ext4: fix block validity checks so they work correctly with meta_bg
[linux-2.6.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/jbd2.h>
29 #include <linux/highuid.h>
30 #include <linux/pagemap.h>
31 #include <linux/quotaops.h>
32 #include <linux/string.h>
33 #include <linux/buffer_head.h>
34 #include <linux/writeback.h>
35 #include <linux/pagevec.h>
36 #include <linux/mpage.h>
37 #include <linux/namei.h>
38 #include <linux/uio.h>
39 #include <linux/bio.h>
40 #include <linux/workqueue.h>
41
42 #include "ext4_jbd2.h"
43 #include "xattr.h"
44 #include "acl.h"
45 #include "ext4_extents.h"
46
47 #include <trace/events/ext4.h>
48
49 #define MPAGE_DA_EXTENT_TAIL 0x01
50
51 static inline int ext4_begin_ordered_truncate(struct inode *inode,
52                                               loff_t new_size)
53 {
54         return jbd2_journal_begin_ordered_truncate(
55                                         EXT4_SB(inode->i_sb)->s_journal,
56                                         &EXT4_I(inode)->jinode,
57                                         new_size);
58 }
59
60 static void ext4_invalidatepage(struct page *page, unsigned long offset);
61
62 /*
63  * Test whether an inode is a fast symlink.
64  */
65 static int ext4_inode_is_fast_symlink(struct inode *inode)
66 {
67         int ea_blocks = EXT4_I(inode)->i_file_acl ?
68                 (inode->i_sb->s_blocksize >> 9) : 0;
69
70         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
71 }
72
73 /*
74  * The ext4 forget function must perform a revoke if we are freeing data
75  * which has been journaled.  Metadata (eg. indirect blocks) must be
76  * revoked in all cases.
77  *
78  * "bh" may be NULL: a metadata block may have been freed from memory
79  * but there may still be a record of it in the journal, and that record
80  * still needs to be revoked.
81  *
82  * If the handle isn't valid we're not journaling, but we still need to
83  * call into ext4_journal_revoke() to put the buffer head.
84  */
85 int ext4_forget(handle_t *handle, int is_metadata, struct inode *inode,
86                 struct buffer_head *bh, ext4_fsblk_t blocknr)
87 {
88         int err;
89
90         might_sleep();
91
92         trace_ext4_forget(inode, is_metadata, blocknr);
93         BUFFER_TRACE(bh, "enter");
94
95         jbd_debug(4, "forgetting bh %p: is_metadata = %d, mode %o, "
96                   "data mode %x\n",
97                   bh, is_metadata, inode->i_mode,
98                   test_opt(inode->i_sb, DATA_FLAGS));
99
100         /* Never use the revoke function if we are doing full data
101          * journaling: there is no need to, and a V1 superblock won't
102          * support it.  Otherwise, only skip the revoke on un-journaled
103          * data blocks. */
104
105         if (test_opt(inode->i_sb, DATA_FLAGS) == EXT4_MOUNT_JOURNAL_DATA ||
106             (!is_metadata && !ext4_should_journal_data(inode))) {
107                 if (bh) {
108                         BUFFER_TRACE(bh, "call jbd2_journal_forget");
109                         return ext4_journal_forget(handle, bh);
110                 }
111                 return 0;
112         }
113
114         /*
115          * data!=journal && (is_metadata || should_journal_data(inode))
116          */
117         BUFFER_TRACE(bh, "call ext4_journal_revoke");
118         err = ext4_journal_revoke(handle, blocknr, bh);
119         if (err)
120                 ext4_abort(inode->i_sb, __func__,
121                            "error %d when attempting revoke", err);
122         BUFFER_TRACE(bh, "exit");
123         return err;
124 }
125
126 /*
127  * Work out how many blocks we need to proceed with the next chunk of a
128  * truncate transaction.
129  */
130 static unsigned long blocks_for_truncate(struct inode *inode)
131 {
132         ext4_lblk_t needed;
133
134         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
135
136         /* Give ourselves just enough room to cope with inodes in which
137          * i_blocks is corrupt: we've seen disk corruptions in the past
138          * which resulted in random data in an inode which looked enough
139          * like a regular file for ext4 to try to delete it.  Things
140          * will go a bit crazy if that happens, but at least we should
141          * try not to panic the whole kernel. */
142         if (needed < 2)
143                 needed = 2;
144
145         /* But we need to bound the transaction so we don't overflow the
146          * journal. */
147         if (needed > EXT4_MAX_TRANS_DATA)
148                 needed = EXT4_MAX_TRANS_DATA;
149
150         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
151 }
152
153 /*
154  * Truncate transactions can be complex and absolutely huge.  So we need to
155  * be able to restart the transaction at a conventient checkpoint to make
156  * sure we don't overflow the journal.
157  *
158  * start_transaction gets us a new handle for a truncate transaction,
159  * and extend_transaction tries to extend the existing one a bit.  If
160  * extend fails, we need to propagate the failure up and restart the
161  * transaction in the top-level truncate loop. --sct
162  */
163 static handle_t *start_transaction(struct inode *inode)
164 {
165         handle_t *result;
166
167         result = ext4_journal_start(inode, blocks_for_truncate(inode));
168         if (!IS_ERR(result))
169                 return result;
170
171         ext4_std_error(inode->i_sb, PTR_ERR(result));
172         return result;
173 }
174
175 /*
176  * Try to extend this transaction for the purposes of truncation.
177  *
178  * Returns 0 if we managed to create more room.  If we can't create more
179  * room, and the transaction must be restarted we return 1.
180  */
181 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
182 {
183         if (!ext4_handle_valid(handle))
184                 return 0;
185         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
186                 return 0;
187         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
188                 return 0;
189         return 1;
190 }
191
192 /*
193  * Restart the transaction associated with *handle.  This does a commit,
194  * so before we call here everything must be consistently dirtied against
195  * this transaction.
196  */
197 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
198                                  int nblocks)
199 {
200         int ret;
201
202         /*
203          * Drop i_data_sem to avoid deadlock with ext4_get_blocks At this
204          * moment, get_block can be called only for blocks inside i_size since
205          * page cache has been already dropped and writes are blocked by
206          * i_mutex. So we can safely drop the i_data_sem here.
207          */
208         BUG_ON(EXT4_JOURNAL(inode) == NULL);
209         jbd_debug(2, "restarting handle %p\n", handle);
210         up_write(&EXT4_I(inode)->i_data_sem);
211         ret = ext4_journal_restart(handle, blocks_for_truncate(inode));
212         down_write(&EXT4_I(inode)->i_data_sem);
213         ext4_discard_preallocations(inode);
214
215         return ret;
216 }
217
218 /*
219  * Called at the last iput() if i_nlink is zero.
220  */
221 void ext4_delete_inode(struct inode *inode)
222 {
223         handle_t *handle;
224         int err;
225
226         if (ext4_should_order_data(inode))
227                 ext4_begin_ordered_truncate(inode, 0);
228         truncate_inode_pages(&inode->i_data, 0);
229
230         if (is_bad_inode(inode))
231                 goto no_delete;
232
233         handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
234         if (IS_ERR(handle)) {
235                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
236                 /*
237                  * If we're going to skip the normal cleanup, we still need to
238                  * make sure that the in-core orphan linked list is properly
239                  * cleaned up.
240                  */
241                 ext4_orphan_del(NULL, inode);
242                 goto no_delete;
243         }
244
245         if (IS_SYNC(inode))
246                 ext4_handle_sync(handle);
247         inode->i_size = 0;
248         err = ext4_mark_inode_dirty(handle, inode);
249         if (err) {
250                 ext4_warning(inode->i_sb, __func__,
251                              "couldn't mark inode dirty (err %d)", err);
252                 goto stop_handle;
253         }
254         if (inode->i_blocks)
255                 ext4_truncate(inode);
256
257         /*
258          * ext4_ext_truncate() doesn't reserve any slop when it
259          * restarts journal transactions; therefore there may not be
260          * enough credits left in the handle to remove the inode from
261          * the orphan list and set the dtime field.
262          */
263         if (!ext4_handle_has_enough_credits(handle, 3)) {
264                 err = ext4_journal_extend(handle, 3);
265                 if (err > 0)
266                         err = ext4_journal_restart(handle, 3);
267                 if (err != 0) {
268                         ext4_warning(inode->i_sb, __func__,
269                                      "couldn't extend journal (err %d)", err);
270                 stop_handle:
271                         ext4_journal_stop(handle);
272                         goto no_delete;
273                 }
274         }
275
276         /*
277          * Kill off the orphan record which ext4_truncate created.
278          * AKPM: I think this can be inside the above `if'.
279          * Note that ext4_orphan_del() has to be able to cope with the
280          * deletion of a non-existent orphan - this is because we don't
281          * know if ext4_truncate() actually created an orphan record.
282          * (Well, we could do this if we need to, but heck - it works)
283          */
284         ext4_orphan_del(handle, inode);
285         EXT4_I(inode)->i_dtime  = get_seconds();
286
287         /*
288          * One subtle ordering requirement: if anything has gone wrong
289          * (transaction abort, IO errors, whatever), then we can still
290          * do these next steps (the fs will already have been marked as
291          * having errors), but we can't free the inode if the mark_dirty
292          * fails.
293          */
294         if (ext4_mark_inode_dirty(handle, inode))
295                 /* If that failed, just do the required in-core inode clear. */
296                 clear_inode(inode);
297         else
298                 ext4_free_inode(handle, inode);
299         ext4_journal_stop(handle);
300         return;
301 no_delete:
302         clear_inode(inode);     /* We must guarantee clearing of inode... */
303 }
304
305 typedef struct {
306         __le32  *p;
307         __le32  key;
308         struct buffer_head *bh;
309 } Indirect;
310
311 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
312 {
313         p->key = *(p->p = v);
314         p->bh = bh;
315 }
316
317 /**
318  *      ext4_block_to_path - parse the block number into array of offsets
319  *      @inode: inode in question (we are only interested in its superblock)
320  *      @i_block: block number to be parsed
321  *      @offsets: array to store the offsets in
322  *      @boundary: set this non-zero if the referred-to block is likely to be
323  *             followed (on disk) by an indirect block.
324  *
325  *      To store the locations of file's data ext4 uses a data structure common
326  *      for UNIX filesystems - tree of pointers anchored in the inode, with
327  *      data blocks at leaves and indirect blocks in intermediate nodes.
328  *      This function translates the block number into path in that tree -
329  *      return value is the path length and @offsets[n] is the offset of
330  *      pointer to (n+1)th node in the nth one. If @block is out of range
331  *      (negative or too large) warning is printed and zero returned.
332  *
333  *      Note: function doesn't find node addresses, so no IO is needed. All
334  *      we need to know is the capacity of indirect blocks (taken from the
335  *      inode->i_sb).
336  */
337
338 /*
339  * Portability note: the last comparison (check that we fit into triple
340  * indirect block) is spelled differently, because otherwise on an
341  * architecture with 32-bit longs and 8Kb pages we might get into trouble
342  * if our filesystem had 8Kb blocks. We might use long long, but that would
343  * kill us on x86. Oh, well, at least the sign propagation does not matter -
344  * i_block would have to be negative in the very beginning, so we would not
345  * get there at all.
346  */
347
348 static int ext4_block_to_path(struct inode *inode,
349                               ext4_lblk_t i_block,
350                               ext4_lblk_t offsets[4], int *boundary)
351 {
352         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
353         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
354         const long direct_blocks = EXT4_NDIR_BLOCKS,
355                 indirect_blocks = ptrs,
356                 double_blocks = (1 << (ptrs_bits * 2));
357         int n = 0;
358         int final = 0;
359
360         if (i_block < direct_blocks) {
361                 offsets[n++] = i_block;
362                 final = direct_blocks;
363         } else if ((i_block -= direct_blocks) < indirect_blocks) {
364                 offsets[n++] = EXT4_IND_BLOCK;
365                 offsets[n++] = i_block;
366                 final = ptrs;
367         } else if ((i_block -= indirect_blocks) < double_blocks) {
368                 offsets[n++] = EXT4_DIND_BLOCK;
369                 offsets[n++] = i_block >> ptrs_bits;
370                 offsets[n++] = i_block & (ptrs - 1);
371                 final = ptrs;
372         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
373                 offsets[n++] = EXT4_TIND_BLOCK;
374                 offsets[n++] = i_block >> (ptrs_bits * 2);
375                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
376                 offsets[n++] = i_block & (ptrs - 1);
377                 final = ptrs;
378         } else {
379                 ext4_warning(inode->i_sb, "ext4_block_to_path",
380                              "block %lu > max in inode %lu",
381                              i_block + direct_blocks +
382                              indirect_blocks + double_blocks, inode->i_ino);
383         }
384         if (boundary)
385                 *boundary = final - 1 - (i_block & (ptrs - 1));
386         return n;
387 }
388
389 static int __ext4_check_blockref(const char *function, struct inode *inode,
390                                  __le32 *p, unsigned int max)
391 {
392         __le32 *bref = p;
393         unsigned int blk;
394
395         while (bref < p+max) {
396                 blk = le32_to_cpu(*bref++);
397                 if (blk &&
398                     unlikely(!ext4_data_block_valid(EXT4_SB(inode->i_sb),
399                                                     blk, 1))) {
400                         ext4_error(inode->i_sb, function,
401                                    "invalid block reference %u "
402                                    "in inode #%lu", blk, inode->i_ino);
403                         return -EIO;
404                 }
405         }
406         return 0;
407 }
408
409
410 #define ext4_check_indirect_blockref(inode, bh)                         \
411         __ext4_check_blockref(__func__, inode, (__le32 *)(bh)->b_data,  \
412                               EXT4_ADDR_PER_BLOCK((inode)->i_sb))
413
414 #define ext4_check_inode_blockref(inode)                                \
415         __ext4_check_blockref(__func__, inode, EXT4_I(inode)->i_data,   \
416                               EXT4_NDIR_BLOCKS)
417
418 /**
419  *      ext4_get_branch - read the chain of indirect blocks leading to data
420  *      @inode: inode in question
421  *      @depth: depth of the chain (1 - direct pointer, etc.)
422  *      @offsets: offsets of pointers in inode/indirect blocks
423  *      @chain: place to store the result
424  *      @err: here we store the error value
425  *
426  *      Function fills the array of triples <key, p, bh> and returns %NULL
427  *      if everything went OK or the pointer to the last filled triple
428  *      (incomplete one) otherwise. Upon the return chain[i].key contains
429  *      the number of (i+1)-th block in the chain (as it is stored in memory,
430  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
431  *      number (it points into struct inode for i==0 and into the bh->b_data
432  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
433  *      block for i>0 and NULL for i==0. In other words, it holds the block
434  *      numbers of the chain, addresses they were taken from (and where we can
435  *      verify that chain did not change) and buffer_heads hosting these
436  *      numbers.
437  *
438  *      Function stops when it stumbles upon zero pointer (absent block)
439  *              (pointer to last triple returned, *@err == 0)
440  *      or when it gets an IO error reading an indirect block
441  *              (ditto, *@err == -EIO)
442  *      or when it reads all @depth-1 indirect blocks successfully and finds
443  *      the whole chain, all way to the data (returns %NULL, *err == 0).
444  *
445  *      Need to be called with
446  *      down_read(&EXT4_I(inode)->i_data_sem)
447  */
448 static Indirect *ext4_get_branch(struct inode *inode, int depth,
449                                  ext4_lblk_t  *offsets,
450                                  Indirect chain[4], int *err)
451 {
452         struct super_block *sb = inode->i_sb;
453         Indirect *p = chain;
454         struct buffer_head *bh;
455
456         *err = 0;
457         /* i_data is not going away, no lock needed */
458         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
459         if (!p->key)
460                 goto no_block;
461         while (--depth) {
462                 bh = sb_getblk(sb, le32_to_cpu(p->key));
463                 if (unlikely(!bh))
464                         goto failure;
465
466                 if (!bh_uptodate_or_lock(bh)) {
467                         if (bh_submit_read(bh) < 0) {
468                                 put_bh(bh);
469                                 goto failure;
470                         }
471                         /* validate block references */
472                         if (ext4_check_indirect_blockref(inode, bh)) {
473                                 put_bh(bh);
474                                 goto failure;
475                         }
476                 }
477
478                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
479                 /* Reader: end */
480                 if (!p->key)
481                         goto no_block;
482         }
483         return NULL;
484
485 failure:
486         *err = -EIO;
487 no_block:
488         return p;
489 }
490
491 /**
492  *      ext4_find_near - find a place for allocation with sufficient locality
493  *      @inode: owner
494  *      @ind: descriptor of indirect block.
495  *
496  *      This function returns the preferred place for block allocation.
497  *      It is used when heuristic for sequential allocation fails.
498  *      Rules are:
499  *        + if there is a block to the left of our position - allocate near it.
500  *        + if pointer will live in indirect block - allocate near that block.
501  *        + if pointer will live in inode - allocate in the same
502  *          cylinder group.
503  *
504  * In the latter case we colour the starting block by the callers PID to
505  * prevent it from clashing with concurrent allocations for a different inode
506  * in the same block group.   The PID is used here so that functionally related
507  * files will be close-by on-disk.
508  *
509  *      Caller must make sure that @ind is valid and will stay that way.
510  */
511 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
512 {
513         struct ext4_inode_info *ei = EXT4_I(inode);
514         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
515         __le32 *p;
516         ext4_fsblk_t bg_start;
517         ext4_fsblk_t last_block;
518         ext4_grpblk_t colour;
519         ext4_group_t block_group;
520         int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
521
522         /* Try to find previous block */
523         for (p = ind->p - 1; p >= start; p--) {
524                 if (*p)
525                         return le32_to_cpu(*p);
526         }
527
528         /* No such thing, so let's try location of indirect block */
529         if (ind->bh)
530                 return ind->bh->b_blocknr;
531
532         /*
533          * It is going to be referred to from the inode itself? OK, just put it
534          * into the same cylinder group then.
535          */
536         block_group = ei->i_block_group;
537         if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
538                 block_group &= ~(flex_size-1);
539                 if (S_ISREG(inode->i_mode))
540                         block_group++;
541         }
542         bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
543         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
544
545         /*
546          * If we are doing delayed allocation, we don't need take
547          * colour into account.
548          */
549         if (test_opt(inode->i_sb, DELALLOC))
550                 return bg_start;
551
552         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
553                 colour = (current->pid % 16) *
554                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
555         else
556                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
557         return bg_start + colour;
558 }
559
560 /**
561  *      ext4_find_goal - find a preferred place for allocation.
562  *      @inode: owner
563  *      @block:  block we want
564  *      @partial: pointer to the last triple within a chain
565  *
566  *      Normally this function find the preferred place for block allocation,
567  *      returns it.
568  *      Because this is only used for non-extent files, we limit the block nr
569  *      to 32 bits.
570  */
571 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
572                                    Indirect *partial)
573 {
574         ext4_fsblk_t goal;
575
576         /*
577          * XXX need to get goal block from mballoc's data structures
578          */
579
580         goal = ext4_find_near(inode, partial);
581         goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
582         return goal;
583 }
584
585 /**
586  *      ext4_blks_to_allocate: Look up the block map and count the number
587  *      of direct blocks need to be allocated for the given branch.
588  *
589  *      @branch: chain of indirect blocks
590  *      @k: number of blocks need for indirect blocks
591  *      @blks: number of data blocks to be mapped.
592  *      @blocks_to_boundary:  the offset in the indirect block
593  *
594  *      return the total number of blocks to be allocate, including the
595  *      direct and indirect blocks.
596  */
597 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
598                                  int blocks_to_boundary)
599 {
600         unsigned int count = 0;
601
602         /*
603          * Simple case, [t,d]Indirect block(s) has not allocated yet
604          * then it's clear blocks on that path have not allocated
605          */
606         if (k > 0) {
607                 /* right now we don't handle cross boundary allocation */
608                 if (blks < blocks_to_boundary + 1)
609                         count += blks;
610                 else
611                         count += blocks_to_boundary + 1;
612                 return count;
613         }
614
615         count++;
616         while (count < blks && count <= blocks_to_boundary &&
617                 le32_to_cpu(*(branch[0].p + count)) == 0) {
618                 count++;
619         }
620         return count;
621 }
622
623 /**
624  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
625  *      @indirect_blks: the number of blocks need to allocate for indirect
626  *                      blocks
627  *
628  *      @new_blocks: on return it will store the new block numbers for
629  *      the indirect blocks(if needed) and the first direct block,
630  *      @blks:  on return it will store the total number of allocated
631  *              direct blocks
632  */
633 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
634                              ext4_lblk_t iblock, ext4_fsblk_t goal,
635                              int indirect_blks, int blks,
636                              ext4_fsblk_t new_blocks[4], int *err)
637 {
638         struct ext4_allocation_request ar;
639         int target, i;
640         unsigned long count = 0, blk_allocated = 0;
641         int index = 0;
642         ext4_fsblk_t current_block = 0;
643         int ret = 0;
644
645         /*
646          * Here we try to allocate the requested multiple blocks at once,
647          * on a best-effort basis.
648          * To build a branch, we should allocate blocks for
649          * the indirect blocks(if not allocated yet), and at least
650          * the first direct block of this branch.  That's the
651          * minimum number of blocks need to allocate(required)
652          */
653         /* first we try to allocate the indirect blocks */
654         target = indirect_blks;
655         while (target > 0) {
656                 count = target;
657                 /* allocating blocks for indirect blocks and direct blocks */
658                 current_block = ext4_new_meta_blocks(handle, inode,
659                                                         goal, &count, err);
660                 if (*err)
661                         goto failed_out;
662
663                 BUG_ON(current_block + count > EXT4_MAX_BLOCK_FILE_PHYS);
664
665                 target -= count;
666                 /* allocate blocks for indirect blocks */
667                 while (index < indirect_blks && count) {
668                         new_blocks[index++] = current_block++;
669                         count--;
670                 }
671                 if (count > 0) {
672                         /*
673                          * save the new block number
674                          * for the first direct block
675                          */
676                         new_blocks[index] = current_block;
677                         printk(KERN_INFO "%s returned more blocks than "
678                                                 "requested\n", __func__);
679                         WARN_ON(1);
680                         break;
681                 }
682         }
683
684         target = blks - count ;
685         blk_allocated = count;
686         if (!target)
687                 goto allocated;
688         /* Now allocate data blocks */
689         memset(&ar, 0, sizeof(ar));
690         ar.inode = inode;
691         ar.goal = goal;
692         ar.len = target;
693         ar.logical = iblock;
694         if (S_ISREG(inode->i_mode))
695                 /* enable in-core preallocation only for regular files */
696                 ar.flags = EXT4_MB_HINT_DATA;
697
698         current_block = ext4_mb_new_blocks(handle, &ar, err);
699         BUG_ON(current_block + ar.len > EXT4_MAX_BLOCK_FILE_PHYS);
700
701         if (*err && (target == blks)) {
702                 /*
703                  * if the allocation failed and we didn't allocate
704                  * any blocks before
705                  */
706                 goto failed_out;
707         }
708         if (!*err) {
709                 if (target == blks) {
710                         /*
711                          * save the new block number
712                          * for the first direct block
713                          */
714                         new_blocks[index] = current_block;
715                 }
716                 blk_allocated += ar.len;
717         }
718 allocated:
719         /* total number of blocks allocated for direct blocks */
720         ret = blk_allocated;
721         *err = 0;
722         return ret;
723 failed_out:
724         for (i = 0; i < index; i++)
725                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
726         return ret;
727 }
728
729 /**
730  *      ext4_alloc_branch - allocate and set up a chain of blocks.
