ext4: truncate the file properly if we fail to copy data from userspace
[linux-2.6] / 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 "ext4_jbd2.h"
41 #include "xattr.h"
42 #include "acl.h"
43 #include "ext4_extents.h"
44
45 #define MPAGE_DA_EXTENT_TAIL 0x01
46
47 static inline int ext4_begin_ordered_truncate(struct inode *inode,
48                                               loff_t new_size)
49 {
50         return jbd2_journal_begin_ordered_truncate(
51                                         EXT4_SB(inode->i_sb)->s_journal,
52                                         &EXT4_I(inode)->jinode,
53                                         new_size);
54 }
55
56 static void ext4_invalidatepage(struct page *page, unsigned long offset);
57
58 /*
59  * Test whether an inode is a fast symlink.
60  */
61 static int ext4_inode_is_fast_symlink(struct inode *inode)
62 {
63         int ea_blocks = EXT4_I(inode)->i_file_acl ?
64                 (inode->i_sb->s_blocksize >> 9) : 0;
65
66         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
67 }
68
69 /*
70  * The ext4 forget function must perform a revoke if we are freeing data
71  * which has been journaled.  Metadata (eg. indirect blocks) must be
72  * revoked in all cases.
73  *
74  * "bh" may be NULL: a metadata block may have been freed from memory
75  * but there may still be a record of it in the journal, and that record
76  * still needs to be revoked.
77  *
78  * If the handle isn't valid we're not journaling so there's nothing to do.
79  */
80 int ext4_forget(handle_t *handle, int is_metadata, struct inode *inode,
81                         struct buffer_head *bh, ext4_fsblk_t blocknr)
82 {
83         int err;
84
85         if (!ext4_handle_valid(handle))
86                 return 0;
87
88         might_sleep();
89
90         BUFFER_TRACE(bh, "enter");
91
92         jbd_debug(4, "forgetting bh %p: is_metadata = %d, mode %o, "
93                   "data mode %lx\n",
94                   bh, is_metadata, inode->i_mode,
95                   test_opt(inode->i_sb, DATA_FLAGS));
96
97         /* Never use the revoke function if we are doing full data
98          * journaling: there is no need to, and a V1 superblock won't
99          * support it.  Otherwise, only skip the revoke on un-journaled
100          * data blocks. */
101
102         if (test_opt(inode->i_sb, DATA_FLAGS) == EXT4_MOUNT_JOURNAL_DATA ||
103             (!is_metadata && !ext4_should_journal_data(inode))) {
104                 if (bh) {
105                         BUFFER_TRACE(bh, "call jbd2_journal_forget");
106                         return ext4_journal_forget(handle, bh);
107                 }
108                 return 0;
109         }
110
111         /*
112          * data!=journal && (is_metadata || should_journal_data(inode))
113          */
114         BUFFER_TRACE(bh, "call ext4_journal_revoke");
115         err = ext4_journal_revoke(handle, blocknr, bh);
116         if (err)
117                 ext4_abort(inode->i_sb, __func__,
118                            "error %d when attempting revoke", err);
119         BUFFER_TRACE(bh, "exit");
120         return err;
121 }
122
123 /*
124  * Work out how many blocks we need to proceed with the next chunk of a
125  * truncate transaction.
126  */
127 static unsigned long blocks_for_truncate(struct inode *inode)
128 {
129         ext4_lblk_t needed;
130
131         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
132
133         /* Give ourselves just enough room to cope with inodes in which
134          * i_blocks is corrupt: we've seen disk corruptions in the past
135          * which resulted in random data in an inode which looked enough
136          * like a regular file for ext4 to try to delete it.  Things
137          * will go a bit crazy if that happens, but at least we should
138          * try not to panic the whole kernel. */
139         if (needed < 2)
140                 needed = 2;
141
142         /* But we need to bound the transaction so we don't overflow the
143          * journal. */
144         if (needed > EXT4_MAX_TRANS_DATA)
145                 needed = EXT4_MAX_TRANS_DATA;
146
147         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
148 }
149
150 /*
151  * Truncate transactions can be complex and absolutely huge.  So we need to
152  * be able to restart the transaction at a conventient checkpoint to make
153  * sure we don't overflow the journal.
154  *
155  * start_transaction gets us a new handle for a truncate transaction,
156  * and extend_transaction tries to extend the existing one a bit.  If
157  * extend fails, we need to propagate the failure up and restart the
158  * transaction in the top-level truncate loop. --sct
159  */
160 static handle_t *start_transaction(struct inode *inode)
161 {
162         handle_t *result;
163
164         result = ext4_journal_start(inode, blocks_for_truncate(inode));
165         if (!IS_ERR(result))
166                 return result;
167
168         ext4_std_error(inode->i_sb, PTR_ERR(result));
169         return result;
170 }
171
172 /*
173  * Try to extend this transaction for the purposes of truncation.
174  *
175  * Returns 0 if we managed to create more room.  If we can't create more
176  * room, and the transaction must be restarted we return 1.
177  */
178 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
179 {
180         if (!ext4_handle_valid(handle))
181                 return 0;
182         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
183                 return 0;
184         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
185                 return 0;
186         return 1;
187 }
188
189 /*
190  * Restart the transaction associated with *handle.  This does a commit,
191  * so before we call here everything must be consistently dirtied against
192  * this transaction.
193  */
194 static int ext4_journal_test_restart(handle_t *handle, struct inode *inode)
195 {
196         BUG_ON(EXT4_JOURNAL(inode) == NULL);
197         jbd_debug(2, "restarting handle %p\n", handle);
198         return ext4_journal_restart(handle, blocks_for_truncate(inode));
199 }
200
201 /*
202  * Called at the last iput() if i_nlink is zero.
203  */
204 void ext4_delete_inode(struct inode *inode)
205 {
206         handle_t *handle;
207         int err;
208
209         if (ext4_should_order_data(inode))
210                 ext4_begin_ordered_truncate(inode, 0);
211         truncate_inode_pages(&inode->i_data, 0);
212
213         if (is_bad_inode(inode))
214                 goto no_delete;
215
216         handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
217         if (IS_ERR(handle)) {
218                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
219                 /*
220                  * If we're going to skip the normal cleanup, we still need to
221                  * make sure that the in-core orphan linked list is properly
222                  * cleaned up.
223                  */
224                 ext4_orphan_del(NULL, inode);
225                 goto no_delete;
226         }
227
228         if (IS_SYNC(inode))
229                 ext4_handle_sync(handle);
230         inode->i_size = 0;
231         err = ext4_mark_inode_dirty(handle, inode);
232         if (err) {
233                 ext4_warning(inode->i_sb, __func__,
234                              "couldn't mark inode dirty (err %d)", err);
235                 goto stop_handle;
236         }
237         if (inode->i_blocks)
238                 ext4_truncate(inode);
239
240         /*
241          * ext4_ext_truncate() doesn't reserve any slop when it
242          * restarts journal transactions; therefore there may not be
243          * enough credits left in the handle to remove the inode from
244          * the orphan list and set the dtime field.
245          */
246         if (!ext4_handle_has_enough_credits(handle, 3)) {
247                 err = ext4_journal_extend(handle, 3);
248                 if (err > 0)
249                         err = ext4_journal_restart(handle, 3);
250                 if (err != 0) {
251                         ext4_warning(inode->i_sb, __func__,
252                                      "couldn't extend journal (err %d)", err);
253                 stop_handle:
254                         ext4_journal_stop(handle);
255                         goto no_delete;
256                 }
257         }
258
259         /*
260          * Kill off the orphan record which ext4_truncate created.
261          * AKPM: I think this can be inside the above `if'.
262          * Note that ext4_orphan_del() has to be able to cope with the
263          * deletion of a non-existent orphan - this is because we don't
264          * know if ext4_truncate() actually created an orphan record.
265          * (Well, we could do this if we need to, but heck - it works)
266          */
267         ext4_orphan_del(handle, inode);
268         EXT4_I(inode)->i_dtime  = get_seconds();
269
270         /*
271          * One subtle ordering requirement: if anything has gone wrong
272          * (transaction abort, IO errors, whatever), then we can still
273          * do these next steps (the fs will already have been marked as
274          * having errors), but we can't free the inode if the mark_dirty
275          * fails.
276          */
277         if (ext4_mark_inode_dirty(handle, inode))
278                 /* If that failed, just do the required in-core inode clear. */
279                 clear_inode(inode);
280         else
281                 ext4_free_inode(handle, inode);
282         ext4_journal_stop(handle);
283         return;
284 no_delete:
285         clear_inode(inode);     /* We must guarantee clearing of inode... */
286 }
287
288 typedef struct {
289         __le32  *p;
290         __le32  key;
291         struct buffer_head *bh;
292 } Indirect;
293
294 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
295 {
296         p->key = *(p->p = v);
297         p->bh = bh;
298 }
299
300 /**
301  *      ext4_block_to_path - parse the block number into array of offsets
302  *      @inode: inode in question (we are only interested in its superblock)
303  *      @i_block: block number to be parsed
304  *      @offsets: array to store the offsets in
305  *      @boundary: set this non-zero if the referred-to block is likely to be
306  *             followed (on disk) by an indirect block.
307  *
308  *      To store the locations of file's data ext4 uses a data structure common
309  *      for UNIX filesystems - tree of pointers anchored in the inode, with
310  *      data blocks at leaves and indirect blocks in intermediate nodes.
311  *      This function translates the block number into path in that tree -
312  *      return value is the path length and @offsets[n] is the offset of
313  *      pointer to (n+1)th node in the nth one. If @block is out of range
314  *      (negative or too large) warning is printed and zero returned.
315  *
316  *      Note: function doesn't find node addresses, so no IO is needed. All
317  *      we need to know is the capacity of indirect blocks (taken from the
318  *      inode->i_sb).
319  */
320
321 /*
322  * Portability note: the last comparison (check that we fit into triple
323  * indirect block) is spelled differently, because otherwise on an
324  * architecture with 32-bit longs and 8Kb pages we might get into trouble
325  * if our filesystem had 8Kb blocks. We might use long long, but that would
326  * kill us on x86. Oh, well, at least the sign propagation does not matter -
327  * i_block would have to be negative in the very beginning, so we would not
328  * get there at all.
329  */
330
331 static int ext4_block_to_path(struct inode *inode,
332                         ext4_lblk_t i_block,
333                         ext4_lblk_t offsets[4], int *boundary)
334 {
335         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
336         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
337         const long direct_blocks = EXT4_NDIR_BLOCKS,
338                 indirect_blocks = ptrs,
339                 double_blocks = (1 << (ptrs_bits * 2));
340         int n = 0;
341         int final = 0;
342
343         if (i_block < 0) {
344                 ext4_warning(inode->i_sb, "ext4_block_to_path", "block < 0");
345         } else if (i_block < direct_blocks) {
346                 offsets[n++] = i_block;
347                 final = direct_blocks;
348         } else if ((i_block -= direct_blocks) < indirect_blocks) {
349                 offsets[n++] = EXT4_IND_BLOCK;
350                 offsets[n++] = i_block;
351                 final = ptrs;
352         } else if ((i_block -= indirect_blocks) < double_blocks) {
353                 offsets[n++] = EXT4_DIND_BLOCK;
354                 offsets[n++] = i_block >> ptrs_bits;
355                 offsets[n++] = i_block & (ptrs - 1);
356                 final = ptrs;
357         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
358                 offsets[n++] = EXT4_TIND_BLOCK;
359                 offsets[n++] = i_block >> (ptrs_bits * 2);
360                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
361                 offsets[n++] = i_block & (ptrs - 1);
362                 final = ptrs;
363         } else {
364                 ext4_warning(inode->i_sb, "ext4_block_to_path",
365                                 "block %lu > max in inode %lu",
366                                 i_block + direct_blocks +
367                                 indirect_blocks + double_blocks, inode->i_ino);
368         }
369         if (boundary)
370                 *boundary = final - 1 - (i_block & (ptrs - 1));
371         return n;
372 }
373
374 static int __ext4_check_blockref(const char *function, struct inode *inode,
375                                  __le32 *p, unsigned int max)
376 {
377         __le32 *bref = p;
378         unsigned int blk;
379
380         while (bref < p+max) {
381                 blk = le32_to_cpu(*bref++);
382                 if (blk && 
383                     unlikely(!ext4_data_block_valid(EXT4_SB(inode->i_sb), 
384                                                     blk, 1))) {
385                         ext4_error(inode->i_sb, function,
386                                    "invalid block reference %u "
387                                    "in inode #%lu", blk, inode->i_ino);
388                         return -EIO;
389                 }
390         }
391         return 0;
392 }
393
394
395 #define ext4_check_indirect_blockref(inode, bh)                         \
396         __ext4_check_blockref(__func__, inode, (__le32 *)(bh)->b_data,  \
397                               EXT4_ADDR_PER_BLOCK((inode)->i_sb))
398
399 #define ext4_check_inode_blockref(inode)                                \
400         __ext4_check_blockref(__func__, inode, EXT4_I(inode)->i_data,   \
401                               EXT4_NDIR_BLOCKS)
402
403 /**
404  *      ext4_get_branch - read the chain of indirect blocks leading to data
405  *      @inode: inode in question
406  *      @depth: depth of the chain (1 - direct pointer, etc.)
407  *      @offsets: offsets of pointers in inode/indirect blocks
408  *      @chain: place to store the result
409  *      @err: here we store the error value
410  *
411  *      Function fills the array of triples <key, p, bh> and returns %NULL
412  *      if everything went OK or the pointer to the last filled triple
413  *      (incomplete one) otherwise. Upon the return chain[i].key contains
414  *      the number of (i+1)-th block in the chain (as it is stored in memory,
415  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
416  *      number (it points into struct inode for i==0 and into the bh->b_data
417  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
418  *      block for i>0 and NULL for i==0. In other words, it holds the block
419  *      numbers of the chain, addresses they were taken from (and where we can
420  *      verify that chain did not change) and buffer_heads hosting these
421  *      numbers.
422  *
423  *      Function stops when it stumbles upon zero pointer (absent block)
424  *              (pointer to last triple returned, *@err == 0)
425  *      or when it gets an IO error reading an indirect block
426  *              (ditto, *@err == -EIO)
427  *      or when it reads all @depth-1 indirect blocks successfully and finds
428  *      the whole chain, all way to the data (returns %NULL, *err == 0).
429  *
430  *      Need to be called with
431  *      down_read(&EXT4_I(inode)->i_data_sem)
432  */
433 static Indirect *ext4_get_branch(struct inode *inode, int depth,
434                                  ext4_lblk_t  *offsets,
435                                  Indirect chain[4], int *err)
436 {
437         struct super_block *sb = inode->i_sb;
438         Indirect *p = chain;
439         struct buffer_head *bh;
440
441         *err = 0;
442         /* i_data is not going away, no lock needed */
443         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
444         if (!p->key)
445                 goto no_block;
446         while (--depth) {
447                 bh = sb_getblk(sb, le32_to_cpu(p->key));
448                 if (unlikely(!bh))
449                         goto failure;
450                   
451                 if (!bh_uptodate_or_lock(bh)) {
452                         if (bh_submit_read(bh) < 0) {
453                                 put_bh(bh);
454                                 goto failure;
455                         }
456                         /* validate block references */
457                         if (ext4_check_indirect_blockref(inode, bh)) {
458                                 put_bh(bh);
459                                 goto failure;
460                         }
461                 }
462                 
463                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
464                 /* Reader: end */
465                 if (!p->key)
466                         goto no_block;
467         }
468         return NULL;
469
470 failure:
471         *err = -EIO;
472 no_block:
473         return p;
474 }
475
476 /**
477  *      ext4_find_near - find a place for allocation with sufficient locality
478  *      @inode: owner
479  *      @ind: descriptor of indirect block.
480  *
481  *      This function returns the preferred place for block allocation.
482  *      It is used when heuristic for sequential allocation fails.
483  *      Rules are:
484  *        + if there is a block to the left of our position - allocate near it.
485  *        + if pointer will live in indirect block - allocate near that block.
486  *        + if pointer will live in inode - allocate in the same
487  *          cylinder group.
488  *
489  * In the latter case we colour the starting block by the callers PID to
490  * prevent it from clashing with concurrent allocations for a different inode
491  * in the same block group.   The PID is used here so that functionally related
492  * files will be close-by on-disk.
493  *
494  *      Caller must make sure that @ind is valid and will stay that way.
495  */
496 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
497 {
498         struct ext4_inode_info *ei = EXT4_I(inode);
499         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
500         __le32 *p;
501         ext4_fsblk_t bg_start;
502         ext4_fsblk_t last_block;
503         ext4_grpblk_t colour;
504         ext4_group_t block_group;
505         int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
506
507         /* Try to find previous block */
508         for (p = ind->p - 1; p >= start; p--) {
509                 if (*p)
510                         return le32_to_cpu(*p);
511         }
512
513         /* No such thing, so let's try location of indirect block */
514         if (ind->bh)
515                 return ind->bh->b_blocknr;
516
517         /*
518          * It is going to be referred to from the inode itself? OK, just put it
519          * into the same cylinder group then.
520          */
521         block_group = ei->i_block_group;
522         if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
523                 block_group &= ~(flex_size-1);
524                 if (S_ISREG(inode->i_mode))
525                         block_group++;
526         }
527         bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
528         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
529
530         /*
531          * If we are doing delayed allocation, we don't need take
532          * colour into account.
533          */
534         if (test_opt(inode->i_sb, DELALLOC))
535                 return bg_start;
536
537         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
538                 colour = (current->pid % 16) *
539                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
540         else
541                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
542         return bg_start + colour;
543 }
544
545 /**
546  *      ext4_find_goal - find a preferred place for allocation.
547  *      @inode: owner
548  *      @block:  block we want
549  *      @partial: pointer to the last triple within a chain
550  *
551  *      Normally this function find the preferred place for block allocation,
552  *      returns it.
553  */
554 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
555                 Indirect *partial)
556 {
557         /*
558          * XXX need to get goal block from mballoc's data structures
559          */
560
561         return ext4_find_near(inode, partial);
562 }
563
564 /**
565  *      ext4_blks_to_allocate: Look up the block map and count the number
566  *      of direct blocks need to be allocated for the given branch.
567  *
568  *      @branch: chain of indirect blocks
569  *      @k: number of blocks need for indirect blocks
570  *      @blks: number of data blocks to be mapped.
571  *      @blocks_to_boundary:  the offset in the indirect block
572  *
573  *      return the total number of blocks to be allocate, including the
574  *      direct and indirect blocks.
575  */
576 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
577                 int blocks_to_boundary)
578 {
579         unsigned int count = 0;
580
581         /*
582          * Simple case, [t,d]Indirect block(s) has not allocated yet
583          * then it's clear blocks on that path have not allocated
584          */
585         if (k > 0) {
586                 /* right now we don't handle cross boundary allocation */
587                 if (blks < blocks_to_boundary + 1)
588                         count += blks;
589                 else
590                         count += blocks_to_boundary + 1;
591                 return count;
592         }
593
594         count++;
595         while (count < blks && count <= blocks_to_boundary &&
596                 le32_to_cpu(*(branch[0].p + count)) == 0) {
597                 count++;
598         }
599         return count;
600 }
601
602 /**
603  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
604  *      @indirect_blks: the number of blocks need to allocate for indirect
605  *                      blocks
606  *
607  *      @new_blocks: on return it will store the new block numbers for
608  *      the indirect blocks(if needed) and the first direct block,
609  *      @blks:  on return it will store the total number of allocated
610  *              direct blocks
611  */
612 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
613                                 ext4_lblk_t iblock, ext4_fsblk_t goal,
614                                 int indirect_blks, int blks,
615                                 ext4_fsblk_t new_blocks[4], int *err)
616 {
617         struct ext4_allocation_request ar;
618         int target, i;
619         unsigned long count = 0, blk_allocated = 0;
620         int index = 0;
621         ext4_fsblk_t current_block = 0;
622         int ret = 0;
623
624         /*
625          * Here we try to allocate the requested multiple blocks at once,
626          * on a best-effort basis.
627          * To build a branch, we should allocate blocks for
628          * the indirect blocks(if not allocated yet), and at least
629          * the first direct block of this branch.  That's the
630          * minimum number of blocks need to allocate(required)
631          */
632         /* first we try to allocate the indirect blocks */
633         target = indirect_blks;
634         while (target > 0) {
635                 count = target;
636                 /* allocating blocks for indirect blocks and direct blocks */
637                 current_block = ext4_new_meta_blocks(handle, inode,
638                                                         goal, &count, err);
639                 if (*err)
640                         goto failed_out;
641
642                 target -= count;
643                 /* allocate blocks for indirect blocks */
644                 while (index < indirect_blks && count) {
645                         new_blocks[index++] = current_block++;
646                         count--;
647                 }
648                 if (count > 0) {
649                         /*
650                          * save the new block number
651                          * for the first direct block
652                          */
653                         new_blocks[index] = current_block;
654                         printk(KERN_INFO "%s returned more blocks than "
655                                                 "requested\n", __func__);
656                         WARN_ON(1);
657                         break;
658                 }
659         }
660
661         target = blks - count ;
662         blk_allocated = count;
663         if (!target)
664                 goto allocated;
665         /* Now allocate data blocks */
666         memset(&ar, 0, sizeof(ar));
667         ar.inode = inode;
668         ar.goal = goal;
669         ar.len = target;
670         ar.logical = iblock;
671         if (S_ISREG(inode->i_mode))
672                 /* enable in-core preallocation only for regular files */
673                 ar.flags = EXT4_MB_HINT_DATA;
674
675         current_block = ext4_mb_new_blocks(handle, &ar, err);
676
677         if (*err && (target == blks)) {
678                 /*
679                  * if the allocation failed and we didn't allocate
680                  * any blocks before
681                  */
682                 goto failed_out;
683         }
684         if (!*err) {
685                 if (target == blks) {
686                 /*
687                  * save the new block number
688                  * for the first direct block
689                  */
690                         new_blocks[index] = current_block;
691                 }
692                 blk_allocated += ar.len;
693         }
694 allocated:
695         /* total number of blocks allocated for direct blocks */
696         ret = blk_allocated;
697         *err = 0;
698         return ret;
699 failed_out:
700         for (i = 0; i < index; i++)
701                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
702         return ret;
703 }
704
705 /**
706  *      ext4_alloc_branch - allocate and set up a chain of blocks.
707  *      @inode: owner
708  *      @indirect_blks: number of allocated indirect blocks
709  *      @blks: number of allocated direct blocks
710  *      @offsets: offsets (in the blocks) to store the pointers to next.
711  *      @branch: place to store the chain in.
712  *
713  *      This function allocates blocks, zeroes out all but the last one,
714  *      links them into chain and (if we are synchronous) writes them to disk.
715  *      In other words, it prepares a branch that can be spliced onto the
716  *      inode. It stores the information about that chain in the branch[], in
717  *      the same format as ext4_get_branch() would do. We are calling it after
718  *      we had read the existing part of chain and partial points to the last
719  *      triple of that (one with zero ->key). Upon the exit we have the same
720  *      picture as after the successful ext4_get_block(), except that in one
721  *      place chain is disconnected - *branch->p is still zero (we did not
722  *      set the last link), but branch->key contains the number that should
723  *      be placed into *branch->p to fill that gap.
724  *
725  *      If allocation fails we free all blocks we've allocated (and forget
726  *      their buffer_heads) and return the error value the from failed
727  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
728  *      as described above and return 0.
