Merge branch 'topic/usb-caiaq' into for-linus
[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 /**
375  *      ext4_get_branch - read the chain of indirect blocks leading to data
376  *      @inode: inode in question
377  *      @depth: depth of the chain (1 - direct pointer, etc.)
378  *      @offsets: offsets of pointers in inode/indirect blocks
379  *      @chain: place to store the result
380  *      @err: here we store the error value
381  *
382  *      Function fills the array of triples <key, p, bh> and returns %NULL
383  *      if everything went OK or the pointer to the last filled triple
384  *      (incomplete one) otherwise. Upon the return chain[i].key contains
385  *      the number of (i+1)-th block in the chain (as it is stored in memory,
386  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
387  *      number (it points into struct inode for i==0 and into the bh->b_data
388  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
389  *      block for i>0 and NULL for i==0. In other words, it holds the block
390  *      numbers of the chain, addresses they were taken from (and where we can
391  *      verify that chain did not change) and buffer_heads hosting these
392  *      numbers.
393  *
394  *      Function stops when it stumbles upon zero pointer (absent block)
395  *              (pointer to last triple returned, *@err == 0)
396  *      or when it gets an IO error reading an indirect block
397  *              (ditto, *@err == -EIO)
398  *      or when it reads all @depth-1 indirect blocks successfully and finds
399  *      the whole chain, all way to the data (returns %NULL, *err == 0).
400  *
401  *      Need to be called with
402  *      down_read(&EXT4_I(inode)->i_data_sem)
403  */
404 static Indirect *ext4_get_branch(struct inode *inode, int depth,
405                                  ext4_lblk_t  *offsets,
406                                  Indirect chain[4], int *err)
407 {
408         struct super_block *sb = inode->i_sb;
409         Indirect *p = chain;
410         struct buffer_head *bh;
411
412         *err = 0;
413         /* i_data is not going away, no lock needed */
414         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
415         if (!p->key)
416                 goto no_block;
417         while (--depth) {
418                 bh = sb_bread(sb, le32_to_cpu(p->key));
419                 if (!bh)
420                         goto failure;
421                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
422                 /* Reader: end */
423                 if (!p->key)
424                         goto no_block;
425         }
426         return NULL;
427
428 failure:
429         *err = -EIO;
430 no_block:
431         return p;
432 }
433
434 /**
435  *      ext4_find_near - find a place for allocation with sufficient locality
436  *      @inode: owner
437  *      @ind: descriptor of indirect block.
438  *
439  *      This function returns the preferred place for block allocation.
440  *      It is used when heuristic for sequential allocation fails.
441  *      Rules are:
442  *        + if there is a block to the left of our position - allocate near it.
443  *        + if pointer will live in indirect block - allocate near that block.
444  *        + if pointer will live in inode - allocate in the same
445  *          cylinder group.
446  *
447  * In the latter case we colour the starting block by the callers PID to
448  * prevent it from clashing with concurrent allocations for a different inode
449  * in the same block group.   The PID is used here so that functionally related
450  * files will be close-by on-disk.
451  *
452  *      Caller must make sure that @ind is valid and will stay that way.
453  */
454 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
455 {
456         struct ext4_inode_info *ei = EXT4_I(inode);
457         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
458         __le32 *p;
459         ext4_fsblk_t bg_start;
460         ext4_fsblk_t last_block;
461         ext4_grpblk_t colour;
462
463         /* Try to find previous block */
464         for (p = ind->p - 1; p >= start; p--) {
465                 if (*p)
466                         return le32_to_cpu(*p);
467         }
468
469         /* No such thing, so let's try location of indirect block */
470         if (ind->bh)
471                 return ind->bh->b_blocknr;
472
473         /*
474          * It is going to be referred to from the inode itself? OK, just put it
475          * into the same cylinder group then.
476          */
477         bg_start = ext4_group_first_block_no(inode->i_sb, ei->i_block_group);
478         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
479
480         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
481                 colour = (current->pid % 16) *
482                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
483         else
484                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
485         return bg_start + colour;
486 }
487
488 /**
489  *      ext4_find_goal - find a preferred place for allocation.
490  *      @inode: owner
491  *      @block:  block we want
492  *      @partial: pointer to the last triple within a chain
493  *
494  *      Normally this function find the preferred place for block allocation,
495  *      returns it.
496  */
497 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
498                 Indirect *partial)
499 {
500         /*
501          * XXX need to get goal block from mballoc's data structures
502          */
503
504         return ext4_find_near(inode, partial);
505 }
506
507 /**
508  *      ext4_blks_to_allocate: Look up the block map and count the number
509  *      of direct blocks need to be allocated for the given branch.
510  *
511  *      @branch: chain of indirect blocks
512  *      @k: number of blocks need for indirect blocks
513  *      @blks: number of data blocks to be mapped.
514  *      @blocks_to_boundary:  the offset in the indirect block
515  *
516  *      return the total number of blocks to be allocate, including the
517  *      direct and indirect blocks.
518  */
519 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
520                 int blocks_to_boundary)
521 {
522         unsigned int count = 0;
523
524         /*
525          * Simple case, [t,d]Indirect block(s) has not allocated yet
526          * then it's clear blocks on that path have not allocated
527          */
528         if (k > 0) {
529                 /* right now we don't handle cross boundary allocation */
530                 if (blks < blocks_to_boundary + 1)
531                         count += blks;
532                 else
533                         count += blocks_to_boundary + 1;
534                 return count;
535         }
536
537         count++;
538         while (count < blks && count <= blocks_to_boundary &&
539                 le32_to_cpu(*(branch[0].p + count)) == 0) {
540                 count++;
541         }
542         return count;
543 }
544
545 /**
546  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
547  *      @indirect_blks: the number of blocks need to allocate for indirect
548  *                      blocks
549  *
550  *      @new_blocks: on return it will store the new block numbers for
551  *      the indirect blocks(if needed) and the first direct block,
552  *      @blks:  on return it will store the total number of allocated
553  *              direct blocks
554  */
555 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
556                                 ext4_lblk_t iblock, ext4_fsblk_t goal,
557                                 int indirect_blks, int blks,
558                                 ext4_fsblk_t new_blocks[4], int *err)
559 {
560         struct ext4_allocation_request ar;
561         int target, i;
562         unsigned long count = 0, blk_allocated = 0;
563         int index = 0;
564         ext4_fsblk_t current_block = 0;
565         int ret = 0;
566
567         /*
568          * Here we try to allocate the requested multiple blocks at once,
569          * on a best-effort basis.
570          * To build a branch, we should allocate blocks for
571          * the indirect blocks(if not allocated yet), and at least
572          * the first direct block of this branch.  That's the
573          * minimum number of blocks need to allocate(required)
574          */
575         /* first we try to allocate the indirect blocks */
576         target = indirect_blks;
577         while (target > 0) {
578                 count = target;
579                 /* allocating blocks for indirect blocks and direct blocks */
580                 current_block = ext4_new_meta_blocks(handle, inode,
581                                                         goal, &count, err);
582                 if (*err)
583                         goto failed_out;
584
585                 target -= count;
586                 /* allocate blocks for indirect blocks */
587                 while (index < indirect_blks && count) {
588                         new_blocks[index++] = current_block++;
589                         count--;
590                 }
591                 if (count > 0) {
592                         /*
593                          * save the new block number
594                          * for the first direct block
595                          */
596                         new_blocks[index] = current_block;
597                         printk(KERN_INFO "%s returned more blocks than "
598                                                 "requested\n", __func__);
599                         WARN_ON(1);
600                         break;
601                 }
602         }
603
604         target = blks - count ;
605         blk_allocated = count;
606         if (!target)
607                 goto allocated;
608         /* Now allocate data blocks */
609         memset(&ar, 0, sizeof(ar));
610         ar.inode = inode;
611         ar.goal = goal;
612         ar.len = target;
613         ar.logical = iblock;
614         if (S_ISREG(inode->i_mode))
615                 /* enable in-core preallocation only for regular files */
616                 ar.flags = EXT4_MB_HINT_DATA;
617
618         current_block = ext4_mb_new_blocks(handle, &ar, err);
619
620         if (*err && (target == blks)) {
621                 /*
622                  * if the allocation failed and we didn't allocate
623                  * any blocks before
624                  */
625                 goto failed_out;
626         }
627         if (!*err) {
628                 if (target == blks) {
629                 /*
630                  * save the new block number
631                  * for the first direct block
632                  */
633                         new_blocks[index] = current_block;
634                 }
635                 blk_allocated += ar.len;
636         }
637 allocated:
638         /* total number of blocks allocated for direct blocks */
639         ret = blk_allocated;
640         *err = 0;
641         return ret;
642 failed_out:
643         for (i = 0; i < index; i++)
644                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
645         return ret;
646 }
647
648 /**
649  *      ext4_alloc_branch - allocate and set up a chain of blocks.
650  *      @inode: owner
651  *      @indirect_blks: number of allocated indirect blocks
652  *      @blks: number of allocated direct blocks
653  *      @offsets: offsets (in the blocks) to store the pointers to next.
654  *      @branch: place to store the chain in.
655  *
656  *      This function allocates blocks, zeroes out all but the last one,
657  *      links them into chain and (if we are synchronous) writes them to disk.
658  *      In other words, it prepares a branch that can be spliced onto the
659  *      inode. It stores the information about that chain in the branch[], in
660  *      the same format as ext4_get_branch() would do. We are calling it after
661  *      we had read the existing part of chain and partial points to the last
662  *      triple of that (one with zero ->key). Upon the exit we have the same
663  *      picture as after the successful ext4_get_block(), except that in one
664  *      place chain is disconnected - *branch->p is still zero (we did not
665  *      set the last link), but branch->key contains the number that should
666  *      be placed into *branch->p to fill that gap.
667  *
668  *      If allocation fails we free all blocks we've allocated (and forget
669  *      their buffer_heads) and return the error value the from failed
670  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
671  *      as described above and return 0.
672  */
673 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
674                                 ext4_lblk_t iblock, int indirect_blks,
675                                 int *blks, ext4_fsblk_t goal,
676                                 ext4_lblk_t *offsets, Indirect *branch)
677 {
678         int blocksize = inode->i_sb->s_blocksize;
679         int i, n = 0;
680         int err = 0;
681         struct buffer_head *bh;
682         int num;
683         ext4_fsblk_t new_blocks[4];
684         ext4_fsblk_t current_block;
685
686         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
687                                 *blks, new_blocks, &err);
688         if (err)
689                 return err;
690
691         branch[0].key = cpu_to_le32(new_blocks[0]);
692         /*
693          * metadata blocks and data blocks are allocated.
694          */
695         for (n = 1; n <= indirect_blks;  n++) {
696                 /*
697                  * Get buffer_head for parent block, zero it out
698                  * and set the pointer to new one, then send
699                  * parent to disk.
700                  */
701                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
702                 branch[n].bh = bh;
703                 lock_buffer(bh);
704                 BUFFER_TRACE(bh, "call get_create_access");
705                 err = ext4_journal_get_create_access(handle, bh);
706                 if (err) {
707                         unlock_buffer(bh);
708                         brelse(bh);
709                         goto failed;
710                 }
711
712                 memset(bh->b_data, 0, blocksize);
713                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
714                 branch[n].key = cpu_to_le32(new_blocks[n]);
715                 *branch[n].p = branch[n].key;
716                 if (n == indirect_blks) {
717                         current_block = new_blocks[n];
718                         /*
719                          * End of chain, update the last new metablock of
720                          * the chain to point to the new allocated
721                          * data blocks numbers
722                          */
723                         for (i=1; i < num; i++)
724                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
725                 }
726                 BUFFER_TRACE(bh, "marking uptodate");
727                 set_buffer_uptodate(bh);
728                 unlock_buffer(bh);
729
730                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
731                 err = ext4_handle_dirty_metadata(handle, inode, bh);
732                 if (err)
733                         goto failed;
734         }
735         *blks = num;
736         return err;
737 failed:
738         /* Allocation failed, free what we already allocated */
739         for (i = 1; i <= n ; i++) {
740                 BUFFER_TRACE(branch[i].bh, "call jbd2_journal_forget");
741                 ext4_journal_forget(handle, branch[i].bh);
742         }
743         for (i = 0; i < indirect_blks; i++)
744                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
745
746         ext4_free_blocks(handle, inode, new_blocks[i], num, 0);
747
748         return err;
749 }
750
751 /**
752  * ext4_splice_branch - splice the allocated branch onto inode.
753  * @inode: owner
754  * @block: (logical) number of block we are adding
755  * @chain: chain of indirect blocks (with a missing link - see
756  *      ext4_alloc_branch)
757  * @where: location of missing link
758  * @num:   number of indirect blocks we are adding
759  * @blks:  number of direct blocks we are adding
760  *
761  * This function fills the missing link and does all housekeeping needed in
762  * inode (->i_blocks, etc.). In case of success we end up with the full
763  * chain to new block and return 0.
764  */
765 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
766                         ext4_lblk_t block, Indirect *where, int num, int blks)
767 {
768         int i;
769         int err = 0;
770         ext4_fsblk_t current_block;
771
772         /*
773          * If we're splicing into a [td]indirect block (as opposed to the
774          * inode) then we need to get write access to the [td]indirect block
775          * before the splice.
776          */
777         if (where->bh) {
778                 BUFFER_TRACE(where->bh, "get_write_access");
779                 err = ext4_journal_get_write_access(handle, where->bh);
780                 if (err)
781                         goto err_out;
782         }
783         /* That's it */
784
785         *where->p = where->key;
786
787         /*
788          * Update the host buffer_head or inode to point to more just allocated
789          * direct blocks blocks
790          */
791         if (num == 0 && blks > 1) {
792                 current_block = le32_to_cpu(where->key) + 1;
793                 for (i = 1; i < blks; i++)
794                         *(where->p + i) = cpu_to_le32(current_block++);
795         }
796
797         /* We are done with atomic stuff, now do the rest of housekeeping */
798
799         inode->i_ctime = ext4_current_time(inode);
800         ext4_mark_inode_dirty(handle, inode);
801
802         /* had we spliced it onto indirect block? */
803         if (where->bh) {
804                 /*
805                  * If we spliced it onto an indirect block, we haven't
806                  * altered the inode.  Note however that if it is being spliced
807                  * onto an indirect block at the very end of the file (the
808                  * file is growing) then we *will* alter the inode to reflect
809                  * the new i_size.  But that is not done here - it is done in
810                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
811                  */
812                 jbd_debug(5, "splicing indirect only\n");
813                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
814                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
815                 if (err)
816                         goto err_out;
817         } else {
818                 /*
819                  * OK, we spliced it into the inode itself on a direct block.
820                  * Inode was dirtied above.
821                  */
822                 jbd_debug(5, "splicing direct\n");
823         }
824         return err;
825
826 err_out:
827         for (i = 1; i <= num; i++) {
828                 BUFFER_TRACE(where[i].bh, "call jbd2_journal_forget");
829                 ext4_journal_forget(handle, where[i].bh);
830                 ext4_free_blocks(handle, inode,
831                                         le32_to_cpu(where[i-1].key), 1, 0);
832         }
833         ext4_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks, 0);
834
835         return err;
836 }
837
838 /*
839  * Allocation strategy is simple: if we have to allocate something, we will
840  * have to go the whole way to leaf. So let's do it before attaching anything
841  * to tree, set linkage between the newborn blocks, write them if sync is
842  * required, recheck the path, free and repeat if check fails, otherwise
843  * set the last missing link (that will protect us from any truncate-generated
844  * removals - all blocks on the path are immune now) and possibly force the
845  * write on the parent block.
846  * That has a nice additional property: no special recovery from the failed
847  * allocations is needed - we simply release blocks and do not touch anything
848  * reachable from inode.
849  *
850  * `handle' can be NULL if create == 0.
851  *
852  * return > 0, # of blocks mapped or allocated.
853  * return = 0, if plain lookup failed.
854  * return < 0, error case.
855  *
856  *
857  * Need to be called with
858  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
859  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
860  */
861 static int ext4_get_blocks_handle(handle_t *handle, struct inode *inode,
862                                   ext4_lblk_t iblock, unsigned int maxblocks,
863                                   struct buffer_head *bh_result,
864                                   int create, int extend_disksize)
865 {
866         int err = -EIO;
867         ext4_lblk_t offsets[4];
868         Indirect chain[4];
869         Indirect *partial;
870         ext4_fsblk_t goal;
871         int indirect_blks;
872         int blocks_to_boundary = 0;
873         int depth;
874         struct ext4_inode_info *ei = EXT4_I(inode);
875         int count = 0;
876         ext4_fsblk_t first_block = 0;
877         loff_t disksize;
878
879
880         J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
881         J_ASSERT(handle != NULL || create == 0);
882         depth = ext4_block_to_path(inode, iblock, offsets,
883                                         &blocks_to_boundary);
884
885         if (depth == 0)
886                 goto out;
887
888         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
889
890         /* Simplest case - block found, no allocation needed */
891         if (!partial) {
892                 first_block = le32_to_cpu(chain[depth - 1].key);
893                 clear_buffer_new(bh_result);
894                 count++;
895                 /*map more blocks*/
896                 while (count < maxblocks && count <= blocks_to_boundary) {
897                         ext4_fsblk_t blk;
898
899                         blk = le32_to_cpu(*(chain[depth-1].p + count));
900
901                         if (blk == first_block + count)
902                                 count++;
903                         else
904                                 break;
905                 }
906                 goto got_it;
907         }
908
909         /* Next simple case - plain lookup or failed read of indirect block */
910         if (!create || err == -EIO)
911                 goto cleanup;
912
913         /*
914          * Okay, we need to do block allocation.
915         */
916         goal = ext4_find_goal(inode, iblock, partial);
917
918         /* the number of blocks need to allocate for [d,t]indirect blocks */
919         indirect_blks = (chain + depth) - partial - 1;
920
921         /*
922          * Next look up the indirect map to count the totoal number of
923          * direct blocks to allocate for this branch.
924          */
925         count = ext4_blks_to_allocate(partial, indirect_blks,
926                                         maxblocks, blocks_to_boundary);
927         /*
928          * Block out ext4_truncate while we alter the tree
929          */
930         err = ext4_alloc_branch(handle, inode, iblock, indirect_blks,
931                                         &count, goal,
932                                         offsets + (partial - chain), partial);
933
934         /*
935          * The ext4_splice_branch call will free and forget any buffers
936          * on the new chain if there is a failure, but that risks using
937          * up transaction credits, especially for bitmaps where the
938          * credits cannot be returned.  Can we handle this somehow?  We
939          * may need to return -EAGAIN upwards in the worst case.  --sct
940          */
941         if (!err)
942                 err = ext4_splice_branch(handle, inode, iblock,
943                                         partial, indirect_blks, count);
944         /*
945          * i_disksize growing is protected by i_data_sem.  Don't forget to
946          * protect it if you're about to implement concurrent
947          * ext4_get_block() -bzzz
948         */
949         if (!err && extend_disksize) {
950                 disksize = ((loff_t) iblock + count) << inode->i_blkbits;
951                 if (disksize > i_size_read(inode))
952                         disksize = i_size_read(inode);
953                 if (disksize > ei->i_disksize)
954                         ei->i_disksize = disksize;
955         }
956         if (err)
957                 goto cleanup;
958
959         set_buffer_new(bh_result);
960 got_it:
961         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
962         if (count > blocks_to_boundary)
963                 set_buffer_boundary(bh_result);
964         err = count;
965         /* Clean up and exit */
966         partial = chain + depth - 1;    /* the whole chain */
967 cleanup:
968         while (partial > chain) {
969                 BUFFER_TRACE(partial->bh, "call brelse");
970                 brelse(partial->bh);
971                 partial--;
972         }
973         BUFFER_TRACE(bh_result, "returned");
974 out:
975         return err;
976 }
977
978 /*
979  * Calculate the number of metadata blocks need to reserve
980  * to allocate @blocks for non extent file based file
981  */
982 static int ext4_indirect_calc_metadata_amount(struct inode *inode, int blocks)
983 {
984         int icap = EXT4_ADDR_PER_BLOCK(inode->i_sb);
985         int ind_blks, dind_blks, tind_blks;
986
987         /* number of new indirect blocks needed */
988         ind_blks = (blocks + icap - 1) / icap;
989
990         dind_blks = (ind_blks + icap - 1) / icap;
991
992         tind_blks = 1;
993
994         return ind_blks + dind_blks + tind_blks;
995 }
996
997 /*
998  * Calculate the number of metadata blocks need to reserve
999  * to allocate given number of blocks
1000  */
1001 static int ext4_calc_metadata_amount(struct inode *inode, int blocks)
1002 {
1003         if (!blocks)
1004                 return 0;
1005
1006         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
1007                 return ext4_ext_calc_metadata_amount(inode, blocks);
1008
1009         return ext4_indirect_calc_metadata_amount(inode, blocks);
1010 }
1011
1012 static void ext4_da_update_reserve_space(struct inode *inode, int used)
1013 {
1014         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1015         int total, mdb, mdb_free;
1016
1017         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1018         /* recalculate the number of metablocks still need to be reserved */
1019         total = EXT4_I(inode)->i_reserved_data_blocks - used;
1020         mdb = ext4_calc_metadata_amount(inode, total);
1021
1022         /* figure out how many metablocks to release */
1023         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1024         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1025
1026         if (mdb_free) {
1027                 /* Account for allocated meta_blocks */
1028                 mdb_free -= EXT4_I(inode)->i_allocated_meta_blocks;
1029
1030                 /* update fs dirty blocks counter */
1031                 percpu_counter_sub(&sbi->s_dirtyblocks_counter, mdb_free);
1032                 EXT4_I(inode)->i_allocated_meta_blocks = 0;
1033                 EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1034         }
1035
1036         /* update per-inode reservations */
1037         BUG_ON(used  > EXT4_I(inode)->i_reserved_data_blocks);
1038         EXT4_I(inode)->i_reserved_data_blocks -= used;
1039
1040         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1041 }
1042
1043 /*
1044  * The ext4_get_blocks_wrap() function try to look up the requested blocks,
1045  * and returns if the blocks are already mapped.
