ext4: Drop EXT4_GET_BLOCKS_UPDATE_RESERVE_SPACE flag
[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, __func__,
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, __func__,
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, "ext4_block_to_path",
327                              "block %lu > max in inode %lu",
328                              i_block + direct_blocks +
329                              indirect_blocks + double_blocks, inode->i_ino);
330         }
331         if (boundary)
332                 *boundary = final - 1 - (i_block & (ptrs - 1));
333         return n;
334 }
335
336 static int __ext4_check_blockref(const char *function, struct inode *inode,
337                                  __le32 *p, unsigned int max)
338 {
339         __le32 *bref = p;
340         unsigned int blk;
341
342         while (bref < p+max) {
343                 blk = le32_to_cpu(*bref++);
344                 if (blk &&
345                     unlikely(!ext4_data_block_valid(EXT4_SB(inode->i_sb),
346                                                     blk, 1))) {
347                         ext4_error(inode->i_sb, function,
348                                    "invalid block reference %u "
349                                    "in inode #%lu", blk, inode->i_ino);
350                         return -EIO;
351                 }
352         }
353         return 0;
354 }
355
356
357 #define ext4_check_indirect_blockref(inode, bh)                         \
358         __ext4_check_blockref(__func__, inode, (__le32 *)(bh)->b_data,  \
359                               EXT4_ADDR_PER_BLOCK((inode)->i_sb))
360
361 #define ext4_check_inode_blockref(inode)                                \
362         __ext4_check_blockref(__func__, inode, EXT4_I(inode)->i_data,   \
363                               EXT4_NDIR_BLOCKS)
364
365 /**
366  *      ext4_get_branch - read the chain of indirect blocks leading to data
367  *      @inode: inode in question
368  *      @depth: depth of the chain (1 - direct pointer, etc.)
369  *      @offsets: offsets of pointers in inode/indirect blocks
370  *      @chain: place to store the result
371  *      @err: here we store the error value
372  *
373  *      Function fills the array of triples <key, p, bh> and returns %NULL
374  *      if everything went OK or the pointer to the last filled triple
375  *      (incomplete one) otherwise. Upon the return chain[i].key contains
376  *      the number of (i+1)-th block in the chain (as it is stored in memory,
377  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
378  *      number (it points into struct inode for i==0 and into the bh->b_data
379  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
380  *      block for i>0 and NULL for i==0. In other words, it holds the block
381  *      numbers of the chain, addresses they were taken from (and where we can
382  *      verify that chain did not change) and buffer_heads hosting these
383  *      numbers.
384  *
385  *      Function stops when it stumbles upon zero pointer (absent block)
386  *              (pointer to last triple returned, *@err == 0)
387  *      or when it gets an IO error reading an indirect block
388  *              (ditto, *@err == -EIO)
389  *      or when it reads all @depth-1 indirect blocks successfully and finds
390  *      the whole chain, all way to the data (returns %NULL, *err == 0).
391  *
392  *      Need to be called with
393  *      down_read(&EXT4_I(inode)->i_data_sem)
394  */
395 static Indirect *ext4_get_branch(struct inode *inode, int depth,
396                                  ext4_lblk_t  *offsets,
397                                  Indirect chain[4], int *err)
398 {
399         struct super_block *sb = inode->i_sb;
400         Indirect *p = chain;
401         struct buffer_head *bh;
402
403         *err = 0;
404         /* i_data is not going away, no lock needed */
405         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
406         if (!p->key)
407                 goto no_block;
408         while (--depth) {
409                 bh = sb_getblk(sb, le32_to_cpu(p->key));
410                 if (unlikely(!bh))
411                         goto failure;
412
413                 if (!bh_uptodate_or_lock(bh)) {
414                         if (bh_submit_read(bh) < 0) {
415                                 put_bh(bh);
416                                 goto failure;
417                         }
418                         /* validate block references */
419                         if (ext4_check_indirect_blockref(inode, bh)) {
420                                 put_bh(bh);
421                                 goto failure;
422                         }
423                 }
424
425                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
426                 /* Reader: end */
427                 if (!p->key)
428                         goto no_block;
429         }
430         return NULL;
431
432 failure:
433         *err = -EIO;
434 no_block:
435         return p;
436 }
437
438 /**
439  *      ext4_find_near - find a place for allocation with sufficient locality
440  *      @inode: owner
441  *      @ind: descriptor of indirect block.
442  *
443  *      This function returns the preferred place for block allocation.
444  *      It is used when heuristic for sequential allocation fails.
445  *      Rules are:
446  *        + if there is a block to the left of our position - allocate near it.
447  *        + if pointer will live in indirect block - allocate near that block.
448  *        + if pointer will live in inode - allocate in the same
449  *          cylinder group.
450  *
451  * In the latter case we colour the starting block by the callers PID to
452  * prevent it from clashing with concurrent allocations for a different inode
453  * in the same block group.   The PID is used here so that functionally related
454  * files will be close-by on-disk.
455  *
456  *      Caller must make sure that @ind is valid and will stay that way.
457  */
458 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
459 {
460         struct ext4_inode_info *ei = EXT4_I(inode);
461         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
462         __le32 *p;
463         ext4_fsblk_t bg_start;
464         ext4_fsblk_t last_block;
465         ext4_grpblk_t colour;
466         ext4_group_t block_group;
467         int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
468
469         /* Try to find previous block */
470         for (p = ind->p - 1; p >= start; p--) {
471                 if (*p)
472                         return le32_to_cpu(*p);
473         }
474
475         /* No such thing, so let's try location of indirect block */
476         if (ind->bh)
477                 return ind->bh->b_blocknr;
478
479         /*
480          * It is going to be referred to from the inode itself? OK, just put it
481          * into the same cylinder group then.
482          */
483         block_group = ei->i_block_group;
484         if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
485                 block_group &= ~(flex_size-1);
486                 if (S_ISREG(inode->i_mode))
487                         block_group++;
488         }
489         bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
490         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
491
492         /*
493          * If we are doing delayed allocation, we don't need take
494          * colour into account.
495          */
496         if (test_opt(inode->i_sb, DELALLOC))
497                 return bg_start;
498
499         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
500                 colour = (current->pid % 16) *
501                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
502         else
503                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
504         return bg_start + colour;
505 }
506
507 /**
508  *      ext4_find_goal - find a preferred place for allocation.
509  *      @inode: owner
510  *      @block:  block we want
511  *      @partial: pointer to the last triple within a chain
512  *
513  *      Normally this function find the preferred place for block allocation,
514  *      returns it.
515  *      Because this is only used for non-extent files, we limit the block nr
516  *      to 32 bits.
517  */
518 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
519                                    Indirect *partial)
520 {
521         ext4_fsblk_t goal;
522
523         /*
524          * XXX need to get goal block from mballoc's data structures
525          */
526
527         goal = ext4_find_near(inode, partial);
528         goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
529         return goal;
530 }
531
532 /**
533  *      ext4_blks_to_allocate: Look up the block map and count the number
534  *      of direct blocks need to be allocated for the given branch.
535  *
536  *      @branch: chain of indirect blocks
537  *      @k: number of blocks need for indirect blocks
538  *      @blks: number of data blocks to be mapped.
539  *      @blocks_to_boundary:  the offset in the indirect block
540  *
541  *      return the total number of blocks to be allocate, including the
542  *      direct and indirect blocks.
543  */
544 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
545                                  int blocks_to_boundary)
546 {
547         unsigned int count = 0;
548
549         /*
550          * Simple case, [t,d]Indirect block(s) has not allocated yet
551          * then it's clear blocks on that path have not allocated
552          */
553         if (k > 0) {
554                 /* right now we don't handle cross boundary allocation */
555                 if (blks < blocks_to_boundary + 1)
556                         count += blks;
557                 else
558                         count += blocks_to_boundary + 1;
559                 return count;
560         }
561
562         count++;
563         while (count < blks && count <= blocks_to_boundary &&
564                 le32_to_cpu(*(branch[0].p + count)) == 0) {
565                 count++;
566         }
567         return count;
568 }
569
570 /**
571  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
572  *      @indirect_blks: the number of blocks need to allocate for indirect
573  *                      blocks
574  *
575  *      @new_blocks: on return it will store the new block numbers for
576  *      the indirect blocks(if needed) and the first direct block,
577  *      @blks:  on return it will store the total number of allocated
578  *              direct blocks
579  */
580 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
581                              ext4_lblk_t iblock, ext4_fsblk_t goal,
582                              int indirect_blks, int blks,
583                              ext4_fsblk_t new_blocks[4], int *err)
584 {
585         struct ext4_allocation_request ar;
586         int target, i;
587         unsigned long count = 0, blk_allocated = 0;
588         int index = 0;
589         ext4_fsblk_t current_block = 0;
590         int ret = 0;
591
592         /*
593          * Here we try to allocate the requested multiple blocks at once,
594          * on a best-effort basis.
595          * To build a branch, we should allocate blocks for
596          * the indirect blocks(if not allocated yet), and at least
597          * the first direct block of this branch.  That's the
598          * minimum number of blocks need to allocate(required)
599          */
600         /* first we try to allocate the indirect blocks */
601         target = indirect_blks;
602         while (target > 0) {
603                 count = target;
604                 /* allocating blocks for indirect blocks and direct blocks */
605                 current_block = ext4_new_meta_blocks(handle, inode,
606                                                         goal, &count, err);
607                 if (*err)
608                         goto failed_out;
609
610                 BUG_ON(current_block + count > EXT4_MAX_BLOCK_FILE_PHYS);
611
612                 target -= count;
613                 /* allocate blocks for indirect blocks */
614                 while (index < indirect_blks && count) {
615                         new_blocks[index++] = current_block++;
616                         count--;
617                 }
618                 if (count > 0) {
619                         /*
620                          * save the new block number
621                          * for the first direct block
622                          */
623                         new_blocks[index] = current_block;
624                         printk(KERN_INFO "%s returned more blocks than "
625                                                 "requested\n", __func__);
626                         WARN_ON(1);
627                         break;
628                 }
629         }
630
631         target = blks - count ;
632         blk_allocated = count;
633         if (!target)
634                 goto allocated;
635         /* Now allocate data blocks */
636         memset(&ar, 0, sizeof(ar));
637         ar.inode = inode;
638         ar.goal = goal;
639         ar.len = target;
640         ar.logical = iblock;
641         if (S_ISREG(inode->i_mode))
642                 /* enable in-core preallocation only for regular files */
643                 ar.flags = EXT4_MB_HINT_DATA;
644
645         current_block = ext4_mb_new_blocks(handle, &ar, err);
646         BUG_ON(current_block + ar.len > EXT4_MAX_BLOCK_FILE_PHYS);
647
648         if (*err && (target == blks)) {
649                 /*
650                  * if the allocation failed and we didn't allocate
651                  * any blocks before
652                  */
653                 goto failed_out;
654         }
655         if (!*err) {
656                 if (target == blks) {
657                         /*
658                          * save the new block number
659                          * for the first direct block
660                          */
661                         new_blocks[index] = current_block;
662                 }
663                 blk_allocated += ar.len;
664         }
665 allocated:
666         /* total number of blocks allocated for direct blocks */
667         ret = blk_allocated;
668         *err = 0;
669         return ret;
670 failed_out:
671         for (i = 0; i < index; i++)
672                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
673         return ret;
674 }
675
676 /**
677  *      ext4_alloc_branch - allocate and set up a chain of blocks.
678  *      @inode: owner
679  *      @indirect_blks: number of allocated indirect blocks
680  *      @blks: number of allocated direct blocks
681  *      @offsets: offsets (in the blocks) to store the pointers to next.
682  *      @branch: place to store the chain in.
683  *
684  *      This function allocates blocks, zeroes out all but the last one,
685  *      links them into chain and (if we are synchronous) writes them to disk.
686  *      In other words, it prepares a branch that can be spliced onto the
687  *      inode. It stores the information about that chain in the branch[], in
688  *      the same format as ext4_get_branch() would do. We are calling it after
689  *      we had read the existing part of chain and partial points to the last
690  *      triple of that (one with zero ->key). Upon the exit we have the same
691  *      picture as after the successful ext4_get_block(), except that in one
692  *      place chain is disconnected - *branch->p is still zero (we did not
693  *      set the last link), but branch->key contains the number that should
694  *      be placed into *branch->p to fill that gap.
695  *
696  *      If allocation fails we free all blocks we've allocated (and forget
697  *      their buffer_heads) and return the error value the from failed
698  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
699  *      as described above and return 0.
700  */
701 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
702                              ext4_lblk_t iblock, int indirect_blks,
703                              int *blks, ext4_fsblk_t goal,
704                              ext4_lblk_t *offsets, Indirect *branch)
705 {
706         int blocksize = inode->i_sb->s_blocksize;
707         int i, n = 0;
708         int err = 0;
709         struct buffer_head *bh;
710         int num;
711         ext4_fsblk_t new_blocks[4];
712         ext4_fsblk_t current_block;
713
714         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
715                                 *blks, new_blocks, &err);
716         if (err)
717                 return err;
718
719         branch[0].key = cpu_to_le32(new_blocks[0]);
720         /*
721          * metadata blocks and data blocks are allocated.
722          */
723         for (n = 1; n <= indirect_blks;  n++) {
724                 /*
725                  * Get buffer_head for parent block, zero it out
726                  * and set the pointer to new one, then send
727                  * parent to disk.
728                  */
729                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
730                 branch[n].bh = bh;
731                 lock_buffer(bh);
732                 BUFFER_TRACE(bh, "call get_create_access");
733                 err = ext4_journal_get_create_access(handle, bh);
734                 if (err) {
735                         /* Don't brelse(bh) here; it's done in
736                          * ext4_journal_forget() below */
737                         unlock_buffer(bh);
738                         goto failed;
739                 }
740
741                 memset(bh->b_data, 0, blocksize);
742                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
743                 branch[n].key = cpu_to_le32(new_blocks[n]);
744                 *branch[n].p = branch[n].key;
745                 if (n == indirect_blks) {
746                         current_block = new_blocks[n];
747                         /*
748                          * End of chain, update the last new metablock of
749                          * the chain to point to the new allocated
750                          * data blocks numbers
751                          */
752                         for (i = 1; i < num; i++)
753                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
754                 }
755                 BUFFER_TRACE(bh, "marking uptodate");
756                 set_buffer_uptodate(bh);
757                 unlock_buffer(bh);
758
759                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
760                 err = ext4_handle_dirty_metadata(handle, inode, bh);
761                 if (err)
762                         goto failed;
763         }
764         *blks = num;
765         return err;
766 failed:
767         /* Allocation failed, free what we already allocated */
768         ext4_free_blocks(handle, inode, 0, new_blocks[0], 1, 0);
769         for (i = 1; i <= n ; i++) {
770                 /* 
771                  * branch[i].bh is newly allocated, so there is no
772                  * need to revoke the block, which is why we don't
773                  * need to set EXT4_FREE_BLOCKS_METADATA.
774                  */
775                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1,
776                                  EXT4_FREE_BLOCKS_FORGET);
777         }
778         for (i = n+1; i < indirect_blks; i++)
779                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
780
781         ext4_free_blocks(handle, inode, 0, new_blocks[i], num, 0);
782
783         return err;
784 }
785
786 /**
787  * ext4_splice_branch - splice the allocated branch onto inode.
788  * @inode: owner
789  * @block: (logical) number of block we are adding
790  * @chain: chain of indirect blocks (with a missing link - see
791  *      ext4_alloc_branch)
792  * @where: location of missing link
793  * @num:   number of indirect blocks we are adding
794  * @blks:  number of direct blocks we are adding
795  *
796  * This function fills the missing link and does all housekeeping needed in
797  * inode (->i_blocks, etc.). In case of success we end up with the full
798  * chain to new block and return 0.
799  */
800 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
801                               ext4_lblk_t block, Indirect *where, int num,
802                               int blks)
803 {
804         int i;
805         int err = 0;
806         ext4_fsblk_t current_block;
807
808         /*
809          * If we're splicing into a [td]indirect block (as opposed to the
810          * inode) then we need to get write access to the [td]indirect block
811          * before the splice.
812          */
813         if (where->bh) {
814                 BUFFER_TRACE(where->bh, "get_write_access");
815                 err = ext4_journal_get_write_access(handle, where->bh);
816                 if (err)
817                         goto err_out;
818         }
819         /* That's it */
820
821         *where->p = where->key;
822
823         /*
824          * Update the host buffer_head or inode to point to more just allocated
825          * direct blocks blocks
826          */
827         if (num == 0 && blks > 1) {
828                 current_block = le32_to_cpu(where->key) + 1;
829                 for (i = 1; i < blks; i++)
830                         *(where->p + i) = cpu_to_le32(current_block++);
831         }
832
833         /* We are done with atomic stuff, now do the rest of housekeeping */
834         /* had we spliced it onto indirect block? */
835         if (where->bh) {
836                 /*
837                  * If we spliced it onto an indirect block, we haven't
838                  * altered the inode.  Note however that if it is being spliced
839                  * onto an indirect block at the very end of the file (the
840                  * file is growing) then we *will* alter the inode to reflect
841                  * the new i_size.  But that is not done here - it is done in
842                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
843                  */
844                 jbd_debug(5, "splicing indirect only\n");
845                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
846                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
847                 if (err)
848                         goto err_out;
849         } else {
850                 /*
851                  * OK, we spliced it into the inode itself on a direct block.
852                  */
853                 ext4_mark_inode_dirty(handle, inode);
854                 jbd_debug(5, "splicing direct\n");
855         }
856         return err;
857
858 err_out:
859         for (i = 1; i <= num; i++) {
860                 /* 
861                  * branch[i].bh is newly allocated, so there is no
862                  * need to revoke the block, which is why we don't
863                  * need to set EXT4_FREE_BLOCKS_METADATA.
864                  */
865                 ext4_free_blocks(handle, inode, where[i].bh, 0, 1,
866                                  EXT4_FREE_BLOCKS_FORGET);
867         }
868         ext4_free_blocks(handle, inode, 0, le32_to_cpu(where[num].key),
869                          blks, 0);
870
871         return err;
872 }
873
874 /*
875  * The ext4_ind_get_blocks() function handles non-extents inodes
876  * (i.e., using the traditional indirect/double-indirect i_blocks
877  * scheme) for ext4_get_blocks().
878  *
879  * Allocation strategy is simple: if we have to allocate something, we will
880  * have to go the whole way to leaf. So let's do it before attaching anything
881  * to tree, set linkage between the newborn blocks, write them if sync is
882  * required, recheck the path, free and repeat if check fails, otherwise
883  * set the last missing link (that will protect us from any truncate-generated
884  * removals - all blocks on the path are immune now) and possibly force the
885  * write on the parent block.
886  * That has a nice additional property: no special recovery from the failed
887  * allocations is needed - we simply release blocks and do not touch anything
888  * reachable from inode.
889  *
890  * `handle' can be NULL if create == 0.
891  *
892  * return > 0, # of blocks mapped or allocated.
893  * return = 0, if plain lookup failed.
894  * return < 0, error case.
895  *
896  * The ext4_ind_get_blocks() function should be called with
897  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
898  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
899  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
900  * blocks.
