536067bcf75b1c579b958df15c21304fae545488
[linux-2.6.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/jbd2.h>
29 #include <linux/highuid.h>
30 #include <linux/pagemap.h>
31 #include <linux/quotaops.h>
32 #include <linux/string.h>
33 #include <linux/buffer_head.h>
34 #include <linux/writeback.h>
35 #include <linux/pagevec.h>
36 #include <linux/mpage.h>
37 #include <linux/namei.h>
38 #include <linux/uio.h>
39 #include <linux/bio.h>
40 #include <linux/workqueue.h>
41
42 #include "ext4_jbd2.h"
43 #include "xattr.h"
44 #include "acl.h"
45 #include "ext4_extents.h"
46
47 #include <trace/events/ext4.h>
48
49 #define MPAGE_DA_EXTENT_TAIL 0x01
50
51 static inline int ext4_begin_ordered_truncate(struct inode *inode,
52                                               loff_t new_size)
53 {
54         return jbd2_journal_begin_ordered_truncate(
55                                         EXT4_SB(inode->i_sb)->s_journal,
56                                         &EXT4_I(inode)->jinode,
57                                         new_size);
58 }
59
60 static void ext4_invalidatepage(struct page *page, unsigned long offset);
61
62 /*
63  * Test whether an inode is a fast symlink.
64  */
65 static int ext4_inode_is_fast_symlink(struct inode *inode)
66 {
67         int ea_blocks = EXT4_I(inode)->i_file_acl ?
68                 (inode->i_sb->s_blocksize >> 9) : 0;
69
70         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
71 }
72
73 /*
74  * Work out how many blocks we need to proceed with the next chunk of a
75  * truncate transaction.
76  */
77 static unsigned long blocks_for_truncate(struct inode *inode)
78 {
79         ext4_lblk_t needed;
80
81         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
82
83         /* Give ourselves just enough room to cope with inodes in which
84          * i_blocks is corrupt: we've seen disk corruptions in the past
85          * which resulted in random data in an inode which looked enough
86          * like a regular file for ext4 to try to delete it.  Things
87          * will go a bit crazy if that happens, but at least we should
88          * try not to panic the whole kernel. */
89         if (needed < 2)
90                 needed = 2;
91
92         /* But we need to bound the transaction so we don't overflow the
93          * journal. */
94         if (needed > EXT4_MAX_TRANS_DATA)
95                 needed = EXT4_MAX_TRANS_DATA;
96
97         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
98 }
99
100 /*
101  * Truncate transactions can be complex and absolutely huge.  So we need to
102  * be able to restart the transaction at a conventient checkpoint to make
103  * sure we don't overflow the journal.
104  *
105  * start_transaction gets us a new handle for a truncate transaction,
106  * and extend_transaction tries to extend the existing one a bit.  If
107  * extend fails, we need to propagate the failure up and restart the
108  * transaction in the top-level truncate loop. --sct
109  */
110 static handle_t *start_transaction(struct inode *inode)
111 {
112         handle_t *result;
113
114         result = ext4_journal_start(inode, blocks_for_truncate(inode));
115         if (!IS_ERR(result))
116                 return result;
117
118         ext4_std_error(inode->i_sb, PTR_ERR(result));
119         return result;
120 }
121
122 /*
123  * Try to extend this transaction for the purposes of truncation.
124  *
125  * Returns 0 if we managed to create more room.  If we can't create more
126  * room, and the transaction must be restarted we return 1.
127  */
128 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
129 {
130         if (!ext4_handle_valid(handle))
131                 return 0;
132         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
133                 return 0;
134         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
135                 return 0;
136         return 1;
137 }
138
139 /*
140  * Restart the transaction associated with *handle.  This does a commit,
141  * so before we call here everything must be consistently dirtied against
142  * this transaction.
143  */
144 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
145                                  int nblocks)
146 {
147         int ret;
148
149         /*
150          * Drop i_data_sem to avoid deadlock with ext4_get_blocks At this
151          * moment, get_block can be called only for blocks inside i_size since
152          * page cache has been already dropped and writes are blocked by
153          * i_mutex. So we can safely drop the i_data_sem here.
154          */
155         BUG_ON(EXT4_JOURNAL(inode) == NULL);
156         jbd_debug(2, "restarting handle %p\n", handle);
157         up_write(&EXT4_I(inode)->i_data_sem);
158         ret = ext4_journal_restart(handle, blocks_for_truncate(inode));
159         down_write(&EXT4_I(inode)->i_data_sem);
160         ext4_discard_preallocations(inode);
161
162         return ret;
163 }
164
165 /*
166  * Called at the last iput() if i_nlink is zero.
167  */
168 void ext4_delete_inode(struct inode *inode)
169 {
170         handle_t *handle;
171         int err;
172
173         if (ext4_should_order_data(inode))
174                 ext4_begin_ordered_truncate(inode, 0);
175         truncate_inode_pages(&inode->i_data, 0);
176
177         if (is_bad_inode(inode))
178                 goto no_delete;
179
180         handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
181         if (IS_ERR(handle)) {
182                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
183                 /*
184                  * If we're going to skip the normal cleanup, we still need to
185                  * make sure that the in-core orphan linked list is properly
186                  * cleaned up.
187                  */
188                 ext4_orphan_del(NULL, inode);
189                 goto no_delete;
190         }
191
192         if (IS_SYNC(inode))
193                 ext4_handle_sync(handle);
194         inode->i_size = 0;
195         err = ext4_mark_inode_dirty(handle, inode);
196         if (err) {
197                 ext4_warning(inode->i_sb,
198                              "couldn't mark inode dirty (err %d)", err);
199                 goto stop_handle;
200         }
201         if (inode->i_blocks)
202                 ext4_truncate(inode);
203
204         /*
205          * ext4_ext_truncate() doesn't reserve any slop when it
206          * restarts journal transactions; therefore there may not be
207          * enough credits left in the handle to remove the inode from
208          * the orphan list and set the dtime field.
209          */
210         if (!ext4_handle_has_enough_credits(handle, 3)) {
211                 err = ext4_journal_extend(handle, 3);
212                 if (err > 0)
213                         err = ext4_journal_restart(handle, 3);
214                 if (err != 0) {
215                         ext4_warning(inode->i_sb,
216                                      "couldn't extend journal (err %d)", err);
217                 stop_handle:
218                         ext4_journal_stop(handle);
219                         goto no_delete;
220                 }
221         }
222
223         /*
224          * Kill off the orphan record which ext4_truncate created.
225          * AKPM: I think this can be inside the above `if'.
226          * Note that ext4_orphan_del() has to be able to cope with the
227          * deletion of a non-existent orphan - this is because we don't
228          * know if ext4_truncate() actually created an orphan record.
229          * (Well, we could do this if we need to, but heck - it works)
230          */
231         ext4_orphan_del(handle, inode);
232         EXT4_I(inode)->i_dtime  = get_seconds();
233
234         /*
235          * One subtle ordering requirement: if anything has gone wrong
236          * (transaction abort, IO errors, whatever), then we can still
237          * do these next steps (the fs will already have been marked as
238          * having errors), but we can't free the inode if the mark_dirty
239          * fails.
240          */
241         if (ext4_mark_inode_dirty(handle, inode))
242                 /* If that failed, just do the required in-core inode clear. */
243                 clear_inode(inode);
244         else
245                 ext4_free_inode(handle, inode);
246         ext4_journal_stop(handle);
247         return;
248 no_delete:
249         clear_inode(inode);     /* We must guarantee clearing of inode... */
250 }
251
252 typedef struct {
253         __le32  *p;
254         __le32  key;
255         struct buffer_head *bh;
256 } Indirect;
257
258 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
259 {
260         p->key = *(p->p = v);
261         p->bh = bh;
262 }
263
264 /**
265  *      ext4_block_to_path - parse the block number into array of offsets
266  *      @inode: inode in question (we are only interested in its superblock)
267  *      @i_block: block number to be parsed
268  *      @offsets: array to store the offsets in
269  *      @boundary: set this non-zero if the referred-to block is likely to be
270  *             followed (on disk) by an indirect block.
271  *
272  *      To store the locations of file's data ext4 uses a data structure common
273  *      for UNIX filesystems - tree of pointers anchored in the inode, with
274  *      data blocks at leaves and indirect blocks in intermediate nodes.
275  *      This function translates the block number into path in that tree -
276  *      return value is the path length and @offsets[n] is the offset of
277  *      pointer to (n+1)th node in the nth one. If @block is out of range
278  *      (negative or too large) warning is printed and zero returned.
279  *
280  *      Note: function doesn't find node addresses, so no IO is needed. All
281  *      we need to know is the capacity of indirect blocks (taken from the
282  *      inode->i_sb).
283  */
284
285 /*
286  * Portability note: the last comparison (check that we fit into triple
287  * indirect block) is spelled differently, because otherwise on an
288  * architecture with 32-bit longs and 8Kb pages we might get into trouble
289  * if our filesystem had 8Kb blocks. We might use long long, but that would
290  * kill us on x86. Oh, well, at least the sign propagation does not matter -
291  * i_block would have to be negative in the very beginning, so we would not
292  * get there at all.
293  */
294
295 static int ext4_block_to_path(struct inode *inode,
296                               ext4_lblk_t i_block,
297                               ext4_lblk_t offsets[4], int *boundary)
298 {
299         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
300         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
301         const long direct_blocks = EXT4_NDIR_BLOCKS,
302                 indirect_blocks = ptrs,
303                 double_blocks = (1 << (ptrs_bits * 2));
304         int n = 0;
305         int final = 0;
306
307         if (i_block < direct_blocks) {
308                 offsets[n++] = i_block;
309                 final = direct_blocks;
310         } else if ((i_block -= direct_blocks) < indirect_blocks) {
311                 offsets[n++] = EXT4_IND_BLOCK;
312                 offsets[n++] = i_block;
313                 final = ptrs;
314         } else if ((i_block -= indirect_blocks) < double_blocks) {
315                 offsets[n++] = EXT4_DIND_BLOCK;
316                 offsets[n++] = i_block >> ptrs_bits;
317                 offsets[n++] = i_block & (ptrs - 1);
318                 final = ptrs;
319         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
320                 offsets[n++] = EXT4_TIND_BLOCK;
321                 offsets[n++] = i_block >> (ptrs_bits * 2);
322                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
323                 offsets[n++] = i_block & (ptrs - 1);
324                 final = ptrs;
325         } else {
326                 ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
327                              i_block + direct_blocks +
328                              indirect_blocks + double_blocks, inode->i_ino);
329         }
330         if (boundary)
331                 *boundary = final - 1 - (i_block & (ptrs - 1));
332         return n;
333 }
334
335 static int __ext4_check_blockref(const char *function, struct inode *inode,
336                                  __le32 *p, unsigned int max)
337 {
338         __le32 *bref = p;
339         unsigned int blk;
340
341         while (bref < p+max) {
342                 blk = le32_to_cpu(*bref++);
343                 if (blk &&
344                     unlikely(!ext4_data_block_valid(EXT4_SB(inode->i_sb),
345                                                     blk, 1))) {
346                         __ext4_error(inode->i_sb, function,
347                                    "invalid block reference %u "
348                                    "in inode #%lu", blk, inode->i_ino);
349                         return -EIO;
350                 }
351         }
352         return 0;
353 }
354
355
356 #define ext4_check_indirect_blockref(inode, bh)                         \
357         __ext4_check_blockref(__func__, inode, (__le32 *)(bh)->b_data,  \
358                               EXT4_ADDR_PER_BLOCK((inode)->i_sb))
359
360 #define ext4_check_inode_blockref(inode)                                \
361         __ext4_check_blockref(__func__, inode, EXT4_I(inode)->i_data,   \
362                               EXT4_NDIR_BLOCKS)
363
364 /**
365  *      ext4_get_branch - read the chain of indirect blocks leading to data
366  *      @inode: inode in question
367  *      @depth: depth of the chain (1 - direct pointer, etc.)
368  *      @offsets: offsets of pointers in inode/indirect blocks
369  *      @chain: place to store the result
370  *      @err: here we store the error value
371  *
372  *      Function fills the array of triples <key, p, bh> and returns %NULL
373  *      if everything went OK or the pointer to the last filled triple
374  *      (incomplete one) otherwise. Upon the return chain[i].key contains
375  *      the number of (i+1)-th block in the chain (as it is stored in memory,
376  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
377  *      number (it points into struct inode for i==0 and into the bh->b_data
378  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
379  *      block for i>0 and NULL for i==0. In other words, it holds the block
380  *      numbers of the chain, addresses they were taken from (and where we can
381  *      verify that chain did not change) and buffer_heads hosting these
382  *      numbers.
383  *
384  *      Function stops when it stumbles upon zero pointer (absent block)
385  *              (pointer to last triple returned, *@err == 0)
386  *      or when it gets an IO error reading an indirect block
387  *              (ditto, *@err == -EIO)
388  *      or when it reads all @depth-1 indirect blocks successfully and finds
389  *      the whole chain, all way to the data (returns %NULL, *err == 0).
390  *
391  *      Need to be called with
392  *      down_read(&EXT4_I(inode)->i_data_sem)
393  */
394 static Indirect *ext4_get_branch(struct inode *inode, int depth,
395                                  ext4_lblk_t  *offsets,
396                                  Indirect chain[4], int *err)
397 {
398         struct super_block *sb = inode->i_sb;
399         Indirect *p = chain;
400         struct buffer_head *bh;
401
402         *err = 0;
403         /* i_data is not going away, no lock needed */
404         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
405         if (!p->key)
406                 goto no_block;
407         while (--depth) {
408                 bh = sb_getblk(sb, le32_to_cpu(p->key));
409                 if (unlikely(!bh))
410                         goto failure;
411
412                 if (!bh_uptodate_or_lock(bh)) {
413                         if (bh_submit_read(bh) < 0) {
414                                 put_bh(bh);
415                                 goto failure;
416                         }
417                         /* validate block references */
418                         if (ext4_check_indirect_blockref(inode, bh)) {
419                                 put_bh(bh);
420                                 goto failure;
421                         }
422                 }
423
424                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
425                 /* Reader: end */
426                 if (!p->key)
427                         goto no_block;
428         }
429         return NULL;
430
431 failure:
432         *err = -EIO;
433 no_block:
434         return p;
435 }
436
437 /**
438  *      ext4_find_near - find a place for allocation with sufficient locality
439  *      @inode: owner
440  *      @ind: descriptor of indirect block.
441  *
442  *      This function returns the preferred place for block allocation.
443  *      It is used when heuristic for sequential allocation fails.
444  *      Rules are:
445  *        + if there is a block to the left of our position - allocate near it.
446  *        + if pointer will live in indirect block - allocate near that block.
447  *        + if pointer will live in inode - allocate in the same
448  *          cylinder group.
449  *
450  * In the latter case we colour the starting block by the callers PID to
451  * prevent it from clashing with concurrent allocations for a different inode
452  * in the same block group.   The PID is used here so that functionally related
453  * files will be close-by on-disk.
454  *
455  *      Caller must make sure that @ind is valid and will stay that way.
456  */
457 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
458 {
459         struct ext4_inode_info *ei = EXT4_I(inode);
460         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
461         __le32 *p;
462         ext4_fsblk_t bg_start;
463         ext4_fsblk_t last_block;
464         ext4_grpblk_t colour;
465         ext4_group_t block_group;
466         int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
467
468         /* Try to find previous block */
469         for (p = ind->p - 1; p >= start; p--) {
470                 if (*p)
471                         return le32_to_cpu(*p);
472         }
473
474         /* No such thing, so let's try location of indirect block */
475         if (ind->bh)
476                 return ind->bh->b_blocknr;
477
478         /*
479          * It is going to be referred to from the inode itself? OK, just put it
480          * into the same cylinder group then.
481          */
482         block_group = ei->i_block_group;
483         if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
484                 block_group &= ~(flex_size-1);
485                 if (S_ISREG(inode->i_mode))
486                         block_group++;
487         }
488         bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
489         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
490
491         /*
492          * If we are doing delayed allocation, we don't need take
493          * colour into account.
494          */
495         if (test_opt(inode->i_sb, DELALLOC))
496                 return bg_start;
497
498         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
499                 colour = (current->pid % 16) *
500                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
501         else
502                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
503         return bg_start + colour;
504 }
505
506 /**
507  *      ext4_find_goal - find a preferred place for allocation.
508  *      @inode: owner
509  *      @block:  block we want
510  *      @partial: pointer to the last triple within a chain
511  *
512  *      Normally this function find the preferred place for block allocation,
513  *      returns it.
514  *      Because this is only used for non-extent files, we limit the block nr
515  *      to 32 bits.
516  */
517 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
518                                    Indirect *partial)
519 {
520         ext4_fsblk_t goal;
521
522         /*
523          * XXX need to get goal block from mballoc's data structures
524          */
525
526         goal = ext4_find_near(inode, partial);
527         goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
528         return goal;
529 }
530
531 /**
532  *      ext4_blks_to_allocate: Look up the block map and count the number
533  *      of direct blocks need to be allocated for the given branch.
534  *
535  *      @branch: chain of indirect blocks
536  *      @k: number of blocks need for indirect blocks
537  *      @blks: number of data blocks to be mapped.
538  *      @blocks_to_boundary:  the offset in the indirect block
539  *
540  *      return the total number of blocks to be allocate, including the
541  *      direct and indirect blocks.
542  */
543 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
544                                  int blocks_to_boundary)
545 {
546         unsigned int count = 0;
547
548         /*
549          * Simple case, [t,d]Indirect block(s) has not allocated yet
550          * then it's clear blocks on that path have not allocated
551          */
552         if (k > 0) {
553                 /* right now we don't handle cross boundary allocation */
554                 if (blks < blocks_to_boundary + 1)
555                         count += blks;
556                 else
557                         count += blocks_to_boundary + 1;
558                 return count;
559         }
560
561         count++;
562         while (count < blks && count <= blocks_to_boundary &&
563                 le32_to_cpu(*(branch[0].p + count)) == 0) {
564                 count++;
565         }
566         return count;
567 }
568
569 /**
570  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
571  *      @indirect_blks: the number of blocks need to allocate for indirect
572  *                      blocks
573  *
574  *      @new_blocks: on return it will store the new block numbers for
575  *      the indirect blocks(if needed) and the first direct block,
576  *      @blks:  on return it will store the total number of allocated
577  *              direct blocks
578  */
579 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
580                              ext4_lblk_t iblock, ext4_fsblk_t goal,
581                              int indirect_blks, int blks,
582                              ext4_fsblk_t new_blocks[4], int *err)
583 {
584         struct ext4_allocation_request ar;
585         int target, i;
586         unsigned long count = 0, blk_allocated = 0;
587         int index = 0;
588         ext4_fsblk_t current_block = 0;
589         int ret = 0;
590
591         /*
592          * Here we try to allocate the requested multiple blocks at once,
593          * on a best-effort basis.
594          * To build a branch, we should allocate blocks for
595          * the indirect blocks(if not allocated yet), and at least
596          * the first direct block of this branch.  That's the
597          * minimum number of blocks need to allocate(required)
598          */
599         /* first we try to allocate the indirect blocks */
600         target = indirect_blks;
601         while (target > 0) {
602                 count = target;
603                 /* allocating blocks for indirect blocks and direct blocks */
604                 current_block = ext4_new_meta_blocks(handle, inode,
605                                                         goal, &count, err);
606                 if (*err)
607                         goto failed_out;
608
609                 BUG_ON(current_block + count > EXT4_MAX_BLOCK_FILE_PHYS);
610
611                 target -= count;
612                 /* allocate blocks for indirect blocks */
613                 while (index < indirect_blks && count) {
614                         new_blocks[index++] = current_block++;
615                         count--;
616                 }
617                 if (count > 0) {
618                         /*
619                          * save the new block number
620                          * for the first direct block
621                          */
622                         new_blocks[index] = current_block;
623                         printk(KERN_INFO "%s returned more blocks than "
624                                                 "requested\n", __func__);
625                         WARN_ON(1);
626                         break;
627                 }
628         }
629
630         target = blks - count ;
631         blk_allocated = count;
632         if (!target)
633                 goto allocated;
634         /* Now allocate data blocks */
635         memset(&ar, 0, sizeof(ar));
636         ar.inode = inode;
637         ar.goal = goal;
638         ar.len = target;
639         ar.logical = iblock;
640         if (S_ISREG(inode->i_mode))
641                 /* enable in-core preallocation only for regular files */
642                 ar.flags = EXT4_MB_HINT_DATA;
643
644         current_block = ext4_mb_new_blocks(handle, &ar, err);
645         BUG_ON(current_block + ar.len > EXT4_MAX_BLOCK_FILE_PHYS);
646
647         if (*err && (target == blks)) {
648                 /*
649                  * if the allocation failed and we didn't allocate
650                  * any blocks before
651                  */
652                 goto failed_out;
653         }
654         if (!*err) {
655                 if (target == blks) {
656                         /*
657                          * save the new block number
658                          * for the first direct block
659                          */
660                         new_blocks[index] = current_block;
661                 }
662                 blk_allocated += ar.len;
663         }
664 allocated:
665         /* total number of blocks allocated for direct blocks */
666         ret = blk_allocated;
667         *err = 0;
668         return ret;
669 failed_out:
670         for (i = 0; i < index; i++)
671                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
672         return ret;
673 }
674
675 /**
676  *      ext4_alloc_branch - allocate and set up a chain of blocks.
677  *      @inode: owner
678  *      @indirect_blks: number of allocated indirect blocks
679  *      @blks: number of allocated direct blocks
680  *      @offsets: offsets (in the blocks) to store the pointers to next.
681  *      @branch: place to store the chain in.
682  *
683  *      This function allocates blocks, zeroes out all but the last one,
684  *      links them into chain and (if we are synchronous) writes them to disk.
685  *      In other words, it prepares a branch that can be spliced onto the
686  *      inode. It stores the information about that chain in the branch[], in
687  *      the same format as ext4_get_branch() would do. We are calling it after
688  *      we had read the existing part of chain and partial points to the last
689  *      triple of that (one with zero ->key). Upon the exit we have the same
690  *      picture as after the successful ext4_get_block(), except that in one
691  *      place chain is disconnected - *branch->p is still zero (we did not
692  *      set the last link), but branch->key contains the number that should
693  *      be placed into *branch->p to fill that gap.
694  *
695  *      If allocation fails we free all blocks we've allocated (and forget
696  *      their buffer_heads) and return the error value the from failed
697  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
698  *      as described above and return 0.
699  */
700 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
701                              ext4_lblk_t iblock, int indirect_blks,
702                              int *blks, ext4_fsblk_t goal,
703                              ext4_lblk_t *offsets, Indirect *branch)
704 {
705         int blocksize = inode->i_sb->s_blocksize;
706         int i, n = 0;
707         int err = 0;
708         struct buffer_head *bh;
709         int num;
710         ext4_fsblk_t new_blocks[4];
711         ext4_fsblk_t current_block;
712
713         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
714                                 *blks, new_blocks, &err);
715         if (err)
716                 return err;
717
718         branch[0].key = cpu_to_le32(new_blocks[0]);
719         /*
720          * metadata blocks and data blocks are allocated.
721          */
722         for (n = 1; n <= indirect_blks;  n++) {
723                 /*
724                  * Get buffer_head for parent block, zero it out
725                  * and set the pointer to new one, then send
726                  * parent to disk.