729  */
730 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
731                                 ext4_lblk_t iblock, int indirect_blks,
732                                 int *blks, ext4_fsblk_t goal,
733                                 ext4_lblk_t *offsets, Indirect *branch)
734 {
735         int blocksize = inode->i_sb->s_blocksize;
736         int i, n = 0;
737         int err = 0;
738         struct buffer_head *bh;
739         int num;
740         ext4_fsblk_t new_blocks[4];
741         ext4_fsblk_t current_block;
742
743         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
744                                 *blks, new_blocks, &err);
745         if (err)
746                 return err;
747
748         branch[0].key = cpu_to_le32(new_blocks[0]);
749         /*
750          * metadata blocks and data blocks are allocated.
751          */
752         for (n = 1; n <= indirect_blks;  n++) {
753                 /*
754                  * Get buffer_head for parent block, zero it out
755                  * and set the pointer to new one, then send
756                  * parent to disk.
757                  */
758                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
759                 branch[n].bh = bh;
760                 lock_buffer(bh);
761                 BUFFER_TRACE(bh, "call get_create_access");
762                 err = ext4_journal_get_create_access(handle, bh);
763                 if (err) {
764                         unlock_buffer(bh);
765                         brelse(bh);
766                         goto failed;
767                 }
768
769                 memset(bh->b_data, 0, blocksize);
770                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
771                 branch[n].key = cpu_to_le32(new_blocks[n]);
772                 *branch[n].p = branch[n].key;
773                 if (n == indirect_blks) {
774                         current_block = new_blocks[n];
775                         /*
776                          * End of chain, update the last new metablock of
777                          * the chain to point to the new allocated
778                          * data blocks numbers
779                          */
780                         for (i=1; i < num; i++)
781                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
782                 }
783                 BUFFER_TRACE(bh, "marking uptodate");
784                 set_buffer_uptodate(bh);
785                 unlock_buffer(bh);
786
787                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
788                 err = ext4_handle_dirty_metadata(handle, inode, bh);
789                 if (err)
790                         goto failed;
791         }
792         *blks = num;
793         return err;
794 failed:
795         /* Allocation failed, free what we already allocated */
796         for (i = 1; i <= n ; i++) {
797                 BUFFER_TRACE(branch[i].bh, "call jbd2_journal_forget");
798                 ext4_journal_forget(handle, branch[i].bh);
799         }
800         for (i = 0; i < indirect_blks; i++)
801                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
802
803         ext4_free_blocks(handle, inode, new_blocks[i], num, 0);
804
805         return err;
806 }
807
808 /**
809  * ext4_splice_branch - splice the allocated branch onto inode.
810  * @inode: owner
811  * @block: (logical) number of block we are adding
812  * @chain: chain of indirect blocks (with a missing link - see
813  *      ext4_alloc_branch)
814  * @where: location of missing link
815  * @num:   number of indirect blocks we are adding
816  * @blks:  number of direct blocks we are adding
817  *
818  * This function fills the missing link and does all housekeeping needed in
819  * inode (->i_blocks, etc.). In case of success we end up with the full
820  * chain to new block and return 0.
821  */
822 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
823                         ext4_lblk_t block, Indirect *where, int num, int blks)
824 {
825         int i;
826         int err = 0;
827         ext4_fsblk_t current_block;
828
829         /*
830          * If we're splicing into a [td]indirect block (as opposed to the
831          * inode) then we need to get write access to the [td]indirect block
832          * before the splice.
833          */
834         if (where->bh) {
835                 BUFFER_TRACE(where->bh, "get_write_access");
836                 err = ext4_journal_get_write_access(handle, where->bh);
837                 if (err)
838                         goto err_out;
839         }
840         /* That's it */
841
842         *where->p = where->key;
843
844         /*
845          * Update the host buffer_head or inode to point to more just allocated
846          * direct blocks blocks
847          */
848         if (num == 0 && blks > 1) {
849                 current_block = le32_to_cpu(where->key) + 1;
850                 for (i = 1; i < blks; i++)
851                         *(where->p + i) = cpu_to_le32(current_block++);
852         }
853
854         /* We are done with atomic stuff, now do the rest of housekeeping */
855
856         inode->i_ctime = ext4_current_time(inode);
857         ext4_mark_inode_dirty(handle, inode);
858
859         /* had we spliced it onto indirect block? */
860         if (where->bh) {
861                 /*
862                  * If we spliced it onto an indirect block, we haven't
863                  * altered the inode.  Note however that if it is being spliced
864                  * onto an indirect block at the very end of the file (the
865                  * file is growing) then we *will* alter the inode to reflect
866                  * the new i_size.  But that is not done here - it is done in
867                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
868                  */
869                 jbd_debug(5, "splicing indirect only\n");
870                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
871                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
872                 if (err)
873                         goto err_out;
874         } else {
875                 /*
876                  * OK, we spliced it into the inode itself on a direct block.
877                  * Inode was dirtied above.
878                  */
879                 jbd_debug(5, "splicing direct\n");
880         }
881         return err;
882
883 err_out:
884         for (i = 1; i <= num; i++) {
885                 BUFFER_TRACE(where[i].bh, "call jbd2_journal_forget");
886                 ext4_journal_forget(handle, where[i].bh);
887                 ext4_free_blocks(handle, inode,
888                                         le32_to_cpu(where[i-1].key), 1, 0);
889         }
890         ext4_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks, 0);
891
892         return err;
893 }
894
895 /*
896  * The ext4_ind_get_blocks() function handles non-extents inodes
897  * (i.e., using the traditional indirect/double-indirect i_blocks
898  * scheme) for ext4_get_blocks().
899  *
900  * Allocation strategy is simple: if we have to allocate something, we will
901  * have to go the whole way to leaf. So let's do it before attaching anything
902  * to tree, set linkage between the newborn blocks, write them if sync is
903  * required, recheck the path, free and repeat if check fails, otherwise
904  * set the last missing link (that will protect us from any truncate-generated
905  * removals - all blocks on the path are immune now) and possibly force the
906  * write on the parent block.
907  * That has a nice additional property: no special recovery from the failed
908  * allocations is needed - we simply release blocks and do not touch anything
909  * reachable from inode.
910  *
911  * `handle' can be NULL if create == 0.
912  *
913  * return > 0, # of blocks mapped or allocated.
914  * return = 0, if plain lookup failed.
915  * return < 0, error case.
916  *
917  * The ext4_ind_get_blocks() function should be called with
918  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
919  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
920  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
921  * blocks.
922  */
923 static int ext4_ind_get_blocks(handle_t *handle, struct inode *inode,
924                                   ext4_lblk_t iblock, unsigned int maxblocks,
925                                   struct buffer_head *bh_result,
926                                   int flags)
927 {
928         int err = -EIO;
929         ext4_lblk_t offsets[4];
930         Indirect chain[4];
931         Indirect *partial;
932         ext4_fsblk_t goal;
933         int indirect_blks;
934         int blocks_to_boundary = 0;
935         int depth;
936         int count = 0;
937         ext4_fsblk_t first_block = 0;
938
939         J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
940         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
941         depth = ext4_block_to_path(inode, iblock, offsets,
942                                         &blocks_to_boundary);
943
944         if (depth == 0)
945                 goto out;
946
947         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
948
949         /* Simplest case - block found, no allocation needed */
950         if (!partial) {
951                 first_block = le32_to_cpu(chain[depth - 1].key);
952                 clear_buffer_new(bh_result);
953                 count++;
954                 /*map more blocks*/
955                 while (count < maxblocks && count <= blocks_to_boundary) {
956                         ext4_fsblk_t blk;
957
958                         blk = le32_to_cpu(*(chain[depth-1].p + count));
959
960                         if (blk == first_block + count)
961                                 count++;
962                         else
963                                 break;
964                 }
965                 goto got_it;
966         }
967
968         /* Next simple case - plain lookup or failed read of indirect block */
969         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
970                 goto cleanup;
971
972         /*
973          * Okay, we need to do block allocation.
974         */
975         goal = ext4_find_goal(inode, iblock, partial);
976
977         /* the number of blocks need to allocate for [d,t]indirect blocks */
978         indirect_blks = (chain + depth) - partial - 1;
979
980         /*
981          * Next look up the indirect map to count the totoal number of
982          * direct blocks to allocate for this branch.
983          */
984         count = ext4_blks_to_allocate(partial, indirect_blks,
985                                         maxblocks, blocks_to_boundary);
986         /*
987          * Block out ext4_truncate while we alter the tree
988          */
989         err = ext4_alloc_branch(handle, inode, iblock, indirect_blks,
990                                         &count, goal,
991                                         offsets + (partial - chain), partial);
992
993         /*
994          * The ext4_splice_branch call will free and forget any buffers
995          * on the new chain if there is a failure, but that risks using
996          * up transaction credits, especially for bitmaps where the
997          * credits cannot be returned.  Can we handle this somehow?  We
998          * may need to return -EAGAIN upwards in the worst case.  --sct
999          */
1000         if (!err)
1001                 err = ext4_splice_branch(handle, inode, iblock,
1002                                         partial, indirect_blks, count);
1003         else 
1004                 goto cleanup;
1005
1006         set_buffer_new(bh_result);
1007 got_it:
1008         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
1009         if (count > blocks_to_boundary)
1010                 set_buffer_boundary(bh_result);
1011         err = count;
1012         /* Clean up and exit */
1013         partial = chain + depth - 1;    /* the whole chain */
1014 cleanup:
1015         while (partial > chain) {
1016                 BUFFER_TRACE(partial->bh, "call brelse");
1017                 brelse(partial->bh);
1018                 partial--;
1019         }
1020         BUFFER_TRACE(bh_result, "returned");
1021 out:
1022         return err;
1023 }
1024
1025 qsize_t ext4_get_reserved_space(struct inode *inode)
1026 {
1027         unsigned long long total;
1028
1029         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1030         total = EXT4_I(inode)->i_reserved_data_blocks +
1031                 EXT4_I(inode)->i_reserved_meta_blocks;
1032         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1033
1034         return total;
1035 }
1036 /*
1037  * Calculate the number of metadata blocks need to reserve
1038  * to allocate @blocks for non extent file based file
1039  */
1040 static int ext4_indirect_calc_metadata_amount(struct inode *inode, int blocks)
1041 {
1042         int icap = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1043         int ind_blks, dind_blks, tind_blks;
1044
1045         /* number of new indirect blocks needed */
1046         ind_blks = (blocks + icap - 1) / icap;
1047
1048         dind_blks = (ind_blks + icap - 1) / icap;
1049
1050         tind_blks = 1;
1051
1052         return ind_blks + dind_blks + tind_blks;
1053 }
1054
1055 /*
1056  * Calculate the number of metadata blocks need to reserve
1057  * to allocate given number of blocks
1058  */
1059 static int ext4_calc_metadata_amount(struct inode *inode, int blocks)
1060 {
1061         if (!blocks)
1062                 return 0;
1063
1064         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
1065                 return ext4_ext_calc_metadata_amount(inode, blocks);
1066
1067         return ext4_indirect_calc_metadata_amount(inode, blocks);
1068 }
1069
1070 static void ext4_da_update_reserve_space(struct inode *inode, int used)
1071 {
1072         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1073         int total, mdb, mdb_free;
1074
1075         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1076         /* recalculate the number of metablocks still need to be reserved */
1077         total = EXT4_I(inode)->i_reserved_data_blocks - used;
1078         mdb = ext4_calc_metadata_amount(inode, total);
1079
1080         /* figure out how many metablocks to release */
1081         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1082         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1083
1084         if (mdb_free) {
1085                 /* Account for allocated meta_blocks */
1086                 mdb_free -= EXT4_I(inode)->i_allocated_meta_blocks;
1087
1088                 /* update fs dirty blocks counter */
1089                 percpu_counter_sub(&sbi->s_dirtyblocks_counter, mdb_free);
1090                 EXT4_I(inode)->i_allocated_meta_blocks = 0;
1091                 EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1092         }
1093
1094         /* update per-inode reservations */
1095         BUG_ON(used  > EXT4_I(inode)->i_reserved_data_blocks);
1096         EXT4_I(inode)->i_reserved_data_blocks -= used;
1097         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1098
1099         /*
1100          * free those over-booking quota for metadata blocks
1101          */
1102         if (mdb_free)
1103                 vfs_dq_release_reservation_block(inode, mdb_free);
1104
1105         /*
1106          * If we have done all the pending block allocations and if
1107          * there aren't any writers on the inode, we can discard the
1108          * inode's preallocations.
1109          */
1110         if (!total && (atomic_read(&inode->i_writecount) == 0))
1111                 ext4_discard_preallocations(inode);
1112 }
1113
1114 static int check_block_validity(struct inode *inode, sector_t logical,
1115                                 sector_t phys, int len)
1116 {
1117         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), phys, len)) {
1118                 ext4_error(inode->i_sb, "check_block_validity",
1119                            "inode #%lu logical block %llu mapped to %llu "
1120                            "(size %d)", inode->i_ino,
1121                            (unsigned long long) logical,
1122                            (unsigned long long) phys, len);
1123                 WARN_ON(1);
1124                 return -EIO;
1125         }
1126         return 0;
1127 }
1128
1129 /*
1130  * The ext4_get_blocks() function tries to look up the requested blocks,
1131  * and returns if the blocks are already mapped.
1132  *
1133  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1134  * and store the allocated blocks in the result buffer head and mark it
1135  * mapped.
1136  *
1137  * If file type is extents based, it will call ext4_ext_get_blocks(),
1138  * Otherwise, call with ext4_ind_get_blocks() to handle indirect mapping
1139  * based files
1140  *
1141  * On success, it returns the number of blocks being mapped or allocate.
1142  * if create==0 and the blocks are pre-allocated and uninitialized block,
1143  * the result buffer head is unmapped. If the create ==1, it will make sure
1144  * the buffer head is mapped.
1145  *
1146  * It returns 0 if plain look up failed (blocks have not been allocated), in
1147  * that casem, buffer head is unmapped
1148  *
1149  * It returns the error in case of allocation failure.
1150  */
1151 int ext4_get_blocks(handle_t *handle, struct inode *inode, sector_t block,
1152                     unsigned int max_blocks, struct buffer_head *bh,
1153                     int flags)
1154 {
1155         int retval;
1156
1157         clear_buffer_mapped(bh);
1158         clear_buffer_unwritten(bh);
1159
1160         /*
1161          * Try to see if we can get the block without requesting a new
1162          * file system block.
1163          */
1164         down_read((&EXT4_I(inode)->i_data_sem));
1165         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1166                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1167                                 bh, 0);
1168         } else {
1169                 retval = ext4_ind_get_blocks(handle, inode, block, max_blocks,
1170                                              bh, 0);
1171         }
1172         up_read((&EXT4_I(inode)->i_data_sem));
1173
1174         if (retval > 0 && buffer_mapped(bh)) {
1175                 int ret = check_block_validity(inode, block, 
1176                                                bh->b_blocknr, retval);
1177                 if (ret != 0)
1178                         return ret;
1179         }
1180
1181         /* If it is only a block(s) look up */
1182         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1183                 return retval;
1184
1185         /*
1186          * Returns if the blocks have already allocated
1187          *
1188          * Note that if blocks have been preallocated
1189          * ext4_ext_get_block() returns th create = 0
1190          * with buffer head unmapped.
1191          */
1192         if (retval > 0 && buffer_mapped(bh))
1193                 return retval;
1194
1195         /*
1196          * When we call get_blocks without the create flag, the
1197          * BH_Unwritten flag could have gotten set if the blocks
1198          * requested were part of a uninitialized extent.  We need to
1199          * clear this flag now that we are committed to convert all or
1200          * part of the uninitialized extent to be an initialized
1201          * extent.  This is because we need to avoid the combination
1202          * of BH_Unwritten and BH_Mapped flags being simultaneously
1203          * set on the buffer_head.
1204          */
1205         clear_buffer_unwritten(bh);
1206
1207         /*
1208          * New blocks allocate and/or writing to uninitialized extent
1209          * will possibly result in updating i_data, so we take
1210          * the write lock of i_data_sem, and call get_blocks()
1211          * with create == 1 flag.
1212          */
1213         down_write((&EXT4_I(inode)->i_data_sem));
1214
1215         /*
1216          * if the caller is from delayed allocation writeout path
1217          * we have already reserved fs blocks for allocation
1218          * let the underlying get_block() function know to
1219          * avoid double accounting
1220          */
1221         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1222                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1223         /*
1224          * We need to check for EXT4 here because migrate
1225          * could have changed the inode type in between
1226          */
1227         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1228                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1229                                               bh, flags);
1230         } else {
1231                 retval = ext4_ind_get_blocks(handle, inode, block,
1232                                              max_blocks, bh, flags);
1233
1234                 if (retval > 0 && buffer_new(bh)) {
1235                         /*
1236                          * We allocated new blocks which will result in
1237                          * i_data's format changing.  Force the migrate
1238                          * to fail by clearing migrate flags
1239                          */
1240                         EXT4_I(inode)->i_flags = EXT4_I(inode)->i_flags &
1241                                                         ~EXT4_EXT_MIGRATE;
1242                 }
1243         }
1244
1245         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1246                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1247
1248         /*
1249          * Update reserved blocks/metadata blocks after successful
1250          * block allocation which had been deferred till now.
1251          */
1252         if ((retval > 0) && (flags & EXT4_GET_BLOCKS_UPDATE_RESERVE_SPACE))
1253                 ext4_da_update_reserve_space(inode, retval);
1254
1255         up_write((&EXT4_I(inode)->i_data_sem));
1256         if (retval > 0 && buffer_mapped(bh)) {
1257                 int ret = check_block_validity(inode, block, 
1258                                                bh->b_blocknr, retval);
1259                 if (ret != 0)
1260                         return ret;
1261         }
1262         return retval;
1263 }
1264
1265 /* Maximum number of blocks we map for direct IO at once. */
1266 #define DIO_MAX_BLOCKS 4096
1267
1268 int ext4_get_block(struct inode *inode, sector_t iblock,
1269                    struct buffer_head *bh_result, int create)
1270 {
1271         handle_t *handle = ext4_journal_current_handle();
1272         int ret = 0, started = 0;
1273         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1274         int dio_credits;
1275
1276         if (create && !handle) {
1277                 /* Direct IO write... */
1278                 if (max_blocks > DIO_MAX_BLOCKS)
1279                         max_blocks = DIO_MAX_BLOCKS;
1280                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
1281                 handle = ext4_journal_start(inode, dio_credits);
1282                 if (IS_ERR(handle)) {
1283                         ret = PTR_ERR(handle);
1284                         goto out;
1285                 }
1286                 started = 1;
1287         }
1288
1289         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
1290                               create ? EXT4_GET_BLOCKS_CREATE : 0);
1291         if (ret > 0) {
1292                 bh_result->b_size = (ret << inode->i_blkbits);
1293                 ret = 0;
1294         }
1295         if (started)
1296                 ext4_journal_stop(handle);
1297 out:
1298         return ret;
1299 }
1300
1301 /*
1302  * `handle' can be NULL if create is zero
1303  */
1304 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1305                                 ext4_lblk_t block, int create, int *errp)
1306 {
1307         struct buffer_head dummy;
1308         int fatal = 0, err;
1309         int flags = 0;
1310
1311         J_ASSERT(handle != NULL || create == 0);
1312
1313         dummy.b_state = 0;
1314         dummy.b_blocknr = -1000;
1315         buffer_trace_init(&dummy.b_history);
1316         if (create)
1317                 flags |= EXT4_GET_BLOCKS_CREATE;
1318         err = ext4_get_blocks(handle, inode, block, 1, &dummy, flags);
1319         /*
1320          * ext4_get_blocks() returns number of blocks mapped. 0 in
1321          * case of a HOLE.
1322          */
1323         if (err > 0) {
1324                 if (err > 1)
1325                         WARN_ON(1);
1326                 err = 0;
1327         }
1328         *errp = err;
1329         if (!err && buffer_mapped(&dummy)) {
1330                 struct buffer_head *bh;
1331                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1332                 if (!bh) {
1333                         *errp = -EIO;
1334                         goto err;
1335                 }
1336                 if (buffer_new(&dummy)) {
1337                         J_ASSERT(create != 0);
1338                         J_ASSERT(handle != NULL);
1339
1340                         /*
1341                          * Now that we do not always journal data, we should
1342                          * keep in mind whether this should always journal the
1343                          * new buffer as metadata.  For now, regular file
1344                          * writes use ext4_get_block instead, so it's not a
1345                          * problem.
1346                          */
1347                         lock_buffer(bh);
1348                         BUFFER_TRACE(bh, "call get_create_access");
1349                         fatal = ext4_journal_get_create_access(handle, bh);
1350                         if (!fatal && !buffer_uptodate(bh)) {
1351                                 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1352                                 set_buffer_uptodate(bh);
1353                         }
1354                         unlock_buffer(bh);
1355                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1356                         err = ext4_handle_dirty_metadata(handle, inode, bh);
1357                         if (!fatal)
1358                                 fatal = err;
1359                 } else {
1360                         BUFFER_TRACE(bh, "not a new buffer");
1361                 }
1362                 if (fatal) {
1363                         *errp = fatal;
1364                         brelse(bh);
1365                         bh = NULL;
1366                 }
1367                 return bh;
1368         }
1369 err:
1370         return NULL;
1371 }
1372
1373 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1374                                ext4_lblk_t block, int create, int *err)
1375 {
1376         struct buffer_head *bh;
1377
1378         bh = ext4_getblk(handle, inode, block, create, err);
1379         if (!bh)
1380                 return bh;
1381         if (buffer_uptodate(bh))
1382                 return bh;
1383         ll_rw_block(READ_META, 1, &bh);
1384         wait_on_buffer(bh);
1385         if (buffer_uptodate(bh))
1386                 return bh;
1387         put_bh(bh);
1388         *err = -EIO;
1389         return NULL;
1390 }
1391
1392 static int walk_page_buffers(handle_t *handle,
1393                              struct buffer_head *head,
1394                              unsigned from,
1395                              unsigned to,
1396                              int *partial,
1397                              int (*fn)(handle_t *handle,
1398                                        struct buffer_head *bh))
1399 {
1400         struct buffer_head *bh;
1401         unsigned block_start, block_end;
1402         unsigned blocksize = head->b_size;
1403         int err, ret = 0;
1404         struct buffer_head *next;
1405
1406         for (bh = head, block_start = 0;
1407              ret == 0 && (bh != head || !block_start);
1408              block_start = block_end, bh = next)
1409         {
1410                 next = bh->b_this_page;
1411                 block_end = block_start + blocksize;
1412                 if (block_end <= from || block_start >= to) {
1413                         if (partial && !buffer_uptodate(bh))
1414                                 *partial = 1;
1415                         continue;
1416                 }
1417                 err = (*fn)(handle, bh);
1418                 if (!ret)
1419                         ret = err;
1420         }
1421         return ret;
1422 }
1423
1424 /*
1425  * To preserve ordering, it is essential that the hole instantiation and
1426  * the data write be encapsulated in a single transaction.  We cannot
1427  * close off a transaction and start a new one between the ext4_get_block()
1428  * and the commit_write().  So doing the jbd2_journal_start at the start of
1429  * prepare_write() is the right place.
1430  *
1431  * Also, this function can nest inside ext4_writepage() ->
1432  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1433  * has generated enough buffer credits to do the whole page.  So we won't
1434  * block on the journal in that case, which is good, because the caller may
1435  * be PF_MEMALLOC.
1436  *
1437  * By accident, ext4 can be reentered when a transaction is open via
1438  * quota file writes.  If we were to commit the transaction while thus
1439  * reentered, there can be a deadlock - we would be holding a quota
1440  * lock, and the commit would never complete if another thread had a
1441  * transaction open and was blocking on the quota lock - a ranking
1442  * violation.