901  */
902 static int ext4_ind_get_blocks(handle_t *handle, struct inode *inode,
903                                ext4_lblk_t iblock, unsigned int maxblocks,
904                                struct buffer_head *bh_result,
905                                int flags)
906 {
907         int err = -EIO;
908         ext4_lblk_t offsets[4];
909         Indirect chain[4];
910         Indirect *partial;
911         ext4_fsblk_t goal;
912         int indirect_blks;
913         int blocks_to_boundary = 0;
914         int depth;
915         int count = 0;
916         ext4_fsblk_t first_block = 0;
917
918         J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
919         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
920         depth = ext4_block_to_path(inode, iblock, offsets,
921                                    &blocks_to_boundary);
922
923         if (depth == 0)
924                 goto out;
925
926         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
927
928         /* Simplest case - block found, no allocation needed */
929         if (!partial) {
930                 first_block = le32_to_cpu(chain[depth - 1].key);
931                 clear_buffer_new(bh_result);
932                 count++;
933                 /*map more blocks*/
934                 while (count < maxblocks && count <= blocks_to_boundary) {
935                         ext4_fsblk_t blk;
936
937                         blk = le32_to_cpu(*(chain[depth-1].p + count));
938
939                         if (blk == first_block + count)
940                                 count++;
941                         else
942                                 break;
943                 }
944                 goto got_it;
945         }
946
947         /* Next simple case - plain lookup or failed read of indirect block */
948         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
949                 goto cleanup;
950
951         /*
952          * Okay, we need to do block allocation.
953         */
954         goal = ext4_find_goal(inode, iblock, partial);
955
956         /* the number of blocks need to allocate for [d,t]indirect blocks */
957         indirect_blks = (chain + depth) - partial - 1;
958
959         /*
960          * Next look up the indirect map to count the totoal number of
961          * direct blocks to allocate for this branch.
962          */
963         count = ext4_blks_to_allocate(partial, indirect_blks,
964                                         maxblocks, blocks_to_boundary);
965         /*
966          * Block out ext4_truncate while we alter the tree
967          */
968         err = ext4_alloc_branch(handle, inode, iblock, indirect_blks,
969                                 &count, goal,
970                                 offsets + (partial - chain), partial);
971
972         /*
973          * The ext4_splice_branch call will free and forget any buffers
974          * on the new chain if there is a failure, but that risks using
975          * up transaction credits, especially for bitmaps where the
976          * credits cannot be returned.  Can we handle this somehow?  We
977          * may need to return -EAGAIN upwards in the worst case.  --sct
978          */
979         if (!err)
980                 err = ext4_splice_branch(handle, inode, iblock,
981                                          partial, indirect_blks, count);
982         if (err)
983                 goto cleanup;
984
985         set_buffer_new(bh_result);
986
987         ext4_update_inode_fsync_trans(handle, inode, 1);
988 got_it:
989         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
990         if (count > blocks_to_boundary)
991                 set_buffer_boundary(bh_result);
992         err = count;
993         /* Clean up and exit */
994         partial = chain + depth - 1;    /* the whole chain */
995 cleanup:
996         while (partial > chain) {
997                 BUFFER_TRACE(partial->bh, "call brelse");
998                 brelse(partial->bh);
999                 partial--;
1000         }
1001         BUFFER_TRACE(bh_result, "returned");
1002 out:
1003         return err;
1004 }
1005
1006 #ifdef CONFIG_QUOTA
1007 qsize_t *ext4_get_reserved_space(struct inode *inode)
1008 {
1009         return &EXT4_I(inode)->i_reserved_quota;
1010 }
1011 #endif
1012
1013 /*
1014  * Calculate the number of metadata blocks need to reserve
1015  * to allocate a new block at @lblocks for non extent file based file
1016  */
1017 static int ext4_indirect_calc_metadata_amount(struct inode *inode,
1018                                               sector_t lblock)
1019 {
1020         struct ext4_inode_info *ei = EXT4_I(inode);
1021         int dind_mask = EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1;
1022         int blk_bits;
1023
1024         if (lblock < EXT4_NDIR_BLOCKS)
1025                 return 0;
1026
1027         lblock -= EXT4_NDIR_BLOCKS;
1028
1029         if (ei->i_da_metadata_calc_len &&
1030             (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
1031                 ei->i_da_metadata_calc_len++;
1032                 return 0;
1033         }
1034         ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
1035         ei->i_da_metadata_calc_len = 1;
1036         blk_bits = roundup_pow_of_two(lblock + 1);
1037         return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
1038 }
1039
1040 /*
1041  * Calculate the number of metadata blocks need to reserve
1042  * to allocate a block located at @lblock
1043  */
1044 static int ext4_calc_metadata_amount(struct inode *inode, sector_t lblock)
1045 {
1046         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
1047                 return ext4_ext_calc_metadata_amount(inode, lblock);
1048
1049         return ext4_indirect_calc_metadata_amount(inode, lblock);
1050 }
1051
1052 /*
1053  * Called with i_data_sem down, which is important since we can call
1054  * ext4_discard_preallocations() from here.
1055  */
1056 void ext4_da_update_reserve_space(struct inode *inode,
1057                                         int used, int quota_claim)
1058 {
1059         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1060         struct ext4_inode_info *ei = EXT4_I(inode);
1061         int mdb_free = 0, allocated_meta_blocks = 0;
1062
1063         spin_lock(&ei->i_block_reservation_lock);
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_I(inode)->i_state &= ~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_I(inode)->i_state |= 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         spin_unlock(&ei->i_block_reservation_lock);
1850
1851         /*
1852          * Make quota reservation here to prevent quota overflow
1853          * later. Real quota accounting is done at pages writeout
1854          * time.
1855          */
1856         if (vfs_dq_reserve_block(inode, md_needed + 1))
1857                 return -EDQUOT;
1858
1859         if (ext4_claim_free_blocks(sbi, md_needed + 1)) {
1860                 vfs_dq_release_reservation_block(inode, md_needed + 1);
1861                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1862                         yield();
1863                         goto repeat;
1864                 }
1865                 return -ENOSPC;
1866         }
1867         spin_lock(&ei->i_block_reservation_lock);
1868         ei->i_reserved_data_blocks++;
1869         ei->i_reserved_meta_blocks += md_needed;
1870         spin_unlock(&ei->i_block_reservation_lock);
1871
1872         return 0;       /* success */
1873 }
1874
1875 static void ext4_da_release_space(struct inode *inode, int to_free)
1876 {
1877         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1878         struct ext4_inode_info *ei = EXT4_I(inode);
1879
1880         if (!to_free)
1881                 return;         /* Nothing to release, exit */
1882
1883         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1884
1885         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1886                 /*
1887                  * if there aren't enough reserved blocks, then the
1888                  * counter is messed up somewhere.  Since this
1889                  * function is called from invalidate page, it's
1890                  * harmless to return without any action.
1891                  */
1892                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1893                          "ino %lu, to_free %d with only %d reserved "
1894                          "data blocks\n", inode->i_ino, to_free,
1895                          ei->i_reserved_data_blocks);
1896                 WARN_ON(1);
1897                 to_free = ei->i_reserved_data_blocks;
1898         }
1899         ei->i_reserved_data_blocks -= to_free;
1900
1901         if (ei->i_reserved_data_blocks == 0) {
1902                 /*
1903                  * We can release all of the reserved metadata blocks
1904                  * only when we have written all of the delayed
1905                  * allocation blocks.
1906                  */
1907                 to_free += ei->i_reserved_meta_blocks;
1908                 ei->i_reserved_meta_blocks = 0;
1909                 ei->i_da_metadata_calc_len = 0;
1910         }
1911
1912         /* update fs dirty blocks counter */
1913         percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1914
1915         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1916
1917         vfs_dq_release_reservation_block(inode, to_free);
1918 }
1919
1920 static void ext4_da_page_release_reservation(struct page *page,
1921                                              unsigned long offset)
1922 {
1923         int to_release = 0;
1924         struct buffer_head *head, *bh;
1925         unsigned int curr_off = 0;
1926
1927         head = page_buffers(page);
1928         bh = head;
1929         do {
1930                 unsigned int next_off = curr_off + bh->b_size;
1931
1932                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1933                         to_release++;
1934                         clear_buffer_delay(bh);
1935                 }
1936                 curr_off = next_off;
1937         } while ((bh = bh->b_this_page) != head);
1938         ext4_da_release_space(page->mapping->host, to_release);
1939 }
1940
1941 /*
1942  * Delayed allocation stuff
1943  */
1944
1945 /*
1946  * mpage_da_submit_io - walks through extent of pages and try to write
1947  * them with writepage() call back
1948  *
1949  * @mpd->inode: inode
1950  * @mpd->first_page: first page of the extent
1951  * @mpd->next_page: page after the last page of the extent
1952  *
1953  * By the time mpage_da_submit_io() is called we expect all blocks
1954  * to be allocated. this may be wrong if allocation failed.
1955  *
1956  * As pages are already locked by write_cache_pages(), we can't use it
1957  */
1958 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1959 {
1960         long pages_skipped;
1961         struct pagevec pvec;
1962         unsigned long index, end;
1963         int ret = 0, err, nr_pages, i;
1964         struct inode *inode = mpd->inode;
1965         struct address_space *mapping = inode->i_mapping;
1966
1967         BUG_ON(mpd->next_page <= mpd->first_page);
1968         /*
1969          * We need to start from the first_page to the next_page - 1
1970          * to make sure we also write the mapped dirty buffer_heads.
1971          * If we look at mpd->b_blocknr we would only be looking
1972          * at the currently mapped buffer_heads.
1973          */
1974         index = mpd->first_page;
1975         end = mpd->next_page - 1;
1976
1977         pagevec_init(&pvec, 0);
1978         while (index <= end) {
1979                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1980                 if (nr_pages == 0)
1981                         break;
1982                 for (i = 0; i < nr_pages; i++) {
1983                         struct page *page = pvec.pages[i];
1984
1985                         index = page->index;
1986                         if (index > end)
1987                                 break;
1988                         index++;
1989
1990                         BUG_ON(!PageLocked(page));
1991                         BUG_ON(PageWriteback(page));
1992
1993                         pages_skipped = mpd->wbc->pages_skipped;
1994                         err = mapping->a_ops->writepage(page, mpd->wbc);
1995                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
1996                                 /*
1997                                  * have successfully written the page
1998                                  * without skipping the same
1999                                  */
2000                                 mpd->pages_written++;
2001                         /*
2002                          * In error case, we have to continue because
2003                          * remaining pages are still locked
2004                          * XXX: unlock and re-dirty them?
2005                          */
2006                         if (ret == 0)
2007                                 ret = err;
2008                 }
2009                 pagevec_release(&pvec);
2010         }
2011         return ret;
2012 }
2013
2014 /*
2015  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
2016  *
2017  * @mpd->inode - inode to walk through
2018  * @exbh->b_blocknr - first block on a disk
2019  * @exbh->b_size - amount of space in bytes
2020  * @logical - first logical block to start assignment with
2021  *
2022  * the function goes through all passed space and put actual disk
2023  * block numbers into buffer heads, dropping BH_Delay and BH_Unwritten
2024  */
2025 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
2026                                  struct buffer_head *exbh)
2027 {
2028         struct inode *inode = mpd->inode;
2029         struct address_space *mapping = inode->i_mapping;
2030         int blocks = exbh->b_size >> inode->i_blkbits;
2031         sector_t pblock = exbh->b_blocknr, cur_logical;
2032         struct buffer_head *head, *bh;
2033         pgoff_t index, end;
2034         struct pagevec pvec;
2035         int nr_pages, i;
2036
2037         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2038         end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2039         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2040
2041         pagevec_init(&pvec, 0);
2042
2043         while (index <= end) {
2044                 /* XXX: optimize tail */
2045                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2046                 if (nr_pages == 0)
2047                         break;
2048                 for (i = 0; i < nr_pages; i++) {
2049                         struct page *page = pvec.pages[i];
2050
2051                         index = page->index;
2052                         if (index > end)
2053                                 break;
2054                         index++;
2055
2056                         BUG_ON(!PageLocked(page));
2057                         BUG_ON(PageWriteback(page));
2058                         BUG_ON(!page_has_buffers(page));
2059
2060                         bh = page_buffers(page);
2061                         head = bh;
2062
2063                         /* skip blocks out of the range */
2064                         do {
2065                                 if (cur_logical >= logical)
2066                                         break;
2067                                 cur_logical++;
2068                         } while ((bh = bh->b_this_page) != head);
2069
2070                         do {
2071                                 if (cur_logical >= logical + blocks)
2072                                         break;
2073
2074                                 if (buffer_delay(bh) ||
2075                                                 buffer_unwritten(bh)) {
2076
2077                                         BUG_ON(bh->b_bdev != inode->i_sb->s_bdev);
2078
2079                                         if (buffer_delay(bh)) {
2080                                                 clear_buffer_delay(bh);
2081                                                 bh->b_blocknr = pblock;
2082                                         } else {
2083                                                 /*
2084                                                  * unwritten already should have
2085                                                  * blocknr assigned. Verify that
2086                                                  */
2087                                                 clear_buffer_unwritten(bh);
2088                                                 BUG_ON(bh->b_blocknr != pblock);
2089                                         }
2090
2091                                 } else if (buffer_mapped(bh))
2092                                         BUG_ON(bh->b_blocknr != pblock);
2093
2094                                 cur_logical++;
2095                                 pblock++;
2096                         } while ((bh = bh->b_this_page) != head);
2097                 }
2098                 pagevec_release(&pvec);
2099         }
2100 }
2101
2102
2103 /*
2104  * __unmap_underlying_blocks - just a helper function to unmap
2105  * set of blocks described by @bh
2106  */
2107 static inline void __unmap_underlying_blocks(struct inode *inode,
2108                                              struct buffer_head *bh)
2109 {
2110         struct block_device *bdev = inode->i_sb->s_bdev;
2111         int blocks, i;
2112
2113         blocks = bh->b_size >> inode->i_blkbits;
2114         for (i = 0; i < blocks; i++)
2115                 unmap_underlying_metadata(bdev, bh->b_blocknr + i);
2116 }
2117
2118 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2119                                         sector_t logical, long blk_cnt)
2120 {
2121         int nr_pages, i;
2122         pgoff_t index, end;
2123         struct pagevec pvec;
2124         struct inode *inode = mpd->inode;
2125         struct address_space *mapping = inode->i_mapping;
2126
2127         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2128         end   = (logical + blk_cnt - 1) >>
2129                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2130         while (index <= end) {
2131                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2132                 if (nr_pages == 0)
2133                         break;
2134                 for (i = 0; i < nr_pages; i++) {
2135                         struct page *page = pvec.pages[i];
2136                         index = page->index;
2137                         if (index > end)
2138                                 break;
2139                         index++;
2140
2141                         BUG_ON(!PageLocked(page));
2142                         BUG_ON(PageWriteback(page));
2143                         block_invalidatepage(page, 0);
2144                         ClearPageUptodate(page);
2145                         unlock_page(page);
2146                 }
2147         }
2148         return;
2149 }
2150
2151 static void ext4_print_free_blocks(struct inode *inode)
2152 {
2153         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2154         printk(KERN_CRIT "Total free blocks count %lld\n",
2155                ext4_count_free_blocks(inode->i_sb));
2156         printk(KERN_CRIT "Free/Dirty block details\n");
2157         printk(KERN_CRIT "free_blocks=%lld\n",
2158                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2159         printk(KERN_CRIT "dirty_blocks=%lld\n",
2160                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2161         printk(KERN_CRIT "Block reservation details\n");
2162         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2163                EXT4_I(inode)->i_reserved_data_blocks);
2164         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2165                EXT4_I(inode)->i_reserved_meta_blocks);
2166         return;
2167 }
2168
2169 /*
2170  * mpage_da_map_blocks - go through given space
2171  *
2172  * @mpd - bh describing space
2173  *
2174  * The function skips space we know is already mapped to disk blocks.
2175  *
2176  */
2177 static int mpage_da_map_blocks(struct mpage_da_data *mpd)
2178 {
2179         int err, blks, get_blocks_flags;
2180         struct buffer_head new;
2181         sector_t next = mpd->b_blocknr;
2182         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2183         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2184         handle_t *handle = NULL;
2185
2186         /*
2187          * We consider only non-mapped and non-allocated blocks
2188          */
2189         if ((mpd->b_state  & (1 << BH_Mapped)) &&
2190                 !(mpd->b_state & (1 << BH_Delay)) &&
2191                 !(mpd->b_state & (1 << BH_Unwritten)))
2192                 return 0;
2193
2194         /*
2195          * If we didn't accumulate anything to write simply return
2196          */
2197         if (!mpd->b_size)
2198                 return 0;
2199
2200         handle = ext4_journal_current_handle();
2201         BUG_ON(!handle);
2202
2203         /*
2204          * Call ext4_get_blocks() to allocate any delayed allocation
2205          * blocks, or to convert an uninitialized extent to be
2206          * initialized (in the case where we have written into
2207          * one or more preallocated blocks).
2208          *
2209          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2210          * indicate that we are on the delayed allocation path.  This
2211          * affects functions in many different parts of the allocation
2212          * call path.  This flag exists primarily because we don't
2213          * want to change *many* call functions, so ext4_get_blocks()
2214          * will set the magic i_delalloc_reserved_flag once the
2215          * inode's allocation semaphore is taken.
2216          *
2217          * If the blocks in questions were delalloc blocks, set
2218          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2219          * variables are updated after the blocks have been allocated.
2220          */
2221         new.b_state = 0;
2222         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
2223         if (mpd->b_state & (1 << BH_Delay))
2224                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2225
2226         blks = ext4_get_blocks(handle, mpd->inode, next, max_blocks,
2227                                &new, get_blocks_flags);
2228         if (blks < 0) {
2229                 err = blks;
2230                 /*
2231                  * If get block returns with error we simply
2232                  * return. Later writepage will redirty the page and
2233                  * writepages will find the dirty page again
2234                  */
2235                 if (err == -EAGAIN)
2236                         return 0;
2237
2238                 if (err == -ENOSPC &&
2239                     ext4_count_free_blocks(mpd->inode->i_sb)) {
2240                         mpd->retval = err;
2241                         return 0;
2242                 }
2243
2244                 /*
2245                  * get block failure will cause us to loop in
2246                  * writepages, because a_ops->writepage won't be able
2247                  * to make progress. The page will be redirtied by
2248                  * writepage and writepages will again try to write
2249                  * the same.
2250                  */
2251                 ext4_msg(mpd->inode->i_sb, KERN_CRIT,
2252                          "delayed block allocation failed for inode %lu at "
2253                          "logical offset %llu with max blocks %zd with "
2254                          "error %d\n", mpd->inode->i_ino,
2255                          (unsigned long long) next,
2256                          mpd->b_size >> mpd->inode->i_blkbits, err);
2257                 printk(KERN_CRIT "This should not happen!!  "
2258                        "Data will be lost\n");
2259                 if (err == -ENOSPC) {
2260                         ext4_print_free_blocks(mpd->inode);
2261                 }
2262                 /* invalidate all the pages */
2263                 ext4_da_block_invalidatepages(mpd, next,
2264                                 mpd->b_size >> mpd->inode->i_blkbits);
2265                 return err;
2266         }
2267         BUG_ON(blks == 0);
2268
2269         new.b_size = (blks << mpd->inode->i_blkbits);
2270
2271         if (buffer_new(&new))
2272                 __unmap_underlying_blocks(mpd->inode, &new);
2273
2274         /*
2275          * If blocks are delayed marked, we need to
2276          * put actual blocknr and drop delayed bit
2277          */
2278         if ((mpd->b_state & (1 << BH_Delay)) ||
2279             (mpd->b_state & (1 << BH_Unwritten)))
2280                 mpage_put_bnr_to_bhs(mpd, next, &new);
2281
2282         if (ext4_should_order_data(mpd->inode)) {
2283                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2284                 if (err)
2285                         return err;
2286         }
2287
2288         /*
2289          * Update on-disk size along with block allocation.
