fs: pass exact type of data dirties to ->dirty_inode
[linux-2.6.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/jbd2.h>
29 #include <linux/highuid.h>
30 #include <linux/pagemap.h>
31 #include <linux/quotaops.h>
32 #include <linux/string.h>
33 #include <linux/buffer_head.h>
34 #include <linux/writeback.h>
35 #include <linux/pagevec.h>
36 #include <linux/mpage.h>
37 #include <linux/namei.h>
38 #include <linux/uio.h>
39 #include <linux/bio.h>
40 #include <linux/workqueue.h>
41 #include <linux/kernel.h>
42 #include <linux/printk.h>
43 #include <linux/slab.h>
44 #include <linux/ratelimit.h>
45
46 #include "ext4_jbd2.h"
47 #include "xattr.h"
48 #include "acl.h"
49 #include "ext4_extents.h"
50
51 #include <trace/events/ext4.h>
52
53 #define MPAGE_DA_EXTENT_TAIL 0x01
54
55 static inline int ext4_begin_ordered_truncate(struct inode *inode,
56                                               loff_t new_size)
57 {
58         trace_ext4_begin_ordered_truncate(inode, new_size);
59         /*
60          * If jinode is zero, then we never opened the file for
61          * writing, so there's no need to call
62          * jbd2_journal_begin_ordered_truncate() since there's no
63          * outstanding writes we need to flush.
64          */
65         if (!EXT4_I(inode)->jinode)
66                 return 0;
67         return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
68                                                    EXT4_I(inode)->jinode,
69                                                    new_size);
70 }
71
72 static void ext4_invalidatepage(struct page *page, unsigned long offset);
73 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
74                                    struct buffer_head *bh_result, int create);
75 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
76 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
77 static int __ext4_journalled_writepage(struct page *page, unsigned int len);
78 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
79
80 /*
81  * Test whether an inode is a fast symlink.
82  */
83 static int ext4_inode_is_fast_symlink(struct inode *inode)
84 {
85         int ea_blocks = EXT4_I(inode)->i_file_acl ?
86                 (inode->i_sb->s_blocksize >> 9) : 0;
87
88         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
89 }
90
91 /*
92  * Work out how many blocks we need to proceed with the next chunk of a
93  * truncate transaction.
94  */
95 static unsigned long blocks_for_truncate(struct inode *inode)
96 {
97         ext4_lblk_t needed;
98
99         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
100
101         /* Give ourselves just enough room to cope with inodes in which
102          * i_blocks is corrupt: we've seen disk corruptions in the past
103          * which resulted in random data in an inode which looked enough
104          * like a regular file for ext4 to try to delete it.  Things
105          * will go a bit crazy if that happens, but at least we should
106          * try not to panic the whole kernel. */
107         if (needed < 2)
108                 needed = 2;
109
110         /* But we need to bound the transaction so we don't overflow the
111          * journal. */
112         if (needed > EXT4_MAX_TRANS_DATA)
113                 needed = EXT4_MAX_TRANS_DATA;
114
115         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
116 }
117
118 /*
119  * Truncate transactions can be complex and absolutely huge.  So we need to
120  * be able to restart the transaction at a conventient checkpoint to make
121  * sure we don't overflow the journal.
122  *
123  * start_transaction gets us a new handle for a truncate transaction,
124  * and extend_transaction tries to extend the existing one a bit.  If
125  * extend fails, we need to propagate the failure up and restart the
126  * transaction in the top-level truncate loop. --sct
127  */
128 static handle_t *start_transaction(struct inode *inode)
129 {
130         handle_t *result;
131
132         result = ext4_journal_start(inode, blocks_for_truncate(inode));
133         if (!IS_ERR(result))
134                 return result;
135
136         ext4_std_error(inode->i_sb, PTR_ERR(result));
137         return result;
138 }
139
140 /*
141  * Try to extend this transaction for the purposes of truncation.
142  *
143  * Returns 0 if we managed to create more room.  If we can't create more
144  * room, and the transaction must be restarted we return 1.
145  */
146 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
147 {
148         if (!ext4_handle_valid(handle))
149                 return 0;
150         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
151                 return 0;
152         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
153                 return 0;
154         return 1;
155 }
156
157 /*
158  * Restart the transaction associated with *handle.  This does a commit,
159  * so before we call here everything must be consistently dirtied against
160  * this transaction.
161  */
162 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
163                                  int nblocks)
164 {
165         int ret;
166
167         /*
168          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
169          * moment, get_block can be called only for blocks inside i_size since
170          * page cache has been already dropped and writes are blocked by
171          * i_mutex. So we can safely drop the i_data_sem here.
172          */
173         BUG_ON(EXT4_JOURNAL(inode) == NULL);
174         jbd_debug(2, "restarting handle %p\n", handle);
175         up_write(&EXT4_I(inode)->i_data_sem);
176         ret = ext4_journal_restart(handle, nblocks);
177         down_write(&EXT4_I(inode)->i_data_sem);
178         ext4_discard_preallocations(inode);
179
180         return ret;
181 }
182
183 /*
184  * Called at the last iput() if i_nlink is zero.
185  */
186 void ext4_evict_inode(struct inode *inode)
187 {
188         handle_t *handle;
189         int err;
190
191         trace_ext4_evict_inode(inode);
192         if (inode->i_nlink) {
193                 truncate_inode_pages(&inode->i_data, 0);
194                 goto no_delete;
195         }
196
197         if (!is_bad_inode(inode))
198                 dquot_initialize(inode);
199
200         if (ext4_should_order_data(inode))
201                 ext4_begin_ordered_truncate(inode, 0);
202         truncate_inode_pages(&inode->i_data, 0);
203
204         if (is_bad_inode(inode))
205                 goto no_delete;
206
207         handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
208         if (IS_ERR(handle)) {
209                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
210                 /*
211                  * If we're going to skip the normal cleanup, we still need to
212                  * make sure that the in-core orphan linked list is properly
213                  * cleaned up.
214                  */
215                 ext4_orphan_del(NULL, inode);
216                 goto no_delete;
217         }
218
219         if (IS_SYNC(inode))
220                 ext4_handle_sync(handle);
221         inode->i_size = 0;
222         err = ext4_mark_inode_dirty(handle, inode);
223         if (err) {
224                 ext4_warning(inode->i_sb,
225                              "couldn't mark inode dirty (err %d)", err);
226                 goto stop_handle;
227         }
228         if (inode->i_blocks)
229                 ext4_truncate(inode);
230
231         /*
232          * ext4_ext_truncate() doesn't reserve any slop when it
233          * restarts journal transactions; therefore there may not be
234          * enough credits left in the handle to remove the inode from
235          * the orphan list and set the dtime field.
236          */
237         if (!ext4_handle_has_enough_credits(handle, 3)) {
238                 err = ext4_journal_extend(handle, 3);
239                 if (err > 0)
240                         err = ext4_journal_restart(handle, 3);
241                 if (err != 0) {
242                         ext4_warning(inode->i_sb,
243                                      "couldn't extend journal (err %d)", err);
244                 stop_handle:
245                         ext4_journal_stop(handle);
246                         ext4_orphan_del(NULL, inode);
247                         goto no_delete;
248                 }
249         }
250
251         /*
252          * Kill off the orphan record which ext4_truncate created.
253          * AKPM: I think this can be inside the above `if'.
254          * Note that ext4_orphan_del() has to be able to cope with the
255          * deletion of a non-existent orphan - this is because we don't
256          * know if ext4_truncate() actually created an orphan record.
257          * (Well, we could do this if we need to, but heck - it works)
258          */
259         ext4_orphan_del(handle, inode);
260         EXT4_I(inode)->i_dtime  = get_seconds();
261
262         /*
263          * One subtle ordering requirement: if anything has gone wrong
264          * (transaction abort, IO errors, whatever), then we can still
265          * do these next steps (the fs will already have been marked as
266          * having errors), but we can't free the inode if the mark_dirty
267          * fails.
268          */
269         if (ext4_mark_inode_dirty(handle, inode))
270                 /* If that failed, just do the required in-core inode clear. */
271                 ext4_clear_inode(inode);
272         else
273                 ext4_free_inode(handle, inode);
274         ext4_journal_stop(handle);
275         return;
276 no_delete:
277         ext4_clear_inode(inode);        /* We must guarantee clearing of inode... */
278 }
279
280 typedef struct {
281         __le32  *p;
282         __le32  key;
283         struct buffer_head *bh;
284 } Indirect;
285
286 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
287 {
288         p->key = *(p->p = v);
289         p->bh = bh;
290 }
291
292 /**
293  *      ext4_block_to_path - parse the block number into array of offsets
294  *      @inode: inode in question (we are only interested in its superblock)
295  *      @i_block: block number to be parsed
296  *      @offsets: array to store the offsets in
297  *      @boundary: set this non-zero if the referred-to block is likely to be
298  *             followed (on disk) by an indirect block.
299  *
300  *      To store the locations of file's data ext4 uses a data structure common
301  *      for UNIX filesystems - tree of pointers anchored in the inode, with
302  *      data blocks at leaves and indirect blocks in intermediate nodes.
303  *      This function translates the block number into path in that tree -
304  *      return value is the path length and @offsets[n] is the offset of
305  *      pointer to (n+1)th node in the nth one. If @block is out of range
306  *      (negative or too large) warning is printed and zero returned.
307  *
308  *      Note: function doesn't find node addresses, so no IO is needed. All
309  *      we need to know is the capacity of indirect blocks (taken from the
310  *      inode->i_sb).
311  */
312
313 /*
314  * Portability note: the last comparison (check that we fit into triple
315  * indirect block) is spelled differently, because otherwise on an
316  * architecture with 32-bit longs and 8Kb pages we might get into trouble
317  * if our filesystem had 8Kb blocks. We might use long long, but that would
318  * kill us on x86. Oh, well, at least the sign propagation does not matter -
319  * i_block would have to be negative in the very beginning, so we would not
320  * get there at all.
321  */
322
323 static int ext4_block_to_path(struct inode *inode,
324                               ext4_lblk_t i_block,
325                               ext4_lblk_t offsets[4], int *boundary)
326 {
327         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
328         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
329         const long direct_blocks = EXT4_NDIR_BLOCKS,
330                 indirect_blocks = ptrs,
331                 double_blocks = (1 << (ptrs_bits * 2));
332         int n = 0;
333         int final = 0;
334
335         if (i_block < direct_blocks) {
336                 offsets[n++] = i_block;
337                 final = direct_blocks;
338         } else if ((i_block -= direct_blocks) < indirect_blocks) {
339                 offsets[n++] = EXT4_IND_BLOCK;
340                 offsets[n++] = i_block;
341                 final = ptrs;
342         } else if ((i_block -= indirect_blocks) < double_blocks) {
343                 offsets[n++] = EXT4_DIND_BLOCK;
344                 offsets[n++] = i_block >> ptrs_bits;
345                 offsets[n++] = i_block & (ptrs - 1);
346                 final = ptrs;
347         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
348                 offsets[n++] = EXT4_TIND_BLOCK;
349                 offsets[n++] = i_block >> (ptrs_bits * 2);
350                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
351                 offsets[n++] = i_block & (ptrs - 1);
352                 final = ptrs;
353         } else {
354                 ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
355                              i_block + direct_blocks +
356                              indirect_blocks + double_blocks, inode->i_ino);
357         }
358         if (boundary)
359                 *boundary = final - 1 - (i_block & (ptrs - 1));
360         return n;
361 }
362
363 static int __ext4_check_blockref(const char *function, unsigned int line,
364                                  struct inode *inode,
365                                  __le32 *p, unsigned int max)
366 {
367         struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
368         __le32 *bref = p;
369         unsigned int blk;
370
371         while (bref < p+max) {
372                 blk = le32_to_cpu(*bref++);
373                 if (blk &&
374                     unlikely(!ext4_data_block_valid(EXT4_SB(inode->i_sb),
375                                                     blk, 1))) {
376                         es->s_last_error_block = cpu_to_le64(blk);
377                         ext4_error_inode(inode, function, line, blk,
378                                          "invalid block");
379                         return -EIO;
380                 }
381         }
382         return 0;
383 }
384
385
386 #define ext4_check_indirect_blockref(inode, bh)                         \
387         __ext4_check_blockref(__func__, __LINE__, inode,                \
388                               (__le32 *)(bh)->b_data,                   \
389                               EXT4_ADDR_PER_BLOCK((inode)->i_sb))
390
391 #define ext4_check_inode_blockref(inode)                                \
392         __ext4_check_blockref(__func__, __LINE__, inode,                \
393                               EXT4_I(inode)->i_data,                    \
394                               EXT4_NDIR_BLOCKS)
395
396 /**
397  *      ext4_get_branch - read the chain of indirect blocks leading to data
398  *      @inode: inode in question
399  *      @depth: depth of the chain (1 - direct pointer, etc.)
400  *      @offsets: offsets of pointers in inode/indirect blocks
401  *      @chain: place to store the result
402  *      @err: here we store the error value
403  *
404  *      Function fills the array of triples <key, p, bh> and returns %NULL
405  *      if everything went OK or the pointer to the last filled triple
406  *      (incomplete one) otherwise. Upon the return chain[i].key contains
407  *      the number of (i+1)-th block in the chain (as it is stored in memory,
408  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
409  *      number (it points into struct inode for i==0 and into the bh->b_data
410  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
411  *      block for i>0 and NULL for i==0. In other words, it holds the block
412  *      numbers of the chain, addresses they were taken from (and where we can
413  *      verify that chain did not change) and buffer_heads hosting these
414  *      numbers.
415  *
416  *      Function stops when it stumbles upon zero pointer (absent block)
417  *              (pointer to last triple returned, *@err == 0)
418  *      or when it gets an IO error reading an indirect block
419  *              (ditto, *@err == -EIO)
420  *      or when it reads all @depth-1 indirect blocks successfully and finds
421  *      the whole chain, all way to the data (returns %NULL, *err == 0).
422  *
423  *      Need to be called with
424  *      down_read(&EXT4_I(inode)->i_data_sem)
425  */
426 static Indirect *ext4_get_branch(struct inode *inode, int depth,
427                                  ext4_lblk_t  *offsets,
428                                  Indirect chain[4], int *err)
429 {
430         struct super_block *sb = inode->i_sb;
431         Indirect *p = chain;
432         struct buffer_head *bh;
433
434         *err = 0;
435         /* i_data is not going away, no lock needed */
436         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
437         if (!p->key)
438                 goto no_block;
439         while (--depth) {
440                 bh = sb_getblk(sb, le32_to_cpu(p->key));
441                 if (unlikely(!bh))
442                         goto failure;
443
444                 if (!bh_uptodate_or_lock(bh)) {
445                         if (bh_submit_read(bh) < 0) {
446                                 put_bh(bh);
447                                 goto failure;
448                         }
449                         /* validate block references */
450                         if (ext4_check_indirect_blockref(inode, bh)) {
451                                 put_bh(bh);
452                                 goto failure;
453                         }
454                 }
455
456                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
457                 /* Reader: end */
458                 if (!p->key)
459                         goto no_block;
460         }
461         return NULL;
462
463 failure:
464         *err = -EIO;
465 no_block:
466         return p;
467 }
468
469 /**
470  *      ext4_find_near - find a place for allocation with sufficient locality
471  *      @inode: owner
472  *      @ind: descriptor of indirect block.
473  *
474  *      This function returns the preferred place for block allocation.
475  *      It is used when heuristic for sequential allocation fails.
476  *      Rules are:
477  *        + if there is a block to the left of our position - allocate near it.
478  *        + if pointer will live in indirect block - allocate near that block.
479  *        + if pointer will live in inode - allocate in the same
480  *          cylinder group.
481  *
482  * In the latter case we colour the starting block by the callers PID to
483  * prevent it from clashing with concurrent allocations for a different inode
484  * in the same block group.   The PID is used here so that functionally related
485  * files will be close-by on-disk.
486  *
487  *      Caller must make sure that @ind is valid and will stay that way.
488  */
489 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
490 {
491         struct ext4_inode_info *ei = EXT4_I(inode);
492         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
493         __le32 *p;
494         ext4_fsblk_t bg_start;
495         ext4_fsblk_t last_block;
496         ext4_grpblk_t colour;
497         ext4_group_t block_group;
498         int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
499
500         /* Try to find previous block */
501         for (p = ind->p - 1; p >= start; p--) {
502                 if (*p)
503                         return le32_to_cpu(*p);
504         }
505
506         /* No such thing, so let's try location of indirect block */
507         if (ind->bh)
508                 return ind->bh->b_blocknr;
509
510         /*
511          * It is going to be referred to from the inode itself? OK, just put it
512          * into the same cylinder group then.
513          */
514         block_group = ei->i_block_group;
515         if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
516                 block_group &= ~(flex_size-1);
517                 if (S_ISREG(inode->i_mode))
518                         block_group++;
519         }
520         bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
521         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
522
523         /*
524          * If we are doing delayed allocation, we don't need take
525          * colour into account.
526          */
527         if (test_opt(inode->i_sb, DELALLOC))
528                 return bg_start;
529
530         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
531                 colour = (current->pid % 16) *
532                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
533         else
534                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
535         return bg_start + colour;
536 }
537
538 /**
539  *      ext4_find_goal - find a preferred place for allocation.
540  *      @inode: owner
541  *      @block:  block we want
542  *      @partial: pointer to the last triple within a chain
543  *
544  *      Normally this function find the preferred place for block allocation,
545  *      returns it.
546  *      Because this is only used for non-extent files, we limit the block nr
547  *      to 32 bits.
548  */
549 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
550                                    Indirect *partial)
551 {
552         ext4_fsblk_t goal;
553
554         /*
555          * XXX need to get goal block from mballoc's data structures
556          */
557
558         goal = ext4_find_near(inode, partial);
559         goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
560         return goal;
561 }
562
563 /**
564  *      ext4_blks_to_allocate - Look up the block map and count the number
565  *      of direct blocks need to be allocated for the given branch.
566  *
567  *      @branch: chain of indirect blocks
568  *      @k: number of blocks need for indirect blocks
569  *      @blks: number of data blocks to be mapped.
570  *      @blocks_to_boundary:  the offset in the indirect block
571  *
572  *      return the total number of blocks to be allocate, including the
573  *      direct and indirect blocks.
574  */
575 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
576                                  int blocks_to_boundary)
577 {
578         unsigned int count = 0;
579
580         /*
581          * Simple case, [t,d]Indirect block(s) has not allocated yet
582          * then it's clear blocks on that path have not allocated
583          */
584         if (k > 0) {
585                 /* right now we don't handle cross boundary allocation */
586                 if (blks < blocks_to_boundary + 1)
587                         count += blks;
588                 else
589                         count += blocks_to_boundary + 1;
590                 return count;
591         }
592
593         count++;
594         while (count < blks && count <= blocks_to_boundary &&
595                 le32_to_cpu(*(branch[0].p + count)) == 0) {
596                 count++;
597         }
598         return count;
599 }
600
601 /**
602  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
603  *      @handle: handle for this transaction
604  *      @inode: inode which needs allocated blocks
605  *      @iblock: the logical block to start allocated at
606  *      @goal: preferred physical block of allocation
607  *      @indirect_blks: the number of blocks need to allocate for indirect
608  *                      blocks
609  *      @blks: number of desired blocks
610  *      @new_blocks: on return it will store the new block numbers for
611  *      the indirect blocks(if needed) and the first direct block,
612  *      @err: on return it will store the error code
613  *
614  *      This function will return the number of blocks allocated as
615  *      requested by the passed-in parameters.
616  */
617 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
618                              ext4_lblk_t iblock, ext4_fsblk_t goal,
619                              int indirect_blks, int blks,
620                              ext4_fsblk_t new_blocks[4], int *err)
621 {
622         struct ext4_allocation_request ar;
623         int target, i;
624         unsigned long count = 0, blk_allocated = 0;
625         int index = 0;
626         ext4_fsblk_t current_block = 0;
627         int ret = 0;
628
629         /*
630          * Here we try to allocate the requested multiple blocks at once,
631          * on a best-effort basis.
632          * To build a branch, we should allocate blocks for
633          * the indirect blocks(if not allocated yet), and at least
634          * the first direct block of this branch.  That's the
635          * minimum number of blocks need to allocate(required)
636          */
637         /* first we try to allocate the indirect blocks */
638         target = indirect_blks;
639         while (target > 0) {
640                 count = target;
641                 /* allocating blocks for indirect blocks and direct blocks */
642                 current_block = ext4_new_meta_blocks(handle, inode, goal,
643                                                      0, &count, err);
644                 if (*err)
645                         goto failed_out;
646
647                 if (unlikely(current_block + count > EXT4_MAX_BLOCK_FILE_PHYS)) {
648                         EXT4_ERROR_INODE(inode,
649                                          "current_block %llu + count %lu > %d!",
650                                          current_block, count,
651                                          EXT4_MAX_BLOCK_FILE_PHYS);
652                         *err = -EIO;
653                         goto failed_out;
654                 }
655
656                 target -= count;
657                 /* allocate blocks for indirect blocks */
658                 while (index < indirect_blks && count) {
659                         new_blocks[index++] = current_block++;
660                         count--;
661                 }
662                 if (count > 0) {
663                         /*
664                          * save the new block number
665                          * for the first direct block
666                          */
667                         new_blocks[index] = current_block;
668                         printk(KERN_INFO "%s returned more blocks than "
669                                                 "requested\n", __func__);
670                         WARN_ON(1);
671                         break;
672                 }
673         }
674
675         target = blks - count ;
676         blk_allocated = count;
677         if (!target)
678                 goto allocated;
679         /* Now allocate data blocks */
680         memset(&ar, 0, sizeof(ar));
681         ar.inode = inode;
682         ar.goal = goal;
683         ar.len = target;
684         ar.logical = iblock;
685         if (S_ISREG(inode->i_mode))
686                 /* enable in-core preallocation only for regular files */
687                 ar.flags = EXT4_MB_HINT_DATA;
688
689         current_block = ext4_mb_new_blocks(handle, &ar, err);
690         if (unlikely(current_block + ar.len > EXT4_MAX_BLOCK_FILE_PHYS)) {
691                 EXT4_ERROR_INODE(inode,
692                                  "current_block %llu + ar.len %d > %d!",
693                                  current_block, ar.len,
694                                  EXT4_MAX_BLOCK_FILE_PHYS);
695                 *err = -EIO;
696                 goto failed_out;
697         }
698
699         if (*err && (target == blks)) {
700                 /*
701                  * if the allocation failed and we didn't allocate
702                  * any blocks before
703                  */
704                 goto failed_out;
705         }
706         if (!*err) {
707                 if (target == blks) {
708                         /*
709                          * save the new block number
710                          * for the first direct block
711                          */
712                         new_blocks[index] = current_block;
713                 }
714                 blk_allocated += ar.len;
715         }
716 allocated:
717         /* total number of blocks allocated for direct blocks */
718         ret = blk_allocated;
719         *err = 0;
720         return ret;
721 failed_out:
722         for (i = 0; i < index; i++)
723                 ext4_free_blocks(handle, inode, NULL, new_blocks[i], 1, 0);
724         return ret;
725 }
726
727 /**
728  *      ext4_alloc_branch - allocate and set up a chain of blocks.