1443  *
1444  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1445  * will _not_ run commit under these circumstances because handle->h_ref
1446  * is elevated.  We'll still have enough credits for the tiny quotafile
1447  * write.
1448  */
1449 static int do_journal_get_write_access(handle_t *handle,
1450                                         struct buffer_head *bh)
1451 {
1452         if (!buffer_mapped(bh) || buffer_freed(bh))
1453                 return 0;
1454         return ext4_journal_get_write_access(handle, bh);
1455 }
1456
1457 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1458                                 loff_t pos, unsigned len, unsigned flags,
1459                                 struct page **pagep, void **fsdata)
1460 {
1461         struct inode *inode = mapping->host;
1462         int ret, needed_blocks;
1463         handle_t *handle;
1464         int retries = 0;
1465         struct page *page;
1466         pgoff_t index;
1467         unsigned from, to;
1468
1469         trace_mark(ext4_write_begin,
1470                    "dev %s ino %lu pos %llu len %u flags %u",
1471                    inode->i_sb->s_id, inode->i_ino,
1472                    (unsigned long long) pos, len, flags);
1473         /*
1474          * Reserve one block more for addition to orphan list in case
1475          * we allocate blocks but write fails for some reason
1476          */
1477         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1478         index = pos >> PAGE_CACHE_SHIFT;
1479         from = pos & (PAGE_CACHE_SIZE - 1);
1480         to = from + len;
1481
1482 retry:
1483         handle = ext4_journal_start(inode, needed_blocks);
1484         if (IS_ERR(handle)) {
1485                 ret = PTR_ERR(handle);
1486                 goto out;
1487         }
1488
1489         /* We cannot recurse into the filesystem as the transaction is already
1490          * started */
1491         flags |= AOP_FLAG_NOFS;
1492
1493         page = grab_cache_page_write_begin(mapping, index, flags);
1494         if (!page) {
1495                 ext4_journal_stop(handle);
1496                 ret = -ENOMEM;
1497                 goto out;
1498         }
1499         *pagep = page;
1500
1501         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1502                                 ext4_get_block);
1503
1504         if (!ret && ext4_should_journal_data(inode)) {
1505                 ret = walk_page_buffers(handle, page_buffers(page),
1506                                 from, to, NULL, do_journal_get_write_access);
1507         }
1508
1509         if (ret) {
1510                 unlock_page(page);
1511                 page_cache_release(page);
1512                 /*
1513                  * block_write_begin may have instantiated a few blocks
1514                  * outside i_size.  Trim these off again. Don't need
1515                  * i_size_read because we hold i_mutex.
1516                  *
1517                  * Add inode to orphan list in case we crash before
1518                  * truncate finishes
1519                  */
1520                 if (pos + len > inode->i_size)
1521                         ext4_orphan_add(handle, inode);
1522
1523                 ext4_journal_stop(handle);
1524                 if (pos + len > inode->i_size) {
1525                         vmtruncate(inode, inode->i_size);
1526                         /* 
1527                          * If vmtruncate failed early the inode might
1528                          * still be on the orphan list; we need to
1529                          * make sure the inode is removed from the
1530                          * orphan list in that case.
1531                          */
1532                         if (inode->i_nlink)
1533                                 ext4_orphan_del(NULL, inode);
1534                 }
1535         }
1536
1537         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1538                 goto retry;
1539 out:
1540         return ret;
1541 }
1542
1543 /* For write_end() in data=journal mode */
1544 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1545 {
1546         if (!buffer_mapped(bh) || buffer_freed(bh))
1547                 return 0;
1548         set_buffer_uptodate(bh);
1549         return ext4_handle_dirty_metadata(handle, NULL, bh);
1550 }
1551
1552 static int ext4_generic_write_end(struct file *file,
1553                                 struct address_space *mapping,
1554                                 loff_t pos, unsigned len, unsigned copied,
1555                                 struct page *page, void *fsdata)
1556 {
1557         int i_size_changed = 0;
1558         struct inode *inode = mapping->host;
1559         handle_t *handle = ext4_journal_current_handle();
1560
1561         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1562
1563         /*
1564          * No need to use i_size_read() here, the i_size
1565          * cannot change under us because we hold i_mutex.
1566          *
1567          * But it's important to update i_size while still holding page lock:
1568          * page writeout could otherwise come in and zero beyond i_size.
1569          */
1570         if (pos + copied > inode->i_size) {
1571                 i_size_write(inode, pos + copied);
1572                 i_size_changed = 1;
1573         }
1574
1575         if (pos + copied >  EXT4_I(inode)->i_disksize) {
1576                 /* We need to mark inode dirty even if
1577                  * new_i_size is less that inode->i_size
1578                  * bu greater than i_disksize.(hint delalloc)
1579                  */
1580                 ext4_update_i_disksize(inode, (pos + copied));
1581                 i_size_changed = 1;
1582         }
1583         unlock_page(page);
1584         page_cache_release(page);
1585
1586         /*
1587          * Don't mark the inode dirty under page lock. First, it unnecessarily
1588          * makes the holding time of page lock longer. Second, it forces lock
1589          * ordering of page lock and transaction start for journaling
1590          * filesystems.
1591          */
1592         if (i_size_changed)
1593                 ext4_mark_inode_dirty(handle, inode);
1594
1595         return copied;
1596 }
1597
1598 /*
1599  * We need to pick up the new inode size which generic_commit_write gave us
1600  * `file' can be NULL - eg, when called from page_symlink().
1601  *
1602  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1603  * buffers are managed internally.
1604  */
1605 static int ext4_ordered_write_end(struct file *file,
1606                                 struct address_space *mapping,
1607                                 loff_t pos, unsigned len, unsigned copied,
1608                                 struct page *page, void *fsdata)
1609 {
1610         handle_t *handle = ext4_journal_current_handle();
1611         struct inode *inode = mapping->host;
1612         int ret = 0, ret2;
1613
1614         trace_mark(ext4_ordered_write_end,
1615                    "dev %s ino %lu pos %llu len %u copied %u",
1616                    inode->i_sb->s_id, inode->i_ino,
1617                    (unsigned long long) pos, len, copied);
1618         ret = ext4_jbd2_file_inode(handle, inode);
1619
1620         if (ret == 0) {
1621                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1622                                                         page, fsdata);
1623                 copied = ret2;
1624                 if (pos + len > inode->i_size)
1625                         /* if we have allocated more blocks and copied
1626                          * less. We will have blocks allocated outside
1627                          * inode->i_size. So truncate them
1628                          */
1629                         ext4_orphan_add(handle, inode);
1630                 if (ret2 < 0)
1631                         ret = ret2;
1632         }
1633         ret2 = ext4_journal_stop(handle);
1634         if (!ret)
1635                 ret = ret2;
1636
1637         if (pos + len > inode->i_size) {
1638                 vmtruncate(inode, inode->i_size);
1639                 /* 
1640                  * If vmtruncate failed early the inode might still be
1641                  * on the orphan list; we need to make sure the inode
1642                  * is removed from the orphan list in that case.
1643                  */
1644                 if (inode->i_nlink)
1645                         ext4_orphan_del(NULL, inode);
1646         }
1647
1648
1649         return ret ? ret : copied;
1650 }
1651
1652 static int ext4_writeback_write_end(struct file *file,
1653                                 struct address_space *mapping,
1654                                 loff_t pos, unsigned len, unsigned copied,
1655                                 struct page *page, void *fsdata)
1656 {
1657         handle_t *handle = ext4_journal_current_handle();
1658         struct inode *inode = mapping->host;
1659         int ret = 0, ret2;
1660
1661         trace_mark(ext4_writeback_write_end,
1662                    "dev %s ino %lu pos %llu len %u copied %u",
1663                    inode->i_sb->s_id, inode->i_ino,
1664                    (unsigned long long) pos, len, copied);
1665         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1666                                                         page, fsdata);
1667         copied = ret2;
1668         if (pos + len > inode->i_size)
1669                 /* if we have allocated more blocks and copied
1670                  * less. We will have blocks allocated outside
1671                  * inode->i_size. So truncate them
1672                  */
1673                 ext4_orphan_add(handle, inode);
1674
1675         if (ret2 < 0)
1676                 ret = ret2;
1677
1678         ret2 = ext4_journal_stop(handle);
1679         if (!ret)
1680                 ret = ret2;
1681
1682         if (pos + len > inode->i_size) {
1683                 vmtruncate(inode, inode->i_size);
1684                 /* 
1685                  * If vmtruncate failed early the inode might still be
1686                  * on the orphan list; we need to make sure the inode
1687                  * is removed from the orphan list in that case.
1688                  */
1689                 if (inode->i_nlink)
1690                         ext4_orphan_del(NULL, inode);
1691         }
1692
1693         return ret ? ret : copied;
1694 }
1695
1696 static int ext4_journalled_write_end(struct file *file,
1697                                 struct address_space *mapping,
1698                                 loff_t pos, unsigned len, unsigned copied,
1699                                 struct page *page, void *fsdata)
1700 {
1701         handle_t *handle = ext4_journal_current_handle();
1702         struct inode *inode = mapping->host;
1703         int ret = 0, ret2;
1704         int partial = 0;
1705         unsigned from, to;
1706         loff_t new_i_size;
1707
1708         trace_mark(ext4_journalled_write_end,
1709                    "dev %s ino %lu pos %llu len %u copied %u",
1710                    inode->i_sb->s_id, inode->i_ino,
1711                    (unsigned long long) pos, len, copied);
1712         from = pos & (PAGE_CACHE_SIZE - 1);
1713         to = from + len;
1714
1715         if (copied < len) {
1716                 if (!PageUptodate(page))
1717                         copied = 0;
1718                 page_zero_new_buffers(page, from+copied, to);
1719         }
1720
1721         ret = walk_page_buffers(handle, page_buffers(page), from,
1722                                 to, &partial, write_end_fn);
1723         if (!partial)
1724                 SetPageUptodate(page);
1725         new_i_size = pos + copied;
1726         if (new_i_size > inode->i_size)
1727                 i_size_write(inode, pos+copied);
1728         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
1729         if (new_i_size > EXT4_I(inode)->i_disksize) {
1730                 ext4_update_i_disksize(inode, new_i_size);
1731                 ret2 = ext4_mark_inode_dirty(handle, inode);
1732                 if (!ret)
1733                         ret = ret2;
1734         }
1735
1736         unlock_page(page);
1737         page_cache_release(page);
1738         if (pos + len > inode->i_size)
1739                 /* if we have allocated more blocks and copied
1740                  * less. We will have blocks allocated outside
1741                  * inode->i_size. So truncate them
1742                  */
1743                 ext4_orphan_add(handle, inode);
1744
1745         ret2 = ext4_journal_stop(handle);
1746         if (!ret)
1747                 ret = ret2;
1748         if (pos + len > inode->i_size) {
1749                 vmtruncate(inode, inode->i_size);
1750                 /* 
1751                  * If vmtruncate failed early the inode might still be
1752                  * on the orphan list; we need to make sure the inode
1753                  * is removed from the orphan list in that case.
1754                  */
1755                 if (inode->i_nlink)
1756                         ext4_orphan_del(NULL, inode);
1757         }
1758
1759         return ret ? ret : copied;
1760 }
1761
1762 static int ext4_da_reserve_space(struct inode *inode, int nrblocks)
1763 {
1764         int retries = 0;
1765         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1766         unsigned long md_needed, mdblocks, total = 0;
1767
1768         /*
1769          * recalculate the amount of metadata blocks to reserve
1770          * in order to allocate nrblocks
1771          * worse case is one extent per block
1772          */
1773 repeat:
1774         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1775         total = EXT4_I(inode)->i_reserved_data_blocks + nrblocks;
1776         mdblocks = ext4_calc_metadata_amount(inode, total);
1777         BUG_ON(mdblocks < EXT4_I(inode)->i_reserved_meta_blocks);
1778
1779         md_needed = mdblocks - EXT4_I(inode)->i_reserved_meta_blocks;
1780         total = md_needed + nrblocks;
1781
1782         /*
1783          * Make quota reservation here to prevent quota overflow
1784          * later. Real quota accounting is done at pages writeout
1785          * time.
1786          */
1787         if (vfs_dq_reserve_block(inode, total)) {
1788                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1789                 return -EDQUOT;
1790         }
1791
1792         if (ext4_claim_free_blocks(sbi, total)) {
1793                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1794                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1795                         yield();
1796                         goto repeat;
1797                 }
1798                 vfs_dq_release_reservation_block(inode, total);
1799                 return -ENOSPC;
1800         }
1801         EXT4_I(inode)->i_reserved_data_blocks += nrblocks;
1802         EXT4_I(inode)->i_reserved_meta_blocks = mdblocks;
1803
1804         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1805         return 0;       /* success */
1806 }
1807
1808 static void ext4_da_release_space(struct inode *inode, int to_free)
1809 {
1810         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1811         int total, mdb, mdb_free, release;
1812
1813         if (!to_free)
1814                 return;         /* Nothing to release, exit */
1815
1816         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1817
1818         if (!EXT4_I(inode)->i_reserved_data_blocks) {
1819                 /*
1820                  * if there is no reserved blocks, but we try to free some
1821                  * then the counter is messed up somewhere.
1822                  * but since this function is called from invalidate
1823                  * page, it's harmless to return without any action
1824                  */
1825                 printk(KERN_INFO "ext4 delalloc try to release %d reserved "
1826                             "blocks for inode %lu, but there is no reserved "
1827                             "data blocks\n", to_free, inode->i_ino);
1828                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1829                 return;
1830         }
1831
1832         /* recalculate the number of metablocks still need to be reserved */
1833         total = EXT4_I(inode)->i_reserved_data_blocks - to_free;
1834         mdb = ext4_calc_metadata_amount(inode, total);
1835
1836         /* figure out how many metablocks to release */
1837         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1838         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1839
1840         release = to_free + mdb_free;
1841
1842         /* update fs dirty blocks counter for truncate case */
1843         percpu_counter_sub(&sbi->s_dirtyblocks_counter, release);
1844
1845         /* update per-inode reservations */
1846         BUG_ON(to_free > EXT4_I(inode)->i_reserved_data_blocks);
1847         EXT4_I(inode)->i_reserved_data_blocks -= to_free;
1848
1849         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1850         EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1851         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1852
1853         vfs_dq_release_reservation_block(inode, release);
1854 }
1855
1856 static void ext4_da_page_release_reservation(struct page *page,
1857                                                 unsigned long offset)
1858 {
1859         int to_release = 0;
1860         struct buffer_head *head, *bh;
1861         unsigned int curr_off = 0;
1862
1863         head = page_buffers(page);
1864         bh = head;
1865         do {
1866                 unsigned int next_off = curr_off + bh->b_size;
1867
1868                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1869                         to_release++;
1870                         clear_buffer_delay(bh);
1871                 }
1872                 curr_off = next_off;
1873         } while ((bh = bh->b_this_page) != head);
1874         ext4_da_release_space(page->mapping->host, to_release);
1875 }
1876
1877 /*
1878  * Delayed allocation stuff
1879  */
1880
1881 struct mpage_da_data {
1882         struct inode *inode;
1883         sector_t b_blocknr;             /* start block number of extent */
1884         size_t b_size;                  /* size of extent */
1885         unsigned long b_state;          /* state of the extent */
1886         unsigned long first_page, next_page;    /* extent of pages */
1887         struct writeback_control *wbc;
1888         int io_done;
1889         int pages_written;
1890         int retval;
1891 };
1892
1893 /*
1894  * mpage_da_submit_io - walks through extent of pages and try to write
1895  * them with writepage() call back
1896  *
1897  * @mpd->inode: inode
1898  * @mpd->first_page: first page of the extent
1899  * @mpd->next_page: page after the last page of the extent
1900  *
1901  * By the time mpage_da_submit_io() is called we expect all blocks
1902  * to be allocated. this may be wrong if allocation failed.
1903  *
1904  * As pages are already locked by write_cache_pages(), we can't use it
1905  */
1906 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1907 {
1908         long pages_skipped;
1909         struct pagevec pvec;
1910         unsigned long index, end;
1911         int ret = 0, err, nr_pages, i;
1912         struct inode *inode = mpd->inode;
1913         struct address_space *mapping = inode->i_mapping;
1914
1915         BUG_ON(mpd->next_page <= mpd->first_page);
1916         /*
1917          * We need to start from the first_page to the next_page - 1
1918          * to make sure we also write the mapped dirty buffer_heads.
1919          * If we look at mpd->b_blocknr we would only be looking
1920          * at the currently mapped buffer_heads.
1921          */
1922         index = mpd->first_page;
1923         end = mpd->next_page - 1;
1924
1925         pagevec_init(&pvec, 0);
1926         while (index <= end) {
1927                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1928                 if (nr_pages == 0)
1929                         break;
1930                 for (i = 0; i < nr_pages; i++) {
1931                         struct page *page = pvec.pages[i];
1932
1933                         index = page->index;
1934                         if (index > end)
1935                                 break;
1936                         index++;
1937
1938                         BUG_ON(!PageLocked(page));
1939                         BUG_ON(PageWriteback(page));
1940
1941                         pages_skipped = mpd->wbc->pages_skipped;
1942                         err = mapping->a_ops->writepage(page, mpd->wbc);
1943                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
1944                                 /*
1945                                  * have successfully written the page
1946                                  * without skipping the same
1947                                  */
1948                                 mpd->pages_written++;
1949                         /*
1950                          * In error case, we have to continue because
1951                          * remaining pages are still locked
1952                          * XXX: unlock and re-dirty them?
1953                          */
1954                         if (ret == 0)
1955                                 ret = err;
1956                 }
1957                 pagevec_release(&pvec);
1958         }
1959         return ret;
1960 }
1961
1962 /*
1963  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
1964  *
1965  * @mpd->inode - inode to walk through
1966  * @exbh->b_blocknr - first block on a disk
1967  * @exbh->b_size - amount of space in bytes
1968  * @logical - first logical block to start assignment with
1969  *
1970  * the function goes through all passed space and put actual disk
1971  * block numbers into buffer heads, dropping BH_Delay and BH_Unwritten
1972  */
1973 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
1974                                  struct buffer_head *exbh)
1975 {
1976         struct inode *inode = mpd->inode;
1977         struct address_space *mapping = inode->i_mapping;
1978         int blocks = exbh->b_size >> inode->i_blkbits;
1979         sector_t pblock = exbh->b_blocknr, cur_logical;
1980         struct buffer_head *head, *bh;
1981         pgoff_t index, end;
1982         struct pagevec pvec;
1983         int nr_pages, i;
1984
1985         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1986         end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1987         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1988
1989         pagevec_init(&pvec, 0);
1990
1991         while (index <= end) {
1992                 /* XXX: optimize tail */
1993                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1994                 if (nr_pages == 0)
1995                         break;
1996                 for (i = 0; i < nr_pages; i++) {
1997                         struct page *page = pvec.pages[i];
1998
1999                         index = page->index;
2000                         if (index > end)
2001                                 break;
2002                         index++;
2003
2004                         BUG_ON(!PageLocked(page));
2005                         BUG_ON(PageWriteback(page));
2006                         BUG_ON(!page_has_buffers(page));
2007
2008                         bh = page_buffers(page);
2009                         head = bh;
2010
2011                         /* skip blocks out of the range */
2012                         do {
2013                                 if (cur_logical >= logical)
2014                                         break;
2015                                 cur_logical++;
2016                         } while ((bh = bh->b_this_page) != head);
2017
2018                         do {
2019                                 if (cur_logical >= logical + blocks)
2020                                         break;
2021
2022                                 if (buffer_delay(bh) ||
2023                                                 buffer_unwritten(bh)) {
2024
2025                                         BUG_ON(bh->b_bdev != inode->i_sb->s_bdev);
2026
2027                                         if (buffer_delay(bh)) {
2028                                                 clear_buffer_delay(bh);
2029                                                 bh->b_blocknr = pblock;
2030                                         } else {
2031                                                 /*
2032                                                  * unwritten already should have
2033                                                  * blocknr assigned. Verify that
2034                                                  */
2035                                                 clear_buffer_unwritten(bh);
2036                                                 BUG_ON(bh->b_blocknr != pblock);
2037                                         }
2038
2039                                 } else if (buffer_mapped(bh))
2040                                         BUG_ON(bh->b_blocknr != pblock);
2041
2042                                 cur_logical++;
2043                                 pblock++;
2044                         } while ((bh = bh->b_this_page) != head);
2045                 }
2046                 pagevec_release(&pvec);
2047         }
2048 }
2049
2050
2051 /*
2052  * __unmap_underlying_blocks - just a helper function to unmap
2053  * set of blocks described by @bh
2054  */
2055 static inline void __unmap_underlying_blocks(struct inode *inode,
2056                                              struct buffer_head *bh)
2057 {
2058         struct block_device *bdev = inode->i_sb->s_bdev;
2059         int blocks, i;
2060
2061         blocks = bh->b_size >> inode->i_blkbits;
2062         for (i = 0; i < blocks; i++)
2063                 unmap_underlying_metadata(bdev, bh->b_blocknr + i);
2064 }
2065
2066 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2067                                         sector_t logical, long blk_cnt)
2068 {
2069         int nr_pages, i;
2070         pgoff_t index, end;
2071         struct pagevec pvec;
2072         struct inode *inode = mpd->inode;
2073         struct address_space *mapping = inode->i_mapping;
2074
2075         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2076         end   = (logical + blk_cnt - 1) >>
2077                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2078         while (index <= end) {
2079                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2080                 if (nr_pages == 0)
2081                         break;
2082                 for (i = 0; i < nr_pages; i++) {
2083                         struct page *page = pvec.pages[i];
2084                         index = page->index;
2085                         if (index > end)
2086                                 break;
2087                         index++;
2088
2089                         BUG_ON(!PageLocked(page));
2090                         BUG_ON(PageWriteback(page));
2091                         block_invalidatepage(page, 0);
2092                         ClearPageUptodate(page);
2093                         unlock_page(page);
2094                 }
2095         }
2096         return;
2097 }
2098
2099 static void ext4_print_free_blocks(struct inode *inode)
2100 {
2101         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2102         printk(KERN_EMERG "Total free blocks count %lld\n",
2103                         ext4_count_free_blocks(inode->i_sb));
2104         printk(KERN_EMERG "Free/Dirty block details\n");
2105         printk(KERN_EMERG "free_blocks=%lld\n",
2106                         (long long)percpu_counter_sum(&sbi->s_freeblocks_counter));
2107         printk(KERN_EMERG "dirty_blocks=%lld\n",
2108                         (long long)percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2109         printk(KERN_EMERG "Block reservation details\n");
2110         printk(KERN_EMERG "i_reserved_data_blocks=%u\n",
2111                         EXT4_I(inode)->i_reserved_data_blocks);
2112         printk(KERN_EMERG "i_reserved_meta_blocks=%u\n",
2113                         EXT4_I(inode)->i_reserved_meta_blocks);
2114         return;
2115 }
2116
2117 /*
2118  * mpage_da_map_blocks - go through given space
2119  *
2120  * @mpd - bh describing space
2121  *
2122  * The function skips space we know is already mapped to disk blocks.
2123  *
2124  */
2125 static int mpage_da_map_blocks(struct mpage_da_data *mpd)
2126 {
2127         int err, blks, get_blocks_flags;
2128         struct buffer_head new;
2129         sector_t next = mpd->b_blocknr;
2130         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2131         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2132         handle_t *handle = NULL;
2133
2134         /*
2135          * We consider only non-mapped and non-allocated blocks
2136          */
2137         if ((mpd->b_state  & (1 << BH_Mapped)) &&
2138                 !(mpd->b_state & (1 << BH_Delay)) &&
2139                 !(mpd->b_state & (1 << BH_Unwritten)))
2140                 return 0;
2141
2142         /*
2143          * If we didn't accumulate anything to write simply return
2144          */
2145         if (!mpd->b_size)
2146                 return 0;
2147
2148         handle = ext4_journal_current_handle();
2149         BUG_ON(!handle);
2150
2151         /*
2152          * Call ext4_get_blocks() to allocate any delayed allocation
2153          * blocks, or to convert an uninitialized extent to be
2154          * initialized (in the case where we have written into
2155          * one or more preallocated blocks).