2290          */
2291         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2292         if (disksize > i_size_read(mpd->inode))
2293                 disksize = i_size_read(mpd->inode);
2294         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2295                 ext4_update_i_disksize(mpd->inode, disksize);
2296                 return ext4_mark_inode_dirty(handle, mpd->inode);
2297         }
2298
2299         return 0;
2300 }
2301
2302 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2303                 (1 << BH_Delay) | (1 << BH_Unwritten))
2304
2305 /*
2306  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2307  *
2308  * @mpd->lbh - extent of blocks
2309  * @logical - logical number of the block in the file
2310  * @bh - bh of the block (used to access block's state)
2311  *
2312  * the function is used to collect contig. blocks in same state
2313  */
2314 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2315                                    sector_t logical, size_t b_size,
2316                                    unsigned long b_state)
2317 {
2318         sector_t next;
2319         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2320
2321         /* check if thereserved journal credits might overflow */
2322         if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
2323                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2324                         /*
2325                          * With non-extent format we are limited by the journal
2326                          * credit available.  Total credit needed to insert
2327                          * nrblocks contiguous blocks is dependent on the
2328                          * nrblocks.  So limit nrblocks.
2329                          */
2330                         goto flush_it;
2331                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2332                                 EXT4_MAX_TRANS_DATA) {
2333                         /*
2334                          * Adding the new buffer_head would make it cross the
2335                          * allowed limit for which we have journal credit
2336                          * reserved. So limit the new bh->b_size
2337                          */
2338                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2339                                                 mpd->inode->i_blkbits;
2340                         /* we will do mpage_da_submit_io in the next loop */
2341                 }
2342         }
2343         /*
2344          * First block in the extent
2345          */
2346         if (mpd->b_size == 0) {
2347                 mpd->b_blocknr = logical;
2348                 mpd->b_size = b_size;
2349                 mpd->b_state = b_state & BH_FLAGS;
2350                 return;
2351         }
2352
2353         next = mpd->b_blocknr + nrblocks;
2354         /*
2355          * Can we merge the block to our big extent?
2356          */
2357         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2358                 mpd->b_size += b_size;
2359                 return;
2360         }
2361
2362 flush_it:
2363         /*
2364          * We couldn't merge the block to our extent, so we
2365          * need to flush current  extent and start new one
2366          */
2367         if (mpage_da_map_blocks(mpd) == 0)
2368                 mpage_da_submit_io(mpd);
2369         mpd->io_done = 1;
2370         return;
2371 }
2372
2373 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
2374 {
2375         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
2376 }
2377
2378 /*
2379  * __mpage_da_writepage - finds extent of pages and blocks
2380  *
2381  * @page: page to consider
2382  * @wbc: not used, we just follow rules
2383  * @data: context
2384  *
2385  * The function finds extents of pages and scan them for all blocks.
2386  */
2387 static int __mpage_da_writepage(struct page *page,
2388                                 struct writeback_control *wbc, void *data)
2389 {
2390         struct mpage_da_data *mpd = data;
2391         struct inode *inode = mpd->inode;
2392         struct buffer_head *bh, *head;
2393         sector_t logical;
2394
2395         if (mpd->io_done) {
2396                 /*
2397                  * Rest of the page in the page_vec
2398                  * redirty then and skip then. We will
2399                  * try to write them again after
2400                  * starting a new transaction
2401                  */
2402                 redirty_page_for_writepage(wbc, page);
2403                 unlock_page(page);
2404                 return MPAGE_DA_EXTENT_TAIL;
2405         }
2406         /*
2407          * Can we merge this page to current extent?
2408          */
2409         if (mpd->next_page != page->index) {
2410                 /*
2411                  * Nope, we can't. So, we map non-allocated blocks
2412                  * and start IO on them using writepage()
2413                  */
2414                 if (mpd->next_page != mpd->first_page) {
2415                         if (mpage_da_map_blocks(mpd) == 0)
2416                                 mpage_da_submit_io(mpd);
2417                         /*
2418                          * skip rest of the page in the page_vec
2419                          */
2420                         mpd->io_done = 1;
2421                         redirty_page_for_writepage(wbc, page);
2422                         unlock_page(page);
2423                         return MPAGE_DA_EXTENT_TAIL;
2424                 }
2425
2426                 /*
2427                  * Start next extent of pages ...
2428                  */
2429                 mpd->first_page = page->index;
2430
2431                 /*
2432                  * ... and blocks
2433                  */
2434                 mpd->b_size = 0;
2435                 mpd->b_state = 0;
2436                 mpd->b_blocknr = 0;
2437         }
2438
2439         mpd->next_page = page->index + 1;
2440         logical = (sector_t) page->index <<
2441                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2442
2443         if (!page_has_buffers(page)) {
2444                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2445                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2446                 if (mpd->io_done)
2447                         return MPAGE_DA_EXTENT_TAIL;
2448         } else {
2449                 /*
2450                  * Page with regular buffer heads, just add all dirty ones
2451                  */
2452                 head = page_buffers(page);
2453                 bh = head;
2454                 do {
2455                         BUG_ON(buffer_locked(bh));
2456                         /*
2457                          * We need to try to allocate
2458                          * unmapped blocks in the same page.
2459                          * Otherwise we won't make progress
2460                          * with the page in ext4_writepage
2461                          */
2462                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2463                                 mpage_add_bh_to_extent(mpd, logical,
2464                                                        bh->b_size,
2465                                                        bh->b_state);
2466                                 if (mpd->io_done)
2467                                         return MPAGE_DA_EXTENT_TAIL;
2468                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2469                                 /*
2470                                  * mapped dirty buffer. We need to update
2471                                  * the b_state because we look at
2472                                  * b_state in mpage_da_map_blocks. We don't
2473                                  * update b_size because if we find an
2474                                  * unmapped buffer_head later we need to
2475                                  * use the b_state flag of that buffer_head.
2476                                  */
2477                                 if (mpd->b_size == 0)
2478                                         mpd->b_state = bh->b_state & BH_FLAGS;
2479                         }
2480                         logical++;
2481                 } while ((bh = bh->b_this_page) != head);
2482         }
2483
2484         return 0;
2485 }
2486
2487 /*
2488  * This is a special get_blocks_t callback which is used by
2489  * ext4_da_write_begin().  It will either return mapped block or
2490  * reserve space for a single block.
2491  *
2492  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2493  * We also have b_blocknr = -1 and b_bdev initialized properly
2494  *
2495  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2496  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2497  * initialized properly.
2498  */
2499 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2500                                   struct buffer_head *bh_result, int create)
2501 {
2502         int ret = 0;
2503         sector_t invalid_block = ~((sector_t) 0xffff);
2504
2505         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2506                 invalid_block = ~0;
2507
2508         BUG_ON(create == 0);
2509         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2510
2511         /*
2512          * first, we need to know whether the block is allocated already
2513          * preallocated blocks are unmapped but should treated
2514          * the same as allocated blocks.
2515          */
2516         ret = ext4_get_blocks(NULL, inode, iblock, 1,  bh_result, 0);
2517         if ((ret == 0) && !buffer_delay(bh_result)) {
2518                 /* the block isn't (pre)allocated yet, let's reserve space */
2519                 /*
2520                  * XXX: __block_prepare_write() unmaps passed block,
2521                  * is it OK?
2522                  */
2523                 ret = ext4_da_reserve_space(inode, iblock);
2524                 if (ret)
2525                         /* not enough space to reserve */
2526                         return ret;
2527
2528                 map_bh(bh_result, inode->i_sb, invalid_block);
2529                 set_buffer_new(bh_result);
2530                 set_buffer_delay(bh_result);
2531         } else if (ret > 0) {
2532                 bh_result->b_size = (ret << inode->i_blkbits);
2533                 if (buffer_unwritten(bh_result)) {
2534                         /* A delayed write to unwritten bh should
2535                          * be marked new and mapped.  Mapped ensures
2536                          * that we don't do get_block multiple times
2537                          * when we write to the same offset and new
2538                          * ensures that we do proper zero out for
2539                          * partial write.
2540                          */
2541                         set_buffer_new(bh_result);
2542                         set_buffer_mapped(bh_result);
2543                 }
2544                 ret = 0;
2545         }
2546
2547         return ret;
2548 }
2549
2550 /*
2551  * This function is used as a standard get_block_t calback function
2552  * when there is no desire to allocate any blocks.  It is used as a
2553  * callback function for block_prepare_write(), nobh_writepage(), and
2554  * block_write_full_page().  These functions should only try to map a
2555  * single block at a time.
2556  *
2557  * Since this function doesn't do block allocations even if the caller
2558  * requests it by passing in create=1, it is critically important that
2559  * any caller checks to make sure that any buffer heads are returned
2560  * by this function are either all already mapped or marked for
2561  * delayed allocation before calling nobh_writepage() or
2562  * block_write_full_page().  Otherwise, b_blocknr could be left
2563  * unitialized, and the page write functions will be taken by
2564  * surprise.
2565  */
2566 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2567                                    struct buffer_head *bh_result, int create)
2568 {
2569         int ret = 0;
2570         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2571
2572         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2573
2574         /*
2575          * we don't want to do block allocation in writepage
2576          * so call get_block_wrap with create = 0
2577          */
2578         ret = ext4_get_blocks(NULL, inode, iblock, max_blocks, bh_result, 0);
2579         if (ret > 0) {
2580                 bh_result->b_size = (ret << inode->i_blkbits);
2581                 ret = 0;
2582         }
2583         return ret;
2584 }
2585
2586 static int bget_one(handle_t *handle, struct buffer_head *bh)
2587 {
2588         get_bh(bh);
2589         return 0;
2590 }
2591
2592 static int bput_one(handle_t *handle, struct buffer_head *bh)
2593 {
2594         put_bh(bh);
2595         return 0;
2596 }
2597
2598 static int __ext4_journalled_writepage(struct page *page,
2599                                        unsigned int len)
2600 {
2601         struct address_space *mapping = page->mapping;
2602         struct inode *inode = mapping->host;
2603         struct buffer_head *page_bufs;
2604         handle_t *handle = NULL;
2605         int ret = 0;
2606         int err;
2607
2608         page_bufs = page_buffers(page);
2609         BUG_ON(!page_bufs);
2610         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2611         /* As soon as we unlock the page, it can go away, but we have
2612          * references to buffers so we are safe */
2613         unlock_page(page);
2614
2615         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2616         if (IS_ERR(handle)) {
2617                 ret = PTR_ERR(handle);
2618                 goto out;
2619         }
2620
2621         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2622                                 do_journal_get_write_access);
2623
2624         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2625                                 write_end_fn);
2626         if (ret == 0)
2627                 ret = err;
2628         err = ext4_journal_stop(handle);
2629         if (!ret)
2630                 ret = err;
2631
2632         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2633         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
2634 out:
2635         return ret;
2636 }
2637
2638 /*
2639  * Note that we don't need to start a transaction unless we're journaling data
2640  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2641  * need to file the inode to the transaction's list in ordered mode because if
2642  * we are writing back data added by write(), the inode is already there and if
2643  * we are writing back data modified via mmap(), noone guarantees in which
2644  * transaction the data will hit the disk. In case we are journaling data, we
2645  * cannot start transaction directly because transaction start ranks above page
2646  * lock so we have to do some magic.
2647  *
2648  * This function can get called via...
2649  *   - ext4_da_writepages after taking page lock (have journal handle)
2650  *   - journal_submit_inode_data_buffers (no journal handle)
2651  *   - shrink_page_list via pdflush (no journal handle)
2652  *   - grab_page_cache when doing write_begin (have journal handle)
2653  *
2654  * We don't do any block allocation in this function. If we have page with
2655  * multiple blocks we need to write those buffer_heads that are mapped. This
2656  * is important for mmaped based write. So if we do with blocksize 1K
2657  * truncate(f, 1024);
2658  * a = mmap(f, 0, 4096);
2659  * a[0] = 'a';
2660  * truncate(f, 4096);
2661  * we have in the page first buffer_head mapped via page_mkwrite call back
2662  * but other bufer_heads would be unmapped but dirty(dirty done via the
2663  * do_wp_page). So writepage should write the first block. If we modify
2664  * the mmap area beyond 1024 we will again get a page_fault and the
2665  * page_mkwrite callback will do the block allocation and mark the
2666  * buffer_heads mapped.
2667  *
2668  * We redirty the page if we have any buffer_heads that is either delay or
2669  * unwritten in the page.
2670  *
2671  * We can get recursively called as show below.
2672  *
2673  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2674  *              ext4_writepage()
2675  *
2676  * But since we don't do any block allocation we should not deadlock.
2677  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2678  */
2679 static int ext4_writepage(struct page *page,
2680                           struct writeback_control *wbc)
2681 {
2682         int ret = 0;
2683         loff_t size;
2684         unsigned int len;
2685         struct buffer_head *page_bufs;
2686         struct inode *inode = page->mapping->host;
2687
2688         trace_ext4_writepage(inode, page);
2689         size = i_size_read(inode);
2690         if (page->index == size >> PAGE_CACHE_SHIFT)
2691                 len = size & ~PAGE_CACHE_MASK;
2692         else
2693                 len = PAGE_CACHE_SIZE;
2694
2695         if (page_has_buffers(page)) {
2696                 page_bufs = page_buffers(page);
2697                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2698                                         ext4_bh_delay_or_unwritten)) {
2699                         /*
2700                          * We don't want to do  block allocation
2701                          * So redirty the page and return
2702                          * We may reach here when we do a journal commit
2703                          * via journal_submit_inode_data_buffers.
2704                          * If we don't have mapping block we just ignore
2705                          * them. We can also reach here via shrink_page_list
2706                          */
2707                         redirty_page_for_writepage(wbc, page);
2708                         unlock_page(page);
2709                         return 0;
2710                 }
2711         } else {
2712                 /*
2713                  * The test for page_has_buffers() is subtle:
2714                  * We know the page is dirty but it lost buffers. That means
2715                  * that at some moment in time after write_begin()/write_end()
2716                  * has been called all buffers have been clean and thus they
2717                  * must have been written at least once. So they are all
2718                  * mapped and we can happily proceed with mapping them
2719                  * and writing the page.
2720                  *
2721                  * Try to initialize the buffer_heads and check whether
2722                  * all are mapped and non delay. We don't want to
2723                  * do block allocation here.
2724                  */
2725                 ret = block_prepare_write(page, 0, len,
2726                                           noalloc_get_block_write);
2727                 if (!ret) {
2728                         page_bufs = page_buffers(page);
2729                         /* check whether all are mapped and non delay */
2730                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2731                                                 ext4_bh_delay_or_unwritten)) {
2732                                 redirty_page_for_writepage(wbc, page);
2733                                 unlock_page(page);
2734                                 return 0;
2735                         }
2736                 } else {
2737                         /*
2738                          * We can't do block allocation here
2739                          * so just redity the page and unlock
2740                          * and return
2741                          */
2742                         redirty_page_for_writepage(wbc, page);
2743                         unlock_page(page);
2744                         return 0;
2745                 }
2746                 /* now mark the buffer_heads as dirty and uptodate */
2747                 block_commit_write(page, 0, len);
2748         }
2749
2750         if (PageChecked(page) && ext4_should_journal_data(inode)) {
2751                 /*
2752                  * It's mmapped pagecache.  Add buffers and journal it.  There
2753                  * doesn't seem much point in redirtying the page here.
2754                  */
2755                 ClearPageChecked(page);
2756                 return __ext4_journalled_writepage(page, len);
2757         }
2758
2759         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2760                 ret = nobh_writepage(page, noalloc_get_block_write, wbc);
2761         else
2762                 ret = block_write_full_page(page, noalloc_get_block_write,
2763                                             wbc);
2764
2765         return ret;
2766 }
2767
2768 /*
2769  * This is called via ext4_da_writepages() to
2770  * calulate the total number of credits to reserve to fit
2771  * a single extent allocation into a single transaction,
2772  * ext4_da_writpeages() will loop calling this before
2773  * the block allocation.
2774  */
2775
2776 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2777 {
2778         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2779
2780         /*
2781          * With non-extent format the journal credit needed to
2782          * insert nrblocks contiguous block is dependent on
2783          * number of contiguous block. So we will limit
2784          * number of contiguous block to a sane value
2785          */
2786         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) &&
2787             (max_blocks > EXT4_MAX_TRANS_DATA))
2788                 max_blocks = EXT4_MAX_TRANS_DATA;
2789
2790         return ext4_chunk_trans_blocks(inode, max_blocks);
2791 }
2792
2793 static int ext4_da_writepages(struct address_space *mapping,
2794                               struct writeback_control *wbc)
2795 {
2796         pgoff_t index;
2797         int range_whole = 0;
2798         handle_t *handle = NULL;
2799         struct mpage_da_data mpd;
2800         struct inode *inode = mapping->host;
2801         int no_nrwrite_index_update;
2802         int pages_written = 0;
2803         long pages_skipped;
2804         unsigned int max_pages;
2805         int range_cyclic, cycled = 1, io_done = 0;
2806         int needed_blocks, ret = 0;
2807         long desired_nr_to_write, nr_to_writebump = 0;
2808         loff_t range_start = wbc->range_start;
2809         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2810
2811         trace_ext4_da_writepages(inode, wbc);
2812
2813         /*
2814          * No pages to write? This is mainly a kludge to avoid starting
2815          * a transaction for special inodes like journal inode on last iput()
2816          * because that could violate lock ordering on umount
2817          */
2818         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2819                 return 0;
2820
2821         /*
2822          * If the filesystem has aborted, it is read-only, so return
2823          * right away instead of dumping stack traces later on that
2824          * will obscure the real source of the problem.  We test
2825          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2826          * the latter could be true if the filesystem is mounted
2827          * read-only, and in that case, ext4_da_writepages should
2828          * *never* be called, so if that ever happens, we would want
2829          * the stack trace.
2830          */
2831         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2832                 return -EROFS;
2833
2834         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2835                 range_whole = 1;
2836
2837         range_cyclic = wbc->range_cyclic;
2838         if (wbc->range_cyclic) {
2839                 index = mapping->writeback_index;
2840                 if (index)
2841                         cycled = 0;
2842                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2843                 wbc->range_end  = LLONG_MAX;
2844                 wbc->range_cyclic = 0;
2845         } else
2846                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2847
2848         /*
2849          * This works around two forms of stupidity.  The first is in
2850          * the writeback code, which caps the maximum number of pages
2851          * written to be 1024 pages.  This is wrong on multiple
2852          * levels; different architectues have a different page size,
2853          * which changes the maximum amount of data which gets
2854          * written.  Secondly, 4 megabytes is way too small.  XFS
2855          * forces this value to be 16 megabytes by multiplying
2856          * nr_to_write parameter by four, and then relies on its
2857          * allocator to allocate larger extents to make them
2858          * contiguous.  Unfortunately this brings us to the second
2859          * stupidity, which is that ext4's mballoc code only allocates
2860          * at most 2048 blocks.  So we force contiguous writes up to
2861          * the number of dirty blocks in the inode, or
2862          * sbi->max_writeback_mb_bump whichever is smaller.
