Merge branch 'topic/jack' into topic/docbook-fix
[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(&EXT4_I(inode)->jinode,
51                                                    new_size);
52 }
53
54 static void ext4_invalidatepage(struct page *page, unsigned long offset);
55
56 /*
57  * Test whether an inode is a fast symlink.
58  */
59 static int ext4_inode_is_fast_symlink(struct inode *inode)
60 {
61         int ea_blocks = EXT4_I(inode)->i_file_acl ?
62                 (inode->i_sb->s_blocksize >> 9) : 0;
63
64         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
65 }
66
67 /*
68  * The ext4 forget function must perform a revoke if we are freeing data
69  * which has been journaled.  Metadata (eg. indirect blocks) must be
70  * revoked in all cases.
71  *
72  * "bh" may be NULL: a metadata block may have been freed from memory
73  * but there may still be a record of it in the journal, and that record
74  * still needs to be revoked.
75  *
76  * If the handle isn't valid we're not journaling so there's nothing to do.
77  */
78 int ext4_forget(handle_t *handle, int is_metadata, struct inode *inode,
79                         struct buffer_head *bh, ext4_fsblk_t blocknr)
80 {
81         int err;
82
83         if (!ext4_handle_valid(handle))
84                 return 0;
85
86         might_sleep();
87
88         BUFFER_TRACE(bh, "enter");
89
90         jbd_debug(4, "forgetting bh %p: is_metadata = %d, mode %o, "
91                   "data mode %lx\n",
92                   bh, is_metadata, inode->i_mode,
93                   test_opt(inode->i_sb, DATA_FLAGS));
94
95         /* Never use the revoke function if we are doing full data
96          * journaling: there is no need to, and a V1 superblock won't
97          * support it.  Otherwise, only skip the revoke on un-journaled
98          * data blocks. */
99
100         if (test_opt(inode->i_sb, DATA_FLAGS) == EXT4_MOUNT_JOURNAL_DATA ||
101             (!is_metadata && !ext4_should_journal_data(inode))) {
102                 if (bh) {
103                         BUFFER_TRACE(bh, "call jbd2_journal_forget");
104                         return ext4_journal_forget(handle, bh);
105                 }
106                 return 0;
107         }
108
109         /*
110          * data!=journal && (is_metadata || should_journal_data(inode))
111          */
112         BUFFER_TRACE(bh, "call ext4_journal_revoke");
113         err = ext4_journal_revoke(handle, blocknr, bh);
114         if (err)
115                 ext4_abort(inode->i_sb, __func__,
116                            "error %d when attempting revoke", err);
117         BUFFER_TRACE(bh, "exit");
118         return err;
119 }
120
121 /*
122  * Work out how many blocks we need to proceed with the next chunk of a
123  * truncate transaction.
124  */
125 static unsigned long blocks_for_truncate(struct inode *inode)
126 {
127         ext4_lblk_t needed;
128
129         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
130
131         /* Give ourselves just enough room to cope with inodes in which
132          * i_blocks is corrupt: we've seen disk corruptions in the past
133          * which resulted in random data in an inode which looked enough
134          * like a regular file for ext4 to try to delete it.  Things
135          * will go a bit crazy if that happens, but at least we should
136          * try not to panic the whole kernel. */
137         if (needed < 2)
138                 needed = 2;
139
140         /* But we need to bound the transaction so we don't overflow the
141          * journal. */
142         if (needed > EXT4_MAX_TRANS_DATA)
143                 needed = EXT4_MAX_TRANS_DATA;
144
145         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
146 }
147
148 /*
149  * Truncate transactions can be complex and absolutely huge.  So we need to
150  * be able to restart the transaction at a conventient checkpoint to make
151  * sure we don't overflow the journal.
152  *
153  * start_transaction gets us a new handle for a truncate transaction,
154  * and extend_transaction tries to extend the existing one a bit.  If
155  * extend fails, we need to propagate the failure up and restart the
156  * transaction in the top-level truncate loop. --sct
157  */
158 static handle_t *start_transaction(struct inode *inode)
159 {
160         handle_t *result;
161
162         result = ext4_journal_start(inode, blocks_for_truncate(inode));
163         if (!IS_ERR(result))
164                 return result;
165
166         ext4_std_error(inode->i_sb, PTR_ERR(result));
167         return result;
168 }
169
170 /*
171  * Try to extend this transaction for the purposes of truncation.
172  *
173  * Returns 0 if we managed to create more room.  If we can't create more
174  * room, and the transaction must be restarted we return 1.
175  */
176 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
177 {
178         if (!ext4_handle_valid(handle))
179                 return 0;
180         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
181                 return 0;
182         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
183                 return 0;
184         return 1;
185 }
186
187 /*
188  * Restart the transaction associated with *handle.  This does a commit,
189  * so before we call here everything must be consistently dirtied against
190  * this transaction.
191  */
192 static int ext4_journal_test_restart(handle_t *handle, struct inode *inode)
193 {
194         BUG_ON(EXT4_JOURNAL(inode) == NULL);
195         jbd_debug(2, "restarting handle %p\n", handle);
196         return ext4_journal_restart(handle, blocks_for_truncate(inode));
197 }
198
199 /*
200  * Called at the last iput() if i_nlink is zero.
201  */
202 void ext4_delete_inode(struct inode *inode)
203 {
204         handle_t *handle;
205         int err;
206
207         if (ext4_should_order_data(inode))
208                 ext4_begin_ordered_truncate(inode, 0);
209         truncate_inode_pages(&inode->i_data, 0);
210
211         if (is_bad_inode(inode))
212                 goto no_delete;
213
214         handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
215         if (IS_ERR(handle)) {
216                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
217                 /*
218                  * If we're going to skip the normal cleanup, we still need to
219                  * make sure that the in-core orphan linked list is properly
220                  * cleaned up.
