ext4: Patch up how we claim metadata blocks for quota purposes
[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  * Calculate the number of metadata blocks need to reserve
1014  * to allocate @blocks for non extent file based file
1015  */
1016 static int ext4_indirect_calc_metadata_amount(struct inode *inode, int blocks)
1017 {
1018         int icap = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1019         int ind_blks, dind_blks, tind_blks;
1020
1021         /* number of new indirect blocks needed */
1022         ind_blks = (blocks + icap - 1) / icap;
1023
1024         dind_blks = (ind_blks + icap - 1) / icap;
1025
1026         tind_blks = 1;
1027
1028         return ind_blks + dind_blks + tind_blks;
1029 }
1030
1031 /*
1032  * Calculate the number of metadata blocks need to reserve
1033  * to allocate given number of blocks
1034  */
1035 static int ext4_calc_metadata_amount(struct inode *inode, int blocks)
1036 {
1037         if (!blocks)
1038                 return 0;
1039
1040         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
1041                 return ext4_ext_calc_metadata_amount(inode, blocks);
1042
1043         return ext4_indirect_calc_metadata_amount(inode, blocks);
1044 }
1045
1046 /*
1047  * Called with i_data_sem down, which is important since we can call
1048  * ext4_discard_preallocations() from here.
1049  */
1050 static void ext4_da_update_reserve_space(struct inode *inode, int used)
1051 {
1052         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1053         struct ext4_inode_info *ei = EXT4_I(inode);
1054         int mdb_free = 0;
1055
1056         spin_lock(&ei->i_block_reservation_lock);
1057         if (unlikely(used > ei->i_reserved_data_blocks)) {
1058                 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
1059                          "with only %d reserved data blocks\n",
1060                          __func__, inode->i_ino, used,
1061                          ei->i_reserved_data_blocks);
1062                 WARN_ON(1);
1063                 used = ei->i_reserved_data_blocks;
1064         }
1065
1066         /* Update per-inode reservations */
1067         ei->i_reserved_data_blocks -= used;
1068         used += ei->i_allocated_meta_blocks;
1069         ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
1070         ei->i_allocated_meta_blocks = 0;
1071         percpu_counter_sub(&sbi->s_dirtyblocks_counter, used);
1072
1073         if (ei->i_reserved_data_blocks == 0) {
1074                 /*
1075                  * We can release all of the reserved metadata blocks
1076                  * only when we have written all of the delayed
1077                  * allocation blocks.
1078                  */
1079                 mdb_free = ei->i_allocated_meta_blocks;
1080                 percpu_counter_sub(&sbi->s_dirtyblocks_counter, mdb_free);
1081                 ei->i_allocated_meta_blocks = 0;
1082         }
1083         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1084
1085         /* Update quota subsystem */
1086         vfs_dq_claim_block(inode, used);
1087         if (mdb_free)
1088                 vfs_dq_release_reservation_block(inode, mdb_free);
1089
1090         /*
1091          * If we have done all the pending block allocations and if
1092          * there aren't any writers on the inode, we can discard the
1093          * inode's preallocations.
1094          */
1095         if ((ei->i_reserved_data_blocks == 0) &&
1096             (atomic_read(&inode->i_writecount) == 0))
1097                 ext4_discard_preallocations(inode);
1098 }
1099
1100 static int check_block_validity(struct inode *inode, const char *msg,
1101                                 sector_t logical, sector_t phys, int len)
1102 {
1103         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), phys, len)) {
1104                 ext4_error(inode->i_sb, msg,
1105                            "inode #%lu logical block %llu mapped to %llu "
1106                            "(size %d)", inode->i_ino,
1107                            (unsigned long long) logical,
1108                            (unsigned long long) phys, len);
1109                 return -EIO;
1110         }
1111         return 0;
1112 }
1113
1114 /*
1115  * Return the number of contiguous dirty pages in a given inode
1116  * starting at page frame idx.
1117  */
1118 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
1119                                     unsigned int max_pages)
1120 {
1121         struct address_space *mapping = inode->i_mapping;
1122         pgoff_t index;
1123         struct pagevec pvec;
1124         pgoff_t num = 0;
1125         int i, nr_pages, done = 0;
1126
1127         if (max_pages == 0)
1128                 return 0;
1129         pagevec_init(&pvec, 0);
1130         while (!done) {
1131                 index = idx;
1132                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
1133                                               PAGECACHE_TAG_DIRTY,
1134                                               (pgoff_t)PAGEVEC_SIZE);
1135                 if (nr_pages == 0)
1136                         break;
1137                 for (i = 0; i < nr_pages; i++) {
1138                         struct page *page = pvec.pages[i];
1139                         struct buffer_head *bh, *head;
1140
1141                         lock_page(page);
1142                         if (unlikely(page->mapping != mapping) ||
1143                             !PageDirty(page) ||
1144                             PageWriteback(page) ||
1145                             page->index != idx) {
1146                                 done = 1;
1147                                 unlock_page(page);
1148                                 break;
1149                         }
1150                         if (page_has_buffers(page)) {
1151                                 bh = head = page_buffers(page);
1152                                 do {
1153                                         if (!buffer_delay(bh) &&
1154                                             !buffer_unwritten(bh))
1155                                                 done = 1;
1156                                         bh = bh->b_this_page;
1157                                 } while (!done && (bh != head));
1158                         }
1159                         unlock_page(page);
1160                         if (done)
1161                                 break;
1162                         idx++;
1163                         num++;
1164                         if (num >= max_pages)
1165                                 break;
1166                 }
1167                 pagevec_release(&pvec);
1168         }
1169         return num;
1170 }
1171
1172 /*
1173  * The ext4_get_blocks() function tries to look up the requested blocks,
1174  * and returns if the blocks are already mapped.
1175  *
1176  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1177  * and store the allocated blocks in the result buffer head and mark it
1178  * mapped.
1179  *
1180  * If file type is extents based, it will call ext4_ext_get_blocks(),
1181  * Otherwise, call with ext4_ind_get_blocks() to handle indirect mapping
1182  * based files
1183  *
1184  * On success, it returns the number of blocks being mapped or allocate.
1185  * if create==0 and the blocks are pre-allocated and uninitialized block,
1186  * the result buffer head is unmapped. If the create ==1, it will make sure
1187  * the buffer head is mapped.
1188  *
1189  * It returns 0 if plain look up failed (blocks have not been allocated), in
1190  * that casem, buffer head is unmapped
1191  *
1192  * It returns the error in case of allocation failure.
1193  */
1194 int ext4_get_blocks(handle_t *handle, struct inode *inode, sector_t block,
1195                     unsigned int max_blocks, struct buffer_head *bh,
1196                     int flags)
1197 {
1198         int retval;
1199
1200         clear_buffer_mapped(bh);
1201         clear_buffer_unwritten(bh);
1202
1203         ext_debug("ext4_get_blocks(): inode %lu, flag %d, max_blocks %u,"
1204                   "logical block %lu\n", inode->i_ino, flags, max_blocks,
1205                   (unsigned long)block);
1206         /*
1207          * Try to see if we can get the block without requesting a new
1208          * file system block.
1209          */
1210         down_read((&EXT4_I(inode)->i_data_sem));
1211         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1212                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1213                                 bh, 0);
1214         } else {
1215                 retval = ext4_ind_get_blocks(handle, inode, block, max_blocks,
1216                                              bh, 0);
1217         }
1218         up_read((&EXT4_I(inode)->i_data_sem));
1219
1220         if (retval > 0 && buffer_mapped(bh)) {
1221                 int ret = check_block_validity(inode, "file system corruption",
1222                                                block, bh->b_blocknr, retval);
1223                 if (ret != 0)
1224                         return ret;
1225         }
1226
1227         /* If it is only a block(s) look up */
1228         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1229                 return retval;
1230
1231         /*
1232          * Returns if the blocks have already allocated
1233          *
1234          * Note that if blocks have been preallocated
1235          * ext4_ext_get_block() returns th create = 0
1236          * with buffer head unmapped.
1237          */
1238         if (retval > 0 && buffer_mapped(bh))
1239                 return retval;
1240
1241         /*
1242          * When we call get_blocks without the create flag, the
1243          * BH_Unwritten flag could have gotten set if the blocks
1244          * requested were part of a uninitialized extent.  We need to
1245          * clear this flag now that we are committed to convert all or
1246          * part of the uninitialized extent to be an initialized
1247          * extent.  This is because we need to avoid the combination
1248          * of BH_Unwritten and BH_Mapped flags being simultaneously
1249          * set on the buffer_head.
1250          */
1251         clear_buffer_unwritten(bh);
1252
1253         /*
1254          * New blocks allocate and/or writing to uninitialized extent
1255          * will possibly result in updating i_data, so we take
1256          * the write lock of i_data_sem, and call get_blocks()
1257          * with create == 1 flag.
1258          */
1259         down_write((&EXT4_I(inode)->i_data_sem));
1260
1261         /*
1262          * if the caller is from delayed allocation writeout path
1263          * we have already reserved fs blocks for allocation
1264          * let the underlying get_block() function know to
1265          * avoid double accounting
1266          */
1267         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1268                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1269         /*
1270          * We need to check for EXT4 here because migrate
1271          * could have changed the inode type in between
1272          */
1273         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1274                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1275                                               bh, flags);
1276         } else {
1277                 retval = ext4_ind_get_blocks(handle, inode, block,
1278                                              max_blocks, bh, flags);
1279
1280                 if (retval > 0 && buffer_new(bh)) {
1281                         /*
1282                          * We allocated new blocks which will result in
1283                          * i_data's format changing.  Force the migrate
1284                          * to fail by clearing migrate flags
1285                          */
1286                         EXT4_I(inode)->i_state &= ~EXT4_STATE_EXT_MIGRATE;
1287                 }
1288         }
1289
1290         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1291                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1292
1293         /*
1294          * Update reserved blocks/metadata blocks after successful
1295          * block allocation which had been deferred till now.
1296          */
1297         if ((retval > 0) && (flags & EXT4_GET_BLOCKS_UPDATE_RESERVE_SPACE))
1298                 ext4_da_update_reserve_space(inode, retval);
1299
1300         up_write((&EXT4_I(inode)->i_data_sem));
1301         if (retval > 0 && buffer_mapped(bh)) {
1302                 int ret = check_block_validity(inode, "file system "
1303                                                "corruption after allocation",
1304                                                block, bh->b_blocknr, retval);
1305                 if (ret != 0)
1306                         return ret;
1307         }
1308         return retval;
1309 }
1310
1311 /* Maximum number of blocks we map for direct IO at once. */
1312 #define DIO_MAX_BLOCKS 4096
1313
1314 int ext4_get_block(struct inode *inode, sector_t iblock,
1315                    struct buffer_head *bh_result, int create)
1316 {
1317         handle_t *handle = ext4_journal_current_handle();
1318         int ret = 0, started = 0;
1319         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1320         int dio_credits;
1321
1322         if (create && !handle) {
1323                 /* Direct IO write... */
1324                 if (max_blocks > DIO_MAX_BLOCKS)
1325                         max_blocks = DIO_MAX_BLOCKS;
1326                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
1327                 handle = ext4_journal_start(inode, dio_credits);
1328                 if (IS_ERR(handle)) {
1329                         ret = PTR_ERR(handle);
1330                         goto out;
1331                 }
1332                 started = 1;
1333         }
1334
1335         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
1336                               create ? EXT4_GET_BLOCKS_CREATE : 0);
1337         if (ret > 0) {
1338                 bh_result->b_size = (ret << inode->i_blkbits);
1339                 ret = 0;
1340         }
1341         if (started)
1342                 ext4_journal_stop(handle);
1343 out:
1344         return ret;
1345 }
1346
1347 /*
1348  * `handle' can be NULL if create is zero
1349  */
1350 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1351                                 ext4_lblk_t block, int create, int *errp)
1352 {
1353         struct buffer_head dummy;
1354         int fatal = 0, err;
1355         int flags = 0;
1356
1357         J_ASSERT(handle != NULL || create == 0);
1358
1359         dummy.b_state = 0;
1360         dummy.b_blocknr = -1000;
1361         buffer_trace_init(&dummy.b_history);
1362         if (create)
1363                 flags |= EXT4_GET_BLOCKS_CREATE;
1364         err = ext4_get_blocks(handle, inode, block, 1, &dummy, flags);
1365         /*
1366          * ext4_get_blocks() returns number of blocks mapped. 0 in
1367          * case of a HOLE.
1368          */
1369         if (err > 0) {
1370                 if (err > 1)
1371                         WARN_ON(1);
1372                 err = 0;
1373         }
1374         *errp = err;
1375         if (!err && buffer_mapped(&dummy)) {
1376                 struct buffer_head *bh;
1377                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1378                 if (!bh) {
1379                         *errp = -EIO;
1380                         goto err;
1381                 }
1382                 if (buffer_new(&dummy)) {
1383                         J_ASSERT(create != 0);
1384                         J_ASSERT(handle != NULL);
1385
1386                         /*
1387                          * Now that we do not always journal data, we should
1388                          * keep in mind whether this should always journal the
1389                          * new buffer as metadata.  For now, regular file
1390                          * writes use ext4_get_block instead, so it's not a
1391                          * problem.
1392                          */
1393                         lock_buffer(bh);
1394                         BUFFER_TRACE(bh, "call get_create_access");
1395                         fatal = ext4_journal_get_create_access(handle, bh);
1396                         if (!fatal && !buffer_uptodate(bh)) {
1397                                 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1398                                 set_buffer_uptodate(bh);
1399                         }
1400                         unlock_buffer(bh);
1401                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1402                         err = ext4_handle_dirty_metadata(handle, inode, bh);
1403                         if (!fatal)
1404                                 fatal = err;
1405                 } else {
1406                         BUFFER_TRACE(bh, "not a new buffer");
1407                 }
1408                 if (fatal) {
1409                         *errp = fatal;
1410                         brelse(bh);
1411                         bh = NULL;
1412                 }
1413                 return bh;
1414         }
1415 err:
1416         return NULL;
1417 }
1418
1419 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1420                                ext4_lblk_t block, int create, int *err)
1421 {
1422         struct buffer_head *bh;
1423
1424         bh = ext4_getblk(handle, inode, block, create, err);
1425         if (!bh)
1426                 return bh;
1427         if (buffer_uptodate(bh))
1428                 return bh;
1429         ll_rw_block(READ_META, 1, &bh);
1430         wait_on_buffer(bh);
1431         if (buffer_uptodate(bh))
1432                 return bh;
1433         put_bh(bh);
1434         *err = -EIO;
1435         return NULL;
1436 }
1437
1438 static int walk_page_buffers(handle_t *handle,
1439                              struct buffer_head *head,
1440                              unsigned from,
1441                              unsigned to,
1442                              int *partial,
1443                              int (*fn)(handle_t *handle,
1444                                        struct buffer_head *bh))
1445 {
1446         struct buffer_head *bh;
1447         unsigned block_start, block_end;
1448         unsigned blocksize = head->b_size;
1449         int err, ret = 0;
1450         struct buffer_head *next;
1451
1452         for (bh = head, block_start = 0;
1453              ret == 0 && (bh != head || !block_start);
1454              block_start = block_end, bh = next) {
1455                 next = bh->b_this_page;
1456                 block_end = block_start + blocksize;
1457                 if (block_end <= from || block_start >= to) {
1458                         if (partial && !buffer_uptodate(bh))
1459                                 *partial = 1;
1460                         continue;
1461                 }
1462                 err = (*fn)(handle, bh);
1463                 if (!ret)
1464                         ret = err;
1465         }
1466         return ret;
1467 }
1468
1469 /*
1470  * To preserve ordering, it is essential that the hole instantiation and
1471  * the data write be encapsulated in a single transaction.  We cannot
1472  * close off a transaction and start a new one between the ext4_get_block()
1473  * and the commit_write().  So doing the jbd2_journal_start at the start of
1474  * prepare_write() is the right place.
1475  *
1476  * Also, this function can nest inside ext4_writepage() ->
1477  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1478  * has generated enough buffer credits to do the whole page.  So we won't
1479  * block on the journal in that case, which is good, because the caller may
1480  * be PF_MEMALLOC.
1481  *
1482  * By accident, ext4 can be reentered when a transaction is open via
1483  * quota file writes.  If we were to commit the transaction while thus
1484  * reentered, there can be a deadlock - we would be holding a quota
1485  * lock, and the commit would never complete if another thread had a
1486  * transaction open and was blocking on the quota lock - a ranking
1487  * violation.
1488  *
1489  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1490  * will _not_ run commit under these circumstances because handle->h_ref
1491  * is elevated.  We'll still have enough credits for the tiny quotafile
1492  * write.
1493  */
1494 static int do_journal_get_write_access(handle_t *handle,
1495                                        struct buffer_head *bh)
1496 {
1497         if (!buffer_mapped(bh) || buffer_freed(bh))
1498                 return 0;
1499         return ext4_journal_get_write_access(handle, bh);
1500 }
1501
1502 /*
1503  * Truncate blocks that were not used by write. We have to truncate the
1504  * pagecache as well so that corresponding buffers get properly unmapped.
1505  */
1506 static void ext4_truncate_failed_write(struct inode *inode)
1507 {
1508         truncate_inode_pages(inode->i_mapping, inode->i_size);
1509         ext4_truncate(inode);
1510 }
1511
1512 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1513                             loff_t pos, unsigned len, unsigned flags,
1514                             struct page **pagep, void **fsdata)
1515 {
1516         struct inode *inode = mapping->host;
1517         int ret, needed_blocks;
1518         handle_t *handle;
1519         int retries = 0;
1520         struct page *page;
1521         pgoff_t index;
1522         unsigned from, to;
1523
1524         trace_ext4_write_begin(inode, pos, len, flags);
1525         /*
1526          * Reserve one block more for addition to orphan list in case
1527          * we allocate blocks but write fails for some reason
1528          */
1529         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1530         index = pos >> PAGE_CACHE_SHIFT;
1531         from = pos & (PAGE_CACHE_SIZE - 1);
1532         to = from + len;
1533
1534 retry:
1535         handle = ext4_journal_start(inode, needed_blocks);
1536         if (IS_ERR(handle)) {
1537                 ret = PTR_ERR(handle);
1538                 goto out;
1539         }
1540
1541         /* We cannot recurse into the filesystem as the transaction is already
1542          * started */
1543         flags |= AOP_FLAG_NOFS;
1544
1545         page = grab_cache_page_write_begin(mapping, index, flags);
1546         if (!page) {
1547                 ext4_journal_stop(handle);
1548                 ret = -ENOMEM;
1549                 goto out;
1550         }
1551         *pagep = page;
1552
1553         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1554                                 ext4_get_block);
1555
1556         if (!ret && ext4_should_journal_data(inode)) {
1557                 ret = walk_page_buffers(handle, page_buffers(page),
1558                                 from, to, NULL, do_journal_get_write_access);
1559         }
1560
1561         if (ret) {
1562                 unlock_page(page);
1563                 page_cache_release(page);
1564                 /*
1565                  * block_write_begin may have instantiated a few blocks
1566                  * outside i_size.  Trim these off again. Don't need
1567                  * i_size_read because we hold i_mutex.
1568                  *
1569                  * Add inode to orphan list in case we crash before
1570                  * truncate finishes
1571                  */
1572                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1573                         ext4_orphan_add(handle, inode);
1574
1575                 ext4_journal_stop(handle);
1576                 if (pos + len > inode->i_size) {
1577                         ext4_truncate_failed_write(inode);
1578                         /*
1579                          * If truncate failed early the inode might
1580                          * still be on the orphan list; we need to
1581                          * make sure the inode is removed from the
1582                          * orphan list in that case.