727                  */
728                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
729                 branch[n].bh = bh;
730                 lock_buffer(bh);
731                 BUFFER_TRACE(bh, "call get_create_access");
732                 err = ext4_journal_get_create_access(handle, bh);
733                 if (err) {
734                         /* Don't brelse(bh) here; it's done in
735                          * ext4_journal_forget() below */
736                         unlock_buffer(bh);
737                         goto failed;
738                 }
739
740                 memset(bh->b_data, 0, blocksize);
741                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
742                 branch[n].key = cpu_to_le32(new_blocks[n]);
743                 *branch[n].p = branch[n].key;
744                 if (n == indirect_blks) {
745                         current_block = new_blocks[n];
746                         /*
747                          * End of chain, update the last new metablock of
748                          * the chain to point to the new allocated
749                          * data blocks numbers
750                          */
751                         for (i = 1; i < num; i++)
752                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
753                 }
754                 BUFFER_TRACE(bh, "marking uptodate");
755                 set_buffer_uptodate(bh);
756                 unlock_buffer(bh);
757
758                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
759                 err = ext4_handle_dirty_metadata(handle, inode, bh);
760                 if (err)
761                         goto failed;
762         }
763         *blks = num;
764         return err;
765 failed:
766         /* Allocation failed, free what we already allocated */
767         ext4_free_blocks(handle, inode, 0, new_blocks[0], 1, 0);
768         for (i = 1; i <= n ; i++) {
769                 /* 
770                  * branch[i].bh is newly allocated, so there is no
771                  * need to revoke the block, which is why we don't
772                  * need to set EXT4_FREE_BLOCKS_METADATA.
773                  */
774                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1,
775                                  EXT4_FREE_BLOCKS_FORGET);
776         }
777         for (i = n+1; i < indirect_blks; i++)
778                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
779
780         ext4_free_blocks(handle, inode, 0, new_blocks[i], num, 0);
781
782         return err;
783 }
784
785 /**
786  * ext4_splice_branch - splice the allocated branch onto inode.
787  * @inode: owner
788  * @block: (logical) number of block we are adding
789  * @chain: chain of indirect blocks (with a missing link - see
790  *      ext4_alloc_branch)
791  * @where: location of missing link
792  * @num:   number of indirect blocks we are adding
793  * @blks:  number of direct blocks we are adding
794  *
795  * This function fills the missing link and does all housekeeping needed in
796  * inode (->i_blocks, etc.). In case of success we end up with the full
797  * chain to new block and return 0.
798  */
799 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
800                               ext4_lblk_t block, Indirect *where, int num,
801                               int blks)
802 {
803         int i;
804         int err = 0;
805         ext4_fsblk_t current_block;
806
807         /*
808          * If we're splicing into a [td]indirect block (as opposed to the
809          * inode) then we need to get write access to the [td]indirect block
810          * before the splice.
811          */
812         if (where->bh) {
813                 BUFFER_TRACE(where->bh, "get_write_access");
814                 err = ext4_journal_get_write_access(handle, where->bh);
815                 if (err)
816                         goto err_out;
817         }
818         /* That's it */
819
820         *where->p = where->key;
821
822         /*
823          * Update the host buffer_head or inode to point to more just allocated
824          * direct blocks blocks
825          */
826         if (num == 0 && blks > 1) {
827                 current_block = le32_to_cpu(where->key) + 1;
828                 for (i = 1; i < blks; i++)
829                         *(where->p + i) = cpu_to_le32(current_block++);
830         }
831
832         /* We are done with atomic stuff, now do the rest of housekeeping */
833         /* had we spliced it onto indirect block? */
834         if (where->bh) {
835                 /*
836                  * If we spliced it onto an indirect block, we haven't
837                  * altered the inode.  Note however that if it is being spliced
838                  * onto an indirect block at the very end of the file (the
839                  * file is growing) then we *will* alter the inode to reflect
840                  * the new i_size.  But that is not done here - it is done in
841                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
842                  */
843                 jbd_debug(5, "splicing indirect only\n");
844                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
845                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
846                 if (err)
847                         goto err_out;
848         } else {
849                 /*
850                  * OK, we spliced it into the inode itself on a direct block.
851                  */
852                 ext4_mark_inode_dirty(handle, inode);
853                 jbd_debug(5, "splicing direct\n");
854         }
855         return err;
856
857 err_out:
858         for (i = 1; i <= num; i++) {
859                 /* 
860                  * branch[i].bh is newly allocated, so there is no
861                  * need to revoke the block, which is why we don't
862                  * need to set EXT4_FREE_BLOCKS_METADATA.
863                  */
864                 ext4_free_blocks(handle, inode, where[i].bh, 0, 1,
865                                  EXT4_FREE_BLOCKS_FORGET);
866         }
867         ext4_free_blocks(handle, inode, 0, le32_to_cpu(where[num].key),
868                          blks, 0);
869
870         return err;
871 }
872
873 /*
874  * The ext4_ind_get_blocks() function handles non-extents inodes
875  * (i.e., using the traditional indirect/double-indirect i_blocks
876  * scheme) for ext4_get_blocks().
877  *
878  * Allocation strategy is simple: if we have to allocate something, we will
879  * have to go the whole way to leaf. So let's do it before attaching anything
880  * to tree, set linkage between the newborn blocks, write them if sync is
881  * required, recheck the path, free and repeat if check fails, otherwise
882  * set the last missing link (that will protect us from any truncate-generated
883  * removals - all blocks on the path are immune now) and possibly force the
884  * write on the parent block.
885  * That has a nice additional property: no special recovery from the failed
886  * allocations is needed - we simply release blocks and do not touch anything
887  * reachable from inode.
888  *
889  * `handle' can be NULL if create == 0.
890  *
891  * return > 0, # of blocks mapped or allocated.
892  * return = 0, if plain lookup failed.
893  * return < 0, error case.
894  *
895  * The ext4_ind_get_blocks() function should be called with
896  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
897  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
898  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
899  * blocks.
900  */
901 static int ext4_ind_get_blocks(handle_t *handle, struct inode *inode,
902                                ext4_lblk_t iblock, unsigned int maxblocks,
903                                struct buffer_head *bh_result,
904                                int flags)
905 {
906         int err = -EIO;
907         ext4_lblk_t offsets[4];
908         Indirect chain[4];
909         Indirect *partial;
910         ext4_fsblk_t goal;
911         int indirect_blks;
912         int blocks_to_boundary = 0;
913         int depth;
914         int count = 0;
915         ext4_fsblk_t first_block = 0;
916
917         J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
918         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
919         depth = ext4_block_to_path(inode, iblock, offsets,
920                                    &blocks_to_boundary);
921
922         if (depth == 0)
923                 goto out;
924
925         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
926
927         /* Simplest case - block found, no allocation needed */
928         if (!partial) {
929                 first_block = le32_to_cpu(chain[depth - 1].key);
930                 clear_buffer_new(bh_result);
931                 count++;
932                 /*map more blocks*/
933                 while (count < maxblocks && count <= blocks_to_boundary) {
934                         ext4_fsblk_t blk;
935
936                         blk = le32_to_cpu(*(chain[depth-1].p + count));
937
938                         if (blk == first_block + count)
939                                 count++;
940                         else
941                                 break;
942                 }
943                 goto got_it;
944         }
945
946         /* Next simple case - plain lookup or failed read of indirect block */
947         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
948                 goto cleanup;
949
950         /*
951          * Okay, we need to do block allocation.
952         */
953         goal = ext4_find_goal(inode, iblock, partial);
954
955         /* the number of blocks need to allocate for [d,t]indirect blocks */
956         indirect_blks = (chain + depth) - partial - 1;
957
958         /*
959          * Next look up the indirect map to count the totoal number of
960          * direct blocks to allocate for this branch.
961          */
962         count = ext4_blks_to_allocate(partial, indirect_blks,
963                                         maxblocks, blocks_to_boundary);
964         /*
965          * Block out ext4_truncate while we alter the tree
966          */
967         err = ext4_alloc_branch(handle, inode, iblock, indirect_blks,
968                                 &count, goal,
969                                 offsets + (partial - chain), partial);
970
971         /*
972          * The ext4_splice_branch call will free and forget any buffers
973          * on the new chain if there is a failure, but that risks using
974          * up transaction credits, especially for bitmaps where the
975          * credits cannot be returned.  Can we handle this somehow?  We
976          * may need to return -EAGAIN upwards in the worst case.  --sct
977          */
978         if (!err)
979                 err = ext4_splice_branch(handle, inode, iblock,
980                                          partial, indirect_blks, count);
981         if (err)
982                 goto cleanup;
983
984         set_buffer_new(bh_result);
985
986         ext4_update_inode_fsync_trans(handle, inode, 1);
987 got_it:
988         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
989         if (count > blocks_to_boundary)
990                 set_buffer_boundary(bh_result);
991         err = count;
992         /* Clean up and exit */
993         partial = chain + depth - 1;    /* the whole chain */
994 cleanup:
995         while (partial > chain) {
996                 BUFFER_TRACE(partial->bh, "call brelse");
997                 brelse(partial->bh);
998                 partial--;
999         }
1000         BUFFER_TRACE(bh_result, "returned");
1001 out:
1002         return err;
1003 }
1004
1005 #ifdef CONFIG_QUOTA
1006 qsize_t *ext4_get_reserved_space(struct inode *inode)
1007 {
1008         return &EXT4_I(inode)->i_reserved_quota;
1009 }
1010 #endif
1011
1012 /*
1013  * Calculate the number of metadata blocks need to reserve
1014  * to allocate a new block at @lblocks for non extent file based file
1015  */
1016 static int ext4_indirect_calc_metadata_amount(struct inode *inode,
1017                                               sector_t lblock)
1018 {
1019         struct ext4_inode_info *ei = EXT4_I(inode);
1020         int dind_mask = EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1;
1021         int blk_bits;
1022
1023         if (lblock < EXT4_NDIR_BLOCKS)
1024                 return 0;
1025
1026         lblock -= EXT4_NDIR_BLOCKS;
1027
1028         if (ei->i_da_metadata_calc_len &&
1029             (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
1030                 ei->i_da_metadata_calc_len++;
1031                 return 0;
1032         }
1033         ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
1034         ei->i_da_metadata_calc_len = 1;
1035         blk_bits = roundup_pow_of_two(lblock + 1);
1036         return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
1037 }
1038
1039 /*
1040  * Calculate the number of metadata blocks need to reserve
1041  * to allocate a block located at @lblock
1042  */
1043 static int ext4_calc_metadata_amount(struct inode *inode, sector_t lblock)
1044 {
1045         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
1046                 return ext4_ext_calc_metadata_amount(inode, lblock);
1047
1048         return ext4_indirect_calc_metadata_amount(inode, lblock);
1049 }
1050
1051 /*
1052  * Called with i_data_sem down, which is important since we can call
1053  * ext4_discard_preallocations() from here.
1054  */
1055 void ext4_da_update_reserve_space(struct inode *inode,
1056                                         int used, int quota_claim)
1057 {
1058         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1059         struct ext4_inode_info *ei = EXT4_I(inode);
1060         int mdb_free = 0, allocated_meta_blocks = 0;
1061
1062         spin_lock(&ei->i_block_reservation_lock);
1063         trace_ext4_da_update_reserve_space(inode, used);
1064         if (unlikely(used > ei->i_reserved_data_blocks)) {
1065                 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
1066                          "with only %d reserved data blocks\n",
1067                          __func__, inode->i_ino, used,
1068                          ei->i_reserved_data_blocks);
1069                 WARN_ON(1);
1070                 used = ei->i_reserved_data_blocks;
1071         }
1072
1073         /* Update per-inode reservations */
1074         ei->i_reserved_data_blocks -= used;
1075         used += ei->i_allocated_meta_blocks;
1076         ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
1077         allocated_meta_blocks = ei->i_allocated_meta_blocks;
1078         ei->i_allocated_meta_blocks = 0;
1079         percpu_counter_sub(&sbi->s_dirtyblocks_counter, used);
1080
1081         if (ei->i_reserved_data_blocks == 0) {
1082                 /*
1083                  * We can release all of the reserved metadata blocks
1084                  * only when we have written all of the delayed
1085                  * allocation blocks.
1086                  */
1087                 mdb_free = ei->i_reserved_meta_blocks;
1088                 ei->i_reserved_meta_blocks = 0;
1089                 ei->i_da_metadata_calc_len = 0;
1090                 percpu_counter_sub(&sbi->s_dirtyblocks_counter, mdb_free);
1091         }
1092         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1093
1094         /* Update quota subsystem */
1095         if (quota_claim) {
1096                 vfs_dq_claim_block(inode, used);
1097                 if (mdb_free)
1098                         vfs_dq_release_reservation_block(inode, mdb_free);
1099         } else {
1100                 /*
1101                  * We did fallocate with an offset that is already delayed
1102                  * allocated. So on delayed allocated writeback we should
1103                  * not update the quota for allocated blocks. But then
1104                  * converting an fallocate region to initialized region would
1105                  * have caused a metadata allocation. So claim quota for
1106                  * that
1107                  */
1108                 if (allocated_meta_blocks)
1109                         vfs_dq_claim_block(inode, allocated_meta_blocks);
1110                 vfs_dq_release_reservation_block(inode, mdb_free + used);
1111         }
1112
1113         /*
1114          * If we have done all the pending block allocations and if
1115          * there aren't any writers on the inode, we can discard the
1116          * inode's preallocations.
1117          */
1118         if ((ei->i_reserved_data_blocks == 0) &&
1119             (atomic_read(&inode->i_writecount) == 0))
1120                 ext4_discard_preallocations(inode);
1121 }
1122
1123 static int check_block_validity(struct inode *inode, const char *msg,
1124                                 sector_t logical, sector_t phys, int len)
1125 {
1126         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), phys, len)) {
1127                 __ext4_error(inode->i_sb, msg,
1128                            "inode #%lu logical block %llu mapped to %llu "
1129                            "(size %d)", inode->i_ino,
1130                            (unsigned long long) logical,
1131                            (unsigned long long) phys, len);
1132                 return -EIO;
1133         }
1134         return 0;
1135 }
1136
1137 /*
1138  * Return the number of contiguous dirty pages in a given inode
1139  * starting at page frame idx.
1140  */
1141 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
1142                                     unsigned int max_pages)
1143 {
1144         struct address_space *mapping = inode->i_mapping;
1145         pgoff_t index;
1146         struct pagevec pvec;
1147         pgoff_t num = 0;
1148         int i, nr_pages, done = 0;
1149
1150         if (max_pages == 0)
1151                 return 0;
1152         pagevec_init(&pvec, 0);
1153         while (!done) {
1154                 index = idx;
1155                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
1156                                               PAGECACHE_TAG_DIRTY,
1157                                               (pgoff_t)PAGEVEC_SIZE);
1158                 if (nr_pages == 0)
1159                         break;
1160                 for (i = 0; i < nr_pages; i++) {
1161                         struct page *page = pvec.pages[i];
1162                         struct buffer_head *bh, *head;
1163
1164                         lock_page(page);
1165                         if (unlikely(page->mapping != mapping) ||
1166                             !PageDirty(page) ||
1167                             PageWriteback(page) ||
1168                             page->index != idx) {
1169                                 done = 1;
1170                                 unlock_page(page);
1171                                 break;
1172                         }
1173                         if (page_has_buffers(page)) {
1174                                 bh = head = page_buffers(page);
1175                                 do {
1176                                         if (!buffer_delay(bh) &&
1177                                             !buffer_unwritten(bh))
1178                                                 done = 1;
1179                                         bh = bh->b_this_page;
1180                                 } while (!done && (bh != head));
1181                         }
1182                         unlock_page(page);
1183                         if (done)
1184                                 break;
1185                         idx++;
1186                         num++;
1187                         if (num >= max_pages)
1188                                 break;
1189                 }
1190                 pagevec_release(&pvec);
1191         }
1192         return num;
1193 }
1194
1195 /*
1196  * The ext4_get_blocks() function tries to look up the requested blocks,
1197  * and returns if the blocks are already mapped.
1198  *
1199  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1200  * and store the allocated blocks in the result buffer head and mark it
1201  * mapped.
1202  *
1203  * If file type is extents based, it will call ext4_ext_get_blocks(),
1204  * Otherwise, call with ext4_ind_get_blocks() to handle indirect mapping
1205  * based files
1206  *
1207  * On success, it returns the number of blocks being mapped or allocate.
1208  * if create==0 and the blocks are pre-allocated and uninitialized block,
1209  * the result buffer head is unmapped. If the create ==1, it will make sure
1210  * the buffer head is mapped.
1211  *
1212  * It returns 0 if plain look up failed (blocks have not been allocated), in
1213  * that casem, buffer head is unmapped
1214  *
1215  * It returns the error in case of allocation failure.
1216  */
1217 int ext4_get_blocks(handle_t *handle, struct inode *inode, sector_t block,
1218                     unsigned int max_blocks, struct buffer_head *bh,
1219                     int flags)
1220 {
1221         int retval;
1222
1223         clear_buffer_mapped(bh);
1224         clear_buffer_unwritten(bh);
1225
1226         ext_debug("ext4_get_blocks(): inode %lu, flag %d, max_blocks %u,"
1227                   "logical block %lu\n", inode->i_ino, flags, max_blocks,
1228                   (unsigned long)block);
1229         /*
1230          * Try to see if we can get the block without requesting a new
1231          * file system block.
1232          */
1233         down_read((&EXT4_I(inode)->i_data_sem));
1234         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1235                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1236                                 bh, 0);
1237         } else {
1238                 retval = ext4_ind_get_blocks(handle, inode, block, max_blocks,
1239                                              bh, 0);
1240         }
1241         up_read((&EXT4_I(inode)->i_data_sem));
1242
1243         if (retval > 0 && buffer_mapped(bh)) {
1244                 int ret = check_block_validity(inode, "file system corruption",
1245                                                block, bh->b_blocknr, retval);
1246                 if (ret != 0)
1247                         return ret;
1248         }
1249
1250         /* If it is only a block(s) look up */
1251         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1252                 return retval;
1253
1254         /*
1255          * Returns if the blocks have already allocated
1256          *
1257          * Note that if blocks have been preallocated
1258          * ext4_ext_get_block() returns th create = 0
1259          * with buffer head unmapped.
1260          */
1261         if (retval > 0 && buffer_mapped(bh))
1262                 return retval;
1263
1264         /*
1265          * When we call get_blocks without the create flag, the
1266          * BH_Unwritten flag could have gotten set if the blocks
1267          * requested were part of a uninitialized extent.  We need to
1268          * clear this flag now that we are committed to convert all or
1269          * part of the uninitialized extent to be an initialized
1270          * extent.  This is because we need to avoid the combination
1271          * of BH_Unwritten and BH_Mapped flags being simultaneously
1272          * set on the buffer_head.
1273          */
1274         clear_buffer_unwritten(bh);
1275
1276         /*
1277          * New blocks allocate and/or writing to uninitialized extent
1278          * will possibly result in updating i_data, so we take
1279          * the write lock of i_data_sem, and call get_blocks()
1280          * with create == 1 flag.
1281          */
1282         down_write((&EXT4_I(inode)->i_data_sem));
1283
1284         /*
1285          * if the caller is from delayed allocation writeout path
1286          * we have already reserved fs blocks for allocation
1287          * let the underlying get_block() function know to
1288          * avoid double accounting
1289          */
1290         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1291                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1292         /*
1293          * We need to check for EXT4 here because migrate
1294          * could have changed the inode type in between
1295          */
1296         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1297                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1298                                               bh, flags);
1299         } else {
1300                 retval = ext4_ind_get_blocks(handle, inode, block,
1301                                              max_blocks, bh, flags);
1302
1303                 if (retval > 0 && buffer_new(bh)) {
1304                         /*
1305                          * We allocated new blocks which will result in
1306                          * i_data's format changing.  Force the migrate
1307                          * to fail by clearing migrate flags
1308                          */
1309                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
1310                 }
1311
1312                 /*
1313                  * Update reserved blocks/metadata blocks after successful
1314                  * block allocation which had been deferred till now. We don't
1315                  * support fallocate for non extent files. So we can update
1316                  * reserve space here.
1317                  */
1318                 if ((retval > 0) &&
1319                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
1320                         ext4_da_update_reserve_space(inode, retval, 1);
1321         }
1322         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1323                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1324
1325         up_write((&EXT4_I(inode)->i_data_sem));
1326         if (retval > 0 && buffer_mapped(bh)) {
1327                 int ret = check_block_validity(inode, "file system "
1328                                                "corruption after allocation",
1329                                                block, bh->b_blocknr, retval);
1330                 if (ret != 0)
1331                         return ret;
1332         }
1333         return retval;
1334 }
1335
1336 /* Maximum number of blocks we map for direct IO at once. */
1337 #define DIO_MAX_BLOCKS 4096
1338
1339 int ext4_get_block(struct inode *inode, sector_t iblock,
1340                    struct buffer_head *bh_result, int create)
1341 {
1342         handle_t *handle = ext4_journal_current_handle();
1343         int ret = 0, started = 0;
1344         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1345         int dio_credits;
1346
1347         if (create && !handle) {
1348                 /* Direct IO write... */
1349                 if (max_blocks > DIO_MAX_BLOCKS)
1350                         max_blocks = DIO_MAX_BLOCKS;
1351                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
1352                 handle = ext4_journal_start(inode, dio_credits);
1353                 if (IS_ERR(handle)) {
1354                         ret = PTR_ERR(handle);
1355                         goto out;
1356                 }
1357                 started = 1;
1358         }
1359
1360         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
1361                               create ? EXT4_GET_BLOCKS_CREATE : 0);
1362         if (ret > 0) {
1363                 bh_result->b_size = (ret << inode->i_blkbits);
1364                 ret = 0;
1365         }
1366         if (started)
1367                 ext4_journal_stop(handle);
1368 out:
1369         return ret;
1370 }
1371
1372 /*
1373  * `handle' can be NULL if create is zero
1374  */
1375 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1376                                 ext4_lblk_t block, int create, int *errp)
1377 {
1378         struct buffer_head dummy;
1379         int fatal = 0, err;
1380         int flags = 0;
1381
1382         J_ASSERT(handle != NULL || create == 0);
1383
1384         dummy.b_state = 0;
1385         dummy.b_blocknr = -1000;
1386         buffer_trace_init(&dummy.b_history);
1387         if (create)
1388                 flags |= EXT4_GET_BLOCKS_CREATE;
1389         err = ext4_get_blocks(handle, inode, block, 1, &dummy, flags);
1390         /*
1391          * ext4_get_blocks() returns number of blocks mapped. 0 in
1392          * case of a HOLE.
1393          */
1394         if (err > 0) {
1395                 if (err > 1)
1396                         WARN_ON(1);
1397                 err = 0;
1398         }
1399         *errp = err;
1400         if (!err && buffer_mapped(&dummy)) {
1401                 struct buffer_head *bh;
1402                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1403                 if (!bh) {
1404                         *errp = -EIO;
1405                         goto err;
1406                 }
1407                 if (buffer_new(&dummy)) {
1408                         J_ASSERT(create != 0);
1409                         J_ASSERT(handle != NULL);
1410
1411                         /*
1412                          * Now that we do not always journal data, we should
1413                          * keep in mind whether this should always journal the
1414                          * new buffer as metadata.  For now, regular file
1415                          * writes use ext4_get_block instead, so it's not a
1416                          * problem.
