ext4: fix ext4_ext_direct_IO()'s return value after converting uninit extents
[linux-2.6.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/jbd2.h>
29 #include <linux/highuid.h>
30 #include <linux/pagemap.h>
31 #include <linux/quotaops.h>
32 #include <linux/string.h>
33 #include <linux/buffer_head.h>
34 #include <linux/writeback.h>
35 #include <linux/pagevec.h>
36 #include <linux/mpage.h>
37 #include <linux/namei.h>
38 #include <linux/uio.h>
39 #include <linux/bio.h>
40 #include <linux/workqueue.h>
41
42 #include "ext4_jbd2.h"
43 #include "xattr.h"
44 #include "acl.h"
45 #include "ext4_extents.h"
46
47 #include <trace/events/ext4.h>
48
49 #define MPAGE_DA_EXTENT_TAIL 0x01
50
51 static inline int ext4_begin_ordered_truncate(struct inode *inode,
52                                               loff_t new_size)
53 {
54         return jbd2_journal_begin_ordered_truncate(
55                                         EXT4_SB(inode->i_sb)->s_journal,
56                                         &EXT4_I(inode)->jinode,
57                                         new_size);
58 }
59
60 static void ext4_invalidatepage(struct page *page, unsigned long offset);
61
62 /*
63  * Test whether an inode is a fast symlink.
64  */
65 static int ext4_inode_is_fast_symlink(struct inode *inode)
66 {
67         int ea_blocks = EXT4_I(inode)->i_file_acl ?
68                 (inode->i_sb->s_blocksize >> 9) : 0;
69
70         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
71 }
72
73 /*
74  * The ext4 forget function must perform a revoke if we are freeing data
75  * which has been journaled.  Metadata (eg. indirect blocks) must be
76  * revoked in all cases.
77  *
78  * "bh" may be NULL: a metadata block may have been freed from memory
79  * but there may still be a record of it in the journal, and that record
80  * still needs to be revoked.
81  *
82  * If the handle isn't valid we're not journaling, but we still need to
83  * call into ext4_journal_revoke() to put the buffer head.
84  */
85 int ext4_forget(handle_t *handle, int is_metadata, struct inode *inode,
86                 struct buffer_head *bh, ext4_fsblk_t blocknr)
87 {
88         int err;
89
90         might_sleep();
91
92         BUFFER_TRACE(bh, "enter");
93
94         jbd_debug(4, "forgetting bh %p: is_metadata = %d, mode %o, "
95                   "data mode %x\n",
96                   bh, is_metadata, inode->i_mode,
97                   test_opt(inode->i_sb, DATA_FLAGS));
98
99         /* Never use the revoke function if we are doing full data
100          * journaling: there is no need to, and a V1 superblock won't
101          * support it.  Otherwise, only skip the revoke on un-journaled
102          * data blocks. */
103
104         if (test_opt(inode->i_sb, DATA_FLAGS) == EXT4_MOUNT_JOURNAL_DATA ||
105             (!is_metadata && !ext4_should_journal_data(inode))) {
106                 if (bh) {
107                         BUFFER_TRACE(bh, "call jbd2_journal_forget");
108                         return ext4_journal_forget(handle, bh);
109                 }
110                 return 0;
111         }
112
113         /*
114          * data!=journal && (is_metadata || should_journal_data(inode))
115          */
116         BUFFER_TRACE(bh, "call ext4_journal_revoke");
117         err = ext4_journal_revoke(handle, blocknr, bh);
118         if (err)
119                 ext4_abort(inode->i_sb, __func__,
120                            "error %d when attempting revoke", err);
121         BUFFER_TRACE(bh, "exit");
122         return err;
123 }
124
125 /*
126  * Work out how many blocks we need to proceed with the next chunk of a
127  * truncate transaction.
128  */
129 static unsigned long blocks_for_truncate(struct inode *inode)
130 {
131         ext4_lblk_t needed;
132
133         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
134
135         /* Give ourselves just enough room to cope with inodes in which
136          * i_blocks is corrupt: we've seen disk corruptions in the past
137          * which resulted in random data in an inode which looked enough
138          * like a regular file for ext4 to try to delete it.  Things
139          * will go a bit crazy if that happens, but at least we should
140          * try not to panic the whole kernel. */
141         if (needed < 2)
142                 needed = 2;
143
144         /* But we need to bound the transaction so we don't overflow the
145          * journal. */
146         if (needed > EXT4_MAX_TRANS_DATA)
147                 needed = EXT4_MAX_TRANS_DATA;
148
149         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
150 }
151
152 /*
153  * Truncate transactions can be complex and absolutely huge.  So we need to
154  * be able to restart the transaction at a conventient checkpoint to make
155  * sure we don't overflow the journal.
156  *
157  * start_transaction gets us a new handle for a truncate transaction,
158  * and extend_transaction tries to extend the existing one a bit.  If
159  * extend fails, we need to propagate the failure up and restart the
160  * transaction in the top-level truncate loop. --sct
161  */
162 static handle_t *start_transaction(struct inode *inode)
163 {
164         handle_t *result;
165
166         result = ext4_journal_start(inode, blocks_for_truncate(inode));
167         if (!IS_ERR(result))
168                 return result;
169
170         ext4_std_error(inode->i_sb, PTR_ERR(result));
171         return result;
172 }
173
174 /*
175  * Try to extend this transaction for the purposes of truncation.
176  *
177  * Returns 0 if we managed to create more room.  If we can't create more
178  * room, and the transaction must be restarted we return 1.
179  */
180 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
181 {
182         if (!ext4_handle_valid(handle))
183                 return 0;
184         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
185                 return 0;
186         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
187                 return 0;
188         return 1;
189 }
190
191 /*
192  * Restart the transaction associated with *handle.  This does a commit,
193  * so before we call here everything must be consistently dirtied against
194  * this transaction.
195  */
196 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
197                                  int nblocks)
198 {
199         int ret;
200
201         /*
202          * Drop i_data_sem to avoid deadlock with ext4_get_blocks At this
203          * moment, get_block can be called only for blocks inside i_size since
204          * page cache has been already dropped and writes are blocked by
205          * i_mutex. So we can safely drop the i_data_sem here.
206          */
207         BUG_ON(EXT4_JOURNAL(inode) == NULL);
208         jbd_debug(2, "restarting handle %p\n", handle);
209         up_write(&EXT4_I(inode)->i_data_sem);
210         ret = ext4_journal_restart(handle, blocks_for_truncate(inode));
211         down_write(&EXT4_I(inode)->i_data_sem);
212         ext4_discard_preallocations(inode);
213
214         return ret;
215 }
216
217 /*
218  * Called at the last iput() if i_nlink is zero.
219  */
220 void ext4_delete_inode(struct inode *inode)
221 {
222         handle_t *handle;
223         int err;
224
225         if (ext4_should_order_data(inode))
226                 ext4_begin_ordered_truncate(inode, 0);
227         truncate_inode_pages(&inode->i_data, 0);
228
229         if (is_bad_inode(inode))
230                 goto no_delete;
231
232         handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
233         if (IS_ERR(handle)) {
234                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
235                 /*
236                  * If we're going to skip the normal cleanup, we still need to
237                  * make sure that the in-core orphan linked list is properly
238                  * cleaned up.
239                  */
240                 ext4_orphan_del(NULL, inode);
241                 goto no_delete;
242         }
243
244         if (IS_SYNC(inode))
245                 ext4_handle_sync(handle);
246         inode->i_size = 0;
247         err = ext4_mark_inode_dirty(handle, inode);
248         if (err) {
249                 ext4_warning(inode->i_sb, __func__,
250                              "couldn't mark inode dirty (err %d)", err);
251                 goto stop_handle;
252         }
253         if (inode->i_blocks)
254                 ext4_truncate(inode);
255
256         /*
257          * ext4_ext_truncate() doesn't reserve any slop when it
258          * restarts journal transactions; therefore there may not be
259          * enough credits left in the handle to remove the inode from
260          * the orphan list and set the dtime field.
261          */
262         if (!ext4_handle_has_enough_credits(handle, 3)) {
263                 err = ext4_journal_extend(handle, 3);
264                 if (err > 0)
265                         err = ext4_journal_restart(handle, 3);
266                 if (err != 0) {
267                         ext4_warning(inode->i_sb, __func__,
268                                      "couldn't extend journal (err %d)", err);
269                 stop_handle:
270                         ext4_journal_stop(handle);
271                         goto no_delete;
272                 }
273         }
274
275         /*
276          * Kill off the orphan record which ext4_truncate created.
277          * AKPM: I think this can be inside the above `if'.
278          * Note that ext4_orphan_del() has to be able to cope with the
279          * deletion of a non-existent orphan - this is because we don't
280          * know if ext4_truncate() actually created an orphan record.
281          * (Well, we could do this if we need to, but heck - it works)
282          */
283         ext4_orphan_del(handle, inode);
284         EXT4_I(inode)->i_dtime  = get_seconds();
285
286         /*
287          * One subtle ordering requirement: if anything has gone wrong
288          * (transaction abort, IO errors, whatever), then we can still
289          * do these next steps (the fs will already have been marked as
290          * having errors), but we can't free the inode if the mark_dirty
291          * fails.
292          */
293         if (ext4_mark_inode_dirty(handle, inode))
294                 /* If that failed, just do the required in-core inode clear. */
295                 clear_inode(inode);
296         else
297                 ext4_free_inode(handle, inode);
298         ext4_journal_stop(handle);
299         return;
300 no_delete:
301         clear_inode(inode);     /* We must guarantee clearing of inode... */
302 }
303
304 typedef struct {
305         __le32  *p;
306         __le32  key;
307         struct buffer_head *bh;
308 } Indirect;
309
310 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
311 {
312         p->key = *(p->p = v);
313         p->bh = bh;
314 }
315
316 /**
317  *      ext4_block_to_path - parse the block number into array of offsets
318  *      @inode: inode in question (we are only interested in its superblock)
319  *      @i_block: block number to be parsed
320  *      @offsets: array to store the offsets in
321  *      @boundary: set this non-zero if the referred-to block is likely to be
322  *             followed (on disk) by an indirect block.
323  *
324  *      To store the locations of file's data ext4 uses a data structure common
325  *      for UNIX filesystems - tree of pointers anchored in the inode, with
326  *      data blocks at leaves and indirect blocks in intermediate nodes.
327  *      This function translates the block number into path in that tree -
328  *      return value is the path length and @offsets[n] is the offset of
329  *      pointer to (n+1)th node in the nth one. If @block is out of range
330  *      (negative or too large) warning is printed and zero returned.
331  *
332  *      Note: function doesn't find node addresses, so no IO is needed. All
333  *      we need to know is the capacity of indirect blocks (taken from the
334  *      inode->i_sb).
335  */
336
337 /*
338  * Portability note: the last comparison (check that we fit into triple
339  * indirect block) is spelled differently, because otherwise on an
340  * architecture with 32-bit longs and 8Kb pages we might get into trouble
341  * if our filesystem had 8Kb blocks. We might use long long, but that would
342  * kill us on x86. Oh, well, at least the sign propagation does not matter -
343  * i_block would have to be negative in the very beginning, so we would not
344  * get there at all.
345  */
346
347 static int ext4_block_to_path(struct inode *inode,
348                               ext4_lblk_t i_block,
349                               ext4_lblk_t offsets[4], int *boundary)
350 {
351         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
352         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
353         const long direct_blocks = EXT4_NDIR_BLOCKS,
354                 indirect_blocks = ptrs,
355                 double_blocks = (1 << (ptrs_bits * 2));
356         int n = 0;
357         int final = 0;
358
359         if (i_block < direct_blocks) {
360                 offsets[n++] = i_block;
361                 final = direct_blocks;
362         } else if ((i_block -= direct_blocks) < indirect_blocks) {
363                 offsets[n++] = EXT4_IND_BLOCK;
364                 offsets[n++] = i_block;
365                 final = ptrs;
366         } else if ((i_block -= indirect_blocks) < double_blocks) {
367                 offsets[n++] = EXT4_DIND_BLOCK;
368                 offsets[n++] = i_block >> ptrs_bits;
369                 offsets[n++] = i_block & (ptrs - 1);
370                 final = ptrs;
371         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
372                 offsets[n++] = EXT4_TIND_BLOCK;
373                 offsets[n++] = i_block >> (ptrs_bits * 2);
374                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
375                 offsets[n++] = i_block & (ptrs - 1);
376                 final = ptrs;
377         } else {
378                 ext4_warning(inode->i_sb, "ext4_block_to_path",
379                              "block %lu > max in inode %lu",
380                              i_block + direct_blocks +
381                              indirect_blocks + double_blocks, inode->i_ino);
382         }
383         if (boundary)
384                 *boundary = final - 1 - (i_block & (ptrs - 1));
385         return n;
386 }
387
388 static int __ext4_check_blockref(const char *function, struct inode *inode,
389                                  __le32 *p, unsigned int max)
390 {
391         __le32 *bref = p;
392         unsigned int blk;
393
394         while (bref < p+max) {
395                 blk = le32_to_cpu(*bref++);
396                 if (blk &&
397                     unlikely(!ext4_data_block_valid(EXT4_SB(inode->i_sb),
398                                                     blk, 1))) {
399                         ext4_error(inode->i_sb, function,
400                                    "invalid block reference %u "
401                                    "in inode #%lu", blk, inode->i_ino);
402                         return -EIO;
403                 }
404         }
405         return 0;
406 }
407
408
409 #define ext4_check_indirect_blockref(inode, bh)                         \
410         __ext4_check_blockref(__func__, inode, (__le32 *)(bh)->b_data,  \
411                               EXT4_ADDR_PER_BLOCK((inode)->i_sb))
412
413 #define ext4_check_inode_blockref(inode)                                \
414         __ext4_check_blockref(__func__, inode, EXT4_I(inode)->i_data,   \
415                               EXT4_NDIR_BLOCKS)
416
417 /**
418  *      ext4_get_branch - read the chain of indirect blocks leading to data
419  *      @inode: inode in question
420  *      @depth: depth of the chain (1 - direct pointer, etc.)
421  *      @offsets: offsets of pointers in inode/indirect blocks
422  *      @chain: place to store the result
423  *      @err: here we store the error value
424  *
425  *      Function fills the array of triples <key, p, bh> and returns %NULL
426  *      if everything went OK or the pointer to the last filled triple
427  *      (incomplete one) otherwise. Upon the return chain[i].key contains
428  *      the number of (i+1)-th block in the chain (as it is stored in memory,
429  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
430  *      number (it points into struct inode for i==0 and into the bh->b_data
431  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
432  *      block for i>0 and NULL for i==0. In other words, it holds the block
433  *      numbers of the chain, addresses they were taken from (and where we can
434  *      verify that chain did not change) and buffer_heads hosting these
435  *      numbers.
436  *
437  *      Function stops when it stumbles upon zero pointer (absent block)
438  *              (pointer to last triple returned, *@err == 0)
439  *      or when it gets an IO error reading an indirect block
440  *              (ditto, *@err == -EIO)
441  *      or when it reads all @depth-1 indirect blocks successfully and finds
442  *      the whole chain, all way to the data (returns %NULL, *err == 0).
443  *
444  *      Need to be called with
445  *      down_read(&EXT4_I(inode)->i_data_sem)
446  */
447 static Indirect *ext4_get_branch(struct inode *inode, int depth,
448                                  ext4_lblk_t  *offsets,
449                                  Indirect chain[4], int *err)
450 {
451         struct super_block *sb = inode->i_sb;
452         Indirect *p = chain;
453         struct buffer_head *bh;
454
455         *err = 0;
456         /* i_data is not going away, no lock needed */
457         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
458         if (!p->key)
459                 goto no_block;
460         while (--depth) {
461                 bh = sb_getblk(sb, le32_to_cpu(p->key));
462                 if (unlikely(!bh))
463                         goto failure;
464
465                 if (!bh_uptodate_or_lock(bh)) {
466                         if (bh_submit_read(bh) < 0) {
467                                 put_bh(bh);
468                                 goto failure;
469                         }
470                         /* validate block references */
471                         if (ext4_check_indirect_blockref(inode, bh)) {
472                                 put_bh(bh);
473                                 goto failure;
474                         }
475                 }
476
477                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
478                 /* Reader: end */
479                 if (!p->key)
480                         goto no_block;
481         }
482         return NULL;
483
484 failure:
485         *err = -EIO;
486 no_block:
487         return p;
488 }
489
490 /**
491  *      ext4_find_near - find a place for allocation with sufficient locality
492  *      @inode: owner
493  *      @ind: descriptor of indirect block.
494  *
495  *      This function returns the preferred place for block allocation.
496  *      It is used when heuristic for sequential allocation fails.
497  *      Rules are:
498  *        + if there is a block to the left of our position - allocate near it.
499  *        + if pointer will live in indirect block - allocate near that block.
500  *        + if pointer will live in inode - allocate in the same
501  *          cylinder group.
502  *
503  * In the latter case we colour the starting block by the callers PID to
504  * prevent it from clashing with concurrent allocations for a different inode
505  * in the same block group.   The PID is used here so that functionally related
506  * files will be close-by on-disk.
507  *
508  *      Caller must make sure that @ind is valid and will stay that way.
509  */
510 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
511 {
512         struct ext4_inode_info *ei = EXT4_I(inode);
513         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
514         __le32 *p;
515         ext4_fsblk_t bg_start;
516         ext4_fsblk_t last_block;
517         ext4_grpblk_t colour;
518         ext4_group_t block_group;
519         int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
520
521         /* Try to find previous block */
522         for (p = ind->p - 1; p >= start; p--) {
523                 if (*p)
524                         return le32_to_cpu(*p);
525         }
526
527         /* No such thing, so let's try location of indirect block */
528         if (ind->bh)
529                 return ind->bh->b_blocknr;
530
531         /*
532          * It is going to be referred to from the inode itself? OK, just put it
533          * into the same cylinder group then.
534          */
535         block_group = ei->i_block_group;
536         if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
537                 block_group &= ~(flex_size-1);
538                 if (S_ISREG(inode->i_mode))
539                         block_group++;
540         }
541         bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
542         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
543
544         /*
545          * If we are doing delayed allocation, we don't need take
546          * colour into account.
547          */
548         if (test_opt(inode->i_sb, DELALLOC))
549                 return bg_start;
550
551         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
552                 colour = (current->pid % 16) *
553                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
554         else
555                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
556         return bg_start + colour;
557 }
558
559 /**
560  *      ext4_find_goal - find a preferred place for allocation.
561  *      @inode: owner
562  *      @block:  block we want
563  *      @partial: pointer to the last triple within a chain
564  *
565  *      Normally this function find the preferred place for block allocation,
566  *      returns it.
567  *      Because this is only used for non-extent files, we limit the block nr
568  *      to 32 bits.
569  */
570 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
571                                    Indirect *partial)
572 {
573         ext4_fsblk_t goal;
574
575         /*
576          * XXX need to get goal block from mballoc's data structures
577          */
578
579         goal = ext4_find_near(inode, partial);
580         goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
581         return goal;
582 }
583
584 /**
585  *      ext4_blks_to_allocate: Look up the block map and count the number
586  *      of direct blocks need to be allocated for the given branch.
587  *
588  *      @branch: chain of indirect blocks
589  *      @k: number of blocks need for indirect blocks
590  *      @blks: number of data blocks to be mapped.
591  *      @blocks_to_boundary:  the offset in the indirect block
592  *
593  *      return the total number of blocks to be allocate, including the
594  *      direct and indirect blocks.
595  */
596 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
597                                  int blocks_to_boundary)
598 {
599         unsigned int count = 0;
600
601         /*
602          * Simple case, [t,d]Indirect block(s) has not allocated yet
603          * then it's clear blocks on that path have not allocated
604          */
605         if (k > 0) {
606                 /* right now we don't handle cross boundary allocation */
607                 if (blks < blocks_to_boundary + 1)
608                         count += blks;
609                 else
610                         count += blocks_to_boundary + 1;
611                 return count;
612         }
613
614         count++;
615         while (count < blks && count <= blocks_to_boundary &&
616                 le32_to_cpu(*(branch[0].p + count)) == 0) {
617                 count++;
618         }
619         return count;
620 }
621
622 /**
623  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
624  *      @indirect_blks: the number of blocks need to allocate for indirect
625  *                      blocks
626  *
627  *      @new_blocks: on return it will store the new block numbers for
628  *      the indirect blocks(if needed) and the first direct block,
629  *      @blks:  on return it will store the total number of allocated
630  *              direct blocks
631  */
632 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
633                              ext4_lblk_t iblock, ext4_fsblk_t goal,
634                              int indirect_blks, int blks,
635                              ext4_fsblk_t new_blocks[4], int *err)
636 {
637         struct ext4_allocation_request ar;
638         int target, i;
639         unsigned long count = 0, blk_allocated = 0;
640         int index = 0;
641         ext4_fsblk_t current_block = 0;
642         int ret = 0;
643
644         /*
645          * Here we try to allocate the requested multiple blocks at once,
646          * on a best-effort basis.
647          * To build a branch, we should allocate blocks for
648          * the indirect blocks(if not allocated yet), and at least
649          * the first direct block of this branch.  That's the
650          * minimum number of blocks need to allocate(required)
651          */
652         /* first we try to allocate the indirect blocks */
653         target = indirect_blks;
654         while (target > 0) {
655                 count = target;
656                 /* allocating blocks for indirect blocks and direct blocks */
657                 current_block = ext4_new_meta_blocks(handle, inode,
658                                                         goal, &count, err);
659                 if (*err)
660                         goto failed_out;
661
662                 BUG_ON(current_block + count > EXT4_MAX_BLOCK_FILE_PHYS);
663
664                 target -= count;
665                 /* allocate blocks for indirect blocks */
666                 while (index < indirect_blks && count) {
667                         new_blocks[index++] = current_block++;
668                         count--;
669                 }
670                 if (count > 0) {
671                         /*
672                          * save the new block number
673                          * for the first direct block
674                          */
675                         new_blocks[index] = current_block;
676                         printk(KERN_INFO "%s returned more blocks than "
677                                                 "requested\n", __func__);
678                         WARN_ON(1);
679                         break;
680                 }
681         }
682
683         target = blks - count ;
684         blk_allocated = count;
685         if (!target)
686                 goto allocated;
687         /* Now allocate data blocks */
688         memset(&ar, 0, sizeof(ar));
689         ar.inode = inode;
690         ar.goal = goal;
691         ar.len = target;
692         ar.logical = iblock;
693         if (S_ISREG(inode->i_mode))
694                 /* enable in-core preallocation only for regular files */
695                 ar.flags = EXT4_MB_HINT_DATA;
696
697         current_block = ext4_mb_new_blocks(handle, &ar, err);
698         BUG_ON(current_block + ar.len > EXT4_MAX_BLOCK_FILE_PHYS);
699
700         if (*err && (target == blks)) {
701                 /*
702                  * if the allocation failed and we didn't allocate
703                  * any blocks before
704                  */
705                 goto failed_out;
706         }
707         if (!*err) {
708                 if (target == blks) {
709                         /*
710                          * save the new block number
711                          * for the first direct block
712                          */
713                         new_blocks[index] = current_block;
714                 }
715                 blk_allocated += ar.len;
716         }
717 allocated:
718         /* total number of blocks allocated for direct blocks */
719         ret = blk_allocated;
720         *err = 0;
721         return ret;
722 failed_out:
723         for (i = 0; i < index; i++)
724                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
725         return ret;
726 }
727
728 /**
729  *      ext4_alloc_branch - allocate and set up a chain of blocks.
730  *      @inode: owner
731  *      @indirect_blks: number of allocated indirect blocks
732  *      @blks: number of allocated direct blocks
733  *      @offsets: offsets (in the blocks) to store the pointers to next.
734  *      @branch: place to store the chain in.
735  *
736  *      This function allocates blocks, zeroes out all but the last one,
737  *      links them into chain and (if we are synchronous) writes them to disk.