1583                          */
1584                         if (inode->i_nlink)
1585                                 ext4_orphan_del(NULL, inode);
1586                 }
1587         }
1588
1589         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1590                 goto retry;
1591 out:
1592         return ret;
1593 }
1594
1595 /* For write_end() in data=journal mode */
1596 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1597 {
1598         if (!buffer_mapped(bh) || buffer_freed(bh))
1599                 return 0;
1600         set_buffer_uptodate(bh);
1601         return ext4_handle_dirty_metadata(handle, NULL, bh);
1602 }
1603
1604 static int ext4_generic_write_end(struct file *file,
1605                                   struct address_space *mapping,
1606                                   loff_t pos, unsigned len, unsigned copied,
1607                                   struct page *page, void *fsdata)
1608 {
1609         int i_size_changed = 0;
1610         struct inode *inode = mapping->host;
1611         handle_t *handle = ext4_journal_current_handle();
1612
1613         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1614
1615         /*
1616          * No need to use i_size_read() here, the i_size
1617          * cannot change under us because we hold i_mutex.
1618          *
1619          * But it's important to update i_size while still holding page lock:
1620          * page writeout could otherwise come in and zero beyond i_size.
1621          */
1622         if (pos + copied > inode->i_size) {
1623                 i_size_write(inode, pos + copied);
1624                 i_size_changed = 1;
1625         }
1626
1627         if (pos + copied >  EXT4_I(inode)->i_disksize) {
1628                 /* We need to mark inode dirty even if
1629                  * new_i_size is less that inode->i_size
1630                  * bu greater than i_disksize.(hint delalloc)
1631                  */
1632                 ext4_update_i_disksize(inode, (pos + copied));
1633                 i_size_changed = 1;
1634         }
1635         unlock_page(page);
1636         page_cache_release(page);
1637
1638         /*
1639          * Don't mark the inode dirty under page lock. First, it unnecessarily
1640          * makes the holding time of page lock longer. Second, it forces lock
1641          * ordering of page lock and transaction start for journaling
1642          * filesystems.
1643          */
1644         if (i_size_changed)
1645                 ext4_mark_inode_dirty(handle, inode);
1646
1647         return copied;
1648 }
1649
1650 /*
1651  * We need to pick up the new inode size which generic_commit_write gave us
1652  * `file' can be NULL - eg, when called from page_symlink().
1653  *
1654  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1655  * buffers are managed internally.
1656  */
1657 static int ext4_ordered_write_end(struct file *file,
1658                                   struct address_space *mapping,
1659                                   loff_t pos, unsigned len, unsigned copied,
1660                                   struct page *page, void *fsdata)
1661 {
1662         handle_t *handle = ext4_journal_current_handle();
1663         struct inode *inode = mapping->host;
1664         int ret = 0, ret2;
1665
1666         trace_ext4_ordered_write_end(inode, pos, len, copied);
1667         ret = ext4_jbd2_file_inode(handle, inode);
1668
1669         if (ret == 0) {
1670                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1671                                                         page, fsdata);
1672                 copied = ret2;
1673                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1674                         /* if we have allocated more blocks and copied
1675                          * less. We will have blocks allocated outside
1676                          * inode->i_size. So truncate them
1677                          */
1678                         ext4_orphan_add(handle, inode);
1679                 if (ret2 < 0)
1680                         ret = ret2;
1681         }
1682         ret2 = ext4_journal_stop(handle);
1683         if (!ret)
1684                 ret = ret2;
1685
1686         if (pos + len > inode->i_size) {
1687                 ext4_truncate_failed_write(inode);
1688                 /*
1689                  * If truncate failed early the inode might still be
1690                  * on the orphan list; we need to make sure the inode
1691                  * is removed from the orphan list in that case.
1692                  */
1693                 if (inode->i_nlink)
1694                         ext4_orphan_del(NULL, inode);
1695         }
1696
1697
1698         return ret ? ret : copied;
1699 }
1700
1701 static int ext4_writeback_write_end(struct file *file,
1702                                     struct address_space *mapping,
1703                                     loff_t pos, unsigned len, unsigned copied,
1704                                     struct page *page, void *fsdata)
1705 {
1706         handle_t *handle = ext4_journal_current_handle();
1707         struct inode *inode = mapping->host;
1708         int ret = 0, ret2;
1709
1710         trace_ext4_writeback_write_end(inode, pos, len, copied);
1711         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1712                                                         page, fsdata);
1713         copied = ret2;
1714         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1715                 /* if we have allocated more blocks and copied
1716                  * less. We will have blocks allocated outside
1717                  * inode->i_size. So truncate them
1718                  */
1719                 ext4_orphan_add(handle, inode);
1720
1721         if (ret2 < 0)
1722                 ret = ret2;
1723
1724         ret2 = ext4_journal_stop(handle);
1725         if (!ret)
1726                 ret = ret2;
1727
1728         if (pos + len > inode->i_size) {
1729                 ext4_truncate_failed_write(inode);
1730                 /*
1731                  * If truncate failed early the inode might still be
1732                  * on the orphan list; we need to make sure the inode
1733                  * is removed from the orphan list in that case.
1734                  */
1735                 if (inode->i_nlink)
1736                         ext4_orphan_del(NULL, inode);
1737         }
1738
1739         return ret ? ret : copied;
1740 }
1741
1742 static int ext4_journalled_write_end(struct file *file,
1743                                      struct address_space *mapping,
1744                                      loff_t pos, unsigned len, unsigned copied,
1745                                      struct page *page, void *fsdata)
1746 {
1747         handle_t *handle = ext4_journal_current_handle();
1748         struct inode *inode = mapping->host;
1749         int ret = 0, ret2;
1750         int partial = 0;
1751         unsigned from, to;
1752         loff_t new_i_size;
1753
1754         trace_ext4_journalled_write_end(inode, pos, len, copied);
1755         from = pos & (PAGE_CACHE_SIZE - 1);
1756         to = from + len;
1757
1758         if (copied < len) {
1759                 if (!PageUptodate(page))
1760                         copied = 0;
1761                 page_zero_new_buffers(page, from+copied, to);
1762         }
1763
1764         ret = walk_page_buffers(handle, page_buffers(page), from,
1765                                 to, &partial, write_end_fn);
1766         if (!partial)
1767                 SetPageUptodate(page);
1768         new_i_size = pos + copied;
1769         if (new_i_size > inode->i_size)
1770                 i_size_write(inode, pos+copied);
1771         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
1772         if (new_i_size > EXT4_I(inode)->i_disksize) {
1773                 ext4_update_i_disksize(inode, new_i_size);
1774                 ret2 = ext4_mark_inode_dirty(handle, inode);
1775                 if (!ret)
1776                         ret = ret2;
1777         }
1778
1779         unlock_page(page);
1780         page_cache_release(page);
1781         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1782                 /* if we have allocated more blocks and copied
1783                  * less. We will have blocks allocated outside
1784                  * inode->i_size. So truncate them
1785                  */
1786                 ext4_orphan_add(handle, inode);
1787
1788         ret2 = ext4_journal_stop(handle);
1789         if (!ret)
1790                 ret = ret2;
1791         if (pos + len > inode->i_size) {
1792                 ext4_truncate_failed_write(inode);
1793                 /*
1794                  * If truncate failed early the inode might still be
1795                  * on the orphan list; we need to make sure the inode
1796                  * is removed from the orphan list in that case.
1797                  */
1798                 if (inode->i_nlink)
1799                         ext4_orphan_del(NULL, inode);
1800         }
1801
1802         return ret ? ret : copied;
1803 }
1804
1805 static int ext4_da_reserve_space(struct inode *inode, int nrblocks)
1806 {
1807         int retries = 0;
1808         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1809         struct ext4_inode_info *ei = EXT4_I(inode);
1810         unsigned long md_needed, md_reserved, total = 0;
1811
1812         /*
1813          * recalculate the amount of metadata blocks to reserve
1814          * in order to allocate nrblocks
1815          * worse case is one extent per block
1816          */
1817 repeat:
1818         spin_lock(&ei->i_block_reservation_lock);
1819         md_reserved = ei->i_reserved_meta_blocks;
1820         md_needed = ext4_calc_metadata_amount(inode, nrblocks);
1821         total = md_needed + nrblocks;
1822         spin_unlock(&ei->i_block_reservation_lock);
1823
1824         /*
1825          * Make quota reservation here to prevent quota overflow
1826          * later. Real quota accounting is done at pages writeout
1827          * time.
1828          */
1829         if (vfs_dq_reserve_block(inode, total)) {
1830                 /* 
1831                  * We tend to badly over-estimate the amount of
1832                  * metadata blocks which are needed, so if we have
1833                  * reserved any metadata blocks, try to force out the
1834                  * inode and see if we have any better luck.
1835                  */
1836                 if (md_reserved && retries++ <= 3)
1837                         goto retry;
1838                 return -EDQUOT;
1839         }
1840
1841         if (ext4_claim_free_blocks(sbi, total)) {
1842                 vfs_dq_release_reservation_block(inode, total);
1843                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1844                 retry:
1845                         if (md_reserved)
1846                                 write_inode_now(inode, (retries == 3));
1847                         yield();
1848                         goto repeat;
1849                 }
1850                 return -ENOSPC;
1851         }
1852         spin_lock(&ei->i_block_reservation_lock);
1853         ei->i_reserved_data_blocks += nrblocks;
1854         ei->i_reserved_meta_blocks += md_needed;
1855         spin_unlock(&ei->i_block_reservation_lock);
1856
1857         return 0;       /* success */
1858 }
1859
1860 static void ext4_da_release_space(struct inode *inode, int to_free)
1861 {
1862         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1863         struct ext4_inode_info *ei = EXT4_I(inode);
1864
1865         if (!to_free)
1866                 return;         /* Nothing to release, exit */
1867
1868         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1869
1870         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1871                 /*
1872                  * if there aren't enough reserved blocks, then the
1873                  * counter is messed up somewhere.  Since this
1874                  * function is called from invalidate page, it's
1875                  * harmless to return without any action.
1876                  */
1877                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1878                          "ino %lu, to_free %d with only %d reserved "
1879                          "data blocks\n", inode->i_ino, to_free,
1880                          ei->i_reserved_data_blocks);
1881                 WARN_ON(1);
1882                 to_free = ei->i_reserved_data_blocks;
1883         }
1884         ei->i_reserved_data_blocks -= to_free;
1885
1886         if (ei->i_reserved_data_blocks == 0) {
1887                 /*
1888                  * We can release all of the reserved metadata blocks
1889                  * only when we have written all of the delayed
1890                  * allocation blocks.
1891                  */
1892                 to_free += ei->i_allocated_meta_blocks;
1893                 ei->i_allocated_meta_blocks = 0;
1894         }
1895
1896         /* update fs dirty blocks counter */
1897         percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1898
1899         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1900
1901         vfs_dq_release_reservation_block(inode, to_free);
1902 }
1903
1904 static void ext4_da_page_release_reservation(struct page *page,
1905                                              unsigned long offset)
1906 {
1907         int to_release = 0;
1908         struct buffer_head *head, *bh;
1909         unsigned int curr_off = 0;
1910
1911         head = page_buffers(page);
1912         bh = head;
1913         do {
1914                 unsigned int next_off = curr_off + bh->b_size;
1915
1916                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1917                         to_release++;
1918                         clear_buffer_delay(bh);
1919                 }
1920                 curr_off = next_off;
1921         } while ((bh = bh->b_this_page) != head);
1922         ext4_da_release_space(page->mapping->host, to_release);
1923 }
1924
1925 /*
1926  * Delayed allocation stuff
1927  */
1928
1929 /*
1930  * mpage_da_submit_io - walks through extent of pages and try to write
1931  * them with writepage() call back
1932  *
1933  * @mpd->inode: inode
1934  * @mpd->first_page: first page of the extent
1935  * @mpd->next_page: page after the last page of the extent
1936  *
1937  * By the time mpage_da_submit_io() is called we expect all blocks
1938  * to be allocated. this may be wrong if allocation failed.
1939  *
1940  * As pages are already locked by write_cache_pages(), we can't use it
1941  */
1942 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1943 {
1944         long pages_skipped;
1945         struct pagevec pvec;
1946         unsigned long index, end;
1947         int ret = 0, err, nr_pages, i;
1948         struct inode *inode = mpd->inode;
1949         struct address_space *mapping = inode->i_mapping;
1950
1951         BUG_ON(mpd->next_page <= mpd->first_page);
1952         /*
1953          * We need to start from the first_page to the next_page - 1
1954          * to make sure we also write the mapped dirty buffer_heads.
1955          * If we look at mpd->b_blocknr we would only be looking
1956          * at the currently mapped buffer_heads.
1957          */
1958         index = mpd->first_page;
1959         end = mpd->next_page - 1;
1960
1961         pagevec_init(&pvec, 0);
1962         while (index <= end) {
1963                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1964                 if (nr_pages == 0)
1965                         break;
1966                 for (i = 0; i < nr_pages; i++) {
1967                         struct page *page = pvec.pages[i];
1968
1969                         index = page->index;
1970                         if (index > end)
1971                                 break;
1972                         index++;
1973
1974                         BUG_ON(!PageLocked(page));
1975                         BUG_ON(PageWriteback(page));
1976
1977                         pages_skipped = mpd->wbc->pages_skipped;
1978                         err = mapping->a_ops->writepage(page, mpd->wbc);
1979                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
1980                                 /*
1981                                  * have successfully written the page
1982                                  * without skipping the same
1983                                  */
1984                                 mpd->pages_written++;
1985                         /*
1986                          * In error case, we have to continue because
1987                          * remaining pages are still locked
1988                          * XXX: unlock and re-dirty them?
1989                          */
1990                         if (ret == 0)
1991                                 ret = err;
1992                 }
1993                 pagevec_release(&pvec);
1994         }
1995         return ret;
1996 }
1997
1998 /*
1999  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
2000  *
2001  * @mpd->inode - inode to walk through
2002  * @exbh->b_blocknr - first block on a disk
2003  * @exbh->b_size - amount of space in bytes
2004  * @logical - first logical block to start assignment with
2005  *
2006  * the function goes through all passed space and put actual disk
2007  * block numbers into buffer heads, dropping BH_Delay and BH_Unwritten
2008  */
2009 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
2010                                  struct buffer_head *exbh)
2011 {
2012         struct inode *inode = mpd->inode;
2013         struct address_space *mapping = inode->i_mapping;
2014         int blocks = exbh->b_size >> inode->i_blkbits;
2015         sector_t pblock = exbh->b_blocknr, cur_logical;
2016         struct buffer_head *head, *bh;
2017         pgoff_t index, end;
2018         struct pagevec pvec;
2019         int nr_pages, i;
2020
2021         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2022         end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2023         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2024
2025         pagevec_init(&pvec, 0);
2026
2027         while (index <= end) {
2028                 /* XXX: optimize tail */
2029                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2030                 if (nr_pages == 0)
2031                         break;
2032                 for (i = 0; i < nr_pages; i++) {
2033                         struct page *page = pvec.pages[i];
2034
2035                         index = page->index;
2036                         if (index > end)
2037                                 break;
2038                         index++;
2039
2040                         BUG_ON(!PageLocked(page));
2041                         BUG_ON(PageWriteback(page));
2042                         BUG_ON(!page_has_buffers(page));
2043
2044                         bh = page_buffers(page);
2045                         head = bh;
2046
2047                         /* skip blocks out of the range */
2048                         do {
2049                                 if (cur_logical >= logical)
2050                                         break;
2051                                 cur_logical++;
2052                         } while ((bh = bh->b_this_page) != head);
2053
2054                         do {
2055                                 if (cur_logical >= logical + blocks)
2056                                         break;
2057
2058                                 if (buffer_delay(bh) ||
2059                                                 buffer_unwritten(bh)) {
2060
2061                                         BUG_ON(bh->b_bdev != inode->i_sb->s_bdev);
2062
2063                                         if (buffer_delay(bh)) {
2064                                                 clear_buffer_delay(bh);
2065                                                 bh->b_blocknr = pblock;
2066                                         } else {
2067                                                 /*
2068                                                  * unwritten already should have
2069                                                  * blocknr assigned. Verify that
2070                                                  */
2071                                                 clear_buffer_unwritten(bh);
2072                                                 BUG_ON(bh->b_blocknr != pblock);
2073                                         }
2074
2075                                 } else if (buffer_mapped(bh))
2076                                         BUG_ON(bh->b_blocknr != pblock);
2077
2078                                 cur_logical++;
2079                                 pblock++;
2080                         } while ((bh = bh->b_this_page) != head);
2081                 }
2082                 pagevec_release(&pvec);
2083         }
2084 }
2085
2086
2087 /*
2088  * __unmap_underlying_blocks - just a helper function to unmap
2089  * set of blocks described by @bh
2090  */
2091 static inline void __unmap_underlying_blocks(struct inode *inode,
2092                                              struct buffer_head *bh)
2093 {
2094         struct block_device *bdev = inode->i_sb->s_bdev;
2095         int blocks, i;
2096
2097         blocks = bh->b_size >> inode->i_blkbits;
2098         for (i = 0; i < blocks; i++)
2099                 unmap_underlying_metadata(bdev, bh->b_blocknr + i);
2100 }
2101
2102 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2103                                         sector_t logical, long blk_cnt)
2104 {
2105         int nr_pages, i;
2106         pgoff_t index, end;
2107         struct pagevec pvec;
2108         struct inode *inode = mpd->inode;
2109         struct address_space *mapping = inode->i_mapping;
2110
2111         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2112         end   = (logical + blk_cnt - 1) >>
2113                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2114         while (index <= end) {
2115                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2116                 if (nr_pages == 0)
2117                         break;
2118                 for (i = 0; i < nr_pages; i++) {
2119                         struct page *page = pvec.pages[i];
2120                         index = page->index;
2121                         if (index > end)
2122                                 break;
2123                         index++;
2124
2125                         BUG_ON(!PageLocked(page));
2126                         BUG_ON(PageWriteback(page));
2127                         block_invalidatepage(page, 0);
2128                         ClearPageUptodate(page);
2129                         unlock_page(page);
2130                 }
2131         }
2132         return;
2133 }
2134
2135 static void ext4_print_free_blocks(struct inode *inode)
2136 {
2137         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2138         printk(KERN_CRIT "Total free blocks count %lld\n",
2139                ext4_count_free_blocks(inode->i_sb));
2140         printk(KERN_CRIT "Free/Dirty block details\n");
2141         printk(KERN_CRIT "free_blocks=%lld\n",
2142                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2143         printk(KERN_CRIT "dirty_blocks=%lld\n",
2144                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2145         printk(KERN_CRIT "Block reservation details\n");
2146         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2147                EXT4_I(inode)->i_reserved_data_blocks);
2148         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2149                EXT4_I(inode)->i_reserved_meta_blocks);
2150         return;
2151 }
2152
2153 /*
2154  * mpage_da_map_blocks - go through given space
2155  *
2156  * @mpd - bh describing space
2157  *
2158  * The function skips space we know is already mapped to disk blocks.
2159  *
2160  */
2161 static int mpage_da_map_blocks(struct mpage_da_data *mpd)
2162 {
2163         int err, blks, get_blocks_flags;
2164         struct buffer_head new;
2165         sector_t next = mpd->b_blocknr;
2166         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2167         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2168         handle_t *handle = NULL;
2169
2170         /*
2171          * We consider only non-mapped and non-allocated blocks
2172          */
2173         if ((mpd->b_state  & (1 << BH_Mapped)) &&
2174                 !(mpd->b_state & (1 << BH_Delay)) &&
2175                 !(mpd->b_state & (1 << BH_Unwritten)))
2176                 return 0;
2177
2178         /*
2179          * If we didn't accumulate anything to write simply return
2180          */
2181         if (!mpd->b_size)
2182                 return 0;
2183
2184         handle = ext4_journal_current_handle();
2185         BUG_ON(!handle);
2186
2187         /*
2188          * Call ext4_get_blocks() to allocate any delayed allocation
2189          * blocks, or to convert an uninitialized extent to be
2190          * initialized (in the case where we have written into
2191          * one or more preallocated blocks).