731  *      @inode: owner
732  *      @indirect_blks: number of allocated indirect blocks
733  *      @blks: number of allocated direct blocks
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                 branch[n].bh = bh;
784                 lock_buffer(bh);
785                 BUFFER_TRACE(bh, "call get_create_access");
786                 err = ext4_journal_get_create_access(handle, bh);
787                 if (err) {
788                         /* Don't brelse(bh) here; it's done in
789                          * ext4_journal_forget() below */
790                         unlock_buffer(bh);
791                         goto failed;
792                 }
793
794                 memset(bh->b_data, 0, blocksize);
795                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
796                 branch[n].key = cpu_to_le32(new_blocks[n]);
797                 *branch[n].p = branch[n].key;
798                 if (n == indirect_blks) {
799                         current_block = new_blocks[n];
800                         /*
801                          * End of chain, update the last new metablock of
802                          * the chain to point to the new allocated
803                          * data blocks numbers
804                          */
805                         for (i = 1; i < num; i++)
806                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
807                 }
808                 BUFFER_TRACE(bh, "marking uptodate");
809                 set_buffer_uptodate(bh);
810                 unlock_buffer(bh);
811
812                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
813                 err = ext4_handle_dirty_metadata(handle, inode, bh);
814                 if (err)
815                         goto failed;
816         }
817         *blks = num;
818         return err;
819 failed:
820         /* Allocation failed, free what we already allocated */
821         for (i = 1; i <= n ; i++) {
822                 BUFFER_TRACE(branch[i].bh, "call jbd2_journal_forget");
823                 ext4_journal_forget(handle, branch[i].bh);
824         }
825         for (i = 0; i < indirect_blks; i++)
826                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
827
828         ext4_free_blocks(handle, inode, new_blocks[i], num, 0);
829
830         return err;
831 }
832
833 /**
834  * ext4_splice_branch - splice the allocated branch onto inode.
835  * @inode: owner
836  * @block: (logical) number of block we are adding
837  * @chain: chain of indirect blocks (with a missing link - see
838  *      ext4_alloc_branch)
839  * @where: location of missing link
840  * @num:   number of indirect blocks we are adding
841  * @blks:  number of direct blocks we are adding
842  *
843  * This function fills the missing link and does all housekeeping needed in
844  * inode (->i_blocks, etc.). In case of success we end up with the full
845  * chain to new block and return 0.
846  */
847 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
848                               ext4_lblk_t block, Indirect *where, int num,
849                               int blks)
850 {
851         int i;
852         int err = 0;
853         ext4_fsblk_t current_block;
854
855         /*
856          * If we're splicing into a [td]indirect block (as opposed to the
857          * inode) then we need to get write access to the [td]indirect block
858          * before the splice.
859          */
860         if (where->bh) {
861                 BUFFER_TRACE(where->bh, "get_write_access");
862                 err = ext4_journal_get_write_access(handle, where->bh);
863                 if (err)
864                         goto err_out;
865         }
866         /* That's it */
867
868         *where->p = where->key;
869
870         /*
871          * Update the host buffer_head or inode to point to more just allocated
872          * direct blocks blocks
873          */
874         if (num == 0 && blks > 1) {
875                 current_block = le32_to_cpu(where->key) + 1;
876                 for (i = 1; i < blks; i++)
877                         *(where->p + i) = cpu_to_le32(current_block++);
878         }
879
880         /* We are done with atomic stuff, now do the rest of housekeeping */
881         /* had we spliced it onto indirect block? */
882         if (where->bh) {
883                 /*
884                  * If we spliced it onto an indirect block, we haven't
885                  * altered the inode.  Note however that if it is being spliced
886                  * onto an indirect block at the very end of the file (the
887                  * file is growing) then we *will* alter the inode to reflect
888                  * the new i_size.  But that is not done here - it is done in
889                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
890                  */
891                 jbd_debug(5, "splicing indirect only\n");
892                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
893                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
894                 if (err)
895                         goto err_out;
896         } else {
897                 /*
898                  * OK, we spliced it into the inode itself on a direct block.
899                  */
900                 ext4_mark_inode_dirty(handle, inode);
901                 jbd_debug(5, "splicing direct\n");
902         }
903         return err;
904
905 err_out:
906         for (i = 1; i <= num; i++) {
907                 BUFFER_TRACE(where[i].bh, "call jbd2_journal_forget");
908                 ext4_journal_forget(handle, where[i].bh);
909                 ext4_free_blocks(handle, inode,
910                                         le32_to_cpu(where[i-1].key), 1, 0);
911         }
912         ext4_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks, 0);
913
914         return err;
915 }
916
917 /*
918  * The ext4_ind_get_blocks() function handles non-extents inodes
919  * (i.e., using the traditional indirect/double-indirect i_blocks
920  * scheme) for ext4_get_blocks().
921  *
922  * Allocation strategy is simple: if we have to allocate something, we will
923  * have to go the whole way to leaf. So let's do it before attaching anything
924  * to tree, set linkage between the newborn blocks, write them if sync is
925  * required, recheck the path, free and repeat if check fails, otherwise
926  * set the last missing link (that will protect us from any truncate-generated
927  * removals - all blocks on the path are immune now) and possibly force the
928  * write on the parent block.
929  * That has a nice additional property: no special recovery from the failed
930  * allocations is needed - we simply release blocks and do not touch anything
931  * reachable from inode.
932  *
933  * `handle' can be NULL if create == 0.
934  *
935  * return > 0, # of blocks mapped or allocated.
936  * return = 0, if plain lookup failed.
937  * return < 0, error case.
938  *
939  * The ext4_ind_get_blocks() function should be called with
940  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
941  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
942  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
943  * blocks.
944  */
945 static int ext4_ind_get_blocks(handle_t *handle, struct inode *inode,
946                                ext4_lblk_t iblock, unsigned int maxblocks,
947                                struct buffer_head *bh_result,
948                                int flags)
949 {
950         int err = -EIO;
951         ext4_lblk_t offsets[4];
952         Indirect chain[4];
953         Indirect *partial;
954         ext4_fsblk_t goal;
955         int indirect_blks;
956         int blocks_to_boundary = 0;
957         int depth;
958         int count = 0;
959         ext4_fsblk_t first_block = 0;
960
961         J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
962         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
963         depth = ext4_block_to_path(inode, iblock, offsets,
964                                    &blocks_to_boundary);
965
966         if (depth == 0)
967                 goto out;
968
969         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
970
971         /* Simplest case - block found, no allocation needed */
972         if (!partial) {
973                 first_block = le32_to_cpu(chain[depth - 1].key);
974                 clear_buffer_new(bh_result);
975                 count++;
976                 /*map more blocks*/
977                 while (count < maxblocks && count <= blocks_to_boundary) {
978                         ext4_fsblk_t blk;
979
980                         blk = le32_to_cpu(*(chain[depth-1].p + count));
981
982                         if (blk == first_block + count)
983                                 count++;
984                         else
985                                 break;
986                 }
987                 goto got_it;
988         }
989
990         /* Next simple case - plain lookup or failed read of indirect block */
991         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
992                 goto cleanup;
993
994         /*
995          * Okay, we need to do block allocation.
996         */
997         goal = ext4_find_goal(inode, iblock, partial);
998
999         /* the number of blocks need to allocate for [d,t]indirect blocks */
1000         indirect_blks = (chain + depth) - partial - 1;
1001
1002         /*
1003          * Next look up the indirect map to count the totoal number of
1004          * direct blocks to allocate for this branch.
1005          */
1006         count = ext4_blks_to_allocate(partial, indirect_blks,
1007                                         maxblocks, blocks_to_boundary);
1008         /*
1009          * Block out ext4_truncate while we alter the tree
1010          */
1011         err = ext4_alloc_branch(handle, inode, iblock, indirect_blks,
1012                                 &count, goal,
1013                                 offsets + (partial - chain), partial);
1014
1015         /*
1016          * The ext4_splice_branch call will free and forget any buffers
1017          * on the new chain if there is a failure, but that risks using
1018          * up transaction credits, especially for bitmaps where the
1019          * credits cannot be returned.  Can we handle this somehow?  We
1020          * may need to return -EAGAIN upwards in the worst case.  --sct
1021          */
1022         if (!err)
1023                 err = ext4_splice_branch(handle, inode, iblock,
1024                                          partial, indirect_blks, count);
1025         else
1026                 goto cleanup;
1027
1028         set_buffer_new(bh_result);
1029 got_it:
1030         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
1031         if (count > blocks_to_boundary)
1032                 set_buffer_boundary(bh_result);
1033         err = count;
1034         /* Clean up and exit */
1035         partial = chain + depth - 1;    /* the whole chain */
1036 cleanup:
1037         while (partial > chain) {
1038                 BUFFER_TRACE(partial->bh, "call brelse");
1039                 brelse(partial->bh);
1040                 partial--;
1041         }
1042         BUFFER_TRACE(bh_result, "returned");
1043 out:
1044         return err;
1045 }
1046
1047 qsize_t ext4_get_reserved_space(struct inode *inode)
1048 {
1049         unsigned long long total;
1050
1051         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1052         total = EXT4_I(inode)->i_reserved_data_blocks +
1053                 EXT4_I(inode)->i_reserved_meta_blocks;
1054         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1055
1056         return total;
1057 }
1058 /*
1059  * Calculate the number of metadata blocks need to reserve
1060  * to allocate @blocks for non extent file based file
1061  */
1062 static int ext4_indirect_calc_metadata_amount(struct inode *inode, int blocks)
1063 {
1064         int icap = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1065         int ind_blks, dind_blks, tind_blks;
1066
1067         /* number of new indirect blocks needed */
1068         ind_blks = (blocks + icap - 1) / icap;
1069
1070         dind_blks = (ind_blks + icap - 1) / icap;
1071
1072         tind_blks = 1;
1073
1074         return ind_blks + dind_blks + tind_blks;
1075 }
1076
1077 /*
1078  * Calculate the number of metadata blocks need to reserve
1079  * to allocate given number of blocks
1080  */
1081 static int ext4_calc_metadata_amount(struct inode *inode, int blocks)
1082 {
1083         if (!blocks)
1084                 return 0;
1085
1086         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
1087                 return ext4_ext_calc_metadata_amount(inode, blocks);
1088
1089         return ext4_indirect_calc_metadata_amount(inode, blocks);
1090 }
1091
1092 static void ext4_da_update_reserve_space(struct inode *inode, int used)
1093 {
1094         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1095         int total, mdb, mdb_free;
1096
1097         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1098         /* recalculate the number of metablocks still need to be reserved */
1099         total = EXT4_I(inode)->i_reserved_data_blocks - used;
1100         mdb = ext4_calc_metadata_amount(inode, total);
1101
1102         /* figure out how many metablocks to release */
1103         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1104         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1105
1106         if (mdb_free) {
1107                 /* Account for allocated meta_blocks */
1108                 mdb_free -= EXT4_I(inode)->i_allocated_meta_blocks;
1109
1110                 /* update fs dirty blocks counter */
1111                 percpu_counter_sub(&sbi->s_dirtyblocks_counter, mdb_free);
1112                 EXT4_I(inode)->i_allocated_meta_blocks = 0;
1113                 EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1114         }
1115
1116         /* update per-inode reservations */
1117         BUG_ON(used  > EXT4_I(inode)->i_reserved_data_blocks);
1118         EXT4_I(inode)->i_reserved_data_blocks -= used;
1119         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1120
1121         /*
1122          * free those over-booking quota for metadata blocks
1123          */
1124         if (mdb_free)
1125                 vfs_dq_release_reservation_block(inode, mdb_free);
1126
1127         /*
1128          * If we have done all the pending block allocations and if
1129          * there aren't any writers on the inode, we can discard the
1130          * inode's preallocations.
1131          */
1132         if (!total && (atomic_read(&inode->i_writecount) == 0))
1133                 ext4_discard_preallocations(inode);
1134 }
1135
1136 static int check_block_validity(struct inode *inode, const char *msg,
1137                                 sector_t logical, sector_t phys, int len)
1138 {
1139         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), phys, len)) {
1140                 ext4_error(inode->i_sb, msg,
1141                            "inode #%lu logical block %llu mapped to %llu "
1142                            "(size %d)", inode->i_ino,
1143                            (unsigned long long) logical,
1144                            (unsigned long long) phys, len);
1145                 return -EIO;
1146         }
1147         return 0;
1148 }
1149
1150 /*
1151  * Return the number of contiguous dirty pages in a given inode
1152  * starting at page frame idx.
1153  */
1154 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
1155                                     unsigned int max_pages)
1156 {
1157         struct address_space *mapping = inode->i_mapping;
1158         pgoff_t index;
1159         struct pagevec pvec;
1160         pgoff_t num = 0;
1161         int i, nr_pages, done = 0;
1162
1163         if (max_pages == 0)
1164                 return 0;
1165         pagevec_init(&pvec, 0);
1166         while (!done) {
1167                 index = idx;
1168                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
1169                                               PAGECACHE_TAG_DIRTY,
1170                                               (pgoff_t)PAGEVEC_SIZE);
1171                 if (nr_pages == 0)
1172                         break;
1173                 for (i = 0; i < nr_pages; i++) {
1174                         struct page *page = pvec.pages[i];
1175                         struct buffer_head *bh, *head;
1176
1177                         lock_page(page);
1178                         if (unlikely(page->mapping != mapping) ||
1179                             !PageDirty(page) ||
1180                             PageWriteback(page) ||
1181                             page->index != idx) {
1182                                 done = 1;
1183                                 unlock_page(page);
1184                                 break;
1185                         }
1186                         if (page_has_buffers(page)) {
1187                                 bh = head = page_buffers(page);
1188                                 do {
1189                                         if (!buffer_delay(bh) &&
1190                                             !buffer_unwritten(bh))
1191                                                 done = 1;
1192                                         bh = bh->b_this_page;
1193                                 } while (!done && (bh != head));
1194                         }
1195                         unlock_page(page);
1196                         if (done)
1197                                 break;
1198                         idx++;
1199                         num++;
1200                         if (num >= max_pages)
1201                                 break;
1202                 }
1203                 pagevec_release(&pvec);
1204         }
1205         return num;
1206 }
1207
1208 /*
1209  * The ext4_get_blocks() function tries to look up the requested blocks,
1210  * and returns if the blocks are already mapped.
1211  *
1212  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1213  * and store the allocated blocks in the result buffer head and mark it
1214  * mapped.
1215  *
1216  * If file type is extents based, it will call ext4_ext_get_blocks(),
1217  * Otherwise, call with ext4_ind_get_blocks() to handle indirect mapping
1218  * based files
1219  *
1220  * On success, it returns the number of blocks being mapped or allocate.
1221  * if create==0 and the blocks are pre-allocated and uninitialized block,
1222  * the result buffer head is unmapped. If the create ==1, it will make sure
1223  * the buffer head is mapped.
1224  *
1225  * It returns 0 if plain look up failed (blocks have not been allocated), in
1226  * that casem, buffer head is unmapped
1227  *
1228  * It returns the error in case of allocation failure.
1229  */
1230 int ext4_get_blocks(handle_t *handle, struct inode *inode, sector_t block,
1231                     unsigned int max_blocks, struct buffer_head *bh,
1232                     int flags)
1233 {
1234         int retval;
1235
1236         clear_buffer_mapped(bh);
1237         clear_buffer_unwritten(bh);
1238
1239         ext_debug("ext4_get_blocks(): inode %lu, flag %d, max_blocks %u,"
1240                   "logical block %lu\n", inode->i_ino, flags, max_blocks,
1241                   (unsigned long)block);
1242         /*
1243          * Try to see if we can get the block without requesting a new
1244          * file system block.
1245          */
1246         down_read((&EXT4_I(inode)->i_data_sem));
1247         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1248                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1249                                 bh, 0);
1250         } else {
1251                 retval = ext4_ind_get_blocks(handle, inode, block, max_blocks,
1252                                              bh, 0);
1253         }
1254         up_read((&EXT4_I(inode)->i_data_sem));
1255
1256         if (retval > 0 && buffer_mapped(bh)) {
1257                 int ret = check_block_validity(inode, "file system corruption",
1258                                                block, bh->b_blocknr, retval);
1259                 if (ret != 0)
1260                         return ret;
1261         }
1262
1263         /* If it is only a block(s) look up */
1264         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1265                 return retval;
1266
1267         /*
1268          * Returns if the blocks have already allocated
1269          *
1270          * Note that if blocks have been preallocated
1271          * ext4_ext_get_block() returns th create = 0
1272          * with buffer head unmapped.
1273          */
1274         if (retval > 0 && buffer_mapped(bh))
1275                 return retval;
1276
1277         /*
1278          * When we call get_blocks without the create flag, the
1279          * BH_Unwritten flag could have gotten set if the blocks
1280          * requested were part of a uninitialized extent.  We need to
1281          * clear this flag now that we are committed to convert all or
1282          * part of the uninitialized extent to be an initialized
1283          * extent.  This is because we need to avoid the combination
1284          * of BH_Unwritten and BH_Mapped flags being simultaneously
1285          * set on the buffer_head.
1286          */
1287         clear_buffer_unwritten(bh);
1288
1289         /*
1290          * New blocks allocate and/or writing to uninitialized extent
1291          * will possibly result in updating i_data, so we take
1292          * the write lock of i_data_sem, and call get_blocks()
1293          * with create == 1 flag.
1294          */
1295         down_write((&EXT4_I(inode)->i_data_sem));
1296
1297         /*
1298          * if the caller is from delayed allocation writeout path
1299          * we have already reserved fs blocks for allocation
1300          * let the underlying get_block() function know to
1301          * avoid double accounting
1302          */
1303         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1304                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1305         /*
1306          * We need to check for EXT4 here because migrate
1307          * could have changed the inode type in between
1308          */
1309         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1310                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1311                                               bh, flags);
1312         } else {
1313                 retval = ext4_ind_get_blocks(handle, inode, block,
1314                                              max_blocks, bh, flags);
1315
1316                 if (retval > 0 && buffer_new(bh)) {
1317                         /*
1318                          * We allocated new blocks which will result in
1319                          * i_data's format changing.  Force the migrate
1320                          * to fail by clearing migrate flags
1321                          */
1322                         EXT4_I(inode)->i_state &= ~EXT4_STATE_EXT_MIGRATE;
1323                 }
1324         }
1325
1326         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1327                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1328
1329         /*
1330          * Update reserved blocks/metadata blocks after successful
1331          * block allocation which had been deferred till now.
1332          */
1333         if ((retval > 0) && (flags & EXT4_GET_BLOCKS_UPDATE_RESERVE_SPACE))
1334                 ext4_da_update_reserve_space(inode, retval);
1335
1336         up_write((&EXT4_I(inode)->i_data_sem));
1337         if (retval > 0 && buffer_mapped(bh)) {
1338                 int ret = check_block_validity(inode, "file system "
1339                                                "corruption after allocation",
1340                                                block, bh->b_blocknr, retval);
1341                 if (ret != 0)
1342                         return ret;
1343         }
1344         return retval;
1345 }
1346
1347 /* Maximum number of blocks we map for direct IO at once. */
1348 #define DIO_MAX_BLOCKS 4096
1349
1350 int ext4_get_block(struct inode *inode, sector_t iblock,
1351                    struct buffer_head *bh_result, int create)
1352 {
1353         handle_t *handle = ext4_journal_current_handle();
1354         int ret = 0, started = 0;
1355         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1356         int dio_credits;
1357
1358         if (create && !handle) {
1359                 /* Direct IO write... */
1360                 if (max_blocks > DIO_MAX_BLOCKS)
1361                         max_blocks = DIO_MAX_BLOCKS;
1362                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
1363                 handle = ext4_journal_start(inode, dio_credits);
1364                 if (IS_ERR(handle)) {
1365                         ret = PTR_ERR(handle);
1366                         goto out;
1367                 }
1368                 started = 1;
1369         }
1370
1371         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
1372                               create ? EXT4_GET_BLOCKS_CREATE : 0);
1373         if (ret > 0) {
1374                 bh_result->b_size = (ret << inode->i_blkbits);
1375                 ret = 0;
1376         }
1377         if (started)
1378                 ext4_journal_stop(handle);
1379 out:
1380         return ret;
1381 }
1382
1383 /*
1384  * `handle' can be NULL if create is zero
1385  */
1386 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1387                                 ext4_lblk_t block, int create, int *errp)
1388 {
1389         struct buffer_head dummy;
1390         int fatal = 0, err;
1391         int flags = 0;
1392
1393         J_ASSERT(handle != NULL || create == 0);
1394
1395         dummy.b_state = 0;
1396         dummy.b_blocknr = -1000;
1397         buffer_trace_init(&dummy.b_history);
1398         if (create)
1399                 flags |= EXT4_GET_BLOCKS_CREATE;
1400         err = ext4_get_blocks(handle, inode, block, 1, &dummy, flags);
1401         /*
1402          * ext4_get_blocks() returns number of blocks mapped. 0 in
1403          * case of a HOLE.
1404          */
1405         if (err > 0) {
1406                 if (err > 1)
1407                         WARN_ON(1);
1408                 err = 0;
1409         }
1410         *errp = err;
1411         if (!err && buffer_mapped(&dummy)) {
1412                 struct buffer_head *bh;
1413                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1414                 if (!bh) {
1415                         *errp = -EIO;
1416                         goto err;
1417                 }
1418                 if (buffer_new(&dummy)) {
1419                         J_ASSERT(create != 0);
1420                         J_ASSERT(handle != NULL);
1421
1422                         /*
1423                          * Now that we do not always journal data, we should
1424                          * keep in mind whether this should always journal the
1425                          * new buffer as metadata.  For now, regular file
1426                          * writes use ext4_get_block instead, so it's not a
1427                          * problem.
1428                          */
1429                         lock_buffer(bh);
1430                         BUFFER_TRACE(bh, "call get_create_access");
1431                         fatal = ext4_journal_get_create_access(handle, bh);
1432                         if (!fatal && !buffer_uptodate(bh)) {
1433                                 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1434                                 set_buffer_uptodate(bh);
1435                         }
1436                         unlock_buffer(bh);
1437                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1438                         err = ext4_handle_dirty_metadata(handle, inode, bh);
1439                         if (!fatal)
1440                                 fatal = err;
1441                 } else {
1442                         BUFFER_TRACE(bh, "not a new buffer");
1443                 }
1444                 if (fatal) {
1445                         *errp = fatal;
1446                         brelse(bh);
1447                         bh = NULL;
1448                 }
1449                 return bh;
1450         }
1451 err:
1452         return NULL;
1453 }
1454
1455 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1456                                ext4_lblk_t block, int create, int *err)
1457 {
1458         struct buffer_head *bh;
1459
1460         bh = ext4_getblk(handle, inode, block, create, err);
1461         if (!bh)
1462                 return bh;
1463         if (buffer_uptodate(bh))
1464                 return bh;
1465         ll_rw_block(READ_META, 1, &bh);
1466         wait_on_buffer(bh);
1467         if (buffer_uptodate(bh))
1468                 return bh;
1469         put_bh(bh);
1470         *err = -EIO;
1471         return NULL;
1472 }
1473
1474 static int walk_page_buffers(handle_t *handle,
1475                              struct buffer_head *head,
1476                              unsigned from,
1477                              unsigned to,
1478                              int *partial,
1479                              int (*fn)(handle_t *handle,
1480                                        struct buffer_head *bh))
1481 {
1482         struct buffer_head *bh;
1483         unsigned block_start, block_end;
1484         unsigned blocksize = head->b_size;
1485         int err, ret = 0;
1486         struct buffer_head *next;
1487
1488         for (bh = head, block_start = 0;
1489              ret == 0 && (bh != head || !block_start);
1490              block_start = block_end, bh = next) {
1491                 next = bh->b_this_page;
1492                 block_end = block_start + blocksize;
1493                 if (block_end <= from || block_start >= to) {
1494                         if (partial && !buffer_uptodate(bh))
1495                                 *partial = 1;
1496                         continue;
1497                 }
1498                 err = (*fn)(handle, bh);
1499                 if (!ret)
1500                         ret = err;
1501         }
1502         return ret;
1503 }
1504
1505 /*
1506  * To preserve ordering, it is essential that the hole instantiation and
1507  * the data write be encapsulated in a single transaction.  We cannot
1508  * close off a transaction and start a new one between the ext4_get_block()
1509  * and the commit_write().  So doing the jbd2_journal_start at the start of
1510  * prepare_write() is the right place.
