode,
3429                            struct buffer_head *this_bh,
3430                            __le32 *first, __le32 *last)
3431 {
3432         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
3433         unsigned long count = 0;            /* Number of blocks in the run */
3434         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
3435                                                corresponding to
3436                                                block_to_free */
3437         ext4_fsblk_t nr;                    /* Current block # */
3438         __le32 *p;                          /* Pointer into inode/ind
3439                                                for current block */
3440         int err;
3441
3442         if (this_bh) {                          /* For indirect block */
3443                 BUFFER_TRACE(this_bh, "get_write_access");
3444                 err = ext4_journal_get_write_access(handle, this_bh);
3445                 /* Important: if we can't update the indirect pointers
3446                  * to the blocks, we can't free them. */
3447                 if (err)
3448                         return;
3449         }
3450
3451         for (p = first; p < last; p++) {
3452                 nr = le32_to_cpu(*p);
3453                 if (nr) {
3454                         /* accumulate blocks to free if they're contiguous */
3455                         if (count == 0) {
3456                                 block_to_free = nr;
3457                                 block_to_free_p = p;
3458                                 count = 1;
3459                         } else if (nr == block_to_free + count) {
3460                                 count++;
3461                         } else {
3462                                 ext4_clear_blocks(handle, inode, this_bh,
3463                                                   block_to_free,
3464                                                   count, block_to_free_p, p);
3465                                 block_to_free = nr;
3466                                 block_to_free_p = p;
3467                                 count = 1;
3468                         }
3469                 }
3470         }
3471
3472         if (count > 0)
3473                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
3474                                   count, block_to_free_p, p);
3475
3476         if (this_bh) {
3477                 BUFFER_TRACE(this_bh, "call ext4_journal_dirty_metadata");
3478
3479                 /*
3480                  * The buffer head should have an attached journal head at this
3481                  * point. However, if the data is corrupted and an indirect
3482                  * block pointed to itself, it would have been detached when
3483                  * the block was cleared. Check for this instead of OOPSing.
3484                  */
3485                 if (bh2jh(this_bh))
3486                         ext4_journal_dirty_metadata(handle, this_bh);
3487                 else
3488                         ext4_error(inode->i_sb, __func__,
3489                                    "circular indirect block detected, "
3490                                    "inode=%lu, block=%llu",
3491                                    inode->i_ino,
3492                                    (unsigned long long) this_bh->b_blocknr);
3493         }
3494 }
3495
3496 /**
3497  *      ext4_free_branches - free an array of branches
3498  *      @handle: JBD handle for this transaction
3499  *      @inode: inode we are dealing with
3500  *      @parent_bh: the buffer_head which contains *@first and *@last
3501  *      @first: array of block numbers
3502  *      @last:  pointer immediately past the end of array
3503  *      @depth: depth of the branches to free
3504  *
3505  *      We are freeing all blocks refered from these branches (numbers are
3506  *      stored as little-endian 32-bit) and updating @inode->i_blocks
3507  *      appropriately.
3508  */
3509 static void ext4_free_branches(handle_t *handle, struct inode *inode,
3510                                struct buffer_head *parent_bh,
3511                                __le32 *first, __le32 *last, int depth)
3512 {
3513         ext4_fsblk_t nr;
3514         __le32 *p;
3515
3516         if (is_handle_aborted(handle))
3517                 return;
3518
3519         if (depth--) {
3520                 struct buffer_head *bh;
3521                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
3522                 p = last;
3523                 while (--p >= first) {
3524                         nr = le32_to_cpu(*p);
3525                         if (!nr)
3526                                 continue;               /* A hole */
3527
3528                         /* Go read the buffer for the next level down */
3529                         bh = sb_bread(inode->i_sb, nr);
3530
3531                         /*
3532                          * A read failure? Report error and clear slot
3533                          * (should be rare).
3534                          */
3535                         if (!bh) {
3536                                 ext4_error(inode->i_sb, "ext4_free_branches",
3537                                            "Read failure, inode=%lu, block=%llu",
3538                                            inode->i_ino, nr);
3539                                 continue;
3540                         }
3541
3542                         /* This zaps the entire block.  Bottom up. */
3543                         BUFFER_TRACE(bh, "free child branches");
3544                         ext4_free_branches(handle, inode, bh,
3545                                         (__le32 *) bh->b_data,
3546                                         (__le32 *) bh->b_data + addr_per_block,
3547                                         depth);
3548
3549                         /*
3550                          * We've probably journalled the indirect block several
3551                          * times during the truncate.  But it's no longer
3552                          * needed and we now drop it from the transaction via
3553                          * jbd2_journal_revoke().
3554                          *
3555                          * That's easy if it's exclusively part of this
3556                          * transaction.  But if it's part of the committing
3557                          * transaction then jbd2_journal_forget() will simply
3558                          * brelse() it.  That means that if the underlying
3559                          * block is reallocated in ext4_get_block(),
3560                          * unmap_underlying_metadata() will find this block
3561                          * and will try to get rid of it.  damn, damn.
3562                          *
3563                          * If this block has already been committed to the
3564                          * journal, a revoke record will be written.  And
3565                          * revoke records must be emitted *before* clearing
3566                          * this block's bit in the bitmaps.
3567                          */
3568                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
3569
3570                         /*
3571                          * Everything below this this pointer has been
3572                          * released.  Now let this top-of-subtree go.
3573                          *
3574                          * We want the freeing of this indirect block to be
3575                          * atomic in the journal with the updating of the
3576                          * bitmap block which owns it.  So make some room in
3577                          * the journal.
3578                          *
3579                          * We zero the parent pointer *after* freeing its
3580                          * pointee in the bitmaps, so if extend_transaction()
3581                          * for some reason fails to put the bitmap changes and
3582                          * the release into the same transaction, recovery
3583                          * will merely complain about releasing a free block,
3584                          * rather than leaking blocks.
3585                          */
3586                         if (is_handle_aborted(handle))
3587                                 return;
3588                         if (try_to_extend_transaction(handle, inode)) {
3589                                 ext4_mark_inode_dirty(handle, inode);
3590                                 ext4_journal_test_restart(handle, inode);
3591                         }
3592
3593                         ext4_free_blocks(handle, inode, nr, 1, 1);
3594
3595                         if (parent_bh) {
3596                                 /*
3597                                  * The block which we have just freed is
3598                                  * pointed to by an indirect block: journal it
3599                                  */
3600                                 BUFFER_TRACE(parent_bh, "get_write_access");
3601                                 if (!ext4_journal_get_write_access(handle,
3602                                                                    parent_bh)){
3603                                         *p = 0;
3604                                         BUFFER_TRACE(parent_bh,
3605                                         "call ext4_journal_dirty_metadata");
3606                                         ext4_journal_dirty_metadata(handle,
3607                                                                     parent_bh);
3608                                 }
3609                         }
3610                 }
3611         } else {
3612                 /* We have reached the bottom of the tree. */
3613                 BUFFER_TRACE(parent_bh, "free data blocks");
3614                 ext4_free_data(handle, inode, parent_bh, first, last);
3615         }
3616 }
3617
3618 int ext4_can_truncate(struct inode *inode)
3619 {
3620         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
3621                 return 0;
3622         if (S_ISREG(inode->i_mode))
3623                 return 1;
3624         if (S_ISDIR(inode->i_mode))
3625                 return 1;
3626         if (S_ISLNK(inode->i_mode))
3627                 return !ext4_inode_is_fast_symlink(inode);
3628         return 0;
3629 }
3630
3631 /*
3632  * ext4_truncate()
3633  *
3634  * We block out ext4_get_block() block instantiations across the entire
3635  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
3636  * simultaneously on behalf of the same inode.