2156          *
2157          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2158          * indicate that we are on the delayed allocation path.  This
2159          * affects functions in many different parts of the allocation
2160          * call path.  This flag exists primarily because we don't
2161          * want to change *many* call functions, so ext4_get_blocks()
2162          * will set the magic i_delalloc_reserved_flag once the
2163          * inode's allocation semaphore is taken.
2164          *
2165          * If the blocks in questions were delalloc blocks, set
2166          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2167          * variables are updated after the blocks have been allocated.
2168          */
2169         new.b_state = 0;
2170         get_blocks_flags = (EXT4_GET_BLOCKS_CREATE |
2171                             EXT4_GET_BLOCKS_DELALLOC_RESERVE);
2172         if (mpd->b_state & (1 << BH_Delay))
2173                 get_blocks_flags |= EXT4_GET_BLOCKS_UPDATE_RESERVE_SPACE;
2174         blks = ext4_get_blocks(handle, mpd->inode, next, max_blocks,
2175                                &new, get_blocks_flags);
2176         if (blks < 0) {
2177                 err = blks;
2178                 /*
2179                  * If get block returns with error we simply
2180                  * return. Later writepage will redirty the page and
2181                  * writepages will find the dirty page again
2182                  */
2183                 if (err == -EAGAIN)
2184                         return 0;
2185
2186                 if (err == -ENOSPC &&
2187                     ext4_count_free_blocks(mpd->inode->i_sb)) {
2188                         mpd->retval = err;
2189                         return 0;
2190                 }
2191
2192                 /*
2193                  * get block failure will cause us to loop in
2194                  * writepages, because a_ops->writepage won't be able
2195                  * to make progress. The page will be redirtied by
2196                  * writepage and writepages will again try to write
2197                  * the same.
2198                  */
2199                 printk(KERN_EMERG "%s block allocation failed for inode %lu "
2200                                   "at logical offset %llu with max blocks "
2201                                   "%zd with error %d\n",
2202                                   __func__, mpd->inode->i_ino,
2203                                   (unsigned long long)next,
2204                                   mpd->b_size >> mpd->inode->i_blkbits, err);
2205                 printk(KERN_EMERG "This should not happen.!! "
2206                                         "Data will be lost\n");
2207                 if (err == -ENOSPC) {
2208                         ext4_print_free_blocks(mpd->inode);
2209                 }
2210                 /* invalidate all the pages */
2211                 ext4_da_block_invalidatepages(mpd, next,
2212                                 mpd->b_size >> mpd->inode->i_blkbits);
2213                 return err;
2214         }
2215         BUG_ON(blks == 0);
2216
2217         new.b_size = (blks << mpd->inode->i_blkbits);
2218
2219         if (buffer_new(&new))
2220                 __unmap_underlying_blocks(mpd->inode, &new);
2221
2222         /*
2223          * If blocks are delayed marked, we need to
2224          * put actual blocknr and drop delayed bit
2225          */
2226         if ((mpd->b_state & (1 << BH_Delay)) ||
2227             (mpd->b_state & (1 << BH_Unwritten)))
2228                 mpage_put_bnr_to_bhs(mpd, next, &new);
2229
2230         if (ext4_should_order_data(mpd->inode)) {
2231                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2232                 if (err)
2233                         return err;
2234         }
2235
2236         /*
2237          * Update on-disk size along with block allocation.
2238          */
2239         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2240         if (disksize > i_size_read(mpd->inode))
2241                 disksize = i_size_read(mpd->inode);
2242         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2243                 ext4_update_i_disksize(mpd->inode, disksize);
2244                 return ext4_mark_inode_dirty(handle, mpd->inode);
2245         }
2246
2247         return 0;
2248 }
2249
2250 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2251                 (1 << BH_Delay) | (1 << BH_Unwritten))
2252
2253 /*
2254  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2255  *
2256  * @mpd->lbh - extent of blocks
2257  * @logical - logical number of the block in the file
2258  * @bh - bh of the block (used to access block's state)
2259  *
2260  * the function is used to collect contig. blocks in same state
2261  */
2262 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2263                                    sector_t logical, size_t b_size,
2264                                    unsigned long b_state)
2265 {
2266         sector_t next;
2267         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2268
2269         /* check if thereserved journal credits might overflow */
2270         if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
2271                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2272                         /*
2273                          * With non-extent format we are limited by the journal
2274                          * credit available.  Total credit needed to insert
2275                          * nrblocks contiguous blocks is dependent on the
2276                          * nrblocks.  So limit nrblocks.
2277                          */
2278                         goto flush_it;
2279                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2280                                 EXT4_MAX_TRANS_DATA) {
2281                         /*
2282                          * Adding the new buffer_head would make it cross the
2283                          * allowed limit for which we have journal credit
2284                          * reserved. So limit the new bh->b_size
2285                          */
2286                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2287                                                 mpd->inode->i_blkbits;
2288                         /* we will do mpage_da_submit_io in the next loop */
2289                 }
2290         }
2291         /*
2292          * First block in the extent
2293          */
2294         if (mpd->b_size == 0) {
2295                 mpd->b_blocknr = logical;
2296                 mpd->b_size = b_size;
2297                 mpd->b_state = b_state & BH_FLAGS;
2298                 return;
2299         }
2300
2301         next = mpd->b_blocknr + nrblocks;
2302         /*
2303          * Can we merge the block to our big extent?
2304          */
2305         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2306                 mpd->b_size += b_size;
2307                 return;
2308         }
2309
2310 flush_it:
2311         /*
2312          * We couldn't merge the block to our extent, so we
2313          * need to flush current  extent and start new one
2314          */
2315         if (mpage_da_map_blocks(mpd) == 0)
2316                 mpage_da_submit_io(mpd);
2317         mpd->io_done = 1;
2318         return;
2319 }
2320
2321 static int ext4_bh_unmapped_or_delay(handle_t *handle, struct buffer_head *bh)
2322 {
2323         /*
2324          * unmapped buffer is possible for holes.
2325          * delay buffer is possible with delayed allocation.
2326          * We also need to consider unwritten buffer as unmapped.
2327          */
2328         return (!buffer_mapped(bh) || buffer_delay(bh) ||
2329                                 buffer_unwritten(bh)) && buffer_dirty(bh);
2330 }
2331
2332 /*
2333  * __mpage_da_writepage - finds extent of pages and blocks
2334  *
2335  * @page: page to consider
2336  * @wbc: not used, we just follow rules
2337  * @data: context
2338  *
2339  * The function finds extents of pages and scan them for all blocks.
2340  */
2341 static int __mpage_da_writepage(struct page *page,
2342                                 struct writeback_control *wbc, void *data)
2343 {
2344         struct mpage_da_data *mpd = data;
2345         struct inode *inode = mpd->inode;
2346         struct buffer_head *bh, *head;
2347         sector_t logical;
2348
2349         if (mpd->io_done) {
2350                 /*
2351                  * Rest of the page in the page_vec
2352                  * redirty then and skip then. We will
2353                  * try to to write them again after
2354                  * starting a new transaction
2355                  */
2356                 redirty_page_for_writepage(wbc, page);
2357                 unlock_page(page);
2358                 return MPAGE_DA_EXTENT_TAIL;
2359         }
2360         /*
2361          * Can we merge this page to current extent?
2362          */
2363         if (mpd->next_page != page->index) {
2364                 /*
2365                  * Nope, we can't. So, we map non-allocated blocks
2366                  * and start IO on them using writepage()
2367                  */
2368                 if (mpd->next_page != mpd->first_page) {
2369                         if (mpage_da_map_blocks(mpd) == 0)
2370                                 mpage_da_submit_io(mpd);
2371                         /*
2372                          * skip rest of the page in the page_vec
2373                          */
2374                         mpd->io_done = 1;
2375                         redirty_page_for_writepage(wbc, page);
2376                         unlock_page(page);
2377                         return MPAGE_DA_EXTENT_TAIL;
2378                 }
2379
2380                 /*
2381                  * Start next extent of pages ...
2382                  */
2383                 mpd->first_page = page->index;
2384
2385                 /*
2386                  * ... and blocks
2387                  */
2388                 mpd->b_size = 0;
2389                 mpd->b_state = 0;
2390                 mpd->b_blocknr = 0;
2391         }
2392
2393         mpd->next_page = page->index + 1;
2394         logical = (sector_t) page->index <<
2395                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2396
2397         if (!page_has_buffers(page)) {
2398                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2399                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2400                 if (mpd->io_done)
2401                         return MPAGE_DA_EXTENT_TAIL;
2402         } else {
2403                 /*
2404                  * Page with regular buffer heads, just add all dirty ones
2405                  */
2406                 head = page_buffers(page);
2407                 bh = head;
2408                 do {
2409                         BUG_ON(buffer_locked(bh));
2410                         /*
2411                          * We need to try to allocate
2412                          * unmapped blocks in the same page.
2413                          * Otherwise we won't make progress
2414                          * with the page in ext4_da_writepage
2415                          */
2416                         if (ext4_bh_unmapped_or_delay(NULL, bh)) {
2417                                 mpage_add_bh_to_extent(mpd, logical,
2418                                                        bh->b_size,
2419                                                        bh->b_state);
2420                                 if (mpd->io_done)
2421                                         return MPAGE_DA_EXTENT_TAIL;
2422                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2423                                 /*
2424                                  * mapped dirty buffer. We need to update
2425                                  * the b_state because we look at
2426                                  * b_state in mpage_da_map_blocks. We don't
2427                                  * update b_size because if we find an
2428                                  * unmapped buffer_head later we need to
2429                                  * use the b_state flag of that buffer_head.
2430                                  */
2431                                 if (mpd->b_size == 0)
2432                                         mpd->b_state = bh->b_state & BH_FLAGS;
2433                         }
2434                         logical++;
2435                 } while ((bh = bh->b_this_page) != head);
2436         }
2437
2438         return 0;
2439 }
2440
2441 /*
2442  * This is a special get_blocks_t callback which is used by
2443  * ext4_da_write_begin().  It will either return mapped block or
2444  * reserve space for a single block.
2445  *
2446  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2447  * We also have b_blocknr = -1 and b_bdev initialized properly
2448  *
2449  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2450  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2451  * initialized properly.
2452  */
2453 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2454                                   struct buffer_head *bh_result, int create)
2455 {
2456         int ret = 0;
2457         sector_t invalid_block = ~((sector_t) 0xffff);
2458
2459         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2460                 invalid_block = ~0;
2461
2462         BUG_ON(create == 0);
2463         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2464
2465         /*
2466          * first, we need to know whether the block is allocated already
2467          * preallocated blocks are unmapped but should treated
2468          * the same as allocated blocks.
2469          */
2470         ret = ext4_get_blocks(NULL, inode, iblock, 1,  bh_result, 0);
2471         if ((ret == 0) && !buffer_delay(bh_result)) {
2472                 /* the block isn't (pre)allocated yet, let's reserve space */
2473                 /*
2474                  * XXX: __block_prepare_write() unmaps passed block,
2475                  * is it OK?
2476                  */
2477                 ret = ext4_da_reserve_space(inode, 1);
2478                 if (ret)
2479                         /* not enough space to reserve */
2480                         return ret;
2481
2482                 map_bh(bh_result, inode->i_sb, invalid_block);
2483                 set_buffer_new(bh_result);
2484                 set_buffer_delay(bh_result);
2485         } else if (ret > 0) {
2486                 bh_result->b_size = (ret << inode->i_blkbits);
2487                 if (buffer_unwritten(bh_result)) {
2488                         /* A delayed write to unwritten bh should
2489                          * be marked new and mapped.  Mapped ensures
2490                          * that we don't do get_block multiple times
2491                          * when we write to the same offset and new
2492                          * ensures that we do proper zero out for
2493                          * partial write.
2494                          */
2495                         set_buffer_new(bh_result);
2496                         set_buffer_mapped(bh_result);
2497                 }
2498                 ret = 0;
2499         }
2500
2501         return ret;
2502 }
2503
2504 /*
2505  * This function is used as a standard get_block_t calback function
2506  * when there is no desire to allocate any blocks.  It is used as a
2507  * callback function for block_prepare_write(), nobh_writepage(), and
2508  * block_write_full_page().  These functions should only try to map a
2509  * single block at a time.
2510  *
2511  * Since this function doesn't do block allocations even if the caller
2512  * requests it by passing in create=1, it is critically important that
2513  * any caller checks to make sure that any buffer heads are returned
2514  * by this function are either all already mapped or marked for
2515  * delayed allocation before calling nobh_writepage() or
2516  * block_write_full_page().  Otherwise, b_blocknr could be left
2517  * unitialized, and the page write functions will be taken by
2518  * surprise.
2519  */
2520 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2521                                    struct buffer_head *bh_result, int create)
2522 {
2523         int ret = 0;
2524         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2525
2526         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2527
2528         /*
2529          * we don't want to do block allocation in writepage
2530          * so call get_block_wrap with create = 0
2531          */
2532         ret = ext4_get_blocks(NULL, inode, iblock, max_blocks, bh_result, 0);
2533         BUG_ON(create && ret == 0);
2534         if (ret > 0) {
2535                 bh_result->b_size = (ret << inode->i_blkbits);
2536                 ret = 0;
2537         }
2538         return ret;
2539 }
2540
2541 /*
2542  * This function can get called via...
2543  *   - ext4_da_writepages after taking page lock (have journal handle)
2544  *   - journal_submit_inode_data_buffers (no journal handle)
2545  *   - shrink_page_list via pdflush (no journal handle)
2546  *   - grab_page_cache when doing write_begin (have journal handle)
2547  */
2548 static int ext4_da_writepage(struct page *page,
2549                                 struct writeback_control *wbc)
2550 {
2551         int ret = 0;
2552         loff_t size;
2553         unsigned int len;
2554         struct buffer_head *page_bufs;
2555         struct inode *inode = page->mapping->host;
2556
2557         trace_mark(ext4_da_writepage,
2558                    "dev %s ino %lu page_index %lu",
2559                    inode->i_sb->s_id, inode->i_ino, page->index);
2560         size = i_size_read(inode);
2561         if (page->index == size >> PAGE_CACHE_SHIFT)
2562                 len = size & ~PAGE_CACHE_MASK;
2563         else
2564                 len = PAGE_CACHE_SIZE;
2565
2566         if (page_has_buffers(page)) {
2567                 page_bufs = page_buffers(page);
2568                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2569                                         ext4_bh_unmapped_or_delay)) {
2570                         /*
2571                          * We don't want to do  block allocation
2572                          * So redirty the page and return
2573                          * We may reach here when we do a journal commit
2574                          * via journal_submit_inode_data_buffers.
2575                          * If we don't have mapping block we just ignore
2576                          * them. We can also reach here via shrink_page_list
2577                          */
2578                         redirty_page_for_writepage(wbc, page);
2579                         unlock_page(page);
2580                         return 0;
2581                 }
2582         } else {
2583                 /*
2584                  * The test for page_has_buffers() is subtle:
2585                  * We know the page is dirty but it lost buffers. That means
2586                  * that at some moment in time after write_begin()/write_end()
2587                  * has been called all buffers have been clean and thus they
2588                  * must have been written at least once. So they are all
2589                  * mapped and we can happily proceed with mapping them
2590                  * and writing the page.
2591                  *
2592                  * Try to initialize the buffer_heads and check whether
2593                  * all are mapped and non delay. We don't want to
2594                  * do block allocation here.
2595                  */
2596                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
2597                                           noalloc_get_block_write);
2598                 if (!ret) {
2599                         page_bufs = page_buffers(page);
2600                         /* check whether all are mapped and non delay */
2601                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2602                                                 ext4_bh_unmapped_or_delay)) {
2603                                 redirty_page_for_writepage(wbc, page);
2604                                 unlock_page(page);
2605                                 return 0;
2606                         }
2607                 } else {
2608                         /*
2609                          * We can't do block allocation here
2610                          * so just redity the page and unlock
2611                          * and return
2612                          */
2613                         redirty_page_for_writepage(wbc, page);
2614                         unlock_page(page);
2615                         return 0;
2616                 }
2617                 /* now mark the buffer_heads as dirty and uptodate */
2618                 block_commit_write(page, 0, PAGE_CACHE_SIZE);
2619         }
2620
2621         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2622                 ret = nobh_writepage(page, noalloc_get_block_write, wbc);
2623         else
2624                 ret = block_write_full_page(page, noalloc_get_block_write,
2625                                             wbc);
2626
2627         return ret;
2628 }
2629
2630 /*
2631  * This is called via ext4_da_writepages() to
2632  * calulate the total number of credits to reserve to fit
2633  * a single extent allocation into a single transaction,
2634  * ext4_da_writpeages() will loop calling this before
2635  * the block allocation.
2636  */
2637
2638 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2639 {
2640         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2641
2642         /*
2643          * With non-extent format the journal credit needed to
2644          * insert nrblocks contiguous block is dependent on
2645          * number of contiguous block. So we will limit
2646          * number of contiguous block to a sane value
2647          */
2648         if (!(inode->i_flags & EXT4_EXTENTS_FL) &&
2649             (max_blocks > EXT4_MAX_TRANS_DATA))
2650                 max_blocks = EXT4_MAX_TRANS_DATA;
2651
2652         return ext4_chunk_trans_blocks(inode, max_blocks);
2653 }
2654
2655 static int ext4_da_writepages(struct address_space *mapping,
2656                               struct writeback_control *wbc)
2657 {
2658         pgoff_t index;
2659         int range_whole = 0;
2660         handle_t *handle = NULL;
2661         struct mpage_da_data mpd;
2662         struct inode *inode = mapping->host;
2663         int no_nrwrite_index_update;
2664         int pages_written = 0;
2665         long pages_skipped;
2666         int range_cyclic, cycled = 1, io_done = 0;
2667         int needed_blocks, ret = 0, nr_to_writebump = 0;
2668         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2669
2670         trace_mark(ext4_da_writepages,
2671                    "dev %s ino %lu nr_t_write %ld "
2672                    "pages_skipped %ld range_start %llu "
2673                    "range_end %llu nonblocking %d "
2674                    "for_kupdate %d for_reclaim %d "
2675                    "for_writepages %d range_cyclic %d",
2676                    inode->i_sb->s_id, inode->i_ino,
2677                    wbc->nr_to_write, wbc->pages_skipped,
2678                    (unsigned long long) wbc->range_start,
2679                    (unsigned long long) wbc->range_end,
2680                    wbc->nonblocking, wbc->for_kupdate,
2681                    wbc->for_reclaim, wbc->for_writepages,
2682                    wbc->range_cyclic);
2683
2684         /*
2685          * No pages to write? This is mainly a kludge to avoid starting
2686          * a transaction for special inodes like journal inode on last iput()
2687          * because that could violate lock ordering on umount
2688          */
2689         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2690                 return 0;
2691
2692         /*
2693          * If the filesystem has aborted, it is read-only, so return
2694          * right away instead of dumping stack traces later on that
2695          * will obscure the real source of the problem.  We test
2696          * EXT4_MOUNT_ABORT instead of sb->s_flag's MS_RDONLY because
2697          * the latter could be true if the filesystem is mounted
2698          * read-only, and in that case, ext4_da_writepages should
2699          * *never* be called, so if that ever happens, we would want
2700          * the stack trace.
2701          */
2702         if (unlikely(sbi->s_mount_opt & EXT4_MOUNT_ABORT))
2703                 return -EROFS;
2704
2705         /*
2706          * Make sure nr_to_write is >= sbi->s_mb_stream_request
2707          * This make sure small files blocks are allocated in
2708          * single attempt. This ensure that small files
2709          * get less fragmented.
2710          */
2711         if (wbc->nr_to_write < sbi->s_mb_stream_request) {
2712                 nr_to_writebump = sbi->s_mb_stream_request - wbc->nr_to_write;
2713                 wbc->nr_to_write = sbi->s_mb_stream_request;
2714         }
2715         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2716                 range_whole = 1;
2717
2718         range_cyclic = wbc->range_cyclic;
2719         if (wbc->range_cyclic) {
2720                 index = mapping->writeback_index;
2721                 if (index)
2722                         cycled = 0;
2723                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2724                 wbc->range_end  = LLONG_MAX;
2725                 wbc->range_cyclic = 0;
2726         } else
2727                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2728
2729         mpd.wbc = wbc;
2730         mpd.inode = mapping->host;
2731
2732         /*
2733          * we don't want write_cache_pages to update
2734          * nr_to_write and writeback_index
2735          */
2736         no_nrwrite_index_update = wbc->no_nrwrite_index_update;
2737         wbc->no_nrwrite_index_update = 1;
2738         pages_skipped = wbc->pages_skipped;
2739
2740 retry:
2741         while (!ret && wbc->nr_to_write > 0) {
2742
2743                 /*
2744                  * we  insert one extent at a time. So we need
2745                  * credit needed for single extent allocation.
2746                  * journalled mode is currently not supported
2747                  * by delalloc
2748                  */
2749                 BUG_ON(ext4_should_journal_data(inode));
2750                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2751
2752                 /* start a new transaction*/
2753                 handle = ext4_journal_start(inode, needed_blocks);
2754                 if (IS_ERR(handle)) {
2755                         ret = PTR_ERR(handle);
2756                         printk(KERN_CRIT "%s: jbd2_start: "
2757                                "%ld pages, ino %lu; err %d\n", __func__,
2758                                 wbc->nr_to_write, inode->i_ino, ret);
2759                         dump_stack();
2760                         goto out_writepages;
2761                 }
2762
2763                 /*
2764                  * Now call __mpage_da_writepage to find the next
2765                  * contiguous region of logical blocks that need
2766                  * blocks to be allocated by ext4.  We don't actually
2767                  * submit the blocks for I/O here, even though
2768                  * write_cache_pages thinks it will, and will set the
2769                  * pages as clean for write before calling
2770                  * __mpage_da_writepage().
2771                  */
2772                 mpd.b_size = 0;
2773                 mpd.b_state = 0;
2774                 mpd.b_blocknr = 0;
2775                 mpd.first_page = 0;
2776                 mpd.next_page = 0;
2777                 mpd.io_done = 0;
2778                 mpd.pages_written = 0;
2779                 mpd.retval = 0;
2780                 ret = write_cache_pages(mapping, wbc, __mpage_da_writepage,
2781                                         &mpd);
2782                 /*
2783                  * If we have a contigous extent of pages and we
2784                  * haven't done the I/O yet, map the blocks and submit
2785                  * them for I/O.