221                  */
222                 ext4_orphan_del(NULL, inode);
223                 goto no_delete;
224         }
225
226         if (IS_SYNC(inode))
227                 ext4_handle_sync(handle);
228         inode->i_size = 0;
229         err = ext4_mark_inode_dirty(handle, inode);
230         if (err) {
231                 ext4_warning(inode->i_sb, __func__,
232                              "couldn't mark inode dirty (err %d)", err);
233                 goto stop_handle;
234         }
235         if (inode->i_blocks)
236                 ext4_truncate(inode);
237
238         /*
239          * ext4_ext_truncate() doesn't reserve any slop when it
240          * restarts journal transactions; therefore there may not be
241          * enough credits left in the handle to remove the inode from
242          * the orphan list and set the dtime field.
243          */
244         if (!ext4_handle_has_enough_credits(handle, 3)) {
245                 err = ext4_journal_extend(handle, 3);
246                 if (err > 0)
247                         err = ext4_journal_restart(handle, 3);
248                 if (err != 0) {
249                         ext4_warning(inode->i_sb, __func__,
250                                      "couldn't extend journal (err %d)", err);
251                 stop_handle:
252                         ext4_journal_stop(handle);
253                         goto no_delete;
254                 }
255         }
256
257         /*
258          * Kill off the orphan record which ext4_truncate created.
259          * AKPM: I think this can be inside the above `if'.
260          * Note that ext4_orphan_del() has to be able to cope with the
261          * deletion of a non-existent orphan - this is because we don't
262          * know if ext4_truncate() actually created an orphan record.
263          * (Well, we could do this if we need to, but heck - it works)
264          */
265         ext4_orphan_del(handle, inode);
266         EXT4_I(inode)->i_dtime  = get_seconds();
267
268         /*
269          * One subtle ordering requirement: if anything has gone wrong
270          * (transaction abort, IO errors, whatever), then we can still
271          * do these next steps (the fs will already have been marked as
272          * having errors), but we can't free the inode if the mark_dirty
273          * fails.
274          */
275         if (ext4_mark_inode_dirty(handle, inode))
276                 /* If that failed, just do the required in-core inode clear. */
277                 clear_inode(inode);
278         else
279                 ext4_free_inode(handle, inode);
280         ext4_journal_stop(handle);
281         return;
282 no_delete:
283         clear_inode(inode);     /* We must guarantee clearing of inode... */
284 }
285
286 typedef struct {
287         __le32  *p;
288         __le32  key;
289         struct buffer_head *bh;
290 } Indirect;
291
292 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
293 {
294         p->key = *(p->p = v);
295         p->bh = bh;
296 }
297
298 /**
299  *      ext4_block_to_path - parse the block number into array of offsets
300  *      @inode: inode in question (we are only interested in its superblock)
301  *      @i_block: block number to be parsed
302  *      @offsets: array to store the offsets in
303  *      @boundary: set this non-zero if the referred-to block is likely to be
304  *             followed (on disk) by an indirect block.
305  *
306  *      To store the locations of file's data ext4 uses a data structure common
307  *      for UNIX filesystems - tree of pointers anchored in the inode, with
308  *      data blocks at leaves and indirect blocks in intermediate nodes.
309  *      This function translates the block number into path in that tree -
310  *      return value is the path length and @offsets[n] is the offset of
311  *      pointer to (n+1)th node in the nth one. If @block is out of range
312  *      (negative or too large) warning is printed and zero returned.
313  *
314  *      Note: function doesn't find node addresses, so no IO is needed. All
315  *      we need to know is the capacity of indirect blocks (taken from the
316  *      inode->i_sb).
317  */
318
319 /*
320  * Portability note: the last comparison (check that we fit into triple
321  * indirect block) is spelled differently, because otherwise on an
322  * architecture with 32-bit longs and 8Kb pages we might get into trouble
323  * if our filesystem had 8Kb blocks. We might use long long, but that would
324  * kill us on x86. Oh, well, at least the sign propagation does not matter -
325  * i_block would have to be negative in the very beginning, so we would not
326  * get there at all.
327  */
328
329 static int ext4_block_to_path(struct inode *inode,
330                         ext4_lblk_t i_block,
331                         ext4_lblk_t offsets[4], int *boundary)
332 {
333         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
334         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
335         const long direct_blocks = EXT4_NDIR_BLOCKS,
336                 indirect_blocks = ptrs,
337                 double_blocks = (1 << (ptrs_bits * 2));
338         int n = 0;
339         int final = 0;
340
341         if (i_block < 0) {
342                 ext4_warning(inode->i_sb, "ext4_block_to_path", "block < 0");
343         } else if (i_block < direct_blocks) {
344                 offsets[n++] = i_block;
345                 final = direct_blocks;
346         } else if ((i_block -= direct_blocks) < indirect_blocks) {
347                 offsets[n++] = EXT4_IND_BLOCK;
348                 offsets[n++] = i_block;
349                 final = ptrs;
350         } else if ((i_block -= indirect_blocks) < double_blocks) {
351                 offsets[n++] = EXT4_DIND_BLOCK;
352                 offsets[n++] = i_block >> ptrs_bits;
353                 offsets[n++] = i_block & (ptrs - 1);
354                 final = ptrs;
355         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
356                 offsets[n++] = EXT4_TIND_BLOCK;
357                 offsets[n++] = i_block >> (ptrs_bits * 2);
358                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
359                 offsets[n++] = i_block & (ptrs - 1);
360                 final = ptrs;
361         } else {
362                 ext4_warning(inode->i_sb, "ext4_block_to_path",
363                                 "block %lu > max",
364                                 i_block + direct_blocks +
365                                 indirect_blocks + double_blocks);
366         }
367         if (boundary)
368                 *boundary = final - 1 - (i_block & (ptrs - 1));
369         return n;
370 }
371
372 /**
373  *      ext4_get_branch - read the chain of indirect blocks leading to data
374  *      @inode: inode in question
375  *      @depth: depth of the chain (1 - direct pointer, etc.)