2192          *
2193          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2194          * indicate that we are on the delayed allocation path.  This
2195          * affects functions in many different parts of the allocation
2196          * call path.  This flag exists primarily because we don't
2197          * want to change *many* call functions, so ext4_get_blocks()
2198          * will set the magic i_delalloc_reserved_flag once the
2199          * inode's allocation semaphore is taken.
2200          *
2201          * If the blocks in questions were delalloc blocks, set
2202          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2203          * variables are updated after the blocks have been allocated.
2204          */
2205         new.b_state = 0;
2206         get_blocks_flags = (EXT4_GET_BLOCKS_CREATE |
2207                             EXT4_GET_BLOCKS_DELALLOC_RESERVE);
2208         if (mpd->b_state & (1 << BH_Delay))
2209                 get_blocks_flags |= EXT4_GET_BLOCKS_UPDATE_RESERVE_SPACE;
2210         blks = ext4_get_blocks(handle, mpd->inode, next, max_blocks,
2211                                &new, get_blocks_flags);
2212         if (blks < 0) {
2213                 err = blks;
2214                 /*
2215                  * If get block returns with error we simply
2216                  * return. Later writepage will redirty the page and
2217                  * writepages will find the dirty page again
2218                  */
2219                 if (err == -EAGAIN)
2220                         return 0;
2221
2222                 if (err == -ENOSPC &&
2223                     ext4_count_free_blocks(mpd->inode->i_sb)) {
2224                         mpd->retval = err;
2225                         return 0;
2226                 }
2227
2228                 /*
2229                  * get block failure will cause us to loop in
2230                  * writepages, because a_ops->writepage won't be able
2231                  * to make progress. The page will be redirtied by
2232                  * writepage and writepages will again try to write
2233                  * the same.
2234                  */
2235                 ext4_msg(mpd->inode->i_sb, KERN_CRIT,
2236                          "delayed block allocation failed for inode %lu at "
2237                          "logical offset %llu with max blocks %zd with "
2238                          "error %d\n", mpd->inode->i_ino,
2239                          (unsigned long long) next,
2240                          mpd->b_size >> mpd->inode->i_blkbits, err);
2241                 printk(KERN_CRIT "This should not happen!!  "
2242                        "Data will be lost\n");
2243                 if (err == -ENOSPC) {
2244                         ext4_print_free_blocks(mpd->inode);
2245                 }
2246                 /* invalidate all the pages */
2247                 ext4_da_block_invalidatepages(mpd, next,
2248                                 mpd->b_size >> mpd->inode->i_blkbits);
2249                 return err;
2250         }
2251         BUG_ON(blks == 0);
2252
2253         new.b_size = (blks << mpd->inode->i_blkbits);
2254
2255         if (buffer_new(&new))
2256                 __unmap_underlying_blocks(mpd->inode, &new);
2257
2258         /*
2259          * If blocks are delayed marked, we need to
2260          * put actual blocknr and drop delayed bit
2261          */
2262         if ((mpd->b_state & (1 << BH_Delay)) ||
2263             (mpd->b_state & (1 << BH_Unwritten)))
2264                 mpage_put_bnr_to_bhs(mpd, next, &new);
2265
2266         if (ext4_should_order_data(mpd->inode)) {
2267                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2268                 if (err)
2269                         return err;
2270         }
2271
2272         /*
2273          * Update on-disk size along with block allocation.
2274          */
2275         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2276         if (disksize > i_size_read(mpd->inode))
2277                 disksize = i_size_read(mpd->inode);
2278         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2279                 ext4_update_i_disksize(mpd->inode, disksize);
2280                 return ext4_mark_inode_dirty(handle, mpd->inode);
2281         }
2282
2283         return 0;
2284 }
2285
2286 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2287                 (1 << BH_Delay) | (1 << BH_Unwritten))
2288
2289 /*
2290  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2291  *
2292  * @mpd->lbh - extent of blocks
2293  * @logical - logical number of the block in the file
2294  * @bh - bh of the block (used to access block's state)
2295  *
2296  * the function is used to collect contig. blocks in same state
2297  */
2298 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2299                                    sector_t logical, size_t b_size,
2300                                    unsigned long b_state)
2301 {
2302         sector_t next;
2303         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2304
2305         /* check if thereserved journal credits might overflow */
2306         if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
2307                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2308                         /*
2309                          * With non-extent format we are limited by the journal
2310                          * credit available.  Total credit needed to insert
2311                          * nrblocks contiguous blocks is dependent on the
2312                          * nrblocks.  So limit nrblocks.
2313                          */
2314                         goto flush_it;
2315                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2316                                 EXT4_MAX_TRANS_DATA) {
2317                         /*
2318                          * Adding the new buffer_head would make it cross the
2319                          * allowed limit for which we have journal credit
2320                          * reserved. So limit the new bh->b_size
2321                          */
2322                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2323                                                 mpd->inode->i_blkbits;
2324                         /* we will do mpage_da_submit_io in the next loop */
2325                 }
2326         }
2327         /*
2328          * First block in the extent
2329          */
2330         if (mpd->b_size == 0) {
2331                 mpd->b_blocknr = logical;
2332                 mpd->b_size = b_size;
2333                 mpd->b_state = b_state & BH_FLAGS;
2334                 return;
2335         }
2336
2337         next = mpd->b_blocknr + nrblocks;
2338         /*
2339          * Can we merge the block to our big extent?
2340          */
2341         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2342                 mpd->b_size += b_size;
2343                 return;
2344         }
2345
2346 flush_it:
2347         /*
2348          * We couldn't merge the block to our extent, so we
2349          * need to flush current  extent and start new one
2350          */
2351         if (mpage_da_map_blocks(mpd) == 0)
2352                 mpage_da_submit_io(mpd);
2353         mpd->io_done = 1;
2354         return;
2355 }
2356
2357 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
2358 {
2359         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
2360 }
2361
2362 /*
2363  * __mpage_da_writepage - finds extent of pages and blocks
2364  *
2365  * @page: page to consider
2366  * @wbc: not used, we just follow rules
2367  * @data: context
2368  *
2369  * The function finds extents of pages and scan them for all blocks.
2370  */
2371 static int __mpage_da_writepage(struct page *page,
2372                                 struct writeback_control *wbc, void *data)
2373 {
2374         struct mpage_da_data *mpd = data;
2375         struct inode *inode = mpd->inode;
2376         struct buffer_head *bh, *head;
2377         sector_t logical;
2378
2379         if (mpd->io_done) {
2380                 /*
2381                  * Rest of the page in the page_vec
2382                  * redirty then and skip then. We will
2383                  * try to write them again after
2384                  * starting a new transaction
2385                  */
2386                 redirty_page_for_writepage(wbc, page);
2387                 unlock_page(page);
2388                 return MPAGE_DA_EXTENT_TAIL;
2389         }
2390         /*
2391          * Can we merge this page to current extent?
2392          */
2393         if (mpd->next_page != page->index) {
2394                 /*
2395                  * Nope, we can't. So, we map non-allocated blocks
2396                  * and start IO on them using writepage()
2397                  */
2398                 if (mpd->next_page != mpd->first_page) {
2399                         if (mpage_da_map_blocks(mpd) == 0)
2400                                 mpage_da_submit_io(mpd);
2401                         /*
2402                          * skip rest of the page in the page_vec
2403                          */
2404                         mpd->io_done = 1;
2405                         redirty_page_for_writepage(wbc, page);
2406                         unlock_page(page);
2407                         return MPAGE_DA_EXTENT_TAIL;
2408                 }
2409
2410                 /*
2411                  * Start next extent of pages ...
2412                  */
2413                 mpd->first_page = page->index;
2414
2415                 /*
2416                  * ... and blocks
2417                  */
2418                 mpd->b_size = 0;
2419                 mpd->b_state = 0;
2420                 mpd->b_blocknr = 0;
2421         }
2422
2423         mpd->next_page = page->index + 1;
2424         logical = (sector_t) page->index <<
2425                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2426
2427         if (!page_has_buffers(page)) {
2428                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2429                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2430                 if (mpd->io_done)
2431                         return MPAGE_DA_EXTENT_TAIL;
2432         } else {
2433                 /*
2434                  * Page with regular buffer heads, just add all dirty ones
2435                  */
2436                 head = page_buffers(page);
2437                 bh = head;
2438                 do {
2439                         BUG_ON(buffer_locked(bh));
2440                         /*
2441                          * We need to try to allocate
2442                          * unmapped blocks in the same page.
2443                          * Otherwise we won't make progress
2444                          * with the page in ext4_writepage
2445                          */
2446                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2447                                 mpage_add_bh_to_extent(mpd, logical,
2448                                                        bh->b_size,
2449                                                        bh->b_state);
2450                                 if (mpd->io_done)
2451                                         return MPAGE_DA_EXTENT_TAIL;
2452                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2453                                 /*
2454                                  * mapped dirty buffer. We need to update
2455                                  * the b_state because we look at
2456                                  * b_state in mpage_da_map_blocks. We don't
2457                                  * update b_size because if we find an
2458                                  * unmapped buffer_head later we need to
2459                                  * use the b_state flag of that buffer_head.
2460                                  */
2461                                 if (mpd->b_size == 0)
2462                                         mpd->b_state = bh->b_state & BH_FLAGS;
2463                         }
2464                         logical++;
2465                 } while ((bh = bh->b_this_page) != head);
2466         }
2467
2468         return 0;
2469 }
2470
2471 /*
2472  * This is a special get_blocks_t callback which is used by
2473  * ext4_da_write_begin().  It will either return mapped block or
2474  * reserve space for a single block.
2475  *
2476  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2477  * We also have b_blocknr = -1 and b_bdev initialized properly
2478  *
2479  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2480  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2481  * initialized properly.
2482  */
2483 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2484                                   struct buffer_head *bh_result, int create)
2485 {
2486         int ret = 0;
2487         sector_t invalid_block = ~((sector_t) 0xffff);
2488
2489         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2490                 invalid_block = ~0;
2491
2492         BUG_ON(create == 0);
2493         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2494
2495         /*
2496          * first, we need to know whether the block is allocated already
2497          * preallocated blocks are unmapped but should treated
2498          * the same as allocated blocks.
2499          */
2500         ret = ext4_get_blocks(NULL, inode, iblock, 1,  bh_result, 0);
2501         if ((ret == 0) && !buffer_delay(bh_result)) {
2502                 /* the block isn't (pre)allocated yet, let's reserve space */
2503                 /*
2504                  * XXX: __block_prepare_write() unmaps passed block,
2505                  * is it OK?
2506                  */
2507                 ret = ext4_da_reserve_space(inode, 1);
2508                 if (ret)
2509                         /* not enough space to reserve */
2510                         return ret;
2511
2512                 map_bh(bh_result, inode->i_sb, invalid_block);
2513                 set_buffer_new(bh_result);
2514                 set_buffer_delay(bh_result);
2515         } else if (ret > 0) {
2516                 bh_result->b_size = (ret << inode->i_blkbits);
2517                 if (buffer_unwritten(bh_result)) {
2518                         /* A delayed write to unwritten bh should
2519                          * be marked new and mapped.  Mapped ensures
2520                          * that we don't do get_block multiple times
2521                          * when we write to the same offset and new
2522                          * ensures that we do proper zero out for
2523                          * partial write.
2524                          */
2525                         set_buffer_new(bh_result);
2526                         set_buffer_mapped(bh_result);
2527                 }
2528                 ret = 0;
2529         }
2530
2531         return ret;
2532 }
2533
2534 /*
2535  * This function is used as a standard get_block_t calback function
2536  * when there is no desire to allocate any blocks.  It is used as a
2537  * callback function for block_prepare_write(), nobh_writepage(), and
2538  * block_write_full_page().  These functions should only try to map a
2539  * single block at a time.
2540  *
2541  * Since this function doesn't do block allocations even if the caller
2542  * requests it by passing in create=1, it is critically important that
2543  * any caller checks to make sure that any buffer heads are returned
2544  * by this function are either all already mapped or marked for
2545  * delayed allocation before calling nobh_writepage() or
2546  * block_write_full_page().  Otherwise, b_blocknr could be left
2547  * unitialized, and the page write functions will be taken by
2548  * surprise.
2549  */
2550 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2551                                    struct buffer_head *bh_result, int create)
2552 {
2553         int ret = 0;
2554         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2555
2556         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2557
2558         /*
2559          * we don't want to do block allocation in writepage
2560          * so call get_block_wrap with create = 0
2561          */
2562         ret = ext4_get_blocks(NULL, inode, iblock, max_blocks, bh_result, 0);
2563         if (ret > 0) {
2564                 bh_result->b_size = (ret << inode->i_blkbits);
2565                 ret = 0;
2566         }
2567         return ret;
2568 }
2569
2570 static int bget_one(handle_t *handle, struct buffer_head *bh)
2571 {
2572         get_bh(bh);
2573         return 0;
2574 }
2575
2576 static int bput_one(handle_t *handle, struct buffer_head *bh)
2577 {
2578         put_bh(bh);
2579         return 0;
2580 }
2581
2582 static int __ext4_journalled_writepage(struct page *page,
2583                                        unsigned int len)
2584 {
2585         struct address_space *mapping = page->mapping;
2586         struct inode *inode = mapping->host;
2587         struct buffer_head *page_bufs;
2588         handle_t *handle = NULL;
2589         int ret = 0;
2590         int err;
2591
2592         page_bufs = page_buffers(page);
2593         BUG_ON(!page_bufs);
2594         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2595         /* As soon as we unlock the page, it can go away, but we have
2596          * references to buffers so we are safe */
2597         unlock_page(page);
2598
2599         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2600         if (IS_ERR(handle)) {
2601                 ret = PTR_ERR(handle);
2602                 goto out;
2603         }
2604
2605         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2606                                 do_journal_get_write_access);
2607
2608         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2609                                 write_end_fn);
2610         if (ret == 0)
2611                 ret = err;
2612         err = ext4_journal_stop(handle);
2613         if (!ret)
2614                 ret = err;
2615
2616         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2617         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
2618 out:
2619         return ret;
2620 }
2621
2622 /*
2623  * Note that we don't need to start a transaction unless we're journaling data
2624  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2625  * need to file the inode to the transaction's list in ordered mode because if
2626  * we are writing back data added by write(), the inode is already there and if
2627  * we are writing back data modified via mmap(), noone guarantees in which
2628  * transaction the data will hit the disk. In case we are journaling data, we
2629  * cannot start transaction directly because transaction start ranks above page
2630  * lock so we have to do some magic.
2631  *
2632  * This function can get called via...
2633  *   - ext4_da_writepages after taking page lock (have journal handle)
2634  *   - journal_submit_inode_data_buffers (no journal handle)
2635  *   - shrink_page_list via pdflush (no journal handle)
2636  *   - grab_page_cache when doing write_begin (have journal handle)
2637  *
2638  * We don't do any block allocation in this function. If we have page with
2639  * multiple blocks we need to write those buffer_heads that are mapped. This
2640  * is important for mmaped based write. So if we do with blocksize 1K
2641  * truncate(f, 1024);
2642  * a = mmap(f, 0, 4096);
2643  * a[0] = 'a';
2644  * truncate(f, 4096);
2645  * we have in the page first buffer_head mapped via page_mkwrite call back
2646  * but other bufer_heads would be unmapped but dirty(dirty done via the
2647  * do_wp_page). So writepage should write the first block. If we modify
2648  * the mmap area beyond 1024 we will again get a page_fault and the
2649  * page_mkwrite callback will do the block allocation and mark the
2650  * buffer_heads mapped.
2651  *
2652  * We redirty the page if we have any buffer_heads that is either delay or
2653  * unwritten in the page.
2654  *
2655  * We can get recursively called as show below.
2656  *
2657  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2658  *              ext4_writepage()
2659  *
2660  * But since we don't do any block allocation we should not deadlock.
2661  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2662  */
2663 static int ext4_writepage(struct page *page,
2664                           struct writeback_control *wbc)
2665 {
2666         int ret = 0;
2667         loff_t size;
2668         unsigned int len;
2669         struct buffer_head *page_bufs;
2670         struct inode *inode = page->mapping->host;
2671
2672         trace_ext4_writepage(inode, page);
2673         size = i_size_read(inode);
2674         if (page->index == size >> PAGE_CACHE_SHIFT)
2675                 len = size & ~PAGE_CACHE_MASK;
2676         else
2677                 len = PAGE_CACHE_SIZE;
2678
2679         if (page_has_buffers(page)) {
2680                 page_bufs = page_buffers(page);
2681                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2682                                         ext4_bh_delay_or_unwritten)) {
2683                         /*
2684                          * We don't want to do  block allocation
2685                          * So redirty the page and return
2686                          * We may reach here when we do a journal commit
2687                          * via journal_submit_inode_data_buffers.
2688                          * If we don't have mapping block we just ignore
2689                          * them. We can also reach here via shrink_page_list
2690                          */
2691                         redirty_page_for_writepage(wbc, page);
2692                         unlock_page(page);
2693                         return 0;
2694                 }
2695         } else {
2696                 /*
2697                  * The test for page_has_buffers() is subtle:
2698                  * We know the page is dirty but it lost buffers. That means
2699                  * that at some moment in time after write_begin()/write_end()
2700                  * has been called all buffers have been clean and thus they
2701                  * must have been written at least once. So they are all
2702                  * mapped and we can happily proceed with mapping them
2703                  * and writing the page.
2704                  *
2705                  * Try to initialize the buffer_heads and check whether
2706                  * all are mapped and non delay. We don't want to
2707                  * do block allocation here.
2708                  */
2709                 ret = block_prepare_write(page, 0, len,
2710                                           noalloc_get_block_write);
2711                 if (!ret) {
2712                         page_bufs = page_buffers(page);
2713                         /* check whether all are mapped and non delay */
2714                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2715                                                 ext4_bh_delay_or_unwritten)) {
2716                                 redirty_page_for_writepage(wbc, page);
2717                                 unlock_page(page);
2718                                 return 0;
2719                         }
2720                 } else {
2721                         /*
2722                          * We can't do block allocation here
2723                          * so just redity the page and unlock
2724                          * and return
2725                          */
2726                         redirty_page_for_writepage(wbc, page);
2727                         unlock_page(page);
2728                         return 0;
2729                 }
2730                 /* now mark the buffer_heads as dirty and uptodate */
2731                 block_commit_write(page, 0, len);
2732         }
2733
2734         if (PageChecked(page) && ext4_should_journal_data(inode)) {
2735                 /*
2736                  * It's mmapped pagecache.  Add buffers and journal it.  There
2737                  * doesn't seem much point in redirtying the page here.
2738                  */
2739                 ClearPageChecked(page);
2740                 return __ext4_journalled_writepage(page, len);
2741         }
2742
2743         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2744                 ret = nobh_writepage(page, noalloc_get_block_write, wbc);
2745         else
2746                 ret = block_write_full_page(page, noalloc_get_block_write,
2747                                             wbc);
2748
2749         return ret;
2750 }
2751
2752 /*
2753  * This is called via ext4_da_writepages() to
2754  * calulate the total number of credits to reserve to fit
2755  * a single extent allocation into a single transaction,
2756  * ext4_da_writpeages() will loop calling this before
2757  * the block allocation.