1511  *
1512  * Also, this function can nest inside ext4_writepage() ->
1513  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1514  * has generated enough buffer credits to do the whole page.  So we won't
1515  * block on the journal in that case, which is good, because the caller may
1516  * be PF_MEMALLOC.
1517  *
1518  * By accident, ext4 can be reentered when a transaction is open via
1519  * quota file writes.  If we were to commit the transaction while thus
1520  * reentered, there can be a deadlock - we would be holding a quota
1521  * lock, and the commit would never complete if another thread had a
1522  * transaction open and was blocking on the quota lock - a ranking
1523  * violation.
1524  *
1525  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1526  * will _not_ run commit under these circumstances because handle->h_ref
1527  * is elevated.  We'll still have enough credits for the tiny quotafile
1528  * write.
1529  */
1530 static int do_journal_get_write_access(handle_t *handle,
1531                                        struct buffer_head *bh)
1532 {
1533         if (!buffer_mapped(bh) || buffer_freed(bh))
1534                 return 0;
1535         return ext4_journal_get_write_access(handle, bh);
1536 }
1537
1538 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1539                             loff_t pos, unsigned len, unsigned flags,
1540                             struct page **pagep, void **fsdata)
1541 {
1542         struct inode *inode = mapping->host;
1543         int ret, needed_blocks;
1544         handle_t *handle;
1545         int retries = 0;
1546         struct page *page;
1547         pgoff_t index;
1548         unsigned from, to;
1549
1550         trace_ext4_write_begin(inode, pos, len, flags);
1551         /*
1552          * Reserve one block more for addition to orphan list in case
1553          * we allocate blocks but write fails for some reason
1554          */
1555         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1556         index = pos >> PAGE_CACHE_SHIFT;
1557         from = pos & (PAGE_CACHE_SIZE - 1);
1558         to = from + len;
1559
1560 retry:
1561         handle = ext4_journal_start(inode, needed_blocks);
1562         if (IS_ERR(handle)) {
1563                 ret = PTR_ERR(handle);
1564                 goto out;
1565         }
1566
1567         /* We cannot recurse into the filesystem as the transaction is already
1568          * started */
1569         flags |= AOP_FLAG_NOFS;
1570
1571         page = grab_cache_page_write_begin(mapping, index, flags);
1572         if (!page) {
1573                 ext4_journal_stop(handle);
1574                 ret = -ENOMEM;
1575                 goto out;
1576         }
1577         *pagep = page;
1578
1579         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1580                                 ext4_get_block);
1581
1582         if (!ret && ext4_should_journal_data(inode)) {
1583                 ret = walk_page_buffers(handle, page_buffers(page),
1584                                 from, to, NULL, do_journal_get_write_access);
1585         }
1586
1587         if (ret) {
1588                 unlock_page(page);
1589                 page_cache_release(page);
1590                 /*
1591                  * block_write_begin may have instantiated a few blocks
1592                  * outside i_size.  Trim these off again. Don't need
1593                  * i_size_read because we hold i_mutex.
1594                  *
1595                  * Add inode to orphan list in case we crash before
1596                  * truncate finishes
1597                  */
1598                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1599                         ext4_orphan_add(handle, inode);
1600
1601                 ext4_journal_stop(handle);
1602                 if (pos + len > inode->i_size) {
1603                         ext4_truncate(inode);
1604                         /*
1605                          * If truncate failed early the inode might
1606                          * still be on the orphan list; we need to
1607                          * make sure the inode is removed from the
1608                          * orphan list in that case.
1609                          */
1610                         if (inode->i_nlink)
1611                                 ext4_orphan_del(NULL, inode);
1612                 }
1613         }
1614
1615         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1616                 goto retry;
1617 out:
1618         return ret;
1619 }
1620
1621 /* For write_end() in data=journal mode */
1622 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1623 {
1624         if (!buffer_mapped(bh) || buffer_freed(bh))
1625                 return 0;
1626         set_buffer_uptodate(bh);
1627         return ext4_handle_dirty_metadata(handle, NULL, bh);
1628 }
1629
1630 static int ext4_generic_write_end(struct file *file,
1631                                   struct address_space *mapping,
1632                                   loff_t pos, unsigned len, unsigned copied,
1633                                   struct page *page, void *fsdata)
1634 {
1635         int i_size_changed = 0;
1636         struct inode *inode = mapping->host;
1637         handle_t *handle = ext4_journal_current_handle();
1638
1639         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1640
1641         /*
1642          * No need to use i_size_read() here, the i_size
1643          * cannot change under us because we hold i_mutex.
1644          *
1645          * But it's important to update i_size while still holding page lock:
1646          * page writeout could otherwise come in and zero beyond i_size.
1647          */
1648         if (pos + copied > inode->i_size) {
1649                 i_size_write(inode, pos + copied);
1650                 i_size_changed = 1;
1651         }
1652
1653         if (pos + copied >  EXT4_I(inode)->i_disksize) {
1654                 /* We need to mark inode dirty even if
1655                  * new_i_size is less that inode->i_size
1656                  * bu greater than i_disksize.(hint delalloc)
1657                  */
1658                 ext4_update_i_disksize(inode, (pos + copied));
1659                 i_size_changed = 1;
1660         }
1661         unlock_page(page);
1662         page_cache_release(page);
1663
1664         /*
1665          * Don't mark the inode dirty under page lock. First, it unnecessarily
1666          * makes the holding time of page lock longer. Second, it forces lock
1667          * ordering of page lock and transaction start for journaling
1668          * filesystems.
1669          */
1670         if (i_size_changed)
1671                 ext4_mark_inode_dirty(handle, inode);
1672
1673         return copied;
1674 }
1675
1676 /*
1677  * We need to pick up the new inode size which generic_commit_write gave us
1678  * `file' can be NULL - eg, when called from page_symlink().
1679  *
1680  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1681  * buffers are managed internally.
1682  */
1683 static int ext4_ordered_write_end(struct file *file,
1684                                   struct address_space *mapping,
1685                                   loff_t pos, unsigned len, unsigned copied,
1686                                   struct page *page, void *fsdata)
1687 {
1688         handle_t *handle = ext4_journal_current_handle();
1689         struct inode *inode = mapping->host;
1690         int ret = 0, ret2;
1691
1692         trace_ext4_ordered_write_end(inode, pos, len, copied);
1693         ret = ext4_jbd2_file_inode(handle, inode);
1694
1695         if (ret == 0) {
1696                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1697                                                         page, fsdata);
1698                 copied = ret2;
1699                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1700                         /* if we have allocated more blocks and copied
1701                          * less. We will have blocks allocated outside
1702                          * inode->i_size. So truncate them
1703                          */
1704                         ext4_orphan_add(handle, inode);
1705                 if (ret2 < 0)
1706                         ret = ret2;
1707         }
1708         ret2 = ext4_journal_stop(handle);
1709         if (!ret)
1710                 ret = ret2;
1711
1712         if (pos + len > inode->i_size) {
1713                 ext4_truncate(inode);
1714                 /*
1715                  * If truncate failed early the inode might still be
1716                  * on the orphan list; we need to make sure the inode
1717                  * is removed from the orphan list in that case.
1718                  */
1719                 if (inode->i_nlink)
1720                         ext4_orphan_del(NULL, inode);
1721         }
1722
1723
1724         return ret ? ret : copied;
1725 }
1726
1727 static int ext4_writeback_write_end(struct file *file,
1728                                     struct address_space *mapping,
1729                                     loff_t pos, unsigned len, unsigned copied,
1730                                     struct page *page, void *fsdata)
1731 {
1732         handle_t *handle = ext4_journal_current_handle();
1733         struct inode *inode = mapping->host;
1734         int ret = 0, ret2;
1735
1736         trace_ext4_writeback_write_end(inode, pos, len, copied);
1737         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1738                                                         page, fsdata);
1739         copied = ret2;
1740         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1741                 /* if we have allocated more blocks and copied
1742                  * less. We will have blocks allocated outside
1743                  * inode->i_size. So truncate them
1744                  */
1745                 ext4_orphan_add(handle, inode);
1746
1747         if (ret2 < 0)
1748                 ret = ret2;
1749
1750         ret2 = ext4_journal_stop(handle);
1751         if (!ret)
1752                 ret = ret2;
1753
1754         if (pos + len > inode->i_size) {
1755                 ext4_truncate(inode);
1756                 /*
1757                  * If truncate failed early the inode might still be
1758                  * on the orphan list; we need to make sure the inode
1759                  * is removed from the orphan list in that case.
1760                  */
1761                 if (inode->i_nlink)
1762                         ext4_orphan_del(NULL, inode);
1763         }
1764
1765         return ret ? ret : copied;
1766 }
1767
1768 static int ext4_journalled_write_end(struct file *file,
1769                                      struct address_space *mapping,
1770                                      loff_t pos, unsigned len, unsigned copied,
1771                                      struct page *page, void *fsdata)
1772 {
1773         handle_t *handle = ext4_journal_current_handle();
1774         struct inode *inode = mapping->host;
1775         int ret = 0, ret2;
1776         int partial = 0;
1777         unsigned from, to;
1778         loff_t new_i_size;
1779
1780         trace_ext4_journalled_write_end(inode, pos, len, copied);
1781         from = pos & (PAGE_CACHE_SIZE - 1);
1782         to = from + len;
1783
1784         if (copied < len) {
1785                 if (!PageUptodate(page))
1786                         copied = 0;
1787                 page_zero_new_buffers(page, from+copied, to);
1788         }
1789
1790         ret = walk_page_buffers(handle, page_buffers(page), from,
1791                                 to, &partial, write_end_fn);
1792         if (!partial)
1793                 SetPageUptodate(page);
1794         new_i_size = pos + copied;
1795         if (new_i_size > inode->i_size)
1796                 i_size_write(inode, pos+copied);
1797         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
1798         if (new_i_size > EXT4_I(inode)->i_disksize) {
1799                 ext4_update_i_disksize(inode, new_i_size);
1800                 ret2 = ext4_mark_inode_dirty(handle, inode);
1801                 if (!ret)
1802                         ret = ret2;
1803         }
1804
1805         unlock_page(page);
1806         page_cache_release(page);
1807         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1808                 /* if we have allocated more blocks and copied
1809                  * less. We will have blocks allocated outside
1810                  * inode->i_size. So truncate them
1811                  */
1812                 ext4_orphan_add(handle, inode);
1813
1814         ret2 = ext4_journal_stop(handle);
1815         if (!ret)
1816                 ret = ret2;
1817         if (pos + len > inode->i_size) {
1818                 ext4_truncate(inode);
1819                 /*
1820                  * If truncate failed early the inode might still be
1821                  * on the orphan list; we need to make sure the inode
1822                  * is removed from the orphan list in that case.
1823                  */
1824                 if (inode->i_nlink)
1825                         ext4_orphan_del(NULL, inode);
1826         }
1827
1828         return ret ? ret : copied;
1829 }
1830
1831 static int ext4_da_reserve_space(struct inode *inode, int nrblocks)
1832 {
1833         int retries = 0;
1834         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1835         unsigned long md_needed, mdblocks, total = 0;
1836
1837         /*
1838          * recalculate the amount of metadata blocks to reserve
1839          * in order to allocate nrblocks
1840          * worse case is one extent per block
1841          */
1842 repeat:
1843         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1844         total = EXT4_I(inode)->i_reserved_data_blocks + nrblocks;
1845         mdblocks = ext4_calc_metadata_amount(inode, total);
1846         BUG_ON(mdblocks < EXT4_I(inode)->i_reserved_meta_blocks);
1847
1848         md_needed = mdblocks - EXT4_I(inode)->i_reserved_meta_blocks;
1849         total = md_needed + nrblocks;
1850
1851         /*
1852          * Make quota reservation here to prevent quota overflow
1853          * later. Real quota accounting is done at pages writeout
1854          * time.
1855          */
1856         if (vfs_dq_reserve_block(inode, total)) {
1857                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1858                 return -EDQUOT;
1859         }
1860
1861         if (ext4_claim_free_blocks(sbi, total)) {
1862                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1863                 vfs_dq_release_reservation_block(inode, total);
1864                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1865                         yield();
1866                         goto repeat;
1867                 }
1868                 return -ENOSPC;
1869         }
1870         EXT4_I(inode)->i_reserved_data_blocks += nrblocks;
1871         EXT4_I(inode)->i_reserved_meta_blocks = mdblocks;
1872
1873         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1874         return 0;       /* success */
1875 }
1876
1877 static void ext4_da_release_space(struct inode *inode, int to_free)
1878 {
1879         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1880         int total, mdb, mdb_free, release;
1881
1882         if (!to_free)
1883                 return;         /* Nothing to release, exit */
1884
1885         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1886
1887         if (!EXT4_I(inode)->i_reserved_data_blocks) {
1888                 /*
1889                  * if there is no reserved blocks, but we try to free some
1890                  * then the counter is messed up somewhere.
1891                  * but since this function is called from invalidate
1892                  * page, it's harmless to return without any action
1893                  */
1894                 printk(KERN_INFO "ext4 delalloc try to release %d reserved "
1895                             "blocks for inode %lu, but there is no reserved "
1896                             "data blocks\n", to_free, inode->i_ino);
1897                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1898                 return;
1899         }
1900
1901         /* recalculate the number of metablocks still need to be reserved */
1902         total = EXT4_I(inode)->i_reserved_data_blocks - to_free;
1903         mdb = ext4_calc_metadata_amount(inode, total);
1904
1905         /* figure out how many metablocks to release */
1906         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1907         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1908
1909         release = to_free + mdb_free;
1910
1911         /* update fs dirty blocks counter for truncate case */
1912         percpu_counter_sub(&sbi->s_dirtyblocks_counter, release);
1913
1914         /* update per-inode reservations */
1915         BUG_ON(to_free > EXT4_I(inode)->i_reserved_data_blocks);
1916         EXT4_I(inode)->i_reserved_data_blocks -= to_free;
1917
1918         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1919         EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1920         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1921
1922         vfs_dq_release_reservation_block(inode, release);
1923 }
1924
1925 static void ext4_da_page_release_reservation(struct page *page,
1926                                              unsigned long offset)
1927 {
1928         int to_release = 0;
1929         struct buffer_head *head, *bh;
1930         unsigned int curr_off = 0;
1931
1932         head = page_buffers(page);
1933         bh = head;
1934         do {
1935                 unsigned int next_off = curr_off + bh->b_size;
1936
1937                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1938                         to_release++;
1939                         clear_buffer_delay(bh);
1940                 }
1941                 curr_off = next_off;
1942         } while ((bh = bh->b_this_page) != head);
1943         ext4_da_release_space(page->mapping->host, to_release);
1944 }
1945
1946 /*
1947  * Delayed allocation stuff
1948  */
1949
1950 /*
1951  * mpage_da_submit_io - walks through extent of pages and try to write
1952  * them with writepage() call back
1953  *
1954  * @mpd->inode: inode
1955  * @mpd->first_page: first page of the extent
1956  * @mpd->next_page: page after the last page of the extent
1957  *
1958  * By the time mpage_da_submit_io() is called we expect all blocks
1959  * to be allocated. this may be wrong if allocation failed.
1960  *
1961  * As pages are already locked by write_cache_pages(), we can't use it
1962  */
1963 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1964 {
1965         long pages_skipped;
1966         struct pagevec pvec;
1967         unsigned long index, end;
1968         int ret = 0, err, nr_pages, i;
1969         struct inode *inode = mpd->inode;
1970         struct address_space *mapping = inode->i_mapping;
1971
1972         BUG_ON(mpd->next_page <= mpd->first_page);
1973         /*
1974          * We need to start from the first_page to the next_page - 1
1975          * to make sure we also write the mapped dirty buffer_heads.
1976          * If we look at mpd->b_blocknr we would only be looking
1977          * at the currently mapped buffer_heads.
1978          */
1979         index = mpd->first_page;
1980         end = mpd->next_page - 1;
1981
1982         pagevec_init(&pvec, 0);
1983         while (index <= end) {
1984                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1985                 if (nr_pages == 0)
1986                         break;
1987                 for (i = 0; i < nr_pages; i++) {
1988                         struct page *page = pvec.pages[i];
1989
1990                         index = page->index;
1991                         if (index > end)
1992                                 break;
1993                         index++;
1994
1995                         BUG_ON(!PageLocked(page));
1996                         BUG_ON(PageWriteback(page));
1997
1998                         pages_skipped = mpd->wbc->pages_skipped;
1999                         err = mapping->a_ops->writepage(page, mpd->wbc);
2000                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
2001                                 /*
2002                                  * have successfully written the page
2003                                  * without skipping the same
2004                                  */
2005                                 mpd->pages_written++;
2006                         /*
2007                          * In error case, we have to continue because
2008                          * remaining pages are still locked
2009                          * XXX: unlock and re-dirty them?
2010                          */
2011                         if (ret == 0)
2012                                 ret = err;
2013                 }
2014                 pagevec_release(&pvec);
2015         }
2016         return ret;
2017 }
2018
2019 /*
2020  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
2021  *
2022  * @mpd->inode - inode to walk through
2023  * @exbh->b_blocknr - first block on a disk
2024  * @exbh->b_size - amount of space in bytes
2025  * @logical - first logical block to start assignment with
2026  *
2027  * the function goes through all passed space and put actual disk
2028  * block numbers into buffer heads, dropping BH_Delay and BH_Unwritten
2029  */
2030 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
2031                                  struct buffer_head *exbh)
2032 {
2033         struct inode *inode = mpd->inode;
2034         struct address_space *mapping = inode->i_mapping;
2035         int blocks = exbh->b_size >> inode->i_blkbits;
2036         sector_t pblock = exbh->b_blocknr, cur_logical;
2037         struct buffer_head *head, *bh;
2038         pgoff_t index, end;
2039         struct pagevec pvec;
2040         int nr_pages, i;
2041
2042         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2043         end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2044         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2045
2046         pagevec_init(&pvec, 0);
2047
2048         while (index <= end) {
2049                 /* XXX: optimize tail */
2050                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2051                 if (nr_pages == 0)
2052                         break;
2053                 for (i = 0; i < nr_pages; i++) {
2054                         struct page *page = pvec.pages[i];
2055
2056                         index = page->index;
2057                         if (index > end)
2058                                 break;
2059                         index++;
2060
2061                         BUG_ON(!PageLocked(page));
2062                         BUG_ON(PageWriteback(page));
2063                         BUG_ON(!page_has_buffers(page));
2064
2065                         bh = page_buffers(page);
2066                         head = bh;
2067
2068                         /* skip blocks out of the range */
2069                         do {
2070                                 if (cur_logical >= logical)
2071                                         break;
2072                                 cur_logical++;
2073                         } while ((bh = bh->b_this_page) != head);
2074
2075                         do {
2076                                 if (cur_logical >= logical + blocks)
2077                                         break;
2078
2079                                 if (buffer_delay(bh) ||
2080                                                 buffer_unwritten(bh)) {
2081
2082                                         BUG_ON(bh->b_bdev != inode->i_sb->s_bdev);
2083
2084                                         if (buffer_delay(bh)) {
2085                                                 clear_buffer_delay(bh);
2086                                                 bh->b_blocknr = pblock;
2087                                         } else {
2088                                                 /*
2089                                                  * unwritten already should have
2090                                                  * blocknr assigned. Verify that
2091                                                  */
2092                                                 clear_buffer_unwritten(bh);
2093                                                 BUG_ON(bh->b_blocknr != pblock);
2094                                         }
2095
2096                                 } else if (buffer_mapped(bh))
2097                                         BUG_ON(bh->b_blocknr != pblock);
2098
2099                                 cur_logical++;
2100                                 pblock++;
2101                         } while ((bh = bh->b_this_page) != head);
2102                 }
2103                 pagevec_release(&pvec);
2104         }
2105 }
2106
2107
2108 /*
2109  * __unmap_underlying_blocks - just a helper function to unmap
2110  * set of blocks described by @bh
2111  */
2112 static inline void __unmap_underlying_blocks(struct inode *inode,
2113                                              struct buffer_head *bh)
2114 {
2115         struct block_device *bdev = inode->i_sb->s_bdev;
2116         int blocks, i;
2117
2118         blocks = bh->b_size >> inode->i_blkbits;
2119         for (i = 0; i < blocks; i++)
2120                 unmap_underlying_metadata(bdev, bh->b_blocknr + i);
2121 }
2122
2123 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2124                                         sector_t logical, long blk_cnt)
2125 {
2126         int nr_pages, i;
2127         pgoff_t index, end;
2128         struct pagevec pvec;
2129         struct inode *inode = mpd->inode;
2130         struct address_space *mapping = inode->i_mapping;
2131
2132         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2133         end   = (logical + blk_cnt - 1) >>
2134                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2135         while (index <= end) {
2136                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2137                 if (nr_pages == 0)
2138                         break;
2139                 for (i = 0; i < nr_pages; i++) {
2140                         struct page *page = pvec.pages[i];
2141                         index = page->index;
2142                         if (index > end)
2143                                 break;
2144                         index++;
2145
2146                         BUG_ON(!PageLocked(page));
2147                         BUG_ON(PageWriteback(page));
2148                         block_invalidatepage(page, 0);
2149                         ClearPageUptodate(page);
2150                         unlock_page(page);
2151                 }
2152         }
2153         return;
2154 }
2155
2156 static void ext4_print_free_blocks(struct inode *inode)
2157 {
2158         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2159         printk(KERN_CRIT "Total free blocks count %lld\n",
2160                ext4_count_free_blocks(inode->i_sb));
2161         printk(KERN_CRIT "Free/Dirty block details\n");
2162         printk(KERN_CRIT "free_blocks=%lld\n",
2163                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2164         printk(KERN_CRIT "dirty_blocks=%lld\n",
2165                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2166         printk(KERN_CRIT "Block reservation details\n");
2167         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2168                EXT4_I(inode)->i_reserved_data_blocks);
2169         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2170                EXT4_I(inode)->i_reserved_meta_blocks);
2171         return;
2172 }
2173
2174 /*
2175  * mpage_da_map_blocks - go through given space
2176  *
2177  * @mpd - bh describing space
2178  *
2179  * The function skips space we know is already mapped to disk blocks.
2180  *
2181  */
2182 static int mpage_da_map_blocks(struct mpage_da_data *mpd)
2183 {
2184         int err, blks, get_blocks_flags;
2185         struct buffer_head new;
2186         sector_t next = mpd->b_blocknr;
2187         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2188         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2189         handle_t *handle = NULL;
2190
2191         /*
2192          * We consider only non-mapped and non-allocated blocks
2193          */
2194         if ((mpd->b_state  & (1 << BH_Mapped)) &&
2195                 !(mpd->b_state & (1 << BH_Delay)) &&
2196                 !(mpd->b_state & (1 << BH_Unwritten)))
2197                 return 0;
2198
2199         /*
2200          * If we didn't accumulate anything to write simply return
2201          */
2202         if (!mpd->b_size)
2203                 return 0;
2204
2205         handle = ext4_journal_current_handle();
2206         BUG_ON(!handle);
2207
2208         /*
2209          * Call ext4_get_blocks() to allocate any delayed allocation
2210          * blocks, or to convert an uninitialized extent to be
2211          * initialized (in the case where we have written into
2212          * one or more preallocated blocks).