376  *      @offsets: offsets of pointers in inode/indirect blocks
377  *      @chain: place to store the result
378  *      @err: here we store the error value
379  *
380  *      Function fills the array of triples <key, p, bh> and returns %NULL
381  *      if everything went OK or the pointer to the last filled triple
382  *      (incomplete one) otherwise. Upon the return chain[i].key contains
383  *      the number of (i+1)-th block in the chain (as it is stored in memory,
384  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
385  *      number (it points into struct inode for i==0 and into the bh->b_data
386  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
387  *      block for i>0 and NULL for i==0. In other words, it holds the block
388  *      numbers of the chain, addresses they were taken from (and where we can
389  *      verify that chain did not change) and buffer_heads hosting these
390  *      numbers.
391  *
392  *      Function stops when it stumbles upon zero pointer (absent block)
393  *              (pointer to last triple returned, *@err == 0)
394  *      or when it gets an IO error reading an indirect block
395  *              (ditto, *@err == -EIO)
396  *      or when it reads all @depth-1 indirect blocks successfully and finds
397  *      the whole chain, all way to the data (returns %NULL, *err == 0).
398  *
399  *      Need to be called with
400  *      down_read(&EXT4_I(inode)->i_data_sem)
401  */
402 static Indirect *ext4_get_branch(struct inode *inode, int depth,
403                                  ext4_lblk_t  *offsets,
404                                  Indirect chain[4], int *err)
405 {
406         struct super_block *sb = inode->i_sb;
407         Indirect *p = chain;
408         struct buffer_head *bh;
409
410         *err = 0;
411         /* i_data is not going away, no lock needed */
412         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
413         if (!p->key)
414                 goto no_block;
415         while (--depth) {
416                 bh = sb_bread(sb, le32_to_cpu(p->key));
417                 if (!bh)
418                         goto failure;
419                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
420                 /* Reader: end */
421                 if (!p->key)
422                         goto no_block;
423         }
424         return NULL;
425
426 failure:
427         *err = -EIO;
428 no_block:
429         return p;
430 }
431
432 /**
433  *      ext4_find_near - find a place for allocation with sufficient locality
434  *      @inode: owner
435  *      @ind: descriptor of indirect block.
436  *
437  *      This function returns the preferred place for block allocation.
438  *      It is used when heuristic for sequential allocation fails.
439  *      Rules are:
440  *        + if there is a block to the left of our position - allocate near it.
441  *        + if pointer will live in indirect block - allocate near that block.
442  *        + if pointer will live in inode - allocate in the same
443  *          cylinder group.
444  *
445  * In the latter case we colour the starting block by the callers PID to
446  * prevent it from clashing with concurrent allocations for a different inode
447  * in the same block group.   The PID is used here so that functionally related
448  * files will be close-by on-disk.
449  *
450  *      Caller must make sure that @ind is valid and will stay that way.
451  */
452 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
453 {
454         struct ext4_inode_info *ei = EXT4_I(inode);
455         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
456         __le32 *p;
457         ext4_fsblk_t bg_start;
458         ext4_fsblk_t last_block;
459         ext4_grpblk_t colour;
460
461         /* Try to find previous block */
462         for (p = ind->p - 1; p >= start; p--) {
463                 if (*p)
464                         return le32_to_cpu(*p);
465         }
466
467         /* No such thing, so let's try location of indirect block */
468         if (ind->bh)
469                 return ind->bh->b_blocknr;
470
471         /*
472          * It is going to be referred to from the inode itself? OK, just put it
473          * into the same cylinder group then.
474          */
475         bg_start = ext4_group_first_block_no(inode->i_sb, ei->i_block_group);
476         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
477
478         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
479                 colour = (current->pid % 16) *
480                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
481         else
482                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
483         return bg_start + colour;
484 }
485
486 /**
487  *      ext4_find_goal - find a preferred place for allocation.
488  *      @inode: owner
489  *      @block:  block we want
490  *      @partial: pointer to the last triple within a chain
491  *
492  *      Normally this function find the preferred place for block allocation,
493  *      returns it.
494  */
495 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
496                 Indirect *partial)
497 {
498         /*
499          * XXX need to get goal block from mballoc's data structures
500          */
501
502         return ext4_find_near(inode, partial);
503 }
504
505 /**
506  *      ext4_blks_to_allocate: Look up the block map and count the number
507  *      of direct blocks need to be allocated for the given branch.
508  *
509  *      @branch: chain of indirect blocks
510  *      @k: number of blocks need for indirect blocks
511  *      @blks: number of data blocks to be mapped.
512  *      @blocks_to_boundary:  the offset in the indirect block
513  *
514  *      return the total number of blocks to be allocate, including the
515  *      direct and indirect blocks.