2758  */
2759
2760 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2761 {
2762         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2763
2764         /*
2765          * With non-extent format the journal credit needed to
2766          * insert nrblocks contiguous block is dependent on
2767          * number of contiguous block. So we will limit
2768          * number of contiguous block to a sane value
2769          */
2770         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) &&
2771             (max_blocks > EXT4_MAX_TRANS_DATA))
2772                 max_blocks = EXT4_MAX_TRANS_DATA;
2773
2774         return ext4_chunk_trans_blocks(inode, max_blocks);
2775 }
2776
2777 static int ext4_da_writepages(struct address_space *mapping,
2778                               struct writeback_control *wbc)
2779 {
2780         pgoff_t index;
2781         int range_whole = 0;
2782         handle_t *handle = NULL;
2783         struct mpage_da_data mpd;
2784         struct inode *inode = mapping->host;
2785         int no_nrwrite_index_update;
2786         int pages_written = 0;
2787         long pages_skipped;
2788         unsigned int max_pages;
2789         int range_cyclic, cycled = 1, io_done = 0;
2790         int needed_blocks, ret = 0;
2791         long desired_nr_to_write, nr_to_writebump = 0;
2792         loff_t range_start = wbc->range_start;
2793         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2794
2795         trace_ext4_da_writepages(inode, wbc);
2796
2797         /*
2798          * No pages to write? This is mainly a kludge to avoid starting
2799          * a transaction for special inodes like journal inode on last iput()
2800          * because that could violate lock ordering on umount
2801          */
2802         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2803                 return 0;
2804
2805         /*
2806          * If the filesystem has aborted, it is read-only, so return
2807          * right away instead of dumping stack traces later on that
2808          * will obscure the real source of the problem.  We test
2809          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2810          * the latter could be true if the filesystem is mounted
2811          * read-only, and in that case, ext4_da_writepages should
2812          * *never* be called, so if that ever happens, we would want
2813          * the stack trace.
2814          */
2815         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2816                 return -EROFS;
2817
2818         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2819                 range_whole = 1;
2820
2821         range_cyclic = wbc->range_cyclic;
2822         if (wbc->range_cyclic) {
2823                 index = mapping->writeback_index;
2824                 if (index)
2825                         cycled = 0;
2826                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2827                 wbc->range_end  = LLONG_MAX;
2828                 wbc->range_cyclic = 0;
2829         } else
2830                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2831
2832         /*
2833          * This works around two forms of stupidity.  The first is in
2834          * the writeback code, which caps the maximum number of pages
2835          * written to be 1024 pages.  This is wrong on multiple
2836          * levels; different architectues have a different page size,
2837          * which changes the maximum amount of data which gets
2838          * written.  Secondly, 4 megabytes is way too small.  XFS
2839          * forces this value to be 16 megabytes by multiplying
2840          * nr_to_write parameter by four, and then relies on its
2841          * allocator to allocate larger extents to make them
2842          * contiguous.  Unfortunately this brings us to the second
2843          * stupidity, which is that ext4's mballoc code only allocates
2844          * at most 2048 blocks.  So we force contiguous writes up to
2845          * the number of dirty blocks in the inode, or
2846          * sbi->max_writeback_mb_bump whichever is smaller.
2847          */
2848         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2849         if (!range_cyclic && range_whole)
2850                 desired_nr_to_write = wbc->nr_to_write * 8;
2851         else
2852                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2853                                                            max_pages);
2854         if (desired_nr_to_write > max_pages)
2855                 desired_nr_to_write = max_pages;
2856
2857         if (wbc->nr_to_write < desired_nr_to_write) {
2858                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2859                 wbc->nr_to_write = desired_nr_to_write;
2860         }
2861
2862         mpd.wbc = wbc;
2863         mpd.inode = mapping->host;
2864
2865         /*
2866          * we don't want write_cache_pages to update
2867          * nr_to_write and writeback_index
2868          */
2869         no_nrwrite_index_update = wbc->no_nrwrite_index_update;
2870         wbc->no_nrwrite_index_update = 1;
2871         pages_skipped = wbc->pages_skipped;
2872
2873 retry:
2874         while (!ret && wbc->nr_to_write > 0) {
2875
2876                 /*
2877                  * we  insert one extent at a time. So we need
2878                  * credit needed for single extent allocation.
2879                  * journalled mode is currently not supported
2880                  * by delalloc
2881                  */
2882                 BUG_ON(ext4_should_journal_data(inode));
2883                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2884
2885                 /* start a new transaction*/
2886                 handle = ext4_journal_start(inode, needed_blocks);
2887                 if (IS_ERR(handle)) {
2888                         ret = PTR_ERR(handle);
2889                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2890                                "%ld pages, ino %lu; err %d\n", __func__,
2891                                 wbc->nr_to_write, inode->i_ino, ret);
2892                         goto out_writepages;
2893                 }
2894
2895                 /*
2896                  * Now call __mpage_da_writepage to find the next
2897                  * contiguous region of logical blocks that need
2898                  * blocks to be allocated by ext4.  We don't actually
2899                  * submit the blocks for I/O here, even though
2900                  * write_cache_pages thinks it will, and will set the
2901                  * pages as clean for write before calling
2902                  * __mpage_da_writepage().
2903                  */
2904                 mpd.b_size = 0;
2905                 mpd.b_state = 0;
2906                 mpd.b_blocknr = 0;
2907                 mpd.first_page = 0;
2908                 mpd.next_page = 0;
2909                 mpd.io_done = 0;
2910                 mpd.pages_written = 0;
2911                 mpd.retval = 0;
2912                 ret = write_cache_pages(mapping, wbc, __mpage_da_writepage,
2913                                         &mpd);
2914                 /*
2915                  * If we have a contiguous extent of pages and we
2916                  * haven't done the I/O yet, map the blocks and submit
2917                  * them for I/O.
2918                  */
2919                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2920                         if (mpage_da_map_blocks(&mpd) == 0)
2921                                 mpage_da_submit_io(&mpd);
2922                         mpd.io_done = 1;
2923                         ret = MPAGE_DA_EXTENT_TAIL;
2924                 }
2925                 trace_ext4_da_write_pages(inode, &mpd);
2926                 wbc->nr_to_write -= mpd.pages_written;
2927
2928                 ext4_journal_stop(handle);
2929
2930                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2931                         /* commit the transaction which would
2932                          * free blocks released in the transaction
2933                          * and try again
2934                          */
2935                         jbd2_journal_force_commit_nested(sbi->s_journal);
2936                         wbc->pages_skipped = pages_skipped;
2937                         ret = 0;
2938                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2939                         /*
2940                          * got one extent now try with
2941                          * rest of the pages
2942                          */
2943                         pages_written += mpd.pages_written;
2944                         wbc->pages_skipped = pages_skipped;
2945                         ret = 0;
2946                         io_done = 1;
2947                 } else if (wbc->nr_to_write)
2948                         /*
2949                          * There is no more writeout needed
2950                          * or we requested for a noblocking writeout
2951                          * and we found the device congested
2952                          */
2953                         break;
2954         }
2955         if (!io_done && !cycled) {
2956                 cycled = 1;
2957                 index = 0;
2958                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2959                 wbc->range_end  = mapping->writeback_index - 1;
2960                 goto retry;
2961         }
2962         if (pages_skipped != wbc->pages_skipped)
2963                 ext4_msg(inode->i_sb, KERN_CRIT,
2964                          "This should not happen leaving %s "
2965                          "with nr_to_write = %ld ret = %d\n",
2966                          __func__, wbc->nr_to_write, ret);
2967
2968         /* Update index */
2969         index += pages_written;
2970         wbc->range_cyclic = range_cyclic;
2971         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2972                 /*
2973                  * set the writeback_index so that range_cyclic
2974                  * mode will write it back later
2975                  */
2976                 mapping->writeback_index = index;
2977
2978 out_writepages:
2979         if (!no_nrwrite_index_update)
2980                 wbc->no_nrwrite_index_update = 0;
2981         wbc->nr_to_write -= nr_to_writebump;
2982         wbc->range_start = range_start;
2983         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
2984         return ret;
2985 }
2986
2987 #define FALL_BACK_TO_NONDELALLOC 1
2988 static int ext4_nonda_switch(struct super_block *sb)
2989 {
2990         s64 free_blocks, dirty_blocks;
2991         struct ext4_sb_info *sbi = EXT4_SB(sb);
2992
2993         /*
2994          * switch to non delalloc mode if we are running low
2995          * on free block. The free block accounting via percpu
2996          * counters can get slightly wrong with percpu_counter_batch getting
2997          * accumulated on each CPU without updating global counters
2998          * Delalloc need an accurate free block accounting. So switch
2999          * to non delalloc when we are near to error range.
3000          */
3001         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
3002         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
3003         if (2 * free_blocks < 3 * dirty_blocks ||
3004                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
3005                 /*
3006                  * free block count is less than 150% of dirty blocks
3007                  * or free blocks is less than watermark
3008                  */
3009                 return 1;
3010         }
3011         /*
3012          * Even if we don't switch but are nearing capacity,
3013          * start pushing delalloc when 1/2 of free blocks are dirty.
3014          */
3015         if (free_blocks < 2 * dirty_blocks)
3016                 writeback_inodes_sb_if_idle(sb);
3017
3018         return 0;
3019 }
3020
3021 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3022                                loff_t pos, unsigned len, unsigned flags,
3023                                struct page **pagep, void **fsdata)
3024 {
3025         int ret, retries = 0;
3026         struct page *page;
3027         pgoff_t index;
3028         unsigned from, to;
3029         struct inode *inode = mapping->host;
3030         handle_t *handle;
3031
3032         index = pos >> PAGE_CACHE_SHIFT;
3033         from = pos & (PAGE_CACHE_SIZE - 1);
3034         to = from + len;
3035
3036         if (ext4_nonda_switch(inode->i_sb)) {
3037                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3038                 return ext4_write_begin(file, mapping, pos,
3039                                         len, flags, pagep, fsdata);
3040         }
3041         *fsdata = (void *)0;
3042         trace_ext4_da_write_begin(inode, pos, len, flags);
3043 retry:
3044         /*
3045          * With delayed allocation, we don't log the i_disksize update
3046          * if there is delayed block allocation. But we still need
3047          * to journalling the i_disksize update if writes to the end
3048          * of file which has an already mapped buffer.
3049          */
3050         handle = ext4_journal_start(inode, 1);
3051         if (IS_ERR(handle)) {
3052                 ret = PTR_ERR(handle);
3053                 goto out;
3054         }
3055         /* We cannot recurse into the filesystem as the transaction is already
3056          * started */
3057         flags |= AOP_FLAG_NOFS;
3058
3059         page = grab_cache_page_write_begin(mapping, index, flags);
3060         if (!page) {
3061                 ext4_journal_stop(handle);
3062                 ret = -ENOMEM;
3063                 goto out;
3064         }
3065         *pagep = page;
3066
3067         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
3068                                 ext4_da_get_block_prep);
3069         if (ret < 0) {
3070                 unlock_page(page);
3071                 ext4_journal_stop(handle);
3072                 page_cache_release(page);
3073                 /*
3074                  * block_write_begin may have instantiated a few blocks
3075                  * outside i_size.  Trim these off again. Don't need
3076                  * i_size_read because we hold i_mutex.
3077                  */
3078                 if (pos + len > inode->i_size)
3079                         ext4_truncate_failed_write(inode);
3080         }
3081
3082         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3083                 goto retry;
3084 out:
3085         return ret;
3086 }
3087
3088 /*
3089  * Check if we should update i_disksize
3090  * when write to the end of file but not require block allocation
3091  */
3092 static int ext4_da_should_update_i_disksize(struct page *page,
3093                                             unsigned long offset)
3094 {
3095         struct buffer_head *bh;
3096         struct inode *inode = page->mapping->host;
3097         unsigned int idx;
3098         int i;
3099
3100         bh = page_buffers(page);
3101         idx = offset >> inode->i_blkbits;
3102
3103         for (i = 0; i < idx; i++)
3104                 bh = bh->b_this_page;
3105
3106         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3107                 return 0;
3108         return 1;
3109 }
3110
3111 static int ext4_da_write_end(struct file *file,
3112                              struct address_space *mapping,
3113                              loff_t pos, unsigned len, unsigned copied,
3114                              struct page *page, void *fsdata)
3115 {
3116         struct inode *inode = mapping->host;
3117         int ret = 0, ret2;
3118         handle_t *handle = ext4_journal_current_handle();
3119         loff_t new_i_size;
3120         unsigned long start, end;
3121         int write_mode = (int)(unsigned long)fsdata;
3122
3123         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
3124                 if (ext4_should_order_data(inode)) {
3125                         return ext4_ordered_write_end(file, mapping, pos,
3126                                         len, copied, page, fsdata);
3127                 } else if (ext4_should_writeback_data(inode)) {
3128                         return ext4_writeback_write_end(file, mapping, pos,
3129                                         len, copied, page, fsdata);
3130                 } else {
3131                         BUG();
3132                 }
3133         }
3134
3135         trace_ext4_da_write_end(inode, pos, len, copied);
3136         start = pos & (PAGE_CACHE_SIZE - 1);
3137         end = start + copied - 1;
3138
3139         /*
3140          * generic_write_end() will run mark_inode_dirty() if i_size
3141          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3142          * into that.
3143          */
3144
3145         new_i_size = pos + copied;
3146         if (new_i_size > EXT4_I(inode)->i_disksize) {
3147                 if (ext4_da_should_update_i_disksize(page, end)) {
3148                         down_write(&EXT4_I(inode)->i_data_sem);
3149                         if (new_i_size > EXT4_I(inode)->i_disksize) {
3150                                 /*
3151                                  * Updating i_disksize when extending file
3152                                  * without needing block allocation
3153                                  */
3154                                 if (ext4_should_order_data(inode))
3155                                         ret = ext4_jbd2_file_inode(handle,
3156                                                                    inode);
3157
3158                                 EXT4_I(inode)->i_disksize = new_i_size;
3159                         }
3160                         up_write(&EXT4_I(inode)->i_data_sem);
3161                         /* We need to mark inode dirty even if
3162                          * new_i_size is less that inode->i_size
3163                          * bu greater than i_disksize.(hint delalloc)
3164                          */
3165                         ext4_mark_inode_dirty(handle, inode);
3166                 }
3167         }
3168         ret2 = generic_write_end(file, mapping, pos, len, copied,
3169                                                         page, fsdata);
3170         copied = ret2;
3171         if (ret2 < 0)
3172                 ret = ret2;
3173         ret2 = ext4_journal_stop(handle);
3174         if (!ret)
3175                 ret = ret2;
3176
3177         return ret ? ret : copied;
3178 }
3179
3180 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
3181 {
3182         /*
3183          * Drop reserved blocks
3184          */
3185         BUG_ON(!PageLocked(page));
3186         if (!page_has_buffers(page))
3187                 goto out;
3188
3189         ext4_da_page_release_reservation(page, offset);
3190
3191 out:
3192         ext4_invalidatepage(page, offset);
3193
3194         return;
3195 }
3196
3197 /*
3198  * Force all delayed allocation blocks to be allocated for a given inode.
3199  */
3200 int ext4_alloc_da_blocks(struct inode *inode)
3201 {
3202         trace_ext4_alloc_da_blocks(inode);
3203
3204         if (!EXT4_I(inode)->i_reserved_data_blocks &&
3205             !EXT4_I(inode)->i_reserved_meta_blocks)
3206                 return 0;
3207
3208         /*
3209          * We do something simple for now.  The filemap_flush() will
3210          * also start triggering a write of the data blocks, which is
3211          * not strictly speaking necessary (and for users of
3212          * laptop_mode, not even desirable).  However, to do otherwise
3213          * would require replicating code paths in:
3214          *
3215          * ext4_da_writepages() ->
3216          *    write_cache_pages() ---> (via passed in callback function)
3217          *        __mpage_da_writepage() -->
3218          *           mpage_add_bh_to_extent()
3219          *           mpage_da_map_blocks()
3220          *
3221          * The problem is that write_cache_pages(), located in
3222          * mm/page-writeback.c, marks pages clean in preparation for
3223          * doing I/O, which is not desirable if we're not planning on
3224          * doing I/O at all.
3225          *
3226          * We could call write_cache_pages(), and then redirty all of
3227          * the pages by calling redirty_page_for_writeback() but that
3228          * would be ugly in the extreme.  So instead we would need to
3229          * replicate parts of the code in the above functions,
3230          * simplifying them becuase we wouldn't actually intend to
3231          * write out the pages, but rather only collect contiguous
3232          * logical block extents, call the multi-block allocator, and
3233          * then update the buffer heads with the block allocations.
3234          *
3235          * For now, though, we'll cheat by calling filemap_flush(),
3236          * which will map the blocks, and start the I/O, but not
3237          * actually wait for the I/O to complete.
3238          */
3239         return filemap_flush(inode->i_mapping);
3240 }
3241
3242 /*
3243  * bmap() is special.  It gets used by applications such as lilo and by
3244  * the swapper to find the on-disk block of a specific piece of data.
3245  *
3246  * Naturally, this is dangerous if the block concerned is still in the
3247  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3248  * filesystem and enables swap, then they may get a nasty shock when the
3249  * data getting swapped to that swapfile suddenly gets overwritten by
3250  * the original zero's written out previously to the journal and
3251  * awaiting writeback in the kernel's buffer cache.
3252  *
3253  * So, if we see any bmap calls here on a modified, data-journaled file,
3254  * take extra steps to flush any blocks which might be in the cache.
3255  */
3256 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3257 {
3258         struct inode *inode = mapping->host;
3259         journal_t *journal;
3260         int err;
3261
3262         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3263                         test_opt(inode->i_sb, DELALLOC)) {
3264                 /*
3265                  * With delalloc we want to sync the file
3266                  * so that we can make sure we allocate
3267                  * blocks for file
3268                  */
3269                 filemap_write_and_wait(mapping);
3270         }
3271
3272         if (EXT4_JOURNAL(inode) && EXT4_I(inode)->i_state & EXT4_STATE_JDATA) {
3273                 /*
3274                  * This is a REALLY heavyweight approach, but the use of
3275                  * bmap on dirty files is expected to be extremely rare:
3276                  * only if we run lilo or swapon on a freshly made file
3277                  * do we expect this to happen.
3278                  *
3279                  * (bmap requires CAP_SYS_RAWIO so this does not
3280                  * represent an unprivileged user DOS attack --- we'd be
3281                  * in trouble if mortal users could trigger this path at
3282                  * will.)
3283                  *
3284                  * NB. EXT4_STATE_JDATA is not set on files other than
3285                  * regular files.  If somebody wants to bmap a directory
3286                  * or symlink and gets confused because the buffer
3287                  * hasn't yet been flushed to disk, they deserve
3288                  * everything they get.
3289                  */
3290
3291                 EXT4_I(inode)->i_state &= ~EXT4_STATE_JDATA;
3292                 journal = EXT4_JOURNAL(inode);
3293                 jbd2_journal_lock_updates(journal);
3294                 err = jbd2_journal_flush(journal);
3295                 jbd2_journal_unlock_updates(journal);
3296
3297                 if (err)
3298                         return 0;
3299         }
3300
3301         return generic_block_bmap(mapping, block, ext4_get_block);
3302 }
3303
3304 static int ext4_readpage(struct file *file, struct page *page)
3305 {
3306         return mpage_readpage(page, ext4_get_block);
3307 }
3308
3309 static int
3310 ext4_readpages(struct file *file, struct address_space *mapping,
3311                 struct list_head *pages, unsigned nr_pages)
3312 {
3313         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3314 }
3315
3316 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3317 {
3318         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3319
3320         /*
3321          * If it's a full truncate we just forget about the pending dirtying
3322          */
3323         if (offset == 0)
3324                 ClearPageChecked(page);
3325
3326         if (journal)
3327                 jbd2_journal_invalidatepage(journal, page, offset);
3328         else
3329                 block_invalidatepage(page, offset);
3330 }
3331
3332 static int ext4_releasepage(struct page *page, gfp_t wait)
3333 {
3334         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3335
3336         WARN_ON(PageChecked(page));
3337         if (!page_has_buffers(page))
3338                 return 0;
3339         if (journal)
3340                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3341         else
3342                 return try_to_free_buffers(page);
3343 }
3344
3345 /*
3346  * O_DIRECT for ext3 (or indirect map) based files
3347  *
3348  * If the O_DIRECT write will extend the file then add this inode to the
3349  * orphan list.  So recovery will truncate it back to the original size
3350  * if the machine crashes during the write.