1417                          */
1418                         lock_buffer(bh);
1419                         BUFFER_TRACE(bh, "call get_create_access");
1420                         fatal = ext4_journal_get_create_access(handle, bh);
1421                         if (!fatal && !buffer_uptodate(bh)) {
1422                                 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1423                                 set_buffer_uptodate(bh);
1424                         }
1425                         unlock_buffer(bh);
1426                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1427                         err = ext4_handle_dirty_metadata(handle, inode, bh);
1428                         if (!fatal)
1429                                 fatal = err;
1430                 } else {
1431                         BUFFER_TRACE(bh, "not a new buffer");
1432                 }
1433                 if (fatal) {
1434                         *errp = fatal;
1435                         brelse(bh);
1436                         bh = NULL;
1437                 }
1438                 return bh;
1439         }
1440 err:
1441         return NULL;
1442 }
1443
1444 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1445                                ext4_lblk_t block, int create, int *err)
1446 {
1447         struct buffer_head *bh;
1448
1449         bh = ext4_getblk(handle, inode, block, create, err);
1450         if (!bh)
1451                 return bh;
1452         if (buffer_uptodate(bh))
1453                 return bh;
1454         ll_rw_block(READ_META, 1, &bh);
1455         wait_on_buffer(bh);
1456         if (buffer_uptodate(bh))
1457                 return bh;
1458         put_bh(bh);
1459         *err = -EIO;
1460         return NULL;
1461 }
1462
1463 static int walk_page_buffers(handle_t *handle,
1464                              struct buffer_head *head,
1465                              unsigned from,
1466                              unsigned to,
1467                              int *partial,
1468                              int (*fn)(handle_t *handle,
1469                                        struct buffer_head *bh))
1470 {
1471         struct buffer_head *bh;
1472         unsigned block_start, block_end;
1473         unsigned blocksize = head->b_size;
1474         int err, ret = 0;
1475         struct buffer_head *next;
1476
1477         for (bh = head, block_start = 0;
1478              ret == 0 && (bh != head || !block_start);
1479              block_start = block_end, bh = next) {
1480                 next = bh->b_this_page;
1481                 block_end = block_start + blocksize;
1482                 if (block_end <= from || block_start >= to) {
1483                         if (partial && !buffer_uptodate(bh))
1484                                 *partial = 1;
1485                         continue;
1486                 }
1487                 err = (*fn)(handle, bh);
1488                 if (!ret)
1489                         ret = err;
1490         }
1491         return ret;
1492 }
1493
1494 /*
1495  * To preserve ordering, it is essential that the hole instantiation and
1496  * the data write be encapsulated in a single transaction.  We cannot
1497  * close off a transaction and start a new one between the ext4_get_block()
1498  * and the commit_write().  So doing the jbd2_journal_start at the start of
1499  * prepare_write() is the right place.
1500  *
1501  * Also, this function can nest inside ext4_writepage() ->
1502  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1503  * has generated enough buffer credits to do the whole page.  So we won't
1504  * block on the journal in that case, which is good, because the caller may
1505  * be PF_MEMALLOC.
1506  *
1507  * By accident, ext4 can be reentered when a transaction is open via
1508  * quota file writes.  If we were to commit the transaction while thus
1509  * reentered, there can be a deadlock - we would be holding a quota
1510  * lock, and the commit would never complete if another thread had a
1511  * transaction open and was blocking on the quota lock - a ranking
1512  * violation.
1513  *
1514  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1515  * will _not_ run commit under these circumstances because handle->h_ref
1516  * is elevated.  We'll still have enough credits for the tiny quotafile
1517  * write.
1518  */
1519 static int do_journal_get_write_access(handle_t *handle,
1520                                        struct buffer_head *bh)
1521 {
1522         if (!buffer_mapped(bh) || buffer_freed(bh))
1523                 return 0;
1524         return ext4_journal_get_write_access(handle, bh);
1525 }
1526
1527 /*
1528  * Truncate blocks that were not used by write. We have to truncate the
1529  * pagecache as well so that corresponding buffers get properly unmapped.
1530  */
1531 static void ext4_truncate_failed_write(struct inode *inode)
1532 {
1533         truncate_inode_pages(inode->i_mapping, inode->i_size);
1534         ext4_truncate(inode);
1535 }
1536
1537 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1538                             loff_t pos, unsigned len, unsigned flags,
1539                             struct page **pagep, void **fsdata)
1540 {
1541         struct inode *inode = mapping->host;
1542         int ret, needed_blocks;
1543         handle_t *handle;
1544         int retries = 0;
1545         struct page *page;
1546         pgoff_t index;
1547         unsigned from, to;
1548
1549         trace_ext4_write_begin(inode, pos, len, flags);
1550         /*
1551          * Reserve one block more for addition to orphan list in case
1552          * we allocate blocks but write fails for some reason
1553          */
1554         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1555         index = pos >> PAGE_CACHE_SHIFT;
1556         from = pos & (PAGE_CACHE_SIZE - 1);
1557         to = from + len;
1558
1559 retry:
1560         handle = ext4_journal_start(inode, needed_blocks);
1561         if (IS_ERR(handle)) {
1562                 ret = PTR_ERR(handle);
1563                 goto out;
1564         }
1565
1566         /* We cannot recurse into the filesystem as the transaction is already
1567          * started */
1568         flags |= AOP_FLAG_NOFS;
1569
1570         page = grab_cache_page_write_begin(mapping, index, flags);
1571         if (!page) {
1572                 ext4_journal_stop(handle);
1573                 ret = -ENOMEM;
1574                 goto out;
1575         }
1576         *pagep = page;
1577
1578         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1579                                 ext4_get_block);
1580
1581         if (!ret && ext4_should_journal_data(inode)) {
1582                 ret = walk_page_buffers(handle, page_buffers(page),
1583                                 from, to, NULL, do_journal_get_write_access);
1584         }
1585
1586         if (ret) {
1587                 unlock_page(page);
1588                 page_cache_release(page);
1589                 /*
1590                  * block_write_begin may have instantiated a few blocks
1591                  * outside i_size.  Trim these off again. Don't need
1592                  * i_size_read because we hold i_mutex.
1593                  *
1594                  * Add inode to orphan list in case we crash before
1595                  * truncate finishes
1596                  */
1597                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1598                         ext4_orphan_add(handle, inode);
1599
1600                 ext4_journal_stop(handle);
1601                 if (pos + len > inode->i_size) {
1602                         ext4_truncate_failed_write(inode);
1603                         /*
1604                          * If truncate failed early the inode might
1605                          * still be on the orphan list; we need to
1606                          * make sure the inode is removed from the
1607                          * orphan list in that case.
1608                          */
1609                         if (inode->i_nlink)
1610                                 ext4_orphan_del(NULL, inode);
1611                 }
1612         }
1613
1614         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1615                 goto retry;
1616 out:
1617         return ret;
1618 }
1619
1620 /* For write_end() in data=journal mode */
1621 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1622 {
1623         if (!buffer_mapped(bh) || buffer_freed(bh))
1624                 return 0;
1625         set_buffer_uptodate(bh);
1626         return ext4_handle_dirty_metadata(handle, NULL, bh);
1627 }
1628
1629 static int ext4_generic_write_end(struct file *file,
1630                                   struct address_space *mapping,
1631                                   loff_t pos, unsigned len, unsigned copied,
1632                                   struct page *page, void *fsdata)
1633 {
1634         int i_size_changed = 0;
1635         struct inode *inode = mapping->host;
1636         handle_t *handle = ext4_journal_current_handle();
1637
1638         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1639
1640         /*
1641          * No need to use i_size_read() here, the i_size
1642          * cannot change under us because we hold i_mutex.
1643          *
1644          * But it's important to update i_size while still holding page lock:
1645          * page writeout could otherwise come in and zero beyond i_size.
1646          */
1647         if (pos + copied > inode->i_size) {
1648                 i_size_write(inode, pos + copied);
1649                 i_size_changed = 1;
1650         }
1651
1652         if (pos + copied >  EXT4_I(inode)->i_disksize) {
1653                 /* We need to mark inode dirty even if
1654                  * new_i_size is less that inode->i_size
1655                  * bu greater than i_disksize.(hint delalloc)
1656                  */
1657                 ext4_update_i_disksize(inode, (pos + copied));
1658                 i_size_changed = 1;
1659         }
1660         unlock_page(page);
1661         page_cache_release(page);
1662
1663         /*
1664          * Don't mark the inode dirty under page lock. First, it unnecessarily
1665          * makes the holding time of page lock longer. Second, it forces lock
1666          * ordering of page lock and transaction start for journaling
1667          * filesystems.
1668          */
1669         if (i_size_changed)
1670                 ext4_mark_inode_dirty(handle, inode);
1671
1672         return copied;
1673 }
1674
1675 /*
1676  * We need to pick up the new inode size which generic_commit_write gave us
1677  * `file' can be NULL - eg, when called from page_symlink().
1678  *
1679  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1680  * buffers are managed internally.
1681  */
1682 static int ext4_ordered_write_end(struct file *file,
1683                                   struct address_space *mapping,
1684                                   loff_t pos, unsigned len, unsigned copied,
1685                                   struct page *page, void *fsdata)
1686 {
1687         handle_t *handle = ext4_journal_current_handle();
1688         struct inode *inode = mapping->host;
1689         int ret = 0, ret2;
1690
1691         trace_ext4_ordered_write_end(inode, pos, len, copied);
1692         ret = ext4_jbd2_file_inode(handle, inode);
1693
1694         if (ret == 0) {
1695                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1696                                                         page, fsdata);
1697                 copied = ret2;
1698                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1699                         /* if we have allocated more blocks and copied
1700                          * less. We will have blocks allocated outside
1701                          * inode->i_size. So truncate them
1702                          */
1703                         ext4_orphan_add(handle, inode);
1704                 if (ret2 < 0)
1705                         ret = ret2;
1706         }
1707         ret2 = ext4_journal_stop(handle);
1708         if (!ret)
1709                 ret = ret2;
1710
1711         if (pos + len > inode->i_size) {
1712                 ext4_truncate_failed_write(inode);
1713                 /*
1714                  * If truncate failed early the inode might still be
1715                  * on the orphan list; we need to make sure the inode
1716                  * is removed from the orphan list in that case.
1717                  */
1718                 if (inode->i_nlink)
1719                         ext4_orphan_del(NULL, inode);
1720         }
1721
1722
1723         return ret ? ret : copied;
1724 }
1725
1726 static int ext4_writeback_write_end(struct file *file,
1727                                     struct address_space *mapping,
1728                                     loff_t pos, unsigned len, unsigned copied,
1729                                     struct page *page, void *fsdata)
1730 {
1731         handle_t *handle = ext4_journal_current_handle();
1732         struct inode *inode = mapping->host;
1733         int ret = 0, ret2;
1734
1735         trace_ext4_writeback_write_end(inode, pos, len, copied);
1736         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1737                                                         page, fsdata);
1738         copied = ret2;
1739         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1740                 /* if we have allocated more blocks and copied
1741                  * less. We will have blocks allocated outside
1742                  * inode->i_size. So truncate them
1743                  */
1744                 ext4_orphan_add(handle, inode);
1745
1746         if (ret2 < 0)
1747                 ret = ret2;
1748
1749         ret2 = ext4_journal_stop(handle);
1750         if (!ret)
1751                 ret = ret2;
1752
1753         if (pos + len > inode->i_size) {
1754                 ext4_truncate_failed_write(inode);
1755                 /*
1756                  * If truncate failed early the inode might still be
1757                  * on the orphan list; we need to make sure the inode
1758                  * is removed from the orphan list in that case.
1759                  */
1760                 if (inode->i_nlink)
1761                         ext4_orphan_del(NULL, inode);
1762         }
1763
1764         return ret ? ret : copied;
1765 }
1766
1767 static int ext4_journalled_write_end(struct file *file,
1768                                      struct address_space *mapping,
1769                                      loff_t pos, unsigned len, unsigned copied,
1770                                      struct page *page, void *fsdata)
1771 {
1772         handle_t *handle = ext4_journal_current_handle();
1773         struct inode *inode = mapping->host;
1774         int ret = 0, ret2;
1775         int partial = 0;
1776         unsigned from, to;
1777         loff_t new_i_size;
1778
1779         trace_ext4_journalled_write_end(inode, pos, len, copied);
1780         from = pos & (PAGE_CACHE_SIZE - 1);
1781         to = from + len;
1782
1783         if (copied < len) {
1784                 if (!PageUptodate(page))
1785                         copied = 0;
1786                 page_zero_new_buffers(page, from+copied, to);
1787         }
1788
1789         ret = walk_page_buffers(handle, page_buffers(page), from,
1790                                 to, &partial, write_end_fn);
1791         if (!partial)
1792                 SetPageUptodate(page);
1793         new_i_size = pos + copied;
1794         if (new_i_size > inode->i_size)
1795                 i_size_write(inode, pos+copied);
1796         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1797         if (new_i_size > EXT4_I(inode)->i_disksize) {
1798                 ext4_update_i_disksize(inode, new_i_size);
1799                 ret2 = ext4_mark_inode_dirty(handle, inode);
1800                 if (!ret)
1801                         ret = ret2;
1802         }
1803
1804         unlock_page(page);
1805         page_cache_release(page);
1806         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1807                 /* if we have allocated more blocks and copied
1808                  * less. We will have blocks allocated outside
1809                  * inode->i_size. So truncate them
1810                  */
1811                 ext4_orphan_add(handle, inode);
1812
1813         ret2 = ext4_journal_stop(handle);
1814         if (!ret)
1815                 ret = ret2;
1816         if (pos + len > inode->i_size) {
1817                 ext4_truncate_failed_write(inode);
1818                 /*
1819                  * If truncate failed early the inode might still be
1820                  * on the orphan list; we need to make sure the inode
1821                  * is removed from the orphan list in that case.
1822                  */
1823                 if (inode->i_nlink)
1824                         ext4_orphan_del(NULL, inode);
1825         }
1826
1827         return ret ? ret : copied;
1828 }
1829
1830 /*
1831  * Reserve a single block located at lblock
1832  */
1833 static int ext4_da_reserve_space(struct inode *inode, sector_t lblock)
1834 {
1835         int retries = 0;
1836         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1837         struct ext4_inode_info *ei = EXT4_I(inode);
1838         unsigned long md_needed, md_reserved;
1839
1840         /*
1841          * recalculate the amount of metadata blocks to reserve
1842          * in order to allocate nrblocks
1843          * worse case is one extent per block
1844          */
1845 repeat:
1846         spin_lock(&ei->i_block_reservation_lock);
1847         md_reserved = ei->i_reserved_meta_blocks;
1848         md_needed = ext4_calc_metadata_amount(inode, lblock);
1849         trace_ext4_da_reserve_space(inode, md_needed);
1850         spin_unlock(&ei->i_block_reservation_lock);
1851
1852         /*
1853          * Make quota reservation here to prevent quota overflow
1854          * later. Real quota accounting is done at pages writeout
1855          * time.
1856          */
1857         if (vfs_dq_reserve_block(inode, md_needed + 1))
1858                 return -EDQUOT;
1859
1860         if (ext4_claim_free_blocks(sbi, md_needed + 1)) {
1861                 vfs_dq_release_reservation_block(inode, md_needed + 1);
1862                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1863                         yield();
1864                         goto repeat;
1865                 }
1866                 return -ENOSPC;
1867         }
1868         spin_lock(&ei->i_block_reservation_lock);
1869         ei->i_reserved_data_blocks++;
1870         ei->i_reserved_meta_blocks += md_needed;
1871         spin_unlock(&ei->i_block_reservation_lock);
1872
1873         return 0;       /* success */
1874 }
1875
1876 static void ext4_da_release_space(struct inode *inode, int to_free)
1877 {
1878         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1879         struct ext4_inode_info *ei = EXT4_I(inode);
1880
1881         if (!to_free)
1882                 return;         /* Nothing to release, exit */
1883
1884         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1885
1886         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1887                 /*
1888                  * if there aren't enough reserved blocks, then the
1889                  * counter is messed up somewhere.  Since this
1890                  * function is called from invalidate page, it's
1891                  * harmless to return without any action.
1892                  */
1893                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1894                          "ino %lu, to_free %d with only %d reserved "
1895                          "data blocks\n", inode->i_ino, to_free,
1896                          ei->i_reserved_data_blocks);
1897                 WARN_ON(1);
1898                 to_free = ei->i_reserved_data_blocks;
1899         }
1900         ei->i_reserved_data_blocks -= to_free;
1901
1902         if (ei->i_reserved_data_blocks == 0) {
1903                 /*
1904                  * We can release all of the reserved metadata blocks
1905                  * only when we have written all of the delayed
1906                  * allocation blocks.
1907                  */
1908                 to_free += ei->i_reserved_meta_blocks;
1909                 ei->i_reserved_meta_blocks = 0;
1910                 ei->i_da_metadata_calc_len = 0;
1911         }
1912
1913         /* update fs dirty blocks counter */
1914         percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1915
1916         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1917
1918         vfs_dq_release_reservation_block(inode, to_free);
1919 }
1920
1921 static void ext4_da_page_release_reservation(struct page *page,
1922                                              unsigned long offset)
1923 {
1924         int to_release = 0;
1925         struct buffer_head *head, *bh;
1926         unsigned int curr_off = 0;
1927
1928         head = page_buffers(page);
1929         bh = head;
1930         do {
1931                 unsigned int next_off = curr_off + bh->b_size;
1932
1933                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1934                         to_release++;
1935                         clear_buffer_delay(bh);
1936                 }
1937                 curr_off = next_off;
1938         } while ((bh = bh->b_this_page) != head);
1939         ext4_da_release_space(page->mapping->host, to_release);
1940 }
1941
1942 /*
1943  * Delayed allocation stuff
1944  */
1945
1946 /*
1947  * mpage_da_submit_io - walks through extent of pages and try to write
1948  * them with writepage() call back
1949  *
1950  * @mpd->inode: inode
1951  * @mpd->first_page: first page of the extent
1952  * @mpd->next_page: page after the last page of the extent
1953  *
1954  * By the time mpage_da_submit_io() is called we expect all blocks
1955  * to be allocated. this may be wrong if allocation failed.
1956  *
1957  * As pages are already locked by write_cache_pages(), we can't use it
1958  */
1959 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1960 {
1961         long pages_skipped;
1962         struct pagevec pvec;
1963         unsigned long index, end;
1964         int ret = 0, err, nr_pages, i;
1965         struct inode *inode = mpd->inode;
1966         struct address_space *mapping = inode->i_mapping;
1967
1968         BUG_ON(mpd->next_page <= mpd->first_page);
1969         /*
1970          * We need to start from the first_page to the next_page - 1
1971          * to make sure we also write the mapped dirty buffer_heads.
1972          * If we look at mpd->b_blocknr we would only be looking
1973          * at the currently mapped buffer_heads.
1974          */
1975         index = mpd->first_page;
1976         end = mpd->next_page - 1;
1977
1978         pagevec_init(&pvec, 0);
1979         while (index <= end) {
1980                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1981                 if (nr_pages == 0)
1982                         break;
1983                 for (i = 0; i < nr_pages; i++) {
1984                         struct page *page = pvec.pages[i];
1985
1986                         index = page->index;
1987                         if (index > end)
1988                                 break;
1989                         index++;
1990
1991                         BUG_ON(!PageLocked(page));
1992                         BUG_ON(PageWriteback(page));
1993
1994                         pages_skipped = mpd->wbc->pages_skipped;
1995                         err = mapping->a_ops->writepage(page, mpd->wbc);
1996                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
1997                                 /*
1998                                  * have successfully written the page
1999                                  * without skipping the same
2000                                  */
2001                                 mpd->pages_written++;
2002                         /*
2003                          * In error case, we have to continue because
2004                          * remaining pages are still locked
2005                          * XXX: unlock and re-dirty them?
2006                          */
2007                         if (ret == 0)
2008                                 ret = err;
2009                 }
2010                 pagevec_release(&pvec);
2011         }
2012         return ret;
2013 }
2014
2015 /*
2016  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
2017  *
2018  * @mpd->inode - inode to walk through
2019  * @exbh->b_blocknr - first block on a disk
2020  * @exbh->b_size - amount of space in bytes
2021  * @logical - first logical block to start assignment with
2022  *
2023  * the function goes through all passed space and put actual disk
2024  * block numbers into buffer heads, dropping BH_Delay and BH_Unwritten
2025  */
2026 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
2027                                  struct buffer_head *exbh)
2028 {
2029         struct inode *inode = mpd->inode;
2030         struct address_space *mapping = inode->i_mapping;
2031         int blocks = exbh->b_size >> inode->i_blkbits;
2032         sector_t pblock = exbh->b_blocknr, cur_logical;
2033         struct buffer_head *head, *bh;
2034         pgoff_t index, end;
2035         struct pagevec pvec;
2036         int nr_pages, i;
2037
2038         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2039         end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2040         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2041
2042         pagevec_init(&pvec, 0);
2043
2044         while (index <= end) {
2045                 /* XXX: optimize tail */
2046                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2047                 if (nr_pages == 0)
2048                         break;
2049                 for (i = 0; i < nr_pages; i++) {
2050                         struct page *page = pvec.pages[i];
2051
2052                         index = page->index;
2053                         if (index > end)
2054                                 break;
2055                         index++;
2056
2057                         BUG_ON(!PageLocked(page));
2058                         BUG_ON(PageWriteback(page));
2059                         BUG_ON(!page_has_buffers(page));
2060
2061                         bh = page_buffers(page);
2062                         head = bh;
2063
2064                         /* skip blocks out of the range */
2065                         do {
2066                                 if (cur_logical >= logical)
2067                                         break;
2068                                 cur_logical++;
2069                         } while ((bh = bh->b_this_page) != head);
2070
2071                         do {
2072                                 if (cur_logical >= logical + blocks)
2073                                         break;
2074
2075                                 if (buffer_delay(bh) ||
2076                                                 buffer_unwritten(bh)) {
2077
2078                                         BUG_ON(bh->b_bdev != inode->i_sb->s_bdev);
2079
2080                                         if (buffer_delay(bh)) {
2081                                                 clear_buffer_delay(bh);
2082                                                 bh->b_blocknr = pblock;
2083                                         } else {
2084                                                 /*
2085                                                  * unwritten already should have
2086                                                  * blocknr assigned. Verify that
2087                                                  */
2088                                                 clear_buffer_unwritten(bh);
2089                                                 BUG_ON(bh->b_blocknr != pblock);
2090                                         }
2091
2092                                 } else if (buffer_mapped(bh))
2093                                         BUG_ON(bh->b_blocknr != pblock);
2094
2095                                 cur_logical++;
2096                                 pblock++;
2097                         } while ((bh = bh->b_this_page) != head);
2098                 }
2099                 pagevec_release(&pvec);
2100         }
2101 }
2102
2103
2104 /*
2105  * __unmap_underlying_blocks - just a helper function to unmap
2106  * set of blocks described by @bh
2107  */
2108 static inline void __unmap_underlying_blocks(struct inode *inode,
2109                                              struct buffer_head *bh)
2110 {
2111         struct block_device *bdev = inode->i_sb->s_bdev;
2112         int blocks, i;
2113
2114         blocks = bh->b_size >> inode->i_blkbits;
2115         for (i = 0; i < blocks; i++)
2116                 unmap_underlying_metadata(bdev, bh->b_blocknr + i);
2117 }
2118
2119 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2120                                         sector_t logical, long blk_cnt)
2121 {
2122         int nr_pages, i;
2123         pgoff_t index, end;
2124         struct pagevec pvec;
2125         struct inode *inode = mpd->inode;
2126         struct address_space *mapping = inode->i_mapping;
2127
2128         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2129         end   = (logical + blk_cnt - 1) >>
2130                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2131         while (index <= end) {
2132                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2133                 if (nr_pages == 0)
2134                         break;
2135                 for (i = 0; i < nr_pages; i++) {
2136                         struct page *page = pvec.pages[i];
2137                         index = page->index;
2138                         if (index > end)
2139                                 break;
2140                         index++;
2141
2142                         BUG_ON(!PageLocked(page));
2143                         BUG_ON(PageWriteback(page));
2144                         block_invalidatepage(page, 0);
2145                         ClearPageUptodate(page);
2146                         unlock_page(page);
2147                 }
2148         }
2149         return;
2150 }
2151
2152 static void ext4_print_free_blocks(struct inode *inode)
2153 {
2154         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2155         printk(KERN_CRIT "Total free blocks count %lld\n",
2156                ext4_count_free_blocks(inode->i_sb));
2157         printk(KERN_CRIT "Free/Dirty block details\n");
2158         printk(KERN_CRIT "free_blocks=%lld\n",
2159                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2160         printk(KERN_CRIT "dirty_blocks=%lld\n",
2161                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2162         printk(KERN_CRIT "Block reservation details\n");
2163         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2164                EXT4_I(inode)->i_reserved_data_blocks);
2165         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2166                EXT4_I(inode)->i_reserved_meta_blocks);
2167         return;
2168 }
2169
2170 /*
2171  * mpage_da_map_blocks - go through given space
2172  *
2173  * @mpd - bh describing space
2174  *
2175  * The function skips space we know is already mapped to disk blocks.