3637  *
3638  * As we work through the truncate and commmit bits of it to the journal there
3639  * is one core, guiding principle: the file's tree must always be consistent on
3640  * disk.  We must be able to restart the truncate after a crash.
3641  *
3642  * The file's tree may be transiently inconsistent in memory (although it
3643  * probably isn't), but whenever we close off and commit a journal transaction,
3644  * the contents of (the filesystem + the journal) must be consistent and
3645  * restartable.  It's pretty simple, really: bottom up, right to left (although
3646  * left-to-right works OK too).
3647  *
3648  * Note that at recovery time, journal replay occurs *before* the restart of
3649  * truncate against the orphan inode list.
3650  *
3651  * The committed inode has the new, desired i_size (which is the same as
3652  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
3653  * that this inode's truncate did not complete and it will again call
3654  * ext4_truncate() to have another go.  So there will be instantiated blocks
3655  * to the right of the truncation point in a crashed ext4 filesystem.  But
3656  * that's fine - as long as they are linked from the inode, the post-crash
3657  * ext4_truncate() run will find them and release them.
3658  */
3659 void ext4_truncate(struct inode *inode)
3660 {
3661         handle_t *handle;
3662         struct ext4_inode_info *ei = EXT4_I(inode);
3663         __le32 *i_data = ei->i_data;
3664         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
3665         struct address_space *mapping = inode->i_mapping;
3666         ext4_lblk_t offsets[4];
3667         Indirect chain[4];
3668         Indirect *partial;
3669         __le32 nr = 0;
3670         int n;
3671         ext4_lblk_t last_block;
3672         unsigned blocksize = inode->i_sb->s_blocksize;
3673
3674         if (!ext4_can_truncate(inode))
3675                 return;
3676
3677         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
3678                 ext4_ext_truncate(inode);
3679                 return;
3680         }
3681
3682         handle = start_transaction(inode);
3683         if (IS_ERR(handle))
3684                 return;         /* AKPM: return what? */
3685
3686         last_block = (inode->i_size + blocksize-1)
3687                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
3688
3689         if (inode->i_size & (blocksize - 1))
3690                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
3691                         goto out_stop;
3692
3693         n = ext4_block_to_path(inode, last_block, offsets, NULL);
3694         if (n == 0)
3695                 goto out_stop;  /* error */
3696
3697         /*
3698          * OK.  This truncate is going to happen.  We add the inode to the
3699          * orphan list, so that if this truncate spans multiple transactions,
3700          * and we crash, we will resume the truncate when the filesystem
3701          * recovers.  It also marks the inode dirty, to catch the new size.
3702          *
3703          * Implication: the file must always be in a sane, consistent
3704          * truncatable state while each transaction commits.
3705          */
3706         if (ext4_orphan_add(handle, inode))
3707                 goto out_stop;
3708
3709         /*
3710          * From here we block out all ext4_get_block() callers who want to
3711          * modify the block allocation tree.
3712          */
3713         down_write(&ei->i_data_sem);
3714
3715         ext4_discard_preallocations(inode);
3716
3717         /*
3718          * The orphan list entry will now protect us from any crash which
3719          * occurs before the truncate completes, so it is now safe to propagate
3720          * the new, shorter inode size (held for now in i_size) into the
3721          * on-disk inode. We do this via i_disksize, which is the value which
3722          * ext4 *really* writes onto the disk inode.
3723          */
3724         ei->i_disksize = inode->i_size;
3725
3726         if (n == 1) {           /* direct blocks */
3727                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
3728                                i_data + EXT4_NDIR_BLOCKS);
3729                 goto do_indirects;
3730         }
3731
3732         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
3733         /* Kill the top of shared branch (not detached) */
3734         if (nr) {
3735                 if (partial == chain) {
3736                         /* Shared branch grows from the inode */
3737                         ext4_free_branches(handle, inode, NULL,
3738                                            &nr, &nr+1, (chain+n-1) - partial);
3739                         *partial->p = 0;
3740                         /*
3741                          * We mark the inode dirty prior to restart,
3742                          * and prior to stop.  No need for it here.
3743                          */
3744                 } else {
3745                         /* Shared branch grows from an indirect block */
3746                         BUFFER_TRACE(partial->bh, "get_write_access");
3747                         ext4_free_branches(handle, inode, partial->bh,
3748                                         partial->p,
3749                                         partial->p+1, (chain+n-1) - partial);
3750                 }
3751         }
3752         /* Clear the ends of indirect blocks on the shared branch */
3753         while (partial > chain) {
3754                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
3755                                    (__le32*)partial->bh->b_data+addr_per_block,
3756                                    (chain+n-1) - partial);
3757                 BUFFER_TRACE(partial->bh, "call brelse");
3758                 brelse (partial->bh);
3759                 partial--;
3760         }
3761 do_indirects:
3762         /* Kill the remaining (whole) subtrees */
3763         switch (offsets[0]) {
3764         default:
3765                 nr = i_data[EXT4_IND_BLOCK];
3766                 if (nr) {
3767                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
3768                         i_data[EXT4_IND_BLOCK] = 0;
3769                 }
3770         case EXT4_IND_BLOCK:
3771                 nr = i_data[EXT4_DIND_BLOCK];
3772                 if (nr) {
3773                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
3774                         i_data[EXT4_DIND_BLOCK] = 0;
3775                 }
3776         case EXT4_DIND_BLOCK:
3777                 nr = i_data[EXT4_TIND_BLOCK];
3778                 if (nr) {
3779                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
3780                         i_data[EXT4_TIND_BLOCK] = 0;
3781                 }
3782         case EXT4_TIND_BLOCK:
3783                 ;
3784         }
3785
3786         up_write(&ei->i_data_sem);
3787         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
3788         ext4_mark_inode_dirty(handle, inode);
3789
3790         /*
3791          * In a multi-transaction truncate, we only make the final transaction
3792          * synchronous
3793          */
3794         if (IS_SYNC(inode))
3795                 handle->h_sync = 1;
3796 out_stop:
3797         /*
3798          * If this was a simple ftruncate(), and the file will remain alive
3799          * then we need to clear up the orphan record which we created above.
3800          * However, if this was a real unlink then we were called by
3801          * ext4_delete_inode(), and we allow that function to clean up the
3802          * orphan info for us.