3351  *
3352  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3353  * crashes then stale disk data _may_ be exposed inside the file. But current
3354  * VFS code falls back into buffered path in that case so we are safe.
3355  */
3356 static ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
3357                               const struct iovec *iov, loff_t offset,
3358                               unsigned long nr_segs)
3359 {
3360         struct file *file = iocb->ki_filp;
3361         struct inode *inode = file->f_mapping->host;
3362         struct ext4_inode_info *ei = EXT4_I(inode);
3363         handle_t *handle;
3364         ssize_t ret;
3365         int orphan = 0;
3366         size_t count = iov_length(iov, nr_segs);
3367         int retries = 0;
3368
3369         if (rw == WRITE) {
3370                 loff_t final_size = offset + count;
3371
3372                 if (final_size > inode->i_size) {
3373                         /* Credits for sb + inode write */
3374                         handle = ext4_journal_start(inode, 2);
3375                         if (IS_ERR(handle)) {
3376                                 ret = PTR_ERR(handle);
3377                                 goto out;
3378                         }
3379                         ret = ext4_orphan_add(handle, inode);
3380                         if (ret) {
3381                                 ext4_journal_stop(handle);
3382                                 goto out;
3383                         }
3384                         orphan = 1;
3385                         ei->i_disksize = inode->i_size;
3386                         ext4_journal_stop(handle);
3387                 }
3388         }
3389
3390 retry:
3391         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
3392                                  offset, nr_segs,
3393                                  ext4_get_block, NULL);
3394         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3395                 goto retry;
3396
3397         if (orphan) {
3398                 int err;
3399
3400                 /* Credits for sb + inode write */
3401                 handle = ext4_journal_start(inode, 2);
3402                 if (IS_ERR(handle)) {
3403                         /* This is really bad luck. We've written the data
3404                          * but cannot extend i_size. Bail out and pretend
3405                          * the write failed... */
3406                         ret = PTR_ERR(handle);
3407                         goto out;
3408                 }
3409                 if (inode->i_nlink)
3410                         ext4_orphan_del(handle, inode);
3411                 if (ret > 0) {
3412                         loff_t end = offset + ret;
3413                         if (end > inode->i_size) {
3414                                 ei->i_disksize = end;
3415                                 i_size_write(inode, end);
3416                                 /*
3417                                  * We're going to return a positive `ret'
3418                                  * here due to non-zero-length I/O, so there's
3419                                  * no way of reporting error returns from
3420                                  * ext4_mark_inode_dirty() to userspace.  So
3421                                  * ignore it.
3422                                  */
3423                                 ext4_mark_inode_dirty(handle, inode);
3424                         }
3425                 }
3426                 err = ext4_journal_stop(handle);
3427                 if (ret == 0)
3428                         ret = err;
3429         }
3430 out:
3431         return ret;
3432 }
3433
3434 static int ext4_get_block_dio_write(struct inode *inode, sector_t iblock,
3435                    struct buffer_head *bh_result, int create)
3436 {
3437         handle_t *handle = NULL;
3438         int ret = 0;
3439         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
3440         int dio_credits;
3441
3442         ext4_debug("ext4_get_block_dio_write: inode %lu, create flag %d\n",
3443                    inode->i_ino, create);
3444         /*
3445          * DIO VFS code passes create = 0 flag for write to
3446          * the middle of file. It does this to avoid block
3447          * allocation for holes, to prevent expose stale data
3448          * out when there is parallel buffered read (which does
3449          * not hold the i_mutex lock) while direct IO write has
3450          * not completed. DIO request on holes finally falls back
3451          * to buffered IO for this reason.
3452          *
3453          * For ext4 extent based file, since we support fallocate,
3454          * new allocated extent as uninitialized, for holes, we
3455          * could fallocate blocks for holes, thus parallel
3456          * buffered IO read will zero out the page when read on
3457          * a hole while parallel DIO write to the hole has not completed.
3458          *
3459          * when we come here, we know it's a direct IO write to
3460          * to the middle of file (<i_size)
3461          * so it's safe to override the create flag from VFS.
3462          */
3463         create = EXT4_GET_BLOCKS_DIO_CREATE_EXT;
3464
3465         if (max_blocks > DIO_MAX_BLOCKS)
3466                 max_blocks = DIO_MAX_BLOCKS;
3467         dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
3468         handle = ext4_journal_start(inode, dio_credits);
3469         if (IS_ERR(handle)) {
3470                 ret = PTR_ERR(handle);
3471                 goto out;
3472         }
3473         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
3474                               create);
3475         if (ret > 0) {
3476                 bh_result->b_size = (ret << inode->i_blkbits);
3477                 ret = 0;
3478         }
3479         ext4_journal_stop(handle);
3480 out:
3481         return ret;
3482 }
3483
3484 static void ext4_free_io_end(ext4_io_end_t *io)
3485 {
3486         BUG_ON(!io);
3487         iput(io->inode);
3488         kfree(io);
3489 }
3490 static void dump_aio_dio_list(struct inode * inode)
3491 {
3492 #ifdef  EXT4_DEBUG
3493         struct list_head *cur, *before, *after;
3494         ext4_io_end_t *io, *io0, *io1;
3495
3496         if (list_empty(&EXT4_I(inode)->i_aio_dio_complete_list)){
3497                 ext4_debug("inode %lu aio dio list is empty\n", inode->i_ino);
3498                 return;
3499         }
3500
3501         ext4_debug("Dump inode %lu aio_dio_completed_IO list \n", inode->i_ino);
3502         list_for_each_entry(io, &EXT4_I(inode)->i_aio_dio_complete_list, list){
3503                 cur = &io->list;
3504                 before = cur->prev;
3505                 io0 = container_of(before, ext4_io_end_t, list);
3506                 after = cur->next;
3507                 io1 = container_of(after, ext4_io_end_t, list);
3508
3509                 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
3510                             io, inode->i_ino, io0, io1);
3511         }
3512 #endif
3513 }
3514
3515 /*
3516  * check a range of space and convert unwritten extents to written.
3517  */
3518 static int ext4_end_aio_dio_nolock(ext4_io_end_t *io)
3519 {
3520         struct inode *inode = io->inode;
3521         loff_t offset = io->offset;
3522         size_t size = io->size;
3523         int ret = 0;
3524
3525         ext4_debug("end_aio_dio_onlock: io 0x%p from inode %lu,list->next 0x%p,"
3526                    "list->prev 0x%p\n",
3527                    io, inode->i_ino, io->list.next, io->list.prev);
3528
3529         if (list_empty(&io->list))
3530                 return ret;
3531
3532         if (io->flag != DIO_AIO_UNWRITTEN)
3533                 return ret;
3534
3535         if (offset + size <= i_size_read(inode))
3536                 ret = ext4_convert_unwritten_extents(inode, offset, size);
3537
3538         if (ret < 0) {
3539                 printk(KERN_EMERG "%s: failed to convert unwritten"
3540                         "extents to written extents, error is %d"
3541                         " io is still on inode %lu aio dio list\n",
3542                        __func__, ret, inode->i_ino);
3543                 return ret;
3544         }
3545
3546         /* clear the DIO AIO unwritten flag */
3547         io->flag = 0;
3548         return ret;
3549 }
3550 /*
3551  * work on completed aio dio IO, to convert unwritten extents to extents
3552  */
3553 static void ext4_end_aio_dio_work(struct work_struct *work)
3554 {
3555         ext4_io_end_t *io  = container_of(work, ext4_io_end_t, work);
3556         struct inode *inode = io->inode;
3557         int ret = 0;
3558
3559         mutex_lock(&inode->i_mutex);
3560         ret = ext4_end_aio_dio_nolock(io);
3561         if (ret >= 0) {
3562                 if (!list_empty(&io->list))
3563                         list_del_init(&io->list);
3564                 ext4_free_io_end(io);
3565         }
3566         mutex_unlock(&inode->i_mutex);
3567 }
3568 /*
3569  * This function is called from ext4_sync_file().
3570  *
3571  * When AIO DIO IO is completed, the work to convert unwritten
3572  * extents to written is queued on workqueue but may not get immediately
3573  * scheduled. When fsync is called, we need to ensure the
3574  * conversion is complete before fsync returns.
3575  * The inode keeps track of a list of completed AIO from DIO path
3576  * that might needs to do the conversion. This function walks through
3577  * the list and convert the related unwritten extents to written.
3578  */
3579 int flush_aio_dio_completed_IO(struct inode *inode)
3580 {
3581         ext4_io_end_t *io;
3582         int ret = 0;
3583         int ret2 = 0;
3584
3585         if (list_empty(&EXT4_I(inode)->i_aio_dio_complete_list))
3586                 return ret;
3587
3588         dump_aio_dio_list(inode);
3589         while (!list_empty(&EXT4_I(inode)->i_aio_dio_complete_list)){
3590                 io = list_entry(EXT4_I(inode)->i_aio_dio_complete_list.next,
3591                                 ext4_io_end_t, list);
3592                 /*
3593                  * Calling ext4_end_aio_dio_nolock() to convert completed
3594                  * IO to written.
3595                  *
3596                  * When ext4_sync_file() is called, run_queue() may already
3597                  * about to flush the work corresponding to this io structure.
3598                  * It will be upset if it founds the io structure related
3599                  * to the work-to-be schedule is freed.
3600                  *
3601                  * Thus we need to keep the io structure still valid here after
3602                  * convertion finished. The io structure has a flag to
3603                  * avoid double converting from both fsync and background work
3604                  * queue work.
3605                  */
3606                 ret = ext4_end_aio_dio_nolock(io);
3607                 if (ret < 0)
3608                         ret2 = ret;
3609                 else
3610                         list_del_init(&io->list);
3611         }
3612         return (ret2 < 0) ? ret2 : 0;
3613 }
3614
3615 static ext4_io_end_t *ext4_init_io_end (struct inode *inode)
3616 {
3617         ext4_io_end_t *io = NULL;
3618
3619         io = kmalloc(sizeof(*io), GFP_NOFS);
3620
3621         if (io) {
3622                 igrab(inode);
3623                 io->inode = inode;
3624                 io->flag = 0;
3625                 io->offset = 0;
3626                 io->size = 0;
3627                 io->error = 0;
3628                 INIT_WORK(&io->work, ext4_end_aio_dio_work);
3629                 INIT_LIST_HEAD(&io->list);
3630         }
3631
3632         return io;
3633 }
3634
3635 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3636                             ssize_t size, void *private)
3637 {
3638         ext4_io_end_t *io_end = iocb->private;
3639         struct workqueue_struct *wq;
3640
3641         /* if not async direct IO or dio with 0 bytes write, just return */
3642         if (!io_end || !size)
3643                 return;
3644
3645         ext_debug("ext4_end_io_dio(): io_end 0x%p"
3646                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
3647                   iocb->private, io_end->inode->i_ino, iocb, offset,
3648                   size);
3649
3650         /* if not aio dio with unwritten extents, just free io and return */
3651         if (io_end->flag != DIO_AIO_UNWRITTEN){
3652                 ext4_free_io_end(io_end);
3653                 iocb->private = NULL;
3654                 return;
3655         }
3656
3657         io_end->offset = offset;
3658         io_end->size = size;
3659         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
3660
3661         /* queue the work to convert unwritten extents to written */
3662         queue_work(wq, &io_end->work);
3663
3664         /* Add the io_end to per-inode completed aio dio list*/
3665         list_add_tail(&io_end->list,
3666                  &EXT4_I(io_end->inode)->i_aio_dio_complete_list);
3667         iocb->private = NULL;
3668 }
3669 /*
3670  * For ext4 extent files, ext4 will do direct-io write to holes,
3671  * preallocated extents, and those write extend the file, no need to
3672  * fall back to buffered IO.
3673  *
3674  * For holes, we fallocate those blocks, mark them as unintialized
3675  * If those blocks were preallocated, we mark sure they are splited, but
3676  * still keep the range to write as unintialized.
3677  *
3678  * The unwrritten extents will be converted to written when DIO is completed.
3679  * For async direct IO, since the IO may still pending when return, we
3680  * set up an end_io call back function, which will do the convertion
3681  * when async direct IO completed.
3682  *
3683  * If the O_DIRECT write will extend the file then add this inode to the
3684  * orphan list.  So recovery will truncate it back to the original size
3685  * if the machine crashes during the write.
3686  *
3687  */
3688 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
3689                               const struct iovec *iov, loff_t offset,
3690                               unsigned long nr_segs)
3691 {
3692         struct file *file = iocb->ki_filp;
3693         struct inode *inode = file->f_mapping->host;
3694         ssize_t ret;
3695         size_t count = iov_length(iov, nr_segs);
3696
3697         loff_t final_size = offset + count;
3698         if (rw == WRITE && final_size <= inode->i_size) {
3699                 /*
3700                  * We could direct write to holes and fallocate.
3701                  *
3702                  * Allocated blocks to fill the hole are marked as uninitialized
3703                  * to prevent paralel buffered read to expose the stale data
3704                  * before DIO complete the data IO.
3705                  *
3706                  * As to previously fallocated extents, ext4 get_block
3707                  * will just simply mark the buffer mapped but still
3708                  * keep the extents uninitialized.
3709                  *
3710                  * for non AIO case, we will convert those unwritten extents
3711                  * to written after return back from blockdev_direct_IO.
3712                  *
3713                  * for async DIO, the conversion needs to be defered when
3714                  * the IO is completed. The ext4 end_io callback function
3715                  * will be called to take care of the conversion work.
3716                  * Here for async case, we allocate an io_end structure to
3717                  * hook to the iocb.
3718                  */
3719                 iocb->private = NULL;
3720                 EXT4_I(inode)->cur_aio_dio = NULL;
3721                 if (!is_sync_kiocb(iocb)) {
3722                         iocb->private = ext4_init_io_end(inode);
3723                         if (!iocb->private)
3724                                 return -ENOMEM;
3725                         /*
3726                          * we save the io structure for current async
3727                          * direct IO, so that later ext4_get_blocks()
3728                          * could flag the io structure whether there
3729                          * is a unwritten extents needs to be converted
3730                          * when IO is completed.
3731                          */
3732                         EXT4_I(inode)->cur_aio_dio = iocb->private;
3733                 }
3734
3735                 ret = blockdev_direct_IO(rw, iocb, inode,
3736                                          inode->i_sb->s_bdev, iov,
3737                                          offset, nr_segs,
3738                                          ext4_get_block_dio_write,
3739                                          ext4_end_io_dio);
3740                 if (iocb->private)
3741                         EXT4_I(inode)->cur_aio_dio = NULL;
3742                 /*
3743                  * The io_end structure takes a reference to the inode,
3744                  * that structure needs to be destroyed and the
3745                  * reference to the inode need to be dropped, when IO is
3746                  * complete, even with 0 byte write, or failed.
3747                  *
3748                  * In the successful AIO DIO case, the io_end structure will be
3749                  * desctroyed and the reference to the inode will be dropped
3750                  * after the end_io call back function is called.
3751                  *
3752                  * In the case there is 0 byte write, or error case, since
3753                  * VFS direct IO won't invoke the end_io call back function,
3754                  * we need to free the end_io structure here.
3755                  */
3756                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
3757                         ext4_free_io_end(iocb->private);
3758                         iocb->private = NULL;
3759                 } else if (ret > 0 && (EXT4_I(inode)->i_state &
3760                                        EXT4_STATE_DIO_UNWRITTEN)) {
3761                         int err;
3762                         /*
3763                          * for non AIO case, since the IO is already
3764                          * completed, we could do the convertion right here
3765                          */
3766                         err = ext4_convert_unwritten_extents(inode,
3767                                                              offset, ret);
3768                         if (err < 0)
3769                                 ret = err;
3770                         EXT4_I(inode)->i_state &= ~EXT4_STATE_DIO_UNWRITTEN;
3771                 }
3772                 return ret;
3773         }
3774
3775         /* for write the the end of file case, we fall back to old way */
3776         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3777 }
3778
3779 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3780                               const struct iovec *iov, loff_t offset,
3781                               unsigned long nr_segs)
3782 {
3783         struct file *file = iocb->ki_filp;
3784         struct inode *inode = file->f_mapping->host;
3785
3786         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
3787                 return ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
3788
3789         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3790 }
3791
3792 /*
3793  * Pages can be marked dirty completely asynchronously from ext4's journalling
3794  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3795  * much here because ->set_page_dirty is called under VFS locks.  The page is
3796  * not necessarily locked.
3797  *
3798  * We cannot just dirty the page and leave attached buffers clean, because the
3799  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3800  * or jbddirty because all the journalling code will explode.
3801  *
3802  * So what we do is to mark the page "pending dirty" and next time writepage
3803  * is called, propagate that into the buffers appropriately.
3804  */
3805 static int ext4_journalled_set_page_dirty(struct page *page)
3806 {
3807         SetPageChecked(page);
3808         return __set_page_dirty_nobuffers(page);
3809 }
3810
3811 static const struct address_space_operations ext4_ordered_aops = {
3812         .readpage               = ext4_readpage,
3813         .readpages              = ext4_readpages,
3814         .writepage              = ext4_writepage,
3815         .sync_page              = block_sync_page,
3816         .write_begin            = ext4_write_begin,
3817         .write_end              = ext4_ordered_write_end,
3818         .bmap                   = ext4_bmap,
3819         .invalidatepage         = ext4_invalidatepage,
3820         .releasepage            = ext4_releasepage,
3821         .direct_IO              = ext4_direct_IO,
3822         .migratepage            = buffer_migrate_page,
3823         .is_partially_uptodate  = block_is_partially_uptodate,
3824         .error_remove_page      = generic_error_remove_page,
3825 };
3826
3827 static const struct address_space_operations ext4_writeback_aops = {
3828         .readpage               = ext4_readpage,
3829         .readpages              = ext4_readpages,
3830         .writepage              = ext4_writepage,
3831         .sync_page              = block_sync_page,
3832         .write_begin            = ext4_write_begin,
3833         .write_end              = ext4_writeback_write_end,
3834         .bmap                   = ext4_bmap,
3835         .invalidatepage         = ext4_invalidatepage,
3836         .releasepage            = ext4_releasepage,
3837         .direct_IO              = ext4_direct_IO,
3838         .migratepage            = buffer_migrate_page,
3839         .is_partially_uptodate  = block_is_partially_uptodate,
3840         .error_remove_page      = generic_error_remove_page,
3841 };
3842
3843 static const struct address_space_operations ext4_journalled_aops = {
3844         .readpage               = ext4_readpage,
3845         .readpages              = ext4_readpages,
3846         .writepage              = ext4_writepage,
3847         .sync_page              = block_sync_page,
3848         .write_begin            = ext4_write_begin,
3849         .write_end              = ext4_journalled_write_end,
3850         .set_page_dirty         = ext4_journalled_set_page_dirty,
3851         .bmap                   = ext4_bmap,
3852         .invalidatepage         = ext4_invalidatepage,
3853         .releasepage            = ext4_releasepage,
3854         .is_partially_uptodate  = block_is_partially_uptodate,
3855         .error_remove_page      = generic_error_remove_page,
3856 };
3857
3858 static const struct address_space_operations ext4_da_aops = {
3859         .readpage               = ext4_readpage,
3860         .readpages              = ext4_readpages,
3861         .writepage              = ext4_writepage,
3862         .writepages             = ext4_da_writepages,
3863         .sync_page              = block_sync_page,
3864         .write_begin            = ext4_da_write_begin,
3865         .write_end              = ext4_da_write_end,
3866         .bmap                   = ext4_bmap,
3867         .invalidatepage         = ext4_da_invalidatepage,
3868         .releasepage            = ext4_releasepage,
3869         .direct_IO              = ext4_direct_IO,
3870         .migratepage            = buffer_migrate_page,
3871         .is_partially_uptodate  = block_is_partially_uptodate,
3872         .error_remove_page      = generic_error_remove_page,
3873 };
3874
3875 void ext4_set_aops(struct inode *inode)
3876 {
3877         if (ext4_should_order_data(inode) &&
3878                 test_opt(inode->i_sb, DELALLOC))
3879                 inode->i_mapping->a_ops = &ext4_da_aops;
3880         else if (ext4_should_order_data(inode))
3881                 inode->i_mapping->a_ops = &ext4_ordered_aops;
3882         else if (ext4_should_writeback_data(inode) &&
3883                  test_opt(inode->i_sb, DELALLOC))
3884                 inode->i_mapping->a_ops = &ext4_da_aops;
3885         else if (ext4_should_writeback_data(inode))
3886                 inode->i_mapping->a_ops = &ext4_writeback_aops;
3887         else
3888                 inode->i_mapping->a_ops = &ext4_journalled_aops;
3889 }
3890
3891 /*
3892  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3893  * up to the end of the block which corresponds to `from'.