2213          *
2214          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2215          * indicate that we are on the delayed allocation path.  This
2216          * affects functions in many different parts of the allocation
2217          * call path.  This flag exists primarily because we don't
2218          * want to change *many* call functions, so ext4_get_blocks()
2219          * will set the magic i_delalloc_reserved_flag once the
2220          * inode's allocation semaphore is taken.
2221          *
2222          * If the blocks in questions were delalloc blocks, set
2223          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2224          * variables are updated after the blocks have been allocated.
2225          */
2226         new.b_state = 0;
2227         get_blocks_flags = (EXT4_GET_BLOCKS_CREATE |
2228                             EXT4_GET_BLOCKS_DELALLOC_RESERVE);
2229         if (mpd->b_state & (1 << BH_Delay))
2230                 get_blocks_flags |= EXT4_GET_BLOCKS_UPDATE_RESERVE_SPACE;
2231         blks = ext4_get_blocks(handle, mpd->inode, next, max_blocks,
2232                                &new, get_blocks_flags);
2233         if (blks < 0) {
2234                 err = blks;
2235                 /*
2236                  * If get block returns with error we simply
2237                  * return. Later writepage will redirty the page and
2238                  * writepages will find the dirty page again
2239                  */
2240                 if (err == -EAGAIN)
2241                         return 0;
2242
2243                 if (err == -ENOSPC &&
2244                     ext4_count_free_blocks(mpd->inode->i_sb)) {
2245                         mpd->retval = err;
2246                         return 0;
2247                 }
2248
2249                 /*
2250                  * get block failure will cause us to loop in
2251                  * writepages, because a_ops->writepage won't be able
2252                  * to make progress. The page will be redirtied by
2253                  * writepage and writepages will again try to write
2254                  * the same.
2255                  */
2256                 ext4_msg(mpd->inode->i_sb, KERN_CRIT,
2257                          "delayed block allocation failed for inode %lu at "
2258                          "logical offset %llu with max blocks %zd with "
2259                          "error %d\n", mpd->inode->i_ino,
2260                          (unsigned long long) next,
2261                          mpd->b_size >> mpd->inode->i_blkbits, err);
2262                 printk(KERN_CRIT "This should not happen!!  "
2263                        "Data will be lost\n");
2264                 if (err == -ENOSPC) {
2265                         ext4_print_free_blocks(mpd->inode);
2266                 }
2267                 /* invalidate all the pages */
2268                 ext4_da_block_invalidatepages(mpd, next,
2269                                 mpd->b_size >> mpd->inode->i_blkbits);
2270                 return err;
2271         }
2272         BUG_ON(blks == 0);
2273
2274         new.b_size = (blks << mpd->inode->i_blkbits);
2275
2276         if (buffer_new(&new))
2277                 __unmap_underlying_blocks(mpd->inode, &new);
2278
2279         /*
2280          * If blocks are delayed marked, we need to
2281          * put actual blocknr and drop delayed bit
2282          */
2283         if ((mpd->b_state & (1 << BH_Delay)) ||
2284             (mpd->b_state & (1 << BH_Unwritten)))
2285                 mpage_put_bnr_to_bhs(mpd, next, &new);
2286
2287         if (ext4_should_order_data(mpd->inode)) {
2288                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2289                 if (err)
2290                         return err;
2291         }
2292
2293         /*
2294          * Update on-disk size along with block allocation.
2295          */
2296         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2297         if (disksize > i_size_read(mpd->inode))
2298                 disksize = i_size_read(mpd->inode);
2299         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2300                 ext4_update_i_disksize(mpd->inode, disksize);
2301                 return ext4_mark_inode_dirty(handle, mpd->inode);
2302         }
2303
2304         return 0;
2305 }
2306
2307 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2308                 (1 << BH_Delay) | (1 << BH_Unwritten))
2309
2310 /*
2311  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2312  *
2313  * @mpd->lbh - extent of blocks
2314  * @logical - logical number of the block in the file
2315  * @bh - bh of the block (used to access block's state)
2316  *
2317  * the function is used to collect contig. blocks in same state
2318  */
2319 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2320                                    sector_t logical, size_t b_size,
2321                                    unsigned long b_state)
2322 {
2323         sector_t next;
2324         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2325
2326         /* check if thereserved journal credits might overflow */
2327         if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
2328                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2329                         /*
2330                          * With non-extent format we are limited by the journal
2331                          * credit available.  Total credit needed to insert
2332                          * nrblocks contiguous blocks is dependent on the
2333                          * nrblocks.  So limit nrblocks.
2334                          */
2335                         goto flush_it;
2336                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2337                                 EXT4_MAX_TRANS_DATA) {
2338                         /*
2339                          * Adding the new buffer_head would make it cross the
2340                          * allowed limit for which we have journal credit
2341                          * reserved. So limit the new bh->b_size
2342                          */
2343                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2344                                                 mpd->inode->i_blkbits;
2345                         /* we will do mpage_da_submit_io in the next loop */
2346                 }
2347         }
2348         /*
2349          * First block in the extent
2350          */
2351         if (mpd->b_size == 0) {
2352                 mpd->b_blocknr = logical;
2353                 mpd->b_size = b_size;
2354                 mpd->b_state = b_state & BH_FLAGS;
2355                 return;
2356         }
2357
2358         next = mpd->b_blocknr + nrblocks;
2359         /*
2360          * Can we merge the block to our big extent?
2361          */
2362         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2363                 mpd->b_size += b_size;
2364                 return;
2365         }
2366
2367 flush_it:
2368         /*
2369          * We couldn't merge the block to our extent, so we
2370          * need to flush current  extent and start new one
2371          */
2372         if (mpage_da_map_blocks(mpd) == 0)
2373                 mpage_da_submit_io(mpd);
2374         mpd->io_done = 1;
2375         return;
2376 }
2377
2378 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
2379 {
2380         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
2381 }
2382
2383 /*
2384  * __mpage_da_writepage - finds extent of pages and blocks
2385  *
2386  * @page: page to consider
2387  * @wbc: not used, we just follow rules
2388  * @data: context
2389  *
2390  * The function finds extents of pages and scan them for all blocks.
2391  */
2392 static int __mpage_da_writepage(struct page *page,
2393                                 struct writeback_control *wbc, void *data)
2394 {
2395         struct mpage_da_data *mpd = data;
2396         struct inode *inode = mpd->inode;
2397         struct buffer_head *bh, *head;
2398         sector_t logical;
2399
2400         if (mpd->io_done) {
2401                 /*
2402                  * Rest of the page in the page_vec
2403                  * redirty then and skip then. We will
2404                  * try to write them again after
2405                  * starting a new transaction
2406                  */
2407                 redirty_page_for_writepage(wbc, page);
2408                 unlock_page(page);
2409                 return MPAGE_DA_EXTENT_TAIL;
2410         }
2411         /*
2412          * Can we merge this page to current extent?
2413          */
2414         if (mpd->next_page != page->index) {
2415                 /*
2416                  * Nope, we can't. So, we map non-allocated blocks
2417                  * and start IO on them using writepage()
2418                  */
2419                 if (mpd->next_page != mpd->first_page) {
2420                         if (mpage_da_map_blocks(mpd) == 0)
2421                                 mpage_da_submit_io(mpd);
2422                         /*
2423                          * skip rest of the page in the page_vec
2424                          */
2425                         mpd->io_done = 1;
2426                         redirty_page_for_writepage(wbc, page);
2427                         unlock_page(page);
2428                         return MPAGE_DA_EXTENT_TAIL;
2429                 }
2430
2431                 /*
2432                  * Start next extent of pages ...
2433                  */
2434                 mpd->first_page = page->index;
2435
2436                 /*
2437                  * ... and blocks
2438                  */
2439                 mpd->b_size = 0;
2440                 mpd->b_state = 0;
2441                 mpd->b_blocknr = 0;
2442         }
2443
2444         mpd->next_page = page->index + 1;
2445         logical = (sector_t) page->index <<
2446                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2447
2448         if (!page_has_buffers(page)) {
2449                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2450                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2451                 if (mpd->io_done)
2452                         return MPAGE_DA_EXTENT_TAIL;
2453         } else {
2454                 /*
2455                  * Page with regular buffer heads, just add all dirty ones
2456                  */
2457                 head = page_buffers(page);
2458                 bh = head;
2459                 do {
2460                         BUG_ON(buffer_locked(bh));
2461                         /*
2462                          * We need to try to allocate
2463                          * unmapped blocks in the same page.
2464                          * Otherwise we won't make progress
2465                          * with the page in ext4_writepage
2466                          */
2467                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2468                                 mpage_add_bh_to_extent(mpd, logical,
2469                                                        bh->b_size,
2470                                                        bh->b_state);
2471                                 if (mpd->io_done)
2472                                         return MPAGE_DA_EXTENT_TAIL;
2473                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2474                                 /*
2475                                  * mapped dirty buffer. We need to update
2476                                  * the b_state because we look at
2477                                  * b_state in mpage_da_map_blocks. We don't
2478                                  * update b_size because if we find an
2479                                  * unmapped buffer_head later we need to
2480                                  * use the b_state flag of that buffer_head.
2481                                  */
2482                                 if (mpd->b_size == 0)
2483                                         mpd->b_state = bh->b_state & BH_FLAGS;
2484                         }
2485                         logical++;
2486                 } while ((bh = bh->b_this_page) != head);
2487         }
2488
2489         return 0;
2490 }
2491
2492 /*
2493  * This is a special get_blocks_t callback which is used by
2494  * ext4_da_write_begin().  It will either return mapped block or
2495  * reserve space for a single block.
2496  *
2497  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2498  * We also have b_blocknr = -1 and b_bdev initialized properly
2499  *
2500  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2501  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2502  * initialized properly.
2503  */
2504 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2505                                   struct buffer_head *bh_result, int create)
2506 {
2507         int ret = 0;
2508         sector_t invalid_block = ~((sector_t) 0xffff);
2509
2510         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2511                 invalid_block = ~0;
2512
2513         BUG_ON(create == 0);
2514         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2515
2516         /*
2517          * first, we need to know whether the block is allocated already
2518          * preallocated blocks are unmapped but should treated
2519          * the same as allocated blocks.
2520          */
2521         ret = ext4_get_blocks(NULL, inode, iblock, 1,  bh_result, 0);
2522         if ((ret == 0) && !buffer_delay(bh_result)) {
2523                 /* the block isn't (pre)allocated yet, let's reserve space */
2524                 /*
2525                  * XXX: __block_prepare_write() unmaps passed block,
2526                  * is it OK?
2527                  */
2528                 ret = ext4_da_reserve_space(inode, 1);
2529                 if (ret)
2530                         /* not enough space to reserve */
2531                         return ret;
2532
2533                 map_bh(bh_result, inode->i_sb, invalid_block);
2534                 set_buffer_new(bh_result);
2535                 set_buffer_delay(bh_result);
2536         } else if (ret > 0) {
2537                 bh_result->b_size = (ret << inode->i_blkbits);
2538                 if (buffer_unwritten(bh_result)) {
2539                         /* A delayed write to unwritten bh should
2540                          * be marked new and mapped.  Mapped ensures
2541                          * that we don't do get_block multiple times
2542                          * when we write to the same offset and new
2543                          * ensures that we do proper zero out for
2544                          * partial write.
2545                          */
2546                         set_buffer_new(bh_result);
2547                         set_buffer_mapped(bh_result);
2548                 }
2549                 ret = 0;
2550         }
2551
2552         return ret;
2553 }
2554
2555 /*
2556  * This function is used as a standard get_block_t calback function
2557  * when there is no desire to allocate any blocks.  It is used as a
2558  * callback function for block_prepare_write(), nobh_writepage(), and
2559  * block_write_full_page().  These functions should only try to map a
2560  * single block at a time.
2561  *
2562  * Since this function doesn't do block allocations even if the caller
2563  * requests it by passing in create=1, it is critically important that
2564  * any caller checks to make sure that any buffer heads are returned
2565  * by this function are either all already mapped or marked for
2566  * delayed allocation before calling nobh_writepage() or
2567  * block_write_full_page().  Otherwise, b_blocknr could be left
2568  * unitialized, and the page write functions will be taken by
2569  * surprise.
2570  */
2571 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2572                                    struct buffer_head *bh_result, int create)
2573 {
2574         int ret = 0;
2575         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2576
2577         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2578
2579         /*
2580          * we don't want to do block allocation in writepage
2581          * so call get_block_wrap with create = 0
2582          */
2583         ret = ext4_get_blocks(NULL, inode, iblock, max_blocks, bh_result, 0);
2584         if (ret > 0) {
2585                 bh_result->b_size = (ret << inode->i_blkbits);
2586                 ret = 0;
2587         }
2588         return ret;
2589 }
2590
2591 static int bget_one(handle_t *handle, struct buffer_head *bh)
2592 {
2593         get_bh(bh);
2594         return 0;
2595 }
2596
2597 static int bput_one(handle_t *handle, struct buffer_head *bh)
2598 {
2599         put_bh(bh);
2600         return 0;
2601 }
2602
2603 static int __ext4_journalled_writepage(struct page *page,
2604                                        struct writeback_control *wbc,
2605                                        unsigned int len)
2606 {
2607         struct address_space *mapping = page->mapping;
2608         struct inode *inode = mapping->host;
2609         struct buffer_head *page_bufs;
2610         handle_t *handle = NULL;
2611         int ret = 0;
2612         int err;
2613
2614         page_bufs = page_buffers(page);
2615         BUG_ON(!page_bufs);
2616         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2617         /* As soon as we unlock the page, it can go away, but we have
2618          * references to buffers so we are safe */
2619         unlock_page(page);
2620
2621         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2622         if (IS_ERR(handle)) {
2623                 ret = PTR_ERR(handle);
2624                 goto out;
2625         }
2626
2627         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2628                                 do_journal_get_write_access);
2629
2630         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2631                                 write_end_fn);
2632         if (ret == 0)
2633                 ret = err;
2634         err = ext4_journal_stop(handle);
2635         if (!ret)
2636                 ret = err;
2637
2638         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2639         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
2640 out:
2641         return ret;
2642 }
2643
2644 /*
2645  * Note that we don't need to start a transaction unless we're journaling data
2646  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2647  * need to file the inode to the transaction's list in ordered mode because if
2648  * we are writing back data added by write(), the inode is already there and if
2649  * we are writing back data modified via mmap(), noone guarantees in which
2650  * transaction the data will hit the disk. In case we are journaling data, we
2651  * cannot start transaction directly because transaction start ranks above page
2652  * lock so we have to do some magic.
2653  *
2654  * This function can get called via...
2655  *   - ext4_da_writepages after taking page lock (have journal handle)
2656  *   - journal_submit_inode_data_buffers (no journal handle)
2657  *   - shrink_page_list via pdflush (no journal handle)
2658  *   - grab_page_cache when doing write_begin (have journal handle)
2659  *
2660  * We don't do any block allocation in this function. If we have page with
2661  * multiple blocks we need to write those buffer_heads that are mapped. This
2662  * is important for mmaped based write. So if we do with blocksize 1K
2663  * truncate(f, 1024);
2664  * a = mmap(f, 0, 4096);
2665  * a[0] = 'a';
2666  * truncate(f, 4096);
2667  * we have in the page first buffer_head mapped via page_mkwrite call back
2668  * but other bufer_heads would be unmapped but dirty(dirty done via the
2669  * do_wp_page). So writepage should write the first block. If we modify
2670  * the mmap area beyond 1024 we will again get a page_fault and the
2671  * page_mkwrite callback will do the block allocation and mark the
2672  * buffer_heads mapped.
2673  *
2674  * We redirty the page if we have any buffer_heads that is either delay or
2675  * unwritten in the page.
2676  *
2677  * We can get recursively called as show below.
2678  *
2679  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2680  *              ext4_writepage()
2681  *
2682  * But since we don't do any block allocation we should not deadlock.
2683  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2684  */
2685 static int ext4_writepage(struct page *page,
2686                           struct writeback_control *wbc)
2687 {
2688         int ret = 0;
2689         loff_t size;
2690         unsigned int len;
2691         struct buffer_head *page_bufs;
2692         struct inode *inode = page->mapping->host;
2693
2694         trace_ext4_writepage(inode, page);
2695         size = i_size_read(inode);
2696         if (page->index == size >> PAGE_CACHE_SHIFT)
2697                 len = size & ~PAGE_CACHE_MASK;
2698         else
2699                 len = PAGE_CACHE_SIZE;
2700
2701         if (page_has_buffers(page)) {
2702                 page_bufs = page_buffers(page);
2703                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2704                                         ext4_bh_delay_or_unwritten)) {
2705                         /*
2706                          * We don't want to do  block allocation
2707                          * So redirty the page and return
2708                          * We may reach here when we do a journal commit
2709                          * via journal_submit_inode_data_buffers.
2710                          * If we don't have mapping block we just ignore
2711                          * them. We can also reach here via shrink_page_list
2712                          */
2713                         redirty_page_for_writepage(wbc, page);
2714                         unlock_page(page);
2715                         return 0;
2716                 }
2717         } else {
2718                 /*
2719                  * The test for page_has_buffers() is subtle:
2720                  * We know the page is dirty but it lost buffers. That means
2721                  * that at some moment in time after write_begin()/write_end()
2722                  * has been called all buffers have been clean and thus they
2723                  * must have been written at least once. So they are all
2724                  * mapped and we can happily proceed with mapping them
2725                  * and writing the page.
2726                  *
2727                  * Try to initialize the buffer_heads and check whether
2728                  * all are mapped and non delay. We don't want to
2729                  * do block allocation here.
2730                  */
2731                 ret = block_prepare_write(page, 0, len,
2732                                           noalloc_get_block_write);
2733                 if (!ret) {
2734                         page_bufs = page_buffers(page);
2735                         /* check whether all are mapped and non delay */
2736                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2737                                                 ext4_bh_delay_or_unwritten)) {
2738                                 redirty_page_for_writepage(wbc, page);
2739                                 unlock_page(page);
2740                                 return 0;
2741                         }
2742                 } else {
2743                         /*
2744                          * We can't do block allocation here
2745                          * so just redity the page and unlock
2746                          * and return
2747                          */
2748                         redirty_page_for_writepage(wbc, page);
2749                         unlock_page(page);
2750                         return 0;
2751                 }
2752                 /* now mark the buffer_heads as dirty and uptodate */
2753                 block_commit_write(page, 0, len);
2754         }
2755
2756         if (PageChecked(page) && ext4_should_journal_data(inode)) {
2757                 /*
2758                  * It's mmapped pagecache.  Add buffers and journal it.  There
2759                  * doesn't seem much point in redirtying the page here.
2760                  */
2761                 ClearPageChecked(page);
2762                 return __ext4_journalled_writepage(page, wbc, len);
2763         }
2764
2765         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2766                 ret = nobh_writepage(page, noalloc_get_block_write, wbc);
2767         else
2768                 ret = block_write_full_page(page, noalloc_get_block_write,
2769                                             wbc);
2770
2771         return ret;
2772 }
2773
2774 /*
2775  * This is called via ext4_da_writepages() to
2776  * calulate the total number of credits to reserve to fit
2777  * a single extent allocation into a single transaction,
2778  * ext4_da_writpeages() will loop calling this before
2779  * the block allocation.
2780  */
2781
2782 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2783 {
2784         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2785
2786         /*
2787          * With non-extent format the journal credit needed to
2788          * insert nrblocks contiguous block is dependent on
2789          * number of contiguous block. So we will limit
2790          * number of contiguous block to a sane value
2791          */
2792         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) &&
2793             (max_blocks > EXT4_MAX_TRANS_DATA))
2794                 max_blocks = EXT4_MAX_TRANS_DATA;
2795
2796         return ext4_chunk_trans_blocks(inode, max_blocks);
2797 }
2798
2799 static int ext4_da_writepages(struct address_space *mapping,
2800                               struct writeback_control *wbc)
2801 {
2802         pgoff_t index;
2803         int range_whole = 0;
2804         handle_t *handle = NULL;
2805         struct mpage_da_data mpd;
2806         struct inode *inode = mapping->host;
2807         int no_nrwrite_index_update;
2808         int pages_written = 0;
2809         long pages_skipped;
2810         unsigned int max_pages;
2811         int range_cyclic, cycled = 1, io_done = 0;
2812         int needed_blocks, ret = 0;
2813         long desired_nr_to_write, nr_to_writebump = 0;
2814         loff_t range_start = wbc->range_start;
2815         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2816
2817         trace_ext4_da_writepages(inode, wbc);
2818
2819         /*
2820          * No pages to write? This is mainly a kludge to avoid starting
2821          * a transaction for special inodes like journal inode on last iput()
2822          * because that could violate lock ordering on umount
2823          */
2824         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2825                 return 0;
2826
2827         /*
2828          * If the filesystem has aborted, it is read-only, so return
2829          * right away instead of dumping stack traces later on that
2830          * will obscure the real source of the problem.  We test
2831          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2832          * the latter could be true if the filesystem is mounted
2833          * read-only, and in that case, ext4_da_writepages should
2834          * *never* be called, so if that ever happens, we would want
2835          * the stack trace.
2836          */
2837         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2838                 return -EROFS;
2839
2840         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2841                 range_whole = 1;
2842
2843         range_cyclic = wbc->range_cyclic;
2844         if (wbc->range_cyclic) {
2845                 index = mapping->writeback_index;
2846                 if (index)
2847                         cycled = 0;
2848                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2849                 wbc->range_end  = LLONG_MAX;
2850                 wbc->range_cyclic = 0;
2851         } else
2852                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2853
2854         /*
2855          * This works around two forms of stupidity.  The first is in
2856          * the writeback code, which caps the maximum number of pages
2857          * written to be 1024 pages.  This is wrong on multiple
2858          * levels; different architectues have a different page size,
2859          * which changes the maximum amount of data which gets
2860          * written.  Secondly, 4 megabytes is way too small.  XFS
2861          * forces this value to be 16 megabytes by multiplying
2862          * nr_to_write parameter by four, and then relies on its
2863          * allocator to allocate larger extents to make them
2864          * contiguous.  Unfortunately this brings us to the second
2865          * stupidity, which is that ext4's mballoc code only allocates
2866          * at most 2048 blocks.  So we force contiguous writes up to
2867          * the number of dirty blocks in the inode, or
2868          * sbi->max_writeback_mb_bump whichever is smaller.
2869          */
2870         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2871         if (!range_cyclic && range_whole)
2872                 desired_nr_to_write = wbc->nr_to_write * 8;
2873         else
2874                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2875                                                            max_pages);
2876         if (desired_nr_to_write > max_pages)
2877                 desired_nr_to_write = max_pages;
2878
2879         if (wbc->nr_to_write < desired_nr_to_write) {
2880                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2881                 wbc->nr_to_write = desired_nr_to_write;
2882         }
2883
2884         mpd.wbc = wbc;
2885         mpd.inode = mapping->host;
2886
2887         /*
2888          * we don't want write_cache_pages to update
2889          * nr_to_write and writeback_index
2890          */
2891         no_nrwrite_index_update = wbc->no_nrwrite_index_update;
2892         wbc->no_nrwrite_index_update = 1;
2893         pages_skipped = wbc->pages_skipped;
2894
2895 retry:
2896         while (!ret && wbc->nr_to_write > 0) {
2897
2898                 /*
2899                  * we  insert one extent at a time. So we need
2900                  * credit needed for single extent allocation.