2176  *
2177  */
2178 static int mpage_da_map_blocks(struct mpage_da_data *mpd)
2179 {
2180         int err, blks, get_blocks_flags;
2181         struct buffer_head new;
2182         sector_t next = mpd->b_blocknr;
2183         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2184         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2185         handle_t *handle = NULL;
2186
2187         /*
2188          * We consider only non-mapped and non-allocated blocks
2189          */
2190         if ((mpd->b_state  & (1 << BH_Mapped)) &&
2191                 !(mpd->b_state & (1 << BH_Delay)) &&
2192                 !(mpd->b_state & (1 << BH_Unwritten)))
2193                 return 0;
2194
2195         /*
2196          * If we didn't accumulate anything to write simply return
2197          */
2198         if (!mpd->b_size)
2199                 return 0;
2200
2201         handle = ext4_journal_current_handle();
2202         BUG_ON(!handle);
2203
2204         /*
2205          * Call ext4_get_blocks() to allocate any delayed allocation
2206          * blocks, or to convert an uninitialized extent to be
2207          * initialized (in the case where we have written into
2208          * one or more preallocated blocks).
2209          *
2210          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2211          * indicate that we are on the delayed allocation path.  This
2212          * affects functions in many different parts of the allocation
2213          * call path.  This flag exists primarily because we don't
2214          * want to change *many* call functions, so ext4_get_blocks()
2215          * will set the magic i_delalloc_reserved_flag once the
2216          * inode's allocation semaphore is taken.
2217          *
2218          * If the blocks in questions were delalloc blocks, set
2219          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2220          * variables are updated after the blocks have been allocated.
2221          */
2222         new.b_state = 0;
2223         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
2224         if (mpd->b_state & (1 << BH_Delay))
2225                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2226
2227         blks = ext4_get_blocks(handle, mpd->inode, next, max_blocks,
2228                                &new, get_blocks_flags);
2229         if (blks < 0) {
2230                 err = blks;
2231                 /*
2232                  * If get block returns with error we simply
2233                  * return. Later writepage will redirty the page and
2234                  * writepages will find the dirty page again
2235                  */
2236                 if (err == -EAGAIN)
2237                         return 0;
2238
2239                 if (err == -ENOSPC &&
2240                     ext4_count_free_blocks(mpd->inode->i_sb)) {
2241                         mpd->retval = err;
2242                         return 0;
2243                 }
2244
2245                 /*
2246                  * get block failure will cause us to loop in
2247                  * writepages, because a_ops->writepage won't be able
2248                  * to make progress. The page will be redirtied by
2249                  * writepage and writepages will again try to write
2250                  * the same.
2251                  */
2252                 ext4_msg(mpd->inode->i_sb, KERN_CRIT,
2253                          "delayed block allocation failed for inode %lu at "
2254                          "logical offset %llu with max blocks %zd with "
2255                          "error %d\n", mpd->inode->i_ino,
2256                          (unsigned long long) next,
2257                          mpd->b_size >> mpd->inode->i_blkbits, err);
2258                 printk(KERN_CRIT "This should not happen!!  "
2259                        "Data will be lost\n");
2260                 if (err == -ENOSPC) {
2261                         ext4_print_free_blocks(mpd->inode);
2262                 }
2263                 /* invalidate all the pages */
2264                 ext4_da_block_invalidatepages(mpd, next,
2265                                 mpd->b_size >> mpd->inode->i_blkbits);
2266                 return err;
2267         }
2268         BUG_ON(blks == 0);
2269
2270         new.b_size = (blks << mpd->inode->i_blkbits);
2271
2272         if (buffer_new(&new))
2273                 __unmap_underlying_blocks(mpd->inode, &new);
2274
2275         /*
2276          * If blocks are delayed marked, we need to
2277          * put actual blocknr and drop delayed bit
2278          */
2279         if ((mpd->b_state & (1 << BH_Delay)) ||
2280             (mpd->b_state & (1 << BH_Unwritten)))
2281                 mpage_put_bnr_to_bhs(mpd, next, &new);
2282
2283         if (ext4_should_order_data(mpd->inode)) {
2284                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2285                 if (err)
2286                         return err;
2287         }
2288
2289         /*
2290          * Update on-disk size along with block allocation.
2291          */
2292         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2293         if (disksize > i_size_read(mpd->inode))
2294                 disksize = i_size_read(mpd->inode);
2295         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2296                 ext4_update_i_disksize(mpd->inode, disksize);
2297                 return ext4_mark_inode_dirty(handle, mpd->inode);
2298         }
2299
2300         return 0;
2301 }
2302
2303 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2304                 (1 << BH_Delay) | (1 << BH_Unwritten))
2305
2306 /*
2307  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2308  *
2309  * @mpd->lbh - extent of blocks
2310  * @logical - logical number of the block in the file
2311  * @bh - bh of the block (used to access block's state)
2312  *
2313  * the function is used to collect contig. blocks in same state
2314  */
2315 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2316                                    sector_t logical, size_t b_size,
2317                                    unsigned long b_state)
2318 {
2319         sector_t next;
2320         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2321
2322         /* check if thereserved journal credits might overflow */
2323         if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
2324                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2325                         /*
2326                          * With non-extent format we are limited by the journal
2327                          * credit available.  Total credit needed to insert
2328                          * nrblocks contiguous blocks is dependent on the
2329                          * nrblocks.  So limit nrblocks.
2330                          */
2331                         goto flush_it;
2332                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2333                                 EXT4_MAX_TRANS_DATA) {
2334                         /*
2335                          * Adding the new buffer_head would make it cross the
2336                          * allowed limit for which we have journal credit
2337                          * reserved. So limit the new bh->b_size
2338                          */
2339                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2340                                                 mpd->inode->i_blkbits;
2341                         /* we will do mpage_da_submit_io in the next loop */
2342                 }
2343         }
2344         /*
2345          * First block in the extent
2346          */
2347         if (mpd->b_size == 0) {
2348                 mpd->b_blocknr = logical;
2349                 mpd->b_size = b_size;
2350                 mpd->b_state = b_state & BH_FLAGS;
2351                 return;
2352         }
2353
2354         next = mpd->b_blocknr + nrblocks;
2355         /*
2356          * Can we merge the block to our big extent?
2357          */
2358         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2359                 mpd->b_size += b_size;
2360                 return;
2361         }
2362
2363 flush_it:
2364         /*
2365          * We couldn't merge the block to our extent, so we
2366          * need to flush current  extent and start new one
2367          */
2368         if (mpage_da_map_blocks(mpd) == 0)
2369                 mpage_da_submit_io(mpd);
2370         mpd->io_done = 1;
2371         return;
2372 }
2373
2374 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
2375 {
2376         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
2377 }
2378
2379 /*
2380  * __mpage_da_writepage - finds extent of pages and blocks
2381  *
2382  * @page: page to consider
2383  * @wbc: not used, we just follow rules
2384  * @data: context
2385  *
2386  * The function finds extents of pages and scan them for all blocks.
2387  */
2388 static int __mpage_da_writepage(struct page *page,
2389                                 struct writeback_control *wbc, void *data)
2390 {
2391         struct mpage_da_data *mpd = data;
2392         struct inode *inode = mpd->inode;
2393         struct buffer_head *bh, *head;
2394         sector_t logical;
2395
2396         if (mpd->io_done) {
2397                 /*
2398                  * Rest of the page in the page_vec
2399                  * redirty then and skip then. We will
2400                  * try to write them again after
2401                  * starting a new transaction
2402                  */
2403                 redirty_page_for_writepage(wbc, page);
2404                 unlock_page(page);
2405                 return MPAGE_DA_EXTENT_TAIL;
2406         }
2407         /*
2408          * Can we merge this page to current extent?
2409          */
2410         if (mpd->next_page != page->index) {
2411                 /*
2412                  * Nope, we can't. So, we map non-allocated blocks
2413                  * and start IO on them using writepage()
2414                  */
2415                 if (mpd->next_page != mpd->first_page) {
2416                         if (mpage_da_map_blocks(mpd) == 0)
2417                                 mpage_da_submit_io(mpd);
2418                         /*
2419                          * skip rest of the page in the page_vec
2420                          */
2421                         mpd->io_done = 1;
2422                         redirty_page_for_writepage(wbc, page);
2423                         unlock_page(page);
2424                         return MPAGE_DA_EXTENT_TAIL;
2425                 }
2426
2427                 /*
2428                  * Start next extent of pages ...
2429                  */
2430                 mpd->first_page = page->index;
2431
2432                 /*
2433                  * ... and blocks
2434                  */
2435                 mpd->b_size = 0;
2436                 mpd->b_state = 0;
2437                 mpd->b_blocknr = 0;
2438         }
2439
2440         mpd->next_page = page->index + 1;
2441         logical = (sector_t) page->index <<
2442                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2443
2444         if (!page_has_buffers(page)) {
2445                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2446                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2447                 if (mpd->io_done)
2448                         return MPAGE_DA_EXTENT_TAIL;
2449         } else {
2450                 /*
2451                  * Page with regular buffer heads, just add all dirty ones
2452                  */
2453                 head = page_buffers(page);
2454                 bh = head;
2455                 do {
2456                         BUG_ON(buffer_locked(bh));
2457                         /*
2458                          * We need to try to allocate
2459                          * unmapped blocks in the same page.
2460                          * Otherwise we won't make progress
2461                          * with the page in ext4_writepage
2462                          */
2463                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2464                                 mpage_add_bh_to_extent(mpd, logical,
2465                                                        bh->b_size,
2466                                                        bh->b_state);
2467                                 if (mpd->io_done)
2468                                         return MPAGE_DA_EXTENT_TAIL;
2469                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2470                                 /*
2471                                  * mapped dirty buffer. We need to update
2472                                  * the b_state because we look at
2473                                  * b_state in mpage_da_map_blocks. We don't
2474                                  * update b_size because if we find an
2475                                  * unmapped buffer_head later we need to
2476                                  * use the b_state flag of that buffer_head.
2477                                  */
2478                                 if (mpd->b_size == 0)
2479                                         mpd->b_state = bh->b_state & BH_FLAGS;
2480                         }
2481                         logical++;
2482                 } while ((bh = bh->b_this_page) != head);
2483         }
2484
2485         return 0;
2486 }
2487
2488 /*
2489  * This is a special get_blocks_t callback which is used by
2490  * ext4_da_write_begin().  It will either return mapped block or
2491  * reserve space for a single block.
2492  *
2493  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2494  * We also have b_blocknr = -1 and b_bdev initialized properly
2495  *
2496  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2497  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2498  * initialized properly.
2499  */
2500 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2501                                   struct buffer_head *bh_result, int create)
2502 {
2503         int ret = 0;
2504         sector_t invalid_block = ~((sector_t) 0xffff);
2505
2506         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2507                 invalid_block = ~0;
2508
2509         BUG_ON(create == 0);
2510         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2511
2512         /*
2513          * first, we need to know whether the block is allocated already
2514          * preallocated blocks are unmapped but should treated
2515          * the same as allocated blocks.
2516          */
2517         ret = ext4_get_blocks(NULL, inode, iblock, 1,  bh_result, 0);
2518         if ((ret == 0) && !buffer_delay(bh_result)) {
2519                 /* the block isn't (pre)allocated yet, let's reserve space */
2520                 /*
2521                  * XXX: __block_prepare_write() unmaps passed block,
2522                  * is it OK?
2523                  */
2524                 ret = ext4_da_reserve_space(inode, iblock);
2525                 if (ret)
2526                         /* not enough space to reserve */
2527                         return ret;
2528
2529                 map_bh(bh_result, inode->i_sb, invalid_block);
2530                 set_buffer_new(bh_result);
2531                 set_buffer_delay(bh_result);
2532         } else if (ret > 0) {
2533                 bh_result->b_size = (ret << inode->i_blkbits);
2534                 if (buffer_unwritten(bh_result)) {
2535                         /* A delayed write to unwritten bh should
2536                          * be marked new and mapped.  Mapped ensures
2537                          * that we don't do get_block multiple times
2538                          * when we write to the same offset and new
2539                          * ensures that we do proper zero out for
2540                          * partial write.
2541                          */
2542                         set_buffer_new(bh_result);
2543                         set_buffer_mapped(bh_result);
2544                 }
2545                 ret = 0;
2546         }
2547
2548         return ret;
2549 }
2550
2551 /*
2552  * This function is used as a standard get_block_t calback function
2553  * when there is no desire to allocate any blocks.  It is used as a
2554  * callback function for block_prepare_write(), nobh_writepage(), and
2555  * block_write_full_page().  These functions should only try to map a
2556  * single block at a time.
2557  *
2558  * Since this function doesn't do block allocations even if the caller
2559  * requests it by passing in create=1, it is critically important that
2560  * any caller checks to make sure that any buffer heads are returned
2561  * by this function are either all already mapped or marked for
2562  * delayed allocation before calling nobh_writepage() or
2563  * block_write_full_page().  Otherwise, b_blocknr could be left
2564  * unitialized, and the page write functions will be taken by
2565  * surprise.
2566  */
2567 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2568                                    struct buffer_head *bh_result, int create)
2569 {
2570         int ret = 0;
2571         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2572
2573         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2574
2575         /*
2576          * we don't want to do block allocation in writepage
2577          * so call get_block_wrap with create = 0
2578          */
2579         ret = ext4_get_blocks(NULL, inode, iblock, max_blocks, bh_result, 0);
2580         if (ret > 0) {
2581                 bh_result->b_size = (ret << inode->i_blkbits);
2582                 ret = 0;
2583         }
2584         return ret;
2585 }
2586
2587 static int bget_one(handle_t *handle, struct buffer_head *bh)
2588 {
2589         get_bh(bh);
2590         return 0;
2591 }
2592
2593 static int bput_one(handle_t *handle, struct buffer_head *bh)
2594 {
2595         put_bh(bh);
2596         return 0;
2597 }
2598
2599 static int __ext4_journalled_writepage(struct page *page,
2600                                        unsigned int len)
2601 {
2602         struct address_space *mapping = page->mapping;
2603         struct inode *inode = mapping->host;
2604         struct buffer_head *page_bufs;
2605         handle_t *handle = NULL;
2606         int ret = 0;
2607         int err;
2608
2609         page_bufs = page_buffers(page);
2610         BUG_ON(!page_bufs);
2611         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2612         /* As soon as we unlock the page, it can go away, but we have
2613          * references to buffers so we are safe */
2614         unlock_page(page);
2615
2616         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2617         if (IS_ERR(handle)) {
2618                 ret = PTR_ERR(handle);
2619                 goto out;
2620         }
2621
2622         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2623                                 do_journal_get_write_access);
2624
2625         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2626                                 write_end_fn);
2627         if (ret == 0)
2628                 ret = err;
2629         err = ext4_journal_stop(handle);
2630         if (!ret)
2631                 ret = err;
2632
2633         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2634         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2635 out:
2636         return ret;
2637 }
2638
2639 /*
2640  * Note that we don't need to start a transaction unless we're journaling data
2641  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2642  * need to file the inode to the transaction's list in ordered mode because if
2643  * we are writing back data added by write(), the inode is already there and if
2644  * we are writing back data modified via mmap(), noone guarantees in which
2645  * transaction the data will hit the disk. In case we are journaling data, we
2646  * cannot start transaction directly because transaction start ranks above page
2647  * lock so we have to do some magic.
2648  *
2649  * This function can get called via...
2650  *   - ext4_da_writepages after taking page lock (have journal handle)
2651  *   - journal_submit_inode_data_buffers (no journal handle)
2652  *   - shrink_page_list via pdflush (no journal handle)
2653  *   - grab_page_cache when doing write_begin (have journal handle)
2654  *
2655  * We don't do any block allocation in this function. If we have page with
2656  * multiple blocks we need to write those buffer_heads that are mapped. This
2657  * is important for mmaped based write. So if we do with blocksize 1K
2658  * truncate(f, 1024);
2659  * a = mmap(f, 0, 4096);
2660  * a[0] = 'a';
2661  * truncate(f, 4096);
2662  * we have in the page first buffer_head mapped via page_mkwrite call back
2663  * but other bufer_heads would be unmapped but dirty(dirty done via the
2664  * do_wp_page). So writepage should write the first block. If we modify
2665  * the mmap area beyond 1024 we will again get a page_fault and the
2666  * page_mkwrite callback will do the block allocation and mark the
2667  * buffer_heads mapped.
2668  *
2669  * We redirty the page if we have any buffer_heads that is either delay or
2670  * unwritten in the page.
2671  *
2672  * We can get recursively called as show below.
2673  *
2674  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2675  *              ext4_writepage()
2676  *
2677  * But since we don't do any block allocation we should not deadlock.
2678  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2679  */
2680 static int ext4_writepage(struct page *page,
2681                           struct writeback_control *wbc)
2682 {
2683         int ret = 0;
2684         loff_t size;
2685         unsigned int len;
2686         struct buffer_head *page_bufs;
2687         struct inode *inode = page->mapping->host;
2688
2689         trace_ext4_writepage(inode, page);
2690         size = i_size_read(inode);
2691         if (page->index == size >> PAGE_CACHE_SHIFT)
2692                 len = size & ~PAGE_CACHE_MASK;
2693         else
2694                 len = PAGE_CACHE_SIZE;
2695
2696         if (page_has_buffers(page)) {
2697                 page_bufs = page_buffers(page);
2698                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2699                                         ext4_bh_delay_or_unwritten)) {
2700                         /*
2701                          * We don't want to do  block allocation
2702                          * So redirty the page and return
2703                          * We may reach here when we do a journal commit
2704                          * via journal_submit_inode_data_buffers.
2705                          * If we don't have mapping block we just ignore
2706                          * them. We can also reach here via shrink_page_list
2707                          */
2708                         redirty_page_for_writepage(wbc, page);
2709                         unlock_page(page);
2710                         return 0;
2711                 }
2712         } else {
2713                 /*
2714                  * The test for page_has_buffers() is subtle:
2715                  * We know the page is dirty but it lost buffers. That means
2716                  * that at some moment in time after write_begin()/write_end()
2717                  * has been called all buffers have been clean and thus they
2718                  * must have been written at least once. So they are all
2719                  * mapped and we can happily proceed with mapping them
2720                  * and writing the page.
2721                  *
2722                  * Try to initialize the buffer_heads and check whether
2723                  * all are mapped and non delay. We don't want to
2724                  * do block allocation here.
2725                  */
2726                 ret = block_prepare_write(page, 0, len,
2727                                           noalloc_get_block_write);
2728                 if (!ret) {
2729                         page_bufs = page_buffers(page);
2730                         /* check whether all are mapped and non delay */
2731                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2732                                                 ext4_bh_delay_or_unwritten)) {
2733                                 redirty_page_for_writepage(wbc, page);
2734                                 unlock_page(page);
2735                                 return 0;
2736                         }
2737                 } else {
2738                         /*
2739                          * We can't do block allocation here
2740                          * so just redity the page and unlock
2741                          * and return
2742                          */
2743                         redirty_page_for_writepage(wbc, page);
2744                         unlock_page(page);
2745                         return 0;
2746                 }
2747                 /* now mark the buffer_heads as dirty and uptodate */
2748                 block_commit_write(page, 0, len);
2749         }
2750
2751         if (PageChecked(page) && ext4_should_journal_data(inode)) {
2752                 /*
2753                  * It's mmapped pagecache.  Add buffers and journal it.  There
2754                  * doesn't seem much point in redirtying the page here.
2755                  */
2756                 ClearPageChecked(page);
2757                 return __ext4_journalled_writepage(page, len);
2758         }
2759
2760         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2761                 ret = nobh_writepage(page, noalloc_get_block_write, wbc);
2762         else
2763                 ret = block_write_full_page(page, noalloc_get_block_write,
2764                                             wbc);
2765
2766         return ret;
2767 }
2768
2769 /*
2770  * This is called via ext4_da_writepages() to
2771  * calulate the total number of credits to reserve to fit
2772  * a single extent allocation into a single transaction,
2773  * ext4_da_writpeages() will loop calling this before
2774  * the block allocation.
2775  */
2776
2777 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2778 {
2779         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2780
2781         /*
2782          * With non-extent format the journal credit needed to
2783          * insert nrblocks contiguous block is dependent on
2784          * number of contiguous block. So we will limit
2785          * number of contiguous block to a sane value
2786          */
2787         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) &&
2788             (max_blocks > EXT4_MAX_TRANS_DATA))
2789                 max_blocks = EXT4_MAX_TRANS_DATA;
2790
2791         return ext4_chunk_trans_blocks(inode, max_blocks);
2792 }
2793
2794 static int ext4_da_writepages(struct address_space *mapping,
2795                               struct writeback_control *wbc)
2796 {
2797         pgoff_t index;
2798         int range_whole = 0;
2799         handle_t *handle = NULL;
2800         struct mpage_da_data mpd;
2801         struct inode *inode = mapping->host;
2802         int no_nrwrite_index_update;
2803         int pages_written = 0;
2804         long pages_skipped;
2805         unsigned int max_pages;
2806         int range_cyclic, cycled = 1, io_done = 0;
2807         int needed_blocks, ret = 0;
2808         long desired_nr_to_write, nr_to_writebump = 0;
2809         loff_t range_start = wbc->range_start;
2810         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2811
2812         trace_ext4_da_writepages(inode, wbc);
2813
2814         /*
2815          * No pages to write? This is mainly a kludge to avoid starting
2816          * a transaction for special inodes like journal inode on last iput()
2817          * because that could violate lock ordering on umount
2818          */
2819         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2820                 return 0;
2821
2822         /*
2823          * If the filesystem has aborted, it is read-only, so return
2824          * right away instead of dumping stack traces later on that
2825          * will obscure the real source of the problem.  We test
2826          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2827          * the latter could be true if the filesystem is mounted
2828          * read-only, and in that case, ext4_da_writepages should
2829          * *never* be called, so if that ever happens, we would want
2830          * the stack trace.
2831          */
2832         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2833                 return -EROFS;
2834
2835         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2836                 range_whole = 1;
2837
2838         range_cyclic = wbc->range_cyclic;
2839         if (wbc->range_cyclic) {
2840                 index = mapping->writeback_index;
2841                 if (index)
2842                         cycled = 0;
2843                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2844                 wbc->range_end  = LLONG_MAX;
2845                 wbc->range_cyclic = 0;
2846         } else
2847                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2848
2849         /*
2850          * This works around two forms of stupidity.  The first is in
2851          * the writeback code, which caps the maximum number of pages
2852          * written to be 1024 pages.  This is wrong on multiple
2853          * levels; different architectues have a different page size,
2854          * which changes the maximum amount of data which gets
2855          * written.  Secondly, 4 megabytes is way too small.  XFS
2856          * forces this value to be 16 megabytes by multiplying
2857          * nr_to_write parameter by four, and then relies on its
2858          * allocator to allocate larger extents to make them
2859          * contiguous.  Unfortunately this brings us to the second
2860          * stupidity, which is that ext4's mballoc code only allocates
2861          * at most 2048 blocks.  So we force contiguous writes up to
2862          * the number of dirty blocks in the inode, or
2863          * sbi->max_writeback_mb_bump whichever is smaller.