2863          */
2864         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2865         if (!range_cyclic && range_whole)
2866                 desired_nr_to_write = wbc->nr_to_write * 8;
2867         else
2868                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2869                                                            max_pages);
2870         if (desired_nr_to_write > max_pages)
2871                 desired_nr_to_write = max_pages;
2872
2873         if (wbc->nr_to_write < desired_nr_to_write) {
2874                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2875                 wbc->nr_to_write = desired_nr_to_write;
2876         }
2877
2878         mpd.wbc = wbc;
2879         mpd.inode = mapping->host;
2880
2881         /*
2882          * we don't want write_cache_pages to update
2883          * nr_to_write and writeback_index
2884          */
2885         no_nrwrite_index_update = wbc->no_nrwrite_index_update;
2886         wbc->no_nrwrite_index_update = 1;
2887         pages_skipped = wbc->pages_skipped;
2888
2889 retry:
2890         while (!ret && wbc->nr_to_write > 0) {
2891
2892                 /*
2893                  * we  insert one extent at a time. So we need
2894                  * credit needed for single extent allocation.
2895                  * journalled mode is currently not supported
2896                  * by delalloc
2897                  */
2898                 BUG_ON(ext4_should_journal_data(inode));
2899                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2900
2901                 /* start a new transaction*/
2902                 handle = ext4_journal_start(inode, needed_blocks);
2903                 if (IS_ERR(handle)) {
2904                         ret = PTR_ERR(handle);
2905                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2906                                "%ld pages, ino %lu; err %d\n", __func__,
2907                                 wbc->nr_to_write, inode->i_ino, ret);
2908                         goto out_writepages;
2909                 }
2910
2911                 /*
2912                  * Now call __mpage_da_writepage to find the next
2913                  * contiguous region of logical blocks that need
2914                  * blocks to be allocated by ext4.  We don't actually
2915                  * submit the blocks for I/O here, even though
2916                  * write_cache_pages thinks it will, and will set the
2917                  * pages as clean for write before calling
2918                  * __mpage_da_writepage().
2919                  */
2920                 mpd.b_size = 0;
2921                 mpd.b_state = 0;
2922                 mpd.b_blocknr = 0;
2923                 mpd.first_page = 0;
2924                 mpd.next_page = 0;
2925                 mpd.io_done = 0;
2926                 mpd.pages_written = 0;
2927                 mpd.retval = 0;
2928                 ret = write_cache_pages(mapping, wbc, __mpage_da_writepage,
2929                                         &mpd);
2930                 /*
2931                  * If we have a contiguous extent of pages and we
2932                  * haven't done the I/O yet, map the blocks and submit
2933                  * them for I/O.
2934                  */
2935                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2936                         if (mpage_da_map_blocks(&mpd) == 0)
2937                                 mpage_da_submit_io(&mpd);
2938                         mpd.io_done = 1;
2939                         ret = MPAGE_DA_EXTENT_TAIL;
2940                 }
2941                 trace_ext4_da_write_pages(inode, &mpd);
2942                 wbc->nr_to_write -= mpd.pages_written;
2943
2944                 ext4_journal_stop(handle);
2945
2946                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2947                         /* commit the transaction which would
2948                          * free blocks released in the transaction
2949                          * and try again
2950                          */
2951                         jbd2_journal_force_commit_nested(sbi->s_journal);
2952                         wbc->pages_skipped = pages_skipped;
2953                         ret = 0;
2954                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2955                         /*
2956                          * got one extent now try with
2957                          * rest of the pages
2958                          */
2959                         pages_written += mpd.pages_written;
2960                         wbc->pages_skipped = pages_skipped;
2961                         ret = 0;
2962                         io_done = 1;
2963                 } else if (wbc->nr_to_write)
2964                         /*
2965                          * There is no more writeout needed
2966                          * or we requested for a noblocking writeout
2967                          * and we found the device congested
2968                          */
2969                         break;
2970         }
2971         if (!io_done && !cycled) {
2972                 cycled = 1;
2973                 index = 0;
2974                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2975                 wbc->range_end  = mapping->writeback_index - 1;
2976                 goto retry;
2977         }
2978         if (pages_skipped != wbc->pages_skipped)
2979                 ext4_msg(inode->i_sb, KERN_CRIT,
2980                          "This should not happen leaving %s "
2981                          "with nr_to_write = %ld ret = %d\n",
2982                          __func__, wbc->nr_to_write, ret);
2983
2984         /* Update index */
2985         index += pages_written;
2986         wbc->range_cyclic = range_cyclic;
2987         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2988                 /*
2989                  * set the writeback_index so that range_cyclic
2990                  * mode will write it back later
2991                  */
2992                 mapping->writeback_index = index;
2993
2994 out_writepages:
2995         if (!no_nrwrite_index_update)
2996                 wbc->no_nrwrite_index_update = 0;
2997         wbc->nr_to_write -= nr_to_writebump;
2998         wbc->range_start = range_start;
2999         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
3000         return ret;
3001 }
3002
3003 #define FALL_BACK_TO_NONDELALLOC 1
3004 static int ext4_nonda_switch(struct super_block *sb)
3005 {
3006         s64 free_blocks, dirty_blocks;
3007         struct ext4_sb_info *sbi = EXT4_SB(sb);
3008
3009         /*
3010          * switch to non delalloc mode if we are running low
3011          * on free block. The free block accounting via percpu
3012          * counters can get slightly wrong with percpu_counter_batch getting
3013          * accumulated on each CPU without updating global counters
3014          * Delalloc need an accurate free block accounting. So switch
3015          * to non delalloc when we are near to error range.
3016          */
3017         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
3018         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
3019         if (2 * free_blocks < 3 * dirty_blocks ||
3020                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
3021                 /*
3022                  * free block count is less than 150% of dirty blocks
3023                  * or free blocks is less than watermark
3024                  */
3025                 return 1;
3026         }
3027         /*
3028          * Even if we don't switch but are nearing capacity,
3029          * start pushing delalloc when 1/2 of free blocks are dirty.
3030          */
3031         if (free_blocks < 2 * dirty_blocks)
3032                 writeback_inodes_sb_if_idle(sb);
3033
3034         return 0;
3035 }
3036
3037 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3038                                loff_t pos, unsigned len, unsigned flags,
3039                                struct page **pagep, void **fsdata)
3040 {
3041         int ret, retries = 0, quota_retries = 0;
3042         struct page *page;
3043         pgoff_t index;
3044         unsigned from, to;
3045         struct inode *inode = mapping->host;
3046         handle_t *handle;
3047
3048         index = pos >> PAGE_CACHE_SHIFT;
3049         from = pos & (PAGE_CACHE_SIZE - 1);
3050         to = from + len;
3051
3052         if (ext4_nonda_switch(inode->i_sb)) {
3053                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3054                 return ext4_write_begin(file, mapping, pos,
3055                                         len, flags, pagep, fsdata);
3056         }
3057         *fsdata = (void *)0;
3058         trace_ext4_da_write_begin(inode, pos, len, flags);
3059 retry:
3060         /*
3061          * With delayed allocation, we don't log the i_disksize update
3062          * if there is delayed block allocation. But we still need
3063          * to journalling the i_disksize update if writes to the end
3064          * of file which has an already mapped buffer.
3065          */
3066         handle = ext4_journal_start(inode, 1);
3067         if (IS_ERR(handle)) {
3068                 ret = PTR_ERR(handle);
3069                 goto out;
3070         }
3071         /* We cannot recurse into the filesystem as the transaction is already
3072          * started */
3073         flags |= AOP_FLAG_NOFS;
3074
3075         page = grab_cache_page_write_begin(mapping, index, flags);
3076         if (!page) {
3077                 ext4_journal_stop(handle);
3078                 ret = -ENOMEM;
3079                 goto out;
3080         }
3081         *pagep = page;
3082
3083         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
3084                                 ext4_da_get_block_prep);
3085         if (ret < 0) {
3086                 unlock_page(page);
3087                 ext4_journal_stop(handle);
3088                 page_cache_release(page);
3089                 /*
3090                  * block_write_begin may have instantiated a few blocks
3091                  * outside i_size.  Trim these off again. Don't need
3092                  * i_size_read because we hold i_mutex.
3093                  */
3094                 if (pos + len > inode->i_size)
3095                         ext4_truncate_failed_write(inode);
3096         }
3097
3098         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3099                 goto retry;
3100
3101         if ((ret == -EDQUOT) &&
3102             EXT4_I(inode)->i_reserved_meta_blocks &&
3103             (quota_retries++ < 3)) {
3104                 /*
3105                  * Since we often over-estimate the number of meta
3106                  * data blocks required, we may sometimes get a
3107                  * spurios out of quota error even though there would
3108                  * be enough space once we write the data blocks and
3109                  * find out how many meta data blocks were _really_
3110                  * required.  So try forcing the inode write to see if
3111                  * that helps.
3112                  */
3113                 write_inode_now(inode, (quota_retries == 3));
3114                 goto retry;
3115         }
3116 out:
3117         return ret;
3118 }
3119
3120 /*
3121  * Check if we should update i_disksize
3122  * when write to the end of file but not require block allocation
3123  */
3124 static int ext4_da_should_update_i_disksize(struct page *page,
3125                                             unsigned long offset)
3126 {
3127         struct buffer_head *bh;
3128         struct inode *inode = page->mapping->host;
3129         unsigned int idx;
3130         int i;
3131
3132         bh = page_buffers(page);
3133         idx = offset >> inode->i_blkbits;
3134
3135         for (i = 0; i < idx; i++)
3136                 bh = bh->b_this_page;
3137
3138         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3139                 return 0;
3140         return 1;
3141 }
3142
3143 static int ext4_da_write_end(struct file *file,
3144                              struct address_space *mapping,
3145                              loff_t pos, unsigned len, unsigned copied,
3146                              struct page *page, void *fsdata)
3147 {
3148         struct inode *inode = mapping->host;
3149         int ret = 0, ret2;
3150         handle_t *handle = ext4_journal_current_handle();
3151         loff_t new_i_size;
3152         unsigned long start, end;
3153         int write_mode = (int)(unsigned long)fsdata;
3154
3155         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
3156                 if (ext4_should_order_data(inode)) {
3157                         return ext4_ordered_write_end(file, mapping, pos,
3158                                         len, copied, page, fsdata);
3159                 } else if (ext4_should_writeback_data(inode)) {
3160                         return ext4_writeback_write_end(file, mapping, pos,
3161                                         len, copied, page, fsdata);
3162                 } else {
3163                         BUG();
3164                 }
3165         }
3166
3167         trace_ext4_da_write_end(inode, pos, len, copied);
3168         start = pos & (PAGE_CACHE_SIZE - 1);
3169         end = start + copied - 1;
3170
3171         /*
3172          * generic_write_end() will run mark_inode_dirty() if i_size
3173          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3174          * into that.
3175          */
3176
3177         new_i_size = pos + copied;
3178         if (new_i_size > EXT4_I(inode)->i_disksize) {
3179                 if (ext4_da_should_update_i_disksize(page, end)) {
3180                         down_write(&EXT4_I(inode)->i_data_sem);
3181                         if (new_i_size > EXT4_I(inode)->i_disksize) {
3182                                 /*
3183                                  * Updating i_disksize when extending file
3184                                  * without needing block allocation
3185                                  */
3186                                 if (ext4_should_order_data(inode))
3187                                         ret = ext4_jbd2_file_inode(handle,
3188                                                                    inode);
3189
3190                                 EXT4_I(inode)->i_disksize = new_i_size;
3191                         }
3192                         up_write(&EXT4_I(inode)->i_data_sem);
3193                         /* We need to mark inode dirty even if
3194                          * new_i_size is less that inode->i_size
3195                          * bu greater than i_disksize.(hint delalloc)
3196                          */
3197                         ext4_mark_inode_dirty(handle, inode);
3198                 }
3199         }
3200         ret2 = generic_write_end(file, mapping, pos, len, copied,
3201                                                         page, fsdata);
3202         copied = ret2;
3203         if (ret2 < 0)
3204                 ret = ret2;
3205         ret2 = ext4_journal_stop(handle);
3206         if (!ret)
3207                 ret = ret2;
3208
3209         return ret ? ret : copied;
3210 }
3211
3212 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
3213 {
3214         /*
3215          * Drop reserved blocks
3216          */
3217         BUG_ON(!PageLocked(page));
3218         if (!page_has_buffers(page))
3219                 goto out;
3220
3221         ext4_da_page_release_reservation(page, offset);
3222
3223 out:
3224         ext4_invalidatepage(page, offset);
3225
3226         return;
3227 }
3228
3229 /*
3230  * Force all delayed allocation blocks to be allocated for a given inode.
3231  */
3232 int ext4_alloc_da_blocks(struct inode *inode)
3233 {
3234         trace_ext4_alloc_da_blocks(inode);
3235
3236         if (!EXT4_I(inode)->i_reserved_data_blocks &&
3237             !EXT4_I(inode)->i_reserved_meta_blocks)
3238                 return 0;
3239
3240         /*
3241          * We do something simple for now.  The filemap_flush() will
3242          * also start triggering a write of the data blocks, which is
3243          * not strictly speaking necessary (and for users of
3244          * laptop_mode, not even desirable).  However, to do otherwise
3245          * would require replicating code paths in:
3246          *
3247          * ext4_da_writepages() ->
3248          *    write_cache_pages() ---> (via passed in callback function)
3249          *        __mpage_da_writepage() -->
3250          *           mpage_add_bh_to_extent()
3251          *           mpage_da_map_blocks()
3252          *
3253          * The problem is that write_cache_pages(), located in
3254          * mm/page-writeback.c, marks pages clean in preparation for
3255          * doing I/O, which is not desirable if we're not planning on
3256          * doing I/O at all.
3257          *
3258          * We could call write_cache_pages(), and then redirty all of
3259          * the pages by calling redirty_page_for_writeback() but that
3260          * would be ugly in the extreme.  So instead we would need to
3261          * replicate parts of the code in the above functions,
3262          * simplifying them becuase we wouldn't actually intend to
3263          * write out the pages, but rather only collect contiguous
3264          * logical block extents, call the multi-block allocator, and
3265          * then update the buffer heads with the block allocations.
3266          *
3267          * For now, though, we'll cheat by calling filemap_flush(),
3268          * which will map the blocks, and start the I/O, but not
3269          * actually wait for the I/O to complete.
3270          */
3271         return filemap_flush(inode->i_mapping);
3272 }
3273
3274 /*
3275  * bmap() is special.  It gets used by applications such as lilo and by
3276  * the swapper to find the on-disk block of a specific piece of data.
3277  *
3278  * Naturally, this is dangerous if the block concerned is still in the
3279  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3280  * filesystem and enables swap, then they may get a nasty shock when the
3281  * data getting swapped to that swapfile suddenly gets overwritten by
3282  * the original zero's written out previously to the journal and
3283  * awaiting writeback in the kernel's buffer cache.
3284  *
3285  * So, if we see any bmap calls here on a modified, data-journaled file,
3286  * take extra steps to flush any blocks which might be in the cache.
3287  */
3288 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3289 {
3290         struct inode *inode = mapping->host;
3291         journal_t *journal;
3292         int err;
3293
3294         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3295                         test_opt(inode->i_sb, DELALLOC)) {
3296                 /*
3297                  * With delalloc we want to sync the file
3298                  * so that we can make sure we allocate
3299                  * blocks for file
3300                  */
3301                 filemap_write_and_wait(mapping);
3302         }
3303
3304         if (EXT4_JOURNAL(inode) && EXT4_I(inode)->i_state & EXT4_STATE_JDATA) {
3305                 /*
3306                  * This is a REALLY heavyweight approach, but the use of
3307                  * bmap on dirty files is expected to be extremely rare:
3308                  * only if we run lilo or swapon on a freshly made file
3309                  * do we expect this to happen.
3310                  *
3311                  * (bmap requires CAP_SYS_RAWIO so this does not
3312                  * represent an unprivileged user DOS attack --- we'd be
3313                  * in trouble if mortal users could trigger this path at
3314                  * will.)
3315                  *
3316                  * NB. EXT4_STATE_JDATA is not set on files other than
3317                  * regular files.  If somebody wants to bmap a directory
3318                  * or symlink and gets confused because the buffer
3319                  * hasn't yet been flushed to disk, they deserve
3320                  * everything they get.
3321                  */
3322
3323                 EXT4_I(inode)->i_state &= ~EXT4_STATE_JDATA;
3324                 journal = EXT4_JOURNAL(inode);
3325                 jbd2_journal_lock_updates(journal);
3326                 err = jbd2_journal_flush(journal);
3327                 jbd2_journal_unlock_updates(journal);
3328
3329                 if (err)
3330                         return 0;
3331         }
3332
3333         return generic_block_bmap(mapping, block, ext4_get_block);
3334 }
3335
3336 static int ext4_readpage(struct file *file, struct page *page)
3337 {
3338         return mpage_readpage(page, ext4_get_block);
3339 }
3340
3341 static int
3342 ext4_readpages(struct file *file, struct address_space *mapping,
3343                 struct list_head *pages, unsigned nr_pages)
3344 {
3345         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3346 }
3347
3348 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3349 {
3350         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3351
3352         /*
3353          * If it's a full truncate we just forget about the pending dirtying
3354          */
3355         if (offset == 0)
3356                 ClearPageChecked(page);
3357
3358         if (journal)
3359                 jbd2_journal_invalidatepage(journal, page, offset);
3360         else
3361                 block_invalidatepage(page, offset);
3362 }
3363
3364 static int ext4_releasepage(struct page *page, gfp_t wait)
3365 {
3366         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3367
3368         WARN_ON(PageChecked(page));
3369         if (!page_has_buffers(page))
3370                 return 0;
3371         if (journal)
3372                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3373         else
3374                 return try_to_free_buffers(page);
3375 }
3376
3377 /*
3378  * O_DIRECT for ext3 (or indirect map) based files
3379  *
3380  * If the O_DIRECT write will extend the file then add this inode to the
3381  * orphan list.  So recovery will truncate it back to the original size
3382  * if the machine crashes during the write.
3383  *
3384  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3385  * crashes then stale disk data _may_ be exposed inside the file. But current
3386  * VFS code falls back into buffered path in that case so we are safe.