1046  *
1047  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1048  * and store the allocated blocks in the result buffer head and mark it
1049  * mapped.
1050  *
1051  * If file type is extents based, it will call ext4_ext_get_blocks(),
1052  * Otherwise, call with ext4_get_blocks_handle() to handle indirect mapping
1053  * based files
1054  *
1055  * On success, it returns the number of blocks being mapped or allocate.
1056  * if create==0 and the blocks are pre-allocated and uninitialized block,
1057  * the result buffer head is unmapped. If the create ==1, it will make sure
1058  * the buffer head is mapped.
1059  *
1060  * It returns 0 if plain look up failed (blocks have not been allocated), in
1061  * that casem, buffer head is unmapped
1062  *
1063  * It returns the error in case of allocation failure.
1064  */
1065 int ext4_get_blocks_wrap(handle_t *handle, struct inode *inode, sector_t block,
1066                         unsigned int max_blocks, struct buffer_head *bh,
1067                         int create, int extend_disksize, int flag)
1068 {
1069         int retval;
1070
1071         clear_buffer_mapped(bh);
1072
1073         /*
1074          * Try to see if we can get  the block without requesting
1075          * for new file system block.
1076          */
1077         down_read((&EXT4_I(inode)->i_data_sem));
1078         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1079                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1080                                 bh, 0, 0);
1081         } else {
1082                 retval = ext4_get_blocks_handle(handle,
1083                                 inode, block, max_blocks, bh, 0, 0);
1084         }
1085         up_read((&EXT4_I(inode)->i_data_sem));
1086
1087         /* If it is only a block(s) look up */
1088         if (!create)
1089                 return retval;
1090
1091         /*
1092          * Returns if the blocks have already allocated
1093          *
1094          * Note that if blocks have been preallocated
1095          * ext4_ext_get_block() returns th create = 0
1096          * with buffer head unmapped.
1097          */
1098         if (retval > 0 && buffer_mapped(bh))
1099                 return retval;
1100
1101         /*
1102          * New blocks allocate and/or writing to uninitialized extent
1103          * will possibly result in updating i_data, so we take
1104          * the write lock of i_data_sem, and call get_blocks()
1105          * with create == 1 flag.
1106          */
1107         down_write((&EXT4_I(inode)->i_data_sem));
1108
1109         /*
1110          * if the caller is from delayed allocation writeout path
1111          * we have already reserved fs blocks for allocation
1112          * let the underlying get_block() function know to
1113          * avoid double accounting
1114          */
1115         if (flag)
1116                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1117         /*
1118          * We need to check for EXT4 here because migrate
1119          * could have changed the inode type in between
1120          */
1121         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1122                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1123                                 bh, create, extend_disksize);
1124         } else {
1125                 retval = ext4_get_blocks_handle(handle, inode, block,
1126                                 max_blocks, bh, create, extend_disksize);
1127
1128                 if (retval > 0 && buffer_new(bh)) {
1129                         /*
1130                          * We allocated new blocks which will result in
1131                          * i_data's format changing.  Force the migrate
1132                          * to fail by clearing migrate flags
1133                          */
1134                         EXT4_I(inode)->i_flags = EXT4_I(inode)->i_flags &
1135                                                         ~EXT4_EXT_MIGRATE;
1136                 }
1137         }
1138
1139         if (flag) {
1140                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1141                 /*
1142                  * Update reserved blocks/metadata blocks
1143                  * after successful block allocation
1144                  * which were deferred till now
1145                  */
1146                 if ((retval > 0) && buffer_delay(bh))
1147                         ext4_da_update_reserve_space(inode, retval);
1148         }
1149
1150         up_write((&EXT4_I(inode)->i_data_sem));
1151         return retval;
1152 }
1153
1154 /* Maximum number of blocks we map for direct IO at once. */
1155 #define DIO_MAX_BLOCKS 4096
1156
1157 int ext4_get_block(struct inode *inode, sector_t iblock,
1158                    struct buffer_head *bh_result, int create)
1159 {
1160         handle_t *handle = ext4_journal_current_handle();
1161         int ret = 0, started = 0;
1162         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1163         int dio_credits;
1164
1165         if (create && !handle) {
1166                 /* Direct IO write... */
1167                 if (max_blocks > DIO_MAX_BLOCKS)
1168                         max_blocks = DIO_MAX_BLOCKS;
1169                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
1170                 handle = ext4_journal_start(inode, dio_credits);
1171                 if (IS_ERR(handle)) {
1172                         ret = PTR_ERR(handle);
1173                         goto out;
1174                 }
1175                 started = 1;
1176         }
1177
1178         ret = ext4_get_blocks_wrap(handle, inode, iblock,
1179                                         max_blocks, bh_result, create, 0, 0);
1180         if (ret > 0) {
1181                 bh_result->b_size = (ret << inode->i_blkbits);
1182                 ret = 0;
1183         }
1184         if (started)
1185                 ext4_journal_stop(handle);
1186 out:
1187         return ret;
1188 }
1189
1190 /*
1191  * `handle' can be NULL if create is zero
1192  */
1193 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1194                                 ext4_lblk_t block, int create, int *errp)
1195 {
1196         struct buffer_head dummy;
1197         int fatal = 0, err;
1198
1199         J_ASSERT(handle != NULL || create == 0);
1200
1201         dummy.b_state = 0;
1202         dummy.b_blocknr = -1000;
1203         buffer_trace_init(&dummy.b_history);
1204         err = ext4_get_blocks_wrap(handle, inode, block, 1,
1205                                         &dummy, create, 1, 0);
1206         /*
1207          * ext4_get_blocks_handle() returns number of blocks
1208          * mapped. 0 in case of a HOLE.
1209          */
1210         if (err > 0) {
1211                 if (err > 1)
1212                         WARN_ON(1);
1213                 err = 0;
1214         }
1215         *errp = err;
1216         if (!err && buffer_mapped(&dummy)) {
1217                 struct buffer_head *bh;
1218                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1219                 if (!bh) {
1220                         *errp = -EIO;
1221                         goto err;
1222                 }
1223                 if (buffer_new(&dummy)) {
1224                         J_ASSERT(create != 0);
1225                         J_ASSERT(handle != NULL);
1226
1227                         /*
1228                          * Now that we do not always journal data, we should
1229                          * keep in mind whether this should always journal the
1230                          * new buffer as metadata.  For now, regular file
1231                          * writes use ext4_get_block instead, so it's not a
1232                          * problem.
1233                          */
1234                         lock_buffer(bh);
1235                         BUFFER_TRACE(bh, "call get_create_access");
1236                         fatal = ext4_journal_get_create_access(handle, bh);
1237                         if (!fatal && !buffer_uptodate(bh)) {
1238                                 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1239                                 set_buffer_uptodate(bh);
1240                         }
1241                         unlock_buffer(bh);
1242                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1243                         err = ext4_handle_dirty_metadata(handle, inode, bh);
1244                         if (!fatal)
1245                                 fatal = err;
1246                 } else {
1247                         BUFFER_TRACE(bh, "not a new buffer");
1248                 }
1249                 if (fatal) {
1250                         *errp = fatal;
1251                         brelse(bh);
1252                         bh = NULL;
1253                 }
1254                 return bh;
1255         }
1256 err:
1257         return NULL;
1258 }
1259
1260 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1261                                ext4_lblk_t block, int create, int *err)
1262 {
1263         struct buffer_head *bh;
1264
1265         bh = ext4_getblk(handle, inode, block, create, err);
1266         if (!bh)
1267                 return bh;
1268         if (buffer_uptodate(bh))
1269                 return bh;
1270         ll_rw_block(READ_META, 1, &bh);
1271         wait_on_buffer(bh);
1272         if (buffer_uptodate(bh))
1273                 return bh;
1274         put_bh(bh);
1275         *err = -EIO;
1276         return NULL;
1277 }
1278
1279 static int walk_page_buffers(handle_t *handle,
1280                              struct buffer_head *head,
1281                              unsigned from,
1282                              unsigned to,
1283                              int *partial,
1284                              int (*fn)(handle_t *handle,
1285                                        struct buffer_head *bh))
1286 {
1287         struct buffer_head *bh;
1288         unsigned block_start, block_end;
1289         unsigned blocksize = head->b_size;
1290         int err, ret = 0;
1291         struct buffer_head *next;
1292
1293         for (bh = head, block_start = 0;
1294              ret == 0 && (bh != head || !block_start);
1295              block_start = block_end, bh = next)
1296         {
1297                 next = bh->b_this_page;
1298                 block_end = block_start + blocksize;
1299                 if (block_end <= from || block_start >= to) {
1300                         if (partial && !buffer_uptodate(bh))
1301                                 *partial = 1;
1302                         continue;
1303                 }
1304                 err = (*fn)(handle, bh);
1305                 if (!ret)
1306                         ret = err;
1307         }
1308         return ret;
1309 }
1310
1311 /*
1312  * To preserve ordering, it is essential that the hole instantiation and
1313  * the data write be encapsulated in a single transaction.  We cannot
1314  * close off a transaction and start a new one between the ext4_get_block()
1315  * and the commit_write().  So doing the jbd2_journal_start at the start of
1316  * prepare_write() is the right place.
1317  *
1318  * Also, this function can nest inside ext4_writepage() ->
1319  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1320  * has generated enough buffer credits to do the whole page.  So we won't
1321  * block on the journal in that case, which is good, because the caller may
1322  * be PF_MEMALLOC.
1323  *
1324  * By accident, ext4 can be reentered when a transaction is open via
1325  * quota file writes.  If we were to commit the transaction while thus
1326  * reentered, there can be a deadlock - we would be holding a quota
1327  * lock, and the commit would never complete if another thread had a
1328  * transaction open and was blocking on the quota lock - a ranking
1329  * violation.
1330  *
1331  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1332  * will _not_ run commit under these circumstances because handle->h_ref
1333  * is elevated.  We'll still have enough credits for the tiny quotafile
1334  * write.
1335  */
1336 static int do_journal_get_write_access(handle_t *handle,
1337                                         struct buffer_head *bh)
1338 {
1339         if (!buffer_mapped(bh) || buffer_freed(bh))
1340                 return 0;
1341         return ext4_journal_get_write_access(handle, bh);
1342 }
1343
1344 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1345                                 loff_t pos, unsigned len, unsigned flags,
1346                                 struct page **pagep, void **fsdata)
1347 {
1348         struct inode *inode = mapping->host;
1349         int ret, needed_blocks = ext4_writepage_trans_blocks(inode);
1350         handle_t *handle;
1351         int retries = 0;
1352         struct page *page;
1353         pgoff_t index;
1354         unsigned from, to;
1355
1356         trace_mark(ext4_write_begin,
1357                    "dev %s ino %lu pos %llu len %u flags %u",
1358                    inode->i_sb->s_id, inode->i_ino,
1359                    (unsigned long long) pos, len, flags);
1360         index = pos >> PAGE_CACHE_SHIFT;
1361         from = pos & (PAGE_CACHE_SIZE - 1);
1362         to = from + len;
1363
1364 retry:
1365         handle = ext4_journal_start(inode, needed_blocks);
1366         if (IS_ERR(handle)) {
1367                 ret = PTR_ERR(handle);
1368                 goto out;
1369         }
1370
1371         /* We cannot recurse into the filesystem as the transaction is already
1372          * started */
1373         flags |= AOP_FLAG_NOFS;
1374
1375         page = grab_cache_page_write_begin(mapping, index, flags);
1376         if (!page) {
1377                 ext4_journal_stop(handle);
1378                 ret = -ENOMEM;
1379                 goto out;
1380         }
1381         *pagep = page;
1382
1383         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1384                                 ext4_get_block);
1385
1386         if (!ret && ext4_should_journal_data(inode)) {
1387                 ret = walk_page_buffers(handle, page_buffers(page),
1388                                 from, to, NULL, do_journal_get_write_access);
1389         }
1390
1391         if (ret) {
1392                 unlock_page(page);
1393                 ext4_journal_stop(handle);
1394                 page_cache_release(page);
1395                 /*
1396                  * block_write_begin may have instantiated a few blocks
1397                  * outside i_size.  Trim these off again. Don't need
1398                  * i_size_read because we hold i_mutex.
1399                  */
1400                 if (pos + len > inode->i_size)
1401                         vmtruncate(inode, inode->i_size);
1402         }
1403
1404         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1405                 goto retry;
1406 out:
1407         return ret;
1408 }
1409
1410 /* For write_end() in data=journal mode */
1411 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1412 {
1413         if (!buffer_mapped(bh) || buffer_freed(bh))
1414                 return 0;
1415         set_buffer_uptodate(bh);
1416         return ext4_handle_dirty_metadata(handle, NULL, bh);
1417 }
1418
1419 /*
1420  * We need to pick up the new inode size which generic_commit_write gave us
1421  * `file' can be NULL - eg, when called from page_symlink().
1422  *
1423  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1424  * buffers are managed internally.
1425  */
1426 static int ext4_ordered_write_end(struct file *file,
1427                                 struct address_space *mapping,
1428                                 loff_t pos, unsigned len, unsigned copied,
1429                                 struct page *page, void *fsdata)
1430 {
1431         handle_t *handle = ext4_journal_current_handle();
1432         struct inode *inode = mapping->host;
1433         int ret = 0, ret2;
1434
1435         trace_mark(ext4_ordered_write_end,
1436                    "dev %s ino %lu pos %llu len %u copied %u",
1437                    inode->i_sb->s_id, inode->i_ino,
1438                    (unsigned long long) pos, len, copied);
1439         ret = ext4_jbd2_file_inode(handle, inode);
1440
1441         if (ret == 0) {
1442                 loff_t new_i_size;
1443
1444                 new_i_size = pos + copied;
1445                 if (new_i_size > EXT4_I(inode)->i_disksize) {
1446                         ext4_update_i_disksize(inode, new_i_size);
1447                         /* We need to mark inode dirty even if
1448                          * new_i_size is less that inode->i_size
1449                          * bu greater than i_disksize.(hint delalloc)
1450                          */
1451                         ext4_mark_inode_dirty(handle, inode);
1452                 }
1453
1454                 ret2 = generic_write_end(file, mapping, pos, len, copied,
1455                                                         page, fsdata);
1456                 copied = ret2;
1457                 if (ret2 < 0)
1458                         ret = ret2;
1459         }
1460         ret2 = ext4_journal_stop(handle);
1461         if (!ret)
1462                 ret = ret2;
1463
1464         return ret ? ret : copied;
1465 }
1466
1467 static int ext4_writeback_write_end(struct file *file,
1468                                 struct address_space *mapping,
1469                                 loff_t pos, unsigned len, unsigned copied,
1470                                 struct page *page, void *fsdata)
1471 {
1472         handle_t *handle = ext4_journal_current_handle();
1473         struct inode *inode = mapping->host;
1474         int ret = 0, ret2;
1475         loff_t new_i_size;
1476
1477         trace_mark(ext4_writeback_write_end,
1478                    "dev %s ino %lu pos %llu len %u copied %u",
1479                    inode->i_sb->s_id, inode->i_ino,
1480                    (unsigned long long) pos, len, copied);
1481         new_i_size = pos + copied;
1482         if (new_i_size > EXT4_I(inode)->i_disksize) {
1483                 ext4_update_i_disksize(inode, new_i_size);
1484                 /* We need to mark inode dirty even if
1485                  * new_i_size is less that inode->i_size
1486                  * bu greater than i_disksize.(hint delalloc)
1487                  */
1488                 ext4_mark_inode_dirty(handle, inode);
1489         }
1490
1491         ret2 = generic_write_end(file, mapping, pos, len, copied,
1492                                                         page, fsdata);
1493         copied = ret2;
1494         if (ret2 < 0)
1495                 ret = ret2;
1496
1497         ret2 = ext4_journal_stop(handle);
1498         if (!ret)
1499                 ret = ret2;
1500
1501         return ret ? ret : copied;
1502 }
1503
1504 static int ext4_journalled_write_end(struct file *file,
1505                                 struct address_space *mapping,
1506                                 loff_t pos, unsigned len, unsigned copied,
1507                                 struct page *page, void *fsdata)
1508 {
1509         handle_t *handle = ext4_journal_current_handle();
1510         struct inode *inode = mapping->host;
1511         int ret = 0, ret2;
1512         int partial = 0;
1513         unsigned from, to;
1514         loff_t new_i_size;
1515
1516         trace_mark(ext4_journalled_write_end,
1517                    "dev %s ino %lu pos %llu len %u copied %u",
1518                    inode->i_sb->s_id, inode->i_ino,
1519                    (unsigned long long) pos, len, copied);
1520         from = pos & (PAGE_CACHE_SIZE - 1);
1521         to = from + len;
1522
1523         if (copied < len) {
1524                 if (!PageUptodate(page))
1525                         copied = 0;
1526                 page_zero_new_buffers(page, from+copied, to);
1527         }
1528
1529         ret = walk_page_buffers(handle, page_buffers(page), from,
1530                                 to, &partial, write_end_fn);
1531         if (!partial)
1532                 SetPageUptodate(page);
1533         new_i_size = pos + copied;
1534         if (new_i_size > inode->i_size)
1535                 i_size_write(inode, pos+copied);
1536         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
1537         if (new_i_size > EXT4_I(inode)->i_disksize) {
1538                 ext4_update_i_disksize(inode, new_i_size);
1539                 ret2 = ext4_mark_inode_dirty(handle, inode);
1540                 if (!ret)
1541                         ret = ret2;
1542         }
1543
1544         unlock_page(page);
1545         ret2 = ext4_journal_stop(handle);
1546         if (!ret)
1547                 ret = ret2;
1548         page_cache_release(page);
1549
1550         return ret ? ret : copied;
1551 }
1552
1553 static int ext4_da_reserve_space(struct inode *inode, int nrblocks)
1554 {
1555         int retries = 0;
1556        struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1557        unsigned long md_needed, mdblocks, total = 0;
1558
1559         /*
1560          * recalculate the amount of metadata blocks to reserve
1561          * in order to allocate nrblocks
1562          * worse case is one extent per block
1563          */
1564 repeat:
1565         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1566         total = EXT4_I(inode)->i_reserved_data_blocks + nrblocks;
1567         mdblocks = ext4_calc_metadata_amount(inode, total);
1568         BUG_ON(mdblocks < EXT4_I(inode)->i_reserved_meta_blocks);
1569
1570         md_needed = mdblocks - EXT4_I(inode)->i_reserved_meta_blocks;
1571         total = md_needed + nrblocks;
1572
1573         if (ext4_claim_free_blocks(sbi, total)) {
1574                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1575                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1576                         yield();
1577                         goto repeat;
1578                 }
1579                 return -ENOSPC;
1580         }
1581         EXT4_I(inode)->i_reserved_data_blocks += nrblocks;
1582         EXT4_I(inode)->i_reserved_meta_blocks = mdblocks;
1583
1584         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1585         return 0;       /* success */
1586 }
1587
1588 static void ext4_da_release_space(struct inode *inode, int to_free)
1589 {
1590         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1591         int total, mdb, mdb_free, release;
1592
1593         if (!to_free)
1594                 return;         /* Nothing to release, exit */
1595
1596         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1597
1598         if (!EXT4_I(inode)->i_reserved_data_blocks) {
1599                 /*
1600                  * if there is no reserved blocks, but we try to free some
1601                  * then the counter is messed up somewhere.