2901                  * journalled mode is currently not supported
2902                  * by delalloc
2903                  */
2904                 BUG_ON(ext4_should_journal_data(inode));
2905                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2906
2907                 /* start a new transaction*/
2908                 handle = ext4_journal_start(inode, needed_blocks);
2909                 if (IS_ERR(handle)) {
2910                         ret = PTR_ERR(handle);
2911                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2912                                "%ld pages, ino %lu; err %d\n", __func__,
2913                                 wbc->nr_to_write, inode->i_ino, ret);
2914                         goto out_writepages;
2915                 }
2916
2917                 /*
2918                  * Now call __mpage_da_writepage to find the next
2919                  * contiguous region of logical blocks that need
2920                  * blocks to be allocated by ext4.  We don't actually
2921                  * submit the blocks for I/O here, even though
2922                  * write_cache_pages thinks it will, and will set the
2923                  * pages as clean for write before calling
2924                  * __mpage_da_writepage().
2925                  */
2926                 mpd.b_size = 0;
2927                 mpd.b_state = 0;
2928                 mpd.b_blocknr = 0;
2929                 mpd.first_page = 0;
2930                 mpd.next_page = 0;
2931                 mpd.io_done = 0;
2932                 mpd.pages_written = 0;
2933                 mpd.retval = 0;
2934                 ret = write_cache_pages(mapping, wbc, __mpage_da_writepage,
2935                                         &mpd);
2936                 /*
2937                  * If we have a contigous extent of pages and we
2938                  * haven't done the I/O yet, map the blocks and submit
2939                  * them for I/O.
2940                  */
2941                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2942                         if (mpage_da_map_blocks(&mpd) == 0)
2943                                 mpage_da_submit_io(&mpd);
2944                         mpd.io_done = 1;
2945                         ret = MPAGE_DA_EXTENT_TAIL;
2946                 }
2947                 trace_ext4_da_write_pages(inode, &mpd);
2948                 wbc->nr_to_write -= mpd.pages_written;
2949
2950                 ext4_journal_stop(handle);
2951
2952                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2953                         /* commit the transaction which would
2954                          * free blocks released in the transaction
2955                          * and try again
2956                          */
2957                         jbd2_journal_force_commit_nested(sbi->s_journal);
2958                         wbc->pages_skipped = pages_skipped;
2959                         ret = 0;
2960                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2961                         /*
2962                          * got one extent now try with
2963                          * rest of the pages
2964                          */
2965                         pages_written += mpd.pages_written;
2966                         wbc->pages_skipped = pages_skipped;
2967                         ret = 0;
2968                         io_done = 1;
2969                 } else if (wbc->nr_to_write)
2970                         /*
2971                          * There is no more writeout needed
2972                          * or we requested for a noblocking writeout
2973                          * and we found the device congested
2974                          */
2975                         break;
2976         }
2977         if (!io_done && !cycled) {
2978                 cycled = 1;
2979                 index = 0;
2980                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2981                 wbc->range_end  = mapping->writeback_index - 1;
2982                 goto retry;
2983         }
2984         if (pages_skipped != wbc->pages_skipped)
2985                 ext4_msg(inode->i_sb, KERN_CRIT,
2986                          "This should not happen leaving %s "
2987                          "with nr_to_write = %ld ret = %d\n",
2988                          __func__, wbc->nr_to_write, ret);
2989
2990         /* Update index */
2991         index += pages_written;
2992         wbc->range_cyclic = range_cyclic;
2993         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2994                 /*
2995                  * set the writeback_index so that range_cyclic
2996                  * mode will write it back later
2997                  */
2998                 mapping->writeback_index = index;
2999
3000 out_writepages:
3001         if (!no_nrwrite_index_update)
3002                 wbc->no_nrwrite_index_update = 0;
3003         if (wbc->nr_to_write > nr_to_writebump)
3004                 wbc->nr_to_write -= nr_to_writebump;
3005         wbc->range_start = range_start;
3006         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
3007         return ret;
3008 }
3009
3010 #define FALL_BACK_TO_NONDELALLOC 1
3011 static int ext4_nonda_switch(struct super_block *sb)
3012 {
3013         s64 free_blocks, dirty_blocks;
3014         struct ext4_sb_info *sbi = EXT4_SB(sb);
3015
3016         /*
3017          * switch to non delalloc mode if we are running low
3018          * on free block. The free block accounting via percpu
3019          * counters can get slightly wrong with percpu_counter_batch getting
3020          * accumulated on each CPU without updating global counters
3021          * Delalloc need an accurate free block accounting. So switch
3022          * to non delalloc when we are near to error range.
3023          */
3024         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
3025         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
3026         if (2 * free_blocks < 3 * dirty_blocks ||
3027                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
3028                 /*
3029                  * free block count is less that 150% of dirty blocks
3030                  * or free blocks is less that watermark
3031                  */
3032                 return 1;
3033         }
3034         return 0;
3035 }
3036
3037 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3038                                loff_t pos, unsigned len, unsigned flags,
3039                                struct page **pagep, void **fsdata)
3040 {
3041         int ret, retries = 0;
3042         struct page *page;
3043         pgoff_t index;
3044         unsigned from, to;
3045         struct inode *inode = mapping->host;
3046         handle_t *handle;
3047
3048         index = pos >> PAGE_CACHE_SHIFT;
3049         from = pos & (PAGE_CACHE_SIZE - 1);
3050         to = from + len;
3051
3052         if (ext4_nonda_switch(inode->i_sb)) {
3053                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3054                 return ext4_write_begin(file, mapping, pos,
3055                                         len, flags, pagep, fsdata);
3056         }
3057         *fsdata = (void *)0;
3058         trace_ext4_da_write_begin(inode, pos, len, flags);
3059 retry:
3060         /*
3061          * With delayed allocation, we don't log the i_disksize update
3062          * if there is delayed block allocation. But we still need
3063          * to journalling the i_disksize update if writes to the end
3064          * of file which has an already mapped buffer.
3065          */
3066         handle = ext4_journal_start(inode, 1);
3067         if (IS_ERR(handle)) {
3068                 ret = PTR_ERR(handle);
3069                 goto out;
3070         }
3071         /* We cannot recurse into the filesystem as the transaction is already
3072          * started */
3073         flags |= AOP_FLAG_NOFS;
3074
3075         page = grab_cache_page_write_begin(mapping, index, flags);
3076         if (!page) {
3077                 ext4_journal_stop(handle);
3078                 ret = -ENOMEM;
3079                 goto out;
3080         }
3081         *pagep = page;
3082
3083         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
3084                                 ext4_da_get_block_prep);
3085         if (ret < 0) {
3086                 unlock_page(page);
3087                 ext4_journal_stop(handle);
3088                 page_cache_release(page);
3089                 /*
3090                  * block_write_begin may have instantiated a few blocks
3091                  * outside i_size.  Trim these off again. Don't need
3092                  * i_size_read because we hold i_mutex.
3093                  */
3094                 if (pos + len > inode->i_size)
3095                         ext4_truncate(inode);
3096         }
3097
3098         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3099                 goto retry;
3100 out:
3101         return ret;
3102 }
3103
3104 /*
3105  * Check if we should update i_disksize
3106  * when write to the end of file but not require block allocation
3107  */
3108 static int ext4_da_should_update_i_disksize(struct page *page,
3109                                             unsigned long offset)
3110 {
3111         struct buffer_head *bh;
3112         struct inode *inode = page->mapping->host;
3113         unsigned int idx;
3114         int i;
3115
3116         bh = page_buffers(page);
3117         idx = offset >> inode->i_blkbits;
3118
3119         for (i = 0; i < idx; i++)
3120                 bh = bh->b_this_page;
3121
3122         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3123                 return 0;
3124         return 1;
3125 }
3126
3127 static int ext4_da_write_end(struct file *file,
3128                              struct address_space *mapping,
3129                              loff_t pos, unsigned len, unsigned copied,
3130                              struct page *page, void *fsdata)
3131 {
3132         struct inode *inode = mapping->host;
3133         int ret = 0, ret2;
3134         handle_t *handle = ext4_journal_current_handle();
3135         loff_t new_i_size;
3136         unsigned long start, end;
3137         int write_mode = (int)(unsigned long)fsdata;
3138
3139         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
3140                 if (ext4_should_order_data(inode)) {
3141                         return ext4_ordered_write_end(file, mapping, pos,
3142                                         len, copied, page, fsdata);
3143                 } else if (ext4_should_writeback_data(inode)) {
3144                         return ext4_writeback_write_end(file, mapping, pos,
3145                                         len, copied, page, fsdata);
3146                 } else {
3147                         BUG();
3148                 }
3149         }
3150
3151         trace_ext4_da_write_end(inode, pos, len, copied);
3152         start = pos & (PAGE_CACHE_SIZE - 1);
3153         end = start + copied - 1;
3154
3155         /*
3156          * generic_write_end() will run mark_inode_dirty() if i_size
3157          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3158          * into that.
3159          */
3160
3161         new_i_size = pos + copied;
3162         if (new_i_size > EXT4_I(inode)->i_disksize) {
3163                 if (ext4_da_should_update_i_disksize(page, end)) {
3164                         down_write(&EXT4_I(inode)->i_data_sem);
3165                         if (new_i_size > EXT4_I(inode)->i_disksize) {
3166                                 /*
3167                                  * Updating i_disksize when extending file
3168                                  * without needing block allocation
3169                                  */
3170                                 if (ext4_should_order_data(inode))
3171                                         ret = ext4_jbd2_file_inode(handle,
3172                                                                    inode);
3173
3174                                 EXT4_I(inode)->i_disksize = new_i_size;
3175                         }
3176                         up_write(&EXT4_I(inode)->i_data_sem);
3177                         /* We need to mark inode dirty even if
3178                          * new_i_size is less that inode->i_size
3179                          * bu greater than i_disksize.(hint delalloc)
3180                          */
3181                         ext4_mark_inode_dirty(handle, inode);
3182                 }
3183         }
3184         ret2 = generic_write_end(file, mapping, pos, len, copied,
3185                                                         page, fsdata);
3186         copied = ret2;
3187         if (ret2 < 0)
3188                 ret = ret2;
3189         ret2 = ext4_journal_stop(handle);
3190         if (!ret)
3191                 ret = ret2;
3192
3193         return ret ? ret : copied;
3194 }
3195
3196 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
3197 {
3198         /*
3199          * Drop reserved blocks
3200          */
3201         BUG_ON(!PageLocked(page));
3202         if (!page_has_buffers(page))
3203                 goto out;
3204
3205         ext4_da_page_release_reservation(page, offset);
3206
3207 out:
3208         ext4_invalidatepage(page, offset);
3209
3210         return;
3211 }
3212
3213 /*
3214  * Force all delayed allocation blocks to be allocated for a given inode.
3215  */
3216 int ext4_alloc_da_blocks(struct inode *inode)
3217 {
3218         trace_ext4_alloc_da_blocks(inode);
3219
3220         if (!EXT4_I(inode)->i_reserved_data_blocks &&
3221             !EXT4_I(inode)->i_reserved_meta_blocks)
3222                 return 0;
3223
3224         /*
3225          * We do something simple for now.  The filemap_flush() will
3226          * also start triggering a write of the data blocks, which is
3227          * not strictly speaking necessary (and for users of
3228          * laptop_mode, not even desirable).  However, to do otherwise
3229          * would require replicating code paths in:
3230          *
3231          * ext4_da_writepages() ->
3232          *    write_cache_pages() ---> (via passed in callback function)
3233          *        __mpage_da_writepage() -->
3234          *           mpage_add_bh_to_extent()
3235          *           mpage_da_map_blocks()
3236          *
3237          * The problem is that write_cache_pages(), located in
3238          * mm/page-writeback.c, marks pages clean in preparation for
3239          * doing I/O, which is not desirable if we're not planning on
3240          * doing I/O at all.
3241          *
3242          * We could call write_cache_pages(), and then redirty all of
3243          * the pages by calling redirty_page_for_writeback() but that
3244          * would be ugly in the extreme.  So instead we would need to
3245          * replicate parts of the code in the above functions,
3246          * simplifying them becuase we wouldn't actually intend to
3247          * write out the pages, but rather only collect contiguous
3248          * logical block extents, call the multi-block allocator, and
3249          * then update the buffer heads with the block allocations.
3250          *
3251          * For now, though, we'll cheat by calling filemap_flush(),
3252          * which will map the blocks, and start the I/O, but not
3253          * actually wait for the I/O to complete.
3254          */
3255         return filemap_flush(inode->i_mapping);
3256 }
3257
3258 /*
3259  * bmap() is special.  It gets used by applications such as lilo and by
3260  * the swapper to find the on-disk block of a specific piece of data.
3261  *
3262  * Naturally, this is dangerous if the block concerned is still in the
3263  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3264  * filesystem and enables swap, then they may get a nasty shock when the
3265  * data getting swapped to that swapfile suddenly gets overwritten by
3266  * the original zero's written out previously to the journal and
3267  * awaiting writeback in the kernel's buffer cache.
3268  *
3269  * So, if we see any bmap calls here on a modified, data-journaled file,
3270  * take extra steps to flush any blocks which might be in the cache.
3271  */
3272 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3273 {
3274         struct inode *inode = mapping->host;
3275         journal_t *journal;
3276         int err;
3277
3278         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3279                         test_opt(inode->i_sb, DELALLOC)) {
3280                 /*
3281                  * With delalloc we want to sync the file
3282                  * so that we can make sure we allocate
3283                  * blocks for file
3284                  */
3285                 filemap_write_and_wait(mapping);
3286         }
3287
3288         if (EXT4_JOURNAL(inode) && EXT4_I(inode)->i_state & EXT4_STATE_JDATA) {
3289                 /*
3290                  * This is a REALLY heavyweight approach, but the use of
3291                  * bmap on dirty files is expected to be extremely rare:
3292                  * only if we run lilo or swapon on a freshly made file
3293                  * do we expect this to happen.
3294                  *
3295                  * (bmap requires CAP_SYS_RAWIO so this does not
3296                  * represent an unprivileged user DOS attack --- we'd be
3297                  * in trouble if mortal users could trigger this path at
3298                  * will.)
3299                  *
3300                  * NB. EXT4_STATE_JDATA is not set on files other than
3301                  * regular files.  If somebody wants to bmap a directory
3302                  * or symlink and gets confused because the buffer
3303                  * hasn't yet been flushed to disk, they deserve
3304                  * everything they get.
3305                  */
3306
3307                 EXT4_I(inode)->i_state &= ~EXT4_STATE_JDATA;
3308                 journal = EXT4_JOURNAL(inode);
3309                 jbd2_journal_lock_updates(journal);
3310                 err = jbd2_journal_flush(journal);
3311                 jbd2_journal_unlock_updates(journal);
3312
3313                 if (err)
3314                         return 0;
3315         }
3316
3317         return generic_block_bmap(mapping, block, ext4_get_block);
3318 }
3319
3320 static int ext4_readpage(struct file *file, struct page *page)
3321 {
3322         return mpage_readpage(page, ext4_get_block);
3323 }
3324
3325 static int
3326 ext4_readpages(struct file *file, struct address_space *mapping,
3327                 struct list_head *pages, unsigned nr_pages)
3328 {
3329         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3330 }
3331
3332 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3333 {
3334         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3335
3336         /*
3337          * If it's a full truncate we just forget about the pending dirtying
3338          */
3339         if (offset == 0)
3340                 ClearPageChecked(page);
3341
3342         if (journal)
3343                 jbd2_journal_invalidatepage(journal, page, offset);
3344         else
3345                 block_invalidatepage(page, offset);
3346 }
3347
3348 static int ext4_releasepage(struct page *page, gfp_t wait)
3349 {
3350         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3351
3352         WARN_ON(PageChecked(page));
3353         if (!page_has_buffers(page))
3354                 return 0;
3355         if (journal)
3356                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3357         else
3358                 return try_to_free_buffers(page);
3359 }
3360
3361 /*
3362  * O_DIRECT for ext3 (or indirect map) based files
3363  *
3364  * If the O_DIRECT write will extend the file then add this inode to the
3365  * orphan list.  So recovery will truncate it back to the original size
3366  * if the machine crashes during the write.
3367  *
3368  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3369  * crashes then stale disk data _may_ be exposed inside the file. But current
3370  * VFS code falls back into buffered path in that case so we are safe.
3371  */
3372 static ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
3373                               const struct iovec *iov, loff_t offset,
3374                               unsigned long nr_segs)
3375 {
3376         struct file *file = iocb->ki_filp;
3377         struct inode *inode = file->f_mapping->host;
3378         struct ext4_inode_info *ei = EXT4_I(inode);
3379         handle_t *handle;
3380         ssize_t ret;
3381         int orphan = 0;
3382         size_t count = iov_length(iov, nr_segs);
3383         int retries = 0;
3384
3385         if (rw == WRITE) {
3386                 loff_t final_size = offset + count;
3387
3388                 if (final_size > inode->i_size) {
3389                         /* Credits for sb + inode write */
3390                         handle = ext4_journal_start(inode, 2);
3391                         if (IS_ERR(handle)) {
3392                                 ret = PTR_ERR(handle);
3393                                 goto out;
3394                         }
3395                         ret = ext4_orphan_add(handle, inode);
3396                         if (ret) {
3397                                 ext4_journal_stop(handle);
3398                                 goto out;
3399                         }
3400                         orphan = 1;
3401                         ei->i_disksize = inode->i_size;
3402                         ext4_journal_stop(handle);
3403                 }
3404         }
3405
3406 retry:
3407         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
3408                                  offset, nr_segs,
3409                                  ext4_get_block, NULL);
3410         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3411                 goto retry;
3412
3413         if (orphan) {
3414                 int err;
3415
3416                 /* Credits for sb + inode write */
3417                 handle = ext4_journal_start(inode, 2);
3418                 if (IS_ERR(handle)) {
3419                         /* This is really bad luck. We've written the data
3420                          * but cannot extend i_size. Bail out and pretend
3421                          * the write failed... */
3422                         ret = PTR_ERR(handle);
3423                         goto out;
3424                 }
3425                 if (inode->i_nlink)
3426                         ext4_orphan_del(handle, inode);
3427                 if (ret > 0) {
3428                         loff_t end = offset + ret;
3429                         if (end > inode->i_size) {
3430                                 ei->i_disksize = end;
3431                                 i_size_write(inode, end);
3432                                 /*
3433                                  * We're going to return a positive `ret'
3434                                  * here due to non-zero-length I/O, so there's
3435                                  * no way of reporting error returns from
3436                                  * ext4_mark_inode_dirty() to userspace.  So
3437                                  * ignore it.
3438                                  */
3439                                 ext4_mark_inode_dirty(handle, inode);
3440                         }
3441                 }
3442                 err = ext4_journal_stop(handle);
3443                 if (ret == 0)
3444                         ret = err;
3445         }
3446 out:
3447         return ret;
3448 }
3449
3450 static int ext4_get_block_dio_write(struct inode *inode, sector_t iblock,
3451                    struct buffer_head *bh_result, int create)
3452 {
3453         handle_t *handle = NULL;
3454         int ret = 0;
3455         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
3456         int dio_credits;
3457
3458         ext4_debug("ext4_get_block_dio_write: inode %lu, create flag %d\n",
3459                    inode->i_ino, create);
3460         /*
3461          * DIO VFS code passes create = 0 flag for write to
3462          * the middle of file. It does this to avoid block
3463          * allocation for holes, to prevent expose stale data
3464          * out when there is parallel buffered read (which does
3465          * not hold the i_mutex lock) while direct IO write has
3466          * not completed. DIO request on holes finally falls back
3467          * to buffered IO for this reason.
3468          *
3469          * For ext4 extent based file, since we support fallocate,
3470          * new allocated extent as uninitialized, for holes, we
3471          * could fallocate blocks for holes, thus parallel
3472          * buffered IO read will zero out the page when read on
3473          * a hole while parallel DIO write to the hole has not completed.
3474          *
3475          * when we come here, we know it's a direct IO write to
3476          * to the middle of file (<i_size)
3477          * so it's safe to override the create flag from VFS.
3478          */
3479         create = EXT4_GET_BLOCKS_DIO_CREATE_EXT;
3480
3481         if (max_blocks > DIO_MAX_BLOCKS)
3482                 max_blocks = DIO_MAX_BLOCKS;
3483         dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
3484         handle = ext4_journal_start(inode, dio_credits);
3485         if (IS_ERR(handle)) {
3486                 ret = PTR_ERR(handle);
3487                 goto out;
3488         }
3489         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
3490                               create);
3491         if (ret > 0) {
3492                 bh_result->b_size = (ret << inode->i_blkbits);
3493                 ret = 0;
3494         }
3495         ext4_journal_stop(handle);
3496 out:
3497         return ret;
3498 }
3499
3500 static void ext4_free_io_end(ext4_io_end_t *io)
3501 {
3502         BUG_ON(!io);
3503         iput(io->inode);
3504         kfree(io);
3505 }
3506 static void dump_aio_dio_list(struct inode * inode)
3507 {
3508 #ifdef  EXT4_DEBUG
3509         struct list_head *cur, *before, *after;
3510         ext4_io_end_t *io, *io0, *io1;
3511
3512         if (list_empty(&EXT4_I(inode)->i_aio_dio_complete_list)){
3513                 ext4_debug("inode %lu aio dio list is empty\n", inode->i_ino);
3514                 return;
3515         }
3516
3517         ext4_debug("Dump inode %lu aio_dio_completed_IO list \n", inode->i_ino);
3518         list_for_each_entry(io, &EXT4_I(inode)->i_aio_dio_complete_list, list){
3519                 cur = &io->list;
3520                 before = cur->prev;
3521                 io0 = container_of(before, ext4_io_end_t, list);
3522                 after = cur->next;
3523                 io1 = container_of(after, ext4_io_end_t, list);
3524
3525                 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
3526                             io, inode->i_ino, io0, io1);
3527         }
3528 #endif
3529 }
3530
3531 /*
3532  * check a range of space and convert unwritten extents to written.
3533  */
3534 static int ext4_end_aio_dio_nolock(ext4_io_end_t *io)
3535 {
3536         struct inode *inode = io->inode;
3537         loff_t offset = io->offset;
3538         size_t size = io->size;
3539         int ret = 0;
3540
3541         ext4_debug("end_aio_dio_onlock: io 0x%p from inode %lu,list->next 0x%p,"
3542                    "list->prev 0x%p\n",
3543                    io, inode->i_ino, io->list.next, io->list.prev);
3544
3545         if (list_empty(&io->list))
3546                 return ret;
3547
3548         if (io->flag != DIO_AIO_UNWRITTEN)
3549                 return ret;
3550
3551         if (offset + size <= i_size_read(inode))
3552                 ret = ext4_convert_unwritten_extents(inode, offset, size);
3553
3554         if (ret < 0) {
3555                 printk(KERN_EMERG "%s: failed to convert unwritten"
3556                         "extents to written extents, error is %d"
3557                         " io is still on inode %lu aio dio list\n",
3558                        __func__, ret, inode->i_ino);
3559                 return ret;
3560         }
3561
3562         /* clear the DIO AIO unwritten flag */
3563         io->flag = 0;
3564         return ret;
3565 }
3566 /*
3567  * work on completed aio dio IO, to convert unwritten extents to extents
3568  */
3569 static void ext4_end_aio_dio_work(struct work_struct *work)
3570 {
3571         ext4_io_end_t *io  = container_of(work, ext4_io_end_t, work);
3572         struct inode *inode = io->inode;
3573         int ret = 0;
3574
3575         mutex_lock(&inode->i_mutex);
3576         ret = ext4_end_aio_dio_nolock(io);
3577         if (ret >= 0) {
3578                 if (!list_empty(&io->list))
3579                         list_del_init(&io->list);
3580                 ext4_free_io_end(io);
3581         }
3582         mutex_unlock(&inode->i_mutex);
3583 }
3584 /*
3585  * This function is called from ext4_sync_file().