3387  */
3388 static ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
3389                               const struct iovec *iov, loff_t offset,
3390                               unsigned long nr_segs)
3391 {
3392         struct file *file = iocb->ki_filp;
3393         struct inode *inode = file->f_mapping->host;
3394         struct ext4_inode_info *ei = EXT4_I(inode);
3395         handle_t *handle;
3396         ssize_t ret;
3397         int orphan = 0;
3398         size_t count = iov_length(iov, nr_segs);
3399         int retries = 0;
3400
3401         if (rw == WRITE) {
3402                 loff_t final_size = offset + count;
3403
3404                 if (final_size > inode->i_size) {
3405                         /* Credits for sb + inode write */
3406                         handle = ext4_journal_start(inode, 2);
3407                         if (IS_ERR(handle)) {
3408                                 ret = PTR_ERR(handle);
3409                                 goto out;
3410                         }
3411                         ret = ext4_orphan_add(handle, inode);
3412                         if (ret) {
3413                                 ext4_journal_stop(handle);
3414                                 goto out;
3415                         }
3416                         orphan = 1;
3417                         ei->i_disksize = inode->i_size;
3418                         ext4_journal_stop(handle);
3419                 }
3420         }
3421
3422 retry:
3423         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
3424                                  offset, nr_segs,
3425                                  ext4_get_block, NULL);
3426         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3427                 goto retry;
3428
3429         if (orphan) {
3430                 int err;
3431
3432                 /* Credits for sb + inode write */
3433                 handle = ext4_journal_start(inode, 2);
3434                 if (IS_ERR(handle)) {
3435                         /* This is really bad luck. We've written the data
3436                          * but cannot extend i_size. Bail out and pretend
3437                          * the write failed... */
3438                         ret = PTR_ERR(handle);
3439                         goto out;
3440                 }
3441                 if (inode->i_nlink)
3442                         ext4_orphan_del(handle, inode);
3443                 if (ret > 0) {
3444                         loff_t end = offset + ret;
3445                         if (end > inode->i_size) {
3446                                 ei->i_disksize = end;
3447                                 i_size_write(inode, end);
3448                                 /*
3449                                  * We're going to return a positive `ret'
3450                                  * here due to non-zero-length I/O, so there's
3451                                  * no way of reporting error returns from
3452                                  * ext4_mark_inode_dirty() to userspace.  So
3453                                  * ignore it.
3454                                  */
3455                                 ext4_mark_inode_dirty(handle, inode);
3456                         }
3457                 }
3458                 err = ext4_journal_stop(handle);
3459                 if (ret == 0)
3460                         ret = err;
3461         }
3462 out:
3463         return ret;
3464 }
3465
3466 static int ext4_get_block_dio_write(struct inode *inode, sector_t iblock,
3467                    struct buffer_head *bh_result, int create)
3468 {
3469         handle_t *handle = NULL;
3470         int ret = 0;
3471         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
3472         int dio_credits;
3473
3474         ext4_debug("ext4_get_block_dio_write: inode %lu, create flag %d\n",
3475                    inode->i_ino, create);
3476         /*
3477          * DIO VFS code passes create = 0 flag for write to
3478          * the middle of file. It does this to avoid block
3479          * allocation for holes, to prevent expose stale data
3480          * out when there is parallel buffered read (which does
3481          * not hold the i_mutex lock) while direct IO write has
3482          * not completed. DIO request on holes finally falls back
3483          * to buffered IO for this reason.
3484          *
3485          * For ext4 extent based file, since we support fallocate,
3486          * new allocated extent as uninitialized, for holes, we
3487          * could fallocate blocks for holes, thus parallel
3488          * buffered IO read will zero out the page when read on
3489          * a hole while parallel DIO write to the hole has not completed.
3490          *
3491          * when we come here, we know it's a direct IO write to
3492          * to the middle of file (<i_size)
3493          * so it's safe to override the create flag from VFS.
3494          */
3495         create = EXT4_GET_BLOCKS_DIO_CREATE_EXT;
3496
3497         if (max_blocks > DIO_MAX_BLOCKS)
3498                 max_blocks = DIO_MAX_BLOCKS;
3499         dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
3500         handle = ext4_journal_start(inode, dio_credits);
3501         if (IS_ERR(handle)) {
3502                 ret = PTR_ERR(handle);
3503                 goto out;
3504         }
3505         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
3506                               create);
3507         if (ret > 0) {
3508                 bh_result->b_size = (ret << inode->i_blkbits);
3509                 ret = 0;
3510         }
3511         ext4_journal_stop(handle);
3512 out:
3513         return ret;
3514 }
3515
3516 static void ext4_free_io_end(ext4_io_end_t *io)
3517 {
3518         BUG_ON(!io);
3519         iput(io->inode);
3520         kfree(io);
3521 }
3522 static void dump_aio_dio_list(struct inode * inode)
3523 {
3524 #ifdef  EXT4_DEBUG
3525         struct list_head *cur, *before, *after;
3526         ext4_io_end_t *io, *io0, *io1;
3527
3528         if (list_empty(&EXT4_I(inode)->i_aio_dio_complete_list)){
3529                 ext4_debug("inode %lu aio dio list is empty\n", inode->i_ino);
3530                 return;
3531         }
3532
3533         ext4_debug("Dump inode %lu aio_dio_completed_IO list \n", inode->i_ino);
3534         list_for_each_entry(io, &EXT4_I(inode)->i_aio_dio_complete_list, list){
3535                 cur = &io->list;
3536                 before = cur->prev;
3537                 io0 = container_of(before, ext4_io_end_t, list);
3538                 after = cur->next;
3539                 io1 = container_of(after, ext4_io_end_t, list);
3540
3541                 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
3542                             io, inode->i_ino, io0, io1);
3543         }
3544 #endif
3545 }
3546
3547 /*
3548  * check a range of space and convert unwritten extents to written.
3549  */
3550 static int ext4_end_aio_dio_nolock(ext4_io_end_t *io)
3551 {
3552         struct inode *inode = io->inode;
3553         loff_t offset = io->offset;
3554         size_t size = io->size;
3555         int ret = 0;
3556
3557         ext4_debug("end_aio_dio_onlock: io 0x%p from inode %lu,list->next 0x%p,"
3558                    "list->prev 0x%p\n",
3559                    io, inode->i_ino, io->list.next, io->list.prev);
3560
3561         if (list_empty(&io->list))
3562                 return ret;
3563
3564         if (io->flag != DIO_AIO_UNWRITTEN)
3565                 return ret;
3566
3567         if (offset + size <= i_size_read(inode))
3568                 ret = ext4_convert_unwritten_extents(inode, offset, size);
3569
3570         if (ret < 0) {
3571                 printk(KERN_EMERG "%s: failed to convert unwritten"
3572                         "extents to written extents, error is %d"
3573                         " io is still on inode %lu aio dio list\n",
3574                        __func__, ret, inode->i_ino);
3575                 return ret;
3576         }
3577
3578         /* clear the DIO AIO unwritten flag */
3579         io->flag = 0;
3580         return ret;
3581 }
3582 /*
3583  * work on completed aio dio IO, to convert unwritten extents to extents
3584  */
3585 static void ext4_end_aio_dio_work(struct work_struct *work)
3586 {
3587         ext4_io_end_t *io  = container_of(work, ext4_io_end_t, work);
3588         struct inode *inode = io->inode;
3589         int ret = 0;
3590
3591         mutex_lock(&inode->i_mutex);
3592         ret = ext4_end_aio_dio_nolock(io);
3593         if (ret >= 0) {
3594                 if (!list_empty(&io->list))
3595                         list_del_init(&io->list);
3596                 ext4_free_io_end(io);
3597         }
3598         mutex_unlock(&inode->i_mutex);
3599 }
3600 /*
3601  * This function is called from ext4_sync_file().
3602  *
3603  * When AIO DIO IO is completed, the work to convert unwritten
3604  * extents to written is queued on workqueue but may not get immediately
3605  * scheduled. When fsync is called, we need to ensure the
3606  * conversion is complete before fsync returns.
3607  * The inode keeps track of a list of completed AIO from DIO path
3608  * that might needs to do the conversion. This function walks through
3609  * the list and convert the related unwritten extents to written.
3610  */
3611 int flush_aio_dio_completed_IO(struct inode *inode)
3612 {
3613         ext4_io_end_t *io;
3614         int ret = 0;
3615         int ret2 = 0;
3616
3617         if (list_empty(&EXT4_I(inode)->i_aio_dio_complete_list))
3618                 return ret;
3619
3620         dump_aio_dio_list(inode);
3621         while (!list_empty(&EXT4_I(inode)->i_aio_dio_complete_list)){
3622                 io = list_entry(EXT4_I(inode)->i_aio_dio_complete_list.next,
3623                                 ext4_io_end_t, list);
3624                 /*
3625                  * Calling ext4_end_aio_dio_nolock() to convert completed
3626                  * IO to written.
3627                  *
3628                  * When ext4_sync_file() is called, run_queue() may already
3629                  * about to flush the work corresponding to this io structure.
3630                  * It will be upset if it founds the io structure related
3631                  * to the work-to-be schedule is freed.
3632                  *
3633                  * Thus we need to keep the io structure still valid here after
3634                  * convertion finished. The io structure has a flag to
3635                  * avoid double converting from both fsync and background work
3636                  * queue work.
3637                  */
3638                 ret = ext4_end_aio_dio_nolock(io);
3639                 if (ret < 0)
3640                         ret2 = ret;
3641                 else
3642                         list_del_init(&io->list);
3643         }
3644         return (ret2 < 0) ? ret2 : 0;
3645 }
3646
3647 static ext4_io_end_t *ext4_init_io_end (struct inode *inode)
3648 {
3649         ext4_io_end_t *io = NULL;
3650
3651         io = kmalloc(sizeof(*io), GFP_NOFS);
3652
3653         if (io) {
3654                 igrab(inode);
3655                 io->inode = inode;
3656                 io->flag = 0;
3657                 io->offset = 0;
3658                 io->size = 0;
3659                 io->error = 0;
3660                 INIT_WORK(&io->work, ext4_end_aio_dio_work);
3661                 INIT_LIST_HEAD(&io->list);
3662         }
3663
3664         return io;
3665 }
3666
3667 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3668                             ssize_t size, void *private)
3669 {
3670         ext4_io_end_t *io_end = iocb->private;
3671         struct workqueue_struct *wq;
3672
3673         /* if not async direct IO or dio with 0 bytes write, just return */
3674         if (!io_end || !size)
3675                 return;
3676
3677         ext_debug("ext4_end_io_dio(): io_end 0x%p"
3678                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
3679                   iocb->private, io_end->inode->i_ino, iocb, offset,
3680                   size);
3681
3682         /* if not aio dio with unwritten extents, just free io and return */
3683         if (io_end->flag != DIO_AIO_UNWRITTEN){
3684                 ext4_free_io_end(io_end);
3685                 iocb->private = NULL;
3686                 return;
3687         }
3688
3689         io_end->offset = offset;
3690         io_end->size = size;
3691         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
3692
3693         /* queue the work to convert unwritten extents to written */
3694         queue_work(wq, &io_end->work);
3695
3696         /* Add the io_end to per-inode completed aio dio list*/
3697         list_add_tail(&io_end->list,
3698                  &EXT4_I(io_end->inode)->i_aio_dio_complete_list);
3699         iocb->private = NULL;
3700 }
3701 /*
3702  * For ext4 extent files, ext4 will do direct-io write to holes,
3703  * preallocated extents, and those write extend the file, no need to
3704  * fall back to buffered IO.
3705  *
3706  * For holes, we fallocate those blocks, mark them as unintialized
3707  * If those blocks were preallocated, we mark sure they are splited, but
3708  * still keep the range to write as unintialized.
3709  *
3710  * The unwrritten extents will be converted to written when DIO is completed.
3711  * For async direct IO, since the IO may still pending when return, we
3712  * set up an end_io call back function, which will do the convertion
3713  * when async direct IO completed.
3714  *
3715  * If the O_DIRECT write will extend the file then add this inode to the
3716  * orphan list.  So recovery will truncate it back to the original size
3717  * if the machine crashes during the write.
3718  *
3719  */
3720 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
3721                               const struct iovec *iov, loff_t offset,
3722                               unsigned long nr_segs)
3723 {
3724         struct file *file = iocb->ki_filp;
3725         struct inode *inode = file->f_mapping->host;
3726         ssize_t ret;
3727         size_t count = iov_length(iov, nr_segs);
3728
3729         loff_t final_size = offset + count;
3730         if (rw == WRITE && final_size <= inode->i_size) {
3731                 /*
3732                  * We could direct write to holes and fallocate.
3733                  *
3734                  * Allocated blocks to fill the hole are marked as uninitialized
3735                  * to prevent paralel buffered read to expose the stale data
3736                  * before DIO complete the data IO.
3737                  *
3738                  * As to previously fallocated extents, ext4 get_block
3739                  * will just simply mark the buffer mapped but still
3740                  * keep the extents uninitialized.
3741                  *
3742                  * for non AIO case, we will convert those unwritten extents
3743                  * to written after return back from blockdev_direct_IO.
3744                  *
3745                  * for async DIO, the conversion needs to be defered when
3746                  * the IO is completed. The ext4 end_io callback function
3747                  * will be called to take care of the conversion work.
3748                  * Here for async case, we allocate an io_end structure to
3749                  * hook to the iocb.
3750                  */
3751                 iocb->private = NULL;
3752                 EXT4_I(inode)->cur_aio_dio = NULL;
3753                 if (!is_sync_kiocb(iocb)) {
3754                         iocb->private = ext4_init_io_end(inode);
3755                         if (!iocb->private)
3756                                 return -ENOMEM;
3757                         /*
3758                          * we save the io structure for current async
3759                          * direct IO, so that later ext4_get_blocks()
3760                          * could flag the io structure whether there
3761                          * is a unwritten extents needs to be converted
3762                          * when IO is completed.
3763                          */
3764                         EXT4_I(inode)->cur_aio_dio = iocb->private;
3765                 }
3766
3767                 ret = blockdev_direct_IO(rw, iocb, inode,
3768                                          inode->i_sb->s_bdev, iov,
3769                                          offset, nr_segs,
3770                                          ext4_get_block_dio_write,
3771                                          ext4_end_io_dio);
3772                 if (iocb->private)
3773                         EXT4_I(inode)->cur_aio_dio = NULL;
3774                 /*
3775                  * The io_end structure takes a reference to the inode,
3776                  * that structure needs to be destroyed and the
3777                  * reference to the inode need to be dropped, when IO is
3778                  * complete, even with 0 byte write, or failed.
3779                  *
3780                  * In the successful AIO DIO case, the io_end structure will be
3781                  * desctroyed and the reference to the inode will be dropped
3782                  * after the end_io call back function is called.
3783                  *
3784                  * In the case there is 0 byte write, or error case, since
3785                  * VFS direct IO won't invoke the end_io call back function,
3786                  * we need to free the end_io structure here.
3787                  */
3788                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
3789                         ext4_free_io_end(iocb->private);
3790                         iocb->private = NULL;
3791                 } else if (ret > 0 && (EXT4_I(inode)->i_state &
3792                                        EXT4_STATE_DIO_UNWRITTEN)) {
3793                         int err;
3794                         /*
3795                          * for non AIO case, since the IO is already
3796                          * completed, we could do the convertion right here
3797                          */
3798                         err = ext4_convert_unwritten_extents(inode,
3799                                                              offset, ret);
3800                         if (err < 0)
3801                                 ret = err;
3802                         EXT4_I(inode)->i_state &= ~EXT4_STATE_DIO_UNWRITTEN;
3803                 }
3804                 return ret;
3805         }
3806
3807         /* for write the the end of file case, we fall back to old way */
3808         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3809 }
3810
3811 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3812                               const struct iovec *iov, loff_t offset,
3813                               unsigned long nr_segs)
3814 {
3815         struct file *file = iocb->ki_filp;
3816         struct inode *inode = file->f_mapping->host;
3817
3818         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
3819                 return ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
3820
3821         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3822 }
3823
3824 /*
3825  * Pages can be marked dirty completely asynchronously from ext4's journalling
3826  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3827  * much here because ->set_page_dirty is called under VFS locks.  The page is
3828  * not necessarily locked.
3829  *
3830  * We cannot just dirty the page and leave attached buffers clean, because the
3831  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3832  * or jbddirty because all the journalling code will explode.
3833  *
3834  * So what we do is to mark the page "pending dirty" and next time writepage
3835  * is called, propagate that into the buffers appropriately.
3836  */
3837 static int ext4_journalled_set_page_dirty(struct page *page)
3838 {
3839         SetPageChecked(page);
3840         return __set_page_dirty_nobuffers(page);
3841 }
3842
3843 static const struct address_space_operations ext4_ordered_aops = {
3844         .readpage               = ext4_readpage,
3845         .readpages              = ext4_readpages,
3846         .writepage              = ext4_writepage,
3847         .sync_page              = block_sync_page,
3848         .write_begin            = ext4_write_begin,
3849         .write_end              = ext4_ordered_write_end,
3850         .bmap                   = ext4_bmap,
3851         .invalidatepage         = ext4_invalidatepage,
3852         .releasepage            = ext4_releasepage,
3853         .direct_IO              = ext4_direct_IO,
3854         .migratepage            = buffer_migrate_page,
3855         .is_partially_uptodate  = block_is_partially_uptodate,
3856         .error_remove_page      = generic_error_remove_page,
3857 };
3858
3859 static const struct address_space_operations ext4_writeback_aops = {
3860         .readpage               = ext4_readpage,
3861         .readpages              = ext4_readpages,
3862         .writepage              = ext4_writepage,
3863         .sync_page              = block_sync_page,
3864         .write_begin            = ext4_write_begin,
3865         .write_end              = ext4_writeback_write_end,
3866         .bmap                   = ext4_bmap,
3867         .invalidatepage         = ext4_invalidatepage,
3868         .releasepage            = ext4_releasepage,
3869         .direct_IO              = ext4_direct_IO,
3870         .migratepage            = buffer_migrate_page,
3871         .is_partially_uptodate  = block_is_partially_uptodate,
3872         .error_remove_page      = generic_error_remove_page,
3873 };
3874
3875 static const struct address_space_operations ext4_journalled_aops = {
3876         .readpage               = ext4_readpage,
3877         .readpages              = ext4_readpages,
3878         .writepage              = ext4_writepage,
3879         .sync_page              = block_sync_page,
3880         .write_begin            = ext4_write_begin,
3881         .write_end              = ext4_journalled_write_end,
3882         .set_page_dirty         = ext4_journalled_set_page_dirty,
3883         .bmap                   = ext4_bmap,
3884         .invalidatepage         = ext4_invalidatepage,
3885         .releasepage            = ext4_releasepage,
3886         .is_partially_uptodate  = block_is_partially_uptodate,
3887         .error_remove_page      = generic_error_remove_page,
3888 };
3889
3890 static const struct address_space_operations ext4_da_aops = {
3891         .readpage               = ext4_readpage,
3892         .readpages              = ext4_readpages,
3893         .writepage              = ext4_writepage,
3894         .writepages             = ext4_da_writepages,
3895         .sync_page              = block_sync_page,
3896         .write_begin            = ext4_da_write_begin,
3897         .write_end              = ext4_da_write_end,
3898         .bmap                   = ext4_bmap,
3899         .invalidatepage         = ext4_da_invalidatepage,
3900         .releasepage            = ext4_releasepage,
3901         .direct_IO              = ext4_direct_IO,
3902         .migratepage            = buffer_migrate_page,
3903         .is_partially_uptodate  = block_is_partially_uptodate,
3904         .error_remove_page      = generic_error_remove_page,
3905 };
3906
3907 void ext4_set_aops(struct inode *inode)
3908 {
3909         if (ext4_should_order_data(inode) &&
3910                 test_opt(inode->i_sb, DELALLOC))
3911                 inode->i_mapping->a_ops = &ext4_da_aops;
3912         else if (ext4_should_order_data(inode))
3913                 inode->i_mapping->a_ops = &ext4_ordered_aops;
3914         else if (ext4_should_writeback_data(inode) &&
3915                  test_opt(inode->i_sb, DELALLOC))
3916                 inode->i_mapping->a_ops = &ext4_da_aops;
3917         else if (ext4_should_writeback_data(inode))
3918                 inode->i_mapping->a_ops = &ext4_writeback_aops;
3919         else
3920                 inode->i_mapping->a_ops = &ext4_journalled_aops;
3921 }
3922
3923 /*
3924  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3925  * up to the end of the block which corresponds to `from'.
