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