729  *      @handle: handle for this transaction
730  *      @inode: owner
731  *      @indirect_blks: number of allocated indirect blocks
732  *      @blks: number of allocated direct blocks
733  *      @goal: preferred place for allocation
734  *      @offsets: offsets (in the blocks) to store the pointers to next.
735  *      @branch: place to store the chain in.
736  *
737  *      This function allocates blocks, zeroes out all but the last one,
738  *      links them into chain and (if we are synchronous) writes them to disk.
739  *      In other words, it prepares a branch that can be spliced onto the
740  *      inode. It stores the information about that chain in the branch[], in
741  *      the same format as ext4_get_branch() would do. We are calling it after
742  *      we had read the existing part of chain and partial points to the last
743  *      triple of that (one with zero ->key). Upon the exit we have the same
744  *      picture as after the successful ext4_get_block(), except that in one
745  *      place chain is disconnected - *branch->p is still zero (we did not
746  *      set the last link), but branch->key contains the number that should
747  *      be placed into *branch->p to fill that gap.
748  *
749  *      If allocation fails we free all blocks we've allocated (and forget
750  *      their buffer_heads) and return the error value the from failed
751  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
752  *      as described above and return 0.
753  */
754 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
755                              ext4_lblk_t iblock, int indirect_blks,
756                              int *blks, ext4_fsblk_t goal,
757                              ext4_lblk_t *offsets, Indirect *branch)
758 {
759         int blocksize = inode->i_sb->s_blocksize;
760         int i, n = 0;
761         int err = 0;
762         struct buffer_head *bh;
763         int num;
764         ext4_fsblk_t new_blocks[4];
765         ext4_fsblk_t current_block;
766
767         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
768                                 *blks, new_blocks, &err);
769         if (err)
770                 return err;
771
772         branch[0].key = cpu_to_le32(new_blocks[0]);
773         /*
774          * metadata blocks and data blocks are allocated.
775          */
776         for (n = 1; n <= indirect_blks;  n++) {
777                 /*
778                  * Get buffer_head for parent block, zero it out
779                  * and set the pointer to new one, then send
780                  * parent to disk.
781                  */
782                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
783                 if (unlikely(!bh)) {
784                         err = -EIO;
785                         goto failed;
786                 }
787
788                 branch[n].bh = bh;
789                 lock_buffer(bh);
790                 BUFFER_TRACE(bh, "call get_create_access");
791                 err = ext4_journal_get_create_access(handle, bh);
792                 if (err) {
793                         /* Don't brelse(bh) here; it's done in
794                          * ext4_journal_forget() below */
795                         unlock_buffer(bh);
796                         goto failed;
797                 }
798
799                 memset(bh->b_data, 0, blocksize);
800                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
801                 branch[n].key = cpu_to_le32(new_blocks[n]);
802                 *branch[n].p = branch[n].key;
803                 if (n == indirect_blks) {
804                         current_block = new_blocks[n];
805                         /*
806                          * End of chain, update the last new metablock of
807                          * the chain to point to the new allocated
808                          * data blocks numbers
809                          */
810                         for (i = 1; i < num; i++)
811                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
812                 }
813                 BUFFER_TRACE(bh, "marking uptodate");
814                 set_buffer_uptodate(bh);
815                 unlock_buffer(bh);
816
817                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
818                 err = ext4_handle_dirty_metadata(handle, inode, bh);
819                 if (err)
820                         goto failed;
821         }
822         *blks = num;
823         return err;
824 failed:
825         /* Allocation failed, free what we already allocated */
826         ext4_free_blocks(handle, inode, NULL, new_blocks[0], 1, 0);
827         for (i = 1; i <= n ; i++) {
828                 /*
829                  * branch[i].bh is newly allocated, so there is no
830                  * need to revoke the block, which is why we don't
831                  * need to set EXT4_FREE_BLOCKS_METADATA.
832                  */
833                 ext4_free_blocks(handle, inode, NULL, new_blocks[i], 1,
834                                  EXT4_FREE_BLOCKS_FORGET);
835         }
836         for (i = n+1; i < indirect_blks; i++)
837                 ext4_free_blocks(handle, inode, NULL, new_blocks[i], 1, 0);
838
839         ext4_free_blocks(handle, inode, NULL, new_blocks[i], num, 0);
840
841         return err;
842 }
843
844 /**
845  * ext4_splice_branch - splice the allocated branch onto inode.
846  * @handle: handle for this transaction
847  * @inode: owner
848  * @block: (logical) number of block we are adding
849  * @chain: chain of indirect blocks (with a missing link - see
850  *      ext4_alloc_branch)
851  * @where: location of missing link
852  * @num:   number of indirect blocks we are adding
853  * @blks:  number of direct blocks we are adding
854  *
855  * This function fills the missing link and does all housekeeping needed in
856  * inode (->i_blocks, etc.). In case of success we end up with the full
857  * chain to new block and return 0.
858  */
859 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
860                               ext4_lblk_t block, Indirect *where, int num,
861                               int blks)
862 {
863         int i;
864         int err = 0;
865         ext4_fsblk_t current_block;
866
867         /*
868          * If we're splicing into a [td]indirect block (as opposed to the
869          * inode) then we need to get write access to the [td]indirect block
870          * before the splice.
871          */
872         if (where->bh) {
873                 BUFFER_TRACE(where->bh, "get_write_access");
874                 err = ext4_journal_get_write_access(handle, where->bh);
875                 if (err)
876                         goto err_out;
877         }
878         /* That's it */
879
880         *where->p = where->key;
881
882         /*
883          * Update the host buffer_head or inode to point to more just allocated
884          * direct blocks blocks
885          */
886         if (num == 0 && blks > 1) {
887                 current_block = le32_to_cpu(where->key) + 1;
888                 for (i = 1; i < blks; i++)
889                         *(where->p + i) = cpu_to_le32(current_block++);
890         }
891
892         /* We are done with atomic stuff, now do the rest of housekeeping */
893         /* had we spliced it onto indirect block? */
894         if (where->bh) {
895                 /*
896                  * If we spliced it onto an indirect block, we haven't
897                  * altered the inode.  Note however that if it is being spliced
898                  * onto an indirect block at the very end of the file (the
899                  * file is growing) then we *will* alter the inode to reflect
900                  * the new i_size.  But that is not done here - it is done in
901                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
902                  */
903                 jbd_debug(5, "splicing indirect only\n");
904                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
905                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
906                 if (err)
907                         goto err_out;
908         } else {
909                 /*
910                  * OK, we spliced it into the inode itself on a direct block.
911                  */
912                 ext4_mark_inode_dirty(handle, inode);
913                 jbd_debug(5, "splicing direct\n");
914         }
915         return err;
916
917 err_out:
918         for (i = 1; i <= num; i++) {
919                 /*
920                  * branch[i].bh is newly allocated, so there is no
921                  * need to revoke the block, which is why we don't
922                  * need to set EXT4_FREE_BLOCKS_METADATA.
923                  */
924                 ext4_free_blocks(handle, inode, where[i].bh, 0, 1,
925                                  EXT4_FREE_BLOCKS_FORGET);
926         }
927         ext4_free_blocks(handle, inode, NULL, le32_to_cpu(where[num].key),
928                          blks, 0);
929
930         return err;
931 }
932
933 /*
934  * The ext4_ind_map_blocks() function handles non-extents inodes
935  * (i.e., using the traditional indirect/double-indirect i_blocks
936  * scheme) for ext4_map_blocks().
937  *
938  * Allocation strategy is simple: if we have to allocate something, we will
939  * have to go the whole way to leaf. So let's do it before attaching anything
940  * to tree, set linkage between the newborn blocks, write them if sync is
941  * required, recheck the path, free and repeat if check fails, otherwise
942  * set the last missing link (that will protect us from any truncate-generated
943  * removals - all blocks on the path are immune now) and possibly force the
944  * write on the parent block.
945  * That has a nice additional property: no special recovery from the failed
946  * allocations is needed - we simply release blocks and do not touch anything
947  * reachable from inode.
948  *
949  * `handle' can be NULL if create == 0.
950  *
951  * return > 0, # of blocks mapped or allocated.
952  * return = 0, if plain lookup failed.
953  * return < 0, error case.
954  *
955  * The ext4_ind_get_blocks() function should be called with
956  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
957  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
958  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
959  * blocks.
960  */
961 static int ext4_ind_map_blocks(handle_t *handle, struct inode *inode,
962                                struct ext4_map_blocks *map,
963                                int flags)
964 {
965         int err = -EIO;
966         ext4_lblk_t offsets[4];
967         Indirect chain[4];
968         Indirect *partial;
969         ext4_fsblk_t goal;
970         int indirect_blks;
971         int blocks_to_boundary = 0;
972         int depth;
973         int count = 0;
974         ext4_fsblk_t first_block = 0;
975
976         trace_ext4_ind_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
977         J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
978         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
979         depth = ext4_block_to_path(inode, map->m_lblk, offsets,
980                                    &blocks_to_boundary);
981
982         if (depth == 0)
983                 goto out;
984
985         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
986
987         /* Simplest case - block found, no allocation needed */
988         if (!partial) {
989                 first_block = le32_to_cpu(chain[depth - 1].key);
990                 count++;
991                 /*map more blocks*/
992                 while (count < map->m_len && count <= blocks_to_boundary) {
993                         ext4_fsblk_t blk;
994
995                         blk = le32_to_cpu(*(chain[depth-1].p + count));
996
997                         if (blk == first_block + count)
998                                 count++;
999                         else
1000                                 break;
1001                 }
1002                 goto got_it;
1003         }
1004
1005         /* Next simple case - plain lookup or failed read of indirect block */
1006         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
1007                 goto cleanup;
1008
1009         /*
1010          * Okay, we need to do block allocation.
1011         */
1012         goal = ext4_find_goal(inode, map->m_lblk, partial);
1013
1014         /* the number of blocks need to allocate for [d,t]indirect blocks */
1015         indirect_blks = (chain + depth) - partial - 1;
1016
1017         /*
1018          * Next look up the indirect map to count the totoal number of
1019          * direct blocks to allocate for this branch.
1020          */
1021         count = ext4_blks_to_allocate(partial, indirect_blks,
1022                                       map->m_len, blocks_to_boundary);
1023         /*
1024          * Block out ext4_truncate while we alter the tree
1025          */
1026         err = ext4_alloc_branch(handle, inode, map->m_lblk, indirect_blks,
1027                                 &count, goal,
1028                                 offsets + (partial - chain), partial);
1029
1030         /*
1031          * The ext4_splice_branch call will free and forget any buffers
1032          * on the new chain if there is a failure, but that risks using
1033          * up transaction credits, especially for bitmaps where the
1034          * credits cannot be returned.  Can we handle this somehow?  We
1035          * may need to return -EAGAIN upwards in the worst case.  --sct
1036          */
1037         if (!err)
1038                 err = ext4_splice_branch(handle, inode, map->m_lblk,
1039                                          partial, indirect_blks, count);
1040         if (err)
1041                 goto cleanup;
1042
1043         map->m_flags |= EXT4_MAP_NEW;
1044
1045         ext4_update_inode_fsync_trans(handle, inode, 1);
1046 got_it:
1047         map->m_flags |= EXT4_MAP_MAPPED;
1048         map->m_pblk = le32_to_cpu(chain[depth-1].key);
1049         map->m_len = count;
1050         if (count > blocks_to_boundary)
1051                 map->m_flags |= EXT4_MAP_BOUNDARY;
1052         err = count;
1053         /* Clean up and exit */
1054         partial = chain + depth - 1;    /* the whole chain */
1055 cleanup:
1056         while (partial > chain) {
1057                 BUFFER_TRACE(partial->bh, "call brelse");
1058                 brelse(partial->bh);
1059                 partial--;
1060         }
1061 out:
1062         trace_ext4_ind_map_blocks_exit(inode, map->m_lblk,
1063                                 map->m_pblk, map->m_len, err);
1064         return err;
1065 }
1066
1067 #ifdef CONFIG_QUOTA
1068 qsize_t *ext4_get_reserved_space(struct inode *inode)
1069 {
1070         return &EXT4_I(inode)->i_reserved_quota;
1071 }
1072 #endif
1073
1074 /*
1075  * Calculate the number of metadata blocks need to reserve
1076  * to allocate a new block at @lblocks for non extent file based file
1077  */
1078 static int ext4_indirect_calc_metadata_amount(struct inode *inode,
1079                                               sector_t lblock)
1080 {
1081         struct ext4_inode_info *ei = EXT4_I(inode);
1082         sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
1083         int blk_bits;
1084
1085         if (lblock < EXT4_NDIR_BLOCKS)
1086                 return 0;
1087
1088         lblock -= EXT4_NDIR_BLOCKS;
1089
1090         if (ei->i_da_metadata_calc_len &&
1091             (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
1092                 ei->i_da_metadata_calc_len++;
1093                 return 0;
1094         }
1095         ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
1096         ei->i_da_metadata_calc_len = 1;
1097         blk_bits = order_base_2(lblock);
1098         return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
1099 }
1100
1101 /*
1102  * Calculate the number of metadata blocks need to reserve
1103  * to allocate a block located at @lblock
1104  */
1105 static int ext4_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
1106 {
1107         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1108                 return ext4_ext_calc_metadata_amount(inode, lblock);
1109
1110         return ext4_indirect_calc_metadata_amount(inode, lblock);
1111 }
1112
1113 /*
1114  * Called with i_data_sem down, which is important since we can call
1115  * ext4_discard_preallocations() from here.
1116  */
1117 void ext4_da_update_reserve_space(struct inode *inode,
1118                                         int used, int quota_claim)
1119 {
1120         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1121         struct ext4_inode_info *ei = EXT4_I(inode);
1122
1123         spin_lock(&ei->i_block_reservation_lock);
1124         trace_ext4_da_update_reserve_space(inode, used);
1125         if (unlikely(used > ei->i_reserved_data_blocks)) {
1126                 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
1127                          "with only %d reserved data blocks\n",
1128                          __func__, inode->i_ino, used,
1129                          ei->i_reserved_data_blocks);
1130                 WARN_ON(1);
1131                 used = ei->i_reserved_data_blocks;
1132         }
1133
1134         /* Update per-inode reservations */
1135         ei->i_reserved_data_blocks -= used;
1136         ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
1137         percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1138                            used + ei->i_allocated_meta_blocks);
1139         ei->i_allocated_meta_blocks = 0;
1140
1141         if (ei->i_reserved_data_blocks == 0) {
1142                 /*
1143                  * We can release all of the reserved metadata blocks
1144                  * only when we have written all of the delayed
1145                  * allocation blocks.
1146                  */
1147                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1148                                    ei->i_reserved_meta_blocks);
1149                 ei->i_reserved_meta_blocks = 0;
1150                 ei->i_da_metadata_calc_len = 0;
1151         }
1152         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1153
1154         /* Update quota subsystem for data blocks */
1155         if (quota_claim)
1156                 dquot_claim_block(inode, used);
1157         else {
1158                 /*
1159                  * We did fallocate with an offset that is already delayed
1160                  * allocated. So on delayed allocated writeback we should
1161                  * not re-claim the quota for fallocated blocks.
1162                  */
1163                 dquot_release_reservation_block(inode, used);
1164         }
1165
1166         /*
1167          * If we have done all the pending block allocations and if
1168          * there aren't any writers on the inode, we can discard the
1169          * inode's preallocations.
1170          */
1171         if ((ei->i_reserved_data_blocks == 0) &&
1172             (atomic_read(&inode->i_writecount) == 0))
1173                 ext4_discard_preallocations(inode);
1174 }
1175
1176 static int __check_block_validity(struct inode *inode, const char *func,
1177                                 unsigned int line,
1178                                 struct ext4_map_blocks *map)
1179 {
1180         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
1181                                    map->m_len)) {
1182                 ext4_error_inode(inode, func, line, map->m_pblk,
1183                                  "lblock %lu mapped to illegal pblock "
1184                                  "(length %d)", (unsigned long) map->m_lblk,
1185                                  map->m_len);
1186                 return -EIO;
1187         }
1188         return 0;
1189 }
1190
1191 #define check_block_validity(inode, map)        \
1192         __check_block_validity((inode), __func__, __LINE__, (map))
1193
1194 /*
1195  * Return the number of contiguous dirty pages in a given inode
1196  * starting at page frame idx.
1197  */
1198 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
1199                                     unsigned int max_pages)
1200 {
1201         struct address_space *mapping = inode->i_mapping;
1202         pgoff_t index;
1203         struct pagevec pvec;
1204         pgoff_t num = 0;
1205         int i, nr_pages, done = 0;
1206
1207         if (max_pages == 0)
1208                 return 0;
1209         pagevec_init(&pvec, 0);
1210         while (!done) {
1211                 index = idx;
1212                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
1213                                               PAGECACHE_TAG_DIRTY,
1214                                               (pgoff_t)PAGEVEC_SIZE);
1215                 if (nr_pages == 0)
1216                         break;
1217                 for (i = 0; i < nr_pages; i++) {
1218                         struct page *page = pvec.pages[i];
1219                         struct buffer_head *bh, *head;
1220
1221                         lock_page(page);
1222                         if (unlikely(page->mapping != mapping) ||
1223                             !PageDirty(page) ||
1224                             PageWriteback(page) ||
1225                             page->index != idx) {
1226                                 done = 1;
1227                                 unlock_page(page);
1228                                 break;
1229                         }
1230                         if (page_has_buffers(page)) {
1231                                 bh = head = page_buffers(page);
1232                                 do {
1233                                         if (!buffer_delay(bh) &&
1234                                             !buffer_unwritten(bh))
1235                                                 done = 1;
1236                                         bh = bh->b_this_page;
1237                                 } while (!done && (bh != head));
1238                         }
1239                         unlock_page(page);
1240                         if (done)
1241                                 break;
1242                         idx++;
1243                         num++;
1244                         if (num >= max_pages) {
1245                                 done = 1;
1246                                 break;
1247                         }
1248                 }
1249                 pagevec_release(&pvec);
1250         }
1251         return num;
1252 }
1253
1254 /*
1255  * The ext4_map_blocks() function tries to look up the requested blocks,
1256  * and returns if the blocks are already mapped.
1257  *
1258  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1259  * and store the allocated blocks in the result buffer head and mark it
1260  * mapped.
1261  *
1262  * If file type is extents based, it will call ext4_ext_map_blocks(),
1263  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
1264  * based files
1265  *
1266  * On success, it returns the number of blocks being mapped or allocate.
1267  * if create==0 and the blocks are pre-allocated and uninitialized block,
1268  * the result buffer head is unmapped. If the create ==1, it will make sure
1269  * the buffer head is mapped.
1270  *
1271  * It returns 0 if plain look up failed (blocks have not been allocated), in
1272  * that casem, buffer head is unmapped
1273  *
1274  * It returns the error in case of allocation failure.
1275  */
1276 int ext4_map_blocks(handle_t *handle, struct inode *inode,
1277                     struct ext4_map_blocks *map, int flags)
1278 {
1279         int retval;
1280
1281         map->m_flags = 0;
1282         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
1283                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
1284                   (unsigned long) map->m_lblk);
1285         /*
1286          * Try to see if we can get the block without requesting a new
1287          * file system block.
1288          */
1289         down_read((&EXT4_I(inode)->i_data_sem));
1290         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
1291                 retval = ext4_ext_map_blocks(handle, inode, map, 0);
1292         } else {
1293                 retval = ext4_ind_map_blocks(handle, inode, map, 0);
1294         }
1295         up_read((&EXT4_I(inode)->i_data_sem));
1296
1297         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
1298                 int ret = check_block_validity(inode, map);
1299                 if (ret != 0)
1300                         return ret;
1301         }
1302
1303         /* If it is only a block(s) look up */
1304         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1305                 return retval;
1306
1307         /*
1308          * Returns if the blocks have already allocated
1309          *
1310          * Note that if blocks have been preallocated
1311          * ext4_ext_get_block() returns th create = 0
1312          * with buffer head unmapped.
1313          */
1314         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
1315                 return retval;
1316
1317         /*
1318          * When we call get_blocks without the create flag, the
1319          * BH_Unwritten flag could have gotten set if the blocks
1320          * requested were part of a uninitialized extent.  We need to
1321          * clear this flag now that we are committed to convert all or
1322          * part of the uninitialized extent to be an initialized
1323          * extent.  This is because we need to avoid the combination
1324          * of BH_Unwritten and BH_Mapped flags being simultaneously
1325          * set on the buffer_head.
1326          */
1327         map->m_flags &= ~EXT4_MAP_UNWRITTEN;
1328
1329         /*
1330          * New blocks allocate and/or writing to uninitialized extent
1331          * will possibly result in updating i_data, so we take
1332          * the write lock of i_data_sem, and call get_blocks()
1333          * with create == 1 flag.
1334          */
1335         down_write((&EXT4_I(inode)->i_data_sem));
1336
1337         /*
1338          * if the caller is from delayed allocation writeout path
1339          * we have already reserved fs blocks for allocation
1340          * let the underlying get_block() function know to
1341          * avoid double accounting
1342          */
1343         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1344                 ext4_set_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
1345         /*
1346          * We need to check for EXT4 here because migrate
1347          * could have changed the inode type in between
1348          */
1349         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
1350                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
1351         } else {
1352                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
1353
1354                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
1355                         /*
1356                          * We allocated new blocks which will result in
1357                          * i_data's format changing.  Force the migrate
1358                          * to fail by clearing migrate flags
1359                          */
1360                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
1361                 }
1362
1363                 /*
1364                  * Update reserved blocks/metadata blocks after successful
1365                  * block allocation which had been deferred till now. We don't
1366                  * support fallocate for non extent files. So we can update
1367                  * reserve space here.
