3ae83137cf34878784f5db6ef2a9d818381c1e84
[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 #include <linux/kernel.h>
42 #include <linux/slab.h>
43 #include <linux/ratelimit.h>
44
45 #include "ext4_jbd2.h"
46 #include "xattr.h"
47 #include "acl.h"
48 #include "ext4_extents.h"
49
50 #include <trace/events/ext4.h>
51
52 #define MPAGE_DA_EXTENT_TAIL 0x01
53
54 static inline int ext4_begin_ordered_truncate(struct inode *inode,
55                                               loff_t new_size)
56 {
57         trace_ext4_begin_ordered_truncate(inode, new_size);
58         return jbd2_journal_begin_ordered_truncate(
59                                         EXT4_SB(inode->i_sb)->s_journal,
60                                         &EXT4_I(inode)->jinode,
61                                         new_size);
62 }
63
64 static void ext4_invalidatepage(struct page *page, unsigned long offset);
65 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
66                                    struct buffer_head *bh_result, int create);
67 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
68 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
69 static int __ext4_journalled_writepage(struct page *page, unsigned int len);
70 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
71
72 /*
73  * Test whether an inode is a fast symlink.
74  */
75 static int ext4_inode_is_fast_symlink(struct inode *inode)
76 {
77         int ea_blocks = EXT4_I(inode)->i_file_acl ?
78                 (inode->i_sb->s_blocksize >> 9) : 0;
79
80         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
81 }
82
83 /*
84  * Work out how many blocks we need to proceed with the next chunk of a
85  * truncate transaction.
86  */
87 static unsigned long blocks_for_truncate(struct inode *inode)
88 {
89         ext4_lblk_t needed;
90
91         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
92
93         /* Give ourselves just enough room to cope with inodes in which
94          * i_blocks is corrupt: we've seen disk corruptions in the past
95          * which resulted in random data in an inode which looked enough
96          * like a regular file for ext4 to try to delete it.  Things
97          * will go a bit crazy if that happens, but at least we should
98          * try not to panic the whole kernel. */
99         if (needed < 2)
100                 needed = 2;
101
102         /* But we need to bound the transaction so we don't overflow the
103          * journal. */
104         if (needed > EXT4_MAX_TRANS_DATA)
105                 needed = EXT4_MAX_TRANS_DATA;
106
107         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
108 }
109
110 /*
111  * Truncate transactions can be complex and absolutely huge.  So we need to
112  * be able to restart the transaction at a conventient checkpoint to make
113  * sure we don't overflow the journal.
114  *
115  * start_transaction gets us a new handle for a truncate transaction,
116  * and extend_transaction tries to extend the existing one a bit.  If
117  * extend fails, we need to propagate the failure up and restart the
118  * transaction in the top-level truncate loop. --sct
119  */
120 static handle_t *start_transaction(struct inode *inode)
121 {
122         handle_t *result;
123
124         result = ext4_journal_start(inode, blocks_for_truncate(inode));
125         if (!IS_ERR(result))
126                 return result;
127
128         ext4_std_error(inode->i_sb, PTR_ERR(result));
129         return result;
130 }
131
132 /*
133  * Try to extend this transaction for the purposes of truncation.
134  *
135  * Returns 0 if we managed to create more room.  If we can't create more
136  * room, and the transaction must be restarted we return 1.
137  */
138 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
139 {
140         if (!ext4_handle_valid(handle))
141                 return 0;
142         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
143                 return 0;
144         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
145                 return 0;
146         return 1;
147 }
148
149 /*
150  * Restart the transaction associated with *handle.  This does a commit,
151  * so before we call here everything must be consistently dirtied against
152  * this transaction.
153  */
154 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
155                                  int nblocks)
156 {
157         int ret;
158
159         /*
160          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
161          * moment, get_block can be called only for blocks inside i_size since
162          * page cache has been already dropped and writes are blocked by
163          * i_mutex. So we can safely drop the i_data_sem here.
164          */
165         BUG_ON(EXT4_JOURNAL(inode) == NULL);
166         jbd_debug(2, "restarting handle %p\n", handle);
167         up_write(&EXT4_I(inode)->i_data_sem);
168         ret = ext4_journal_restart(handle, blocks_for_truncate(inode));
169         down_write(&EXT4_I(inode)->i_data_sem);
170         ext4_discard_preallocations(inode);
171
172         return ret;
173 }
174
175 /*
176  * Called at the last iput() if i_nlink is zero.
177  */
178 void ext4_evict_inode(struct inode *inode)
179 {
180         handle_t *handle;
181         int err;
182
183         trace_ext4_evict_inode(inode);
184         if (inode->i_nlink) {
185                 truncate_inode_pages(&inode->i_data, 0);
186                 goto no_delete;
187         }
188
189         if (!is_bad_inode(inode))
190                 dquot_initialize(inode);
191
192         if (ext4_should_order_data(inode))
193                 ext4_begin_ordered_truncate(inode, 0);
194         truncate_inode_pages(&inode->i_data, 0);
195
196         if (is_bad_inode(inode))
197                 goto no_delete;
198
199         handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
200         if (IS_ERR(handle)) {
201                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
202                 /*
203                  * If we're going to skip the normal cleanup, we still need to
204                  * make sure that the in-core orphan linked list is properly
205                  * cleaned up.
206                  */
207                 ext4_orphan_del(NULL, inode);
208                 goto no_delete;
209         }
210
211         if (IS_SYNC(inode))
212                 ext4_handle_sync(handle);
213         inode->i_size = 0;
214         err = ext4_mark_inode_dirty(handle, inode);
215         if (err) {
216                 ext4_warning(inode->i_sb,
217                              "couldn't mark inode dirty (err %d)", err);
218                 goto stop_handle;
219         }
220         if (inode->i_blocks)
221                 ext4_truncate(inode);
222
223         /*
224          * ext4_ext_truncate() doesn't reserve any slop when it
225          * restarts journal transactions; therefore there may not be
226          * enough credits left in the handle to remove the inode from
227          * the orphan list and set the dtime field.
228          */
229         if (!ext4_handle_has_enough_credits(handle, 3)) {
230                 err = ext4_journal_extend(handle, 3);
231                 if (err > 0)
232                         err = ext4_journal_restart(handle, 3);
233                 if (err != 0) {
234                         ext4_warning(inode->i_sb,
235                                      "couldn't extend journal (err %d)", err);
236                 stop_handle:
237                         ext4_journal_stop(handle);
238                         ext4_orphan_del(NULL, inode);
239                         goto no_delete;
240                 }
241         }
242
243         /*
244          * Kill off the orphan record which ext4_truncate created.
245          * AKPM: I think this can be inside the above `if'.
246          * Note that ext4_orphan_del() has to be able to cope with the
247          * deletion of a non-existent orphan - this is because we don't
248          * know if ext4_truncate() actually created an orphan record.
249          * (Well, we could do this if we need to, but heck - it works)
250          */
251         ext4_orphan_del(handle, inode);
252         EXT4_I(inode)->i_dtime  = get_seconds();
253
254         /*
255          * One subtle ordering requirement: if anything has gone wrong
256          * (transaction abort, IO errors, whatever), then we can still
257          * do these next steps (the fs will already have been marked as
258          * having errors), but we can't free the inode if the mark_dirty
259          * fails.
260          */
261         if (ext4_mark_inode_dirty(handle, inode))
262                 /* If that failed, just do the required in-core inode clear. */
263                 ext4_clear_inode(inode);
264         else
265                 ext4_free_inode(handle, inode);
266         ext4_journal_stop(handle);
267         return;
268 no_delete:
269         ext4_clear_inode(inode);        /* We must guarantee clearing of inode... */
270 }
271
272 typedef struct {
273         __le32  *p;
274         __le32  key;
275         struct buffer_head *bh;
276 } Indirect;
277
278 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
279 {
280         p->key = *(p->p = v);
281         p->bh = bh;
282 }
283
284 /**
285  *      ext4_block_to_path - parse the block number into array of offsets
286  *      @inode: inode in question (we are only interested in its superblock)
287  *      @i_block: block number to be parsed
288  *      @offsets: array to store the offsets in
289  *      @boundary: set this non-zero if the referred-to block is likely to be
290  *             followed (on disk) by an indirect block.
291  *
292  *      To store the locations of file's data ext4 uses a data structure common
293  *      for UNIX filesystems - tree of pointers anchored in the inode, with
294  *      data blocks at leaves and indirect blocks in intermediate nodes.
295  *      This function translates the block number into path in that tree -
296  *      return value is the path length and @offsets[n] is the offset of
297  *      pointer to (n+1)th node in the nth one. If @block is out of range
298  *      (negative or too large) warning is printed and zero returned.
299  *
300  *      Note: function doesn't find node addresses, so no IO is needed. All
301  *      we need to know is the capacity of indirect blocks (taken from the
302  *      inode->i_sb).
303  */
304
305 /*
306  * Portability note: the last comparison (check that we fit into triple
307  * indirect block) is spelled differently, because otherwise on an
308  * architecture with 32-bit longs and 8Kb pages we might get into trouble
309  * if our filesystem had 8Kb blocks. We might use long long, but that would
310  * kill us on x86. Oh, well, at least the sign propagation does not matter -
311  * i_block would have to be negative in the very beginning, so we would not
312  * get there at all.
313  */
314
315 static int ext4_block_to_path(struct inode *inode,
316                               ext4_lblk_t i_block,
317                               ext4_lblk_t offsets[4], int *boundary)
318 {
319         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
320         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
321         const long direct_blocks = EXT4_NDIR_BLOCKS,
322                 indirect_blocks = ptrs,
323                 double_blocks = (1 << (ptrs_bits * 2));
324         int n = 0;
325         int final = 0;
326
327         if (i_block < direct_blocks) {
328                 offsets[n++] = i_block;
329                 final = direct_blocks;
330         } else if ((i_block -= direct_blocks) < indirect_blocks) {
331                 offsets[n++] = EXT4_IND_BLOCK;
332                 offsets[n++] = i_block;
333                 final = ptrs;
334         } else if ((i_block -= indirect_blocks) < double_blocks) {
335                 offsets[n++] = EXT4_DIND_BLOCK;
336                 offsets[n++] = i_block >> ptrs_bits;
337                 offsets[n++] = i_block & (ptrs - 1);
338                 final = ptrs;
339         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
340                 offsets[n++] = EXT4_TIND_BLOCK;
341                 offsets[n++] = i_block >> (ptrs_bits * 2);
342                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
343                 offsets[n++] = i_block & (ptrs - 1);
344                 final = ptrs;
345         } else {
346                 ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
347                              i_block + direct_blocks +
348                              indirect_blocks + double_blocks, inode->i_ino);
349         }
350         if (boundary)
351                 *boundary = final - 1 - (i_block & (ptrs - 1));
352         return n;
353 }
354
355 static int __ext4_check_blockref(const char *function, unsigned int line,
356                                  struct inode *inode,
357                                  __le32 *p, unsigned int max)
358 {
359         struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
360         __le32 *bref = p;
361         unsigned int blk;
362
363         while (bref < p+max) {
364                 blk = le32_to_cpu(*bref++);
365                 if (blk &&
366                     unlikely(!ext4_data_block_valid(EXT4_SB(inode->i_sb),
367                                                     blk, 1))) {
368                         es->s_last_error_block = cpu_to_le64(blk);
369                         ext4_error_inode(inode, function, line, blk,
370                                          "invalid block");
371                         return -EIO;
372                 }
373         }
374         return 0;
375 }
376
377
378 #define ext4_check_indirect_blockref(inode, bh)                         \
379         __ext4_check_blockref(__func__, __LINE__, inode,                \
380                               (__le32 *)(bh)->b_data,                   \
381                               EXT4_ADDR_PER_BLOCK((inode)->i_sb))
382
383 #define ext4_check_inode_blockref(inode)                                \
384         __ext4_check_blockref(__func__, __LINE__, inode,                \
385                               EXT4_I(inode)->i_data,                    \
386                               EXT4_NDIR_BLOCKS)
387
388 /**
389  *      ext4_get_branch - read the chain of indirect blocks leading to data
390  *      @inode: inode in question
391  *      @depth: depth of the chain (1 - direct pointer, etc.)
392  *      @offsets: offsets of pointers in inode/indirect blocks
393  *      @chain: place to store the result
394  *      @err: here we store the error value
395  *
396  *      Function fills the array of triples <key, p, bh> and returns %NULL
397  *      if everything went OK or the pointer to the last filled triple
398  *      (incomplete one) otherwise. Upon the return chain[i].key contains
399  *      the number of (i+1)-th block in the chain (as it is stored in memory,
400  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
401  *      number (it points into struct inode for i==0 and into the bh->b_data
402  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
403  *      block for i>0 and NULL for i==0. In other words, it holds the block
404  *      numbers of the chain, addresses they were taken from (and where we can
405  *      verify that chain did not change) and buffer_heads hosting these
406  *      numbers.
407  *
408  *      Function stops when it stumbles upon zero pointer (absent block)
409  *              (pointer to last triple returned, *@err == 0)
410  *      or when it gets an IO error reading an indirect block
411  *              (ditto, *@err == -EIO)
412  *      or when it reads all @depth-1 indirect blocks successfully and finds
413  *      the whole chain, all way to the data (returns %NULL, *err == 0).
414  *
415  *      Need to be called with
416  *      down_read(&EXT4_I(inode)->i_data_sem)
417  */
418 static Indirect *ext4_get_branch(struct inode *inode, int depth,
419                                  ext4_lblk_t  *offsets,
420                                  Indirect chain[4], int *err)
421 {
422         struct super_block *sb = inode->i_sb;
423         Indirect *p = chain;
424         struct buffer_head *bh;
425
426         *err = 0;
427         /* i_data is not going away, no lock needed */
428         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
429         if (!p->key)
430                 goto no_block;
431         while (--depth) {
432                 bh = sb_getblk(sb, le32_to_cpu(p->key));
433                 if (unlikely(!bh))
434                         goto failure;
435
436                 if (!bh_uptodate_or_lock(bh)) {
437                         if (bh_submit_read(bh) < 0) {
438                                 put_bh(bh);
439                                 goto failure;
440                         }
441                         /* validate block references */
442                         if (ext4_check_indirect_blockref(inode, bh)) {
443                                 put_bh(bh);
444                                 goto failure;
445                         }
446                 }
447
448                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
449                 /* Reader: end */
450                 if (!p->key)
451                         goto no_block;
452         }
453         return NULL;
454
455 failure:
456         *err = -EIO;
457 no_block:
458         return p;
459 }
460
461 /**
462  *      ext4_find_near - find a place for allocation with sufficient locality
463  *      @inode: owner
464  *      @ind: descriptor of indirect block.
465  *
466  *      This function returns the preferred place for block allocation.
467  *      It is used when heuristic for sequential allocation fails.
468  *      Rules are:
469  *        + if there is a block to the left of our position - allocate near it.
470  *        + if pointer will live in indirect block - allocate near that block.
471  *        + if pointer will live in inode - allocate in the same
472  *          cylinder group.
473  *
474  * In the latter case we colour the starting block by the callers PID to
475  * prevent it from clashing with concurrent allocations for a different inode
476  * in the same block group.   The PID is used here so that functionally related
477  * files will be close-by on-disk.
478  *
479  *      Caller must make sure that @ind is valid and will stay that way.
480  */
481 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
482 {
483         struct ext4_inode_info *ei = EXT4_I(inode);
484         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
485         __le32 *p;
486         ext4_fsblk_t bg_start;
487         ext4_fsblk_t last_block;
488         ext4_grpblk_t colour;
489         ext4_group_t block_group;
490         int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
491
492         /* Try to find previous block */
493         for (p = ind->p - 1; p >= start; p--) {
494                 if (*p)
495                         return le32_to_cpu(*p);
496         }
497
498         /* No such thing, so let's try location of indirect block */
499         if (ind->bh)
500                 return ind->bh->b_blocknr;
501
502         /*
503          * It is going to be referred to from the inode itself? OK, just put it
504          * into the same cylinder group then.
505          */
506         block_group = ei->i_block_group;
507         if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
508                 block_group &= ~(flex_size-1);
509                 if (S_ISREG(inode->i_mode))
510                         block_group++;
511         }
512         bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
513         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
514
515         /*
516          * If we are doing delayed allocation, we don't need take
517          * colour into account.
518          */
519         if (test_opt(inode->i_sb, DELALLOC))
520                 return bg_start;
521
522         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
523                 colour = (current->pid % 16) *
524                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
525         else
526                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
527         return bg_start + colour;
528 }
529
530 /**
531  *      ext4_find_goal - find a preferred place for allocation.
532  *      @inode: owner
533  *      @block:  block we want
534  *      @partial: pointer to the last triple within a chain
535  *
536  *      Normally this function find the preferred place for block allocation,
537  *      returns it.
538  *      Because this is only used for non-extent files, we limit the block nr
539  *      to 32 bits.
540  */
541 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
542                                    Indirect *partial)
543 {
544         ext4_fsblk_t goal;
545
546         /*
547          * XXX need to get goal block from mballoc's data structures
548          */
549
550         goal = ext4_find_near(inode, partial);
551         goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
552         return goal;
553 }
554
555 /**
556  *      ext4_blks_to_allocate - Look up the block map and count the number
557  *      of direct blocks need to be allocated for the given branch.
558  *
559  *      @branch: chain of indirect blocks
560  *      @k: number of blocks need for indirect blocks
561  *      @blks: number of data blocks to be mapped.
562  *      @blocks_to_boundary:  the offset in the indirect block
563  *
564  *      return the total number of blocks to be allocate, including the
565  *      direct and indirect blocks.
566  */
567 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
568                                  int blocks_to_boundary)
569 {
570         unsigned int count = 0;
571
572         /*
573          * Simple case, [t,d]Indirect block(s) has not allocated yet
574          * then it's clear blocks on that path have not allocated
575          */
576         if (k > 0) {
577                 /* right now we don't handle cross boundary allocation */
578                 if (blks < blocks_to_boundary + 1)
579                         count += blks;
580                 else
581                         count += blocks_to_boundary + 1;
582                 return count;
583         }
584
585         count++;
586         while (count < blks && count <= blocks_to_boundary &&
587                 le32_to_cpu(*(branch[0].p + count)) == 0) {
588                 count++;
589         }
590         return count;
591 }
592
593 /**
594  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
595  *      @handle: handle for this transaction
596  *      @inode: inode which needs allocated blocks
597  *      @iblock: the logical block to start allocated at
598  *      @goal: preferred physical block of allocation
599  *      @indirect_blks: the number of blocks need to allocate for indirect
600  *                      blocks
601  *      @blks: number of desired blocks
602  *      @new_blocks: on return it will store the new block numbers for
603  *      the indirect blocks(if needed) and the first direct block,
604  *      @err: on return it will store the error code
605  *
606  *      This function will return the number of blocks allocated as
607  *      requested by the passed-in parameters.
608  */
609 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
610                              ext4_lblk_t iblock, ext4_fsblk_t goal,
611                              int indirect_blks, int blks,
612                              ext4_fsblk_t new_blocks[4], int *err)
613 {
614         struct ext4_allocation_request ar;
615         int target, i;
616         unsigned long count = 0, blk_allocated = 0;
617         int index = 0;
618         ext4_fsblk_t current_block = 0;
619         int ret = 0;
620
621         /*
622          * Here we try to allocate the requested multiple blocks at once,
623          * on a best-effort basis.
624          * To build a branch, we should allocate blocks for
625          * the indirect blocks(if not allocated yet), and at least
626          * the first direct block of this branch.  That's the
627          * minimum number of blocks need to allocate(required)
628          */
629         /* first we try to allocate the indirect blocks */
630         target = indirect_blks;
631         while (target > 0) {
632                 count = target;
633                 /* allocating blocks for indirect blocks and direct blocks */
634                 current_block = ext4_new_meta_blocks(handle, inode,
635                                                         goal, &count, err);
636                 if (*err)
637                         goto failed_out;
638
639                 if (unlikely(current_block + count > EXT4_MAX_BLOCK_FILE_PHYS)) {
640                         EXT4_ERROR_INODE(inode,
641                                          "current_block %llu + count %lu > %d!",
642                                          current_block, count,
643                                          EXT4_MAX_BLOCK_FILE_PHYS);
644                         *err = -EIO;
645                         goto failed_out;
646                 }
647
648                 target -= count;
649                 /* allocate blocks for indirect blocks */
650                 while (index < indirect_blks && count) {
651                         new_blocks[index++] = current_block++;
652                         count--;
653                 }
654                 if (count > 0) {
655                         /*
656                          * save the new block number
657                          * for the first direct block
658                          */
659                         new_blocks[index] = current_block;
660                         printk(KERN_INFO "%s returned more blocks than "
661                                                 "requested\n", __func__);
662                         WARN_ON(1);
663                         break;
664                 }
665         }
666
667         target = blks - count ;
668         blk_allocated = count;
669         if (!target)
670                 goto allocated;
671         /* Now allocate data blocks */
672         memset(&ar, 0, sizeof(ar));
673         ar.inode = inode;
674         ar.goal = goal;
675         ar.len = target;
676         ar.logical = iblock;
677         if (S_ISREG(inode->i_mode))
678                 /* enable in-core preallocation only for regular files */
679                 ar.flags = EXT4_MB_HINT_DATA;
680
681         current_block = ext4_mb_new_blocks(handle, &ar, err);
682         if (unlikely(current_block + ar.len > EXT4_MAX_BLOCK_FILE_PHYS)) {
683                 EXT4_ERROR_INODE(inode,
684                                  "current_block %llu + ar.len %d > %d!",
685                                  current_block, ar.len,
686                                  EXT4_MAX_BLOCK_FILE_PHYS);
687                 *err = -EIO;
688                 goto failed_out;
689         }
690
691         if (*err && (target == blks)) {
692                 /*
693                  * if the allocation failed and we didn't allocate
694                  * any blocks before
695                  */
696                 goto failed_out;
697         }
698         if (!*err) {
699                 if (target == blks) {
700                         /*
701                          * save the new block number
702                          * for the first direct block
703                          */
704                         new_blocks[index] = current_block;
705                 }
706                 blk_allocated += ar.len;
707         }
708 allocated:
709         /* total number of blocks allocated for direct blocks */
710         ret = blk_allocated;
711         *err = 0;
712         return ret;
713 failed_out:
714         for (i = 0; i < index; i++)
715                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
716         return ret;
717 }
718
719 /**
720  *      ext4_alloc_branch - allocate and set up a chain of blocks.
721  *      @handle: handle for this transaction
722  *      @inode: owner
723  *      @indirect_blks: number of allocated indirect blocks
724  *      @blks: number of allocated direct blocks
725  *      @goal: preferred place for allocation
726  *      @offsets: offsets (in the blocks) to store the pointers to next.
727  *      @branch: place to store the chain in.
728  *
729  *      This function allocates blocks, zeroes out all but the last one,
730  *      links them into chain and (if we are synchronous) writes them to disk.
731  *      In other words, it prepares a branch that can be spliced onto the
732  *      inode. It stores the information about that chain in the branch[], in
733  *      the same format as ext4_get_branch() would do. We are calling it after
734  *      we had read the existing part of chain and partial points to the last
735  *      triple of that (one with zero ->key). Upon the exit we have the same
736  *      picture as after the successful ext4_get_block(), except that in one
737  *      place chain is disconnected - *branch->p is still zero (we did not
738  *      set the last link), but branch->key contains the number that should
739  *      be placed into *branch->p to fill that gap.