3803          */
3804         if (inode->i_nlink)
3805                 ext4_orphan_del(handle, inode);
3806
3807         ext4_journal_stop(handle);
3808 }
3809
3810 /*
3811  * ext4_get_inode_loc returns with an extra refcount against the inode's
3812  * underlying buffer_head on success. If 'in_mem' is true, we have all
3813  * data in memory that is needed to recreate the on-disk version of this
3814  * inode.
3815  */
3816 static int __ext4_get_inode_loc(struct inode *inode,
3817                                 struct ext4_iloc *iloc, int in_mem)
3818 {
3819         struct ext4_group_desc  *gdp;
3820         struct buffer_head      *bh;
3821         struct super_block      *sb = inode->i_sb;
3822         ext4_fsblk_t            block;
3823         int                     inodes_per_block, inode_offset;
3824
3825         iloc->bh = 0;
3826         if (!ext4_valid_inum(sb, inode->i_ino))
3827                 return -EIO;
3828
3829         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
3830         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
3831         if (!gdp)
3832                 return -EIO;
3833
3834         /*
3835          * Figure out the offset within the block group inode table
3836          */
3837         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
3838         inode_offset = ((inode->i_ino - 1) %
3839                         EXT4_INODES_PER_GROUP(sb));
3840         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
3841         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
3842
3843         bh = sb_getblk(sb, block);
3844         if (!bh) {
3845                 ext4_error(sb, "ext4_get_inode_loc", "unable to read "
3846                            "inode block - inode=%lu, block=%llu",
3847                            inode->i_ino, block);
3848                 return -EIO;
3849         }
3850         if (!buffer_uptodate(bh)) {
3851                 lock_buffer(bh);
3852
3853                 /*
3854                  * If the buffer has the write error flag, we have failed
3855                  * to write out another inode in the same block.  In this
3856                  * case, we don't have to read the block because we may
3857                  * read the old inode data successfully.
3858                  */
3859                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
3860                         set_buffer_uptodate(bh);
3861
3862                 if (buffer_uptodate(bh)) {
3863                         /* someone brought it uptodate while we waited */
3864                         unlock_buffer(bh);
3865                         goto has_buffer;
3866                 }
3867
3868                 /*
3869                  * If we have all information of the inode in memory and this
3870                  * is the only valid inode in the block, we need not read the
3871                  * block.
3872                  */
3873                 if (in_mem) {
3874                         struct buffer_head *bitmap_bh;
3875                         int i, start;
3876
3877                         start = inode_offset & ~(inodes_per_block - 1);
3878
3879                         /* Is the inode bitmap in cache? */
3880                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
3881                         if (!bitmap_bh)
3882                                 goto make_io;
3883
3884                         /*
3885                          * If the inode bitmap isn't in cache then the
3886                          * optimisation may end up performing two reads instead
3887                          * of one, so skip it.
3888                          */
3889                         if (!buffer_uptodate(bitmap_bh)) {
3890                                 brelse(bitmap_bh);
3891                                 goto make_io;
3892                         }
3893                         for (i = start; i < start + inodes_per_block; i++) {
3894                                 if (i == inode_offset)
3895                                         continue;
3896                                 if (ext4_test_bit(i, bitmap_bh->b_data))
3897                                         break;
3898                         }
3899                         brelse(bitmap_bh);
3900                         if (i == start + inodes_per_block) {
3901                                 /* all other inodes are free, so skip I/O */
3902                                 memset(bh->b_data, 0, bh->b_size);
3903                                 set_buffer_uptodate(bh);
3904                                 unlock_buffer(bh);
3905                                 goto has_buffer;
3906                         }
3907                 }
3908
3909 make_io:
3910                 /*
3911                  * If we need to do any I/O, try to pre-readahead extra
3912                  * blocks from the inode table.
3913                  */
3914                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
3915                         ext4_fsblk_t b, end, table;
3916                         unsigned num;
3917
3918                         table = ext4_inode_table(sb, gdp);
3919                         /* Make sure s_inode_readahead_blks is a power of 2 */
3920                         while (EXT4_SB(sb)->s_inode_readahead_blks &
3921                                (EXT4_SB(sb)->s_inode_readahead_blks-1))
3922                                 EXT4_SB(sb)->s_inode_readahead_blks = 
3923                                    (EXT4_SB(sb)->s_inode_readahead_blks &
3924                                     (EXT4_SB(sb)->s_inode_readahead_blks-1));
3925                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
3926                         if (table > b)
3927                                 b = table;
3928                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
3929                         num = EXT4_INODES_PER_GROUP(sb);
3930                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
3931                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
3932                                 num -= le16_to_cpu(gdp->bg_itable_unused);
3933                         table += num / inodes_per_block;
3934                         if (end > table)
3935                                 end = table;
3936                         while (b <= end)
3937                                 sb_breadahead(sb, b++);
3938                 }
3939
3940                 /*
3941                  * There are other valid inodes in the buffer, this inode
3942                  * has in-inode xattrs, or we don't have this inode in memory.
3943                  * Read the block from disk.