1368                  */
1369                 if ((retval > 0) &&
1370                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
1371                         ext4_da_update_reserve_space(inode, retval, 1);
1372         }
1373         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1374                 ext4_clear_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
1375
1376         up_write((&EXT4_I(inode)->i_data_sem));
1377         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
1378                 int ret = check_block_validity(inode, map);
1379                 if (ret != 0)
1380                         return ret;
1381         }
1382         return retval;
1383 }
1384
1385 /* Maximum number of blocks we map for direct IO at once. */
1386 #define DIO_MAX_BLOCKS 4096
1387
1388 static int _ext4_get_block(struct inode *inode, sector_t iblock,
1389                            struct buffer_head *bh, int flags)
1390 {
1391         handle_t *handle = ext4_journal_current_handle();
1392         struct ext4_map_blocks map;
1393         int ret = 0, started = 0;
1394         int dio_credits;
1395
1396         map.m_lblk = iblock;
1397         map.m_len = bh->b_size >> inode->i_blkbits;
1398
1399         if (flags && !handle) {
1400                 /* Direct IO write... */
1401                 if (map.m_len > DIO_MAX_BLOCKS)
1402                         map.m_len = DIO_MAX_BLOCKS;
1403                 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
1404                 handle = ext4_journal_start(inode, dio_credits);
1405                 if (IS_ERR(handle)) {
1406                         ret = PTR_ERR(handle);
1407                         return ret;
1408                 }
1409                 started = 1;
1410         }
1411
1412         ret = ext4_map_blocks(handle, inode, &map, flags);
1413         if (ret > 0) {
1414                 map_bh(bh, inode->i_sb, map.m_pblk);
1415                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
1416                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
1417                 ret = 0;
1418         }
1419         if (started)
1420                 ext4_journal_stop(handle);
1421         return ret;
1422 }
1423
1424 int ext4_get_block(struct inode *inode, sector_t iblock,
1425                    struct buffer_head *bh, int create)
1426 {
1427         return _ext4_get_block(inode, iblock, bh,
1428                                create ? EXT4_GET_BLOCKS_CREATE : 0);
1429 }
1430
1431 /*
1432  * `handle' can be NULL if create is zero
1433  */
1434 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1435                                 ext4_lblk_t block, int create, int *errp)
1436 {
1437         struct ext4_map_blocks map;
1438         struct buffer_head *bh;
1439         int fatal = 0, err;
1440
1441         J_ASSERT(handle != NULL || create == 0);
1442
1443         map.m_lblk = block;
1444         map.m_len = 1;
1445         err = ext4_map_blocks(handle, inode, &map,
1446                               create ? EXT4_GET_BLOCKS_CREATE : 0);
1447
1448         if (err < 0)
1449                 *errp = err;
1450         if (err <= 0)
1451                 return NULL;
1452         *errp = 0;
1453
1454         bh = sb_getblk(inode->i_sb, map.m_pblk);
1455         if (!bh) {
1456                 *errp = -EIO;
1457                 return NULL;
1458         }
1459         if (map.m_flags & EXT4_MAP_NEW) {
1460                 J_ASSERT(create != 0);
1461                 J_ASSERT(handle != NULL);
1462
1463                 /*
1464                  * Now that we do not always journal data, we should
1465                  * keep in mind whether this should always journal the
1466                  * new buffer as metadata.  For now, regular file
1467                  * writes use ext4_get_block instead, so it's not a
1468                  * problem.
1469                  */
1470                 lock_buffer(bh);
1471                 BUFFER_TRACE(bh, "call get_create_access");
1472                 fatal = ext4_journal_get_create_access(handle, bh);
1473                 if (!fatal && !buffer_uptodate(bh)) {
1474                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1475                         set_buffer_uptodate(bh);
1476                 }
1477                 unlock_buffer(bh);
1478                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1479                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1480                 if (!fatal)
1481                         fatal = err;
1482         } else {
1483                 BUFFER_TRACE(bh, "not a new buffer");
1484         }
1485         if (fatal) {
1486                 *errp = fatal;
1487                 brelse(bh);
1488                 bh = NULL;
1489         }
1490         return bh;
1491 }
1492
1493 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1494                                ext4_lblk_t block, int create, int *err)
1495 {
1496         struct buffer_head *bh;
1497
1498         bh = ext4_getblk(handle, inode, block, create, err);
1499         if (!bh)
1500                 return bh;
1501         if (buffer_uptodate(bh))
1502                 return bh;
1503         ll_rw_block(READ_META, 1, &bh);
1504         wait_on_buffer(bh);
1505         if (buffer_uptodate(bh))
1506                 return bh;
1507         put_bh(bh);
1508         *err = -EIO;
1509         return NULL;
1510 }
1511
1512 static int walk_page_buffers(handle_t *handle,
1513                              struct buffer_head *head,
1514                              unsigned from,
1515                              unsigned to,
1516                              int *partial,
1517                              int (*fn)(handle_t *handle,
1518                                        struct buffer_head *bh))
1519 {
1520         struct buffer_head *bh;
1521         unsigned block_start, block_end;
1522         unsigned blocksize = head->b_size;
1523         int err, ret = 0;
1524         struct buffer_head *next;
1525
1526         for (bh = head, block_start = 0;
1527              ret == 0 && (bh != head || !block_start);
1528              block_start = block_end, bh = next) {
1529                 next = bh->b_this_page;
1530                 block_end = block_start + blocksize;
1531                 if (block_end <= from || block_start >= to) {
1532                         if (partial && !buffer_uptodate(bh))
1533                                 *partial = 1;
1534                         continue;
1535                 }
1536                 err = (*fn)(handle, bh);
1537                 if (!ret)
1538                         ret = err;
1539         }
1540         return ret;
1541 }
1542
1543 /*
1544  * To preserve ordering, it is essential that the hole instantiation and
1545  * the data write be encapsulated in a single transaction.  We cannot
1546  * close off a transaction and start a new one between the ext4_get_block()
1547  * and the commit_write().  So doing the jbd2_journal_start at the start of
1548  * prepare_write() is the right place.
1549  *
1550  * Also, this function can nest inside ext4_writepage() ->
1551  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1552  * has generated enough buffer credits to do the whole page.  So we won't
1553  * block on the journal in that case, which is good, because the caller may
1554  * be PF_MEMALLOC.
1555  *
1556  * By accident, ext4 can be reentered when a transaction is open via
1557  * quota file writes.  If we were to commit the transaction while thus
1558  * reentered, there can be a deadlock - we would be holding a quota
1559  * lock, and the commit would never complete if another thread had a
1560  * transaction open and was blocking on the quota lock - a ranking
1561  * violation.
1562  *
1563  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1564  * will _not_ run commit under these circumstances because handle->h_ref
1565  * is elevated.  We'll still have enough credits for the tiny quotafile
1566  * write.
1567  */
1568 static int do_journal_get_write_access(handle_t *handle,
1569                                        struct buffer_head *bh)
1570 {
1571         int dirty = buffer_dirty(bh);
1572         int ret;
1573
1574         if (!buffer_mapped(bh) || buffer_freed(bh))
1575                 return 0;
1576         /*
1577          * __block_write_begin() could have dirtied some buffers. Clean
1578          * the dirty bit as jbd2_journal_get_write_access() could complain
1579          * otherwise about fs integrity issues. Setting of the dirty bit
1580          * by __block_write_begin() isn't a real problem here as we clear
1581          * the bit before releasing a page lock and thus writeback cannot
1582          * ever write the buffer.
1583          */
1584         if (dirty)
1585                 clear_buffer_dirty(bh);
1586         ret = ext4_journal_get_write_access(handle, bh);
1587         if (!ret && dirty)
1588                 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1589         return ret;
1590 }
1591
1592 /*
1593  * Truncate blocks that were not used by write. We have to truncate the
1594  * pagecache as well so that corresponding buffers get properly unmapped.
1595  */
1596 static void ext4_truncate_failed_write(struct inode *inode)
1597 {
1598         truncate_inode_pages(inode->i_mapping, inode->i_size);
1599         ext4_truncate(inode);
1600 }
1601
1602 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
1603                    struct buffer_head *bh_result, int create);
1604 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1605                             loff_t pos, unsigned len, unsigned flags,
1606                             struct page **pagep, void **fsdata)
1607 {
1608         struct inode *inode = mapping->host;
1609         int ret, needed_blocks;
1610         handle_t *handle;
1611         int retries = 0;
1612         struct page *page;
1613         pgoff_t index;
1614         unsigned from, to;
1615
1616         trace_ext4_write_begin(inode, pos, len, flags);
1617         /*
1618          * Reserve one block more for addition to orphan list in case
1619          * we allocate blocks but write fails for some reason
1620          */
1621         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1622         index = pos >> PAGE_CACHE_SHIFT;
1623         from = pos & (PAGE_CACHE_SIZE - 1);
1624         to = from + len;
1625
1626 retry:
1627         handle = ext4_journal_start(inode, needed_blocks);
1628         if (IS_ERR(handle)) {
1629                 ret = PTR_ERR(handle);
1630                 goto out;
1631         }
1632
1633         /* We cannot recurse into the filesystem as the transaction is already
1634          * started */
1635         flags |= AOP_FLAG_NOFS;
1636
1637         page = grab_cache_page_write_begin(mapping, index, flags);
1638         if (!page) {
1639                 ext4_journal_stop(handle);
1640                 ret = -ENOMEM;
1641                 goto out;
1642         }
1643         *pagep = page;
1644
1645         if (ext4_should_dioread_nolock(inode))
1646                 ret = __block_write_begin(page, pos, len, ext4_get_block_write);
1647         else
1648                 ret = __block_write_begin(page, pos, len, ext4_get_block);
1649
1650         if (!ret && ext4_should_journal_data(inode)) {
1651                 ret = walk_page_buffers(handle, page_buffers(page),
1652                                 from, to, NULL, do_journal_get_write_access);
1653         }
1654
1655         if (ret) {
1656                 unlock_page(page);
1657                 page_cache_release(page);
1658                 /*
1659                  * __block_write_begin may have instantiated a few blocks
1660                  * outside i_size.  Trim these off again. Don't need
1661                  * i_size_read because we hold i_mutex.
1662                  *
1663                  * Add inode to orphan list in case we crash before
1664                  * truncate finishes
1665                  */
1666                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1667                         ext4_orphan_add(handle, inode);
1668
1669                 ext4_journal_stop(handle);
1670                 if (pos + len > inode->i_size) {
1671                         ext4_truncate_failed_write(inode);
1672                         /*
1673                          * If truncate failed early the inode might
1674                          * still be on the orphan list; we need to
1675                          * make sure the inode is removed from the
1676                          * orphan list in that case.
1677                          */
1678                         if (inode->i_nlink)
1679                                 ext4_orphan_del(NULL, inode);
1680                 }
1681         }
1682
1683         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1684                 goto retry;
1685 out:
1686         return ret;
1687 }
1688
1689 /* For write_end() in data=journal mode */
1690 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1691 {
1692         if (!buffer_mapped(bh) || buffer_freed(bh))
1693                 return 0;
1694         set_buffer_uptodate(bh);
1695         return ext4_handle_dirty_metadata(handle, NULL, bh);
1696 }
1697
1698 static int ext4_generic_write_end(struct file *file,
1699                                   struct address_space *mapping,
1700                                   loff_t pos, unsigned len, unsigned copied,
1701                                   struct page *page, void *fsdata)
1702 {
1703         int i_size_changed = 0;
1704         struct inode *inode = mapping->host;
1705         handle_t *handle = ext4_journal_current_handle();
1706
1707         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1708
1709         /*
1710          * No need to use i_size_read() here, the i_size
1711          * cannot change under us because we hold i_mutex.
1712          *
1713          * But it's important to update i_size while still holding page lock:
1714          * page writeout could otherwise come in and zero beyond i_size.
1715          */
1716         if (pos + copied > inode->i_size) {
1717                 i_size_write(inode, pos + copied);
1718                 i_size_changed = 1;
1719         }
1720
1721         if (pos + copied >  EXT4_I(inode)->i_disksize) {
1722                 /* We need to mark inode dirty even if
1723                  * new_i_size is less that inode->i_size
1724                  * bu greater than i_disksize.(hint delalloc)
1725                  */
1726                 ext4_update_i_disksize(inode, (pos + copied));
1727                 i_size_changed = 1;
1728         }
1729         unlock_page(page);
1730         page_cache_release(page);
1731
1732         /*
1733          * Don't mark the inode dirty under page lock. First, it unnecessarily
1734          * makes the holding time of page lock longer. Second, it forces lock
1735          * ordering of page lock and transaction start for journaling
1736          * filesystems.
1737          */
1738         if (i_size_changed)
1739                 ext4_mark_inode_dirty(handle, inode);
1740
1741         return copied;
1742 }
1743
1744 /*
1745  * We need to pick up the new inode size which generic_commit_write gave us
1746  * `file' can be NULL - eg, when called from page_symlink().
1747  *
1748  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1749  * buffers are managed internally.
1750  */
1751 static int ext4_ordered_write_end(struct file *file,
1752                                   struct address_space *mapping,
1753                                   loff_t pos, unsigned len, unsigned copied,
1754                                   struct page *page, void *fsdata)
1755 {
1756         handle_t *handle = ext4_journal_current_handle();
1757         struct inode *inode = mapping->host;
1758         int ret = 0, ret2;
1759
1760         trace_ext4_ordered_write_end(inode, pos, len, copied);
1761         ret = ext4_jbd2_file_inode(handle, inode);
1762
1763         if (ret == 0) {
1764                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1765                                                         page, fsdata);
1766                 copied = ret2;
1767                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1768                         /* if we have allocated more blocks and copied
1769                          * less. We will have blocks allocated outside
1770                          * inode->i_size. So truncate them
1771                          */
1772                         ext4_orphan_add(handle, inode);
1773                 if (ret2 < 0)
1774                         ret = ret2;
1775         }
1776         ret2 = ext4_journal_stop(handle);
1777         if (!ret)
1778                 ret = ret2;
1779
1780         if (pos + len > inode->i_size) {
1781                 ext4_truncate_failed_write(inode);
1782                 /*
1783                  * If truncate failed early the inode might still be
1784                  * on the orphan list; we need to make sure the inode
1785                  * is removed from the orphan list in that case.
1786                  */
1787                 if (inode->i_nlink)
1788                         ext4_orphan_del(NULL, inode);
1789         }
1790
1791
1792         return ret ? ret : copied;
1793 }
1794
1795 static int ext4_writeback_write_end(struct file *file,
1796                                     struct address_space *mapping,
1797                                     loff_t pos, unsigned len, unsigned copied,
1798                                     struct page *page, void *fsdata)
1799 {
1800         handle_t *handle = ext4_journal_current_handle();
1801         struct inode *inode = mapping->host;
1802         int ret = 0, ret2;
1803
1804         trace_ext4_writeback_write_end(inode, pos, len, copied);
1805         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1806                                                         page, fsdata);
1807         copied = ret2;
1808         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1809                 /* if we have allocated more blocks and copied
1810                  * less. We will have blocks allocated outside
1811                  * inode->i_size. So truncate them
1812                  */
1813                 ext4_orphan_add(handle, inode);
1814
1815         if (ret2 < 0)
1816                 ret = ret2;
1817
1818         ret2 = ext4_journal_stop(handle);
1819         if (!ret)
1820                 ret = ret2;
1821
1822         if (pos + len > inode->i_size) {
1823                 ext4_truncate_failed_write(inode);
1824                 /*
1825                  * If truncate failed early the inode might still be
1826                  * on the orphan list; we need to make sure the inode
1827                  * is removed from the orphan list in that case.
1828                  */
1829                 if (inode->i_nlink)
1830                         ext4_orphan_del(NULL, inode);
1831         }
1832
1833         return ret ? ret : copied;
1834 }
1835
1836 static int ext4_journalled_write_end(struct file *file,
1837                                      struct address_space *mapping,
1838                                      loff_t pos, unsigned len, unsigned copied,
1839                                      struct page *page, void *fsdata)
1840 {
1841         handle_t *handle = ext4_journal_current_handle();
1842         struct inode *inode = mapping->host;
1843         int ret = 0, ret2;
1844         int partial = 0;
1845         unsigned from, to;
1846         loff_t new_i_size;
1847
1848         trace_ext4_journalled_write_end(inode, pos, len, copied);
1849         from = pos & (PAGE_CACHE_SIZE - 1);
1850         to = from + len;
1851
1852         if (copied < len) {
1853                 if (!PageUptodate(page))
1854                         copied = 0;
1855                 page_zero_new_buffers(page, from+copied, to);
1856         }
1857
1858         ret = walk_page_buffers(handle, page_buffers(page), from,
1859                                 to, &partial, write_end_fn);
1860         if (!partial)
1861                 SetPageUptodate(page);
1862         new_i_size = pos + copied;
1863         if (new_i_size > inode->i_size)
1864                 i_size_write(inode, pos+copied);
1865         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1866         if (new_i_size > EXT4_I(inode)->i_disksize) {
1867                 ext4_update_i_disksize(inode, new_i_size);
1868                 ret2 = ext4_mark_inode_dirty(handle, inode);
1869                 if (!ret)
1870                         ret = ret2;
1871         }
1872
1873         unlock_page(page);
1874         page_cache_release(page);
1875         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1876                 /* if we have allocated more blocks and copied
1877                  * less. We will have blocks allocated outside
1878                  * inode->i_size. So truncate them
1879                  */
1880                 ext4_orphan_add(handle, inode);
1881
1882         ret2 = ext4_journal_stop(handle);
1883         if (!ret)
1884                 ret = ret2;
1885         if (pos + len > inode->i_size) {
1886                 ext4_truncate_failed_write(inode);
1887                 /*
1888                  * If truncate failed early the inode might still be
1889                  * on the orphan list; we need to make sure the inode
1890                  * is removed from the orphan list in that case.
1891                  */
1892                 if (inode->i_nlink)
1893                         ext4_orphan_del(NULL, inode);
1894         }
1895
1896         return ret ? ret : copied;
1897 }
1898
1899 /*
1900  * Reserve a single block located at lblock
1901  */
1902 static int ext4_da_reserve_space(struct inode *inode, ext4_lblk_t lblock)
1903 {
1904         int retries = 0;
1905         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1906         struct ext4_inode_info *ei = EXT4_I(inode);
1907         unsigned long md_needed;
1908         int ret;
1909
1910         /*
1911          * recalculate the amount of metadata blocks to reserve
1912          * in order to allocate nrblocks
1913          * worse case is one extent per block
1914          */
1915 repeat:
1916         spin_lock(&ei->i_block_reservation_lock);
1917         md_needed = ext4_calc_metadata_amount(inode, lblock);
1918         trace_ext4_da_reserve_space(inode, md_needed);
1919         spin_unlock(&ei->i_block_reservation_lock);
1920
1921         /*
1922          * We will charge metadata quota at writeout time; this saves
1923          * us from metadata over-estimation, though we may go over by
1924          * a small amount in the end.  Here we just reserve for data.
1925          */
1926         ret = dquot_reserve_block(inode, 1);
1927         if (ret)
1928                 return ret;
1929         /*
1930          * We do still charge estimated metadata to the sb though;
1931          * we cannot afford to run out of free blocks.
1932          */
1933         if (ext4_claim_free_blocks(sbi, md_needed + 1, 0)) {
1934                 dquot_release_reservation_block(inode, 1);
1935                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1936                         yield();
1937                         goto repeat;
1938                 }
1939                 return -ENOSPC;
1940         }
1941         spin_lock(&ei->i_block_reservation_lock);
1942         ei->i_reserved_data_blocks++;
1943         ei->i_reserved_meta_blocks += md_needed;
1944         spin_unlock(&ei->i_block_reservation_lock);
1945
1946         return 0;       /* success */
1947 }
1948
1949 static void ext4_da_release_space(struct inode *inode, int to_free)
1950 {
1951         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1952         struct ext4_inode_info *ei = EXT4_I(inode);
1953
1954         if (!to_free)
1955                 return;         /* Nothing to release, exit */
1956
1957         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1958
1959         trace_ext4_da_release_space(inode, to_free);
1960         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1961                 /*
1962                  * if there aren't enough reserved blocks, then the
1963                  * counter is messed up somewhere.  Since this
1964                  * function is called from invalidate page, it's
1965                  * harmless to return without any action.
1966                  */
1967                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1968                          "ino %lu, to_free %d with only %d reserved "
1969                          "data blocks\n", inode->i_ino, to_free,
1970                          ei->i_reserved_data_blocks);
1971                 WARN_ON(1);
1972                 to_free = ei->i_reserved_data_blocks;
1973         }
1974         ei->i_reserved_data_blocks -= to_free;
1975
1976         if (ei->i_reserved_data_blocks == 0) {
1977                 /*
1978                  * We can release all of the reserved metadata blocks
1979                  * only when we have written all of the delayed
1980                  * allocation blocks.
1981                  */
1982                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1983                                    ei->i_reserved_meta_blocks);
1984                 ei->i_reserved_meta_blocks = 0;
1985                 ei->i_da_metadata_calc_len = 0;
1986         }
1987
1988         /* update fs dirty data blocks counter */
1989         percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1990
1991         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1992
1993         dquot_release_reservation_block(inode, to_free);
1994 }
1995
1996 static void ext4_da_page_release_reservation(struct page *page,
1997                                              unsigned long offset)
1998 {
1999         int to_release = 0;
2000         struct buffer_head *head, *bh;
2001         unsigned int curr_off = 0;
2002
2003         head = page_buffers(page);
2004         bh = head;
2005         do {
2006                 unsigned int next_off = curr_off + bh->b_size;
2007
2008                 if ((offset <= curr_off) && (buffer_delay(bh))) {
2009                         to_release++;
2010                         clear_buffer_delay(bh);
2011                 }
2012                 curr_off = next_off;
2013         } while ((bh = bh->b_this_page) != head);
2014         ext4_da_release_space(page->mapping->host, to_release);
2015 }
2016
2017 /*
2018  * Delayed allocation stuff
2019  */
2020
2021 /*
2022  * mpage_da_submit_io - walks through extent of pages and try to write
2023  * them with writepage() call back
2024  *
2025  * @mpd->inode: inode
2026  * @mpd->first_page: first page of the extent
2027  * @mpd->next_page: page after the last page of the extent
2028  *
2029  * By the time mpage_da_submit_io() is called we expect all blocks
2030  * to be allocated. this may be wrong if allocation failed.
2031  *
2032  * As pages are already locked by write_cache_pages(), we can't use it
2033  */
2034 static int mpage_da_submit_io(struct mpage_da_data *mpd,
2035                               struct ext4_map_blocks *map)
2036 {
2037         struct pagevec pvec;
2038         unsigned long index, end;
2039         int ret = 0, err, nr_pages, i;
2040         struct inode *inode = mpd->inode;
2041         struct address_space *mapping = inode->i_mapping;
2042         loff_t size = i_size_read(inode);
2043         unsigned int len, block_start;
2044         struct buffer_head *bh, *page_bufs = NULL;
2045         int journal_data = ext4_should_journal_data(inode);
2046         sector_t pblock = 0, cur_logical = 0;
2047         struct ext4_io_submit io_submit;
2048
2049         BUG_ON(mpd->next_page <= mpd->first_page);
2050         memset(&io_submit, 0, sizeof(io_submit));
2051         /*
2052          * We need to start from the first_page to the next_page - 1
2053          * to make sure we also write the mapped dirty buffer_heads.
2054          * If we look at mpd->b_blocknr we would only be looking
2055          * at the currently mapped buffer_heads.
2056          */
2057         index = mpd->first_page;
2058         end = mpd->next_page - 1;
2059
2060         pagevec_init(&pvec, 0);
2061         while (index <= end) {
2062                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2063                 if (nr_pages == 0)
2064                         break;
2065                 for (i = 0; i < nr_pages; i++) {
2066                         int commit_write = 0, skip_page = 0;
2067                         struct page *page = pvec.pages[i];
2068
2069                         index = page->index;
2070                         if (index > end)
2071                                 break;
2072
2073                         if (index == size >> PAGE_CACHE_SHIFT)
2074                                 len = size & ~PAGE_CACHE_MASK;
2075                         else
2076                                 len = PAGE_CACHE_SIZE;
2077                         if (map) {
2078                                 cur_logical = index << (PAGE_CACHE_SHIFT -
2079                                                         inode->i_blkbits);
2080                                 pblock = map->m_pblk + (cur_logical -
2081                                                         map->m_lblk);
2082                         }
2083                         index++;
2084
2085                         BUG_ON(!PageLocked(page));
2086                         BUG_ON(PageWriteback(page));
2087
2088                         /*
2089                          * If the page does not have buffers (for
2090                          * whatever reason), try to create them using
2091                          * __block_write_begin.  If this fails,
2092                          * skip the page and move on.