2786                  */
2787                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2788                         if (mpage_da_map_blocks(&mpd) == 0)
2789                                 mpage_da_submit_io(&mpd);
2790                         mpd.io_done = 1;
2791                         ret = MPAGE_DA_EXTENT_TAIL;
2792                 }
2793                 wbc->nr_to_write -= mpd.pages_written;
2794
2795                 ext4_journal_stop(handle);
2796
2797                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2798                         /* commit the transaction which would
2799                          * free blocks released in the transaction
2800                          * and try again
2801                          */
2802                         jbd2_journal_force_commit_nested(sbi->s_journal);
2803                         wbc->pages_skipped = pages_skipped;
2804                         ret = 0;
2805                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2806                         /*
2807                          * got one extent now try with
2808                          * rest of the pages
2809                          */
2810                         pages_written += mpd.pages_written;
2811                         wbc->pages_skipped = pages_skipped;
2812                         ret = 0;
2813                         io_done = 1;
2814                 } else if (wbc->nr_to_write)
2815                         /*
2816                          * There is no more writeout needed
2817                          * or we requested for a noblocking writeout
2818                          * and we found the device congested
2819                          */
2820                         break;
2821         }
2822         if (!io_done && !cycled) {
2823                 cycled = 1;
2824                 index = 0;
2825                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2826                 wbc->range_end  = mapping->writeback_index - 1;
2827                 goto retry;
2828         }
2829         if (pages_skipped != wbc->pages_skipped)
2830                 printk(KERN_EMERG "This should not happen leaving %s "
2831                                 "with nr_to_write = %ld ret = %d\n",
2832                                 __func__, wbc->nr_to_write, ret);
2833
2834         /* Update index */
2835         index += pages_written;
2836         wbc->range_cyclic = range_cyclic;
2837         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2838                 /*
2839                  * set the writeback_index so that range_cyclic
2840                  * mode will write it back later
2841                  */
2842                 mapping->writeback_index = index;
2843
2844 out_writepages:
2845         if (!no_nrwrite_index_update)
2846                 wbc->no_nrwrite_index_update = 0;
2847         wbc->nr_to_write -= nr_to_writebump;
2848         trace_mark(ext4_da_writepage_result,
2849                    "dev %s ino %lu ret %d pages_written %d "
2850                    "pages_skipped %ld congestion %d "
2851                    "more_io %d no_nrwrite_index_update %d",
2852                    inode->i_sb->s_id, inode->i_ino, ret,
2853                    pages_written, wbc->pages_skipped,
2854                    wbc->encountered_congestion, wbc->more_io,
2855                    wbc->no_nrwrite_index_update);
2856         return ret;
2857 }
2858
2859 #define FALL_BACK_TO_NONDELALLOC 1
2860 static int ext4_nonda_switch(struct super_block *sb)
2861 {
2862         s64 free_blocks, dirty_blocks;
2863         struct ext4_sb_info *sbi = EXT4_SB(sb);
2864
2865         /*
2866          * switch to non delalloc mode if we are running low
2867          * on free block. The free block accounting via percpu
2868          * counters can get slightly wrong with percpu_counter_batch getting
2869          * accumulated on each CPU without updating global counters
2870          * Delalloc need an accurate free block accounting. So switch
2871          * to non delalloc when we are near to error range.
2872          */
2873         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
2874         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
2875         if (2 * free_blocks < 3 * dirty_blocks ||
2876                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
2877                 /*
2878                  * free block count is less that 150% of dirty blocks
2879                  * or free blocks is less that watermark
2880                  */
2881                 return 1;
2882         }
2883         return 0;
2884 }
2885
2886 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2887                                 loff_t pos, unsigned len, unsigned flags,
2888                                 struct page **pagep, void **fsdata)
2889 {
2890         int ret, retries = 0;
2891         struct page *page;
2892         pgoff_t index;
2893         unsigned from, to;
2894         struct inode *inode = mapping->host;
2895         handle_t *handle;
2896
2897         index = pos >> PAGE_CACHE_SHIFT;
2898         from = pos & (PAGE_CACHE_SIZE - 1);
2899         to = from + len;
2900
2901         if (ext4_nonda_switch(inode->i_sb)) {
2902                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2903                 return ext4_write_begin(file, mapping, pos,
2904                                         len, flags, pagep, fsdata);
2905         }
2906         *fsdata = (void *)0;
2907
2908         trace_mark(ext4_da_write_begin,
2909                    "dev %s ino %lu pos %llu len %u flags %u",
2910                    inode->i_sb->s_id, inode->i_ino,
2911                    (unsigned long long) pos, len, flags);
2912 retry:
2913         /*
2914          * With delayed allocation, we don't log the i_disksize update
2915          * if there is delayed block allocation. But we still need
2916          * to journalling the i_disksize update if writes to the end
2917          * of file which has an already mapped buffer.
2918          */
2919         handle = ext4_journal_start(inode, 1);
2920         if (IS_ERR(handle)) {
2921                 ret = PTR_ERR(handle);
2922                 goto out;
2923         }
2924         /* We cannot recurse into the filesystem as the transaction is already
2925          * started */
2926         flags |= AOP_FLAG_NOFS;
2927
2928         page = grab_cache_page_write_begin(mapping, index, flags);
2929         if (!page) {
2930                 ext4_journal_stop(handle);
2931                 ret = -ENOMEM;
2932                 goto out;
2933         }
2934         *pagep = page;
2935
2936         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
2937                                 ext4_da_get_block_prep);
2938         if (ret < 0) {
2939                 unlock_page(page);
2940                 ext4_journal_stop(handle);
2941                 page_cache_release(page);
2942                 /*
2943                  * block_write_begin may have instantiated a few blocks
2944                  * outside i_size.  Trim these off again. Don't need
2945                  * i_size_read because we hold i_mutex.
2946                  */
2947                 if (pos + len > inode->i_size)
2948                         vmtruncate(inode, inode->i_size);
2949         }
2950
2951         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
2952                 goto retry;
2953 out:
2954         return ret;
2955 }
2956
2957 /*
2958  * Check if we should update i_disksize
2959  * when write to the end of file but not require block allocation
2960  */
2961 static int ext4_da_should_update_i_disksize(struct page *page,
2962                                          unsigned long offset)
2963 {
2964         struct buffer_head *bh;
2965         struct inode *inode = page->mapping->host;
2966         unsigned int idx;
2967         int i;
2968
2969         bh = page_buffers(page);
2970         idx = offset >> inode->i_blkbits;
2971
2972         for (i = 0; i < idx; i++)
2973                 bh = bh->b_this_page;
2974
2975         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
2976                 return 0;
2977         return 1;
2978 }
2979
2980 static int ext4_da_write_end(struct file *file,
2981                                 struct address_space *mapping,
2982                                 loff_t pos, unsigned len, unsigned copied,
2983                                 struct page *page, void *fsdata)
2984 {
2985         struct inode *inode = mapping->host;
2986         int ret = 0, ret2;
2987         handle_t *handle = ext4_journal_current_handle();
2988         loff_t new_i_size;
2989         unsigned long start, end;
2990         int write_mode = (int)(unsigned long)fsdata;
2991
2992         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
2993                 if (ext4_should_order_data(inode)) {
2994                         return ext4_ordered_write_end(file, mapping, pos,
2995                                         len, copied, page, fsdata);
2996                 } else if (ext4_should_writeback_data(inode)) {
2997                         return ext4_writeback_write_end(file, mapping, pos,
2998                                         len, copied, page, fsdata);
2999                 } else {
3000                         BUG();
3001                 }
3002         }
3003
3004         trace_mark(ext4_da_write_end,
3005                    "dev %s ino %lu pos %llu len %u copied %u",
3006                    inode->i_sb->s_id, inode->i_ino,
3007                    (unsigned long long) pos, len, copied);
3008         start = pos & (PAGE_CACHE_SIZE - 1);
3009         end = start + copied - 1;
3010
3011         /*
3012          * generic_write_end() will run mark_inode_dirty() if i_size
3013          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3014          * into that.
3015          */
3016
3017         new_i_size = pos + copied;
3018         if (new_i_size > EXT4_I(inode)->i_disksize) {
3019                 if (ext4_da_should_update_i_disksize(page, end)) {
3020                         down_write(&EXT4_I(inode)->i_data_sem);
3021                         if (new_i_size > EXT4_I(inode)->i_disksize) {
3022                                 /*
3023                                  * Updating i_disksize when extending file
3024                                  * without needing block allocation
3025                                  */
3026                                 if (ext4_should_order_data(inode))
3027                                         ret = ext4_jbd2_file_inode(handle,
3028                                                                    inode);
3029
3030                                 EXT4_I(inode)->i_disksize = new_i_size;
3031                         }
3032                         up_write(&EXT4_I(inode)->i_data_sem);
3033                         /* We need to mark inode dirty even if
3034                          * new_i_size is less that inode->i_size
3035                          * bu greater than i_disksize.(hint delalloc)
3036                          */
3037                         ext4_mark_inode_dirty(handle, inode);
3038                 }
3039         }
3040         ret2 = generic_write_end(file, mapping, pos, len, copied,
3041                                                         page, fsdata);
3042         copied = ret2;
3043         if (ret2 < 0)
3044                 ret = ret2;
3045         ret2 = ext4_journal_stop(handle);
3046         if (!ret)
3047                 ret = ret2;
3048
3049         return ret ? ret : copied;
3050 }
3051
3052 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
3053 {
3054         /*
3055          * Drop reserved blocks
3056          */
3057         BUG_ON(!PageLocked(page));
3058         if (!page_has_buffers(page))
3059                 goto out;
3060
3061         ext4_da_page_release_reservation(page, offset);
3062
3063 out:
3064         ext4_invalidatepage(page, offset);
3065
3066         return;
3067 }
3068
3069 /*
3070  * Force all delayed allocation blocks to be allocated for a given inode.
3071  */
3072 int ext4_alloc_da_blocks(struct inode *inode)
3073 {
3074         if (!EXT4_I(inode)->i_reserved_data_blocks &&
3075             !EXT4_I(inode)->i_reserved_meta_blocks)
3076                 return 0;
3077
3078         /*
3079          * We do something simple for now.  The filemap_flush() will
3080          * also start triggering a write of the data blocks, which is
3081          * not strictly speaking necessary (and for users of
3082          * laptop_mode, not even desirable).  However, to do otherwise
3083          * would require replicating code paths in:
3084          * 
3085          * ext4_da_writepages() ->
3086          *    write_cache_pages() ---> (via passed in callback function)
3087          *        __mpage_da_writepage() -->
3088          *           mpage_add_bh_to_extent()
3089          *           mpage_da_map_blocks()
3090          *
3091          * The problem is that write_cache_pages(), located in
3092          * mm/page-writeback.c, marks pages clean in preparation for
3093          * doing I/O, which is not desirable if we're not planning on
3094          * doing I/O at all.
3095          *
3096          * We could call write_cache_pages(), and then redirty all of
3097          * the pages by calling redirty_page_for_writeback() but that
3098          * would be ugly in the extreme.  So instead we would need to
3099          * replicate parts of the code in the above functions,
3100          * simplifying them becuase we wouldn't actually intend to
3101          * write out the pages, but rather only collect contiguous
3102          * logical block extents, call the multi-block allocator, and
3103          * then update the buffer heads with the block allocations.
3104          * 
3105          * For now, though, we'll cheat by calling filemap_flush(),
3106          * which will map the blocks, and start the I/O, but not
3107          * actually wait for the I/O to complete.
3108          */
3109         return filemap_flush(inode->i_mapping);
3110 }
3111
3112 /*
3113  * bmap() is special.  It gets used by applications such as lilo and by
3114  * the swapper to find the on-disk block of a specific piece of data.
3115  *
3116  * Naturally, this is dangerous if the block concerned is still in the
3117  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3118  * filesystem and enables swap, then they may get a nasty shock when the
3119  * data getting swapped to that swapfile suddenly gets overwritten by
3120  * the original zero's written out previously to the journal and
3121  * awaiting writeback in the kernel's buffer cache.
3122  *
3123  * So, if we see any bmap calls here on a modified, data-journaled file,
3124  * take extra steps to flush any blocks which might be in the cache.
3125  */
3126 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3127 {
3128         struct inode *inode = mapping->host;
3129         journal_t *journal;
3130         int err;
3131
3132         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3133                         test_opt(inode->i_sb, DELALLOC)) {
3134                 /*
3135                  * With delalloc we want to sync the file
3136                  * so that we can make sure we allocate
3137                  * blocks for file
3138                  */
3139                 filemap_write_and_wait(mapping);
3140         }
3141
3142         if (EXT4_JOURNAL(inode) && EXT4_I(inode)->i_state & EXT4_STATE_JDATA) {
3143                 /*
3144                  * This is a REALLY heavyweight approach, but the use of
3145                  * bmap on dirty files is expected to be extremely rare:
3146                  * only if we run lilo or swapon on a freshly made file
3147                  * do we expect this to happen.
3148                  *
3149                  * (bmap requires CAP_SYS_RAWIO so this does not
3150                  * represent an unprivileged user DOS attack --- we'd be
3151                  * in trouble if mortal users could trigger this path at
3152                  * will.)
3153                  *
3154                  * NB. EXT4_STATE_JDATA is not set on files other than
3155                  * regular files.  If somebody wants to bmap a directory
3156                  * or symlink and gets confused because the buffer
3157                  * hasn't yet been flushed to disk, they deserve
3158                  * everything they get.
3159                  */
3160
3161                 EXT4_I(inode)->i_state &= ~EXT4_STATE_JDATA;
3162                 journal = EXT4_JOURNAL(inode);
3163                 jbd2_journal_lock_updates(journal);
3164                 err = jbd2_journal_flush(journal);
3165                 jbd2_journal_unlock_updates(journal);
3166
3167                 if (err)
3168                         return 0;
3169         }
3170
3171         return generic_block_bmap(mapping, block, ext4_get_block);
3172 }
3173
3174 static int bget_one(handle_t *handle, struct buffer_head *bh)
3175 {
3176         get_bh(bh);
3177         return 0;
3178 }
3179
3180 static int bput_one(handle_t *handle, struct buffer_head *bh)
3181 {
3182         put_bh(bh);
3183         return 0;
3184 }
3185
3186 /*
3187  * Note that we don't need to start a transaction unless we're journaling data
3188  * because we should have holes filled from ext4_page_mkwrite(). We even don't
3189  * need to file the inode to the transaction's list in ordered mode because if
3190  * we are writing back data added by write(), the inode is already there and if
3191  * we are writing back data modified via mmap(), noone guarantees in which
3192  * transaction the data will hit the disk. In case we are journaling data, we
3193  * cannot start transaction directly because transaction start ranks above page
3194  * lock so we have to do some magic.
3195  *
3196  * In all journaling modes block_write_full_page() will start the I/O.
3197  *
3198  * Problem:
3199  *
3200  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
3201  *              ext4_writepage()
3202  *
3203  * Similar for:
3204  *
3205  *      ext4_file_write() -> generic_file_write() -> __alloc_pages() -> ...
3206  *
3207  * Same applies to ext4_get_block().  We will deadlock on various things like
3208  * lock_journal and i_data_sem
3209  *
3210  * Setting PF_MEMALLOC here doesn't work - too many internal memory
3211  * allocations fail.
3212  *
3213  * 16May01: If we're reentered then journal_current_handle() will be
3214  *          non-zero. We simply *return*.
3215  *
3216  * 1 July 2001: @@@ FIXME:
3217  *   In journalled data mode, a data buffer may be metadata against the
3218  *   current transaction.  But the same file is part of a shared mapping
3219  *   and someone does a writepage() on it.
3220  *
3221  *   We will move the buffer onto the async_data list, but *after* it has
3222  *   been dirtied. So there's a small window where we have dirty data on
3223  *   BJ_Metadata.
3224  *
3225  *   Note that this only applies to the last partial page in the file.  The
3226  *   bit which block_write_full_page() uses prepare/commit for.  (That's
3227  *   broken code anyway: it's wrong for msync()).
3228  *
3229  *   It's a rare case: affects the final partial page, for journalled data
3230  *   where the file is subject to bith write() and writepage() in the same
3231  *   transction.  To fix it we'll need a custom block_write_full_page().
3232  *   We'll probably need that anyway for journalling writepage() output.
3233  *
3234  * We don't honour synchronous mounts for writepage().  That would be
3235  * disastrous.  Any write() or metadata operation will sync the fs for
3236  * us.
3237  *
3238  */
3239 static int __ext4_normal_writepage(struct page *page,
3240                                 struct writeback_control *wbc)
3241 {
3242         struct inode *inode = page->mapping->host;
3243
3244         if (test_opt(inode->i_sb, NOBH))
3245                 return nobh_writepage(page, noalloc_get_block_write, wbc);
3246         else
3247                 return block_write_full_page(page, noalloc_get_block_write,
3248                                              wbc);
3249 }
3250
3251 static int ext4_normal_writepage(struct page *page,
3252                                 struct writeback_control *wbc)
3253 {
3254         struct inode *inode = page->mapping->host;
3255         loff_t size = i_size_read(inode);
3256         loff_t len;
3257
3258         trace_mark(ext4_normal_writepage,
3259                    "dev %s ino %lu page_index %lu",
3260                    inode->i_sb->s_id, inode->i_ino, page->index);
3261         J_ASSERT(PageLocked(page));
3262         if (page->index == size >> PAGE_CACHE_SHIFT)
3263                 len = size & ~PAGE_CACHE_MASK;
3264         else
3265                 len = PAGE_CACHE_SIZE;
3266
3267         if (page_has_buffers(page)) {
3268                 /* if page has buffers it should all be mapped
3269                  * and allocated. If there are not buffers attached
3270                  * to the page we know the page is dirty but it lost
3271                  * buffers. That means that at some moment in time
3272                  * after write_begin() / write_end() has been called
3273                  * all buffers have been clean and thus they must have been
3274                  * written at least once. So they are all mapped and we can
3275                  * happily proceed with mapping them and writing the page.
3276                  */
3277                 BUG_ON(walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
3278                                         ext4_bh_unmapped_or_delay));
3279         }
3280
3281         if (!ext4_journal_current_handle())
3282                 return __ext4_normal_writepage(page, wbc);
3283
3284         redirty_page_for_writepage(wbc, page);
3285         unlock_page(page);
3286         return 0;
3287 }
3288
3289 static int __ext4_journalled_writepage(struct page *page,
3290                                 struct writeback_control *wbc)
3291 {
3292         struct address_space *mapping = page->mapping;
3293         struct inode *inode = mapping->host;
3294         struct buffer_head *page_bufs;
3295         handle_t *handle = NULL;
3296         int ret = 0;
3297         int err;
3298
3299         ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
3300                                   noalloc_get_block_write);
3301         if (ret != 0)
3302                 goto out_unlock;
3303
3304         page_bufs = page_buffers(page);
3305         walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE, NULL,
3306                                                                 bget_one);
3307         /* As soon as we unlock the page, it can go away, but we have
3308          * references to buffers so we are safe */
3309         unlock_page(page);
3310
3311         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
3312         if (IS_ERR(handle)) {
3313                 ret = PTR_ERR(handle);
3314                 goto out;
3315         }
3316
3317         ret = walk_page_buffers(handle, page_bufs, 0,
3318                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
3319
3320         err = walk_page_buffers(handle, page_bufs, 0,
3321                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
3322         if (ret == 0)
3323                 ret = err;
3324         err = ext4_journal_stop(handle);
3325         if (!ret)
3326                 ret = err;
3327
3328         walk_page_buffers(handle, page_bufs, 0,
3329                                 PAGE_CACHE_SIZE, NULL, bput_one);
3330         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
3331         goto out;
3332
3333 out_unlock:
3334         unlock_page(page);
3335 out:
3336         return ret;
3337 }
3338
3339 static int ext4_journalled_writepage(struct page *page,
3340                                 struct writeback_control *wbc)
3341 {
3342         struct inode *inode = page->mapping->host;
3343         loff_t size = i_size_read(inode);
3344         loff_t len;
3345
3346         trace_mark(ext4_journalled_writepage,
3347                    "dev %s ino %lu page_index %lu",
3348                    inode->i_sb->s_id, inode->i_ino, page->index);
3349         J_ASSERT(PageLocked(page));
3350         if (page->index == size >> PAGE_CACHE_SHIFT)
3351                 len = size & ~PAGE_CACHE_MASK;
3352         else
3353                 len = PAGE_CACHE_SIZE;
3354
3355         if (page_has_buffers(page)) {
3356                 /* if page has buffers it should all be mapped
3357                  * and allocated. If there are not buffers attached
3358                  * to the page we know the page is dirty but it lost
3359                  * buffers. That means that at some moment in time
3360                  * after write_begin() / write_end() has been called
3361                  * all buffers have been clean and thus they must have been
3362                  * written at least once. So they are all mapped and we can
3363                  * happily proceed with mapping them and writing the page.
3364                  */
3365                 BUG_ON(walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
3366                                         ext4_bh_unmapped_or_delay));
3367         }
3368
3369         if (ext4_journal_current_handle())
3370                 goto no_write;
3371
3372         if (PageChecked(page)) {
3373                 /*
3374                  * It's mmapped pagecache.  Add buffers and journal it.  There
3375                  * doesn't seem much point in redirtying the page here.
3376                  */
3377                 ClearPageChecked(page);
3378                 return __ext4_journalled_writepage(page, wbc);
3379         } else {
3380                 /*
3381                  * It may be a page full of checkpoint-mode buffers.  We don't
3382                  * really know unless we go poke around in the buffer_heads.
3383                  * But block_write_full_page will do the right thing.
3384                  */
3385                 return block_write_full_page(page, noalloc_get_block_write,
3386                                              wbc);
3387         }
3388 no_write:
3389         redirty_page_for_writepage(wbc, page);
3390         unlock_page(page);
3391         return 0;
3392 }
3393
3394 static int ext4_readpage(struct file *file, struct page *page)
3395 {
3396         return mpage_readpage(page, ext4_get_block);
3397 }
3398
3399 static int
3400 ext4_readpages(struct file *file, struct address_space *mapping,
3401                 struct list_head *pages, unsigned nr_pages)
3402 {
3403         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3404 }
3405
3406 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3407 {
3408         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3409
3410         /*
3411          * If it's a full truncate we just forget about the pending dirtying
3412          */
3413         if (offset == 0)
3414                 ClearPageChecked(page);
3415
3416         if (journal)
3417                 jbd2_journal_invalidatepage(journal, page, offset);
3418         else
3419                 block_invalidatepage(page, offset);
3420 }
3421
3422 static int ext4_releasepage(struct page *page, gfp_t wait)
3423 {
3424         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3425
3426         WARN_ON(PageChecked(page));
3427         if (!page_has_buffers(page))
3428                 return 0;
3429         if (journal)
3430                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3431         else
3432                 return try_to_free_buffers(page);
3433 }
3434
3435 /*
3436  * If the O_DIRECT write will extend the file then add this inode to the
3437  * orphan list.  So recovery will truncate it back to the original size
3438  * if the machine crashes during the write.
3439  *
3440  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3441  * crashes then stale disk data _may_ be exposed inside the file. But current
3442  * VFS code falls back into buffered path in that case so we are safe.
3443  */
3444 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3445                         const struct iovec *iov, loff_t offset,
3446                         unsigned long nr_segs)
3447 {
3448         struct file *file = iocb->ki_filp;
3449         struct inode *inode = file->f_mapping->host;
3450         struct ext4_inode_info *ei = EXT4_I(inode);
3451         handle_t *handle;
3452         ssize_t ret;
3453         int orphan = 0;
3454         size_t count = iov_length(iov, nr_segs);
3455
3456         if (rw == WRITE) {
3457                 loff_t final_size = offset + count;
3458
3459                 if (final_size > inode->i_size) {
3460                         /* Credits for sb + inode write */
3461                         handle = ext4_journal_start(inode, 2);
3462                         if (IS_ERR(handle)) {
3463                                 ret = PTR_ERR(handle);
3464                                 goto out;
3465                         }
3466                         ret = ext4_orphan_add(handle, inode);
3467                         if (ret) {
3468                                 ext4_journal_stop(handle);
3469                                 goto out;
3470                         }
3471                         orphan = 1;
3472                         ei->i_disksize = inode->i_size;
3473                         ext4_journal_stop(handle);
3474                 }
3475         }
3476
3477         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
3478                                  offset, nr_segs,
3479                                  ext4_get_block, NULL);
3480
3481         if (orphan) {
3482                 int err;
3483
3484                 /* Credits for sb + inode write */
3485                 handle = ext4_journal_start(inode, 2);
3486                 if (IS_ERR(handle)) {
3487                         /* This is really bad luck. We've written the data
3488                          * but cannot extend i_size. Bail out and pretend
3489                          * the write failed... */
3490                         ret = PTR_ERR(handle);
3491                         goto out;
3492                 }
3493                 if (inode->i_nlink)
3494                         ext4_orphan_del(handle, inode);
3495                 if (ret > 0) {
3496                         loff_t end = offset + ret;
3497                         if (end > inode->i_size) {
3498                                 ei->i_disksize = end;
3499                                 i_size_write(inode, end);
3500                                 /*
3501                                  * We're going to return a positive `ret'
3502                                  * here due to non-zero-length I/O, so there's
3503                                  * no way of reporting error returns from
3504                                  * ext4_mark_inode_dirty() to userspace.  So
3505                                  * ignore it.