3944                  */
3945                 get_bh(bh);
3946                 bh->b_end_io = end_buffer_read_sync;
3947                 submit_bh(READ_META, bh);
3948                 wait_on_buffer(bh);
3949                 if (!buffer_uptodate(bh)) {
3950                         ext4_error(sb, __func__,
3951                                    "unable to read inode block - inode=%lu, "
3952                                    "block=%llu", inode->i_ino, block);
3953                         brelse(bh);
3954                         return -EIO;
3955                 }
3956         }
3957 has_buffer:
3958         iloc->bh = bh;
3959         return 0;
3960 }
3961
3962 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
3963 {
3964         /* We have all inode data except xattrs in memory here. */
3965         return __ext4_get_inode_loc(inode, iloc,
3966                 !(EXT4_I(inode)->i_state & EXT4_STATE_XATTR));
3967 }
3968
3969 void ext4_set_inode_flags(struct inode *inode)
3970 {
3971         unsigned int flags = EXT4_I(inode)->i_flags;
3972
3973         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
3974         if (flags & EXT4_SYNC_FL)
3975                 inode->i_flags |= S_SYNC;
3976         if (flags & EXT4_APPEND_FL)
3977                 inode->i_flags |= S_APPEND;
3978         if (flags & EXT4_IMMUTABLE_FL)
3979                 inode->i_flags |= S_IMMUTABLE;
3980         if (flags & EXT4_NOATIME_FL)
3981                 inode->i_flags |= S_NOATIME;
3982         if (flags & EXT4_DIRSYNC_FL)
3983                 inode->i_flags |= S_DIRSYNC;
3984 }
3985
3986 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
3987 void ext4_get_inode_flags(struct ext4_inode_info *ei)
3988 {
3989         unsigned int flags = ei->vfs_inode.i_flags;
3990
3991         ei->i_flags &= ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
3992                         EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|EXT4_DIRSYNC_FL);
3993         if (flags & S_SYNC)
3994                 ei->i_flags |= EXT4_SYNC_FL;
3995         if (flags & S_APPEND)
3996                 ei->i_flags |= EXT4_APPEND_FL;
3997         if (flags & S_IMMUTABLE)
3998                 ei->i_flags |= EXT4_IMMUTABLE_FL;
3999         if (flags & S_NOATIME)
4000                 ei->i_flags |= EXT4_NOATIME_FL;
4001         if (flags & S_DIRSYNC)
4002                 ei->i_flags |= EXT4_DIRSYNC_FL;
4003 }
4004 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4005                                         struct ext4_inode_info *ei)
4006 {
4007         blkcnt_t i_blocks ;
4008         struct inode *inode = &(ei->vfs_inode);
4009         struct super_block *sb = inode->i_sb;
4010
4011         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4012                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4013                 /* we are using combined 48 bit field */
4014                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4015                                         le32_to_cpu(raw_inode->i_blocks_lo);
4016                 if (ei->i_flags & EXT4_HUGE_FILE_FL) {
4017                         /* i_blocks represent file system block size */
4018                         return i_blocks  << (inode->i_blkbits - 9);
4019                 } else {
4020                         return i_blocks;
4021                 }
4022         } else {
4023                 return le32_to_cpu(raw_inode->i_blocks_lo);
4024         }
4025 }
4026
4027 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4028 {
4029         struct ext4_iloc iloc;
4030         struct ext4_inode *raw_inode;
4031         struct ext4_inode_info *ei;
4032         struct buffer_head *bh;
4033         struct inode *inode;
4034         long ret;
4035         int block;
4036
4037         inode = iget_locked(sb, ino);
4038         if (!inode)
4039                 return ERR_PTR(-ENOMEM);
4040         if (!(inode->i_state & I_NEW))
4041                 return inode;
4042
4043         ei = EXT4_I(inode);
4044 #ifdef CONFIG_EXT4_FS_POSIX_ACL
4045         ei->i_acl = EXT4_ACL_NOT_CACHED;
4046         ei->i_default_acl = EXT4_ACL_NOT_CACHED;
4047 #endif
4048
4049         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4050         if (ret < 0)
4051                 goto bad_inode;
4052         bh = iloc.bh;
4053         raw_inode = ext4_raw_inode(&iloc);
4054         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4055         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4056         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4057         if (!(test_opt(inode->i_sb, NO_UID32))) {
4058                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4059                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4060         }
4061         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
4062
4063         ei->i_state = 0;
4064         ei->i_dir_start_lookup = 0;
4065         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4066         /* We now have enough fields to check if the inode was active or not.
4067          * This is needed because nfsd might try to access dead inodes
4068          * the test is that same one that e2fsck uses
4069          * NeilBrown 1999oct15
4070          */
4071         if (inode->i_nlink == 0) {
4072                 if (inode->i_mode == 0 ||
4073                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
4074                         /* this inode is deleted */
4075                         brelse(bh);
4076                         ret = -ESTALE;
4077                         goto bad_inode;
4078                 }
4079                 /* The only unlinked inodes we let through here have
4080                  * valid i_mode and are being read by the orphan
4081                  * recovery code: that's fine, we're about to complete
4082                  * the process of deleting those. */
4083         }
4084         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4085         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4086         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4087         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
4088             cpu_to_le32(EXT4_OS_HURD)) {
4089                 ei->i_file_acl |=
4090                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4091         }
4092         inode->i_size = ext4_isize(raw_inode);
4093         ei->i_disksize = inode->i_size;
4094         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4095         ei->i_block_group = iloc.block_group;
4096         /*
4097          * NOTE! The in-memory inode i_data array is in little-endian order
4098          * even on big-endian machines: we do NOT byteswap the block numbers!
4099          */
4100         for (block = 0; block < EXT4_N_BLOCKS; block++)
4101                 ei->i_data[block] = raw_inode->i_block[block];
4102         INIT_LIST_HEAD(&ei->i_orphan);
4103
4104         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4105                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4106                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4107                     EXT4_INODE_SIZE(inode->i_sb)) {
4108                         brelse(bh);
4109                         ret = -EIO;
4110                         goto bad_inode;
4111                 }
4112                 if (ei->i_extra_isize == 0) {
4113                         /* The extra space is currently unused. Use it. */
4114                         ei->i_extra_isize = sizeof(struct ext4_inode) -
4115                                             EXT4_GOOD_OLD_INODE_SIZE;
4116                 } else {
4117                         __le32 *magic = (void *)raw_inode +
4118                                         EXT4_GOOD_OLD_INODE_SIZE +
4119                                         ei->i_extra_isize;
4120                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
4121                                  ei->i_state |= EXT4_STATE_XATTR;
4122                 }
4123         } else
4124                 ei->i_extra_isize = 0;
4125
4126         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4127         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4128         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4129         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4130
4131         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4132         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4133                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4134                         inode->i_version |=
4135                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4136         }
4137
4138         if (S_ISREG(inode->i_mode)) {
4139                 inode->i_op = &ext4_file_inode_operations;
4140                 inode->i_fop = &ext4_file_operations;
4141                 ext4_set_aops(inode);
4142         } else if (S_ISDIR(inode->i_mode)) {
4143                 inode->i_op = &ext4_dir_inode_operations;
4144                 inode->i_fop = &ext4_dir_operations;
4145         } else if (S_ISLNK(inode->i_mode)) {
4146                 if (ext4_inode_is_fast_symlink(inode))
4147                         inode->i_op = &ext4_fast_symlink_inode_operations;
4148                 else {
4149                         inode->i_op = &ext4_symlink_inode_operations;
4150                         ext4_set_aops(inode);
4151                 }
4152         } else {
4153                 inode->i_op = &ext4_special_inode_operations;
4154                 if (raw_inode->i_block[0])
4155                         init_special_inode(inode, inode->i_mode,
4156                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4157                 else
4158                         init_special_inode(inode, inode->i_mode,
4159                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4160         }
4161         brelse(iloc.bh);
4162         ext4_set_inode_flags(inode);
4163         unlock_new_inode(inode);
4164         return inode;
4165
4166 bad_inode:
4167         iget_failed(inode);
4168         return ERR_PTR(ret);
4169 }
4170
4171 static int ext4_inode_blocks_set(handle_t *handle,
4172                                 struct ext4_inode *raw_inode,
4173                                 struct ext4_inode_info *ei)
4174 {
4175         struct inode *inode = &(ei->vfs_inode);
4176         u64 i_blocks = inode->i_blocks;
4177         struct super_block *sb = inode->i_sb;
4178         int err = 0;
4179
4180         if (i_blocks <= ~0U) {
4181                 /*
4182                  * i_blocks can be represnted in a 32 bit variable
4183                  * as multiple of 512 bytes
4184                  */
4185                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4186                 raw_inode->i_blocks_high = 0;
4187                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4188         } else if (i_blocks <= 0xffffffffffffULL) {
4189                 /*
4190                  * i_blocks can be represented in a 48 bit variable
4191                  * as multiple of 512 bytes
4192                  */
4193                 err = ext4_update_rocompat_feature(handle, sb,
4194                                             EXT4_FEATURE_RO_COMPAT_HUGE_FILE);
4195                 if (err)
4196                         goto  err_out;
4197                 /* i_block is stored in the split  48 bit fields */
4198                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4199                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4200                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4201         } else {
4202                 /*
4203                  * i_blocks should be represented in a 48 bit variable
4204                  * as multiple of  file system block size
4205                  */
4206                 err = ext4_update_rocompat_feature(handle, sb,
4207                                             EXT4_FEATURE_RO_COMPAT_HUGE_FILE);
4208                 if (err)
4209                         goto  err_out;
4210                 ei->i_flags |= EXT4_HUGE_FILE_FL;
4211                 /* i_block is stored in file system block size */
4212                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4213                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4214                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4215         }
4216 err_out:
4217         return err;
4218 }
4219
4220 /*
4221  * Post the struct inode info into an on-disk inode location in the
4222  * buffer-cache.  This gobbles the caller's reference to the
4223  * buffer_head in the inode location struct.