1602                  * but since this function is called from invalidate
1603                  * page, it's harmless to return without any action
1604                  */
1605                 printk(KERN_INFO "ext4 delalloc try to release %d reserved "
1606                             "blocks for inode %lu, but there is no reserved "
1607                             "data blocks\n", to_free, inode->i_ino);
1608                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1609                 return;
1610         }
1611
1612         /* recalculate the number of metablocks still need to be reserved */
1613         total = EXT4_I(inode)->i_reserved_data_blocks - to_free;
1614         mdb = ext4_calc_metadata_amount(inode, total);
1615
1616         /* figure out how many metablocks to release */
1617         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1618         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1619
1620         release = to_free + mdb_free;
1621
1622         /* update fs dirty blocks counter for truncate case */
1623         percpu_counter_sub(&sbi->s_dirtyblocks_counter, release);
1624
1625         /* update per-inode reservations */
1626         BUG_ON(to_free > EXT4_I(inode)->i_reserved_data_blocks);
1627         EXT4_I(inode)->i_reserved_data_blocks -= to_free;
1628
1629         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1630         EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1631         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1632 }
1633
1634 static void ext4_da_page_release_reservation(struct page *page,
1635                                                 unsigned long offset)
1636 {
1637         int to_release = 0;
1638         struct buffer_head *head, *bh;
1639         unsigned int curr_off = 0;
1640
1641         head = page_buffers(page);
1642         bh = head;
1643         do {
1644                 unsigned int next_off = curr_off + bh->b_size;
1645
1646                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1647                         to_release++;
1648                         clear_buffer_delay(bh);
1649                 }
1650                 curr_off = next_off;
1651         } while ((bh = bh->b_this_page) != head);
1652         ext4_da_release_space(page->mapping->host, to_release);
1653 }
1654
1655 /*
1656  * Delayed allocation stuff
1657  */
1658
1659 struct mpage_da_data {
1660         struct inode *inode;
1661         struct buffer_head lbh;                 /* extent of blocks */
1662         unsigned long first_page, next_page;    /* extent of pages */
1663         get_block_t *get_block;
1664         struct writeback_control *wbc;
1665         int io_done;
1666         int pages_written;
1667         int retval;
1668 };
1669
1670 /*
1671  * mpage_da_submit_io - walks through extent of pages and try to write
1672  * them with writepage() call back
1673  *
1674  * @mpd->inode: inode
1675  * @mpd->first_page: first page of the extent
1676  * @mpd->next_page: page after the last page of the extent
1677  * @mpd->get_block: the filesystem's block mapper function
1678  *
1679  * By the time mpage_da_submit_io() is called we expect all blocks
1680  * to be allocated. this may be wrong if allocation failed.
1681  *
1682  * As pages are already locked by write_cache_pages(), we can't use it
1683  */
1684 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1685 {
1686         long pages_skipped;
1687         struct pagevec pvec;
1688         unsigned long index, end;
1689         int ret = 0, err, nr_pages, i;
1690         struct inode *inode = mpd->inode;
1691         struct address_space *mapping = inode->i_mapping;
1692
1693         BUG_ON(mpd->next_page <= mpd->first_page);
1694         /*
1695          * We need to start from the first_page to the next_page - 1
1696          * to make sure we also write the mapped dirty buffer_heads.
1697          * If we look at mpd->lbh.b_blocknr we would only be looking
1698          * at the currently mapped buffer_heads.
1699          */
1700         index = mpd->first_page;
1701         end = mpd->next_page - 1;
1702
1703         pagevec_init(&pvec, 0);
1704         while (index <= end) {
1705                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1706                 if (nr_pages == 0)
1707                         break;
1708                 for (i = 0; i < nr_pages; i++) {
1709                         struct page *page = pvec.pages[i];
1710
1711                         index = page->index;
1712                         if (index > end)
1713                                 break;
1714                         index++;
1715
1716                         BUG_ON(!PageLocked(page));
1717                         BUG_ON(PageWriteback(page));
1718
1719                         pages_skipped = mpd->wbc->pages_skipped;
1720                         err = mapping->a_ops->writepage(page, mpd->wbc);
1721                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
1722                                 /*
1723                                  * have successfully written the page
1724                                  * without skipping the same
1725                                  */
1726                                 mpd->pages_written++;
1727                         /*
1728                          * In error case, we have to continue because
1729                          * remaining pages are still locked
1730                          * XXX: unlock and re-dirty them?
1731                          */
1732                         if (ret == 0)
1733                                 ret = err;
1734                 }
1735                 pagevec_release(&pvec);
1736         }
1737         return ret;
1738 }
1739
1740 /*
1741  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
1742  *
1743  * @mpd->inode - inode to walk through
1744  * @exbh->b_blocknr - first block on a disk
1745  * @exbh->b_size - amount of space in bytes
1746  * @logical - first logical block to start assignment with
1747  *
1748  * the function goes through all passed space and put actual disk
1749  * block numbers into buffer heads, dropping BH_Delay
1750  */
1751 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
1752                                  struct buffer_head *exbh)
1753 {
1754         struct inode *inode = mpd->inode;
1755         struct address_space *mapping = inode->i_mapping;
1756         int blocks = exbh->b_size >> inode->i_blkbits;
1757         sector_t pblock = exbh->b_blocknr, cur_logical;
1758         struct buffer_head *head, *bh;
1759         pgoff_t index, end;
1760         struct pagevec pvec;
1761         int nr_pages, i;
1762
1763         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1764         end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1765         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1766
1767         pagevec_init(&pvec, 0);
1768
1769         while (index <= end) {
1770                 /* XXX: optimize tail */
1771                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1772                 if (nr_pages == 0)
1773                         break;
1774                 for (i = 0; i < nr_pages; i++) {
1775                         struct page *page = pvec.pages[i];
1776
1777                         index = page->index;
1778                         if (index > end)
1779                                 break;
1780                         index++;
1781
1782                         BUG_ON(!PageLocked(page));
1783                         BUG_ON(PageWriteback(page));
1784                         BUG_ON(!page_has_buffers(page));
1785
1786                         bh = page_buffers(page);
1787                         head = bh;
1788
1789                         /* skip blocks out of the range */
1790                         do {
1791                                 if (cur_logical >= logical)
1792                                         break;
1793                                 cur_logical++;
1794                         } while ((bh = bh->b_this_page) != head);
1795
1796                         do {
1797                                 if (cur_logical >= logical + blocks)
1798                                         break;
1799                                 if (buffer_delay(bh)) {
1800                                         bh->b_blocknr = pblock;
1801                                         clear_buffer_delay(bh);
1802                                         bh->b_bdev = inode->i_sb->s_bdev;
1803                                 } else if (buffer_unwritten(bh)) {
1804                                         bh->b_blocknr = pblock;
1805                                         clear_buffer_unwritten(bh);
1806                                         set_buffer_mapped(bh);
1807                                         set_buffer_new(bh);
1808                                         bh->b_bdev = inode->i_sb->s_bdev;
1809                                 } else if (buffer_mapped(bh))
1810                                         BUG_ON(bh->b_blocknr != pblock);
1811
1812                                 cur_logical++;
1813                                 pblock++;
1814                         } while ((bh = bh->b_this_page) != head);
1815                 }
1816                 pagevec_release(&pvec);
1817         }
1818 }
1819
1820
1821 /*
1822  * __unmap_underlying_blocks - just a helper function to unmap
1823  * set of blocks described by @bh
1824  */
1825 static inline void __unmap_underlying_blocks(struct inode *inode,
1826                                              struct buffer_head *bh)
1827 {
1828         struct block_device *bdev = inode->i_sb->s_bdev;
1829         int blocks, i;
1830
1831         blocks = bh->b_size >> inode->i_blkbits;
1832         for (i = 0; i < blocks; i++)
1833                 unmap_underlying_metadata(bdev, bh->b_blocknr + i);
1834 }
1835
1836 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
1837                                         sector_t logical, long blk_cnt)
1838 {
1839         int nr_pages, i;
1840         pgoff_t index, end;
1841         struct pagevec pvec;
1842         struct inode *inode = mpd->inode;
1843         struct address_space *mapping = inode->i_mapping;
1844
1845         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1846         end   = (logical + blk_cnt - 1) >>
1847                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
1848         while (index <= end) {
1849                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1850                 if (nr_pages == 0)
1851                         break;
1852                 for (i = 0; i < nr_pages; i++) {
1853                         struct page *page = pvec.pages[i];
1854                         index = page->index;
1855                         if (index > end)
1856                                 break;
1857                         index++;
1858
1859                         BUG_ON(!PageLocked(page));
1860                         BUG_ON(PageWriteback(page));
1861                         block_invalidatepage(page, 0);
1862                         ClearPageUptodate(page);
1863                         unlock_page(page);
1864                 }
1865         }
1866         return;
1867 }
1868
1869 static void ext4_print_free_blocks(struct inode *inode)
1870 {
1871         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1872         printk(KERN_EMERG "Total free blocks count %lld\n",
1873                         ext4_count_free_blocks(inode->i_sb));
1874         printk(KERN_EMERG "Free/Dirty block details\n");
1875         printk(KERN_EMERG "free_blocks=%lld\n",
1876                         (long long)percpu_counter_sum(&sbi->s_freeblocks_counter));
1877         printk(KERN_EMERG "dirty_blocks=%lld\n",
1878                         (long long)percpu_counter_sum(&sbi->s_dirtyblocks_counter));
1879         printk(KERN_EMERG "Block reservation details\n");
1880         printk(KERN_EMERG "i_reserved_data_blocks=%u\n",
1881                         EXT4_I(inode)->i_reserved_data_blocks);
1882         printk(KERN_EMERG "i_reserved_meta_blocks=%u\n",
1883                         EXT4_I(inode)->i_reserved_meta_blocks);
1884         return;
1885 }
1886
1887 /*
1888  * mpage_da_map_blocks - go through given space
1889  *
1890  * @mpd->lbh - bh describing space
1891  * @mpd->get_block - the filesystem's block mapper function
1892  *
1893  * The function skips space we know is already mapped to disk blocks.
1894  *
1895  */
1896 static int  mpage_da_map_blocks(struct mpage_da_data *mpd)
1897 {
1898         int err = 0;
1899         struct buffer_head new;
1900         struct buffer_head *lbh = &mpd->lbh;
1901         sector_t next;
1902
1903         /*
1904          * We consider only non-mapped and non-allocated blocks
1905          */
1906         if (buffer_mapped(lbh) && !buffer_delay(lbh))
1907                 return 0;
1908         new.b_state = lbh->b_state;
1909         new.b_blocknr = 0;
1910         new.b_size = lbh->b_size;
1911         next = lbh->b_blocknr;
1912         /*
1913          * If we didn't accumulate anything
1914          * to write simply return
1915          */
1916         if (!new.b_size)
1917                 return 0;
1918         err = mpd->get_block(mpd->inode, next, &new, 1);
1919         if (err) {
1920
1921                 /* If get block returns with error
1922                  * we simply return. Later writepage
1923                  * will redirty the page and writepages
1924                  * will find the dirty page again
1925                  */
1926                 if (err == -EAGAIN)
1927                         return 0;
1928
1929                 if (err == -ENOSPC &&
1930                                 ext4_count_free_blocks(mpd->inode->i_sb)) {
1931                         mpd->retval = err;
1932                         return 0;
1933                 }
1934
1935                 /*
1936                  * get block failure will cause us
1937                  * to loop in writepages. Because
1938                  * a_ops->writepage won't be able to
1939                  * make progress. The page will be redirtied
1940                  * by writepage and writepages will again
1941                  * try to write the same.
1942                  */
1943                 printk(KERN_EMERG "%s block allocation failed for inode %lu "
1944                                   "at logical offset %llu with max blocks "
1945                                   "%zd with error %d\n",
1946                                   __func__, mpd->inode->i_ino,
1947                                   (unsigned long long)next,
1948                                   lbh->b_size >> mpd->inode->i_blkbits, err);
1949                 printk(KERN_EMERG "This should not happen.!! "
1950                                         "Data will be lost\n");
1951                 if (err == -ENOSPC) {
1952                         ext4_print_free_blocks(mpd->inode);
1953                 }
1954                 /* invlaidate all the pages */
1955                 ext4_da_block_invalidatepages(mpd, next,
1956                                 lbh->b_size >> mpd->inode->i_blkbits);
1957                 return err;
1958         }
1959         BUG_ON(new.b_size == 0);
1960
1961         if (buffer_new(&new))
1962                 __unmap_underlying_blocks(mpd->inode, &new);
1963
1964         /*
1965          * If blocks are delayed marked, we need to
1966          * put actual blocknr and drop delayed bit
1967          */
1968         if (buffer_delay(lbh) || buffer_unwritten(lbh))
1969                 mpage_put_bnr_to_bhs(mpd, next, &new);
1970
1971         return 0;
1972 }
1973
1974 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
1975                 (1 << BH_Delay) | (1 << BH_Unwritten))
1976
1977 /*
1978  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
1979  *
1980  * @mpd->lbh - extent of blocks
1981  * @logical - logical number of the block in the file
1982  * @bh - bh of the block (used to access block's state)
1983  *
1984  * the function is used to collect contig. blocks in same state
1985  */
1986 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
1987                                    sector_t logical, struct buffer_head *bh)
1988 {
1989         sector_t next;
1990         size_t b_size = bh->b_size;
1991         struct buffer_head *lbh = &mpd->lbh;
1992         int nrblocks = lbh->b_size >> mpd->inode->i_blkbits;
1993
1994         /* check if thereserved journal credits might overflow */
1995         if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
1996                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
1997                         /*
1998                          * With non-extent format we are limited by the journal
1999                          * credit available.  Total credit needed to insert
2000                          * nrblocks contiguous blocks is dependent on the
2001                          * nrblocks.  So limit nrblocks.
2002                          */
2003                         goto flush_it;
2004                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2005                                 EXT4_MAX_TRANS_DATA) {
2006                         /*
2007                          * Adding the new buffer_head would make it cross the
2008                          * allowed limit for which we have journal credit
2009                          * reserved. So limit the new bh->b_size
2010                          */
2011                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2012                                                 mpd->inode->i_blkbits;
2013                         /* we will do mpage_da_submit_io in the next loop */
2014                 }
2015         }
2016         /*
2017          * First block in the extent
2018          */
2019         if (lbh->b_size == 0) {
2020                 lbh->b_blocknr = logical;
2021                 lbh->b_size = b_size;
2022                 lbh->b_state = bh->b_state & BH_FLAGS;
2023                 return;
2024         }
2025
2026         next = lbh->b_blocknr + nrblocks;
2027         /*
2028          * Can we merge the block to our big extent?
2029          */
2030         if (logical == next && (bh->b_state & BH_FLAGS) == lbh->b_state) {
2031                 lbh->b_size += b_size;
2032                 return;
2033         }
2034
2035 flush_it:
2036         /*
2037          * We couldn't merge the block to our extent, so we
2038          * need to flush current  extent and start new one
2039          */
2040         if (mpage_da_map_blocks(mpd) == 0)
2041                 mpage_da_submit_io(mpd);
2042         mpd->io_done = 1;
2043         return;
2044 }
2045
2046 /*
2047  * __mpage_da_writepage - finds extent of pages and blocks
2048  *
2049  * @page: page to consider
2050  * @wbc: not used, we just follow rules
2051  * @data: context
2052  *
2053  * The function finds extents of pages and scan them for all blocks.
2054  */
2055 static int __mpage_da_writepage(struct page *page,
2056                                 struct writeback_control *wbc, void *data)
2057 {
2058         struct mpage_da_data *mpd = data;
2059         struct inode *inode = mpd->inode;
2060         struct buffer_head *bh, *head, fake;
2061         sector_t logical;
2062
2063         if (mpd->io_done) {
2064                 /*
2065                  * Rest of the page in the page_vec
2066                  * redirty then and skip then. We will
2067                  * try to to write them again after
2068                  * starting a new transaction
2069                  */
2070                 redirty_page_for_writepage(wbc, page);
2071                 unlock_page(page);
2072                 return MPAGE_DA_EXTENT_TAIL;
2073         }
2074         /*
2075          * Can we merge this page to current extent?
2076          */
2077         if (mpd->next_page != page->index) {
2078                 /*
2079                  * Nope, we can't. So, we map non-allocated blocks
2080                  * and start IO on them using writepage()
2081                  */
2082                 if (mpd->next_page != mpd->first_page) {
2083                         if (mpage_da_map_blocks(mpd) == 0)
2084                                 mpage_da_submit_io(mpd);
2085                         /*
2086                          * skip rest of the page in the page_vec
2087                          */
2088                         mpd->io_done = 1;
2089                         redirty_page_for_writepage(wbc, page);
2090                         unlock_page(page);
2091                         return MPAGE_DA_EXTENT_TAIL;
2092                 }
2093
2094                 /*
2095                  * Start next extent of pages ...
2096                  */
2097                 mpd->first_page = page->index;
2098
2099                 /*
2100                  * ... and blocks
2101                  */
2102                 mpd->lbh.b_size = 0;
2103                 mpd->lbh.b_state = 0;
2104                 mpd->lbh.b_blocknr = 0;
2105         }
2106
2107         mpd->next_page = page->index + 1;
2108         logical = (sector_t) page->index <<
2109                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2110
2111         if (!page_has_buffers(page)) {
2112                 /*
2113                  * There is no attached buffer heads yet (mmap?)
2114                  * we treat the page asfull of dirty blocks
2115                  */
2116                 bh = &fake;
2117                 bh->b_size = PAGE_CACHE_SIZE;
2118                 bh->b_state = 0;
2119                 set_buffer_dirty(bh);
2120                 set_buffer_uptodate(bh);
2121                 mpage_add_bh_to_extent(mpd, logical, bh);
2122                 if (mpd->io_done)
2123                         return MPAGE_DA_EXTENT_TAIL;
2124         } else {
2125                 /*
2126                  * Page with regular buffer heads, just add all dirty ones
2127                  */
2128                 head = page_buffers(page);
2129                 bh = head;
2130                 do {
2131                         BUG_ON(buffer_locked(bh));
2132                         /*
2133                          * We need to try to allocate
2134                          * unmapped blocks in the same page.
2135                          * Otherwise we won't make progress
2136                          * with the page in ext4_da_writepage
2137                          */
2138                         if (buffer_dirty(bh) &&
2139                                 (!buffer_mapped(bh) || buffer_delay(bh))) {
2140                                 mpage_add_bh_to_extent(mpd, logical, bh);
2141                                 if (mpd->io_done)
2142                                         return MPAGE_DA_EXTENT_TAIL;
2143                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2144                                 /*
2145                                  * mapped dirty buffer. We need to update
2146                                  * the b_state because we look at
2147                                  * b_state in mpage_da_map_blocks. We don't
2148                                  * update b_size because if we find an
2149                                  * unmapped buffer_head later we need to
2150                                  * use the b_state flag of that buffer_head.
2151                                  */
2152                                 if (mpd->lbh.b_size == 0)
2153                                         mpd->lbh.b_state =
2154                                                 bh->b_state & BH_FLAGS;
2155                         }
2156                         logical++;
2157                 } while ((bh = bh->b_this_page) != head);
2158         }
2159
2160         return 0;
2161 }
2162
2163 /*
2164  * mpage_da_writepages - walk the list of dirty pages of the given
2165  * address space, allocates non-allocated blocks, maps newly-allocated
2166  * blocks to existing bhs and issue IO them
2167  *
2168  * @mapping: address space structure to write
2169  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2170  * @get_block: the filesystem's block mapper function.
2171  *
2172  * This is a library function, which implements the writepages()
2173  * address_space_operation.
2174  */
2175 static int mpage_da_writepages(struct address_space *mapping,
2176                                struct writeback_control *wbc,
2177                                struct mpage_da_data *mpd)
2178 {
2179         int ret;
2180
2181         if (!mpd->get_block)
2182                 return generic_writepages(mapping, wbc);
2183
2184         mpd->lbh.b_size = 0;
2185         mpd->lbh.b_state = 0;
2186         mpd->lbh.b_blocknr = 0;
2187         mpd->first_page = 0;
2188         mpd->next_page = 0;
2189         mpd->io_done = 0;
2190         mpd->pages_written = 0;
2191         mpd->retval = 0;
2192
2193         ret = write_cache_pages(mapping, wbc, __mpage_da_writepage, mpd);
2194         /*
2195          * Handle last extent of pages
2196          */
2197         if (!mpd->io_done && mpd->next_page != mpd->first_page) {
2198                 if (mpage_da_map_blocks(mpd) == 0)
2199                         mpage_da_submit_io(mpd);
2200
2201                 mpd->io_done = 1;
2202                 ret = MPAGE_DA_EXTENT_TAIL;
2203         }
2204         wbc->nr_to_write -= mpd->pages_written;
2205         return ret;
2206 }
2207
2208 /*
2209  * this is a special callback for ->write_begin() only
2210  * it's intention is to return mapped block or reserve space
2211  */
2212 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2213                                   struct buffer_head *bh_result, int create)
2214 {
2215         int ret = 0;
2216
2217         BUG_ON(create == 0);
2218         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2219
2220         /*
2221          * first, we need to know whether the block is allocated already
2222          * preallocated blocks are unmapped but should treated
2223          * the same as allocated blocks.
2224          */
2225         ret = ext4_get_blocks_wrap(NULL, inode, iblock, 1,  bh_result, 0, 0, 0);
2226         if ((ret == 0) && !buffer_delay(bh_result)) {
2227                 /* the block isn't (pre)allocated yet, let's reserve space */
2228                 /*
2229                  * XXX: __block_prepare_write() unmaps passed block,
2230                  * is it OK?