3926  * This required during truncate. We need to physically zero the tail end
3927  * of that block so it doesn't yield old data if the file is later grown.
3928  */
3929 int ext4_block_truncate_page(handle_t *handle,
3930                 struct address_space *mapping, loff_t from)
3931 {
3932         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3933         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3934         unsigned blocksize, length, pos;
3935         ext4_lblk_t iblock;
3936         struct inode *inode = mapping->host;
3937         struct buffer_head *bh;
3938         struct page *page;
3939         int err = 0;
3940
3941         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
3942                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
3943         if (!page)
3944                 return -EINVAL;
3945
3946         blocksize = inode->i_sb->s_blocksize;
3947         length = blocksize - (offset & (blocksize - 1));
3948         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3949
3950         /*
3951          * For "nobh" option,  we can only work if we don't need to
3952          * read-in the page - otherwise we create buffers to do the IO.
3953          */
3954         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
3955              ext4_should_writeback_data(inode) && PageUptodate(page)) {
3956                 zero_user(page, offset, length);
3957                 set_page_dirty(page);
3958                 goto unlock;
3959         }
3960
3961         if (!page_has_buffers(page))
3962                 create_empty_buffers(page, blocksize, 0);
3963
3964         /* Find the buffer that contains "offset" */
3965         bh = page_buffers(page);
3966         pos = blocksize;
3967         while (offset >= pos) {
3968                 bh = bh->b_this_page;
3969                 iblock++;
3970                 pos += blocksize;
3971         }
3972
3973         err = 0;
3974         if (buffer_freed(bh)) {
3975                 BUFFER_TRACE(bh, "freed: skip");
3976                 goto unlock;
3977         }
3978
3979         if (!buffer_mapped(bh)) {
3980                 BUFFER_TRACE(bh, "unmapped");
3981                 ext4_get_block(inode, iblock, bh, 0);
3982                 /* unmapped? It's a hole - nothing to do */
3983                 if (!buffer_mapped(bh)) {
3984                         BUFFER_TRACE(bh, "still unmapped");
3985                         goto unlock;
3986                 }
3987         }
3988
3989         /* Ok, it's mapped. Make sure it's up-to-date */
3990         if (PageUptodate(page))
3991                 set_buffer_uptodate(bh);
3992
3993         if (!buffer_uptodate(bh)) {
3994                 err = -EIO;
3995                 ll_rw_block(READ, 1, &bh);
3996                 wait_on_buffer(bh);
3997                 /* Uhhuh. Read error. Complain and punt. */
3998                 if (!buffer_uptodate(bh))
3999                         goto unlock;
4000         }
4001
4002         if (ext4_should_journal_data(inode)) {
4003                 BUFFER_TRACE(bh, "get write access");
4004                 err = ext4_journal_get_write_access(handle, bh);
4005                 if (err)
4006                         goto unlock;
4007         }
4008
4009         zero_user(page, offset, length);
4010
4011         BUFFER_TRACE(bh, "zeroed end of block");
4012
4013         err = 0;
4014         if (ext4_should_journal_data(inode)) {
4015                 err = ext4_handle_dirty_metadata(handle, inode, bh);
4016         } else {
4017                 if (ext4_should_order_data(inode))
4018                         err = ext4_jbd2_file_inode(handle, inode);
4019                 mark_buffer_dirty(bh);
4020         }
4021
4022 unlock:
4023         unlock_page(page);
4024         page_cache_release(page);
4025         return err;
4026 }
4027
4028 /*
4029  * Probably it should be a library function... search for first non-zero word
4030  * or memcmp with zero_page, whatever is better for particular architecture.
4031  * Linus?
4032  */
4033 static inline int all_zeroes(__le32 *p, __le32 *q)
4034 {
4035         while (p < q)
4036                 if (*p++)
4037                         return 0;
4038         return 1;
4039 }
4040
4041 /**
4042  *      ext4_find_shared - find the indirect blocks for partial truncation.
4043  *      @inode:   inode in question
4044  *      @depth:   depth of the affected branch
4045  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
4046  *      @chain:   place to store the pointers to partial indirect blocks
4047  *      @top:     place to the (detached) top of branch
4048  *
4049  *      This is a helper function used by ext4_truncate().
4050  *
4051  *      When we do truncate() we may have to clean the ends of several
4052  *      indirect blocks but leave the blocks themselves alive. Block is
4053  *      partially truncated if some data below the new i_size is refered
4054  *      from it (and it is on the path to the first completely truncated
4055  *      data block, indeed).  We have to free the top of that path along
4056  *      with everything to the right of the path. Since no allocation
4057  *      past the truncation point is possible until ext4_truncate()
4058  *      finishes, we may safely do the latter, but top of branch may
4059  *      require special attention - pageout below the truncation point
4060  *      might try to populate it.
4061  *
4062  *      We atomically detach the top of branch from the tree, store the
4063  *      block number of its root in *@top, pointers to buffer_heads of
4064  *      partially truncated blocks - in @chain[].bh and pointers to
4065  *      their last elements that should not be removed - in
4066  *      @chain[].p. Return value is the pointer to last filled element
4067  *      of @chain.
4068  *
4069  *      The work left to caller to do the actual freeing of subtrees:
4070  *              a) free the subtree starting from *@top
4071  *              b) free the subtrees whose roots are stored in
4072  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
4073  *              c) free the subtrees growing from the inode past the @chain[0].
4074  *                      (no partially truncated stuff there).  */
4075
4076 static Indirect *ext4_find_shared(struct inode *inode, int depth,
4077                                   ext4_lblk_t offsets[4], Indirect chain[4],
4078                                   __le32 *top)
4079 {
4080         Indirect *partial, *p;
4081         int k, err;
4082
4083         *top = 0;
4084         /* Make k index the deepest non-null offset + 1 */
4085         for (k = depth; k > 1 && !offsets[k-1]; k--)
4086                 ;
4087         partial = ext4_get_branch(inode, k, offsets, chain, &err);
4088         /* Writer: pointers */
4089         if (!partial)
4090                 partial = chain + k-1;
4091         /*
4092          * If the branch acquired continuation since we've looked at it -
4093          * fine, it should all survive and (new) top doesn't belong to us.
4094          */
4095         if (!partial->key && *partial->p)
4096                 /* Writer: end */
4097                 goto no_top;
4098         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
4099                 ;
4100         /*
4101          * OK, we've found the last block that must survive. The rest of our
4102          * branch should be detached before unlocking. However, if that rest
4103          * of branch is all ours and does not grow immediately from the inode
4104          * it's easier to cheat and just decrement partial->p.
4105          */
4106         if (p == chain + k - 1 && p > chain) {
4107                 p->p--;
4108         } else {
4109                 *top = *p->p;
4110                 /* Nope, don't do this in ext4.  Must leave the tree intact */
4111 #if 0
4112                 *p->p = 0;
4113 #endif
4114         }
4115         /* Writer: end */
4116
4117         while (partial > p) {
4118                 brelse(partial->bh);
4119                 partial--;
4120         }
4121 no_top:
4122         return partial;
4123 }
4124
4125 /*
4126  * Zero a number of block pointers in either an inode or an indirect block.
4127  * If we restart the transaction we must again get write access to the
4128  * indirect block for further modification.
4129  *
4130  * We release `count' blocks on disk, but (last - first) may be greater
4131  * than `count' because there can be holes in there.
4132  */
4133 static void ext4_clear_blocks(handle_t *handle, struct inode *inode,
4134                               struct buffer_head *bh,
4135                               ext4_fsblk_t block_to_free,
4136                               unsigned long count, __le32 *first,
4137                               __le32 *last)
4138 {
4139         __le32 *p;
4140         int     flags = EXT4_FREE_BLOCKS_FORGET;
4141
4142         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
4143                 flags |= EXT4_FREE_BLOCKS_METADATA;
4144
4145         if (try_to_extend_transaction(handle, inode)) {
4146                 if (bh) {
4147                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4148                         ext4_handle_dirty_metadata(handle, inode, bh);
4149                 }
4150                 ext4_mark_inode_dirty(handle, inode);
4151                 ext4_truncate_restart_trans(handle, inode,
4152                                             blocks_for_truncate(inode));
4153                 if (bh) {
4154                         BUFFER_TRACE(bh, "retaking write access");
4155                         ext4_journal_get_write_access(handle, bh);
4156                 }
4157         }
4158
4159         for (p = first; p < last; p++)
4160                 *p = 0;
4161
4162         ext4_free_blocks(handle, inode, 0, block_to_free, count, flags);
4163 }
4164
4165 /**
4166  * ext4_free_data - free a list of data blocks
4167  * @handle:     handle for this transaction
4168  * @inode:      inode we are dealing with
4169  * @this_bh:    indirect buffer_head which contains *@first and *@last
4170  * @first:      array of block numbers
4171  * @last:       points immediately past the end of array
4172  *
4173  * We are freeing all blocks refered from that array (numbers are stored as
4174  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
4175  *
4176  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
4177  * blocks are contiguous then releasing them at one time will only affect one
4178  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
4179  * actually use a lot of journal space.
4180  *
4181  * @this_bh will be %NULL if @first and @last point into the inode's direct
4182  * block pointers.
4183  */
4184 static void ext4_free_data(handle_t *handle, struct inode *inode,
4185                            struct buffer_head *this_bh,
4186                            __le32 *first, __le32 *last)
4187 {
4188         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
4189         unsigned long count = 0;            /* Number of blocks in the run */
4190         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
4191                                                corresponding to
4192                                                block_to_free */
4193         ext4_fsblk_t nr;                    /* Current block # */
4194         __le32 *p;                          /* Pointer into inode/ind
4195                                                for current block */
4196         int err;
4197
4198         if (this_bh) {                          /* For indirect block */
4199                 BUFFER_TRACE(this_bh, "get_write_access");
4200                 err = ext4_journal_get_write_access(handle, this_bh);
4201                 /* Important: if we can't update the indirect pointers
4202                  * to the blocks, we can't free them. */
4203                 if (err)
4204                         return;
4205         }
4206
4207         for (p = first; p < last; p++) {
4208                 nr = le32_to_cpu(*p);
4209                 if (nr) {
4210                         /* accumulate blocks to free if they're contiguous */
4211                         if (count == 0) {
4212                                 block_to_free = nr;
4213                                 block_to_free_p = p;
4214                                 count = 1;
4215                         } else if (nr == block_to_free + count) {
4216                                 count++;
4217                         } else {
4218                                 ext4_clear_blocks(handle, inode, this_bh,
4219                                                   block_to_free,
4220                                                   count, block_to_free_p, p);
4221                                 block_to_free = nr;
4222                                 block_to_free_p = p;
4223                                 count = 1;
4224                         }
4225                 }
4226         }
4227
4228         if (count > 0)
4229                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
4230                                   count, block_to_free_p, p);
4231
4232         if (this_bh) {
4233                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
4234
4235                 /*
4236                  * The buffer head should have an attached journal head at this
4237                  * point. However, if the data is corrupted and an indirect
4238                  * block pointed to itself, it would have been detached when
4239                  * the block was cleared. Check for this instead of OOPSing.
4240                  */
4241                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
4242                         ext4_handle_dirty_metadata(handle, inode, this_bh);
4243                 else
4244                         ext4_error(inode->i_sb, __func__,
4245                                    "circular indirect block detected, "
4246                                    "inode=%lu, block=%llu",
4247                                    inode->i_ino,
4248                                    (unsigned long long) this_bh->b_blocknr);
4249         }
4250 }
4251
4252 /**
4253  *      ext4_free_branches - free an array of branches
4254  *      @handle: JBD handle for this transaction
4255  *      @inode: inode we are dealing with
4256  *      @parent_bh: the buffer_head which contains *@first and *@last
4257  *      @first: array of block numbers
4258  *      @last:  pointer immediately past the end of array
4259  *      @depth: depth of the branches to free
4260  *
4261  *      We are freeing all blocks refered from these branches (numbers are
4262  *      stored as little-endian 32-bit) and updating @inode->i_blocks
4263  *      appropriately.
4264  */
4265 static void ext4_free_branches(handle_t *handle, struct inode *inode,
4266                                struct buffer_head *parent_bh,
4267                                __le32 *first, __le32 *last, int depth)
4268 {
4269         ext4_fsblk_t nr;
4270         __le32 *p;
4271
4272         if (ext4_handle_is_aborted(handle))
4273                 return;
4274
4275         if (depth--) {
4276                 struct buffer_head *bh;
4277                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4278                 p = last;
4279                 while (--p >= first) {
4280                         nr = le32_to_cpu(*p);
4281                         if (!nr)
4282                                 continue;               /* A hole */
4283
4284                         /* Go read the buffer for the next level down */
4285                         bh = sb_bread(inode->i_sb, nr);
4286
4287                         /*
4288                          * A read failure? Report error and clear slot
4289                          * (should be rare).
4290                          */
4291                         if (!bh) {
4292                                 ext4_error(inode->i_sb, "ext4_free_branches",
4293                                            "Read failure, inode=%lu, block=%llu",
4294                                            inode->i_ino, nr);
4295                                 continue;
4296                         }
4297
4298                         /* This zaps the entire block.  Bottom up. */
4299                         BUFFER_TRACE(bh, "free child branches");
4300                         ext4_free_branches(handle, inode, bh,
4301                                         (__le32 *) bh->b_data,
4302                                         (__le32 *) bh->b_data + addr_per_block,
4303                                         depth);
4304
4305                         /*
4306                          * We've probably journalled the indirect block several
4307                          * times during the truncate.  But it's no longer
4308                          * needed and we now drop it from the transaction via
4309                          * jbd2_journal_revoke().
4310                          *
4311                          * That's easy if it's exclusively part of this
4312                          * transaction.  But if it's part of the committing
4313                          * transaction then jbd2_journal_forget() will simply
4314                          * brelse() it.  That means that if the underlying
4315                          * block is reallocated in ext4_get_block(),
4316                          * unmap_underlying_metadata() will find this block
4317                          * and will try to get rid of it.  damn, damn.
4318                          *
4319                          * If this block has already been committed to the
4320                          * journal, a revoke record will be written.  And
4321                          * revoke records must be emitted *before* clearing
4322                          * this block's bit in the bitmaps.
4323                          */
4324                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
4325
4326                         /*
4327                          * Everything below this this pointer has been
4328                          * released.  Now let this top-of-subtree go.
4329                          *
4330                          * We want the freeing of this indirect block to be
4331                          * atomic in the journal with the updating of the
4332                          * bitmap block which owns it.  So make some room in
4333                          * the journal.
4334                          *
4335                          * We zero the parent pointer *after* freeing its
4336                          * pointee in the bitmaps, so if extend_transaction()
4337                          * for some reason fails to put the bitmap changes and
4338                          * the release into the same transaction, recovery
4339                          * will merely complain about releasing a free block,
4340                          * rather than leaking blocks.
4341                          */
4342                         if (ext4_handle_is_aborted(handle))
4343                                 return;
4344                         if (try_to_extend_transaction(handle, inode)) {
4345                                 ext4_mark_inode_dirty(handle, inode);
4346                                 ext4_truncate_restart_trans(handle, inode,
4347                                             blocks_for_truncate(inode));
4348                         }
4349
4350                         ext4_free_blocks(handle, inode, 0, nr, 1,
4351                                          EXT4_FREE_BLOCKS_METADATA);
4352
4353                         if (parent_bh) {
4354                                 /*
4355                                  * The block which we have just freed is
4356                                  * pointed to by an indirect block: journal it
4357                                  */
4358                                 BUFFER_TRACE(parent_bh, "get_write_access");
4359                                 if (!ext4_journal_get_write_access(handle,
4360                                                                    parent_bh)){
4361                                         *p = 0;
4362                                         BUFFER_TRACE(parent_bh,
4363                                         "call ext4_handle_dirty_metadata");
4364                                         ext4_handle_dirty_metadata(handle,
4365                                                                    inode,
4366                                                                    parent_bh);
4367                                 }
4368                         }
4369                 }
4370         } else {
4371                 /* We have reached the bottom of the tree. */
4372                 BUFFER_TRACE(parent_bh, "free data blocks");
4373                 ext4_free_data(handle, inode, parent_bh, first, last);
4374         }
4375 }
4376
4377 int ext4_can_truncate(struct inode *inode)
4378 {
4379         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4380                 return 0;
4381         if (S_ISREG(inode->i_mode))
4382                 return 1;
4383         if (S_ISDIR(inode->i_mode))
4384                 return 1;
4385         if (S_ISLNK(inode->i_mode))
4386                 return !ext4_inode_is_fast_symlink(inode);
4387         return 0;
4388 }
4389
4390 /*
4391  * ext4_truncate()
4392  *
4393  * We block out ext4_get_block() block instantiations across the entire
4394  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4395  * simultaneously on behalf of the same inode.
4396  *
4397  * As we work through the truncate and commmit bits of it to the journal there
4398  * is one core, guiding principle: the file's tree must always be consistent on
4399  * disk.  We must be able to restart the truncate after a crash.
4400  *
4401  * The file's tree may be transiently inconsistent in memory (although it
4402  * probably isn't), but whenever we close off and commit a journal transaction,
4403  * the contents of (the filesystem + the journal) must be consistent and
4404  * restartable.  It's pretty simple, really: bottom up, right to left (although
4405  * left-to-right works OK too).
4406  *
4407  * Note that at recovery time, journal replay occurs *before* the restart of
4408  * truncate against the orphan inode list.
4409  *
4410  * The committed inode has the new, desired i_size (which is the same as
4411  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4412  * that this inode's truncate did not complete and it will again call
4413  * ext4_truncate() to have another go.  So there will be instantiated blocks
4414  * to the right of the truncation point in a crashed ext4 filesystem.  But
4415  * that's fine - as long as they are linked from the inode, the post-crash
4416  * ext4_truncate() run will find them and release them.
4417  */
4418 void ext4_truncate(struct inode *inode)
4419 {
4420         handle_t *handle;
4421         struct ext4_inode_info *ei = EXT4_I(inode);
4422         __le32 *i_data = ei->i_data;
4423         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4424         struct address_space *mapping = inode->i_mapping;
4425         ext4_lblk_t offsets[4];
4426         Indirect chain[4];
4427         Indirect *partial;
4428         __le32 nr = 0;
4429         int n;
4430         ext4_lblk_t last_block;
4431         unsigned blocksize = inode->i_sb->s_blocksize;
4432
4433         if (!ext4_can_truncate(inode))
4434                 return;
4435
4436         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4437                 ei->i_state |= EXT4_STATE_DA_ALLOC_CLOSE;
4438
4439         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
4440                 ext4_ext_truncate(inode);
4441                 return;
4442         }
4443
4444         handle = start_transaction(inode);
4445         if (IS_ERR(handle))
4446                 return;         /* AKPM: return what? */
4447
4448         last_block = (inode->i_size + blocksize-1)
4449                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4450
4451         if (inode->i_size & (blocksize - 1))
4452                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4453                         goto out_stop;
4454
4455         n = ext4_block_to_path(inode, last_block, offsets, NULL);
4456         if (n == 0)
4457                 goto out_stop;  /* error */
4458
4459         /*
4460          * OK.  This truncate is going to happen.  We add the inode to the
4461          * orphan list, so that if this truncate spans multiple transactions,
4462          * and we crash, we will resume the truncate when the filesystem
4463          * recovers.  It also marks the inode dirty, to catch the new size.