740  *
741  *      If allocation fails we free all blocks we've allocated (and forget
742  *      their buffer_heads) and return the error value the from failed
743  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
744  *      as described above and return 0.
745  */
746 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
747                              ext4_lblk_t iblock, int indirect_blks,
748                              int *blks, ext4_fsblk_t goal,
749                              ext4_lblk_t *offsets, Indirect *branch)
750 {
751         int blocksize = inode->i_sb->s_blocksize;
752         int i, n = 0;
753         int err = 0;
754         struct buffer_head *bh;
755         int num;
756         ext4_fsblk_t new_blocks[4];
757         ext4_fsblk_t current_block;
758
759         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
760                                 *blks, new_blocks, &err);
761         if (err)
762                 return err;
763
764         branch[0].key = cpu_to_le32(new_blocks[0]);
765         /*
766          * metadata blocks and data blocks are allocated.
767          */
768         for (n = 1; n <= indirect_blks;  n++) {
769                 /*
770                  * Get buffer_head for parent block, zero it out
771                  * and set the pointer to new one, then send
772                  * parent to disk.
773                  */
774                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
775                 if (unlikely(!bh)) {
776                         err = -EIO;
777                         goto failed;
778                 }
779
780                 branch[n].bh = bh;
781                 lock_buffer(bh);
782                 BUFFER_TRACE(bh, "call get_create_access");
783                 err = ext4_journal_get_create_access(handle, bh);
784                 if (err) {
785                         /* Don't brelse(bh) here; it's done in
786                          * ext4_journal_forget() below */
787                         unlock_buffer(bh);
788                         goto failed;
789                 }
790
791                 memset(bh->b_data, 0, blocksize);
792                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
793                 branch[n].key = cpu_to_le32(new_blocks[n]);
794                 *branch[n].p = branch[n].key;
795                 if (n == indirect_blks) {
796                         current_block = new_blocks[n];
797                         /*
798                          * End of chain, update the last new metablock of
799                          * the chain to point to the new allocated
800                          * data blocks numbers
801                          */
802                         for (i = 1; i < num; i++)
803                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
804                 }
805                 BUFFER_TRACE(bh, "marking uptodate");
806                 set_buffer_uptodate(bh);
807                 unlock_buffer(bh);
808
809                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
810                 err = ext4_handle_dirty_metadata(handle, inode, bh);
811                 if (err)
812                         goto failed;
813         }
814         *blks = num;
815         return err;
816 failed:
817         /* Allocation failed, free what we already allocated */
818         ext4_free_blocks(handle, inode, 0, new_blocks[0], 1, 0);
819         for (i = 1; i <= n ; i++) {
820                 /*
821                  * branch[i].bh is newly allocated, so there is no
822                  * need to revoke the block, which is why we don't
823                  * need to set EXT4_FREE_BLOCKS_METADATA.
824                  */
825                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1,
826                                  EXT4_FREE_BLOCKS_FORGET);
827         }
828         for (i = n+1; i < indirect_blks; i++)
829                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
830
831         ext4_free_blocks(handle, inode, 0, new_blocks[i], num, 0);
832
833         return err;
834 }
835
836 /**
837  * ext4_splice_branch - splice the allocated branch onto inode.
838  * @handle: handle for this transaction
839  * @inode: owner
840  * @block: (logical) number of block we are adding
841  * @chain: chain of indirect blocks (with a missing link - see
842  *      ext4_alloc_branch)
843  * @where: location of missing link
844  * @num:   number of indirect blocks we are adding
845  * @blks:  number of direct blocks we are adding
846  *
847  * This function fills the missing link and does all housekeeping needed in
848  * inode (->i_blocks, etc.). In case of success we end up with the full
849  * chain to new block and return 0.
850  */
851 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
852                               ext4_lblk_t block, Indirect *where, int num,
853                               int blks)
854 {
855         int i;
856         int err = 0;
857         ext4_fsblk_t current_block;
858
859         /*
860          * If we're splicing into a [td]indirect block (as opposed to the
861          * inode) then we need to get write access to the [td]indirect block
862          * before the splice.
863          */
864         if (where->bh) {
865                 BUFFER_TRACE(where->bh, "get_write_access");
866                 err = ext4_journal_get_write_access(handle, where->bh);
867                 if (err)
868                         goto err_out;
869         }
870         /* That's it */
871
872         *where->p = where->key;
873
874         /*
875          * Update the host buffer_head or inode to point to more just allocated
876          * direct blocks blocks
877          */
878         if (num == 0 && blks > 1) {
879                 current_block = le32_to_cpu(where->key) + 1;
880                 for (i = 1; i < blks; i++)
881                         *(where->p + i) = cpu_to_le32(current_block++);
882         }
883
884         /* We are done with atomic stuff, now do the rest of housekeeping */
885         /* had we spliced it onto indirect block? */
886         if (where->bh) {
887                 /*
888                  * If we spliced it onto an indirect block, we haven't
889                  * altered the inode.  Note however that if it is being spliced
890                  * onto an indirect block at the very end of the file (the
891                  * file is growing) then we *will* alter the inode to reflect
892                  * the new i_size.  But that is not done here - it is done in
893                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
894                  */
895                 jbd_debug(5, "splicing indirect only\n");
896                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
897                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
898                 if (err)
899                         goto err_out;
900         } else {
901                 /*
902                  * OK, we spliced it into the inode itself on a direct block.
903                  */
904                 ext4_mark_inode_dirty(handle, inode);
905                 jbd_debug(5, "splicing direct\n");
906         }
907         return err;
908
909 err_out:
910         for (i = 1; i <= num; i++) {
911                 /*
912                  * branch[i].bh is newly allocated, so there is no
913                  * need to revoke the block, which is why we don't
914                  * need to set EXT4_FREE_BLOCKS_METADATA.
915                  */
916                 ext4_free_blocks(handle, inode, where[i].bh, 0, 1,
917                                  EXT4_FREE_BLOCKS_FORGET);
918         }
919         ext4_free_blocks(handle, inode, 0, le32_to_cpu(where[num].key),
920                          blks, 0);
921
922         return err;
923 }
924
925 /*
926  * The ext4_ind_map_blocks() function handles non-extents inodes
927  * (i.e., using the traditional indirect/double-indirect i_blocks
928  * scheme) for ext4_map_blocks().
929  *
930  * Allocation strategy is simple: if we have to allocate something, we will
931  * have to go the whole way to leaf. So let's do it before attaching anything
932  * to tree, set linkage between the newborn blocks, write them if sync is
933  * required, recheck the path, free and repeat if check fails, otherwise
934  * set the last missing link (that will protect us from any truncate-generated
935  * removals - all blocks on the path are immune now) and possibly force the
936  * write on the parent block.
937  * That has a nice additional property: no special recovery from the failed
938  * allocations is needed - we simply release blocks and do not touch anything
939  * reachable from inode.
940  *
941  * `handle' can be NULL if create == 0.
942  *
943  * return > 0, # of blocks mapped or allocated.
944  * return = 0, if plain lookup failed.
945  * return < 0, error case.
946  *
947  * The ext4_ind_get_blocks() function should be called with
948  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
949  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
950  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
951  * blocks.
952  */
953 static int ext4_ind_map_blocks(handle_t *handle, struct inode *inode,
954                                struct ext4_map_blocks *map,
955                                int flags)
956 {
957         int err = -EIO;
958         ext4_lblk_t offsets[4];
959         Indirect chain[4];
960         Indirect *partial;
961         ext4_fsblk_t goal;
962         int indirect_blks;
963         int blocks_to_boundary = 0;
964         int depth;
965         int count = 0;
966         ext4_fsblk_t first_block = 0;
967
968         J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
969         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
970         depth = ext4_block_to_path(inode, map->m_lblk, offsets,
971                                    &blocks_to_boundary);
972
973         if (depth == 0)
974                 goto out;
975
976         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
977
978         /* Simplest case - block found, no allocation needed */
979         if (!partial) {
980                 first_block = le32_to_cpu(chain[depth - 1].key);
981                 count++;
982                 /*map more blocks*/
983                 while (count < map->m_len && count <= blocks_to_boundary) {
984                         ext4_fsblk_t blk;
985
986                         blk = le32_to_cpu(*(chain[depth-1].p + count));
987
988                         if (blk == first_block + count)
989                                 count++;
990                         else
991                                 break;
992                 }
993                 goto got_it;
994         }
995
996         /* Next simple case - plain lookup or failed read of indirect block */
997         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
998                 goto cleanup;
999
1000         /*
1001          * Okay, we need to do block allocation.
1002         */
1003         goal = ext4_find_goal(inode, map->m_lblk, partial);
1004
1005         /* the number of blocks need to allocate for [d,t]indirect blocks */
1006         indirect_blks = (chain + depth) - partial - 1;
1007
1008         /*
1009          * Next look up the indirect map to count the totoal number of
1010          * direct blocks to allocate for this branch.
1011          */
1012         count = ext4_blks_to_allocate(partial, indirect_blks,
1013                                       map->m_len, blocks_to_boundary);
1014         /*
1015          * Block out ext4_truncate while we alter the tree
1016          */
1017         err = ext4_alloc_branch(handle, inode, map->m_lblk, indirect_blks,
1018                                 &count, goal,
1019                                 offsets + (partial - chain), partial);
1020
1021         /*
1022          * The ext4_splice_branch call will free and forget any buffers
1023          * on the new chain if there is a failure, but that risks using
1024          * up transaction credits, especially for bitmaps where the
1025          * credits cannot be returned.  Can we handle this somehow?  We
1026          * may need to return -EAGAIN upwards in the worst case.  --sct
1027          */
1028         if (!err)
1029                 err = ext4_splice_branch(handle, inode, map->m_lblk,
1030                                          partial, indirect_blks, count);
1031         if (err)
1032                 goto cleanup;
1033
1034         map->m_flags |= EXT4_MAP_NEW;
1035
1036         ext4_update_inode_fsync_trans(handle, inode, 1);
1037 got_it:
1038         map->m_flags |= EXT4_MAP_MAPPED;
1039         map->m_pblk = le32_to_cpu(chain[depth-1].key);
1040         map->m_len = count;
1041         if (count > blocks_to_boundary)
1042                 map->m_flags |= EXT4_MAP_BOUNDARY;
1043         err = count;
1044         /* Clean up and exit */
1045         partial = chain + depth - 1;    /* the whole chain */
1046 cleanup:
1047         while (partial > chain) {
1048                 BUFFER_TRACE(partial->bh, "call brelse");
1049                 brelse(partial->bh);
1050                 partial--;
1051         }
1052 out:
1053         return err;
1054 }
1055
1056 #ifdef CONFIG_QUOTA
1057 qsize_t *ext4_get_reserved_space(struct inode *inode)
1058 {
1059         return &EXT4_I(inode)->i_reserved_quota;
1060 }
1061 #endif
1062
1063 /*
1064  * Calculate the number of metadata blocks need to reserve
1065  * to allocate a new block at @lblocks for non extent file based file
1066  */
1067 static int ext4_indirect_calc_metadata_amount(struct inode *inode,
1068                                               sector_t lblock)
1069 {
1070         struct ext4_inode_info *ei = EXT4_I(inode);
1071         sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
1072         int blk_bits;
1073
1074         if (lblock < EXT4_NDIR_BLOCKS)
1075                 return 0;
1076
1077         lblock -= EXT4_NDIR_BLOCKS;
1078
1079         if (ei->i_da_metadata_calc_len &&
1080             (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
1081                 ei->i_da_metadata_calc_len++;
1082                 return 0;
1083         }
1084         ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
1085         ei->i_da_metadata_calc_len = 1;
1086         blk_bits = order_base_2(lblock);
1087         return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
1088 }
1089
1090 /*
1091  * Calculate the number of metadata blocks need to reserve
1092  * to allocate a block located at @lblock
1093  */
1094 static int ext4_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
1095 {
1096         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1097                 return ext4_ext_calc_metadata_amount(inode, lblock);
1098
1099         return ext4_indirect_calc_metadata_amount(inode, lblock);
1100 }
1101
1102 /*
1103  * Called with i_data_sem down, which is important since we can call
1104  * ext4_discard_preallocations() from here.
1105  */
1106 void ext4_da_update_reserve_space(struct inode *inode,
1107                                         int used, int quota_claim)
1108 {
1109         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1110         struct ext4_inode_info *ei = EXT4_I(inode);
1111
1112         spin_lock(&ei->i_block_reservation_lock);
1113         trace_ext4_da_update_reserve_space(inode, used);
1114         if (unlikely(used > ei->i_reserved_data_blocks)) {
1115                 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
1116                          "with only %d reserved data blocks\n",
1117                          __func__, inode->i_ino, used,
1118                          ei->i_reserved_data_blocks);
1119                 WARN_ON(1);
1120                 used = ei->i_reserved_data_blocks;
1121         }
1122
1123         /* Update per-inode reservations */
1124         ei->i_reserved_data_blocks -= used;
1125         ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
1126         percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1127                            used + ei->i_allocated_meta_blocks);
1128         ei->i_allocated_meta_blocks = 0;
1129
1130         if (ei->i_reserved_data_blocks == 0) {
1131                 /*
1132                  * We can release all of the reserved metadata blocks
1133                  * only when we have written all of the delayed
1134                  * allocation blocks.
1135                  */
1136                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1137                                    ei->i_reserved_meta_blocks);
1138                 ei->i_reserved_meta_blocks = 0;
1139                 ei->i_da_metadata_calc_len = 0;
1140         }
1141         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1142
1143         /* Update quota subsystem for data blocks */
1144         if (quota_claim)
1145                 dquot_claim_block(inode, used);
1146         else {
1147                 /*
1148                  * We did fallocate with an offset that is already delayed
1149                  * allocated. So on delayed allocated writeback we should
1150                  * not re-claim the quota for fallocated blocks.
1151                  */
1152                 dquot_release_reservation_block(inode, used);
1153         }
1154
1155         /*
1156          * If we have done all the pending block allocations and if
1157          * there aren't any writers on the inode, we can discard the
1158          * inode's preallocations.
1159          */
1160         if ((ei->i_reserved_data_blocks == 0) &&
1161             (atomic_read(&inode->i_writecount) == 0))
1162                 ext4_discard_preallocations(inode);
1163 }
1164
1165 static int __check_block_validity(struct inode *inode, const char *func,
1166                                 unsigned int line,
1167                                 struct ext4_map_blocks *map)
1168 {
1169         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
1170                                    map->m_len)) {
1171                 ext4_error_inode(inode, func, line, map->m_pblk,
1172                                  "lblock %lu mapped to illegal pblock "
1173                                  "(length %d)", (unsigned long) map->m_lblk,
1174                                  map->m_len);
1175                 return -EIO;
1176         }
1177         return 0;
1178 }
1179
1180 #define check_block_validity(inode, map)        \
1181         __check_block_validity((inode), __func__, __LINE__, (map))
1182
1183 /*
1184  * Return the number of contiguous dirty pages in a given inode
1185  * starting at page frame idx.
1186  */
1187 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
1188                                     unsigned int max_pages)
1189 {
1190         struct address_space *mapping = inode->i_mapping;
1191         pgoff_t index;
1192         struct pagevec pvec;
1193         pgoff_t num = 0;
1194         int i, nr_pages, done = 0;
1195
1196         if (max_pages == 0)
1197                 return 0;
1198         pagevec_init(&pvec, 0);
1199         while (!done) {
1200                 index = idx;
1201                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
1202                                               PAGECACHE_TAG_DIRTY,
1203                                               (pgoff_t)PAGEVEC_SIZE);
1204                 if (nr_pages == 0)
1205                         break;
1206                 for (i = 0; i < nr_pages; i++) {
1207                         struct page *page = pvec.pages[i];
1208                         struct buffer_head *bh, *head;
1209
1210                         lock_page(page);
1211                         if (unlikely(page->mapping != mapping) ||
1212                             !PageDirty(page) ||
1213                             PageWriteback(page) ||
1214                             page->index != idx) {
1215                                 done = 1;
1216                                 unlock_page(page);
1217                                 break;
1218                         }
1219                         if (page_has_buffers(page)) {
1220                                 bh = head = page_buffers(page);
1221                                 do {
1222                                         if (!buffer_delay(bh) &&
1223                                             !buffer_unwritten(bh))
1224                                                 done = 1;
1225                                         bh = bh->b_this_page;
1226                                 } while (!done && (bh != head));
1227                         }
1228                         unlock_page(page);
1229                         if (done)
1230                                 break;
1231                         idx++;
1232                         num++;
1233                         if (num >= max_pages) {
1234                                 done = 1;
1235                                 break;
1236                         }
1237                 }
1238                 pagevec_release(&pvec);
1239         }
1240         return num;
1241 }
1242
1243 /*
1244  * The ext4_map_blocks() function tries to look up the requested blocks,
1245  * and returns if the blocks are already mapped.
1246  *
1247  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1248  * and store the allocated blocks in the result buffer head and mark it
1249  * mapped.
1250  *
1251  * If file type is extents based, it will call ext4_ext_map_blocks(),
1252  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
1253  * based files
1254  *
1255  * On success, it returns the number of blocks being mapped or allocate.
1256  * if create==0 and the blocks are pre-allocated and uninitialized block,
1257  * the result buffer head is unmapped. If the create ==1, it will make sure
1258  * the buffer head is mapped.
1259  *
1260  * It returns 0 if plain look up failed (blocks have not been allocated), in
1261  * that casem, buffer head is unmapped
1262  *
1263  * It returns the error in case of allocation failure.
1264  */
1265 int ext4_map_blocks(handle_t *handle, struct inode *inode,
1266                     struct ext4_map_blocks *map, int flags)
1267 {
1268         int retval;
1269
1270         map->m_flags = 0;
1271         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
1272                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
1273                   (unsigned long) map->m_lblk);
1274         /*
1275          * Try to see if we can get the block without requesting a new
1276          * file system block.
1277          */
1278         down_read((&EXT4_I(inode)->i_data_sem));
1279         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
1280                 retval = ext4_ext_map_blocks(handle, inode, map, 0);
1281         } else {
1282                 retval = ext4_ind_map_blocks(handle, inode, map, 0);
1283         }
1284         up_read((&EXT4_I(inode)->i_data_sem));
1285
1286         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
1287                 int ret = check_block_validity(inode, map);
1288                 if (ret != 0)
1289                         return ret;
1290         }
1291
1292         /* If it is only a block(s) look up */
1293         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1294                 return retval;
1295
1296         /*
1297          * Returns if the blocks have already allocated
1298          *
1299          * Note that if blocks have been preallocated
1300          * ext4_ext_get_block() returns th create = 0
1301          * with buffer head unmapped.
1302          */
1303         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
1304                 return retval;
1305
1306         /*
1307          * When we call get_blocks without the create flag, the
1308          * BH_Unwritten flag could have gotten set if the blocks
1309          * requested were part of a uninitialized extent.  We need to
1310          * clear this flag now that we are committed to convert all or
1311          * part of the uninitialized extent to be an initialized
1312          * extent.  This is because we need to avoid the combination
1313          * of BH_Unwritten and BH_Mapped flags being simultaneously
1314          * set on the buffer_head.
1315          */
1316         map->m_flags &= ~EXT4_MAP_UNWRITTEN;
1317
1318         /*
1319          * New blocks allocate and/or writing to uninitialized extent
1320          * will possibly result in updating i_data, so we take
1321          * the write lock of i_data_sem, and call get_blocks()
1322          * with create == 1 flag.
1323          */
1324         down_write((&EXT4_I(inode)->i_data_sem));
1325
1326         /*
1327          * if the caller is from delayed allocation writeout path
1328          * we have already reserved fs blocks for allocation
1329          * let the underlying get_block() function know to
1330          * avoid double accounting
1331          */
1332         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1333                 ext4_set_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
1334         /*
1335          * We need to check for EXT4 here because migrate
1336          * could have changed the inode type in between
1337          */
1338         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
1339                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
1340         } else {
1341                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
1342
1343                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
1344                         /*
1345                          * We allocated new blocks which will result in
1346                          * i_data's format changing.  Force the migrate
1347                          * to fail by clearing migrate flags
1348                          */
1349                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
1350                 }
1351
1352                 /*
1353                  * Update reserved blocks/metadata blocks after successful
1354                  * block allocation which had been deferred till now. We don't
1355                  * support fallocate for non extent files. So we can update
1356                  * reserve space here.
1357                  */
1358                 if ((retval > 0) &&
1359                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
1360                         ext4_da_update_reserve_space(inode, retval, 1);
1361         }
1362         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1363                 ext4_clear_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
1364
1365         up_write((&EXT4_I(inode)->i_data_sem));
1366         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
1367                 int ret = check_block_validity(inode, map);
1368                 if (ret != 0)
1369                         return ret;
1370         }
1371         return retval;
1372 }
1373
1374 /* Maximum number of blocks we map for direct IO at once. */
1375 #define DIO_MAX_BLOCKS 4096
1376
1377 static int _ext4_get_block(struct inode *inode, sector_t iblock,
1378                            struct buffer_head *bh, int flags)
1379 {
1380         handle_t *handle = ext4_journal_current_handle();
1381         struct ext4_map_blocks map;
1382         int ret = 0, started = 0;
1383         int dio_credits;
1384
1385         map.m_lblk = iblock;
1386         map.m_len = bh->b_size >> inode->i_blkbits;
1387
1388         if (flags && !handle) {
1389                 /* Direct IO write... */
1390                 if (map.m_len > DIO_MAX_BLOCKS)
1391                         map.m_len = DIO_MAX_BLOCKS;
1392                 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
1393                 handle = ext4_journal_start(inode, dio_credits);
1394                 if (IS_ERR(handle)) {
1395                         ret = PTR_ERR(handle);
1396                         return ret;
1397                 }
1398                 started = 1;
1399         }
1400
1401         ret = ext4_map_blocks(handle, inode, &map, flags);
1402         if (ret > 0) {
1403                 map_bh(bh, inode->i_sb, map.m_pblk);
1404                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
1405                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
1406                 ret = 0;
1407         }
1408         if (started)
1409                 ext4_journal_stop(handle);
1410         return ret;
1411 }
1412
1413 int ext4_get_block(struct inode *inode, sector_t iblock,
1414                    struct buffer_head *bh, int create)
1415 {
1416         return _ext4_get_block(inode, iblock, bh,
1417                                create ? EXT4_GET_BLOCKS_CREATE : 0);
1418 }
1419
1420 /*
1421  * `handle' can be NULL if create is zero
1422  */
1423 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1424                                 ext4_lblk_t block, int create, int *errp)
1425 {
1426         struct ext4_map_blocks map;
1427         struct buffer_head *bh;
1428         int fatal = 0, err;
1429
1430         J_ASSERT(handle != NULL || create == 0);
1431
1432         map.m_lblk = block;
1433         map.m_len = 1;
1434         err = ext4_map_blocks(handle, inode, &map,
1435                               create ? EXT4_GET_BLOCKS_CREATE : 0);
1436
1437         if (err < 0)
1438                 *errp = err;
1439         if (err <= 0)
1440                 return NULL;
1441         *errp = 0;
1442
1443         bh = sb_getblk(inode->i_sb, map.m_pblk);
1444         if (!bh) {
1445                 *errp = -EIO;
1446                 return NULL;
1447         }
1448         if (map.m_flags & EXT4_MAP_NEW) {
1449                 J_ASSERT(create != 0);
1450                 J_ASSERT(handle != NULL);
1451
1452                 /*
1453                  * Now that we do not always journal data, we should
1454                  * keep in mind whether this should always journal the
1455                  * new buffer as metadata.  For now, regular file
1456                  * writes use ext4_get_block instead, so it's not a
1457                  * problem.