3506                                  */
3507                                 ext4_mark_inode_dirty(handle, inode);
3508                         }
3509                 }
3510                 err = ext4_journal_stop(handle);
3511                 if (ret == 0)
3512                         ret = err;
3513         }
3514 out:
3515         return ret;
3516 }
3517
3518 /*
3519  * Pages can be marked dirty completely asynchronously from ext4's journalling
3520  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3521  * much here because ->set_page_dirty is called under VFS locks.  The page is
3522  * not necessarily locked.
3523  *
3524  * We cannot just dirty the page and leave attached buffers clean, because the
3525  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3526  * or jbddirty because all the journalling code will explode.
3527  *
3528  * So what we do is to mark the page "pending dirty" and next time writepage
3529  * is called, propagate that into the buffers appropriately.
3530  */
3531 static int ext4_journalled_set_page_dirty(struct page *page)
3532 {
3533         SetPageChecked(page);
3534         return __set_page_dirty_nobuffers(page);
3535 }
3536
3537 static const struct address_space_operations ext4_ordered_aops = {
3538         .readpage               = ext4_readpage,
3539         .readpages              = ext4_readpages,
3540         .writepage              = ext4_normal_writepage,
3541         .sync_page              = block_sync_page,
3542         .write_begin            = ext4_write_begin,
3543         .write_end              = ext4_ordered_write_end,
3544         .bmap                   = ext4_bmap,
3545         .invalidatepage         = ext4_invalidatepage,
3546         .releasepage            = ext4_releasepage,
3547         .direct_IO              = ext4_direct_IO,
3548         .migratepage            = buffer_migrate_page,
3549         .is_partially_uptodate  = block_is_partially_uptodate,
3550 };
3551
3552 static const struct address_space_operations ext4_writeback_aops = {
3553         .readpage               = ext4_readpage,
3554         .readpages              = ext4_readpages,
3555         .writepage              = ext4_normal_writepage,
3556         .sync_page              = block_sync_page,
3557         .write_begin            = ext4_write_begin,
3558         .write_end              = ext4_writeback_write_end,
3559         .bmap                   = ext4_bmap,
3560         .invalidatepage         = ext4_invalidatepage,
3561         .releasepage            = ext4_releasepage,
3562         .direct_IO              = ext4_direct_IO,
3563         .migratepage            = buffer_migrate_page,
3564         .is_partially_uptodate  = block_is_partially_uptodate,
3565 };
3566
3567 static const struct address_space_operations ext4_journalled_aops = {
3568         .readpage               = ext4_readpage,
3569         .readpages              = ext4_readpages,
3570         .writepage              = ext4_journalled_writepage,
3571         .sync_page              = block_sync_page,
3572         .write_begin            = ext4_write_begin,
3573         .write_end              = ext4_journalled_write_end,
3574         .set_page_dirty         = ext4_journalled_set_page_dirty,
3575         .bmap                   = ext4_bmap,
3576         .invalidatepage         = ext4_invalidatepage,
3577         .releasepage            = ext4_releasepage,
3578         .is_partially_uptodate  = block_is_partially_uptodate,
3579 };
3580
3581 static const struct address_space_operations ext4_da_aops = {
3582         .readpage               = ext4_readpage,
3583         .readpages              = ext4_readpages,
3584         .writepage              = ext4_da_writepage,
3585         .writepages             = ext4_da_writepages,
3586         .sync_page              = block_sync_page,
3587         .write_begin            = ext4_da_write_begin,
3588         .write_end              = ext4_da_write_end,
3589         .bmap                   = ext4_bmap,
3590         .invalidatepage         = ext4_da_invalidatepage,
3591         .releasepage            = ext4_releasepage,
3592         .direct_IO              = ext4_direct_IO,
3593         .migratepage            = buffer_migrate_page,
3594         .is_partially_uptodate  = block_is_partially_uptodate,
3595 };
3596
3597 void ext4_set_aops(struct inode *inode)
3598 {
3599         if (ext4_should_order_data(inode) &&
3600                 test_opt(inode->i_sb, DELALLOC))
3601                 inode->i_mapping->a_ops = &ext4_da_aops;
3602         else if (ext4_should_order_data(inode))
3603                 inode->i_mapping->a_ops = &ext4_ordered_aops;
3604         else if (ext4_should_writeback_data(inode) &&
3605                  test_opt(inode->i_sb, DELALLOC))
3606                 inode->i_mapping->a_ops = &ext4_da_aops;
3607         else if (ext4_should_writeback_data(inode))
3608                 inode->i_mapping->a_ops = &ext4_writeback_aops;
3609         else
3610                 inode->i_mapping->a_ops = &ext4_journalled_aops;
3611 }
3612
3613 /*
3614  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3615  * up to the end of the block which corresponds to `from'.
3616  * This required during truncate. We need to physically zero the tail end
3617  * of that block so it doesn't yield old data if the file is later grown.
3618  */
3619 int ext4_block_truncate_page(handle_t *handle,
3620                 struct address_space *mapping, loff_t from)
3621 {
3622         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3623         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3624         unsigned blocksize, length, pos;
3625         ext4_lblk_t iblock;
3626         struct inode *inode = mapping->host;
3627         struct buffer_head *bh;
3628         struct page *page;
3629         int err = 0;
3630
3631         page = grab_cache_page(mapping, from >> PAGE_CACHE_SHIFT);
3632         if (!page)
3633                 return -EINVAL;
3634
3635         blocksize = inode->i_sb->s_blocksize;
3636         length = blocksize - (offset & (blocksize - 1));
3637         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3638
3639         /*
3640          * For "nobh" option,  we can only work if we don't need to
3641          * read-in the page - otherwise we create buffers to do the IO.
3642          */
3643         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
3644              ext4_should_writeback_data(inode) && PageUptodate(page)) {
3645                 zero_user(page, offset, length);
3646                 set_page_dirty(page);
3647                 goto unlock;
3648         }
3649
3650         if (!page_has_buffers(page))
3651                 create_empty_buffers(page, blocksize, 0);
3652
3653         /* Find the buffer that contains "offset" */
3654         bh = page_buffers(page);
3655         pos = blocksize;
3656         while (offset >= pos) {
3657                 bh = bh->b_this_page;
3658                 iblock++;
3659                 pos += blocksize;
3660         }
3661
3662         err = 0;
3663         if (buffer_freed(bh)) {
3664                 BUFFER_TRACE(bh, "freed: skip");
3665                 goto unlock;
3666         }
3667
3668         if (!buffer_mapped(bh)) {
3669                 BUFFER_TRACE(bh, "unmapped");
3670                 ext4_get_block(inode, iblock, bh, 0);
3671                 /* unmapped? It's a hole - nothing to do */
3672                 if (!buffer_mapped(bh)) {
3673                         BUFFER_TRACE(bh, "still unmapped");
3674                         goto unlock;
3675                 }
3676         }
3677
3678         /* Ok, it's mapped. Make sure it's up-to-date */
3679         if (PageUptodate(page))
3680                 set_buffer_uptodate(bh);
3681
3682         if (!buffer_uptodate(bh)) {
3683                 err = -EIO;
3684                 ll_rw_block(READ, 1, &bh);
3685                 wait_on_buffer(bh);
3686                 /* Uhhuh. Read error. Complain and punt. */
3687                 if (!buffer_uptodate(bh))
3688                         goto unlock;
3689         }
3690
3691         if (ext4_should_journal_data(inode)) {
3692                 BUFFER_TRACE(bh, "get write access");
3693                 err = ext4_journal_get_write_access(handle, bh);
3694                 if (err)
3695                         goto unlock;
3696         }
3697
3698         zero_user(page, offset, length);
3699
3700         BUFFER_TRACE(bh, "zeroed end of block");
3701
3702         err = 0;
3703         if (ext4_should_journal_data(inode)) {
3704                 err = ext4_handle_dirty_metadata(handle, inode, bh);
3705         } else {
3706                 if (ext4_should_order_data(inode))
3707                         err = ext4_jbd2_file_inode(handle, inode);
3708                 mark_buffer_dirty(bh);
3709         }
3710
3711 unlock:
3712         unlock_page(page);
3713         page_cache_release(page);
3714         return err;
3715 }
3716
3717 /*
3718  * Probably it should be a library function... search for first non-zero word
3719  * or memcmp with zero_page, whatever is better for particular architecture.
3720  * Linus?
3721  */
3722 static inline int all_zeroes(__le32 *p, __le32 *q)
3723 {
3724         while (p < q)
3725                 if (*p++)
3726                         return 0;
3727         return 1;
3728 }
3729
3730 /**
3731  *      ext4_find_shared - find the indirect blocks for partial truncation.
3732  *      @inode:   inode in question
3733  *      @depth:   depth of the affected branch
3734  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
3735  *      @chain:   place to store the pointers to partial indirect blocks
3736  *      @top:     place to the (detached) top of branch
3737  *
3738  *      This is a helper function used by ext4_truncate().
3739  *
3740  *      When we do truncate() we may have to clean the ends of several
3741  *      indirect blocks but leave the blocks themselves alive. Block is
3742  *      partially truncated if some data below the new i_size is refered
3743  *      from it (and it is on the path to the first completely truncated
3744  *      data block, indeed).  We have to free the top of that path along
3745  *      with everything to the right of the path. Since no allocation
3746  *      past the truncation point is possible until ext4_truncate()
3747  *      finishes, we may safely do the latter, but top of branch may
3748  *      require special attention - pageout below the truncation point
3749  *      might try to populate it.
3750  *
3751  *      We atomically detach the top of branch from the tree, store the
3752  *      block number of its root in *@top, pointers to buffer_heads of
3753  *      partially truncated blocks - in @chain[].bh and pointers to
3754  *      their last elements that should not be removed - in
3755  *      @chain[].p. Return value is the pointer to last filled element
3756  *      of @chain.
3757  *
3758  *      The work left to caller to do the actual freeing of subtrees:
3759  *              a) free the subtree starting from *@top
3760  *              b) free the subtrees whose roots are stored in
3761  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
3762  *              c) free the subtrees growing from the inode past the @chain[0].
3763  *                      (no partially truncated stuff there).  */
3764
3765 static Indirect *ext4_find_shared(struct inode *inode, int depth,
3766                         ext4_lblk_t offsets[4], Indirect chain[4], __le32 *top)
3767 {
3768         Indirect *partial, *p;
3769         int k, err;
3770
3771         *top = 0;
3772         /* Make k index the deepest non-null offest + 1 */
3773         for (k = depth; k > 1 && !offsets[k-1]; k--)
3774                 ;
3775         partial = ext4_get_branch(inode, k, offsets, chain, &err);
3776         /* Writer: pointers */
3777         if (!partial)
3778                 partial = chain + k-1;
3779         /*
3780          * If the branch acquired continuation since we've looked at it -
3781          * fine, it should all survive and (new) top doesn't belong to us.
3782          */
3783         if (!partial->key && *partial->p)
3784                 /* Writer: end */
3785                 goto no_top;
3786         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
3787                 ;
3788         /*
3789          * OK, we've found the last block that must survive. The rest of our
3790          * branch should be detached before unlocking. However, if that rest
3791          * of branch is all ours and does not grow immediately from the inode
3792          * it's easier to cheat and just decrement partial->p.
3793          */
3794         if (p == chain + k - 1 && p > chain) {
3795                 p->p--;
3796         } else {
3797                 *top = *p->p;
3798                 /* Nope, don't do this in ext4.  Must leave the tree intact */
3799 #if 0
3800                 *p->p = 0;
3801 #endif
3802         }
3803         /* Writer: end */
3804
3805         while (partial > p) {
3806                 brelse(partial->bh);
3807                 partial--;
3808         }
3809 no_top:
3810         return partial;
3811 }
3812
3813 /*
3814  * Zero a number of block pointers in either an inode or an indirect block.
3815  * If we restart the transaction we must again get write access to the
3816  * indirect block for further modification.
3817  *
3818  * We release `count' blocks on disk, but (last - first) may be greater
3819  * than `count' because there can be holes in there.
3820  */
3821 static void ext4_clear_blocks(handle_t *handle, struct inode *inode,
3822                 struct buffer_head *bh, ext4_fsblk_t block_to_free,
3823                 unsigned long count, __le32 *first, __le32 *last)
3824 {
3825         __le32 *p;
3826         if (try_to_extend_transaction(handle, inode)) {
3827                 if (bh) {
3828                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
3829                         ext4_handle_dirty_metadata(handle, inode, bh);
3830                 }
3831                 ext4_mark_inode_dirty(handle, inode);
3832                 ext4_journal_test_restart(handle, inode);
3833                 if (bh) {
3834                         BUFFER_TRACE(bh, "retaking write access");
3835                         ext4_journal_get_write_access(handle, bh);
3836                 }
3837         }
3838
3839         /*
3840          * Any buffers which are on the journal will be in memory. We find
3841          * them on the hash table so jbd2_journal_revoke() will run jbd2_journal_forget()
3842          * on them.  We've already detached each block from the file, so
3843          * bforget() in jbd2_journal_forget() should be safe.
3844          *
3845          * AKPM: turn on bforget in jbd2_journal_forget()!!!
3846          */
3847         for (p = first; p < last; p++) {
3848                 u32 nr = le32_to_cpu(*p);
3849                 if (nr) {
3850                         struct buffer_head *tbh;
3851
3852                         *p = 0;
3853                         tbh = sb_find_get_block(inode->i_sb, nr);
3854                         ext4_forget(handle, 0, inode, tbh, nr);
3855                 }
3856         }
3857
3858         ext4_free_blocks(handle, inode, block_to_free, count, 0);
3859 }
3860
3861 /**
3862  * ext4_free_data - free a list of data blocks
3863  * @handle:     handle for this transaction
3864  * @inode:      inode we are dealing with
3865  * @this_bh:    indirect buffer_head which contains *@first and *@last
3866  * @first:      array of block numbers
3867  * @last:       points immediately past the end of array
3868  *
3869  * We are freeing all blocks refered from that array (numbers are stored as
3870  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
3871  *
3872  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
3873  * blocks are contiguous then releasing them at one time will only affect one
3874  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
3875  * actually use a lot of journal space.
3876  *
3877  * @this_bh will be %NULL if @first and @last point into the inode's direct
3878  * block pointers.
3879  */
3880 static void ext4_free_data(handle_t *handle, struct inode *inode,
3881                            struct buffer_head *this_bh,
3882                            __le32 *first, __le32 *last)
3883 {
3884         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
3885         unsigned long count = 0;            /* Number of blocks in the run */
3886         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
3887                                                corresponding to
3888                                                block_to_free */
3889         ext4_fsblk_t nr;                    /* Current block # */
3890         __le32 *p;                          /* Pointer into inode/ind
3891                                                for current block */
3892         int err;
3893
3894         if (this_bh) {                          /* For indirect block */
3895                 BUFFER_TRACE(this_bh, "get_write_access");
3896                 err = ext4_journal_get_write_access(handle, this_bh);
3897                 /* Important: if we can't update the indirect pointers
3898                  * to the blocks, we can't free them. */
3899                 if (err)
3900                         return;
3901         }
3902
3903         for (p = first; p < last; p++) {
3904                 nr = le32_to_cpu(*p);
3905                 if (nr) {
3906                         /* accumulate blocks to free if they're contiguous */
3907                         if (count == 0) {
3908                                 block_to_free = nr;
3909                                 block_to_free_p = p;
3910                                 count = 1;
3911                         } else if (nr == block_to_free + count) {
3912                                 count++;
3913                         } else {
3914                                 ext4_clear_blocks(handle, inode, this_bh,
3915                                                   block_to_free,
3916                                                   count, block_to_free_p, p);
3917                                 block_to_free = nr;
3918                                 block_to_free_p = p;
3919                                 count = 1;
3920                         }
3921                 }
3922         }
3923
3924         if (count > 0)
3925                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
3926                                   count, block_to_free_p, p);
3927
3928         if (this_bh) {
3929                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
3930
3931                 /*
3932                  * The buffer head should have an attached journal head at this
3933                  * point. However, if the data is corrupted and an indirect
3934                  * block pointed to itself, it would have been detached when
3935                  * the block was cleared. Check for this instead of OOPSing.
3936                  */
3937                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
3938                         ext4_handle_dirty_metadata(handle, inode, this_bh);
3939                 else
3940                         ext4_error(inode->i_sb, __func__,
3941                                    "circular indirect block detected, "
3942                                    "inode=%lu, block=%llu",
3943                                    inode->i_ino,
3944                                    (unsigned long long) this_bh->b_blocknr);
3945         }
3946 }
3947
3948 /**
3949  *      ext4_free_branches - free an array of branches
3950  *      @handle: JBD handle for this transaction
3951  *      @inode: inode we are dealing with
3952  *      @parent_bh: the buffer_head which contains *@first and *@last
3953  *      @first: array of block numbers
3954  *      @last:  pointer immediately past the end of array
3955  *      @depth: depth of the branches to free
3956  *
3957  *      We are freeing all blocks refered from these branches (numbers are
3958  *      stored as little-endian 32-bit) and updating @inode->i_blocks
3959  *      appropriately.
3960  */
3961 static void ext4_free_branches(handle_t *handle, struct inode *inode,
3962                                struct buffer_head *parent_bh,
3963                                __le32 *first, __le32 *last, int depth)
3964 {
3965         ext4_fsblk_t nr;
3966         __le32 *p;
3967
3968         if (ext4_handle_is_aborted(handle))
3969                 return;
3970
3971         if (depth--) {
3972                 struct buffer_head *bh;
3973                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
3974                 p = last;
3975                 while (--p >= first) {
3976                         nr = le32_to_cpu(*p);
3977                         if (!nr)
3978                                 continue;               /* A hole */
3979
3980                         /* Go read the buffer for the next level down */
3981                         bh = sb_bread(inode->i_sb, nr);
3982
3983                         /*
3984                          * A read failure? Report error and clear slot
3985                          * (should be rare).
3986                          */
3987                         if (!bh) {
3988                                 ext4_error(inode->i_sb, "ext4_free_branches",
3989                                            "Read failure, inode=%lu, block=%llu",
3990                                            inode->i_ino, nr);
3991                                 continue;
3992                         }
3993
3994                         /* This zaps the entire block.  Bottom up. */
3995                         BUFFER_TRACE(bh, "free child branches");
3996                         ext4_free_branches(handle, inode, bh,
3997                                         (__le32 *) bh->b_data,
3998                                         (__le32 *) bh->b_data + addr_per_block,
3999                                         depth);
4000
4001                         /*
4002                          * We've probably journalled the indirect block several
4003                          * times during the truncate.  But it's no longer
4004                          * needed and we now drop it from the transaction via
4005                          * jbd2_journal_revoke().
4006                          *
4007                          * That's easy if it's exclusively part of this
4008                          * transaction.  But if it's part of the committing
4009                          * transaction then jbd2_journal_forget() will simply
4010                          * brelse() it.  That means that if the underlying
4011                          * block is reallocated in ext4_get_block(),
4012                          * unmap_underlying_metadata() will find this block
4013                          * and will try to get rid of it.  damn, damn.
4014                          *
4015                          * If this block has already been committed to the
4016                          * journal, a revoke record will be written.  And
4017                          * revoke records must be emitted *before* clearing
4018                          * this block's bit in the bitmaps.
4019                          */
4020                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
4021
4022                         /*
4023                          * Everything below this this pointer has been
4024                          * released.  Now let this top-of-subtree go.
4025                          *
4026                          * We want the freeing of this indirect block to be
4027                          * atomic in the journal with the updating of the
4028                          * bitmap block which owns it.  So make some room in
4029                          * the journal.
4030                          *
4031                          * We zero the parent pointer *after* freeing its
4032                          * pointee in the bitmaps, so if extend_transaction()
4033                          * for some reason fails to put the bitmap changes and
4034                          * the release into the same transaction, recovery
4035                          * will merely complain about releasing a free block,
4036                          * rather than leaking blocks.
4037                          */
4038                         if (ext4_handle_is_aborted(handle))
4039                                 return;
4040                         if (try_to_extend_transaction(handle, inode)) {
4041                                 ext4_mark_inode_dirty(handle, inode);
4042                                 ext4_journal_test_restart(handle, inode);
4043                         }
4044
4045                         ext4_free_blocks(handle, inode, nr, 1, 1);
4046
4047                         if (parent_bh) {
4048                                 /*
4049                                  * The block which we have just freed is
4050                                  * pointed to by an indirect block: journal it
4051                                  */
4052                                 BUFFER_TRACE(parent_bh, "get_write_access");
4053                                 if (!ext4_journal_get_write_access(handle,
4054                                                                    parent_bh)){
4055                                         *p = 0;
4056                                         BUFFER_TRACE(parent_bh,
4057                                         "call ext4_handle_dirty_metadata");
4058                                         ext4_handle_dirty_metadata(handle,
4059                                                                    inode,
4060                                                                    parent_bh);
4061                                 }
4062                         }
4063                 }
4064         } else {
4065                 /* We have reached the bottom of the tree. */
4066                 BUFFER_TRACE(parent_bh, "free data blocks");
4067                 ext4_free_data(handle, inode, parent_bh, first, last);
4068         }
4069 }
4070
4071 int ext4_can_truncate(struct inode *inode)
4072 {
4073         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4074                 return 0;
4075         if (S_ISREG(inode->i_mode))
4076                 return 1;
4077         if (S_ISDIR(inode->i_mode))
4078                 return 1;
4079         if (S_ISLNK(inode->i_mode))
4080                 return !ext4_inode_is_fast_symlink(inode);
4081         return 0;
4082 }
4083
4084 /*
4085  * ext4_truncate()
4086  *
4087  * We block out ext4_get_block() block instantiations across the entire
4088  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4089  * simultaneously on behalf of the same inode.
4090  *
4091  * As we work through the truncate and commmit bits of it to the journal there
4092  * is one core, guiding principle: the file's tree must always be consistent on
4093  * disk.  We must be able to restart the truncate after a crash.
4094  *
4095  * The file's tree may be transiently inconsistent in memory (although it
4096  * probably isn't), but whenever we close off and commit a journal transaction,
4097  * the contents of (the filesystem + the journal) must be consistent and
4098  * restartable.  It's pretty simple, really: bottom up, right to left (although
4099  * left-to-right works OK too).
4100  *
4101  * Note that at recovery time, journal replay occurs *before* the restart of
4102  * truncate against the orphan inode list.