2093                          */
2094                         if (!page_has_buffers(page)) {
2095                                 if (__block_write_begin(page, 0, len,
2096                                                 noalloc_get_block_write)) {
2097                                 skip_page:
2098                                         unlock_page(page);
2099                                         continue;
2100                                 }
2101                                 commit_write = 1;
2102                         }
2103
2104                         bh = page_bufs = page_buffers(page);
2105                         block_start = 0;
2106                         do {
2107                                 if (!bh)
2108                                         goto skip_page;
2109                                 if (map && (cur_logical >= map->m_lblk) &&
2110                                     (cur_logical <= (map->m_lblk +
2111                                                      (map->m_len - 1)))) {
2112                                         if (buffer_delay(bh)) {
2113                                                 clear_buffer_delay(bh);
2114                                                 bh->b_blocknr = pblock;
2115                                         }
2116                                         if (buffer_unwritten(bh) ||
2117                                             buffer_mapped(bh))
2118                                                 BUG_ON(bh->b_blocknr != pblock);
2119                                         if (map->m_flags & EXT4_MAP_UNINIT)
2120                                                 set_buffer_uninit(bh);
2121                                         clear_buffer_unwritten(bh);
2122                                 }
2123
2124                                 /* skip page if block allocation undone */
2125                                 if (buffer_delay(bh) || buffer_unwritten(bh))
2126                                         skip_page = 1;
2127                                 bh = bh->b_this_page;
2128                                 block_start += bh->b_size;
2129                                 cur_logical++;
2130                                 pblock++;
2131                         } while (bh != page_bufs);
2132
2133                         if (skip_page)
2134                                 goto skip_page;
2135
2136                         if (commit_write)
2137                                 /* mark the buffer_heads as dirty & uptodate */
2138                                 block_commit_write(page, 0, len);
2139
2140                         clear_page_dirty_for_io(page);
2141                         /*
2142                          * Delalloc doesn't support data journalling,
2143                          * but eventually maybe we'll lift this
2144                          * restriction.
2145                          */
2146                         if (unlikely(journal_data && PageChecked(page)))
2147                                 err = __ext4_journalled_writepage(page, len);
2148                         else if (test_opt(inode->i_sb, MBLK_IO_SUBMIT))
2149                                 err = ext4_bio_write_page(&io_submit, page,
2150                                                           len, mpd->wbc);
2151                         else
2152                                 err = block_write_full_page(page,
2153                                         noalloc_get_block_write, mpd->wbc);
2154
2155                         if (!err)
2156                                 mpd->pages_written++;
2157                         /*
2158                          * In error case, we have to continue because
2159                          * remaining pages are still locked
2160                          */
2161                         if (ret == 0)
2162                                 ret = err;
2163                 }
2164                 pagevec_release(&pvec);
2165         }
2166         ext4_io_submit(&io_submit);
2167         return ret;
2168 }
2169
2170 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd)
2171 {
2172         int nr_pages, i;
2173         pgoff_t index, end;
2174         struct pagevec pvec;
2175         struct inode *inode = mpd->inode;
2176         struct address_space *mapping = inode->i_mapping;
2177
2178         index = mpd->first_page;
2179         end   = mpd->next_page - 1;
2180         while (index <= end) {
2181                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2182                 if (nr_pages == 0)
2183                         break;
2184                 for (i = 0; i < nr_pages; i++) {
2185                         struct page *page = pvec.pages[i];
2186                         if (page->index > end)
2187                                 break;
2188                         BUG_ON(!PageLocked(page));
2189                         BUG_ON(PageWriteback(page));
2190                         block_invalidatepage(page, 0);
2191                         ClearPageUptodate(page);
2192                         unlock_page(page);
2193                 }
2194                 index = pvec.pages[nr_pages - 1]->index + 1;
2195                 pagevec_release(&pvec);
2196         }
2197         return;
2198 }
2199
2200 static void ext4_print_free_blocks(struct inode *inode)
2201 {
2202         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2203         printk(KERN_CRIT "Total free blocks count %lld\n",
2204                ext4_count_free_blocks(inode->i_sb));
2205         printk(KERN_CRIT "Free/Dirty block details\n");
2206         printk(KERN_CRIT "free_blocks=%lld\n",
2207                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2208         printk(KERN_CRIT "dirty_blocks=%lld\n",
2209                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2210         printk(KERN_CRIT "Block reservation details\n");
2211         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2212                EXT4_I(inode)->i_reserved_data_blocks);
2213         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2214                EXT4_I(inode)->i_reserved_meta_blocks);
2215         return;
2216 }
2217
2218 /*
2219  * mpage_da_map_and_submit - go through given space, map them
2220  *       if necessary, and then submit them for I/O
2221  *
2222  * @mpd - bh describing space
2223  *
2224  * The function skips space we know is already mapped to disk blocks.
2225  *
2226  */
2227 static void mpage_da_map_and_submit(struct mpage_da_data *mpd)
2228 {
2229         int err, blks, get_blocks_flags;
2230         struct ext4_map_blocks map, *mapp = NULL;
2231         sector_t next = mpd->b_blocknr;
2232         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2233         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2234         handle_t *handle = NULL;
2235
2236         /*
2237          * If the blocks are mapped already, or we couldn't accumulate
2238          * any blocks, then proceed immediately to the submission stage.
2239          */
2240         if ((mpd->b_size == 0) ||
2241             ((mpd->b_state  & (1 << BH_Mapped)) &&
2242              !(mpd->b_state & (1 << BH_Delay)) &&
2243              !(mpd->b_state & (1 << BH_Unwritten))))
2244                 goto submit_io;
2245
2246         handle = ext4_journal_current_handle();
2247         BUG_ON(!handle);
2248
2249         /*
2250          * Call ext4_map_blocks() to allocate any delayed allocation
2251          * blocks, or to convert an uninitialized extent to be
2252          * initialized (in the case where we have written into
2253          * one or more preallocated blocks).
2254          *
2255          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2256          * indicate that we are on the delayed allocation path.  This
2257          * affects functions in many different parts of the allocation
2258          * call path.  This flag exists primarily because we don't
2259          * want to change *many* call functions, so ext4_map_blocks()
2260          * will set the EXT4_STATE_DELALLOC_RESERVED flag once the
2261          * inode's allocation semaphore is taken.
2262          *
2263          * If the blocks in questions were delalloc blocks, set
2264          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2265          * variables are updated after the blocks have been allocated.
2266          */
2267         map.m_lblk = next;
2268         map.m_len = max_blocks;
2269         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
2270         if (ext4_should_dioread_nolock(mpd->inode))
2271                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2272         if (mpd->b_state & (1 << BH_Delay))
2273                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2274
2275         blks = ext4_map_blocks(handle, mpd->inode, &map, get_blocks_flags);
2276         if (blks < 0) {
2277                 struct super_block *sb = mpd->inode->i_sb;
2278
2279                 err = blks;
2280                 /*
2281                  * If get block returns EAGAIN or ENOSPC and there
2282                  * appears to be free blocks we will just let
2283                  * mpage_da_submit_io() unlock all of the pages.
2284                  */
2285                 if (err == -EAGAIN)
2286                         goto submit_io;
2287
2288                 if (err == -ENOSPC &&
2289                     ext4_count_free_blocks(sb)) {
2290                         mpd->retval = err;
2291                         goto submit_io;
2292                 }
2293
2294                 /*
2295                  * get block failure will cause us to loop in
2296                  * writepages, because a_ops->writepage won't be able
2297                  * to make progress. The page will be redirtied by
2298                  * writepage and writepages will again try to write
2299                  * the same.
2300                  */
2301                 if (!(EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2302                         ext4_msg(sb, KERN_CRIT,
2303                                  "delayed block allocation failed for inode %lu "
2304                                  "at logical offset %llu with max blocks %zd "
2305                                  "with error %d", mpd->inode->i_ino,
2306                                  (unsigned long long) next,
2307                                  mpd->b_size >> mpd->inode->i_blkbits, err);
2308                         ext4_msg(sb, KERN_CRIT,
2309                                 "This should not happen!! Data will be lost\n");
2310                         if (err == -ENOSPC)
2311                                 ext4_print_free_blocks(mpd->inode);
2312                 }
2313                 /* invalidate all the pages */
2314                 ext4_da_block_invalidatepages(mpd);
2315
2316                 /* Mark this page range as having been completed */
2317                 mpd->io_done = 1;
2318                 return;
2319         }
2320         BUG_ON(blks == 0);
2321
2322         mapp = &map;
2323         if (map.m_flags & EXT4_MAP_NEW) {
2324                 struct block_device *bdev = mpd->inode->i_sb->s_bdev;
2325                 int i;
2326
2327                 for (i = 0; i < map.m_len; i++)
2328                         unmap_underlying_metadata(bdev, map.m_pblk + i);
2329         }
2330
2331         if (ext4_should_order_data(mpd->inode)) {
2332                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2333                 if (err)
2334                         /* This only happens if the journal is aborted */
2335                         return;
2336         }
2337
2338         /*
2339          * Update on-disk size along with block allocation.
2340          */
2341         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2342         if (disksize > i_size_read(mpd->inode))
2343                 disksize = i_size_read(mpd->inode);
2344         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2345                 ext4_update_i_disksize(mpd->inode, disksize);
2346                 err = ext4_mark_inode_dirty(handle, mpd->inode);
2347                 if (err)
2348                         ext4_error(mpd->inode->i_sb,
2349                                    "Failed to mark inode %lu dirty",
2350                                    mpd->inode->i_ino);
2351         }
2352
2353 submit_io:
2354         mpage_da_submit_io(mpd, mapp);
2355         mpd->io_done = 1;
2356 }
2357
2358 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2359                 (1 << BH_Delay) | (1 << BH_Unwritten))
2360
2361 /*
2362  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2363  *
2364  * @mpd->lbh - extent of blocks
2365  * @logical - logical number of the block in the file
2366  * @bh - bh of the block (used to access block's state)
2367  *
2368  * the function is used to collect contig. blocks in same state
2369  */
2370 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2371                                    sector_t logical, size_t b_size,
2372                                    unsigned long b_state)
2373 {
2374         sector_t next;
2375         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2376
2377         /*
2378          * XXX Don't go larger than mballoc is willing to allocate
2379          * This is a stopgap solution.  We eventually need to fold
2380          * mpage_da_submit_io() into this function and then call
2381          * ext4_map_blocks() multiple times in a loop
2382          */
2383         if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
2384                 goto flush_it;
2385
2386         /* check if thereserved journal credits might overflow */
2387         if (!(ext4_test_inode_flag(mpd->inode, EXT4_INODE_EXTENTS))) {
2388                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2389                         /*
2390                          * With non-extent format we are limited by the journal
2391                          * credit available.  Total credit needed to insert
2392                          * nrblocks contiguous blocks is dependent on the
2393                          * nrblocks.  So limit nrblocks.
2394                          */
2395                         goto flush_it;
2396                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2397                                 EXT4_MAX_TRANS_DATA) {
2398                         /*
2399                          * Adding the new buffer_head would make it cross the
2400                          * allowed limit for which we have journal credit
2401                          * reserved. So limit the new bh->b_size
2402                          */
2403                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2404                                                 mpd->inode->i_blkbits;
2405                         /* we will do mpage_da_submit_io in the next loop */
2406                 }
2407         }
2408         /*
2409          * First block in the extent
2410          */
2411         if (mpd->b_size == 0) {
2412                 mpd->b_blocknr = logical;
2413                 mpd->b_size = b_size;
2414                 mpd->b_state = b_state & BH_FLAGS;
2415                 return;
2416         }
2417
2418         next = mpd->b_blocknr + nrblocks;
2419         /*
2420          * Can we merge the block to our big extent?
2421          */
2422         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2423                 mpd->b_size += b_size;
2424                 return;
2425         }
2426
2427 flush_it:
2428         /*
2429          * We couldn't merge the block to our extent, so we
2430          * need to flush current  extent and start new one
2431          */
2432         mpage_da_map_and_submit(mpd);
2433         return;
2434 }
2435
2436 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
2437 {
2438         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
2439 }
2440
2441 /*
2442  * This is a special get_blocks_t callback which is used by
2443  * ext4_da_write_begin().  It will either return mapped block or
2444  * reserve space for a single block.
2445  *
2446  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2447  * We also have b_blocknr = -1 and b_bdev initialized properly
2448  *
2449  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2450  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2451  * initialized properly.
2452  */
2453 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2454                                   struct buffer_head *bh, int create)
2455 {
2456         struct ext4_map_blocks map;
2457         int ret = 0;
2458         sector_t invalid_block = ~((sector_t) 0xffff);
2459
2460         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2461                 invalid_block = ~0;
2462
2463         BUG_ON(create == 0);
2464         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
2465
2466         map.m_lblk = iblock;
2467         map.m_len = 1;
2468
2469         /*
2470          * first, we need to know whether the block is allocated already
2471          * preallocated blocks are unmapped but should treated
2472          * the same as allocated blocks.
2473          */
2474         ret = ext4_map_blocks(NULL, inode, &map, 0);
2475         if (ret < 0)
2476                 return ret;
2477         if (ret == 0) {
2478                 if (buffer_delay(bh))
2479                         return 0; /* Not sure this could or should happen */
2480                 /*
2481                  * XXX: __block_write_begin() unmaps passed block, is it OK?
2482                  */
2483                 ret = ext4_da_reserve_space(inode, iblock);
2484                 if (ret)
2485                         /* not enough space to reserve */
2486                         return ret;
2487
2488                 map_bh(bh, inode->i_sb, invalid_block);
2489                 set_buffer_new(bh);
2490                 set_buffer_delay(bh);
2491                 return 0;
2492         }
2493
2494         map_bh(bh, inode->i_sb, map.m_pblk);
2495         bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
2496
2497         if (buffer_unwritten(bh)) {
2498                 /* A delayed write to unwritten bh should be marked
2499                  * new and mapped.  Mapped ensures that we don't do
2500                  * get_block multiple times when we write to the same
2501                  * offset and new ensures that we do proper zero out
2502                  * for partial write.
2503                  */
2504                 set_buffer_new(bh);
2505                 set_buffer_mapped(bh);
2506         }
2507         return 0;
2508 }
2509
2510 /*
2511  * This function is used as a standard get_block_t calback function
2512  * when there is no desire to allocate any blocks.  It is used as a
2513  * callback function for block_write_begin() and block_write_full_page().
2514  * These functions should only try to map a single block at a time.
2515  *
2516  * Since this function doesn't do block allocations even if the caller
2517  * requests it by passing in create=1, it is critically important that
2518  * any caller checks to make sure that any buffer heads are returned
2519  * by this function are either all already mapped or marked for
2520  * delayed allocation before calling  block_write_full_page().  Otherwise,
2521  * b_blocknr could be left unitialized, and the page write functions will
2522  * be taken by surprise.
2523  */
2524 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2525                                    struct buffer_head *bh_result, int create)
2526 {
2527         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2528         return _ext4_get_block(inode, iblock, bh_result, 0);
2529 }
2530
2531 static int bget_one(handle_t *handle, struct buffer_head *bh)
2532 {
2533         get_bh(bh);
2534         return 0;
2535 }
2536
2537 static int bput_one(handle_t *handle, struct buffer_head *bh)
2538 {
2539         put_bh(bh);
2540         return 0;
2541 }
2542
2543 static int __ext4_journalled_writepage(struct page *page,
2544                                        unsigned int len)
2545 {
2546         struct address_space *mapping = page->mapping;
2547         struct inode *inode = mapping->host;
2548         struct buffer_head *page_bufs;
2549         handle_t *handle = NULL;
2550         int ret = 0;
2551         int err;
2552
2553         ClearPageChecked(page);
2554         page_bufs = page_buffers(page);
2555         BUG_ON(!page_bufs);
2556         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2557         /* As soon as we unlock the page, it can go away, but we have
2558          * references to buffers so we are safe */
2559         unlock_page(page);
2560
2561         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2562         if (IS_ERR(handle)) {
2563                 ret = PTR_ERR(handle);
2564                 goto out;
2565         }
2566
2567         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2568                                 do_journal_get_write_access);
2569
2570         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2571                                 write_end_fn);
2572         if (ret == 0)
2573                 ret = err;
2574         err = ext4_journal_stop(handle);
2575         if (!ret)
2576                 ret = err;
2577
2578         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2579         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2580 out:
2581         return ret;
2582 }
2583
2584 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
2585 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
2586
2587 /*
2588  * Note that we don't need to start a transaction unless we're journaling data
2589  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2590  * need to file the inode to the transaction's list in ordered mode because if
2591  * we are writing back data added by write(), the inode is already there and if
2592  * we are writing back data modified via mmap(), no one guarantees in which
2593  * transaction the data will hit the disk. In case we are journaling data, we
2594  * cannot start transaction directly because transaction start ranks above page
2595  * lock so we have to do some magic.
2596  *
2597  * This function can get called via...
2598  *   - ext4_da_writepages after taking page lock (have journal handle)
2599  *   - journal_submit_inode_data_buffers (no journal handle)
2600  *   - shrink_page_list via pdflush (no journal handle)
2601  *   - grab_page_cache when doing write_begin (have journal handle)
2602  *
2603  * We don't do any block allocation in this function. If we have page with
2604  * multiple blocks we need to write those buffer_heads that are mapped. This
2605  * is important for mmaped based write. So if we do with blocksize 1K
2606  * truncate(f, 1024);
2607  * a = mmap(f, 0, 4096);
2608  * a[0] = 'a';
2609  * truncate(f, 4096);
2610  * we have in the page first buffer_head mapped via page_mkwrite call back
2611  * but other bufer_heads would be unmapped but dirty(dirty done via the
2612  * do_wp_page). So writepage should write the first block. If we modify
2613  * the mmap area beyond 1024 we will again get a page_fault and the
2614  * page_mkwrite callback will do the block allocation and mark the
2615  * buffer_heads mapped.
2616  *
2617  * We redirty the page if we have any buffer_heads that is either delay or
2618  * unwritten in the page.
2619  *
2620  * We can get recursively called as show below.
2621  *
2622  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2623  *              ext4_writepage()
2624  *
2625  * But since we don't do any block allocation we should not deadlock.
2626  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2627  */
2628 static int ext4_writepage(struct page *page,
2629                           struct writeback_control *wbc)
2630 {
2631         int ret = 0, commit_write = 0;
2632         loff_t size;
2633         unsigned int len;
2634         struct buffer_head *page_bufs = NULL;
2635         struct inode *inode = page->mapping->host;
2636
2637         trace_ext4_writepage(inode, page);
2638         size = i_size_read(inode);
2639         if (page->index == size >> PAGE_CACHE_SHIFT)
2640                 len = size & ~PAGE_CACHE_MASK;
2641         else
2642                 len = PAGE_CACHE_SIZE;
2643
2644         /*
2645          * If the page does not have buffers (for whatever reason),
2646          * try to create them using __block_write_begin.  If this
2647          * fails, redirty the page and move on.
2648          */
2649         if (!page_has_buffers(page)) {
2650                 if (__block_write_begin(page, 0, len,
2651                                         noalloc_get_block_write)) {
2652                 redirty_page:
2653                         redirty_page_for_writepage(wbc, page);
2654                         unlock_page(page);
2655                         return 0;
2656                 }
2657                 commit_write = 1;
2658         }
2659         page_bufs = page_buffers(page);
2660         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2661                               ext4_bh_delay_or_unwritten)) {
2662                 /*
2663                  * We don't want to do block allocation, so redirty
2664                  * the page and return.  We may reach here when we do
2665                  * a journal commit via journal_submit_inode_data_buffers.
2666                  * We can also reach here via shrink_page_list
2667                  */
2668                 goto redirty_page;
2669         }
2670         if (commit_write)
2671                 /* now mark the buffer_heads as dirty and uptodate */
2672                 block_commit_write(page, 0, len);
2673
2674         if (PageChecked(page) && ext4_should_journal_data(inode))
2675                 /*
2676                  * It's mmapped pagecache.  Add buffers and journal it.  There
2677                  * doesn't seem much point in redirtying the page here.
2678                  */
2679                 return __ext4_journalled_writepage(page, len);
2680
2681         if (buffer_uninit(page_bufs)) {
2682                 ext4_set_bh_endio(page_bufs, inode);
2683                 ret = block_write_full_page_endio(page, noalloc_get_block_write,
2684                                             wbc, ext4_end_io_buffer_write);
2685         } else
2686                 ret = block_write_full_page(page, noalloc_get_block_write,
2687                                             wbc);
2688
2689         return ret;
2690 }
2691
2692 /*
2693  * This is called via ext4_da_writepages() to
2694  * calculate the total number of credits to reserve to fit
2695  * a single extent allocation into a single transaction,
2696  * ext4_da_writpeages() will loop calling this before
2697  * the block allocation.
2698  */
2699
2700 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2701 {
2702         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2703
2704         /*
2705          * With non-extent format the journal credit needed to
2706          * insert nrblocks contiguous block is dependent on
2707          * number of contiguous block. So we will limit
2708          * number of contiguous block to a sane value
2709          */
2710         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) &&
2711             (max_blocks > EXT4_MAX_TRANS_DATA))
2712                 max_blocks = EXT4_MAX_TRANS_DATA;
2713
2714         return ext4_chunk_trans_blocks(inode, max_blocks);
2715 }
2716
2717 /*
2718  * write_cache_pages_da - walk the list of dirty pages of the given
2719  * address space and accumulate pages that need writing, and call
2720  * mpage_da_map_and_submit to map a single contiguous memory region
2721  * and then write them.
2722  */
2723 static int write_cache_pages_da(struct address_space *mapping,
2724                                 struct writeback_control *wbc,
2725                                 struct mpage_da_data *mpd,
2726                                 pgoff_t *done_index)
2727 {
2728         struct buffer_head      *bh, *head;
2729         struct inode            *inode = mapping->host;
2730         struct pagevec          pvec;
2731         unsigned int            nr_pages;
2732         sector_t                logical;
2733         pgoff_t                 index, end;
2734         long                    nr_to_write = wbc->nr_to_write;
2735         int                     i, tag, ret = 0;
2736
2737         memset(mpd, 0, sizeof(struct mpage_da_data));
2738         mpd->wbc = wbc;
2739         mpd->inode = inode;
2740         pagevec_init(&pvec, 0);
2741         index = wbc->range_start >> PAGE_CACHE_SHIFT;
2742         end = wbc->range_end >> PAGE_CACHE_SHIFT;
2743
2744         if (wbc->sync_mode == WB_SYNC_ALL)
2745                 tag = PAGECACHE_TAG_TOWRITE;
2746         else
2747                 tag = PAGECACHE_TAG_DIRTY;
2748
2749         *done_index = index;
2750         while (index <= end) {
2751                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2752                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2753                 if (nr_pages == 0)
2754                         return 0;
2755
2756                 for (i = 0; i < nr_pages; i++) {
2757                         struct page *page = pvec.pages[i];
2758
2759                         /*
2760                          * At this point, the page may be truncated or
2761                          * invalidated (changing page->mapping to NULL), or
2762                          * even swizzled back from swapper_space to tmpfs file
2763                          * mapping. However, page->index will not change
2764                          * because we have a reference on the page.