1458                  */
1459                 lock_buffer(bh);
1460                 BUFFER_TRACE(bh, "call get_create_access");
1461                 fatal = ext4_journal_get_create_access(handle, bh);
1462                 if (!fatal && !buffer_uptodate(bh)) {
1463                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1464                         set_buffer_uptodate(bh);
1465                 }
1466                 unlock_buffer(bh);
1467                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1468                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1469                 if (!fatal)
1470                         fatal = err;
1471         } else {
1472                 BUFFER_TRACE(bh, "not a new buffer");
1473         }
1474         if (fatal) {
1475                 *errp = fatal;
1476                 brelse(bh);
1477                 bh = NULL;
1478         }
1479         return bh;
1480 }
1481
1482 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1483                                ext4_lblk_t block, int create, int *err)
1484 {
1485         struct buffer_head *bh;
1486
1487         bh = ext4_getblk(handle, inode, block, create, err);
1488         if (!bh)
1489                 return bh;
1490         if (buffer_uptodate(bh))
1491                 return bh;
1492         ll_rw_block(READ_META, 1, &bh);
1493         wait_on_buffer(bh);
1494         if (buffer_uptodate(bh))
1495                 return bh;
1496         put_bh(bh);
1497         *err = -EIO;
1498         return NULL;
1499 }
1500
1501 static int walk_page_buffers(handle_t *handle,
1502                              struct buffer_head *head,
1503                              unsigned from,
1504                              unsigned to,
1505                              int *partial,
1506                              int (*fn)(handle_t *handle,
1507                                        struct buffer_head *bh))
1508 {
1509         struct buffer_head *bh;
1510         unsigned block_start, block_end;
1511         unsigned blocksize = head->b_size;
1512         int err, ret = 0;
1513         struct buffer_head *next;
1514
1515         for (bh = head, block_start = 0;
1516              ret == 0 && (bh != head || !block_start);
1517              block_start = block_end, bh = next) {
1518                 next = bh->b_this_page;
1519                 block_end = block_start + blocksize;
1520                 if (block_end <= from || block_start >= to) {
1521                         if (partial && !buffer_uptodate(bh))
1522                                 *partial = 1;
1523                         continue;
1524                 }
1525                 err = (*fn)(handle, bh);
1526                 if (!ret)
1527                         ret = err;
1528         }
1529         return ret;
1530 }
1531
1532 /*
1533  * To preserve ordering, it is essential that the hole instantiation and
1534  * the data write be encapsulated in a single transaction.  We cannot
1535  * close off a transaction and start a new one between the ext4_get_block()
1536  * and the commit_write().  So doing the jbd2_journal_start at the start of
1537  * prepare_write() is the right place.
1538  *
1539  * Also, this function can nest inside ext4_writepage() ->
1540  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1541  * has generated enough buffer credits to do the whole page.  So we won't
1542  * block on the journal in that case, which is good, because the caller may
1543  * be PF_MEMALLOC.
1544  *
1545  * By accident, ext4 can be reentered when a transaction is open via
1546  * quota file writes.  If we were to commit the transaction while thus
1547  * reentered, there can be a deadlock - we would be holding a quota
1548  * lock, and the commit would never complete if another thread had a
1549  * transaction open and was blocking on the quota lock - a ranking
1550  * violation.
1551  *
1552  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1553  * will _not_ run commit under these circumstances because handle->h_ref
1554  * is elevated.  We'll still have enough credits for the tiny quotafile
1555  * write.
1556  */
1557 static int do_journal_get_write_access(handle_t *handle,
1558                                        struct buffer_head *bh)
1559 {
1560         int dirty = buffer_dirty(bh);
1561         int ret;
1562
1563         if (!buffer_mapped(bh) || buffer_freed(bh))
1564                 return 0;
1565         /*
1566          * __block_write_begin() could have dirtied some buffers. Clean
1567          * the dirty bit as jbd2_journal_get_write_access() could complain
1568          * otherwise about fs integrity issues. Setting of the dirty bit
1569          * by __block_write_begin() isn't a real problem here as we clear
1570          * the bit before releasing a page lock and thus writeback cannot
1571          * ever write the buffer.
1572          */
1573         if (dirty)
1574                 clear_buffer_dirty(bh);
1575         ret = ext4_journal_get_write_access(handle, bh);
1576         if (!ret && dirty)
1577                 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1578         return ret;
1579 }
1580
1581 /*
1582  * Truncate blocks that were not used by write. We have to truncate the
1583  * pagecache as well so that corresponding buffers get properly unmapped.
1584  */
1585 static void ext4_truncate_failed_write(struct inode *inode)
1586 {
1587         truncate_inode_pages(inode->i_mapping, inode->i_size);
1588         ext4_truncate(inode);
1589 }
1590
1591 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
1592                    struct buffer_head *bh_result, int create);
1593 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1594                             loff_t pos, unsigned len, unsigned flags,
1595                             struct page **pagep, void **fsdata)
1596 {
1597         struct inode *inode = mapping->host;
1598         int ret, needed_blocks;
1599         handle_t *handle;
1600         int retries = 0;
1601         struct page *page;
1602         pgoff_t index;
1603         unsigned from, to;
1604
1605         trace_ext4_write_begin(inode, pos, len, flags);
1606         /*
1607          * Reserve one block more for addition to orphan list in case
1608          * we allocate blocks but write fails for some reason
1609          */
1610         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1611         index = pos >> PAGE_CACHE_SHIFT;
1612         from = pos & (PAGE_CACHE_SIZE - 1);
1613         to = from + len;
1614
1615 retry:
1616         handle = ext4_journal_start(inode, needed_blocks);
1617         if (IS_ERR(handle)) {
1618                 ret = PTR_ERR(handle);
1619                 goto out;
1620         }
1621
1622         /* We cannot recurse into the filesystem as the transaction is already
1623          * started */
1624         flags |= AOP_FLAG_NOFS;
1625
1626         page = grab_cache_page_write_begin(mapping, index, flags);
1627         if (!page) {
1628                 ext4_journal_stop(handle);
1629                 ret = -ENOMEM;
1630                 goto out;
1631         }
1632         *pagep = page;
1633
1634         if (ext4_should_dioread_nolock(inode))
1635                 ret = __block_write_begin(page, pos, len, ext4_get_block_write);
1636         else
1637                 ret = __block_write_begin(page, pos, len, ext4_get_block);
1638
1639         if (!ret && ext4_should_journal_data(inode)) {
1640                 ret = walk_page_buffers(handle, page_buffers(page),
1641                                 from, to, NULL, do_journal_get_write_access);
1642         }
1643
1644         if (ret) {
1645                 unlock_page(page);
1646                 page_cache_release(page);
1647                 /*
1648                  * __block_write_begin may have instantiated a few blocks
1649                  * outside i_size.  Trim these off again. Don't need
1650                  * i_size_read because we hold i_mutex.
1651                  *
1652                  * Add inode to orphan list in case we crash before
1653                  * truncate finishes
1654                  */
1655                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1656                         ext4_orphan_add(handle, inode);
1657
1658                 ext4_journal_stop(handle);
1659                 if (pos + len > inode->i_size) {
1660                         ext4_truncate_failed_write(inode);
1661                         /*
1662                          * If truncate failed early the inode might
1663                          * still be on the orphan list; we need to
1664                          * make sure the inode is removed from the
1665                          * orphan list in that case.
1666                          */
1667                         if (inode->i_nlink)
1668                                 ext4_orphan_del(NULL, inode);
1669                 }
1670         }
1671
1672         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1673                 goto retry;
1674 out:
1675         return ret;
1676 }
1677
1678 /* For write_end() in data=journal mode */
1679 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1680 {
1681         if (!buffer_mapped(bh) || buffer_freed(bh))
1682                 return 0;
1683         set_buffer_uptodate(bh);
1684         return ext4_handle_dirty_metadata(handle, NULL, bh);
1685 }
1686
1687 static int ext4_generic_write_end(struct file *file,
1688                                   struct address_space *mapping,
1689                                   loff_t pos, unsigned len, unsigned copied,
1690                                   struct page *page, void *fsdata)
1691 {
1692         int i_size_changed = 0;
1693         struct inode *inode = mapping->host;
1694         handle_t *handle = ext4_journal_current_handle();
1695
1696         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1697
1698         /*
1699          * No need to use i_size_read() here, the i_size
1700          * cannot change under us because we hold i_mutex.
1701          *
1702          * But it's important to update i_size while still holding page lock:
1703          * page writeout could otherwise come in and zero beyond i_size.
1704          */
1705         if (pos + copied > inode->i_size) {
1706                 i_size_write(inode, pos + copied);
1707                 i_size_changed = 1;
1708         }
1709
1710         if (pos + copied >  EXT4_I(inode)->i_disksize) {
1711                 /* We need to mark inode dirty even if
1712                  * new_i_size is less that inode->i_size
1713                  * bu greater than i_disksize.(hint delalloc)
1714                  */
1715                 ext4_update_i_disksize(inode, (pos + copied));
1716                 i_size_changed = 1;
1717         }
1718         unlock_page(page);
1719         page_cache_release(page);
1720
1721         /*
1722          * Don't mark the inode dirty under page lock. First, it unnecessarily
1723          * makes the holding time of page lock longer. Second, it forces lock
1724          * ordering of page lock and transaction start for journaling
1725          * filesystems.
1726          */
1727         if (i_size_changed)
1728                 ext4_mark_inode_dirty(handle, inode);
1729
1730         return copied;
1731 }
1732
1733 /*
1734  * We need to pick up the new inode size which generic_commit_write gave us
1735  * `file' can be NULL - eg, when called from page_symlink().
1736  *
1737  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1738  * buffers are managed internally.
1739  */
1740 static int ext4_ordered_write_end(struct file *file,
1741                                   struct address_space *mapping,
1742                                   loff_t pos, unsigned len, unsigned copied,
1743                                   struct page *page, void *fsdata)
1744 {
1745         handle_t *handle = ext4_journal_current_handle();
1746         struct inode *inode = mapping->host;
1747         int ret = 0, ret2;
1748
1749         trace_ext4_ordered_write_end(inode, pos, len, copied);
1750         ret = ext4_jbd2_file_inode(handle, inode);
1751
1752         if (ret == 0) {
1753                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1754                                                         page, fsdata);
1755                 copied = ret2;
1756                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1757                         /* if we have allocated more blocks and copied
1758                          * less. We will have blocks allocated outside
1759                          * inode->i_size. So truncate them
1760                          */
1761                         ext4_orphan_add(handle, inode);
1762                 if (ret2 < 0)
1763                         ret = ret2;
1764         }
1765         ret2 = ext4_journal_stop(handle);
1766         if (!ret)
1767                 ret = ret2;
1768
1769         if (pos + len > inode->i_size) {
1770                 ext4_truncate_failed_write(inode);
1771                 /*
1772                  * If truncate failed early the inode might still be
1773                  * on the orphan list; we need to make sure the inode
1774                  * is removed from the orphan list in that case.
1775                  */
1776                 if (inode->i_nlink)
1777                         ext4_orphan_del(NULL, inode);
1778         }
1779
1780
1781         return ret ? ret : copied;
1782 }
1783
1784 static int ext4_writeback_write_end(struct file *file,
1785                                     struct address_space *mapping,
1786                                     loff_t pos, unsigned len, unsigned copied,
1787                                     struct page *page, void *fsdata)
1788 {
1789         handle_t *handle = ext4_journal_current_handle();
1790         struct inode *inode = mapping->host;
1791         int ret = 0, ret2;
1792
1793         trace_ext4_writeback_write_end(inode, pos, len, copied);
1794         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1795                                                         page, fsdata);
1796         copied = ret2;
1797         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1798                 /* if we have allocated more blocks and copied
1799                  * less. We will have blocks allocated outside
1800                  * inode->i_size. So truncate them
1801                  */
1802                 ext4_orphan_add(handle, inode);
1803
1804         if (ret2 < 0)
1805                 ret = ret2;
1806
1807         ret2 = ext4_journal_stop(handle);
1808         if (!ret)
1809                 ret = ret2;
1810
1811         if (pos + len > inode->i_size) {
1812                 ext4_truncate_failed_write(inode);
1813                 /*
1814                  * If truncate failed early the inode might still be
1815                  * on the orphan list; we need to make sure the inode
1816                  * is removed from the orphan list in that case.
1817                  */
1818                 if (inode->i_nlink)
1819                         ext4_orphan_del(NULL, inode);
1820         }
1821
1822         return ret ? ret : copied;
1823 }
1824
1825 static int ext4_journalled_write_end(struct file *file,
1826                                      struct address_space *mapping,
1827                                      loff_t pos, unsigned len, unsigned copied,
1828                                      struct page *page, void *fsdata)
1829 {
1830         handle_t *handle = ext4_journal_current_handle();
1831         struct inode *inode = mapping->host;
1832         int ret = 0, ret2;
1833         int partial = 0;
1834         unsigned from, to;
1835         loff_t new_i_size;
1836
1837         trace_ext4_journalled_write_end(inode, pos, len, copied);
1838         from = pos & (PAGE_CACHE_SIZE - 1);
1839         to = from + len;
1840
1841         if (copied < len) {
1842                 if (!PageUptodate(page))
1843                         copied = 0;
1844                 page_zero_new_buffers(page, from+copied, to);
1845         }
1846
1847         ret = walk_page_buffers(handle, page_buffers(page), from,
1848                                 to, &partial, write_end_fn);
1849         if (!partial)
1850                 SetPageUptodate(page);
1851         new_i_size = pos + copied;
1852         if (new_i_size > inode->i_size)
1853                 i_size_write(inode, pos+copied);
1854         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1855         if (new_i_size > EXT4_I(inode)->i_disksize) {
1856                 ext4_update_i_disksize(inode, new_i_size);
1857                 ret2 = ext4_mark_inode_dirty(handle, inode);
1858                 if (!ret)
1859                         ret = ret2;
1860         }
1861
1862         unlock_page(page);
1863         page_cache_release(page);
1864         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1865                 /* if we have allocated more blocks and copied
1866                  * less. We will have blocks allocated outside
1867                  * inode->i_size. So truncate them
1868                  */
1869                 ext4_orphan_add(handle, inode);
1870
1871         ret2 = ext4_journal_stop(handle);
1872         if (!ret)
1873                 ret = ret2;
1874         if (pos + len > inode->i_size) {
1875                 ext4_truncate_failed_write(inode);
1876                 /*
1877                  * If truncate failed early the inode might still be
1878                  * on the orphan list; we need to make sure the inode
1879                  * is removed from the orphan list in that case.
1880                  */
1881                 if (inode->i_nlink)
1882                         ext4_orphan_del(NULL, inode);
1883         }
1884
1885         return ret ? ret : copied;
1886 }
1887
1888 /*
1889  * Reserve a single block located at lblock
1890  */
1891 static int ext4_da_reserve_space(struct inode *inode, ext4_lblk_t lblock)
1892 {
1893         int retries = 0;
1894         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1895         struct ext4_inode_info *ei = EXT4_I(inode);
1896         unsigned long md_needed;
1897         int ret;
1898
1899         /*
1900          * recalculate the amount of metadata blocks to reserve
1901          * in order to allocate nrblocks
1902          * worse case is one extent per block
1903          */
1904 repeat:
1905         spin_lock(&ei->i_block_reservation_lock);
1906         md_needed = ext4_calc_metadata_amount(inode, lblock);
1907         trace_ext4_da_reserve_space(inode, md_needed);
1908         spin_unlock(&ei->i_block_reservation_lock);
1909
1910         /*
1911          * We will charge metadata quota at writeout time; this saves
1912          * us from metadata over-estimation, though we may go over by
1913          * a small amount in the end.  Here we just reserve for data.
1914          */
1915         ret = dquot_reserve_block(inode, 1);
1916         if (ret)
1917                 return ret;
1918         /*
1919          * We do still charge estimated metadata to the sb though;
1920          * we cannot afford to run out of free blocks.
1921          */
1922         if (ext4_claim_free_blocks(sbi, md_needed + 1)) {
1923                 dquot_release_reservation_block(inode, 1);
1924                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1925                         yield();
1926                         goto repeat;
1927                 }
1928                 return -ENOSPC;
1929         }
1930         spin_lock(&ei->i_block_reservation_lock);
1931         ei->i_reserved_data_blocks++;
1932         ei->i_reserved_meta_blocks += md_needed;
1933         spin_unlock(&ei->i_block_reservation_lock);
1934
1935         return 0;       /* success */
1936 }
1937
1938 static void ext4_da_release_space(struct inode *inode, int to_free)
1939 {
1940         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1941         struct ext4_inode_info *ei = EXT4_I(inode);
1942
1943         if (!to_free)
1944                 return;         /* Nothing to release, exit */
1945
1946         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1947
1948         trace_ext4_da_release_space(inode, to_free);
1949         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1950                 /*
1951                  * if there aren't enough reserved blocks, then the
1952                  * counter is messed up somewhere.  Since this
1953                  * function is called from invalidate page, it's
1954                  * harmless to return without any action.
1955                  */
1956                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1957                          "ino %lu, to_free %d with only %d reserved "
1958                          "data blocks\n", inode->i_ino, to_free,
1959                          ei->i_reserved_data_blocks);
1960                 WARN_ON(1);
1961                 to_free = ei->i_reserved_data_blocks;
1962         }
1963         ei->i_reserved_data_blocks -= to_free;
1964
1965         if (ei->i_reserved_data_blocks == 0) {
1966                 /*
1967                  * We can release all of the reserved metadata blocks
1968                  * only when we have written all of the delayed
1969                  * allocation blocks.
1970                  */
1971                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1972                                    ei->i_reserved_meta_blocks);
1973                 ei->i_reserved_meta_blocks = 0;
1974                 ei->i_da_metadata_calc_len = 0;
1975         }
1976
1977         /* update fs dirty data blocks counter */
1978         percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1979
1980         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1981
1982         dquot_release_reservation_block(inode, to_free);
1983 }
1984
1985 static void ext4_da_page_release_reservation(struct page *page,
1986                                              unsigned long offset)
1987 {
1988         int to_release = 0;
1989         struct buffer_head *head, *bh;
1990         unsigned int curr_off = 0;
1991
1992         head = page_buffers(page);
1993         bh = head;
1994         do {
1995                 unsigned int next_off = curr_off + bh->b_size;
1996
1997                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1998                         to_release++;
1999                         clear_buffer_delay(bh);
2000                 }
2001                 curr_off = next_off;
2002         } while ((bh = bh->b_this_page) != head);
2003         ext4_da_release_space(page->mapping->host, to_release);
2004 }
2005
2006 /*
2007  * Delayed allocation stuff
2008  */
2009
2010 /*
2011  * mpage_da_submit_io - walks through extent of pages and try to write
2012  * them with writepage() call back
2013  *
2014  * @mpd->inode: inode
2015  * @mpd->first_page: first page of the extent
2016  * @mpd->next_page: page after the last page of the extent
2017  *
2018  * By the time mpage_da_submit_io() is called we expect all blocks
2019  * to be allocated. this may be wrong if allocation failed.
2020  *
2021  * As pages are already locked by write_cache_pages(), we can't use it
2022  */
2023 static int mpage_da_submit_io(struct mpage_da_data *mpd,
2024                               struct ext4_map_blocks *map)
2025 {
2026         struct pagevec pvec;
2027         unsigned long index, end;
2028         int ret = 0, err, nr_pages, i;
2029         struct inode *inode = mpd->inode;
2030         struct address_space *mapping = inode->i_mapping;
2031         loff_t size = i_size_read(inode);
2032         unsigned int len, block_start;
2033         struct buffer_head *bh, *page_bufs = NULL;
2034         int journal_data = ext4_should_journal_data(inode);
2035         sector_t pblock = 0, cur_logical = 0;
2036         struct ext4_io_submit io_submit;
2037
2038         BUG_ON(mpd->next_page <= mpd->first_page);
2039         memset(&io_submit, 0, sizeof(io_submit));
2040         /*
2041          * We need to start from the first_page to the next_page - 1
2042          * to make sure we also write the mapped dirty buffer_heads.
2043          * If we look at mpd->b_blocknr we would only be looking
2044          * at the currently mapped buffer_heads.
2045          */
2046         index = mpd->first_page;
2047         end = mpd->next_page - 1;
2048
2049         pagevec_init(&pvec, 0);
2050         while (index <= end) {
2051                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2052                 if (nr_pages == 0)
2053                         break;
2054                 for (i = 0; i < nr_pages; i++) {
2055                         int commit_write = 0, redirty_page = 0;
2056                         struct page *page = pvec.pages[i];
2057
2058                         index = page->index;
2059                         if (index > end)
2060                                 break;
2061
2062                         if (index == size >> PAGE_CACHE_SHIFT)
2063                                 len = size & ~PAGE_CACHE_MASK;
2064                         else
2065                                 len = PAGE_CACHE_SIZE;
2066                         if (map) {
2067                                 cur_logical = index << (PAGE_CACHE_SHIFT -
2068                                                         inode->i_blkbits);
2069                                 pblock = map->m_pblk + (cur_logical -
2070                                                         map->m_lblk);
2071                         }
2072                         index++;
2073
2074                         BUG_ON(!PageLocked(page));
2075                         BUG_ON(PageWriteback(page));
2076
2077                         /*
2078                          * If the page does not have buffers (for
2079                          * whatever reason), try to create them using
2080                          * __block_write_begin.  If this fails,
2081                          * redirty the page and move on.
2082                          */
2083                         if (!page_has_buffers(page)) {
2084                                 if (__block_write_begin(page, 0, len,
2085                                                 noalloc_get_block_write)) {
2086                                 redirty_page:
2087                                         redirty_page_for_writepage(mpd->wbc,
2088                                                                    page);
2089                                         unlock_page(page);
2090                                         continue;
2091                                 }
2092                                 commit_write = 1;
2093                         }
2094
2095                         bh = page_bufs = page_buffers(page);
2096                         block_start = 0;
2097                         do {
2098                                 if (!bh)
2099                                         goto redirty_page;
2100                                 if (map && (cur_logical >= map->m_lblk) &&
2101                                     (cur_logical <= (map->m_lblk +
2102                                                      (map->m_len - 1)))) {
2103                                         if (buffer_delay(bh)) {
2104                                                 clear_buffer_delay(bh);
2105                                                 bh->b_blocknr = pblock;
2106                                         }
2107                                         if (buffer_unwritten(bh) ||
2108                                             buffer_mapped(bh))
2109                                                 BUG_ON(bh->b_blocknr != pblock);
2110                                         if (map->m_flags & EXT4_MAP_UNINIT)
2111                                                 set_buffer_uninit(bh);
2112                                         clear_buffer_unwritten(bh);
2113                                 }
2114
2115                                 /* redirty page if block allocation undone */
2116                                 if (buffer_delay(bh) || buffer_unwritten(bh))
2117                                         redirty_page = 1;
2118                                 bh = bh->b_this_page;
2119                                 block_start += bh->b_size;
2120                                 cur_logical++;
2121                                 pblock++;
2122                         } while (bh != page_bufs);
2123
2124                         if (redirty_page)
2125                                 goto redirty_page;
2126
2127                         if (commit_write)
2128                                 /* mark the buffer_heads as dirty & uptodate */
2129                                 block_commit_write(page, 0, len);
2130
2131                         /*
2132                          * Delalloc doesn't support data journalling,
2133                          * but eventually maybe we'll lift this
2134                          * restriction.