2231                  */
2232                 ret = ext4_da_reserve_space(inode, 1);
2233                 if (ret)
2234                         /* not enough space to reserve */
2235                         return ret;
2236
2237                 map_bh(bh_result, inode->i_sb, 0);
2238                 set_buffer_new(bh_result);
2239                 set_buffer_delay(bh_result);
2240         } else if (ret > 0) {
2241                 bh_result->b_size = (ret << inode->i_blkbits);
2242                 ret = 0;
2243         }
2244
2245         return ret;
2246 }
2247 #define         EXT4_DELALLOC_RSVED     1
2248 static int ext4_da_get_block_write(struct inode *inode, sector_t iblock,
2249                                    struct buffer_head *bh_result, int create)
2250 {
2251         int ret;
2252         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2253         loff_t disksize = EXT4_I(inode)->i_disksize;
2254         handle_t *handle = NULL;
2255
2256         handle = ext4_journal_current_handle();
2257         BUG_ON(!handle);
2258         ret = ext4_get_blocks_wrap(handle, inode, iblock, max_blocks,
2259                         bh_result, create, 0, EXT4_DELALLOC_RSVED);
2260         if (ret > 0) {
2261
2262                 bh_result->b_size = (ret << inode->i_blkbits);
2263
2264                 if (ext4_should_order_data(inode)) {
2265                         int retval;
2266                         retval = ext4_jbd2_file_inode(handle, inode);
2267                         if (retval)
2268                                 /*
2269                                  * Failed to add inode for ordered
2270                                  * mode. Don't update file size
2271                                  */
2272                                 return retval;
2273                 }
2274
2275                 /*
2276                  * Update on-disk size along with block allocation
2277                  * we don't use 'extend_disksize' as size may change
2278                  * within already allocated block -bzzz
2279                  */
2280                 disksize = ((loff_t) iblock + ret) << inode->i_blkbits;
2281                 if (disksize > i_size_read(inode))
2282                         disksize = i_size_read(inode);
2283                 if (disksize > EXT4_I(inode)->i_disksize) {
2284                         ext4_update_i_disksize(inode, disksize);
2285                         ret = ext4_mark_inode_dirty(handle, inode);
2286                         return ret;
2287                 }
2288                 ret = 0;
2289         }
2290         return ret;
2291 }
2292
2293 static int ext4_bh_unmapped_or_delay(handle_t *handle, struct buffer_head *bh)
2294 {
2295         /*
2296          * unmapped buffer is possible for holes.
2297          * delay buffer is possible with delayed allocation
2298          */
2299         return ((!buffer_mapped(bh) || buffer_delay(bh)) && buffer_dirty(bh));
2300 }
2301
2302 static int ext4_normal_get_block_write(struct inode *inode, sector_t iblock,
2303                                    struct buffer_head *bh_result, int create)
2304 {
2305         int ret = 0;
2306         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2307
2308         /*
2309          * we don't want to do block allocation in writepage
2310          * so call get_block_wrap with create = 0
2311          */
2312         ret = ext4_get_blocks_wrap(NULL, inode, iblock, max_blocks,
2313                                    bh_result, 0, 0, 0);
2314         if (ret > 0) {
2315                 bh_result->b_size = (ret << inode->i_blkbits);
2316                 ret = 0;
2317         }
2318         return ret;
2319 }
2320
2321 /*
2322  * get called vi ext4_da_writepages after taking page lock (have journal handle)
2323  * get called via journal_submit_inode_data_buffers (no journal handle)
2324  * get called via shrink_page_list via pdflush (no journal handle)
2325  * or grab_page_cache when doing write_begin (have journal handle)
2326  */
2327 static int ext4_da_writepage(struct page *page,
2328                                 struct writeback_control *wbc)
2329 {
2330         int ret = 0;
2331         loff_t size;
2332         unsigned int len;
2333         struct buffer_head *page_bufs;
2334         struct inode *inode = page->mapping->host;
2335
2336         trace_mark(ext4_da_writepage,
2337                    "dev %s ino %lu page_index %lu",
2338                    inode->i_sb->s_id, inode->i_ino, page->index);
2339         size = i_size_read(inode);
2340         if (page->index == size >> PAGE_CACHE_SHIFT)
2341                 len = size & ~PAGE_CACHE_MASK;
2342         else
2343                 len = PAGE_CACHE_SIZE;
2344
2345         if (page_has_buffers(page)) {
2346                 page_bufs = page_buffers(page);
2347                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2348                                         ext4_bh_unmapped_or_delay)) {
2349                         /*
2350                          * We don't want to do  block allocation
2351                          * So redirty the page and return
2352                          * We may reach here when we do a journal commit
2353                          * via journal_submit_inode_data_buffers.
2354                          * If we don't have mapping block we just ignore
2355                          * them. We can also reach here via shrink_page_list
2356                          */
2357                         redirty_page_for_writepage(wbc, page);
2358                         unlock_page(page);
2359                         return 0;
2360                 }
2361         } else {
2362                 /*
2363                  * The test for page_has_buffers() is subtle:
2364                  * We know the page is dirty but it lost buffers. That means
2365                  * that at some moment in time after write_begin()/write_end()
2366                  * has been called all buffers have been clean and thus they
2367                  * must have been written at least once. So they are all
2368                  * mapped and we can happily proceed with mapping them
2369                  * and writing the page.
2370                  *
2371                  * Try to initialize the buffer_heads and check whether
2372                  * all are mapped and non delay. We don't want to
2373                  * do block allocation here.
2374                  */
2375                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
2376                                                 ext4_normal_get_block_write);
2377                 if (!ret) {
2378                         page_bufs = page_buffers(page);
2379                         /* check whether all are mapped and non delay */
2380                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2381                                                 ext4_bh_unmapped_or_delay)) {
2382                                 redirty_page_for_writepage(wbc, page);
2383                                 unlock_page(page);
2384                                 return 0;
2385                         }
2386                 } else {
2387                         /*
2388                          * We can't do block allocation here
2389                          * so just redity the page and unlock
2390                          * and return
2391                          */
2392                         redirty_page_for_writepage(wbc, page);
2393                         unlock_page(page);
2394                         return 0;
2395                 }
2396                 /* now mark the buffer_heads as dirty and uptodate */
2397                 block_commit_write(page, 0, PAGE_CACHE_SIZE);
2398         }
2399
2400         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2401                 ret = nobh_writepage(page, ext4_normal_get_block_write, wbc);
2402         else
2403                 ret = block_write_full_page(page,
2404                                                 ext4_normal_get_block_write,
2405                                                 wbc);
2406
2407         return ret;
2408 }
2409
2410 /*
2411  * This is called via ext4_da_writepages() to
2412  * calulate the total number of credits to reserve to fit
2413  * a single extent allocation into a single transaction,
2414  * ext4_da_writpeages() will loop calling this before
2415  * the block allocation.
2416  */
2417
2418 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2419 {
2420         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2421
2422         /*
2423          * With non-extent format the journal credit needed to
2424          * insert nrblocks contiguous block is dependent on
2425          * number of contiguous block. So we will limit
2426          * number of contiguous block to a sane value
2427          */
2428         if (!(inode->i_flags & EXT4_EXTENTS_FL) &&
2429             (max_blocks > EXT4_MAX_TRANS_DATA))
2430                 max_blocks = EXT4_MAX_TRANS_DATA;
2431
2432         return ext4_chunk_trans_blocks(inode, max_blocks);
2433 }
2434
2435 static int ext4_da_writepages(struct address_space *mapping,
2436                               struct writeback_control *wbc)
2437 {
2438         pgoff_t index;
2439         int range_whole = 0;
2440         handle_t *handle = NULL;
2441         struct mpage_da_data mpd;
2442         struct inode *inode = mapping->host;
2443         int no_nrwrite_index_update;
2444         int pages_written = 0;
2445         long pages_skipped;
2446         int range_cyclic, cycled = 1, io_done = 0;
2447         int needed_blocks, ret = 0, nr_to_writebump = 0;
2448         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2449
2450         trace_mark(ext4_da_writepages,
2451                    "dev %s ino %lu nr_t_write %ld "
2452                    "pages_skipped %ld range_start %llu "
2453                    "range_end %llu nonblocking %d "
2454                    "for_kupdate %d for_reclaim %d "
2455                    "for_writepages %d range_cyclic %d",
2456                    inode->i_sb->s_id, inode->i_ino,
2457                    wbc->nr_to_write, wbc->pages_skipped,
2458                    (unsigned long long) wbc->range_start,
2459                    (unsigned long long) wbc->range_end,
2460                    wbc->nonblocking, wbc->for_kupdate,
2461                    wbc->for_reclaim, wbc->for_writepages,
2462                    wbc->range_cyclic);
2463
2464         /*
2465          * No pages to write? This is mainly a kludge to avoid starting
2466          * a transaction for special inodes like journal inode on last iput()
2467          * because that could violate lock ordering on umount
2468          */
2469         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2470                 return 0;
2471
2472         /*
2473          * If the filesystem has aborted, it is read-only, so return
2474          * right away instead of dumping stack traces later on that
2475          * will obscure the real source of the problem.  We test
2476          * EXT4_MOUNT_ABORT instead of sb->s_flag's MS_RDONLY because
2477          * the latter could be true if the filesystem is mounted
2478          * read-only, and in that case, ext4_da_writepages should
2479          * *never* be called, so if that ever happens, we would want
2480          * the stack trace.
2481          */
2482         if (unlikely(sbi->s_mount_opt & EXT4_MOUNT_ABORT))
2483                 return -EROFS;
2484
2485         /*
2486          * Make sure nr_to_write is >= sbi->s_mb_stream_request
2487          * This make sure small files blocks are allocated in
2488          * single attempt. This ensure that small files
2489          * get less fragmented.
2490          */
2491         if (wbc->nr_to_write < sbi->s_mb_stream_request) {
2492                 nr_to_writebump = sbi->s_mb_stream_request - wbc->nr_to_write;
2493                 wbc->nr_to_write = sbi->s_mb_stream_request;
2494         }
2495         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2496                 range_whole = 1;
2497
2498         range_cyclic = wbc->range_cyclic;
2499         if (wbc->range_cyclic) {
2500                 index = mapping->writeback_index;
2501                 if (index)
2502                         cycled = 0;
2503                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2504                 wbc->range_end  = LLONG_MAX;
2505                 wbc->range_cyclic = 0;
2506         } else
2507                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2508
2509         mpd.wbc = wbc;
2510         mpd.inode = mapping->host;
2511
2512         /*
2513          * we don't want write_cache_pages to update
2514          * nr_to_write and writeback_index
2515          */
2516         no_nrwrite_index_update = wbc->no_nrwrite_index_update;
2517         wbc->no_nrwrite_index_update = 1;
2518         pages_skipped = wbc->pages_skipped;
2519
2520 retry:
2521         while (!ret && wbc->nr_to_write > 0) {
2522
2523                 /*
2524                  * we  insert one extent at a time. So we need
2525                  * credit needed for single extent allocation.
2526                  * journalled mode is currently not supported
2527                  * by delalloc
2528                  */
2529                 BUG_ON(ext4_should_journal_data(inode));
2530                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2531
2532                 /* start a new transaction*/
2533                 handle = ext4_journal_start(inode, needed_blocks);
2534                 if (IS_ERR(handle)) {
2535                         ret = PTR_ERR(handle);
2536                         printk(KERN_CRIT "%s: jbd2_start: "
2537                                "%ld pages, ino %lu; err %d\n", __func__,
2538                                 wbc->nr_to_write, inode->i_ino, ret);
2539                         dump_stack();
2540                         goto out_writepages;
2541                 }
2542                 mpd.get_block = ext4_da_get_block_write;
2543                 ret = mpage_da_writepages(mapping, wbc, &mpd);
2544
2545                 ext4_journal_stop(handle);
2546
2547                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2548                         /* commit the transaction which would
2549                          * free blocks released in the transaction
2550                          * and try again
2551                          */
2552                         jbd2_journal_force_commit_nested(sbi->s_journal);
2553                         wbc->pages_skipped = pages_skipped;
2554                         ret = 0;
2555                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2556                         /*
2557                          * got one extent now try with
2558                          * rest of the pages
2559                          */
2560                         pages_written += mpd.pages_written;
2561                         wbc->pages_skipped = pages_skipped;
2562                         ret = 0;
2563                         io_done = 1;
2564                 } else if (wbc->nr_to_write)
2565                         /*
2566                          * There is no more writeout needed
2567                          * or we requested for a noblocking writeout
2568                          * and we found the device congested
2569                          */
2570                         break;
2571         }
2572         if (!io_done && !cycled) {
2573                 cycled = 1;
2574                 index = 0;
2575                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2576                 wbc->range_end  = mapping->writeback_index - 1;
2577                 goto retry;
2578         }
2579         if (pages_skipped != wbc->pages_skipped)
2580                 printk(KERN_EMERG "This should not happen leaving %s "
2581                                 "with nr_to_write = %ld ret = %d\n",
2582                                 __func__, wbc->nr_to_write, ret);
2583
2584         /* Update index */
2585         index += pages_written;
2586         wbc->range_cyclic = range_cyclic;
2587         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2588                 /*
2589                  * set the writeback_index so that range_cyclic
2590                  * mode will write it back later
2591                  */
2592                 mapping->writeback_index = index;
2593
2594 out_writepages:
2595         if (!no_nrwrite_index_update)
2596                 wbc->no_nrwrite_index_update = 0;
2597         wbc->nr_to_write -= nr_to_writebump;
2598         trace_mark(ext4_da_writepage_result,
2599                    "dev %s ino %lu ret %d pages_written %d "
2600                    "pages_skipped %ld congestion %d "
2601                    "more_io %d no_nrwrite_index_update %d",
2602                    inode->i_sb->s_id, inode->i_ino, ret,
2603                    pages_written, wbc->pages_skipped,
2604                    wbc->encountered_congestion, wbc->more_io,
2605                    wbc->no_nrwrite_index_update);
2606         return ret;
2607 }
2608
2609 #define FALL_BACK_TO_NONDELALLOC 1
2610 static int ext4_nonda_switch(struct super_block *sb)
2611 {
2612         s64 free_blocks, dirty_blocks;
2613         struct ext4_sb_info *sbi = EXT4_SB(sb);
2614
2615         /*
2616          * switch to non delalloc mode if we are running low
2617          * on free block. The free block accounting via percpu
2618          * counters can get slightly wrong with percpu_counter_batch getting
2619          * accumulated on each CPU without updating global counters
2620          * Delalloc need an accurate free block accounting. So switch
2621          * to non delalloc when we are near to error range.
2622          */
2623         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
2624         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
2625         if (2 * free_blocks < 3 * dirty_blocks ||
2626                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
2627                 /*
2628                  * free block count is less that 150% of dirty blocks
2629                  * or free blocks is less that watermark
2630                  */
2631                 return 1;
2632         }
2633         return 0;
2634 }
2635
2636 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2637                                 loff_t pos, unsigned len, unsigned flags,
2638                                 struct page **pagep, void **fsdata)
2639 {
2640         int ret, retries = 0;
2641         struct page *page;
2642         pgoff_t index;
2643         unsigned from, to;
2644         struct inode *inode = mapping->host;
2645         handle_t *handle;
2646
2647         index = pos >> PAGE_CACHE_SHIFT;
2648         from = pos & (PAGE_CACHE_SIZE - 1);
2649         to = from + len;
2650
2651         if (ext4_nonda_switch(inode->i_sb)) {
2652                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2653                 return ext4_write_begin(file, mapping, pos,
2654                                         len, flags, pagep, fsdata);
2655         }
2656         *fsdata = (void *)0;
2657
2658         trace_mark(ext4_da_write_begin,
2659                    "dev %s ino %lu pos %llu len %u flags %u",
2660                    inode->i_sb->s_id, inode->i_ino,
2661                    (unsigned long long) pos, len, flags);
2662 retry:
2663         /*
2664          * With delayed allocation, we don't log the i_disksize update
2665          * if there is delayed block allocation. But we still need
2666          * to journalling the i_disksize update if writes to the end
2667          * of file which has an already mapped buffer.
2668          */
2669         handle = ext4_journal_start(inode, 1);
2670         if (IS_ERR(handle)) {
2671                 ret = PTR_ERR(handle);
2672                 goto out;
2673         }
2674         /* We cannot recurse into the filesystem as the transaction is already
2675          * started */
2676         flags |= AOP_FLAG_NOFS;
2677
2678         page = grab_cache_page_write_begin(mapping, index, flags);
2679         if (!page) {
2680                 ext4_journal_stop(handle);
2681                 ret = -ENOMEM;
2682                 goto out;
2683         }
2684         *pagep = page;
2685
2686         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
2687                                                         ext4_da_get_block_prep);
2688         if (ret < 0) {
2689                 unlock_page(page);
2690                 ext4_journal_stop(handle);
2691                 page_cache_release(page);
2692                 /*
2693                  * block_write_begin may have instantiated a few blocks
2694                  * outside i_size.  Trim these off again. Don't need
2695                  * i_size_read because we hold i_mutex.
2696                  */
2697                 if (pos + len > inode->i_size)
2698                         vmtruncate(inode, inode->i_size);
2699         }
2700
2701         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
2702                 goto retry;
2703 out:
2704         return ret;
2705 }
2706
2707 /*
2708  * Check if we should update i_disksize
2709  * when write to the end of file but not require block allocation
2710  */
2711 static int ext4_da_should_update_i_disksize(struct page *page,
2712                                          unsigned long offset)
2713 {
2714         struct buffer_head *bh;
2715         struct inode *inode = page->mapping->host;
2716         unsigned int idx;
2717         int i;
2718
2719         bh = page_buffers(page);
2720         idx = offset >> inode->i_blkbits;
2721
2722         for (i = 0; i < idx; i++)
2723                 bh = bh->b_this_page;
2724
2725         if (!buffer_mapped(bh) || (buffer_delay(bh)))
2726                 return 0;
2727         return 1;
2728 }
2729
2730 static int ext4_da_write_end(struct file *file,
2731                                 struct address_space *mapping,
2732                                 loff_t pos, unsigned len, unsigned copied,
2733                                 struct page *page, void *fsdata)
2734 {
2735         struct inode *inode = mapping->host;
2736         int ret = 0, ret2;
2737         handle_t *handle = ext4_journal_current_handle();
2738         loff_t new_i_size;
2739         unsigned long start, end;
2740         int write_mode = (int)(unsigned long)fsdata;
2741
2742         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
2743                 if (ext4_should_order_data(inode)) {
2744                         return ext4_ordered_write_end(file, mapping, pos,
2745                                         len, copied, page, fsdata);
2746                 } else if (ext4_should_writeback_data(inode)) {
2747                         return ext4_writeback_write_end(file, mapping, pos,
2748                                         len, copied, page, fsdata);
2749                 } else {
2750                         BUG();
2751                 }
2752         }
2753
2754         trace_mark(ext4_da_write_end,
2755                    "dev %s ino %lu pos %llu len %u copied %u",
2756                    inode->i_sb->s_id, inode->i_ino,
2757                    (unsigned long long) pos, len, copied);
2758         start = pos & (PAGE_CACHE_SIZE - 1);
2759         end = start + copied - 1;
2760
2761         /*
2762          * generic_write_end() will run mark_inode_dirty() if i_size
2763          * changes.  So let's piggyback the i_disksize mark_inode_dirty
2764          * into that.
2765          */
2766
2767         new_i_size = pos + copied;
2768         if (new_i_size > EXT4_I(inode)->i_disksize) {
2769                 if (ext4_da_should_update_i_disksize(page, end)) {
2770                         down_write(&EXT4_I(inode)->i_data_sem);
2771                         if (new_i_size > EXT4_I(inode)->i_disksize) {
2772                                 /*
2773                                  * Updating i_disksize when extending file
2774                                  * without needing block allocation
2775                                  */
2776                                 if (ext4_should_order_data(inode))
2777                                         ret = ext4_jbd2_file_inode(handle,
2778                                                                    inode);
2779
2780                                 EXT4_I(inode)->i_disksize = new_i_size;
2781                         }
2782                         up_write(&EXT4_I(inode)->i_data_sem);
2783                         /* We need to mark inode dirty even if
2784                          * new_i_size is less that inode->i_size
2785                          * bu greater than i_disksize.(hint delalloc)
2786                          */
2787                         ext4_mark_inode_dirty(handle, inode);
2788                 }
2789         }
2790         ret2 = generic_write_end(file, mapping, pos, len, copied,
2791                                                         page, fsdata);
2792         copied = ret2;
2793         if (ret2 < 0)
2794                 ret = ret2;
2795         ret2 = ext4_journal_stop(handle);
2796         if (!ret)
2797                 ret = ret2;
2798
2799         return ret ? ret : copied;
2800 }
2801
2802 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
2803 {
2804         /*
2805          * Drop reserved blocks
2806          */
2807         BUG_ON(!PageLocked(page));
2808         if (!page_has_buffers(page))
2809                 goto out;
2810
2811         ext4_da_page_release_reservation(page, offset);
2812
2813 out:
2814         ext4_invalidatepage(page, offset);
2815
2816         return;
2817 }
2818
2819
2820 /*
2821  * bmap() is special.  It gets used by applications such as lilo and by
2822  * the swapper to find the on-disk block of a specific piece of data.