2864          */
2865         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2866         if (!range_cyclic && range_whole)
2867                 desired_nr_to_write = wbc->nr_to_write * 8;
2868         else
2869                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2870                                                            max_pages);
2871         if (desired_nr_to_write > max_pages)
2872                 desired_nr_to_write = max_pages;
2873
2874         if (wbc->nr_to_write < desired_nr_to_write) {
2875                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2876                 wbc->nr_to_write = desired_nr_to_write;
2877         }
2878
2879         mpd.wbc = wbc;
2880         mpd.inode = mapping->host;
2881
2882         /*
2883          * we don't want write_cache_pages to update
2884          * nr_to_write and writeback_index
2885          */
2886         no_nrwrite_index_update = wbc->no_nrwrite_index_update;
2887         wbc->no_nrwrite_index_update = 1;
2888         pages_skipped = wbc->pages_skipped;
2889
2890 retry:
2891         while (!ret && wbc->nr_to_write > 0) {
2892
2893                 /*
2894                  * we  insert one extent at a time. So we need
2895                  * credit needed for single extent allocation.
2896                  * journalled mode is currently not supported
2897                  * by delalloc
2898                  */
2899                 BUG_ON(ext4_should_journal_data(inode));
2900                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2901
2902                 /* start a new transaction*/
2903                 handle = ext4_journal_start(inode, needed_blocks);
2904                 if (IS_ERR(handle)) {
2905                         ret = PTR_ERR(handle);
2906                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2907                                "%ld pages, ino %lu; err %d\n", __func__,
2908                                 wbc->nr_to_write, inode->i_ino, ret);
2909                         goto out_writepages;
2910                 }
2911
2912                 /*
2913                  * Now call __mpage_da_writepage to find the next
2914                  * contiguous region of logical blocks that need
2915                  * blocks to be allocated by ext4.  We don't actually
2916                  * submit the blocks for I/O here, even though
2917                  * write_cache_pages thinks it will, and will set the
2918                  * pages as clean for write before calling
2919                  * __mpage_da_writepage().
2920                  */
2921                 mpd.b_size = 0;
2922                 mpd.b_state = 0;
2923                 mpd.b_blocknr = 0;
2924                 mpd.first_page = 0;
2925                 mpd.next_page = 0;
2926                 mpd.io_done = 0;
2927                 mpd.pages_written = 0;
2928                 mpd.retval = 0;
2929                 ret = write_cache_pages(mapping, wbc, __mpage_da_writepage,
2930                                         &mpd);
2931                 /*
2932                  * If we have a contiguous extent of pages and we
2933                  * haven't done the I/O yet, map the blocks and submit
2934                  * them for I/O.
2935                  */
2936                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2937                         if (mpage_da_map_blocks(&mpd) == 0)
2938                                 mpage_da_submit_io(&mpd);
2939                         mpd.io_done = 1;
2940                         ret = MPAGE_DA_EXTENT_TAIL;
2941                 }
2942                 trace_ext4_da_write_pages(inode, &mpd);
2943                 wbc->nr_to_write -= mpd.pages_written;
2944
2945                 ext4_journal_stop(handle);
2946
2947                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2948                         /* commit the transaction which would
2949                          * free blocks released in the transaction
2950                          * and try again
2951                          */
2952                         jbd2_journal_force_commit_nested(sbi->s_journal);
2953                         wbc->pages_skipped = pages_skipped;
2954                         ret = 0;
2955                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2956                         /*
2957                          * got one extent now try with
2958                          * rest of the pages
2959                          */
2960                         pages_written += mpd.pages_written;
2961                         wbc->pages_skipped = pages_skipped;
2962                         ret = 0;
2963                         io_done = 1;
2964                 } else if (wbc->nr_to_write)
2965                         /*
2966                          * There is no more writeout needed
2967                          * or we requested for a noblocking writeout
2968                          * and we found the device congested
2969                          */
2970                         break;
2971         }
2972         if (!io_done && !cycled) {
2973                 cycled = 1;
2974                 index = 0;
2975                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2976                 wbc->range_end  = mapping->writeback_index - 1;
2977                 goto retry;
2978         }
2979         if (pages_skipped != wbc->pages_skipped)
2980                 ext4_msg(inode->i_sb, KERN_CRIT,
2981                          "This should not happen leaving %s "
2982                          "with nr_to_write = %ld ret = %d\n",
2983                          __func__, wbc->nr_to_write, ret);
2984
2985         /* Update index */
2986         index += pages_written;
2987         wbc->range_cyclic = range_cyclic;
2988         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2989                 /*
2990                  * set the writeback_index so that range_cyclic
2991                  * mode will write it back later
2992                  */
2993                 mapping->writeback_index = index;
2994
2995 out_writepages:
2996         if (!no_nrwrite_index_update)
2997                 wbc->no_nrwrite_index_update = 0;
2998         wbc->nr_to_write -= nr_to_writebump;
2999         wbc->range_start = range_start;
3000         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
3001         return ret;
3002 }
3003
3004 #define FALL_BACK_TO_NONDELALLOC 1
3005 static int ext4_nonda_switch(struct super_block *sb)
3006 {
3007         s64 free_blocks, dirty_blocks;
3008         struct ext4_sb_info *sbi = EXT4_SB(sb);
3009
3010         /*
3011          * switch to non delalloc mode if we are running low
3012          * on free block. The free block accounting via percpu
3013          * counters can get slightly wrong with percpu_counter_batch getting
3014          * accumulated on each CPU without updating global counters
3015          * Delalloc need an accurate free block accounting. So switch
3016          * to non delalloc when we are near to error range.
3017          */
3018         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
3019         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
3020         if (2 * free_blocks < 3 * dirty_blocks ||
3021                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
3022                 /*
3023                  * free block count is less than 150% of dirty blocks
3024                  * or free blocks is less than watermark
3025                  */
3026                 return 1;
3027         }
3028         /*
3029          * Even if we don't switch but are nearing capacity,
3030          * start pushing delalloc when 1/2 of free blocks are dirty.
3031          */
3032         if (free_blocks < 2 * dirty_blocks)
3033                 writeback_inodes_sb_if_idle(sb);
3034
3035         return 0;
3036 }
3037
3038 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3039                                loff_t pos, unsigned len, unsigned flags,
3040                                struct page **pagep, void **fsdata)
3041 {
3042         int ret, retries = 0, quota_retries = 0;
3043         struct page *page;
3044         pgoff_t index;
3045         unsigned from, to;
3046         struct inode *inode = mapping->host;
3047         handle_t *handle;
3048
3049         index = pos >> PAGE_CACHE_SHIFT;
3050         from = pos & (PAGE_CACHE_SIZE - 1);
3051         to = from + len;
3052
3053         if (ext4_nonda_switch(inode->i_sb)) {
3054                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3055                 return ext4_write_begin(file, mapping, pos,
3056                                         len, flags, pagep, fsdata);
3057         }
3058         *fsdata = (void *)0;
3059         trace_ext4_da_write_begin(inode, pos, len, flags);
3060 retry:
3061         /*
3062          * With delayed allocation, we don't log the i_disksize update
3063          * if there is delayed block allocation. But we still need
3064          * to journalling the i_disksize update if writes to the end
3065          * of file which has an already mapped buffer.
3066          */
3067         handle = ext4_journal_start(inode, 1);
3068         if (IS_ERR(handle)) {
3069                 ret = PTR_ERR(handle);
3070                 goto out;
3071         }
3072         /* We cannot recurse into the filesystem as the transaction is already
3073          * started */
3074         flags |= AOP_FLAG_NOFS;
3075
3076         page = grab_cache_page_write_begin(mapping, index, flags);
3077         if (!page) {
3078                 ext4_journal_stop(handle);
3079                 ret = -ENOMEM;
3080                 goto out;
3081         }
3082         *pagep = page;
3083
3084         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
3085                                 ext4_da_get_block_prep);
3086         if (ret < 0) {
3087                 unlock_page(page);
3088                 ext4_journal_stop(handle);
3089                 page_cache_release(page);
3090                 /*
3091                  * block_write_begin may have instantiated a few blocks
3092                  * outside i_size.  Trim these off again. Don't need
3093                  * i_size_read because we hold i_mutex.
3094                  */
3095                 if (pos + len > inode->i_size)
3096                         ext4_truncate_failed_write(inode);
3097         }
3098
3099         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3100                 goto retry;
3101
3102         if ((ret == -EDQUOT) &&
3103             EXT4_I(inode)->i_reserved_meta_blocks &&
3104             (quota_retries++ < 3)) {
3105                 /*
3106                  * Since we often over-estimate the number of meta
3107                  * data blocks required, we may sometimes get a
3108                  * spurios out of quota error even though there would
3109                  * be enough space once we write the data blocks and
3110                  * find out how many meta data blocks were _really_
3111                  * required.  So try forcing the inode write to see if
3112                  * that helps.
3113                  */
3114                 write_inode_now(inode, (quota_retries == 3));
3115                 goto retry;
3116         }
3117 out:
3118         return ret;
3119 }
3120
3121 /*
3122  * Check if we should update i_disksize
3123  * when write to the end of file but not require block allocation
3124  */
3125 static int ext4_da_should_update_i_disksize(struct page *page,
3126                                             unsigned long offset)
3127 {
3128         struct buffer_head *bh;
3129         struct inode *inode = page->mapping->host;
3130         unsigned int idx;
3131         int i;
3132
3133         bh = page_buffers(page);
3134         idx = offset >> inode->i_blkbits;
3135
3136         for (i = 0; i < idx; i++)
3137                 bh = bh->b_this_page;
3138
3139         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3140                 return 0;
3141         return 1;
3142 }
3143
3144 static int ext4_da_write_end(struct file *file,
3145                              struct address_space *mapping,
3146                              loff_t pos, unsigned len, unsigned copied,
3147                              struct page *page, void *fsdata)
3148 {
3149         struct inode *inode = mapping->host;
3150         int ret = 0, ret2;
3151         handle_t *handle = ext4_journal_current_handle();
3152         loff_t new_i_size;
3153         unsigned long start, end;
3154         int write_mode = (int)(unsigned long)fsdata;
3155
3156         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
3157                 if (ext4_should_order_data(inode)) {
3158                         return ext4_ordered_write_end(file, mapping, pos,
3159                                         len, copied, page, fsdata);
3160                 } else if (ext4_should_writeback_data(inode)) {
3161                         return ext4_writeback_write_end(file, mapping, pos,
3162                                         len, copied, page, fsdata);
3163                 } else {
3164                         BUG();
3165                 }
3166         }
3167
3168         trace_ext4_da_write_end(inode, pos, len, copied);
3169         start = pos & (PAGE_CACHE_SIZE - 1);
3170         end = start + copied - 1;
3171
3172         /*
3173          * generic_write_end() will run mark_inode_dirty() if i_size
3174          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3175          * into that.
3176          */
3177
3178         new_i_size = pos + copied;
3179         if (new_i_size > EXT4_I(inode)->i_disksize) {
3180                 if (ext4_da_should_update_i_disksize(page, end)) {
3181                         down_write(&EXT4_I(inode)->i_data_sem);
3182                         if (new_i_size > EXT4_I(inode)->i_disksize) {
3183                                 /*
3184                                  * Updating i_disksize when extending file
3185                                  * without needing block allocation
3186                                  */
3187                                 if (ext4_should_order_data(inode))
3188                                         ret = ext4_jbd2_file_inode(handle,
3189                                                                    inode);
3190
3191                                 EXT4_I(inode)->i_disksize = new_i_size;
3192                         }
3193                         up_write(&EXT4_I(inode)->i_data_sem);
3194                         /* We need to mark inode dirty even if
3195                          * new_i_size is less that inode->i_size
3196                          * bu greater than i_disksize.(hint delalloc)
3197                          */
3198                         ext4_mark_inode_dirty(handle, inode);
3199                 }
3200         }
3201         ret2 = generic_write_end(file, mapping, pos, len, copied,
3202                                                         page, fsdata);
3203         copied = ret2;
3204         if (ret2 < 0)
3205                 ret = ret2;
3206         ret2 = ext4_journal_stop(handle);
3207         if (!ret)
3208                 ret = ret2;
3209
3210         return ret ? ret : copied;
3211 }
3212
3213 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
3214 {
3215         /*
3216          * Drop reserved blocks
3217          */
3218         BUG_ON(!PageLocked(page));
3219         if (!page_has_buffers(page))
3220                 goto out;
3221
3222         ext4_da_page_release_reservation(page, offset);
3223
3224 out:
3225         ext4_invalidatepage(page, offset);
3226
3227         return;
3228 }
3229
3230 /*
3231  * Force all delayed allocation blocks to be allocated for a given inode.
3232  */
3233 int ext4_alloc_da_blocks(struct inode *inode)
3234 {
3235         trace_ext4_alloc_da_blocks(inode);
3236
3237         if (!EXT4_I(inode)->i_reserved_data_blocks &&
3238             !EXT4_I(inode)->i_reserved_meta_blocks)
3239                 return 0;
3240
3241         /*
3242          * We do something simple for now.  The filemap_flush() will
3243          * also start triggering a write of the data blocks, which is
3244          * not strictly speaking necessary (and for users of
3245          * laptop_mode, not even desirable).  However, to do otherwise
3246          * would require replicating code paths in:
3247          *
3248          * ext4_da_writepages() ->
3249          *    write_cache_pages() ---> (via passed in callback function)
3250          *        __mpage_da_writepage() -->
3251          *           mpage_add_bh_to_extent()
3252          *           mpage_da_map_blocks()
3253          *
3254          * The problem is that write_cache_pages(), located in
3255          * mm/page-writeback.c, marks pages clean in preparation for
3256          * doing I/O, which is not desirable if we're not planning on
3257          * doing I/O at all.
3258          *
3259          * We could call write_cache_pages(), and then redirty all of
3260          * the pages by calling redirty_page_for_writeback() but that
3261          * would be ugly in the extreme.  So instead we would need to
3262          * replicate parts of the code in the above functions,
3263          * simplifying them becuase we wouldn't actually intend to
3264          * write out the pages, but rather only collect contiguous
3265          * logical block extents, call the multi-block allocator, and
3266          * then update the buffer heads with the block allocations.
3267          *
3268          * For now, though, we'll cheat by calling filemap_flush(),
3269          * which will map the blocks, and start the I/O, but not
3270          * actually wait for the I/O to complete.
3271          */
3272         return filemap_flush(inode->i_mapping);
3273 }
3274
3275 /*
3276  * bmap() is special.  It gets used by applications such as lilo and by
3277  * the swapper to find the on-disk block of a specific piece of data.
3278  *
3279  * Naturally, this is dangerous if the block concerned is still in the
3280  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3281  * filesystem and enables swap, then they may get a nasty shock when the
3282  * data getting swapped to that swapfile suddenly gets overwritten by
3283  * the original zero's written out previously to the journal and
3284  * awaiting writeback in the kernel's buffer cache.
3285  *
3286  * So, if we see any bmap calls here on a modified, data-journaled file,
3287  * take extra steps to flush any blocks which might be in the cache.
3288  */
3289 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3290 {
3291         struct inode *inode = mapping->host;
3292         journal_t *journal;
3293         int err;
3294
3295         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3296                         test_opt(inode->i_sb, DELALLOC)) {
3297                 /*
3298                  * With delalloc we want to sync the file
3299                  * so that we can make sure we allocate
3300                  * blocks for file
3301                  */
3302                 filemap_write_and_wait(mapping);
3303         }
3304
3305         if (EXT4_JOURNAL(inode) &&
3306             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3307                 /*
3308                  * This is a REALLY heavyweight approach, but the use of
3309                  * bmap on dirty files is expected to be extremely rare:
3310                  * only if we run lilo or swapon on a freshly made file
3311                  * do we expect this to happen.
3312                  *
3313                  * (bmap requires CAP_SYS_RAWIO so this does not
3314                  * represent an unprivileged user DOS attack --- we'd be
3315                  * in trouble if mortal users could trigger this path at
3316                  * will.)
3317                  *
3318                  * NB. EXT4_STATE_JDATA is not set on files other than
3319                  * regular files.  If somebody wants to bmap a directory
3320                  * or symlink and gets confused because the buffer
3321                  * hasn't yet been flushed to disk, they deserve
3322                  * everything they get.
3323                  */
3324
3325                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3326                 journal = EXT4_JOURNAL(inode);
3327                 jbd2_journal_lock_updates(journal);
3328                 err = jbd2_journal_flush(journal);
3329                 jbd2_journal_unlock_updates(journal);
3330
3331                 if (err)
3332                         return 0;
3333         }
3334
3335         return generic_block_bmap(mapping, block, ext4_get_block);
3336 }
3337
3338 static int ext4_readpage(struct file *file, struct page *page)
3339 {
3340         return mpage_readpage(page, ext4_get_block);
3341 }
3342
3343 static int
3344 ext4_readpages(struct file *file, struct address_space *mapping,
3345                 struct list_head *pages, unsigned nr_pages)
3346 {
3347         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3348 }
3349
3350 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3351 {
3352         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3353
3354         /*
3355          * If it's a full truncate we just forget about the pending dirtying
3356          */
3357         if (offset == 0)
3358                 ClearPageChecked(page);
3359
3360         if (journal)
3361                 jbd2_journal_invalidatepage(journal, page, offset);
3362         else
3363                 block_invalidatepage(page, offset);
3364 }
3365
3366 static int ext4_releasepage(struct page *page, gfp_t wait)
3367 {
3368         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3369
3370         WARN_ON(PageChecked(page));
3371         if (!page_has_buffers(page))
3372                 return 0;
3373         if (journal)
3374                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3375         else
3376                 return try_to_free_buffers(page);
3377 }
3378
3379 /*
3380  * O_DIRECT for ext3 (or indirect map) based files
3381  *
3382  * If the O_DIRECT write will extend the file then add this inode to the
3383  * orphan list.  So recovery will truncate it back to the original size
3384  * if the machine crashes during the write.
3385  *
3386  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3387  * crashes then stale disk data _may_ be exposed inside the file. But current
3388  * VFS code falls back into buffered path in that case so we are safe.
3389  */
3390 static ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
3391                               const struct iovec *iov, loff_t offset,
3392                               unsigned long nr_segs)
3393 {
3394         struct file *file = iocb->ki_filp;
3395         struct inode *inode = file->f_mapping->host;
3396         struct ext4_inode_info *ei = EXT4_I(inode);
3397         handle_t *handle;
3398         ssize_t ret;
3399         int orphan = 0;
3400         size_t count = iov_length(iov, nr_segs);
3401         int retries = 0;
3402
3403         if (rw == WRITE) {
3404                 loff_t final_size = offset + count;
3405
3406                 if (final_size > inode->i_size) {
3407                         /* Credits for sb + inode write */
3408                         handle = ext4_journal_start(inode, 2);
3409                         if (IS_ERR(handle)) {
3410                                 ret = PTR_ERR(handle);
3411                                 goto out;
3412                         }
3413                         ret = ext4_orphan_add(handle, inode);
3414                         if (ret) {
3415                                 ext4_journal_stop(handle);
3416                                 goto out;
3417                         }
3418                         orphan = 1;
3419                         ei->i_disksize = inode->i_size;
3420                         ext4_journal_stop(handle);
3421                 }
3422         }
3423
3424 retry:
3425         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
3426                                  offset, nr_segs,
3427                                  ext4_get_block, NULL);
3428         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3429                 goto retry;
3430
3431         if (orphan) {
3432                 int err;
3433
3434                 /* Credits for sb + inode write */
3435                 handle = ext4_journal_start(inode, 2);
3436                 if (IS_ERR(handle)) {
3437                         /* This is really bad luck. We've written the data
3438                          * but cannot extend i_size. Bail out and pretend
3439                          * the write failed... */
3440                         ret = PTR_ERR(handle);
3441                         goto out;
3442                 }
3443                 if (inode->i_nlink)
3444                         ext4_orphan_del(handle, inode);
3445                 if (ret > 0) {
3446                         loff_t end = offset + ret;
3447                         if (end > inode->i_size) {
3448                                 ei->i_disksize = end;
3449                                 i_size_write(inode, end);
3450                                 /*
3451                                  * We're going to return a positive `ret'
3452                                  * here due to non-zero-length I/O, so there's
3453                                  * no way of reporting error returns from
3454                                  * ext4_mark_inode_dirty() to userspace.  So
3455                                  * ignore it.
3456                                  */
3457                                 ext4_mark_inode_dirty(handle, inode);
3458                         }
3459                 }
3460                 err = ext4_journal_stop(handle);
3461                 if (ret == 0)
3462                         ret = err;
3463         }
3464 out:
3465         return ret;
3466 }
3467
3468 static int ext4_get_block_dio_write(struct inode *inode, sector_t iblock,
3469                    struct buffer_head *bh_result, int create)
3470 {
3471         handle_t *handle = NULL;
3472         int ret = 0;
3473         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
3474         int dio_credits;
3475
3476         ext4_debug("ext4_get_block_dio_write: inode %lu, create flag %d\n",
3477                    inode->i_ino, create);
3478         /*
3479          * DIO VFS code passes create = 0 flag for write to
3480          * the middle of file. It does this to avoid block
3481          * allocation for holes, to prevent expose stale data
3482          * out when there is parallel buffered read (which does
3483          * not hold the i_mutex lock) while direct IO write has
3484          * not completed. DIO request on holes finally falls back
3485          * to buffered IO for this reason.
3486          *
3487          * For ext4 extent based file, since we support fallocate,
3488          * new allocated extent as uninitialized, for holes, we
3489          * could fallocate blocks for holes, thus parallel
3490          * buffered IO read will zero out the page when read on
3491          * a hole while parallel DIO write to the hole has not completed.
3492          *
3493          * when we come here, we know it's a direct IO write to
3494          * to the middle of file (<i_size)
3495          * so it's safe to override the create flag from VFS.
3496          */
3497         create = EXT4_GET_BLOCKS_DIO_CREATE_EXT;
3498
3499         if (max_blocks > DIO_MAX_BLOCKS)
3500                 max_blocks = DIO_MAX_BLOCKS;
3501         dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
3502         handle = ext4_journal_start(inode, dio_credits);
3503         if (IS_ERR(handle)) {
3504                 ret = PTR_ERR(handle);
3505                 goto out;
3506         }
3507         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
3508                               create);
3509         if (ret > 0) {
3510                 bh_result->b_size = (ret << inode->i_blkbits);
3511                 ret = 0;
3512         }
3513         ext4_journal_stop(handle);
3514 out:
3515         return ret;
3516 }
3517
3518 static void ext4_free_io_end(ext4_io_end_t *io)
3519 {
3520         BUG_ON(!io);
3521         iput(io->inode);
3522         kfree(io);
3523 }
3524 static void dump_aio_dio_list(struct inode * inode)
3525 {
3526 #ifdef  EXT4_DEBUG
3527         struct list_head *cur, *before, *after;
3528         ext4_io_end_t *io, *io0, *io1;
3529
3530         if (list_empty(&EXT4_I(inode)->i_aio_dio_complete_list)){
3531                 ext4_debug("inode %lu aio dio list is empty\n", inode->i_ino);
3532                 return;
3533         }
3534
3535         ext4_debug("Dump inode %lu aio_dio_completed_IO list \n", inode->i_ino);
3536         list_for_each_entry(io, &EXT4_I(inode)->i_aio_dio_complete_list, list){
3537                 cur = &io->list;
3538                 before = cur->prev;
3539                 io0 = container_of(before, ext4_io_end_t, list);
3540                 after = cur->next;
3541                 io1 = container_of(after, ext4_io_end_t, list);
3542
3543                 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
3544                             io, inode->i_ino, io0, io1);
3545         }
3546 #endif
3547 }
3548
3549 /*
3550  * check a range of space and convert unwritten extents to written.