2135                          */
2136                         if (unlikely(journal_data && PageChecked(page)))
2137                                 err = __ext4_journalled_writepage(page, len);
2138                         else if (test_opt(inode->i_sb, MBLK_IO_SUBMIT))
2139                                 err = ext4_bio_write_page(&io_submit, page,
2140                                                           len, mpd->wbc);
2141                         else
2142                                 err = block_write_full_page(page,
2143                                         noalloc_get_block_write, mpd->wbc);
2144
2145                         if (!err)
2146                                 mpd->pages_written++;
2147                         /*
2148                          * In error case, we have to continue because
2149                          * remaining pages are still locked
2150                          */
2151                         if (ret == 0)
2152                                 ret = err;
2153                 }
2154                 pagevec_release(&pvec);
2155         }
2156         ext4_io_submit(&io_submit);
2157         return ret;
2158 }
2159
2160 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2161                                         sector_t logical, long blk_cnt)
2162 {
2163         int nr_pages, i;
2164         pgoff_t index, end;
2165         struct pagevec pvec;
2166         struct inode *inode = mpd->inode;
2167         struct address_space *mapping = inode->i_mapping;
2168
2169         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2170         end   = (logical + blk_cnt - 1) >>
2171                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2172         while (index <= end) {
2173                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2174                 if (nr_pages == 0)
2175                         break;
2176                 for (i = 0; i < nr_pages; i++) {
2177                         struct page *page = pvec.pages[i];
2178                         if (page->index > end)
2179                                 break;
2180                         BUG_ON(!PageLocked(page));
2181                         BUG_ON(PageWriteback(page));
2182                         block_invalidatepage(page, 0);
2183                         ClearPageUptodate(page);
2184                         unlock_page(page);
2185                 }
2186                 index = pvec.pages[nr_pages - 1]->index + 1;
2187                 pagevec_release(&pvec);
2188         }
2189         return;
2190 }
2191
2192 static void ext4_print_free_blocks(struct inode *inode)
2193 {
2194         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2195         printk(KERN_CRIT "Total free blocks count %lld\n",
2196                ext4_count_free_blocks(inode->i_sb));
2197         printk(KERN_CRIT "Free/Dirty block details\n");
2198         printk(KERN_CRIT "free_blocks=%lld\n",
2199                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2200         printk(KERN_CRIT "dirty_blocks=%lld\n",
2201                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2202         printk(KERN_CRIT "Block reservation details\n");
2203         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2204                EXT4_I(inode)->i_reserved_data_blocks);
2205         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2206                EXT4_I(inode)->i_reserved_meta_blocks);
2207         return;
2208 }
2209
2210 /*
2211  * mpage_da_map_and_submit - go through given space, map them
2212  *       if necessary, and then submit them for I/O
2213  *
2214  * @mpd - bh describing space
2215  *
2216  * The function skips space we know is already mapped to disk blocks.
2217  *
2218  */
2219 static void mpage_da_map_and_submit(struct mpage_da_data *mpd)
2220 {
2221         int err, blks, get_blocks_flags;
2222         struct ext4_map_blocks map, *mapp = NULL;
2223         sector_t next = mpd->b_blocknr;
2224         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2225         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2226         handle_t *handle = NULL;
2227
2228         /*
2229          * If the blocks are mapped already, or we couldn't accumulate
2230          * any blocks, then proceed immediately to the submission stage.
2231          */
2232         if ((mpd->b_size == 0) ||
2233             ((mpd->b_state  & (1 << BH_Mapped)) &&
2234              !(mpd->b_state & (1 << BH_Delay)) &&
2235              !(mpd->b_state & (1 << BH_Unwritten))))
2236                 goto submit_io;
2237
2238         handle = ext4_journal_current_handle();
2239         BUG_ON(!handle);
2240
2241         /*
2242          * Call ext4_map_blocks() to allocate any delayed allocation
2243          * blocks, or to convert an uninitialized extent to be
2244          * initialized (in the case where we have written into
2245          * one or more preallocated blocks).
2246          *
2247          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2248          * indicate that we are on the delayed allocation path.  This
2249          * affects functions in many different parts of the allocation
2250          * call path.  This flag exists primarily because we don't
2251          * want to change *many* call functions, so ext4_map_blocks()
2252          * will set the EXT4_STATE_DELALLOC_RESERVED flag once the
2253          * inode's allocation semaphore is taken.
2254          *
2255          * If the blocks in questions were delalloc blocks, set
2256          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2257          * variables are updated after the blocks have been allocated.
2258          */
2259         map.m_lblk = next;
2260         map.m_len = max_blocks;
2261         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
2262         if (ext4_should_dioread_nolock(mpd->inode))
2263                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2264         if (mpd->b_state & (1 << BH_Delay))
2265                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2266
2267         blks = ext4_map_blocks(handle, mpd->inode, &map, get_blocks_flags);
2268         if (blks < 0) {
2269                 struct super_block *sb = mpd->inode->i_sb;
2270
2271                 err = blks;
2272                 /*
2273                  * If get block returns EAGAIN or ENOSPC and there
2274                  * appears to be free blocks we will call
2275                  * ext4_writepage() for all of the pages which will
2276                  * just redirty the pages.
2277                  */
2278                 if (err == -EAGAIN)
2279                         goto submit_io;
2280
2281                 if (err == -ENOSPC &&
2282                     ext4_count_free_blocks(sb)) {
2283                         mpd->retval = err;
2284                         goto submit_io;
2285                 }
2286
2287                 /*
2288                  * get block failure will cause us to loop in
2289                  * writepages, because a_ops->writepage won't be able
2290                  * to make progress. The page will be redirtied by
2291                  * writepage and writepages will again try to write
2292                  * the same.
2293                  */
2294                 if (!(EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2295                         ext4_msg(sb, KERN_CRIT,
2296                                  "delayed block allocation failed for inode %lu "
2297                                  "at logical offset %llu with max blocks %zd "
2298                                  "with error %d", mpd->inode->i_ino,
2299                                  (unsigned long long) next,
2300                                  mpd->b_size >> mpd->inode->i_blkbits, err);
2301                         ext4_msg(sb, KERN_CRIT,
2302                                 "This should not happen!! Data will be lost\n");
2303                         if (err == -ENOSPC)
2304                                 ext4_print_free_blocks(mpd->inode);
2305                 }
2306                 /* invalidate all the pages */
2307                 ext4_da_block_invalidatepages(mpd, next,
2308                                 mpd->b_size >> mpd->inode->i_blkbits);
2309                 return;
2310         }
2311         BUG_ON(blks == 0);
2312
2313         mapp = &map;
2314         if (map.m_flags & EXT4_MAP_NEW) {
2315                 struct block_device *bdev = mpd->inode->i_sb->s_bdev;
2316                 int i;
2317
2318                 for (i = 0; i < map.m_len; i++)
2319                         unmap_underlying_metadata(bdev, map.m_pblk + i);
2320         }
2321
2322         if (ext4_should_order_data(mpd->inode)) {
2323                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2324                 if (err)
2325                         /* This only happens if the journal is aborted */
2326                         return;
2327         }
2328
2329         /*
2330          * Update on-disk size along with block allocation.
2331          */
2332         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2333         if (disksize > i_size_read(mpd->inode))
2334                 disksize = i_size_read(mpd->inode);
2335         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2336                 ext4_update_i_disksize(mpd->inode, disksize);
2337                 err = ext4_mark_inode_dirty(handle, mpd->inode);
2338                 if (err)
2339                         ext4_error(mpd->inode->i_sb,
2340                                    "Failed to mark inode %lu dirty",
2341                                    mpd->inode->i_ino);
2342         }
2343
2344 submit_io:
2345         mpage_da_submit_io(mpd, mapp);
2346         mpd->io_done = 1;
2347 }
2348
2349 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2350                 (1 << BH_Delay) | (1 << BH_Unwritten))
2351
2352 /*
2353  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2354  *
2355  * @mpd->lbh - extent of blocks
2356  * @logical - logical number of the block in the file
2357  * @bh - bh of the block (used to access block's state)
2358  *
2359  * the function is used to collect contig. blocks in same state
2360  */
2361 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2362                                    sector_t logical, size_t b_size,
2363                                    unsigned long b_state)
2364 {
2365         sector_t next;
2366         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2367
2368         /*
2369          * XXX Don't go larger than mballoc is willing to allocate
2370          * This is a stopgap solution.  We eventually need to fold
2371          * mpage_da_submit_io() into this function and then call
2372          * ext4_map_blocks() multiple times in a loop
2373          */
2374         if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
2375                 goto flush_it;
2376
2377         /* check if thereserved journal credits might overflow */
2378         if (!(ext4_test_inode_flag(mpd->inode, EXT4_INODE_EXTENTS))) {
2379                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2380                         /*
2381                          * With non-extent format we are limited by the journal
2382                          * credit available.  Total credit needed to insert
2383                          * nrblocks contiguous blocks is dependent on the
2384                          * nrblocks.  So limit nrblocks.
2385                          */
2386                         goto flush_it;
2387                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2388                                 EXT4_MAX_TRANS_DATA) {
2389                         /*
2390                          * Adding the new buffer_head would make it cross the
2391                          * allowed limit for which we have journal credit
2392                          * reserved. So limit the new bh->b_size
2393                          */
2394                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2395                                                 mpd->inode->i_blkbits;
2396                         /* we will do mpage_da_submit_io in the next loop */
2397                 }
2398         }
2399         /*
2400          * First block in the extent
2401          */
2402         if (mpd->b_size == 0) {
2403                 mpd->b_blocknr = logical;
2404                 mpd->b_size = b_size;
2405                 mpd->b_state = b_state & BH_FLAGS;
2406                 return;
2407         }
2408
2409         next = mpd->b_blocknr + nrblocks;
2410         /*
2411          * Can we merge the block to our big extent?
2412          */
2413         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2414                 mpd->b_size += b_size;
2415                 return;
2416         }
2417
2418 flush_it:
2419         /*
2420          * We couldn't merge the block to our extent, so we
2421          * need to flush current  extent and start new one
2422          */
2423         mpage_da_map_and_submit(mpd);
2424         return;
2425 }
2426
2427 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
2428 {
2429         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
2430 }
2431
2432 /*
2433  * __mpage_da_writepage - finds extent of pages and blocks
2434  *
2435  * @page: page to consider
2436  * @wbc: not used, we just follow rules
2437  * @data: context
2438  *
2439  * The function finds extents of pages and scan them for all blocks.
2440  */
2441 static int __mpage_da_writepage(struct page *page,
2442                                 struct writeback_control *wbc,
2443                                 struct mpage_da_data *mpd)
2444 {
2445         struct inode *inode = mpd->inode;
2446         struct buffer_head *bh, *head;
2447         sector_t logical;
2448
2449         /*
2450          * Can we merge this page to current extent?
2451          */
2452         if (mpd->next_page != page->index) {
2453                 /*
2454                  * Nope, we can't. So, we map non-allocated blocks
2455                  * and start IO on them
2456                  */
2457                 if (mpd->next_page != mpd->first_page) {
2458                         mpage_da_map_and_submit(mpd);
2459                         /*
2460                          * skip rest of the page in the page_vec
2461                          */
2462                         redirty_page_for_writepage(wbc, page);
2463                         unlock_page(page);
2464                         return MPAGE_DA_EXTENT_TAIL;
2465                 }
2466
2467                 /*
2468                  * Start next extent of pages ...
2469                  */
2470                 mpd->first_page = page->index;
2471
2472                 /*
2473                  * ... and blocks
2474                  */
2475                 mpd->b_size = 0;
2476                 mpd->b_state = 0;
2477                 mpd->b_blocknr = 0;
2478         }
2479
2480         mpd->next_page = page->index + 1;
2481         logical = (sector_t) page->index <<
2482                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2483
2484         if (!page_has_buffers(page)) {
2485                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2486                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2487                 if (mpd->io_done)
2488                         return MPAGE_DA_EXTENT_TAIL;
2489         } else {
2490                 /*
2491                  * Page with regular buffer heads, just add all dirty ones
2492                  */
2493                 head = page_buffers(page);
2494                 bh = head;
2495                 do {
2496                         BUG_ON(buffer_locked(bh));
2497                         /*
2498                          * We need to try to allocate
2499                          * unmapped blocks in the same page.
2500                          * Otherwise we won't make progress
2501                          * with the page in ext4_writepage
2502                          */
2503                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2504                                 mpage_add_bh_to_extent(mpd, logical,
2505                                                        bh->b_size,
2506                                                        bh->b_state);
2507                                 if (mpd->io_done)
2508                                         return MPAGE_DA_EXTENT_TAIL;
2509                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2510                                 /*
2511                                  * mapped dirty buffer. We need to update
2512                                  * the b_state because we look at
2513                                  * b_state in mpage_da_map_blocks. We don't
2514                                  * update b_size because if we find an
2515                                  * unmapped buffer_head later we need to
2516                                  * use the b_state flag of that buffer_head.
2517                                  */
2518                                 if (mpd->b_size == 0)
2519                                         mpd->b_state = bh->b_state & BH_FLAGS;
2520                         }
2521                         logical++;
2522                 } while ((bh = bh->b_this_page) != head);
2523         }
2524
2525         return 0;
2526 }
2527
2528 /*
2529  * This is a special get_blocks_t callback which is used by
2530  * ext4_da_write_begin().  It will either return mapped block or
2531  * reserve space for a single block.
2532  *
2533  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2534  * We also have b_blocknr = -1 and b_bdev initialized properly
2535  *
2536  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2537  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2538  * initialized properly.
2539  */
2540 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2541                                   struct buffer_head *bh, int create)
2542 {
2543         struct ext4_map_blocks map;
2544         int ret = 0;
2545         sector_t invalid_block = ~((sector_t) 0xffff);
2546
2547         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2548                 invalid_block = ~0;
2549
2550         BUG_ON(create == 0);
2551         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
2552
2553         map.m_lblk = iblock;
2554         map.m_len = 1;
2555
2556         /*
2557          * first, we need to know whether the block is allocated already
2558          * preallocated blocks are unmapped but should treated
2559          * the same as allocated blocks.
2560          */
2561         ret = ext4_map_blocks(NULL, inode, &map, 0);
2562         if (ret < 0)
2563                 return ret;
2564         if (ret == 0) {
2565                 if (buffer_delay(bh))
2566                         return 0; /* Not sure this could or should happen */
2567                 /*
2568                  * XXX: __block_write_begin() unmaps passed block, is it OK?
2569                  */
2570                 ret = ext4_da_reserve_space(inode, iblock);
2571                 if (ret)
2572                         /* not enough space to reserve */
2573                         return ret;
2574
2575                 map_bh(bh, inode->i_sb, invalid_block);
2576                 set_buffer_new(bh);
2577                 set_buffer_delay(bh);
2578                 return 0;
2579         }
2580
2581         map_bh(bh, inode->i_sb, map.m_pblk);
2582         bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
2583
2584         if (buffer_unwritten(bh)) {
2585                 /* A delayed write to unwritten bh should be marked
2586                  * new and mapped.  Mapped ensures that we don't do
2587                  * get_block multiple times when we write to the same
2588                  * offset and new ensures that we do proper zero out
2589                  * for partial write.
2590                  */
2591                 set_buffer_new(bh);
2592                 set_buffer_mapped(bh);
2593         }
2594         return 0;
2595 }
2596
2597 /*
2598  * This function is used as a standard get_block_t calback function
2599  * when there is no desire to allocate any blocks.  It is used as a
2600  * callback function for block_write_begin() and block_write_full_page().
2601  * These functions should only try to map a single block at a time.
2602  *
2603  * Since this function doesn't do block allocations even if the caller
2604  * requests it by passing in create=1, it is critically important that
2605  * any caller checks to make sure that any buffer heads are returned
2606  * by this function are either all already mapped or marked for
2607  * delayed allocation before calling  block_write_full_page().  Otherwise,
2608  * b_blocknr could be left unitialized, and the page write functions will
2609  * be taken by surprise.
2610  */
2611 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2612                                    struct buffer_head *bh_result, int create)
2613 {
2614         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2615         return _ext4_get_block(inode, iblock, bh_result, 0);
2616 }
2617
2618 static int bget_one(handle_t *handle, struct buffer_head *bh)
2619 {
2620         get_bh(bh);
2621         return 0;
2622 }
2623
2624 static int bput_one(handle_t *handle, struct buffer_head *bh)
2625 {
2626         put_bh(bh);
2627         return 0;
2628 }
2629
2630 static int __ext4_journalled_writepage(struct page *page,
2631                                        unsigned int len)
2632 {
2633         struct address_space *mapping = page->mapping;
2634         struct inode *inode = mapping->host;
2635         struct buffer_head *page_bufs;
2636         handle_t *handle = NULL;
2637         int ret = 0;
2638         int err;
2639
2640         ClearPageChecked(page);
2641         page_bufs = page_buffers(page);
2642         BUG_ON(!page_bufs);
2643         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2644         /* As soon as we unlock the page, it can go away, but we have
2645          * references to buffers so we are safe */
2646         unlock_page(page);
2647
2648         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2649         if (IS_ERR(handle)) {
2650                 ret = PTR_ERR(handle);
2651                 goto out;
2652         }
2653
2654         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2655                                 do_journal_get_write_access);
2656
2657         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2658                                 write_end_fn);
2659         if (ret == 0)
2660                 ret = err;
2661         err = ext4_journal_stop(handle);
2662         if (!ret)
2663                 ret = err;
2664
2665         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2666         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2667 out:
2668         return ret;
2669 }
2670
2671 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
2672 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
2673
2674 /*
2675  * Note that we don't need to start a transaction unless we're journaling data
2676  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2677  * need to file the inode to the transaction's list in ordered mode because if
2678  * we are writing back data added by write(), the inode is already there and if
2679  * we are writing back data modified via mmap(), noone guarantees in which
2680  * transaction the data will hit the disk. In case we are journaling data, we
2681  * cannot start transaction directly because transaction start ranks above page
2682  * lock so we have to do some magic.
2683  *
2684  * This function can get called via...
2685  *   - ext4_da_writepages after taking page lock (have journal handle)
2686  *   - journal_submit_inode_data_buffers (no journal handle)
2687  *   - shrink_page_list via pdflush (no journal handle)
2688  *   - grab_page_cache when doing write_begin (have journal handle)
2689  *
2690  * We don't do any block allocation in this function. If we have page with
2691  * multiple blocks we need to write those buffer_heads that are mapped. This
2692  * is important for mmaped based write. So if we do with blocksize 1K
2693  * truncate(f, 1024);
2694  * a = mmap(f, 0, 4096);
2695  * a[0] = 'a';
2696  * truncate(f, 4096);
2697  * we have in the page first buffer_head mapped via page_mkwrite call back
2698  * but other bufer_heads would be unmapped but dirty(dirty done via the
2699  * do_wp_page). So writepage should write the first block. If we modify
2700  * the mmap area beyond 1024 we will again get a page_fault and the
2701  * page_mkwrite callback will do the block allocation and mark the
2702  * buffer_heads mapped.
2703  *
2704  * We redirty the page if we have any buffer_heads that is either delay or
2705  * unwritten in the page.
2706  *
2707  * We can get recursively called as show below.
2708  *
2709  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2710  *              ext4_writepage()
2711  *
2712  * But since we don't do any block allocation we should not deadlock.
2713  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2714  */
2715 static int ext4_writepage(struct page *page,
2716                           struct writeback_control *wbc)
2717 {
2718         int ret = 0, commit_write = 0;
2719         loff_t size;
2720         unsigned int len;
2721         struct buffer_head *page_bufs = NULL;
2722         struct inode *inode = page->mapping->host;
2723
2724         trace_ext4_writepage(inode, page);
2725         size = i_size_read(inode);
2726         if (page->index == size >> PAGE_CACHE_SHIFT)
2727                 len = size & ~PAGE_CACHE_MASK;
2728         else
2729                 len = PAGE_CACHE_SIZE;
2730
2731         /*
2732          * If the page does not have buffers (for whatever reason),
2733          * try to create them using __block_write_begin.  If this
2734          * fails, redirty the page and move on.
2735          */
2736         if (!page_has_buffers(page)) {
2737                 if (__block_write_begin(page, 0, len,
2738                                         noalloc_get_block_write)) {
2739                 redirty_page:
2740                         redirty_page_for_writepage(wbc, page);
2741                         unlock_page(page);
2742                         return 0;
2743                 }
2744                 commit_write = 1;
2745         }
2746         page_bufs = page_buffers(page);
2747         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2748                               ext4_bh_delay_or_unwritten)) {
2749                 /*
2750                  * We don't want to do block allocation, so redirty
2751                  * the page and return.  We may reach here when we do
2752                  * a journal commit via journal_submit_inode_data_buffers.
2753                  * We can also reach here via shrink_page_list
2754                  */
2755                 goto redirty_page;
2756         }
2757         if (commit_write)
2758                 /* now mark the buffer_heads as dirty and uptodate */
2759                 block_commit_write(page, 0, len);
2760
2761         if (PageChecked(page) && ext4_should_journal_data(inode))
2762                 /*
2763                  * It's mmapped pagecache.  Add buffers and journal it.  There
2764                  * doesn't seem much point in redirtying the page here.
2765                  */
2766                 return __ext4_journalled_writepage(page, len);
2767
2768         if (buffer_uninit(page_bufs)) {
2769                 ext4_set_bh_endio(page_bufs, inode);
2770                 ret = block_write_full_page_endio(page, noalloc_get_block_write,
2771                                             wbc, ext4_end_io_buffer_write);
2772         } else
2773                 ret = block_write_full_page(page, noalloc_get_block_write,
2774                                             wbc);
2775
2776         return ret;
2777 }
2778
2779 /*
2780  * This is called via ext4_da_writepages() to
2781  * calulate the total number of credits to reserve to fit
2782  * a single extent allocation into a single transaction,
2783  * ext4_da_writpeages() will loop calling this before
2784  * the block allocation.
2785  */
2786
2787 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2788 {
2789         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2790
2791         /*
2792          * With non-extent format the journal credit needed to
2793          * insert nrblocks contiguous block is dependent on
2794          * number of contiguous block. So we will limit
2795          * number of contiguous block to a sane value
2796          */
2797         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) &&
2798             (max_blocks > EXT4_MAX_TRANS_DATA))
2799                 max_blocks = EXT4_MAX_TRANS_DATA;
2800
2801         return ext4_chunk_trans_blocks(inode, max_blocks);
2802 }
2803
2804 /*
2805  * write_cache_pages_da - walk the list of dirty pages of the given
2806  * address space and call the callback function (which usually writes
2807  * the pages).
2808  *
2809  * This is a forked version of write_cache_pages().  Differences:
2810  *      Range cyclic is ignored.