2823  *
2824  * Naturally, this is dangerous if the block concerned is still in the
2825  * journal.  If somebody makes a swapfile on an ext4 data-journaling
2826  * filesystem and enables swap, then they may get a nasty shock when the
2827  * data getting swapped to that swapfile suddenly gets overwritten by
2828  * the original zero's written out previously to the journal and
2829  * awaiting writeback in the kernel's buffer cache.
2830  *
2831  * So, if we see any bmap calls here on a modified, data-journaled file,
2832  * take extra steps to flush any blocks which might be in the cache.
2833  */
2834 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
2835 {
2836         struct inode *inode = mapping->host;
2837         journal_t *journal;
2838         int err;
2839
2840         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
2841                         test_opt(inode->i_sb, DELALLOC)) {
2842                 /*
2843                  * With delalloc we want to sync the file
2844                  * so that we can make sure we allocate
2845                  * blocks for file
2846                  */
2847                 filemap_write_and_wait(mapping);
2848         }
2849
2850         if (EXT4_JOURNAL(inode) && EXT4_I(inode)->i_state & EXT4_STATE_JDATA) {
2851                 /*
2852                  * This is a REALLY heavyweight approach, but the use of
2853                  * bmap on dirty files is expected to be extremely rare:
2854                  * only if we run lilo or swapon on a freshly made file
2855                  * do we expect this to happen.
2856                  *
2857                  * (bmap requires CAP_SYS_RAWIO so this does not
2858                  * represent an unprivileged user DOS attack --- we'd be
2859                  * in trouble if mortal users could trigger this path at
2860                  * will.)
2861                  *
2862                  * NB. EXT4_STATE_JDATA is not set on files other than
2863                  * regular files.  If somebody wants to bmap a directory
2864                  * or symlink and gets confused because the buffer
2865                  * hasn't yet been flushed to disk, they deserve
2866                  * everything they get.
2867                  */
2868
2869                 EXT4_I(inode)->i_state &= ~EXT4_STATE_JDATA;
2870                 journal = EXT4_JOURNAL(inode);
2871                 jbd2_journal_lock_updates(journal);
2872                 err = jbd2_journal_flush(journal);
2873                 jbd2_journal_unlock_updates(journal);
2874
2875                 if (err)
2876                         return 0;
2877         }
2878
2879         return generic_block_bmap(mapping, block, ext4_get_block);
2880 }
2881
2882 static int bget_one(handle_t *handle, struct buffer_head *bh)
2883 {
2884         get_bh(bh);
2885         return 0;
2886 }
2887
2888 static int bput_one(handle_t *handle, struct buffer_head *bh)
2889 {
2890         put_bh(bh);
2891         return 0;
2892 }
2893
2894 /*
2895  * Note that we don't need to start a transaction unless we're journaling data
2896  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2897  * need to file the inode to the transaction's list in ordered mode because if
2898  * we are writing back data added by write(), the inode is already there and if
2899  * we are writing back data modified via mmap(), noone guarantees in which
2900  * transaction the data will hit the disk. In case we are journaling data, we
2901  * cannot start transaction directly because transaction start ranks above page
2902  * lock so we have to do some magic.
2903  *
2904  * In all journaling modes block_write_full_page() will start the I/O.
2905  *
2906  * Problem:
2907  *
2908  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2909  *              ext4_writepage()
2910  *
2911  * Similar for:
2912  *
2913  *      ext4_file_write() -> generic_file_write() -> __alloc_pages() -> ...
2914  *
2915  * Same applies to ext4_get_block().  We will deadlock on various things like
2916  * lock_journal and i_data_sem
2917  *
2918  * Setting PF_MEMALLOC here doesn't work - too many internal memory
2919  * allocations fail.
2920  *
2921  * 16May01: If we're reentered then journal_current_handle() will be
2922  *          non-zero. We simply *return*.
2923  *
2924  * 1 July 2001: @@@ FIXME:
2925  *   In journalled data mode, a data buffer may be metadata against the
2926  *   current transaction.  But the same file is part of a shared mapping
2927  *   and someone does a writepage() on it.
2928  *
2929  *   We will move the buffer onto the async_data list, but *after* it has
2930  *   been dirtied. So there's a small window where we have dirty data on
2931  *   BJ_Metadata.
2932  *
2933  *   Note that this only applies to the last partial page in the file.  The
2934  *   bit which block_write_full_page() uses prepare/commit for.  (That's
2935  *   broken code anyway: it's wrong for msync()).
2936  *
2937  *   It's a rare case: affects the final partial page, for journalled data
2938  *   where the file is subject to bith write() and writepage() in the same
2939  *   transction.  To fix it we'll need a custom block_write_full_page().
2940  *   We'll probably need that anyway for journalling writepage() output.
2941  *
2942  * We don't honour synchronous mounts for writepage().  That would be
2943  * disastrous.  Any write() or metadata operation will sync the fs for
2944  * us.
2945  *
2946  */
2947 static int __ext4_normal_writepage(struct page *page,
2948                                 struct writeback_control *wbc)
2949 {
2950         struct inode *inode = page->mapping->host;
2951
2952         if (test_opt(inode->i_sb, NOBH))
2953                 return nobh_writepage(page,
2954                                         ext4_normal_get_block_write, wbc);
2955         else
2956                 return block_write_full_page(page,
2957                                                 ext4_normal_get_block_write,
2958                                                 wbc);
2959 }
2960
2961 static int ext4_normal_writepage(struct page *page,
2962                                 struct writeback_control *wbc)
2963 {
2964         struct inode *inode = page->mapping->host;
2965         loff_t size = i_size_read(inode);
2966         loff_t len;
2967
2968         trace_mark(ext4_normal_writepage,
2969                    "dev %s ino %lu page_index %lu",
2970                    inode->i_sb->s_id, inode->i_ino, page->index);
2971         J_ASSERT(PageLocked(page));
2972         if (page->index == size >> PAGE_CACHE_SHIFT)
2973                 len = size & ~PAGE_CACHE_MASK;
2974         else
2975                 len = PAGE_CACHE_SIZE;
2976
2977         if (page_has_buffers(page)) {
2978                 /* if page has buffers it should all be mapped
2979                  * and allocated. If there are not buffers attached
2980                  * to the page we know the page is dirty but it lost
2981                  * buffers. That means that at some moment in time
2982                  * after write_begin() / write_end() has been called
2983                  * all buffers have been clean and thus they must have been
2984                  * written at least once. So they are all mapped and we can
2985                  * happily proceed with mapping them and writing the page.
2986                  */
2987                 BUG_ON(walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
2988                                         ext4_bh_unmapped_or_delay));
2989         }
2990
2991         if (!ext4_journal_current_handle())
2992                 return __ext4_normal_writepage(page, wbc);
2993
2994         redirty_page_for_writepage(wbc, page);
2995         unlock_page(page);
2996         return 0;
2997 }
2998
2999 static int __ext4_journalled_writepage(struct page *page,
3000                                 struct writeback_control *wbc)
3001 {
3002         struct address_space *mapping = page->mapping;
3003         struct inode *inode = mapping->host;
3004         struct buffer_head *page_bufs;
3005         handle_t *handle = NULL;
3006         int ret = 0;
3007         int err;
3008
3009         ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
3010                                         ext4_normal_get_block_write);
3011         if (ret != 0)
3012                 goto out_unlock;
3013
3014         page_bufs = page_buffers(page);
3015         walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE, NULL,
3016                                                                 bget_one);
3017         /* As soon as we unlock the page, it can go away, but we have
3018          * references to buffers so we are safe */
3019         unlock_page(page);
3020
3021         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
3022         if (IS_ERR(handle)) {
3023                 ret = PTR_ERR(handle);
3024                 goto out;
3025         }
3026
3027         ret = walk_page_buffers(handle, page_bufs, 0,
3028                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
3029
3030         err = walk_page_buffers(handle, page_bufs, 0,
3031                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
3032         if (ret == 0)
3033                 ret = err;
3034         err = ext4_journal_stop(handle);
3035         if (!ret)
3036                 ret = err;
3037
3038         walk_page_buffers(handle, page_bufs, 0,
3039                                 PAGE_CACHE_SIZE, NULL, bput_one);
3040         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
3041         goto out;
3042
3043 out_unlock:
3044         unlock_page(page);
3045 out:
3046         return ret;
3047 }
3048
3049 static int ext4_journalled_writepage(struct page *page,
3050                                 struct writeback_control *wbc)
3051 {
3052         struct inode *inode = page->mapping->host;
3053         loff_t size = i_size_read(inode);
3054         loff_t len;
3055
3056         trace_mark(ext4_journalled_writepage,
3057                    "dev %s ino %lu page_index %lu",
3058                    inode->i_sb->s_id, inode->i_ino, page->index);
3059         J_ASSERT(PageLocked(page));
3060         if (page->index == size >> PAGE_CACHE_SHIFT)
3061                 len = size & ~PAGE_CACHE_MASK;
3062         else
3063                 len = PAGE_CACHE_SIZE;
3064
3065         if (page_has_buffers(page)) {
3066                 /* if page has buffers it should all be mapped
3067                  * and allocated. If there are not buffers attached
3068                  * to the page we know the page is dirty but it lost
3069                  * buffers. That means that at some moment in time
3070                  * after write_begin() / write_end() has been called
3071                  * all buffers have been clean and thus they must have been
3072                  * written at least once. So they are all mapped and we can
3073                  * happily proceed with mapping them and writing the page.
3074                  */
3075                 BUG_ON(walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
3076                                         ext4_bh_unmapped_or_delay));
3077         }
3078
3079         if (ext4_journal_current_handle())
3080                 goto no_write;
3081
3082         if (PageChecked(page)) {
3083                 /*
3084                  * It's mmapped pagecache.  Add buffers and journal it.  There
3085                  * doesn't seem much point in redirtying the page here.
3086                  */
3087                 ClearPageChecked(page);
3088                 return __ext4_journalled_writepage(page, wbc);
3089         } else {
3090                 /*
3091                  * It may be a page full of checkpoint-mode buffers.  We don't
3092                  * really know unless we go poke around in the buffer_heads.
3093                  * But block_write_full_page will do the right thing.
3094                  */
3095                 return block_write_full_page(page,
3096                                                 ext4_normal_get_block_write,
3097                                                 wbc);
3098         }
3099 no_write:
3100         redirty_page_for_writepage(wbc, page);
3101         unlock_page(page);
3102         return 0;
3103 }
3104
3105 static int ext4_readpage(struct file *file, struct page *page)
3106 {
3107         return mpage_readpage(page, ext4_get_block);
3108 }
3109
3110 static int
3111 ext4_readpages(struct file *file, struct address_space *mapping,
3112                 struct list_head *pages, unsigned nr_pages)
3113 {
3114         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3115 }
3116
3117 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3118 {
3119         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3120
3121         /*
3122          * If it's a full truncate we just forget about the pending dirtying
3123          */
3124         if (offset == 0)
3125                 ClearPageChecked(page);
3126
3127         if (journal)
3128                 jbd2_journal_invalidatepage(journal, page, offset);
3129         else
3130                 block_invalidatepage(page, offset);
3131 }
3132
3133 static int ext4_releasepage(struct page *page, gfp_t wait)
3134 {
3135         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3136
3137         WARN_ON(PageChecked(page));
3138         if (!page_has_buffers(page))
3139                 return 0;
3140         if (journal)
3141                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3142         else
3143                 return try_to_free_buffers(page);
3144 }
3145
3146 /*
3147  * If the O_DIRECT write will extend the file then add this inode to the
3148  * orphan list.  So recovery will truncate it back to the original size
3149  * if the machine crashes during the write.
3150  *
3151  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3152  * crashes then stale disk data _may_ be exposed inside the file. But current
3153  * VFS code falls back into buffered path in that case so we are safe.
3154  */
3155 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3156                         const struct iovec *iov, loff_t offset,
3157                         unsigned long nr_segs)
3158 {
3159         struct file *file = iocb->ki_filp;
3160         struct inode *inode = file->f_mapping->host;
3161         struct ext4_inode_info *ei = EXT4_I(inode);
3162         handle_t *handle;
3163         ssize_t ret;
3164         int orphan = 0;
3165         size_t count = iov_length(iov, nr_segs);
3166
3167         if (rw == WRITE) {
3168                 loff_t final_size = offset + count;
3169
3170                 if (final_size > inode->i_size) {
3171                         /* Credits for sb + inode write */
3172                         handle = ext4_journal_start(inode, 2);
3173                         if (IS_ERR(handle)) {
3174                                 ret = PTR_ERR(handle);
3175                                 goto out;
3176                         }
3177                         ret = ext4_orphan_add(handle, inode);
3178                         if (ret) {
3179                                 ext4_journal_stop(handle);
3180                                 goto out;
3181                         }
3182                         orphan = 1;
3183                         ei->i_disksize = inode->i_size;
3184                         ext4_journal_stop(handle);
3185                 }
3186         }
3187
3188         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
3189                                  offset, nr_segs,
3190                                  ext4_get_block, NULL);
3191
3192         if (orphan) {
3193                 int err;
3194
3195                 /* Credits for sb + inode write */
3196                 handle = ext4_journal_start(inode, 2);
3197                 if (IS_ERR(handle)) {
3198                         /* This is really bad luck. We've written the data
3199                          * but cannot extend i_size. Bail out and pretend
3200                          * the write failed... */
3201                         ret = PTR_ERR(handle);
3202                         goto out;
3203                 }
3204                 if (inode->i_nlink)
3205                         ext4_orphan_del(handle, inode);
3206                 if (ret > 0) {
3207                         loff_t end = offset + ret;
3208                         if (end > inode->i_size) {
3209                                 ei->i_disksize = end;
3210                                 i_size_write(inode, end);
3211                                 /*
3212                                  * We're going to return a positive `ret'
3213                                  * here due to non-zero-length I/O, so there's
3214                                  * no way of reporting error returns from
3215                                  * ext4_mark_inode_dirty() to userspace.  So
3216                                  * ignore it.
3217                                  */
3218                                 ext4_mark_inode_dirty(handle, inode);
3219                         }
3220                 }
3221                 err = ext4_journal_stop(handle);
3222                 if (ret == 0)
3223                         ret = err;
3224         }
3225 out:
3226         return ret;
3227 }
3228
3229 /*
3230  * Pages can be marked dirty completely asynchronously from ext4's journalling
3231  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3232  * much here because ->set_page_dirty is called under VFS locks.  The page is
3233  * not necessarily locked.
3234  *
3235  * We cannot just dirty the page and leave attached buffers clean, because the
3236  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3237  * or jbddirty because all the journalling code will explode.
3238  *
3239  * So what we do is to mark the page "pending dirty" and next time writepage
3240  * is called, propagate that into the buffers appropriately.
3241  */
3242 static int ext4_journalled_set_page_dirty(struct page *page)
3243 {
3244         SetPageChecked(page);
3245         return __set_page_dirty_nobuffers(page);
3246 }
3247
3248 static const struct address_space_operations ext4_ordered_aops = {
3249         .readpage               = ext4_readpage,
3250         .readpages              = ext4_readpages,
3251         .writepage              = ext4_normal_writepage,
3252         .sync_page              = block_sync_page,
3253         .write_begin            = ext4_write_begin,
3254         .write_end              = ext4_ordered_write_end,
3255         .bmap                   = ext4_bmap,
3256         .invalidatepage         = ext4_invalidatepage,
3257         .releasepage            = ext4_releasepage,
3258         .direct_IO              = ext4_direct_IO,
3259         .migratepage            = buffer_migrate_page,
3260         .is_partially_uptodate  = block_is_partially_uptodate,
3261 };
3262
3263 static const struct address_space_operations ext4_writeback_aops = {
3264         .readpage               = ext4_readpage,
3265         .readpages              = ext4_readpages,
3266         .writepage              = ext4_normal_writepage,
3267         .sync_page              = block_sync_page,
3268         .write_begin            = ext4_write_begin,
3269         .write_end              = ext4_writeback_write_end,
3270         .bmap                   = ext4_bmap,
3271         .invalidatepage         = ext4_invalidatepage,
3272         .releasepage            = ext4_releasepage,
3273         .direct_IO              = ext4_direct_IO,
3274         .migratepage            = buffer_migrate_page,
3275         .is_partially_uptodate  = block_is_partially_uptodate,
3276 };
3277
3278 static const struct address_space_operations ext4_journalled_aops = {
3279         .readpage               = ext4_readpage,
3280         .readpages              = ext4_readpages,
3281         .writepage              = ext4_journalled_writepage,
3282         .sync_page              = block_sync_page,
3283         .write_begin            = ext4_write_begin,
3284         .write_end              = ext4_journalled_write_end,
3285         .set_page_dirty         = ext4_journalled_set_page_dirty,
3286         .bmap                   = ext4_bmap,
3287         .invalidatepage         = ext4_invalidatepage,
3288         .releasepage            = ext4_releasepage,
3289         .is_partially_uptodate  = block_is_partially_uptodate,
3290 };
3291
3292 static const struct address_space_operations ext4_da_aops = {
3293         .readpage               = ext4_readpage,
3294         .readpages              = ext4_readpages,
3295         .writepage              = ext4_da_writepage,
3296         .writepages             = ext4_da_writepages,
3297         .sync_page              = block_sync_page,
3298         .write_begin            = ext4_da_write_begin,
3299         .write_end              = ext4_da_write_end,
3300         .bmap                   = ext4_bmap,
3301         .invalidatepage         = ext4_da_invalidatepage,
3302         .releasepage            = ext4_releasepage,
3303         .direct_IO              = ext4_direct_IO,
3304         .migratepage            = buffer_migrate_page,
3305         .is_partially_uptodate  = block_is_partially_uptodate,
3306 };
3307
3308 void ext4_set_aops(struct inode *inode)
3309 {
3310         if (ext4_should_order_data(inode) &&
3311                 test_opt(inode->i_sb, DELALLOC))
3312                 inode->i_mapping->a_ops = &ext4_da_aops;
3313         else if (ext4_should_order_data(inode))
3314                 inode->i_mapping->a_ops = &ext4_ordered_aops;
3315         else if (ext4_should_writeback_data(inode) &&
3316                  test_opt(inode->i_sb, DELALLOC))
3317                 inode->i_mapping->a_ops = &ext4_da_aops;
3318         else if (ext4_should_writeback_data(inode))
3319                 inode->i_mapping->a_ops = &ext4_writeback_aops;
3320         else
3321                 inode->i_mapping->a_ops = &ext4_journalled_aops;
3322 }
3323
3324 /*
3325  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3326  * up to the end of the block which corresponds to `from'.
3327  * This required during truncate. We need to physically zero the tail end
3328  * of that block so it doesn't yield old data if the file is later grown.
3329  */
3330 int ext4_block_truncate_page(handle_t *handle,
3331                 struct address_space *mapping, loff_t from)
3332 {
3333         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3334         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3335         unsigned blocksize, length, pos;
3336         ext4_lblk_t iblock;
3337         struct inode *inode = mapping->host;
3338         struct buffer_head *bh;
3339         struct page *page;
3340         int err = 0;
3341
3342         page = grab_cache_page(mapping, from >> PAGE_CACHE_SHIFT);
3343         if (!page)
3344                 return -EINVAL;
3345
3346         blocksize = inode->i_sb->s_blocksize;
3347         length = blocksize - (offset & (blocksize - 1));
3348         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3349
3350         /*
3351          * For "nobh" option,  we can only work if we don't need to
3352          * read-in the page - otherwise we create buffers to do the IO.