2765                          */
2766                         if (page->index > end)
2767                                 goto out;
2768
2769                         *done_index = page->index + 1;
2770
2771                         /*
2772                          * If we can't merge this page, and we have
2773                          * accumulated an contiguous region, write it
2774                          */
2775                         if ((mpd->next_page != page->index) &&
2776                             (mpd->next_page != mpd->first_page)) {
2777                                 mpage_da_map_and_submit(mpd);
2778                                 goto ret_extent_tail;
2779                         }
2780
2781                         lock_page(page);
2782
2783                         /*
2784                          * If the page is no longer dirty, or its
2785                          * mapping no longer corresponds to inode we
2786                          * are writing (which means it has been
2787                          * truncated or invalidated), or the page is
2788                          * already under writeback and we are not
2789                          * doing a data integrity writeback, skip the page
2790                          */
2791                         if (!PageDirty(page) ||
2792                             (PageWriteback(page) &&
2793                              (wbc->sync_mode == WB_SYNC_NONE)) ||
2794                             unlikely(page->mapping != mapping)) {
2795                                 unlock_page(page);
2796                                 continue;
2797                         }
2798
2799                         wait_on_page_writeback(page);
2800                         BUG_ON(PageWriteback(page));
2801
2802                         if (mpd->next_page != page->index)
2803                                 mpd->first_page = page->index;
2804                         mpd->next_page = page->index + 1;
2805                         logical = (sector_t) page->index <<
2806                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2807
2808                         if (!page_has_buffers(page)) {
2809                                 mpage_add_bh_to_extent(mpd, logical,
2810                                                        PAGE_CACHE_SIZE,
2811                                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2812                                 if (mpd->io_done)
2813                                         goto ret_extent_tail;
2814                         } else {
2815                                 /*
2816                                  * Page with regular buffer heads,
2817                                  * just add all dirty ones
2818                                  */
2819                                 head = page_buffers(page);
2820                                 bh = head;
2821                                 do {
2822                                         BUG_ON(buffer_locked(bh));
2823                                         /*
2824                                          * We need to try to allocate
2825                                          * unmapped blocks in the same page.
2826                                          * Otherwise we won't make progress
2827                                          * with the page in ext4_writepage
2828                                          */
2829                                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2830                                                 mpage_add_bh_to_extent(mpd, logical,
2831                                                                        bh->b_size,
2832                                                                        bh->b_state);
2833                                                 if (mpd->io_done)
2834                                                         goto ret_extent_tail;
2835                                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2836                                                 /*
2837                                                  * mapped dirty buffer. We need
2838                                                  * to update the b_state
2839                                                  * because we look at b_state
2840                                                  * in mpage_da_map_blocks.  We
2841                                                  * don't update b_size because
2842                                                  * if we find an unmapped
2843                                                  * buffer_head later we need to
2844                                                  * use the b_state flag of that
2845                                                  * buffer_head.
2846                                                  */
2847                                                 if (mpd->b_size == 0)
2848                                                         mpd->b_state = bh->b_state & BH_FLAGS;
2849                                         }
2850                                         logical++;
2851                                 } while ((bh = bh->b_this_page) != head);
2852                         }
2853
2854                         if (nr_to_write > 0) {
2855                                 nr_to_write--;
2856                                 if (nr_to_write == 0 &&
2857                                     wbc->sync_mode == WB_SYNC_NONE)
2858                                         /*
2859                                          * We stop writing back only if we are
2860                                          * not doing integrity sync. In case of
2861                                          * integrity sync we have to keep going
2862                                          * because someone may be concurrently
2863                                          * dirtying pages, and we might have
2864                                          * synced a lot of newly appeared dirty
2865                                          * pages, but have not synced all of the
2866                                          * old dirty pages.
2867                                          */
2868                                         goto out;
2869                         }
2870                 }
2871                 pagevec_release(&pvec);
2872                 cond_resched();
2873         }
2874         return 0;
2875 ret_extent_tail:
2876         ret = MPAGE_DA_EXTENT_TAIL;
2877 out:
2878         pagevec_release(&pvec);
2879         cond_resched();
2880         return ret;
2881 }
2882
2883
2884 static int ext4_da_writepages(struct address_space *mapping,
2885                               struct writeback_control *wbc)
2886 {
2887         pgoff_t index;
2888         int range_whole = 0;
2889         handle_t *handle = NULL;
2890         struct mpage_da_data mpd;
2891         struct inode *inode = mapping->host;
2892         int pages_written = 0;
2893         unsigned int max_pages;
2894         int range_cyclic, cycled = 1, io_done = 0;
2895         int needed_blocks, ret = 0;
2896         long desired_nr_to_write, nr_to_writebump = 0;
2897         loff_t range_start = wbc->range_start;
2898         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2899         pgoff_t done_index = 0;
2900         pgoff_t end;
2901
2902         trace_ext4_da_writepages(inode, wbc);
2903
2904         /*
2905          * No pages to write? This is mainly a kludge to avoid starting
2906          * a transaction for special inodes like journal inode on last iput()
2907          * because that could violate lock ordering on umount
2908          */
2909         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2910                 return 0;
2911
2912         /*
2913          * If the filesystem has aborted, it is read-only, so return
2914          * right away instead of dumping stack traces later on that
2915          * will obscure the real source of the problem.  We test
2916          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2917          * the latter could be true if the filesystem is mounted
2918          * read-only, and in that case, ext4_da_writepages should
2919          * *never* be called, so if that ever happens, we would want
2920          * the stack trace.
2921          */
2922         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2923                 return -EROFS;
2924
2925         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2926                 range_whole = 1;
2927
2928         range_cyclic = wbc->range_cyclic;
2929         if (wbc->range_cyclic) {
2930                 index = mapping->writeback_index;
2931                 if (index)
2932                         cycled = 0;
2933                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2934                 wbc->range_end  = LLONG_MAX;
2935                 wbc->range_cyclic = 0;
2936                 end = -1;
2937         } else {
2938                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2939                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2940         }
2941
2942         /*
2943          * This works around two forms of stupidity.  The first is in
2944          * the writeback code, which caps the maximum number of pages
2945          * written to be 1024 pages.  This is wrong on multiple
2946          * levels; different architectues have a different page size,
2947          * which changes the maximum amount of data which gets
2948          * written.  Secondly, 4 megabytes is way too small.  XFS
2949          * forces this value to be 16 megabytes by multiplying
2950          * nr_to_write parameter by four, and then relies on its
2951          * allocator to allocate larger extents to make them
2952          * contiguous.  Unfortunately this brings us to the second
2953          * stupidity, which is that ext4's mballoc code only allocates
2954          * at most 2048 blocks.  So we force contiguous writes up to
2955          * the number of dirty blocks in the inode, or
2956          * sbi->max_writeback_mb_bump whichever is smaller.
2957          */
2958         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2959         if (!range_cyclic && range_whole) {
2960                 if (wbc->nr_to_write == LONG_MAX)
2961                         desired_nr_to_write = wbc->nr_to_write;
2962                 else
2963                         desired_nr_to_write = wbc->nr_to_write * 8;
2964         } else
2965                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2966                                                            max_pages);
2967         if (desired_nr_to_write > max_pages)
2968                 desired_nr_to_write = max_pages;
2969
2970         if (wbc->nr_to_write < desired_nr_to_write) {
2971                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2972                 wbc->nr_to_write = desired_nr_to_write;
2973         }
2974
2975 retry:
2976         if (wbc->sync_mode == WB_SYNC_ALL)
2977                 tag_pages_for_writeback(mapping, index, end);
2978
2979         while (!ret && wbc->nr_to_write > 0) {
2980
2981                 /*
2982                  * we  insert one extent at a time. So we need
2983                  * credit needed for single extent allocation.
2984                  * journalled mode is currently not supported
2985                  * by delalloc
2986                  */
2987                 BUG_ON(ext4_should_journal_data(inode));
2988                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2989
2990                 /* start a new transaction*/
2991                 handle = ext4_journal_start(inode, needed_blocks);
2992                 if (IS_ERR(handle)) {
2993                         ret = PTR_ERR(handle);
2994                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2995                                "%ld pages, ino %lu; err %d", __func__,
2996                                 wbc->nr_to_write, inode->i_ino, ret);
2997                         goto out_writepages;
2998                 }
2999
3000                 /*
3001                  * Now call write_cache_pages_da() to find the next
3002                  * contiguous region of logical blocks that need
3003                  * blocks to be allocated by ext4 and submit them.
3004                  */
3005                 ret = write_cache_pages_da(mapping, wbc, &mpd, &done_index);
3006                 /*
3007                  * If we have a contiguous extent of pages and we
3008                  * haven't done the I/O yet, map the blocks and submit
3009                  * them for I/O.
3010                  */
3011                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
3012                         mpage_da_map_and_submit(&mpd);
3013                         ret = MPAGE_DA_EXTENT_TAIL;
3014                 }
3015                 trace_ext4_da_write_pages(inode, &mpd);
3016                 wbc->nr_to_write -= mpd.pages_written;
3017
3018                 ext4_journal_stop(handle);
3019
3020                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
3021                         /* commit the transaction which would
3022                          * free blocks released in the transaction
3023                          * and try again
3024                          */
3025                         jbd2_journal_force_commit_nested(sbi->s_journal);
3026                         ret = 0;
3027                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
3028                         /*
3029                          * got one extent now try with
3030                          * rest of the pages
3031                          */
3032                         pages_written += mpd.pages_written;
3033                         ret = 0;
3034                         io_done = 1;
3035                 } else if (wbc->nr_to_write)
3036                         /*
3037                          * There is no more writeout needed
3038                          * or we requested for a noblocking writeout
3039                          * and we found the device congested
3040                          */
3041                         break;
3042         }
3043         if (!io_done && !cycled) {
3044                 cycled = 1;
3045                 index = 0;
3046                 wbc->range_start = index << PAGE_CACHE_SHIFT;
3047                 wbc->range_end  = mapping->writeback_index - 1;
3048                 goto retry;
3049         }
3050
3051         /* Update index */
3052         wbc->range_cyclic = range_cyclic;
3053         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
3054                 /*
3055                  * set the writeback_index so that range_cyclic
3056                  * mode will write it back later
3057                  */
3058                 mapping->writeback_index = done_index;
3059
3060 out_writepages:
3061         wbc->nr_to_write -= nr_to_writebump;
3062         wbc->range_start = range_start;
3063         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
3064         return ret;
3065 }
3066
3067 #define FALL_BACK_TO_NONDELALLOC 1
3068 static int ext4_nonda_switch(struct super_block *sb)
3069 {
3070         s64 free_blocks, dirty_blocks;
3071         struct ext4_sb_info *sbi = EXT4_SB(sb);
3072
3073         /*
3074          * switch to non delalloc mode if we are running low
3075          * on free block. The free block accounting via percpu
3076          * counters can get slightly wrong with percpu_counter_batch getting
3077          * accumulated on each CPU without updating global counters
3078          * Delalloc need an accurate free block accounting. So switch
3079          * to non delalloc when we are near to error range.
3080          */
3081         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
3082         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
3083         if (2 * free_blocks < 3 * dirty_blocks ||
3084                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
3085                 /*
3086                  * free block count is less than 150% of dirty blocks
3087                  * or free blocks is less than watermark
3088                  */
3089                 return 1;
3090         }
3091         /*
3092          * Even if we don't switch but are nearing capacity,
3093          * start pushing delalloc when 1/2 of free blocks are dirty.
3094          */
3095         if (free_blocks < 2 * dirty_blocks)
3096                 writeback_inodes_sb_if_idle(sb);
3097
3098         return 0;
3099 }
3100
3101 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3102                                loff_t pos, unsigned len, unsigned flags,
3103                                struct page **pagep, void **fsdata)
3104 {
3105         int ret, retries = 0;
3106         struct page *page;
3107         pgoff_t index;
3108         struct inode *inode = mapping->host;
3109         handle_t *handle;
3110
3111         index = pos >> PAGE_CACHE_SHIFT;
3112
3113         if (ext4_nonda_switch(inode->i_sb)) {
3114                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3115                 return ext4_write_begin(file, mapping, pos,
3116                                         len, flags, pagep, fsdata);
3117         }
3118         *fsdata = (void *)0;
3119         trace_ext4_da_write_begin(inode, pos, len, flags);
3120 retry:
3121         /*
3122          * With delayed allocation, we don't log the i_disksize update
3123          * if there is delayed block allocation. But we still need
3124          * to journalling the i_disksize update if writes to the end
3125          * of file which has an already mapped buffer.
3126          */
3127         handle = ext4_journal_start(inode, 1);
3128         if (IS_ERR(handle)) {
3129                 ret = PTR_ERR(handle);
3130                 goto out;
3131         }
3132         /* We cannot recurse into the filesystem as the transaction is already
3133          * started */
3134         flags |= AOP_FLAG_NOFS;
3135
3136         page = grab_cache_page_write_begin(mapping, index, flags);
3137         if (!page) {
3138                 ext4_journal_stop(handle);
3139                 ret = -ENOMEM;
3140                 goto out;
3141         }
3142         *pagep = page;
3143
3144         ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
3145         if (ret < 0) {
3146                 unlock_page(page);
3147                 ext4_journal_stop(handle);
3148                 page_cache_release(page);
3149                 /*
3150                  * block_write_begin may have instantiated a few blocks
3151                  * outside i_size.  Trim these off again. Don't need
3152                  * i_size_read because we hold i_mutex.
3153                  */
3154                 if (pos + len > inode->i_size)
3155                         ext4_truncate_failed_write(inode);
3156         }
3157
3158         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3159                 goto retry;
3160 out:
3161         return ret;
3162 }
3163
3164 /*
3165  * Check if we should update i_disksize
3166  * when write to the end of file but not require block allocation
3167  */
3168 static int ext4_da_should_update_i_disksize(struct page *page,
3169                                             unsigned long offset)
3170 {
3171         struct buffer_head *bh;
3172         struct inode *inode = page->mapping->host;
3173         unsigned int idx;
3174         int i;
3175
3176         bh = page_buffers(page);
3177         idx = offset >> inode->i_blkbits;
3178
3179         for (i = 0; i < idx; i++)
3180                 bh = bh->b_this_page;
3181
3182         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3183                 return 0;
3184         return 1;
3185 }
3186
3187 static int ext4_da_write_end(struct file *file,
3188                              struct address_space *mapping,
3189                              loff_t pos, unsigned len, unsigned copied,
3190                              struct page *page, void *fsdata)
3191 {
3192         struct inode *inode = mapping->host;
3193         int ret = 0, ret2;
3194         handle_t *handle = ext4_journal_current_handle();
3195         loff_t new_i_size;
3196         unsigned long start, end;
3197         int write_mode = (int)(unsigned long)fsdata;
3198
3199         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
3200                 if (ext4_should_order_data(inode)) {
3201                         return ext4_ordered_write_end(file, mapping, pos,
3202                                         len, copied, page, fsdata);
3203                 } else if (ext4_should_writeback_data(inode)) {
3204                         return ext4_writeback_write_end(file, mapping, pos,
3205                                         len, copied, page, fsdata);
3206                 } else {
3207                         BUG();
3208                 }
3209         }
3210
3211         trace_ext4_da_write_end(inode, pos, len, copied);
3212         start = pos & (PAGE_CACHE_SIZE - 1);
3213         end = start + copied - 1;
3214
3215         /*
3216          * generic_write_end() will run mark_inode_dirty() if i_size
3217          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3218          * into that.
3219          */
3220
3221         new_i_size = pos + copied;
3222         if (new_i_size > EXT4_I(inode)->i_disksize) {
3223                 if (ext4_da_should_update_i_disksize(page, end)) {
3224                         down_write(&EXT4_I(inode)->i_data_sem);
3225                         if (new_i_size > EXT4_I(inode)->i_disksize) {
3226                                 /*
3227                                  * Updating i_disksize when extending file
3228                                  * without needing block allocation
3229                                  */
3230                                 if (ext4_should_order_data(inode))
3231                                         ret = ext4_jbd2_file_inode(handle,
3232                                                                    inode);
3233
3234                                 EXT4_I(inode)->i_disksize = new_i_size;
3235                         }
3236                         up_write(&EXT4_I(inode)->i_data_sem);
3237                         /* We need to mark inode dirty even if
3238                          * new_i_size is less that inode->i_size
3239                          * bu greater than i_disksize.(hint delalloc)
3240                          */
3241                         ext4_mark_inode_dirty(handle, inode);
3242                 }
3243         }
3244         ret2 = generic_write_end(file, mapping, pos, len, copied,
3245                                                         page, fsdata);
3246         copied = ret2;
3247         if (ret2 < 0)
3248                 ret = ret2;
3249         ret2 = ext4_journal_stop(handle);
3250         if (!ret)
3251                 ret = ret2;
3252
3253         return ret ? ret : copied;
3254 }
3255
3256 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
3257 {
3258         /*
3259          * Drop reserved blocks
3260          */
3261         BUG_ON(!PageLocked(page));
3262         if (!page_has_buffers(page))
3263                 goto out;
3264
3265         ext4_da_page_release_reservation(page, offset);
3266
3267 out:
3268         ext4_invalidatepage(page, offset);
3269
3270         return;
3271 }
3272
3273 /*
3274  * Force all delayed allocation blocks to be allocated for a given inode.
3275  */
3276 int ext4_alloc_da_blocks(struct inode *inode)
3277 {
3278         trace_ext4_alloc_da_blocks(inode);
3279
3280         if (!EXT4_I(inode)->i_reserved_data_blocks &&
3281             !EXT4_I(inode)->i_reserved_meta_blocks)
3282                 return 0;
3283
3284         /*
3285          * We do something simple for now.  The filemap_flush() will
3286          * also start triggering a write of the data blocks, which is
3287          * not strictly speaking necessary (and for users of
3288          * laptop_mode, not even desirable).  However, to do otherwise
3289          * would require replicating code paths in:
3290          *
3291          * ext4_da_writepages() ->
3292          *    write_cache_pages() ---> (via passed in callback function)
3293          *        __mpage_da_writepage() -->
3294          *           mpage_add_bh_to_extent()
3295          *           mpage_da_map_blocks()
3296          *
3297          * The problem is that write_cache_pages(), located in
3298          * mm/page-writeback.c, marks pages clean in preparation for
3299          * doing I/O, which is not desirable if we're not planning on
3300          * doing I/O at all.
3301          *
3302          * We could call write_cache_pages(), and then redirty all of
3303          * the pages by calling redirty_page_for_writepage() but that
3304          * would be ugly in the extreme.  So instead we would need to
3305          * replicate parts of the code in the above functions,
3306          * simplifying them because we wouldn't actually intend to
3307          * write out the pages, but rather only collect contiguous
3308          * logical block extents, call the multi-block allocator, and
3309          * then update the buffer heads with the block allocations.
3310          *
3311          * For now, though, we'll cheat by calling filemap_flush(),
3312          * which will map the blocks, and start the I/O, but not
3313          * actually wait for the I/O to complete.
3314          */
3315         return filemap_flush(inode->i_mapping);
3316 }
3317
3318 /*
3319  * bmap() is special.  It gets used by applications such as lilo and by
3320  * the swapper to find the on-disk block of a specific piece of data.
3321  *
3322  * Naturally, this is dangerous if the block concerned is still in the
3323  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3324  * filesystem and enables swap, then they may get a nasty shock when the
3325  * data getting swapped to that swapfile suddenly gets overwritten by
3326  * the original zero's written out previously to the journal and
3327  * awaiting writeback in the kernel's buffer cache.
3328  *
3329  * So, if we see any bmap calls here on a modified, data-journaled file,
3330  * take extra steps to flush any blocks which might be in the cache.
3331  */
3332 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3333 {
3334         struct inode *inode = mapping->host;
3335         journal_t *journal;
3336         int err;
3337
3338         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3339                         test_opt(inode->i_sb, DELALLOC)) {
3340                 /*
3341                  * With delalloc we want to sync the file
3342                  * so that we can make sure we allocate
3343                  * blocks for file
3344                  */
3345                 filemap_write_and_wait(mapping);
3346         }
3347
3348         if (EXT4_JOURNAL(inode) &&
3349             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3350                 /*
3351                  * This is a REALLY heavyweight approach, but the use of
3352                  * bmap on dirty files is expected to be extremely rare:
3353                  * only if we run lilo or swapon on a freshly made file
3354                  * do we expect this to happen.
3355                  *
3356                  * (bmap requires CAP_SYS_RAWIO so this does not
3357                  * represent an unprivileged user DOS attack --- we'd be
3358                  * in trouble if mortal users could trigger this path at
3359                  * will.)
3360                  *
3361                  * NB. EXT4_STATE_JDATA is not set on files other than
3362                  * regular files.  If somebody wants to bmap a directory
3363                  * or symlink and gets confused because the buffer
3364                  * hasn't yet been flushed to disk, they deserve
3365                  * everything they get.
3366                  */
3367
3368                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3369                 journal = EXT4_JOURNAL(inode);
3370                 jbd2_journal_lock_updates(journal);
3371                 err = jbd2_journal_flush(journal);
3372                 jbd2_journal_unlock_updates(journal);
3373
3374                 if (err)
3375                         return 0;
3376         }
3377
3378         return generic_block_bmap(mapping, block, ext4_get_block);
3379 }
3380
3381 static int ext4_readpage(struct file *file, struct page *page)
3382 {
3383         trace_ext4_readpage(page);
3384         return mpage_readpage(page, ext4_get_block);
3385 }
3386
3387 static int
3388 ext4_readpages(struct file *file, struct address_space *mapping,
3389                 struct list_head *pages, unsigned nr_pages)
3390 {
3391         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3392 }
3393
3394 static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
3395 {
3396         struct buffer_head *head, *bh;
3397         unsigned int curr_off = 0;
3398
3399         if (!page_has_buffers(page))
3400                 return;
3401         head = bh = page_buffers(page);
3402         do {
3403                 if (offset <= curr_off && test_clear_buffer_uninit(bh)
3404                                         && bh->b_private) {
3405                         ext4_free_io_end(bh->b_private);
3406                         bh->b_private = NULL;
3407                         bh->b_end_io = NULL;
3408                 }
3409                 curr_off = curr_off + bh->b_size;
3410                 bh = bh->b_this_page;
3411         } while (bh != head);
3412 }
3413
3414 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3415 {
3416         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3417
3418         trace_ext4_invalidatepage(page, offset);
3419
3420         /*
3421          * free any io_end structure allocated for buffers to be discarded
3422          */
3423         if (ext4_should_dioread_nolock(page->mapping->host))
3424                 ext4_invalidatepage_free_endio(page, offset);
3425         /*
3426          * If it's a full truncate we just forget about the pending dirtying
3427          */
3428         if (offset == 0)
3429                 ClearPageChecked(page);
3430
3431         if (journal)
3432                 jbd2_journal_invalidatepage(journal, page, offset);
3433         else
3434                 block_invalidatepage(page, offset);
3435 }
3436
3437 static int ext4_releasepage(struct page *page, gfp_t wait)
3438 {
3439         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3440
3441         trace_ext4_releasepage(page);
3442
3443         WARN_ON(PageChecked(page));
3444         if (!page_has_buffers(page))
3445                 return 0;
3446         if (journal)
3447                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3448         else
3449                 return try_to_free_buffers(page);
3450 }
3451
3452 /*
3453  * O_DIRECT for ext3 (or indirect map) based files
3454  *
3455  * If the O_DIRECT write will extend the file then add this inode to the
3456  * orphan list.  So recovery will truncate it back to the original size
3457  * if the machine crashes during the write.