2811  *      no_nrwrite_index_update is always presumed true
2812  */
2813 static int write_cache_pages_da(struct address_space *mapping,
2814                                 struct writeback_control *wbc,
2815                                 struct mpage_da_data *mpd,
2816                                 pgoff_t *done_index)
2817 {
2818         int ret = 0;
2819         int done = 0;
2820         struct pagevec pvec;
2821         unsigned nr_pages;
2822         pgoff_t index;
2823         pgoff_t end;            /* Inclusive */
2824         long nr_to_write = wbc->nr_to_write;
2825         int tag;
2826
2827         pagevec_init(&pvec, 0);
2828         index = wbc->range_start >> PAGE_CACHE_SHIFT;
2829         end = wbc->range_end >> PAGE_CACHE_SHIFT;
2830
2831         if (wbc->sync_mode == WB_SYNC_ALL)
2832                 tag = PAGECACHE_TAG_TOWRITE;
2833         else
2834                 tag = PAGECACHE_TAG_DIRTY;
2835
2836         *done_index = index;
2837         while (!done && (index <= end)) {
2838                 int i;
2839
2840                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2841                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2842                 if (nr_pages == 0)
2843                         break;
2844
2845                 for (i = 0; i < nr_pages; i++) {
2846                         struct page *page = pvec.pages[i];
2847
2848                         /*
2849                          * At this point, the page may be truncated or
2850                          * invalidated (changing page->mapping to NULL), or
2851                          * even swizzled back from swapper_space to tmpfs file
2852                          * mapping. However, page->index will not change
2853                          * because we have a reference on the page.
2854                          */
2855                         if (page->index > end) {
2856                                 done = 1;
2857                                 break;
2858                         }
2859
2860                         *done_index = page->index + 1;
2861
2862                         lock_page(page);
2863
2864                         /*
2865                          * Page truncated or invalidated. We can freely skip it
2866                          * then, even for data integrity operations: the page
2867                          * has disappeared concurrently, so there could be no
2868                          * real expectation of this data interity operation
2869                          * even if there is now a new, dirty page at the same
2870                          * pagecache address.
2871                          */
2872                         if (unlikely(page->mapping != mapping)) {
2873 continue_unlock:
2874                                 unlock_page(page);
2875                                 continue;
2876                         }
2877
2878                         if (!PageDirty(page)) {
2879                                 /* someone wrote it for us */
2880                                 goto continue_unlock;
2881                         }
2882
2883                         if (PageWriteback(page)) {
2884                                 if (wbc->sync_mode != WB_SYNC_NONE)
2885                                         wait_on_page_writeback(page);
2886                                 else
2887                                         goto continue_unlock;
2888                         }
2889
2890                         BUG_ON(PageWriteback(page));
2891                         if (!clear_page_dirty_for_io(page))
2892                                 goto continue_unlock;
2893
2894                         ret = __mpage_da_writepage(page, wbc, mpd);
2895                         if (unlikely(ret)) {
2896                                 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2897                                         unlock_page(page);
2898                                         ret = 0;
2899                                 } else {
2900                                         done = 1;
2901                                         break;
2902                                 }
2903                         }
2904
2905                         if (nr_to_write > 0) {
2906                                 nr_to_write--;
2907                                 if (nr_to_write == 0 &&
2908                                     wbc->sync_mode == WB_SYNC_NONE) {
2909                                         /*
2910                                          * We stop writing back only if we are
2911                                          * not doing integrity sync. In case of
2912                                          * integrity sync we have to keep going
2913                                          * because someone may be concurrently
2914                                          * dirtying pages, and we might have
2915                                          * synced a lot of newly appeared dirty
2916                                          * pages, but have not synced all of the
2917                                          * old dirty pages.
2918                                          */
2919                                         done = 1;
2920                                         break;
2921                                 }
2922                         }
2923                 }
2924                 pagevec_release(&pvec);
2925                 cond_resched();
2926         }
2927         return ret;
2928 }
2929
2930
2931 static int ext4_da_writepages(struct address_space *mapping,
2932                               struct writeback_control *wbc)
2933 {
2934         pgoff_t index;
2935         int range_whole = 0;
2936         handle_t *handle = NULL;
2937         struct mpage_da_data mpd;
2938         struct inode *inode = mapping->host;
2939         int pages_written = 0;
2940         long pages_skipped;
2941         unsigned int max_pages;
2942         int range_cyclic, cycled = 1, io_done = 0;
2943         int needed_blocks, ret = 0;
2944         long desired_nr_to_write, nr_to_writebump = 0;
2945         loff_t range_start = wbc->range_start;
2946         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2947         pgoff_t done_index = 0;
2948         pgoff_t end;
2949
2950         trace_ext4_da_writepages(inode, wbc);
2951
2952         /*
2953          * No pages to write? This is mainly a kludge to avoid starting
2954          * a transaction for special inodes like journal inode on last iput()
2955          * because that could violate lock ordering on umount
2956          */
2957         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2958                 return 0;
2959
2960         /*
2961          * If the filesystem has aborted, it is read-only, so return
2962          * right away instead of dumping stack traces later on that
2963          * will obscure the real source of the problem.  We test
2964          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2965          * the latter could be true if the filesystem is mounted
2966          * read-only, and in that case, ext4_da_writepages should
2967          * *never* be called, so if that ever happens, we would want
2968          * the stack trace.
2969          */
2970         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2971                 return -EROFS;
2972
2973         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2974                 range_whole = 1;
2975
2976         range_cyclic = wbc->range_cyclic;
2977         if (wbc->range_cyclic) {
2978                 index = mapping->writeback_index;
2979                 if (index)
2980                         cycled = 0;
2981                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2982                 wbc->range_end  = LLONG_MAX;
2983                 wbc->range_cyclic = 0;
2984                 end = -1;
2985         } else {
2986                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2987                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2988         }
2989
2990         /*
2991          * This works around two forms of stupidity.  The first is in
2992          * the writeback code, which caps the maximum number of pages
2993          * written to be 1024 pages.  This is wrong on multiple
2994          * levels; different architectues have a different page size,
2995          * which changes the maximum amount of data which gets
2996          * written.  Secondly, 4 megabytes is way too small.  XFS
2997          * forces this value to be 16 megabytes by multiplying
2998          * nr_to_write parameter by four, and then relies on its
2999          * allocator to allocate larger extents to make them
3000          * contiguous.  Unfortunately this brings us to the second
3001          * stupidity, which is that ext4's mballoc code only allocates
3002          * at most 2048 blocks.  So we force contiguous writes up to
3003          * the number of dirty blocks in the inode, or
3004          * sbi->max_writeback_mb_bump whichever is smaller.
3005          */
3006         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
3007         if (!range_cyclic && range_whole) {
3008                 if (wbc->nr_to_write == LONG_MAX)
3009                         desired_nr_to_write = wbc->nr_to_write;
3010                 else
3011                         desired_nr_to_write = wbc->nr_to_write * 8;
3012         } else
3013                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
3014                                                            max_pages);
3015         if (desired_nr_to_write > max_pages)
3016                 desired_nr_to_write = max_pages;
3017
3018         if (wbc->nr_to_write < desired_nr_to_write) {
3019                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
3020                 wbc->nr_to_write = desired_nr_to_write;
3021         }
3022
3023         mpd.wbc = wbc;
3024         mpd.inode = mapping->host;
3025
3026         pages_skipped = wbc->pages_skipped;
3027
3028 retry:
3029         if (wbc->sync_mode == WB_SYNC_ALL)
3030                 tag_pages_for_writeback(mapping, index, end);
3031
3032         while (!ret && wbc->nr_to_write > 0) {
3033
3034                 /*
3035                  * we  insert one extent at a time. So we need
3036                  * credit needed for single extent allocation.
3037                  * journalled mode is currently not supported
3038                  * by delalloc
3039                  */
3040                 BUG_ON(ext4_should_journal_data(inode));
3041                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
3042
3043                 /* start a new transaction*/
3044                 handle = ext4_journal_start(inode, needed_blocks);
3045                 if (IS_ERR(handle)) {
3046                         ret = PTR_ERR(handle);
3047                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
3048                                "%ld pages, ino %lu; err %d", __func__,
3049                                 wbc->nr_to_write, inode->i_ino, ret);
3050                         goto out_writepages;
3051                 }
3052
3053                 /*
3054                  * Now call __mpage_da_writepage to find the next
3055                  * contiguous region of logical blocks that need
3056                  * blocks to be allocated by ext4.  We don't actually
3057                  * submit the blocks for I/O here, even though
3058                  * write_cache_pages thinks it will, and will set the
3059                  * pages as clean for write before calling
3060                  * __mpage_da_writepage().
3061                  */
3062                 mpd.b_size = 0;
3063                 mpd.b_state = 0;
3064                 mpd.b_blocknr = 0;
3065                 mpd.first_page = 0;
3066                 mpd.next_page = 0;
3067                 mpd.io_done = 0;
3068                 mpd.pages_written = 0;
3069                 mpd.retval = 0;
3070                 ret = write_cache_pages_da(mapping, wbc, &mpd, &done_index);
3071                 /*
3072                  * If we have a contiguous extent of pages and we
3073                  * haven't done the I/O yet, map the blocks and submit
3074                  * them for I/O.
3075                  */
3076                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
3077                         mpage_da_map_and_submit(&mpd);
3078                         ret = MPAGE_DA_EXTENT_TAIL;
3079                 }
3080                 trace_ext4_da_write_pages(inode, &mpd);
3081                 wbc->nr_to_write -= mpd.pages_written;
3082
3083                 ext4_journal_stop(handle);
3084
3085                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
3086                         /* commit the transaction which would
3087                          * free blocks released in the transaction
3088                          * and try again
3089                          */
3090                         jbd2_journal_force_commit_nested(sbi->s_journal);
3091                         wbc->pages_skipped = pages_skipped;
3092                         ret = 0;
3093                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
3094                         /*
3095                          * got one extent now try with
3096                          * rest of the pages
3097                          */
3098                         pages_written += mpd.pages_written;
3099                         wbc->pages_skipped = pages_skipped;
3100                         ret = 0;
3101                         io_done = 1;
3102                 } else if (wbc->nr_to_write)
3103                         /*
3104                          * There is no more writeout needed
3105                          * or we requested for a noblocking writeout
3106                          * and we found the device congested
3107                          */
3108                         break;
3109         }
3110         if (!io_done && !cycled) {
3111                 cycled = 1;
3112                 index = 0;
3113                 wbc->range_start = index << PAGE_CACHE_SHIFT;
3114                 wbc->range_end  = mapping->writeback_index - 1;
3115                 goto retry;
3116         }
3117         if (pages_skipped != wbc->pages_skipped)
3118                 ext4_msg(inode->i_sb, KERN_CRIT,
3119                          "This should not happen leaving %s "
3120                          "with nr_to_write = %ld ret = %d",
3121                          __func__, wbc->nr_to_write, ret);
3122
3123         /* Update index */
3124         wbc->range_cyclic = range_cyclic;
3125         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
3126                 /*
3127                  * set the writeback_index so that range_cyclic
3128                  * mode will write it back later
3129                  */
3130                 mapping->writeback_index = done_index;
3131
3132 out_writepages:
3133         wbc->nr_to_write -= nr_to_writebump;
3134         wbc->range_start = range_start;
3135         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
3136         return ret;
3137 }
3138
3139 #define FALL_BACK_TO_NONDELALLOC 1
3140 static int ext4_nonda_switch(struct super_block *sb)
3141 {
3142         s64 free_blocks, dirty_blocks;
3143         struct ext4_sb_info *sbi = EXT4_SB(sb);
3144
3145         /*
3146          * switch to non delalloc mode if we are running low
3147          * on free block. The free block accounting via percpu
3148          * counters can get slightly wrong with percpu_counter_batch getting
3149          * accumulated on each CPU without updating global counters
3150          * Delalloc need an accurate free block accounting. So switch
3151          * to non delalloc when we are near to error range.
3152          */
3153         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
3154         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
3155         if (2 * free_blocks < 3 * dirty_blocks ||
3156                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
3157                 /*
3158                  * free block count is less than 150% of dirty blocks
3159                  * or free blocks is less than watermark
3160                  */
3161                 return 1;
3162         }
3163         /*
3164          * Even if we don't switch but are nearing capacity,
3165          * start pushing delalloc when 1/2 of free blocks are dirty.
3166          */
3167         if (free_blocks < 2 * dirty_blocks)
3168                 writeback_inodes_sb_if_idle(sb);
3169
3170         return 0;
3171 }
3172
3173 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3174                                loff_t pos, unsigned len, unsigned flags,
3175                                struct page **pagep, void **fsdata)
3176 {
3177         int ret, retries = 0;
3178         struct page *page;
3179         pgoff_t index;
3180         struct inode *inode = mapping->host;
3181         handle_t *handle;
3182
3183         index = pos >> PAGE_CACHE_SHIFT;
3184
3185         if (ext4_nonda_switch(inode->i_sb)) {
3186                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3187                 return ext4_write_begin(file, mapping, pos,
3188                                         len, flags, pagep, fsdata);
3189         }
3190         *fsdata = (void *)0;
3191         trace_ext4_da_write_begin(inode, pos, len, flags);
3192 retry:
3193         /*
3194          * With delayed allocation, we don't log the i_disksize update
3195          * if there is delayed block allocation. But we still need
3196          * to journalling the i_disksize update if writes to the end
3197          * of file which has an already mapped buffer.
3198          */
3199         handle = ext4_journal_start(inode, 1);
3200         if (IS_ERR(handle)) {
3201                 ret = PTR_ERR(handle);
3202                 goto out;
3203         }
3204         /* We cannot recurse into the filesystem as the transaction is already
3205          * started */
3206         flags |= AOP_FLAG_NOFS;
3207
3208         page = grab_cache_page_write_begin(mapping, index, flags);
3209         if (!page) {
3210                 ext4_journal_stop(handle);
3211                 ret = -ENOMEM;
3212                 goto out;
3213         }
3214         *pagep = page;
3215
3216         ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
3217         if (ret < 0) {
3218                 unlock_page(page);
3219                 ext4_journal_stop(handle);
3220                 page_cache_release(page);
3221                 /*
3222                  * block_write_begin may have instantiated a few blocks
3223                  * outside i_size.  Trim these off again. Don't need
3224                  * i_size_read because we hold i_mutex.
3225                  */
3226                 if (pos + len > inode->i_size)
3227                         ext4_truncate_failed_write(inode);
3228         }
3229
3230         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3231                 goto retry;
3232 out:
3233         return ret;
3234 }
3235
3236 /*
3237  * Check if we should update i_disksize
3238  * when write to the end of file but not require block allocation
3239  */
3240 static int ext4_da_should_update_i_disksize(struct page *page,
3241                                             unsigned long offset)
3242 {
3243         struct buffer_head *bh;
3244         struct inode *inode = page->mapping->host;
3245         unsigned int idx;
3246         int i;
3247
3248         bh = page_buffers(page);
3249         idx = offset >> inode->i_blkbits;
3250
3251         for (i = 0; i < idx; i++)
3252                 bh = bh->b_this_page;
3253
3254         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3255                 return 0;
3256         return 1;
3257 }
3258
3259 static int ext4_da_write_end(struct file *file,
3260                              struct address_space *mapping,
3261                              loff_t pos, unsigned len, unsigned copied,
3262                              struct page *page, void *fsdata)
3263 {
3264         struct inode *inode = mapping->host;
3265         int ret = 0, ret2;
3266         handle_t *handle = ext4_journal_current_handle();
3267         loff_t new_i_size;
3268         unsigned long start, end;
3269         int write_mode = (int)(unsigned long)fsdata;
3270
3271         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
3272                 if (ext4_should_order_data(inode)) {
3273                         return ext4_ordered_write_end(file, mapping, pos,
3274                                         len, copied, page, fsdata);
3275                 } else if (ext4_should_writeback_data(inode)) {
3276                         return ext4_writeback_write_end(file, mapping, pos,
3277                                         len, copied, page, fsdata);
3278                 } else {
3279                         BUG();
3280                 }
3281         }
3282
3283         trace_ext4_da_write_end(inode, pos, len, copied);
3284         start = pos & (PAGE_CACHE_SIZE - 1);
3285         end = start + copied - 1;
3286
3287         /*
3288          * generic_write_end() will run mark_inode_dirty() if i_size
3289          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3290          * into that.
3291          */
3292
3293         new_i_size = pos + copied;
3294         if (new_i_size > EXT4_I(inode)->i_disksize) {
3295                 if (ext4_da_should_update_i_disksize(page, end)) {
3296                         down_write(&EXT4_I(inode)->i_data_sem);
3297                         if (new_i_size > EXT4_I(inode)->i_disksize) {
3298                                 /*
3299                                  * Updating i_disksize when extending file
3300                                  * without needing block allocation
3301                                  */
3302                                 if (ext4_should_order_data(inode))
3303                                         ret = ext4_jbd2_file_inode(handle,
3304                                                                    inode);
3305
3306                                 EXT4_I(inode)->i_disksize = new_i_size;
3307                         }
3308                         up_write(&EXT4_I(inode)->i_data_sem);
3309                         /* We need to mark inode dirty even if
3310                          * new_i_size is less that inode->i_size
3311                          * bu greater than i_disksize.(hint delalloc)
3312                          */
3313                         ext4_mark_inode_dirty(handle, inode);
3314                 }
3315         }
3316         ret2 = generic_write_end(file, mapping, pos, len, copied,
3317                                                         page, fsdata);
3318         copied = ret2;
3319         if (ret2 < 0)
3320                 ret = ret2;
3321         ret2 = ext4_journal_stop(handle);
3322         if (!ret)
3323                 ret = ret2;
3324
3325         return ret ? ret : copied;
3326 }
3327
3328 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
3329 {
3330         /*
3331          * Drop reserved blocks
3332          */
3333         BUG_ON(!PageLocked(page));
3334         if (!page_has_buffers(page))
3335                 goto out;
3336
3337         ext4_da_page_release_reservation(page, offset);
3338
3339 out:
3340         ext4_invalidatepage(page, offset);
3341
3342         return;
3343 }
3344
3345 /*
3346  * Force all delayed allocation blocks to be allocated for a given inode.
3347  */
3348 int ext4_alloc_da_blocks(struct inode *inode)
3349 {
3350         trace_ext4_alloc_da_blocks(inode);
3351
3352         if (!EXT4_I(inode)->i_reserved_data_blocks &&
3353             !EXT4_I(inode)->i_reserved_meta_blocks)
3354                 return 0;
3355
3356         /*
3357          * We do something simple for now.  The filemap_flush() will
3358          * also start triggering a write of the data blocks, which is
3359          * not strictly speaking necessary (and for users of
3360          * laptop_mode, not even desirable).  However, to do otherwise
3361          * would require replicating code paths in:
3362          *
3363          * ext4_da_writepages() ->
3364          *    write_cache_pages() ---> (via passed in callback function)
3365          *        __mpage_da_writepage() -->
3366          *           mpage_add_bh_to_extent()
3367          *           mpage_da_map_blocks()
3368          *
3369          * The problem is that write_cache_pages(), located in
3370          * mm/page-writeback.c, marks pages clean in preparation for
3371          * doing I/O, which is not desirable if we're not planning on
3372          * doing I/O at all.
3373          *
3374          * We could call write_cache_pages(), and then redirty all of
3375          * the pages by calling redirty_page_for_writeback() but that
3376          * would be ugly in the extreme.  So instead we would need to
3377          * replicate parts of the code in the above functions,
3378          * simplifying them becuase we wouldn't actually intend to
3379          * write out the pages, but rather only collect contiguous
3380          * logical block extents, call the multi-block allocator, and
3381          * then update the buffer heads with the block allocations.
3382          *
3383          * For now, though, we'll cheat by calling filemap_flush(),
3384          * which will map the blocks, and start the I/O, but not
3385          * actually wait for the I/O to complete.
3386          */
3387         return filemap_flush(inode->i_mapping);
3388 }
3389
3390 /*
3391  * bmap() is special.  It gets used by applications such as lilo and by
3392  * the swapper to find the on-disk block of a specific piece of data.
3393  *
3394  * Naturally, this is dangerous if the block concerned is still in the
3395  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3396  * filesystem and enables swap, then they may get a nasty shock when the
3397  * data getting swapped to that swapfile suddenly gets overwritten by
3398  * the original zero's written out previously to the journal and
3399  * awaiting writeback in the kernel's buffer cache.
3400  *
3401  * So, if we see any bmap calls here on a modified, data-journaled file,
3402  * take extra steps to flush any blocks which might be in the cache.
3403  */
3404 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3405 {
3406         struct inode *inode = mapping->host;
3407         journal_t *journal;
3408         int err;
3409
3410         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3411                         test_opt(inode->i_sb, DELALLOC)) {
3412                 /*
3413                  * With delalloc we want to sync the file
3414                  * so that we can make sure we allocate
3415                  * blocks for file
3416                  */
3417                 filemap_write_and_wait(mapping);
3418         }
3419
3420         if (EXT4_JOURNAL(inode) &&
3421             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3422                 /*
3423                  * This is a REALLY heavyweight approach, but the use of
3424                  * bmap on dirty files is expected to be extremely rare:
3425                  * only if we run lilo or swapon on a freshly made file
3426                  * do we expect this to happen.
3427                  *
3428                  * (bmap requires CAP_SYS_RAWIO so this does not
3429                  * represent an unprivileged user DOS attack --- we'd be
3430                  * in trouble if mortal users could trigger this path at
3431                  * will.)
3432                  *
3433                  * NB. EXT4_STATE_JDATA is not set on files other than
3434                  * regular files.  If somebody wants to bmap a directory
3435                  * or symlink and gets confused because the buffer
3436                  * hasn't yet been flushed to disk, they deserve
3437                  * everything they get.
3438                  */
3439
3440                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3441                 journal = EXT4_JOURNAL(inode);
3442                 jbd2_journal_lock_updates(journal);
3443                 err = jbd2_journal_flush(journal);
3444                 jbd2_journal_unlock_updates(journal);
3445
3446                 if (err)
3447                         return 0;
3448         }
3449
3450         return generic_block_bmap(mapping, block, ext4_get_block);
3451 }
3452
3453 static int ext4_readpage(struct file *file, struct page *page)
3454 {
3455         return mpage_readpage(page, ext4_get_block);
3456 }
3457
3458 static int
3459 ext4_readpages(struct file *file, struct address_space *mapping,
3460                 struct list_head *pages, unsigned nr_pages)
3461 {
3462         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3463 }
3464
3465 static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
3466 {
3467         struct buffer_head *head, *bh;
3468         unsigned int curr_off = 0;
3469
3470         if (!page_has_buffers(page))
3471                 return;
3472         head = bh = page_buffers(page);
3473         do {
3474                 if (offset <= curr_off && test_clear_buffer_uninit(bh)
3475                                         && bh->b_private) {
3476                         ext4_free_io_end(bh->b_private);
3477                         bh->b_private = NULL;
3478                         bh->b_end_io = NULL;
3479                 }
3480                 curr_off = curr_off + bh->b_size;
3481                 bh = bh->b_this_page;
3482         } while (bh != head);
3483 }
3484
3485 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3486 {
3487         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3488
3489         /*
3490          * free any io_end structure allocated for buffers to be discarded
3491          */
3492         if (ext4_should_dioread_nolock(page->mapping->host))
3493                 ext4_invalidatepage_free_endio(page, offset);
3494         /*
3495          * If it's a full truncate we just forget about the pending dirtying
3496          */
3497         if (offset == 0)
3498                 ClearPageChecked(page);
3499
3500         if (journal)
3501                 jbd2_journal_invalidatepage(journal, page, offset);
3502         else
3503                 block_invalidatepage(page, offset);
3504 }
3505
3506 static int ext4_releasepage(struct page *page, gfp_t wait)
3507 {
3508         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3509
3510         WARN_ON(PageChecked(page));
3511         if (!page_has_buffers(page))
3512                 return 0;
3513         if (journal)
3514                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3515         else
3516                 return try_to_free_buffers(page);
3517 }
3518
3519 /*
3520  * O_DIRECT for ext3 (or indirect map) based files
3521  *
3522  * If the O_DIRECT write will extend the file then add this inode to the
3523  * orphan list.  So recovery will truncate it back to the original size
3524  * if the machine crashes during the write.