738  *      In other words, it prepares a branch that can be spliced onto the
739  *      inode. It stores the information about that chain in the branch[], in
740  *      the same format as ext4_get_branch() would do. We are calling it after
741  *      we had read the existing part of chain and partial points to the last
742  *      triple of that (one with zero ->key). Upon the exit we have the same
743  *      picture as after the successful ext4_get_block(), except that in one
744  *      place chain is disconnected - *branch->p is still zero (we did not
745  *      set the last link), but branch->key contains the number that should
746  *      be placed into *branch->p to fill that gap.
747  *
748  *      If allocation fails we free all blocks we've allocated (and forget
749  *      their buffer_heads) and return the error value the from failed
750  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
751  *      as described above and return 0.
752  */
753 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
754                              ext4_lblk_t iblock, int indirect_blks,
755                              int *blks, ext4_fsblk_t goal,
756                              ext4_lblk_t *offsets, Indirect *branch)
757 {
758         int blocksize = inode->i_sb->s_blocksize;
759         int i, n = 0;
760         int err = 0;
761         struct buffer_head *bh;
762         int num;
763         ext4_fsblk_t new_blocks[4];
764         ext4_fsblk_t current_block;
765
766         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
767                                 *blks, new_blocks, &err);
768         if (err)
769                 return err;
770
771         branch[0].key = cpu_to_le32(new_blocks[0]);
772         /*
773          * metadata blocks and data blocks are allocated.
774          */
775         for (n = 1; n <= indirect_blks;  n++) {
776                 /*
777                  * Get buffer_head for parent block, zero it out
778                  * and set the pointer to new one, then send
779                  * parent to disk.
780                  */
781                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
782                 branch[n].bh = bh;
783                 lock_buffer(bh);
784                 BUFFER_TRACE(bh, "call get_create_access");
785                 err = ext4_journal_get_create_access(handle, bh);
786                 if (err) {
787                         /* Don't brelse(bh) here; it's done in
788                          * ext4_journal_forget() below */
789                         unlock_buffer(bh);
790                         goto failed;
791                 }
792
793                 memset(bh->b_data, 0, blocksize);
794                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
795                 branch[n].key = cpu_to_le32(new_blocks[n]);
796                 *branch[n].p = branch[n].key;
797                 if (n == indirect_blks) {
798                         current_block = new_blocks[n];
799                         /*
800                          * End of chain, update the last new metablock of
801                          * the chain to point to the new allocated
802                          * data blocks numbers
803                          */
804                         for (i = 1; i < num; i++)
805                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
806                 }
807                 BUFFER_TRACE(bh, "marking uptodate");
808                 set_buffer_uptodate(bh);
809                 unlock_buffer(bh);
810
811                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
812                 err = ext4_handle_dirty_metadata(handle, inode, bh);
813                 if (err)
814                         goto failed;
815         }
816         *blks = num;
817         return err;
818 failed:
819         /* Allocation failed, free what we already allocated */
820         for (i = 1; i <= n ; i++) {
821                 BUFFER_TRACE(branch[i].bh, "call jbd2_journal_forget");
822                 ext4_journal_forget(handle, branch[i].bh);
823         }
824         for (i = 0; i < indirect_blks; i++)
825                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
826
827         ext4_free_blocks(handle, inode, new_blocks[i], num, 0);
828
829         return err;
830 }
831
832 /**
833  * ext4_splice_branch - splice the allocated branch onto inode.
834  * @inode: owner
835  * @block: (logical) number of block we are adding
836  * @chain: chain of indirect blocks (with a missing link - see
837  *      ext4_alloc_branch)
838  * @where: location of missing link
839  * @num:   number of indirect blocks we are adding
840  * @blks:  number of direct blocks we are adding
841  *
842  * This function fills the missing link and does all housekeeping needed in
843  * inode (->i_blocks, etc.). In case of success we end up with the full
844  * chain to new block and return 0.
845  */
846 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
847                               ext4_lblk_t block, Indirect *where, int num,
848                               int blks)
849 {
850         int i;
851         int err = 0;
852         ext4_fsblk_t current_block;
853
854         /*
855          * If we're splicing into a [td]indirect block (as opposed to the
856          * inode) then we need to get write access to the [td]indirect block
857          * before the splice.
858          */
859         if (where->bh) {
860                 BUFFER_TRACE(where->bh, "get_write_access");
861                 err = ext4_journal_get_write_access(handle, where->bh);
862                 if (err)
863                         goto err_out;
864         }
865         /* That's it */
866
867         *where->p = where->key;
868
869         /*
870          * Update the host buffer_head or inode to point to more just allocated
871          * direct blocks blocks
872          */
873         if (num == 0 && blks > 1) {
874                 current_block = le32_to_cpu(where->key) + 1;
875                 for (i = 1; i < blks; i++)
876                         *(where->p + i) = cpu_to_le32(current_block++);
877         }
878
879         /* We are done with atomic stuff, now do the rest of housekeeping */
880         /* had we spliced it onto indirect block? */
881         if (where->bh) {
882                 /*
883                  * If we spliced it onto an indirect block, we haven't
884                  * altered the inode.  Note however that if it is being spliced
885                  * onto an indirect block at the very end of the file (the
886                  * file is growing) then we *will* alter the inode to reflect
887                  * the new i_size.  But that is not done here - it is done in
888                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
889                  */
890                 jbd_debug(5, "splicing indirect only\n");
891                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
892                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
893                 if (err)
894                         goto err_out;
895         } else {
896                 /*
897                  * OK, we spliced it into the inode itself on a direct block.
898                  */
899                 ext4_mark_inode_dirty(handle, inode);
900                 jbd_debug(5, "splicing direct\n");
901         }
902         return err;
903
904 err_out:
905         for (i = 1; i <= num; i++) {
906                 BUFFER_TRACE(where[i].bh, "call jbd2_journal_forget");
907                 ext4_journal_forget(handle, where[i].bh);
908                 ext4_free_blocks(handle, inode,
909                                         le32_to_cpu(where[i-1].key), 1, 0);
910         }
911         ext4_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks, 0);
912
913         return err;
914 }
915
916 /*
917  * The ext4_ind_get_blocks() function handles non-extents inodes
918  * (i.e., using the traditional indirect/double-indirect i_blocks
919  * scheme) for ext4_get_blocks().
920  *
921  * Allocation strategy is simple: if we have to allocate something, we will
922  * have to go the whole way to leaf. So let's do it before attaching anything
923  * to tree, set linkage between the newborn blocks, write them if sync is
924  * required, recheck the path, free and repeat if check fails, otherwise
925  * set the last missing link (that will protect us from any truncate-generated
926  * removals - all blocks on the path are immune now) and possibly force the
927  * write on the parent block.
928  * That has a nice additional property: no special recovery from the failed
929  * allocations is needed - we simply release blocks and do not touch anything
930  * reachable from inode.
931  *
932  * `handle' can be NULL if create == 0.
933  *
934  * return > 0, # of blocks mapped or allocated.
935  * return = 0, if plain lookup failed.
936  * return < 0, error case.
937  *
938  * The ext4_ind_get_blocks() function should be called with
939  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
940  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
941  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
942  * blocks.
943  */
944 static int ext4_ind_get_blocks(handle_t *handle, struct inode *inode,
945                                ext4_lblk_t iblock, unsigned int maxblocks,
946                                struct buffer_head *bh_result,
947                                int flags)
948 {
949         int err = -EIO;
950         ext4_lblk_t offsets[4];
951         Indirect chain[4];
952         Indirect *partial;
953         ext4_fsblk_t goal;
954         int indirect_blks;
955         int blocks_to_boundary = 0;
956         int depth;
957         int count = 0;
958         ext4_fsblk_t first_block = 0;
959
960         J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
961         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
962         depth = ext4_block_to_path(inode, iblock, offsets,
963                                    &blocks_to_boundary);
964
965         if (depth == 0)
966                 goto out;
967
968         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
969
970         /* Simplest case - block found, no allocation needed */
971         if (!partial) {
972                 first_block = le32_to_cpu(chain[depth - 1].key);
973                 clear_buffer_new(bh_result);
974                 count++;
975                 /*map more blocks*/
976                 while (count < maxblocks && count <= blocks_to_boundary) {
977                         ext4_fsblk_t blk;
978
979                         blk = le32_to_cpu(*(chain[depth-1].p + count));
980
981                         if (blk == first_block + count)
982                                 count++;
983                         else
984                                 break;
985                 }
986                 goto got_it;
987         }
988
989         /* Next simple case - plain lookup or failed read of indirect block */
990         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
991                 goto cleanup;
992
993         /*
994          * Okay, we need to do block allocation.
995         */
996         goal = ext4_find_goal(inode, iblock, partial);
997
998         /* the number of blocks need to allocate for [d,t]indirect blocks */
999         indirect_blks = (chain + depth) - partial - 1;
1000
1001         /*
1002          * Next look up the indirect map to count the totoal number of
1003          * direct blocks to allocate for this branch.
1004          */
1005         count = ext4_blks_to_allocate(partial, indirect_blks,
1006                                         maxblocks, blocks_to_boundary);
1007         /*
1008          * Block out ext4_truncate while we alter the tree
1009          */
1010         err = ext4_alloc_branch(handle, inode, iblock, indirect_blks,
1011                                 &count, goal,
1012                                 offsets + (partial - chain), partial);
1013
1014         /*
1015          * The ext4_splice_branch call will free and forget any buffers
1016          * on the new chain if there is a failure, but that risks using
1017          * up transaction credits, especially for bitmaps where the
1018          * credits cannot be returned.  Can we handle this somehow?  We
1019          * may need to return -EAGAIN upwards in the worst case.  --sct
1020          */
1021         if (!err)
1022                 err = ext4_splice_branch(handle, inode, iblock,
1023                                          partial, indirect_blks, count);
1024         else
1025                 goto cleanup;
1026
1027         set_buffer_new(bh_result);
1028 got_it:
1029         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
1030         if (count > blocks_to_boundary)
1031                 set_buffer_boundary(bh_result);
1032         err = count;
1033         /* Clean up and exit */
1034         partial = chain + depth - 1;    /* the whole chain */
1035 cleanup:
1036         while (partial > chain) {
1037                 BUFFER_TRACE(partial->bh, "call brelse");
1038                 brelse(partial->bh);
1039                 partial--;
1040         }
1041         BUFFER_TRACE(bh_result, "returned");
1042 out:
1043         return err;
1044 }
1045
1046 qsize_t ext4_get_reserved_space(struct inode *inode)
1047 {
1048         unsigned long long total;
1049
1050         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1051         total = EXT4_I(inode)->i_reserved_data_blocks +
1052                 EXT4_I(inode)->i_reserved_meta_blocks;
1053         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1054
1055         return total;
1056 }
1057 /*
1058  * Calculate the number of metadata blocks need to reserve
1059  * to allocate @blocks for non extent file based file
1060  */
1061 static int ext4_indirect_calc_metadata_amount(struct inode *inode, int blocks)
1062 {
1063         int icap = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1064         int ind_blks, dind_blks, tind_blks;
1065
1066         /* number of new indirect blocks needed */
1067         ind_blks = (blocks + icap - 1) / icap;
1068
1069         dind_blks = (ind_blks + icap - 1) / icap;
1070
1071         tind_blks = 1;
1072
1073         return ind_blks + dind_blks + tind_blks;
1074 }
1075
1076 /*
1077  * Calculate the number of metadata blocks need to reserve
1078  * to allocate given number of blocks
1079  */
1080 static int ext4_calc_metadata_amount(struct inode *inode, int blocks)
1081 {
1082         if (!blocks)
1083                 return 0;
1084
1085         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
1086                 return ext4_ext_calc_metadata_amount(inode, blocks);
1087
1088         return ext4_indirect_calc_metadata_amount(inode, blocks);
1089 }
1090
1091 static void ext4_da_update_reserve_space(struct inode *inode, int used)
1092 {
1093         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1094         int total, mdb, mdb_free;
1095
1096         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1097         /* recalculate the number of metablocks still need to be reserved */
1098         total = EXT4_I(inode)->i_reserved_data_blocks - used;
1099         mdb = ext4_calc_metadata_amount(inode, total);
1100
1101         /* figure out how many metablocks to release */
1102         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1103         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1104
1105         if (mdb_free) {
1106                 /* Account for allocated meta_blocks */
1107                 mdb_free -= EXT4_I(inode)->i_allocated_meta_blocks;
1108
1109                 /* update fs dirty blocks counter */
1110                 percpu_counter_sub(&sbi->s_dirtyblocks_counter, mdb_free);
1111                 EXT4_I(inode)->i_allocated_meta_blocks = 0;
1112                 EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1113         }
1114
1115         /* update per-inode reservations */
1116         BUG_ON(used  > EXT4_I(inode)->i_reserved_data_blocks);
1117         EXT4_I(inode)->i_reserved_data_blocks -= used;
1118         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1119
1120         /*
1121          * free those over-booking quota for metadata blocks
1122          */
1123         if (mdb_free)
1124                 vfs_dq_release_reservation_block(inode, mdb_free);
1125
1126         /*
1127          * If we have done all the pending block allocations and if
1128          * there aren't any writers on the inode, we can discard the
1129          * inode's preallocations.
1130          */
1131         if (!total && (atomic_read(&inode->i_writecount) == 0))
1132                 ext4_discard_preallocations(inode);
1133 }
1134
1135 static int check_block_validity(struct inode *inode, const char *msg,
1136                                 sector_t logical, sector_t phys, int len)
1137 {
1138         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), phys, len)) {
1139                 ext4_error(inode->i_sb, msg,
1140                            "inode #%lu logical block %llu mapped to %llu "
1141                            "(size %d)", inode->i_ino,
1142                            (unsigned long long) logical,
1143                            (unsigned long long) phys, len);
1144                 return -EIO;
1145         }
1146         return 0;
1147 }
1148
1149 /*
1150  * Return the number of contiguous dirty pages in a given inode
1151  * starting at page frame idx.
1152  */
1153 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
1154                                     unsigned int max_pages)
1155 {
1156         struct address_space *mapping = inode->i_mapping;
1157         pgoff_t index;
1158         struct pagevec pvec;
1159         pgoff_t num = 0;
1160         int i, nr_pages, done = 0;
1161
1162         if (max_pages == 0)
1163                 return 0;
1164         pagevec_init(&pvec, 0);
1165         while (!done) {
1166                 index = idx;
1167                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
1168                                               PAGECACHE_TAG_DIRTY,
1169                                               (pgoff_t)PAGEVEC_SIZE);
1170                 if (nr_pages == 0)
1171                         break;
1172                 for (i = 0; i < nr_pages; i++) {
1173                         struct page *page = pvec.pages[i];
1174                         struct buffer_head *bh, *head;
1175
1176                         lock_page(page);
1177                         if (unlikely(page->mapping != mapping) ||
1178                             !PageDirty(page) ||
1179                             PageWriteback(page) ||
1180                             page->index != idx) {
1181                                 done = 1;
1182                                 unlock_page(page);
1183                                 break;
1184                         }
1185                         if (page_has_buffers(page)) {
1186                                 bh = head = page_buffers(page);
1187                                 do {
1188                                         if (!buffer_delay(bh) &&
1189                                             !buffer_unwritten(bh))
1190                                                 done = 1;
1191                                         bh = bh->b_this_page;
1192                                 } while (!done && (bh != head));
1193                         }
1194                         unlock_page(page);
1195                         if (done)
1196                                 break;
1197                         idx++;
1198                         num++;
1199                         if (num >= max_pages)
1200                                 break;
1201                 }
1202                 pagevec_release(&pvec);
1203         }
1204         return num;
1205 }
1206
1207 /*
1208  * The ext4_get_blocks() function tries to look up the requested blocks,
1209  * and returns if the blocks are already mapped.
1210  *
1211  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1212  * and store the allocated blocks in the result buffer head and mark it
1213  * mapped.
1214  *
1215  * If file type is extents based, it will call ext4_ext_get_blocks(),
1216  * Otherwise, call with ext4_ind_get_blocks() to handle indirect mapping
1217  * based files
1218  *
1219  * On success, it returns the number of blocks being mapped or allocate.
1220  * if create==0 and the blocks are pre-allocated and uninitialized block,
1221  * the result buffer head is unmapped. If the create ==1, it will make sure
1222  * the buffer head is mapped.
1223  *
1224  * It returns 0 if plain look up failed (blocks have not been allocated), in
1225  * that casem, buffer head is unmapped
1226  *
1227  * It returns the error in case of allocation failure.
1228  */
1229 int ext4_get_blocks(handle_t *handle, struct inode *inode, sector_t block,
1230                     unsigned int max_blocks, struct buffer_head *bh,
1231                     int flags)
1232 {
1233         int retval;
1234
1235         clear_buffer_mapped(bh);
1236         clear_buffer_unwritten(bh);
1237
1238         ext_debug("ext4_get_blocks(): inode %lu, flag %d, max_blocks %u,"
1239                   "logical block %lu\n", inode->i_ino, flags, max_blocks,
1240                   (unsigned long)block);
1241         /*
1242          * Try to see if we can get the block without requesting a new
1243          * file system block.
1244          */
1245         down_read((&EXT4_I(inode)->i_data_sem));
1246         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1247                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1248                                 bh, 0);
1249         } else {
1250                 retval = ext4_ind_get_blocks(handle, inode, block, max_blocks,
1251                                              bh, 0);
1252         }
1253         up_read((&EXT4_I(inode)->i_data_sem));
1254
1255         if (retval > 0 && buffer_mapped(bh)) {
1256                 int ret = check_block_validity(inode, "file system corruption",
1257                                                block, bh->b_blocknr, retval);
1258                 if (ret != 0)
1259                         return ret;
1260         }
1261
1262         /* If it is only a block(s) look up */
1263         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1264                 return retval;
1265
1266         /*
1267          * Returns if the blocks have already allocated
1268          *
1269          * Note that if blocks have been preallocated
1270          * ext4_ext_get_block() returns th create = 0
1271          * with buffer head unmapped.
1272          */
1273         if (retval > 0 && buffer_mapped(bh))
1274                 return retval;
1275
1276         /*
1277          * When we call get_blocks without the create flag, the
1278          * BH_Unwritten flag could have gotten set if the blocks
1279          * requested were part of a uninitialized extent.  We need to
1280          * clear this flag now that we are committed to convert all or
1281          * part of the uninitialized extent to be an initialized
1282          * extent.  This is because we need to avoid the combination
1283          * of BH_Unwritten and BH_Mapped flags being simultaneously
1284          * set on the buffer_head.
1285          */
1286         clear_buffer_unwritten(bh);
1287
1288         /*
1289          * New blocks allocate and/or writing to uninitialized extent
1290          * will possibly result in updating i_data, so we take
1291          * the write lock of i_data_sem, and call get_blocks()
1292          * with create == 1 flag.
1293          */
1294         down_write((&EXT4_I(inode)->i_data_sem));
1295
1296         /*
1297          * if the caller is from delayed allocation writeout path
1298          * we have already reserved fs blocks for allocation
1299          * let the underlying get_block() function know to
1300          * avoid double accounting
1301          */
1302         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1303                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1304         /*
1305          * We need to check for EXT4 here because migrate
1306          * could have changed the inode type in between
1307          */
1308         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1309                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1310                                               bh, flags);
1311         } else {
1312                 retval = ext4_ind_get_blocks(handle, inode, block,
1313                                              max_blocks, bh, flags);
1314
1315                 if (retval > 0 && buffer_new(bh)) {
1316                         /*
1317                          * We allocated new blocks which will result in
1318                          * i_data's format changing.  Force the migrate
1319                          * to fail by clearing migrate flags
1320                          */
1321                         EXT4_I(inode)->i_state &= ~EXT4_STATE_EXT_MIGRATE;
1322                 }
1323         }
1324
1325         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1326                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1327
1328         /*
1329          * Update reserved blocks/metadata blocks after successful
1330          * block allocation which had been deferred till now.
1331          */
1332         if ((retval > 0) && (flags & EXT4_GET_BLOCKS_UPDATE_RESERVE_SPACE))
1333                 ext4_da_update_reserve_space(inode, retval);
1334
1335         up_write((&EXT4_I(inode)->i_data_sem));
1336         if (retval > 0 && buffer_mapped(bh)) {
1337                 int ret = check_block_validity(inode, "file system "
1338                                                "corruption after allocation",
1339                                                block, bh->b_blocknr, retval);
1340                 if (ret != 0)
1341                         return ret;
1342         }
1343         return retval;
1344 }
1345
1346 /* Maximum number of blocks we map for direct IO at once. */
1347 #define DIO_MAX_BLOCKS 4096
1348
1349 int ext4_get_block(struct inode *inode, sector_t iblock,
1350                    struct buffer_head *bh_result, int create)
1351 {
1352         handle_t *handle = ext4_journal_current_handle();
1353         int ret = 0, started = 0;
1354         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1355         int dio_credits;
1356
1357         if (create && !handle) {
1358                 /* Direct IO write... */
1359                 if (max_blocks > DIO_MAX_BLOCKS)
1360                         max_blocks = DIO_MAX_BLOCKS;
1361                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
1362                 handle = ext4_journal_start(inode, dio_credits);
1363                 if (IS_ERR(handle)) {
1364                         ret = PTR_ERR(handle);
1365                         goto out;
1366                 }
1367                 started = 1;
1368         }
1369
1370         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
1371                               create ? EXT4_GET_BLOCKS_CREATE : 0);
1372         if (ret > 0) {
1373                 bh_result->b_size = (ret << inode->i_blkbits);
1374                 ret = 0;
1375         }
1376         if (started)
1377                 ext4_journal_stop(handle);
1378 out:
1379         return ret;
1380 }
1381
1382 /*
1383  * `handle' can be NULL if create is zero
1384  */
1385 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1386                                 ext4_lblk_t block, int create, int *errp)
1387 {
1388         struct buffer_head dummy;
1389         int fatal = 0, err;
1390         int flags = 0;
1391
1392         J_ASSERT(handle != NULL || create == 0);
1393
1394         dummy.b_state = 0;
1395         dummy.b_blocknr = -1000;
1396         buffer_trace_init(&dummy.b_history);
1397         if (create)
1398                 flags |= EXT4_GET_BLOCKS_CREATE;
1399         err = ext4_get_blocks(handle, inode, block, 1, &dummy, flags);
1400         /*
1401          * ext4_get_blocks() returns number of blocks mapped. 0 in
1402          * case of a HOLE.
1403          */
1404         if (err > 0) {
1405                 if (err > 1)
1406                         WARN_ON(1);
1407                 err = 0;
1408         }
1409         *errp = err;
1410         if (!err && buffer_mapped(&dummy)) {
1411                 struct buffer_head *bh;
1412                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1413                 if (!bh) {
1414                         *errp = -EIO;
1415                         goto err;
1416                 }
1417                 if (buffer_new(&dummy)) {
1418                         J_ASSERT(create != 0);
1419                         J_ASSERT(handle != NULL);
1420
1421                         /*
1422                          * Now that we do not always journal data, we should
1423                          * keep in mind whether this should always journal the
1424                          * new buffer as metadata.  For now, regular file
1425                          * writes use ext4_get_block instead, so it's not a
1426                          * problem.