3458  *
3459  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3460  * crashes then stale disk data _may_ be exposed inside the file. But current
3461  * VFS code falls back into buffered path in that case so we are safe.
3462  */
3463 static ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
3464                               const struct iovec *iov, loff_t offset,
3465                               unsigned long nr_segs)
3466 {
3467         struct file *file = iocb->ki_filp;
3468         struct inode *inode = file->f_mapping->host;
3469         struct ext4_inode_info *ei = EXT4_I(inode);
3470         handle_t *handle;
3471         ssize_t ret;
3472         int orphan = 0;
3473         size_t count = iov_length(iov, nr_segs);
3474         int retries = 0;
3475
3476         if (rw == WRITE) {
3477                 loff_t final_size = offset + count;
3478
3479                 if (final_size > inode->i_size) {
3480                         /* Credits for sb + inode write */
3481                         handle = ext4_journal_start(inode, 2);
3482                         if (IS_ERR(handle)) {
3483                                 ret = PTR_ERR(handle);
3484                                 goto out;
3485                         }
3486                         ret = ext4_orphan_add(handle, inode);
3487                         if (ret) {
3488                                 ext4_journal_stop(handle);
3489                                 goto out;
3490                         }
3491                         orphan = 1;
3492                         ei->i_disksize = inode->i_size;
3493                         ext4_journal_stop(handle);
3494                 }
3495         }
3496
3497 retry:
3498         if (rw == READ && ext4_should_dioread_nolock(inode))
3499                 ret = __blockdev_direct_IO(rw, iocb, inode,
3500                                  inode->i_sb->s_bdev, iov,
3501                                  offset, nr_segs,
3502                                  ext4_get_block, NULL, NULL, 0);
3503         else {
3504                 ret = blockdev_direct_IO(rw, iocb, inode,
3505                                  inode->i_sb->s_bdev, iov,
3506                                  offset, nr_segs,
3507                                  ext4_get_block, NULL);
3508
3509                 if (unlikely((rw & WRITE) && ret < 0)) {
3510                         loff_t isize = i_size_read(inode);
3511                         loff_t end = offset + iov_length(iov, nr_segs);
3512
3513                         if (end > isize)
3514                                 ext4_truncate_failed_write(inode);
3515                 }
3516         }
3517         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3518                 goto retry;
3519
3520         if (orphan) {
3521                 int err;
3522
3523                 /* Credits for sb + inode write */
3524                 handle = ext4_journal_start(inode, 2);
3525                 if (IS_ERR(handle)) {
3526                         /* This is really bad luck. We've written the data
3527                          * but cannot extend i_size. Bail out and pretend
3528                          * the write failed... */
3529                         ret = PTR_ERR(handle);
3530                         if (inode->i_nlink)
3531                                 ext4_orphan_del(NULL, inode);
3532
3533                         goto out;
3534                 }
3535                 if (inode->i_nlink)
3536                         ext4_orphan_del(handle, inode);
3537                 if (ret > 0) {
3538                         loff_t end = offset + ret;
3539                         if (end > inode->i_size) {
3540                                 ei->i_disksize = end;
3541                                 i_size_write(inode, end);
3542                                 /*
3543                                  * We're going to return a positive `ret'
3544                                  * here due to non-zero-length I/O, so there's
3545                                  * no way of reporting error returns from
3546                                  * ext4_mark_inode_dirty() to userspace.  So
3547                                  * ignore it.
3548                                  */
3549                                 ext4_mark_inode_dirty(handle, inode);
3550                         }
3551                 }
3552                 err = ext4_journal_stop(handle);
3553                 if (ret == 0)
3554                         ret = err;
3555         }
3556 out:
3557         return ret;
3558 }
3559
3560 /*
3561  * ext4_get_block used when preparing for a DIO write or buffer write.
3562  * We allocate an uinitialized extent if blocks haven't been allocated.
3563  * The extent will be converted to initialized after the IO is complete.
3564  */
3565 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
3566                    struct buffer_head *bh_result, int create)
3567 {
3568         ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
3569                    inode->i_ino, create);
3570         return _ext4_get_block(inode, iblock, bh_result,
3571                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
3572 }
3573
3574 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3575                             ssize_t size, void *private, int ret,
3576                             bool is_async)
3577 {
3578         ext4_io_end_t *io_end = iocb->private;
3579         struct workqueue_struct *wq;
3580         unsigned long flags;
3581         struct ext4_inode_info *ei;
3582
3583         /* if not async direct IO or dio with 0 bytes write, just return */
3584         if (!io_end || !size)
3585                 goto out;
3586
3587         ext_debug("ext4_end_io_dio(): io_end 0x%p"
3588                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
3589                   iocb->private, io_end->inode->i_ino, iocb, offset,
3590                   size);
3591
3592         /* if not aio dio with unwritten extents, just free io and return */
3593         if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
3594                 ext4_free_io_end(io_end);
3595                 iocb->private = NULL;
3596 out:
3597                 if (is_async)
3598                         aio_complete(iocb, ret, 0);
3599                 return;
3600         }
3601
3602         io_end->offset = offset;
3603         io_end->size = size;
3604         if (is_async) {
3605                 io_end->iocb = iocb;
3606                 io_end->result = ret;
3607         }
3608         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
3609
3610         /* Add the io_end to per-inode completed aio dio list*/
3611         ei = EXT4_I(io_end->inode);
3612         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
3613         list_add_tail(&io_end->list, &ei->i_completed_io_list);
3614         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
3615
3616         /* queue the work to convert unwritten extents to written */
3617         queue_work(wq, &io_end->work);
3618         iocb->private = NULL;
3619 }
3620
3621 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
3622 {
3623         ext4_io_end_t *io_end = bh->b_private;
3624         struct workqueue_struct *wq;
3625         struct inode *inode;
3626         unsigned long flags;
3627
3628         if (!test_clear_buffer_uninit(bh) || !io_end)
3629                 goto out;
3630
3631         if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
3632                 printk("sb umounted, discard end_io request for inode %lu\n",
3633                         io_end->inode->i_ino);
3634                 ext4_free_io_end(io_end);
3635                 goto out;
3636         }
3637
3638         io_end->flag = EXT4_IO_END_UNWRITTEN;
3639         inode = io_end->inode;
3640
3641         /* Add the io_end to per-inode completed io list*/
3642         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
3643         list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
3644         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
3645
3646         wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
3647         /* queue the work to convert unwritten extents to written */
3648         queue_work(wq, &io_end->work);
3649 out:
3650         bh->b_private = NULL;
3651         bh->b_end_io = NULL;
3652         clear_buffer_uninit(bh);
3653         end_buffer_async_write(bh, uptodate);
3654 }
3655
3656 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
3657 {
3658         ext4_io_end_t *io_end;
3659         struct page *page = bh->b_page;
3660         loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
3661         size_t size = bh->b_size;
3662
3663 retry:
3664         io_end = ext4_init_io_end(inode, GFP_ATOMIC);
3665         if (!io_end) {
3666                 pr_warn_ratelimited("%s: allocation fail\n", __func__);
3667                 schedule();
3668                 goto retry;
3669         }
3670         io_end->offset = offset;
3671         io_end->size = size;
3672         /*
3673          * We need to hold a reference to the page to make sure it
3674          * doesn't get evicted before ext4_end_io_work() has a chance
3675          * to convert the extent from written to unwritten.
3676          */
3677         io_end->page = page;
3678         get_page(io_end->page);
3679
3680         bh->b_private = io_end;
3681         bh->b_end_io = ext4_end_io_buffer_write;
3682         return 0;
3683 }
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 uninitialized
3691  * If those blocks were preallocated, we mark sure they are splited, but
3692  * still keep the range to write as uninitialized.
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 conversion
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 parallel 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, GFP_NOFS);
3739                         if (!iocb->private)
3740                                 return -ENOMEM;
3741                         /*
3742                          * we save the io structure for current async
3743                          * direct IO, so that later ext4_map_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_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 && ext4_test_inode_state(inode,
3776                                                 EXT4_STATE_DIO_UNWRITTEN)) {
3777                         int err;
3778                         /*
3779                          * for non AIO case, since the IO is already
3780                          * completed, we could do the conversion right here
3781                          */
3782                         err = ext4_convert_unwritten_extents(inode,
3783                                                              offset, ret);
3784                         if (err < 0)
3785                                 ret = err;
3786                         ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3787                 }
3788                 return ret;
3789         }
3790
3791         /* for write the the end of file case, we fall back to old way */
3792         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3793 }
3794
3795 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3796                               const struct iovec *iov, loff_t offset,
3797                               unsigned long nr_segs)
3798 {
3799         struct file *file = iocb->ki_filp;
3800         struct inode *inode = file->f_mapping->host;
3801         ssize_t ret;
3802
3803         trace_ext4_direct_IO_enter(inode, offset, iov_length(iov, nr_segs), rw);
3804         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3805                 ret = ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
3806         else
3807                 ret = ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3808         trace_ext4_direct_IO_exit(inode, offset,
3809                                 iov_length(iov, nr_segs), rw, ret);
3810         return ret;
3811 }
3812
3813 /*
3814  * Pages can be marked dirty completely asynchronously from ext4's journalling
3815  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3816  * much here because ->set_page_dirty is called under VFS locks.  The page is
3817  * not necessarily locked.
3818  *
3819  * We cannot just dirty the page and leave attached buffers clean, because the
3820  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3821  * or jbddirty because all the journalling code will explode.
3822  *
3823  * So what we do is to mark the page "pending dirty" and next time writepage
3824  * is called, propagate that into the buffers appropriately.
3825  */
3826 static int ext4_journalled_set_page_dirty(struct page *page)
3827 {
3828         SetPageChecked(page);
3829         return __set_page_dirty_nobuffers(page);
3830 }
3831
3832 static const struct address_space_operations ext4_ordered_aops = {
3833         .readpage               = ext4_readpage,
3834         .readpages              = ext4_readpages,
3835         .writepage              = ext4_writepage,
3836         .write_begin            = ext4_write_begin,
3837         .write_end              = ext4_ordered_write_end,
3838         .bmap                   = ext4_bmap,
3839         .invalidatepage         = ext4_invalidatepage,
3840         .releasepage            = ext4_releasepage,
3841         .direct_IO              = ext4_direct_IO,
3842         .migratepage            = buffer_migrate_page,
3843         .is_partially_uptodate  = block_is_partially_uptodate,
3844         .error_remove_page      = generic_error_remove_page,
3845 };
3846
3847 static const struct address_space_operations ext4_writeback_aops = {
3848         .readpage               = ext4_readpage,
3849         .readpages              = ext4_readpages,
3850         .writepage              = ext4_writepage,
3851         .write_begin            = ext4_write_begin,
3852         .write_end              = ext4_writeback_write_end,
3853         .bmap                   = ext4_bmap,
3854         .invalidatepage         = ext4_invalidatepage,
3855         .releasepage            = ext4_releasepage,
3856         .direct_IO              = ext4_direct_IO,
3857         .migratepage            = buffer_migrate_page,
3858         .is_partially_uptodate  = block_is_partially_uptodate,
3859         .error_remove_page      = generic_error_remove_page,
3860 };
3861
3862 static const struct address_space_operations ext4_journalled_aops = {
3863         .readpage               = ext4_readpage,
3864         .readpages              = ext4_readpages,
3865         .writepage              = ext4_writepage,
3866         .write_begin            = ext4_write_begin,
3867         .write_end              = ext4_journalled_write_end,
3868         .set_page_dirty         = ext4_journalled_set_page_dirty,
3869         .bmap                   = ext4_bmap,
3870         .invalidatepage         = ext4_invalidatepage,
3871         .releasepage            = ext4_releasepage,
3872         .is_partially_uptodate  = block_is_partially_uptodate,
3873         .error_remove_page      = generic_error_remove_page,
3874 };
3875
3876 static const struct address_space_operations ext4_da_aops = {
3877         .readpage               = ext4_readpage,
3878         .readpages              = ext4_readpages,
3879         .writepage              = ext4_writepage,
3880         .writepages             = ext4_da_writepages,
3881         .write_begin            = ext4_da_write_begin,
3882         .write_end              = ext4_da_write_end,
3883         .bmap                   = ext4_bmap,
3884         .invalidatepage         = ext4_da_invalidatepage,
3885         .releasepage            = ext4_releasepage,
3886         .direct_IO              = ext4_direct_IO,
3887         .migratepage            = buffer_migrate_page,
3888         .is_partially_uptodate  = block_is_partially_uptodate,
3889         .error_remove_page      = generic_error_remove_page,
3890 };
3891
3892 void ext4_set_aops(struct inode *inode)
3893 {
3894         if (ext4_should_order_data(inode) &&
3895                 test_opt(inode->i_sb, DELALLOC))
3896                 inode->i_mapping->a_ops = &ext4_da_aops;
3897         else if (ext4_should_order_data(inode))
3898                 inode->i_mapping->a_ops = &ext4_ordered_aops;
3899         else if (ext4_should_writeback_data(inode) &&
3900                  test_opt(inode->i_sb, DELALLOC))
3901                 inode->i_mapping->a_ops = &ext4_da_aops;
3902         else if (ext4_should_writeback_data(inode))
3903                 inode->i_mapping->a_ops = &ext4_writeback_aops;
3904         else
3905                 inode->i_mapping->a_ops = &ext4_journalled_aops;
3906 }
3907
3908 /*
3909  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3910  * up to the end of the block which corresponds to `from'.
3911  * This required during truncate. We need to physically zero the tail end
3912  * of that block so it doesn't yield old data if the file is later grown.
3913  */
3914 int ext4_block_truncate_page(handle_t *handle,
3915                 struct address_space *mapping, loff_t from)
3916 {
3917         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3918         unsigned length;
3919         unsigned blocksize;
3920         struct inode *inode = mapping->host;
3921
3922         blocksize = inode->i_sb->s_blocksize;
3923         length = blocksize - (offset & (blocksize - 1));
3924
3925         return ext4_block_zero_page_range(handle, mapping, from, length);
3926 }
3927
3928 /*
3929  * ext4_block_zero_page_range() zeros out a mapping of length 'length'
3930  * starting from file offset 'from'.  The range to be zero'd must
3931  * be contained with in one block.  If the specified range exceeds
3932  * the end of the block it will be shortened to end of the block
3933  * that cooresponds to 'from'
3934  */
3935 int ext4_block_zero_page_range(handle_t *handle,
3936                 struct address_space *mapping, loff_t from, loff_t length)
3937 {
3938         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3939         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3940         unsigned blocksize, max, pos;
3941         ext4_lblk_t iblock;
3942         struct inode *inode = mapping->host;
3943         struct buffer_head *bh;
3944         struct page *page;
3945         int err = 0;
3946
3947         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
3948                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
3949         if (!page)
3950                 return -EINVAL;
3951
3952         blocksize = inode->i_sb->s_blocksize;
3953         max = blocksize - (offset & (blocksize - 1));
3954
3955         /*
3956          * correct length if it does not fall between
3957          * 'from' and the end of the block
3958          */
3959         if (length > max || length < 0)
3960                 length = max;
3961
3962         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3963
3964         if (!page_has_buffers(page))
3965                 create_empty_buffers(page, blocksize, 0);
3966
3967         /* Find the buffer that contains "offset" */
3968         bh = page_buffers(page);
3969         pos = blocksize;
3970         while (offset >= pos) {
3971                 bh = bh->b_this_page;
3972                 iblock++;
3973                 pos += blocksize;
3974         }
3975
3976         err = 0;
3977         if (buffer_freed(bh)) {
3978                 BUFFER_TRACE(bh, "freed: skip");
3979                 goto unlock;
3980         }
3981
3982         if (!buffer_mapped(bh)) {
3983                 BUFFER_TRACE(bh, "unmapped");
3984                 ext4_get_block(inode, iblock, bh, 0);
3985                 /* unmapped? It's a hole - nothing to do */
3986                 if (!buffer_mapped(bh)) {
3987                         BUFFER_TRACE(bh, "still unmapped");
3988                         goto unlock;
3989                 }
3990         }
3991
3992         /* Ok, it's mapped. Make sure it's up-to-date */
3993         if (PageUptodate(page))
3994                 set_buffer_uptodate(bh);
3995
3996         if (!buffer_uptodate(bh)) {
3997                 err = -EIO;
3998                 ll_rw_block(READ, 1, &bh);
3999                 wait_on_buffer(bh);
4000                 /* Uhhuh. Read error. Complain and punt. */
4001                 if (!buffer_uptodate(bh))
4002                         goto unlock;
4003         }
4004
4005         if (ext4_should_journal_data(inode)) {
4006                 BUFFER_TRACE(bh, "get write access");
4007                 err = ext4_journal_get_write_access(handle, bh);
4008                 if (err)
4009                         goto unlock;
4010         }
4011
4012         zero_user(page, offset, length);
4013
4014         BUFFER_TRACE(bh, "zeroed end of block");
4015
4016         err = 0;
4017         if (ext4_should_journal_data(inode)) {
4018                 err = ext4_handle_dirty_metadata(handle, inode, bh);
4019         } else {
4020                 if (ext4_should_order_data(inode) && EXT4_I(inode)->jinode)
4021                         err = ext4_jbd2_file_inode(handle, inode);
4022                 mark_buffer_dirty(bh);
4023         }
4024
4025 unlock:
4026         unlock_page(page);
4027         page_cache_release(page);
4028         return err;
4029 }
4030
4031 /*
4032  * Probably it should be a library function... search for first non-zero word
4033  * or memcmp with zero_page, whatever is better for particular architecture.
4034  * Linus?
4035  */
4036 static inline int all_zeroes(__le32 *p, __le32 *q)
4037 {
4038         while (p < q)
4039                 if (*p++)
4040                         return 0;
4041         return 1;
4042 }
4043
4044 /**
4045  *      ext4_find_shared - find the indirect blocks for partial truncation.
4046  *      @inode:   inode in question
4047  *      @depth:   depth of the affected branch
4048  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
4049  *      @chain:   place to store the pointers to partial indirect blocks
4050  *      @top:     place to the (detached) top of branch
4051  *
4052  *      This is a helper function used by ext4_truncate().
4053  *
4054  *      When we do truncate() we may have to clean the ends of several
4055  *      indirect blocks but leave the blocks themselves alive. Block is
4056  *      partially truncated if some data below the new i_size is referred
4057  *      from it (and it is on the path to the first completely truncated
4058  *      data block, indeed).  We have to free the top of that path along
4059  *      with everything to the right of the path. Since no allocation
4060  *      past the truncation point is possible until ext4_truncate()
4061  *      finishes, we may safely do the latter, but top of branch may
4062  *      require special attention - pageout below the truncation point
4063  *      might try to populate it.
4064  *
4065  *      We atomically detach the top of branch from the tree, store the
4066  *      block number of its root in *@top, pointers to buffer_heads of
4067  *      partially truncated blocks - in @chain[].bh and pointers to
4068  *      their last elements that should not be removed - in
4069  *      @chain[].p. Return value is the pointer to last filled element
4070  *      of @chain.
4071  *
4072  *      The work left to caller to do the actual freeing of subtrees:
4073  *              a) free the subtree starting from *@top
4074  *              b) free the subtrees whose roots are stored in
4075  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
4076  *              c) free the subtrees growing from the inode past the @chain[0].
4077  *                      (no partially truncated stuff there).  */
4078
4079 static Indirect *ext4_find_shared(struct inode *inode, int depth,
4080                                   ext4_lblk_t offsets[4], Indirect chain[4],
4081                                   __le32 *top)
4082 {
4083         Indirect *partial, *p;
4084         int k, err;
4085
4086         *top = 0;
4087         /* Make k index the deepest non-null offset + 1 */
4088         for (k = depth; k > 1 && !offsets[k-1]; k--)
4089                 ;
4090         partial = ext4_get_branch(inode, k, offsets, chain, &err);
4091         /* Writer: pointers */
4092         if (!partial)
4093                 partial = chain + k-1;
4094         /*
4095          * If the branch acquired continuation since we've looked at it -
4096          * fine, it should all survive and (new) top doesn't belong to us.
4097          */
4098         if (!partial->key && *partial->p)
4099                 /* Writer: end */
4100                 goto no_top;
4101         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
4102                 ;
4103         /*
4104          * OK, we've found the last block that must survive. The rest of our
4105          * branch should be detached before unlocking. However, if that rest
4106          * of branch is all ours and does not grow immediately from the inode
4107          * it's easier to cheat and just decrement partial->p.
4108          */
4109         if (p == chain + k - 1 && p > chain) {
4110                 p->p--;
4111         } else {
4112                 *top = *p->p;
4113                 /* Nope, don't do this in ext4.  Must leave the tree intact */
4114 #if 0
4115                 *p->p = 0;
4116 #endif
4117         }
4118         /* Writer: end */
4119
4120         while (partial > p) {
4121                 brelse(partial->bh);
4122                 partial--;
4123         }
4124 no_top:
4125         return partial;
4126 }
4127
4128 /*
4129  * Zero a number of block pointers in either an inode or an indirect block.
4130  * If we restart the transaction we must again get write access to the
4131  * indirect block for further modification.
4132  *
4133  * We release `count' blocks on disk, but (last - first) may be greater
4134  * than `count' because there can be holes in there.
4135  *
4136  * Return 0 on success, 1 on invalid block range
4137  * and < 0 on fatal error.