3551  */
3552 static int ext4_end_aio_dio_nolock(ext4_io_end_t *io)
3553 {
3554         struct inode *inode = io->inode;
3555         loff_t offset = io->offset;
3556         ssize_t size = io->size;
3557         int ret = 0;
3558
3559         ext4_debug("end_aio_dio_onlock: io 0x%p from inode %lu,list->next 0x%p,"
3560                    "list->prev 0x%p\n",
3561                    io, inode->i_ino, io->list.next, io->list.prev);
3562
3563         if (list_empty(&io->list))
3564                 return ret;
3565
3566         if (io->flag != DIO_AIO_UNWRITTEN)
3567                 return ret;
3568
3569         if (offset + size <= i_size_read(inode))
3570                 ret = ext4_convert_unwritten_extents(inode, offset, size);
3571
3572         if (ret < 0) {
3573                 printk(KERN_EMERG "%s: failed to convert unwritten"
3574                         "extents to written extents, error is %d"
3575                         " io is still on inode %lu aio dio list\n",
3576                        __func__, ret, inode->i_ino);
3577                 return ret;
3578         }
3579
3580         /* clear the DIO AIO unwritten flag */
3581         io->flag = 0;
3582         return ret;
3583 }
3584 /*
3585  * work on completed aio dio IO, to convert unwritten extents to extents
3586  */
3587 static void ext4_end_aio_dio_work(struct work_struct *work)
3588 {
3589         ext4_io_end_t *io  = container_of(work, ext4_io_end_t, work);
3590         struct inode *inode = io->inode;
3591         int ret = 0;
3592
3593         mutex_lock(&inode->i_mutex);
3594         ret = ext4_end_aio_dio_nolock(io);
3595         if (ret >= 0) {
3596                 if (!list_empty(&io->list))
3597                         list_del_init(&io->list);
3598                 ext4_free_io_end(io);
3599         }
3600         mutex_unlock(&inode->i_mutex);
3601 }
3602 /*
3603  * This function is called from ext4_sync_file().
3604  *
3605  * When AIO DIO IO is completed, the work to convert unwritten
3606  * extents to written is queued on workqueue but may not get immediately
3607  * scheduled. When fsync is called, we need to ensure the
3608  * conversion is complete before fsync returns.
3609  * The inode keeps track of a list of completed AIO from DIO path
3610  * that might needs to do the conversion. This function walks through
3611  * the list and convert the related unwritten extents to written.
3612  */
3613 int flush_aio_dio_completed_IO(struct inode *inode)
3614 {
3615         ext4_io_end_t *io;
3616         int ret = 0;
3617         int ret2 = 0;
3618
3619         if (list_empty(&EXT4_I(inode)->i_aio_dio_complete_list))
3620                 return ret;
3621
3622         dump_aio_dio_list(inode);
3623         while (!list_empty(&EXT4_I(inode)->i_aio_dio_complete_list)){
3624                 io = list_entry(EXT4_I(inode)->i_aio_dio_complete_list.next,
3625                                 ext4_io_end_t, list);
3626                 /*
3627                  * Calling ext4_end_aio_dio_nolock() to convert completed
3628                  * IO to written.
3629                  *
3630                  * When ext4_sync_file() is called, run_queue() may already
3631                  * about to flush the work corresponding to this io structure.
3632                  * It will be upset if it founds the io structure related
3633                  * to the work-to-be schedule is freed.
3634                  *
3635                  * Thus we need to keep the io structure still valid here after
3636                  * convertion finished. The io structure has a flag to
3637                  * avoid double converting from both fsync and background work
3638                  * queue work.
3639                  */
3640                 ret = ext4_end_aio_dio_nolock(io);
3641                 if (ret < 0)
3642                         ret2 = ret;
3643                 else
3644                         list_del_init(&io->list);
3645         }
3646         return (ret2 < 0) ? ret2 : 0;
3647 }
3648
3649 static ext4_io_end_t *ext4_init_io_end (struct inode *inode)
3650 {
3651         ext4_io_end_t *io = NULL;
3652
3653         io = kmalloc(sizeof(*io), GFP_NOFS);
3654
3655         if (io) {
3656                 igrab(inode);
3657                 io->inode = inode;
3658                 io->flag = 0;
3659                 io->offset = 0;
3660                 io->size = 0;
3661                 io->error = 0;
3662                 INIT_WORK(&io->work, ext4_end_aio_dio_work);
3663                 INIT_LIST_HEAD(&io->list);
3664         }
3665
3666         return io;
3667 }
3668
3669 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3670                             ssize_t size, void *private)
3671 {
3672         ext4_io_end_t *io_end = iocb->private;
3673         struct workqueue_struct *wq;
3674
3675         /* if not async direct IO or dio with 0 bytes write, just return */
3676         if (!io_end || !size)
3677                 return;
3678
3679         ext_debug("ext4_end_io_dio(): io_end 0x%p"
3680                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
3681                   iocb->private, io_end->inode->i_ino, iocb, offset,
3682                   size);
3683
3684         /* if not aio dio with unwritten extents, just free io and return */
3685         if (io_end->flag != DIO_AIO_UNWRITTEN){
3686                 ext4_free_io_end(io_end);
3687                 iocb->private = NULL;
3688                 return;
3689         }
3690
3691         io_end->offset = offset;
3692         io_end->size = size;
3693         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
3694
3695         /* queue the work to convert unwritten extents to written */
3696         queue_work(wq, &io_end->work);
3697
3698         /* Add the io_end to per-inode completed aio dio list*/
3699         list_add_tail(&io_end->list,
3700                  &EXT4_I(io_end->inode)->i_aio_dio_complete_list);
3701         iocb->private = NULL;
3702 }
3703 /*
3704  * For ext4 extent files, ext4 will do direct-io write to holes,
3705  * preallocated extents, and those write extend the file, no need to
3706  * fall back to buffered IO.
3707  *
3708  * For holes, we fallocate those blocks, mark them as unintialized
3709  * If those blocks were preallocated, we mark sure they are splited, but
3710  * still keep the range to write as unintialized.
3711  *
3712  * The unwrritten extents will be converted to written when DIO is completed.
3713  * For async direct IO, since the IO may still pending when return, we
3714  * set up an end_io call back function, which will do the convertion
3715  * when async direct IO completed.
3716  *
3717  * If the O_DIRECT write will extend the file then add this inode to the
3718  * orphan list.  So recovery will truncate it back to the original size
3719  * if the machine crashes during the write.
3720  *
3721  */
3722 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
3723                               const struct iovec *iov, loff_t offset,
3724                               unsigned long nr_segs)
3725 {
3726         struct file *file = iocb->ki_filp;
3727         struct inode *inode = file->f_mapping->host;
3728         ssize_t ret;
3729         size_t count = iov_length(iov, nr_segs);
3730
3731         loff_t final_size = offset + count;
3732         if (rw == WRITE && final_size <= inode->i_size) {
3733                 /*
3734                  * We could direct write to holes and fallocate.
3735                  *
3736                  * Allocated blocks to fill the hole are marked as uninitialized
3737                  * to prevent paralel buffered read to expose the stale data
3738                  * before DIO complete the data IO.
3739                  *
3740                  * As to previously fallocated extents, ext4 get_block
3741                  * will just simply mark the buffer mapped but still
3742                  * keep the extents uninitialized.
3743                  *
3744                  * for non AIO case, we will convert those unwritten extents
3745                  * to written after return back from blockdev_direct_IO.
3746                  *
3747                  * for async DIO, the conversion needs to be defered when
3748                  * the IO is completed. The ext4 end_io callback function
3749                  * will be called to take care of the conversion work.
3750                  * Here for async case, we allocate an io_end structure to
3751                  * hook to the iocb.
3752                  */
3753                 iocb->private = NULL;
3754                 EXT4_I(inode)->cur_aio_dio = NULL;
3755                 if (!is_sync_kiocb(iocb)) {
3756                         iocb->private = ext4_init_io_end(inode);
3757                         if (!iocb->private)
3758                                 return -ENOMEM;
3759                         /*
3760                          * we save the io structure for current async
3761                          * direct IO, so that later ext4_get_blocks()
3762                          * could flag the io structure whether there
3763                          * is a unwritten extents needs to be converted
3764                          * when IO is completed.
3765                          */
3766                         EXT4_I(inode)->cur_aio_dio = iocb->private;
3767                 }
3768
3769                 ret = blockdev_direct_IO(rw, iocb, inode,
3770                                          inode->i_sb->s_bdev, iov,
3771                                          offset, nr_segs,
3772                                          ext4_get_block_dio_write,
3773                                          ext4_end_io_dio);
3774                 if (iocb->private)
3775                         EXT4_I(inode)->cur_aio_dio = NULL;
3776                 /*
3777                  * The io_end structure takes a reference to the inode,
3778                  * that structure needs to be destroyed and the
3779                  * reference to the inode need to be dropped, when IO is
3780                  * complete, even with 0 byte write, or failed.
3781                  *
3782                  * In the successful AIO DIO case, the io_end structure will be
3783                  * desctroyed and the reference to the inode will be dropped
3784                  * after the end_io call back function is called.
3785                  *
3786                  * In the case there is 0 byte write, or error case, since
3787                  * VFS direct IO won't invoke the end_io call back function,
3788                  * we need to free the end_io structure here.
3789                  */
3790                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
3791                         ext4_free_io_end(iocb->private);
3792                         iocb->private = NULL;
3793                 } else if (ret > 0 && ext4_test_inode_state(inode,
3794                                                 EXT4_STATE_DIO_UNWRITTEN)) {
3795                         int err;
3796                         /*
3797                          * for non AIO case, since the IO is already
3798                          * completed, we could do the convertion right here
3799                          */
3800                         err = ext4_convert_unwritten_extents(inode,
3801                                                              offset, ret);
3802                         if (err < 0)
3803                                 ret = err;
3804                         ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3805                 }
3806                 return ret;
3807         }
3808
3809         /* for write the the end of file case, we fall back to old way */
3810         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3811 }
3812
3813 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3814                               const struct iovec *iov, loff_t offset,
3815                               unsigned long nr_segs)
3816 {
3817         struct file *file = iocb->ki_filp;
3818         struct inode *inode = file->f_mapping->host;
3819
3820         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
3821                 return ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
3822
3823         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3824 }
3825
3826 /*
3827  * Pages can be marked dirty completely asynchronously from ext4's journalling
3828  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3829  * much here because ->set_page_dirty is called under VFS locks.  The page is
3830  * not necessarily locked.
3831  *
3832  * We cannot just dirty the page and leave attached buffers clean, because the
3833  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3834  * or jbddirty because all the journalling code will explode.
3835  *
3836  * So what we do is to mark the page "pending dirty" and next time writepage
3837  * is called, propagate that into the buffers appropriately.
3838  */
3839 static int ext4_journalled_set_page_dirty(struct page *page)
3840 {
3841         SetPageChecked(page);
3842         return __set_page_dirty_nobuffers(page);
3843 }
3844
3845 static const struct address_space_operations ext4_ordered_aops = {
3846         .readpage               = ext4_readpage,
3847         .readpages              = ext4_readpages,
3848         .writepage              = ext4_writepage,
3849         .sync_page              = block_sync_page,
3850         .write_begin            = ext4_write_begin,
3851         .write_end              = ext4_ordered_write_end,
3852         .bmap                   = ext4_bmap,
3853         .invalidatepage         = ext4_invalidatepage,
3854         .releasepage            = ext4_releasepage,
3855         .direct_IO              = ext4_direct_IO,
3856         .migratepage            = buffer_migrate_page,
3857         .is_partially_uptodate  = block_is_partially_uptodate,
3858         .error_remove_page      = generic_error_remove_page,
3859 };
3860
3861 static const struct address_space_operations ext4_writeback_aops = {
3862         .readpage               = ext4_readpage,
3863         .readpages              = ext4_readpages,
3864         .writepage              = ext4_writepage,
3865         .sync_page              = block_sync_page,
3866         .write_begin            = ext4_write_begin,
3867         .write_end              = ext4_writeback_write_end,
3868         .bmap                   = ext4_bmap,
3869         .invalidatepage         = ext4_invalidatepage,
3870         .releasepage            = ext4_releasepage,
3871         .direct_IO              = ext4_direct_IO,
3872         .migratepage            = buffer_migrate_page,
3873         .is_partially_uptodate  = block_is_partially_uptodate,
3874         .error_remove_page      = generic_error_remove_page,
3875 };
3876
3877 static const struct address_space_operations ext4_journalled_aops = {
3878         .readpage               = ext4_readpage,
3879         .readpages              = ext4_readpages,
3880         .writepage              = ext4_writepage,
3881         .sync_page              = block_sync_page,
3882         .write_begin            = ext4_write_begin,
3883         .write_end              = ext4_journalled_write_end,
3884         .set_page_dirty         = ext4_journalled_set_page_dirty,
3885         .bmap                   = ext4_bmap,
3886         .invalidatepage         = ext4_invalidatepage,
3887         .releasepage            = ext4_releasepage,
3888         .is_partially_uptodate  = block_is_partially_uptodate,
3889         .error_remove_page      = generic_error_remove_page,
3890 };
3891
3892 static const struct address_space_operations ext4_da_aops = {
3893         .readpage               = ext4_readpage,
3894         .readpages              = ext4_readpages,
3895         .writepage              = ext4_writepage,
3896         .writepages             = ext4_da_writepages,
3897         .sync_page              = block_sync_page,
3898         .write_begin            = ext4_da_write_begin,
3899         .write_end              = ext4_da_write_end,
3900         .bmap                   = ext4_bmap,
3901         .invalidatepage         = ext4_da_invalidatepage,
3902         .releasepage            = ext4_releasepage,
3903         .direct_IO              = ext4_direct_IO,
3904         .migratepage            = buffer_migrate_page,
3905         .is_partially_uptodate  = block_is_partially_uptodate,
3906         .error_remove_page      = generic_error_remove_page,
3907 };
3908
3909 void ext4_set_aops(struct inode *inode)
3910 {
3911         if (ext4_should_order_data(inode) &&
3912                 test_opt(inode->i_sb, DELALLOC))
3913                 inode->i_mapping->a_ops = &ext4_da_aops;
3914         else if (ext4_should_order_data(inode))
3915                 inode->i_mapping->a_ops = &ext4_ordered_aops;
3916         else if (ext4_should_writeback_data(inode) &&
3917                  test_opt(inode->i_sb, DELALLOC))
3918                 inode->i_mapping->a_ops = &ext4_da_aops;
3919         else if (ext4_should_writeback_data(inode))
3920                 inode->i_mapping->a_ops = &ext4_writeback_aops;
3921         else
3922                 inode->i_mapping->a_ops = &ext4_journalled_aops;
3923 }
3924
3925 /*
3926  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3927  * up to the end of the block which corresponds to `from'.
3928  * This required during truncate. We need to physically zero the tail end
3929  * of that block so it doesn't yield old data if the file is later grown.
3930  */
3931 int ext4_block_truncate_page(handle_t *handle,
3932                 struct address_space *mapping, loff_t from)
3933 {
3934         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3935         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3936         unsigned blocksize, length, pos;
3937         ext4_lblk_t iblock;
3938         struct inode *inode = mapping->host;
3939         struct buffer_head *bh;
3940         struct page *page;
3941         int err = 0;
3942
3943         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
3944                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
3945         if (!page)
3946                 return -EINVAL;
3947
3948         blocksize = inode->i_sb->s_blocksize;
3949         length = blocksize - (offset & (blocksize - 1));
3950         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3951
3952         /*
3953          * For "nobh" option,  we can only work if we don't need to
3954          * read-in the page - otherwise we create buffers to do the IO.
3955          */
3956         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
3957              ext4_should_writeback_data(inode) && PageUptodate(page)) {
3958                 zero_user(page, offset, length);
3959                 set_page_dirty(page);
3960                 goto unlock;
3961         }
3962
3963         if (!page_has_buffers(page))
3964                 create_empty_buffers(page, blocksize, 0);
3965
3966         /* Find the buffer that contains "offset" */
3967         bh = page_buffers(page);
3968         pos = blocksize;
3969         while (offset >= pos) {
3970                 bh = bh->b_this_page;
3971                 iblock++;
3972                 pos += blocksize;
3973         }
3974
3975         err = 0;
3976         if (buffer_freed(bh)) {
3977                 BUFFER_TRACE(bh, "freed: skip");
3978                 goto unlock;
3979         }
3980
3981         if (!buffer_mapped(bh)) {
3982                 BUFFER_TRACE(bh, "unmapped");
3983                 ext4_get_block(inode, iblock, bh, 0);
3984                 /* unmapped? It's a hole - nothing to do */
3985                 if (!buffer_mapped(bh)) {
3986                         BUFFER_TRACE(bh, "still unmapped");
3987                         goto unlock;
3988                 }
3989         }
3990
3991         /* Ok, it's mapped. Make sure it's up-to-date */
3992         if (PageUptodate(page))
3993                 set_buffer_uptodate(bh);
3994
3995         if (!buffer_uptodate(bh)) {
3996                 err = -EIO;
3997                 ll_rw_block(READ, 1, &bh);
3998                 wait_on_buffer(bh);
3999                 /* Uhhuh. Read error. Complain and punt. */
4000                 if (!buffer_uptodate(bh))
4001                         goto unlock;
4002         }
4003
4004         if (ext4_should_journal_data(inode)) {
4005                 BUFFER_TRACE(bh, "get write access");
4006                 err = ext4_journal_get_write_access(handle, bh);
4007                 if (err)
4008                         goto unlock;
4009         }
4010
4011         zero_user(page, offset, length);
4012
4013         BUFFER_TRACE(bh, "zeroed end of block");
4014
4015         err = 0;
4016         if (ext4_should_journal_data(inode)) {
4017                 err = ext4_handle_dirty_metadata(handle, inode, bh);
4018         } else {
4019                 if (ext4_should_order_data(inode))
4020                         err = ext4_jbd2_file_inode(handle, inode);
4021                 mark_buffer_dirty(bh);
4022         }
4023
4024 unlock:
4025         unlock_page(page);
4026         page_cache_release(page);
4027         return err;
4028 }
4029
4030 /*
4031  * Probably it should be a library function... search for first non-zero word
4032  * or memcmp with zero_page, whatever is better for particular architecture.
4033  * Linus?
4034  */
4035 static inline int all_zeroes(__le32 *p, __le32 *q)
4036 {
4037         while (p < q)
4038                 if (*p++)
4039                         return 0;
4040         return 1;
4041 }
4042
4043 /**
4044  *      ext4_find_shared - find the indirect blocks for partial truncation.
4045  *      @inode:   inode in question
4046  *      @depth:   depth of the affected branch
4047  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
4048  *      @chain:   place to store the pointers to partial indirect blocks
4049  *      @top:     place to the (detached) top of branch
4050  *
4051  *      This is a helper function used by ext4_truncate().
4052  *
4053  *      When we do truncate() we may have to clean the ends of several
4054  *      indirect blocks but leave the blocks themselves alive. Block is
4055  *      partially truncated if some data below the new i_size is refered
4056  *      from it (and it is on the path to the first completely truncated
4057  *      data block, indeed).  We have to free the top of that path along
4058  *      with everything to the right of the path. Since no allocation
4059  *      past the truncation point is possible until ext4_truncate()
4060  *      finishes, we may safely do the latter, but top of branch may
4061  *      require special attention - pageout below the truncation point
4062  *      might try to populate it.
4063  *
4064  *      We atomically detach the top of branch from the tree, store the
4065  *      block number of its root in *@top, pointers to buffer_heads of
4066  *      partially truncated blocks - in @chain[].bh and pointers to
4067  *      their last elements that should not be removed - in
4068  *      @chain[].p. Return value is the pointer to last filled element
4069  *      of @chain.
4070  *
4071  *      The work left to caller to do the actual freeing of subtrees:
4072  *              a) free the subtree starting from *@top
4073  *              b) free the subtrees whose roots are stored in
4074  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
4075  *              c) free the subtrees growing from the inode past the @chain[0].
4076  *                      (no partially truncated stuff there).  */
4077
4078 static Indirect *ext4_find_shared(struct inode *inode, int depth,
4079                                   ext4_lblk_t offsets[4], Indirect chain[4],
4080                                   __le32 *top)
4081 {
4082         Indirect *partial, *p;
4083         int k, err;
4084
4085         *top = 0;
4086         /* Make k index the deepest non-null offset + 1 */
4087         for (k = depth; k > 1 && !offsets[k-1]; k--)
4088                 ;
4089         partial = ext4_get_branch(inode, k, offsets, chain, &err);
4090         /* Writer: pointers */
4091         if (!partial)
4092                 partial = chain + k-1;
4093         /*
4094          * If the branch acquired continuation since we've looked at it -
4095          * fine, it should all survive and (new) top doesn't belong to us.
4096          */
4097         if (!partial->key && *partial->p)
4098                 /* Writer: end */
4099                 goto no_top;
4100         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
4101                 ;
4102         /*
4103          * OK, we've found the last block that must survive. The rest of our
4104          * branch should be detached before unlocking. However, if that rest
4105          * of branch is all ours and does not grow immediately from the inode
4106          * it's easier to cheat and just decrement partial->p.
4107          */
4108         if (p == chain + k - 1 && p > chain) {
4109                 p->p--;
4110         } else {
4111                 *top = *p->p;
4112                 /* Nope, don't do this in ext4.  Must leave the tree intact */
4113 #if 0
4114                 *p->p = 0;
4115 #endif
4116         }
4117         /* Writer: end */
4118
4119         while (partial > p) {
4120                 brelse(partial->bh);
4121                 partial--;
4122         }
4123 no_top:
4124         return partial;
4125 }
4126
4127 /*
4128  * Zero a number of block pointers in either an inode or an indirect block.
4129  * If we restart the transaction we must again get write access to the
4130  * indirect block for further modification.
4131  *
4132  * We release `count' blocks on disk, but (last - first) may be greater
4133  * than `count' because there can be holes in there.
4134  */
4135 static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
4136                              struct buffer_head *bh,
4137                              ext4_fsblk_t block_to_free,
4138                              unsigned long count, __le32 *first,
4139                              __le32 *last)
4140 {
4141         __le32 *p;
4142         int     flags = EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_VALIDATED;
4143
4144         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
4145                 flags |= EXT4_FREE_BLOCKS_METADATA;
4146
4147         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
4148                                    count)) {
4149                 ext4_error(inode->i_sb, "inode #%lu: "
4150                            "attempt to clear blocks %llu len %lu, invalid",
4151                            inode->i_ino, (unsigned long long) block_to_free,
4152                            count);
4153                 return 1;
4154         }
4155
4156         if (try_to_extend_transaction(handle, inode)) {
4157                 if (bh) {
4158                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4159                         ext4_handle_dirty_metadata(handle, inode, bh);
4160                 }
4161                 ext4_mark_inode_dirty(handle, inode);
4162                 ext4_truncate_restart_trans(handle, inode,
4163                                             blocks_for_truncate(inode));
4164                 if (bh) {
4165                         BUFFER_TRACE(bh, "retaking write access");
4166                         ext4_journal_get_write_access(handle, bh);
4167                 }
4168         }
4169
4170         for (p = first; p < last; p++)
4171                 *p = 0;
4172
4173         ext4_free_blocks(handle, inode, 0, block_to_free, count, flags);
4174         return 0;
4175 }
4176
4177 /**
4178  * ext4_free_data - free a list of data blocks
4179  * @handle:     handle for this transaction
4180  * @inode:      inode we are dealing with
4181  * @this_bh:    indirect buffer_head which contains *@first and *@last
4182  * @first:      array of block numbers
4183  * @last:       points immediately past the end of array
4184  *
4185  * We are freeing all blocks refered from that array (numbers are stored as
4186  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
4187  *
4188  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
4189  * blocks are contiguous then releasing them at one time will only affect one
4190  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
4191  * actually use a lot of journal space.