1427                          */
1428                         lock_buffer(bh);
1429                         BUFFER_TRACE(bh, "call get_create_access");
1430                         fatal = ext4_journal_get_create_access(handle, bh);
1431                         if (!fatal && !buffer_uptodate(bh)) {
1432                                 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1433                                 set_buffer_uptodate(bh);
1434                         }
1435                         unlock_buffer(bh);
1436                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1437                         err = ext4_handle_dirty_metadata(handle, inode, bh);
1438                         if (!fatal)
1439                                 fatal = err;
1440                 } else {
1441                         BUFFER_TRACE(bh, "not a new buffer");
1442                 }
1443                 if (fatal) {
1444                         *errp = fatal;
1445                         brelse(bh);
1446                         bh = NULL;
1447                 }
1448                 return bh;
1449         }
1450 err:
1451         return NULL;
1452 }
1453
1454 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1455                                ext4_lblk_t block, int create, int *err)
1456 {
1457         struct buffer_head *bh;
1458
1459         bh = ext4_getblk(handle, inode, block, create, err);
1460         if (!bh)
1461                 return bh;
1462         if (buffer_uptodate(bh))
1463                 return bh;
1464         ll_rw_block(READ_META, 1, &bh);
1465         wait_on_buffer(bh);
1466         if (buffer_uptodate(bh))
1467                 return bh;
1468         put_bh(bh);
1469         *err = -EIO;
1470         return NULL;
1471 }
1472
1473 static int walk_page_buffers(handle_t *handle,
1474                              struct buffer_head *head,
1475                              unsigned from,
1476                              unsigned to,
1477                              int *partial,
1478                              int (*fn)(handle_t *handle,
1479                                        struct buffer_head *bh))
1480 {
1481         struct buffer_head *bh;
1482         unsigned block_start, block_end;
1483         unsigned blocksize = head->b_size;
1484         int err, ret = 0;
1485         struct buffer_head *next;
1486
1487         for (bh = head, block_start = 0;
1488              ret == 0 && (bh != head || !block_start);
1489              block_start = block_end, bh = next) {
1490                 next = bh->b_this_page;
1491                 block_end = block_start + blocksize;
1492                 if (block_end <= from || block_start >= to) {
1493                         if (partial && !buffer_uptodate(bh))
1494                                 *partial = 1;
1495                         continue;
1496                 }
1497                 err = (*fn)(handle, bh);
1498                 if (!ret)
1499                         ret = err;
1500         }
1501         return ret;
1502 }
1503
1504 /*
1505  * To preserve ordering, it is essential that the hole instantiation and
1506  * the data write be encapsulated in a single transaction.  We cannot
1507  * close off a transaction and start a new one between the ext4_get_block()
1508  * and the commit_write().  So doing the jbd2_journal_start at the start of
1509  * prepare_write() is the right place.
1510  *
1511  * Also, this function can nest inside ext4_writepage() ->
1512  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1513  * has generated enough buffer credits to do the whole page.  So we won't
1514  * block on the journal in that case, which is good, because the caller may
1515  * be PF_MEMALLOC.
1516  *
1517  * By accident, ext4 can be reentered when a transaction is open via
1518  * quota file writes.  If we were to commit the transaction while thus
1519  * reentered, there can be a deadlock - we would be holding a quota
1520  * lock, and the commit would never complete if another thread had a
1521  * transaction open and was blocking on the quota lock - a ranking
1522  * violation.
1523  *
1524  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1525  * will _not_ run commit under these circumstances because handle->h_ref
1526  * is elevated.  We'll still have enough credits for the tiny quotafile
1527  * write.
1528  */
1529 static int do_journal_get_write_access(handle_t *handle,
1530                                        struct buffer_head *bh)
1531 {
1532         if (!buffer_mapped(bh) || buffer_freed(bh))
1533                 return 0;
1534         return ext4_journal_get_write_access(handle, bh);
1535 }
1536
1537 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1538                             loff_t pos, unsigned len, unsigned flags,
1539                             struct page **pagep, void **fsdata)
1540 {
1541         struct inode *inode = mapping->host;
1542         int ret, needed_blocks;
1543         handle_t *handle;
1544         int retries = 0;
1545         struct page *page;
1546         pgoff_t index;
1547         unsigned from, to;
1548
1549         trace_ext4_write_begin(inode, pos, len, flags);
1550         /*
1551          * Reserve one block more for addition to orphan list in case
1552          * we allocate blocks but write fails for some reason
1553          */
1554         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1555         index = pos >> PAGE_CACHE_SHIFT;
1556         from = pos & (PAGE_CACHE_SIZE - 1);
1557         to = from + len;
1558
1559 retry:
1560         handle = ext4_journal_start(inode, needed_blocks);
1561         if (IS_ERR(handle)) {
1562                 ret = PTR_ERR(handle);
1563                 goto out;
1564         }
1565
1566         /* We cannot recurse into the filesystem as the transaction is already
1567          * started */
1568         flags |= AOP_FLAG_NOFS;
1569
1570         page = grab_cache_page_write_begin(mapping, index, flags);
1571         if (!page) {
1572                 ext4_journal_stop(handle);
1573                 ret = -ENOMEM;
1574                 goto out;
1575         }
1576         *pagep = page;
1577
1578         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1579                                 ext4_get_block);
1580
1581         if (!ret && ext4_should_journal_data(inode)) {
1582                 ret = walk_page_buffers(handle, page_buffers(page),
1583                                 from, to, NULL, do_journal_get_write_access);
1584         }
1585
1586         if (ret) {
1587                 unlock_page(page);
1588                 page_cache_release(page);
1589                 /*
1590                  * block_write_begin may have instantiated a few blocks
1591                  * outside i_size.  Trim these off again. Don't need
1592                  * i_size_read because we hold i_mutex.
1593                  *
1594                  * Add inode to orphan list in case we crash before
1595                  * truncate finishes
1596                  */
1597                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1598                         ext4_orphan_add(handle, inode);
1599
1600                 ext4_journal_stop(handle);
1601                 if (pos + len > inode->i_size) {
1602                         ext4_truncate(inode);
1603                         /*
1604                          * If truncate failed early the inode might
1605                          * still be on the orphan list; we need to
1606                          * make sure the inode is removed from the
1607                          * orphan list in that case.
1608                          */
1609                         if (inode->i_nlink)
1610                                 ext4_orphan_del(NULL, inode);
1611                 }
1612         }
1613
1614         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1615                 goto retry;
1616 out:
1617         return ret;
1618 }
1619
1620 /* For write_end() in data=journal mode */
1621 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1622 {
1623         if (!buffer_mapped(bh) || buffer_freed(bh))
1624                 return 0;
1625         set_buffer_uptodate(bh);
1626         return ext4_handle_dirty_metadata(handle, NULL, bh);
1627 }
1628
1629 static int ext4_generic_write_end(struct file *file,
1630                                   struct address_space *mapping,
1631                                   loff_t pos, unsigned len, unsigned copied,
1632                                   struct page *page, void *fsdata)
1633 {
1634         int i_size_changed = 0;
1635         struct inode *inode = mapping->host;
1636         handle_t *handle = ext4_journal_current_handle();
1637
1638         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1639
1640         /*
1641          * No need to use i_size_read() here, the i_size
1642          * cannot change under us because we hold i_mutex.
1643          *
1644          * But it's important to update i_size while still holding page lock:
1645          * page writeout could otherwise come in and zero beyond i_size.
1646          */
1647         if (pos + copied > inode->i_size) {
1648                 i_size_write(inode, pos + copied);
1649                 i_size_changed = 1;
1650         }
1651
1652         if (pos + copied >  EXT4_I(inode)->i_disksize) {
1653                 /* We need to mark inode dirty even if
1654                  * new_i_size is less that inode->i_size
1655                  * bu greater than i_disksize.(hint delalloc)
1656                  */
1657                 ext4_update_i_disksize(inode, (pos + copied));
1658                 i_size_changed = 1;
1659         }
1660         unlock_page(page);
1661         page_cache_release(page);
1662
1663         /*
1664          * Don't mark the inode dirty under page lock. First, it unnecessarily
1665          * makes the holding time of page lock longer. Second, it forces lock
1666          * ordering of page lock and transaction start for journaling
1667          * filesystems.
1668          */
1669         if (i_size_changed)
1670                 ext4_mark_inode_dirty(handle, inode);
1671
1672         return copied;
1673 }
1674
1675 /*
1676  * We need to pick up the new inode size which generic_commit_write gave us
1677  * `file' can be NULL - eg, when called from page_symlink().
1678  *
1679  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1680  * buffers are managed internally.
1681  */
1682 static int ext4_ordered_write_end(struct file *file,
1683                                   struct address_space *mapping,
1684                                   loff_t pos, unsigned len, unsigned copied,
1685                                   struct page *page, void *fsdata)
1686 {
1687         handle_t *handle = ext4_journal_current_handle();
1688         struct inode *inode = mapping->host;
1689         int ret = 0, ret2;
1690
1691         trace_ext4_ordered_write_end(inode, pos, len, copied);
1692         ret = ext4_jbd2_file_inode(handle, inode);
1693
1694         if (ret == 0) {
1695                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1696                                                         page, fsdata);
1697                 copied = ret2;
1698                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1699                         /* if we have allocated more blocks and copied
1700                          * less. We will have blocks allocated outside
1701                          * inode->i_size. So truncate them
1702                          */
1703                         ext4_orphan_add(handle, inode);
1704                 if (ret2 < 0)
1705                         ret = ret2;
1706         }
1707         ret2 = ext4_journal_stop(handle);
1708         if (!ret)
1709                 ret = ret2;
1710
1711         if (pos + len > inode->i_size) {
1712                 ext4_truncate(inode);
1713                 /*
1714                  * If truncate failed early the inode might still be
1715                  * on the orphan list; we need to make sure the inode
1716                  * is removed from the orphan list in that case.
1717                  */
1718                 if (inode->i_nlink)
1719                         ext4_orphan_del(NULL, inode);
1720         }
1721
1722
1723         return ret ? ret : copied;
1724 }
1725
1726 static int ext4_writeback_write_end(struct file *file,
1727                                     struct address_space *mapping,
1728                                     loff_t pos, unsigned len, unsigned copied,
1729                                     struct page *page, void *fsdata)
1730 {
1731         handle_t *handle = ext4_journal_current_handle();
1732         struct inode *inode = mapping->host;
1733         int ret = 0, ret2;
1734
1735         trace_ext4_writeback_write_end(inode, pos, len, copied);
1736         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1737                                                         page, fsdata);
1738         copied = ret2;
1739         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1740                 /* if we have allocated more blocks and copied
1741                  * less. We will have blocks allocated outside
1742                  * inode->i_size. So truncate them
1743                  */
1744                 ext4_orphan_add(handle, inode);
1745
1746         if (ret2 < 0)
1747                 ret = ret2;
1748
1749         ret2 = ext4_journal_stop(handle);
1750         if (!ret)
1751                 ret = ret2;
1752
1753         if (pos + len > inode->i_size) {
1754                 ext4_truncate(inode);
1755                 /*
1756                  * If truncate failed early the inode might still be
1757                  * on the orphan list; we need to make sure the inode
1758                  * is removed from the orphan list in that case.
1759                  */
1760                 if (inode->i_nlink)
1761                         ext4_orphan_del(NULL, inode);
1762         }
1763
1764         return ret ? ret : copied;
1765 }
1766
1767 static int ext4_journalled_write_end(struct file *file,
1768                                      struct address_space *mapping,
1769                                      loff_t pos, unsigned len, unsigned copied,
1770                                      struct page *page, void *fsdata)
1771 {
1772         handle_t *handle = ext4_journal_current_handle();
1773         struct inode *inode = mapping->host;
1774         int ret = 0, ret2;
1775         int partial = 0;
1776         unsigned from, to;
1777         loff_t new_i_size;
1778
1779         trace_ext4_journalled_write_end(inode, pos, len, copied);
1780         from = pos & (PAGE_CACHE_SIZE - 1);
1781         to = from + len;
1782
1783         if (copied < len) {
1784                 if (!PageUptodate(page))
1785                         copied = 0;
1786                 page_zero_new_buffers(page, from+copied, to);
1787         }
1788
1789         ret = walk_page_buffers(handle, page_buffers(page), from,
1790                                 to, &partial, write_end_fn);
1791         if (!partial)
1792                 SetPageUptodate(page);
1793         new_i_size = pos + copied;
1794         if (new_i_size > inode->i_size)
1795                 i_size_write(inode, pos+copied);
1796         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
1797         if (new_i_size > EXT4_I(inode)->i_disksize) {
1798                 ext4_update_i_disksize(inode, new_i_size);
1799                 ret2 = ext4_mark_inode_dirty(handle, inode);
1800                 if (!ret)
1801                         ret = ret2;
1802         }
1803
1804         unlock_page(page);
1805         page_cache_release(page);
1806         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1807                 /* if we have allocated more blocks and copied
1808                  * less. We will have blocks allocated outside
1809                  * inode->i_size. So truncate them
1810                  */
1811                 ext4_orphan_add(handle, inode);
1812
1813         ret2 = ext4_journal_stop(handle);
1814         if (!ret)
1815                 ret = ret2;
1816         if (pos + len > inode->i_size) {
1817                 ext4_truncate(inode);
1818                 /*
1819                  * If truncate failed early the inode might still be
1820                  * on the orphan list; we need to make sure the inode
1821                  * is removed from the orphan list in that case.
1822                  */
1823                 if (inode->i_nlink)
1824                         ext4_orphan_del(NULL, inode);
1825         }
1826
1827         return ret ? ret : copied;
1828 }
1829
1830 static int ext4_da_reserve_space(struct inode *inode, int nrblocks)
1831 {
1832         int retries = 0;
1833         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1834         unsigned long md_needed, mdblocks, total = 0;
1835
1836         /*
1837          * recalculate the amount of metadata blocks to reserve
1838          * in order to allocate nrblocks
1839          * worse case is one extent per block
1840          */
1841 repeat:
1842         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1843         total = EXT4_I(inode)->i_reserved_data_blocks + nrblocks;
1844         mdblocks = ext4_calc_metadata_amount(inode, total);
1845         BUG_ON(mdblocks < EXT4_I(inode)->i_reserved_meta_blocks);
1846
1847         md_needed = mdblocks - EXT4_I(inode)->i_reserved_meta_blocks;
1848         total = md_needed + nrblocks;
1849
1850         /*
1851          * Make quota reservation here to prevent quota overflow
1852          * later. Real quota accounting is done at pages writeout
1853          * time.
1854          */
1855         if (vfs_dq_reserve_block(inode, total)) {
1856                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1857                 return -EDQUOT;
1858         }
1859
1860         if (ext4_claim_free_blocks(sbi, total)) {
1861                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1862                 vfs_dq_release_reservation_block(inode, total);
1863                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1864                         yield();
1865                         goto repeat;
1866                 }
1867                 return -ENOSPC;
1868         }
1869         EXT4_I(inode)->i_reserved_data_blocks += nrblocks;
1870         EXT4_I(inode)->i_reserved_meta_blocks = mdblocks;
1871
1872         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1873         return 0;       /* success */
1874 }
1875
1876 static void ext4_da_release_space(struct inode *inode, int to_free)
1877 {
1878         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1879         int total, mdb, mdb_free, release;
1880
1881         if (!to_free)
1882                 return;         /* Nothing to release, exit */
1883
1884         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1885
1886         if (!EXT4_I(inode)->i_reserved_data_blocks) {
1887                 /*
1888                  * if there is no reserved blocks, but we try to free some
1889                  * then the counter is messed up somewhere.
1890                  * but since this function is called from invalidate
1891                  * page, it's harmless to return without any action
1892                  */
1893                 printk(KERN_INFO "ext4 delalloc try to release %d reserved "
1894                             "blocks for inode %lu, but there is no reserved "
1895                             "data blocks\n", to_free, inode->i_ino);
1896                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1897                 return;
1898         }
1899
1900         /* recalculate the number of metablocks still need to be reserved */
1901         total = EXT4_I(inode)->i_reserved_data_blocks - to_free;
1902         mdb = ext4_calc_metadata_amount(inode, total);
1903
1904         /* figure out how many metablocks to release */
1905         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1906         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1907
1908         release = to_free + mdb_free;
1909
1910         /* update fs dirty blocks counter for truncate case */
1911         percpu_counter_sub(&sbi->s_dirtyblocks_counter, release);
1912
1913         /* update per-inode reservations */
1914         BUG_ON(to_free > EXT4_I(inode)->i_reserved_data_blocks);
1915         EXT4_I(inode)->i_reserved_data_blocks -= to_free;
1916
1917         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1918         EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1919         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1920
1921         vfs_dq_release_reservation_block(inode, release);
1922 }
1923
1924 static void ext4_da_page_release_reservation(struct page *page,
1925                                              unsigned long offset)
1926 {
1927         int to_release = 0;
1928         struct buffer_head *head, *bh;
1929         unsigned int curr_off = 0;
1930
1931         head = page_buffers(page);
1932         bh = head;
1933         do {
1934                 unsigned int next_off = curr_off + bh->b_size;
1935
1936                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1937                         to_release++;
1938                         clear_buffer_delay(bh);
1939                 }
1940                 curr_off = next_off;
1941         } while ((bh = bh->b_this_page) != head);
1942         ext4_da_release_space(page->mapping->host, to_release);
1943 }
1944
1945 /*
1946  * Delayed allocation stuff
1947  */
1948
1949 /*
1950  * mpage_da_submit_io - walks through extent of pages and try to write
1951  * them with writepage() call back
1952  *
1953  * @mpd->inode: inode
1954  * @mpd->first_page: first page of the extent
1955  * @mpd->next_page: page after the last page of the extent
1956  *
1957  * By the time mpage_da_submit_io() is called we expect all blocks
1958  * to be allocated. this may be wrong if allocation failed.
1959  *
1960  * As pages are already locked by write_cache_pages(), we can't use it
1961  */
1962 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1963 {
1964         long pages_skipped;
1965         struct pagevec pvec;
1966         unsigned long index, end;
1967         int ret = 0, err, nr_pages, i;
1968         struct inode *inode = mpd->inode;
1969         struct address_space *mapping = inode->i_mapping;
1970
1971         BUG_ON(mpd->next_page <= mpd->first_page);
1972         /*
1973          * We need to start from the first_page to the next_page - 1
1974          * to make sure we also write the mapped dirty buffer_heads.
1975          * If we look at mpd->b_blocknr we would only be looking
1976          * at the currently mapped buffer_heads.
1977          */
1978         index = mpd->first_page;
1979         end = mpd->next_page - 1;
1980
1981         pagevec_init(&pvec, 0);
1982         while (index <= end) {
1983                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1984                 if (nr_pages == 0)
1985                         break;
1986                 for (i = 0; i < nr_pages; i++) {
1987                         struct page *page = pvec.pages[i];
1988
1989                         index = page->index;
1990                         if (index > end)
1991                                 break;
1992                         index++;
1993
1994                         BUG_ON(!PageLocked(page));
1995                         BUG_ON(PageWriteback(page));
1996
1997                         pages_skipped = mpd->wbc->pages_skipped;
1998                         err = mapping->a_ops->writepage(page, mpd->wbc);
1999                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
2000                                 /*
2001                                  * have successfully written the page
2002                                  * without skipping the same
2003                                  */
2004                                 mpd->pages_written++;
2005                         /*
2006                          * In error case, we have to continue because
2007                          * remaining pages are still locked
2008                          * XXX: unlock and re-dirty them?
2009                          */
2010                         if (ret == 0)
2011                                 ret = err;
2012                 }
2013                 pagevec_release(&pvec);
2014         }
2015         return ret;
2016 }
2017
2018 /*
2019  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
2020  *
2021  * @mpd->inode - inode to walk through
2022  * @exbh->b_blocknr - first block on a disk
2023  * @exbh->b_size - amount of space in bytes
2024  * @logical - first logical block to start assignment with
2025  *
2026  * the function goes through all passed space and put actual disk
2027  * block numbers into buffer heads, dropping BH_Delay and BH_Unwritten
2028  */
2029 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
2030                                  struct buffer_head *exbh)
2031 {
2032         struct inode *inode = mpd->inode;
2033         struct address_space *mapping = inode->i_mapping;
2034         int blocks = exbh->b_size >> inode->i_blkbits;
2035         sector_t pblock = exbh->b_blocknr, cur_logical;
2036         struct buffer_head *head, *bh;
2037         pgoff_t index, end;
2038         struct pagevec pvec;
2039         int nr_pages, i;
2040
2041         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2042         end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2043         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2044
2045         pagevec_init(&pvec, 0);
2046
2047         while (index <= end) {
2048                 /* XXX: optimize tail */
2049                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2050                 if (nr_pages == 0)
2051                         break;
2052                 for (i = 0; i < nr_pages; i++) {
2053                         struct page *page = pvec.pages[i];
2054
2055                         index = page->index;
2056                         if (index > end)
2057                                 break;
2058                         index++;
2059
2060                         BUG_ON(!PageLocked(page));
2061                         BUG_ON(PageWriteback(page));
2062                         BUG_ON(!page_has_buffers(page));
2063
2064                         bh = page_buffers(page);
2065                         head = bh;
2066
2067                         /* skip blocks out of the range */
2068                         do {
2069                                 if (cur_logical >= logical)
2070                                         break;
2071                                 cur_logical++;
2072                         } while ((bh = bh->b_this_page) != head);
2073
2074                         do {
2075                                 if (cur_logical >= logical + blocks)
2076                                         break;
2077
2078                                 if (buffer_delay(bh) ||
2079                                                 buffer_unwritten(bh)) {
2080
2081                                         BUG_ON(bh->b_bdev != inode->i_sb->s_bdev);
2082
2083                                         if (buffer_delay(bh)) {
2084                                                 clear_buffer_delay(bh);
2085                                                 bh->b_blocknr = pblock;
2086                                         } else {
2087                                                 /*
2088                                                  * unwritten already should have
2089                                                  * blocknr assigned. Verify that
2090                                                  */
2091                                                 clear_buffer_unwritten(bh);
2092                                                 BUG_ON(bh->b_blocknr != pblock);
2093                                         }
2094
2095                                 } else if (buffer_mapped(bh))
2096                                         BUG_ON(bh->b_blocknr != pblock);
2097
2098                                 cur_logical++;
2099                                 pblock++;
2100                         } while ((bh = bh->b_this_page) != head);
2101                 }
2102                 pagevec_release(&pvec);
2103         }
2104 }
2105
2106
2107 /*
2108  * __unmap_underlying_blocks - just a helper function to unmap
2109  * set of blocks described by @bh
2110  */
2111 static inline void __unmap_underlying_blocks(struct inode *inode,
2112                                              struct buffer_head *bh)
2113 {
2114         struct block_device *bdev = inode->i_sb->s_bdev;
2115         int blocks, i;
2116
2117         blocks = bh->b_size >> inode->i_blkbits;
2118         for (i = 0; i < blocks; i++)
2119                 unmap_underlying_metadata(bdev, bh->b_blocknr + i);
2120 }
2121
2122 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2123                                         sector_t logical, long blk_cnt)
2124 {
2125         int nr_pages, i;
2126         pgoff_t index, end;
2127         struct pagevec pvec;
2128         struct inode *inode = mpd->inode;
2129         struct address_space *mapping = inode->i_mapping;
2130
2131         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2132         end   = (logical + blk_cnt - 1) >>
2133                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2134         while (index <= end) {
2135                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2136                 if (nr_pages == 0)
2137                         break;
2138                 for (i = 0; i < nr_pages; i++) {
2139                         struct page *page = pvec.pages[i];
2140                         index = page->index;
2141                         if (index > end)
2142                                 break;
2143                         index++;
2144
2145                         BUG_ON(!PageLocked(page));
2146                         BUG_ON(PageWriteback(page));
2147                         block_invalidatepage(page, 0);
2148                         ClearPageUptodate(page);
2149                         unlock_page(page);
2150                 }
2151         }
2152         return;
2153 }
2154
2155 static void ext4_print_free_blocks(struct inode *inode)
2156 {
2157         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2158         printk(KERN_CRIT "Total free blocks count %lld\n",
2159                ext4_count_free_blocks(inode->i_sb));
2160         printk(KERN_CRIT "Free/Dirty block details\n");
2161         printk(KERN_CRIT "free_blocks=%lld\n",
2162                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2163         printk(KERN_CRIT "dirty_blocks=%lld\n",
2164                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2165         printk(KERN_CRIT "Block reservation details\n");
2166         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2167                EXT4_I(inode)->i_reserved_data_blocks);
2168         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2169                EXT4_I(inode)->i_reserved_meta_blocks);
2170         return;
2171 }
2172
2173 /*
2174  * mpage_da_map_blocks - go through given space
2175  *
2176  * @mpd - bh describing space
2177  *
2178  * The function skips space we know is already mapped to disk blocks.