4103  *
4104  * The committed inode has the new, desired i_size (which is the same as
4105  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4106  * that this inode's truncate did not complete and it will again call
4107  * ext4_truncate() to have another go.  So there will be instantiated blocks
4108  * to the right of the truncation point in a crashed ext4 filesystem.  But
4109  * that's fine - as long as they are linked from the inode, the post-crash
4110  * ext4_truncate() run will find them and release them.
4111  */
4112 void ext4_truncate(struct inode *inode)
4113 {
4114         handle_t *handle;
4115         struct ext4_inode_info *ei = EXT4_I(inode);
4116         __le32 *i_data = ei->i_data;
4117         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4118         struct address_space *mapping = inode->i_mapping;
4119         ext4_lblk_t offsets[4];
4120         Indirect chain[4];
4121         Indirect *partial;
4122         __le32 nr = 0;
4123         int n;
4124         ext4_lblk_t last_block;
4125         unsigned blocksize = inode->i_sb->s_blocksize;
4126
4127         if (!ext4_can_truncate(inode))
4128                 return;
4129
4130         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4131                 ei->i_state |= EXT4_STATE_DA_ALLOC_CLOSE;
4132
4133         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
4134                 ext4_ext_truncate(inode);
4135                 return;
4136         }
4137
4138         handle = start_transaction(inode);
4139         if (IS_ERR(handle))
4140                 return;         /* AKPM: return what? */
4141
4142         last_block = (inode->i_size + blocksize-1)
4143                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4144
4145         if (inode->i_size & (blocksize - 1))
4146                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4147                         goto out_stop;
4148
4149         n = ext4_block_to_path(inode, last_block, offsets, NULL);
4150         if (n == 0)
4151                 goto out_stop;  /* error */
4152
4153         /*
4154          * OK.  This truncate is going to happen.  We add the inode to the
4155          * orphan list, so that if this truncate spans multiple transactions,
4156          * and we crash, we will resume the truncate when the filesystem
4157          * recovers.  It also marks the inode dirty, to catch the new size.
4158          *
4159          * Implication: the file must always be in a sane, consistent
4160          * truncatable state while each transaction commits.
4161          */
4162         if (ext4_orphan_add(handle, inode))
4163                 goto out_stop;
4164
4165         /*
4166          * From here we block out all ext4_get_block() callers who want to
4167          * modify the block allocation tree.
4168          */
4169         down_write(&ei->i_data_sem);
4170
4171         ext4_discard_preallocations(inode);
4172
4173         /*
4174          * The orphan list entry will now protect us from any crash which
4175          * occurs before the truncate completes, so it is now safe to propagate
4176          * the new, shorter inode size (held for now in i_size) into the
4177          * on-disk inode. We do this via i_disksize, which is the value which
4178          * ext4 *really* writes onto the disk inode.
4179          */
4180         ei->i_disksize = inode->i_size;
4181
4182         if (n == 1) {           /* direct blocks */
4183                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4184                                i_data + EXT4_NDIR_BLOCKS);
4185                 goto do_indirects;
4186         }
4187
4188         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4189         /* Kill the top of shared branch (not detached) */
4190         if (nr) {
4191                 if (partial == chain) {
4192                         /* Shared branch grows from the inode */
4193                         ext4_free_branches(handle, inode, NULL,
4194                                            &nr, &nr+1, (chain+n-1) - partial);
4195                         *partial->p = 0;
4196                         /*
4197                          * We mark the inode dirty prior to restart,
4198                          * and prior to stop.  No need for it here.
4199                          */
4200                 } else {
4201                         /* Shared branch grows from an indirect block */
4202                         BUFFER_TRACE(partial->bh, "get_write_access");
4203                         ext4_free_branches(handle, inode, partial->bh,
4204                                         partial->p,
4205                                         partial->p+1, (chain+n-1) - partial);
4206                 }
4207         }
4208         /* Clear the ends of indirect blocks on the shared branch */
4209         while (partial > chain) {
4210                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4211                                    (__le32*)partial->bh->b_data+addr_per_block,
4212                                    (chain+n-1) - partial);
4213                 BUFFER_TRACE(partial->bh, "call brelse");
4214                 brelse (partial->bh);
4215                 partial--;
4216         }
4217 do_indirects:
4218         /* Kill the remaining (whole) subtrees */
4219         switch (offsets[0]) {
4220         default:
4221                 nr = i_data[EXT4_IND_BLOCK];
4222                 if (nr) {
4223                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4224                         i_data[EXT4_IND_BLOCK] = 0;
4225                 }
4226         case EXT4_IND_BLOCK:
4227                 nr = i_data[EXT4_DIND_BLOCK];
4228                 if (nr) {
4229                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4230                         i_data[EXT4_DIND_BLOCK] = 0;
4231                 }
4232         case EXT4_DIND_BLOCK:
4233                 nr = i_data[EXT4_TIND_BLOCK];
4234                 if (nr) {
4235                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4236                         i_data[EXT4_TIND_BLOCK] = 0;
4237                 }
4238         case EXT4_TIND_BLOCK:
4239                 ;
4240         }
4241
4242         up_write(&ei->i_data_sem);
4243         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4244         ext4_mark_inode_dirty(handle, inode);
4245
4246         /*
4247          * In a multi-transaction truncate, we only make the final transaction
4248          * synchronous
4249          */
4250         if (IS_SYNC(inode))
4251                 ext4_handle_sync(handle);
4252 out_stop:
4253         /*
4254          * If this was a simple ftruncate(), and the file will remain alive
4255          * then we need to clear up the orphan record which we created above.
4256          * However, if this was a real unlink then we were called by
4257          * ext4_delete_inode(), and we allow that function to clean up the
4258          * orphan info for us.
4259          */
4260         if (inode->i_nlink)
4261                 ext4_orphan_del(handle, inode);
4262
4263         ext4_journal_stop(handle);
4264 }
4265
4266 /*
4267  * ext4_get_inode_loc returns with an extra refcount against the inode's
4268  * underlying buffer_head on success. If 'in_mem' is true, we have all
4269  * data in memory that is needed to recreate the on-disk version of this
4270  * inode.
4271  */
4272 static int __ext4_get_inode_loc(struct inode *inode,
4273                                 struct ext4_iloc *iloc, int in_mem)
4274 {
4275         struct ext4_group_desc  *gdp;
4276         struct buffer_head      *bh;
4277         struct super_block      *sb = inode->i_sb;
4278         ext4_fsblk_t            block;
4279         int                     inodes_per_block, inode_offset;
4280
4281         iloc->bh = NULL;
4282         if (!ext4_valid_inum(sb, inode->i_ino))
4283                 return -EIO;
4284
4285         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4286         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4287         if (!gdp)
4288                 return -EIO;
4289
4290         /*
4291          * Figure out the offset within the block group inode table
4292          */
4293         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4294         inode_offset = ((inode->i_ino - 1) %
4295                         EXT4_INODES_PER_GROUP(sb));
4296         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4297         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4298
4299         bh = sb_getblk(sb, block);
4300         if (!bh) {
4301                 ext4_error(sb, "ext4_get_inode_loc", "unable to read "
4302                            "inode block - inode=%lu, block=%llu",
4303                            inode->i_ino, block);
4304                 return -EIO;
4305         }
4306         if (!buffer_uptodate(bh)) {
4307                 lock_buffer(bh);
4308
4309                 /*
4310                  * If the buffer has the write error flag, we have failed
4311                  * to write out another inode in the same block.  In this
4312                  * case, we don't have to read the block because we may
4313                  * read the old inode data successfully.
4314                  */
4315                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4316                         set_buffer_uptodate(bh);
4317
4318                 if (buffer_uptodate(bh)) {
4319                         /* someone brought it uptodate while we waited */
4320                         unlock_buffer(bh);
4321                         goto has_buffer;
4322                 }
4323
4324                 /*
4325                  * If we have all information of the inode in memory and this
4326                  * is the only valid inode in the block, we need not read the
4327                  * block.
4328                  */
4329                 if (in_mem) {
4330                         struct buffer_head *bitmap_bh;
4331                         int i, start;
4332
4333                         start = inode_offset & ~(inodes_per_block - 1);
4334
4335                         /* Is the inode bitmap in cache? */
4336                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4337                         if (!bitmap_bh)
4338                                 goto make_io;
4339
4340                         /*
4341                          * If the inode bitmap isn't in cache then the
4342                          * optimisation may end up performing two reads instead
4343                          * of one, so skip it.
4344                          */
4345                         if (!buffer_uptodate(bitmap_bh)) {
4346                                 brelse(bitmap_bh);
4347                                 goto make_io;
4348                         }
4349                         for (i = start; i < start + inodes_per_block; i++) {
4350                                 if (i == inode_offset)
4351                                         continue;
4352                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4353                                         break;
4354                         }
4355                         brelse(bitmap_bh);
4356                         if (i == start + inodes_per_block) {
4357                                 /* all other inodes are free, so skip I/O */
4358                                 memset(bh->b_data, 0, bh->b_size);
4359                                 set_buffer_uptodate(bh);
4360                                 unlock_buffer(bh);
4361                                 goto has_buffer;
4362                         }
4363                 }
4364
4365 make_io:
4366                 /*
4367                  * If we need to do any I/O, try to pre-readahead extra
4368                  * blocks from the inode table.
4369                  */
4370                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4371                         ext4_fsblk_t b, end, table;
4372                         unsigned num;
4373
4374                         table = ext4_inode_table(sb, gdp);
4375                         /* s_inode_readahead_blks is always a power of 2 */
4376                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4377                         if (table > b)
4378                                 b = table;
4379                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4380                         num = EXT4_INODES_PER_GROUP(sb);
4381                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4382                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4383                                 num -= ext4_itable_unused_count(sb, gdp);
4384                         table += num / inodes_per_block;
4385                         if (end > table)
4386                                 end = table;
4387                         while (b <= end)
4388                                 sb_breadahead(sb, b++);
4389                 }
4390
4391                 /*
4392                  * There are other valid inodes in the buffer, this inode
4393                  * has in-inode xattrs, or we don't have this inode in memory.
4394                  * Read the block from disk.
4395                  */
4396                 get_bh(bh);
4397                 bh->b_end_io = end_buffer_read_sync;
4398                 submit_bh(READ_META, bh);
4399                 wait_on_buffer(bh);
4400                 if (!buffer_uptodate(bh)) {
4401                         ext4_error(sb, __func__,
4402                                    "unable to read inode block - inode=%lu, "
4403                                    "block=%llu", inode->i_ino, block);
4404                         brelse(bh);
4405                         return -EIO;
4406                 }
4407         }
4408 has_buffer:
4409         iloc->bh = bh;
4410         return 0;
4411 }
4412
4413 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4414 {
4415         /* We have all inode data except xattrs in memory here. */
4416         return __ext4_get_inode_loc(inode, iloc,
4417                 !(EXT4_I(inode)->i_state & EXT4_STATE_XATTR));
4418 }
4419
4420 void ext4_set_inode_flags(struct inode *inode)
4421 {
4422         unsigned int flags = EXT4_I(inode)->i_flags;
4423
4424         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4425         if (flags & EXT4_SYNC_FL)
4426                 inode->i_flags |= S_SYNC;
4427         if (flags & EXT4_APPEND_FL)
4428                 inode->i_flags |= S_APPEND;
4429         if (flags & EXT4_IMMUTABLE_FL)
4430                 inode->i_flags |= S_IMMUTABLE;
4431         if (flags & EXT4_NOATIME_FL)
4432                 inode->i_flags |= S_NOATIME;
4433         if (flags & EXT4_DIRSYNC_FL)
4434                 inode->i_flags |= S_DIRSYNC;
4435 }
4436
4437 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4438 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4439 {
4440         unsigned int flags = ei->vfs_inode.i_flags;
4441
4442         ei->i_flags &= ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4443                         EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|EXT4_DIRSYNC_FL);
4444         if (flags & S_SYNC)
4445                 ei->i_flags |= EXT4_SYNC_FL;
4446         if (flags & S_APPEND)
4447                 ei->i_flags |= EXT4_APPEND_FL;
4448         if (flags & S_IMMUTABLE)
4449                 ei->i_flags |= EXT4_IMMUTABLE_FL;
4450         if (flags & S_NOATIME)
4451                 ei->i_flags |= EXT4_NOATIME_FL;
4452         if (flags & S_DIRSYNC)
4453                 ei->i_flags |= EXT4_DIRSYNC_FL;
4454 }
4455 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4456                                         struct ext4_inode_info *ei)
4457 {
4458         blkcnt_t i_blocks ;
4459         struct inode *inode = &(ei->vfs_inode);
4460         struct super_block *sb = inode->i_sb;
4461
4462         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4463                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4464                 /* we are using combined 48 bit field */
4465                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4466                                         le32_to_cpu(raw_inode->i_blocks_lo);
4467                 if (ei->i_flags & EXT4_HUGE_FILE_FL) {
4468                         /* i_blocks represent file system block size */
4469                         return i_blocks  << (inode->i_blkbits - 9);
4470                 } else {
4471                         return i_blocks;
4472                 }
4473         } else {
4474                 return le32_to_cpu(raw_inode->i_blocks_lo);
4475         }
4476 }
4477
4478 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4479 {
4480         struct ext4_iloc iloc;
4481         struct ext4_inode *raw_inode;
4482         struct ext4_inode_info *ei;
4483         struct buffer_head *bh;
4484         struct inode *inode;
4485         long ret;
4486         int block;
4487
4488         inode = iget_locked(sb, ino);
4489         if (!inode)
4490                 return ERR_PTR(-ENOMEM);
4491         if (!(inode->i_state & I_NEW))
4492                 return inode;
4493
4494         ei = EXT4_I(inode);
4495 #ifdef CONFIG_EXT4_FS_POSIX_ACL
4496         ei->i_acl = EXT4_ACL_NOT_CACHED;
4497         ei->i_default_acl = EXT4_ACL_NOT_CACHED;
4498 #endif
4499
4500         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4501         if (ret < 0)
4502                 goto bad_inode;
4503         bh = iloc.bh;
4504         raw_inode = ext4_raw_inode(&iloc);
4505         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4506         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4507         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4508         if (!(test_opt(inode->i_sb, NO_UID32))) {
4509                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4510                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4511         }
4512         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
4513
4514         ei->i_state = 0;
4515         ei->i_dir_start_lookup = 0;
4516         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4517         /* We now have enough fields to check if the inode was active or not.
4518          * This is needed because nfsd might try to access dead inodes
4519          * the test is that same one that e2fsck uses
4520          * NeilBrown 1999oct15
4521          */
4522         if (inode->i_nlink == 0) {
4523                 if (inode->i_mode == 0 ||
4524                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
4525                         /* this inode is deleted */
4526                         brelse(bh);
4527                         ret = -ESTALE;
4528                         goto bad_inode;
4529                 }
4530                 /* The only unlinked inodes we let through here have
4531                  * valid i_mode and are being read by the orphan
4532                  * recovery code: that's fine, we're about to complete
4533                  * the process of deleting those. */
4534         }
4535         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4536         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4537         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4538         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
4539                 ei->i_file_acl |=
4540                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4541         inode->i_size = ext4_isize(raw_inode);
4542         ei->i_disksize = inode->i_size;
4543         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4544         ei->i_block_group = iloc.block_group;
4545         ei->i_last_alloc_group = ~0;
4546         /*
4547          * NOTE! The in-memory inode i_data array is in little-endian order
4548          * even on big-endian machines: we do NOT byteswap the block numbers!
4549          */
4550         for (block = 0; block < EXT4_N_BLOCKS; block++)
4551                 ei->i_data[block] = raw_inode->i_block[block];
4552         INIT_LIST_HEAD(&ei->i_orphan);
4553
4554         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4555                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4556                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4557                     EXT4_INODE_SIZE(inode->i_sb)) {
4558                         brelse(bh);
4559                         ret = -EIO;
4560                         goto bad_inode;
4561                 }
4562                 if (ei->i_extra_isize == 0) {
4563                         /* The extra space is currently unused. Use it. */
4564                         ei->i_extra_isize = sizeof(struct ext4_inode) -
4565                                             EXT4_GOOD_OLD_INODE_SIZE;
4566                 } else {
4567                         __le32 *magic = (void *)raw_inode +
4568                                         EXT4_GOOD_OLD_INODE_SIZE +
4569                                         ei->i_extra_isize;
4570                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
4571                                  ei->i_state |= EXT4_STATE_XATTR;
4572                 }
4573         } else
4574                 ei->i_extra_isize = 0;
4575
4576         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4577         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4578         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4579         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4580
4581         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4582         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4583                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4584                         inode->i_version |=
4585                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4586         }
4587
4588         ret = 0;
4589         if (ei->i_file_acl &&
4590             ((ei->i_file_acl < 
4591               (le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block) +
4592                EXT4_SB(sb)->s_gdb_count)) ||
4593              (ei->i_file_acl >= ext4_blocks_count(EXT4_SB(sb)->s_es)))) {
4594                 ext4_error(sb, __func__,
4595                            "bad extended attribute block %llu in inode #%lu",
4596                            ei->i_file_acl, inode->i_ino);
4597                 ret = -EIO;
4598                 goto bad_inode;
4599         } else if (ei->i_flags & EXT4_EXTENTS_FL) {
4600                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4601                     (S_ISLNK(inode->i_mode) &&
4602                      !ext4_inode_is_fast_symlink(inode)))
4603                         /* Validate extent which is part of inode */
4604                         ret = ext4_ext_check_inode(inode);
4605         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4606                    (S_ISLNK(inode->i_mode) &&
4607                     !ext4_inode_is_fast_symlink(inode))) {
4608                 /* Validate block references which are part of inode */
4609                 ret = ext4_check_inode_blockref(inode);
4610         }
4611         if (ret) {
4612                 brelse(bh);
4613                 goto bad_inode;
4614         }
4615
4616         if (S_ISREG(inode->i_mode)) {
4617                 inode->i_op = &ext4_file_inode_operations;
4618                 inode->i_fop = &ext4_file_operations;
4619                 ext4_set_aops(inode);
4620         } else if (S_ISDIR(inode->i_mode)) {
4621                 inode->i_op = &ext4_dir_inode_operations;
4622                 inode->i_fop = &ext4_dir_operations;
4623         } else if (S_ISLNK(inode->i_mode)) {
4624                 if (ext4_inode_is_fast_symlink(inode)) {
4625                         inode->i_op = &ext4_fast_symlink_inode_operations;
4626                         nd_terminate_link(ei->i_data, inode->i_size,
4627                                 sizeof(ei->i_data) - 1);
4628                 } else {
4629                         inode->i_op = &ext4_symlink_inode_operations;
4630                         ext4_set_aops(inode);
4631                 }
4632         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
4633               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
4634                 inode->i_op = &ext4_special_inode_operations;
4635                 if (raw_inode->i_block[0])
4636                         init_special_inode(inode, inode->i_mode,
4637                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4638                 else
4639                         init_special_inode(inode, inode->i_mode,
4640                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4641         } else {
4642                 brelse(bh);
4643                 ret = -EIO;
4644                 ext4_error(inode->i_sb, __func__, 
4645                            "bogus i_mode (%o) for inode=%lu",
4646                            inode->i_mode, inode->i_ino);
4647                 goto bad_inode;
4648         }
4649         brelse(iloc.bh);
4650         ext4_set_inode_flags(inode);
4651         unlock_new_inode(inode);
4652         return inode;
4653
4654 bad_inode:
4655         iget_failed(inode);
4656         return ERR_PTR(ret);
4657 }
4658
4659 static int ext4_inode_blocks_set(handle_t *handle,
4660                                 struct ext4_inode *raw_inode,
4661                                 struct ext4_inode_info *ei)
4662 {
4663         struct inode *inode = &(ei->vfs_inode);
4664         u64 i_blocks = inode->i_blocks;
4665         struct super_block *sb = inode->i_sb;
4666
4667         if (i_blocks <= ~0U) {
4668                 /*
4669                  * i_blocks can be represnted in a 32 bit variable
4670                  * as multiple of 512 bytes
4671                  */
4672                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4673                 raw_inode->i_blocks_high = 0;
4674                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4675                 return 0;
4676         }
4677         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
4678                 return -EFBIG;
4679
4680         if (i_blocks <= 0xffffffffffffULL) {
4681                 /*
4682                  * i_blocks can be represented in a 48 bit variable
4683                  * as multiple of 512 bytes
4684                  */
4685                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4686                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4687                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4688         } else {
4689                 ei->i_flags |= EXT4_HUGE_FILE_FL;
4690                 /* i_block is stored in file system block size */
4691                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4692                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4693                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4694         }
4695         return 0;
4696 }
4697
4698 /*
4699  * Post the struct inode info into an on-disk inode location in the
4700  * buffer-cache.  This gobbles the caller's reference to the
4701  * buffer_head in the inode location struct.
4702  *
4703  * The caller must have write access to iloc->bh.
4704  */
4705 static int ext4_do_update_inode(handle_t *handle,
4706                                 struct inode *inode,
4707                                 struct ext4_iloc *iloc)
4708 {
4709         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
4710         struct ext4_inode_info *ei = EXT4_I(inode);
4711         struct buffer_head *bh = iloc->bh;
4712         int err = 0, rc, block;
4713
4714         /* For fields not not tracking in the in-memory inode,
4715          * initialise them to zero for new inodes. */
4716         if (ei->i_state & EXT4_STATE_NEW)
4717                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
4718
4719         ext4_get_inode_flags(ei);
4720         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
4721         if (!(test_opt(inode->i_sb, NO_UID32))) {
4722                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
4723                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
4724 /*
4725  * Fix up interoperability with old kernels. Otherwise, old inodes get
4726  * re-used with the upper 16 bits of the uid/gid intact
4727  */
4728                 if (!ei->i_dtime) {
4729                         raw_inode->i_uid_high =
4730                                 cpu_to_le16(high_16_bits(inode->i_uid));
4731                         raw_inode->i_gid_high =
4732                                 cpu_to_le16(high_16_bits(inode->i_gid));
4733                 } else {
4734                         raw_inode->i_uid_high = 0;
4735                         raw_inode->i_gid_high = 0;
4736                 }
4737         } else {
4738                 raw_inode->i_uid_low =
4739                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
4740                 raw_inode->i_gid_low =
4741                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
4742                 raw_inode->i_uid_high = 0;
4743                 raw_inode->i_gid_high = 0;
4744         }
4745         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
4746
4747         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
4748         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
4749         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
4750         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
4751
4752         if (ext4_inode_blocks_set(handle, raw_inode, ei))
4753                 goto out_brelse;
4754         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
4755         /* clear the migrate flag in the raw_inode */
4756         raw_inode->i_flags = cpu_to_le32(ei->i_flags & ~EXT4_EXT_MIGRATE);
4757         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
4758             cpu_to_le32(EXT4_OS_HURD))
4759                 raw_inode->i_file_acl_high =
4760                         cpu_to_le16(ei->i_file_acl >> 32);
4761         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
4762         ext4_isize_set(raw_inode, ei->i_disksize);
4763         if (ei->i_disksize > 0x7fffffffULL) {
4764                 struct super_block *sb = inode->i_sb;
4765                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
4766                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
4767                                 EXT4_SB(sb)->s_es->s_rev_level ==
4768                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
4769                         /* If this is the first large file
4770                          * created, add a flag to the superblock.