4464          *
4465          * Implication: the file must always be in a sane, consistent
4466          * truncatable state while each transaction commits.
4467          */
4468         if (ext4_orphan_add(handle, inode))
4469                 goto out_stop;
4470
4471         /*
4472          * From here we block out all ext4_get_block() callers who want to
4473          * modify the block allocation tree.
4474          */
4475         down_write(&ei->i_data_sem);
4476
4477         ext4_discard_preallocations(inode);
4478
4479         /*
4480          * The orphan list entry will now protect us from any crash which
4481          * occurs before the truncate completes, so it is now safe to propagate
4482          * the new, shorter inode size (held for now in i_size) into the
4483          * on-disk inode. We do this via i_disksize, which is the value which
4484          * ext4 *really* writes onto the disk inode.
4485          */
4486         ei->i_disksize = inode->i_size;
4487
4488         if (n == 1) {           /* direct blocks */
4489                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4490                                i_data + EXT4_NDIR_BLOCKS);
4491                 goto do_indirects;
4492         }
4493
4494         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4495         /* Kill the top of shared branch (not detached) */
4496         if (nr) {
4497                 if (partial == chain) {
4498                         /* Shared branch grows from the inode */
4499                         ext4_free_branches(handle, inode, NULL,
4500                                            &nr, &nr+1, (chain+n-1) - partial);
4501                         *partial->p = 0;
4502                         /*
4503                          * We mark the inode dirty prior to restart,
4504                          * and prior to stop.  No need for it here.
4505                          */
4506                 } else {
4507                         /* Shared branch grows from an indirect block */
4508                         BUFFER_TRACE(partial->bh, "get_write_access");
4509                         ext4_free_branches(handle, inode, partial->bh,
4510                                         partial->p,
4511                                         partial->p+1, (chain+n-1) - partial);
4512                 }
4513         }
4514         /* Clear the ends of indirect blocks on the shared branch */
4515         while (partial > chain) {
4516                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4517                                    (__le32*)partial->bh->b_data+addr_per_block,
4518                                    (chain+n-1) - partial);
4519                 BUFFER_TRACE(partial->bh, "call brelse");
4520                 brelse(partial->bh);
4521                 partial--;
4522         }
4523 do_indirects:
4524         /* Kill the remaining (whole) subtrees */
4525         switch (offsets[0]) {
4526         default:
4527                 nr = i_data[EXT4_IND_BLOCK];
4528                 if (nr) {
4529                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4530                         i_data[EXT4_IND_BLOCK] = 0;
4531                 }
4532         case EXT4_IND_BLOCK:
4533                 nr = i_data[EXT4_DIND_BLOCK];
4534                 if (nr) {
4535                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4536                         i_data[EXT4_DIND_BLOCK] = 0;
4537                 }
4538         case EXT4_DIND_BLOCK:
4539                 nr = i_data[EXT4_TIND_BLOCK];
4540                 if (nr) {
4541                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4542                         i_data[EXT4_TIND_BLOCK] = 0;
4543                 }
4544         case EXT4_TIND_BLOCK:
4545                 ;
4546         }
4547
4548         up_write(&ei->i_data_sem);
4549         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4550         ext4_mark_inode_dirty(handle, inode);
4551
4552         /*
4553          * In a multi-transaction truncate, we only make the final transaction
4554          * synchronous
4555          */
4556         if (IS_SYNC(inode))
4557                 ext4_handle_sync(handle);
4558 out_stop:
4559         /*
4560          * If this was a simple ftruncate(), and the file will remain alive
4561          * then we need to clear up the orphan record which we created above.
4562          * However, if this was a real unlink then we were called by
4563          * ext4_delete_inode(), and we allow that function to clean up the
4564          * orphan info for us.
4565          */
4566         if (inode->i_nlink)
4567                 ext4_orphan_del(handle, inode);
4568
4569         ext4_journal_stop(handle);
4570 }
4571
4572 /*
4573  * ext4_get_inode_loc returns with an extra refcount against the inode's
4574  * underlying buffer_head on success. If 'in_mem' is true, we have all
4575  * data in memory that is needed to recreate the on-disk version of this
4576  * inode.
4577  */
4578 static int __ext4_get_inode_loc(struct inode *inode,
4579                                 struct ext4_iloc *iloc, int in_mem)
4580 {
4581         struct ext4_group_desc  *gdp;
4582         struct buffer_head      *bh;
4583         struct super_block      *sb = inode->i_sb;
4584         ext4_fsblk_t            block;
4585         int                     inodes_per_block, inode_offset;
4586
4587         iloc->bh = NULL;
4588         if (!ext4_valid_inum(sb, inode->i_ino))
4589                 return -EIO;
4590
4591         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4592         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4593         if (!gdp)
4594                 return -EIO;
4595
4596         /*
4597          * Figure out the offset within the block group inode table
4598          */
4599         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4600         inode_offset = ((inode->i_ino - 1) %
4601                         EXT4_INODES_PER_GROUP(sb));
4602         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4603         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4604
4605         bh = sb_getblk(sb, block);
4606         if (!bh) {
4607                 ext4_error(sb, "ext4_get_inode_loc", "unable to read "
4608                            "inode block - inode=%lu, block=%llu",
4609                            inode->i_ino, block);
4610                 return -EIO;
4611         }
4612         if (!buffer_uptodate(bh)) {
4613                 lock_buffer(bh);
4614
4615                 /*
4616                  * If the buffer has the write error flag, we have failed
4617                  * to write out another inode in the same block.  In this
4618                  * case, we don't have to read the block because we may
4619                  * read the old inode data successfully.
4620                  */
4621                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4622                         set_buffer_uptodate(bh);
4623
4624                 if (buffer_uptodate(bh)) {
4625                         /* someone brought it uptodate while we waited */
4626                         unlock_buffer(bh);
4627                         goto has_buffer;
4628                 }
4629
4630                 /*
4631                  * If we have all information of the inode in memory and this
4632                  * is the only valid inode in the block, we need not read the
4633                  * block.
4634                  */
4635                 if (in_mem) {
4636                         struct buffer_head *bitmap_bh;
4637                         int i, start;
4638
4639                         start = inode_offset & ~(inodes_per_block - 1);
4640
4641                         /* Is the inode bitmap in cache? */
4642                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4643                         if (!bitmap_bh)
4644                                 goto make_io;
4645
4646                         /*
4647                          * If the inode bitmap isn't in cache then the
4648                          * optimisation may end up performing two reads instead
4649                          * of one, so skip it.
4650                          */
4651                         if (!buffer_uptodate(bitmap_bh)) {
4652                                 brelse(bitmap_bh);
4653                                 goto make_io;
4654                         }
4655                         for (i = start; i < start + inodes_per_block; i++) {
4656                                 if (i == inode_offset)
4657                                         continue;
4658                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4659                                         break;
4660                         }
4661                         brelse(bitmap_bh);
4662                         if (i == start + inodes_per_block) {
4663                                 /* all other inodes are free, so skip I/O */
4664                                 memset(bh->b_data, 0, bh->b_size);
4665                                 set_buffer_uptodate(bh);
4666                                 unlock_buffer(bh);
4667                                 goto has_buffer;
4668                         }
4669                 }
4670
4671 make_io:
4672                 /*
4673                  * If we need to do any I/O, try to pre-readahead extra
4674                  * blocks from the inode table.
4675                  */
4676                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4677                         ext4_fsblk_t b, end, table;
4678                         unsigned num;
4679
4680                         table = ext4_inode_table(sb, gdp);
4681                         /* s_inode_readahead_blks is always a power of 2 */
4682                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4683                         if (table > b)
4684                                 b = table;
4685                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4686                         num = EXT4_INODES_PER_GROUP(sb);
4687                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4688                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4689                                 num -= ext4_itable_unused_count(sb, gdp);
4690                         table += num / inodes_per_block;
4691                         if (end > table)
4692                                 end = table;
4693                         while (b <= end)
4694                                 sb_breadahead(sb, b++);
4695                 }
4696
4697                 /*
4698                  * There are other valid inodes in the buffer, this inode
4699                  * has in-inode xattrs, or we don't have this inode in memory.
4700                  * Read the block from disk.
4701                  */
4702                 get_bh(bh);
4703                 bh->b_end_io = end_buffer_read_sync;
4704                 submit_bh(READ_META, bh);
4705                 wait_on_buffer(bh);
4706                 if (!buffer_uptodate(bh)) {
4707                         ext4_error(sb, __func__,
4708                                    "unable to read inode block - inode=%lu, "
4709                                    "block=%llu", inode->i_ino, block);
4710                         brelse(bh);
4711                         return -EIO;
4712                 }
4713         }
4714 has_buffer:
4715         iloc->bh = bh;
4716         return 0;
4717 }
4718
4719 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4720 {
4721         /* We have all inode data except xattrs in memory here. */
4722         return __ext4_get_inode_loc(inode, iloc,
4723                 !(EXT4_I(inode)->i_state & EXT4_STATE_XATTR));
4724 }
4725
4726 void ext4_set_inode_flags(struct inode *inode)
4727 {
4728         unsigned int flags = EXT4_I(inode)->i_flags;
4729
4730         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4731         if (flags & EXT4_SYNC_FL)
4732                 inode->i_flags |= S_SYNC;
4733         if (flags & EXT4_APPEND_FL)
4734                 inode->i_flags |= S_APPEND;
4735         if (flags & EXT4_IMMUTABLE_FL)
4736                 inode->i_flags |= S_IMMUTABLE;
4737         if (flags & EXT4_NOATIME_FL)
4738                 inode->i_flags |= S_NOATIME;
4739         if (flags & EXT4_DIRSYNC_FL)
4740                 inode->i_flags |= S_DIRSYNC;
4741 }
4742
4743 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4744 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4745 {
4746         unsigned int flags = ei->vfs_inode.i_flags;
4747
4748         ei->i_flags &= ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4749                         EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|EXT4_DIRSYNC_FL);
4750         if (flags & S_SYNC)
4751                 ei->i_flags |= EXT4_SYNC_FL;
4752         if (flags & S_APPEND)
4753                 ei->i_flags |= EXT4_APPEND_FL;
4754         if (flags & S_IMMUTABLE)
4755                 ei->i_flags |= EXT4_IMMUTABLE_FL;
4756         if (flags & S_NOATIME)
4757                 ei->i_flags |= EXT4_NOATIME_FL;
4758         if (flags & S_DIRSYNC)
4759                 ei->i_flags |= EXT4_DIRSYNC_FL;
4760 }
4761
4762 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4763                                   struct ext4_inode_info *ei)
4764 {
4765         blkcnt_t i_blocks ;
4766         struct inode *inode = &(ei->vfs_inode);
4767         struct super_block *sb = inode->i_sb;
4768
4769         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4770                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4771                 /* we are using combined 48 bit field */
4772                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4773                                         le32_to_cpu(raw_inode->i_blocks_lo);
4774                 if (ei->i_flags & EXT4_HUGE_FILE_FL) {
4775                         /* i_blocks represent file system block size */
4776                         return i_blocks  << (inode->i_blkbits - 9);
4777                 } else {
4778                         return i_blocks;
4779                 }
4780         } else {
4781                 return le32_to_cpu(raw_inode->i_blocks_lo);
4782         }
4783 }
4784
4785 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4786 {
4787         struct ext4_iloc iloc;
4788         struct ext4_inode *raw_inode;
4789         struct ext4_inode_info *ei;
4790         struct inode *inode;
4791         journal_t *journal = EXT4_SB(sb)->s_journal;
4792         long ret;
4793         int block;
4794
4795         inode = iget_locked(sb, ino);
4796         if (!inode)
4797                 return ERR_PTR(-ENOMEM);
4798         if (!(inode->i_state & I_NEW))
4799                 return inode;
4800
4801         ei = EXT4_I(inode);
4802         iloc.bh = 0;
4803
4804         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4805         if (ret < 0)
4806                 goto bad_inode;
4807         raw_inode = ext4_raw_inode(&iloc);
4808         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4809         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4810         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4811         if (!(test_opt(inode->i_sb, NO_UID32))) {
4812                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4813                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4814         }
4815         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
4816
4817         ei->i_state = 0;
4818         ei->i_dir_start_lookup = 0;
4819         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4820         /* We now have enough fields to check if the inode was active or not.
4821          * This is needed because nfsd might try to access dead inodes
4822          * the test is that same one that e2fsck uses
4823          * NeilBrown 1999oct15
4824          */
4825         if (inode->i_nlink == 0) {
4826                 if (inode->i_mode == 0 ||
4827                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
4828                         /* this inode is deleted */
4829                         ret = -ESTALE;
4830                         goto bad_inode;
4831                 }
4832                 /* The only unlinked inodes we let through here have
4833                  * valid i_mode and are being read by the orphan
4834                  * recovery code: that's fine, we're about to complete
4835                  * the process of deleting those. */
4836         }
4837         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4838         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4839         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4840         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
4841                 ei->i_file_acl |=
4842                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4843         inode->i_size = ext4_isize(raw_inode);
4844         ei->i_disksize = inode->i_size;
4845 #ifdef CONFIG_QUOTA
4846         ei->i_reserved_quota = 0;
4847 #endif
4848         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4849         ei->i_block_group = iloc.block_group;
4850         ei->i_last_alloc_group = ~0;
4851         /*
4852          * NOTE! The in-memory inode i_data array is in little-endian order
4853          * even on big-endian machines: we do NOT byteswap the block numbers!
4854          */
4855         for (block = 0; block < EXT4_N_BLOCKS; block++)
4856                 ei->i_data[block] = raw_inode->i_block[block];
4857         INIT_LIST_HEAD(&ei->i_orphan);
4858
4859         /*
4860          * Set transaction id's of transactions that have to be committed
4861          * to finish f[data]sync. We set them to currently running transaction
4862          * as we cannot be sure that the inode or some of its metadata isn't
4863          * part of the transaction - the inode could have been reclaimed and
4864          * now it is reread from disk.
4865          */
4866         if (journal) {
4867                 transaction_t *transaction;
4868                 tid_t tid;
4869
4870                 spin_lock(&journal->j_state_lock);
4871                 if (journal->j_running_transaction)
4872                         transaction = journal->j_running_transaction;
4873                 else
4874                         transaction = journal->j_committing_transaction;
4875                 if (transaction)
4876                         tid = transaction->t_tid;
4877                 else
4878                         tid = journal->j_commit_sequence;
4879                 spin_unlock(&journal->j_state_lock);
4880                 ei->i_sync_tid = tid;
4881                 ei->i_datasync_tid = tid;
4882         }
4883
4884         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4885                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4886                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4887                     EXT4_INODE_SIZE(inode->i_sb)) {
4888                         ret = -EIO;
4889                         goto bad_inode;
4890                 }
4891                 if (ei->i_extra_isize == 0) {
4892                         /* The extra space is currently unused. Use it. */
4893                         ei->i_extra_isize = sizeof(struct ext4_inode) -
4894                                             EXT4_GOOD_OLD_INODE_SIZE;
4895                 } else {
4896                         __le32 *magic = (void *)raw_inode +
4897                                         EXT4_GOOD_OLD_INODE_SIZE +
4898                                         ei->i_extra_isize;
4899                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
4900                                 ei->i_state |= EXT4_STATE_XATTR;
4901                 }
4902         } else
4903                 ei->i_extra_isize = 0;
4904
4905         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4906         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4907         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4908         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4909
4910         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4911         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4912                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4913                         inode->i_version |=
4914                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4915         }
4916
4917         ret = 0;
4918         if (ei->i_file_acl &&
4919             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
4920                 ext4_error(sb, __func__,
4921                            "bad extended attribute block %llu in inode #%lu",
4922                            ei->i_file_acl, inode->i_ino);
4923                 ret = -EIO;
4924                 goto bad_inode;
4925         } else if (ei->i_flags & EXT4_EXTENTS_FL) {
4926                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4927                     (S_ISLNK(inode->i_mode) &&
4928                      !ext4_inode_is_fast_symlink(inode)))
4929                         /* Validate extent which is part of inode */
4930                         ret = ext4_ext_check_inode(inode);
4931         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4932                    (S_ISLNK(inode->i_mode) &&
4933                     !ext4_inode_is_fast_symlink(inode))) {
4934                 /* Validate block references which are part of inode */
4935                 ret = ext4_check_inode_blockref(inode);
4936         }
4937         if (ret)
4938                 goto bad_inode;
4939
4940         if (S_ISREG(inode->i_mode)) {
4941                 inode->i_op = &ext4_file_inode_operations;
4942                 inode->i_fop = &ext4_file_operations;
4943                 ext4_set_aops(inode);
4944         } else if (S_ISDIR(inode->i_mode)) {
4945                 inode->i_op = &ext4_dir_inode_operations;
4946                 inode->i_fop = &ext4_dir_operations;
4947         } else if (S_ISLNK(inode->i_mode)) {
4948                 if (ext4_inode_is_fast_symlink(inode)) {
4949                         inode->i_op = &ext4_fast_symlink_inode_operations;
4950                         nd_terminate_link(ei->i_data, inode->i_size,
4951                                 sizeof(ei->i_data) - 1);
4952                 } else {
4953                         inode->i_op = &ext4_symlink_inode_operations;
4954                         ext4_set_aops(inode);
4955                 }
4956         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
4957               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
4958                 inode->i_op = &ext4_special_inode_operations;
4959                 if (raw_inode->i_block[0])
4960                         init_special_inode(inode, inode->i_mode,
4961                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4962                 else
4963                         init_special_inode(inode, inode->i_mode,
4964                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4965         } else {
4966                 ret = -EIO;
4967                 ext4_error(inode->i_sb, __func__,
4968                            "bogus i_mode (%o) for inode=%lu",
4969                            inode->i_mode, inode->i_ino);
4970                 goto bad_inode;
4971         }
4972         brelse(iloc.bh);
4973         ext4_set_inode_flags(inode);
4974         unlock_new_inode(inode);
4975         return inode;
4976
4977 bad_inode:
4978         brelse(iloc.bh);
4979         iget_failed(inode);
4980         return ERR_PTR(ret);
4981 }
4982
4983 static int ext4_inode_blocks_set(handle_t *handle,
4984                                 struct ext4_inode *raw_inode,
4985                                 struct ext4_inode_info *ei)
4986 {
4987         struct inode *inode = &(ei->vfs_inode);
4988         u64 i_blocks = inode->i_blocks;
4989         struct super_block *sb = inode->i_sb;
4990
4991         if (i_blocks <= ~0U) {
4992                 /*
4993                  * i_blocks can be represnted in a 32 bit variable
4994                  * as multiple of 512 bytes
4995                  */
4996                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4997                 raw_inode->i_blocks_high = 0;
4998                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4999                 return 0;
5000         }
5001         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
5002                 return -EFBIG;
5003
5004         if (i_blocks <= 0xffffffffffffULL) {
5005                 /*
5006                  * i_blocks can be represented in a 48 bit variable
5007                  * as multiple of 512 bytes
5008                  */
5009                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5010                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5011                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
5012         } else {
5013                 ei->i_flags |= EXT4_HUGE_FILE_FL;
5014                 /* i_block is stored in file system block size */
5015                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
5016                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5017                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5018         }
5019         return 0;
5020 }
5021
5022 /*
5023  * Post the struct inode info into an on-disk inode location in the
5024  * buffer-cache.  This gobbles the caller's reference to the
5025  * buffer_head in the inode location struct.