3525  *
3526  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3527  * crashes then stale disk data _may_ be exposed inside the file. But current
3528  * VFS code falls back into buffered path in that case so we are safe.
3529  */
3530 static ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
3531                               const struct iovec *iov, loff_t offset,
3532                               unsigned long nr_segs)
3533 {
3534         struct file *file = iocb->ki_filp;
3535         struct inode *inode = file->f_mapping->host;
3536         struct ext4_inode_info *ei = EXT4_I(inode);
3537         handle_t *handle;
3538         ssize_t ret;
3539         int orphan = 0;
3540         size_t count = iov_length(iov, nr_segs);
3541         int retries = 0;
3542
3543         if (rw == WRITE) {
3544                 loff_t final_size = offset + count;
3545
3546                 if (final_size > inode->i_size) {
3547                         /* Credits for sb + inode write */
3548                         handle = ext4_journal_start(inode, 2);
3549                         if (IS_ERR(handle)) {
3550                                 ret = PTR_ERR(handle);
3551                                 goto out;
3552                         }
3553                         ret = ext4_orphan_add(handle, inode);
3554                         if (ret) {
3555                                 ext4_journal_stop(handle);
3556                                 goto out;
3557                         }
3558                         orphan = 1;
3559                         ei->i_disksize = inode->i_size;
3560                         ext4_journal_stop(handle);
3561                 }
3562         }
3563
3564 retry:
3565         if (rw == READ && ext4_should_dioread_nolock(inode))
3566                 ret = __blockdev_direct_IO(rw, iocb, inode,
3567                                  inode->i_sb->s_bdev, iov,
3568                                  offset, nr_segs,
3569                                  ext4_get_block, NULL, NULL, 0);
3570         else {
3571                 ret = blockdev_direct_IO(rw, iocb, inode,
3572                                  inode->i_sb->s_bdev, iov,
3573                                  offset, nr_segs,
3574                                  ext4_get_block, NULL);
3575
3576                 if (unlikely((rw & WRITE) && ret < 0)) {
3577                         loff_t isize = i_size_read(inode);
3578                         loff_t end = offset + iov_length(iov, nr_segs);
3579
3580                         if (end > isize)
3581                                 vmtruncate(inode, isize);
3582                 }
3583         }
3584         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3585                 goto retry;
3586
3587         if (orphan) {
3588                 int err;
3589
3590                 /* Credits for sb + inode write */
3591                 handle = ext4_journal_start(inode, 2);
3592                 if (IS_ERR(handle)) {
3593                         /* This is really bad luck. We've written the data
3594                          * but cannot extend i_size. Bail out and pretend
3595                          * the write failed... */
3596                         ret = PTR_ERR(handle);
3597                         if (inode->i_nlink)
3598                                 ext4_orphan_del(NULL, inode);
3599
3600                         goto out;
3601                 }
3602                 if (inode->i_nlink)
3603                         ext4_orphan_del(handle, inode);
3604                 if (ret > 0) {
3605                         loff_t end = offset + ret;
3606                         if (end > inode->i_size) {
3607                                 ei->i_disksize = end;
3608                                 i_size_write(inode, end);
3609                                 /*
3610                                  * We're going to return a positive `ret'
3611                                  * here due to non-zero-length I/O, so there's
3612                                  * no way of reporting error returns from
3613                                  * ext4_mark_inode_dirty() to userspace.  So
3614                                  * ignore it.
3615                                  */
3616                                 ext4_mark_inode_dirty(handle, inode);
3617                         }
3618                 }
3619                 err = ext4_journal_stop(handle);
3620                 if (ret == 0)
3621                         ret = err;
3622         }
3623 out:
3624         return ret;
3625 }
3626
3627 /*
3628  * ext4_get_block used when preparing for a DIO write or buffer write.
3629  * We allocate an uinitialized extent if blocks haven't been allocated.
3630  * The extent will be converted to initialized after the IO is complete.
3631  */
3632 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
3633                    struct buffer_head *bh_result, int create)
3634 {
3635         ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
3636                    inode->i_ino, create);
3637         return _ext4_get_block(inode, iblock, bh_result,
3638                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
3639 }
3640
3641 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3642                             ssize_t size, void *private, int ret,
3643                             bool is_async)
3644 {
3645         ext4_io_end_t *io_end = iocb->private;
3646         struct workqueue_struct *wq;
3647         unsigned long flags;
3648         struct ext4_inode_info *ei;
3649
3650         /* if not async direct IO or dio with 0 bytes write, just return */
3651         if (!io_end || !size)
3652                 goto out;
3653
3654         ext_debug("ext4_end_io_dio(): io_end 0x%p"
3655                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
3656                   iocb->private, io_end->inode->i_ino, iocb, offset,
3657                   size);
3658
3659         /* if not aio dio with unwritten extents, just free io and return */
3660         if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
3661                 ext4_free_io_end(io_end);
3662                 iocb->private = NULL;
3663 out:
3664                 if (is_async)
3665                         aio_complete(iocb, ret, 0);
3666                 return;
3667         }
3668
3669         io_end->offset = offset;
3670         io_end->size = size;
3671         if (is_async) {
3672                 io_end->iocb = iocb;
3673                 io_end->result = ret;
3674         }
3675         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
3676
3677         /* Add the io_end to per-inode completed aio dio list*/
3678         ei = EXT4_I(io_end->inode);
3679         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3680         list_add_tail(&io_end->list, &ei->i_completed_io_list);
3681         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3682
3683         /* queue the work to convert unwritten extents to written */
3684         queue_work(wq, &io_end->work);
3685         iocb->private = NULL;
3686 }
3687
3688 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
3689 {
3690         ext4_io_end_t *io_end = bh->b_private;
3691         struct workqueue_struct *wq;
3692         struct inode *inode;
3693         unsigned long flags;
3694
3695         if (!test_clear_buffer_uninit(bh) || !io_end)
3696                 goto out;
3697
3698         if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
3699                 printk("sb umounted, discard end_io request for inode %lu\n",
3700                         io_end->inode->i_ino);
3701                 ext4_free_io_end(io_end);
3702                 goto out;
3703         }
3704
3705         io_end->flag = EXT4_IO_END_UNWRITTEN;
3706         inode = io_end->inode;
3707
3708         /* Add the io_end to per-inode completed io list*/
3709         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3710         list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
3711         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3712
3713         wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
3714         /* queue the work to convert unwritten extents to written */
3715         queue_work(wq, &io_end->work);
3716 out:
3717         bh->b_private = NULL;
3718         bh->b_end_io = NULL;
3719         clear_buffer_uninit(bh);
3720         end_buffer_async_write(bh, uptodate);
3721 }
3722
3723 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
3724 {
3725         ext4_io_end_t *io_end;
3726         struct page *page = bh->b_page;
3727         loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
3728         size_t size = bh->b_size;
3729
3730 retry:
3731         io_end = ext4_init_io_end(inode, GFP_ATOMIC);
3732         if (!io_end) {
3733                 pr_warning_ratelimited("%s: allocation fail\n", __func__);
3734                 schedule();
3735                 goto retry;
3736         }
3737         io_end->offset = offset;
3738         io_end->size = size;
3739         /*
3740          * We need to hold a reference to the page to make sure it
3741          * doesn't get evicted before ext4_end_io_work() has a chance
3742          * to convert the extent from written to unwritten.
3743          */
3744         io_end->page = page;
3745         get_page(io_end->page);
3746
3747         bh->b_private = io_end;
3748         bh->b_end_io = ext4_end_io_buffer_write;
3749         return 0;
3750 }
3751
3752 /*
3753  * For ext4 extent files, ext4 will do direct-io write to holes,
3754  * preallocated extents, and those write extend the file, no need to
3755  * fall back to buffered IO.
3756  *
3757  * For holes, we fallocate those blocks, mark them as unintialized
3758  * If those blocks were preallocated, we mark sure they are splited, but
3759  * still keep the range to write as unintialized.
3760  *
3761  * The unwrritten extents will be converted to written when DIO is completed.
3762  * For async direct IO, since the IO may still pending when return, we
3763  * set up an end_io call back function, which will do the convertion
3764  * when async direct IO completed.
3765  *
3766  * If the O_DIRECT write will extend the file then add this inode to the
3767  * orphan list.  So recovery will truncate it back to the original size
3768  * if the machine crashes during the write.
3769  *
3770  */
3771 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
3772                               const struct iovec *iov, loff_t offset,
3773                               unsigned long nr_segs)
3774 {
3775         struct file *file = iocb->ki_filp;
3776         struct inode *inode = file->f_mapping->host;
3777         ssize_t ret;
3778         size_t count = iov_length(iov, nr_segs);
3779
3780         loff_t final_size = offset + count;
3781         if (rw == WRITE && final_size <= inode->i_size) {
3782                 /*
3783                  * We could direct write to holes and fallocate.
3784                  *
3785                  * Allocated blocks to fill the hole are marked as uninitialized
3786                  * to prevent paralel buffered read to expose the stale data
3787                  * before DIO complete the data IO.
3788                  *
3789                  * As to previously fallocated extents, ext4 get_block
3790                  * will just simply mark the buffer mapped but still
3791                  * keep the extents uninitialized.
3792                  *
3793                  * for non AIO case, we will convert those unwritten extents
3794                  * to written after return back from blockdev_direct_IO.
3795                  *
3796                  * for async DIO, the conversion needs to be defered when
3797                  * the IO is completed. The ext4 end_io callback function
3798                  * will be called to take care of the conversion work.
3799                  * Here for async case, we allocate an io_end structure to
3800                  * hook to the iocb.
3801                  */
3802                 iocb->private = NULL;
3803                 EXT4_I(inode)->cur_aio_dio = NULL;
3804                 if (!is_sync_kiocb(iocb)) {
3805                         iocb->private = ext4_init_io_end(inode, GFP_NOFS);
3806                         if (!iocb->private)
3807                                 return -ENOMEM;
3808                         /*
3809                          * we save the io structure for current async
3810                          * direct IO, so that later ext4_map_blocks()
3811                          * could flag the io structure whether there
3812                          * is a unwritten extents needs to be converted
3813                          * when IO is completed.
3814                          */
3815                         EXT4_I(inode)->cur_aio_dio = iocb->private;
3816                 }
3817
3818                 ret = blockdev_direct_IO(rw, iocb, inode,
3819                                          inode->i_sb->s_bdev, iov,
3820                                          offset, nr_segs,
3821                                          ext4_get_block_write,
3822                                          ext4_end_io_dio);
3823                 if (iocb->private)
3824                         EXT4_I(inode)->cur_aio_dio = NULL;
3825                 /*
3826                  * The io_end structure takes a reference to the inode,
3827                  * that structure needs to be destroyed and the
3828                  * reference to the inode need to be dropped, when IO is
3829                  * complete, even with 0 byte write, or failed.
3830                  *
3831                  * In the successful AIO DIO case, the io_end structure will be
3832                  * desctroyed and the reference to the inode will be dropped
3833                  * after the end_io call back function is called.
3834                  *
3835                  * In the case there is 0 byte write, or error case, since
3836                  * VFS direct IO won't invoke the end_io call back function,
3837                  * we need to free the end_io structure here.
3838                  */
3839                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
3840                         ext4_free_io_end(iocb->private);
3841                         iocb->private = NULL;
3842                 } else if (ret > 0 && ext4_test_inode_state(inode,
3843                                                 EXT4_STATE_DIO_UNWRITTEN)) {
3844                         int err;
3845                         /*
3846                          * for non AIO case, since the IO is already
3847                          * completed, we could do the convertion right here
3848                          */
3849                         err = ext4_convert_unwritten_extents(inode,
3850                                                              offset, ret);
3851                         if (err < 0)
3852                                 ret = err;
3853                         ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3854                 }
3855                 return ret;
3856         }
3857
3858         /* for write the the end of file case, we fall back to old way */
3859         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3860 }
3861
3862 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3863                               const struct iovec *iov, loff_t offset,
3864                               unsigned long nr_segs)
3865 {
3866         struct file *file = iocb->ki_filp;
3867         struct inode *inode = file->f_mapping->host;
3868
3869         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3870                 return ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
3871
3872         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3873 }
3874
3875 /*
3876  * Pages can be marked dirty completely asynchronously from ext4's journalling
3877  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3878  * much here because ->set_page_dirty is called under VFS locks.  The page is
3879  * not necessarily locked.
3880  *
3881  * We cannot just dirty the page and leave attached buffers clean, because the
3882  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3883  * or jbddirty because all the journalling code will explode.
3884  *
3885  * So what we do is to mark the page "pending dirty" and next time writepage
3886  * is called, propagate that into the buffers appropriately.
3887  */
3888 static int ext4_journalled_set_page_dirty(struct page *page)
3889 {
3890         SetPageChecked(page);
3891         return __set_page_dirty_nobuffers(page);
3892 }
3893
3894 static const struct address_space_operations ext4_ordered_aops = {
3895         .readpage               = ext4_readpage,
3896         .readpages              = ext4_readpages,
3897         .writepage              = ext4_writepage,
3898         .sync_page              = block_sync_page,
3899         .write_begin            = ext4_write_begin,
3900         .write_end              = ext4_ordered_write_end,
3901         .bmap                   = ext4_bmap,
3902         .invalidatepage         = ext4_invalidatepage,
3903         .releasepage            = ext4_releasepage,
3904         .direct_IO              = ext4_direct_IO,
3905         .migratepage            = buffer_migrate_page,
3906         .is_partially_uptodate  = block_is_partially_uptodate,
3907         .error_remove_page      = generic_error_remove_page,
3908 };
3909
3910 static const struct address_space_operations ext4_writeback_aops = {
3911         .readpage               = ext4_readpage,
3912         .readpages              = ext4_readpages,
3913         .writepage              = ext4_writepage,
3914         .sync_page              = block_sync_page,
3915         .write_begin            = ext4_write_begin,
3916         .write_end              = ext4_writeback_write_end,
3917         .bmap                   = ext4_bmap,
3918         .invalidatepage         = ext4_invalidatepage,
3919         .releasepage            = ext4_releasepage,
3920         .direct_IO              = ext4_direct_IO,
3921         .migratepage            = buffer_migrate_page,
3922         .is_partially_uptodate  = block_is_partially_uptodate,
3923         .error_remove_page      = generic_error_remove_page,
3924 };
3925
3926 static const struct address_space_operations ext4_journalled_aops = {
3927         .readpage               = ext4_readpage,
3928         .readpages              = ext4_readpages,
3929         .writepage              = ext4_writepage,
3930         .sync_page              = block_sync_page,
3931         .write_begin            = ext4_write_begin,
3932         .write_end              = ext4_journalled_write_end,
3933         .set_page_dirty         = ext4_journalled_set_page_dirty,
3934         .bmap                   = ext4_bmap,
3935         .invalidatepage         = ext4_invalidatepage,
3936         .releasepage            = ext4_releasepage,
3937         .is_partially_uptodate  = block_is_partially_uptodate,
3938         .error_remove_page      = generic_error_remove_page,
3939 };
3940
3941 static const struct address_space_operations ext4_da_aops = {
3942         .readpage               = ext4_readpage,
3943         .readpages              = ext4_readpages,
3944         .writepage              = ext4_writepage,
3945         .writepages             = ext4_da_writepages,
3946         .sync_page              = block_sync_page,
3947         .write_begin            = ext4_da_write_begin,
3948         .write_end              = ext4_da_write_end,
3949         .bmap                   = ext4_bmap,
3950         .invalidatepage         = ext4_da_invalidatepage,
3951         .releasepage            = ext4_releasepage,
3952         .direct_IO              = ext4_direct_IO,
3953         .migratepage            = buffer_migrate_page,
3954         .is_partially_uptodate  = block_is_partially_uptodate,
3955         .error_remove_page      = generic_error_remove_page,
3956 };
3957
3958 void ext4_set_aops(struct inode *inode)
3959 {
3960         if (ext4_should_order_data(inode) &&
3961                 test_opt(inode->i_sb, DELALLOC))
3962                 inode->i_mapping->a_ops = &ext4_da_aops;
3963         else if (ext4_should_order_data(inode))
3964                 inode->i_mapping->a_ops = &ext4_ordered_aops;
3965         else if (ext4_should_writeback_data(inode) &&
3966                  test_opt(inode->i_sb, DELALLOC))
3967                 inode->i_mapping->a_ops = &ext4_da_aops;
3968         else if (ext4_should_writeback_data(inode))
3969                 inode->i_mapping->a_ops = &ext4_writeback_aops;
3970         else
3971                 inode->i_mapping->a_ops = &ext4_journalled_aops;
3972 }
3973
3974 /*
3975  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3976  * up to the end of the block which corresponds to `from'.
3977  * This required during truncate. We need to physically zero the tail end
3978  * of that block so it doesn't yield old data if the file is later grown.
3979  */
3980 int ext4_block_truncate_page(handle_t *handle,
3981                 struct address_space *mapping, loff_t from)
3982 {
3983         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3984         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3985         unsigned blocksize, length, pos;
3986         ext4_lblk_t iblock;
3987         struct inode *inode = mapping->host;
3988         struct buffer_head *bh;
3989         struct page *page;
3990         int err = 0;
3991
3992         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
3993                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
3994         if (!page)
3995                 return -EINVAL;
3996
3997         blocksize = inode->i_sb->s_blocksize;
3998         length = blocksize - (offset & (blocksize - 1));
3999         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
4000
4001         if (!page_has_buffers(page))
4002                 create_empty_buffers(page, blocksize, 0);
4003
4004         /* Find the buffer that contains "offset" */
4005         bh = page_buffers(page);
4006         pos = blocksize;
4007         while (offset >= pos) {
4008                 bh = bh->b_this_page;
4009                 iblock++;
4010                 pos += blocksize;
4011         }
4012
4013         err = 0;
4014         if (buffer_freed(bh)) {
4015                 BUFFER_TRACE(bh, "freed: skip");
4016                 goto unlock;
4017         }
4018
4019         if (!buffer_mapped(bh)) {
4020                 BUFFER_TRACE(bh, "unmapped");
4021                 ext4_get_block(inode, iblock, bh, 0);
4022                 /* unmapped? It's a hole - nothing to do */
4023                 if (!buffer_mapped(bh)) {
4024                         BUFFER_TRACE(bh, "still unmapped");
4025                         goto unlock;
4026                 }
4027         }
4028
4029         /* Ok, it's mapped. Make sure it's up-to-date */
4030         if (PageUptodate(page))
4031                 set_buffer_uptodate(bh);
4032
4033         if (!buffer_uptodate(bh)) {
4034                 err = -EIO;
4035                 ll_rw_block(READ, 1, &bh);
4036                 wait_on_buffer(bh);
4037                 /* Uhhuh. Read error. Complain and punt. */
4038                 if (!buffer_uptodate(bh))
4039                         goto unlock;
4040         }
4041
4042         if (ext4_should_journal_data(inode)) {
4043                 BUFFER_TRACE(bh, "get write access");
4044                 err = ext4_journal_get_write_access(handle, bh);
4045                 if (err)
4046                         goto unlock;
4047         }
4048
4049         zero_user(page, offset, length);
4050
4051         BUFFER_TRACE(bh, "zeroed end of block");
4052
4053         err = 0;
4054         if (ext4_should_journal_data(inode)) {
4055                 err = ext4_handle_dirty_metadata(handle, inode, bh);
4056         } else {
4057                 if (ext4_should_order_data(inode))
4058                         err = ext4_jbd2_file_inode(handle, inode);
4059                 mark_buffer_dirty(bh);
4060         }
4061
4062 unlock:
4063         unlock_page(page);
4064         page_cache_release(page);
4065         return err;
4066 }
4067
4068 /*
4069  * Probably it should be a library function... search for first non-zero word
4070  * or memcmp with zero_page, whatever is better for particular architecture.
4071  * Linus?
4072  */
4073 static inline int all_zeroes(__le32 *p, __le32 *q)
4074 {
4075         while (p < q)
4076                 if (*p++)
4077                         return 0;
4078         return 1;
4079 }
4080
4081 /**
4082  *      ext4_find_shared - find the indirect blocks for partial truncation.
4083  *      @inode:   inode in question
4084  *      @depth:   depth of the affected branch
4085  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
4086  *      @chain:   place to store the pointers to partial indirect blocks
4087  *      @top:     place to the (detached) top of branch
4088  *
4089  *      This is a helper function used by ext4_truncate().
4090  *
4091  *      When we do truncate() we may have to clean the ends of several
4092  *      indirect blocks but leave the blocks themselves alive. Block is
4093  *      partially truncated if some data below the new i_size is refered
4094  *      from it (and it is on the path to the first completely truncated
4095  *      data block, indeed).  We have to free the top of that path along
4096  *      with everything to the right of the path. Since no allocation
4097  *      past the truncation point is possible until ext4_truncate()
4098  *      finishes, we may safely do the latter, but top of branch may
4099  *      require special attention - pageout below the truncation point
4100  *      might try to populate it.
4101  *
4102  *      We atomically detach the top of branch from the tree, store the
4103  *      block number of its root in *@top, pointers to buffer_heads of
4104  *      partially truncated blocks - in @chain[].bh and pointers to
4105  *      their last elements that should not be removed - in
4106  *      @chain[].p. Return value is the pointer to last filled element
4107  *      of @chain.
4108  *
4109  *      The work left to caller to do the actual freeing of subtrees:
4110  *              a) free the subtree starting from *@top
4111  *              b) free the subtrees whose roots are stored in
4112  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
4113  *              c) free the subtrees growing from the inode past the @chain[0].
4114  *                      (no partially truncated stuff there).  */
4115
4116 static Indirect *ext4_find_shared(struct inode *inode, int depth,
4117                                   ext4_lblk_t offsets[4], Indirect chain[4],
4118                                   __le32 *top)
4119 {
4120         Indirect *partial, *p;
4121         int k, err;
4122
4123         *top = 0;
4124         /* Make k index the deepest non-null offset + 1 */
4125         for (k = depth; k > 1 && !offsets[k-1]; k--)
4126                 ;
4127         partial = ext4_get_branch(inode, k, offsets, chain, &err);
4128         /* Writer: pointers */
4129         if (!partial)
4130                 partial = chain + k-1;
4131         /*
4132          * If the branch acquired continuation since we've looked at it -
4133          * fine, it should all survive and (new) top doesn't belong to us.
4134          */
4135         if (!partial->key && *partial->p)
4136                 /* Writer: end */
4137                 goto no_top;
4138         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
4139                 ;
4140         /*
4141          * OK, we've found the last block that must survive. The rest of our
4142          * branch should be detached before unlocking. However, if that rest
4143          * of branch is all ours and does not grow immediately from the inode
4144          * it's easier to cheat and just decrement partial->p.