4771                          */
4772                         err = ext4_journal_get_write_access(handle,
4773                                         EXT4_SB(sb)->s_sbh);
4774                         if (err)
4775                                 goto out_brelse;
4776                         ext4_update_dynamic_rev(sb);
4777                         EXT4_SET_RO_COMPAT_FEATURE(sb,
4778                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
4779                         sb->s_dirt = 1;
4780                         ext4_handle_sync(handle);
4781                         err = ext4_handle_dirty_metadata(handle, inode,
4782                                         EXT4_SB(sb)->s_sbh);
4783                 }
4784         }
4785         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
4786         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
4787                 if (old_valid_dev(inode->i_rdev)) {
4788                         raw_inode->i_block[0] =
4789                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
4790                         raw_inode->i_block[1] = 0;
4791                 } else {
4792                         raw_inode->i_block[0] = 0;
4793                         raw_inode->i_block[1] =
4794                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
4795                         raw_inode->i_block[2] = 0;
4796                 }
4797         } else for (block = 0; block < EXT4_N_BLOCKS; block++)
4798                 raw_inode->i_block[block] = ei->i_data[block];
4799
4800         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
4801         if (ei->i_extra_isize) {
4802                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4803                         raw_inode->i_version_hi =
4804                         cpu_to_le32(inode->i_version >> 32);
4805                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
4806         }
4807
4808         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4809         rc = ext4_handle_dirty_metadata(handle, inode, bh);
4810         if (!err)
4811                 err = rc;
4812         ei->i_state &= ~EXT4_STATE_NEW;
4813
4814 out_brelse:
4815         brelse(bh);
4816         ext4_std_error(inode->i_sb, err);
4817         return err;
4818 }
4819
4820 /*
4821  * ext4_write_inode()
4822  *
4823  * We are called from a few places:
4824  *
4825  * - Within generic_file_write() for O_SYNC files.
4826  *   Here, there will be no transaction running. We wait for any running
4827  *   trasnaction to commit.
4828  *
4829  * - Within sys_sync(), kupdate and such.
4830  *   We wait on commit, if tol to.
4831  *
4832  * - Within prune_icache() (PF_MEMALLOC == true)
4833  *   Here we simply return.  We can't afford to block kswapd on the
4834  *   journal commit.
4835  *
4836  * In all cases it is actually safe for us to return without doing anything,
4837  * because the inode has been copied into a raw inode buffer in
4838  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
4839  * knfsd.
4840  *
4841  * Note that we are absolutely dependent upon all inode dirtiers doing the
4842  * right thing: they *must* call mark_inode_dirty() after dirtying info in
4843  * which we are interested.
4844  *
4845  * It would be a bug for them to not do this.  The code:
4846  *
4847  *      mark_inode_dirty(inode)
4848  *      stuff();
4849  *      inode->i_size = expr;
4850  *
4851  * is in error because a kswapd-driven write_inode() could occur while
4852  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
4853  * will no longer be on the superblock's dirty inode list.
4854  */
4855 int ext4_write_inode(struct inode *inode, int wait)
4856 {
4857         if (current->flags & PF_MEMALLOC)
4858                 return 0;
4859
4860         if (ext4_journal_current_handle()) {
4861                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
4862                 dump_stack();
4863                 return -EIO;
4864         }
4865
4866         if (!wait)
4867                 return 0;
4868
4869         return ext4_force_commit(inode->i_sb);
4870 }
4871
4872 /*
4873  * ext4_setattr()
4874  *
4875  * Called from notify_change.
4876  *
4877  * We want to trap VFS attempts to truncate the file as soon as
4878  * possible.  In particular, we want to make sure that when the VFS
4879  * shrinks i_size, we put the inode on the orphan list and modify
4880  * i_disksize immediately, so that during the subsequent flushing of
4881  * dirty pages and freeing of disk blocks, we can guarantee that any
4882  * commit will leave the blocks being flushed in an unused state on
4883  * disk.  (On recovery, the inode will get truncated and the blocks will
4884  * be freed, so we have a strong guarantee that no future commit will
4885  * leave these blocks visible to the user.)
4886  *
4887  * Another thing we have to assure is that if we are in ordered mode
4888  * and inode is still attached to the committing transaction, we must
4889  * we start writeout of all the dirty pages which are being truncated.
4890  * This way we are sure that all the data written in the previous
4891  * transaction are already on disk (truncate waits for pages under
4892  * writeback).
4893  *
4894  * Called with inode->i_mutex down.
4895  */
4896 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
4897 {
4898         struct inode *inode = dentry->d_inode;
4899         int error, rc = 0;
4900         const unsigned int ia_valid = attr->ia_valid;
4901
4902         error = inode_change_ok(inode, attr);
4903         if (error)
4904                 return error;
4905
4906         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
4907                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
4908                 handle_t *handle;
4909
4910                 /* (user+group)*(old+new) structure, inode write (sb,
4911                  * inode block, ? - but truncate inode update has it) */
4912                 handle = ext4_journal_start(inode, 2*(EXT4_QUOTA_INIT_BLOCKS(inode->i_sb)+
4913                                         EXT4_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
4914                 if (IS_ERR(handle)) {
4915                         error = PTR_ERR(handle);
4916                         goto err_out;
4917                 }
4918                 error = vfs_dq_transfer(inode, attr) ? -EDQUOT : 0;
4919                 if (error) {
4920                         ext4_journal_stop(handle);
4921                         return error;
4922                 }
4923                 /* Update corresponding info in inode so that everything is in
4924                  * one transaction */
4925                 if (attr->ia_valid & ATTR_UID)
4926                         inode->i_uid = attr->ia_uid;
4927                 if (attr->ia_valid & ATTR_GID)
4928                         inode->i_gid = attr->ia_gid;
4929                 error = ext4_mark_inode_dirty(handle, inode);
4930                 ext4_journal_stop(handle);
4931         }
4932
4933         if (attr->ia_valid & ATTR_SIZE) {
4934                 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)) {
4935                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4936
4937                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
4938                                 error = -EFBIG;
4939                                 goto err_out;
4940                         }
4941                 }
4942         }
4943
4944         if (S_ISREG(inode->i_mode) &&
4945             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
4946                 handle_t *handle;
4947
4948                 handle = ext4_journal_start(inode, 3);
4949                 if (IS_ERR(handle)) {
4950                         error = PTR_ERR(handle);
4951                         goto err_out;
4952                 }
4953
4954                 error = ext4_orphan_add(handle, inode);
4955                 EXT4_I(inode)->i_disksize = attr->ia_size;
4956                 rc = ext4_mark_inode_dirty(handle, inode);
4957                 if (!error)
4958                         error = rc;
4959                 ext4_journal_stop(handle);
4960
4961                 if (ext4_should_order_data(inode)) {
4962                         error = ext4_begin_ordered_truncate(inode,
4963                                                             attr->ia_size);
4964                         if (error) {
4965                                 /* Do as much error cleanup as possible */
4966                                 handle = ext4_journal_start(inode, 3);
4967                                 if (IS_ERR(handle)) {
4968                                         ext4_orphan_del(NULL, inode);
4969                                         goto err_out;
4970                                 }
4971                                 ext4_orphan_del(handle, inode);
4972                                 ext4_journal_stop(handle);
4973                                 goto err_out;
4974                         }
4975                 }
4976         }
4977
4978         rc = inode_setattr(inode, attr);
4979
4980         /* If inode_setattr's call to ext4_truncate failed to get a
4981          * transaction handle at all, we need to clean up the in-core
4982          * orphan list manually. */
4983         if (inode->i_nlink)
4984                 ext4_orphan_del(NULL, inode);
4985
4986         if (!rc && (ia_valid & ATTR_MODE))
4987                 rc = ext4_acl_chmod(inode);
4988
4989 err_out:
4990         ext4_std_error(inode->i_sb, error);
4991         if (!error)
4992                 error = rc;
4993         return error;
4994 }
4995
4996 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
4997                  struct kstat *stat)
4998 {
4999         struct inode *inode;
5000         unsigned long delalloc_blocks;
5001
5002         inode = dentry->d_inode;
5003         generic_fillattr(inode, stat);
5004
5005         /*
5006          * We can't update i_blocks if the block allocation is delayed
5007          * otherwise in the case of system crash before the real block
5008          * allocation is done, we will have i_blocks inconsistent with
5009          * on-disk file blocks.
5010          * We always keep i_blocks updated together with real
5011          * allocation. But to not confuse with user, stat
5012          * will return the blocks that include the delayed allocation
5013          * blocks for this file.
5014          */
5015         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
5016         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
5017         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
5018
5019         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
5020         return 0;
5021 }
5022
5023 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
5024                                       int chunk)
5025 {
5026         int indirects;
5027
5028         /* if nrblocks are contiguous */
5029         if (chunk) {
5030                 /*
5031                  * With N contiguous data blocks, it need at most
5032                  * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) indirect blocks
5033                  * 2 dindirect blocks
5034                  * 1 tindirect block
5035                  */
5036                 indirects = nrblocks / EXT4_ADDR_PER_BLOCK(inode->i_sb);
5037                 return indirects + 3;
5038         }
5039         /*
5040          * if nrblocks are not contiguous, worse case, each block touch
5041          * a indirect block, and each indirect block touch a double indirect
5042          * block, plus a triple indirect block
5043          */
5044         indirects = nrblocks * 2 + 1;
5045         return indirects;
5046 }
5047
5048 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5049 {
5050         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL))
5051                 return ext4_indirect_trans_blocks(inode, nrblocks, chunk);
5052         return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
5053 }
5054
5055 /*
5056  * Account for index blocks, block groups bitmaps and block group
5057  * descriptor blocks if modify datablocks and index blocks
5058  * worse case, the indexs blocks spread over different block groups
5059  *
5060  * If datablocks are discontiguous, they are possible to spread over
5061  * different block groups too. If they are contiugous, with flexbg,
5062  * they could still across block group boundary.
5063  *
5064  * Also account for superblock, inode, quota and xattr blocks
5065  */
5066 int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5067 {
5068         ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
5069         int gdpblocks;
5070         int idxblocks;
5071         int ret = 0;
5072
5073         /*
5074          * How many index blocks need to touch to modify nrblocks?
5075          * The "Chunk" flag indicating whether the nrblocks is
5076          * physically contiguous on disk
5077          *
5078          * For Direct IO and fallocate, they calls get_block to allocate
5079          * one single extent at a time, so they could set the "Chunk" flag
5080          */
5081         idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
5082
5083         ret = idxblocks;
5084
5085         /*
5086          * Now let's see how many group bitmaps and group descriptors need
5087          * to account
5088          */
5089         groups = idxblocks;
5090         if (chunk)
5091                 groups += 1;
5092         else
5093                 groups += nrblocks;
5094
5095         gdpblocks = groups;
5096         if (groups > ngroups)
5097                 groups = ngroups;
5098         if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
5099                 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
5100
5101         /* bitmaps and block group descriptor blocks */
5102         ret += groups + gdpblocks;
5103
5104         /* Blocks for super block, inode, quota and xattr blocks */
5105         ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
5106
5107         return ret;
5108 }
5109
5110 /*
5111  * Calulate the total number of credits to reserve to fit
5112  * the modification of a single pages into a single transaction,
5113  * which may include multiple chunks of block allocations.
5114  *
5115  * This could be called via ext4_write_begin()
5116  *
5117  * We need to consider the worse case, when
5118  * one new block per extent.
5119  */
5120 int ext4_writepage_trans_blocks(struct inode *inode)
5121 {
5122         int bpp = ext4_journal_blocks_per_page(inode);
5123         int ret;
5124
5125         ret = ext4_meta_trans_blocks(inode, bpp, 0);
5126
5127         /* Account for data blocks for journalled mode */
5128         if (ext4_should_journal_data(inode))
5129                 ret += bpp;
5130         return ret;
5131 }
5132
5133 /*
5134  * Calculate the journal credits for a chunk of data modification.
5135  *
5136  * This is called from DIO, fallocate or whoever calling
5137  * ext4_get_blocks() to map/allocate a chunk of contigous disk blocks.
5138  *
5139  * journal buffers for data blocks are not included here, as DIO
5140  * and fallocate do no need to journal data buffers.
5141  */
5142 int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
5143 {
5144         return ext4_meta_trans_blocks(inode, nrblocks, 1);
5145 }
5146
5147 /*
5148  * The caller must have previously called ext4_reserve_inode_write().
5149  * Give this, we know that the caller already has write access to iloc->bh.
5150  */
5151 int ext4_mark_iloc_dirty(handle_t *handle,
5152                 struct inode *inode, struct ext4_iloc *iloc)
5153 {
5154         int err = 0;
5155
5156         if (test_opt(inode->i_sb, I_VERSION))
5157                 inode_inc_iversion(inode);
5158
5159         /* the do_update_inode consumes one bh->b_count */
5160         get_bh(iloc->bh);
5161
5162         /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
5163         err = ext4_do_update_inode(handle, inode, iloc);
5164         put_bh(iloc->bh);
5165         return err;
5166 }
5167
5168 /*
5169  * On success, We end up with an outstanding reference count against
5170  * iloc->bh.  This _must_ be cleaned up later.
5171  */
5172
5173 int
5174 ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
5175                          struct ext4_iloc *iloc)
5176 {
5177         int err;
5178
5179         err = ext4_get_inode_loc(inode, iloc);
5180         if (!err) {
5181                 BUFFER_TRACE(iloc->bh, "get_write_access");
5182                 err = ext4_journal_get_write_access(handle, iloc->bh);
5183                 if (err) {
5184                         brelse(iloc->bh);
5185                         iloc->bh = NULL;
5186                 }
5187         }
5188         ext4_std_error(inode->i_sb, err);
5189         return err;
5190 }
5191
5192 /*
5193  * Expand an inode by new_extra_isize bytes.
5194  * Returns 0 on success or negative error number on failure.
5195  */
5196 static int ext4_expand_extra_isize(struct inode *inode,
5197                                    unsigned int new_extra_isize,
5198                                    struct ext4_iloc iloc,
5199                                    handle_t *handle)
5200 {
5201         struct ext4_inode *raw_inode;
5202         struct ext4_xattr_ibody_header *header;
5203         struct ext4_xattr_entry *entry;
5204
5205         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
5206                 return 0;
5207
5208         raw_inode = ext4_raw_inode(&iloc);
5209
5210         header = IHDR(inode, raw_inode);
5211         entry = IFIRST(header);
5212
5213         /* No extended attributes present */
5214         if (!(EXT4_I(inode)->i_state & EXT4_STATE_XATTR) ||
5215                 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
5216                 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
5217                         new_extra_isize);
5218                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
5219                 return 0;
5220         }
5221
5222         /* try to expand with EAs present */
5223         return ext4_expand_extra_isize_ea(inode, new_extra_isize,
5224                                           raw_inode, handle);
5225 }
5226
5227 /*
5228  * What we do here is to mark the in-core inode as clean with respect to inode
5229  * dirtiness (it may still be data-dirty).
5230  * This means that the in-core inode may be reaped by prune_icache
5231  * without having to perform any I/O.  This is a very good thing,
5232  * because *any* task may call prune_icache - even ones which
5233  * have a transaction open against a different journal.
5234  *
5235  * Is this cheating?  Not really.  Sure, we haven't written the
5236  * inode out, but prune_icache isn't a user-visible syncing function.
5237  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
5238  * we start and wait on commits.
5239  *
5240  * Is this efficient/effective?  Well, we're being nice to the system
5241  * by cleaning up our inodes proactively so they can be reaped
5242  * without I/O.  But we are potentially leaving up to five seconds'
5243  * worth of inodes floating about which prune_icache wants us to
5244  * write out.  One way to fix that would be to get prune_icache()
5245  * to do a write_super() to free up some memory.  It has the desired
5246  * effect.
5247  */
5248 int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
5249 {
5250         struct ext4_iloc iloc;
5251         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5252         static unsigned int mnt_count;
5253         int err, ret;
5254
5255         might_sleep();
5256         err = ext4_reserve_inode_write(handle, inode, &iloc);
5257         if (ext4_handle_valid(handle) &&
5258             EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
5259             !(EXT4_I(inode)->i_state & EXT4_STATE_NO_EXPAND)) {
5260                 /*
5261                  * We need extra buffer credits since we may write into EA block
5262                  * with this same handle. If journal_extend fails, then it will
5263                  * only result in a minor loss of functionality for that inode.
5264                  * If this is felt to be critical, then e2fsck should be run to
5265                  * force a large enough s_min_extra_isize.
5266                  */
5267                 if ((jbd2_journal_extend(handle,
5268                              EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
5269                         ret = ext4_expand_extra_isize(inode,
5270                                                       sbi->s_want_extra_isize,
5271                                                       iloc, handle);
5272                         if (ret) {
5273                                 EXT4_I(inode)->i_state |= EXT4_STATE_NO_EXPAND;
5274                                 if (mnt_count !=
5275                                         le16_to_cpu(sbi->s_es->s_mnt_count)) {
5276                                         ext4_warning(inode->i_sb, __func__,
5277                                         "Unable to expand inode %lu. Delete"
5278                                         " some EAs or run e2fsck.",
5279                                         inode->i_ino);
5280                                         mnt_count =
5281                                           le16_to_cpu(sbi->s_es->s_mnt_count);
5282                                 }
5283                         }
5284                 }
5285         }
5286         if (!err)
5287                 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
5288         return err;
5289 }
5290
5291 /*
5292  * ext4_dirty_inode() is called from __mark_inode_dirty()
5293  *
5294  * We're really interested in the case where a file is being extended.
5295  * i_size has been changed by generic_commit_write() and we thus need
5296  * to include the updated inode in the current transaction.
5297  *
5298  * Also, vfs_dq_alloc_block() will always dirty the inode when blocks
5299  * are allocated to the file.
5300  *
5301  * If the inode is marked synchronous, we don't honour that here - doing
5302  * so would cause a commit on atime updates, which we don't bother doing.
5303  * We handle synchronous inodes at the highest possible level.
5304  */
5305 void ext4_dirty_inode(struct inode *inode)
5306 {
5307         handle_t *current_handle = ext4_journal_current_handle();
5308         handle_t *handle;
5309
5310         if (!ext4_handle_valid(current_handle)) {
5311                 ext4_mark_inode_dirty(current_handle, inode);
5312                 return;
5313         }
5314
5315         handle = ext4_journal_start(inode, 2);
5316         if (IS_ERR(handle))
5317                 goto out;
5318         if (current_handle &&
5319                 current_handle->h_transaction != handle->h_transaction) {
5320                 /* This task has a transaction open against a different fs */
5321                 printk(KERN_EMERG "%s: transactions do not match!\n",
5322                        __func__);
5323         } else {
5324                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
5325                                 current_handle);
5326                 ext4_mark_inode_dirty(handle, inode);
5327         }
5328         ext4_journal_stop(handle);
5329 out:
5330         return;
5331 }
5332
5333 #if 0
5334 /*
5335  * Bind an inode's backing buffer_head into this transaction, to prevent
5336  * it from being flushed to disk early.  Unlike
5337  * ext4_reserve_inode_write, this leaves behind no bh reference and
5338  * returns no iloc structure, so the caller needs to repeat the iloc
5339  * lookup to mark the inode dirty later.
5340  */
5341 static int ext4_pin_inode(handle_t *handle, struct inode *inode)
5342 {
5343         struct ext4_iloc iloc;
5344
5345         int err = 0;
5346         if (handle) {
5347                 err = ext4_get_inode_loc(inode, &iloc);
5348                 if (!err) {
5349                         BUFFER_TRACE(iloc.bh, "get_write_access");
5350                         err = jbd2_journal_get_write_access(handle, iloc.bh);
5351                         if (!err)
5352                                 err = ext4_handle_dirty_metadata(handle,
5353                                                                  inode,
5354                                                                  iloc.bh);
5355                         brelse(iloc.bh);
5356                 }
5357         }
5358         ext4_std_error(inode->i_sb, err);
5359         return err;
5360 }
5361 #endif
5362
5363 int ext4_change_inode_journal_flag(struct inode *inode, int val)
5364 {
5365         journal_t *journal;
5366         handle_t *handle;
5367         int err;
5368
5369         /*
5370          * We have to be very careful here: changing a data block's
5371          * journaling status dynamically is dangerous.  If we write a
5372          * data block to the journal, change the status and then delete
5373          * that block, we risk forgetting to revoke the old log record
5374          * from the journal and so a subsequent replay can corrupt data.
5375          * So, first we make sure that the journal is empty and that
5376          * nobody is changing anything.
5377          */
5378
5379         journal = EXT4_JOURNAL(inode);
5380         if (!journal)
5381                 return 0;
5382         if (is_journal_aborted(journal))
5383                 return -EROFS;
5384
5385         jbd2_journal_lock_updates(journal);
5386         jbd2_journal_flush(journal);
5387
5388         /*
5389          * OK, there are no updates running now, and all cached data is
5390          * synced to disk.  We are now in a completely consistent state
5391          * which doesn't have anything in the journal, and we know that
5392          * no filesystem updates are running, so it is safe to modify
5393          * the inode's in-core data-journaling state flag now.
5394          */
5395
5396         if (val)
5397                 EXT4_I(inode)->i_flags |= EXT4_JOURNAL_DATA_FL;
5398         else
5399                 EXT4_I(inode)->i_flags &= ~EXT4_JOURNAL_DATA_FL;
5400         ext4_set_aops(inode);
5401
5402         jbd2_journal_unlock_updates(journal);
5403
5404         /* Finally we can mark the inode as dirty. */
5405
5406         handle = ext4_journal_start(inode, 1);
5407         if (IS_ERR(handle))
5408                 return PTR_ERR(handle);
5409
5410         err = ext4_mark_inode_dirty(handle, inode);
5411         ext4_handle_sync(handle);
5412         ext4_journal_stop(handle);
5413         ext4_std_error(inode->i_sb, err);
5414
5415         return err;
5416 }
5417
5418 static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
5419 {
5420         return !buffer_mapped(bh);
5421 }
5422
5423 int ext4_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
5424 {
5425         struct page *page = vmf->page;
5426         loff_t size;
5427         unsigned long len;
5428         int ret = -EINVAL;
5429         void *fsdata;
5430         struct file *file = vma->vm_file;
5431         struct inode *inode = file->f_path.dentry->d_inode;
5432         struct address_space *mapping = inode->i_mapping;
5433
5434         /*
5435          * Get i_alloc_sem to stop truncates messing with the inode. We cannot
5436          * get i_mutex because we are already holding mmap_sem.
5437          */
5438         down_read(&inode->i_alloc_sem);
5439         size = i_size_read(inode);
5440         if (page->mapping != mapping || size <= page_offset(page)
5441             || !PageUptodate(page)) {
5442                 /* page got truncated from under us? */
5443                 goto out_unlock;
5444         }
5445         ret = 0;
5446         if (PageMappedToDisk(page))
5447                 goto out_unlock;
5448
5449         if (page->index == size >> PAGE_CACHE_SHIFT)
5450                 len = size & ~PAGE_CACHE_MASK;
5451         else
5452                 len = PAGE_CACHE_SIZE;
5453
5454         if (page_has_buffers(page)) {
5455                 /* return if we have all the buffers mapped */
5456                 if (!walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
5457                                        ext4_bh_unmapped))
5458                         goto out_unlock;
5459         }
5460         /*
5461          * OK, we need to fill the hole... Do write_begin write_end
5462          * to do block allocation/reservation.We are not holding
5463          * inode.i__mutex here. That allow * parallel write_begin,
5464          * write_end call. lock_page prevent this from happening
5465          * on the same page though
5466          */
5467         ret = mapping->a_ops->write_begin(file, mapping, page_offset(page),
5468                         len, AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
5469         if (ret < 0)
5470                 goto out_unlock;
5471         ret = mapping->a_ops->write_end(file, mapping, page_offset(page),
5472                         len, len, page, fsdata);
5473         if (ret < 0)
5474                 goto out_unlock;
5475         ret = 0;
5476 out_unlock:
5477         if (ret)
5478                 ret = VM_FAULT_SIGBUS;
5479         up_read(&inode->i_alloc_sem);
5480         return ret;
5481 }