3353          */
3354         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
3355              ext4_should_writeback_data(inode) && PageUptodate(page)) {
3356                 zero_user(page, offset, length);
3357                 set_page_dirty(page);
3358                 goto unlock;
3359         }
3360
3361         if (!page_has_buffers(page))
3362                 create_empty_buffers(page, blocksize, 0);
3363
3364         /* Find the buffer that contains "offset" */
3365         bh = page_buffers(page);
3366         pos = blocksize;
3367         while (offset >= pos) {
3368                 bh = bh->b_this_page;
3369                 iblock++;
3370                 pos += blocksize;
3371         }
3372
3373         err = 0;
3374         if (buffer_freed(bh)) {
3375                 BUFFER_TRACE(bh, "freed: skip");
3376                 goto unlock;
3377         }
3378
3379         if (!buffer_mapped(bh)) {
3380                 BUFFER_TRACE(bh, "unmapped");
3381                 ext4_get_block(inode, iblock, bh, 0);
3382                 /* unmapped? It's a hole - nothing to do */
3383                 if (!buffer_mapped(bh)) {
3384                         BUFFER_TRACE(bh, "still unmapped");
3385                         goto unlock;
3386                 }
3387         }
3388
3389         /* Ok, it's mapped. Make sure it's up-to-date */
3390         if (PageUptodate(page))
3391                 set_buffer_uptodate(bh);
3392
3393         if (!buffer_uptodate(bh)) {
3394                 err = -EIO;
3395                 ll_rw_block(READ, 1, &bh);
3396                 wait_on_buffer(bh);
3397                 /* Uhhuh. Read error. Complain and punt. */
3398                 if (!buffer_uptodate(bh))
3399                         goto unlock;
3400         }
3401
3402         if (ext4_should_journal_data(inode)) {
3403                 BUFFER_TRACE(bh, "get write access");
3404                 err = ext4_journal_get_write_access(handle, bh);
3405                 if (err)
3406                         goto unlock;
3407         }
3408
3409         zero_user(page, offset, length);
3410
3411         BUFFER_TRACE(bh, "zeroed end of block");
3412
3413         err = 0;
3414         if (ext4_should_journal_data(inode)) {
3415                 err = ext4_handle_dirty_metadata(handle, inode, bh);
3416         } else {
3417                 if (ext4_should_order_data(inode))
3418                         err = ext4_jbd2_file_inode(handle, inode);
3419                 mark_buffer_dirty(bh);
3420         }
3421
3422 unlock:
3423         unlock_page(page);
3424         page_cache_release(page);
3425         return err;
3426 }
3427
3428 /*
3429  * Probably it should be a library function... search for first non-zero word
3430  * or memcmp with zero_page, whatever is better for particular architecture.
3431  * Linus?
3432  */
3433 static inline int all_zeroes(__le32 *p, __le32 *q)
3434 {
3435         while (p < q)
3436                 if (*p++)
3437                         return 0;
3438         return 1;
3439 }
3440
3441 /**
3442  *      ext4_find_shared - find the indirect blocks for partial truncation.
3443  *      @inode:   inode in question
3444  *      @depth:   depth of the affected branch
3445  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
3446  *      @chain:   place to store the pointers to partial indirect blocks
3447  *      @top:     place to the (detached) top of branch
3448  *
3449  *      This is a helper function used by ext4_truncate().
3450  *
3451  *      When we do truncate() we may have to clean the ends of several
3452  *      indirect blocks but leave the blocks themselves alive. Block is
3453  *      partially truncated if some data below the new i_size is refered
3454  *      from it (and it is on the path to the first completely truncated
3455  *      data block, indeed).  We have to free the top of that path along
3456  *      with everything to the right of the path. Since no allocation
3457  *      past the truncation point is possible until ext4_truncate()
3458  *      finishes, we may safely do the latter, but top of branch may
3459  *      require special attention - pageout below the truncation point
3460  *      might try to populate it.
3461  *
3462  *      We atomically detach the top of branch from the tree, store the
3463  *      block number of its root in *@top, pointers to buffer_heads of
3464  *      partially truncated blocks - in @chain[].bh and pointers to
3465  *      their last elements that should not be removed - in
3466  *      @chain[].p. Return value is the pointer to last filled element
3467  *      of @chain.
3468  *
3469  *      The work left to caller to do the actual freeing of subtrees:
3470  *              a) free the subtree starting from *@top
3471  *              b) free the subtrees whose roots are stored in
3472  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
3473  *              c) free the subtrees growing from the inode past the @chain[0].
3474  *                      (no partially truncated stuff there).  */
3475
3476 static Indirect *ext4_find_shared(struct inode *inode, int depth,
3477                         ext4_lblk_t offsets[4], Indirect chain[4], __le32 *top)
3478 {
3479         Indirect *partial, *p;
3480         int k, err;
3481
3482         *top = 0;
3483         /* Make k index the deepest non-null offest + 1 */
3484         for (k = depth; k > 1 && !offsets[k-1]; k--)
3485                 ;
3486         partial = ext4_get_branch(inode, k, offsets, chain, &err);
3487         /* Writer: pointers */
3488         if (!partial)
3489                 partial = chain + k-1;
3490         /*
3491          * If the branch acquired continuation since we've looked at it -
3492          * fine, it should all survive and (new) top doesn't belong to us.
3493          */
3494         if (!partial->key && *partial->p)
3495                 /* Writer: end */
3496                 goto no_top;
3497         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
3498                 ;
3499         /*
3500          * OK, we've found the last block that must survive. The rest of our
3501          * branch should be detached before unlocking. However, if that rest
3502          * of branch is all ours and does not grow immediately from the inode
3503          * it's easier to cheat and just decrement partial->p.
3504          */
3505         if (p == chain + k - 1 && p > chain) {
3506                 p->p--;
3507         } else {
3508                 *top = *p->p;
3509                 /* Nope, don't do this in ext4.  Must leave the tree intact */
3510 #if 0
3511                 *p->p = 0;
3512 #endif
3513         }
3514         /* Writer: end */
3515
3516         while (partial > p) {
3517                 brelse(partial->bh);
3518                 partial--;
3519         }
3520 no_top:
3521         return partial;
3522 }
3523
3524 /*
3525  * Zero a number of block pointers in either an inode or an indirect block.
3526  * If we restart the transaction we must again get write access to the
3527  * indirect block for further modification.
3528  *
3529  * We release `count' blocks on disk, but (last - first) may be greater
3530  * than `count' because there can be holes in there.
3531  */
3532 static void ext4_clear_blocks(handle_t *handle, struct inode *inode,
3533                 struct buffer_head *bh, ext4_fsblk_t block_to_free,
3534                 unsigned long count, __le32 *first, __le32 *last)
3535 {
3536         __le32 *p;
3537         if (try_to_extend_transaction(handle, inode)) {
3538                 if (bh) {
3539                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
3540                         ext4_handle_dirty_metadata(handle, inode, bh);
3541                 }
3542                 ext4_mark_inode_dirty(handle, inode);
3543                 ext4_journal_test_restart(handle, inode);
3544                 if (bh) {
3545                         BUFFER_TRACE(bh, "retaking write access");
3546                         ext4_journal_get_write_access(handle, bh);
3547                 }
3548         }
3549
3550         /*
3551          * Any buffers which are on the journal will be in memory. We find
3552          * them on the hash table so jbd2_journal_revoke() will run jbd2_journal_forget()
3553          * on them.  We've already detached each block from the file, so
3554          * bforget() in jbd2_journal_forget() should be safe.
3555          *
3556          * AKPM: turn on bforget in jbd2_journal_forget()!!!
3557          */
3558         for (p = first; p < last; p++) {
3559                 u32 nr = le32_to_cpu(*p);
3560                 if (nr) {
3561                         struct buffer_head *tbh;
3562
3563                         *p = 0;
3564                         tbh = sb_find_get_block(inode->i_sb, nr);
3565                         ext4_forget(handle, 0, inode, tbh, nr);
3566                 }
3567         }
3568
3569         ext4_free_blocks(handle, inode, block_to_free, count, 0);
3570 }
3571
3572 /**
3573  * ext4_free_data - free a list of data blocks
3574  * @handle:     handle for this transaction
3575  * @inode:      inode we are dealing with
3576  * @this_bh:    indirect buffer_head which contains *@first and *@last
3577  * @first:      array of block numbers
3578  * @last:       points immediately past the end of array
3579  *
3580  * We are freeing all blocks refered from that array (numbers are stored as
3581  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
3582  *
3583  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
3584  * blocks are contiguous then releasing them at one time will only affect one
3585  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
3586  * actually use a lot of journal space.
3587  *
3588  * @this_bh will be %NULL if @first and @last point into the inode's direct
3589  * block pointers.
3590  */
3591 static void ext4_free_data(handle_t *handle, struct inode *inode,
3592                            struct buffer_head *this_bh,
3593                            __le32 *first, __le32 *last)
3594 {
3595         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
3596         unsigned long count = 0;            /* Number of blocks in the run */
3597         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
3598                                                corresponding to
3599                                                block_to_free */
3600         ext4_fsblk_t nr;                    /* Current block # */
3601         __le32 *p;                          /* Pointer into inode/ind
3602                                                for current block */
3603         int err;
3604
3605         if (this_bh) {                          /* For indirect block */
3606                 BUFFER_TRACE(this_bh, "get_write_access");
3607                 err = ext4_journal_get_write_access(handle, this_bh);
3608                 /* Important: if we can't update the indirect pointers
3609                  * to the blocks, we can't free them. */
3610                 if (err)
3611                         return;
3612         }
3613
3614         for (p = first; p < last; p++) {
3615                 nr = le32_to_cpu(*p);
3616                 if (nr) {
3617                         /* accumulate blocks to free if they're contiguous */
3618                         if (count == 0) {
3619                                 block_to_free = nr;
3620                                 block_to_free_p = p;
3621                                 count = 1;
3622                         } else if (nr == block_to_free + count) {
3623                                 count++;
3624                         } else {
3625                                 ext4_clear_blocks(handle, inode, this_bh,
3626                                                   block_to_free,
3627                                                   count, block_to_free_p, p);
3628                                 block_to_free = nr;
3629                                 block_to_free_p = p;
3630                                 count = 1;
3631                         }
3632                 }
3633         }
3634
3635         if (count > 0)
3636                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
3637                                   count, block_to_free_p, p);
3638
3639         if (this_bh) {
3640                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
3641
3642                 /*
3643                  * The buffer head should have an attached journal head at this
3644                  * point. However, if the data is corrupted and an indirect
3645                  * block pointed to itself, it would have been detached when
3646                  * the block was cleared. Check for this instead of OOPSing.
3647                  */
3648                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
3649                         ext4_handle_dirty_metadata(handle, inode, this_bh);
3650                 else
3651                         ext4_error(inode->i_sb, __func__,
3652                                    "circular indirect block detected, "
3653                                    "inode=%lu, block=%llu",
3654                                    inode->i_ino,
3655                                    (unsigned long long) this_bh->b_blocknr);
3656         }
3657 }
3658
3659 /**
3660  *      ext4_free_branches - free an array of branches
3661  *      @handle: JBD handle for this transaction
3662  *      @inode: inode we are dealing with
3663  *      @parent_bh: the buffer_head which contains *@first and *@last
3664  *      @first: array of block numbers
3665  *      @last:  pointer immediately past the end of array
3666  *      @depth: depth of the branches to free
3667  *
3668  *      We are freeing all blocks refered from these branches (numbers are
3669  *      stored as little-endian 32-bit) and updating @inode->i_blocks
3670  *      appropriately.
3671  */
3672 static void ext4_free_branches(handle_t *handle, struct inode *inode,
3673                                struct buffer_head *parent_bh,
3674                                __le32 *first, __le32 *last, int depth)
3675 {
3676         ext4_fsblk_t nr;
3677         __le32 *p;
3678
3679         if (ext4_handle_is_aborted(handle))
3680                 return;
3681
3682         if (depth--) {
3683                 struct buffer_head *bh;
3684                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
3685                 p = last;
3686                 while (--p >= first) {
3687                         nr = le32_to_cpu(*p);
3688                         if (!nr)
3689                                 continue;               /* A hole */
3690
3691                         /* Go read the buffer for the next level down */
3692                         bh = sb_bread(inode->i_sb, nr);
3693
3694                         /*
3695                          * A read failure? Report error and clear slot
3696                          * (should be rare).
3697                          */
3698                         if (!bh) {
3699                                 ext4_error(inode->i_sb, "ext4_free_branches",
3700                                            "Read failure, inode=%lu, block=%llu",
3701                                            inode->i_ino, nr);
3702                                 continue;
3703                         }
3704
3705                         /* This zaps the entire block.  Bottom up. */
3706                         BUFFER_TRACE(bh, "free child branches");
3707                         ext4_free_branches(handle, inode, bh,
3708                                         (__le32 *) bh->b_data,
3709                                         (__le32 *) bh->b_data + addr_per_block,
3710                                         depth);
3711
3712                         /*
3713                          * We've probably journalled the indirect block several
3714                          * times during the truncate.  But it's no longer
3715                          * needed and we now drop it from the transaction via
3716                          * jbd2_journal_revoke().
3717                          *
3718                          * That's easy if it's exclusively part of this
3719                          * transaction.  But if it's part of the committing
3720                          * transaction then jbd2_journal_forget() will simply
3721                          * brelse() it.  That means that if the underlying
3722                          * block is reallocated in ext4_get_block(),
3723                          * unmap_underlying_metadata() will find this block
3724                          * and will try to get rid of it.  damn, damn.
3725                          *
3726                          * If this block has already been committed to the
3727                          * journal, a revoke record will be written.  And
3728                          * revoke records must be emitted *before* clearing
3729                          * this block's bit in the bitmaps.
3730                          */
3731                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
3732
3733                         /*
3734                          * Everything below this this pointer has been
3735                          * released.  Now let this top-of-subtree go.
3736                          *
3737                          * We want the freeing of this indirect block to be
3738                          * atomic in the journal with the updating of the
3739                          * bitmap block which owns it.  So make some room in
3740                          * the journal.
3741                          *
3742                          * We zero the parent pointer *after* freeing its
3743                          * pointee in the bitmaps, so if extend_transaction()
3744                          * for some reason fails to put the bitmap changes and
3745                          * the release into the same transaction, recovery
3746                          * will merely complain about releasing a free block,
3747                          * rather than leaking blocks.
3748                          */
3749                         if (ext4_handle_is_aborted(handle))
3750                                 return;
3751                         if (try_to_extend_transaction(handle, inode)) {
3752                                 ext4_mark_inode_dirty(handle, inode);
3753                                 ext4_journal_test_restart(handle, inode);
3754                         }
3755
3756                         ext4_free_blocks(handle, inode, nr, 1, 1);
3757
3758                         if (parent_bh) {
3759                                 /*
3760                                  * The block which we have just freed is
3761                                  * pointed to by an indirect block: journal it
3762                                  */
3763                                 BUFFER_TRACE(parent_bh, "get_write_access");
3764                                 if (!ext4_journal_get_write_access(handle,
3765                                                                    parent_bh)){
3766                                         *p = 0;
3767                                         BUFFER_TRACE(parent_bh,
3768                                         "call ext4_handle_dirty_metadata");
3769                                         ext4_handle_dirty_metadata(handle,
3770                                                                    inode,
3771                                                                    parent_bh);
3772                                 }
3773                         }
3774                 }
3775         } else {
3776                 /* We have reached the bottom of the tree. */
3777                 BUFFER_TRACE(parent_bh, "free data blocks");
3778                 ext4_free_data(handle, inode, parent_bh, first, last);
3779         }
3780 }
3781
3782 int ext4_can_truncate(struct inode *inode)
3783 {
3784         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
3785                 return 0;
3786         if (S_ISREG(inode->i_mode))
3787                 return 1;
3788         if (S_ISDIR(inode->i_mode))
3789                 return 1;
3790         if (S_ISLNK(inode->i_mode))
3791                 return !ext4_inode_is_fast_symlink(inode);
3792         return 0;
3793 }
3794
3795 /*
3796  * ext4_truncate()
3797  *
3798  * We block out ext4_get_block() block instantiations across the entire
3799  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
3800  * simultaneously on behalf of the same inode.
3801  *
3802  * As we work through the truncate and commmit bits of it to the journal there
3803  * is one core, guiding principle: the file's tree must always be consistent on
3804  * disk.  We must be able to restart the truncate after a crash.
3805  *
3806  * The file's tree may be transiently inconsistent in memory (although it
3807  * probably isn't), but whenever we close off and commit a journal transaction,
3808  * the contents of (the filesystem + the journal) must be consistent and
3809  * restartable.  It's pretty simple, really: bottom up, right to left (although
3810  * left-to-right works OK too).
3811  *
3812  * Note that at recovery time, journal replay occurs *before* the restart of
3813  * truncate against the orphan inode list.
3814  *
3815  * The committed inode has the new, desired i_size (which is the same as
3816  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
3817  * that this inode's truncate did not complete and it will again call
3818  * ext4_truncate() to have another go.  So there will be instantiated blocks
3819  * to the right of the truncation point in a crashed ext4 filesystem.  But
3820  * that's fine - as long as they are linked from the inode, the post-crash
3821  * ext4_truncate() run will find them and release them.
3822  */
3823 void ext4_truncate(struct inode *inode)
3824 {
3825         handle_t *handle;
3826         struct ext4_inode_info *ei = EXT4_I(inode);
3827         __le32 *i_data = ei->i_data;
3828         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
3829         struct address_space *mapping = inode->i_mapping;
3830         ext4_lblk_t offsets[4];
3831         Indirect chain[4];
3832         Indirect *partial;
3833         __le32 nr = 0;
3834         int n;
3835         ext4_lblk_t last_block;
3836         unsigned blocksize = inode->i_sb->s_blocksize;
3837
3838         if (!ext4_can_truncate(inode))
3839                 return;
3840
3841         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
3842                 ext4_ext_truncate(inode);
3843                 return;
3844         }
3845
3846         handle = start_transaction(inode);
3847         if (IS_ERR(handle))
3848                 return;         /* AKPM: return what? */
3849
3850         last_block = (inode->i_size + blocksize-1)
3851                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
3852
3853         if (inode->i_size & (blocksize - 1))
3854                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
3855                         goto out_stop;
3856
3857         n = ext4_block_to_path(inode, last_block, offsets, NULL);
3858         if (n == 0)
3859                 goto out_stop;  /* error */
3860
3861         /*
3862          * OK.  This truncate is going to happen.  We add the inode to the
3863          * orphan list, so that if this truncate spans multiple transactions,
3864          * and we crash, we will resume the truncate when the filesystem
3865          * recovers.  It also marks the inode dirty, to catch the new size.
3866          *
3867          * Implication: the file must always be in a sane, consistent
3868          * truncatable state while each transaction commits.
3869          */
3870         if (ext4_orphan_add(handle, inode))
3871                 goto out_stop;
3872
3873         /*
3874          * From here we block out all ext4_get_block() callers who want to
3875          * modify the block allocation tree.
3876          */
3877         down_write(&ei->i_data_sem);
3878
3879         ext4_discard_preallocations(inode);
3880
3881         /*
3882          * The orphan list entry will now protect us from any crash which
3883          * occurs before the truncate completes, so it is now safe to propagate
3884          * the new, shorter inode size (held for now in i_size) into the
3885          * on-disk inode. We do this via i_disksize, which is the value which
3886          * ext4 *really* writes onto the disk inode.
3887          */
3888         ei->i_disksize = inode->i_size;
3889
3890         if (n == 1) {           /* direct blocks */
3891                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
3892                                i_data + EXT4_NDIR_BLOCKS);
3893                 goto do_indirects;
3894         }
3895
3896         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
3897         /* Kill the top of shared branch (not detached) */
3898         if (nr) {
3899                 if (partial == chain) {
3900                         /* Shared branch grows from the inode */
3901                         ext4_free_branches(handle, inode, NULL,
3902                                            &nr, &nr+1, (chain+n-1) - partial);
3903                         *partial->p = 0;
3904                         /*
3905                          * We mark the inode dirty prior to restart,
3906                          * and prior to stop.  No need for it here.
3907                          */
3908                 } else {
3909                         /* Shared branch grows from an indirect block */
3910                         BUFFER_TRACE(partial->bh, "get_write_access");
3911                         ext4_free_branches(handle, inode, partial->bh,
3912                                         partial->p,
3913                                         partial->p+1, (chain+n-1) - partial);
3914                 }
3915         }
3916         /* Clear the ends of indirect blocks on the shared branch */
3917         while (partial > chain) {
3918                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
3919                                    (__le32*)partial->bh->b_data+addr_per_block,
3920                                    (chain+n-1) - partial);
3921                 BUFFER_TRACE(partial->bh, "call brelse");
3922                 brelse (partial->bh);
3923                 partial--;
3924         }
3925 do_indirects:
3926         /* Kill the remaining (whole) subtrees */
3927         switch (offsets[0]) {
3928         default:
3929                 nr = i_data[EXT4_IND_BLOCK];
3930                 if (nr) {
3931                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
3932                         i_data[EXT4_IND_BLOCK] = 0;
3933                 }
3934         case EXT4_IND_BLOCK:
3935                 nr = i_data[EXT4_DIND_BLOCK];
3936                 if (nr) {
3937                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
3938                         i_data[EXT4_DIND_BLOCK] = 0;
3939                 }
3940         case EXT4_DIND_BLOCK:
3941                 nr = i_data[EXT4_TIND_BLOCK];
3942                 if (nr) {
3943                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
3944                         i_data[EXT4_TIND_BLOCK] = 0;
3945                 }
3946         case EXT4_TIND_BLOCK:
3947                 ;
3948         }
3949
3950         up_write(&ei->i_data_sem);
3951         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
3952         ext4_mark_inode_dirty(handle, inode);
3953
3954         /*
3955          * In a multi-transaction truncate, we only make the final transaction
3956          * synchronous
3957          */
3958         if (IS_SYNC(inode))
3959                 ext4_handle_sync(handle);
3960 out_stop:
3961         /*
3962          * If this was a simple ftruncate(), and the file will remain alive
3963          * then we need to clear up the orphan record which we created above.