516  */
517 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
518                 int blocks_to_boundary)
519 {
520         unsigned int count = 0;
521
522         /*
523          * Simple case, [t,d]Indirect block(s) has not allocated yet
524          * then it's clear blocks on that path have not allocated
525          */
526         if (k > 0) {
527                 /* right now we don't handle cross boundary allocation */
528                 if (blks < blocks_to_boundary + 1)
529                         count += blks;
530                 else
531                         count += blocks_to_boundary + 1;
532                 return count;
533         }
534
535         count++;
536         while (count < blks && count <= blocks_to_boundary &&
537                 le32_to_cpu(*(branch[0].p + count)) == 0) {
538                 count++;
539         }
540         return count;
541 }
542
543 /**
544  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
545  *      @indirect_blks: the number of blocks need to allocate for indirect
546  *                      blocks
547  *
548  *      @new_blocks: on return it will store the new block numbers for
549  *      the indirect blocks(if needed) and the first direct block,
550  *      @blks:  on return it will store the total number of allocated
551  *              direct blocks
552  */
553 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
554                                 ext4_lblk_t iblock, ext4_fsblk_t goal,
555                                 int indirect_blks, int blks,
556                                 ext4_fsblk_t new_blocks[4], int *err)
557 {
558         struct ext4_allocation_request ar;
559         int target, i;
560         unsigned long count = 0, blk_allocated = 0;
561         int index = 0;
562         ext4_fsblk_t current_block = 0;
563         int ret = 0;
564
565         /*
566          * Here we try to allocate the requested multiple blocks at once,
567          * on a best-effort basis.
568          * To build a branch, we should allocate blocks for
569          * the indirect blocks(if not allocated yet), and at least
570          * the first direct block of this branch.  That's the
571          * minimum number of blocks need to allocate(required)
572          */
573         /* first we try to allocate the indirect blocks */
574         target = indirect_blks;
575         while (target > 0) {
576                 count = target;
577                 /* allocating blocks for indirect blocks and direct blocks */
578                 current_block = ext4_new_meta_blocks(handle, inode,
579                                                         goal, &count, err);
580                 if (*err)
581                         goto failed_out;
582
583                 target -= count;
584                 /* allocate blocks for indirect blocks */
585                 while (index < indirect_blks && count) {
586                         new_blocks[index++] = current_block++;
587                         count--;
588                 }
589                 if (count > 0) {
590                         /*
591                          * save the new block number
592                          * for the first direct block
593                          */
594                         new_blocks[index] = current_block;
595                         printk(KERN_INFO "%s returned more blocks than "
596                                                 "requested\n", __func__);
597                         WARN_ON(1);
598                         break;
599                 }
600         }
601
602         target = blks - count ;
603         blk_allocated = count;
604         if (!target)
605                 goto allocated;
606         /* Now allocate data blocks */
607         memset(&ar, 0, sizeof(ar));
608         ar.inode = inode;
609         ar.goal = goal;
610         ar.len = target;
611         ar.logical = iblock;
612         if (S_ISREG(inode->i_mode))
613                 /* enable in-core preallocation only for regular files */
614                 ar.flags = EXT4_MB_HINT_DATA;
615
616         current_block = ext4_mb_new_blocks(handle, &ar, err);
617
618         if (*err && (target == blks)) {
619                 /*
620                  * if the allocation failed and we didn't allocate
621                  * any blocks before
622                  */
623                 goto failed_out;
624         }
625         if (!*err) {
626                 if (target == blks) {
627                 /*
628                  * save the new block number
629                  * for the first direct block
630                  */
631                         new_blocks[index] = current_block;
632                 }
633                 blk_allocated += ar.len;
634         }
635 allocated:
636         /* total number of blocks allocated for direct blocks */
637         ret = blk_allocated;
638         *err = 0;
639         return ret;
640 failed_out:
641         for (i = 0; i < index; i++)
642                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
643         return ret;
644 }
645
646 /**
647  *      ext4_alloc_branch - allocate and set up a chain of blocks.
648  *      @inode: owner
649  *      @indirect_blks: number of allocated indirect blocks
650  *      @blks: number of allocated direct blocks
651  *      @offsets: offsets (in the blocks) to store the pointers to next.
652  *      @branch: place to store the chain in.
653  *
654  *      This function allocates blocks, zeroes out all but the last one,
655  *      links them into chain and (if we are synchronous) writes them to disk.
656  *      In other words, it prepares a branch that can be spliced onto the
657  *      inode. It stores the information about that chain in the branch[], in
658  *      the same format as ext4_get_branch() would do. We are calling it after
659  *      we had read the existing part of chain and partial points to the last
660  *      triple of that (one with zero ->key). Upon the exit we have the same
661  *      picture as after the successful ext4_get_block(), except that in one
662  *      place chain is disconnected - *branch->p is still zero (we did not
663  *      set the last link), but branch->key contains the number that should
664  *      be placed into *branch->p to fill that gap.
665  *
666  *      If allocation fails we free all blocks we've allocated (and forget
667  *      their buffer_heads) and return the error value the from failed
668  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
669  *      as described above and return 0.
670  */
671 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
672                                 ext4_lblk_t iblock, int indirect_blks,
673                                 int *blks, ext4_fsblk_t goal,
674                                 ext4_lblk_t *offsets, Indirect *branch)
675 {
676         int blocksize = inode->i_sb->s_blocksize;
677         int i, n = 0;
678         int err = 0;
679         struct buffer_head *bh;
680         int num;
681         ext4_fsblk_t new_blocks[4];
682         ext4_fsblk_t current_block;
683
684         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
685                                 *blks, new_blocks, &err);
686         if (err)
687                 return err;
688
689         branch[0].key = cpu_to_le32(new_blocks[0]);
690         /*
691          * metadata blocks and data blocks are allocated.