3586  *
3587  * When AIO DIO IO is completed, the work to convert unwritten
3588  * extents to written is queued on workqueue but may not get immediately
3589  * scheduled. When fsync is called, we need to ensure the
3590  * conversion is complete before fsync returns.
3591  * The inode keeps track of a list of completed AIO from DIO path
3592  * that might needs to do the conversion. This function walks through
3593  * the list and convert the related unwritten extents to written.
3594  */
3595 int flush_aio_dio_completed_IO(struct inode *inode)
3596 {
3597         ext4_io_end_t *io;
3598         int ret = 0;
3599         int ret2 = 0;
3600
3601         if (list_empty(&EXT4_I(inode)->i_aio_dio_complete_list))
3602                 return ret;
3603
3604         dump_aio_dio_list(inode);
3605         while (!list_empty(&EXT4_I(inode)->i_aio_dio_complete_list)){
3606                 io = list_entry(EXT4_I(inode)->i_aio_dio_complete_list.next,
3607                                 ext4_io_end_t, list);
3608                 /*
3609                  * Calling ext4_end_aio_dio_nolock() to convert completed
3610                  * IO to written.
3611                  *
3612                  * When ext4_sync_file() is called, run_queue() may already
3613                  * about to flush the work corresponding to this io structure.
3614                  * It will be upset if it founds the io structure related
3615                  * to the work-to-be schedule is freed.
3616                  *
3617                  * Thus we need to keep the io structure still valid here after
3618                  * convertion finished. The io structure has a flag to
3619                  * avoid double converting from both fsync and background work
3620                  * queue work.
3621                  */
3622                 ret = ext4_end_aio_dio_nolock(io);
3623                 if (ret < 0)
3624                         ret2 = ret;
3625                 else
3626                         list_del_init(&io->list);
3627         }
3628         return (ret2 < 0) ? ret2 : 0;
3629 }
3630
3631 static ext4_io_end_t *ext4_init_io_end (struct inode *inode)
3632 {
3633         ext4_io_end_t *io = NULL;
3634
3635         io = kmalloc(sizeof(*io), GFP_NOFS);
3636
3637         if (io) {
3638                 igrab(inode);
3639                 io->inode = inode;
3640                 io->flag = 0;
3641                 io->offset = 0;
3642                 io->size = 0;
3643                 io->error = 0;
3644                 INIT_WORK(&io->work, ext4_end_aio_dio_work);
3645                 INIT_LIST_HEAD(&io->list);
3646         }
3647
3648         return io;
3649 }
3650
3651 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3652                             ssize_t size, void *private)
3653 {
3654         ext4_io_end_t *io_end = iocb->private;
3655         struct workqueue_struct *wq;
3656
3657         /* if not async direct IO or dio with 0 bytes write, just return */
3658         if (!io_end || !size)
3659                 return;
3660
3661         ext_debug("ext4_end_io_dio(): io_end 0x%p"
3662                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
3663                   iocb->private, io_end->inode->i_ino, iocb, offset,
3664                   size);
3665
3666         /* if not aio dio with unwritten extents, just free io and return */
3667         if (io_end->flag != DIO_AIO_UNWRITTEN){
3668                 ext4_free_io_end(io_end);
3669                 iocb->private = NULL;
3670                 return;
3671         }
3672
3673         io_end->offset = offset;
3674         io_end->size = size;
3675         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
3676
3677         /* queue the work to convert unwritten extents to written */
3678         queue_work(wq, &io_end->work);
3679
3680         /* Add the io_end to per-inode completed aio dio list*/
3681         list_add_tail(&io_end->list,
3682                  &EXT4_I(io_end->inode)->i_aio_dio_complete_list);
3683         iocb->private = NULL;
3684 }
3685 /*
3686  * For ext4 extent files, ext4 will do direct-io write to holes,
3687  * preallocated extents, and those write extend the file, no need to
3688  * fall back to buffered IO.
3689  *
3690  * For holes, we fallocate those blocks, mark them as unintialized
3691  * If those blocks were preallocated, we mark sure they are splited, but
3692  * still keep the range to write as unintialized.
3693  *
3694  * The unwrritten extents will be converted to written when DIO is completed.
3695  * For async direct IO, since the IO may still pending when return, we
3696  * set up an end_io call back function, which will do the convertion
3697  * when async direct IO completed.
3698  *
3699  * If the O_DIRECT write will extend the file then add this inode to the
3700  * orphan list.  So recovery will truncate it back to the original size
3701  * if the machine crashes during the write.
3702  *
3703  */
3704 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
3705                               const struct iovec *iov, loff_t offset,
3706                               unsigned long nr_segs)
3707 {
3708         struct file *file = iocb->ki_filp;
3709         struct inode *inode = file->f_mapping->host;
3710         ssize_t ret;
3711         size_t count = iov_length(iov, nr_segs);
3712
3713         loff_t final_size = offset + count;
3714         if (rw == WRITE && final_size <= inode->i_size) {
3715                 /*
3716                  * We could direct write to holes and fallocate.
3717                  *
3718                  * Allocated blocks to fill the hole are marked as uninitialized
3719                  * to prevent paralel buffered read to expose the stale data
3720                  * before DIO complete the data IO.
3721                  *
3722                  * As to previously fallocated extents, ext4 get_block
3723                  * will just simply mark the buffer mapped but still
3724                  * keep the extents uninitialized.
3725                  *
3726                  * for non AIO case, we will convert those unwritten extents
3727                  * to written after return back from blockdev_direct_IO.
3728                  *
3729                  * for async DIO, the conversion needs to be defered when
3730                  * the IO is completed. The ext4 end_io callback function
3731                  * will be called to take care of the conversion work.
3732                  * Here for async case, we allocate an io_end structure to
3733                  * hook to the iocb.
3734                  */
3735                 iocb->private = NULL;
3736                 EXT4_I(inode)->cur_aio_dio = NULL;
3737                 if (!is_sync_kiocb(iocb)) {
3738                         iocb->private = ext4_init_io_end(inode);
3739                         if (!iocb->private)
3740                                 return -ENOMEM;
3741                         /*
3742                          * we save the io structure for current async
3743                          * direct IO, so that later ext4_get_blocks()
3744                          * could flag the io structure whether there
3745                          * is a unwritten extents needs to be converted
3746                          * when IO is completed.
3747                          */
3748                         EXT4_I(inode)->cur_aio_dio = iocb->private;
3749                 }
3750
3751                 ret = blockdev_direct_IO(rw, iocb, inode,
3752                                          inode->i_sb->s_bdev, iov,
3753                                          offset, nr_segs,
3754                                          ext4_get_block_dio_write,
3755                                          ext4_end_io_dio);
3756                 if (iocb->private)
3757                         EXT4_I(inode)->cur_aio_dio = NULL;
3758                 /*
3759                  * The io_end structure takes a reference to the inode,
3760                  * that structure needs to be destroyed and the
3761                  * reference to the inode need to be dropped, when IO is
3762                  * complete, even with 0 byte write, or failed.
3763                  *
3764                  * In the successful AIO DIO case, the io_end structure will be
3765                  * desctroyed and the reference to the inode will be dropped
3766                  * after the end_io call back function is called.
3767                  *
3768                  * In the case there is 0 byte write, or error case, since
3769                  * VFS direct IO won't invoke the end_io call back function,
3770                  * we need to free the end_io structure here.
3771                  */
3772                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
3773                         ext4_free_io_end(iocb->private);
3774                         iocb->private = NULL;
3775                 } else if (ret > 0 && (EXT4_I(inode)->i_state &
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 convertion right here
3781                          */
3782                         err = ext4_convert_unwritten_extents(inode,
3783                                                              offset, ret);
3784                         if (err < 0)
3785                                 ret = err;
3786                         EXT4_I(inode)->i_state &= ~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
3802         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
3803                 return ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
3804
3805         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3806 }
3807
3808 /*
3809  * Pages can be marked dirty completely asynchronously from ext4's journalling
3810  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3811  * much here because ->set_page_dirty is called under VFS locks.  The page is
3812  * not necessarily locked.
3813  *
3814  * We cannot just dirty the page and leave attached buffers clean, because the
3815  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3816  * or jbddirty because all the journalling code will explode.
3817  *
3818  * So what we do is to mark the page "pending dirty" and next time writepage
3819  * is called, propagate that into the buffers appropriately.
3820  */
3821 static int ext4_journalled_set_page_dirty(struct page *page)
3822 {
3823         SetPageChecked(page);
3824         return __set_page_dirty_nobuffers(page);
3825 }
3826
3827 static const struct address_space_operations ext4_ordered_aops = {
3828         .readpage               = ext4_readpage,
3829         .readpages              = ext4_readpages,
3830         .writepage              = ext4_writepage,
3831         .sync_page              = block_sync_page,
3832         .write_begin            = ext4_write_begin,
3833         .write_end              = ext4_ordered_write_end,
3834         .bmap                   = ext4_bmap,
3835         .invalidatepage         = ext4_invalidatepage,
3836         .releasepage            = ext4_releasepage,
3837         .direct_IO              = ext4_direct_IO,
3838         .migratepage            = buffer_migrate_page,
3839         .is_partially_uptodate  = block_is_partially_uptodate,
3840         .error_remove_page      = generic_error_remove_page,
3841 };
3842
3843 static const struct address_space_operations ext4_writeback_aops = {
3844         .readpage               = ext4_readpage,
3845         .readpages              = ext4_readpages,
3846         .writepage              = ext4_writepage,
3847         .sync_page              = block_sync_page,
3848         .write_begin            = ext4_write_begin,
3849         .write_end              = ext4_writeback_write_end,
3850         .bmap                   = ext4_bmap,
3851         .invalidatepage         = ext4_invalidatepage,
3852         .releasepage            = ext4_releasepage,
3853         .direct_IO              = ext4_direct_IO,
3854         .migratepage            = buffer_migrate_page,
3855         .is_partially_uptodate  = block_is_partially_uptodate,
3856         .error_remove_page      = generic_error_remove_page,
3857 };
3858
3859 static const struct address_space_operations ext4_journalled_aops = {
3860         .readpage               = ext4_readpage,
3861         .readpages              = ext4_readpages,
3862         .writepage              = ext4_writepage,
3863         .sync_page              = block_sync_page,
3864         .write_begin            = ext4_write_begin,
3865         .write_end              = ext4_journalled_write_end,
3866         .set_page_dirty         = ext4_journalled_set_page_dirty,
3867         .bmap                   = ext4_bmap,
3868         .invalidatepage         = ext4_invalidatepage,
3869         .releasepage            = ext4_releasepage,
3870         .is_partially_uptodate  = block_is_partially_uptodate,
3871         .error_remove_page      = generic_error_remove_page,
3872 };
3873
3874 static const struct address_space_operations ext4_da_aops = {
3875         .readpage               = ext4_readpage,
3876         .readpages              = ext4_readpages,
3877         .writepage              = ext4_writepage,
3878         .writepages             = ext4_da_writepages,
3879         .sync_page              = block_sync_page,
3880         .write_begin            = ext4_da_write_begin,
3881         .write_end              = ext4_da_write_end,
3882         .bmap                   = ext4_bmap,
3883         .invalidatepage         = ext4_da_invalidatepage,
3884         .releasepage            = ext4_releasepage,
3885         .direct_IO              = ext4_direct_IO,
3886         .migratepage            = buffer_migrate_page,
3887         .is_partially_uptodate  = block_is_partially_uptodate,
3888         .error_remove_page      = generic_error_remove_page,
3889 };
3890
3891 void ext4_set_aops(struct inode *inode)
3892 {
3893         if (ext4_should_order_data(inode) &&
3894                 test_opt(inode->i_sb, DELALLOC))
3895                 inode->i_mapping->a_ops = &ext4_da_aops;
3896         else if (ext4_should_order_data(inode))
3897                 inode->i_mapping->a_ops = &ext4_ordered_aops;
3898         else if (ext4_should_writeback_data(inode) &&
3899                  test_opt(inode->i_sb, DELALLOC))
3900                 inode->i_mapping->a_ops = &ext4_da_aops;
3901         else if (ext4_should_writeback_data(inode))
3902                 inode->i_mapping->a_ops = &ext4_writeback_aops;
3903         else
3904                 inode->i_mapping->a_ops = &ext4_journalled_aops;
3905 }
3906
3907 /*
3908  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3909  * up to the end of the block which corresponds to `from'.
3910  * This required during truncate. We need to physically zero the tail end
3911  * of that block so it doesn't yield old data if the file is later grown.
3912  */
3913 int ext4_block_truncate_page(handle_t *handle,
3914                 struct address_space *mapping, loff_t from)
3915 {
3916         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3917         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3918         unsigned blocksize, length, pos;
3919         ext4_lblk_t iblock;
3920         struct inode *inode = mapping->host;
3921         struct buffer_head *bh;
3922         struct page *page;
3923         int err = 0;
3924
3925         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
3926                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
3927         if (!page)
3928                 return -EINVAL;
3929
3930         blocksize = inode->i_sb->s_blocksize;
3931         length = blocksize - (offset & (blocksize - 1));
3932         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3933
3934         /*
3935          * For "nobh" option,  we can only work if we don't need to
3936          * read-in the page - otherwise we create buffers to do the IO.
3937          */
3938         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
3939              ext4_should_writeback_data(inode) && PageUptodate(page)) {
3940                 zero_user(page, offset, length);
3941                 set_page_dirty(page);
3942                 goto unlock;
3943         }
3944
3945         if (!page_has_buffers(page))
3946                 create_empty_buffers(page, blocksize, 0);
3947
3948         /* Find the buffer that contains "offset" */
3949         bh = page_buffers(page);
3950         pos = blocksize;
3951         while (offset >= pos) {
3952                 bh = bh->b_this_page;
3953                 iblock++;
3954                 pos += blocksize;
3955         }
3956
3957         err = 0;
3958         if (buffer_freed(bh)) {
3959                 BUFFER_TRACE(bh, "freed: skip");
3960                 goto unlock;
3961         }
3962
3963         if (!buffer_mapped(bh)) {
3964                 BUFFER_TRACE(bh, "unmapped");
3965                 ext4_get_block(inode, iblock, bh, 0);
3966                 /* unmapped? It's a hole - nothing to do */
3967                 if (!buffer_mapped(bh)) {
3968                         BUFFER_TRACE(bh, "still unmapped");
3969                         goto unlock;
3970                 }
3971         }
3972
3973         /* Ok, it's mapped. Make sure it's up-to-date */
3974         if (PageUptodate(page))
3975                 set_buffer_uptodate(bh);
3976
3977         if (!buffer_uptodate(bh)) {
3978                 err = -EIO;
3979                 ll_rw_block(READ, 1, &bh);
3980                 wait_on_buffer(bh);
3981                 /* Uhhuh. Read error. Complain and punt. */
3982                 if (!buffer_uptodate(bh))
3983                         goto unlock;
3984         }
3985
3986         if (ext4_should_journal_data(inode)) {
3987                 BUFFER_TRACE(bh, "get write access");
3988                 err = ext4_journal_get_write_access(handle, bh);
3989                 if (err)
3990                         goto unlock;
3991         }
3992
3993         zero_user(page, offset, length);
3994
3995         BUFFER_TRACE(bh, "zeroed end of block");
3996
3997         err = 0;
3998         if (ext4_should_journal_data(inode)) {
3999                 err = ext4_handle_dirty_metadata(handle, inode, bh);
4000         } else {
4001                 if (ext4_should_order_data(inode))
4002                         err = ext4_jbd2_file_inode(handle, inode);
4003                 mark_buffer_dirty(bh);
4004         }
4005
4006 unlock:
4007         unlock_page(page);
4008         page_cache_release(page);
4009         return err;
4010 }
4011
4012 /*
4013  * Probably it should be a library function... search for first non-zero word
4014  * or memcmp with zero_page, whatever is better for particular architecture.
4015  * Linus?
4016  */
4017 static inline int all_zeroes(__le32 *p, __le32 *q)
4018 {
4019         while (p < q)
4020                 if (*p++)
4021                         return 0;
4022         return 1;
4023 }
4024
4025 /**
4026  *      ext4_find_shared - find the indirect blocks for partial truncation.
4027  *      @inode:   inode in question
4028  *      @depth:   depth of the affected branch
4029  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
4030  *      @chain:   place to store the pointers to partial indirect blocks
4031  *      @top:     place to the (detached) top of branch
4032  *
4033  *      This is a helper function used by ext4_truncate().
4034  *
4035  *      When we do truncate() we may have to clean the ends of several
4036  *      indirect blocks but leave the blocks themselves alive. Block is
4037  *      partially truncated if some data below the new i_size is refered
4038  *      from it (and it is on the path to the first completely truncated
4039  *      data block, indeed).  We have to free the top of that path along
4040  *      with everything to the right of the path. Since no allocation
4041  *      past the truncation point is possible until ext4_truncate()
4042  *      finishes, we may safely do the latter, but top of branch may
4043  *      require special attention - pageout below the truncation point
4044  *      might try to populate it.
4045  *
4046  *      We atomically detach the top of branch from the tree, store the
4047  *      block number of its root in *@top, pointers to buffer_heads of
4048  *      partially truncated blocks - in @chain[].bh and pointers to
4049  *      their last elements that should not be removed - in
4050  *      @chain[].p. Return value is the pointer to last filled element
4051  *      of @chain.
4052  *
4053  *      The work left to caller to do the actual freeing of subtrees:
4054  *              a) free the subtree starting from *@top
4055  *              b) free the subtrees whose roots are stored in
4056  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
4057  *              c) free the subtrees growing from the inode past the @chain[0].
4058  *                      (no partially truncated stuff there).  */
4059
4060 static Indirect *ext4_find_shared(struct inode *inode, int depth,
4061                                   ext4_lblk_t offsets[4], Indirect chain[4],
4062                                   __le32 *top)
4063 {
4064         Indirect *partial, *p;
4065         int k, err;
4066
4067         *top = 0;
4068         /* Make k index the deepest non-null offest + 1 */
4069         for (k = depth; k > 1 && !offsets[k-1]; k--)
4070                 ;
4071         partial = ext4_get_branch(inode, k, offsets, chain, &err);
4072         /* Writer: pointers */
4073         if (!partial)
4074                 partial = chain + k-1;
4075         /*
4076          * If the branch acquired continuation since we've looked at it -
4077          * fine, it should all survive and (new) top doesn't belong to us.
4078          */
4079         if (!partial->key && *partial->p)
4080                 /* Writer: end */
4081                 goto no_top;
4082         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
4083                 ;
4084         /*
4085          * OK, we've found the last block that must survive. The rest of our
4086          * branch should be detached before unlocking. However, if that rest
4087          * of branch is all ours and does not grow immediately from the inode
4088          * it's easier to cheat and just decrement partial->p.
4089          */
4090         if (p == chain + k - 1 && p > chain) {
4091                 p->p--;
4092         } else {
4093                 *top = *p->p;
4094                 /* Nope, don't do this in ext4.  Must leave the tree intact */
4095 #if 0
4096                 *p->p = 0;
4097 #endif
4098         }
4099         /* Writer: end */
4100
4101         while (partial > p) {
4102                 brelse(partial->bh);
4103                 partial--;
4104         }
4105 no_top:
4106         return partial;
4107 }
4108
4109 /*
4110  * Zero a number of block pointers in either an inode or an indirect block.
4111  * If we restart the transaction we must again get write access to the
4112  * indirect block for further modification.
4113  *
4114  * We release `count' blocks on disk, but (last - first) may be greater
4115  * than `count' because there can be holes in there.
4116  */
4117 static void ext4_clear_blocks(handle_t *handle, struct inode *inode,
4118                               struct buffer_head *bh,
4119                               ext4_fsblk_t block_to_free,
4120                               unsigned long count, __le32 *first,
4121                               __le32 *last)
4122 {
4123         __le32 *p;
4124         int     is_metadata = S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode);
4125
4126         if (try_to_extend_transaction(handle, inode)) {
4127                 if (bh) {
4128                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4129                         ext4_handle_dirty_metadata(handle, inode, bh);
4130                 }
4131                 ext4_mark_inode_dirty(handle, inode);
4132                 ext4_truncate_restart_trans(handle, inode,
4133                                             blocks_for_truncate(inode));
4134                 if (bh) {
4135                         BUFFER_TRACE(bh, "retaking write access");
4136                         ext4_journal_get_write_access(handle, bh);
4137                 }
4138         }
4139
4140         /*
4141          * Any buffers which are on the journal will be in memory. We
4142          * find them on the hash table so jbd2_journal_revoke() will
4143          * run jbd2_journal_forget() on them.  We've already detached
4144          * each block from the file, so bforget() in
4145          * jbd2_journal_forget() should be safe.
4146          *
4147          * AKPM: turn on bforget in jbd2_journal_forget()!!!
4148          */
4149         for (p = first; p < last; p++) {
4150                 u32 nr = le32_to_cpu(*p);
4151                 if (nr) {
4152                         struct buffer_head *tbh;
4153
4154                         *p = 0;
4155                         tbh = sb_find_get_block(inode->i_sb, nr);
4156                         ext4_forget(handle, is_metadata, inode, tbh, nr);
4157                 }
4158         }
4159
4160         ext4_free_blocks(handle, inode, block_to_free, count, is_metadata);
4161 }
4162
4163 /**
4164  * ext4_free_data - free a list of data blocks
4165  * @handle:     handle for this transaction
4166  * @inode:      inode we are dealing with
4167  * @this_bh:    indirect buffer_head which contains *@first and *@last
4168  * @first:      array of block numbers
4169  * @last:       points immediately past the end of array
4170  *
4171  * We are freeing all blocks refered from that array (numbers are stored as
4172  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
4173  *
4174  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
4175  * blocks are contiguous then releasing them at one time will only affect one
4176  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
4177  * actually use a lot of journal space.
4178  *
4179  * @this_bh will be %NULL if @first and @last point into the inode's direct
4180  * block pointers.
4181  */
4182 static void ext4_free_data(handle_t *handle, struct inode *inode,
4183                            struct buffer_head *this_bh,
4184                            __le32 *first, __le32 *last)
4185 {
4186         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
4187         unsigned long count = 0;            /* Number of blocks in the run */
4188         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
4189                                                corresponding to
4190                                                block_to_free */
4191         ext4_fsblk_t nr;                    /* Current block # */
4192         __le32 *p;                          /* Pointer into inode/ind
4193                                                for current block */
4194         int err;
4195
4196         if (this_bh) {                          /* For indirect block */
4197                 BUFFER_TRACE(this_bh, "get_write_access");
4198                 err = ext4_journal_get_write_access(handle, this_bh);
4199                 /* Important: if we can't update the indirect pointers
4200                  * to the blocks, we can't free them. */
4201                 if (err)
4202                         return;
4203         }
4204
4205         for (p = first; p < last; p++) {
4206                 nr = le32_to_cpu(*p);
4207                 if (nr) {
4208                         /* accumulate blocks to free if they're contiguous */
4209                         if (count == 0) {
4210                                 block_to_free = nr;
4211                                 block_to_free_p = p;
4212                                 count = 1;
4213                         } else if (nr == block_to_free + count) {
4214                                 count++;
4215                         } else {
4216                                 ext4_clear_blocks(handle, inode, this_bh,
4217                                                   block_to_free,
4218                                                   count, block_to_free_p, p);
4219                                 block_to_free = nr;
4220                                 block_to_free_p = p;
4221                                 count = 1;
4222                         }
4223                 }
4224         }
4225
4226         if (count > 0)
4227                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
4228                                   count, block_to_free_p, p);
4229
4230         if (this_bh) {
4231                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
4232
4233                 /*
4234                  * The buffer head should have an attached journal head at this
4235                  * point. However, if the data is corrupted and an indirect
4236                  * block pointed to itself, it would have been detached when
4237                  * the block was cleared. Check for this instead of OOPSing.