4224  *
4225  * The caller must have write access to iloc->bh.
4226  */
4227 static int ext4_do_update_inode(handle_t *handle,
4228                                 struct inode *inode,
4229                                 struct ext4_iloc *iloc)
4230 {
4231         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
4232         struct ext4_inode_info *ei = EXT4_I(inode);
4233         struct buffer_head *bh = iloc->bh;
4234         int err = 0, rc, block;
4235
4236         /* For fields not not tracking in the in-memory inode,
4237          * initialise them to zero for new inodes. */
4238         if (ei->i_state & EXT4_STATE_NEW)
4239                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
4240
4241         ext4_get_inode_flags(ei);
4242         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
4243         if (!(test_opt(inode->i_sb, NO_UID32))) {
4244                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
4245                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
4246 /*
4247  * Fix up interoperability with old kernels. Otherwise, old inodes get
4248  * re-used with the upper 16 bits of the uid/gid intact
4249  */
4250                 if (!ei->i_dtime) {
4251                         raw_inode->i_uid_high =
4252                                 cpu_to_le16(high_16_bits(inode->i_uid));
4253                         raw_inode->i_gid_high =
4254                                 cpu_to_le16(high_16_bits(inode->i_gid));
4255                 } else {
4256                         raw_inode->i_uid_high = 0;
4257                         raw_inode->i_gid_high = 0;
4258                 }
4259         } else {
4260                 raw_inode->i_uid_low =
4261                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
4262                 raw_inode->i_gid_low =
4263                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
4264                 raw_inode->i_uid_high = 0;
4265                 raw_inode->i_gid_high = 0;
4266         }
4267         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
4268
4269         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
4270         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
4271         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
4272         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
4273
4274         if (ext4_inode_blocks_set(handle, raw_inode, ei))
4275                 goto out_brelse;
4276         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
4277         /* clear the migrate flag in the raw_inode */
4278         raw_inode->i_flags = cpu_to_le32(ei->i_flags & ~EXT4_EXT_MIGRATE);
4279         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
4280             cpu_to_le32(EXT4_OS_HURD))
4281                 raw_inode->i_file_acl_high =
4282                         cpu_to_le16(ei->i_file_acl >> 32);
4283         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
4284         ext4_isize_set(raw_inode, ei->i_disksize);
4285         if (ei->i_disksize > 0x7fffffffULL) {
4286                 struct super_block *sb = inode->i_sb;
4287                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
4288                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
4289                                 EXT4_SB(sb)->s_es->s_rev_level ==
4290                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
4291                         /* If this is the first large file
4292                          * created, add a flag to the superblock.
4293                          */
4294                         err = ext4_journal_get_write_access(handle,
4295                                         EXT4_SB(sb)->s_sbh);
4296                         if (err)
4297                                 goto out_brelse;
4298                         ext4_update_dynamic_rev(sb);
4299                         EXT4_SET_RO_COMPAT_FEATURE(sb,
4300                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
4301                         sb->s_dirt = 1;
4302                         handle->h_sync = 1;
4303                         err = ext4_journal_dirty_metadata(handle,
4304                                         EXT4_SB(sb)->s_sbh);
4305                 }
4306         }
4307         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
4308         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
4309                 if (old_valid_dev(inode->i_rdev)) {
4310                         raw_inode->i_block[0] =
4311                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
4312                         raw_inode->i_block[1] = 0;
4313                 } else {
4314                         raw_inode->i_block[0] = 0;
4315                         raw_inode->i_block[1] =
4316                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
4317                         raw_inode->i_block[2] = 0;
4318                 }
4319         } else for (block = 0; block < EXT4_N_BLOCKS; block++)
4320                 raw_inode->i_block[block] = ei->i_data[block];
4321
4322         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
4323         if (ei->i_extra_isize) {
4324                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4325                         raw_inode->i_version_hi =
4326                         cpu_to_le32(inode->i_version >> 32);
4327                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
4328         }
4329
4330
4331         BUFFER_TRACE(bh, "call ext4_journal_dirty_metadata");
4332         rc = ext4_journal_dirty_metadata(handle, bh);
4333         if (!err)
4334                 err = rc;
4335         ei->i_state &= ~EXT4_STATE_NEW;
4336
4337 out_brelse:
4338         brelse(bh);
4339         ext4_std_error(inode->i_sb, err);
4340         return err;
4341 }
4342
4343 /*
4344  * ext4_write_inode()
4345  *
4346  * We are called from a few places:
4347  *
4348  * - Within generic_file_write() for O_SYNC files.
4349  *   Here, there will be no transaction running. We wait for any running
4350  *   trasnaction to commit.
4351  *
4352  * - Within sys_sync(), kupdate and such.
4353  *   We wait on commit, if tol to.
4354  *
4355  * - Within prune_icache() (PF_MEMALLOC == true)
4356  *   Here we simply return.  We can't afford to block kswapd on the
4357  *   journal commit.
4358  *
4359  * In all cases it is actually safe for us to return without doing anything,
4360  * because the inode has been copied into a raw inode buffer in
4361  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
4362  * knfsd.
4363  *
4364  * Note that we are absolutely dependent upon all inode dirtiers doing the
4365  * right thing: they *must* call mark_inode_dirty() after dirtying info in
4366  * which we are interested.
4367  *
4368  * It would be a bug for them to not do this.  The code:
4369  *
4370  *      mark_inode_dirty(inode)
4371  *      stuff();
4372  *      inode->i_size = expr;
4373  *
4374  * is in error because a kswapd-driven write_inode() could occur while
4375  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
4376  * will no longer be on the superblock's dirty inode list.
4377  */
4378 int ext4_write_inode(struct inode *inode, int wait)
4379 {
4380         if (current->flags & PF_MEMALLOC)
4381                 return 0;
4382
4383         if (ext4_journal_current_handle()) {
4384                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
4385                 dump_stack();
4386                 return -EIO;
4387         }
4388
4389         if (!wait)
4390                 return 0;
4391
4392         return ext4_force_commit(inode->i_sb);
4393 }
4394
4395 /*
4396  * ext4_setattr()
4397  *
4398  * Called from notify_change.