3894  * This required during truncate. We need to physically zero the tail end
3895  * of that block so it doesn't yield old data if the file is later grown.
3896  */
3897 int ext4_block_truncate_page(handle_t *handle,
3898                 struct address_space *mapping, loff_t from)
3899 {
3900         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3901         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3902         unsigned blocksize, length, pos;
3903         ext4_lblk_t iblock;
3904         struct inode *inode = mapping->host;
3905         struct buffer_head *bh;
3906         struct page *page;
3907         int err = 0;
3908
3909         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
3910                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
3911         if (!page)
3912                 return -EINVAL;
3913
3914         blocksize = inode->i_sb->s_blocksize;
3915         length = blocksize - (offset & (blocksize - 1));
3916         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3917
3918         /*
3919          * For "nobh" option,  we can only work if we don't need to
3920          * read-in the page - otherwise we create buffers to do the IO.
3921          */
3922         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
3923              ext4_should_writeback_data(inode) && PageUptodate(page)) {
3924                 zero_user(page, offset, length);
3925                 set_page_dirty(page);
3926                 goto unlock;
3927         }
3928
3929         if (!page_has_buffers(page))
3930                 create_empty_buffers(page, blocksize, 0);
3931
3932         /* Find the buffer that contains "offset" */
3933         bh = page_buffers(page);
3934         pos = blocksize;
3935         while (offset >= pos) {
3936                 bh = bh->b_this_page;
3937                 iblock++;
3938                 pos += blocksize;
3939         }
3940
3941         err = 0;
3942         if (buffer_freed(bh)) {
3943                 BUFFER_TRACE(bh, "freed: skip");
3944                 goto unlock;
3945         }
3946
3947         if (!buffer_mapped(bh)) {
3948                 BUFFER_TRACE(bh, "unmapped");
3949                 ext4_get_block(inode, iblock, bh, 0);
3950                 /* unmapped? It's a hole - nothing to do */
3951                 if (!buffer_mapped(bh)) {
3952                         BUFFER_TRACE(bh, "still unmapped");
3953                         goto unlock;
3954                 }
3955         }
3956
3957         /* Ok, it's mapped. Make sure it's up-to-date */
3958         if (PageUptodate(page))
3959                 set_buffer_uptodate(bh);
3960
3961         if (!buffer_uptodate(bh)) {
3962                 err = -EIO;
3963                 ll_rw_block(READ, 1, &bh);
3964                 wait_on_buffer(bh);
3965                 /* Uhhuh. Read error. Complain and punt. */
3966                 if (!buffer_uptodate(bh))
3967                         goto unlock;
3968         }
3969
3970         if (ext4_should_journal_data(inode)) {
3971                 BUFFER_TRACE(bh, "get write access");
3972                 err = ext4_journal_get_write_access(handle, bh);
3973                 if (err)
3974                         goto unlock;
3975         }
3976
3977         zero_user(page, offset, length);
3978
3979         BUFFER_TRACE(bh, "zeroed end of block");
3980
3981         err = 0;
3982         if (ext4_should_journal_data(inode)) {
3983                 err = ext4_handle_dirty_metadata(handle, inode, bh);
3984         } else {
3985                 if (ext4_should_order_data(inode))
3986                         err = ext4_jbd2_file_inode(handle, inode);
3987                 mark_buffer_dirty(bh);
3988         }
3989
3990 unlock:
3991         unlock_page(page);
3992         page_cache_release(page);
3993         return err;
3994 }
3995
3996 /*
3997  * Probably it should be a library function... search for first non-zero word
3998  * or memcmp with zero_page, whatever is better for particular architecture.
3999  * Linus?
4000  */
4001 static inline int all_zeroes(__le32 *p, __le32 *q)
4002 {
4003         while (p < q)
4004                 if (*p++)
4005                         return 0;
4006         return 1;
4007 }
4008
4009 /**
4010  *      ext4_find_shared - find the indirect blocks for partial truncation.
4011  *      @inode:   inode in question
4012  *      @depth:   depth of the affected branch
4013  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
4014  *      @chain:   place to store the pointers to partial indirect blocks
4015  *      @top:     place to the (detached) top of branch
4016  *
4017  *      This is a helper function used by ext4_truncate().
4018  *
4019  *      When we do truncate() we may have to clean the ends of several
4020  *      indirect blocks but leave the blocks themselves alive. Block is
4021  *      partially truncated if some data below the new i_size is refered
4022  *      from it (and it is on the path to the first completely truncated
4023  *      data block, indeed).  We have to free the top of that path along
4024  *      with everything to the right of the path. Since no allocation
4025  *      past the truncation point is possible until ext4_truncate()
4026  *      finishes, we may safely do the latter, but top of branch may
4027  *      require special attention - pageout below the truncation point
4028  *      might try to populate it.
4029  *
4030  *      We atomically detach the top of branch from the tree, store the
4031  *      block number of its root in *@top, pointers to buffer_heads of
4032  *      partially truncated blocks - in @chain[].bh and pointers to
4033  *      their last elements that should not be removed - in
4034  *      @chain[].p. Return value is the pointer to last filled element
4035  *      of @chain.
4036  *
4037  *      The work left to caller to do the actual freeing of subtrees:
4038  *              a) free the subtree starting from *@top
4039  *              b) free the subtrees whose roots are stored in
4040  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
4041  *              c) free the subtrees growing from the inode past the @chain[0].
4042  *                      (no partially truncated stuff there).  */
4043
4044 static Indirect *ext4_find_shared(struct inode *inode, int depth,
4045                                   ext4_lblk_t offsets[4], Indirect chain[4],
4046                                   __le32 *top)
4047 {
4048         Indirect *partial, *p;
4049         int k, err;
4050
4051         *top = 0;
4052         /* Make k index the deepest non-null offset + 1 */
4053         for (k = depth; k > 1 && !offsets[k-1]; k--)
4054                 ;
4055         partial = ext4_get_branch(inode, k, offsets, chain, &err);
4056         /* Writer: pointers */
4057         if (!partial)
4058                 partial = chain + k-1;
4059         /*
4060          * If the branch acquired continuation since we've looked at it -
4061          * fine, it should all survive and (new) top doesn't belong to us.
4062          */
4063         if (!partial->key && *partial->p)
4064                 /* Writer: end */
4065                 goto no_top;
4066         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
4067                 ;
4068         /*
4069          * OK, we've found the last block that must survive. The rest of our
4070          * branch should be detached before unlocking. However, if that rest
4071          * of branch is all ours and does not grow immediately from the inode
4072          * it's easier to cheat and just decrement partial->p.
4073          */
4074         if (p == chain + k - 1 && p > chain) {
4075                 p->p--;
4076         } else {
4077                 *top = *p->p;
4078                 /* Nope, don't do this in ext4.  Must leave the tree intact */
4079 #if 0
4080                 *p->p = 0;
4081 #endif
4082         }
4083         /* Writer: end */
4084
4085         while (partial > p) {
4086                 brelse(partial->bh);
4087                 partial--;
4088         }
4089 no_top:
4090         return partial;
4091 }
4092
4093 /*
4094  * Zero a number of block pointers in either an inode or an indirect block.
4095  * If we restart the transaction we must again get write access to the
4096  * indirect block for further modification.
4097  *
4098  * We release `count' blocks on disk, but (last - first) may be greater
4099  * than `count' because there can be holes in there.
4100  */
4101 static void ext4_clear_blocks(handle_t *handle, struct inode *inode,
4102                               struct buffer_head *bh,
4103                               ext4_fsblk_t block_to_free,
4104                               unsigned long count, __le32 *first,
4105                               __le32 *last)
4106 {
4107         __le32 *p;
4108         int     flags = EXT4_FREE_BLOCKS_FORGET;
4109
4110         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
4111                 flags |= EXT4_FREE_BLOCKS_METADATA;
4112
4113         if (try_to_extend_transaction(handle, inode)) {
4114                 if (bh) {
4115                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4116                         ext4_handle_dirty_metadata(handle, inode, bh);
4117                 }
4118                 ext4_mark_inode_dirty(handle, inode);
4119                 ext4_truncate_restart_trans(handle, inode,
4120                                             blocks_for_truncate(inode));
4121                 if (bh) {
4122                         BUFFER_TRACE(bh, "retaking write access");
4123                         ext4_journal_get_write_access(handle, bh);
4124                 }
4125         }
4126
4127         for (p = first; p < last; p++)
4128                 *p = 0;
4129
4130         ext4_free_blocks(handle, inode, 0, block_to_free, count, flags);
4131 }
4132
4133 /**
4134  * ext4_free_data - free a list of data blocks
4135  * @handle:     handle for this transaction
4136  * @inode:      inode we are dealing with
4137  * @this_bh:    indirect buffer_head which contains *@first and *@last
4138  * @first:      array of block numbers
4139  * @last:       points immediately past the end of array
4140  *
4141  * We are freeing all blocks refered from that array (numbers are stored as
4142  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
4143  *
4144  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
4145  * blocks are contiguous then releasing them at one time will only affect one
4146  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
4147  * actually use a lot of journal space.
4148  *
4149  * @this_bh will be %NULL if @first and @last point into the inode's direct
4150  * block pointers.
4151  */
4152 static void ext4_free_data(handle_t *handle, struct inode *inode,
4153                            struct buffer_head *this_bh,
4154                            __le32 *first, __le32 *last)
4155 {
4156         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
4157         unsigned long count = 0;            /* Number of blocks in the run */
4158         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
4159                                                corresponding to
4160                                                block_to_free */
4161         ext4_fsblk_t nr;                    /* Current block # */
4162         __le32 *p;                          /* Pointer into inode/ind
4163                                                for current block */
4164         int err;
4165
4166         if (this_bh) {                          /* For indirect block */
4167                 BUFFER_TRACE(this_bh, "get_write_access");
4168                 err = ext4_journal_get_write_access(handle, this_bh);
4169                 /* Important: if we can't update the indirect pointers
4170                  * to the blocks, we can't free them. */
4171                 if (err)
4172                         return;
4173         }
4174
4175         for (p = first; p < last; p++) {
4176                 nr = le32_to_cpu(*p);
4177                 if (nr) {
4178                         /* accumulate blocks to free if they're contiguous */
4179                         if (count == 0) {
4180                                 block_to_free = nr;
4181                                 block_to_free_p = p;
4182                                 count = 1;
4183                         } else if (nr == block_to_free + count) {
4184                                 count++;
4185                         } else {
4186                                 ext4_clear_blocks(handle, inode, this_bh,
4187                                                   block_to_free,
4188                                                   count, block_to_free_p, p);
4189                                 block_to_free = nr;
4190                                 block_to_free_p = p;
4191                                 count = 1;
4192                         }
4193                 }
4194         }
4195
4196         if (count > 0)
4197                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
4198                                   count, block_to_free_p, p);
4199
4200         if (this_bh) {
4201                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
4202
4203                 /*
4204                  * The buffer head should have an attached journal head at this
4205                  * point. However, if the data is corrupted and an indirect
4206                  * block pointed to itself, it would have been detached when
4207                  * the block was cleared. Check for this instead of OOPSing.
4208                  */
4209                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
4210                         ext4_handle_dirty_metadata(handle, inode, this_bh);
4211                 else
4212                         ext4_error(inode->i_sb, __func__,
4213                                    "circular indirect block detected, "
4214                                    "inode=%lu, block=%llu",
4215                                    inode->i_ino,
4216                                    (unsigned long long) this_bh->b_blocknr);
4217         }
4218 }
4219
4220 /**
4221  *      ext4_free_branches - free an array of branches
4222  *      @handle: JBD handle for this transaction
4223  *      @inode: inode we are dealing with
4224  *      @parent_bh: the buffer_head which contains *@first and *@last
4225  *      @first: array of block numbers
4226  *      @last:  pointer immediately past the end of array
4227  *      @depth: depth of the branches to free
4228  *
4229  *      We are freeing all blocks refered from these branches (numbers are
4230  *      stored as little-endian 32-bit) and updating @inode->i_blocks
4231  *      appropriately.
4232  */
4233 static void ext4_free_branches(handle_t *handle, struct inode *inode,
4234                                struct buffer_head *parent_bh,
4235                                __le32 *first, __le32 *last, int depth)
4236 {
4237         ext4_fsblk_t nr;
4238         __le32 *p;
4239
4240         if (ext4_handle_is_aborted(handle))
4241                 return;
4242
4243         if (depth--) {
4244                 struct buffer_head *bh;
4245                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4246                 p = last;
4247                 while (--p >= first) {
4248                         nr = le32_to_cpu(*p);
4249                         if (!nr)
4250                                 continue;               /* A hole */
4251
4252                         /* Go read the buffer for the next level down */
4253                         bh = sb_bread(inode->i_sb, nr);
4254
4255                         /*
4256                          * A read failure? Report error and clear slot
4257                          * (should be rare).
4258                          */
4259                         if (!bh) {
4260                                 ext4_error(inode->i_sb, "ext4_free_branches",
4261                                            "Read failure, inode=%lu, block=%llu",
4262                                            inode->i_ino, nr);
4263                                 continue;
4264                         }
4265
4266                         /* This zaps the entire block.  Bottom up. */
4267                         BUFFER_TRACE(bh, "free child branches");
4268                         ext4_free_branches(handle, inode, bh,
4269                                         (__le32 *) bh->b_data,
4270                                         (__le32 *) bh->b_data + addr_per_block,
4271                                         depth);
4272
4273                         /*
4274                          * We've probably journalled the indirect block several
4275                          * times during the truncate.  But it's no longer
4276                          * needed and we now drop it from the transaction via
4277                          * jbd2_journal_revoke().
4278                          *
4279                          * That's easy if it's exclusively part of this
4280                          * transaction.  But if it's part of the committing
4281                          * transaction then jbd2_journal_forget() will simply
4282                          * brelse() it.  That means that if the underlying
4283                          * block is reallocated in ext4_get_block(),
4284                          * unmap_underlying_metadata() will find this block
4285                          * and will try to get rid of it.  damn, damn.
4286                          *
4287                          * If this block has already been committed to the
4288                          * journal, a revoke record will be written.  And
4289                          * revoke records must be emitted *before* clearing
4290                          * this block's bit in the bitmaps.
4291                          */
4292                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
4293
4294                         /*
4295                          * Everything below this this pointer has been
4296                          * released.  Now let this top-of-subtree go.
4297                          *
4298                          * We want the freeing of this indirect block to be
4299                          * atomic in the journal with the updating of the
4300                          * bitmap block which owns it.  So make some room in
4301                          * the journal.
4302                          *
4303                          * We zero the parent pointer *after* freeing its
4304                          * pointee in the bitmaps, so if extend_transaction()
4305                          * for some reason fails to put the bitmap changes and
4306                          * the release into the same transaction, recovery
4307                          * will merely complain about releasing a free block,
4308                          * rather than leaking blocks.
4309                          */
4310                         if (ext4_handle_is_aborted(handle))
4311                                 return;
4312                         if (try_to_extend_transaction(handle, inode)) {
4313                                 ext4_mark_inode_dirty(handle, inode);
4314                                 ext4_truncate_restart_trans(handle, inode,
4315                                             blocks_for_truncate(inode));
4316                         }
4317
4318                         ext4_free_blocks(handle, inode, 0, nr, 1,
4319                                          EXT4_FREE_BLOCKS_METADATA);
4320
4321                         if (parent_bh) {
4322                                 /*
4323                                  * The block which we have just freed is
4324                                  * pointed to by an indirect block: journal it
4325                                  */
4326                                 BUFFER_TRACE(parent_bh, "get_write_access");
4327                                 if (!ext4_journal_get_write_access(handle,
4328                                                                    parent_bh)){
4329                                         *p = 0;
4330                                         BUFFER_TRACE(parent_bh,
4331                                         "call ext4_handle_dirty_metadata");
4332                                         ext4_handle_dirty_metadata(handle,
4333                                                                    inode,
4334                                                                    parent_bh);
4335                                 }
4336                         }
4337                 }
4338         } else {
4339                 /* We have reached the bottom of the tree. */
4340                 BUFFER_TRACE(parent_bh, "free data blocks");
4341                 ext4_free_data(handle, inode, parent_bh, first, last);
4342         }
4343 }
4344
4345 int ext4_can_truncate(struct inode *inode)
4346 {
4347         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4348                 return 0;
4349         if (S_ISREG(inode->i_mode))
4350                 return 1;
4351         if (S_ISDIR(inode->i_mode))
4352                 return 1;
4353         if (S_ISLNK(inode->i_mode))
4354                 return !ext4_inode_is_fast_symlink(inode);
4355         return 0;
4356 }
4357
4358 /*
4359  * ext4_truncate()
4360  *
4361  * We block out ext4_get_block() block instantiations across the entire
4362  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4363  * simultaneously on behalf of the same inode.
4364  *
4365  * As we work through the truncate and commmit bits of it to the journal there
4366  * is one core, guiding principle: the file's tree must always be consistent on
4367  * disk.  We must be able to restart the truncate after a crash.
4368  *
4369  * The file's tree may be transiently inconsistent in memory (although it
4370  * probably isn't), but whenever we close off and commit a journal transaction,
4371  * the contents of (the filesystem + the journal) must be consistent and
4372  * restartable.  It's pretty simple, really: bottom up, right to left (although
4373  * left-to-right works OK too).
4374  *
4375  * Note that at recovery time, journal replay occurs *before* the restart of
4376  * truncate against the orphan inode list.
4377  *
4378  * The committed inode has the new, desired i_size (which is the same as
4379  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4380  * that this inode's truncate did not complete and it will again call
4381  * ext4_truncate() to have another go.  So there will be instantiated blocks
4382  * to the right of the truncation point in a crashed ext4 filesystem.  But
4383  * that's fine - as long as they are linked from the inode, the post-crash
4384  * ext4_truncate() run will find them and release them.
4385  */
4386 void ext4_truncate(struct inode *inode)
4387 {
4388         handle_t *handle;
4389         struct ext4_inode_info *ei = EXT4_I(inode);
4390         __le32 *i_data = ei->i_data;
4391         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4392         struct address_space *mapping = inode->i_mapping;
4393         ext4_lblk_t offsets[4];
4394         Indirect chain[4];
4395         Indirect *partial;
4396         __le32 nr = 0;
4397         int n;
4398         ext4_lblk_t last_block;
4399         unsigned blocksize = inode->i_sb->s_blocksize;
4400
4401         if (!ext4_can_truncate(inode))
4402                 return;
4403
4404         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4405                 ei->i_state |= EXT4_STATE_DA_ALLOC_CLOSE;
4406
4407         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
4408                 ext4_ext_truncate(inode);
4409                 return;
4410         }
4411
4412         handle = start_transaction(inode);
4413         if (IS_ERR(handle))
4414                 return;         /* AKPM: return what? */
4415
4416         last_block = (inode->i_size + blocksize-1)
4417                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4418
4419         if (inode->i_size & (blocksize - 1))
4420                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4421                         goto out_stop;
4422
4423         n = ext4_block_to_path(inode, last_block, offsets, NULL);
4424         if (n == 0)
4425                 goto out_stop;  /* error */
4426
4427         /*
4428          * OK.  This truncate is going to happen.  We add the inode to the
4429          * orphan list, so that if this truncate spans multiple transactions,
4430          * and we crash, we will resume the truncate when the filesystem
4431          * recovers.  It also marks the inode dirty, to catch the new size.