692          */
693         for (n = 1; n <= indirect_blks;  n++) {
694                 /*
695                  * Get buffer_head for parent block, zero it out
696                  * and set the pointer to new one, then send
697                  * parent to disk.
698                  */
699                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
700                 branch[n].bh = bh;
701                 lock_buffer(bh);
702                 BUFFER_TRACE(bh, "call get_create_access");
703                 err = ext4_journal_get_create_access(handle, bh);
704                 if (err) {
705                         unlock_buffer(bh);
706                         brelse(bh);
707                         goto failed;
708                 }
709
710                 memset(bh->b_data, 0, blocksize);
711                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
712                 branch[n].key = cpu_to_le32(new_blocks[n]);
713                 *branch[n].p = branch[n].key;
714                 if (n == indirect_blks) {
715                         current_block = new_blocks[n];
716                         /*
717                          * End of chain, update the last new metablock of
718                          * the chain to point to the new allocated
719                          * data blocks numbers
720                          */
721                         for (i=1; i < num; i++)
722                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
723                 }
724                 BUFFER_TRACE(bh, "marking uptodate");
725                 set_buffer_uptodate(bh);
726                 unlock_buffer(bh);
727
728                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
729                 err = ext4_handle_dirty_metadata(handle, inode, bh);
730                 if (err)
731                         goto failed;
732         }
733         *blks = num;
734         return err;
735 failed:
736         /* Allocation failed, free what we already allocated */
737         for (i = 1; i <= n ; i++) {
738                 BUFFER_TRACE(branch[i].bh, "call jbd2_journal_forget");
739                 ext4_journal_forget(handle, branch[i].bh);
740         }
741         for (i = 0; i < indirect_blks; i++)
742                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
743
744         ext4_free_blocks(handle, inode, new_blocks[i], num, 0);
745
746         return err;
747 }
748
749 /**
750  * ext4_splice_branch - splice the allocated branch onto inode.
751  * @inode: owner
752  * @block: (logical) number of block we are adding
753  * @chain: chain of indirect blocks (with a missing link - see
754  *      ext4_alloc_branch)
755  * @where: location of missing link
756  * @num:   number of indirect blocks we are adding
757  * @blks:  number of direct blocks we are adding
758  *
759  * This function fills the missing link and does all housekeeping needed in
760  * inode (->i_blocks, etc.). In case of success we end up with the full
761  * chain to new block and return 0.
762  */
763 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
764                         ext4_lblk_t block, Indirect *where, int num, int blks)
765 {
766         int i;
767         int err = 0;
768         ext4_fsblk_t current_block;
769
770         /*
771          * If we're splicing into a [td]indirect block (as opposed to the
772          * inode) then we need to get write access to the [td]indirect block
773          * before the splice.
774          */
775         if (where->bh) {
776                 BUFFER_TRACE(where->bh, "get_write_access");
777                 err = ext4_journal_get_write_access(handle, where->bh);
778                 if (err)
779                         goto err_out;
780         }
781         /* That's it */
782
783         *where->p = where->key;
784
785         /*
786          * Update the host buffer_head or inode to point to more just allocated
787          * direct blocks blocks
788          */
789         if (num == 0 && blks > 1) {
790                 current_block = le32_to_cpu(where->key) + 1;
791                 for (i = 1; i < blks; i++)
792                         *(where->p + i) = cpu_to_le32(current_block++);
793         }
794
795         /* We are done with atomic stuff, now do the rest of housekeeping */
796
797         inode->i_ctime = ext4_current_time(inode);
798         ext4_mark_inode_dirty(handle, inode);
799
800         /* had we spliced it onto indirect block? */
801         if (where->bh) {
802                 /*
803                  * If we spliced it onto an indirect block, we haven't
804                  * altered the inode.  Note however that if it is being spliced
805                  * onto an indirect block at the very end of the file (the
806                  * file is growing) then we *will* alter the inode to reflect
807                  * the new i_size.  But that is not done here - it is done in
808                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
809                  */
810                 jbd_debug(5, "splicing indirect only\n");
811                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
812                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
813                 if (err)
814                         goto err_out;
815         } else {
816                 /*
817                  * OK, we spliced it into the inode itself on a direct block.
818                  * Inode was dirtied above.
819                  */
820                 jbd_debug(5, "splicing direct\n");
821         }
822         return err;
823
824 err_out:
825         for (i = 1; i <= num; i++) {
826                 BUFFER_TRACE(where[i].bh, "call jbd2_journal_forget");
827                 ext4_journal_forget(handle, where[i].bh);
828                 ext4_free_blocks(handle, inode,
829                                         le32_to_cpu(where[i-1].key), 1, 0);
830         }
831         ext4_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks, 0);
832
833         return err;
834 }
835
836 /*
837  * Allocation strategy is simple: if we have to allocate something, we will
838  * have to go the whole way to leaf. So let's do it before attaching anything
839  * to tree, set linkage between the newborn blocks, write them if sync is
840  * required, recheck the path, free and repeat if check fails, otherwise
841  * set the last missing link (that will protect us from any truncate-generated
842  * removals - all blocks on the path are immune now) and possibly force the
843  * write on the parent block.
844  * That has a nice additional property: no special recovery from the failed
845  * allocations is needed - we simply release blocks and do not touch anything
846  * reachable from inode.
847  *
848  * `handle' can be NULL if create == 0.
849  *
850  * return > 0, # of blocks mapped or allocated.
851  * return = 0, if plain lookup failed.
852  * return < 0, error case.