4138  */
4139 static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
4140                              struct buffer_head *bh,
4141                              ext4_fsblk_t block_to_free,
4142                              unsigned long count, __le32 *first,
4143                              __le32 *last)
4144 {
4145         __le32 *p;
4146         int     flags = EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_VALIDATED;
4147         int     err;
4148
4149         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
4150                 flags |= EXT4_FREE_BLOCKS_METADATA;
4151
4152         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
4153                                    count)) {
4154                 EXT4_ERROR_INODE(inode, "attempt to clear invalid "
4155                                  "blocks %llu len %lu",
4156                                  (unsigned long long) block_to_free, count);
4157                 return 1;
4158         }
4159
4160         if (try_to_extend_transaction(handle, inode)) {
4161                 if (bh) {
4162                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4163                         err = ext4_handle_dirty_metadata(handle, inode, bh);
4164                         if (unlikely(err))
4165                                 goto out_err;
4166                 }
4167                 err = ext4_mark_inode_dirty(handle, inode);
4168                 if (unlikely(err))
4169                         goto out_err;
4170                 err = ext4_truncate_restart_trans(handle, inode,
4171                                                   blocks_for_truncate(inode));
4172                 if (unlikely(err))
4173                         goto out_err;
4174                 if (bh) {
4175                         BUFFER_TRACE(bh, "retaking write access");
4176                         err = ext4_journal_get_write_access(handle, bh);
4177                         if (unlikely(err))
4178                                 goto out_err;
4179                 }
4180         }
4181
4182         for (p = first; p < last; p++)
4183                 *p = 0;
4184
4185         ext4_free_blocks(handle, inode, NULL, block_to_free, count, flags);
4186         return 0;
4187 out_err:
4188         ext4_std_error(inode->i_sb, err);
4189         return err;
4190 }
4191
4192 /**
4193  * ext4_free_data - free a list of data blocks
4194  * @handle:     handle for this transaction
4195  * @inode:      inode we are dealing with
4196  * @this_bh:    indirect buffer_head which contains *@first and *@last
4197  * @first:      array of block numbers
4198  * @last:       points immediately past the end of array
4199  *
4200  * We are freeing all blocks referred from that array (numbers are stored as
4201  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
4202  *
4203  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
4204  * blocks are contiguous then releasing them at one time will only affect one
4205  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
4206  * actually use a lot of journal space.
4207  *
4208  * @this_bh will be %NULL if @first and @last point into the inode's direct
4209  * block pointers.
4210  */
4211 static void ext4_free_data(handle_t *handle, struct inode *inode,
4212                            struct buffer_head *this_bh,
4213                            __le32 *first, __le32 *last)
4214 {
4215         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
4216         unsigned long count = 0;            /* Number of blocks in the run */
4217         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
4218                                                corresponding to
4219                                                block_to_free */
4220         ext4_fsblk_t nr;                    /* Current block # */
4221         __le32 *p;                          /* Pointer into inode/ind
4222                                                for current block */
4223         int err = 0;
4224
4225         if (this_bh) {                          /* For indirect block */
4226                 BUFFER_TRACE(this_bh, "get_write_access");
4227                 err = ext4_journal_get_write_access(handle, this_bh);
4228                 /* Important: if we can't update the indirect pointers
4229                  * to the blocks, we can't free them. */
4230                 if (err)
4231                         return;
4232         }
4233
4234         for (p = first; p < last; p++) {
4235                 nr = le32_to_cpu(*p);
4236                 if (nr) {
4237                         /* accumulate blocks to free if they're contiguous */
4238                         if (count == 0) {
4239                                 block_to_free = nr;
4240                                 block_to_free_p = p;
4241                                 count = 1;
4242                         } else if (nr == block_to_free + count) {
4243                                 count++;
4244                         } else {
4245                                 err = ext4_clear_blocks(handle, inode, this_bh,
4246                                                         block_to_free, count,
4247                                                         block_to_free_p, p);
4248                                 if (err)
4249                                         break;
4250                                 block_to_free = nr;
4251                                 block_to_free_p = p;
4252                                 count = 1;
4253                         }
4254                 }
4255         }
4256
4257         if (!err && count > 0)
4258                 err = ext4_clear_blocks(handle, inode, this_bh, block_to_free,
4259                                         count, block_to_free_p, p);
4260         if (err < 0)
4261                 /* fatal error */
4262                 return;
4263
4264         if (this_bh) {
4265                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
4266
4267                 /*
4268                  * The buffer head should have an attached journal head at this
4269                  * point. However, if the data is corrupted and an indirect
4270                  * block pointed to itself, it would have been detached when
4271                  * the block was cleared. Check for this instead of OOPSing.
4272                  */
4273                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
4274                         ext4_handle_dirty_metadata(handle, inode, this_bh);
4275                 else
4276                         EXT4_ERROR_INODE(inode,
4277                                          "circular indirect block detected at "
4278                                          "block %llu",
4279                                 (unsigned long long) this_bh->b_blocknr);
4280         }
4281 }
4282
4283 /**
4284  *      ext4_free_branches - free an array of branches
4285  *      @handle: JBD handle for this transaction
4286  *      @inode: inode we are dealing with
4287  *      @parent_bh: the buffer_head which contains *@first and *@last
4288  *      @first: array of block numbers
4289  *      @last:  pointer immediately past the end of array
4290  *      @depth: depth of the branches to free
4291  *
4292  *      We are freeing all blocks referred from these branches (numbers are
4293  *      stored as little-endian 32-bit) and updating @inode->i_blocks
4294  *      appropriately.
4295  */
4296 static void ext4_free_branches(handle_t *handle, struct inode *inode,
4297                                struct buffer_head *parent_bh,
4298                                __le32 *first, __le32 *last, int depth)
4299 {
4300         ext4_fsblk_t nr;
4301         __le32 *p;
4302
4303         if (ext4_handle_is_aborted(handle))
4304                 return;
4305
4306         if (depth--) {
4307                 struct buffer_head *bh;
4308                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4309                 p = last;
4310                 while (--p >= first) {
4311                         nr = le32_to_cpu(*p);
4312                         if (!nr)
4313                                 continue;               /* A hole */
4314
4315                         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
4316                                                    nr, 1)) {
4317                                 EXT4_ERROR_INODE(inode,
4318                                                  "invalid indirect mapped "
4319                                                  "block %lu (level %d)",
4320                                                  (unsigned long) nr, depth);
4321                                 break;
4322                         }
4323
4324                         /* Go read the buffer for the next level down */
4325                         bh = sb_bread(inode->i_sb, nr);
4326
4327                         /*
4328                          * A read failure? Report error and clear slot
4329                          * (should be rare).
4330                          */
4331                         if (!bh) {
4332                                 EXT4_ERROR_INODE_BLOCK(inode, nr,
4333                                                        "Read failure");
4334                                 continue;
4335                         }
4336
4337                         /* This zaps the entire block.  Bottom up. */
4338                         BUFFER_TRACE(bh, "free child branches");
4339                         ext4_free_branches(handle, inode, bh,
4340                                         (__le32 *) bh->b_data,
4341                                         (__le32 *) bh->b_data + addr_per_block,
4342                                         depth);
4343                         brelse(bh);
4344
4345                         /*
4346                          * Everything below this this pointer has been
4347                          * released.  Now let this top-of-subtree go.
4348                          *
4349                          * We want the freeing of this indirect block to be
4350                          * atomic in the journal with the updating of the
4351                          * bitmap block which owns it.  So make some room in
4352                          * the journal.
4353                          *
4354                          * We zero the parent pointer *after* freeing its
4355                          * pointee in the bitmaps, so if extend_transaction()
4356                          * for some reason fails to put the bitmap changes and
4357                          * the release into the same transaction, recovery
4358                          * will merely complain about releasing a free block,
4359                          * rather than leaking blocks.
4360                          */
4361                         if (ext4_handle_is_aborted(handle))
4362                                 return;
4363                         if (try_to_extend_transaction(handle, inode)) {
4364                                 ext4_mark_inode_dirty(handle, inode);
4365                                 ext4_truncate_restart_trans(handle, inode,
4366                                             blocks_for_truncate(inode));
4367                         }
4368
4369                         /*
4370                          * The forget flag here is critical because if
4371                          * we are journaling (and not doing data
4372                          * journaling), we have to make sure a revoke
4373                          * record is written to prevent the journal
4374                          * replay from overwriting the (former)
4375                          * indirect block if it gets reallocated as a
4376                          * data block.  This must happen in the same
4377                          * transaction where the data blocks are
4378                          * actually freed.
4379                          */
4380                         ext4_free_blocks(handle, inode, NULL, nr, 1,
4381                                          EXT4_FREE_BLOCKS_METADATA|
4382                                          EXT4_FREE_BLOCKS_FORGET);
4383
4384                         if (parent_bh) {
4385                                 /*
4386                                  * The block which we have just freed is
4387                                  * pointed to by an indirect block: journal it
4388                                  */
4389                                 BUFFER_TRACE(parent_bh, "get_write_access");
4390                                 if (!ext4_journal_get_write_access(handle,
4391                                                                    parent_bh)){
4392                                         *p = 0;
4393                                         BUFFER_TRACE(parent_bh,
4394                                         "call ext4_handle_dirty_metadata");
4395                                         ext4_handle_dirty_metadata(handle,
4396                                                                    inode,
4397                                                                    parent_bh);
4398                                 }
4399                         }
4400                 }
4401         } else {
4402                 /* We have reached the bottom of the tree. */
4403                 BUFFER_TRACE(parent_bh, "free data blocks");
4404                 ext4_free_data(handle, inode, parent_bh, first, last);
4405         }
4406 }
4407
4408 int ext4_can_truncate(struct inode *inode)
4409 {
4410         if (S_ISREG(inode->i_mode))
4411                 return 1;
4412         if (S_ISDIR(inode->i_mode))
4413                 return 1;
4414         if (S_ISLNK(inode->i_mode))
4415                 return !ext4_inode_is_fast_symlink(inode);
4416         return 0;
4417 }
4418
4419 /*
4420  * ext4_punch_hole: punches a hole in a file by releaseing the blocks
4421  * associated with the given offset and length
4422  *
4423  * @inode:  File inode
4424  * @offset: The offset where the hole will begin
4425  * @len:    The length of the hole
4426  *
4427  * Returns: 0 on sucess or negative on failure
4428  */
4429
4430 int ext4_punch_hole(struct file *file, loff_t offset, loff_t length)
4431 {
4432         struct inode *inode = file->f_path.dentry->d_inode;
4433         if (!S_ISREG(inode->i_mode))
4434                 return -ENOTSUPP;
4435
4436         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
4437                 /* TODO: Add support for non extent hole punching */
4438                 return -ENOTSUPP;
4439         }
4440
4441         return ext4_ext_punch_hole(file, offset, length);
4442 }
4443
4444 /*
4445  * ext4_truncate()
4446  *
4447  * We block out ext4_get_block() block instantiations across the entire
4448  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4449  * simultaneously on behalf of the same inode.
4450  *
4451  * As we work through the truncate and commmit bits of it to the journal there
4452  * is one core, guiding principle: the file's tree must always be consistent on
4453  * disk.  We must be able to restart the truncate after a crash.
4454  *
4455  * The file's tree may be transiently inconsistent in memory (although it
4456  * probably isn't), but whenever we close off and commit a journal transaction,
4457  * the contents of (the filesystem + the journal) must be consistent and
4458  * restartable.  It's pretty simple, really: bottom up, right to left (although
4459  * left-to-right works OK too).
4460  *
4461  * Note that at recovery time, journal replay occurs *before* the restart of
4462  * truncate against the orphan inode list.
4463  *
4464  * The committed inode has the new, desired i_size (which is the same as
4465  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4466  * that this inode's truncate did not complete and it will again call
4467  * ext4_truncate() to have another go.  So there will be instantiated blocks
4468  * to the right of the truncation point in a crashed ext4 filesystem.  But
4469  * that's fine - as long as they are linked from the inode, the post-crash
4470  * ext4_truncate() run will find them and release them.
4471  */
4472 void ext4_truncate(struct inode *inode)
4473 {
4474         handle_t *handle;
4475         struct ext4_inode_info *ei = EXT4_I(inode);
4476         __le32 *i_data = ei->i_data;
4477         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4478         struct address_space *mapping = inode->i_mapping;
4479         ext4_lblk_t offsets[4];
4480         Indirect chain[4];
4481         Indirect *partial;
4482         __le32 nr = 0;
4483         int n = 0;
4484         ext4_lblk_t last_block, max_block;
4485         unsigned blocksize = inode->i_sb->s_blocksize;
4486
4487         trace_ext4_truncate_enter(inode);
4488
4489         if (!ext4_can_truncate(inode))
4490                 return;
4491
4492         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4493
4494         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4495                 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
4496
4497         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
4498                 ext4_ext_truncate(inode);
4499                 trace_ext4_truncate_exit(inode);
4500                 return;
4501         }
4502
4503         handle = start_transaction(inode);
4504         if (IS_ERR(handle))
4505                 return;         /* AKPM: return what? */
4506
4507         last_block = (inode->i_size + blocksize-1)
4508                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4509         max_block = (EXT4_SB(inode->i_sb)->s_bitmap_maxbytes + blocksize-1)
4510                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4511
4512         if (inode->i_size & (blocksize - 1))
4513                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4514                         goto out_stop;
4515
4516         if (last_block != max_block) {
4517                 n = ext4_block_to_path(inode, last_block, offsets, NULL);
4518                 if (n == 0)
4519                         goto out_stop;  /* error */
4520         }
4521
4522         /*
4523          * OK.  This truncate is going to happen.  We add the inode to the
4524          * orphan list, so that if this truncate spans multiple transactions,
4525          * and we crash, we will resume the truncate when the filesystem
4526          * recovers.  It also marks the inode dirty, to catch the new size.
4527          *
4528          * Implication: the file must always be in a sane, consistent
4529          * truncatable state while each transaction commits.
4530          */
4531         if (ext4_orphan_add(handle, inode))
4532                 goto out_stop;
4533
4534         /*
4535          * From here we block out all ext4_get_block() callers who want to
4536          * modify the block allocation tree.
4537          */
4538         down_write(&ei->i_data_sem);
4539
4540         ext4_discard_preallocations(inode);
4541
4542         /*
4543          * The orphan list entry will now protect us from any crash which
4544          * occurs before the truncate completes, so it is now safe to propagate
4545          * the new, shorter inode size (held for now in i_size) into the
4546          * on-disk inode. We do this via i_disksize, which is the value which
4547          * ext4 *really* writes onto the disk inode.
4548          */
4549         ei->i_disksize = inode->i_size;
4550
4551         if (last_block == max_block) {
4552                 /*
4553                  * It is unnecessary to free any data blocks if last_block is
4554                  * equal to the indirect block limit.
4555                  */
4556                 goto out_unlock;
4557         } else if (n == 1) {            /* direct blocks */
4558                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4559                                i_data + EXT4_NDIR_BLOCKS);
4560                 goto do_indirects;
4561         }
4562
4563         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4564         /* Kill the top of shared branch (not detached) */
4565         if (nr) {
4566                 if (partial == chain) {
4567                         /* Shared branch grows from the inode */
4568                         ext4_free_branches(handle, inode, NULL,
4569                                            &nr, &nr+1, (chain+n-1) - partial);
4570                         *partial->p = 0;
4571                         /*
4572                          * We mark the inode dirty prior to restart,
4573                          * and prior to stop.  No need for it here.
4574                          */
4575                 } else {
4576                         /* Shared branch grows from an indirect block */
4577                         BUFFER_TRACE(partial->bh, "get_write_access");
4578                         ext4_free_branches(handle, inode, partial->bh,
4579                                         partial->p,
4580                                         partial->p+1, (chain+n-1) - partial);
4581                 }
4582         }
4583         /* Clear the ends of indirect blocks on the shared branch */
4584         while (partial > chain) {
4585                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4586                                    (__le32*)partial->bh->b_data+addr_per_block,
4587                                    (chain+n-1) - partial);
4588                 BUFFER_TRACE(partial->bh, "call brelse");
4589                 brelse(partial->bh);
4590                 partial--;
4591         }
4592 do_indirects:
4593         /* Kill the remaining (whole) subtrees */
4594         switch (offsets[0]) {
4595         default:
4596                 nr = i_data[EXT4_IND_BLOCK];
4597                 if (nr) {
4598                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4599                         i_data[EXT4_IND_BLOCK] = 0;
4600                 }
4601         case EXT4_IND_BLOCK:
4602                 nr = i_data[EXT4_DIND_BLOCK];
4603                 if (nr) {
4604                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4605                         i_data[EXT4_DIND_BLOCK] = 0;
4606                 }
4607         case EXT4_DIND_BLOCK:
4608                 nr = i_data[EXT4_TIND_BLOCK];
4609                 if (nr) {
4610                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4611                         i_data[EXT4_TIND_BLOCK] = 0;
4612                 }
4613         case EXT4_TIND_BLOCK:
4614                 ;
4615         }
4616
4617 out_unlock:
4618         up_write(&ei->i_data_sem);
4619         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4620         ext4_mark_inode_dirty(handle, inode);
4621
4622         /*
4623          * In a multi-transaction truncate, we only make the final transaction
4624          * synchronous
4625          */
4626         if (IS_SYNC(inode))
4627                 ext4_handle_sync(handle);
4628 out_stop:
4629         /*
4630          * If this was a simple ftruncate(), and the file will remain alive
4631          * then we need to clear up the orphan record which we created above.
4632          * However, if this was a real unlink then we were called by
4633          * ext4_delete_inode(), and we allow that function to clean up the
4634          * orphan info for us.
4635          */
4636         if (inode->i_nlink)
4637                 ext4_orphan_del(handle, inode);
4638
4639         ext4_journal_stop(handle);
4640         trace_ext4_truncate_exit(inode);
4641 }
4642
4643 /*
4644  * ext4_get_inode_loc returns with an extra refcount against the inode's
4645  * underlying buffer_head on success. If 'in_mem' is true, we have all
4646  * data in memory that is needed to recreate the on-disk version of this
4647  * inode.
4648  */
4649 static int __ext4_get_inode_loc(struct inode *inode,
4650                                 struct ext4_iloc *iloc, int in_mem)
4651 {
4652         struct ext4_group_desc  *gdp;
4653         struct buffer_head      *bh;
4654         struct super_block      *sb = inode->i_sb;
4655         ext4_fsblk_t            block;
4656         int                     inodes_per_block, inode_offset;
4657
4658         iloc->bh = NULL;
4659         if (!ext4_valid_inum(sb, inode->i_ino))
4660                 return -EIO;
4661
4662         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4663         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4664         if (!gdp)
4665                 return -EIO;
4666
4667         /*
4668          * Figure out the offset within the block group inode table
4669          */
4670         inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
4671         inode_offset = ((inode->i_ino - 1) %
4672                         EXT4_INODES_PER_GROUP(sb));
4673         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4674         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4675
4676         bh = sb_getblk(sb, block);
4677         if (!bh) {
4678                 EXT4_ERROR_INODE_BLOCK(inode, block,
4679                                        "unable to read itable block");
4680                 return -EIO;
4681         }
4682         if (!buffer_uptodate(bh)) {
4683                 lock_buffer(bh);
4684
4685                 /*
4686                  * If the buffer has the write error flag, we have failed
4687                  * to write out another inode in the same block.  In this
4688                  * case, we don't have to read the block because we may
4689                  * read the old inode data successfully.
4690                  */
4691                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4692                         set_buffer_uptodate(bh);
4693
4694                 if (buffer_uptodate(bh)) {
4695                         /* someone brought it uptodate while we waited */
4696                         unlock_buffer(bh);
4697                         goto has_buffer;
4698                 }
4699
4700                 /*
4701                  * If we have all information of the inode in memory and this
4702                  * is the only valid inode in the block, we need not read the
4703                  * block.
4704                  */
4705                 if (in_mem) {
4706                         struct buffer_head *bitmap_bh;
4707                         int i, start;
4708
4709                         start = inode_offset & ~(inodes_per_block - 1);
4710
4711                         /* Is the inode bitmap in cache? */
4712                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4713                         if (!bitmap_bh)
4714                                 goto make_io;
4715
4716                         /*
4717                          * If the inode bitmap isn't in cache then the
4718                          * optimisation may end up performing two reads instead
4719                          * of one, so skip it.
4720                          */
4721                         if (!buffer_uptodate(bitmap_bh)) {
4722                                 brelse(bitmap_bh);
4723                                 goto make_io;
4724                         }
4725                         for (i = start; i < start + inodes_per_block; i++) {
4726                                 if (i == inode_offset)
4727                                         continue;
4728                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4729                                         break;
4730                         }
4731                         brelse(bitmap_bh);
4732                         if (i == start + inodes_per_block) {
4733                                 /* all other inodes are free, so skip I/O */
4734                                 memset(bh->b_data, 0, bh->b_size);
4735                                 set_buffer_uptodate(bh);
4736                                 unlock_buffer(bh);
4737                                 goto has_buffer;
4738                         }
4739                 }
4740
4741 make_io:
4742                 /*
4743                  * If we need to do any I/O, try to pre-readahead extra
4744                  * blocks from the inode table.
4745                  */
4746                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4747                         ext4_fsblk_t b, end, table;
4748                         unsigned num;
4749
4750                         table = ext4_inode_table(sb, gdp);
4751                         /* s_inode_readahead_blks is always a power of 2 */
4752                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4753                         if (table > b)
4754                                 b = table;
4755                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4756                         num = EXT4_INODES_PER_GROUP(sb);
4757                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4758                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4759                                 num -= ext4_itable_unused_count(sb, gdp);
4760                         table += num / inodes_per_block;
4761                         if (end > table)
4762                                 end = table;
4763                         while (b <= end)
4764                                 sb_breadahead(sb, b++);
4765                 }
4766
4767                 /*
4768                  * There are other valid inodes in the buffer, this inode
4769                  * has in-inode xattrs, or we don't have this inode in memory.
4770                  * Read the block from disk.