2179  *
2180  */
2181 static int mpage_da_map_blocks(struct mpage_da_data *mpd)
2182 {
2183         int err, blks, get_blocks_flags;
2184         struct buffer_head new;
2185         sector_t next = mpd->b_blocknr;
2186         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2187         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2188         handle_t *handle = NULL;
2189
2190         /*
2191          * We consider only non-mapped and non-allocated blocks
2192          */
2193         if ((mpd->b_state  & (1 << BH_Mapped)) &&
2194                 !(mpd->b_state & (1 << BH_Delay)) &&
2195                 !(mpd->b_state & (1 << BH_Unwritten)))
2196                 return 0;
2197
2198         /*
2199          * If we didn't accumulate anything to write simply return
2200          */
2201         if (!mpd->b_size)
2202                 return 0;
2203
2204         handle = ext4_journal_current_handle();
2205         BUG_ON(!handle);
2206
2207         /*
2208          * Call ext4_get_blocks() to allocate any delayed allocation
2209          * blocks, or to convert an uninitialized extent to be
2210          * initialized (in the case where we have written into
2211          * one or more preallocated blocks).
2212          *
2213          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2214          * indicate that we are on the delayed allocation path.  This
2215          * affects functions in many different parts of the allocation
2216          * call path.  This flag exists primarily because we don't
2217          * want to change *many* call functions, so ext4_get_blocks()
2218          * will set the magic i_delalloc_reserved_flag once the
2219          * inode's allocation semaphore is taken.
2220          *
2221          * If the blocks in questions were delalloc blocks, set
2222          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2223          * variables are updated after the blocks have been allocated.
2224          */
2225         new.b_state = 0;
2226         get_blocks_flags = (EXT4_GET_BLOCKS_CREATE |
2227                             EXT4_GET_BLOCKS_DELALLOC_RESERVE);
2228         if (mpd->b_state & (1 << BH_Delay))
2229                 get_blocks_flags |= EXT4_GET_BLOCKS_UPDATE_RESERVE_SPACE;
2230         blks = ext4_get_blocks(handle, mpd->inode, next, max_blocks,
2231                                &new, get_blocks_flags);
2232         if (blks < 0) {
2233                 err = blks;
2234                 /*
2235                  * If get block returns with error we simply
2236                  * return. Later writepage will redirty the page and
2237                  * writepages will find the dirty page again
2238                  */
2239                 if (err == -EAGAIN)
2240                         return 0;
2241
2242                 if (err == -ENOSPC &&
2243                     ext4_count_free_blocks(mpd->inode->i_sb)) {
2244                         mpd->retval = err;
2245                         return 0;
2246                 }
2247
2248                 /*
2249                  * get block failure will cause us to loop in
2250                  * writepages, because a_ops->writepage won't be able
2251                  * to make progress. The page will be redirtied by
2252                  * writepage and writepages will again try to write
2253                  * the same.
2254                  */
2255                 ext4_msg(mpd->inode->i_sb, KERN_CRIT,
2256                          "delayed block allocation failed for inode %lu at "
2257                          "logical offset %llu with max blocks %zd with "
2258                          "error %d\n", mpd->inode->i_ino,
2259                          (unsigned long long) next,
2260                          mpd->b_size >> mpd->inode->i_blkbits, err);
2261                 printk(KERN_CRIT "This should not happen!!  "
2262                        "Data will be lost\n");
2263                 if (err == -ENOSPC) {
2264                         ext4_print_free_blocks(mpd->inode);
2265                 }
2266                 /* invalidate all the pages */
2267                 ext4_da_block_invalidatepages(mpd, next,
2268                                 mpd->b_size >> mpd->inode->i_blkbits);
2269                 return err;
2270         }
2271         BUG_ON(blks == 0);
2272
2273         new.b_size = (blks << mpd->inode->i_blkbits);
2274
2275         if (buffer_new(&new))
2276                 __unmap_underlying_blocks(mpd->inode, &new);
2277
2278         /*
2279          * If blocks are delayed marked, we need to
2280          * put actual blocknr and drop delayed bit
2281          */
2282         if ((mpd->b_state & (1 << BH_Delay)) ||
2283             (mpd->b_state & (1 << BH_Unwritten)))
2284                 mpage_put_bnr_to_bhs(mpd, next, &new);
2285
2286         if (ext4_should_order_data(mpd->inode)) {
2287                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2288                 if (err)
2289                         return err;
2290         }
2291
2292         /*
2293          * Update on-disk size along with block allocation.
2294          */
2295         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2296         if (disksize > i_size_read(mpd->inode))
2297                 disksize = i_size_read(mpd->inode);
2298         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2299                 ext4_update_i_disksize(mpd->inode, disksize);
2300                 return ext4_mark_inode_dirty(handle, mpd->inode);
2301         }
2302
2303         return 0;
2304 }
2305
2306 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2307                 (1 << BH_Delay) | (1 << BH_Unwritten))
2308
2309 /*
2310  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2311  *
2312  * @mpd->lbh - extent of blocks
2313  * @logical - logical number of the block in the file
2314  * @bh - bh of the block (used to access block's state)
2315  *
2316  * the function is used to collect contig. blocks in same state
2317  */
2318 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2319                                    sector_t logical, size_t b_size,
2320                                    unsigned long b_state)
2321 {
2322         sector_t next;
2323         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2324
2325         /* check if thereserved journal credits might overflow */
2326         if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
2327                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2328                         /*
2329                          * With non-extent format we are limited by the journal
2330                          * credit available.  Total credit needed to insert
2331                          * nrblocks contiguous blocks is dependent on the
2332                          * nrblocks.  So limit nrblocks.
2333                          */
2334                         goto flush_it;
2335                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2336                                 EXT4_MAX_TRANS_DATA) {
2337                         /*
2338                          * Adding the new buffer_head would make it cross the
2339                          * allowed limit for which we have journal credit
2340                          * reserved. So limit the new bh->b_size
2341                          */
2342                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2343                                                 mpd->inode->i_blkbits;
2344                         /* we will do mpage_da_submit_io in the next loop */
2345                 }
2346         }
2347         /*
2348          * First block in the extent
2349          */
2350         if (mpd->b_size == 0) {
2351                 mpd->b_blocknr = logical;
2352                 mpd->b_size = b_size;
2353                 mpd->b_state = b_state & BH_FLAGS;
2354                 return;
2355         }
2356
2357         next = mpd->b_blocknr + nrblocks;
2358         /*
2359          * Can we merge the block to our big extent?
2360          */
2361         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2362                 mpd->b_size += b_size;
2363                 return;
2364         }
2365
2366 flush_it:
2367         /*
2368          * We couldn't merge the block to our extent, so we
2369          * need to flush current  extent and start new one
2370          */
2371         if (mpage_da_map_blocks(mpd) == 0)
2372                 mpage_da_submit_io(mpd);
2373         mpd->io_done = 1;
2374         return;
2375 }
2376
2377 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
2378 {
2379         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
2380 }
2381
2382 /*
2383  * __mpage_da_writepage - finds extent of pages and blocks
2384  *
2385  * @page: page to consider
2386  * @wbc: not used, we just follow rules
2387  * @data: context
2388  *
2389  * The function finds extents of pages and scan them for all blocks.
2390  */
2391 static int __mpage_da_writepage(struct page *page,
2392                                 struct writeback_control *wbc, void *data)
2393 {
2394         struct mpage_da_data *mpd = data;
2395         struct inode *inode = mpd->inode;
2396         struct buffer_head *bh, *head;
2397         sector_t logical;
2398
2399         if (mpd->io_done) {
2400                 /*
2401                  * Rest of the page in the page_vec
2402                  * redirty then and skip then. We will
2403                  * try to write them again after
2404                  * starting a new transaction
2405                  */
2406                 redirty_page_for_writepage(wbc, page);
2407                 unlock_page(page);
2408                 return MPAGE_DA_EXTENT_TAIL;
2409         }
2410         /*
2411          * Can we merge this page to current extent?
2412          */
2413         if (mpd->next_page != page->index) {
2414                 /*
2415                  * Nope, we can't. So, we map non-allocated blocks
2416                  * and start IO on them using writepage()
2417                  */
2418                 if (mpd->next_page != mpd->first_page) {
2419                         if (mpage_da_map_blocks(mpd) == 0)
2420                                 mpage_da_submit_io(mpd);
2421                         /*
2422                          * skip rest of the page in the page_vec
2423                          */
2424                         mpd->io_done = 1;
2425                         redirty_page_for_writepage(wbc, page);
2426                         unlock_page(page);
2427                         return MPAGE_DA_EXTENT_TAIL;
2428                 }
2429
2430                 /*
2431                  * Start next extent of pages ...
2432                  */
2433                 mpd->first_page = page->index;
2434
2435                 /*
2436                  * ... and blocks
2437                  */
2438                 mpd->b_size = 0;
2439                 mpd->b_state = 0;
2440                 mpd->b_blocknr = 0;
2441         }
2442
2443         mpd->next_page = page->index + 1;
2444         logical = (sector_t) page->index <<
2445                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2446
2447         if (!page_has_buffers(page)) {
2448                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2449                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2450                 if (mpd->io_done)
2451                         return MPAGE_DA_EXTENT_TAIL;
2452         } else {
2453                 /*
2454                  * Page with regular buffer heads, just add all dirty ones
2455                  */
2456                 head = page_buffers(page);
2457                 bh = head;
2458                 do {
2459                         BUG_ON(buffer_locked(bh));
2460                         /*
2461                          * We need to try to allocate
2462                          * unmapped blocks in the same page.
2463                          * Otherwise we won't make progress
2464                          * with the page in ext4_writepage
2465                          */
2466                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2467                                 mpage_add_bh_to_extent(mpd, logical,
2468                                                        bh->b_size,
2469                                                        bh->b_state);
2470                                 if (mpd->io_done)
2471                                         return MPAGE_DA_EXTENT_TAIL;
2472                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2473                                 /*
2474                                  * mapped dirty buffer. We need to update
2475                                  * the b_state because we look at
2476                                  * b_state in mpage_da_map_blocks. We don't
2477                                  * update b_size because if we find an
2478                                  * unmapped buffer_head later we need to
2479                                  * use the b_state flag of that buffer_head.
2480                                  */
2481                                 if (mpd->b_size == 0)
2482                                         mpd->b_state = bh->b_state & BH_FLAGS;
2483                         }
2484                         logical++;
2485                 } while ((bh = bh->b_this_page) != head);
2486         }
2487
2488         return 0;
2489 }
2490
2491 /*
2492  * This is a special get_blocks_t callback which is used by
2493  * ext4_da_write_begin().  It will either return mapped block or
2494  * reserve space for a single block.
2495  *
2496  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2497  * We also have b_blocknr = -1 and b_bdev initialized properly
2498  *
2499  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2500  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2501  * initialized properly.
2502  */
2503 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2504                                   struct buffer_head *bh_result, int create)
2505 {
2506         int ret = 0;
2507         sector_t invalid_block = ~((sector_t) 0xffff);
2508
2509         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2510                 invalid_block = ~0;
2511
2512         BUG_ON(create == 0);
2513         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2514
2515         /*
2516          * first, we need to know whether the block is allocated already
2517          * preallocated blocks are unmapped but should treated
2518          * the same as allocated blocks.
2519          */
2520         ret = ext4_get_blocks(NULL, inode, iblock, 1,  bh_result, 0);
2521         if ((ret == 0) && !buffer_delay(bh_result)) {
2522                 /* the block isn't (pre)allocated yet, let's reserve space */
2523                 /*
2524                  * XXX: __block_prepare_write() unmaps passed block,
2525                  * is it OK?
2526                  */
2527                 ret = ext4_da_reserve_space(inode, 1);
2528                 if (ret)
2529                         /* not enough space to reserve */
2530                         return ret;
2531
2532                 map_bh(bh_result, inode->i_sb, invalid_block);
2533                 set_buffer_new(bh_result);
2534                 set_buffer_delay(bh_result);
2535         } else if (ret > 0) {
2536                 bh_result->b_size = (ret << inode->i_blkbits);
2537                 if (buffer_unwritten(bh_result)) {
2538                         /* A delayed write to unwritten bh should
2539                          * be marked new and mapped.  Mapped ensures
2540                          * that we don't do get_block multiple times
2541                          * when we write to the same offset and new
2542                          * ensures that we do proper zero out for
2543                          * partial write.
2544                          */
2545                         set_buffer_new(bh_result);
2546                         set_buffer_mapped(bh_result);
2547                 }
2548                 ret = 0;
2549         }
2550
2551         return ret;
2552 }
2553
2554 /*
2555  * This function is used as a standard get_block_t calback function
2556  * when there is no desire to allocate any blocks.  It is used as a
2557  * callback function for block_prepare_write(), nobh_writepage(), and
2558  * block_write_full_page().  These functions should only try to map a
2559  * single block at a time.
2560  *
2561  * Since this function doesn't do block allocations even if the caller
2562  * requests it by passing in create=1, it is critically important that
2563  * any caller checks to make sure that any buffer heads are returned
2564  * by this function are either all already mapped or marked for
2565  * delayed allocation before calling nobh_writepage() or
2566  * block_write_full_page().  Otherwise, b_blocknr could be left
2567  * unitialized, and the page write functions will be taken by
2568  * surprise.
2569  */
2570 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2571                                    struct buffer_head *bh_result, int create)
2572 {
2573         int ret = 0;
2574         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2575
2576         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2577
2578         /*
2579          * we don't want to do block allocation in writepage
2580          * so call get_block_wrap with create = 0
2581          */
2582         ret = ext4_get_blocks(NULL, inode, iblock, max_blocks, bh_result, 0);
2583         if (ret > 0) {
2584                 bh_result->b_size = (ret << inode->i_blkbits);
2585                 ret = 0;
2586         }
2587         return ret;
2588 }
2589
2590 static int bget_one(handle_t *handle, struct buffer_head *bh)
2591 {
2592         get_bh(bh);
2593         return 0;
2594 }
2595
2596 static int bput_one(handle_t *handle, struct buffer_head *bh)
2597 {
2598         put_bh(bh);
2599         return 0;
2600 }
2601
2602 static int __ext4_journalled_writepage(struct page *page,
2603                                        struct writeback_control *wbc,
2604                                        unsigned int len)
2605 {
2606         struct address_space *mapping = page->mapping;
2607         struct inode *inode = mapping->host;
2608         struct buffer_head *page_bufs;
2609         handle_t *handle = NULL;
2610         int ret = 0;
2611         int err;
2612
2613         page_bufs = page_buffers(page);
2614         BUG_ON(!page_bufs);
2615         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2616         /* As soon as we unlock the page, it can go away, but we have
2617          * references to buffers so we are safe */
2618         unlock_page(page);
2619
2620         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2621         if (IS_ERR(handle)) {
2622                 ret = PTR_ERR(handle);
2623                 goto out;
2624         }
2625
2626         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2627                                 do_journal_get_write_access);
2628
2629         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2630                                 write_end_fn);
2631         if (ret == 0)
2632                 ret = err;
2633         err = ext4_journal_stop(handle);
2634         if (!ret)
2635                 ret = err;
2636
2637         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2638         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
2639 out:
2640         return ret;
2641 }
2642
2643 /*
2644  * Note that we don't need to start a transaction unless we're journaling data
2645  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2646  * need to file the inode to the transaction's list in ordered mode because if
2647  * we are writing back data added by write(), the inode is already there and if
2648  * we are writing back data modified via mmap(), noone guarantees in which
2649  * transaction the data will hit the disk. In case we are journaling data, we
2650  * cannot start transaction directly because transaction start ranks above page
2651  * lock so we have to do some magic.
2652  *
2653  * This function can get called via...
2654  *   - ext4_da_writepages after taking page lock (have journal handle)
2655  *   - journal_submit_inode_data_buffers (no journal handle)
2656  *   - shrink_page_list via pdflush (no journal handle)
2657  *   - grab_page_cache when doing write_begin (have journal handle)
2658  *
2659  * We don't do any block allocation in this function. If we have page with
2660  * multiple blocks we need to write those buffer_heads that are mapped. This
2661  * is important for mmaped based write. So if we do with blocksize 1K
2662  * truncate(f, 1024);
2663  * a = mmap(f, 0, 4096);
2664  * a[0] = 'a';
2665  * truncate(f, 4096);
2666  * we have in the page first buffer_head mapped via page_mkwrite call back
2667  * but other bufer_heads would be unmapped but dirty(dirty done via the
2668  * do_wp_page). So writepage should write the first block. If we modify
2669  * the mmap area beyond 1024 we will again get a page_fault and the
2670  * page_mkwrite callback will do the block allocation and mark the
2671  * buffer_heads mapped.
2672  *
2673  * We redirty the page if we have any buffer_heads that is either delay or
2674  * unwritten in the page.
2675  *
2676  * We can get recursively called as show below.
2677  *
2678  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2679  *              ext4_writepage()
2680  *
2681  * But since we don't do any block allocation we should not deadlock.
2682  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2683  */
2684 static int ext4_writepage(struct page *page,
2685                           struct writeback_control *wbc)
2686 {
2687         int ret = 0;
2688         loff_t size;
2689         unsigned int len;
2690         struct buffer_head *page_bufs;
2691         struct inode *inode = page->mapping->host;
2692
2693         trace_ext4_writepage(inode, page);
2694         size = i_size_read(inode);
2695         if (page->index == size >> PAGE_CACHE_SHIFT)
2696                 len = size & ~PAGE_CACHE_MASK;
2697         else
2698                 len = PAGE_CACHE_SIZE;
2699
2700         if (page_has_buffers(page)) {
2701                 page_bufs = page_buffers(page);
2702                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2703                                         ext4_bh_delay_or_unwritten)) {
2704                         /*
2705                          * We don't want to do  block allocation
2706                          * So redirty the page and return
2707                          * We may reach here when we do a journal commit
2708                          * via journal_submit_inode_data_buffers.
2709                          * If we don't have mapping block we just ignore
2710                          * them. We can also reach here via shrink_page_list
2711                          */
2712                         redirty_page_for_writepage(wbc, page);
2713                         unlock_page(page);
2714                         return 0;
2715                 }
2716         } else {
2717                 /*
2718                  * The test for page_has_buffers() is subtle:
2719                  * We know the page is dirty but it lost buffers. That means
2720                  * that at some moment in time after write_begin()/write_end()
2721                  * has been called all buffers have been clean and thus they
2722                  * must have been written at least once. So they are all
2723                  * mapped and we can happily proceed with mapping them
2724                  * and writing the page.
2725                  *
2726                  * Try to initialize the buffer_heads and check whether
2727                  * all are mapped and non delay. We don't want to
2728                  * do block allocation here.
2729                  */
2730                 ret = block_prepare_write(page, 0, len,
2731                                           noalloc_get_block_write);
2732                 if (!ret) {
2733                         page_bufs = page_buffers(page);
2734                         /* check whether all are mapped and non delay */
2735                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2736                                                 ext4_bh_delay_or_unwritten)) {
2737                                 redirty_page_for_writepage(wbc, page);
2738                                 unlock_page(page);
2739                                 return 0;
2740                         }
2741                 } else {
2742                         /*
2743                          * We can't do block allocation here
2744                          * so just redity the page and unlock
2745                          * and return
2746                          */
2747                         redirty_page_for_writepage(wbc, page);
2748                         unlock_page(page);
2749                         return 0;
2750                 }
2751                 /* now mark the buffer_heads as dirty and uptodate */
2752                 block_commit_write(page, 0, len);
2753         }
2754
2755         if (PageChecked(page) && ext4_should_journal_data(inode)) {
2756                 /*
2757                  * It's mmapped pagecache.  Add buffers and journal it.  There
2758                  * doesn't seem much point in redirtying the page here.
2759                  */
2760                 ClearPageChecked(page);
2761                 return __ext4_journalled_writepage(page, wbc, len);
2762         }
2763
2764         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2765                 ret = nobh_writepage(page, noalloc_get_block_write, wbc);
2766         else
2767                 ret = block_write_full_page(page, noalloc_get_block_write,
2768                                             wbc);
2769
2770         return ret;
2771 }
2772
2773 /*
2774  * This is called via ext4_da_writepages() to
2775  * calulate the total number of credits to reserve to fit
2776  * a single extent allocation into a single transaction,
2777  * ext4_da_writpeages() will loop calling this before
2778  * the block allocation.
2779  */
2780
2781 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2782 {
2783         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2784
2785         /*
2786          * With non-extent format the journal credit needed to
2787          * insert nrblocks contiguous block is dependent on
2788          * number of contiguous block. So we will limit
2789          * number of contiguous block to a sane value
2790          */
2791         if (!(inode->i_flags & EXT4_EXTENTS_FL) &&
2792             (max_blocks > EXT4_MAX_TRANS_DATA))
2793                 max_blocks = EXT4_MAX_TRANS_DATA;
2794
2795         return ext4_chunk_trans_blocks(inode, max_blocks);
2796 }
2797
2798 static int ext4_da_writepages(struct address_space *mapping,
2799                               struct writeback_control *wbc)
2800 {
2801         pgoff_t index;
2802         int range_whole = 0;
2803         handle_t *handle = NULL;
2804         struct mpage_da_data mpd;
2805         struct inode *inode = mapping->host;
2806         int no_nrwrite_index_update;
2807         int pages_written = 0;
2808         long pages_skipped;
2809         unsigned int max_pages;
2810         int range_cyclic, cycled = 1, io_done = 0;
2811         int needed_blocks, ret = 0;
2812         long desired_nr_to_write, nr_to_writebump = 0;
2813         loff_t range_start = wbc->range_start;
2814         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2815
2816         trace_ext4_da_writepages(inode, wbc);
2817
2818         /*
2819          * No pages to write? This is mainly a kludge to avoid starting
2820          * a transaction for special inodes like journal inode on last iput()
2821          * because that could violate lock ordering on umount
2822          */
2823         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2824                 return 0;
2825
2826         /*
2827          * If the filesystem has aborted, it is read-only, so return
2828          * right away instead of dumping stack traces later on that
2829          * will obscure the real source of the problem.  We test
2830          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2831          * the latter could be true if the filesystem is mounted
2832          * read-only, and in that case, ext4_da_writepages should
2833          * *never* be called, so if that ever happens, we would want
2834          * the stack trace.
2835          */
2836         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2837                 return -EROFS;
2838
2839         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2840                 range_whole = 1;
2841
2842         range_cyclic = wbc->range_cyclic;
2843         if (wbc->range_cyclic) {
2844                 index = mapping->writeback_index;
2845                 if (index)
2846                         cycled = 0;
2847                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2848                 wbc->range_end  = LLONG_MAX;
2849                 wbc->range_cyclic = 0;
2850         } else
2851                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2852
2853         /*
2854          * This works around two forms of stupidity.  The first is in
2855          * the writeback code, which caps the maximum number of pages
2856          * written to be 1024 pages.  This is wrong on multiple
2857          * levels; different architectues have a different page size,
2858          * which changes the maximum amount of data which gets
2859          * written.  Secondly, 4 megabytes is way too small.  XFS
2860          * forces this value to be 16 megabytes by multiplying
2861          * nr_to_write parameter by four, and then relies on its
2862          * allocator to allocate larger extents to make them
2863          * contiguous.  Unfortunately this brings us to the second
2864          * stupidity, which is that ext4's mballoc code only allocates
2865          * at most 2048 blocks.  So we force contiguous writes up to
2866          * the number of dirty blocks in the inode, or
2867          * sbi->max_writeback_mb_bump whichever is smaller.