853  *
854  *
855  * Need to be called with
856  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
857  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
858  */
859 static int ext4_get_blocks_handle(handle_t *handle, struct inode *inode,
860                                   ext4_lblk_t iblock, unsigned int maxblocks,
861                                   struct buffer_head *bh_result,
862                                   int create, int extend_disksize)
863 {
864         int err = -EIO;
865         ext4_lblk_t offsets[4];
866         Indirect chain[4];
867         Indirect *partial;
868         ext4_fsblk_t goal;
869         int indirect_blks;
870         int blocks_to_boundary = 0;
871         int depth;
872         struct ext4_inode_info *ei = EXT4_I(inode);
873         int count = 0;
874         ext4_fsblk_t first_block = 0;
875         loff_t disksize;
876
877
878         J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
879         J_ASSERT(handle != NULL || create == 0);
880         depth = ext4_block_to_path(inode, iblock, offsets,
881                                         &blocks_to_boundary);
882
883         if (depth == 0)
884                 goto out;
885
886         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
887
888         /* Simplest case - block found, no allocation needed */
889         if (!partial) {
890                 first_block = le32_to_cpu(chain[depth - 1].key);
891                 clear_buffer_new(bh_result);
892                 count++;
893                 /*map more blocks*/
894                 while (count < maxblocks && count <= blocks_to_boundary) {
895                         ext4_fsblk_t blk;
896
897                         blk = le32_to_cpu(*(chain[depth-1].p + count));
898
899                         if (blk == first_block + count)
900                                 count++;
901                         else
902                                 break;
903                 }
904                 goto got_it;
905         }
906
907         /* Next simple case - plain lookup or failed read of indirect block */
908         if (!create || err == -EIO)
909                 goto cleanup;
910
911         /*
912          * Okay, we need to do block allocation.
913         */
914         goal = ext4_find_goal(inode, iblock, partial);
915
916         /* the number of blocks need to allocate for [d,t]indirect blocks */
917         indirect_blks = (chain + depth) - partial - 1;
918
919         /*
920          * Next look up the indirect map to count the totoal number of
921          * direct blocks to allocate for this branch.
922          */
923         count = ext4_blks_to_allocate(partial, indirect_blks,
924                                         maxblocks, blocks_to_boundary);
925         /*
926          * Block out ext4_truncate while we alter the tree
927          */
928         err = ext4_alloc_branch(handle, inode, iblock, indirect_blks,
929                                         &count, goal,
930                                         offsets + (partial - chain), partial);
931
932         /*
933          * The ext4_splice_branch call will free and forget any buffers
934          * on the new chain if there is a failure, but that risks using
935          * up transaction credits, especially for bitmaps where the
936          * credits cannot be returned.  Can we handle this somehow?  We
937          * may need to return -EAGAIN upwards in the worst case.  --sct
938          */
939         if (!err)
940                 err = ext4_splice_branch(handle, inode, iblock,
941                                         partial, indirect_blks, count);
942         /*
943          * i_disksize growing is protected by i_data_sem.  Don't forget to
944          * protect it if you're about to implement concurrent
945          * ext4_get_block() -bzzz
946         */
947         if (!err && extend_disksize) {
948                 disksize = ((loff_t) iblock + count) << inode->i_blkbits;
949                 if (disksize > i_size_read(inode))
950                         disksize = i_size_read(inode);
951                 if (disksize > ei->i_disksize)
952                         ei->i_disksize = disksize;
953         }
954         if (err)
955                 goto cleanup;
956
957         set_buffer_new(bh_result);
958 got_it:
959         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
960         if (count > blocks_to_boundary)
961                 set_buffer_boundary(bh_result);
962         err = count;
963         /* Clean up and exit */
964         partial = chain + depth - 1;    /* the whole chain */
965 cleanup:
966         while (partial > chain) {
967                 BUFFER_TRACE(partial->bh, "call brelse");
968                 brelse(partial->bh);
969                 partial--;
970         }
971         BUFFER_TRACE(bh_result, "returned");
972 out:
973         return err;
974 }
975
976 /*
977  * Calculate the number of metadata blocks need to reserve
978  * to allocate @blocks for non extent file based file
979  */
980 static int ext4_indirect_calc_metadata_amount(struct inode *inode, int blocks)
981 {
982         int icap = EXT4_ADDR_PER_BLOCK(inode->i_sb);
983         int ind_blks, dind_blks, tind_blks;
984
985         /* number of new indirect blocks needed */
986         ind_blks = (blocks + icap - 1) / icap;
987
988         dind_blks = (ind_blks + icap - 1) / icap;
989
990         tind_blks = 1;
991
992         return ind_blks + dind_blks + tind_blks;
993 }
994
995 /*
996  * Calculate the number of metadata blocks need to reserve
997  * to allocate given number of blocks
998  */
999 static int ext4_calc_metadata_amount(struct inode *inode, int blocks)
1000 {
1001         if (!blocks)
1002                 return 0;
1003
1004         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
1005                 return ext4_ext_calc_metadata_amount(inode, blocks);
1006
1007         return ext4_indirect_calc_metadata_amount(inode, blocks);
1008 }
1009
1010 static void ext4_da_update_reserve_space(struct inode *inode, int used)
1011 {
1012         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1013         int total, mdb, mdb_free;
1014
1015         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1016         /* recalculate the number of metablocks still need to be reserved */
1017         total = EXT4_I(inode)->i_reserved_data_blocks - used;
1018         mdb = ext4_calc_metadata_amount(inode, total);
1019
1020         /* figure out how many metablocks to release */
1021         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1022         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1023
1024         if (mdb_free) {
1025                 /* Account for allocated meta_blocks */
1026                 mdb_free -= EXT4_I(inode)->i_allocated_meta_blocks;
1027
1028                 /* update fs dirty blocks counter */
1029                 percpu_counter_sub(&sbi->s_dirtyblocks_counter, mdb_free);
1030                 EXT4_I(inode)->i_allocated_meta_blocks = 0;
1031                 EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1032         }
1033
1034         /* update per-inode reservations */
1035         BUG_ON(used  > EXT4_I(inode)->i_reserved_data_blocks);
1036         EXT4_I(inode)->i_reserved_data_blocks -= used;
1037
1038         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1039 }
1040
1041 /*
1042  * The ext4_get_blocks_wrap() function try to look up the requested blocks,
1043  * and returns if the blocks are already mapped.