4238                  */
4239                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
4240                         ext4_handle_dirty_metadata(handle, inode, this_bh);
4241                 else
4242                         ext4_error(inode->i_sb, __func__,
4243                                    "circular indirect block detected, "
4244                                    "inode=%lu, block=%llu",
4245                                    inode->i_ino,
4246                                    (unsigned long long) this_bh->b_blocknr);
4247         }
4248 }
4249
4250 /**
4251  *      ext4_free_branches - free an array of branches
4252  *      @handle: JBD handle for this transaction
4253  *      @inode: inode we are dealing with
4254  *      @parent_bh: the buffer_head which contains *@first and *@last
4255  *      @first: array of block numbers
4256  *      @last:  pointer immediately past the end of array
4257  *      @depth: depth of the branches to free
4258  *
4259  *      We are freeing all blocks refered from these branches (numbers are
4260  *      stored as little-endian 32-bit) and updating @inode->i_blocks
4261  *      appropriately.
4262  */
4263 static void ext4_free_branches(handle_t *handle, struct inode *inode,
4264                                struct buffer_head *parent_bh,
4265                                __le32 *first, __le32 *last, int depth)
4266 {
4267         ext4_fsblk_t nr;
4268         __le32 *p;
4269
4270         if (ext4_handle_is_aborted(handle))
4271                 return;
4272
4273         if (depth--) {
4274                 struct buffer_head *bh;
4275                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4276                 p = last;
4277                 while (--p >= first) {
4278                         nr = le32_to_cpu(*p);
4279                         if (!nr)
4280                                 continue;               /* A hole */
4281
4282                         /* Go read the buffer for the next level down */
4283                         bh = sb_bread(inode->i_sb, nr);
4284
4285                         /*
4286                          * A read failure? Report error and clear slot
4287                          * (should be rare).
4288                          */
4289                         if (!bh) {
4290                                 ext4_error(inode->i_sb, "ext4_free_branches",
4291                                            "Read failure, inode=%lu, block=%llu",
4292                                            inode->i_ino, nr);
4293                                 continue;
4294                         }
4295
4296                         /* This zaps the entire block.  Bottom up. */
4297                         BUFFER_TRACE(bh, "free child branches");
4298                         ext4_free_branches(handle, inode, bh,
4299                                         (__le32 *) bh->b_data,
4300                                         (__le32 *) bh->b_data + addr_per_block,
4301                                         depth);
4302
4303                         /*
4304                          * We've probably journalled the indirect block several
4305                          * times during the truncate.  But it's no longer
4306                          * needed and we now drop it from the transaction via
4307                          * jbd2_journal_revoke().
4308                          *
4309                          * That's easy if it's exclusively part of this
4310                          * transaction.  But if it's part of the committing
4311                          * transaction then jbd2_journal_forget() will simply
4312                          * brelse() it.  That means that if the underlying
4313                          * block is reallocated in ext4_get_block(),
4314                          * unmap_underlying_metadata() will find this block
4315                          * and will try to get rid of it.  damn, damn.
4316                          *
4317                          * If this block has already been committed to the
4318                          * journal, a revoke record will be written.  And
4319                          * revoke records must be emitted *before* clearing
4320                          * this block's bit in the bitmaps.
4321                          */
4322                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
4323
4324                         /*
4325                          * Everything below this this pointer has been
4326                          * released.  Now let this top-of-subtree go.
4327                          *
4328                          * We want the freeing of this indirect block to be
4329                          * atomic in the journal with the updating of the
4330                          * bitmap block which owns it.  So make some room in
4331                          * the journal.
4332                          *
4333                          * We zero the parent pointer *after* freeing its
4334                          * pointee in the bitmaps, so if extend_transaction()
4335                          * for some reason fails to put the bitmap changes and
4336                          * the release into the same transaction, recovery
4337                          * will merely complain about releasing a free block,
4338                          * rather than leaking blocks.
4339                          */
4340                         if (ext4_handle_is_aborted(handle))
4341                                 return;
4342                         if (try_to_extend_transaction(handle, inode)) {
4343                                 ext4_mark_inode_dirty(handle, inode);
4344                                 ext4_truncate_restart_trans(handle, inode,
4345                                             blocks_for_truncate(inode));
4346                         }
4347
4348                         ext4_free_blocks(handle, inode, nr, 1, 1);
4349
4350                         if (parent_bh) {
4351                                 /*
4352                                  * The block which we have just freed is
4353                                  * pointed to by an indirect block: journal it
4354                                  */
4355                                 BUFFER_TRACE(parent_bh, "get_write_access");
4356                                 if (!ext4_journal_get_write_access(handle,
4357                                                                    parent_bh)){
4358                                         *p = 0;
4359                                         BUFFER_TRACE(parent_bh,
4360                                         "call ext4_handle_dirty_metadata");
4361                                         ext4_handle_dirty_metadata(handle,
4362                                                                    inode,
4363                                                                    parent_bh);
4364                                 }
4365                         }
4366                 }
4367         } else {
4368                 /* We have reached the bottom of the tree. */
4369                 BUFFER_TRACE(parent_bh, "free data blocks");
4370                 ext4_free_data(handle, inode, parent_bh, first, last);
4371         }
4372 }
4373
4374 int ext4_can_truncate(struct inode *inode)
4375 {
4376         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4377                 return 0;
4378         if (S_ISREG(inode->i_mode))
4379                 return 1;
4380         if (S_ISDIR(inode->i_mode))
4381                 return 1;
4382         if (S_ISLNK(inode->i_mode))
4383                 return !ext4_inode_is_fast_symlink(inode);
4384         return 0;
4385 }
4386
4387 /*
4388  * ext4_truncate()
4389  *
4390  * We block out ext4_get_block() block instantiations across the entire
4391  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4392  * simultaneously on behalf of the same inode.
4393  *
4394  * As we work through the truncate and commmit bits of it to the journal there
4395  * is one core, guiding principle: the file's tree must always be consistent on
4396  * disk.  We must be able to restart the truncate after a crash.
4397  *
4398  * The file's tree may be transiently inconsistent in memory (although it
4399  * probably isn't), but whenever we close off and commit a journal transaction,
4400  * the contents of (the filesystem + the journal) must be consistent and
4401  * restartable.  It's pretty simple, really: bottom up, right to left (although
4402  * left-to-right works OK too).
4403  *
4404  * Note that at recovery time, journal replay occurs *before* the restart of
4405  * truncate against the orphan inode list.
4406  *
4407  * The committed inode has the new, desired i_size (which is the same as
4408  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4409  * that this inode's truncate did not complete and it will again call
4410  * ext4_truncate() to have another go.  So there will be instantiated blocks
4411  * to the right of the truncation point in a crashed ext4 filesystem.  But
4412  * that's fine - as long as they are linked from the inode, the post-crash
4413  * ext4_truncate() run will find them and release them.
4414  */
4415 void ext4_truncate(struct inode *inode)
4416 {
4417         handle_t *handle;
4418         struct ext4_inode_info *ei = EXT4_I(inode);
4419         __le32 *i_data = ei->i_data;
4420         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4421         struct address_space *mapping = inode->i_mapping;
4422         ext4_lblk_t offsets[4];
4423         Indirect chain[4];
4424         Indirect *partial;
4425         __le32 nr = 0;
4426         int n;
4427         ext4_lblk_t last_block;
4428         unsigned blocksize = inode->i_sb->s_blocksize;
4429
4430         if (!ext4_can_truncate(inode))
4431                 return;
4432
4433         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4434                 ei->i_state |= EXT4_STATE_DA_ALLOC_CLOSE;
4435
4436         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
4437                 ext4_ext_truncate(inode);
4438                 return;
4439         }
4440
4441         handle = start_transaction(inode);
4442         if (IS_ERR(handle))
4443                 return;         /* AKPM: return what? */
4444
4445         last_block = (inode->i_size + blocksize-1)
4446                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4447
4448         if (inode->i_size & (blocksize - 1))
4449                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4450                         goto out_stop;
4451
4452         n = ext4_block_to_path(inode, last_block, offsets, NULL);
4453         if (n == 0)
4454                 goto out_stop;  /* error */
4455
4456         /*
4457          * OK.  This truncate is going to happen.  We add the inode to the
4458          * orphan list, so that if this truncate spans multiple transactions,
4459          * and we crash, we will resume the truncate when the filesystem
4460          * recovers.  It also marks the inode dirty, to catch the new size.
4461          *
4462          * Implication: the file must always be in a sane, consistent
4463          * truncatable state while each transaction commits.
4464          */
4465         if (ext4_orphan_add(handle, inode))
4466                 goto out_stop;
4467
4468         /*
4469          * From here we block out all ext4_get_block() callers who want to
4470          * modify the block allocation tree.
4471          */
4472         down_write(&ei->i_data_sem);
4473
4474         ext4_discard_preallocations(inode);
4475
4476         /*
4477          * The orphan list entry will now protect us from any crash which
4478          * occurs before the truncate completes, so it is now safe to propagate
4479          * the new, shorter inode size (held for now in i_size) into the
4480          * on-disk inode. We do this via i_disksize, which is the value which
4481          * ext4 *really* writes onto the disk inode.
4482          */
4483         ei->i_disksize = inode->i_size;
4484
4485         if (n == 1) {           /* direct blocks */
4486                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4487                                i_data + EXT4_NDIR_BLOCKS);
4488                 goto do_indirects;
4489         }
4490
4491         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4492         /* Kill the top of shared branch (not detached) */
4493         if (nr) {
4494                 if (partial == chain) {
4495                         /* Shared branch grows from the inode */
4496                         ext4_free_branches(handle, inode, NULL,
4497                                            &nr, &nr+1, (chain+n-1) - partial);
4498                         *partial->p = 0;
4499                         /*
4500                          * We mark the inode dirty prior to restart,
4501                          * and prior to stop.  No need for it here.
4502                          */
4503                 } else {
4504                         /* Shared branch grows from an indirect block */
4505                         BUFFER_TRACE(partial->bh, "get_write_access");
4506                         ext4_free_branches(handle, inode, partial->bh,
4507                                         partial->p,
4508                                         partial->p+1, (chain+n-1) - partial);
4509                 }
4510         }
4511         /* Clear the ends of indirect blocks on the shared branch */
4512         while (partial > chain) {
4513                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4514                                    (__le32*)partial->bh->b_data+addr_per_block,
4515                                    (chain+n-1) - partial);
4516                 BUFFER_TRACE(partial->bh, "call brelse");
4517                 brelse(partial->bh);
4518                 partial--;
4519         }
4520 do_indirects:
4521         /* Kill the remaining (whole) subtrees */
4522         switch (offsets[0]) {
4523         default:
4524                 nr = i_data[EXT4_IND_BLOCK];
4525                 if (nr) {
4526                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4527                         i_data[EXT4_IND_BLOCK] = 0;
4528                 }
4529         case EXT4_IND_BLOCK:
4530                 nr = i_data[EXT4_DIND_BLOCK];
4531                 if (nr) {
4532                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4533                         i_data[EXT4_DIND_BLOCK] = 0;
4534                 }
4535         case EXT4_DIND_BLOCK:
4536                 nr = i_data[EXT4_TIND_BLOCK];
4537                 if (nr) {
4538                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4539                         i_data[EXT4_TIND_BLOCK] = 0;
4540                 }
4541         case EXT4_TIND_BLOCK:
4542                 ;
4543         }
4544
4545         up_write(&ei->i_data_sem);
4546         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4547         ext4_mark_inode_dirty(handle, inode);
4548
4549         /*
4550          * In a multi-transaction truncate, we only make the final transaction
4551          * synchronous
4552          */
4553         if (IS_SYNC(inode))
4554                 ext4_handle_sync(handle);
4555 out_stop:
4556         /*
4557          * If this was a simple ftruncate(), and the file will remain alive
4558          * then we need to clear up the orphan record which we created above.
4559          * However, if this was a real unlink then we were called by
4560          * ext4_delete_inode(), and we allow that function to clean up the
4561          * orphan info for us.
4562          */
4563         if (inode->i_nlink)
4564                 ext4_orphan_del(handle, inode);
4565
4566         ext4_journal_stop(handle);
4567 }
4568
4569 /*
4570  * ext4_get_inode_loc returns with an extra refcount against the inode's
4571  * underlying buffer_head on success. If 'in_mem' is true, we have all
4572  * data in memory that is needed to recreate the on-disk version of this
4573  * inode.
4574  */
4575 static int __ext4_get_inode_loc(struct inode *inode,
4576                                 struct ext4_iloc *iloc, int in_mem)
4577 {
4578         struct ext4_group_desc  *gdp;
4579         struct buffer_head      *bh;
4580         struct super_block      *sb = inode->i_sb;
4581         ext4_fsblk_t            block;
4582         int                     inodes_per_block, inode_offset;
4583
4584         iloc->bh = NULL;
4585         if (!ext4_valid_inum(sb, inode->i_ino))
4586                 return -EIO;
4587
4588         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4589         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4590         if (!gdp)
4591                 return -EIO;
4592
4593         /*
4594          * Figure out the offset within the block group inode table
4595          */
4596         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4597         inode_offset = ((inode->i_ino - 1) %
4598                         EXT4_INODES_PER_GROUP(sb));
4599         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4600         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4601
4602         bh = sb_getblk(sb, block);
4603         if (!bh) {
4604                 ext4_error(sb, "ext4_get_inode_loc", "unable to read "
4605                            "inode block - inode=%lu, block=%llu",
4606                            inode->i_ino, block);
4607                 return -EIO;
4608         }
4609         if (!buffer_uptodate(bh)) {
4610                 lock_buffer(bh);
4611
4612                 /*
4613                  * If the buffer has the write error flag, we have failed
4614                  * to write out another inode in the same block.  In this
4615                  * case, we don't have to read the block because we may
4616                  * read the old inode data successfully.
4617                  */
4618                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4619                         set_buffer_uptodate(bh);
4620
4621                 if (buffer_uptodate(bh)) {
4622                         /* someone brought it uptodate while we waited */
4623                         unlock_buffer(bh);
4624                         goto has_buffer;
4625                 }
4626
4627                 /*
4628                  * If we have all information of the inode in memory and this
4629                  * is the only valid inode in the block, we need not read the
4630                  * block.
4631                  */
4632                 if (in_mem) {
4633                         struct buffer_head *bitmap_bh;
4634                         int i, start;
4635
4636                         start = inode_offset & ~(inodes_per_block - 1);
4637
4638                         /* Is the inode bitmap in cache? */
4639                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4640                         if (!bitmap_bh)
4641                                 goto make_io;
4642
4643                         /*
4644                          * If the inode bitmap isn't in cache then the
4645                          * optimisation may end up performing two reads instead
4646                          * of one, so skip it.
4647                          */
4648                         if (!buffer_uptodate(bitmap_bh)) {
4649                                 brelse(bitmap_bh);
4650                                 goto make_io;
4651                         }
4652                         for (i = start; i < start + inodes_per_block; i++) {
4653                                 if (i == inode_offset)
4654                                         continue;
4655                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4656                                         break;
4657                         }
4658                         brelse(bitmap_bh);
4659                         if (i == start + inodes_per_block) {
4660                                 /* all other inodes are free, so skip I/O */
4661                                 memset(bh->b_data, 0, bh->b_size);
4662                                 set_buffer_uptodate(bh);
4663                                 unlock_buffer(bh);
4664                                 goto has_buffer;
4665                         }
4666                 }
4667
4668 make_io:
4669                 /*
4670                  * If we need to do any I/O, try to pre-readahead extra
4671                  * blocks from the inode table.
4672                  */
4673                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4674                         ext4_fsblk_t b, end, table;
4675                         unsigned num;
4676
4677                         table = ext4_inode_table(sb, gdp);
4678                         /* s_inode_readahead_blks is always a power of 2 */
4679                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4680                         if (table > b)
4681                                 b = table;
4682                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4683                         num = EXT4_INODES_PER_GROUP(sb);
4684                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4685                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4686                                 num -= ext4_itable_unused_count(sb, gdp);
4687                         table += num / inodes_per_block;
4688                         if (end > table)
4689                                 end = table;
4690                         while (b <= end)
4691                                 sb_breadahead(sb, b++);
4692                 }
4693
4694                 /*
4695                  * There are other valid inodes in the buffer, this inode
4696                  * has in-inode xattrs, or we don't have this inode in memory.
4697                  * Read the block from disk.
4698                  */
4699                 get_bh(bh);
4700                 bh->b_end_io = end_buffer_read_sync;
4701                 submit_bh(READ_META, bh);
4702                 wait_on_buffer(bh);
4703                 if (!buffer_uptodate(bh)) {
4704                         ext4_error(sb, __func__,
4705                                    "unable to read inode block - inode=%lu, "
4706                                    "block=%llu", inode->i_ino, block);
4707                         brelse(bh);
4708                         return -EIO;
4709                 }
4710         }
4711 has_buffer:
4712         iloc->bh = bh;
4713         return 0;
4714 }
4715
4716 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4717 {
4718         /* We have all inode data except xattrs in memory here. */
4719         return __ext4_get_inode_loc(inode, iloc,
4720                 !(EXT4_I(inode)->i_state & EXT4_STATE_XATTR));
4721 }
4722
4723 void ext4_set_inode_flags(struct inode *inode)
4724 {
4725         unsigned int flags = EXT4_I(inode)->i_flags;
4726
4727         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4728         if (flags & EXT4_SYNC_FL)
4729                 inode->i_flags |= S_SYNC;
4730         if (flags & EXT4_APPEND_FL)
4731                 inode->i_flags |= S_APPEND;
4732         if (flags & EXT4_IMMUTABLE_FL)
4733                 inode->i_flags |= S_IMMUTABLE;
4734         if (flags & EXT4_NOATIME_FL)
4735                 inode->i_flags |= S_NOATIME;
4736         if (flags & EXT4_DIRSYNC_FL)
4737                 inode->i_flags |= S_DIRSYNC;
4738 }
4739
4740 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4741 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4742 {
4743         unsigned int flags = ei->vfs_inode.i_flags;
4744
4745         ei->i_flags &= ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4746                         EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|EXT4_DIRSYNC_FL);
4747         if (flags & S_SYNC)
4748                 ei->i_flags |= EXT4_SYNC_FL;
4749         if (flags & S_APPEND)
4750                 ei->i_flags |= EXT4_APPEND_FL;
4751         if (flags & S_IMMUTABLE)
4752                 ei->i_flags |= EXT4_IMMUTABLE_FL;
4753         if (flags & S_NOATIME)
4754                 ei->i_flags |= EXT4_NOATIME_FL;
4755         if (flags & S_DIRSYNC)
4756                 ei->i_flags |= EXT4_DIRSYNC_FL;
4757 }
4758
4759 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4760                                   struct ext4_inode_info *ei)
4761 {
4762         blkcnt_t i_blocks ;
4763         struct inode *inode = &(ei->vfs_inode);
4764         struct super_block *sb = inode->i_sb;
4765
4766         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4767                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4768                 /* we are using combined 48 bit field */
4769                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4770                                         le32_to_cpu(raw_inode->i_blocks_lo);
4771                 if (ei->i_flags & EXT4_HUGE_FILE_FL) {
4772                         /* i_blocks represent file system block size */
4773                         return i_blocks  << (inode->i_blkbits - 9);
4774                 } else {
4775                         return i_blocks;
4776                 }
4777         } else {
4778                 return le32_to_cpu(raw_inode->i_blocks_lo);
4779         }
4780 }
4781
4782 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4783 {
4784         struct ext4_iloc iloc;
4785         struct ext4_inode *raw_inode;
4786         struct ext4_inode_info *ei;
4787         struct inode *inode;
4788         long ret;
4789         int block;
4790
4791         inode = iget_locked(sb, ino);
4792         if (!inode)
4793                 return ERR_PTR(-ENOMEM);
4794         if (!(inode->i_state & I_NEW))
4795                 return inode;
4796
4797         ei = EXT4_I(inode);
4798         iloc.bh = 0;
4799
4800         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4801         if (ret < 0)
4802                 goto bad_inode;
4803         raw_inode = ext4_raw_inode(&iloc);
4804         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4805         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4806         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4807         if (!(test_opt(inode->i_sb, NO_UID32))) {
4808                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4809                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4810         }
4811         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
4812
4813         ei->i_state = 0;
4814         ei->i_dir_start_lookup = 0;
4815         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4816         /* We now have enough fields to check if the inode was active or not.
4817          * This is needed because nfsd might try to access dead inodes
4818          * the test is that same one that e2fsck uses
4819          * NeilBrown 1999oct15
4820          */
4821         if (inode->i_nlink == 0) {
4822                 if (inode->i_mode == 0 ||
4823                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
4824                         /* this inode is deleted */
4825                         ret = -ESTALE;
4826                         goto bad_inode;
4827                 }
4828                 /* The only unlinked inodes we let through here have
4829                  * valid i_mode and are being read by the orphan
4830                  * recovery code: that's fine, we're about to complete
4831                  * the process of deleting those. */
4832         }
4833         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4834         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4835         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4836         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
4837                 ei->i_file_acl |=
4838                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4839         inode->i_size = ext4_isize(raw_inode);
4840         ei->i_disksize = inode->i_size;
4841         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4842         ei->i_block_group = iloc.block_group;
4843         ei->i_last_alloc_group = ~0;
4844         /*
4845          * NOTE! The in-memory inode i_data array is in little-endian order
4846          * even on big-endian machines: we do NOT byteswap the block numbers!