4192  *
4193  * @this_bh will be %NULL if @first and @last point into the inode's direct
4194  * block pointers.
4195  */
4196 static void ext4_free_data(handle_t *handle, struct inode *inode,
4197                            struct buffer_head *this_bh,
4198                            __le32 *first, __le32 *last)
4199 {
4200         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
4201         unsigned long count = 0;            /* Number of blocks in the run */
4202         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
4203                                                corresponding to
4204                                                block_to_free */
4205         ext4_fsblk_t nr;                    /* Current block # */
4206         __le32 *p;                          /* Pointer into inode/ind
4207                                                for current block */
4208         int err;
4209
4210         if (this_bh) {                          /* For indirect block */
4211                 BUFFER_TRACE(this_bh, "get_write_access");
4212                 err = ext4_journal_get_write_access(handle, this_bh);
4213                 /* Important: if we can't update the indirect pointers
4214                  * to the blocks, we can't free them. */
4215                 if (err)
4216                         return;
4217         }
4218
4219         for (p = first; p < last; p++) {
4220                 nr = le32_to_cpu(*p);
4221                 if (nr) {
4222                         /* accumulate blocks to free if they're contiguous */
4223                         if (count == 0) {
4224                                 block_to_free = nr;
4225                                 block_to_free_p = p;
4226                                 count = 1;
4227                         } else if (nr == block_to_free + count) {
4228                                 count++;
4229                         } else {
4230                                 if (ext4_clear_blocks(handle, inode, this_bh,
4231                                                       block_to_free, count,
4232                                                       block_to_free_p, p))
4233                                         break;
4234                                 block_to_free = nr;
4235                                 block_to_free_p = p;
4236                                 count = 1;
4237                         }
4238                 }
4239         }
4240
4241         if (count > 0)
4242                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
4243                                   count, block_to_free_p, p);
4244
4245         if (this_bh) {
4246                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
4247
4248                 /*
4249                  * The buffer head should have an attached journal head at this
4250                  * point. However, if the data is corrupted and an indirect
4251                  * block pointed to itself, it would have been detached when
4252                  * the block was cleared. Check for this instead of OOPSing.
4253                  */
4254                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
4255                         ext4_handle_dirty_metadata(handle, inode, this_bh);
4256                 else
4257                         ext4_error(inode->i_sb,
4258                                    "circular indirect block detected, "
4259                                    "inode=%lu, block=%llu",
4260                                    inode->i_ino,
4261                                    (unsigned long long) this_bh->b_blocknr);
4262         }
4263 }
4264
4265 /**
4266  *      ext4_free_branches - free an array of branches
4267  *      @handle: JBD handle for this transaction
4268  *      @inode: inode we are dealing with
4269  *      @parent_bh: the buffer_head which contains *@first and *@last
4270  *      @first: array of block numbers
4271  *      @last:  pointer immediately past the end of array
4272  *      @depth: depth of the branches to free
4273  *
4274  *      We are freeing all blocks refered from these branches (numbers are
4275  *      stored as little-endian 32-bit) and updating @inode->i_blocks
4276  *      appropriately.
4277  */
4278 static void ext4_free_branches(handle_t *handle, struct inode *inode,
4279                                struct buffer_head *parent_bh,
4280                                __le32 *first, __le32 *last, int depth)
4281 {
4282         ext4_fsblk_t nr;
4283         __le32 *p;
4284
4285         if (ext4_handle_is_aborted(handle))
4286                 return;
4287
4288         if (depth--) {
4289                 struct buffer_head *bh;
4290                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4291                 p = last;
4292                 while (--p >= first) {
4293                         nr = le32_to_cpu(*p);
4294                         if (!nr)
4295                                 continue;               /* A hole */
4296
4297                         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
4298                                                    nr, 1)) {
4299                                 ext4_error(inode->i_sb,
4300                                            "indirect mapped block in inode "
4301                                            "#%lu invalid (level %d, blk #%lu)",
4302                                            inode->i_ino, depth,
4303                                            (unsigned long) nr);
4304                                 break;
4305                         }
4306
4307                         /* Go read the buffer for the next level down */
4308                         bh = sb_bread(inode->i_sb, nr);
4309
4310                         /*
4311                          * A read failure? Report error and clear slot
4312                          * (should be rare).
4313                          */
4314                         if (!bh) {
4315                                 ext4_error(inode->i_sb,
4316                                            "Read failure, inode=%lu, block=%llu",
4317                                            inode->i_ino, nr);
4318                                 continue;
4319                         }
4320
4321                         /* This zaps the entire block.  Bottom up. */
4322                         BUFFER_TRACE(bh, "free child branches");
4323                         ext4_free_branches(handle, inode, bh,
4324                                         (__le32 *) bh->b_data,
4325                                         (__le32 *) bh->b_data + addr_per_block,
4326                                         depth);
4327
4328                         /*
4329                          * We've probably journalled the indirect block several
4330                          * times during the truncate.  But it's no longer
4331                          * needed and we now drop it from the transaction via
4332                          * jbd2_journal_revoke().
4333                          *
4334                          * That's easy if it's exclusively part of this
4335                          * transaction.  But if it's part of the committing
4336                          * transaction then jbd2_journal_forget() will simply
4337                          * brelse() it.  That means that if the underlying
4338                          * block is reallocated in ext4_get_block(),
4339                          * unmap_underlying_metadata() will find this block
4340                          * and will try to get rid of it.  damn, damn.
4341                          *
4342                          * If this block has already been committed to the
4343                          * journal, a revoke record will be written.  And
4344                          * revoke records must be emitted *before* clearing
4345                          * this block's bit in the bitmaps.
4346                          */
4347                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
4348
4349                         /*
4350                          * Everything below this this pointer has been
4351                          * released.  Now let this top-of-subtree go.
4352                          *
4353                          * We want the freeing of this indirect block to be
4354                          * atomic in the journal with the updating of the
4355                          * bitmap block which owns it.  So make some room in
4356                          * the journal.
4357                          *
4358                          * We zero the parent pointer *after* freeing its
4359                          * pointee in the bitmaps, so if extend_transaction()
4360                          * for some reason fails to put the bitmap changes and
4361                          * the release into the same transaction, recovery
4362                          * will merely complain about releasing a free block,
4363                          * rather than leaking blocks.
4364                          */
4365                         if (ext4_handle_is_aborted(handle))
4366                                 return;
4367                         if (try_to_extend_transaction(handle, inode)) {
4368                                 ext4_mark_inode_dirty(handle, inode);
4369                                 ext4_truncate_restart_trans(handle, inode,
4370                                             blocks_for_truncate(inode));
4371                         }
4372
4373                         ext4_free_blocks(handle, inode, 0, nr, 1,
4374                                          EXT4_FREE_BLOCKS_METADATA);
4375
4376                         if (parent_bh) {
4377                                 /*
4378                                  * The block which we have just freed is
4379                                  * pointed to by an indirect block: journal it
4380                                  */
4381                                 BUFFER_TRACE(parent_bh, "get_write_access");
4382                                 if (!ext4_journal_get_write_access(handle,
4383                                                                    parent_bh)){
4384                                         *p = 0;
4385                                         BUFFER_TRACE(parent_bh,
4386                                         "call ext4_handle_dirty_metadata");
4387                                         ext4_handle_dirty_metadata(handle,
4388                                                                    inode,
4389                                                                    parent_bh);
4390                                 }
4391                         }
4392                 }
4393         } else {
4394                 /* We have reached the bottom of the tree. */
4395                 BUFFER_TRACE(parent_bh, "free data blocks");
4396                 ext4_free_data(handle, inode, parent_bh, first, last);
4397         }
4398 }
4399
4400 int ext4_can_truncate(struct inode *inode)
4401 {
4402         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4403                 return 0;
4404         if (S_ISREG(inode->i_mode))
4405                 return 1;
4406         if (S_ISDIR(inode->i_mode))
4407                 return 1;
4408         if (S_ISLNK(inode->i_mode))
4409                 return !ext4_inode_is_fast_symlink(inode);
4410         return 0;
4411 }
4412
4413 /*
4414  * ext4_truncate()
4415  *
4416  * We block out ext4_get_block() block instantiations across the entire
4417  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4418  * simultaneously on behalf of the same inode.
4419  *
4420  * As we work through the truncate and commmit bits of it to the journal there
4421  * is one core, guiding principle: the file's tree must always be consistent on
4422  * disk.  We must be able to restart the truncate after a crash.
4423  *
4424  * The file's tree may be transiently inconsistent in memory (although it
4425  * probably isn't), but whenever we close off and commit a journal transaction,
4426  * the contents of (the filesystem + the journal) must be consistent and
4427  * restartable.  It's pretty simple, really: bottom up, right to left (although
4428  * left-to-right works OK too).
4429  *
4430  * Note that at recovery time, journal replay occurs *before* the restart of
4431  * truncate against the orphan inode list.
4432  *
4433  * The committed inode has the new, desired i_size (which is the same as
4434  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4435  * that this inode's truncate did not complete and it will again call
4436  * ext4_truncate() to have another go.  So there will be instantiated blocks
4437  * to the right of the truncation point in a crashed ext4 filesystem.  But
4438  * that's fine - as long as they are linked from the inode, the post-crash
4439  * ext4_truncate() run will find them and release them.
4440  */
4441 void ext4_truncate(struct inode *inode)
4442 {
4443         handle_t *handle;
4444         struct ext4_inode_info *ei = EXT4_I(inode);
4445         __le32 *i_data = ei->i_data;
4446         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4447         struct address_space *mapping = inode->i_mapping;
4448         ext4_lblk_t offsets[4];
4449         Indirect chain[4];
4450         Indirect *partial;
4451         __le32 nr = 0;
4452         int n;
4453         ext4_lblk_t last_block;
4454         unsigned blocksize = inode->i_sb->s_blocksize;
4455
4456         if (!ext4_can_truncate(inode))
4457                 return;
4458
4459         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4460                 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4461
4462         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
4463                 ext4_ext_truncate(inode);
4464                 return;
4465         }
4466
4467         handle = start_transaction(inode);
4468         if (IS_ERR(handle))
4469                 return;         /* AKPM: return what? */
4470
4471         last_block = (inode->i_size + blocksize-1)
4472                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4473
4474         if (inode->i_size & (blocksize - 1))
4475                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4476                         goto out_stop;
4477
4478         n = ext4_block_to_path(inode, last_block, offsets, NULL);
4479         if (n == 0)
4480                 goto out_stop;  /* error */
4481
4482         /*
4483          * OK.  This truncate is going to happen.  We add the inode to the
4484          * orphan list, so that if this truncate spans multiple transactions,
4485          * and we crash, we will resume the truncate when the filesystem
4486          * recovers.  It also marks the inode dirty, to catch the new size.
4487          *
4488          * Implication: the file must always be in a sane, consistent
4489          * truncatable state while each transaction commits.
4490          */
4491         if (ext4_orphan_add(handle, inode))
4492                 goto out_stop;
4493
4494         /*
4495          * From here we block out all ext4_get_block() callers who want to
4496          * modify the block allocation tree.
4497          */
4498         down_write(&ei->i_data_sem);
4499
4500         ext4_discard_preallocations(inode);
4501
4502         /*
4503          * The orphan list entry will now protect us from any crash which
4504          * occurs before the truncate completes, so it is now safe to propagate
4505          * the new, shorter inode size (held for now in i_size) into the
4506          * on-disk inode. We do this via i_disksize, which is the value which
4507          * ext4 *really* writes onto the disk inode.
4508          */
4509         ei->i_disksize = inode->i_size;
4510
4511         if (n == 1) {           /* direct blocks */
4512                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4513                                i_data + EXT4_NDIR_BLOCKS);
4514                 goto do_indirects;
4515         }
4516
4517         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4518         /* Kill the top of shared branch (not detached) */
4519         if (nr) {
4520                 if (partial == chain) {
4521                         /* Shared branch grows from the inode */
4522                         ext4_free_branches(handle, inode, NULL,
4523                                            &nr, &nr+1, (chain+n-1) - partial);
4524                         *partial->p = 0;
4525                         /*
4526                          * We mark the inode dirty prior to restart,
4527                          * and prior to stop.  No need for it here.
4528                          */
4529                 } else {
4530                         /* Shared branch grows from an indirect block */
4531                         BUFFER_TRACE(partial->bh, "get_write_access");
4532                         ext4_free_branches(handle, inode, partial->bh,
4533                                         partial->p,
4534                                         partial->p+1, (chain+n-1) - partial);
4535                 }
4536         }
4537         /* Clear the ends of indirect blocks on the shared branch */
4538         while (partial > chain) {
4539                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4540                                    (__le32*)partial->bh->b_data+addr_per_block,
4541                                    (chain+n-1) - partial);
4542                 BUFFER_TRACE(partial->bh, "call brelse");
4543                 brelse(partial->bh);
4544                 partial--;
4545         }
4546 do_indirects:
4547         /* Kill the remaining (whole) subtrees */
4548         switch (offsets[0]) {
4549         default:
4550                 nr = i_data[EXT4_IND_BLOCK];
4551                 if (nr) {
4552                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4553                         i_data[EXT4_IND_BLOCK] = 0;
4554                 }
4555         case EXT4_IND_BLOCK:
4556                 nr = i_data[EXT4_DIND_BLOCK];
4557                 if (nr) {
4558                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4559                         i_data[EXT4_DIND_BLOCK] = 0;
4560                 }
4561         case EXT4_DIND_BLOCK:
4562                 nr = i_data[EXT4_TIND_BLOCK];
4563                 if (nr) {
4564                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4565                         i_data[EXT4_TIND_BLOCK] = 0;
4566                 }
4567         case EXT4_TIND_BLOCK:
4568                 ;
4569         }
4570
4571         up_write(&ei->i_data_sem);
4572         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4573         ext4_mark_inode_dirty(handle, inode);
4574
4575         /*
4576          * In a multi-transaction truncate, we only make the final transaction
4577          * synchronous
4578          */
4579         if (IS_SYNC(inode))
4580                 ext4_handle_sync(handle);
4581 out_stop:
4582         /*
4583          * If this was a simple ftruncate(), and the file will remain alive
4584          * then we need to clear up the orphan record which we created above.
4585          * However, if this was a real unlink then we were called by
4586          * ext4_delete_inode(), and we allow that function to clean up the
4587          * orphan info for us.
4588          */
4589         if (inode->i_nlink)
4590                 ext4_orphan_del(handle, inode);
4591
4592         ext4_journal_stop(handle);
4593 }
4594
4595 /*
4596  * ext4_get_inode_loc returns with an extra refcount against the inode's
4597  * underlying buffer_head on success. If 'in_mem' is true, we have all
4598  * data in memory that is needed to recreate the on-disk version of this
4599  * inode.
4600  */
4601 static int __ext4_get_inode_loc(struct inode *inode,
4602                                 struct ext4_iloc *iloc, int in_mem)
4603 {
4604         struct ext4_group_desc  *gdp;
4605         struct buffer_head      *bh;
4606         struct super_block      *sb = inode->i_sb;
4607         ext4_fsblk_t            block;
4608         int                     inodes_per_block, inode_offset;
4609
4610         iloc->bh = NULL;
4611         if (!ext4_valid_inum(sb, inode->i_ino))
4612                 return -EIO;
4613
4614         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4615         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4616         if (!gdp)
4617                 return -EIO;
4618
4619         /*
4620          * Figure out the offset within the block group inode table
4621          */
4622         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4623         inode_offset = ((inode->i_ino - 1) %
4624                         EXT4_INODES_PER_GROUP(sb));
4625         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4626         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4627
4628         bh = sb_getblk(sb, block);
4629         if (!bh) {
4630                 ext4_error(sb, "unable to read inode block - "
4631                            "inode=%lu, block=%llu", inode->i_ino, block);
4632                 return -EIO;
4633         }
4634         if (!buffer_uptodate(bh)) {
4635                 lock_buffer(bh);
4636
4637                 /*
4638                  * If the buffer has the write error flag, we have failed
4639                  * to write out another inode in the same block.  In this
4640                  * case, we don't have to read the block because we may
4641                  * read the old inode data successfully.
4642                  */
4643                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4644                         set_buffer_uptodate(bh);
4645
4646                 if (buffer_uptodate(bh)) {
4647                         /* someone brought it uptodate while we waited */
4648                         unlock_buffer(bh);
4649                         goto has_buffer;
4650                 }
4651
4652                 /*
4653                  * If we have all information of the inode in memory and this
4654                  * is the only valid inode in the block, we need not read the
4655                  * block.
4656                  */
4657                 if (in_mem) {
4658                         struct buffer_head *bitmap_bh;
4659                         int i, start;
4660
4661                         start = inode_offset & ~(inodes_per_block - 1);
4662
4663                         /* Is the inode bitmap in cache? */
4664                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4665                         if (!bitmap_bh)
4666                                 goto make_io;
4667
4668                         /*
4669                          * If the inode bitmap isn't in cache then the
4670                          * optimisation may end up performing two reads instead
4671                          * of one, so skip it.
4672                          */
4673                         if (!buffer_uptodate(bitmap_bh)) {
4674                                 brelse(bitmap_bh);
4675                                 goto make_io;
4676                         }
4677                         for (i = start; i < start + inodes_per_block; i++) {
4678                                 if (i == inode_offset)
4679                                         continue;
4680                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4681                                         break;
4682                         }
4683                         brelse(bitmap_bh);
4684                         if (i == start + inodes_per_block) {
4685                                 /* all other inodes are free, so skip I/O */
4686                                 memset(bh->b_data, 0, bh->b_size);
4687                                 set_buffer_uptodate(bh);
4688                                 unlock_buffer(bh);
4689                                 goto has_buffer;
4690                         }
4691                 }
4692
4693 make_io:
4694                 /*
4695                  * If we need to do any I/O, try to pre-readahead extra
4696                  * blocks from the inode table.
4697                  */
4698                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4699                         ext4_fsblk_t b, end, table;
4700                         unsigned num;
4701
4702                         table = ext4_inode_table(sb, gdp);
4703                         /* s_inode_readahead_blks is always a power of 2 */
4704                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4705                         if (table > b)
4706                                 b = table;
4707                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4708                         num = EXT4_INODES_PER_GROUP(sb);
4709                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4710                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4711                                 num -= ext4_itable_unused_count(sb, gdp);
4712                         table += num / inodes_per_block;
4713                         if (end > table)
4714                                 end = table;
4715                         while (b <= end)
4716                                 sb_breadahead(sb, b++);
4717                 }
4718
4719                 /*
4720                  * There are other valid inodes in the buffer, this inode
4721                  * has in-inode xattrs, or we don't have this inode in memory.
4722                  * Read the block from disk.
4723                  */
4724                 get_bh(bh);
4725                 bh->b_end_io = end_buffer_read_sync;
4726                 submit_bh(READ_META, bh);
4727                 wait_on_buffer(bh);
4728                 if (!buffer_uptodate(bh)) {
4729                         ext4_error(sb, "unable to read inode block - inode=%lu,"
4730                                    " block=%llu", inode->i_ino, block);
4731                         brelse(bh);
4732                         return -EIO;
4733                 }
4734         }
4735 has_buffer:
4736         iloc->bh = bh;
4737         return 0;
4738 }
4739
4740 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4741 {
4742         /* We have all inode data except xattrs in memory here. */
4743         return __ext4_get_inode_loc(inode, iloc,
4744                 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
4745 }
4746
4747 void ext4_set_inode_flags(struct inode *inode)
4748 {
4749         unsigned int flags = EXT4_I(inode)->i_flags;
4750
4751         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4752         if (flags & EXT4_SYNC_FL)
4753                 inode->i_flags |= S_SYNC;
4754         if (flags & EXT4_APPEND_FL)
4755                 inode->i_flags |= S_APPEND;
4756         if (flags & EXT4_IMMUTABLE_FL)
4757                 inode->i_flags |= S_IMMUTABLE;
4758         if (flags & EXT4_NOATIME_FL)
4759                 inode->i_flags |= S_NOATIME;
4760         if (flags & EXT4_DIRSYNC_FL)
4761                 inode->i_flags |= S_DIRSYNC;
4762 }
4763
4764 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4765 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4766 {
4767         unsigned int flags = ei->vfs_inode.i_flags;
4768
4769         ei->i_flags &= ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4770                         EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|EXT4_DIRSYNC_FL);
4771         if (flags & S_SYNC)
4772                 ei->i_flags |= EXT4_SYNC_FL;
4773         if (flags & S_APPEND)
4774                 ei->i_flags |= EXT4_APPEND_FL;
4775         if (flags & S_IMMUTABLE)
4776                 ei->i_flags |= EXT4_IMMUTABLE_FL;
4777         if (flags & S_NOATIME)
4778                 ei->i_flags |= EXT4_NOATIME_FL;
4779         if (flags & S_DIRSYNC)
4780                 ei->i_flags |= EXT4_DIRSYNC_FL;
4781 }
4782
4783 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4784                                   struct ext4_inode_info *ei)
4785 {
4786         blkcnt_t i_blocks ;
4787         struct inode *inode = &(ei->vfs_inode);
4788         struct super_block *sb = inode->i_sb;
4789
4790         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4791                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4792                 /* we are using combined 48 bit field */
4793                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4794                                         le32_to_cpu(raw_inode->i_blocks_lo);
4795                 if (ei->i_flags & EXT4_HUGE_FILE_FL) {
4796                         /* i_blocks represent file system block size */
4797                         return i_blocks  << (inode->i_blkbits - 9);
4798                 } else {
4799                         return i_blocks;
4800                 }
4801         } else {
4802                 return le32_to_cpu(raw_inode->i_blocks_lo);
4803         }
4804 }
4805
4806 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4807 {
4808         struct ext4_iloc iloc;
4809         struct ext4_inode *raw_inode;
4810         struct ext4_inode_info *ei;
4811         struct inode *inode;
4812         journal_t *journal = EXT4_SB(sb)->s_journal;
4813         long ret;
4814         int block;
4815
4816         inode = iget_locked(sb, ino);
4817         if (!inode)
4818                 return ERR_PTR(-ENOMEM);
4819         if (!(inode->i_state & I_NEW))
4820                 return inode;
4821
4822         ei = EXT4_I(inode);
4823         iloc.bh = 0;
4824
4825         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4826         if (ret < 0)
4827                 goto bad_inode;
4828         raw_inode = ext4_raw_inode(&iloc);
4829         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4830         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4831         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4832         if (!(test_opt(inode->i_sb, NO_UID32))) {
4833                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4834                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4835         }
4836         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
4837
4838         ei->i_state_flags = 0;
4839         ei->i_dir_start_lookup = 0;
4840         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4841         /* We now have enough fields to check if the inode was active or not.
4842          * This is needed because nfsd might try to access dead inodes
4843          * the test is that same one that e2fsck uses
4844          * NeilBrown 1999oct15
4845          */
4846         if (inode->i_nlink == 0) {
4847                 if (inode->i_mode == 0 ||
4848                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
4849                         /* this inode is deleted */
4850                         ret = -ESTALE;
4851                         goto bad_inode;
4852                 }
4853                 /* The only unlinked inodes we let through here have
4854                  * valid i_mode and are being read by the orphan
4855                  * recovery code: that's fine, we're about to complete
4856                  * the process of deleting those. */
4857         }
4858         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4859         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4860         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4861         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
4862                 ei->i_file_acl |=
4863                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4864         inode->i_size = ext4_isize(raw_inode);
4865         ei->i_disksize = inode->i_size;
4866 #ifdef CONFIG_QUOTA
4867         ei->i_reserved_quota = 0;
4868 #endif
4869         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4870         ei->i_block_group = iloc.block_group;
4871         ei->i_last_alloc_group = ~0;
4872         /*
4873          * NOTE! The in-memory inode i_data array is in little-endian order
4874          * even on big-endian machines: we do NOT byteswap the block numbers!