5026  *
5027  * The caller must have write access to iloc->bh.
5028  */
5029 static int ext4_do_update_inode(handle_t *handle,
5030                                 struct inode *inode,
5031                                 struct ext4_iloc *iloc)
5032 {
5033         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5034         struct ext4_inode_info *ei = EXT4_I(inode);
5035         struct buffer_head *bh = iloc->bh;
5036         int err = 0, rc, block;
5037
5038         /* For fields not not tracking in the in-memory inode,
5039          * initialise them to zero for new inodes. */
5040         if (ei->i_state & EXT4_STATE_NEW)
5041                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5042
5043         ext4_get_inode_flags(ei);
5044         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5045         if (!(test_opt(inode->i_sb, NO_UID32))) {
5046                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
5047                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
5048 /*
5049  * Fix up interoperability with old kernels. Otherwise, old inodes get
5050  * re-used with the upper 16 bits of the uid/gid intact
5051  */
5052                 if (!ei->i_dtime) {
5053                         raw_inode->i_uid_high =
5054                                 cpu_to_le16(high_16_bits(inode->i_uid));
5055                         raw_inode->i_gid_high =
5056                                 cpu_to_le16(high_16_bits(inode->i_gid));
5057                 } else {
5058                         raw_inode->i_uid_high = 0;
5059                         raw_inode->i_gid_high = 0;
5060                 }
5061         } else {
5062                 raw_inode->i_uid_low =
5063                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
5064                 raw_inode->i_gid_low =
5065                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
5066                 raw_inode->i_uid_high = 0;
5067                 raw_inode->i_gid_high = 0;
5068         }
5069         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5070
5071         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5072         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5073         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5074         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5075
5076         if (ext4_inode_blocks_set(handle, raw_inode, ei))
5077                 goto out_brelse;
5078         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5079         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
5080         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
5081             cpu_to_le32(EXT4_OS_HURD))
5082                 raw_inode->i_file_acl_high =
5083                         cpu_to_le16(ei->i_file_acl >> 32);
5084         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5085         ext4_isize_set(raw_inode, ei->i_disksize);
5086         if (ei->i_disksize > 0x7fffffffULL) {
5087                 struct super_block *sb = inode->i_sb;
5088                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
5089                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
5090                                 EXT4_SB(sb)->s_es->s_rev_level ==
5091                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
5092                         /* If this is the first large file
5093                          * created, add a flag to the superblock.
5094                          */
5095                         err = ext4_journal_get_write_access(handle,
5096                                         EXT4_SB(sb)->s_sbh);
5097                         if (err)
5098                                 goto out_brelse;
5099                         ext4_update_dynamic_rev(sb);
5100                         EXT4_SET_RO_COMPAT_FEATURE(sb,
5101                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
5102                         sb->s_dirt = 1;
5103                         ext4_handle_sync(handle);
5104                         err = ext4_handle_dirty_metadata(handle, inode,
5105                                         EXT4_SB(sb)->s_sbh);
5106                 }
5107         }
5108         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5109         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5110                 if (old_valid_dev(inode->i_rdev)) {
5111                         raw_inode->i_block[0] =
5112                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
5113                         raw_inode->i_block[1] = 0;
5114                 } else {
5115                         raw_inode->i_block[0] = 0;
5116                         raw_inode->i_block[1] =
5117                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
5118                         raw_inode->i_block[2] = 0;
5119                 }
5120         } else
5121                 for (block = 0; block < EXT4_N_BLOCKS; block++)
5122                         raw_inode->i_block[block] = ei->i_data[block];
5123
5124         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
5125         if (ei->i_extra_isize) {
5126                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5127                         raw_inode->i_version_hi =
5128                         cpu_to_le32(inode->i_version >> 32);
5129                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
5130         }
5131
5132         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5133         rc = ext4_handle_dirty_metadata(handle, inode, bh);
5134         if (!err)
5135                 err = rc;
5136         ei->i_state &= ~EXT4_STATE_NEW;
5137
5138         ext4_update_inode_fsync_trans(handle, inode, 0);
5139 out_brelse:
5140         brelse(bh);
5141         ext4_std_error(inode->i_sb, err);
5142         return err;
5143 }
5144
5145 /*
5146  * ext4_write_inode()
5147  *
5148  * We are called from a few places:
5149  *
5150  * - Within generic_file_write() for O_SYNC files.
5151  *   Here, there will be no transaction running. We wait for any running
5152  *   trasnaction to commit.
5153  *
5154  * - Within sys_sync(), kupdate and such.
5155  *   We wait on commit, if tol to.
5156  *
5157  * - Within prune_icache() (PF_MEMALLOC == true)
5158  *   Here we simply return.  We can't afford to block kswapd on the
5159  *   journal commit.
5160  *
5161  * In all cases it is actually safe for us to return without doing anything,
5162  * because the inode has been copied into a raw inode buffer in
5163  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
5164  * knfsd.
5165  *
5166  * Note that we are absolutely dependent upon all inode dirtiers doing the
5167  * right thing: they *must* call mark_inode_dirty() after dirtying info in
5168  * which we are interested.
5169  *
5170  * It would be a bug for them to not do this.  The code:
5171  *
5172  *      mark_inode_dirty(inode)
5173  *      stuff();
5174  *      inode->i_size = expr;
5175  *
5176  * is in error because a kswapd-driven write_inode() could occur while
5177  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
5178  * will no longer be on the superblock's dirty inode list.
5179  */
5180 int ext4_write_inode(struct inode *inode, int wait)
5181 {
5182         int err;
5183
5184         if (current->flags & PF_MEMALLOC)
5185                 return 0;
5186
5187         if (EXT4_SB(inode->i_sb)->s_journal) {
5188                 if (ext4_journal_current_handle()) {
5189                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5190                         dump_stack();
5191                         return -EIO;
5192                 }
5193
5194                 if (!wait)
5195                         return 0;
5196
5197                 err = ext4_force_commit(inode->i_sb);
5198         } else {
5199                 struct ext4_iloc iloc;
5200
5201                 err = ext4_get_inode_loc(inode, &iloc);
5202                 if (err)
5203                         return err;
5204                 if (wait)
5205                         sync_dirty_buffer(iloc.bh);
5206                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5207                         ext4_error(inode->i_sb, __func__,
5208                                    "IO error syncing inode, "
5209                                    "inode=%lu, block=%llu",
5210                                    inode->i_ino,
5211                                    (unsigned long long)iloc.bh->b_blocknr);
5212                         err = -EIO;
5213                 }
5214         }
5215         return err;
5216 }
5217
5218 /*
5219  * ext4_setattr()
5220  *
5221  * Called from notify_change.
5222  *
5223  * We want to trap VFS attempts to truncate the file as soon as
5224  * possible.  In particular, we want to make sure that when the VFS
5225  * shrinks i_size, we put the inode on the orphan list and modify
5226  * i_disksize immediately, so that during the subsequent flushing of
5227  * dirty pages and freeing of disk blocks, we can guarantee that any
5228  * commit will leave the blocks being flushed in an unused state on
5229  * disk.  (On recovery, the inode will get truncated and the blocks will
5230  * be freed, so we have a strong guarantee that no future commit will
5231  * leave these blocks visible to the user.)
5232  *
5233  * Another thing we have to assure is that if we are in ordered mode
5234  * and inode is still attached to the committing transaction, we must
5235  * we start writeout of all the dirty pages which are being truncated.
5236  * This way we are sure that all the data written in the previous
5237  * transaction are already on disk (truncate waits for pages under
5238  * writeback).
5239  *
5240  * Called with inode->i_mutex down.
5241  */
5242 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5243 {
5244         struct inode *inode = dentry->d_inode;
5245         int error, rc = 0;
5246         const unsigned int ia_valid = attr->ia_valid;
5247
5248         error = inode_change_ok(inode, attr);
5249         if (error)
5250                 return error;
5251
5252         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
5253                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
5254                 handle_t *handle;
5255
5256                 /* (user+group)*(old+new) structure, inode write (sb,
5257                  * inode block, ? - but truncate inode update has it) */
5258                 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
5259                                         EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
5260                 if (IS_ERR(handle)) {
5261                         error = PTR_ERR(handle);
5262                         goto err_out;
5263                 }
5264                 error = vfs_dq_transfer(inode, attr) ? -EDQUOT : 0;
5265                 if (error) {
5266                         ext4_journal_stop(handle);
5267                         return error;
5268                 }
5269                 /* Update corresponding info in inode so that everything is in
5270                  * one transaction */
5271                 if (attr->ia_valid & ATTR_UID)
5272                         inode->i_uid = attr->ia_uid;
5273                 if (attr->ia_valid & ATTR_GID)
5274                         inode->i_gid = attr->ia_gid;
5275                 error = ext4_mark_inode_dirty(handle, inode);
5276                 ext4_journal_stop(handle);
5277         }
5278
5279         if (attr->ia_valid & ATTR_SIZE) {
5280                 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)) {
5281                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5282
5283                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
5284                                 error = -EFBIG;
5285                                 goto err_out;
5286                         }
5287                 }
5288         }
5289
5290         if (S_ISREG(inode->i_mode) &&
5291             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
5292                 handle_t *handle;
5293
5294                 handle = ext4_journal_start(inode, 3);
5295                 if (IS_ERR(handle)) {
5296                         error = PTR_ERR(handle);
5297                         goto err_out;
5298                 }
5299
5300                 error = ext4_orphan_add(handle, inode);
5301                 EXT4_I(inode)->i_disksize = attr->ia_size;
5302                 rc = ext4_mark_inode_dirty(handle, inode);
5303                 if (!error)
5304                         error = rc;
5305                 ext4_journal_stop(handle);
5306
5307                 if (ext4_should_order_data(inode)) {
5308                         error = ext4_begin_ordered_truncate(inode,
5309                                                             attr->ia_size);
5310                         if (error) {
5311                                 /* Do as much error cleanup as possible */
5312                                 handle = ext4_journal_start(inode, 3);
5313                                 if (IS_ERR(handle)) {
5314                                         ext4_orphan_del(NULL, inode);
5315                                         goto err_out;
5316                                 }
5317                                 ext4_orphan_del(handle, inode);
5318                                 ext4_journal_stop(handle);
5319                                 goto err_out;
5320                         }
5321                 }
5322         }
5323
5324         rc = inode_setattr(inode, attr);
5325
5326         /* If inode_setattr's call to ext4_truncate failed to get a
5327          * transaction handle at all, we need to clean up the in-core
5328          * orphan list manually. */
5329         if (inode->i_nlink)
5330                 ext4_orphan_del(NULL, inode);
5331
5332         if (!rc && (ia_valid & ATTR_MODE))
5333                 rc = ext4_acl_chmod(inode);
5334
5335 err_out:
5336         ext4_std_error(inode->i_sb, error);
5337         if (!error)
5338                 error = rc;
5339         return error;
5340 }
5341
5342 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
5343                  struct kstat *stat)
5344 {
5345         struct inode *inode;
5346         unsigned long delalloc_blocks;
5347
5348         inode = dentry->d_inode;
5349         generic_fillattr(inode, stat);
5350
5351         /*
5352          * We can't update i_blocks if the block allocation is delayed
5353          * otherwise in the case of system crash before the real block
5354          * allocation is done, we will have i_blocks inconsistent with
5355          * on-disk file blocks.
5356          * We always keep i_blocks updated together with real
5357          * allocation. But to not confuse with user, stat
5358          * will return the blocks that include the delayed allocation
5359          * blocks for this file.
5360          */
5361         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
5362         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
5363         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
5364
5365         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
5366         return 0;
5367 }
5368
5369 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
5370                                       int chunk)
5371 {
5372         int indirects;
5373
5374         /* if nrblocks are contiguous */
5375         if (chunk) {
5376                 /*
5377                  * With N contiguous data blocks, it need at most
5378                  * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) indirect blocks
5379                  * 2 dindirect blocks
5380                  * 1 tindirect block
5381                  */
5382                 indirects = nrblocks / EXT4_ADDR_PER_BLOCK(inode->i_sb);
5383                 return indirects + 3;
5384         }
5385         /*
5386          * if nrblocks are not contiguous, worse case, each block touch
5387          * a indirect block, and each indirect block touch a double indirect
5388          * block, plus a triple indirect block
5389          */
5390         indirects = nrblocks * 2 + 1;
5391         return indirects;
5392 }
5393
5394 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5395 {
5396         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL))
5397                 return ext4_indirect_trans_blocks(inode, nrblocks, chunk);
5398         return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
5399 }
5400
5401 /*
5402  * Account for index blocks, block groups bitmaps and block group
5403  * descriptor blocks if modify datablocks and index blocks
5404  * worse case, the indexs blocks spread over different block groups
5405  *
5406  * If datablocks are discontiguous, they are possible to spread over
5407  * different block groups too. If they are contiuguous, with flexbg,
5408  * they could still across block group boundary.
5409  *
5410  * Also account for superblock, inode, quota and xattr blocks
5411  */
5412 int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5413 {
5414         ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
5415         int gdpblocks;
5416         int idxblocks;
5417         int ret = 0;
5418
5419         /*
5420          * How many index blocks need to touch to modify nrblocks?
5421          * The "Chunk" flag indicating whether the nrblocks is
5422          * physically contiguous on disk
5423          *
5424          * For Direct IO and fallocate, they calls get_block to allocate
5425          * one single extent at a time, so they could set the "Chunk" flag
5426          */
5427         idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
5428
5429         ret = idxblocks;
5430
5431         /*
5432          * Now let's see how many group bitmaps and group descriptors need
5433          * to account
5434          */
5435         groups = idxblocks;
5436         if (chunk)
5437                 groups += 1;
5438         else
5439                 groups += nrblocks;
5440
5441         gdpblocks = groups;
5442         if (groups > ngroups)
5443                 groups = ngroups;
5444         if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
5445                 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
5446
5447         /* bitmaps and block group descriptor blocks */
5448         ret += groups + gdpblocks;
5449
5450         /* Blocks for super block, inode, quota and xattr blocks */
5451         ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
5452
5453         return ret;
5454 }
5455
5456 /*
5457  * Calulate the total number of credits to reserve to fit
5458  * the modification of a single pages into a single transaction,
5459  * which may include multiple chunks of block allocations.
5460  *
5461  * This could be called via ext4_write_begin()
5462  *
5463  * We need to consider the worse case, when
5464  * one new block per extent.
5465  */
5466 int ext4_writepage_trans_blocks(struct inode *inode)
5467 {
5468         int bpp = ext4_journal_blocks_per_page(inode);
5469         int ret;
5470
5471         ret = ext4_meta_trans_blocks(inode, bpp, 0);
5472
5473         /* Account for data blocks for journalled mode */
5474         if (ext4_should_journal_data(inode))
5475                 ret += bpp;
5476         return ret;
5477 }
5478
5479 /*
5480  * Calculate the journal credits for a chunk of data modification.
5481  *
5482  * This is called from DIO, fallocate or whoever calling
5483  * ext4_get_blocks() to map/allocate a chunk of contiguous disk blocks.
5484  *
5485  * journal buffers for data blocks are not included here, as DIO
5486  * and fallocate do no need to journal data buffers.
5487  */
5488 int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
5489 {
5490         return ext4_meta_trans_blocks(inode, nrblocks, 1);
5491 }
5492
5493 /*
5494  * The caller must have previously called ext4_reserve_inode_write().
5495  * Give this, we know that the caller already has write access to iloc->bh.
5496  */
5497 int ext4_mark_iloc_dirty(handle_t *handle,
5498                          struct inode *inode, struct ext4_iloc *iloc)
5499 {
5500         int err = 0;
5501
5502         if (test_opt(inode->i_sb, I_VERSION))
5503                 inode_inc_iversion(inode);
5504
5505         /* the do_update_inode consumes one bh->b_count */
5506         get_bh(iloc->bh);
5507
5508         /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
5509         err = ext4_do_update_inode(handle, inode, iloc);
5510         put_bh(iloc->bh);
5511         return err;
5512 }
5513
5514 /*
5515  * On success, We end up with an outstanding reference count against
5516  * iloc->bh.  This _must_ be cleaned up later.
5517  */
5518
5519 int
5520 ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
5521                          struct ext4_iloc *iloc)
5522 {
5523         int err;
5524
5525         err = ext4_get_inode_loc(inode, iloc);
5526         if (!err) {
5527                 BUFFER_TRACE(iloc->bh, "get_write_access");
5528                 err = ext4_journal_get_write_access(handle, iloc->bh);
5529                 if (err) {
5530                         brelse(iloc->bh);
5531                         iloc->bh = NULL;
5532                 }
5533         }
5534         ext4_std_error(inode->i_sb, err);
5535         return err;
5536 }
5537
5538 /*
5539  * Expand an inode by new_extra_isize bytes.
5540  * Returns 0 on success or negative error number on failure.
5541  */
5542 static int ext4_expand_extra_isize(struct inode *inode,
5543                                    unsigned int new_extra_isize,
5544                                    struct ext4_iloc iloc,
5545                                    handle_t *handle)
5546 {
5547         struct ext4_inode *raw_inode;
5548         struct ext4_xattr_ibody_header *header;
5549         struct ext4_xattr_entry *entry;
5550
5551         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
5552                 return 0;
5553
5554         raw_inode = ext4_raw_inode(&iloc);
5555
5556         header = IHDR(inode, raw_inode);
5557         entry = IFIRST(header);
5558
5559         /* No extended attributes present */
5560         if (!(EXT4_I(inode)->i_state & EXT4_STATE_XATTR) ||
5561                 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
5562                 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
5563                         new_extra_isize);
5564                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
5565                 return 0;
5566         }
5567
5568         /* try to expand with EAs present */
5569         return ext4_expand_extra_isize_ea(inode, new_extra_isize,
5570                                           raw_inode, handle);
5571 }
5572
5573 /*
5574  * What we do here is to mark the in-core inode as clean with respect to inode
5575  * dirtiness (it may still be data-dirty).
5576  * This means that the in-core inode may be reaped by prune_icache
5577  * without having to perform any I/O.  This is a very good thing,
5578  * because *any* task may call prune_icache - even ones which
5579  * have a transaction open against a different journal.
5580  *
5581  * Is this cheating?  Not really.  Sure, we haven't written the
5582  * inode out, but prune_icache isn't a user-visible syncing function.
5583  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
5584  * we start and wait on commits.
5585  *
5586  * Is this efficient/effective?  Well, we're being nice to the system
5587  * by cleaning up our inodes proactively so they can be reaped
5588  * without I/O.  But we are potentially leaving up to five seconds'
5589  * worth of inodes floating about which prune_icache wants us to
5590  * write out.  One way to fix that would be to get prune_icache()
5591  * to do a write_super() to