4771                  */
4772                 trace_ext4_load_inode(inode);
4773                 get_bh(bh);
4774                 bh->b_end_io = end_buffer_read_sync;
4775                 submit_bh(READ_META, bh);
4776                 wait_on_buffer(bh);
4777                 if (!buffer_uptodate(bh)) {
4778                         EXT4_ERROR_INODE_BLOCK(inode, block,
4779                                                "unable to read itable block");
4780                         brelse(bh);
4781                         return -EIO;
4782                 }
4783         }
4784 has_buffer:
4785         iloc->bh = bh;
4786         return 0;
4787 }
4788
4789 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4790 {
4791         /* We have all inode data except xattrs in memory here. */
4792         return __ext4_get_inode_loc(inode, iloc,
4793                 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
4794 }
4795
4796 void ext4_set_inode_flags(struct inode *inode)
4797 {
4798         unsigned int flags = EXT4_I(inode)->i_flags;
4799
4800         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4801         if (flags & EXT4_SYNC_FL)
4802                 inode->i_flags |= S_SYNC;
4803         if (flags & EXT4_APPEND_FL)
4804                 inode->i_flags |= S_APPEND;
4805         if (flags & EXT4_IMMUTABLE_FL)
4806                 inode->i_flags |= S_IMMUTABLE;
4807         if (flags & EXT4_NOATIME_FL)
4808                 inode->i_flags |= S_NOATIME;
4809         if (flags & EXT4_DIRSYNC_FL)
4810                 inode->i_flags |= S_DIRSYNC;
4811 }
4812
4813 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4814 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4815 {
4816         unsigned int vfs_fl;
4817         unsigned long old_fl, new_fl;
4818
4819         do {
4820                 vfs_fl = ei->vfs_inode.i_flags;
4821                 old_fl = ei->i_flags;
4822                 new_fl = old_fl & ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4823                                 EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|
4824                                 EXT4_DIRSYNC_FL);
4825                 if (vfs_fl & S_SYNC)
4826                         new_fl |= EXT4_SYNC_FL;
4827                 if (vfs_fl & S_APPEND)
4828                         new_fl |= EXT4_APPEND_FL;
4829                 if (vfs_fl & S_IMMUTABLE)
4830                         new_fl |= EXT4_IMMUTABLE_FL;
4831                 if (vfs_fl & S_NOATIME)
4832                         new_fl |= EXT4_NOATIME_FL;
4833                 if (vfs_fl & S_DIRSYNC)
4834                         new_fl |= EXT4_DIRSYNC_FL;
4835         } while (cmpxchg(&ei->i_flags, old_fl, new_fl) != old_fl);
4836 }
4837
4838 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4839                                   struct ext4_inode_info *ei)
4840 {
4841         blkcnt_t i_blocks ;
4842         struct inode *inode = &(ei->vfs_inode);
4843         struct super_block *sb = inode->i_sb;
4844
4845         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4846                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4847                 /* we are using combined 48 bit field */
4848                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4849                                         le32_to_cpu(raw_inode->i_blocks_lo);
4850                 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
4851                         /* i_blocks represent file system block size */
4852                         return i_blocks  << (inode->i_blkbits - 9);
4853                 } else {
4854                         return i_blocks;
4855                 }
4856         } else {
4857                 return le32_to_cpu(raw_inode->i_blocks_lo);
4858         }
4859 }
4860
4861 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4862 {
4863         struct ext4_iloc iloc;
4864         struct ext4_inode *raw_inode;
4865         struct ext4_inode_info *ei;
4866         struct inode *inode;
4867         journal_t *journal = EXT4_SB(sb)->s_journal;
4868         long ret;
4869         int block;
4870
4871         inode = iget_locked(sb, ino);
4872         if (!inode)
4873                 return ERR_PTR(-ENOMEM);
4874         if (!(inode->i_state & I_NEW))
4875                 return inode;
4876
4877         ei = EXT4_I(inode);
4878         iloc.bh = NULL;
4879
4880         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4881         if (ret < 0)
4882                 goto bad_inode;
4883         raw_inode = ext4_raw_inode(&iloc);
4884         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4885         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4886         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4887         if (!(test_opt(inode->i_sb, NO_UID32))) {
4888                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4889                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4890         }
4891         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
4892
4893         ext4_clear_state_flags(ei);     /* Only relevant on 32-bit archs */
4894         ei->i_dir_start_lookup = 0;
4895         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4896         /* We now have enough fields to check if the inode was active or not.
4897          * This is needed because nfsd might try to access dead inodes
4898          * the test is that same one that e2fsck uses
4899          * NeilBrown 1999oct15
4900          */
4901         if (inode->i_nlink == 0) {
4902                 if (inode->i_mode == 0 ||
4903                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
4904                         /* this inode is deleted */
4905                         ret = -ESTALE;
4906                         goto bad_inode;
4907                 }
4908                 /* The only unlinked inodes we let through here have
4909                  * valid i_mode and are being read by the orphan
4910                  * recovery code: that's fine, we're about to complete
4911                  * the process of deleting those. */
4912         }
4913         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4914         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4915         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4916         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
4917                 ei->i_file_acl |=
4918                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4919         inode->i_size = ext4_isize(raw_inode);
4920         ei->i_disksize = inode->i_size;
4921 #ifdef CONFIG_QUOTA
4922         ei->i_reserved_quota = 0;
4923 #endif
4924         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4925         ei->i_block_group = iloc.block_group;
4926         ei->i_last_alloc_group = ~0;
4927         /*
4928          * NOTE! The in-memory inode i_data array is in little-endian order
4929          * even on big-endian machines: we do NOT byteswap the block numbers!
4930          */
4931         for (block = 0; block < EXT4_N_BLOCKS; block++)
4932                 ei->i_data[block] = raw_inode->i_block[block];
4933         INIT_LIST_HEAD(&ei->i_orphan);
4934
4935         /*
4936          * Set transaction id's of transactions that have to be committed
4937          * to finish f[data]sync. We set them to currently running transaction
4938          * as we cannot be sure that the inode or some of its metadata isn't
4939          * part of the transaction - the inode could have been reclaimed and
4940          * now it is reread from disk.
4941          */
4942         if (journal) {
4943                 transaction_t *transaction;
4944                 tid_t tid;
4945
4946                 read_lock(&journal->j_state_lock);
4947                 if (journal->j_running_transaction)
4948                         transaction = journal->j_running_transaction;
4949                 else
4950                         transaction = journal->j_committing_transaction;
4951                 if (transaction)
4952                         tid = transaction->t_tid;
4953                 else
4954                         tid = journal->j_commit_sequence;
4955                 read_unlock(&journal->j_state_lock);
4956                 ei->i_sync_tid = tid;
4957                 ei->i_datasync_tid = tid;
4958         }
4959
4960         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4961                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4962                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4963                     EXT4_INODE_SIZE(inode->i_sb)) {
4964                         ret = -EIO;
4965                         goto bad_inode;
4966                 }
4967                 if (ei->i_extra_isize == 0) {
4968                         /* The extra space is currently unused. Use it. */
4969                         ei->i_extra_isize = sizeof(struct ext4_inode) -
4970                                             EXT4_GOOD_OLD_INODE_SIZE;
4971                 } else {
4972                         __le32 *magic = (void *)raw_inode +
4973                                         EXT4_GOOD_OLD_INODE_SIZE +
4974                                         ei->i_extra_isize;
4975                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
4976                                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
4977                 }
4978         } else
4979                 ei->i_extra_isize = 0;
4980
4981         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4982         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4983         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4984         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4985
4986         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4987         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4988                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4989                         inode->i_version |=
4990                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4991         }
4992
4993         ret = 0;
4994         if (ei->i_file_acl &&
4995             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
4996                 EXT4_ERROR_INODE(inode, "bad extended attribute block %llu",
4997                                  ei->i_file_acl);
4998                 ret = -EIO;
4999                 goto bad_inode;
5000         } else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5001                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5002                     (S_ISLNK(inode->i_mode) &&
5003                      !ext4_inode_is_fast_symlink(inode)))
5004                         /* Validate extent which is part of inode */
5005                         ret = ext4_ext_check_inode(inode);
5006         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
5007                    (S_ISLNK(inode->i_mode) &&
5008                     !ext4_inode_is_fast_symlink(inode))) {
5009                 /* Validate block references which are part of inode */
5010                 ret = ext4_check_inode_blockref(inode);
5011         }
5012         if (ret)
5013                 goto bad_inode;
5014
5015         if (S_ISREG(inode->i_mode)) {
5016                 inode->i_op = &ext4_file_inode_operations;
5017                 inode->i_fop = &ext4_file_operations;
5018                 ext4_set_aops(inode);
5019         } else if (S_ISDIR(inode->i_mode)) {
5020                 inode->i_op = &ext4_dir_inode_operations;
5021                 inode->i_fop = &ext4_dir_operations;
5022         } else if (S_ISLNK(inode->i_mode)) {
5023                 if (ext4_inode_is_fast_symlink(inode)) {
5024                         inode->i_op = &ext4_fast_symlink_inode_operations;
5025                         nd_terminate_link(ei->i_data, inode->i_size,
5026                                 sizeof(ei->i_data) - 1);
5027                 } else {
5028                         inode->i_op = &ext4_symlink_inode_operations;
5029                         ext4_set_aops(inode);
5030                 }
5031         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
5032               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
5033                 inode->i_op = &ext4_special_inode_operations;
5034                 if (raw_inode->i_block[0])
5035                         init_special_inode(inode, inode->i_mode,
5036                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
5037                 else
5038                         init_special_inode(inode, inode->i_mode,
5039                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
5040         } else {
5041                 ret = -EIO;
5042                 EXT4_ERROR_INODE(inode, "bogus i_mode (%o)", inode->i_mode);
5043                 goto bad_inode;
5044         }
5045         brelse(iloc.bh);
5046         ext4_set_inode_flags(inode);
5047         unlock_new_inode(inode);
5048         return inode;
5049
5050 bad_inode:
5051         brelse(iloc.bh);
5052         iget_failed(inode);
5053         return ERR_PTR(ret);
5054 }
5055
5056 static int ext4_inode_blocks_set(handle_t *handle,
5057                                 struct ext4_inode *raw_inode,
5058                                 struct ext4_inode_info *ei)
5059 {
5060         struct inode *inode = &(ei->vfs_inode);
5061         u64 i_blocks = inode->i_blocks;
5062         struct super_block *sb = inode->i_sb;
5063
5064         if (i_blocks <= ~0U) {
5065                 /*
5066                  * i_blocks can be represnted in a 32 bit variable
5067                  * as multiple of 512 bytes
5068                  */
5069                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5070                 raw_inode->i_blocks_high = 0;
5071                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5072                 return 0;
5073         }
5074         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
5075                 return -EFBIG;
5076
5077         if (i_blocks <= 0xffffffffffffULL) {
5078                 /*
5079                  * i_blocks can be represented in a 48 bit variable
5080                  * as multiple of 512 bytes
5081                  */
5082                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5083                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5084                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5085         } else {
5086                 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
5087                 /* i_block is stored in file system block size */
5088                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
5089                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
5090                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
5091         }
5092         return 0;
5093 }
5094
5095 /*
5096  * Post the struct inode info into an on-disk inode location in the
5097  * buffer-cache.  This gobbles the caller's reference to the
5098  * buffer_head in the inode location struct.
5099  *
5100  * The caller must have write access to iloc->bh.
5101  */
5102 static int ext4_do_update_inode(handle_t *handle,
5103                                 struct inode *inode,
5104                                 struct ext4_iloc *iloc)
5105 {
5106         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5107         struct ext4_inode_info *ei = EXT4_I(inode);
5108         struct buffer_head *bh = iloc->bh;
5109         int err = 0, rc, block;
5110
5111         /* For fields not not tracking in the in-memory inode,
5112          * initialise them to zero for new inodes. */
5113         if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
5114                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5115
5116         ext4_get_inode_flags(ei);
5117         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5118         if (!(test_opt(inode->i_sb, NO_UID32))) {
5119                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
5120                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
5121 /*
5122  * Fix up interoperability with old kernels. Otherwise, old inodes get
5123  * re-used with the upper 16 bits of the uid/gid intact
5124  */
5125                 if (!ei->i_dtime) {
5126                         raw_inode->i_uid_high =
5127                                 cpu_to_le16(high_16_bits(inode->i_uid));
5128                         raw_inode->i_gid_high =
5129                                 cpu_to_le16(high_16_bits(inode->i_gid));
5130                 } else {
5131                         raw_inode->i_uid_high = 0;
5132                         raw_inode->i_gid_high = 0;
5133                 }
5134         } else {
5135                 raw_inode->i_uid_low =
5136                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
5137                 raw_inode->i_gid_low =
5138                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
5139                 raw_inode->i_uid_high = 0;
5140                 raw_inode->i_gid_high = 0;
5141         }
5142         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5143
5144         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5145         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5146         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5147         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5148
5149         if (ext4_inode_blocks_set(handle, raw_inode, ei))
5150                 goto out_brelse;
5151         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5152         raw_inode->i_flags = cpu_to_le32(ei->i_flags & 0xFFFFFFFF);
5153         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
5154             cpu_to_le32(EXT4_OS_HURD))
5155                 raw_inode->i_file_acl_high =
5156                         cpu_to_le16(ei->i_file_acl >> 32);
5157         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5158         ext4_isize_set(raw_inode, ei->i_disksize);
5159         if (ei->i_disksize > 0x7fffffffULL) {
5160                 struct super_block *sb = inode->i_sb;
5161                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
5162                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
5163                                 EXT4_SB(sb)->s_es->s_rev_level ==
5164                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
5165                         /* If this is the first large file
5166                          * created, add a flag to the superblock.
5167                          */
5168                         err = ext4_journal_get_write_access(handle,
5169                                         EXT4_SB(sb)->s_sbh);
5170                         if (err)
5171                                 goto out_brelse;
5172                         ext4_update_dynamic_rev(sb);
5173                         EXT4_SET_RO_COMPAT_FEATURE(sb,
5174                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
5175                         sb->s_dirt = 1;
5176                         ext4_handle_sync(handle);
5177                         err = ext4_handle_dirty_metadata(handle, NULL,
5178                                         EXT4_SB(sb)->s_sbh);
5179                 }
5180         }
5181         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5182         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5183                 if (old_valid_dev(inode->i_rdev)) {
5184                         raw_inode->i_block[0] =
5185                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
5186                         raw_inode->i_block[1] = 0;
5187                 } else {
5188                         raw_inode->i_block[0] = 0;
5189                         raw_inode->i_block[1] =
5190                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
5191                         raw_inode->i_block[2] = 0;
5192                 }
5193         } else
5194                 for (block = 0; block < EXT4_N_BLOCKS; block++)
5195                         raw_inode->i_block[block] = ei->i_data[block];
5196
5197         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
5198         if (ei->i_extra_isize) {
5199                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5200                         raw_inode->i_version_hi =
5201                         cpu_to_le32(inode->i_version >> 32);
5202                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
5203         }
5204
5205         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5206         rc = ext4_handle_dirty_metadata(handle, NULL, bh);
5207         if (!err)
5208                 err = rc;
5209         ext4_clear_inode_state(inode, EXT4_STATE_NEW);
5210
5211         ext4_update_inode_fsync_trans(handle, inode, 0);
5212 out_brelse:
5213         brelse(bh);
5214         ext4_std_error(inode->i_sb, err);
5215         return err;
5216 }
5217
5218 /*
5219  * ext4_write_inode()
5220  *
5221  * We are called from a few places:
5222  *
5223  * - Within generic_file_write() for O_SYNC files.
5224  *   Here, there will be no transaction running. We wait for any running
5225  *   trasnaction to commit.
5226  *
5227  * - Within sys_sync(), kupdate and such.
5228  *   We wait on commit, if tol to.
5229  *
5230  * - Within prune_icache() (PF_MEMALLOC == true)
5231  *   Here we simply return.  We can't afford to block kswapd on the
5232  *   journal commit.
5233  *
5234  * In all cases it is actually safe for us to return without doing anything,
5235  * because the inode has been copied into a raw inode buffer in
5236  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
5237  * knfsd.
5238  *
5239  * Note that we are absolutely dependent upon all inode dirtiers doing the
5240  * right thing: they *must* call mark_inode_dirty() after dirtying info in
5241  * which we are interested.
5242  *
5243  * It would be a bug for them to not do this.  The code:
5244  *
5245  *      mark_inode_dirty(inode)
5246  *      stuff();
5247  *      inode->i_size = expr;
5248  *
5249  * is in error because a kswapd-driven write_inode() could occur while
5250  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
5251  * will no longer be on the superblock's dirty inode list.
5252  */
5253 int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
5254 {
5255         int err;
5256
5257         if (current->flags & PF_MEMALLOC)
5258                 return 0;
5259
5260         if (EXT4_SB(inode->i_sb)->s_journal) {
5261                 if (ext4_journal_current_handle()) {
5262                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5263                         dump_stack();
5264                         return -EIO;
5265                 }
5266
5267                 if (wbc->sync_mode != WB_SYNC_ALL)
5268                         return 0;
5269
5270                 err = ext4_force_commit(inode->i_sb);
5271         } else {
5272                 struct ext4_iloc iloc;
5273
5274                 err = __ext4_get_inode_loc(inode, &iloc, 0);
5275                 if (err)
5276                         return err;
5277                 if (wbc->sync_mode == WB_SYNC_ALL)
5278                         sync_dirty_buffer(iloc.bh);
5279                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5280                         EXT4_ERROR_INODE_BLOCK(inode, iloc.bh->b_blocknr,
5281                                          "IO error syncing inode");
5282                         err = -EIO;
5283                 }
5284                 brelse(iloc.bh);
5285         }
5286         return err;
5287 }
5288
5289 /*
5290  * ext4_setattr()
5291  *
5292  * Called from notify_change.
5293  *
5294  * We want to trap VFS attempts to truncate the file as soon as
5295  * possible.  In particular, we want to make sure that when the VFS
5296  * shrinks i_size, we put the inode on the orphan list and modify
5297  * i_disksize immediately, so that during the subsequent flushing of
5298  * dirty pages and freeing of disk blocks, we can guarantee that any
5299  * commit will leave the blocks being flushed in an unused state on
5300  * disk.  (On recovery, the inode will get truncated and the blocks will
5301  * be freed, so we have a strong guarantee that no future commit will
5302  * leave these blocks visible to the user.)
5303  *
5304  * Another thing we have to assure is that if we are in ordered mode
5305  * and inode is still attached to the committing transaction, we must
5306  * we start writeout of all the dirty pages which are being truncated.
5307  * This way we are sure that all the data written in the previous
5308  * transaction are already on disk (truncate waits for pages under
5309  * writeback).
5310  *
5311  * Called with inode->i_mutex down.
5312  */
5313 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5314 {
5315         struct inode *inode = dentry->d_inode;
5316         int error, rc = 0;
5317         int orphan = 0;
5318         const unsigned int ia_valid = attr->ia_valid;
5319
5320         error = inode_change_ok(inode, attr);
5321         if (error)
5322                 return error;
5323
5324         if (is_quota_modification(inode, attr))
5325                 dquot_initialize(inode);
5326         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
5327                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
5328                 handle_t *handle;
5329
5330                 /* (user+group)*(old+new) structure, inode write (sb,
5331                  * inode block, ? - but truncate inode update has it) */
5332                 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
5333                                         EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
5334                 if (IS_ERR(handle)) {
5335                         error = PTR_ERR(handle);
5336                         goto err_out;
5337                 }
5338                 error = dquot_transfer(inode, attr);
5339                 if (error) {
5340                         ext4_journal_stop(handle);
5341                         return error;
5342                 }
5343                 /* Update corresponding info in inode so that everything is in
5344                  * one transaction */
5345                 if (attr->ia_valid & ATTR_UID)
5346                         inode->i_uid = attr->ia_uid;
5347                 if (attr->ia_valid & ATTR_GID)
5348                         inode->i_gid = attr->ia_gid;
5349                 error = ext4_mark_inode_dirty(handle, inode);
5350                 ext4_journal_stop(handle);
5351         }
5352
5353         if (attr->ia_valid & ATTR_SIZE) {
5354                 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
5355                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5356
5357                         if (attr->ia_size > sbi->s_bitmap_maxbytes)
5358                                 return -EFBIG;
5359                 }
5360         }
5361
5362         if (S_ISREG(inode->i_mode) &&
5363             attr->ia_valid & ATTR_SIZE &&
5364             (attr->ia_size < inode->i_size)) {
5365                 handle_t *handle;
5366
5367                 handle = ext4_journal_start(inode, 3);
5368                 if (IS_ERR(handle)) {
5369                         error = PTR_ERR(handle);
5370                         goto err_out;
5371                 }
5372                 if (ext4_handle_valid(handle)) {
5373                         error = ext4_orphan_add(handle, inode);
5374                         orphan = 1;
5375                 }
5376                 EXT4_I(inode)->i_disksize = attr->ia_size;
5377                 rc = ext4_mark_inode_dirty(handle, inode);
5378                 if (!error)
5379                         error = rc;
5380                 ext4_journal_stop(handle);
5381
5382                 if (ext4_should_order_data(inode)) {
5383                         error = ext4_begin_ordered_truncate(inode,
5384                                                             attr->ia_size);
5385                         if (error) {
5386                                 /* Do as much error cleanup as possible */
5387                                 handle = ext4_journal_start(inode, 3);
5388                                 if (IS_ERR(handle)) {
5389                                         ext4_orphan_del(NULL, inode);
5390                                         goto err_out;
5391                                 }
5392                                 ext4_orphan_del(handle, inode);
5393                                 orphan = 0;
5394                                 ext4_journal_stop(handle);
5395                                 goto err_out;
5396                         }
5397                 }
5398         }
5399
5400         if (attr->ia_valid & ATTR_SIZE) {
5401                 if (attr->ia_size != i_size_read(inode)) {
5402                         truncate_setsize(inode, attr->ia_size);
5403                         ext4_truncate(inode);
5404                 } else if (ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
5405                         ext4_truncate(inode);
5406         }
5407
5408         if (!rc) {
5409                 setattr_copy(inode, attr);
5410                 mark_inode_dirty(inode);
5411         }
5412
5413         /*
5414          * If the call to ext4_truncate failed to get a transaction handle at
5415          * all, we need to clean up the in-core orphan list manually.
5416          */
5417         if (orphan && inode->i_nlink)
5418                 ext4_orphan_del(NULL, inode);
5419
5420         if (!rc && (ia_valid & ATTR_MODE))
5421                 rc = ext4_acl_chmod(inode);
5422
5423 err_out:
5424         ext4_std_error(inode->i_sb, error);
5425         if (!error)
5426                 error = rc;
5427         return error;
5428 }
5429
5430 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
5431                  struct kstat *stat)
5432 {
5433         struct inode *inode;
5434         unsigned long delalloc_blocks;
5435
5436         inode = dentry->d_inode;
5437         generic_fillattr(inode, stat);
5438
5439         /*
5440          * We can't update i_blocks if the block allocation is delayed
5441          * otherwise in the case of system crash before the real block
5442          * allocation is done, we will have i_blocks inconsistent with
5443          * on-disk file blocks.
5444          * We always keep i_blocks updated together with real
5445          * allocation. But to not confuse with user, stat
5446          * will return the blocks that include the delayed allocation
5447          * blocks for this file.
5448          */
5449         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
5450
5451         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
5452         return 0;
5453 }
5454
5455 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
5456                                       int chunk)
5457 {
5458         int indirects;
5459
5460         /* if nrblocks are contiguous */
5461         if (chunk) {
5462                 /*
5463                  * With N contiguous data blocks, we need at most
5464                  * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) + 1 indirect blocks,
5465                  * 2 dindirect blocks, and 1 tindirect block
5466                  */
5467                 return DIV_ROUND_UP(nrblocks,
5468                                     EXT4_ADDR_PER_BLOCK(inode->i_sb)) + 4;
5469         }
5470         /*
5471          * if nrblocks are not contiguous, worse case, each block touch
5472          * a indirect block, and each indirect block touch a double indirect
5473          * block, plus a triple indirect block
5474          */
5475         indirects = nrblocks * 2 + 1;
5476         return indirects;
5477 }
5478
5479 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5480 {
5481         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5482                 return ext4_indirect_trans_blocks(inode, nrblocks, chunk);
5483         return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
5484 }
5485
5486 /*
5487  * Account for index blocks, block groups bitmaps and block group
5488  * descriptor blocks if modify datablocks and index blocks
5489  * worse case, the indexs blocks spread over different block groups
5490  *
5491  * If datablocks are discontiguous, they are possible to spread over
5492  * different block groups too. If they are contiuguous, with flexbg,
5493  * they could still across block group boundary.
5494  *
5495  * Also account for superblock, inode, quota and xattr blocks
5496  */
5497 static int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5498 {
5499         ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
5500         int gdpblocks;
5501         int idxblocks;
5502         int ret = 0;
5503
5504         /*
5505          * How many index blocks need to touch to modify nrblocks?
5506          * The "Chunk" flag indicating whether the nrblocks is
5507          * physically conti