4875          */
4876         for (block = 0; block < EXT4_N_BLOCKS; block++)
4877                 ei->i_data[block] = raw_inode->i_block[block];
4878         INIT_LIST_HEAD(&ei->i_orphan);
4879
4880         /*
4881          * Set transaction id's of transactions that have to be committed
4882          * to finish f[data]sync. We set them to currently running transaction
4883          * as we cannot be sure that the inode or some of its metadata isn't
4884          * part of the transaction - the inode could have been reclaimed and
4885          * now it is reread from disk.
4886          */
4887         if (journal) {
4888                 transaction_t *transaction;
4889                 tid_t tid;
4890
4891                 spin_lock(&journal->j_state_lock);
4892                 if (journal->j_running_transaction)
4893                         transaction = journal->j_running_transaction;
4894                 else
4895                         transaction = journal->j_committing_transaction;
4896                 if (transaction)
4897                         tid = transaction->t_tid;
4898                 else
4899                         tid = journal->j_commit_sequence;
4900                 spin_unlock(&journal->j_state_lock);
4901                 ei->i_sync_tid = tid;
4902                 ei->i_datasync_tid = tid;
4903         }
4904
4905         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4906                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4907                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4908                     EXT4_INODE_SIZE(inode->i_sb)) {
4909                         ret = -EIO;
4910                         goto bad_inode;
4911                 }
4912                 if (ei->i_extra_isize == 0) {
4913                         /* The extra space is currently unused. Use it. */
4914                         ei->i_extra_isize = sizeof(struct ext4_inode) -
4915                                             EXT4_GOOD_OLD_INODE_SIZE;
4916                 } else {
4917                         __le32 *magic = (void *)raw_inode +
4918                                         EXT4_GOOD_OLD_INODE_SIZE +
4919                                         ei->i_extra_isize;
4920                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
4921                                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
4922                 }
4923         } else
4924                 ei->i_extra_isize = 0;
4925
4926         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4927         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4928         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4929         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4930
4931         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4932         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4933                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4934                         inode->i_version |=
4935                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4936         }
4937
4938         ret = 0;
4939         if (ei->i_file_acl &&
4940             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
4941                 ext4_error(sb, "bad extended attribute block %llu inode #%lu",
4942                            ei->i_file_acl, inode->i_ino);
4943                 ret = -EIO;
4944                 goto bad_inode;
4945         } else if (ei->i_flags & EXT4_EXTENTS_FL) {
4946                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4947                     (S_ISLNK(inode->i_mode) &&
4948                      !ext4_inode_is_fast_symlink(inode)))
4949                         /* Validate extent which is part of inode */
4950                         ret = ext4_ext_check_inode(inode);
4951         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4952                    (S_ISLNK(inode->i_mode) &&
4953                     !ext4_inode_is_fast_symlink(inode))) {
4954                 /* Validate block references which are part of inode */
4955                 ret = ext4_check_inode_blockref(inode);
4956         }
4957         if (ret)
4958                 goto bad_inode;
4959
4960         if (S_ISREG(inode->i_mode)) {
4961                 inode->i_op = &ext4_file_inode_operations;
4962                 inode->i_fop = &ext4_file_operations;
4963                 ext4_set_aops(inode);
4964         } else if (S_ISDIR(inode->i_mode)) {
4965                 inode->i_op = &ext4_dir_inode_operations;
4966                 inode->i_fop = &ext4_dir_operations;
4967         } else if (S_ISLNK(inode->i_mode)) {
4968                 if (ext4_inode_is_fast_symlink(inode)) {
4969                         inode->i_op = &ext4_fast_symlink_inode_operations;
4970                         nd_terminate_link(ei->i_data, inode->i_size,
4971                                 sizeof(ei->i_data) - 1);
4972                 } else {
4973                         inode->i_op = &ext4_symlink_inode_operations;
4974                         ext4_set_aops(inode);
4975                 }
4976         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
4977               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
4978                 inode->i_op = &ext4_special_inode_operations;
4979                 if (raw_inode->i_block[0])
4980                         init_special_inode(inode, inode->i_mode,
4981                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4982                 else
4983                         init_special_inode(inode, inode->i_mode,
4984                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4985         } else {
4986                 ret = -EIO;
4987                 ext4_error(inode->i_sb, "bogus i_mode (%o) for inode=%lu",
4988                            inode->i_mode, inode->i_ino);
4989                 goto bad_inode;
4990         }
4991         brelse(iloc.bh);
4992         ext4_set_inode_flags(inode);
4993         unlock_new_inode(inode);
4994         return inode;
4995
4996 bad_inode:
4997         brelse(iloc.bh);
4998         iget_failed(inode);
4999         return ERR_PTR(ret);
5000 }
5001
5002 static int ext4_inode_blocks_set(handle_t *handle,
5003                                 struct ext4_inode *raw_inode,
5004                                 struct ext4_inode_info *ei)
5005 {
5006         struct inode *inode = &(ei->vfs_inode);
5007         u64 i_blocks = inode->i_blocks;
5008         struct super_block *sb = inode->i_sb;
5009
5010         if (i_blocks <= ~0U) {
5011                 /*
5012                  * i_blocks can be represnted in a 32 bit variable
5013                  * as multiple of 512 bytes
5014                  */
5015                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5016                 raw_inode->i_blocks_high = 0;
5017                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
5018                 return 0;
5019         }
5020         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
5021                 return -EFBIG;
5022
5023         if (i_blocks <= 0xffffffffffffULL) {
5024                 /*
5025                  * i_blocks can be represented in a 48 bit variable
5026                  * as multiple of 512 bytes
5027                  */
5028                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5029                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5030                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
5031         } else {
5032                 ei->i_flags |= EXT4_HUGE_FILE_FL;
5033                 /* i_block is stored in file system block size */
5034                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
5035                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5036                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5037         }
5038         return 0;
5039 }
5040
5041 /*
5042  * Post the struct inode info into an on-disk inode location in the
5043  * buffer-cache.  This gobbles the caller's reference to the
5044  * buffer_head in the inode location struct.
5045  *
5046  * The caller must have write access to iloc->bh.
5047  */
5048 static int ext4_do_update_inode(handle_t *handle,
5049                                 struct inode *inode,
5050                                 struct ext4_iloc *iloc)
5051 {
5052         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5053         struct ext4_inode_info *ei = EXT4_I(inode);
5054         struct buffer_head *bh = iloc->bh;
5055         int err = 0, rc, block;
5056
5057         /* For fields not not tracking in the in-memory inode,
5058          * initialise them to zero for new inodes. */
5059         if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5060                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5061
5062         ext4_get_inode_flags(ei);
5063         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5064         if (!(test_opt(inode->i_sb, NO_UID32))) {
5065                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
5066                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
5067 /*
5068  * Fix up interoperability with old kernels. Otherwise, old inodes get
5069  * re-used with the upper 16 bits of the uid/gid intact
5070  */
5071                 if (!ei->i_dtime) {
5072                         raw_inode->i_uid_high =
5073                                 cpu_to_le16(high_16_bits(inode->i_uid));
5074                         raw_inode->i_gid_high =
5075                                 cpu_to_le16(high_16_bits(inode->i_gid));
5076                 } else {
5077                         raw_inode->i_uid_high = 0;
5078                         raw_inode->i_gid_high = 0;
5079                 }
5080         } else {
5081                 raw_inode->i_uid_low =
5082                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
5083                 raw_inode->i_gid_low =
5084                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
5085                 raw_inode->i_uid_high = 0;
5086                 raw_inode->i_gid_high = 0;
5087         }
5088         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5089
5090         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5091         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5092         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5093         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5094
5095         if (ext4_inode_blocks_set(handle, raw_inode, ei))
5096                 goto out_brelse;
5097         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5098         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
5099         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
5100             cpu_to_le32(EXT4_OS_HURD))
5101                 raw_inode->i_file_acl_high =
5102                         cpu_to_le16(ei->i_file_acl >> 32);
5103         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5104         ext4_isize_set(raw_inode, ei->i_disksize);
5105         if (ei->i_disksize > 0x7fffffffULL) {
5106                 struct super_block *sb = inode->i_sb;
5107                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
5108                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
5109                                 EXT4_SB(sb)->s_es->s_rev_level ==
5110                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
5111                         /* If this is the first large file
5112                          * created, add a flag to the superblock.
5113                          */
5114                         err = ext4_journal_get_write_access(handle,
5115                                         EXT4_SB(sb)->s_sbh);
5116                         if (err)
5117                                 goto out_brelse;
5118                         ext4_update_dynamic_rev(sb);
5119                         EXT4_SET_RO_COMPAT_FEATURE(sb,
5120                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
5121                         sb->s_dirt = 1;
5122                         ext4_handle_sync(handle);
5123                         err = ext4_handle_dirty_metadata(handle, inode,
5124                                         EXT4_SB(sb)->s_sbh);
5125                 }
5126         }
5127         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5128         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5129                 if (old_valid_dev(inode->i_rdev)) {
5130                         raw_inode->i_block[0] =
5131                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
5132                         raw_inode->i_block[1] = 0;
5133                 } else {
5134                         raw_inode->i_block[0] = 0;
5135                         raw_inode->i_block[1] =
5136                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
5137                         raw_inode->i_block[2] = 0;
5138                 }
5139         } else
5140                 for (block = 0; block < EXT4_N_BLOCKS; block++)
5141                         raw_inode->i_block[block] = ei->i_data[block];
5142
5143         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
5144         if (ei->i_extra_isize) {
5145                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5146                         raw_inode->i_version_hi =
5147                         cpu_to_le32(inode->i_version >> 32);
5148                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
5149         }
5150
5151         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5152         rc = ext4_handle_dirty_metadata(handle, inode, bh);
5153         if (!err)
5154                 err = rc;
5155         ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5156
5157         ext4_update_inode_fsync_trans(handle, inode, 0);
5158 out_brelse:
5159         brelse(bh);
5160         ext4_std_error(inode->i_sb, err);
5161         return err;
5162 }
5163
5164 /*
5165  * ext4_write_inode()
5166  *
5167  * We are called from a few places:
5168  *
5169  * - Within generic_file_write() for O_SYNC files.
5170  *   Here, there will be no transaction running. We wait for any running
5171  *   trasnaction to commit.
5172  *
5173  * - Within sys_sync(), kupdate and such.
5174  *   We wait on commit, if tol to.
5175  *
5176  * - Within prune_icache() (PF_MEMALLOC == true)
5177  *   Here we simply return.  We can't afford to block kswapd on the
5178  *   journal commit.
5179  *
5180  * In all cases it is actually safe for us to return without doing anything,
5181  * because the inode has been copied into a raw inode buffer in
5182  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
5183  * knfsd.
5184  *
5185  * Note that we are absolutely dependent upon all inode dirtiers doing the
5186  * right thing: they *must* call mark_inode_dirty() after dirtying info in
5187  * which we are interested.
5188  *
5189  * It would be a bug for them to not do this.  The code:
5190  *
5191  *      mark_inode_dirty(inode)
5192  *      stuff();
5193  *      inode->i_size = expr;
5194  *
5195  * is in error because a kswapd-driven write_inode() could occur while
5196  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
5197  * will no longer be on the superblock's dirty inode list.
5198  */
5199 int ext4_write_inode(struct inode *inode, int wait)
5200 {
5201         int err;
5202
5203         if (current->flags & PF_MEMALLOC)
5204                 return 0;
5205
5206         if (EXT4_SB(inode->i_sb)->s_journal) {
5207                 if (ext4_journal_current_handle()) {
5208                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5209                         dump_stack();
5210                         return -EIO;
5211                 }
5212
5213                 if (!wait)
5214                         return 0;
5215
5216                 err = ext4_force_commit(inode->i_sb);
5217         } else {
5218                 struct ext4_iloc iloc;
5219
5220                 err = ext4_get_inode_loc(inode, &iloc);
5221                 if (err)
5222                         return err;
5223                 if (wait)
5224                         sync_dirty_buffer(iloc.bh);
5225                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5226                         ext4_error(inode->i_sb, "IO error syncing inode, "
5227                                    "inode=%lu, block=%llu", inode->i_ino,
5228                                    (unsigned long long)iloc.bh->b_blocknr);
5229                         err = -EIO;
5230                 }
5231         }
5232         return err;
5233 }
5234
5235 /*
5236  * ext4_setattr()
5237  *
5238  * Called from notify_change.
5239  *
5240  * We want to trap VFS attempts to truncate the file as soon as
5241  * possible.  In particular, we want to make sure that when the VFS
5242  * shrinks i_size, we put the inode on the orphan list and modify
5243  * i_disksize immediately, so that during the subsequent flushing of
5244  * dirty pages and freeing of disk blocks, we can guarantee that any
5245  * commit will leave the blocks being flushed in an unused state on
5246  * disk.  (On recovery, the inode will get truncated and the blocks will
5247  * be freed, so we have a strong guarantee that no future commit will
5248  * leave these blocks visible to the user.)
5249  *
5250  * Another thing we have to assure is that if we are in ordered mode
5251  * and inode is still attached to the committing transaction, we must
5252  * we start writeout of all the dirty pages which are being truncated.
5253  * This way we are sure that all the data written in the previous
5254  * transaction are already on disk (truncate waits for pages under
5255  * writeback).
5256  *
5257  * Called with inode->i_mutex down.
5258  */
5259 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5260 {
5261         struct inode *inode = dentry->d_inode;
5262         int error, rc = 0;
5263         const unsigned int ia_valid = attr->ia_valid;
5264
5265         error = inode_change_ok(inode, attr);
5266         if (error)
5267                 return error;
5268
5269         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
5270                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
5271                 handle_t *handle;
5272
5273                 /* (user+group)*(old+new) structure, inode write (sb,
5274                  * inode block, ? - but truncate inode update has it) */
5275                 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
5276                                         EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
5277                 if (IS_ERR(handle)) {
5278                         error = PTR_ERR(handle);
5279                         goto err_out;
5280                 }
5281                 error = vfs_dq_transfer(inode, attr) ? -EDQUOT : 0;
5282                 if (error) {
5283                         ext4_journal_stop(handle);
5284                         return error;
5285                 }
5286                 /* Update corresponding info in inode so that everything is in
5287                  * one transaction */
5288                 if (attr->ia_valid & ATTR_UID)
5289                         inode->i_uid = attr->ia_uid;
5290                 if (attr->ia_valid & ATTR_GID)
5291                         inode->i_gid = attr->ia_gid;
5292                 error = ext4_mark_inode_dirty(handle, inode);
5293                 ext4_journal_stop(handle);
5294         }
5295
5296         if (attr->ia_valid & ATTR_SIZE) {
5297                 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)) {
5298                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5299
5300                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
5301                                 error = -EFBIG;
5302                                 goto err_out;
5303                         }
5304                 }
5305         }
5306
5307         if (S_ISREG(inode->i_mode) &&
5308             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
5309                 handle_t *handle;
5310
5311                 handle = ext4_journal_start(inode, 3);
5312                 if (IS_ERR(handle)) {
5313                         error = PTR_ERR(handle);
5314                         goto err_out;
5315                 }
5316
5317                 error = ext4_orphan_add(handle, inode);
5318                 EXT4_I(inode)->i_disksize = attr->ia_size;
5319                 rc = ext4_mark_inode_dirty(handle, inode);
5320                 if (!error)
5321                         error = rc;
5322                 ext4_journal_stop(handle);
5323
5324                 if (ext4_should_order_data(inode)) {
5325                         error = ext4_begin_ordered_truncate(inode,
5326                                                             attr->ia_size);
5327                         if (error) {
5328                                 /* Do as much error cleanup as possible */
5329                                 handle = ext4_journal_start(inode, 3);
5330                                 if (IS_ERR(handle)) {
5331                                         ext4_orphan_del(NULL, inode);
5332                                         goto err_out;
5333                                 }
5334                                 ext4_orphan_del(handle, inode);
5335                                 ext4_journal_stop(handle);
5336                                 goto err_out;
5337                         }
5338                 }
5339         }
5340
5341         rc = inode_setattr(inode, attr);
5342
5343         /* If inode_setattr's call to ext4_truncate failed to get a
5344          * transaction handle at all, we need to clean up the in-core
5345          * orphan list manually. */
5346         if (inode->i_nlink)
5347                 ext4_orphan_del(NULL, inode);
5348
5349         if (!rc && (ia_valid & ATTR_MODE))
5350                 rc = ext4_acl_chmod(inode);
5351
5352 err_out:
5353         ext4_std_error(inode->i_sb, error);
5354         if (!error)
5355                 error = rc;
5356         return error;
5357 }
5358
5359 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
5360                  struct kstat *stat)
5361 {
5362         struct inode *inode;
5363         unsigned long delalloc_blocks;
5364
5365         inode = dentry->d_inode;
5366         generic_fillattr(inode, stat);
5367
5368         /*
5369          * We can't update i_blocks if the block allocation is delayed
5370          * otherwise in the case of system crash before the real block
5371          * allocation is done, we will have i_blocks inconsistent with
5372          * on-disk file blocks.
5373          * We always keep i_blocks updated together with real
5374          * allocation. But to not confuse with user, stat
5375          * will return the blocks that include the delayed allocation
5376          * blocks for this file.
5377          */
5378         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
5379         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
5380         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
5381
5382         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
5383         return 0;
5384 }
5385
5386 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
5387                                       int chunk)
5388 {
5389         int indirects;
5390
5391         /* if nrblocks are contiguous */
5392         if (chunk) {
5393                 /*
5394                  * With N contiguous data blocks, it need at most
5395                  * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) indirect blocks
5396                  * 2 dindirect blocks
5397                  * 1 tindirect block
5398                  */
5399                 indirects = nrblocks / EXT4_ADDR_PER_BLOCK(inode->i_sb);
5400                 return indirects + 3;
5401         }
5402         /*
5403          * if nrblocks are not contiguous, worse case, each block touch
5404          * a indirect block, and each indirect block touch a double indirect
5405          * block, plus a triple indirect block
5406          */
5407         indirects = nrblocks * 2 + 1;
5408         return indirects;
5409 }
5410
5411 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5412 {
5413         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL))
5414                 return ext4_indirect_trans_blocks(inode, nrblocks, chunk);
5415         return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
5416 }
5417
5418 /*
5419  * Account for index blocks, block groups bitmaps and block group
5420  * descriptor blocks if modify datablocks and index blocks
5421  * worse case, the indexs blocks spread over different block groups
5422  *
5423  * If datablocks are discontiguous, they are possible to spread over
5424  * different block groups too. If they are contiuguous, with flexbg,
5425  * they could still across block group boundary.
5426  *
5427  * Also account for superblock, inode, quota and xattr blocks
5428  */
5429 int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5430 {
5431         ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
5432         int gdpblocks;
5433         int idxblocks;
5434         int ret = 0;
5435
5436         /*
5437          * How many index blocks need to touch to modify nrblocks?
5438          * The "Chunk" flag indicating whether the nrblocks is
5439          * physically contiguous on disk
5440          *
5441          * For Direct IO and fallocate, they calls get_block to allocate
5442          * one single extent at a time, so they could set the "Chunk" flag
5443          */
5444         idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
5445
5446         ret = idxblocks;
5447
5448         /*
5449          * Now let's see how many group bitmaps and group descriptors need
5450          * to account
5451          */
5452         groups = idxblocks;
5453         if (chunk)
5454                 groups += 1;
5455         else
5456                 groups += nrblocks;
5457
5458         gdpblocks = groups;
5459         if (groups > ngroups)
5460                 groups = ngroups;
5461         if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
5462                 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
5463
5464         /* bitmaps and block group descriptor blocks */
5465         ret += groups + gdpblocks;
5466
5467         /* Blocks for super block, inode, quota and xattr blocks */
5468         ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
5469
5470         return ret;
5471 }
5472
5473 /*
5474  * Calulate the total number of credits to reserve to fit
5475  * the modification of a single pages into a single transaction,
5476  * which may include multiple chunks of block allocations.
5477  *
5478  * This could be called via ext4_write_begin()
5479  *
5480  * We need to consider the worse case, when
5481  * one new block per extent.
5482  */
5483 int ext4_writepage_trans_blocks(struct inode *inode)
5484 {
5485         int bpp = ext4_journal_blocks_per_page(inode);
5486         int ret;
5487
5488         ret = ext4_meta_trans_blocks(inode, bpp, 0);
5489
5490         /* Account for data blocks for journalled mode */
5491         if (ext4_should_journal_data(inode))
5492                 ret += bpp;
5493         return ret;
5494 }
5495
5496 /*
5497  * Calculate the journal credits for a chunk of data modification.
5498  *
5499  * This is called from DIO, fallocate or whoever calling
5500  * ext4_get_blocks() to map/allocate a chunk of contiguous disk blocks.
5501  *
5502  * journal buffers for data blocks are not included here, as DIO
5503  * and fallocate do no need to journal data buffers.
5504  */
5505 int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
5506 {
5507         return ext4_meta_trans_blocks(inode, nrblocks, 1);
5508 }
5509
5510 /*
5511  * The caller must have previously called ext4_reserve_inode_write().
5512  * Give this, we know that the caller already has write access to iloc->bh.
5513  */
5514 int ext4_mark_iloc_dirty(handle_t *handle,
5515                          struct inode *inode, struct ext4_iloc *iloc)
5516 {
5517         int err = 0;
5518
5519         if (test_opt(inode->i_sb, I_VERSION))
5520                 inode_inc_iversion(inode);
5521
5522         /* the do_update_inode consumes one bh->b_count */
5523         get_bh(iloc->bh);
5524
5525         /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
5526         err = ext4_do_update_inode(handle, inode, iloc);
5527         put_bh(iloc->bh);
5528         return err;
5529 }
5530
5531 /*
5532  * On success, We end up with an outstanding reference count against
5533  * iloc->bh.  This _must_ be cleaned up later.
5534  */
5535
5536 int
5537 ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
5538                          struct ext4_iloc *iloc)
5539 {
5540         int err;
5541
5542         err = ext4_get_inode_loc(inode, iloc);
5543         if (!err) {
5544                 BUFFER_TRACE(iloc->bh, "get_write_access");
5545                 err = ext4_journal_get_write_access(handle, iloc->bh);
5546                 if (err) {
5547                         brelse(iloc->bh);
5548                         iloc->bh = NULL;
5549                 }
5550         }
5551         ext4_std_error(inode->i_sb, err);
5552         return err;
5553 }
5554
5555 /*
5556  * Expand an inode by new_extra_isize bytes.
5557  * Returns 0 on success or negative error number on failure.
5558  */
5559 static int ext4_expand_extra_isize(struct inode *inode,
5560                                    unsigned int new_extra_isize,
5561                                    struct ext4_iloc iloc,
5562                                    handle_t *handle)
5563 {
5564         struct ext4_inode *raw_inode;
5565         struct ext4_xattr_ibody_header *header;
5566         struct ext4_xattr_entry *entry;
5567
5568         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
5569                 return 0;
5570
5571         raw_inode = ext4_raw_inode(&iloc);
5572
5573         header = IHDR(inode, raw_inode);
5574         entry = IFIRST(header);
5575
5576         /* No extended attributes present */
5577         if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR) ||
5578             header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
5579                 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
5580                         new_extra_isize);
5581                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
5582                 return 0;
5583         }
5584
5585         /* try to expand with EAs present */
5586         return ext4_expand_extra_isize_ea(inode, new_extra_isize,
5587                                           raw_inode, handle);
5588 }
5589
5590 /*
5591  * What we do here is to mark the in-core inode as clean with respect to inode
5592  * dirtiness (it m