4145          */
4146         if (p == chain + k - 1 && p > chain) {
4147                 p->p--;
4148         } else {
4149                 *top = *p->p;
4150                 /* Nope, don't do this in ext4.  Must leave the tree intact */
4151 #if 0
4152                 *p->p = 0;
4153 #endif
4154         }
4155         /* Writer: end */
4156
4157         while (partial > p) {
4158                 brelse(partial->bh);
4159                 partial--;
4160         }
4161 no_top:
4162         return partial;
4163 }
4164
4165 /*
4166  * Zero a number of block pointers in either an inode or an indirect block.
4167  * If we restart the transaction we must again get write access to the
4168  * indirect block for further modification.
4169  *
4170  * We release `count' blocks on disk, but (last - first) may be greater
4171  * than `count' because there can be holes in there.
4172  */
4173 static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
4174                              struct buffer_head *bh,
4175                              ext4_fsblk_t block_to_free,
4176                              unsigned long count, __le32 *first,
4177                              __le32 *last)
4178 {
4179         __le32 *p;
4180         int     flags = EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_VALIDATED;
4181
4182         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
4183                 flags |= EXT4_FREE_BLOCKS_METADATA;
4184
4185         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
4186                                    count)) {
4187                 EXT4_ERROR_INODE(inode, "attempt to clear invalid "
4188                                  "blocks %llu len %lu",
4189                                  (unsigned long long) block_to_free, count);
4190                 return 1;
4191         }
4192
4193         if (try_to_extend_transaction(handle, inode)) {
4194                 if (bh) {
4195                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4196                         ext4_handle_dirty_metadata(handle, inode, bh);
4197                 }
4198                 ext4_mark_inode_dirty(handle, inode);
4199                 ext4_truncate_restart_trans(handle, inode,
4200                                             blocks_for_truncate(inode));
4201                 if (bh) {
4202                         BUFFER_TRACE(bh, "retaking write access");
4203                         ext4_journal_get_write_access(handle, bh);
4204                 }
4205         }
4206
4207         for (p = first; p < last; p++)
4208                 *p = 0;
4209
4210         ext4_free_blocks(handle, inode, 0, block_to_free, count, flags);
4211         return 0;
4212 }
4213
4214 /**
4215  * ext4_free_data - free a list of data blocks
4216  * @handle:     handle for this transaction
4217  * @inode:      inode we are dealing with
4218  * @this_bh:    indirect buffer_head which contains *@first and *@last
4219  * @first:      array of block numbers
4220  * @last:       points immediately past the end of array
4221  *
4222  * We are freeing all blocks refered from that array (numbers are stored as
4223  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
4224  *
4225  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
4226  * blocks are contiguous then releasing them at one time will only affect one
4227  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
4228  * actually use a lot of journal space.
4229  *
4230  * @this_bh will be %NULL if @first and @last point into the inode's direct
4231  * block pointers.
4232  */
4233 static void ext4_free_data(handle_t *handle, struct inode *inode,
4234                            struct buffer_head *this_bh,
4235                            __le32 *first, __le32 *last)
4236 {
4237         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
4238         unsigned long count = 0;            /* Number of blocks in the run */
4239         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
4240                                                corresponding to
4241                                                block_to_free */
4242         ext4_fsblk_t nr;                    /* Current block # */
4243         __le32 *p;                          /* Pointer into inode/ind
4244                                                for current block */
4245         int err;
4246
4247         if (this_bh) {                          /* For indirect block */
4248                 BUFFER_TRACE(this_bh, "get_write_access");
4249                 err = ext4_journal_get_write_access(handle, this_bh);
4250                 /* Important: if we can't update the indirect pointers
4251                  * to the blocks, we can't free them. */
4252                 if (err)
4253                         return;
4254         }
4255
4256         for (p = first; p < last; p++) {
4257                 nr = le32_to_cpu(*p);
4258                 if (nr) {
4259                         /* accumulate blocks to free if they're contiguous */
4260                         if (count == 0) {
4261                                 block_to_free = nr;
4262                                 block_to_free_p = p;
4263                                 count = 1;
4264                         } else if (nr == block_to_free + count) {
4265                                 count++;
4266                         } else {
4267                                 if (ext4_clear_blocks(handle, inode, this_bh,
4268                                                       block_to_free, count,
4269                                                       block_to_free_p, p))
4270                                         break;
4271                                 block_to_free = nr;
4272                                 block_to_free_p = p;
4273                                 count = 1;
4274                         }
4275                 }
4276         }
4277
4278         if (count > 0)
4279                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
4280                                   count, block_to_free_p, p);
4281
4282         if (this_bh) {
4283                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
4284
4285                 /*
4286                  * The buffer head should have an attached journal head at this
4287                  * point. However, if the data is corrupted and an indirect
4288                  * block pointed to itself, it would have been detached when
4289                  * the block was cleared. Check for this instead of OOPSing.
4290                  */
4291                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
4292                         ext4_handle_dirty_metadata(handle, inode, this_bh);
4293                 else
4294                         EXT4_ERROR_INODE(inode,
4295                                          "circular indirect block detected at "
4296                                          "block %llu",
4297                                 (unsigned long long) this_bh->b_blocknr);
4298         }
4299 }
4300
4301 /**
4302  *      ext4_free_branches - free an array of branches
4303  *      @handle: JBD handle for this transaction
4304  *      @inode: inode we are dealing with
4305  *      @parent_bh: the buffer_head which contains *@first and *@last
4306  *      @first: array of block numbers
4307  *      @last:  pointer immediately past the end of array
4308  *      @depth: depth of the branches to free
4309  *
4310  *      We are freeing all blocks refered from these branches (numbers are
4311  *      stored as little-endian 32-bit) and updating @inode->i_blocks
4312  *      appropriately.
4313  */
4314 static void ext4_free_branches(handle_t *handle, struct inode *inode,
4315                                struct buffer_head *parent_bh,
4316                                __le32 *first, __le32 *last, int depth)
4317 {
4318         ext4_fsblk_t nr;
4319         __le32 *p;
4320
4321         if (ext4_handle_is_aborted(handle))
4322                 return;
4323
4324         if (depth--) {
4325                 struct buffer_head *bh;
4326                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4327                 p = last;
4328                 while (--p >= first) {
4329                         nr = le32_to_cpu(*p);
4330                         if (!nr)
4331                                 continue;               /* A hole */
4332
4333                         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
4334                                                    nr, 1)) {
4335                                 EXT4_ERROR_INODE(inode,
4336                                                  "invalid indirect mapped "
4337                                                  "block %lu (level %d)",
4338                                                  (unsigned long) nr, depth);
4339                                 break;
4340                         }
4341
4342                         /* Go read the buffer for the next level down */
4343                         bh = sb_bread(inode->i_sb, nr);
4344
4345                         /*
4346                          * A read failure? Report error and clear slot
4347                          * (should be rare).
4348                          */
4349                         if (!bh) {
4350                                 EXT4_ERROR_INODE_BLOCK(inode, nr,
4351                                                        "Read failure");
4352                                 continue;
4353                         }
4354
4355                         /* This zaps the entire block.  Bottom up. */
4356                         BUFFER_TRACE(bh, "free child branches");
4357                         ext4_free_branches(handle, inode, bh,
4358                                         (__le32 *) bh->b_data,
4359                                         (__le32 *) bh->b_data + addr_per_block,
4360                                         depth);
4361
4362                         /*
4363                          * Everything below this this pointer has been
4364                          * released.  Now let this top-of-subtree go.
4365                          *
4366                          * We want the freeing of this indirect block to be
4367                          * atomic in the journal with the updating of the
4368                          * bitmap block which owns it.  So make some room in
4369                          * the journal.
4370                          *
4371                          * We zero the parent pointer *after* freeing its
4372                          * pointee in the bitmaps, so if extend_transaction()
4373                          * for some reason fails to put the bitmap changes and
4374                          * the release into the same transaction, recovery
4375                          * will merely complain about releasing a free block,
4376                          * rather than leaking blocks.
4377                          */
4378                         if (ext4_handle_is_aborted(handle))
4379                                 return;
4380                         if (try_to_extend_transaction(handle, inode)) {
4381                                 ext4_mark_inode_dirty(handle, inode);
4382                                 ext4_truncate_restart_trans(handle, inode,
4383                                             blocks_for_truncate(inode));
4384                         }
4385
4386                         /*
4387                          * The forget flag here is critical because if
4388                          * we are journaling (and not doing data
4389                          * journaling), we have to make sure a revoke
4390                          * record is written to prevent the journal
4391                          * replay from overwriting the (former)
4392                          * indirect block if it gets reallocated as a
4393                          * data block.  This must happen in the same
4394                          * transaction where the data blocks are
4395                          * actually freed.
4396                          */
4397                         ext4_free_blocks(handle, inode, 0, nr, 1,
4398                                          EXT4_FREE_BLOCKS_METADATA|
4399                                          EXT4_FREE_BLOCKS_FORGET);
4400
4401                         if (parent_bh) {
4402                                 /*
4403                                  * The block which we have just freed is
4404                                  * pointed to by an indirect block: journal it
4405                                  */
4406                                 BUFFER_TRACE(parent_bh, "get_write_access");
4407                                 if (!ext4_journal_get_write_access(handle,
4408                                                                    parent_bh)){
4409                                         *p = 0;
4410                                         BUFFER_TRACE(parent_bh,
4411                                         "call ext4_handle_dirty_metadata");
4412                                         ext4_handle_dirty_metadata(handle,
4413                                                                    inode,
4414                                                                    parent_bh);
4415                                 }
4416                         }
4417                 }
4418         } else {
4419                 /* We have reached the bottom of the tree. */
4420                 BUFFER_TRACE(parent_bh, "free data blocks");
4421                 ext4_free_data(handle, inode, parent_bh, first, last);
4422         }
4423 }
4424
4425 int ext4_can_truncate(struct inode *inode)
4426 {
4427         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4428                 return 0;
4429         if (S_ISREG(inode->i_mode))
4430                 return 1;
4431         if (S_ISDIR(inode->i_mode))
4432                 return 1;
4433         if (S_ISLNK(inode->i_mode))
4434                 return !ext4_inode_is_fast_symlink(inode);
4435         return 0;
4436 }
4437
4438 /*
4439  * ext4_truncate()
4440  *
4441  * We block out ext4_get_block() block instantiations across the entire
4442  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4443  * simultaneously on behalf of the same inode.
4444  *
4445  * As we work through the truncate and commmit bits of it to the journal there
4446  * is one core, guiding principle: the file's tree must always be consistent on
4447  * disk.  We must be able to restart the truncate after a crash.
4448  *
4449  * The file's tree may be transiently inconsistent in memory (although it
4450  * probably isn't), but whenever we close off and commit a journal transaction,
4451  * the contents of (the filesystem + the journal) must be consistent and
4452  * restartable.  It's pretty simple, really: bottom up, right to left (although
4453  * left-to-right works OK too).
4454  *
4455  * Note that at recovery time, journal replay occurs *before* the restart of
4456  * truncate against the orphan inode list.
4457  *
4458  * The committed inode has the new, desired i_size (which is the same as
4459  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4460  * that this inode's truncate did not complete and it will again call
4461  * ext4_truncate() to have another go.  So there will be instantiated blocks
4462  * to the right of the truncation point in a crashed ext4 filesystem.  But
4463  * that's fine - as long as they are linked from the inode, the post-crash
4464  * ext4_truncate() run will find them and release them.
4465  */
4466 void ext4_truncate(struct inode *inode)
4467 {
4468         handle_t *handle;
4469         struct ext4_inode_info *ei = EXT4_I(inode);
4470         __le32 *i_data = ei->i_data;
4471         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4472         struct address_space *mapping = inode->i_mapping;
4473         ext4_lblk_t offsets[4];
4474         Indirect chain[4];
4475         Indirect *partial;
4476         __le32 nr = 0;
4477         int n;
4478         ext4_lblk_t last_block;
4479         unsigned blocksize = inode->i_sb->s_blocksize;
4480
4481         if (!ext4_can_truncate(inode))
4482                 return;
4483
4484         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4485
4486         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4487                 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4488
4489         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
4490                 ext4_ext_truncate(inode);
4491                 return;
4492         }
4493
4494         handle = start_transaction(inode);
4495         if (IS_ERR(handle))
4496                 return;         /* AKPM: return what? */
4497
4498         last_block = (inode->i_size + blocksize-1)
4499                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4500
4501         if (inode->i_size & (blocksize - 1))
4502                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4503                         goto out_stop;
4504
4505         n = ext4_block_to_path(inode, last_block, offsets, NULL);
4506         if (n == 0)
4507                 goto out_stop;  /* error */
4508
4509         /*
4510          * OK.  This truncate is going to happen.  We add the inode to the
4511          * orphan list, so that if this truncate spans multiple transactions,
4512          * and we crash, we will resume the truncate when the filesystem
4513          * recovers.  It also marks the inode dirty, to catch the new size.
4514          *
4515          * Implication: the file must always be in a sane, consistent
4516          * truncatable state while each transaction commits.
4517          */
4518         if (ext4_orphan_add(handle, inode))
4519                 goto out_stop;
4520
4521         /*
4522          * From here we block out all ext4_get_block() callers who want to
4523          * modify the block allocation tree.
4524          */
4525         down_write(&ei->i_data_sem);
4526
4527         ext4_discard_preallocations(inode);
4528
4529         /*
4530          * The orphan list entry will now protect us from any crash which
4531          * occurs before the truncate completes, so it is now safe to propagate
4532          * the new, shorter inode size (held for now in i_size) into the
4533          * on-disk inode. We do this via i_disksize, which is the value which
4534          * ext4 *really* writes onto the disk inode.
4535          */
4536         ei->i_disksize = inode->i_size;
4537
4538         if (n == 1) {           /* direct blocks */
4539                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4540                                i_data + EXT4_NDIR_BLOCKS);
4541                 goto do_indirects;
4542         }
4543
4544         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4545         /* Kill the top of shared branch (not detached) */
4546         if (nr) {
4547                 if (partial == chain) {
4548                         /* Shared branch grows from the inode */
4549                         ext4_free_branches(handle, inode, NULL,
4550                                            &nr, &nr+1, (chain+n-1) - partial);
4551                         *partial->p = 0;
4552                         /*
4553                          * We mark the inode dirty prior to restart,
4554                          * and prior to stop.  No need for it here.
4555                          */
4556                 } else {
4557                         /* Shared branch grows from an indirect block */
4558                         BUFFER_TRACE(partial->bh, "get_write_access");
4559                         ext4_free_branches(handle, inode, partial->bh,
4560                                         partial->p,
4561                                         partial->p+1, (chain+n-1) - partial);
4562                 }
4563         }
4564         /* Clear the ends of indirect blocks on the shared branch */
4565         while (partial > chain) {
4566                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4567                                    (__le32*)partial->bh->b_data+addr_per_block,
4568                                    (chain+n-1) - partial);
4569                 BUFFER_TRACE(partial->bh, "call brelse");
4570                 brelse(partial->bh);
4571                 partial--;
4572         }
4573 do_indirects:
4574         /* Kill the remaining (whole) subtrees */
4575         switch (offsets[0]) {
4576         default:
4577                 nr = i_data[EXT4_IND_BLOCK];
4578                 if (nr) {
4579                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4580                         i_data[EXT4_IND_BLOCK] = 0;
4581                 }
4582         case EXT4_IND_BLOCK:
4583                 nr = i_data[EXT4_DIND_BLOCK];
4584                 if (nr) {
4585                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4586                         i_data[EXT4_DIND_BLOCK] = 0;
4587                 }
4588         case EXT4_DIND_BLOCK:
4589                 nr = i_data[EXT4_TIND_BLOCK];
4590                 if (nr) {
4591                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4592                         i_data[EXT4_TIND_BLOCK] = 0;
4593                 }
4594         case EXT4_TIND_BLOCK:
4595                 ;
4596         }
4597
4598         up_write(&ei->i_data_sem);
4599         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4600         ext4_mark_inode_dirty(handle, inode);
4601
4602         /*
4603          * In a multi-transaction truncate, we only make the final transaction
4604          * synchronous
4605          */
4606         if (IS_SYNC(inode))
4607                 ext4_handle_sync(handle);
4608 out_stop:
4609         /*
4610          * If this was a simple ftruncate(), and the file will remain alive
4611          * then we need to clear up the orphan record which we created above.
4612          * However, if this was a real unlink then we were called by
4613          * ext4_delete_inode(), and we allow that function to clean up the
4614          * orphan info for us.
4615          */
4616         if (inode->i_nlink)
4617                 ext4_orphan_del(handle, inode);
4618
4619         ext4_journal_stop(handle);
4620 }
4621
4622 /*
4623  * ext4_get_inode_loc returns with an extra refcount against the inode's
4624  * underlying buffer_head on success. If 'in_mem' is true, we have all
4625  * data in memory that is needed to recreate the on-disk version of this
4626  * inode.
4627  */
4628 static int __ext4_get_inode_loc(struct inode *inode,
4629                                 struct ext4_iloc *iloc, int in_mem)
4630 {
4631         struct ext4_group_desc  *gdp;
4632         struct buffer_head      *bh;
4633         struct super_block      *sb = inode->i_sb;
4634         ext4_fsblk_t            block;
4635         int                     inodes_per_block, inode_offset;
4636
4637         iloc->bh = NULL;
4638         if (!ext4_valid_inum(sb, inode->i_ino))
4639                 return -EIO;
4640
4641         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4642         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4643         if (!gdp)
4644                 return -EIO;
4645
4646         /*
4647          * Figure out the offset within the block group inode table
4648          */
4649         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4650         inode_offset = ((inode->i_ino - 1) %
4651                         EXT4_INODES_PER_GROUP(sb));
4652         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4653         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4654
4655         bh = sb_getblk(sb, block);
4656         if (!bh) {
4657                 EXT4_ERROR_INODE_BLOCK(inode, block,
4658                                        "unable to read itable block");
4659                 return -EIO;
4660         }
4661         if (!buffer_uptodate(bh)) {
4662                 lock_buffer(bh);
4663
4664                 /*
4665                  * If the buffer has the write error flag, we have failed
4666                  * to write out another inode in the same block.  In this
4667                  * case, we don't have to read the block because we may
4668                  * read the old inode data successfully.
4669                  */
4670                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4671                         set_buffer_uptodate(bh);
4672
4673                 if (buffer_uptodate(bh)) {
4674                         /* someone brought it uptodate while we waited */
4675                         unlock_buffer(bh);
4676                         goto has_buffer;
4677                 }
4678
4679                 /*
4680                  * If we have all information of the inode in memory and this
4681                  * is the only valid inode in the block, we need not read the
4682                  * block.
4683                  */
4684                 if (in_mem) {
4685                         struct buffer_head *bitmap_bh;
4686                         int i, start;
4687
4688                         start = inode_offset & ~(inodes_per_block - 1);
4689
4690                         /* Is the inode bitmap in cache? */
4691                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4692                         if (!bitmap_bh)
4693                                 goto make_io;
4694
4695                         /*
4696                          * If the inode bitmap isn't in cache then the
4697                          * optimisation may end up performing two reads instead
4698                          * of one, so skip it.
4699                          */
4700                         if (!buffer_uptodate(bitmap_bh)) {
4701                                 brelse(bitmap_bh);
4702                                 goto make_io;
4703                         }
4704                         for (i = start; i < start + inodes_per_block; i++) {
4705                                 if (i == inode_offset)
4706                                         continue;
4707                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4708                                         break;
4709                         }
4710                         brelse(bitmap_bh);
4711                         if (i == start + inodes_per_block) {
4712                                 /* all other inodes are free, so skip I/O */
4713                                 memset(bh->b_data, 0, bh->b_size);
4714                                 set_buffer_uptodate(bh);
4715                                 unlock_buffer(bh);
4716                                 goto has_buffer;
4717                         }
4718                 }
4719
4720 make_io:
4721                 /*
4722                  * If we need to do any I/O, try to pre-readahead extra
4723                  * blocks from the inode table.
4724                  */
4725                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4726                         ext4_fsblk_t b, end, table;
4727                         unsigned num;
4728
4729                         table = ext4_inode_table(sb, gdp);
4730                         /* s_inode_readahead_blks is always a power of 2 */
4731                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4732                         if (table > b)
4733                                 b = table;
4734                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4735                         num = EXT4_INODES_PER_GROUP(sb);
4736                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4737                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4738                                 num -= ext4_itable_unused_count(sb, gdp);
4739                         table += num / inodes_per_block;
4740                         if (end > table)
4741                                 end = table;
4742                         while (b <= end)
4743                                 sb_breadahead(sb, b++);
4744                 }
4745
4746                 /*
4747                  * There are other valid inodes in the buffer, this inode
4748                  * has in-inode xattrs, or we don't have this inode in memory.
4749                  * Read the block from disk.
4750                  */
4751                 get_bh(bh);
4752                 bh->b_end_io = end_buffer_read_sync;
4753                 submit_bh(READ_META, bh);
4754                 wait_on_buffer(bh);
4755                 if (!buffer_uptodate(bh)) {
4756                         EXT4_ERROR_INODE_BLOCK(inode, block,
4757                                                "unable to read itable block");
4758                         brelse(bh);
4759                         return -EIO;
4760                 }
4761         }
4762 has_buffer:
4763         iloc->bh = bh;
4764         return 0;
4765 }
4766
4767 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4768 {
4769         /* We have all inode data except xattrs in memory here. */
4770         return __ext4_get_inode_loc(inode, iloc,
4771                 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
4772 }
4773
4774 void ext4_set_inode_flags(struct inode *inode)
4775 {
4776         unsigned int flags = EXT4_I(inode)->i_flags;
4777
4778         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4779         if (flags & EXT4_SYNC_FL)
4780                 inode->i_flags |= S_SYNC;
4781         if (flags & EXT4_APPEND_FL)
4782                 inode->i_flags |= S_APPEND;
4783         if (flags & EXT4_IMMUTABLE_FL)
4784                 inode->i_flags |= S_IMMUTABLE;
4785         if (flags & EXT4_NOATIME_FL)
4786                 inode->i_flags |= S_NOATIME;
4787         if (flags & EXT4_DIRSYNC_FL)
4788                 inode->i_flags |= S_DIRSYNC;
4789 }
4790
4791 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4792 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4793 {
4794         unsigned int vfs_fl;
4795         unsigned long old_fl, new_fl;
4796
4797         do {
4798                 vfs_fl = ei->vfs_inode.i_flags;
4799                 old_fl = ei->i_flags;
4800                 new_fl = old_fl & ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4801                                 EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|
4802                                 EXT4_DIRSYNC_FL);
4803                 if (vfs_fl & S_SYNC)
4804                         new_fl |= EXT4_SYNC_FL;
4805                 if (vfs_fl & S_APPEND)
4806                         new_fl |= EXT4_APPEND_FL;
4807                 if (vfs_fl & S_IMMUTABLE)
4808                         new_fl |= EXT4_IMMUTABLE_FL;
4809                 if (vfs_fl & S_NOATIME)
4810                         new_fl |= EXT4_NOATIME_FL;
4811                 if (vfs_fl & S_DIRSYNC)
4812                         new_fl |= EXT4_DIRSYNC_FL;
4813         } while (cmpxchg(&ei->i_flags, old_fl, new_fl) != old_fl);
4814 }
4815
4816 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4817                                   struct ext4_inode_info *ei)
4818 {
4819         blkcnt_t i_blocks ;
4820         struct inode *inode = &(ei->vfs_inode);
4821         struct super_block *sb = inode->i_sb;
4822
4823         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4824                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4825                 /* we are using combined 48 bit field */
4826                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4827                                         le32_to_cpu(raw_inode->i_blocks_lo);
4828                 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
4829                         /* i_blocks represent file system block size */
4830                         return i_blocks  << (inode->i_blkbits - 9);
4831                 } else {
4832                         return i_blocks;
4833                 }
4834         } else {
4835                 return le32_to_cpu(raw_inode->i_blocks_lo);
4836         }
4837 }
4838
4839 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4840 {
4841         struct ext4_iloc iloc;
4842         struct ext4_inode *raw_inode;
4843         struct ext4_inode_info *ei;
4844         struct inode *inode;
4845         journal_t *journal = EXT4_SB(sb)->s_journal;
4846         long ret;
4847         int block;
4848
4849         inode = iget_locked(sb, ino);
4850         if (!inode)
4851                 return ERR_PTR(-ENOMEM);
4852         if (!(inode->i_state & I_NEW))
4853                 return inode;
4854
4855         ei = EXT4_I(inode);
4856         iloc.bh = 0;
4857
4858         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4859         if (ret < 0)
4860                 goto bad_inode;
4861         raw_inode = ext4_raw_inode(&iloc);
4862         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4863         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4864         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4865         if (!(test_opt(inode->i_sb, NO_UID32))) {
4866                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4867                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4868         }
4869         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
4870
4871         ei->i_state_flags = 0;
4872         ei->i_dir_start_lookup = 0;
4873         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4874         /* We now have enough fields to check if the inode was active or not.