4432          *
4433          * Implication: the file must always be in a sane, consistent
4434          * truncatable state while each transaction commits.
4435          */
4436         if (ext4_orphan_add(handle, inode))
4437                 goto out_stop;
4438
4439         /*
4440          * From here we block out all ext4_get_block() callers who want to
4441          * modify the block allocation tree.
4442          */
4443         down_write(&ei->i_data_sem);
4444
4445         ext4_discard_preallocations(inode);
4446
4447         /*
4448          * The orphan list entry will now protect us from any crash which
4449          * occurs before the truncate completes, so it is now safe to propagate
4450          * the new, shorter inode size (held for now in i_size) into the
4451          * on-disk inode. We do this via i_disksize, which is the value which
4452          * ext4 *really* writes onto the disk inode.
4453          */
4454         ei->i_disksize = inode->i_size;
4455
4456         if (n == 1) {           /* direct blocks */
4457                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4458                                i_data + EXT4_NDIR_BLOCKS);
4459                 goto do_indirects;
4460         }
4461
4462         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4463         /* Kill the top of shared branch (not detached) */
4464         if (nr) {
4465                 if (partial == chain) {
4466                         /* Shared branch grows from the inode */
4467                         ext4_free_branches(handle, inode, NULL,
4468                                            &nr, &nr+1, (chain+n-1) - partial);
4469                         *partial->p = 0;
4470                         /*
4471                          * We mark the inode dirty prior to restart,
4472                          * and prior to stop.  No need for it here.
4473                          */
4474                 } else {
4475                         /* Shared branch grows from an indirect block */
4476                         BUFFER_TRACE(partial->bh, "get_write_access");
4477                         ext4_free_branches(handle, inode, partial->bh,
4478                                         partial->p,
4479                                         partial->p+1, (chain+n-1) - partial);
4480                 }
4481         }
4482         /* Clear the ends of indirect blocks on the shared branch */
4483         while (partial > chain) {
4484                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4485                                    (__le32*)partial->bh->b_data+addr_per_block,
4486                                    (chain+n-1) - partial);
4487                 BUFFER_TRACE(partial->bh, "call brelse");
4488                 brelse(partial->bh);
4489                 partial--;
4490         }
4491 do_indirects:
4492         /* Kill the remaining (whole) subtrees */
4493         switch (offsets[0]) {
4494         default:
4495                 nr = i_data[EXT4_IND_BLOCK];
4496                 if (nr) {
4497                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4498                         i_data[EXT4_IND_BLOCK] = 0;
4499                 }
4500         case EXT4_IND_BLOCK:
4501                 nr = i_data[EXT4_DIND_BLOCK];
4502                 if (nr) {
4503                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4504                         i_data[EXT4_DIND_BLOCK] = 0;
4505                 }
4506         case EXT4_DIND_BLOCK:
4507                 nr = i_data[EXT4_TIND_BLOCK];
4508                 if (nr) {
4509                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4510                         i_data[EXT4_TIND_BLOCK] = 0;
4511                 }
4512         case EXT4_TIND_BLOCK:
4513                 ;
4514         }
4515
4516         up_write(&ei->i_data_sem);
4517         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4518         ext4_mark_inode_dirty(handle, inode);
4519
4520         /*
4521          * In a multi-transaction truncate, we only make the final transaction
4522          * synchronous
4523          */
4524         if (IS_SYNC(inode))
4525                 ext4_handle_sync(handle);
4526 out_stop:
4527         /*
4528          * If this was a simple ftruncate(), and the file will remain alive
4529          * then we need to clear up the orphan record which we created above.
4530          * However, if this was a real unlink then we were called by
4531          * ext4_delete_inode(), and we allow that function to clean up the
4532          * orphan info for us.
4533          */
4534         if (inode->i_nlink)
4535                 ext4_orphan_del(handle, inode);
4536
4537         ext4_journal_stop(handle);
4538 }
4539
4540 /*
4541  * ext4_get_inode_loc returns with an extra refcount against the inode's
4542  * underlying buffer_head on success. If 'in_mem' is true, we have all
4543  * data in memory that is needed to recreate the on-disk version of this
4544  * inode.
4545  */
4546 static int __ext4_get_inode_loc(struct inode *inode,
4547                                 struct ext4_iloc *iloc, int in_mem)
4548 {
4549         struct ext4_group_desc  *gdp;
4550         struct buffer_head      *bh;
4551         struct super_block      *sb = inode->i_sb;
4552         ext4_fsblk_t            block;
4553         int                     inodes_per_block, inode_offset;
4554
4555         iloc->bh = NULL;
4556         if (!ext4_valid_inum(sb, inode->i_ino))
4557                 return -EIO;
4558
4559         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4560         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4561         if (!gdp)
4562                 return -EIO;
4563
4564         /*
4565          * Figure out the offset within the block group inode table
4566          */
4567         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4568         inode_offset = ((inode->i_ino - 1) %
4569                         EXT4_INODES_PER_GROUP(sb));
4570         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4571         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4572
4573         bh = sb_getblk(sb, block);
4574         if (!bh) {
4575                 ext4_error(sb, "ext4_get_inode_loc", "unable to read "
4576                            "inode block - inode=%lu, block=%llu",
4577                            inode->i_ino, block);
4578                 return -EIO;
4579         }
4580         if (!buffer_uptodate(bh)) {
4581                 lock_buffer(bh);
4582
4583                 /*
4584                  * If the buffer has the write error flag, we have failed
4585                  * to write out another inode in the same block.  In this
4586                  * case, we don't have to read the block because we may
4587                  * read the old inode data successfully.
4588                  */
4589                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4590                         set_buffer_uptodate(bh);
4591
4592                 if (buffer_uptodate(bh)) {
4593                         /* someone brought it uptodate while we waited */
4594                         unlock_buffer(bh);
4595                         goto has_buffer;
4596                 }
4597
4598                 /*
4599                  * If we have all information of the inode in memory and this
4600                  * is the only valid inode in the block, we need not read the
4601                  * block.
4602                  */
4603                 if (in_mem) {
4604                         struct buffer_head *bitmap_bh;
4605                         int i, start;
4606
4607                         start = inode_offset & ~(inodes_per_block - 1);
4608
4609                         /* Is the inode bitmap in cache? */
4610                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4611                         if (!bitmap_bh)
4612                                 goto make_io;
4613
4614                         /*
4615                          * If the inode bitmap isn't in cache then the
4616                          * optimisation may end up performing two reads instead
4617                          * of one, so skip it.
4618                          */
4619                         if (!buffer_uptodate(bitmap_bh)) {
4620                                 brelse(bitmap_bh);
4621                                 goto make_io;
4622                         }
4623                         for (i = start; i < start + inodes_per_block; i++) {
4624                                 if (i == inode_offset)
4625                                         continue;
4626                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4627                                         break;
4628                         }
4629                         brelse(bitmap_bh);
4630                         if (i == start + inodes_per_block) {
4631                                 /* all other inodes are free, so skip I/O */
4632                                 memset(bh->b_data, 0, bh->b_size);
4633                                 set_buffer_uptodate(bh);
4634                                 unlock_buffer(bh);
4635                                 goto has_buffer;
4636                         }
4637                 }
4638
4639 make_io:
4640                 /*
4641                  * If we need to do any I/O, try to pre-readahead extra
4642                  * blocks from the inode table.
4643                  */
4644                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4645                         ext4_fsblk_t b, end, table;
4646                         unsigned num;
4647
4648                         table = ext4_inode_table(sb, gdp);
4649                         /* s_inode_readahead_blks is always a power of 2 */
4650                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4651                         if (table > b)
4652                                 b = table;
4653                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4654                         num = EXT4_INODES_PER_GROUP(sb);
4655                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4656                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4657                                 num -= ext4_itable_unused_count(sb, gdp);
4658                         table += num / inodes_per_block;
4659                         if (end > table)
4660                                 end = table;
4661                         while (b <= end)
4662                                 sb_breadahead(sb, b++);
4663                 }
4664
4665                 /*
4666                  * There are other valid inodes in the buffer, this inode
4667                  * has in-inode xattrs, or we don't have this inode in memory.
4668                  * Read the block from disk.
4669                  */
4670                 get_bh(bh);
4671                 bh->b_end_io = end_buffer_read_sync;
4672                 submit_bh(READ_META, bh);
4673                 wait_on_buffer(bh);
4674                 if (!buffer_uptodate(bh)) {
4675                         ext4_error(sb, __func__,
4676                                    "unable to read inode block - inode=%lu, "
4677                                    "block=%llu", inode->i_ino, block);
4678                         brelse(bh);
4679                         return -EIO;
4680                 }
4681         }
4682 has_buffer:
4683         iloc->bh = bh;
4684         return 0;
4685 }
4686
4687 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4688 {
4689         /* We have all inode data except xattrs in memory here. */
4690         return __ext4_get_inode_loc(inode, iloc,
4691                 !(EXT4_I(inode)->i_state & EXT4_STATE_XATTR));
4692 }
4693
4694 void ext4_set_inode_flags(struct inode *inode)
4695 {
4696         unsigned int flags = EXT4_I(inode)->i_flags;
4697
4698         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4699         if (flags & EXT4_SYNC_FL)
4700                 inode->i_flags |= S_SYNC;
4701         if (flags & EXT4_APPEND_FL)
4702                 inode->i_flags |= S_APPEND;
4703         if (flags & EXT4_IMMUTABLE_FL)
4704                 inode->i_flags |= S_IMMUTABLE;
4705         if (flags & EXT4_NOATIME_FL)
4706                 inode->i_flags |= S_NOATIME;
4707         if (flags & EXT4_DIRSYNC_FL)
4708                 inode->i_flags |= S_DIRSYNC;
4709 }
4710
4711 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4712 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4713 {
4714         unsigned int flags = ei->vfs_inode.i_flags;
4715
4716         ei->i_flags &= ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4717                         EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|EXT4_DIRSYNC_FL);
4718         if (flags & S_SYNC)
4719                 ei->i_flags |= EXT4_SYNC_FL;
4720         if (flags & S_APPEND)
4721                 ei->i_flags |= EXT4_APPEND_FL;
4722         if (flags & S_IMMUTABLE)
4723                 ei->i_flags |= EXT4_IMMUTABLE_FL;
4724         if (flags & S_NOATIME)
4725                 ei->i_flags |= EXT4_NOATIME_FL;
4726         if (flags & S_DIRSYNC)
4727                 ei->i_flags |= EXT4_DIRSYNC_FL;
4728 }
4729
4730 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4731                                   struct ext4_inode_info *ei)
4732 {
4733         blkcnt_t i_blocks ;
4734         struct inode *inode = &(ei->vfs_inode);
4735         struct super_block *sb = inode->i_sb;
4736
4737         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4738                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4739                 /* we are using combined 48 bit field */
4740                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4741                                         le32_to_cpu(raw_inode->i_blocks_lo);
4742                 if (ei->i_flags & EXT4_HUGE_FILE_FL) {
4743                         /* i_blocks represent file system block size */
4744                         return i_blocks  << (inode->i_blkbits - 9);
4745                 } else {
4746                         return i_blocks;
4747                 }
4748         } else {
4749                 return le32_to_cpu(raw_inode->i_blocks_lo);
4750         }
4751 }
4752
4753 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4754 {
4755         struct ext4_iloc iloc;
4756         struct ext4_inode *raw_inode;
4757         struct ext4_inode_info *ei;
4758         struct inode *inode;
4759         journal_t *journal = EXT4_SB(sb)->s_journal;
4760         long ret;
4761         int block;
4762
4763         inode = iget_locked(sb, ino);
4764         if (!inode)
4765                 return ERR_PTR(-ENOMEM);
4766         if (!(inode->i_state & I_NEW))
4767                 return inode;
4768
4769         ei = EXT4_I(inode);
4770         iloc.bh = 0;
4771
4772         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4773         if (ret < 0)
4774                 goto bad_inode;
4775         raw_inode = ext4_raw_inode(&iloc);
4776         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4777         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4778         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4779         if (!(test_opt(inode->i_sb, NO_UID32))) {
4780                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4781                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4782         }
4783         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
4784
4785         ei->i_state = 0;
4786         ei->i_dir_start_lookup = 0;
4787         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4788         /* We now have enough fields to check if the inode was active or not.
4789          * This is needed because nfsd might try to access dead inodes
4790          * the test is that same one that e2fsck uses
4791          * NeilBrown 1999oct15
4792          */
4793         if (inode->i_nlink == 0) {
4794                 if (inode->i_mode == 0 ||
4795                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
4796                         /* this inode is deleted */
4797                         ret = -ESTALE;
4798                         goto bad_inode;
4799                 }
4800                 /* The only unlinked inodes we let through here have
4801                  * valid i_mode and are being read by the orphan
4802                  * recovery code: that's fine, we're about to complete
4803                  * the process of deleting those. */
4804         }
4805         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4806         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4807         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4808         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
4809                 ei->i_file_acl |=
4810                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4811         inode->i_size = ext4_isize(raw_inode);
4812         ei->i_disksize = inode->i_size;
4813 #ifdef CONFIG_QUOTA
4814         ei->i_reserved_quota = 0;
4815 #endif
4816         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4817         ei->i_block_group = iloc.block_group;
4818         ei->i_last_alloc_group = ~0;
4819         /*
4820          * NOTE! The in-memory inode i_data array is in little-endian order
4821          * even on big-endian machines: we do NOT byteswap the block numbers!
4822          */
4823         for (block = 0; block < EXT4_N_BLOCKS; block++)
4824                 ei->i_data[block] = raw_inode->i_block[block];
4825         INIT_LIST_HEAD(&ei->i_orphan);
4826
4827         /*
4828          * Set transaction id's of transactions that have to be committed
4829          * to finish f[data]sync. We set them to currently running transaction
4830          * as we cannot be sure that the inode or some of its metadata isn't
4831          * part of the transaction - the inode could have been reclaimed and
4832          * now it is reread from disk.
4833          */
4834         if (journal) {
4835                 transaction_t *transaction;
4836                 tid_t tid;
4837
4838                 spin_lock(&journal->j_state_lock);
4839                 if (journal->j_running_transaction)
4840                         transaction = journal->j_running_transaction;
4841                 else
4842                         transaction = journal->j_committing_transaction;
4843                 if (transaction)
4844                         tid = transaction->t_tid;
4845                 else
4846                         tid = journal->j_commit_sequence;
4847                 spin_unlock(&journal->j_state_lock);
4848                 ei->i_sync_tid = tid;
4849                 ei->i_datasync_tid = tid;
4850         }
4851
4852         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4853                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4854                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4855                     EXT4_INODE_SIZE(inode->i_sb)) {
4856                         ret = -EIO;
4857                         goto bad_inode;
4858                 }
4859                 if (ei->i_extra_isize == 0) {
4860                         /* The extra space is currently unused. Use it. */
4861                         ei->i_extra_isize = sizeof(struct ext4_inode) -
4862                                             EXT4_GOOD_OLD_INODE_SIZE;
4863                 } else {
4864                         __le32 *magic = (void *)raw_inode +
4865                                         EXT4_GOOD_OLD_INODE_SIZE +
4866                                         ei->i_extra_isize;
4867                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
4868                                 ei->i_state |= EXT4_STATE_XATTR;
4869                 }
4870         } else
4871                 ei->i_extra_isize = 0;
4872
4873         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4874         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4875         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4876         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4877
4878         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4879         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4880                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4881                         inode->i_version |=
4882                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4883         }
4884
4885         ret = 0;
4886         if (ei->i_file_acl &&
4887             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
4888                 ext4_error(sb, __func__,
4889                            "bad extended attribute block %llu in inode #%lu",
4890                            ei->i_file_acl, inode->i_ino);
4891                 ret = -EIO;
4892                 goto bad_inode;
4893         } else if (ei->i_flags & EXT4_EXTENTS_FL) {
4894                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4895                     (S_ISLNK(inode->i_mode) &&
4896                      !ext4_inode_is_fast_symlink(inode)))
4897                         /* Validate extent which is part of inode */
4898                         ret = ext4_ext_check_inode(inode);
4899         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4900                    (S_ISLNK(inode->i_mode) &&
4901                     !ext4_inode_is_fast_symlink(inode))) {
4902                 /* Validate block references which are part of inode */
4903                 ret = ext4_check_inode_blockref(inode);
4904         }
4905         if (ret)
4906                 goto bad_inode;
4907
4908         if (S_ISREG(inode->i_mode)) {
4909                 inode->i_op = &ext4_file_inode_operations;
4910                 inode->i_fop = &ext4_file_operations;
4911                 ext4_set_aops(inode);
4912         } else if (S_ISDIR(inode->i_mode)) {
4913                 inode->i_op = &ext4_dir_inode_operations;
4914                 inode->i_fop = &ext4_dir_operations;
4915         } else if (S_ISLNK(inode->i_mode)) {
4916                 if (ext4_inode_is_fast_symlink(inode)) {
4917                         inode->i_op = &ext4_fast_symlink_inode_operations;
4918                         nd_terminate_link(ei->i_data, inode->i_size,
4919                                 sizeof(ei->i_data) - 1);
4920                 } else {
4921                         inode->i_op = &ext4_symlink_inode_operations;
4922                         ext4_set_aops(inode);
4923                 }
4924         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
4925               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
4926                 inode->i_op = &ext4_special_inode_operations;
4927                 if (raw_inode->i_block[0])
4928                         init_special_inode(inode, inode->i_mode,
4929                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4930                 else
4931                         init_special_inode(inode, inode->i_mode,
4932                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4933         } else {
4934                 ret = -EIO;
4935                 ext4_error(inode->i_sb, __func__,
4936                            "bogus i_mode (%o) for inode=%lu",
4937                            inode->i_mode, inode->i_ino);
4938                 goto bad_inode;
4939         }
4940         brelse(iloc.bh);
4941         ext4_set_inode_flags(inode);
4942         unlock_new_inode(inode);
4943         return inode;
4944
4945 bad_inode:
4946         brelse(iloc.bh);
4947         iget_failed(inode);
4948         return ERR_PTR(ret);
4949 }
4950
4951 static int ext4_inode_blocks_set(handle_t *handle,
4952                                 struct ext4_inode *raw_inode,
4953                                 struct ext4_inode_info *ei)
4954 {
4955         struct inode *inode = &(ei->vfs_inode);
4956         u64 i_blocks = inode->i_blocks;
4957         struct super_block *sb = inode->i_sb;
4958
4959         if (i_blocks <= ~0U) {
4960                 /*
4961                  * i_blocks can be represnted in a 32 bit variable
4962                  * as multiple of 512 bytes
4963                  */
4964                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4965                 raw_inode->i_blocks_high = 0;
4966                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4967                 return 0;
4968         }
4969         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
4970                 return -EFBIG;
4971
4972         if (i_blocks <= 0xffffffffffffULL) {
4973                 /*
4974                  * i_blocks can be represented in a 48 bit variable
4975                  * as multiple of 512 bytes
4976                  */
4977                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4978                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4979                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4980         } else {
4981                 ei->i_flags |= EXT4_HUGE_FILE_FL;
4982                 /* i_block is stored in file system block size */
4983                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4984                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4985                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4986         }
4987         return 0;
4988 }
4989
4990 /*
4991  * Post the struct inode info into an on-disk inode location in the
4992  * buffer-cache.  This gobbles the caller's reference to the
4993  * buffer_head in the inode location struct.
4994  *
4995  * The caller must have write access to iloc->bh.