4399  *
4400  * We want to trap VFS attempts to truncate the file as soon as
4401  * possible.  In particular, we want to make sure that when the VFS
4402  * shrinks i_size, we put the inode on the orphan list and modify
4403  * i_disksize immediately, so that during the subsequent flushing of
4404  * dirty pages and freeing of disk blocks, we can guarantee that any
4405  * commit will leave the blocks being flushed in an unused state on
4406  * disk.  (On recovery, the inode will get truncated and the blocks will
4407  * be freed, so we have a strong guarantee that no future commit will
4408  * leave these blocks visible to the user.)
4409  *
4410  * Another thing we have to assure is that if we are in ordered mode
4411  * and inode is still attached to the committing transaction, we must
4412  * we start writeout of all the dirty pages which are being truncated.
4413  * This way we are sure that all the data written in the previous
4414  * transaction are already on disk (truncate waits for pages under
4415  * writeback).
4416  *
4417  * Called with inode->i_mutex down.
4418  */
4419 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
4420 {
4421         struct inode *inode = dentry->d_inode;
4422         int error, rc = 0;
4423         const unsigned int ia_valid = attr->ia_valid;
4424
4425         error = inode_change_ok(inode, attr);
4426         if (error)
4427                 return error;
4428
4429         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
4430                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
4431                 handle_t *handle;
4432
4433                 /* (user+group)*(old+new) structure, inode write (sb,
4434                  * inode block, ? - but truncate inode update has it) */
4435                 handle = ext4_journal_start(inode, 2*(EXT4_QUOTA_INIT_BLOCKS(inode->i_sb)+
4436                                         EXT4_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
4437                 if (IS_ERR(handle)) {
4438                         error = PTR_ERR(handle);
4439                         goto err_out;
4440                 }
4441                 error = DQUOT_TRANSFER(inode, attr) ? -EDQUOT : 0;
4442                 if (error) {
4443                         ext4_journal_stop(handle);
4444                         return error;
4445                 }
4446                 /* Update corresponding info in inode so that everything is in
4447                  * one transaction */
4448                 if (attr->ia_valid & ATTR_UID)
4449                         inode->i_uid = attr->ia_uid;
4450                 if (attr->ia_valid & ATTR_GID)
4451                         inode->i_gid = attr->ia_gid;
4452                 error = ext4_mark_inode_dirty(handle, inode);
4453                 ext4_journal_stop(handle);
4454         }
4455
4456         if (attr->ia_valid & ATTR_SIZE) {
4457                 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)) {
4458                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4459
4460                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
4461                                 error = -EFBIG;
4462                                 goto err_out;
4463                         }
4464                 }
4465         }
4466
4467         if (S_ISREG(inode->i_mode) &&
4468             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
4469                 handle_t *handle;
4470
4471                 handle = ext4_journal_start(inode, 3);
4472                 if (IS_ERR(handle)) {
4473                         error = PTR_ERR(handle);
4474                         goto err_out;
4475                 }
4476
4477                 error = ext4_orphan_add(handle, inode);
4478                 EXT4_I(inode)->i_disksize = attr->ia_size;
4479                 rc = ext4_mark_inode_dirty(handle, inode);
4480                 if (!error)
4481                         error = rc;
4482                 ext4_journal_stop(handle);
4483
4484                 if (ext4_should_order_data(inode)) {
4485                         error = ext4_begin_ordered_truncate(inode,
4486                                                             attr->ia_size);
4487                         if (error) {
4488                                 /* Do as much error cleanup as possible */
4489                                 handle = ext4_journal_start(inode, 3);
4490                                 if (IS_ERR(handle)) {
4491                                         ext4_orphan_del(NULL, inode);
4492                                         goto err_out;
4493                                 }
4494                                 ext4_orphan_del(handle, inode);
4495                                 ext4_journal_stop(handle);
4496                                 goto err_out;
4497                         }
4498                 }
4499         }
4500
4501         rc = inode_setattr(inode, attr);
4502
4503         /* If inode_setattr's call to ext4_truncate failed to get a
4504          * transaction handle at all, we need to clean up the in-core
4505          * orphan list manually. */
4506         if (inode->i_nlink)
4507                 ext4_orphan_del(NULL, inode);
4508
4509         if (!rc && (ia_valid & ATTR_MODE))
4510                 rc = ext4_acl_chmod(inode);
4511
4512 err_out:
4513         ext4_std_error(inode->i_sb, error);
4514         if (!error)
4515                 error = rc;
4516         return error;
4517 }
4518
4519 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
4520                  struct kstat *stat)
4521 {
4522         struct inode *inode;
4523         unsigned long delalloc_blocks;
4524
4525         inode = dentry->d_inode;
4526         generic_fillattr(inode, stat);
4527
4528         /*
4529          * We can't update i_blocks if the block allocation is delayed
4530          * otherwise in the case of system crash before the real block
4531          * allocation is done, we will have i_blocks inconsistent with
4532          * on-disk file blocks.
4533          * We always keep i_blocks updated together with real
4534          * allocation. But to not confuse with user, stat
4535          * will return the blocks that include the delayed allocation
4536          * blocks for this file.
4537          */
4538         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
4539         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
4540         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
4541
4542         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
4543         return 0;
4544 }
4545
4546 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
4547                                       int chunk)
4548 {
4549         int indirects;
4550
4551         /* if nrblocks are contiguous */
4552         if (chunk) {
4553                 /*
4554                  * With N contiguous data blocks, it need at most
4555                  * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) indirect blocks
4556                  * 2 dindirect blocks
4557                  * 1 tindirect block
4558                  */
4559                 indirects = nrblocks / EXT4_ADDR_PER_BLOCK(inode->i_sb);
4560                 return indirects + 3;
4561         }
4562         /*
4563          * if nrblocks are not contiguous, worse case, each block touch
4564          * a indirect block, and each indirect block touch a double indirect
4565          * block, plus a triple indirect block
4566          */
4567         indirects = nrblocks * 2 + 1;
4568         return indirects;
4569 }
4570
4571 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4572 {
4573         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL))
4574                 return ext4_indirect_trans_blocks(inode, nrblocks, 0);
4575         return ext4_ext_index_trans_blocks(inode, nrblocks, 0);
4576 }
4577 /*
4578  * Account for index blocks, block groups bitmaps and block group
4579  * descriptor blocks if modify datablocks and index blocks
4580  * worse case, the indexs blocks spread over different block groups
4581  *
4582  * If datablocks are discontiguous, they are possible to spread over
4583  * different block groups too. If they are contiugous, with flexbg,
4584  * they could still across block group boundary.
4585  *
4586  * Also account for superblock, inode, quota and xattr blocks
4587  */
4588 int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4589 {
4590         int groups, gdpblocks;
4591         int idxblocks;
4592         int ret = 0;
4593
4594         /*
4595          * How many index blocks need to touch to modify nrblocks?