2868          */
2869         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2870         if (!range_cyclic && range_whole)
2871                 desired_nr_to_write = wbc->nr_to_write * 8;
2872         else
2873                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2874                                                            max_pages);
2875         if (desired_nr_to_write > max_pages)
2876                 desired_nr_to_write = max_pages;
2877
2878         if (wbc->nr_to_write < desired_nr_to_write) {
2879                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2880                 wbc->nr_to_write = desired_nr_to_write;
2881         }
2882
2883         mpd.wbc = wbc;
2884         mpd.inode = mapping->host;
2885
2886         /*
2887          * we don't want write_cache_pages to update
2888          * nr_to_write and writeback_index
2889          */
2890         no_nrwrite_index_update = wbc->no_nrwrite_index_update;
2891         wbc->no_nrwrite_index_update = 1;
2892         pages_skipped = wbc->pages_skipped;
2893
2894 retry:
2895         while (!ret && wbc->nr_to_write > 0) {
2896
2897                 /*
2898                  * we  insert one extent at a time. So we need
2899                  * credit needed for single extent allocation.
2900                  * journalled mode is currently not supported
2901                  * by delalloc
2902                  */
2903                 BUG_ON(ext4_should_journal_data(inode));
2904                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2905
2906                 /* start a new transaction*/
2907                 handle = ext4_journal_start(inode, needed_blocks);
2908                 if (IS_ERR(handle)) {
2909                         ret = PTR_ERR(handle);
2910                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2911                                "%ld pages, ino %lu; err %d\n", __func__,
2912                                 wbc->nr_to_write, inode->i_ino, ret);
2913                         goto out_writepages;
2914                 }
2915
2916                 /*
2917                  * Now call __mpage_da_writepage to find the next
2918                  * contiguous region of logical blocks that need
2919                  * blocks to be allocated by ext4.  We don't actually
2920                  * submit the blocks for I/O here, even though
2921                  * write_cache_pages thinks it will, and will set the
2922                  * pages as clean for write before calling
2923                  * __mpage_da_writepage().
2924                  */
2925                 mpd.b_size = 0;
2926                 mpd.b_state = 0;
2927                 mpd.b_blocknr = 0;
2928                 mpd.first_page = 0;
2929                 mpd.next_page = 0;
2930                 mpd.io_done = 0;
2931                 mpd.pages_written = 0;
2932                 mpd.retval = 0;
2933                 ret = write_cache_pages(mapping, wbc, __mpage_da_writepage,
2934                                         &mpd);
2935                 /*
2936                  * If we have a contigous extent of pages and we
2937                  * haven't done the I/O yet, map the blocks and submit
2938                  * them for I/O.
2939                  */
2940                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2941                         if (mpage_da_map_blocks(&mpd) == 0)
2942                                 mpage_da_submit_io(&mpd);
2943                         mpd.io_done = 1;
2944                         ret = MPAGE_DA_EXTENT_TAIL;
2945                 }
2946                 trace_ext4_da_write_pages(inode, &mpd);
2947                 wbc->nr_to_write -= mpd.pages_written;
2948
2949                 ext4_journal_stop(handle);
2950
2951                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2952                         /* commit the transaction which would
2953                          * free blocks released in the transaction
2954                          * and try again
2955                          */
2956                         jbd2_journal_force_commit_nested(sbi->s_journal);
2957                         wbc->pages_skipped = pages_skipped;
2958                         ret = 0;
2959                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2960                         /*
2961                          * got one extent now try with
2962                          * rest of the pages
2963                          */
2964                         pages_written += mpd.pages_written;
2965                         wbc->pages_skipped = pages_skipped;
2966                         ret = 0;
2967                         io_done = 1;
2968                 } else if (wbc->nr_to_write)
2969                         /*
2970                          * There is no more writeout needed
2971                          * or we requested for a noblocking writeout
2972                          * and we found the device congested
2973                          */
2974                         break;
2975         }
2976         if (!io_done && !cycled) {
2977                 cycled = 1;
2978                 index = 0;
2979                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2980                 wbc->range_end  = mapping->writeback_index - 1;
2981                 goto retry;
2982         }
2983         if (pages_skipped != wbc->pages_skipped)
2984                 ext4_msg(inode->i_sb, KERN_CRIT,
2985                          "This should not happen leaving %s "
2986                          "with nr_to_write = %ld ret = %d\n",
2987                          __func__, wbc->nr_to_write, ret);
2988
2989         /* Update index */
2990         index += pages_written;
2991         wbc->range_cyclic = range_cyclic;
2992         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2993                 /*
2994                  * set the writeback_index so that range_cyclic
2995                  * mode will write it back later
2996                  */
2997                 mapping->writeback_index = index;
2998
2999 out_writepages:
3000         if (!no_nrwrite_index_update)
3001                 wbc->no_nrwrite_index_update = 0;
3002         if (wbc->nr_to_write > nr_to_writebump)
3003                 wbc->nr_to_write -= nr_to_writebump;
3004         wbc->range_start = range_start;
3005         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
3006         return ret;
3007 }
3008
3009 #define FALL_BACK_TO_NONDELALLOC 1
3010 static int ext4_nonda_switch(struct super_block *sb)
3011 {
3012         s64 free_blocks, dirty_blocks;
3013         struct ext4_sb_info *sbi = EXT4_SB(sb);
3014
3015         /*
3016          * switch to non delalloc mode if we are running low
3017          * on free block. The free block accounting via percpu
3018          * counters can get slightly wrong with percpu_counter_batch getting
3019          * accumulated on each CPU without updating global counters
3020          * Delalloc need an accurate free block accounting. So switch
3021          * to non delalloc when we are near to error range.
3022          */
3023         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
3024         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
3025         if (2 * free_blocks < 3 * dirty_blocks ||
3026                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
3027                 /*
3028                  * free block count is less that 150% of dirty blocks
3029                  * or free blocks is less that watermark
3030                  */
3031                 return 1;
3032         }
3033         return 0;
3034 }
3035
3036 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3037                                loff_t pos, unsigned len, unsigned flags,
3038                                struct page **pagep, void **fsdata)
3039 {
3040         int ret, retries = 0;
3041         struct page *page;
3042         pgoff_t index;
3043         unsigned from, to;
3044         struct inode *inode = mapping->host;
3045         handle_t *handle;
3046
3047         index = pos >> PAGE_CACHE_SHIFT;
3048         from = pos & (PAGE_CACHE_SIZE - 1);
3049         to = from + len;
3050
3051         if (ext4_nonda_switch(inode->i_sb)) {
3052                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3053                 return ext4_write_begin(file, mapping, pos,
3054                                         len, flags, pagep, fsdata);
3055         }
3056         *fsdata = (void *)0;
3057         trace_ext4_da_write_begin(inode, pos, len, flags);
3058 retry:
3059         /*
3060          * With delayed allocation, we don't log the i_disksize update
3061          * if there is delayed block allocation. But we still need
3062          * to journalling the i_disksize update if writes to the end
3063          * of file which has an already mapped buffer.
3064          */
3065         handle = ext4_journal_start(inode, 1);
3066         if (IS_ERR(handle)) {
3067                 ret = PTR_ERR(handle);
3068                 goto out;
3069         }
3070         /* We cannot recurse into the filesystem as the transaction is already
3071          * started */
3072         flags |= AOP_FLAG_NOFS;
3073
3074         page = grab_cache_page_write_begin(mapping, index, flags);
3075         if (!page) {
3076                 ext4_journal_stop(handle);
3077                 ret = -ENOMEM;
3078                 goto out;
3079         }
3080         *pagep = page;
3081
3082         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
3083                                 ext4_da_get_block_prep);
3084         if (ret < 0) {
3085                 unlock_page(page);
3086                 ext4_journal_stop(handle);
3087                 page_cache_release(page);
3088                 /*
3089                  * block_write_begin may have instantiated a few blocks
3090                  * outside i_size.  Trim these off again. Don't need
3091                  * i_size_read because we hold i_mutex.
3092                  */
3093                 if (pos + len > inode->i_size)
3094                         ext4_truncate(inode);
3095         }
3096
3097         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3098                 goto retry;
3099 out:
3100         return ret;
3101 }
3102
3103 /*
3104  * Check if we should update i_disksize
3105  * when write to the end of file but not require block allocation
3106  */
3107 static int ext4_da_should_update_i_disksize(struct page *page,
3108                                             unsigned long offset)
3109 {
3110         struct buffer_head *bh;
3111         struct inode *inode = page->mapping->host;
3112         unsigned int idx;
3113         int i;
3114
3115         bh = page_buffers(page);
3116         idx = offset >> inode->i_blkbits;
3117
3118         for (i = 0; i < idx; i++)
3119                 bh = bh->b_this_page;
3120
3121         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3122                 return 0;
3123         return 1;
3124 }
3125
3126 static int ext4_da_write_end(struct file *file,
3127                              struct address_space *mapping,
3128                              loff_t pos, unsigned len, unsigned copied,
3129                              struct page *page, void *fsdata)
3130 {
3131         struct inode *inode = mapping->host;
3132         int ret = 0, ret2;
3133         handle_t *handle = ext4_journal_current_handle();
3134         loff_t new_i_size;
3135         unsigned long start, end;
3136         int write_mode = (int)(unsigned long)fsdata;
3137
3138         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
3139                 if (ext4_should_order_data(inode)) {
3140                         return ext4_ordered_write_end(file, mapping, pos,
3141                                         len, copied, page, fsdata);
3142                 } else if (ext4_should_writeback_data(inode)) {
3143                         return ext4_writeback_write_end(file, mapping, pos,
3144                                         len, copied, page, fsdata);
3145                 } else {
3146                         BUG();
3147                 }
3148         }
3149
3150         trace_ext4_da_write_end(inode, pos, len, copied);
3151         start = pos & (PAGE_CACHE_SIZE - 1);
3152         end = start + copied - 1;
3153
3154         /*
3155          * generic_write_end() will run mark_inode_dirty() if i_size
3156          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3157          * into that.
3158          */
3159
3160         new_i_size = pos + copied;
3161         if (new_i_size > EXT4_I(inode)->i_disksize) {
3162                 if (ext4_da_should_update_i_disksize(page, end)) {
3163                         down_write(&EXT4_I(inode)->i_data_sem);
3164                         if (new_i_size > EXT4_I(inode)->i_disksize) {
3165                                 /*
3166                                  * Updating i_disksize when extending file
3167                                  * without needing block allocation
3168                                  */
3169                                 if (ext4_should_order_data(inode))
3170                                         ret = ext4_jbd2_file_inode(handle,
3171                                                                    inode);
3172
3173                                 EXT4_I(inode)->i_disksize = new_i_size;
3174                         }
3175                         up_write(&EXT4_I(inode)->i_data_sem);
3176                         /* We need to mark inode dirty even if
3177                          * new_i_size is less that inode->i_size
3178                          * bu greater than i_disksize.(hint delalloc)
3179                          */
3180                         ext4_mark_inode_dirty(handle, inode);
3181                 }
3182         }
3183         ret2 = generic_write_end(file, mapping, pos, len, copied,
3184                                                         page, fsdata);
3185         copied = ret2;
3186         if (ret2 < 0)
3187                 ret = ret2;
3188         ret2 = ext4_journal_stop(handle);
3189         if (!ret)
3190                 ret = ret2;
3191
3192         return ret ? ret : copied;
3193 }
3194
3195 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
3196 {
3197         /*
3198          * Drop reserved blocks
3199          */
3200         BUG_ON(!PageLocked(page));
3201         if (!page_has_buffers(page))
3202                 goto out;
3203
3204         ext4_da_page_release_reservation(page, offset);
3205
3206 out:
3207         ext4_invalidatepage(page, offset);
3208
3209         return;
3210 }
3211
3212 /*
3213  * Force all delayed allocation blocks to be allocated for a given inode.
3214  */
3215 int ext4_alloc_da_blocks(struct inode *inode)
3216 {
3217         trace_ext4_alloc_da_blocks(inode);
3218
3219         if (!EXT4_I(inode)->i_reserved_data_blocks &&
3220             !EXT4_I(inode)->i_reserved_meta_blocks)
3221                 return 0;
3222
3223         /*
3224          * We do something simple for now.  The filemap_flush() will
3225          * also start triggering a write of the data blocks, which is
3226          * not strictly speaking necessary (and for users of
3227          * laptop_mode, not even desirable).  However, to do otherwise
3228          * would require replicating code paths in:
3229          *
3230          * ext4_da_writepages() ->
3231          *    write_cache_pages() ---> (via passed in callback function)
3232          *        __mpage_da_writepage() -->
3233          *           mpage_add_bh_to_extent()
3234          *           mpage_da_map_blocks()
3235          *
3236          * The problem is that write_cache_pages(), located in
3237          * mm/page-writeback.c, marks pages clean in preparation for
3238          * doing I/O, which is not desirable if we're not planning on
3239          * doing I/O at all.
3240          *
3241          * We could call write_cache_pages(), and then redirty all of
3242          * the pages by calling redirty_page_for_writeback() but that
3243          * would be ugly in the extreme.  So instead we would need to
3244          * replicate parts of the code in the above functions,
3245          * simplifying them becuase we wouldn't actually intend to
3246          * write out the pages, but rather only collect contiguous
3247          * logical block extents, call the multi-block allocator, and
3248          * then update the buffer heads with the block allocations.
3249          *
3250          * For now, though, we'll cheat by calling filemap_flush(),
3251          * which will map the blocks, and start the I/O, but not
3252          * actually wait for the I/O to complete.
3253          */
3254         return filemap_flush(inode->i_mapping);
3255 }
3256
3257 /*
3258  * bmap() is special.  It gets used by applications such as lilo and by
3259  * the swapper to find the on-disk block of a specific piece of data.
3260  *
3261  * Naturally, this is dangerous if the block concerned is still in the
3262  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3263  * filesystem and enables swap, then they may get a nasty shock when the
3264  * data getting swapped to that swapfile suddenly gets overwritten by
3265  * the original zero's written out previously to the journal and
3266  * awaiting writeback in the kernel's buffer cache.
3267  *
3268  * So, if we see any bmap calls here on a modified, data-journaled file,
3269  * take extra steps to flush any blocks which might be in the cache.
3270  */
3271 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3272 {
3273         struct inode *inode = mapping->host;
3274         journal_t *journal;
3275         int err;
3276
3277         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3278                         test_opt(inode->i_sb, DELALLOC)) {
3279                 /*
3280                  * With delalloc we want to sync the file
3281                  * so that we can make sure we allocate
3282                  * blocks for file
3283                  */
3284                 filemap_write_and_wait(mapping);
3285         }
3286
3287         if (EXT4_JOURNAL(inode) && EXT4_I(inode)->i_state & EXT4_STATE_JDATA) {
3288                 /*
3289                  * This is a REALLY heavyweight approach, but the use of
3290                  * bmap on dirty files is expected to be extremely rare:
3291                  * only if we run lilo or swapon on a freshly made file
3292                  * do we expect this to happen.
3293                  *
3294                  * (bmap requires CAP_SYS_RAWIO so this does not
3295                  * represent an unprivileged user DOS attack --- we'd be
3296                  * in trouble if mortal users could trigger this path at
3297                  * will.)
3298                  *
3299                  * NB. EXT4_STATE_JDATA is not set on files other than
3300                  * regular files.  If somebody wants to bmap a directory
3301                  * or symlink and gets confused because the buffer
3302                  * hasn't yet been flushed to disk, they deserve
3303                  * everything they get.
3304                  */
3305
3306                 EXT4_I(inode)->i_state &= ~EXT4_STATE_JDATA;
3307                 journal = EXT4_JOURNAL(inode);
3308                 jbd2_journal_lock_updates(journal);
3309                 err = jbd2_journal_flush(journal);
3310                 jbd2_journal_unlock_updates(journal);
3311
3312                 if (err)
3313                         return 0;
3314         }
3315
3316         return generic_block_bmap(mapping, block, ext4_get_block);
3317 }
3318
3319 static int ext4_readpage(struct file *file, struct page *page)
3320 {
3321         return mpage_readpage(page, ext4_get_block);
3322 }
3323
3324 static int
3325 ext4_readpages(struct file *file, struct address_space *mapping,
3326                 struct list_head *pages, unsigned nr_pages)
3327 {
3328         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3329 }
3330
3331 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3332 {
3333         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3334
3335         /*
3336          * If it's a full truncate we just forget about the pending dirtying
3337          */
3338         if (offset == 0)
3339                 ClearPageChecked(page);
3340
3341         if (journal)
3342                 jbd2_journal_invalidatepage(journal, page, offset);
3343         else
3344                 block_invalidatepage(page, offset);
3345 }
3346
3347 static int ext4_releasepage(struct page *page, gfp_t wait)
3348 {
3349         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3350
3351         WARN_ON(PageChecked(page));
3352         if (!page_has_buffers(page))
3353                 return 0;
3354         if (journal)
3355                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3356         else
3357                 return try_to_free_buffers(page);
3358 }
3359
3360 /*
3361  * O_DIRECT for ext3 (or indirect map) based files
3362  *
3363  * If the O_DIRECT write will extend the file then add this inode to the
3364  * orphan list.  So recovery will truncate it back to the original size
3365  * if the machine crashes during the write.
3366  *
3367  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3368  * crashes then stale disk data _may_ be exposed inside the file. But current
3369  * VFS code falls back into buffered path in that case so we are safe.
3370  */
3371 static ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
3372                               const struct iovec *iov, loff_t offset,
3373                               unsigned long nr_segs)
3374 {
3375         struct file *file = iocb->ki_filp;
3376         struct inode *inode = file->f_mapping->host;
3377         struct ext4_inode_info *ei = EXT4_I(inode);
3378         handle_t *handle;
3379         ssize_t ret;
3380         int orphan = 0;
3381         size_t count = iov_length(iov, nr_segs);
3382         int retries = 0;
3383
3384         if (rw == WRITE) {
3385                 loff_t final_size = offset + count;
3386
3387                 if (final_size > inode->i_size) {
3388                         /* Credits for sb + inode write */
3389                         handle = ext4_journal_start(inode, 2);
3390                         if (IS_ERR(handle)) {
3391                                 ret = PTR_ERR(handle);
3392                                 goto out;
3393                         }
3394                         ret = ext4_orphan_add(handle, inode);
3395                         if (ret) {
3396                                 ext4_journal_stop(handle);
3397                                 goto out;
3398                         }
3399                         orphan = 1;
3400                         ei->i_disksize = inode->i_size;
3401                         ext4_journal_stop(handle);
3402                 }
3403         }
3404
3405 retry:
3406         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
3407                                  offset, nr_segs,
3408                                  ext4_get_block, NULL);
3409         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3410                 goto retry;
3411
3412         if (orphan) {
3413                 int err;
3414
3415                 /* Credits for sb + inode write */
3416                 handle = ext4_journal_start(inode, 2);
3417                 if (IS_ERR(handle)) {
3418                         /* This is really bad luck. We've written the data
3419                          * but cannot extend i_size. Bail out and pretend
3420                          * the write failed... */
3421                         ret = PTR_ERR(handle);
3422                         goto out;
3423                 }
3424                 if (inode->i_nlink)
3425                         ext4_orphan_del(handle, inode);
3426                 if (ret > 0) {
3427                         loff_t end = offset + ret;
3428                         if (end > inode->i_size) {
3429                                 ei->i_disksize = end;
3430                                 i_size_write(inode, end);
3431                                 /*
3432                                  * We're going to return a positive `ret'
3433                                  * here due to non-zero-length I/O, so there's
3434                                  * no way of reporting error returns from
3435                                  * ext4_mark_inode_dirty() to userspace.  So
3436                                  * ignore it.
3437                                  */
3438                                 ext4_mark_inode_dirty(handle, inode);
3439                         }
3440                 }
3441                 err = ext4_journal_stop(handle);
3442                 if (ret == 0)
3443                         ret = err;
3444         }
3445 out:
3446         return ret;
3447 }
3448
3449 /* Maximum number of blocks we map for direct IO at once. */
3450
3451 static int ext4_get_block_dio_write(struct inode *inode, sector_t iblock,
3452                    struct buffer_head *bh_result, int create)
3453 {
3454         handle_t *handle = NULL;
3455         int ret = 0;
3456         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
3457         int dio_credits;
3458
3459         ext4_debug("ext4_get_block_dio_write: inode %lu, create flag %d\n",
3460                    inode->i_ino, create);
3461         /*
3462          * DIO VFS code passes create = 0 flag for write to
3463          * the middle of file. It does this to avoid block
3464          * allocation for holes, to prevent expose stale data
3465          * out when there is parallel buffered read (which does
3466          * not hold the i_mutex lock) while direct IO write has
3467          * not completed. DIO request on holes finally falls back
3468          * to buffered IO for this reason.
3469          *
3470          * For ext4 extent based file, since we support fallocate,
3471          * new allocated extent as uninitialized, for holes, we
3472          * could fallocate blocks for holes, thus parallel
3473          * buffered IO read will zero out the page when read on
3474          * a hole while parallel DIO write to the hole has not completed.
3475          *
3476          * when we come here, we know it's a direct IO write to
3477          * to the middle of file (<i_size)
3478          * so it's safe to override the create flag from VFS.
3479          */
3480         create = EXT4_GET_BLOCKS_DIO_CREATE_EXT;
3481
3482         if (max_blocks > DIO_MAX_BLOCKS)
3483                 max_blocks = DIO_MAX_BLOCKS;
3484         dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
3485         handle = ext4_journal_start(inode, dio_credits);
3486         if (IS_ERR(handle)) {
3487                 ret = PTR_ERR(handle);
3488                 goto out;
3489         }
3490         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
3491                               create);
3492         if (ret > 0) {
3493                 bh_result->b_size = (ret << inode->i_blkbits);
3494                 ret = 0;
3495         }
3496         ext4_journal_stop(handle);
3497 out:
3498         return ret;
3499 }
3500
3501 static void ext4_free_io_end(ext4_io_end_t *io)
3502 {
3503         BUG_ON(!io);
3504         iput(io->inode);
3505         kfree(io);
3506 }
3507 static void dump_aio_dio_list(struct inode * inode)
3508 {
3509 #ifdef  EXT4_DEBUG
3510         struct list_head *cur, *before, *after;
3511         ext4_io_end_t *io, *io0, *io1;
3512
3513         if (list_empty(&EXT4_I(inode)->i_aio_dio_complete_list)){
3514                 ext4_debug("inode %lu aio dio list is empty\n", inode->i_ino);
3515                 return;
3516         }
3517
3518         ext4_debug("Dump inode %lu aio_dio_completed_IO list \n", inode->i_ino);
3519         list_for_each_entry(io, &EXT4_I(inode)->i_aio_dio_complete_list, list){
3520                 cur = &io->list;
3521                 before = cur->prev;
3522                 io0 = container_of(before, ext4_io_end_t, list);
3523                 after = cur->next;
3524                 io1 = container_of(after, ext4_io_end_t, list);
3525
3526                 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
3527                             io, inode->i_ino, io0, io1);
3528         }
3529 #endif
3530 }
3531
3532 /*
3533  * check a range of space and convert unwritten extents to written.
3534  */
3535 static int ext4_end_aio_dio_nolock(ext4_io_end_t *io)
3536 {
3537         struct inode *inode = io->inode;
3538         loff_t offset = io->offset;
3539         size_t size = io->size;
3540         int ret = 0;
3541
3542         ext4_debug("end_aio_dio_onlock: io 0x%p from inode %lu,list->next 0x%p,"
3543                    "list->prev 0x%p\n",
3544                    io, inode->i_ino, io->list.next, io->list.prev);
3545
3546         if (list_empty(&io->list))
3547                 return ret;
3548
3549         if (io->flag != DIO_AIO_UNWRITTEN)
3550                 return ret;
3551
3552         if (offset + size <= i_size_read(inode))
3553                 ret = ext4_convert_unwritten_extents(inode, offset, size);
3554
3555         if (ret < 0) {
3556                 printk(KERN_EMERG "%s: failed to convert unwritten"
3557                         "extents to written extents, error is %d"
3558                         " io is still on inode %lu aio dio list\n",
3559                        __func__, ret, inode->i_ino);
3560                 return ret;
3561         }
3562
3563         /* clear the DIO AIO unwritten flag */
3564         io->flag = 0;
3565         return ret;
3566 }
3567 /*
3568  * work on completed aio dio IO, to convert unwritten extents to extents
3569  */
3570 static void ext4_end_aio_dio_work(struct work_struct *work)
3571 {
3572         ext4_io_end_t *io  = container_of(work, ext4_io_end_t, work);
3573         struct inode *inode = io->inode;
3574         int ret = 0;
3575
3576         mutex_lock(&inode->i_mutex);
3577         ret = ext4_end_aio_dio_nolock(io);
3578         if (ret >= 0) {
3579                 if (!list_empty(&io->list))
3580                         list_del_init(&io->list);
3581                 ext4_free_io_end(io);
3582         }
3583         mutex_unlock(&inode->i_mutex);
3584 }
3585 /*
3586  * This function is called from ext4_sync_file().