3964          * However, if this was a real unlink then we were called by
3965          * ext4_delete_inode(), and we allow that function to clean up the
3966          * orphan info for us.
3967          */
3968         if (inode->i_nlink)
3969                 ext4_orphan_del(handle, inode);
3970
3971         ext4_journal_stop(handle);
3972 }
3973
3974 /*
3975  * ext4_get_inode_loc returns with an extra refcount against the inode's
3976  * underlying buffer_head on success. If 'in_mem' is true, we have all
3977  * data in memory that is needed to recreate the on-disk version of this
3978  * inode.
3979  */
3980 static int __ext4_get_inode_loc(struct inode *inode,
3981                                 struct ext4_iloc *iloc, int in_mem)
3982 {
3983         struct ext4_group_desc  *gdp;
3984         struct buffer_head      *bh;
3985         struct super_block      *sb = inode->i_sb;
3986         ext4_fsblk_t            block;
3987         int                     inodes_per_block, inode_offset;
3988
3989         iloc->bh = NULL;
3990         if (!ext4_valid_inum(sb, inode->i_ino))
3991                 return -EIO;
3992
3993         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
3994         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
3995         if (!gdp)
3996                 return -EIO;
3997
3998         /*
3999          * Figure out the offset within the block group inode table
4000          */
4001         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4002         inode_offset = ((inode->i_ino - 1) %
4003                         EXT4_INODES_PER_GROUP(sb));
4004         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4005         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4006
4007         bh = sb_getblk(sb, block);
4008         if (!bh) {
4009                 ext4_error(sb, "ext4_get_inode_loc", "unable to read "
4010                            "inode block - inode=%lu, block=%llu",
4011                            inode->i_ino, block);
4012                 return -EIO;
4013         }
4014         if (!buffer_uptodate(bh)) {
4015                 lock_buffer(bh);
4016
4017                 /*
4018                  * If the buffer has the write error flag, we have failed
4019                  * to write out another inode in the same block.  In this
4020                  * case, we don't have to read the block because we may
4021                  * read the old inode data successfully.
4022                  */
4023                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4024                         set_buffer_uptodate(bh);
4025
4026                 if (buffer_uptodate(bh)) {
4027                         /* someone brought it uptodate while we waited */
4028                         unlock_buffer(bh);
4029                         goto has_buffer;
4030                 }
4031
4032                 /*
4033                  * If we have all information of the inode in memory and this
4034                  * is the only valid inode in the block, we need not read the
4035                  * block.
4036                  */
4037                 if (in_mem) {
4038                         struct buffer_head *bitmap_bh;
4039                         int i, start;
4040
4041                         start = inode_offset & ~(inodes_per_block - 1);
4042
4043                         /* Is the inode bitmap in cache? */
4044                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4045                         if (!bitmap_bh)
4046                                 goto make_io;
4047
4048                         /*
4049                          * If the inode bitmap isn't in cache then the
4050                          * optimisation may end up performing two reads instead
4051                          * of one, so skip it.
4052                          */
4053                         if (!buffer_uptodate(bitmap_bh)) {
4054                                 brelse(bitmap_bh);
4055                                 goto make_io;
4056                         }
4057                         for (i = start; i < start + inodes_per_block; i++) {
4058                                 if (i == inode_offset)
4059                                         continue;
4060                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4061                                         break;
4062                         }
4063                         brelse(bitmap_bh);
4064                         if (i == start + inodes_per_block) {
4065                                 /* all other inodes are free, so skip I/O */
4066                                 memset(bh->b_data, 0, bh->b_size);
4067                                 set_buffer_uptodate(bh);
4068                                 unlock_buffer(bh);
4069                                 goto has_buffer;
4070                         }
4071                 }
4072
4073 make_io:
4074                 /*
4075                  * If we need to do any I/O, try to pre-readahead extra
4076                  * blocks from the inode table.
4077                  */
4078                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4079                         ext4_fsblk_t b, end, table;
4080                         unsigned num;
4081
4082                         table = ext4_inode_table(sb, gdp);
4083                         /* Make sure s_inode_readahead_blks is a power of 2 */
4084                         while (EXT4_SB(sb)->s_inode_readahead_blks &
4085                                (EXT4_SB(sb)->s_inode_readahead_blks-1))
4086                                 EXT4_SB(sb)->s_inode_readahead_blks = 
4087                                    (EXT4_SB(sb)->s_inode_readahead_blks &
4088                                     (EXT4_SB(sb)->s_inode_readahead_blks-1));
4089                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4090                         if (table > b)
4091                                 b = table;
4092                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4093                         num = EXT4_INODES_PER_GROUP(sb);
4094                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4095                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4096                                 num -= ext4_itable_unused_count(sb, gdp);
4097                         table += num / inodes_per_block;
4098                         if (end > table)
4099                                 end = table;
4100                         while (b <= end)
4101                                 sb_breadahead(sb, b++);
4102                 }
4103
4104                 /*
4105                  * There are other valid inodes in the buffer, this inode
4106                  * has in-inode xattrs, or we don't have this inode in memory.
4107                  * Read the block from disk.
4108                  */
4109                 get_bh(bh);
4110                 bh->b_end_io = end_buffer_read_sync;
4111                 submit_bh(READ_META, bh);
4112                 wait_on_buffer(bh);
4113                 if (!buffer_uptodate(bh)) {
4114                         ext4_error(sb, __func__,
4115                                    "unable to read inode block - inode=%lu, "
4116                                    "block=%llu", inode->i_ino, block);
4117                         brelse(bh);
4118                         return -EIO;
4119                 }
4120         }
4121 has_buffer:
4122         iloc->bh = bh;
4123         return 0;
4124 }
4125
4126 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4127 {
4128         /* We have all inode data except xattrs in memory here. */
4129         return __ext4_get_inode_loc(inode, iloc,
4130                 !(EXT4_I(inode)->i_state & EXT4_STATE_XATTR));
4131 }
4132
4133 void ext4_set_inode_flags(struct inode *inode)
4134 {
4135         unsigned int flags = EXT4_I(inode)->i_flags;
4136
4137         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4138         if (flags & EXT4_SYNC_FL)
4139                 inode->i_flags |= S_SYNC;
4140         if (flags & EXT4_APPEND_FL)
4141                 inode->i_flags |= S_APPEND;
4142         if (flags & EXT4_IMMUTABLE_FL)
4143                 inode->i_flags |= S_IMMUTABLE;
4144         if (flags & EXT4_NOATIME_FL)
4145                 inode->i_flags |= S_NOATIME;
4146         if (flags & EXT4_DIRSYNC_FL)
4147                 inode->i_flags |= S_DIRSYNC;
4148 }
4149
4150 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4151 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4152 {
4153         unsigned int flags = ei->vfs_inode.i_flags;
4154
4155         ei->i_flags &= ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4156                         EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|EXT4_DIRSYNC_FL);
4157         if (flags & S_SYNC)
4158                 ei->i_flags |= EXT4_SYNC_FL;
4159         if (flags & S_APPEND)
4160                 ei->i_flags |= EXT4_APPEND_FL;
4161         if (flags & S_IMMUTABLE)
4162                 ei->i_flags |= EXT4_IMMUTABLE_FL;
4163         if (flags & S_NOATIME)
4164                 ei->i_flags |= EXT4_NOATIME_FL;
4165         if (flags & S_DIRSYNC)
4166                 ei->i_flags |= EXT4_DIRSYNC_FL;
4167 }
4168 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4169                                         struct ext4_inode_info *ei)
4170 {
4171         blkcnt_t i_blocks ;
4172         struct inode *inode = &(ei->vfs_inode);
4173         struct super_block *sb = inode->i_sb;
4174
4175         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4176                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4177                 /* we are using combined 48 bit field */
4178                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4179                                         le32_to_cpu(raw_inode->i_blocks_lo);
4180                 if (ei->i_flags & EXT4_HUGE_FILE_FL) {
4181                         /* i_blocks represent file system block size */
4182                         return i_blocks  << (inode->i_blkbits - 9);
4183                 } else {
4184                         return i_blocks;
4185                 }
4186         } else {
4187                 return le32_to_cpu(raw_inode->i_blocks_lo);
4188         }
4189 }
4190
4191 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4192 {
4193         struct ext4_iloc iloc;
4194         struct ext4_inode *raw_inode;
4195         struct ext4_inode_info *ei;
4196         struct buffer_head *bh;
4197         struct inode *inode;
4198         long ret;
4199         int block;
4200
4201         inode = iget_locked(sb, ino);
4202         if (!inode)
4203                 return ERR_PTR(-ENOMEM);
4204         if (!(inode->i_state & I_NEW))
4205                 return inode;
4206
4207         ei = EXT4_I(inode);
4208 #ifdef CONFIG_EXT4_FS_POSIX_ACL
4209         ei->i_acl = EXT4_ACL_NOT_CACHED;
4210         ei->i_default_acl = EXT4_ACL_NOT_CACHED;
4211 #endif
4212
4213         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4214         if (ret < 0)
4215                 goto bad_inode;
4216         bh = iloc.bh;
4217         raw_inode = ext4_raw_inode(&iloc);
4218         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4219         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4220         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4221         if (!(test_opt(inode->i_sb, NO_UID32))) {
4222                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4223                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4224         }
4225         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
4226
4227         ei->i_state = 0;
4228         ei->i_dir_start_lookup = 0;
4229         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4230         /* We now have enough fields to check if the inode was active or not.
4231          * This is needed because nfsd might try to access dead inodes
4232          * the test is that same one that e2fsck uses
4233          * NeilBrown 1999oct15
4234          */
4235         if (inode->i_nlink == 0) {
4236                 if (inode->i_mode == 0 ||
4237                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
4238                         /* this inode is deleted */
4239                         brelse(bh);
4240                         ret = -ESTALE;
4241                         goto bad_inode;
4242                 }
4243                 /* The only unlinked inodes we let through here have
4244                  * valid i_mode and are being read by the orphan
4245                  * recovery code: that's fine, we're about to complete
4246                  * the process of deleting those. */
4247         }
4248         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4249         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4250         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4251         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
4252             cpu_to_le32(EXT4_OS_HURD)) {
4253                 ei->i_file_acl |=
4254                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4255         }
4256         inode->i_size = ext4_isize(raw_inode);
4257         ei->i_disksize = inode->i_size;
4258         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4259         ei->i_block_group = iloc.block_group;
4260         /*
4261          * NOTE! The in-memory inode i_data array is in little-endian order
4262          * even on big-endian machines: we do NOT byteswap the block numbers!
4263          */
4264         for (block = 0; block < EXT4_N_BLOCKS; block++)
4265                 ei->i_data[block] = raw_inode->i_block[block];
4266         INIT_LIST_HEAD(&ei->i_orphan);
4267
4268         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4269                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4270                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4271                     EXT4_INODE_SIZE(inode->i_sb)) {
4272                         brelse(bh);
4273                         ret = -EIO;
4274                         goto bad_inode;
4275                 }
4276                 if (ei->i_extra_isize == 0) {
4277                         /* The extra space is currently unused. Use it. */
4278                         ei->i_extra_isize = sizeof(struct ext4_inode) -
4279                                             EXT4_GOOD_OLD_INODE_SIZE;
4280                 } else {
4281                         __le32 *magic = (void *)raw_inode +
4282                                         EXT4_GOOD_OLD_INODE_SIZE +
4283                                         ei->i_extra_isize;
4284                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
4285                                  ei->i_state |= EXT4_STATE_XATTR;
4286                 }
4287         } else
4288                 ei->i_extra_isize = 0;
4289
4290         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4291         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4292         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4293         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4294
4295         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4296         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4297                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4298                         inode->i_version |=
4299                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4300         }
4301
4302         if (S_ISREG(inode->i_mode)) {
4303                 inode->i_op = &ext4_file_inode_operations;
4304                 inode->i_fop = &ext4_file_operations;
4305                 ext4_set_aops(inode);
4306         } else if (S_ISDIR(inode->i_mode)) {
4307                 inode->i_op = &ext4_dir_inode_operations;
4308                 inode->i_fop = &ext4_dir_operations;
4309         } else if (S_ISLNK(inode->i_mode)) {
4310                 if (ext4_inode_is_fast_symlink(inode)) {
4311                         inode->i_op = &ext4_fast_symlink_inode_operations;
4312                         nd_terminate_link(ei->i_data, inode->i_size,
4313                                 sizeof(ei->i_data) - 1);
4314                 } else {
4315                         inode->i_op = &ext4_symlink_inode_operations;
4316                         ext4_set_aops(inode);
4317                 }
4318         } else {
4319                 inode->i_op = &ext4_special_inode_operations;
4320                 if (raw_inode->i_block[0])
4321                         init_special_inode(inode, inode->i_mode,
4322                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4323                 else
4324                         init_special_inode(inode, inode->i_mode,
4325                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4326         }
4327         brelse(iloc.bh);
4328         ext4_set_inode_flags(inode);
4329         unlock_new_inode(inode);
4330         return inode;
4331
4332 bad_inode:
4333         iget_failed(inode);
4334         return ERR_PTR(ret);
4335 }
4336
4337 static int ext4_inode_blocks_set(handle_t *handle,
4338                                 struct ext4_inode *raw_inode,
4339                                 struct ext4_inode_info *ei)
4340 {
4341         struct inode *inode = &(ei->vfs_inode);
4342         u64 i_blocks = inode->i_blocks;
4343         struct super_block *sb = inode->i_sb;
4344
4345         if (i_blocks <= ~0U) {
4346                 /*
4347                  * i_blocks can be represnted in a 32 bit variable
4348                  * as multiple of 512 bytes
4349                  */
4350                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4351                 raw_inode->i_blocks_high = 0;
4352                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4353                 return 0;
4354         }
4355         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
4356                 return -EFBIG;
4357
4358         if (i_blocks <= 0xffffffffffffULL) {
4359                 /*
4360                  * i_blocks can be represented in a 48 bit variable
4361                  * as multiple of 512 bytes
4362                  */
4363                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4364                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4365                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4366         } else {
4367                 ei->i_flags |= EXT4_HUGE_FILE_FL;
4368                 /* i_block is stored in file system block size */
4369                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4370                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4371                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4372         }
4373         return 0;
4374 }
4375
4376 /*
4377  * Post the struct inode info into an on-disk inode location in the
4378  * buffer-cache.  This gobbles the caller's reference to the
4379  * buffer_head in the inode location struct.
4380  *
4381  * The caller must have write access to iloc->bh.
4382  */
4383 static int ext4_do_update_inode(handle_t *handle,
4384                                 struct inode *inode,
4385                                 struct ext4_iloc *iloc)
4386 {
4387         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
4388         struct ext4_inode_info *ei = EXT4_I(inode);
4389         struct buffer_head *bh = iloc->bh;
4390         int err = 0, rc, block;
4391
4392         /* For fields not not tracking in the in-memory inode,
4393          * initialise them to zero for new inodes. */
4394         if (ei->i_state & EXT4_STATE_NEW)
4395                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
4396
4397         ext4_get_inode_flags(ei);
4398         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
4399         if (!(test_opt(inode->i_sb, NO_UID32))) {
4400                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
4401                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
4402 /*
4403  * Fix up interoperability with old kernels. Otherwise, old inodes get
4404  * re-used with the upper 16 bits of the uid/gid intact
4405  */
4406                 if (!ei->i_dtime) {
4407                         raw_inode->i_uid_high =
4408                                 cpu_to_le16(high_16_bits(inode->i_uid));
4409                         raw_inode->i_gid_high =
4410                                 cpu_to_le16(high_16_bits(inode->i_gid));
4411                 } else {
4412                         raw_inode->i_uid_high = 0;
4413                         raw_inode->i_gid_high = 0;
4414                 }
4415         } else {
4416                 raw_inode->i_uid_low =
4417                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
4418                 raw_inode->i_gid_low =
4419                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
4420                 raw_inode->i_uid_high = 0;
4421                 raw_inode->i_gid_high = 0;
4422         }
4423         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
4424
4425         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
4426         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
4427         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
4428         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
4429
4430         if (ext4_inode_blocks_set(handle, raw_inode, ei))
4431                 goto out_brelse;
4432         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
4433         /* clear the migrate flag in the raw_inode */
4434         raw_inode->i_flags = cpu_to_le32(ei->i_flags & ~EXT4_EXT_MIGRATE);
4435         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
4436             cpu_to_le32(EXT4_OS_HURD))
4437                 raw_inode->i_file_acl_high =
4438                         cpu_to_le16(ei->i_file_acl >> 32);
4439         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
4440         ext4_isize_set(raw_inode, ei->i_disksize);
4441         if (ei->i_disksize > 0x7fffffffULL) {
4442                 struct super_block *sb = inode->i_sb;
4443                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
4444                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
4445                                 EXT4_SB(sb)->s_es->s_rev_level ==
4446                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
4447                         /* If this is the first large file
4448                          * created, add a flag to the superblock.
4449                          */
4450                         err = ext4_journal_get_write_access(handle,
4451                                         EXT4_SB(sb)->s_sbh);
4452                         if (err)
4453                                 goto out_brelse;
4454                         ext4_update_dynamic_rev(sb);
4455                         EXT4_SET_RO_COMPAT_FEATURE(sb,
4456                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
4457                         sb->s_dirt = 1;
4458                         ext4_handle_sync(handle);
4459                         err = ext4_handle_dirty_metadata(handle, inode,
4460                                         EXT4_SB(sb)->s_sbh);
4461                 }
4462         }
4463         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
4464         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
4465                 if (old_valid_dev(inode->i_rdev)) {
4466                         raw_inode->i_block[0] =
4467                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
4468                         raw_inode->i_block[1] = 0;
4469                 } else {
4470                         raw_inode->i_block[0] = 0;
4471                         raw_inode->i_block[1] =
4472                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
4473                         raw_inode->i_block[2] = 0;
4474                 }
4475         } else for (block = 0; block < EXT4_N_BLOCKS; block++)
4476                 raw_inode->i_block[block] = ei->i_data[block];
4477
4478         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
4479         if (ei->i_extra_isize) {
4480                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4481                         raw_inode->i_version_hi =
4482                         cpu_to_le32(inode->i_version >> 32);
4483                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
4484         }
4485
4486         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4487         rc = ext4_handle_dirty_metadata(handle, inode, bh);
4488         if (!err)
4489                 err = rc;
4490         ei->i_state &= ~EXT4_STATE_NEW;
4491
4492 out_brelse:
4493         brelse(bh);
4494         ext4_std_error(inode->i_sb, err);
4495         return err;
4496 }
4497
4498 /*
4499  * ext4_write_inode()
4500  *
4501  * We are called from a few places:
4502  *
4503  * - Within generic_file_write() for O_SYNC files.
4504  *   Here, there will be no transaction running. We wait for any running
4505  *   trasnaction to commit.
4506  *
4507  * - Within sys_sync(), kupdate and such.
4508  *   We wait on commit, if tol to.
4509  *
4510  * - Within prune_icache() (PF_MEMALLOC == true)
4511  *   Here we simply return.  We can't afford to block kswapd on the
4512  *   journal commit.
4513  *
4514  * In all cases it is actually safe for us to return without doing anything,
4515  * because the inode has been copied into a raw inode buffer in
4516  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
4517  * knfsd.
4518  *
4519  * Note that we are absolutely dependent upon all inode dirtiers doing the
4520  * right thing: they *must* call mark_inode_dirty() after dirtying info in
4521  * which we are interested.
4522  *
4523  * It would be a bug for them to not do this.  The code:
4524  *
4525  *      mark_inode_dirty(inode)
4526  *      stuff();
4527  *      inode->i_size = expr;
4528  *
4529  * is in error because a kswapd-driven write_inode() could occur while
4530  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
4531  * will no longer be on the superblock's dirty inode list.
4532  */
4533 int ext4_write_inode(struct inode *inode, int wait)
4534 {
4535         if (current->flags & PF_MEMALLOC)
4536                 return 0;
4537
4538         if (ext4_journal_current_handle()) {
4539                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
4540                 dump_stack();
4541                 return -EIO;
4542         }
4543
4544         if (!wait)
4545                 return 0;
4546
4547         return ext4_force_commit(inode->i_sb);
4548 }
4549
4550 int __ext4_write_dirty_metadata(struct inode *inode, struct buffer_head *bh)
4551 {
4552         int err = 0;
4553
4554         mark_buffer_dirty(bh);
4555         if (inode && inode_needs_sync(inode)) {
4556                 sync_dirty_buffer(bh);
4557                 if (buffer_req(bh) && !buffer_uptodate(bh)) {
4558                         ext4_error(inode->i_sb, __func__,
4559                                    "IO error syncing inode, "
4560                                    "inode=%lu, block=%llu",
4561                                    inode->i_ino,
4562                                    (unsigned long long)bh->b_blocknr);
4563                         err = -EIO;
4564                 }
4565         }
4566         return err;
4567 }
4568
4569 /*
4570  * ext4_setattr()
4571  *
4572  * Called from notify_change.