4596          * The "Chunk" flag indicating whether the nrblocks is
4597          * physically contiguous on disk
4598          *
4599          * For Direct IO and fallocate, they calls get_block to allocate
4600          * one single extent at a time, so they could set the "Chunk" flag
4601          */
4602         idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
4603
4604         ret = idxblocks;
4605
4606         /*
4607          * Now let's see how many group bitmaps and group descriptors need
4608          * to account
4609          */
4610         groups = idxblocks;
4611         if (chunk)
4612                 groups += 1;
4613         else
4614                 groups += nrblocks;
4615
4616         gdpblocks = groups;
4617         if (groups > EXT4_SB(inode->i_sb)->s_groups_count)
4618                 groups = EXT4_SB(inode->i_sb)->s_groups_count;
4619         if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
4620                 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
4621
4622         /* bitmaps and block group descriptor blocks */
4623         ret += groups + gdpblocks;
4624
4625         /* Blocks for super block, inode, quota and xattr blocks */
4626         ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
4627
4628         return ret;
4629 }
4630
4631 /*
4632  * Calulate the total number of credits to reserve to fit
4633  * the modification of a single pages into a single transaction,
4634  * which may include multiple chunks of block allocations.
4635  *
4636  * This could be called via ext4_write_begin()
4637  *
4638  * We need to consider the worse case, when
4639  * one new block per extent.
4640  */
4641 int ext4_writepage_trans_blocks(struct inode *inode)
4642 {
4643         int bpp = ext4_journal_blocks_per_page(inode);
4644         int ret;
4645
4646         ret = ext4_meta_trans_blocks(inode, bpp, 0);
4647
4648         /* Account for data blocks for journalled mode */
4649         if (ext4_should_journal_data(inode))
4650                 ret += bpp;
4651         return ret;
4652 }
4653
4654 /*
4655  * Calculate the journal credits for a chunk of data modification.
4656  *
4657  * This is called from DIO, fallocate or whoever calling
4658  * ext4_get_blocks_wrap() to map/allocate a chunk of contigous disk blocks.
4659  *
4660  * journal buffers for data blocks are not included here, as DIO
4661  * and fallocate do no need to journal data buffers.
4662  */
4663 int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
4664 {
4665         return ext4_meta_trans_blocks(inode, nrblocks, 1);
4666 }
4667
4668 /*
4669  * The caller must have previously called ext4_reserve_inode_write().
4670  * Give this, we know that the caller already has write access to iloc->bh.
4671  */
4672 int ext4_mark_iloc_dirty(handle_t *handle,
4673                 struct inode *inode, struct ext4_iloc *iloc)
4674 {
4675         int err = 0;
4676
4677         if (test_opt(inode->i_sb, I_VERSION))
4678                 inode_inc_iversion(inode);
4679
4680         /* the do_update_inode consumes one bh->b_count */
4681         get_bh(iloc->bh);
4682
4683         /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
4684         err = ext4_do_update_inode(handle, inode, iloc);
4685         put_bh(iloc->bh);
4686         return err;
4687 }
4688
4689 /*
4690  * On success, We end up with an outstanding reference count against
4691  * iloc->bh.  This _must_ be cleaned up later.
4692  */
4693
4694 int
4695 ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
4696                          struct ext4_iloc *iloc)
4697 {
4698         int err = 0;
4699         if (handle) {
4700                 err = ext4_get_inode_loc(inode, iloc);
4701                 if (!err) {
4702                         BUFFER_TRACE(iloc->bh, "get_write_access");
4703                         err = ext4_journal_get_write_access(handle, iloc->bh);
4704                         if (err) {
4705                                 brelse(iloc->bh);
4706                                 iloc->bh = NULL;
4707                         }
4708                 }
4709         }
4710         ext4_std_error(inode->i_sb, err);
4711         return err;
4712 }
4713
4714 /*
4715  * Expand an inode by new_extra_isize bytes.
4716  * Returns 0 on success or negative error number on failure.
4717  */
4718 static int ext4_expand_extra_isize(struct inode *inode,
4719                                    unsigned int new_extra_isize,
4720                                    struct ext4_iloc iloc,
4721                                    handle_t *handle)
4722 {
4723         struct ext4_inode *raw_inode;
4724         struct ext4_xattr_ibody_header *header;
4725         struct ext4_xattr_entry *entry;
4726
4727         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
4728                 return 0;
4729
4730         raw_inode = ext4_raw_inode(&iloc);
4731
4732         header = IHDR(inode, raw_inode);
4733         entry = IFIRST(header);
4734
4735         /* No extended attributes present */
4736         if (!(EXT4_I(inode)->i_state & EXT4_STATE_XATTR) ||
4737                 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
4738                 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
4739                         new_extra_isize);
4740                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
4741                 return 0;
4742         }
4743
4744         /* try to expand with EAs present */
4745         return ext4_expand_extra_isize_ea(inode, new_extra_isize,
4746                                           raw_inode, handle);
4747 }
4748
4749 /*
4750  * What we do here is to mark the in-core inode as clean with respect to inode
4751  * dirtiness (it may still be data-dirty).
4752  * This means that the in-core inode may be reaped by prune_icache
4753  * without having to perform any I/O.  This is a very good thing,
4754  * because *any* task may call prune_icache - even ones which
4755  * have a transaction open against a different journal.
4756  *
4757  * Is this cheating?  Not really.  Sure, we haven't written the
4758  * inode out, but prune_icache isn't a user-visible syncing function.
4759  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
4760  * we start and wait on commits.
4761  *
4762  * Is this efficient/effective?  Well, we're being nice to the system
4763  * by cleaning up our inodes proactively so they can be reaped
4764  * without I/O.  But we are potentially leaving up to five seconds'
4765  * worth of inodes floating about which prune_icache wants us to
4766  * write out.  One way to fix that would be to get prune_icache()
4767  * to do a write_super() to free up some memory.  It has the desired
4768  * effect.
4769  */
4770 int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
4771 {
4772         struct ext4_iloc iloc;
4773         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4774         static unsigned int mnt_count;
4775         int err, ret;
4776
4777         might_sleep();
4778         err = ext4_reserve_inode_write(handle, inode, &iloc);
4779         if (EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
4780             !(EXT4_I(inode)->i_state & EXT4_STATE_NO_EXPAND)) {
4781                 /*
4782                  * We need extra buffer credits since we may write into EA block
4783                  * with this same handle. If journal_extend fails, then it will
4784                  * only result in a minor loss of functionality for that inode.
4785                  * If this is felt to be critical, then e2fsck should be run to
4786                  * force a large enough s_min_extra_isize.
4787                  */
4788                 if ((jbd2_journal_extend(handle,
4789                              EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
4790                         ret = ext4_expand_extra_isize(inode,
4791                                                       sbi->s_want_extra_isize,
4792                                                       iloc, handle);
4793                         if (ret) {
4794                                 EXT4_I(inode)->i_state |= EXT4_STATE_NO_EXPAND;
4795                                 if (mnt_count !=
4796                                         le16_to_cpu(sbi->s_es->s_mnt_count)) {
4797                                         ext4_warning(inode->i_sb, __func__,
4798                                         "Unable to expand inode %lu. Delete"
4799                                         " some EAs or run e2fsck.",
4800                                         inode->i_ino);
4801                                         mnt_count =
4802                                           le16_to_cpu(sbi->s_es->s_mnt_count);
4803                                 }
4804                         }
4805                 }
4806         }
4807         if (!err)
4808                 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
4809         return err;
4810 }
4811
4812 /*
4813  * ext4_dirty_inode() is called from __mark_inode_dirty()
4814  *
4815  * We're really interested in the case where a file is being extended.