3587  *
3588  * When AIO DIO IO is completed, the work to convert unwritten
3589  * extents to written is queued on workqueue but may not get immediately
3590  * scheduled. When fsync is called, we need to ensure the
3591  * conversion is complete before fsync returns.
3592  * The inode keeps track of a list of completed AIO from DIO path
3593  * that might needs to do the conversion. This function walks through
3594  * the list and convert the related unwritten extents to written.
3595  */
3596 int flush_aio_dio_completed_IO(struct inode *inode)
3597 {
3598         ext4_io_end_t *io;
3599         int ret = 0;
3600         int ret2 = 0;
3601
3602         if (list_empty(&EXT4_I(inode)->i_aio_dio_complete_list))
3603                 return ret;
3604
3605         dump_aio_dio_list(inode);
3606         while (!list_empty(&EXT4_I(inode)->i_aio_dio_complete_list)){
3607                 io = list_entry(EXT4_I(inode)->i_aio_dio_complete_list.next,
3608                                 ext4_io_end_t, list);
3609                 /*
3610                  * Calling ext4_end_aio_dio_nolock() to convert completed
3611                  * IO to written.
3612                  *
3613                  * When ext4_sync_file() is called, run_queue() may already
3614                  * about to flush the work corresponding to this io structure.
3615                  * It will be upset if it founds the io structure related
3616                  * to the work-to-be schedule is freed.
3617                  *
3618                  * Thus we need to keep the io structure still valid here after
3619                  * convertion finished. The io structure has a flag to
3620                  * avoid double converting from both fsync and background work
3621                  * queue work.
3622                  */
3623                 ret = ext4_end_aio_dio_nolock(io);
3624                 if (ret < 0)
3625                         ret2 = ret;
3626                 else
3627                         list_del_init(&io->list);
3628         }
3629         return (ret2 < 0) ? ret2 : 0;
3630 }
3631
3632 static ext4_io_end_t *ext4_init_io_end (struct inode *inode)
3633 {
3634         ext4_io_end_t *io = NULL;
3635
3636         io = kmalloc(sizeof(*io), GFP_NOFS);
3637
3638         if (io) {
3639                 igrab(inode);
3640                 io->inode = inode;
3641                 io->flag = 0;
3642                 io->offset = 0;
3643                 io->size = 0;
3644                 io->error = 0;
3645                 INIT_WORK(&io->work, ext4_end_aio_dio_work);
3646                 INIT_LIST_HEAD(&io->list);
3647         }
3648
3649         return io;
3650 }
3651
3652 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3653                             ssize_t size, void *private)
3654 {
3655         ext4_io_end_t *io_end = iocb->private;
3656         struct workqueue_struct *wq;
3657
3658         ext_debug("ext4_end_io_dio(): io_end 0x%p"
3659                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
3660                   iocb->private, io_end->inode->i_ino, iocb, offset,
3661                   size);
3662         /* if not async direct IO or dio with 0 bytes write, just return */
3663         if (!io_end || !size)
3664                 return;
3665
3666         /* if not aio dio with unwritten extents, just free io and return */
3667         if (io_end->flag != DIO_AIO_UNWRITTEN){
3668                 ext4_free_io_end(io_end);
3669                 iocb->private = NULL;
3670                 return;
3671         }
3672
3673         io_end->offset = offset;
3674         io_end->size = size;
3675         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
3676
3677         /* queue the work to convert unwritten extents to written */
3678         queue_work(wq, &io_end->work);
3679
3680         /* Add the io_end to per-inode completed aio dio list*/
3681         list_add_tail(&io_end->list,
3682                  &EXT4_I(io_end->inode)->i_aio_dio_complete_list);
3683         iocb->private = NULL;
3684 }
3685 /*
3686  * For ext4 extent files, ext4 will do direct-io write to holes,
3687  * preallocated extents, and those write extend the file, no need to
3688  * fall back to buffered IO.
3689  *
3690  * For holes, we fallocate those blocks, mark them as unintialized
3691  * If those blocks were preallocated, we mark sure they are splited, but
3692  * still keep the range to write as unintialized.
3693  *
3694  * The unwrritten extents will be converted to written when DIO is completed.
3695  * For async direct IO, since the IO may still pending when return, we
3696  * set up an end_io call back function, which will do the convertion
3697  * when async direct IO completed.
3698  *
3699  * If the O_DIRECT write will extend the file then add this inode to the
3700  * orphan list.  So recovery will truncate it back to the original size
3701  * if the machine crashes during the write.
3702  *
3703  */
3704 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
3705                               const struct iovec *iov, loff_t offset,
3706                               unsigned long nr_segs)
3707 {
3708         struct file *file = iocb->ki_filp;
3709         struct inode *inode = file->f_mapping->host;
3710         ssize_t ret;
3711         size_t count = iov_length(iov, nr_segs);
3712
3713         loff_t final_size = offset + count;
3714         if (rw == WRITE && final_size <= inode->i_size) {
3715                 /*
3716                  * We could direct write to holes and fallocate.
3717                  *
3718                  * Allocated blocks to fill the hole are marked as uninitialized
3719                  * to prevent paralel buffered read to expose the stale data
3720                  * before DIO complete the data IO.
3721                  *
3722                  * As to previously fallocated extents, ext4 get_block
3723                  * will just simply mark the buffer mapped but still
3724                  * keep the extents uninitialized.
3725                  *
3726                  * for non AIO case, we will convert those unwritten extents
3727                  * to written after return back from blockdev_direct_IO.
3728                  *
3729                  * for async DIO, the conversion needs to be defered when
3730                  * the IO is completed. The ext4 end_io callback function
3731                  * will be called to take care of the conversion work.
3732                  * Here for async case, we allocate an io_end structure to
3733                  * hook to the iocb.
3734                  */
3735                 iocb->private = NULL;
3736                 EXT4_I(inode)->cur_aio_dio = NULL;
3737                 if (!is_sync_kiocb(iocb)) {
3738                         iocb->private = ext4_init_io_end(inode);
3739                         if (!iocb->private)
3740                                 return -ENOMEM;
3741                         /*
3742                          * we save the io structure for current async
3743                          * direct IO, so that later ext4_get_blocks()
3744                          * could flag the io structure whether there
3745                          * is a unwritten extents needs to be converted
3746                          * when IO is completed.
3747                          */
3748                         EXT4_I(inode)->cur_aio_dio = iocb->private;
3749                 }
3750
3751                 ret = blockdev_direct_IO(rw, iocb, inode,
3752                                          inode->i_sb->s_bdev, iov,
3753                                          offset, nr_segs,
3754                                          ext4_get_block_dio_write,
3755                                          ext4_end_io_dio);
3756                 if (iocb->private)
3757                         EXT4_I(inode)->cur_aio_dio = NULL;
3758                 /*
3759                  * The io_end structure takes a reference to the inode,
3760                  * that structure needs to be destroyed and the
3761                  * reference to the inode need to be dropped, when IO is
3762                  * complete, even with 0 byte write, or failed.
3763                  *
3764                  * In the successful AIO DIO case, the io_end structure will be
3765                  * desctroyed and the reference to the inode will be dropped
3766                  * after the end_io call back function is called.
3767                  *
3768                  * In the case there is 0 byte write, or error case, since
3769                  * VFS direct IO won't invoke the end_io call back function,
3770                  * we need to free the end_io structure here.
3771                  */
3772                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
3773                         ext4_free_io_end(iocb->private);
3774                         iocb->private = NULL;
3775                 } else if (ret > 0) {
3776                         int err;
3777                         /*
3778                          * for non AIO case, since the IO is already
3779                          * completed, we could do the convertion right here
3780                          */
3781                         err = ext4_convert_unwritten_extents(inode,
3782                                                              offset, ret);
3783                         if (err < 0)
3784                                 ret = err;
3785                 }
3786                 return ret;
3787         }
3788
3789         /* for write the the end of file case, we fall back to old way */
3790         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3791 }
3792
3793 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3794                               const struct iovec *iov, loff_t offset,
3795                               unsigned long nr_segs)
3796 {
3797         struct file *file = iocb->ki_filp;
3798         struct inode *inode = file->f_mapping->host;
3799
3800         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
3801                 return ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
3802
3803         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3804 }
3805
3806 /*
3807  * Pages can be marked dirty completely asynchronously from ext4's journalling
3808  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3809  * much here because ->set_page_dirty is called under VFS locks.  The page is
3810  * not necessarily locked.
3811  *
3812  * We cannot just dirty the page and leave attached buffers clean, because the
3813  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3814  * or jbddirty because all the journalling code will explode.
3815  *
3816  * So what we do is to mark the page "pending dirty" and next time writepage
3817  * is called, propagate that into the buffers appropriately.
3818  */
3819 static int ext4_journalled_set_page_dirty(struct page *page)
3820 {
3821         SetPageChecked(page);
3822         return __set_page_dirty_nobuffers(page);
3823 }
3824
3825 static const struct address_space_operations ext4_ordered_aops = {
3826         .readpage               = ext4_readpage,
3827         .readpages              = ext4_readpages,
3828         .writepage              = ext4_writepage,
3829         .sync_page              = block_sync_page,
3830         .write_begin            = ext4_write_begin,
3831         .write_end              = ext4_ordered_write_end,
3832         .bmap                   = ext4_bmap,
3833         .invalidatepage         = ext4_invalidatepage,
3834         .releasepage            = ext4_releasepage,
3835         .direct_IO              = ext4_direct_IO,
3836         .migratepage            = buffer_migrate_page,
3837         .is_partially_uptodate  = block_is_partially_uptodate,
3838         .error_remove_page      = generic_error_remove_page,
3839 };
3840
3841 static const struct address_space_operations ext4_writeback_aops = {
3842         .readpage               = ext4_readpage,
3843         .readpages              = ext4_readpages,
3844         .writepage              = ext4_writepage,
3845         .sync_page              = block_sync_page,
3846         .write_begin            = ext4_write_begin,
3847         .write_end              = ext4_writeback_write_end,
3848         .bmap                   = ext4_bmap,
3849         .invalidatepage         = ext4_invalidatepage,
3850         .releasepage            = ext4_releasepage,
3851         .direct_IO              = ext4_direct_IO,
3852         .migratepage            = buffer_migrate_page,
3853         .is_partially_uptodate  = block_is_partially_uptodate,
3854         .error_remove_page      = generic_error_remove_page,
3855 };
3856
3857 static const struct address_space_operations ext4_journalled_aops = {
3858         .readpage               = ext4_readpage,
3859         .readpages              = ext4_readpages,
3860         .writepage              = ext4_writepage,
3861         .sync_page              = block_sync_page,
3862         .write_begin            = ext4_write_begin,
3863         .write_end              = ext4_journalled_write_end,
3864         .set_page_dirty         = ext4_journalled_set_page_dirty,
3865         .bmap                   = ext4_bmap,
3866         .invalidatepage         = ext4_invalidatepage,
3867         .releasepage            = ext4_releasepage,
3868         .is_partially_uptodate  = block_is_partially_uptodate,
3869         .error_remove_page      = generic_error_remove_page,
3870 };
3871
3872 static const struct address_space_operations ext4_da_aops = {
3873         .readpage               = ext4_readpage,
3874         .readpages              = ext4_readpages,
3875         .writepage              = ext4_writepage,
3876         .writepages             = ext4_da_writepages,
3877         .sync_page              = block_sync_page,
3878         .write_begin            = ext4_da_write_begin,
3879         .write_end              = ext4_da_write_end,
3880         .bmap                   = ext4_bmap,
3881         .invalidatepage         = ext4_da_invalidatepage,
3882         .releasepage            = ext4_releasepage,
3883         .direct_IO              = ext4_direct_IO,
3884         .migratepage            = buffer_migrate_page,
3885         .is_partially_uptodate  = block_is_partially_uptodate,
3886         .error_remove_page      = generic_error_remove_page,
3887 };
3888
3889 void ext4_set_aops(struct inode *inode)
3890 {
3891         if (ext4_should_order_data(inode) &&
3892                 test_opt(inode->i_sb, DELALLOC))
3893                 inode->i_mapping->a_ops = &ext4_da_aops;
3894         else if (ext4_should_order_data(inode))
3895                 inode->i_mapping->a_ops = &ext4_ordered_aops;
3896         else if (ext4_should_writeback_data(inode) &&
3897                  test_opt(inode->i_sb, DELALLOC))
3898                 inode->i_mapping->a_ops = &ext4_da_aops;
3899         else if (ext4_should_writeback_data(inode))
3900                 inode->i_mapping->a_ops = &ext4_writeback_aops;
3901         else
3902                 inode->i_mapping->a_ops = &ext4_journalled_aops;
3903 }
3904
3905 /*
3906  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3907  * up to the end of the block which corresponds to `from'.
3908  * This required during truncate. We need to physically zero the tail end
3909  * of that block so it doesn't yield old data if the file is later grown.
3910  */
3911 int ext4_block_truncate_page(handle_t *handle,
3912                 struct address_space *mapping, loff_t from)
3913 {
3914         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3915         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3916         unsigned blocksize, length, pos;
3917         ext4_lblk_t iblock;
3918         struct inode *inode = mapping->host;
3919         struct buffer_head *bh;
3920         struct page *page;
3921         int err = 0;
3922
3923         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
3924                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
3925         if (!page)
3926                 return -EINVAL;
3927
3928         blocksize = inode->i_sb->s_blocksize;
3929         length = blocksize - (offset & (blocksize - 1));
3930         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3931
3932         /*
3933          * For "nobh" option,  we can only work if we don't need to
3934          * read-in the page - otherwise we create buffers to do the IO.
3935          */
3936         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
3937              ext4_should_writeback_data(inode) && PageUptodate(page)) {
3938                 zero_user(page, offset, length);
3939                 set_page_dirty(page);
3940                 goto unlock;
3941         }
3942
3943         if (!page_has_buffers(page))
3944                 create_empty_buffers(page, blocksize, 0);
3945
3946         /* Find the buffer that contains "offset" */
3947         bh = page_buffers(page);
3948         pos = blocksize;
3949         while (offset >= pos) {
3950                 bh = bh->b_this_page;
3951                 iblock++;
3952                 pos += blocksize;
3953         }
3954
3955         err = 0;
3956         if (buffer_freed(bh)) {
3957                 BUFFER_TRACE(bh, "freed: skip");
3958                 goto unlock;
3959         }
3960
3961         if (!buffer_mapped(bh)) {
3962                 BUFFER_TRACE(bh, "unmapped");
3963                 ext4_get_block(inode, iblock, bh, 0);
3964                 /* unmapped? It's a hole - nothing to do */
3965                 if (!buffer_mapped(bh)) {
3966                         BUFFER_TRACE(bh, "still unmapped");
3967                         goto unlock;
3968                 }
3969         }
3970
3971         /* Ok, it's mapped. Make sure it's up-to-date */
3972         if (PageUptodate(page))
3973                 set_buffer_uptodate(bh);
3974
3975         if (!buffer_uptodate(bh)) {
3976                 err = -EIO;
3977                 ll_rw_block(READ, 1, &bh);
3978                 wait_on_buffer(bh);
3979                 /* Uhhuh. Read error. Complain and punt. */
3980                 if (!buffer_uptodate(bh))
3981                         goto unlock;
3982         }
3983
3984         if (ext4_should_journal_data(inode)) {
3985                 BUFFER_TRACE(bh, "get write access");
3986                 err = ext4_journal_get_write_access(handle, bh);
3987                 if (err)
3988                         goto unlock;
3989         }
3990
3991         zero_user(page, offset, length);
3992
3993         BUFFER_TRACE(bh, "zeroed end of block");
3994
3995         err = 0;
3996         if (ext4_should_journal_data(inode)) {
3997                 err = ext4_handle_dirty_metadata(handle, inode, bh);
3998         } else {
3999                 if (ext4_should_order_data(inode))
4000                         err = ext4_jbd2_file_inode(handle, inode);
4001                 mark_buffer_dirty(bh);
4002         }
4003
4004 unlock:
4005         unlock_page(page);
4006         page_cache_release(page);
4007         return err;
4008 }
4009
4010 /*
4011  * Probably it should be a library function... search for first non-zero word
4012  * or memcmp with zero_page, whatever is better for particular architecture.
4013  * Linus?
4014  */
4015 static inline int all_zeroes(__le32 *p, __le32 *q)
4016 {
4017         while (p < q)
4018                 if (*p++)
4019                         return 0;
4020         return 1;
4021 }
4022
4023 /**
4024  *      ext4_find_shared - find the indirect blocks for partial truncation.
4025  *      @inode:   inode in question
4026  *      @depth:   depth of the affected branch
4027  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
4028  *      @chain:   place to store the pointers to partial indirect blocks
4029  *      @top:     place to the (detached) top of branch
4030  *
4031  *      This is a helper function used by ext4_truncate().
4032  *
4033  *      When we do truncate() we may have to clean the ends of several
4034  *      indirect blocks but leave the blocks themselves alive. Block is
4035  *      partially truncated if some data below the new i_size is refered
4036  *      from it (and it is on the path to the first completely truncated
4037  *      data block, indeed).  We have to free the top of that path along
4038  *      with everything to the right of the path. Since no allocation
4039  *      past the truncation point is possible until ext4_truncate()
4040  *      finishes, we may safely do the latter, but top of branch may
4041  *      require special attention - pageout below the truncation point
4042  *      might try to populate it.
4043  *
4044  *      We atomically detach the top of branch from the tree, store the
4045  *      block number of its root in *@top, pointers to buffer_heads of
4046  *      partially truncated blocks - in @chain[].bh and pointers to
4047  *      their last elements that should not be removed - in
4048  *      @chain[].p. Return value is the pointer to last filled element
4049  *      of @chain.
4050  *
4051  *      The work left to caller to do the actual freeing of subtrees:
4052  *              a) free the subtree starting from *@top
4053  *              b) free the subtrees whose roots are stored in
4054  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
4055  *              c) free the subtrees growing from the inode past the @chain[0].
4056  *                      (no partially truncated stuff there).  */
4057
4058 static Indirect *ext4_find_shared(struct inode *inode, int depth,
4059                                   ext4_lblk_t offsets[4], Indirect chain[4],
4060                                   __le32 *top)
4061 {
4062         Indirect *partial, *p;
4063         int k, err;
4064
4065         *top = 0;
4066         /* Make k index the deepest non-null offest + 1 */
4067         for (k = depth; k > 1 && !offsets[k-1]; k--)
4068                 ;
4069         partial = ext4_get_branch(inode, k, offsets, chain, &err);
4070         /* Writer: pointers */
4071         if (!partial)
4072                 partial = chain + k-1;
4073         /*
4074          * If the branch acquired continuation since we've looked at it -
4075          * fine, it should all survive and (new) top doesn't belong to us.
4076          */
4077         if (!partial->key && *partial->p)
4078                 /* Writer: end */
4079                 goto no_top;
4080         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
4081                 ;
4082         /*
4083          * OK, we've found the last block that must survive. The rest of our
4084          * branch should be detached before unlocking. However, if that rest
4085          * of branch is all ours and does not grow immediately from the inode
4086          * it's easier to cheat and just decrement partial->p.
4087          */
4088         if (p == chain + k - 1 && p > chain) {
4089                 p->p--;
4090         } else {
4091                 *top = *p->p;
4092                 /* Nope, don't do this in ext4.  Must leave the tree intact */
4093 #if 0
4094                 *p->p = 0;
4095 #endif
4096         }
4097         /* Writer: end */
4098
4099         while (partial > p) {
4100                 brelse(partial->bh);
4101                 partial--;
4102         }
4103 no_top:
4104         return partial;
4105 }
4106
4107 /*
4108  * Zero a number of block pointers in either an inode or an indirect block.
4109  * If we restart the transaction we must again get write access to the
4110  * indirect block for further modification.
4111  *
4112  * We release `count' blocks on disk, but (last - first) may be greater
4113  * than `count' because there can be holes in there.
4114  */
4115 static void ext4_clear_blocks(handle_t *handle, struct inode *inode,
4116                               struct buffer_head *bh,
4117                               ext4_fsblk_t block_to_free,
4118                               unsigned long count, __le32 *first,
4119                               __le32 *last)
4120 {
4121         __le32 *p;
4122         if (try_to_extend_transaction(handle, inode)) {
4123                 if (bh) {
4124                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4125                         ext4_handle_dirty_metadata(handle, inode, bh);
4126                 }
4127                 ext4_mark_inode_dirty(handle, inode);
4128                 ext4_truncate_restart_trans(handle, inode,
4129                                             blocks_for_truncate(inode));
4130                 if (bh) {
4131                         BUFFER_TRACE(bh, "retaking write access");
4132                         ext4_journal_get_write_access(handle, bh);
4133                 }
4134         }
4135
4136         /*
4137          * Any buffers which are on the journal will be in memory. We
4138          * find them on the hash table so jbd2_journal_revoke() will
4139          * run jbd2_journal_forget() on them.  We've already detached
4140          * each block from the file, so bforget() in
4141          * jbd2_journal_forget() should be safe.
4142          *
4143          * AKPM: turn on bforget in jbd2_journal_forget()!!!
4144          */
4145         for (p = first; p < last; p++) {
4146                 u32 nr = le32_to_cpu(*p);
4147                 if (nr) {
4148                         struct buffer_head *tbh;
4149
4150                         *p = 0;
4151                         tbh = sb_find_get_block(inode->i_sb, nr);
4152                         ext4_forget(handle, 0, inode, tbh, nr);
4153                 }
4154         }
4155
4156         ext4_free_blocks(handle, inode, block_to_free, count, 0);
4157 }
4158
4159 /**
4160  * ext4_free_data - free a list of data blocks
4161  * @handle:     handle for this transaction
4162  * @inode:      inode we are dealing with
4163  * @this_bh:    indirect buffer_head which contains *@first and *@last
4164  * @first:      array of block numbers
4165  * @last:       points immediately past the end of array
4166  *
4167  * We are freeing all blocks refered from that array (numbers are stored as
4168  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
4169  *
4170  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
4171  * blocks are contiguous then releasing them at one time will only affect one
4172  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
4173  * actually use a lot of journal space.
4174  *
4175  * @this_bh will be %NULL if @first and @last point into the inode's direct
4176  * block pointers.
4177  */
4178 static void ext4_free_data(handle_t *handle, struct inode *inode,
4179                            struct buffer_head *this_bh,
4180                            __le32 *first, __le32 *last)
4181 {
4182         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
4183         unsigned long count = 0;            /* Number of blocks in the run */
4184         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
4185                                                corresponding to
4186                                                block_to_free */
4187         ext4_fsblk_t nr;                    /* Current block # */
4188         __le32 *p;                          /* Pointer into inode/ind
4189                                                for current block */
4190         int err;
4191
4192         if (this_bh) {                          /* For indirect block */
4193                 BUFFER_TRACE(this_bh, "get_write_access");
4194                 err = ext4_journal_get_write_access(handle, this_bh);
4195                 /* Important: if we can't update the indirect pointers
4196                  * to the blocks, we can't free them. */
4197                 if (err)
4198                         return;
4199         }
4200
4201         for (p = first; p < last; p++) {
4202                 nr = le32_to_cpu(*p);
4203                 if (nr) {
4204                         /* accumulate blocks to free if they're contiguous */
4205                         if (count == 0) {
4206                                 block_to_free = nr;
4207                                 block_to_free_p = p;
4208                                 count = 1;
4209                         } else if (nr == block_to_free + count) {
4210                                 count++;
4211                         } else {
4212                                 ext4_clear_blocks(handle, inode, this_bh,
4213                                                   block_to_free,
4214                                                   count, block_to_free_p, p);
4215                                 block_to_free = nr;
4216                                 block_to_free_p = p;
4217                                 count = 1;
4218                         }
4219                 }
4220         }
4221
4222         if (count > 0)
4223                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
4224                                   count, block_to_free_p, p);
4225
4226         if (this_bh) {
4227                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
4228
4229                 /*
4230                  * The buffer head should have an attached journal head at this
4231                  * point. However, if the data is corrupted and an indirect
4232                  * block pointed to itself, it would have been detached when
4233                  * the block was cleared. Check for this instead of OOPSing.