1044  *
1045  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1046  * and store the allocated blocks in the result buffer head and mark it
1047  * mapped.
1048  *
1049  * If file type is extents based, it will call ext4_ext_get_blocks(),
1050  * Otherwise, call with ext4_get_blocks_handle() to handle indirect mapping
1051  * based files
1052  *
1053  * On success, it returns the number of blocks being mapped or allocate.
1054  * if create==0 and the blocks are pre-allocated and uninitialized block,
1055  * the result buffer head is unmapped. If the create ==1, it will make sure
1056  * the buffer head is mapped.
1057  *
1058  * It returns 0 if plain look up failed (blocks have not been allocated), in
1059  * that casem, buffer head is unmapped
1060  *
1061  * It returns the error in case of allocation failure.
1062  */
1063 int ext4_get_blocks_wrap(handle_t *handle, struct inode *inode, sector_t block,
1064                         unsigned int max_blocks, struct buffer_head *bh,
1065                         int create, int extend_disksize, int flag)
1066 {
1067         int retval;
1068
1069         clear_buffer_mapped(bh);
1070
1071         /*
1072          * Try to see if we can get  the block without requesting
1073          * for new file system block.
1074          */
1075         down_read((&EXT4_I(inode)->i_data_sem));
1076         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1077                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1078                                 bh, 0, 0);
1079         } else {
1080                 retval = ext4_get_blocks_handle(handle,
1081                                 inode, block, max_blocks, bh, 0, 0);
1082         }
1083         up_read((&EXT4_I(inode)->i_data_sem));
1084
1085         /* If it is only a block(s) look up */
1086         if (!create)
1087                 return retval;
1088
1089         /*
1090          * Returns if the blocks have already allocated
1091          *
1092          * Note that if blocks have been preallocated
1093          * ext4_ext_get_block() returns th create = 0
1094          * with buffer head unmapped.
1095          */
1096         if (retval > 0 && buffer_mapped(bh))
1097                 return retval;
1098
1099         /*
1100          * New blocks allocate and/or writing to uninitialized extent
1101          * will possibly result in updating i_data, so we take
1102          * the write lock of i_data_sem, and call get_blocks()
1103          * with create == 1 flag.
1104          */
1105         down_write((&EXT4_I(inode)->i_data_sem));
1106
1107         /*
1108          * if the caller is from delayed allocation writeout path
1109          * we have already reserved fs blocks for allocation
1110          * let the underlying get_block() function know to
1111          * avoid double accounting
1112          */
1113         if (flag)
1114                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1115         /*
1116          * We need to check for EXT4 here because migrate
1117          * could have changed the inode type in between
1118          */
1119         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1120                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1121                                 bh, create, extend_disksize);
1122         } else {
1123                 retval = ext4_get_blocks_handle(handle, inode, block,
1124                                 max_blocks, bh, create, extend_disksize);
1125
1126                 if (retval > 0 && buffer_new(bh)) {
1127                         /*
1128                          * We allocated new blocks which will result in
1129                          * i_data's format changing.  Force the migrate
1130                          * to fail by clearing migrate flags
1131                          */
1132                         EXT4_I(inode)->i_flags = EXT4_I(inode)->i_flags &
1133                                                         ~EXT4_EXT_MIGRATE;
1134                 }
1135         }
1136
1137         if (flag) {
1138                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1139                 /*
1140                  * Update reserved blocks/metadata blocks
1141                  * after successful block allocation
1142                  * which were deferred till now
1143                  */
1144                 if ((retval > 0) && buffer_delay(bh))
1145                         ext4_da_update_reserve_space(inode, retval);
1146         }
1147
1148         up_write((&EXT4_I(inode)->i_data_sem));
1149         return retval;
1150 }
1151
1152 /* Maximum number of blocks we map for direct IO at once. */
1153 #define DIO_MAX_BLOCKS 4096
1154
1155 int ext4_get_block(struct inode *inode, sector_t iblock,
1156                    struct buffer_head *bh_result, int create)
1157 {
1158         handle_t *handle = ext4_journal_current_handle();
1159         int ret = 0, started = 0;
1160         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1161         int dio_credits;
1162
1163         if (create && !handle) {
1164                 /* Direct IO write... */
1165                 if (max_blocks > DIO_MAX_BLOCKS)
1166                         max_blocks = DIO_MAX_BLOCKS;
1167                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
1168                 handle = ext4_journal_start(inode, dio_credits);
1169                 if (IS_ERR(handle)) {
1170                         ret = PTR_ERR(handle);
1171                         goto out;
1172                 }
1173                 started = 1;
1174         }
1175
1176         ret = ext4_get_blocks_wrap(handle, inode, iblock,
1177                                         max_blocks, bh_result, create, 0, 0);
1178         if (ret > 0) {
1179                 bh_result->b_size = (ret << inode->i_blkbits);
1180                 ret = 0;
1181         }
1182         if (started)
1183                 ext4_journal_stop(handle);
1184 out:
1185         return ret;
1186 }
1187
1188 /*
1189  * `handle' can be NULL if create is zero
1190  */
1191 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1192                                 ext4_lblk_t block, int create, int *errp)
1193 {
1194         struct buffer_head dummy;
1195         int fatal = 0, err;
1196
1197         J_ASSERT(handle != NULL || create == 0);
1198
1199         dummy.b_state = 0;
1200         dummy.b_blocknr = -1000;
1201         buffer_trace_init(&dummy.b_history);
1202         err = ext4_get_blocks_wrap(handle, inode, block, 1,
1203                                         &dummy, create, 1, 0);
1204         /*
1205          * ext4_get_blocks_handle() returns number of blocks
1206          * mapped. 0 in case of a HOLE.