4816  * i_size has been changed by generic_commit_write() and we thus need
4817  * to include the updated inode in the current transaction.
4818  *
4819  * Also, DQUOT_ALLOC_SPACE() will always dirty the inode when blocks
4820  * are allocated to the file.
4821  *
4822  * If the inode is marked synchronous, we don't honour that here - doing
4823  * so would cause a commit on atime updates, which we don't bother doing.
4824  * We handle synchronous inodes at the highest possible level.
4825  */
4826 void ext4_dirty_inode(struct inode *inode)
4827 {
4828         handle_t *current_handle = ext4_journal_current_handle();
4829         handle_t *handle;
4830
4831         handle = ext4_journal_start(inode, 2);
4832         if (IS_ERR(handle))
4833                 goto out;
4834         if (current_handle &&
4835                 current_handle->h_transaction != handle->h_transaction) {
4836                 /* This task has a transaction open against a different fs */
4837                 printk(KERN_EMERG "%s: transactions do not match!\n",
4838                        __func__);
4839         } else {
4840                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
4841                                 current_handle);
4842                 ext4_mark_inode_dirty(handle, inode);
4843         }
4844         ext4_journal_stop(handle);
4845 out:
4846         return;
4847 }
4848
4849 #if 0
4850 /*
4851  * Bind an inode's backing buffer_head into this transaction, to prevent
4852  * it from being flushed to disk early.  Unlike
4853  * ext4_reserve_inode_write, this leaves behind no bh reference and
4854  * returns no iloc structure, so the caller needs to repeat the iloc
4855  * lookup to mark the inode dirty later.
4856  */
4857 static int ext4_pin_inode(handle_t *handle, struct inode *inode)
4858 {
4859         struct ext4_iloc iloc;
4860
4861         int err = 0;
4862         if (handle) {
4863                 err = ext4_get_inode_loc(inode, &iloc);
4864                 if (!err) {
4865                         BUFFER_TRACE(iloc.bh, "get_write_access");
4866                         err = jbd2_journal_get_write_access(handle, iloc.bh);
4867                         if (!err)
4868                                 err = ext4_journal_dirty_metadata(handle,
4869                                                                   iloc.bh);
4870                         brelse(iloc.bh);
4871                 }
4872         }
4873         ext4_std_error(inode->i_sb, err);
4874         return err;
4875 }
4876 #endif
4877
4878 int ext4_change_inode_journal_flag(struct inode *inode, int val)
4879 {
4880         journal_t *journal;
4881         handle_t *handle;
4882         int err;
4883
4884         /*
4885          * We have to be very careful here: changing a data block's
4886          * journaling status dynamically is dangerous.  If we write a
4887          * data block to the journal, change the status and then delete
4888          * that block, we risk forgetting to revoke the old log record
4889          * from the journal and so a subsequent replay can corrupt data.
4890          * So, first we make sure that the journal is empty and that
4891          * nobody is changing anything.
4892          */
4893
4894         journal = EXT4_JOURNAL(inode);
4895         if (is_journal_aborted(journal))
4896                 return -EROFS;
4897
4898         jbd2_journal_lock_updates(journal);
4899         jbd2_journal_flush(journal);
4900
4901         /*
4902          * OK, there are no updates running now, and all cached data is
4903          * synced to disk.  We are now in a completely consistent state
4904          * which doesn't have anything in the journal, and we know that
4905          * no filesystem updates are running, so it is safe to modify
4906          * the inode's in-core data-journaling state flag now.
4907          */
4908
4909         if (val)
4910                 EXT4_I(inode)->i_flags |= EXT4_JOURNAL_DATA_FL;
4911         else
4912                 EXT4_I(inode)->i_flags &= ~EXT4_JOURNAL_DATA_FL;
4913         ext4_set_aops(inode);
4914
4915         jbd2_journal_unlock_updates(journal);
4916
4917         /* Finally we can mark the inode as dirty. */
4918
4919         handle = ext4_journal_start(inode, 1);
4920         if (IS_ERR(handle))
4921                 return PTR_ERR(handle);
4922
4923         err = ext4_mark_inode_dirty(handle, inode);
4924         handle->h_sync = 1;
4925         ext4_journal_stop(handle);
4926         ext4_std_error(inode->i_sb, err);
4927
4928         return err;
4929 }
4930
4931 static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
4932 {
4933         return !buffer_mapped(bh);
4934 }
4935
4936 int ext4_page_mkwrite(struct vm_area_struct *vma, struct page *page)
4937 {
4938         loff_t size;
4939         unsigned long len;
4940         int ret = -EINVAL;
4941         void *fsdata;
4942         struct file *file = vma->vm_file;
4943         struct inode *inode = file->f_path.dentry->d_inode;
4944         struct address_space *mapping = inode->i_mapping;
4945
4946         /*
4947          * Get i_alloc_sem to stop truncates messing with the inode. We cannot
4948          * get i_mutex because we are already holding mmap_sem.
4949          */
4950         down_read(&inode->i_alloc_sem);
4951         size = i_size_read(inode);
4952         if (page->mapping != mapping || size <= page_offset(page)
4953             || !PageUptodate(page)) {
4954                 /* page got truncated from under us? */
4955                 goto out_unlock;
4956         }
4957         ret = 0;
4958         if (PageMappedToDisk(page))
4959                 goto out_unlock;
4960
4961         if (page->index == size >> PAGE_CACHE_SHIFT)
4962                 len = size & ~PAGE_CACHE_MASK;
4963         else
4964                 len = PAGE_CACHE_SIZE;
4965
4966         if (page_has_buffers(page)) {
4967                 /* return if we have all the buffers mapped */
4968                 if (!walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
4969                                        ext4_bh_unmapped))
4970                         goto out_unlock;
4971         }
4972         /*
4973          * OK, we need to fill the hole... Do write_begin write_end
4974          * to do block allocation/reservation.We are not holding
4975          * inode.i__mutex here. That allow * parallel write_begin,
4976          * write_end call. lock_page prevent this from happening
4977          * on the same page though
4978          */
4979         ret = mapping->a_ops->write_begin(file, mapping, page_offset(page),
4980                         len, AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
4981         if (ret < 0)
4982                 goto out_unlock;
4983         ret = mapping->a_ops->write_end(file, mapping, page_offset(page),
4984                         len, len, page, fsdata);
4985         if (ret < 0)
4986                 goto out_unlock;
4987         ret = 0;
4988 out_unlock:
4989         up_read(&inode->i_alloc_sem);
4990         return ret;
4991 }