4234                  */
4235                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
4236                         ext4_handle_dirty_metadata(handle, inode, this_bh);
4237                 else
4238                         ext4_error(inode->i_sb, __func__,
4239                                    "circular indirect block detected, "
4240                                    "inode=%lu, block=%llu",
4241                                    inode->i_ino,
4242                                    (unsigned long long) this_bh->b_blocknr);
4243         }
4244 }
4245
4246 /**
4247  *      ext4_free_branches - free an array of branches
4248  *      @handle: JBD handle for this transaction
4249  *      @inode: inode we are dealing with
4250  *      @parent_bh: the buffer_head which contains *@first and *@last
4251  *      @first: array of block numbers
4252  *      @last:  pointer immediately past the end of array
4253  *      @depth: depth of the branches to free
4254  *
4255  *      We are freeing all blocks refered from these branches (numbers are
4256  *      stored as little-endian 32-bit) and updating @inode->i_blocks
4257  *      appropriately.
4258  */
4259 static void ext4_free_branches(handle_t *handle, struct inode *inode,
4260                                struct buffer_head *parent_bh,
4261                                __le32 *first, __le32 *last, int depth)
4262 {
4263         ext4_fsblk_t nr;
4264         __le32 *p;
4265
4266         if (ext4_handle_is_aborted(handle))
4267                 return;
4268
4269         if (depth--) {
4270                 struct buffer_head *bh;
4271                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4272                 p = last;
4273                 while (--p >= first) {
4274                         nr = le32_to_cpu(*p);
4275                         if (!nr)
4276                                 continue;               /* A hole */
4277
4278                         /* Go read the buffer for the next level down */
4279                         bh = sb_bread(inode->i_sb, nr);
4280
4281                         /*
4282                          * A read failure? Report error and clear slot
4283                          * (should be rare).
4284                          */
4285                         if (!bh) {
4286                                 ext4_error(inode->i_sb, "ext4_free_branches",
4287                                            "Read failure, inode=%lu, block=%llu",
4288                                            inode->i_ino, nr);
4289                                 continue;
4290                         }
4291
4292                         /* This zaps the entire block.  Bottom up. */
4293                         BUFFER_TRACE(bh, "free child branches");
4294                         ext4_free_branches(handle, inode, bh,
4295                                         (__le32 *) bh->b_data,
4296                                         (__le32 *) bh->b_data + addr_per_block,
4297                                         depth);
4298
4299                         /*
4300                          * We've probably journalled the indirect block several
4301                          * times during the truncate.  But it's no longer
4302                          * needed and we now drop it from the transaction via
4303                          * jbd2_journal_revoke().
4304                          *
4305                          * That's easy if it's exclusively part of this
4306                          * transaction.  But if it's part of the committing
4307                          * transaction then jbd2_journal_forget() will simply
4308                          * brelse() it.  That means that if the underlying
4309                          * block is reallocated in ext4_get_block(),
4310                          * unmap_underlying_metadata() will find this block
4311                          * and will try to get rid of it.  damn, damn.
4312                          *
4313                          * If this block has already been committed to the
4314                          * journal, a revoke record will be written.  And
4315                          * revoke records must be emitted *before* clearing
4316                          * this block's bit in the bitmaps.
4317                          */
4318                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
4319
4320                         /*
4321                          * Everything below this this pointer has been
4322                          * released.  Now let this top-of-subtree go.
4323                          *
4324                          * We want the freeing of this indirect block to be
4325                          * atomic in the journal with the updating of the
4326                          * bitmap block which owns it.  So make some room in
4327                          * the journal.
4328                          *
4329                          * We zero the parent pointer *after* freeing its
4330                          * pointee in the bitmaps, so if extend_transaction()
4331                          * for some reason fails to put the bitmap changes and
4332                          * the release into the same transaction, recovery
4333                          * will merely complain about releasing a free block,
4334                          * rather than leaking blocks.
4335                          */
4336                         if (ext4_handle_is_aborted(handle))
4337                                 return;
4338                         if (try_to_extend_transaction(handle, inode)) {
4339                                 ext4_mark_inode_dirty(handle, inode);
4340                                 ext4_truncate_restart_trans(handle, inode,
4341                                             blocks_for_truncate(inode));
4342                         }
4343
4344                         ext4_free_blocks(handle, inode, nr, 1, 1);
4345
4346                         if (parent_bh) {
4347                                 /*
4348                                  * The block which we have just freed is
4349                                  * pointed to by an indirect block: journal it
4350                                  */
4351                                 BUFFER_TRACE(parent_bh, "get_write_access");
4352                                 if (!ext4_journal_get_write_access(handle,
4353                                                                    parent_bh)){
4354                                         *p = 0;
4355                                         BUFFER_TRACE(parent_bh,
4356                                         "call ext4_handle_dirty_metadata");
4357                                         ext4_handle_dirty_metadata(handle,
4358                                                                    inode,
4359                                                                    parent_bh);
4360                                 }
4361                         }
4362                 }
4363         } else {
4364                 /* We have reached the bottom of the tree. */
4365                 BUFFER_TRACE(parent_bh, "free data blocks");
4366                 ext4_free_data(handle, inode, parent_bh, first, last);
4367         }
4368 }
4369
4370 int ext4_can_truncate(struct inode *inode)
4371 {
4372         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4373                 return 0;
4374         if (S_ISREG(inode->i_mode))
4375                 return 1;
4376         if (S_ISDIR(inode->i_mode))
4377                 return 1;
4378         if (S_ISLNK(inode->i_mode))
4379                 return !ext4_inode_is_fast_symlink(inode);
4380         return 0;
4381 }
4382
4383 /*
4384  * ext4_truncate()
4385  *
4386  * We block out ext4_get_block() block instantiations across the entire
4387  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4388  * simultaneously on behalf of the same inode.
4389  *
4390  * As we work through the truncate and commmit bits of it to the journal there
4391  * is one core, guiding principle: the file's tree must always be consistent on
4392  * disk.  We must be able to restart the truncate after a crash.
4393  *
4394  * The file's tree may be transiently inconsistent in memory (although it
4395  * probably isn't), but whenever we close off and commit a journal transaction,
4396  * the contents of (the filesystem + the journal) must be consistent and
4397  * restartable.  It's pretty simple, really: bottom up, right to left (although
4398  * left-to-right works OK too).
4399  *
4400  * Note that at recovery time, journal replay occurs *before* the restart of
4401  * truncate against the orphan inode list.
4402  *
4403  * The committed inode has the new, desired i_size (which is the same as
4404  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4405  * that this inode's truncate did not complete and it will again call
4406  * ext4_truncate() to have another go.  So there will be instantiated blocks
4407  * to the right of the truncation point in a crashed ext4 filesystem.  But
4408  * that's fine - as long as they are linked from the inode, the post-crash
4409  * ext4_truncate() run will find them and release them.
4410  */
4411 void ext4_truncate(struct inode *inode)
4412 {
4413         handle_t *handle;
4414         struct ext4_inode_info *ei = EXT4_I(inode);
4415         __le32 *i_data = ei->i_data;
4416         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4417         struct address_space *mapping = inode->i_mapping;
4418         ext4_lblk_t offsets[4];
4419         Indirect chain[4];
4420         Indirect *partial;
4421         __le32 nr = 0;
4422         int n;
4423         ext4_lblk_t last_block;
4424         unsigned blocksize = inode->i_sb->s_blocksize;
4425
4426         if (!ext4_can_truncate(inode))
4427                 return;
4428
4429         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4430                 ei->i_state |= EXT4_STATE_DA_ALLOC_CLOSE;
4431
4432         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
4433                 ext4_ext_truncate(inode);
4434                 return;
4435         }
4436
4437         handle = start_transaction(inode);
4438         if (IS_ERR(handle))
4439                 return;         /* AKPM: return what? */
4440
4441         last_block = (inode->i_size + blocksize-1)
4442                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4443
4444         if (inode->i_size & (blocksize - 1))
4445                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4446                         goto out_stop;
4447
4448         n = ext4_block_to_path(inode, last_block, offsets, NULL);
4449         if (n == 0)
4450                 goto out_stop;  /* error */
4451
4452         /*
4453          * OK.  This truncate is going to happen.  We add the inode to the
4454          * orphan list, so that if this truncate spans multiple transactions,
4455          * and we crash, we will resume the truncate when the filesystem
4456          * recovers.  It also marks the inode dirty, to catch the new size.
4457          *
4458          * Implication: the file must always be in a sane, consistent
4459          * truncatable state while each transaction commits.
4460          */
4461         if (ext4_orphan_add(handle, inode))
4462                 goto out_stop;
4463
4464         /*
4465          * From here we block out all ext4_get_block() callers who want to
4466          * modify the block allocation tree.
4467          */
4468         down_write(&ei->i_data_sem);
4469
4470         ext4_discard_preallocations(inode);
4471
4472         /*
4473          * The orphan list entry will now protect us from any crash which
4474          * occurs before the truncate completes, so it is now safe to propagate
4475          * the new, shorter inode size (held for now in i_size) into the
4476          * on-disk inode. We do this via i_disksize, which is the value which
4477          * ext4 *really* writes onto the disk inode.
4478          */
4479         ei->i_disksize = inode->i_size;
4480
4481         if (n == 1) {           /* direct blocks */
4482                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4483                                i_data + EXT4_NDIR_BLOCKS);
4484                 goto do_indirects;
4485         }
4486
4487         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4488         /* Kill the top of shared branch (not detached) */
4489         if (nr) {
4490                 if (partial == chain) {
4491                         /* Shared branch grows from the inode */
4492                         ext4_free_branches(handle, inode, NULL,
4493                                            &nr, &nr+1, (chain+n-1) - partial);
4494                         *partial->p = 0;
4495                         /*
4496                          * We mark the inode dirty prior to restart,
4497                          * and prior to stop.  No need for it here.
4498                          */
4499                 } else {
4500                         /* Shared branch grows from an indirect block */
4501                         BUFFER_TRACE(partial->bh, "get_write_access");
4502                         ext4_free_branches(handle, inode, partial->bh,
4503                                         partial->p,
4504                                         partial->p+1, (chain+n-1) - partial);
4505                 }
4506         }
4507         /* Clear the ends of indirect blocks on the shared branch */
4508         while (partial > chain) {
4509                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4510                                    (__le32*)partial->bh->b_data+addr_per_block,
4511                                    (chain+n-1) - partial);
4512                 BUFFER_TRACE(partial->bh, "call brelse");
4513                 brelse(partial->bh);
4514                 partial--;
4515         }
4516 do_indirects:
4517         /* Kill the remaining (whole) subtrees */
4518         switch (offsets[0]) {
4519         default:
4520                 nr = i_data[EXT4_IND_BLOCK];
4521                 if (nr) {
4522                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4523                         i_data[EXT4_IND_BLOCK] = 0;
4524                 }
4525         case EXT4_IND_BLOCK:
4526                 nr = i_data[EXT4_DIND_BLOCK];
4527                 if (nr) {
4528                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4529                         i_data[EXT4_DIND_BLOCK] = 0;
4530                 }
4531         case EXT4_DIND_BLOCK:
4532                 nr = i_data[EXT4_TIND_BLOCK];
4533                 if (nr) {
4534                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4535                         i_data[EXT4_TIND_BLOCK] = 0;
4536                 }
4537         case EXT4_TIND_BLOCK:
4538                 ;
4539         }
4540
4541         up_write(&ei->i_data_sem);
4542         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4543         ext4_mark_inode_dirty(handle, inode);
4544
4545         /*
4546          * In a multi-transaction truncate, we only make the final transaction
4547          * synchronous
4548          */
4549         if (IS_SYNC(inode))
4550                 ext4_handle_sync(handle);
4551 out_stop:
4552         /*
4553          * If this was a simple ftruncate(), and the file will remain alive
4554          * then we need to clear up the orphan record which we created above.
4555          * However, if this was a real unlink then we were called by
4556          * ext4_delete_inode(), and we allow that function to clean up the
4557          * orphan info for us.
4558          */
4559         if (inode->i_nlink)
4560                 ext4_orphan_del(handle, inode);
4561
4562         ext4_journal_stop(handle);
4563 }
4564
4565 /*
4566  * ext4_get_inode_loc returns with an extra refcount against the inode's
4567  * underlying buffer_head on success. If 'in_mem' is true, we have all
4568  * data in memory that is needed to recreate the on-disk version of this
4569  * inode.
4570  */
4571 static int __ext4_get_inode_loc(struct inode *inode,
4572                                 struct ext4_iloc *iloc, int in_mem)
4573 {
4574         struct ext4_group_desc  *gdp;
4575         struct buffer_head      *bh;
4576         struct super_block      *sb = inode->i_sb;
4577         ext4_fsblk_t            block;
4578         int                     inodes_per_block, inode_offset;
4579
4580         iloc->bh = NULL;
4581         if (!ext4_valid_inum(sb, inode->i_ino))
4582                 return -EIO;
4583
4584         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4585         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4586         if (!gdp)
4587                 return -EIO;
4588
4589         /*
4590          * Figure out the offset within the block group inode table
4591          */
4592         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4593         inode_offset = ((inode->i_ino - 1) %
4594                         EXT4_INODES_PER_GROUP(sb));
4595         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4596         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4597
4598         bh = sb_getblk(sb, block);
4599         if (!bh) {
4600                 ext4_error(sb, "ext4_get_inode_loc", "unable to read "
4601                            "inode block - inode=%lu, block=%llu",
4602                            inode->i_ino, block);
4603                 return -EIO;
4604         }
4605         if (!buffer_uptodate(bh)) {
4606                 lock_buffer(bh);
4607
4608                 /*
4609                  * If the buffer has the write error flag, we have failed
4610                  * to write out another inode in the same block.  In this
4611                  * case, we don't have to read the block because we may
4612                  * read the old inode data successfully.
4613                  */
4614                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4615                         set_buffer_uptodate(bh);
4616
4617                 if (buffer_uptodate(bh)) {
4618                         /* someone brought it uptodate while we waited */
4619                         unlock_buffer(bh);
4620                         goto has_buffer;
4621                 }
4622
4623                 /*
4624                  * If we have all information of the inode in memory and this
4625                  * is the only valid inode in the block, we need not read the
4626                  * block.
4627                  */
4628                 if (in_mem) {
4629                         struct buffer_head *bitmap_bh;
4630                         int i, start;
4631
4632                         start = inode_offset & ~(inodes_per_block - 1);
4633
4634                         /* Is the inode bitmap in cache? */
4635                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4636                         if (!bitmap_bh)
4637                                 goto make_io;
4638
4639                         /*
4640                          * If the inode bitmap isn't in cache then the
4641                          * optimisation may end up performing two reads instead
4642                          * of one, so skip it.
4643                          */
4644                         if (!buffer_uptodate(bitmap_bh)) {
4645                                 brelse(bitmap_bh);
4646                                 goto make_io;
4647                         }
4648                         for (i = start; i < start + inodes_per_block; i++) {
4649                                 if (i == inode_offset)
4650                                         continue;
4651                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4652                                         break;
4653                         }
4654                         brelse(bitmap_bh);
4655                         if (i == start + inodes_per_block) {
4656                                 /* all other inodes are free, so skip I/O */
4657                                 memset(bh->b_data, 0, bh->b_size);
4658                                 set_buffer_uptodate(bh);
4659                                 unlock_buffer(bh);
4660                                 goto has_buffer;
4661                         }
4662                 }
4663
4664 make_io:
4665                 /*
4666                  * If we need to do any I/O, try to pre-readahead extra
4667                  * blocks from the inode table.
4668                  */
4669                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4670                         ext4_fsblk_t b, end, table;
4671                         unsigned num;
4672
4673                         table = ext4_inode_table(sb, gdp);
4674                         /* s_inode_readahead_blks is always a power of 2 */
4675                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4676                         if (table > b)
4677                                 b = table;
4678                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4679                         num = EXT4_INODES_PER_GROUP(sb);
4680                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4681                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4682                                 num -= ext4_itable_unused_count(sb, gdp);
4683                         table += num / inodes_per_block;
4684                         if (end > table)
4685                                 end = table;
4686                         while (b <= end)
4687                                 sb_breadahead(sb, b++);
4688                 }
4689
4690                 /*
4691                  * There are other valid inodes in the buffer, this inode
4692                  * has in-inode xattrs, or we don't have this inode in memory.
4693                  * Read the block from disk.
4694                  */
4695                 get_bh(bh);
4696                 bh->b_end_io = end_buffer_read_sync;
4697                 submit_bh(READ_META, bh);
4698                 wait_on_buffer(bh);
4699                 if (!buffer_uptodate(bh)) {
4700                         ext4_error(sb, __func__,
4701                                    "unable to read inode block - inode=%lu, "
4702                                    "block=%llu", inode->i_ino, block);
4703                         brelse(bh);
4704                         return -EIO;
4705                 }
4706         }
4707 has_buffer:
4708         iloc->bh = bh;
4709         return 0;
4710 }
4711
4712 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4713 {
4714         /* We have all inode data except xattrs in memory here. */
4715         return __ext4_get_inode_loc(inode, iloc,
4716                 !(EXT4_I(inode)->i_state & EXT4_STATE_XATTR));
4717 }
4718
4719 void ext4_set_inode_flags(struct inode *inode)
4720 {
4721         unsigned int flags = EXT4_I(inode)->i_flags;
4722
4723         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4724         if (flags & EXT4_SYNC_FL)
4725                 inode->i_flags |= S_SYNC;
4726         if (flags & EXT4_APPEND_FL)
4727                 inode->i_flags |= S_APPEND;
4728         if (flags & EXT4_IMMUTABLE_FL)
4729                 inode->i_flags |= S_IMMUTABLE;
4730         if (flags & EXT4_NOATIME_FL)
4731                 inode->i_flags |= S_NOATIME;
4732         if (flags & EXT4_DIRSYNC_FL)
4733                 inode->i_flags |= S_DIRSYNC;
4734 }
4735
4736 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4737 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4738 {
4739         unsigned int flags = ei->vfs_inode.i_flags;
4740
4741         ei->i_flags &= ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4742                         EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|EXT4_DIRSYNC_FL);
4743         if (flags & S_SYNC)
4744                 ei->i_flags |= EXT4_SYNC_FL;
4745         if (flags & S_APPEND)
4746                 ei->i_flags |= EXT4_APPEND_FL;
4747         if (flags & S_IMMUTABLE)
4748                 ei->i_flags |= EXT4_IMMUTABLE_FL;
4749         if (flags & S_NOATIME)
4750                 ei->i_flags |= EXT4_NOATIME_FL;
4751         if (flags & S_DIRSYNC)
4752                 ei->i_flags |= EXT4_DIRSYNC_FL;
4753 }
4754
4755 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4756                                   struct ext4_inode_info *ei)
4757 {
4758         blkcnt_t i_blocks ;
4759         struct inode *inode = &(ei->vfs_inode);
4760         struct super_block *sb = inode->i_sb;
4761
4762         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4763                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4764                 /* we are using combined 48 bit field */
4765                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4766                                         le32_to_cpu(raw_inode->i_blocks_lo);
4767                 if (ei->i_flags & EXT4_HUGE_FILE_FL) {
4768                         /* i_blocks represent file system block size */
4769                         return i_blocks  << (inode->i_blkbits - 9);
4770                 } else {
4771                         return i_blocks;
4772                 }
4773         } else {
4774                 return le32_to_cpu(raw_inode->i_blocks_lo);
4775         }
4776 }
4777
4778 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4779 {
4780         struct ext4_iloc iloc;
4781         struct ext4_inode *raw_inode;
4782         struct ext4_inode_info *ei;
4783         struct buffer_head *bh;
4784         struct inode *inode;
4785         long ret;
4786         int block;
4787
4788         inode = iget_locked(sb, ino);
4789         if (!inode)
4790                 return ERR_PTR(-ENOMEM);
4791         if (!(inode->i_state & I_NEW))
4792                 return inode;
4793
4794         ei = EXT4_I(inode);
4795
4796         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4797         if (ret < 0)
4798                 goto bad_inode;
4799         bh = iloc.bh;
4800         raw_inode = ext4_raw_inode(&iloc);
4801         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4802         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4803         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4804         if (!(test_opt(inode->i_sb, NO_UID32))) {
4805                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4806                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4807         }
4808         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
4809
4810         ei->i_state = 0;
4811         ei->i_dir_start_lookup = 0;
4812         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4813         /* We now have enough fields to check if the inode was active or not.
4814          * This is needed because nfsd might try to access dead inodes
4815          * the test is that same one that e2fsck uses
4816          * NeilBrown 1999oct15
4817          */
4818         if (inode->i_nlink == 0) {
4819                 if (inode->i_mode == 0 ||
4820                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
4821                         /* this inode is deleted */
4822                         brelse(bh);
4823                         ret = -ESTALE;
4824                         goto bad_inode;
4825                 }
4826                 /* The only unlinked inodes we let through here have
4827                  * valid i_mode and are being read by the orphan
4828                  * recovery code: that's fine, we're about to complete
4829                  * the process of deleting those. */
4830         }
4831         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4832         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4833         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4834         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
4835                 ei->i_file_acl |=
4836                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4837         inode->i_size = ext4_isize(raw_inode);
4838         ei->i_disksize = inode->i_size;
4839         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4840         ei->i_block_group = iloc.block_group;
4841         ei->i_last_alloc_group = ~0;
4842         /*
4843          * NOTE! The in-memory inode i_data array is in little-endian order
4844          * even on big-endian machines: we do NOT byteswap the block numbers!