4847          */
4848         for (block = 0; block < EXT4_N_BLOCKS; block++)
4849                 ei->i_data[block] = raw_inode->i_block[block];
4850         INIT_LIST_HEAD(&ei->i_orphan);
4851
4852         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4853                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4854                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4855                     EXT4_INODE_SIZE(inode->i_sb)) {
4856                         ret = -EIO;
4857                         goto bad_inode;
4858                 }
4859                 if (ei->i_extra_isize == 0) {
4860                         /* The extra space is currently unused. Use it. */
4861                         ei->i_extra_isize = sizeof(struct ext4_inode) -
4862                                             EXT4_GOOD_OLD_INODE_SIZE;
4863                 } else {
4864                         __le32 *magic = (void *)raw_inode +
4865                                         EXT4_GOOD_OLD_INODE_SIZE +
4866                                         ei->i_extra_isize;
4867                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
4868                                 ei->i_state |= EXT4_STATE_XATTR;
4869                 }
4870         } else
4871                 ei->i_extra_isize = 0;
4872
4873         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4874         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4875         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4876         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4877
4878         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4879         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4880                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4881                         inode->i_version |=
4882                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4883         }
4884
4885         ret = 0;
4886         if (ei->i_file_acl &&
4887             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
4888                 ext4_error(sb, __func__,
4889                            "bad extended attribute block %llu in inode #%lu",
4890                            ei->i_file_acl, inode->i_ino);
4891                 ret = -EIO;
4892                 goto bad_inode;
4893         } else if (ei->i_flags & EXT4_EXTENTS_FL) {
4894                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4895                     (S_ISLNK(inode->i_mode) &&
4896                      !ext4_inode_is_fast_symlink(inode)))
4897                         /* Validate extent which is part of inode */
4898                         ret = ext4_ext_check_inode(inode);
4899         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4900                    (S_ISLNK(inode->i_mode) &&
4901                     !ext4_inode_is_fast_symlink(inode))) {
4902                 /* Validate block references which are part of inode */
4903                 ret = ext4_check_inode_blockref(inode);
4904         }
4905         if (ret)
4906                 goto bad_inode;
4907
4908         if (S_ISREG(inode->i_mode)) {
4909                 inode->i_op = &ext4_file_inode_operations;
4910                 inode->i_fop = &ext4_file_operations;
4911                 ext4_set_aops(inode);
4912         } else if (S_ISDIR(inode->i_mode)) {
4913                 inode->i_op = &ext4_dir_inode_operations;
4914                 inode->i_fop = &ext4_dir_operations;
4915         } else if (S_ISLNK(inode->i_mode)) {
4916                 if (ext4_inode_is_fast_symlink(inode)) {
4917                         inode->i_op = &ext4_fast_symlink_inode_operations;
4918                         nd_terminate_link(ei->i_data, inode->i_size,
4919                                 sizeof(ei->i_data) - 1);
4920                 } else {
4921                         inode->i_op = &ext4_symlink_inode_operations;
4922                         ext4_set_aops(inode);
4923                 }
4924         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
4925               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
4926                 inode->i_op = &ext4_special_inode_operations;
4927                 if (raw_inode->i_block[0])
4928                         init_special_inode(inode, inode->i_mode,
4929                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4930                 else
4931                         init_special_inode(inode, inode->i_mode,
4932                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4933         } else {
4934                 ret = -EIO;
4935                 ext4_error(inode->i_sb, __func__,
4936                            "bogus i_mode (%o) for inode=%lu",
4937                            inode->i_mode, inode->i_ino);
4938                 goto bad_inode;
4939         }
4940         brelse(iloc.bh);
4941         ext4_set_inode_flags(inode);
4942         unlock_new_inode(inode);
4943         return inode;
4944
4945 bad_inode:
4946         brelse(iloc.bh);
4947         iget_failed(inode);
4948         return ERR_PTR(ret);
4949 }
4950
4951 static int ext4_inode_blocks_set(handle_t *handle,
4952                                 struct ext4_inode *raw_inode,
4953                                 struct ext4_inode_info *ei)
4954 {
4955         struct inode *inode = &(ei->vfs_inode);
4956         u64 i_blocks = inode->i_blocks;
4957         struct super_block *sb = inode->i_sb;
4958
4959         if (i_blocks <= ~0U) {
4960                 /*
4961                  * i_blocks can be represnted in a 32 bit variable
4962                  * as multiple of 512 bytes
4963                  */
4964                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4965                 raw_inode->i_blocks_high = 0;
4966                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4967                 return 0;
4968         }
4969         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
4970                 return -EFBIG;
4971
4972         if (i_blocks <= 0xffffffffffffULL) {
4973                 /*
4974                  * i_blocks can be represented in a 48 bit variable
4975                  * as multiple of 512 bytes
4976                  */
4977                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4978                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4979                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4980         } else {
4981                 ei->i_flags |= EXT4_HUGE_FILE_FL;
4982                 /* i_block is stored in file system block size */
4983                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4984                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4985                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4986         }
4987         return 0;
4988 }
4989
4990 /*
4991  * Post the struct inode info into an on-disk inode location in the
4992  * buffer-cache.  This gobbles the caller's reference to the
4993  * buffer_head in the inode location struct.
4994  *
4995  * The caller must have write access to iloc->bh.
4996  */
4997 static int ext4_do_update_inode(handle_t *handle,
4998                                 struct inode *inode,
4999                                 struct ext4_iloc *iloc)
5000 {
5001         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
5002         struct ext4_inode_info *ei = EXT4_I(inode);
5003         struct buffer_head *bh = iloc->bh;
5004         int err = 0, rc, block;
5005
5006         /* For fields not not tracking in the in-memory inode,
5007          * initialise them to zero for new inodes. */
5008         if (ei->i_state & EXT4_STATE_NEW)
5009                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5010
5011         ext4_get_inode_flags(ei);
5012         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5013         if (!(test_opt(inode->i_sb, NO_UID32))) {
5014                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
5015                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
5016 /*
5017  * Fix up interoperability with old kernels. Otherwise, old inodes get
5018  * re-used with the upper 16 bits of the uid/gid intact
5019  */
5020                 if (!ei->i_dtime) {
5021                         raw_inode->i_uid_high =
5022                                 cpu_to_le16(high_16_bits(inode->i_uid));
5023                         raw_inode->i_gid_high =
5024                                 cpu_to_le16(high_16_bits(inode->i_gid));
5025                 } else {
5026                         raw_inode->i_uid_high = 0;
5027                         raw_inode->i_gid_high = 0;
5028                 }
5029         } else {
5030                 raw_inode->i_uid_low =
5031                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
5032                 raw_inode->i_gid_low =
5033                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
5034                 raw_inode->i_uid_high = 0;
5035                 raw_inode->i_gid_high = 0;
5036         }
5037         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5038
5039         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5040         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5041         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5042         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5043
5044         if (ext4_inode_blocks_set(handle, raw_inode, ei))
5045                 goto out_brelse;
5046         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5047         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
5048         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
5049             cpu_to_le32(EXT4_OS_HURD))
5050                 raw_inode->i_file_acl_high =
5051                         cpu_to_le16(ei->i_file_acl >> 32);
5052         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5053         ext4_isize_set(raw_inode, ei->i_disksize);
5054         if (ei->i_disksize > 0x7fffffffULL) {
5055                 struct super_block *sb = inode->i_sb;
5056                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
5057                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
5058                                 EXT4_SB(sb)->s_es->s_rev_level ==
5059                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
5060                         /* If this is the first large file
5061                          * created, add a flag to the superblock.
5062                          */
5063                         err = ext4_journal_get_write_access(handle,
5064                                         EXT4_SB(sb)->s_sbh);
5065                         if (err)
5066                                 goto out_brelse;
5067                         ext4_update_dynamic_rev(sb);
5068                         EXT4_SET_RO_COMPAT_FEATURE(sb,
5069                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
5070                         sb->s_dirt = 1;
5071                         ext4_handle_sync(handle);
5072                         err = ext4_handle_dirty_metadata(handle, inode,
5073                                         EXT4_SB(sb)->s_sbh);
5074                 }
5075         }
5076         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5077         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5078                 if (old_valid_dev(inode->i_rdev)) {
5079                         raw_inode->i_block[0] =
5080                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
5081                         raw_inode->i_block[1] = 0;
5082                 } else {
5083                         raw_inode->i_block[0] = 0;
5084                         raw_inode->i_block[1] =
5085                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
5086                         raw_inode->i_block[2] = 0;
5087                 }
5088         } else
5089                 for (block = 0; block < EXT4_N_BLOCKS; block++)
5090                         raw_inode->i_block[block] = ei->i_data[block];
5091
5092         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
5093         if (ei->i_extra_isize) {
5094                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5095                         raw_inode->i_version_hi =
5096                         cpu_to_le32(inode->i_version >> 32);
5097                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
5098         }
5099
5100         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5101         rc = ext4_handle_dirty_metadata(handle, inode, bh);
5102         if (!err)
5103                 err = rc;
5104         ei->i_state &= ~EXT4_STATE_NEW;
5105
5106 out_brelse:
5107         brelse(bh);
5108         ext4_std_error(inode->i_sb, err);
5109         return err;
5110 }
5111
5112 /*
5113  * ext4_write_inode()
5114  *
5115  * We are called from a few places:
5116  *
5117  * - Within generic_file_write() for O_SYNC files.
5118  *   Here, there will be no transaction running. We wait for any running
5119  *   trasnaction to commit.
5120  *
5121  * - Within sys_sync(), kupdate and such.
5122  *   We wait on commit, if tol to.
5123  *
5124  * - Within prune_icache() (PF_MEMALLOC == true)
5125  *   Here we simply return.  We can't afford to block kswapd on the
5126  *   journal commit.
5127  *
5128  * In all cases it is actually safe for us to return without doing anything,
5129  * because the inode has been copied into a raw inode buffer in
5130  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
5131  * knfsd.
5132  *
5133  * Note that we are absolutely dependent upon all inode dirtiers doing the
5134  * right thing: they *must* call mark_inode_dirty() after dirtying info in
5135  * which we are interested.
5136  *
5137  * It would be a bug for them to not do this.  The code:
5138  *
5139  *      mark_inode_dirty(inode)
5140  *      stuff();
5141  *      inode->i_size = expr;
5142  *
5143  * is in error because a kswapd-driven write_inode() could occur while
5144  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
5145  * will no longer be on the superblock's dirty inode list.
5146  */
5147 int ext4_write_inode(struct inode *inode, int wait)
5148 {
5149         int err;
5150
5151         if (current->flags & PF_MEMALLOC)
5152                 return 0;
5153
5154         if (EXT4_SB(inode->i_sb)->s_journal) {
5155                 if (ext4_journal_current_handle()) {
5156                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5157                         dump_stack();
5158                         return -EIO;
5159                 }
5160
5161                 if (!wait)
5162                         return 0;
5163
5164                 err = ext4_force_commit(inode->i_sb);
5165         } else {
5166                 struct ext4_iloc iloc;
5167
5168                 err = ext4_get_inode_loc(inode, &iloc);
5169                 if (err)
5170                         return err;
5171                 if (wait)
5172                         sync_dirty_buffer(iloc.bh);
5173                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5174                         ext4_error(inode->i_sb, __func__,
5175                                    "IO error syncing inode, "
5176                                    "inode=%lu, block=%llu",
5177                                    inode->i_ino,
5178                                    (unsigned long long)iloc.bh->b_blocknr);
5179                         err = -EIO;
5180                 }
5181         }
5182         return err;
5183 }
5184
5185 /*
5186  * ext4_setattr()
5187  *
5188  * Called from notify_change.
5189  *
5190  * We want to trap VFS attempts to truncate the file as soon as
5191  * possible.  In particular, we want to make sure that when the VFS
5192  * shrinks i_size, we put the inode on the orphan list and modify
5193  * i_disksize immediately, so that during the subsequent flushing of
5194  * dirty pages and freeing of disk blocks, we can guarantee that any
5195  * commit will leave the blocks being flushed in an unused state on
5196  * disk.  (On recovery, the inode will get truncated and the blocks will
5197  * be freed, so we have a strong guarantee that no future commit will
5198  * leave these blocks visible to the user.)
5199  *
5200  * Another thing we have to assure is that if we are in ordered mode
5201  * and inode is still attached to the committing transaction, we must
5202  * we start writeout of all the dirty pages which are being truncated.
5203  * This way we are sure that all the data written in the previous
5204  * transaction are already on disk (truncate waits for pages under
5205  * writeback).
5206  *
5207  * Called with inode->i_mutex down.
5208  */
5209 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5210 {
5211         struct inode *inode = dentry->d_inode;
5212         int error, rc = 0;
5213         const unsigned int ia_valid = attr->ia_valid;
5214
5215         error = inode_change_ok(inode, attr);
5216         if (error)
5217                 return error;
5218
5219         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
5220                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
5221                 handle_t *handle;
5222
5223                 /* (user+group)*(old+new) structure, inode write (sb,
5224                  * inode block, ? - but truncate inode update has it) */
5225                 handle = ext4_journal_start(inode, 2*(EXT4_QUOTA_INIT_BLOCKS(inode->i_sb)+
5226                                         EXT4_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
5227                 if (IS_ERR(handle)) {
5228                         error = PTR_ERR(handle);
5229                         goto err_out;
5230                 }
5231                 error = vfs_dq_transfer(inode, attr) ? -EDQUOT : 0;
5232                 if (error) {
5233                         ext4_journal_stop(handle);
5234                         return error;
5235                 }
5236                 /* Update corresponding info in inode so that everything is in
5237                  * one transaction */
5238                 if (attr->ia_valid & ATTR_UID)
5239                         inode->i_uid = attr->ia_uid;
5240                 if (attr->ia_valid & ATTR_GID)
5241                         inode->i_gid = attr->ia_gid;
5242                 error = ext4_mark_inode_dirty(handle, inode);
5243                 ext4_journal_stop(handle);
5244         }
5245
5246         if (attr->ia_valid & ATTR_SIZE) {
5247                 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)) {
5248                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5249
5250                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
5251                                 error = -EFBIG;
5252                                 goto err_out;
5253                         }
5254                 }
5255         }
5256
5257         if (S_ISREG(inode->i_mode) &&
5258             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
5259                 handle_t *handle;
5260
5261                 handle = ext4_journal_start(inode, 3);
5262                 if (IS_ERR(handle)) {
5263                         error = PTR_ERR(handle);
5264                         goto err_out;
5265                 }
5266
5267                 error = ext4_orphan_add(handle, inode);
5268                 EXT4_I(inode)->i_disksize = attr->ia_size;
5269                 rc = ext4_mark_inode_dirty(handle, inode);
5270                 if (!error)
5271                         error = rc;
5272                 ext4_journal_stop(handle);
5273
5274                 if (ext4_should_order_data(inode)) {
5275                         error = ext4_begin_ordered_truncate(inode,
5276                                                             attr->ia_size);
5277                         if (error) {
5278                                 /* Do as much error cleanup as possible */
5279                                 handle = ext4_journal_start(inode, 3);
5280                                 if (IS_ERR(handle)) {
5281                                         ext4_orphan_del(NULL, inode);
5282                                         goto err_out;
5283                                 }
5284                                 ext4_orphan_del(handle, inode);
5285                                 ext4_journal_stop(handle);
5286                                 goto err_out;
5287                         }
5288                 }
5289         }
5290
5291         rc = inode_setattr(inode, attr);
5292
5293         /* If inode_setattr's call to ext4_truncate failed to get a
5294          * transaction handle at all, we need to clean up the in-core
5295          * orphan list manually. */
5296         if (inode->i_nlink)
5297                 ext4_orphan_del(NULL, inode);
5298
5299         if (!rc && (ia_valid & ATTR_MODE))
5300                 rc = ext4_acl_chmod(inode);
5301
5302 err_out:
5303         ext4_std_error(inode->i_sb, error);
5304         if (!error)
5305                 error = rc;
5306         return error;
5307 }
5308
5309 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
5310                  struct kstat *stat)
5311 {
5312         struct inode *inode;
5313         unsigned long delalloc_blocks;
5314
5315         inode = dentry->d_inode;
5316         generic_fillattr(inode, stat);
5317
5318         /*
5319          * We can't update i_blocks if the block allocation is delayed
5320          * otherwise in the case of system crash before the real block
5321          * allocation is done, we will have i_blocks inconsistent with
5322          * on-disk file blocks.
5323          * We always keep i_blocks updated together with real
5324          * allocation. But to not confuse with user, stat
5325          * will return the blocks that include the delayed allocation
5326          * blocks for this file.
5327          */
5328         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
5329         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
5330         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
5331
5332         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
5333         return 0;
5334 }
5335
5336 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
5337                                       int chunk)
5338 {
5339         int indirects;
5340
5341         /* if nrblocks are contiguous */
5342         if (chunk) {
5343                 /*
5344                  * With N contiguous data blocks, it need at most
5345                  * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) indirect blocks
5346                  * 2 dindirect blocks
5347                  * 1 tindirect block
5348                  */
5349                 indirects = nrblocks / EXT4_ADDR_PER_BLOCK(inode->i_sb);
5350                 return indirects + 3;
5351         }
5352         /*
5353          * if nrblocks are not contiguous, worse case, each block touch
5354          * a indirect block, and each indirect block touch a double indirect
5355          * block, plus a triple indirect block
5356          */
5357         indirects = nrblocks * 2 + 1;
5358         return indirects;
5359 }
5360
5361 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5362 {
5363         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL))
5364                 return ext4_indirect_trans_blocks(inode, nrblocks, chunk);
5365         return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
5366 }
5367
5368 /*
5369  * Account for index blocks, block groups bitmaps and block group
5370  * descriptor blocks if modify datablocks and index blocks
5371  * worse case, the indexs blocks spread over different block groups
5372  *
5373  * If datablocks are discontiguous, they are possible to spread over
5374  * different block groups too. If they are contiugous, with flexbg,
5375  * they could still across block group boundary.
5376  *
5377  * Also account for superblock, inode, quota and xattr blocks
5378  */
5379 int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5380 {
5381         ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
5382         int gdpblocks;
5383         int idxblocks;
5384         int ret = 0;
5385
5386         /*
5387          * How many index blocks need to touch to modify nrblocks?
5388          * The "Chunk" flag indicating whether the nrblocks is
5389          * physically contiguous on disk
5390          *
5391          * For Direct IO and fallocate, they calls get_block to allocate
5392          * one single extent at a time, so they could set the "Chunk" flag
5393          */
5394         idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
5395
5396         ret = idxblocks;
5397
5398         /*
5399          * Now let's see how many group bitmaps and group descriptors need
5400          * to account
5401          */
5402         groups = idxblocks;
5403         if (chunk)
5404                 groups += 1;
5405         else
5406                 groups += nrblocks;
5407
5408         gdpblocks = groups;
5409         if (groups > ngroups)
5410                 groups = ngroups;
5411         if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
5412                 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
5413
5414         /* bitmaps and block group descriptor blocks */
5415         ret += groups + gdpblocks;
5416
5417         /* Blocks for super block, inode, quota and xattr blocks */
5418         ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
5419
5420         return ret;
5421 }
5422
5423 /*
5424  * Calulate the total number of credits to reserve to fit
5425  * the modification of a single pages into a single transaction,
5426  * which may include multiple chunks of block allocations.
5427  *
5428  * This could be called via ext4_write_begin()
5429  *
5430  * We need to consider the worse case, when
5431  * one new block per extent.
5432  */
5433 int ext4_writepage_trans_blocks(struct inode *inode)
5434 {
5435         int bpp = ext4_journal_blocks_per_page(inode);
5436         int ret;
5437
5438         ret = ext4_meta_trans_blocks(inode, bpp, 0);
5439
5440         /* Account for data blocks for journalled mode */
5441         if (ext4_should_journal_data(inode))
5442                 ret += bpp;
5443         return ret;
5444 }
5445
5446 /*
5447  * Calculate the journal credits for a chunk of data modification.
5448  *
5449  * This is called from DIO, fallocate or whoever calling
5450  * ext4_get_blocks() to map/allocate a chunk of contigous disk blocks.
5451  *
5452  * journal buffers for data blocks are not included here, as DIO
5453  * and fallocate do no need to journal data buffers.
5454  */
5455 int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
5456 {
5457         return ext4_meta_trans_blocks(inode, nrblocks, 1);
5458 }
5459
5460 /*
5461  * The caller must have previously called ext4_reserve_inode_write().
5462  * Give this, we know that the caller already has write access to iloc->bh.
5463  */
5464 int ext4_mark_iloc_dirty(handle_t *handle,
5465                          struct inode *inode, struct ext4_iloc *iloc)
5466 {
5467         int err = 0;
5468
5469         if (test_opt(inode->i_sb, I_VERSION))
5470                 inode_inc_iversion(inode);
5471
5472         /* the do_update_inode consumes one bh->b_count */
5473         get_bh(iloc->bh);
5474
5475         /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
5476         err = ext4_do_update_inode(handle, inode, iloc);
5477         put_bh(iloc->bh);
5478         return err;
5479 }
5480
5481 /*
5482  * On success, We end up with an outstanding reference count against
5483  * iloc->bh.  This _must_ be cleaned up later.
5484  */
5485
5486 int
5487 ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
5488                          struct ext4_iloc *iloc)
5489 {
5490         int err;
5491
5492         err = ext4_get_inode_loc(inode, iloc);
5493         if (!err) {
5494                 BUFFER_TRACE(iloc->bh, "get_write_access");
5495                 err = ext4_journal_get_write_access(handle, iloc->bh);
5496                 if (err) {
5497                         brelse(iloc->bh);
5498                         iloc->bh = NULL;
5499                 }
5500         }
5501         ext4_std_error(inode->i_sb, err);
5502         return err;
5503 }
5504
5505 /*
5506  * Expand an inode by new_extra_isize bytes.
5507  * Returns 0 on success or negative error number on failure.
5508  */
5509 static int ext4_expand_extra_isize(struct inode *inode,
5510                                    unsigned int new_extra_isize,
5511                                    struct ext4_iloc iloc,
5512                                    handle_t *handle)
5513 {
5514         struct ext4_inode *raw_inode;
5515         struct ext4_xattr_ibody_header *header;
5516         struct ext4_xattr_entry *entry;
5517
5518         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
5519                 return 0;
5520
5521         raw_inode = ext4_raw_inode(&iloc);
5522
5523         header = IHDR(inode, raw_inode);
5524         entry = IFIRST(header);
5525
5526         /* No extended attributes present */
5527         if (!(EXT4_I(inode)->i_state & EXT4_STATE_XATTR) ||
5528                 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
5529                 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
5530                         new_extra_isize);
5531                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
5532                 return 0;
5533         }
5534
5535         /* try to expand with EAs present */
5536         return ext4_expand_extra_isize_ea(inode, new_extra_isize,
5537                                           raw_inode, handle);
5538 }
5539
5540 /*
5541  * What we do here is to mark the in-core inode as clean with respect to inode
5542  * dirtiness (it may still be data-dirty).
5543  * This means that the in-core inode may be reaped by prune_icache
5544  * without having to perform any I/O.  This is a very good thing,
5545  * because *any* task may call prune_icache - even ones which
5546  * have a transaction open against a different journal.
5547  *
5548  * Is this cheating?  Not really.  Sure, we haven't written the
5549  * inode out, but prune_icache isn't a user-visible syncing function.
5550  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
5551  * we start and wait on commits.
5552  *
5553  * Is this efficient/effective?  Well, we're being nice to the system
5554  * by cleaning up our inodes proactively so they can be reaped
5555  * without I/O.  But we are potentially leaving up to five seconds'
5556  * worth of inodes floating about which prune_icache wants us to
5557  * write out.  One way to fix that would be to get prune_icache()
5558  * to do a write_super() to free up some memory.  It has the desired
5559  * effect.
5560  */
5561 int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
5562 {
5563         struct ext4_iloc iloc;
5564         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5565         static unsigned int mnt_count;
5566         int err, ret;
5567
5568         might_sleep();
5569         err = ext4_reserve_inode_write(handle, inode, &iloc);
5570         if (ext4_handle_valid(handle) &&
5571             EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
5572             !(EXT4_I(inode)->i_state & EXT4_STATE_NO_EXPAND)) {
5573                 /*
5574                  * We need extra buffer credits since we may write into EA block
5575                  * with this same handle. If journal_extend fails, then it will
5576                  * only result in a minor loss of functionality for that inode.
5577                  * If this is felt to be critical, then e2fsck should be run to
5578                  * force a large enough s_min_extra_isize.
5579                  */
5580                 if ((jbd2_journal_extend(handle,
5581                              EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
5582                         ret = ext4_expand_extra_isize(inode,
5583                                                       sbi->s_want_extra_isize,
5584                                                       iloc, handle);
5585                         if (ret) {
5586                                 EXT4_I(inode)->i_state |= EXT4_STATE_NO_EXPAND;
5587                                 if (mnt_count !=
5588                                         le16_to_cpu(sbi->s_es->s_mnt_count)) {
5589                                         ext4_warning(inode->i_sb, __func__,
5590                                         "Unable to expand inode %lu. Delete"
5591                                         " some EAs or run e2fsck.",
5592                                         inode->i_ino);
5593                                         mnt_count =
5594                                     &