4996  */
4997 static int ext4_do_update_inode(handle_t *handle,
4998                                 struct inode *inode,
4999                                 struct ext4_iloc *iloc)
5000 {
5001         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5002         struct ext4_inode_info *ei = EXT4_I(inode);
5003         struct buffer_head *bh = iloc->bh;
5004         int err = 0, rc, block;
5005
5006         /* For fields not not tracking in the in-memory inode,
5007          * initialise them to zero for new inodes. */
5008         if (ei->i_state & EXT4_STATE_NEW)
5009                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5010
5011         ext4_get_inode_flags(ei);
5012         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5013         if (!(test_opt(inode->i_sb, NO_UID32))) {
5014                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
5015                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
5016 /*
5017  * Fix up interoperability with old kernels. Otherwise, old inodes get
5018  * re-used with the upper 16 bits of the uid/gid intact
5019  */
5020                 if (!ei->i_dtime) {
5021                         raw_inode->i_uid_high =
5022                                 cpu_to_le16(high_16_bits(inode->i_uid));
5023                         raw_inode->i_gid_high =
5024                                 cpu_to_le16(high_16_bits(inode->i_gid));
5025                 } else {
5026                         raw_inode->i_uid_high = 0;
5027                         raw_inode->i_gid_high = 0;
5028                 }
5029         } else {
5030                 raw_inode->i_uid_low =
5031                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
5032                 raw_inode->i_gid_low =
5033                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
5034                 raw_inode->i_uid_high = 0;
5035                 raw_inode->i_gid_high = 0;
5036         }
5037         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5038
5039         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5040         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5041         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5042         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5043
5044         if (ext4_inode_blocks_set(handle, raw_inode, ei))
5045                 goto out_brelse;
5046         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5047         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
5048         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
5049             cpu_to_le32(EXT4_OS_HURD))
5050                 raw_inode->i_file_acl_high =
5051                         cpu_to_le16(ei->i_file_acl >> 32);
5052         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5053         ext4_isize_set(raw_inode, ei->i_disksize);
5054         if (ei->i_disksize > 0x7fffffffULL) {
5055                 struct super_block *sb = inode->i_sb;
5056                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
5057                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
5058                                 EXT4_SB(sb)->s_es->s_rev_level ==
5059                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
5060                         /* If this is the first large file
5061                          * created, add a flag to the superblock.
5062                          */
5063                         err = ext4_journal_get_write_access(handle,
5064                                         EXT4_SB(sb)->s_sbh);
5065                         if (err)
5066                                 goto out_brelse;
5067                         ext4_update_dynamic_rev(sb);
5068                         EXT4_SET_RO_COMPAT_FEATURE(sb,
5069                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
5070                         sb->s_dirt = 1;
5071                         ext4_handle_sync(handle);
5072                         err = ext4_handle_dirty_metadata(handle, inode,
5073                                         EXT4_SB(sb)->s_sbh);
5074                 }
5075         }
5076         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5077         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5078                 if (old_valid_dev(inode->i_rdev)) {
5079                         raw_inode->i_block[0] =
5080                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
5081                         raw_inode->i_block[1] = 0;
5082                 } else {
5083                         raw_inode->i_block[0] = 0;
5084                         raw_inode->i_block[1] =
5085                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
5086                         raw_inode->i_block[2] = 0;
5087                 }
5088         } else
5089                 for (block = 0; block < EXT4_N_BLOCKS; block++)
5090                         raw_inode->i_block[block] = ei->i_data[block];
5091
5092         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
5093         if (ei->i_extra_isize) {
5094                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5095                         raw_inode->i_version_hi =
5096                         cpu_to_le32(inode->i_version >> 32);
5097                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
5098         }
5099
5100         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5101         rc = ext4_handle_dirty_metadata(handle, inode, bh);
5102         if (!err)
5103                 err = rc;
5104         ei->i_state &= ~EXT4_STATE_NEW;
5105
5106         ext4_update_inode_fsync_trans(handle, inode, 0);
5107 out_brelse:
5108         brelse(bh);
5109         ext4_std_error(inode->i_sb, err);
5110         return err;
5111 }
5112
5113 /*
5114  * ext4_write_inode()
5115  *
5116  * We are called from a few places:
5117  *
5118  * - Within generic_file_write() for O_SYNC files.
5119  *   Here, there will be no transaction running. We wait for any running
5120  *   trasnaction to commit.
5121  *
5122  * - Within sys_sync(), kupdate and such.
5123  *   We wait on commit, if tol to.
5124  *
5125  * - Within prune_icache() (PF_MEMALLOC == true)
5126  *   Here we simply return.  We can't afford to block kswapd on the
5127  *   journal commit.
5128  *
5129  * In all cases it is actually safe for us to return without doing anything,
5130  * because the inode has been copied into a raw inode buffer in
5131  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
5132  * knfsd.
5133  *
5134  * Note that we are absolutely dependent upon all inode dirtiers doing the
5135  * right thing: they *must* call mark_inode_dirty() after dirtying info in
5136  * which we are interested.
5137  *
5138  * It would be a bug for them to not do this.  The code:
5139  *
5140  *      mark_inode_dirty(inode)
5141  *      stuff();
5142  *      inode->i_size = expr;
5143  *
5144  * is in error because a kswapd-driven write_inode() could occur while
5145  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
5146  * will no longer be on the superblock's dirty inode list.
5147  */
5148 int ext4_write_inode(struct inode *inode, int wait)
5149 {
5150         int err;
5151
5152         if (current->flags & PF_MEMALLOC)
5153                 return 0;
5154
5155         if (EXT4_SB(inode->i_sb)->s_journal) {
5156                 if (ext4_journal_current_handle()) {
5157                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5158                         dump_stack();
5159                         return -EIO;
5160                 }
5161
5162                 if (!wait)
5163                         return 0;
5164
5165                 err = ext4_force_commit(inode->i_sb);
5166         } else {
5167                 struct ext4_iloc iloc;
5168
5169                 err = ext4_get_inode_loc(inode, &iloc);
5170                 if (err)
5171                         return err;
5172                 if (wait)
5173                         sync_dirty_buffer(iloc.bh);
5174                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5175                         ext4_error(inode->i_sb, __func__,
5176                                    "IO error syncing inode, "
5177                                    "inode=%lu, block=%llu",
5178                                    inode->i_ino,
5179                                    (unsigned long long)iloc.bh->b_blocknr);
5180                         err = -EIO;
5181                 }
5182         }
5183         return err;
5184 }
5185
5186 /*
5187  * ext4_setattr()
5188  *
5189  * Called from notify_change.
5190  *
5191  * We want to trap VFS attempts to truncate the file as soon as
5192  * possible.  In particular, we want to make sure that when the VFS
5193  * shrinks i_size, we put the inode on the orphan list and modify
5194  * i_disksize immediately, so that during the subsequent flushing of
5195  * dirty pages and freeing of disk blocks, we can guarantee that any
5196  * commit will leave the blocks being flushed in an unused state on
5197  * disk.  (On recovery, the inode will get truncated and the blocks will
5198  * be freed, so we have a strong guarantee that no future commit will
5199  * leave these blocks visible to the user.)
5200  *
5201  * Another thing we have to assure is that if we are in ordered mode
5202  * and inode is still attached to the committing transaction, we must
5203  * we start writeout of all the dirty pages which are being truncated.
5204  * This way we are sure that all the data written in the previous
5205  * transaction are already on disk (truncate waits for pages under
5206  * writeback).
5207  *
5208  * Called with inode->i_mutex down.
5209  */
5210 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5211 {
5212         struct inode *inode = dentry->d_inode;
5213         int error, rc = 0;
5214         const unsigned int ia_valid = attr->ia_valid;
5215
5216         error = inode_change_ok(inode, attr);
5217         if (error)
5218                 return error;
5219
5220         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
5221                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
5222                 handle_t *handle;
5223
5224                 /* (user+group)*(old+new) structure, inode write (sb,
5225                  * inode block, ? - but truncate inode update has it) */
5226                 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
5227                                         EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
5228                 if (IS_ERR(handle)) {
5229                         error = PTR_ERR(handle);
5230                         goto err_out;
5231                 }
5232                 error = vfs_dq_transfer(inode, attr) ? -EDQUOT : 0;
5233                 if (error) {
5234                         ext4_journal_stop(handle);
5235                         return error;
5236                 }
5237                 /* Update corresponding info in inode so that everything is in
5238                  * one transaction */
5239                 if (attr->ia_valid & ATTR_UID)
5240                         inode->i_uid = attr->ia_uid;
5241                 if (attr->ia_valid & ATTR_GID)
5242                         inode->i_gid = attr->ia_gid;
5243                 error = ext4_mark_inode_dirty(handle, inode);
5244                 ext4_journal_stop(handle);
5245         }
5246
5247         if (attr->ia_valid & ATTR_SIZE) {
5248                 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)) {
5249                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5250
5251                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
5252                                 error = -EFBIG;
5253                                 goto err_out;
5254                         }
5255                 }
5256         }
5257
5258         if (S_ISREG(inode->i_mode) &&
5259             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
5260                 handle_t *handle;
5261
5262                 handle = ext4_journal_start(inode, 3);
5263                 if (IS_ERR(handle)) {
5264                         error = PTR_ERR(handle);
5265                         goto err_out;
5266                 }
5267
5268                 error = ext4_orphan_add(handle, inode);
5269                 EXT4_I(inode)->i_disksize = attr->ia_size;
5270                 rc = ext4_mark_inode_dirty(handle, inode);
5271                 if (!error)
5272                         error = rc;
5273                 ext4_journal_stop(handle);
5274
5275                 if (ext4_should_order_data(inode)) {
5276                         error = ext4_begin_ordered_truncate(inode,
5277                                                             attr->ia_size);
5278                         if (error) {
5279                                 /* Do as much error cleanup as possible */
5280                                 handle = ext4_journal_start(inode, 3);
5281                                 if (IS_ERR(handle)) {
5282                                         ext4_orphan_del(NULL, inode);
5283                                         goto err_out;
5284                                 }
5285                                 ext4_orphan_del(handle, inode);
5286                                 ext4_journal_stop(handle);
5287                                 goto err_out;
5288                         }
5289                 }
5290         }
5291
5292         rc = inode_setattr(inode, attr);
5293
5294         /* If inode_setattr's call to ext4_truncate failed to get a
5295          * transaction handle at all, we need to clean up the in-core
5296          * orphan list manually. */
5297         if (inode->i_nlink)
5298                 ext4_orphan_del(NULL, inode);
5299
5300         if (!rc && (ia_valid & ATTR_MODE))
5301                 rc = ext4_acl_chmod(inode);
5302
5303 err_out:
5304         ext4_std_error(inode->i_sb, error);
5305         if (!error)
5306                 error = rc;
5307         return error;
5308 }
5309
5310 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
5311                  struct kstat *stat)
5312 {
5313         struct inode *inode;
5314         unsigned long delalloc_blocks;
5315
5316         inode = dentry->d_inode;
5317         generic_fillattr(inode, stat);
5318
5319         /*
5320          * We can't update i_blocks if the block allocation is delayed
5321          * otherwise in the case of system crash before the real block
5322          * allocation is done, we will have i_blocks inconsistent with
5323          * on-disk file blocks.
5324          * We always keep i_blocks updated together with real
5325          * allocation. But to not confuse with user, stat
5326          * will return the blocks that include the delayed allocation
5327          * blocks for this file.
5328          */
5329         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
5330         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
5331         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
5332
5333         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
5334         return 0;
5335 }
5336
5337 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
5338                                       int chunk)
5339 {
5340         int indirects;
5341
5342         /* if nrblocks are contiguous */
5343         if (chunk) {
5344                 /*
5345                  * With N contiguous data blocks, it need at most
5346                  * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) indirect blocks
5347                  * 2 dindirect blocks
5348                  * 1 tindirect block
5349                  */
5350                 indirects = nrblocks / EXT4_ADDR_PER_BLOCK(inode->i_sb);
5351                 return indirects + 3;
5352         }
5353         /*
5354          * if nrblocks are not contiguous, worse case, each block touch
5355          * a indirect block, and each indirect block touch a double indirect
5356          * block, plus a triple indirect block
5357          */
5358         indirects = nrblocks * 2 + 1;
5359         return indirects;
5360 }
5361
5362 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5363 {
5364         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL))
5365                 return ext4_indirect_trans_blocks(inode, nrblocks, chunk);
5366         return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
5367 }
5368
5369 /*
5370  * Account for index blocks, block groups bitmaps and block group
5371  * descriptor blocks if modify datablocks and index blocks
5372  * worse case, the indexs blocks spread over different block groups
5373  *
5374  * If datablocks are discontiguous, they are possible to spread over
5375  * different block groups too. If they are contiuguous, with flexbg,
5376  * they could still across block group boundary.
5377  *
5378  * Also account for superblock, inode, quota and xattr blocks
5379  */
5380 int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5381 {
5382         ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
5383         int gdpblocks;
5384         int idxblocks;
5385         int ret = 0;
5386
5387         /*
5388          * How many index blocks need to touch to modify nrblocks?
5389          * The "Chunk" flag indicating whether the nrblocks is
5390          * physically contiguous on disk
5391          *
5392          * For Direct IO and fallocate, they calls get_block to allocate
5393          * one single extent at a time, so they could set the "Chunk" flag
5394          */
5395         idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
5396
5397         ret = idxblocks;
5398
5399         /*
5400          * Now let's see how many group bitmaps and group descriptors need
5401          * to account
5402          */
5403         groups = idxblocks;
5404         if (chunk)
5405                 groups += 1;
5406         else
5407                 groups += nrblocks;
5408
5409         gdpblocks = groups;
5410         if (groups > ngroups)
5411                 groups = ngroups;
5412         if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
5413                 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
5414
5415         /* bitmaps and block group descriptor blocks */
5416         ret += groups + gdpblocks;
5417
5418         /* Blocks for super block, inode, quota and xattr blocks */
5419         ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
5420
5421         return ret;
5422 }
5423
5424 /*
5425  * Calulate the total number of credits to reserve to fit
5426  * the modification of a single pages into a single transaction,
5427  * which may include multiple chunks of block allocations.
5428  *
5429  * This could be called via ext4_write_begin()
5430  *
5431  * We need to consider the worse case, when
5432  * one new block per extent.
5433  */
5434 int ext4_writepage_trans_blocks(struct inode *inode)
5435 {
5436         int bpp = ext4_journal_blocks_per_page(inode);
5437         int ret;
5438
5439         ret = ext4_meta_trans_blocks(inode, bpp, 0);
5440
5441         /* Account for data blocks for journalled mode */
5442         if (ext4_should_journal_data(inode))
5443                 ret += bpp;
5444         return ret;
5445 }
5446
5447 /*
5448  * Calculate the journal credits for a chunk of data modification.
5449  *
5450  * This is called from DIO, fallocate or whoever calling
5451  * ext4_get_blocks() to map/allocate a chunk of contiguous disk blocks.
5452  *
5453  * journal buffers for data blocks are not included here, as DIO
5454  * and fallocate do no need to journal data buffers.
5455  */
5456 int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
5457 {
5458         return ext4_meta_trans_blocks(inode, nrblocks, 1);
5459 }
5460
5461 /*
5462  * The caller must have previously called ext4_reserve_inode_write().
5463  * Give this, we know that the caller already has write access to iloc->bh.
5464  */
5465 int ext4_mark_iloc_dirty(handle_t *handle,
5466                          struct inode *inode, struct ext4_iloc *iloc)
5467 {
5468         int err = 0;
5469
5470         if (test_opt(inode->i_sb, I_VERSION))
5471                 inode_inc_iversion(inode);
5472
5473         /* the do_update_inode consumes one bh->b_count */
5474         get_bh(iloc->bh);
5475
5476         /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
5477         err = ext4_do_update_inode(handle, inode, iloc);
5478         put_bh(iloc->bh);
5479         return err;
5480 }
5481
5482 /*
5483  * On success, We end up with an outstanding reference count against
5484  * iloc->bh.  This _must_ be cleaned up later.
5485  */
5486
5487 int
5488 ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
5489                          struct ext4_iloc *iloc)
5490 {
5491         int err;
5492
5493         err = ext4_get_inode_loc(inode, iloc);
5494         if (!err) {
5495                 BUFFER_TRACE(iloc->bh, "get_write_access");
5496                 err = ext4_journal_get_write_access(handle, iloc->bh);
5497                 if (err) {
5498                         brelse(iloc->bh);
5499                         iloc->bh = NULL;
5500                 }
5501         }
5502         ext4_std_error(inode->i_sb, err);
5503         return err;
5504 }
5505
5506 /*
5507  * Expand an inode by new_extra_isize bytes.
5508  * Returns 0 on success or negative error number on failure.
5509  */
5510 static int ext4_expand_extra_isize(struct inode *inode,
5511                                    unsigned int new_extra_isize,
5512                                    struct ext4_iloc iloc,
5513                                    handle_t *handle)
5514 {
5515         struct ext4_inode *raw_inode;
5516         struct ext4_xattr_ibody_header *header;
5517         struct ext4_xattr_entry *entry;
5518
5519         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
5520                 return 0;
5521
5522         raw_inode = ext4_raw_inode(&iloc);
5523
5524         header = IHDR(inode, raw_inode);
5525         entry = IFIRST(header);
5526
5527         /* No extended attributes present */
5528         if (!(EXT4_I(inode)->i_state & EXT4_STATE_XATTR) ||
5529                 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
5530                 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
5531                         new_extra_isize);
5532                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
5533                 return 0;
5534         }
5535
5536         /* try to expand with EAs present */
5537         return ext4_expand_extra_isize_ea(inode, new_extra_isize,
5538                                           raw_inode, handle);
5539 }
5540
5541 /*
5542  * What we do here is to mark the in-core inode as clean with respect to inode
5543  * dirtiness (it may still be data-dirty).
5544  * This means that the in-core inode may be reaped by prune_icache
5545  * without having to perform any I/O.  This is a very good thing,
5546  * because *any* task may call prune_icache - even ones which
5547  * have a transaction open against a different journal.
5548  *
5549  * Is this cheating?  Not really.  Sure, we haven't written the
5550  * inode out, but prune_icache isn't a user-visible syncing function.
5551  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
5552  * we start and wait on commits.
5553  *
5554  * Is this efficient/effective?  Well, we're being nice to the system
5555  * by cleaning up our inodes proactively so they can be reaped
5556  * without I/O.  But we are potentially leaving up to five seconds'
5557  * worth of inodes floating about which prune_icache wants us to
5558  * write out.  One way to fix that would be to get prune_icache()
5559  * to do a write_super() to free up some memory.  It has the desired
5560  * effect.
5561  */
5562 int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
5563 {
5564         struct ext4_iloc iloc;
5565         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5566         static unsigned int mnt_count;
5567         int err, ret;
5568
5569         might_sleep();
5570         err = ext4_reserve_inode_write(handle, inode, &iloc);
5571         if (ext4_handle_valid(handle) &&
5572             EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
5573             !(EXT4_I(inode)->i_state & EXT4_STATE_NO_EXPAND)) {
5574                 /*
5575                  * We need extra buffer credits since we may write into EA block
5576                  * with this same handle. If journal_extend fails, then it will
5577                  * only result in a minor loss of functionality for that inode.
5578                  * If this is felt to be critical, then e2fsck should be run to
5579                  * force a large enough s_min_extra_isize.
5580                  */
5581                 if ((jbd2_journal_extend(handle,
5582                              EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
5583                         ret = ext4_expand_extra_isize(inode,
5584                                                       sbi->s_want_extra_isize,
5585                                                       iloc, handle);
5586                         if (ret) {
5587                                 EXT4_I(inode)->i_state |= EXT4_STATE_NO_EXPAND;
5588                                 if (mnt_count !=
5589                                         le16_to_cpu(sbi->s_es->s_mnt_count)) {
5590                                        &nbs