4845          */
4846         for (block = 0; block < EXT4_N_BLOCKS; block++)
4847                 ei->i_data[block] = raw_inode->i_block[block];
4848         INIT_LIST_HEAD(&ei->i_orphan);
4849
4850         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4851                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4852                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4853                     EXT4_INODE_SIZE(inode->i_sb)) {
4854                         brelse(bh);
4855                         ret = -EIO;
4856                         goto bad_inode;
4857                 }
4858                 if (ei->i_extra_isize == 0) {
4859                         /* The extra space is currently unused. Use it. */
4860                         ei->i_extra_isize = sizeof(struct ext4_inode) -
4861                                             EXT4_GOOD_OLD_INODE_SIZE;
4862                 } else {
4863                         __le32 *magic = (void *)raw_inode +
4864                                         EXT4_GOOD_OLD_INODE_SIZE +
4865                                         ei->i_extra_isize;
4866                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
4867                                 ei->i_state |= EXT4_STATE_XATTR;
4868                 }
4869         } else
4870                 ei->i_extra_isize = 0;
4871
4872         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4873         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4874         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4875         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4876
4877         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4878         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4879                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4880                         inode->i_version |=
4881                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4882         }
4883
4884         ret = 0;
4885         if (ei->i_file_acl &&
4886             ((ei->i_file_acl <
4887               (le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block) +
4888                EXT4_SB(sb)->s_gdb_count)) ||
4889              (ei->i_file_acl >= ext4_blocks_count(EXT4_SB(sb)->s_es)))) {
4890                 ext4_error(sb, __func__,
4891                            "bad extended attribute block %llu in inode #%lu",
4892                            ei->i_file_acl, inode->i_ino);
4893                 ret = -EIO;
4894                 goto bad_inode;
4895         } else if (ei->i_flags & EXT4_EXTENTS_FL) {
4896                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4897                     (S_ISLNK(inode->i_mode) &&
4898                      !ext4_inode_is_fast_symlink(inode)))
4899                         /* Validate extent which is part of inode */
4900                         ret = ext4_ext_check_inode(inode);
4901         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4902                    (S_ISLNK(inode->i_mode) &&
4903                     !ext4_inode_is_fast_symlink(inode))) {
4904                 /* Validate block references which are part of inode */
4905                 ret = ext4_check_inode_blockref(inode);
4906         }
4907         if (ret) {
4908                 brelse(bh);
4909                 goto bad_inode;
4910         }
4911
4912         if (S_ISREG(inode->i_mode)) {
4913                 inode->i_op = &ext4_file_inode_operations;
4914                 inode->i_fop = &ext4_file_operations;
4915                 ext4_set_aops(inode);
4916         } else if (S_ISDIR(inode->i_mode)) {
4917                 inode->i_op = &ext4_dir_inode_operations;
4918                 inode->i_fop = &ext4_dir_operations;
4919         } else if (S_ISLNK(inode->i_mode)) {
4920                 if (ext4_inode_is_fast_symlink(inode)) {
4921                         inode->i_op = &ext4_fast_symlink_inode_operations;
4922                         nd_terminate_link(ei->i_data, inode->i_size,
4923                                 sizeof(ei->i_data) - 1);
4924                 } else {
4925                         inode->i_op = &ext4_symlink_inode_operations;
4926                         ext4_set_aops(inode);
4927                 }
4928         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
4929               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
4930                 inode->i_op = &ext4_special_inode_operations;
4931                 if (raw_inode->i_block[0])
4932                         init_special_inode(inode, inode->i_mode,
4933                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4934                 else
4935                         init_special_inode(inode, inode->i_mode,
4936                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4937         } else {
4938                 brelse(bh);
4939                 ret = -EIO;
4940                 ext4_error(inode->i_sb, __func__,
4941                            "bogus i_mode (%o) for inode=%lu",
4942                            inode->i_mode, inode->i_ino);
4943                 goto bad_inode;
4944         }
4945         brelse(iloc.bh);
4946         ext4_set_inode_flags(inode);
4947         unlock_new_inode(inode);
4948         return inode;
4949
4950 bad_inode:
4951         iget_failed(inode);
4952         return ERR_PTR(ret);
4953 }
4954
4955 static int ext4_inode_blocks_set(handle_t *handle,
4956                                 struct ext4_inode *raw_inode,
4957                                 struct ext4_inode_info *ei)
4958 {
4959         struct inode *inode = &(ei->vfs_inode);
4960         u64 i_blocks = inode->i_blocks;
4961         struct super_block *sb = inode->i_sb;
4962
4963         if (i_blocks <= ~0U) {
4964                 /*
4965                  * i_blocks can be represnted in a 32 bit variable
4966                  * as multiple of 512 bytes
4967                  */
4968                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4969                 raw_inode->i_blocks_high = 0;
4970                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4971                 return 0;
4972         }
4973         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
4974                 return -EFBIG;
4975
4976         if (i_blocks <= 0xffffffffffffULL) {
4977                 /*
4978                  * i_blocks can be represented in a 48 bit variable
4979                  * as multiple of 512 bytes
4980                  */
4981                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4982                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4983                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4984         } else {
4985                 ei->i_flags |= EXT4_HUGE_FILE_FL;
4986                 /* i_block is stored in file system block size */
4987                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4988                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4989                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4990         }
4991         return 0;
4992 }
4993
4994 /*
4995  * Post the struct inode info into an on-disk inode location in the
4996  * buffer-cache.  This gobbles the caller's reference to the
4997  * buffer_head in the inode location struct.
4998  *
4999  * The caller must have write access to iloc->bh.
5000  */
5001 static int ext4_do_update_inode(handle_t *handle,
5002                                 struct inode *inode,
5003                                 struct ext4_iloc *iloc)
5004 {
5005         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5006         struct ext4_inode_info *ei = EXT4_I(inode);
5007         struct buffer_head *bh = iloc->bh;
5008         int err = 0, rc, block;
5009
5010         /* For fields not not tracking in the in-memory inode,
5011          * initialise them to zero for new inodes. */
5012         if (ei->i_state & EXT4_STATE_NEW)
5013                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5014
5015         ext4_get_inode_flags(ei);
5016         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5017         if (!(test_opt(inode->i_sb, NO_UID32))) {
5018                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
5019                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
5020 /*
5021  * Fix up interoperability with old kernels. Otherwise, old inodes get
5022  * re-used with the upper 16 bits of the uid/gid intact
5023  */
5024                 if (!ei->i_dtime) {
5025                         raw_inode->i_uid_high =
5026                                 cpu_to_le16(high_16_bits(inode->i_uid));
5027                         raw_inode->i_gid_high =
5028                                 cpu_to_le16(high_16_bits(inode->i_gid));
5029                 } else {
5030                         raw_inode->i_uid_high = 0;
5031                         raw_inode->i_gid_high = 0;
5032                 }
5033         } else {
5034                 raw_inode->i_uid_low =
5035                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
5036                 raw_inode->i_gid_low =
5037                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
5038                 raw_inode->i_uid_high = 0;
5039                 raw_inode->i_gid_high = 0;
5040         }
5041         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5042
5043         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5044         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5045         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5046         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5047
5048         if (ext4_inode_blocks_set(handle, raw_inode, ei))
5049                 goto out_brelse;
5050         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5051         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
5052         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
5053             cpu_to_le32(EXT4_OS_HURD))
5054                 raw_inode->i_file_acl_high =
5055                         cpu_to_le16(ei->i_file_acl >> 32);
5056         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5057         ext4_isize_set(raw_inode, ei->i_disksize);
5058         if (ei->i_disksize > 0x7fffffffULL) {
5059                 struct super_block *sb = inode->i_sb;
5060                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
5061                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
5062                                 EXT4_SB(sb)->s_es->s_rev_level ==
5063                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
5064                         /* If this is the first large file
5065                          * created, add a flag to the superblock.
5066                          */
5067                         err = ext4_journal_get_write_access(handle,
5068                                         EXT4_SB(sb)->s_sbh);
5069                         if (err)
5070                                 goto out_brelse;
5071                         ext4_update_dynamic_rev(sb);
5072                         EXT4_SET_RO_COMPAT_FEATURE(sb,
5073                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
5074                         sb->s_dirt = 1;
5075                         ext4_handle_sync(handle);
5076                         err = ext4_handle_dirty_metadata(handle, inode,
5077                                         EXT4_SB(sb)->s_sbh);
5078                 }
5079         }
5080         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5081         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5082                 if (old_valid_dev(inode->i_rdev)) {
5083                         raw_inode->i_block[0] =
5084                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
5085                         raw_inode->i_block[1] = 0;
5086                 } else {
5087                         raw_inode->i_block[0] = 0;
5088                         raw_inode->i_block[1] =
5089                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
5090                         raw_inode->i_block[2] = 0;
5091                 }
5092         } else
5093                 for (block = 0; block < EXT4_N_BLOCKS; block++)
5094                         raw_inode->i_block[block] = ei->i_data[block];
5095
5096         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
5097         if (ei->i_extra_isize) {
5098                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5099                         raw_inode->i_version_hi =
5100                         cpu_to_le32(inode->i_version >> 32);
5101                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
5102         }
5103
5104         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5105         rc = ext4_handle_dirty_metadata(handle, inode, bh);
5106         if (!err)
5107                 err = rc;
5108         ei->i_state &= ~EXT4_STATE_NEW;
5109
5110 out_brelse:
5111         brelse(bh);
5112         ext4_std_error(inode->i_sb, err);
5113         return err;
5114 }
5115
5116 /*
5117  * ext4_write_inode()
5118  *
5119  * We are called from a few places:
5120  *
5121  * - Within generic_file_write() for O_SYNC files.
5122  *   Here, there will be no transaction running. We wait for any running
5123  *   trasnaction to commit.
5124  *
5125  * - Within sys_sync(), kupdate and such.
5126  *   We wait on commit, if tol to.
5127  *
5128  * - Within prune_icache() (PF_MEMALLOC == true)
5129  *   Here we simply return.  We can't afford to block kswapd on the
5130  *   journal commit.
5131  *
5132  * In all cases it is actually safe for us to return without doing anything,
5133  * because the inode has been copied into a raw inode buffer in
5134  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
5135  * knfsd.
5136  *
5137  * Note that we are absolutely dependent upon all inode dirtiers doing the
5138  * right thing: they *must* call mark_inode_dirty() after dirtying info in
5139  * which we are interested.
5140  *
5141  * It would be a bug for them to not do this.  The code:
5142  *
5143  *      mark_inode_dirty(inode)
5144  *      stuff();
5145  *      inode->i_size = expr;
5146  *
5147  * is in error because a kswapd-driven write_inode() could occur while
5148  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
5149  * will no longer be on the superblock's dirty inode list.
5150  */
5151 int ext4_write_inode(struct inode *inode, int wait)
5152 {
5153         int err;
5154
5155         if (current->flags & PF_MEMALLOC)
5156                 return 0;
5157
5158         if (EXT4_SB(inode->i_sb)->s_journal) {
5159                 if (ext4_journal_current_handle()) {
5160                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5161                         dump_stack();
5162                         return -EIO;
5163                 }
5164
5165                 if (!wait)
5166                         return 0;
5167
5168                 err = ext4_force_commit(inode->i_sb);
5169         } else {
5170                 struct ext4_iloc iloc;
5171
5172                 err = ext4_get_inode_loc(inode, &iloc);
5173                 if (err)
5174                         return err;
5175                 if (wait)
5176                         sync_dirty_buffer(iloc.bh);
5177                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5178                         ext4_error(inode->i_sb, __func__,
5179                                    "IO error syncing inode, "
5180                                    "inode=%lu, block=%llu",
5181                                    inode->i_ino,
5182                                    (unsigned long long)iloc.bh->b_blocknr);
5183                         err = -EIO;
5184                 }
5185         }
5186         return err;
5187 }
5188
5189 /*
5190  * ext4_setattr()
5191  *
5192  * Called from notify_change.
5193  *
5194  * We want to trap VFS attempts to truncate the file as soon as
5195  * possible.  In particular, we want to make sure that when the VFS
5196  * shrinks i_size, we put the inode on the orphan list and modify
5197  * i_disksize immediately, so that during the subsequent flushing of
5198  * dirty pages and freeing of disk blocks, we can guarantee that any
5199  * commit will leave the blocks being flushed in an unused state on
5200  * disk.  (On recovery, the inode will get truncated and the blocks will
5201  * be freed, so we have a strong guarantee that no future commit will
5202  * leave these blocks visible to the user.)
5203  *
5204  * Another thing we have to assure is that if we are in ordered mode
5205  * and inode is still attached to the committing transaction, we must
5206  * we start writeout of all the dirty pages which are being truncated.
5207  * This way we are sure that all the data written in the previous
5208  * transaction are already on disk (truncate waits for pages under
5209  * writeback).
5210  *
5211  * Called with inode->i_mutex down.
5212  */
5213 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5214 {
5215         struct inode *inode = dentry->d_inode;
5216         int error, rc = 0;
5217         const unsigned int ia_valid = attr->ia_valid;
5218
5219         error = inode_change_ok(inode, attr);
5220         if (error)
5221                 return error;
5222
5223         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
5224                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
5225                 handle_t *handle;
5226
5227                 /* (user+group)*(old+new) structure, inode write (sb,
5228                  * inode block, ? - but truncate inode update has it) */
5229                 handle = ext4_journal_start(inode, 2*(EXT4_QUOTA_INIT_BLOCKS(inode->i_sb)+
5230                                         EXT4_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
5231                 if (IS_ERR(handle)) {
5232                         error = PTR_ERR(handle);
5233                         goto err_out;
5234                 }
5235                 error = vfs_dq_transfer(inode, attr) ? -EDQUOT : 0;
5236                 if (error) {
5237                         ext4_journal_stop(handle);
5238                         return error;
5239                 }
5240                 /* Update corresponding info in inode so that everything is in
5241                  * one transaction */
5242                 if (attr->ia_valid & ATTR_UID)
5243                         inode->i_uid = attr->ia_uid;
5244                 if (attr->ia_valid & ATTR_GID)
5245                         inode->i_gid = attr->ia_gid;
5246                 error = ext4_mark_inode_dirty(handle, inode);
5247                 ext4_journal_stop(handle);
5248         }
5249
5250         if (attr->ia_valid & ATTR_SIZE) {
5251                 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)) {
5252                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5253
5254                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
5255                                 error = -EFBIG;
5256                                 goto err_out;
5257                         }
5258                 }
5259         }
5260
5261         if (S_ISREG(inode->i_mode) &&
5262             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
5263                 handle_t *handle;
5264
5265                 handle = ext4_journal_start(inode, 3);
5266                 if (IS_ERR(handle)) {
5267                         error = PTR_ERR(handle);
5268                         goto err_out;
5269                 }
5270
5271                 error = ext4_orphan_add(handle, inode);
5272                 EXT4_I(inode)->i_disksize = attr->ia_size;
5273                 rc = ext4_mark_inode_dirty(handle, inode);
5274                 if (!error)
5275                         error = rc;
5276                 ext4_journal_stop(handle);
5277
5278                 if (ext4_should_order_data(inode)) {
5279                         error = ext4_begin_ordered_truncate(inode,
5280                                                             attr->ia_size);
5281                         if (error) {
5282                                 /* Do as much error cleanup as possible */
5283                                 handle = ext4_journal_start(inode, 3);
5284                                 if (IS_ERR(handle)) {
5285                                         ext4_orphan_del(NULL, inode);
5286                                         goto err_out;
5287                                 }
5288                                 ext4_orphan_del(handle, inode);
5289                                 ext4_journal_stop(handle);
5290                                 goto err_out;
5291                         }
5292                 }
5293         }
5294
5295         rc = inode_setattr(inode, attr);
5296
5297         /* If inode_setattr's call to ext4_truncate failed to get a
5298          * transaction handle at all, we need to clean up the in-core
5299          * orphan list manually. */
5300         if (inode->i_nlink)
5301                 ext4_orphan_del(NULL, inode);
5302
5303         if (!rc && (ia_valid & ATTR_MODE))
5304                 rc = ext4_acl_chmod(inode);
5305
5306 err_out:
5307         ext4_std_error(inode->i_sb, error);
5308         if (!error)
5309                 error = rc;
5310         return error;
5311 }
5312
5313 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
5314                  struct kstat *stat)
5315 {
5316         struct inode *inode;
5317         unsigned long delalloc_blocks;
5318
5319         inode = dentry->d_inode;
5320         generic_fillattr(inode, stat);
5321
5322         /*
5323          * We can't update i_blocks if the block allocation is delayed
5324          * otherwise in the case of system crash before the real block
5325          * allocation is done, we will have i_blocks inconsistent with
5326          * on-disk file blocks.
5327          * We always keep i_blocks updated together with real
5328          * allocation. But to not confuse with user, stat
5329          * will return the blocks that include the delayed allocation
5330          * blocks for this file.
5331          */
5332         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
5333         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
5334         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
5335
5336         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
5337         return 0;
5338 }
5339
5340 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
5341                                       int chunk)
5342 {
5343         int indirects;
5344
5345         /* if nrblocks are contiguous */
5346         if (chunk) {
5347                 /*
5348                  * With N contiguous data blocks, it need at most
5349                  * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) indirect blocks
5350                  * 2 dindirect blocks
5351                  * 1 tindirect block
5352                  */
5353                 indirects = nrblocks / EXT4_ADDR_PER_BLOCK(inode->i_sb);
5354                 return indirects + 3;
5355         }
5356         /*
5357          * if nrblocks are not contiguous, worse case, each block touch
5358          * a indirect block, and each indirect block touch a double indirect
5359          * block, plus a triple indirect block
5360          */
5361         indirects = nrblocks * 2 + 1;
5362         return indirects;
5363 }
5364
5365 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5366 {
5367         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL))
5368                 return ext4_indirect_trans_blocks(inode, nrblocks, chunk);
5369         return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
5370 }
5371
5372 /*
5373  * Account for index blocks, block groups bitmaps and block group
5374  * descriptor blocks if modify datablocks and index blocks
5375  * worse case, the indexs blocks spread over different block groups
5376  *
5377  * If datablocks are discontiguous, they are possible to spread over
5378  * different block groups too. If they are contiugous, with flexbg,
5379  * they could still across block group boundary.
5380  *
5381  * Also account for superblock, inode, quota and xattr blocks
5382  */
5383 int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5384 {
5385         ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
5386         int gdpblocks;
5387         int idxblocks;
5388         int ret = 0;
5389
5390         /*
5391          * How many index blocks need to touch to modify nrblocks?
5392          * The "Chunk" flag indicating whether the nrblocks is
5393          * physically contiguous on disk
5394          *
5395          * For Direct IO and fallocate, they calls get_block to allocate
5396          * one single extent at a time, so they could set the "Chunk" flag
5397          */
5398         idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
5399
5400         ret = idxblocks;
5401
5402         /*
5403          * Now let's see how many group bitmaps and group descriptors need
5404          * to account
5405          */
5406         groups = idxblocks;
5407         if (chunk)
5408                 groups += 1;
5409         else
5410                 groups += nrblocks;
5411
5412         gdpblocks = groups;
5413         if (groups > ngroups)
5414                 groups = ngroups;
5415         if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
5416                 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
5417
5418         /* bitmaps and block group descriptor blocks */
5419         ret += groups + gdpblocks;
5420
5421         /* Blocks for super block, inode, quota and xattr blocks */
5422         ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
5423
5424         return ret;
5425 }
5426
5427 /*
5428  * Calulate the total number of credits to reserve to fit
5429  * the modification of a single pages into a single transaction,
5430  * which may include multiple chunks of block allocations.
5431  *
5432  * This could be called via ext4_write_begin()
5433  *
5434  * We need to consider the worse case, when
5435  * one new block per extent.
5436  */
5437 int ext4_writepage_trans_blocks(struct inode *inode)
5438 {
5439         int bpp = ext4_journal_blocks_per_page(inode);
5440         int ret;
5441
5442         ret = ext4_meta_trans_blocks(inode, bpp, 0);
5443
5444         /* Account for data blocks for journalled mode */
5445         if (ext4_should_journal_data(inode))
5446                 ret += bpp;
5447         return ret;
5448 }
5449
5450 /*
5451  * Calculate the journal credits for a chunk of data modification.
5452  *
5453  * This is called from DIO, fallocate or whoever calling
5454  * ext4_get_blocks() to map/allocate a chunk of contigous disk blocks.
5455  *
5456  * journal buffers for data blocks are not included here, as DIO
5457  * and fallocate do no need to journal data buffers.
5458  */
5459 int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
5460 {
5461         return ext4_meta_trans_blocks(inode, nrblocks, 1);
5462 }
5463
5464 /*
5465  * The caller must have previously called ext4_reserve_inode_write().
5466  * Give this, we know that the caller already has write access to iloc->bh.
5467  */
5468 int ext4_mark_iloc_dirty(handle_t *handle,
5469                          struct inode *inode, struct ext4_iloc *iloc)
5470 {
5471         int err = 0;
5472
5473         if (test_opt(inode->i_sb, I_VERSION))
5474                 inode_inc_iversion(inode);
5475
5476         /* the do_update_inode consumes one bh->b_count */
5477         get_bh(iloc->bh);
5478
5479         /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
5480         err = ext4_do_update_inode(handle, inode, iloc);
5481         put_bh(iloc->bh);
5482         return err;
5483 }
5484
5485 /*
5486  * On success, We end up with an outstanding reference count against
5487  * iloc->bh.  This _must_ be cleaned up later.
5488  */
5489
5490 int
5491 ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
5492                          struct ext4_iloc *iloc)
5493 {
5494         int err;
5495
5496         err = ext4_get_inode_loc(inode, iloc);
5497         if (!err) {
5498                 BUFFER_TRACE(iloc->bh, "get_write_access");
5499                 err = ext4_journal_get_write_access(handle, iloc->bh);
5500                 if (err) {
5501                         brelse(iloc->bh);
5502                         iloc->bh = NULL;
5503                 }
5504         }
5505         ext4_std_error(inode->i_sb, err);
5506         return err;
5507 }
5508
5509 /*
5510  * Expand an inode by new_extra_isize bytes.
5511  * Returns 0 on success or negative error number on failure.
5512  */
5513 static int ext4_expand_extra_isize(struct inode *inode,
5514                                    unsigned int new_extra_isize,
5515                                    struct ext4_iloc iloc,
5516                                    handle_t *handle)
5517 {
5518         struct ext4_inode *raw_inode;
5519         struct ext4_xattr_ibody_header *header;
5520         struct ext4_xattr_entry *entry;
5521
5522         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
5523                 return 0;
5524
5525         raw_inode = ext4_raw_inode(&iloc);
5526
5527         header = IHDR(inode, raw_inode);
5528         entry = IFIRST(header);
5529
5530         /* No extended attributes present */
5531         if (!(EXT4_I(inode)->i_state & EXT4_STATE_XATTR) ||
5532                 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
5533                 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
5534                         new_extra_isize);
5535                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
5536                 return 0;
5537         }
5538
5539         /* try to expand with EAs present */
5540         return ext4_expand_extra_isize_ea(inode, new_extra_isize,
5541                                           raw_inode, handle);
5542 }
5543
5544 /*
5545  * What we do here is to mark the in-core inode as clean with respect to inode
5546  * dirtiness (it may still be data-dirty).
5547  * This means that the in-core inode may be reaped by prune_icache
5548  * without having to perform any I/O.  This is a very good thing,
5549  * because *any* task may call prune_icache - even ones which
5550  * have a transaction open against a different journal.
5551  *
5552  * Is this cheating?  Not really.  Sure, we haven't written the
5553  * inode out, but prune_icache isn't a user-visible syncing function.
5554  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
5555  * we start and wait on commits.
5556  *
5557  * Is this efficient/effective?  Well, we're being nice to the system
5558  * by cleaning up our inodes proactively so they can be reaped
5559  * without I/O.  But we are potentially leaving up to five seconds'
5560  * worth of inodes floating about which prune_icache wants us to
5561  * write out.  One way to fix that would be to get prune_icache()
5562  * to do a write_super() to free up some memory.  It has the desired
5563  * effect.
5564  */
5565 int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
5566 {
5567         struct ext4_iloc iloc;
5568         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5569         static unsigned int mnt_count;
5570         int err, ret;
5571
5572         might_sleep();
5573         err = ext4_reserve_inode_write(handle, inode, &iloc);
5574         if (ext4_handle_valid(handle) &&
5575             EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
5576             !(EXT4_I(inode)->i_state & EXT4_STATE_NO_EXPAND)) {
5577                 /*
5578                  * We need extra buffer credits since we may write into EA block
5579                  * with this same handle. If journal_extend fails, then it will
5580                  * only result in a minor loss of functionality for that inode.
5581                  * If this is felt to be critical, then e2fsck should be run to
5582                  * force a large enough s_min_extra_isize.
5583                  */
5584                 if ((jbd2_journal_extend(handle,
5585                              EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
5586                         ret = ext4_expand_extra_isize(inode,
5587                                                       sbi->s_want_extra_isize,
5588                                                       iloc, handle);
5589                         if (ret) {
5590                                 EXT4_I(inode)->i_state |= EXT4_STATE_NO_EXPAND;
5591                                 if (mnt_count !=
5592                                         le16_to_cpu(sbi->s_es->s_mnt_count)) {
5593                                         ext4_warning(inode->i_sb, __func__,
5594                                         "Unable to expand inode %lu. Delete"
5595                                         " some EAs or run e2fsck.",
5596                                         inode->i_ino);