4573  *
4574  * We want to trap VFS attempts to truncate the file as soon as
4575  * possible.  In particular, we want to make sure that when the VFS
4576  * shrinks i_size, we put the inode on the orphan list and modify
4577  * i_disksize immediately, so that during the subsequent flushing of
4578  * dirty pages and freeing of disk blocks, we can guarantee that any
4579  * commit will leave the blocks being flushed in an unused state on
4580  * disk.  (On recovery, the inode will get truncated and the blocks will
4581  * be freed, so we have a strong guarantee that no future commit will
4582  * leave these blocks visible to the user.)
4583  *
4584  * Another thing we have to assure is that if we are in ordered mode
4585  * and inode is still attached to the committing transaction, we must
4586  * we start writeout of all the dirty pages which are being truncated.
4587  * This way we are sure that all the data written in the previous
4588  * transaction are already on disk (truncate waits for pages under
4589  * writeback).
4590  *
4591  * Called with inode->i_mutex down.
4592  */
4593 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
4594 {
4595         struct inode *inode = dentry->d_inode;
4596         int error, rc = 0;
4597         const unsigned int ia_valid = attr->ia_valid;
4598
4599         error = inode_change_ok(inode, attr);
4600         if (error)
4601                 return error;
4602
4603         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
4604                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
4605                 handle_t *handle;
4606
4607                 /* (user+group)*(old+new) structure, inode write (sb,
4608                  * inode block, ? - but truncate inode update has it) */
4609                 handle = ext4_journal_start(inode, 2*(EXT4_QUOTA_INIT_BLOCKS(inode->i_sb)+
4610                                         EXT4_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
4611                 if (IS_ERR(handle)) {
4612                         error = PTR_ERR(handle);
4613                         goto err_out;
4614                 }
4615                 error = DQUOT_TRANSFER(inode, attr) ? -EDQUOT : 0;
4616                 if (error) {
4617                         ext4_journal_stop(handle);
4618                         return error;
4619                 }
4620                 /* Update corresponding info in inode so that everything is in
4621                  * one transaction */
4622                 if (attr->ia_valid & ATTR_UID)
4623                         inode->i_uid = attr->ia_uid;
4624                 if (attr->ia_valid & ATTR_GID)
4625                         inode->i_gid = attr->ia_gid;
4626                 error = ext4_mark_inode_dirty(handle, inode);
4627                 ext4_journal_stop(handle);
4628         }
4629
4630         if (attr->ia_valid & ATTR_SIZE) {
4631                 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)) {
4632                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4633
4634                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
4635                                 error = -EFBIG;
4636                                 goto err_out;
4637                         }
4638                 }
4639         }
4640
4641         if (S_ISREG(inode->i_mode) &&
4642             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
4643                 handle_t *handle;
4644
4645                 handle = ext4_journal_start(inode, 3);
4646                 if (IS_ERR(handle)) {
4647                         error = PTR_ERR(handle);
4648                         goto err_out;
4649                 }
4650
4651                 error = ext4_orphan_add(handle, inode);
4652                 EXT4_I(inode)->i_disksize = attr->ia_size;
4653                 rc = ext4_mark_inode_dirty(handle, inode);
4654                 if (!error)
4655                         error = rc;
4656                 ext4_journal_stop(handle);
4657
4658                 if (ext4_should_order_data(inode)) {
4659                         error = ext4_begin_ordered_truncate(inode,
4660                                                             attr->ia_size);
4661                         if (error) {
4662                                 /* Do as much error cleanup as possible */
4663                                 handle = ext4_journal_start(inode, 3);
4664                                 if (IS_ERR(handle)) {
4665                                         ext4_orphan_del(NULL, inode);
4666                                         goto err_out;
4667                                 }
4668                                 ext4_orphan_del(handle, inode);
4669                                 ext4_journal_stop(handle);
4670                                 goto err_out;
4671                         }
4672                 }
4673         }
4674
4675         rc = inode_setattr(inode, attr);
4676
4677         /* If inode_setattr's call to ext4_truncate failed to get a
4678          * transaction handle at all, we need to clean up the in-core
4679          * orphan list manually. */
4680         if (inode->i_nlink)
4681                 ext4_orphan_del(NULL, inode);
4682
4683         if (!rc && (ia_valid & ATTR_MODE))
4684                 rc = ext4_acl_chmod(inode);
4685
4686 err_out:
4687         ext4_std_error(inode->i_sb, error);
4688         if (!error)
4689                 error = rc;
4690         return error;
4691 }
4692
4693 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
4694                  struct kstat *stat)
4695 {
4696         struct inode *inode;
4697         unsigned long delalloc_blocks;
4698
4699         inode = dentry->d_inode;
4700         generic_fillattr(inode, stat);
4701
4702         /*
4703          * We can't update i_blocks if the block allocation is delayed
4704          * otherwise in the case of system crash before the real block
4705          * allocation is done, we will have i_blocks inconsistent with
4706          * on-disk file blocks.
4707          * We always keep i_blocks updated together with real
4708          * allocation. But to not confuse with user, stat
4709          * will return the blocks that include the delayed allocation
4710          * blocks for this file.
4711          */
4712         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
4713         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
4714         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
4715
4716         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
4717         return 0;
4718 }
4719
4720 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
4721                                       int chunk)
4722 {
4723         int indirects;
4724
4725         /* if nrblocks are contiguous */
4726         if (chunk) {
4727                 /*
4728                  * With N contiguous data blocks, it need at most
4729                  * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) indirect blocks
4730                  * 2 dindirect blocks
4731                  * 1 tindirect block
4732                  */
4733                 indirects = nrblocks / EXT4_ADDR_PER_BLOCK(inode->i_sb);
4734                 return indirects + 3;
4735         }
4736         /*
4737          * if nrblocks are not contiguous, worse case, each block touch
4738          * a indirect block, and each indirect block touch a double indirect
4739          * block, plus a triple indirect block
4740          */
4741         indirects = nrblocks * 2 + 1;
4742         return indirects;
4743 }
4744
4745 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4746 {
4747         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL))
4748                 return ext4_indirect_trans_blocks(inode, nrblocks, chunk);
4749         return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
4750 }
4751
4752 /*
4753  * Account for index blocks, block groups bitmaps and block group
4754  * descriptor blocks if modify datablocks and index blocks
4755  * worse case, the indexs blocks spread over different block groups
4756  *
4757  * If datablocks are discontiguous, they are possible to spread over
4758  * different block groups too. If they are contiugous, with flexbg,
4759  * they could still across block group boundary.
4760  *
4761  * Also account for superblock, inode, quota and xattr blocks
4762  */
4763 int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4764 {
4765         int groups, gdpblocks;
4766         int idxblocks;
4767         int ret = 0;
4768
4769         /*
4770          * How many index blocks need to touch to modify nrblocks?
4771          * The "Chunk" flag indicating whether the nrblocks is
4772          * physically contiguous on disk
4773          *
4774          * For Direct IO and fallocate, they calls get_block to allocate
4775          * one single extent at a time, so they could set the "Chunk" flag
4776          */
4777         idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
4778
4779         ret = idxblocks;
4780
4781         /*
4782          * Now let's see how many group bitmaps and group descriptors need
4783          * to account
4784          */
4785         groups = idxblocks;
4786         if (chunk)
4787                 groups += 1;
4788         else
4789                 groups += nrblocks;
4790
4791         gdpblocks = groups;
4792         if (groups > EXT4_SB(inode->i_sb)->s_groups_count)
4793                 groups = EXT4_SB(inode->i_sb)->s_groups_count;
4794         if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
4795                 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
4796
4797         /* bitmaps and block group descriptor blocks */
4798         ret += groups + gdpblocks;
4799
4800         /* Blocks for super block, inode, quota and xattr blocks */
4801         ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
4802
4803         return ret;
4804 }
4805
4806 /*
4807  * Calulate the total number of credits to reserve to fit
4808  * the modification of a single pages into a single transaction,
4809  * which may include multiple chunks of block allocations.
4810  *
4811  * This could be called via ext4_write_begin()
4812  *
4813  * We need to consider the worse case, when
4814  * one new block per extent.
4815  */
4816 int ext4_writepage_trans_blocks(struct inode *inode)
4817 {
4818         int bpp = ext4_journal_blocks_per_page(inode);
4819         int ret;
4820
4821         ret = ext4_meta_trans_blocks(inode, bpp, 0);
4822
4823         /* Account for data blocks for journalled mode */
4824         if (ext4_should_journal_data(inode))
4825                 ret += bpp;
4826         return ret;
4827 }
4828
4829 /*
4830  * Calculate the journal credits for a chunk of data modification.
4831  *
4832  * This is called from DIO, fallocate or whoever calling
4833  * ext4_get_blocks_wrap() to map/allocate a chunk of contigous disk blocks.
4834  *
4835  * journal buffers for data blocks are not included here, as DIO
4836  * and fallocate do no need to journal data buffers.
4837  */
4838 int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
4839 {
4840         return ext4_meta_trans_blocks(inode, nrblocks, 1);
4841 }
4842
4843 /*
4844  * The caller must have previously called ext4_reserve_inode_write().
4845  * Give this, we know that the caller already has write access to iloc->bh.
4846  */
4847 int ext4_mark_iloc_dirty(handle_t *handle,
4848                 struct inode *inode, struct ext4_iloc *iloc)
4849 {
4850         int err = 0;
4851
4852         if (test_opt(inode->i_sb, I_VERSION))
4853                 inode_inc_iversion(inode);
4854
4855         /* the do_update_inode consumes one bh->b_count */
4856         get_bh(iloc->bh);
4857
4858         /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
4859         err = ext4_do_update_inode(handle, inode, iloc);
4860         put_bh(iloc->bh);
4861         return err;
4862 }
4863
4864 /*
4865  * On success, We end up with an outstanding reference count against
4866  * iloc->bh.  This _must_ be cleaned up later.
4867  */
4868
4869 int
4870 ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
4871                          struct ext4_iloc *iloc)
4872 {
4873         int err;
4874
4875         err = ext4_get_inode_loc(inode, iloc);
4876         if (!err) {
4877                 BUFFER_TRACE(iloc->bh, "get_write_access");
4878                 err = ext4_journal_get_write_access(handle, iloc->bh);
4879                 if (err) {
4880                         brelse(iloc->bh);
4881                         iloc->bh = NULL;
4882                 }
4883         }
4884         ext4_std_error(inode->i_sb, err);
4885         return err;
4886 }
4887
4888 /*
4889  * Expand an inode by new_extra_isize bytes.
4890  * Returns 0 on success or negative error number on failure.
4891  */
4892 static int ext4_expand_extra_isize(struct inode *inode,
4893                                    unsigned int new_extra_isize,
4894                                    struct ext4_iloc iloc,
4895                                    handle_t *handle)
4896 {
4897         struct ext4_inode *raw_inode;
4898         struct ext4_xattr_ibody_header *header;
4899         struct ext4_xattr_entry *entry;
4900
4901         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
4902                 return 0;
4903
4904         raw_inode = ext4_raw_inode(&iloc);
4905
4906         header = IHDR(inode, raw_inode);
4907         entry = IFIRST(header);
4908
4909         /* No extended attributes present */
4910         if (!(EXT4_I(inode)->i_state & EXT4_STATE_XATTR) ||
4911                 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
4912                 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
4913                         new_extra_isize);
4914                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
4915                 return 0;
4916         }
4917
4918         /* try to expand with EAs present */
4919         return ext4_expand_extra_isize_ea(inode, new_extra_isize,
4920                                           raw_inode, handle);
4921 }
4922
4923 /*
4924  * What we do here is to mark the in-core inode as clean with respect to inode
4925  * dirtiness (it may still be data-dirty).
4926  * This means that the in-core inode may be reaped by prune_icache
4927  * without having to perform any I/O.  This is a very good thing,
4928  * because *any* task may call prune_icache - even ones which
4929  * have a transaction open against a different journal.
4930  *
4931  * Is this cheating?  Not really.  Sure, we haven't written the
4932  * inode out, but prune_icache isn't a user-visible syncing function.
4933  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
4934  * we start and wait on commits.
4935  *
4936  * Is this efficient/effective?  Well, we're being nice to the system
4937  * by cleaning up our inodes proactively so they can be reaped
4938  * without I/O.  But we are potentially leaving up to five seconds'
4939  * worth of inodes floating about which prune_icache wants us to
4940  * write out.  One way to fix that would be to get prune_icache()
4941  * to do a write_super() to free up some memory.  It has the desired
4942  * effect.
4943  */
4944 int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
4945 {
4946         struct ext4_iloc iloc;
4947         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4948         static unsigned int mnt_count;
4949         int err, ret;
4950
4951         might_sleep();
4952         err = ext4_reserve_inode_write(handle, inode, &iloc);
4953         if (ext4_handle_valid(handle) &&
4954             EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
4955             !(EXT4_I(inode)->i_state & EXT4_STATE_NO_EXPAND)) {
4956                 /*
4957                  * We need extra buffer credits since we may write into EA block
4958                  * with this same handle. If journal_extend fails, then it will
4959                  * only result in a minor loss of functionality for that inode.
4960                  * If this is felt to be critical, then e2fsck should be run to
4961                  * force a large enough s_min_extra_isize.
4962                  */
4963                 if ((jbd2_journal_extend(handle,
4964                              EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
4965                         ret = ext4_expand_extra_isize(inode,
4966                                                       sbi->s_want_extra_isize,
4967                                                       iloc, handle);
4968                         if (ret) {
4969                                 EXT4_I(inode)->i_state |= EXT4_STATE_NO_EXPAND;
4970                                 if (mnt_count !=
4971                                         le16_to_cpu(sbi->s_es->s_mnt_count)) {
4972                                         ext4_warning(inode->i_sb, __func__,
4973                                         "Unable to expand inode %lu. Delete"
4974                                         " some EAs or run e2fsck.",
4975                                         inode->i_ino);
4976                                         mnt_count =
4977                                           le16_to_cpu(sbi->s_es->s_mnt_count);
4978                                 }
4979                         }
4980                 }
4981         }
4982         if (!err)
4983                 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
4984         return err;
4985 }
4986
4987 /*
4988  * ext4_dirty_inode() is called from __mark_inode_dirty()
4989  *
4990  * We're really interested in the case where a file is being extended.
4991  * i_size has been changed by generic_commit_write() and we thus need
4992  * to include the updated inode in the current transaction.
4993  *
4994  * Also, DQUOT_ALLOC_SPACE() will always dirty the inode when blocks
4995  * are allocated to the file.
4996  *
4997  * If the inode is marked synchronous, we don't honour that here - doing
4998  * so would cause a commit on atime updates, which we don't bother doing.
4999  * We handle synchronous inodes at the highest possible level.
5000  */
5001 void ext4_dirty_inode(struct inode *inode)
5002 {
5003         handle_t *current_handle = ext4_journal_current_handle();
5004         handle_t *handle;
5005
5006         if (!ext4_handle_valid(current_handle)) {
5007                 ext4_mark_inode_dirty(current_handle, inode);
5008                 return;
5009         }
5010
5011         handle = ext4_journal_start(inode, 2);
5012         if (IS_ERR(handle))
5013                 goto out;
5014         if (current_handle &&
5015                 current_handle->h_transaction != handle->h_transaction) {
5016                 /* This task has a transaction open against a different fs */
5017                 printk(KERN_EMERG "%s: transactions do not match!\n",
5018                        __func__);
5019         } else {
5020                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
5021                                 current_handle);
5022                 ext4_mark_inode_dirty(handle, inode);
5023         }
5024         ext4_journal_stop(handle);
5025 out:
5026         return;
5027 }
5028
5029 #if 0
5030 /*
5031  * Bind an inode's backing buffer_head into this transaction, to prevent
5032  * it from being flushed to disk early.  Unlike
5033  * ext4_reserve_inode_write, this leaves behind no bh reference and
5034  * returns no iloc structure, so the caller needs to repeat the iloc
5035  * lookup to mark the inode dirty later.
5036  */
5037 static int ext4_pin_inode(handle_t *handle, struct inode *inode)
5038 {
5039         struct ext4_iloc iloc;
5040
5041         int err = 0;
5042         if (handle) {
5043                 err = ext4_get_inode_loc(inode, &iloc);
5044                 if (!err) {
5045                         BUFFER_TRACE(iloc.bh, "get_write_access");
5046                         err = jbd2_journal_get_write_access(handle, iloc.bh);
5047                         if (!err)
5048                                 err = ext4_handle_dirty_metadata(handle,
5049                                                                  inode,
5050                                                                  iloc.bh);
5051                         brelse(iloc.bh);
5052                 }
5053         }
5054         ext4_std_error(inode->i_sb, err);
5055         return err;
5056 }
5057 #endif
5058
5059 int ext4_change_inode_journal_flag(struct inode *inode, int val)
5060 {
5061         journal_t *journal;
5062         handle_t *handle;
5063         int err;
5064
5065         /*
5066          * We have to be very careful here: changing a data block's
5067          * journaling status dynamically is dangerous.  If we write a
5068          * data block to the journal, change the status and then delete
5069          * that block, we risk forgetting to revoke the old log record
5070          * from the journal and so a subsequent replay can corrupt data.
5071          * So, first we make sure that the journal is empty and that
5072          * nobody is changing anything.
5073          */
5074
5075         journal = EXT4_JOURNAL(inode);
5076         if (!journal)
5077                 return 0;
5078         if (is_journal_aborted(journal))
5079                 return -EROFS;
5080
5081         jbd2_journal_lock_updates(journal);
5082         jbd2_journal_flush(journal);
5083
5084         /*
5085          * OK, there are no updates running now, and all cached data is
5086          * synced to disk.  We are now in a completely consistent state
5087          * which doesn't have anything in the journal, and we know that
5088          * no filesystem updates are running, so it is safe to modify
5089          * the inode's in-core data-journaling state flag now.
5090          */
5091
5092         if (val)
5093                 EXT4_I(inode)->i_flags |= EXT4_JOURNAL_DATA_FL;
5094         else
5095                 EXT4_I(inode)->i_flags &= ~EXT4_JOURNAL_DATA_FL;
5096         ext4_set_aops(inode);
5097
5098         jbd2_journal_unlock_updates(journal);
5099
5100         /* Finally we can mark the inode as dirty. */
5101
5102         handle = ext4_journal_start(inode, 1);
5103         if (IS_ERR(handle))
5104                 return PTR_ERR(handle);
5105
5106         err = ext4_mark_inode_dirty(handle, inode);
5107         ext4_handle_sync(handle);
5108         ext4_journal_stop(handle);
5109         ext4_std_error(inode->i_sb, err);
5110
5111         return err;
5112 }
5113
5114 static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
5115 {
5116         return !buffer_mapped(bh);
5117 }
5118
5119 int ext4_page_mkwrite(struct vm_area_struct *vma, struct page *page)
5120 {
5121         loff_t size;
5122         unsigned long len;
5123         int ret = -EINVAL;
5124         void *fsdata;
5125         struct file *file = vma->vm_file;
5126         struct inode *inode = file->f_path.dentry->d_inode;
5127         struct address_space *mapping = inode->i_mapping;
5128
5129         /*
5130          * Get i_alloc_sem to stop truncates messing with the inode. We cannot
5131          * get i_mutex because we are already holding mmap_sem.
5132          */
5133         down_read(&inode->i_alloc_sem);
5134         size = i_size_read(inode);
5135         if (page->mapping != mapping || size <= page_offset(page)
5136             || !PageUptodate(page)) {
5137                 /* page got truncated from under us? */
5138                 goto out_unlock;
5139         }
5140         ret = 0;
5141         if (PageMappedToDisk(page))
5142                 goto out_unlock;
5143
5144         if (page->index == size >> PAGE_CACHE_SHIFT)
5145                 len = size & ~PAGE_CACHE_MASK;
5146         else
5147                 len = PAGE_CACHE_SIZE;
5148
5149         if (page_has_buffers(page)) {
5150                 /* return if we have all the buffers mapped */
5151                 if (!walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
5152                                        ext4_bh_unmapped))
5153                         goto out_unlock;
5154         }
5155         /*
5156          * OK, we need to fill the hole... Do write_begin write_end
5157          * to do block allocation/reservation.We are not holding
5158          * inode.i__mutex here. That allow * parallel write_begin,
5159          * write_end call. lock_page prevent this from happening
5160          * on the same page though
5161          */
5162         ret = mapping->a_ops->write_begin(file, mapping, page_offset(page),
5163                         len, AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
5164         if (ret < 0)
5165                 goto out_unlock;
5166         ret = mapping->a_ops->write_end(file, mapping, page_offset(page),
5167                         len, len, page, fsdata);
5168         if (ret < 0)
5169                 goto out_unlock;
5170         ret = 0;
5171 out_unlock:
5172         up_read(&inode->i_alloc_sem);
5173         return ret;
5174 }