4875          * This is needed because nfsd might try to access dead inodes
4876          * the test is that same one that e2fsck uses
4877          * NeilBrown 1999oct15
4878          */
4879         if (inode->i_nlink == 0) {
4880                 if (inode->i_mode == 0 ||
4881                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
4882                         /* this inode is deleted */
4883                         ret = -ESTALE;
4884                         goto bad_inode;
4885                 }
4886                 /* The only unlinked inodes we let through here have
4887                  * valid i_mode and are being read by the orphan
4888                  * recovery code: that's fine, we're about to complete
4889                  * the process of deleting those. */
4890         }
4891         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4892         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4893         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4894         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
4895                 ei->i_file_acl |=
4896                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4897         inode->i_size = ext4_isize(raw_inode);
4898         ei->i_disksize = inode->i_size;
4899 #ifdef CONFIG_QUOTA
4900         ei->i_reserved_quota = 0;
4901 #endif
4902         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4903         ei->i_block_group = iloc.block_group;
4904         ei->i_last_alloc_group = ~0;
4905         /*
4906          * NOTE! The in-memory inode i_data array is in little-endian order
4907          * even on big-endian machines: we do NOT byteswap the block numbers!
4908          */
4909         for (block = 0; block < EXT4_N_BLOCKS; block++)
4910                 ei->i_data[block] = raw_inode->i_block[block];
4911         INIT_LIST_HEAD(&ei->i_orphan);
4912
4913         /*
4914          * Set transaction id's of transactions that have to be committed
4915          * to finish f[data]sync. We set them to currently running transaction
4916          * as we cannot be sure that the inode or some of its metadata isn't
4917          * part of the transaction - the inode could have been reclaimed and
4918          * now it is reread from disk.
4919          */
4920         if (journal) {
4921                 transaction_t *transaction;
4922                 tid_t tid;
4923
4924                 read_lock(&journal->j_state_lock);
4925                 if (journal->j_running_transaction)
4926                         transaction = journal->j_running_transaction;
4927                 else
4928                         transaction = journal->j_committing_transaction;
4929                 if (transaction)
4930                         tid = transaction->t_tid;
4931                 else
4932                         tid = journal->j_commit_sequence;
4933                 read_unlock(&journal->j_state_lock);
4934                 ei->i_sync_tid = tid;
4935                 ei->i_datasync_tid = tid;
4936         }
4937
4938         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4939                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4940                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4941                     EXT4_INODE_SIZE(inode->i_sb)) {
4942                         ret = -EIO;
4943                         goto bad_inode;
4944                 }
4945                 if (ei->i_extra_isize == 0) {
4946                         /* The extra space is currently unused. Use it. */
4947                         ei->i_extra_isize = sizeof(struct ext4_inode) -
4948                                             EXT4_GOOD_OLD_INODE_SIZE;
4949                 } else {
4950                         __le32 *magic = (void *)raw_inode +
4951                                         EXT4_GOOD_OLD_INODE_SIZE +
4952                                         ei->i_extra_isize;
4953                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
4954                                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
4955                 }
4956         } else
4957                 ei->i_extra_isize = 0;
4958
4959         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4960         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4961         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4962         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4963
4964         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4965         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4966                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4967                         inode->i_version |=
4968                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4969         }
4970
4971         ret = 0;
4972         if (ei->i_file_acl &&
4973             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
4974                 EXT4_ERROR_INODE(inode, "bad extended attribute block %llu",
4975                                  ei->i_file_acl);
4976                 ret = -EIO;
4977                 goto bad_inode;
4978         } else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
4979                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4980                     (S_ISLNK(inode->i_mode) &&
4981                      !ext4_inode_is_fast_symlink(inode)))
4982                         /* Validate extent which is part of inode */
4983                         ret = ext4_ext_check_inode(inode);
4984         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4985                    (S_ISLNK(inode->i_mode) &&
4986                     !ext4_inode_is_fast_symlink(inode))) {
4987                 /* Validate block references which are part of inode */
4988                 ret = ext4_check_inode_blockref(inode);
4989         }
4990         if (ret)
4991                 goto bad_inode;
4992
4993         if (S_ISREG(inode->i_mode)) {
4994                 inode->i_op = &ext4_file_inode_operations;
4995                 inode->i_fop = &ext4_file_operations;
4996                 ext4_set_aops(inode);
4997         } else if (S_ISDIR(inode->i_mode)) {
4998                 inode->i_op = &ext4_dir_inode_operations;
4999                 inode->i_fop = &ext4_dir_operations;
5000         } else if (S_ISLNK(inode->i_mode)) {
5001                 if (ext4_inode_is_fast_symlink(inode)) {
5002                         inode->i_op = &ext4_fast_symlink_inode_operations;
5003                         nd_terminate_link(ei->i_data, inode->i_size,
5004                                 sizeof(ei->i_data) - 1);
5005                 } else {
5006                         inode->i_op = &ext4_symlink_inode_operations;
5007                         ext4_set_aops(inode);
5008                 }
5009         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
5010               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
5011                 inode->i_op = &ext4_special_inode_operations;
5012                 if (raw_inode->i_block[0])
5013                         init_special_inode(inode, inode->i_mode,
5014                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
5015                 else
5016                         init_special_inode(inode, inode->i_mode,
5017                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
5018         } else {
5019                 ret = -EIO;
5020                 EXT4_ERROR_INODE(inode, "bogus i_mode (%o)", inode->i_mode);
5021                 goto bad_inode;
5022         }
5023         brelse(iloc.bh);
5024         ext4_set_inode_flags(inode);
5025         unlock_new_inode(inode);
5026         return inode;
5027
5028 bad_inode:
5029         brelse(iloc.bh);
5030         iget_failed(inode);
5031         return ERR_PTR(ret);
5032 }
5033
5034 static int ext4_inode_blocks_set(handle_t *handle,
5035                                 struct ext4_inode *raw_inode,
5036                                 struct ext4_inode_info *ei)
5037 {
5038         struct inode *inode = &(ei->vfs_inode);
5039         u64 i_blocks = inode->i_blocks;
5040         struct super_block *sb = inode->i_sb;
5041
5042         if (i_blocks <= ~0U) {
5043                 /*
5044                  * i_blocks can be represnted in a 32 bit variable
5045                  * as multiple of 512 bytes
5046                  */
5047                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5048                 raw_inode->i_blocks_high = 0;
5049                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5050                 return 0;
5051         }
5052         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
5053                 return -EFBIG;
5054
5055         if (i_blocks <= 0xffffffffffffULL) {
5056                 /*
5057                  * i_blocks can be represented in a 48 bit variable
5058                  * as multiple of 512 bytes
5059                  */
5060                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5061                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5062                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5063         } else {
5064                 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5065                 /* i_block is stored in file system block size */
5066                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
5067                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5068                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5069         }
5070         return 0;
5071 }
5072
5073 /*
5074  * Post the struct inode info into an on-disk inode location in the
5075  * buffer-cache.  This gobbles the caller's reference to the
5076  * buffer_head in the inode location struct.
5077  *
5078  * The caller must have write access to iloc->bh.
5079  */
5080 static int ext4_do_update_inode(handle_t *handle,
5081                                 struct inode *inode,
5082                                 struct ext4_iloc *iloc)
5083 {
5084         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5085         struct ext4_inode_info *ei = EXT4_I(inode);
5086         struct buffer_head *bh = iloc->bh;
5087         int err = 0, rc, block;
5088
5089         /* For fields not not tracking in the in-memory inode,
5090          * initialise them to zero for new inodes. */
5091         if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5092                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5093
5094         ext4_get_inode_flags(ei);
5095         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5096         if (!(test_opt(inode->i_sb, NO_UID32))) {
5097                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
5098                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
5099 /*
5100  * Fix up interoperability with old kernels. Otherwise, old inodes get
5101  * re-used with the upper 16 bits of the uid/gid intact
5102  */
5103                 if (!ei->i_dtime) {
5104                         raw_inode->i_uid_high =
5105                                 cpu_to_le16(high_16_bits(inode->i_uid));
5106                         raw_inode->i_gid_high =
5107                                 cpu_to_le16(high_16_bits(inode->i_gid));
5108                 } else {
5109                         raw_inode->i_uid_high = 0;
5110                         raw_inode->i_gid_high = 0;
5111                 }
5112         } else {
5113                 raw_inode->i_uid_low =
5114                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
5115                 raw_inode->i_gid_low =
5116                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
5117                 raw_inode->i_uid_high = 0;
5118                 raw_inode->i_gid_high = 0;
5119         }
5120         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5121
5122         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5123         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5124         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5125         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5126
5127         if (ext4_inode_blocks_set(handle, raw_inode, ei))
5128                 goto out_brelse;
5129         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5130         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
5131         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
5132             cpu_to_le32(EXT4_OS_HURD))
5133                 raw_inode->i_file_acl_high =
5134                         cpu_to_le16(ei->i_file_acl >> 32);
5135         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5136         ext4_isize_set(raw_inode, ei->i_disksize);
5137         if (ei->i_disksize > 0x7fffffffULL) {
5138                 struct super_block *sb = inode->i_sb;
5139                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
5140                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
5141                                 EXT4_SB(sb)->s_es->s_rev_level ==
5142                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
5143                         /* If this is the first large file
5144                          * created, add a flag to the superblock.
5145                          */
5146                         err = ext4_journal_get_write_access(handle,
5147                                         EXT4_SB(sb)->s_sbh);
5148                         if (err)
5149                                 goto out_brelse;
5150                         ext4_update_dynamic_rev(sb);
5151                         EXT4_SET_RO_COMPAT_FEATURE(sb,
5152                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
5153                         sb->s_dirt = 1;
5154                         ext4_handle_sync(handle);
5155                         err = ext4_handle_dirty_metadata(handle, NULL,
5156                                         EXT4_SB(sb)->s_sbh);
5157                 }
5158         }
5159         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5160         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5161                 if (old_valid_dev(inode->i_rdev)) {
5162                         raw_inode->i_block[0] =
5163                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
5164                         raw_inode->i_block[1] = 0;
5165                 } else {
5166                         raw_inode->i_block[0] = 0;
5167                         raw_inode->i_block[1] =
5168                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
5169                         raw_inode->i_block[2] = 0;
5170                 }
5171         } else
5172                 for (block = 0; block < EXT4_N_BLOCKS; block++)
5173                         raw_inode->i_block[block] = ei->i_data[block];
5174
5175         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
5176         if (ei->i_extra_isize) {
5177                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5178                         raw_inode->i_version_hi =
5179                         cpu_to_le32(inode->i_version >> 32);
5180                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
5181         }
5182
5183         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5184         rc = ext4_handle_dirty_metadata(handle, NULL, bh);
5185         if (!err)
5186                 err = rc;
5187         ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5188
5189         ext4_update_inode_fsync_trans(handle, inode, 0);
5190 out_brelse:
5191         brelse(bh);
5192         ext4_std_error(inode->i_sb, err);
5193         return err;
5194 }
5195
5196 /*
5197  * ext4_write_inode()
5198  *
5199  * We are called from a few places:
5200  *
5201  * - Within generic_file_write() for O_SYNC files.
5202  *   Here, there will be no transaction running. We wait for any running
5203  *   trasnaction to commit.
5204  *
5205  * - Within sys_sync(), kupdate and such.
5206  *   We wait on commit, if tol to.
5207  *
5208  * - Within prune_icache() (PF_MEMALLOC == true)
5209  *   Here we simply return.  We can't afford to block kswapd on the
5210  *   journal commit.
5211  *
5212  * In all cases it is actually safe for us to return without doing anything,
5213  * because the inode has been copied into a raw inode buffer in
5214  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
5215  * knfsd.
5216  *
5217  * Note that we are absolutely dependent upon all inode dirtiers doing the
5218  * right thing: they *must* call mark_inode_dirty() after dirtying info in
5219  * which we are interested.
5220  *
5221  * It would be a bug for them to not do this.  The code:
5222  *
5223  *      mark_inode_dirty(inode)
5224  *      stuff();
5225  *      inode->i_size = expr;
5226  *
5227  * is in error because a kswapd-driven write_inode() could occur while
5228  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
5229  * will no longer be on the superblock's dirty inode list.
5230  */
5231 int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5232 {
5233         int err;
5234
5235         if (current->flags & PF_MEMALLOC)
5236                 return 0;
5237
5238         if (EXT4_SB(inode->i_sb)->s_journal) {
5239                 if (ext4_journal_current_handle()) {
5240                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5241                         dump_stack();
5242                         return -EIO;
5243                 }
5244
5245                 if (wbc->sync_mode != WB_SYNC_ALL)
5246                         return 0;
5247
5248                 err = ext4_force_commit(inode->i_sb);
5249         } else {
5250                 struct ext4_iloc iloc;
5251
5252                 err = __ext4_get_inode_loc(inode, &iloc, 0);
5253                 if (err)
5254                         return err;
5255                 if (wbc->sync_mode == WB_SYNC_ALL)
5256                         sync_dirty_buffer(iloc.bh);
5257                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5258                         EXT4_ERROR_INODE_BLOCK(inode, iloc.bh->b_blocknr,
5259                                          "IO error syncing inode");
5260                         err = -EIO;
5261                 }
5262                 brelse(iloc.bh);
5263         }
5264         return err;
5265 }
5266
5267 /*
5268  * ext4_setattr()
5269  *
5270  * Called from notify_change.
5271  *
5272  * We want to trap VFS attempts to truncate the file as soon as
5273  * possible.  In particular, we want to make sure that when the VFS
5274  * shrinks i_size, we put the inode on the orphan list and modify
5275  * i_disksize immediately, so that during the subsequent flushing of
5276  * dirty pages and freeing of disk blocks, we can guarantee that any
5277  * commit will leave the blocks being flushed in an unused state on
5278  * disk.  (On recovery, the inode will get truncated and the blocks will
5279  * be freed, so we have a strong guarantee that no future commit will
5280  * leave these blocks visible to the user.)
5281  *
5282  * Another thing we have to assure is that if we are in ordered mode
5283  * and inode is still attached to the committing transaction, we must
5284  * we start writeout of all the dirty pages which are being truncated.
5285  * This way we are sure that all the data written in the previous
5286  * transaction are already on disk (truncate waits for pages under
5287  * writeback).
5288  *
5289  * Called with inode->i_mutex down.
5290  */
5291 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5292 {
5293         struct inode *inode = dentry->d_inode;
5294         int error, rc = 0;
5295         int orphan = 0;
5296         const unsigned int ia_valid = attr->ia_valid;
5297
5298         error = inode_change_ok(inode, attr);
5299         if (error)
5300                 return error;
5301
5302         if (is_quota_modification(inode, attr))
5303                 dquot_initialize(inode);
5304         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
5305                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
5306                 handle_t *handle;
5307
5308                 /* (user+group)*(old+new) structure, inode write (sb,
5309                  * inode block, ? - but truncate inode update has it) */
5310                 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
5311                                         EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
5312                 if (IS_ERR(handle)) {
5313                         error = PTR_ERR(handle);
5314                         goto err_out;
5315                 }
5316                 error = dquot_transfer(inode, attr);
5317                 if (error) {
5318                         ext4_journal_stop(handle);
5319                         return error;
5320                 }
5321                 /* Update corresponding info in inode so that everything is in
5322                  * one transaction */
5323                 if (attr->ia_valid & ATTR_UID)
5324                         inode->i_uid = attr->ia_uid;
5325                 if (attr->ia_valid & ATTR_GID)
5326                         inode->i_gid = attr->ia_gid;
5327                 error = ext4_mark_inode_dirty(handle, inode);
5328                 ext4_journal_stop(handle);
5329         }
5330
5331         if (attr->ia_valid & ATTR_SIZE) {
5332                 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
5333                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5334
5335                         if (attr->ia_size > sbi->s_bitmap_maxbytes)
5336                                 return -EFBIG;
5337                 }
5338         }
5339
5340         if (S_ISREG(inode->i_mode) &&
5341             attr->ia_valid & ATTR_SIZE &&
5342             (attr->ia_size < inode->i_size ||
5343              (ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS)))) {
5344                 handle_t *handle;
5345
5346                 handle = ext4_journal_start(inode, 3);
5347                 if (IS_ERR(handle)) {
5348                         error = PTR_ERR(handle);
5349                         goto err_out;
5350                 }
5351                 if (ext4_handle_valid(handle)) {
5352                         error = ext4_orphan_add(handle, inode);
5353                         orphan = 1;
5354                 }
5355                 EXT4_I(inode)->i_disksize = attr->ia_size;
5356                 rc = ext4_mark_inode_dirty(handle, inode);
5357                 if (!error)
5358                         error = rc;
5359                 ext4_journal_stop(handle);
5360
5361                 if (ext4_should_order_data(inode)) {
5362                         error = ext4_begin_ordered_truncate(inode,
5363                                                             attr->ia_size);
5364                         if (error) {
5365                                 /* Do as much error cleanup as possible */
5366                                 handle = ext4_journal_start(inode, 3);
5367                                 if (IS_ERR(handle)) {
5368                                         ext4_orphan_del(NULL, inode);
5369                                         goto err_out;
5370                                 }
5371                                 ext4_orphan_del(handle, inode);
5372                                 orphan = 0;
5373                                 ext4_journal_stop(handle);
5374                                 goto err_out;
5375                         }
5376                 }
5377                 /* ext4_truncate will clear the flag */
5378                 if ((ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS)))
5379                         ext4_truncate(inode);
5380         }
5381
5382         if ((attr->ia_valid & ATTR_SIZE) &&
5383             attr->ia_size != i_size_read(inode))
5384                 rc = vmtruncate(inode, attr->ia_size);
5385
5386         if (!rc) {
5387                 setattr_copy(inode, attr);
5388                 mark_inode_dirty(inode);
5389         }
5390
5391         /*
5392          * If the call to ext4_truncate failed to get a transaction handle at
5393          * all, we need to clean up the in-core orphan list manually.
5394          */
5395         if (orphan && inode->i_nlink)
5396                 ext4_orphan_del(NULL, inode);
5397
5398         if (!rc && (ia_valid & ATTR_MODE))
5399                 rc = ext4_acl_chmod(inode);
5400
5401 err_out:
5402         ext4_std_error(inode->i_sb, error);
5403         if (!error)
5404                 error = rc;
5405         return error;
5406 }
5407
5408 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
5409                  struct kstat *stat)
5410 {
5411         struct inode *inode;
5412         unsigned long delalloc_blocks;
5413
5414         inode = dentry->d_inode;
5415         generic_fillattr(inode, stat);
5416
5417         /*
5418          * We can't update i_blocks if the block allocation is delayed
5419          * otherwise in the case of system crash before the real block
5420          * allocation is done, we will have i_blocks inconsistent with
5421          * on-disk file blocks.
5422          * We always keep i_blocks updated together with real
5423          * allocation. But to not confuse with user, stat
5424          * will return the blocks that include the delayed allocation
5425          * blocks for this file.
5426          */
5427         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
5428
5429         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
5430         return 0;
5431 }
5432
5433 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
5434                                       int chunk)
5435 {
5436         int indirects;
5437
5438         /* if nrblocks are contiguous */
5439         if (chunk) {
5440                 /*
5441                  * With N contiguous data blocks, it need at most
5442                  * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) indirect blocks
5443                  * 2 dindirect blocks
5444                  * 1 tindirect block
5445                  */
5446                 indirects = nrblocks / EXT4_ADDR_PER_BLOCK(inode->i_sb);
5447                 return indirects + 3;
5448         }
5449         /*
5450          * if nrblocks are not contiguous, worse case, each block touch
5451          * a indirect block, and each indirect block touch a double indirect
5452          * block, plus a triple indirect block
5453          */
5454         indirects = nrblocks * 2 + 1;
5455         return indirects;
5456 }
5457
5458 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5459 {
5460         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5461                 return ext4_indirect_trans_blocks(inode, nrblocks, chunk);
5462         return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
5463 }
5464
5465 /*
5466  * Account for index blocks, block groups bitmaps and block group
5467  * descriptor blocks if modify datablocks and index blocks
5468  * worse case, the indexs blocks spread over different block groups
5469  *
5470  * If datablocks are discontiguous, they are possible to spread over
5471  * different block groups too. If they are contiuguous, with flexbg,
5472  * they could still across block group boundary.
5473  *
5474  * Also account for superblock, inode, quota and xattr blocks
5475  */
5476 static int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5477 {
5478         ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
5479         int gdpblocks;
5480         int idxblocks;
5481         int ret = 0;
5482
5483         /*
5484          * How many index blocks need to touch to modify nrblocks?
5485          * The "Chunk" flag indicating whether the nrblocks is
5486          * physically contiguous on disk
5487          *
5488          * For Direct IO and fallocate, they calls get_block to allocate
5489          * one single extent at a time, so they could set the "Chunk" flag
5490          */
5491         idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
5492
5493         ret = idxblocks;
5494
5495         /*
5496          * Now let's see how many group bitmaps and group descriptors need
5497          * to account
5498          */
5499         groups = idxblocks;
5500         if (chunk)
5501                 groups += 1;
5502         else
5503                 groups += nrblocks;
5504
5505         gdpblocks = groups;
5506         if (groups > ngroups)
5507                 groups = ngroups;
5508         if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
5509                 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
5510
5511         /* bitmaps and block group descriptor blocks */
5512         ret += groups + gdpblocks;
5513
5514         /* Blocks for super block, inode, quota and xattr blocks */
5515         ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
5516
5517         return ret;
5518 }
5519
5520 /*
5521  * Calulate the total number of credits to reserve to fit
5522  * the modification of a single pages into a single transaction,
5523  * which may include multiple chunks of block allocations.
5524  *
5525  * This could be called via ext4_write_begin()
5526  *
5527  * We need to consider the worse case, when
5528  * one new block per extent.
5529  */
5530 int ext4_writepage_trans_blocks(struct inode *inode)
5531 {
5532         int bpp = ext4_journal_blocks_per_page(inode);
5533         int ret;
5534