1207          */
1208         if (err > 0) {
1209                 if (err > 1)
1210                         WARN_ON(1);
1211                 err = 0;
1212         }
1213         *errp = err;
1214         if (!err && buffer_mapped(&dummy)) {
1215                 struct buffer_head *bh;
1216                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1217                 if (!bh) {
1218                         *errp = -EIO;
1219                         goto err;
1220                 }
1221                 if (buffer_new(&dummy)) {
1222                         J_ASSERT(create != 0);
1223                         J_ASSERT(handle != NULL);
1224
1225                         /*
1226                          * Now that we do not always journal data, we should
1227                          * keep in mind whether this should always journal the
1228                          * new buffer as metadata.  For now, regular file
1229                          * writes use ext4_get_block instead, so it's not a
1230                          * problem.
1231                          */
1232                         lock_buffer(bh);
1233                         BUFFER_TRACE(bh, "call get_create_access");
1234                         fatal = ext4_journal_get_create_access(handle, bh);
1235                         if (!fatal && !buffer_uptodate(bh)) {
1236                                 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1237                                 set_buffer_uptodate(bh);
1238                         }
1239                         unlock_buffer(bh);
1240                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1241                         err = ext4_handle_dirty_metadata(handle, inode, bh);
1242                         if (!fatal)
1243                                 fatal = err;
1244                 } else {
1245                         BUFFER_TRACE(bh, "not a new buffer");
1246                 }
1247                 if (fatal) {
1248                         *errp = fatal;
1249                         brelse(bh);
1250                         bh = NULL;
1251                 }
1252                 return bh;
1253         }
1254 err:
1255         return NULL;
1256 }
1257
1258 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1259                                ext4_lblk_t block, int create, int *err)
1260 {
1261         struct buffer_head *bh;
1262
1263         bh = ext4_getblk(handle, inode, block, create, err);
1264         if (!bh)
1265                 return bh;
1266         if (buffer_uptodate(bh))
1267                 return bh;
1268         ll_rw_block(READ_META, 1, &bh);
1269         wait_on_buffer(bh);
1270         if (buffer_uptodate(bh))
1271                 return bh;
1272         put_bh(bh);
1273         *err = -EIO;
1274         return NULL;
1275 }
1276
1277 static int walk_page_buffers(handle_t *handle,
1278                              struct buffer_head *head,
1279                              unsigned from,
1280                              unsigned to,
1281                              int *partial,
1282                              int (*fn)(handle_t *handle,
1283                                        struct buffer_head *bh))
1284 {
1285         struct buffer_head *bh;
1286         unsigned block_start, block_end;
1287         unsigned blocksize = head->b_size;
1288         int err, ret = 0;
1289         struct buffer_head *next;
1290
1291         for (bh = head, block_start = 0;
1292              ret == 0 && (bh != head || !block_start);
1293              block_start = block_end, bh = next)
1294         {
1295                 next = bh->b_this_page;
1296                 block_end = block_start + blocksize;
1297                 if (block_end <= from || block_start >= to) {
1298                         if (partial && !buffer_uptodate(bh))
1299                                 *partial = 1;
1300                         continue;
1301                 }
1302                 err = (*fn)(handle, bh);
1303                 if (!ret)
1304                         ret = err;
1305         }
1306         return ret;
1307 }
1308
1309 /*
1310  * To preserve ordering, it is essential that the hole instantiation and
1311  * the data write be encapsulated in a single transaction.  We cannot
1312  * close off a transaction and start a new one between the ext4_get_block()
1313  * and the commit_write().  So doing the jbd2_journal_start at the start of
1314  * prepare_write() is the right place.
1315  *
1316  * Also, this function can nest inside ext4_writepage() ->
1317  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1318  * has generated enough buffer credits to do the whole page.  So we won't
1319  * block on the journal in that case, which is good, because the caller may
1320  * be PF_MEMALLOC.
1321  *
1322  * By accident, ext4 can be reentered when a transaction is open via
1323  * quota file writes.  If we were to commit the transaction while thus
1324  * reentered, there can be a deadlock - we would be holding a quota
1325  * lock, and the commit would never complete if another thread had a
1326  * transaction open and was blocking on the quota lock - a ranking
1327  * violation.
1328  *
1329  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1330  * will _not_ run commit under these circumstances because handle->h_ref
1331  * is elevated.  We'll still have enough credits for the tiny quotafile
1332  * write.
1333  */
1334 static int do_journal_get_write_access(handle_t *handle,
1335                                         struct buffer_head *bh)
1336 {
1337         if (!buffer_mapped(bh) || buffer_freed(bh))
1338                 return 0;
1339         return ext4_journal_get_write_access(handle, bh);
1340 }
1341
1342 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1343                                 loff_t pos, unsigned len, unsigned flags,
1344                                 struct page **pagep, void **fsdata)
1345 {
1346         struct inode *inode = mapping->host;
1347         int ret, needed_blocks = ext4_writepage_trans_blocks(inode);
1348         handle_t *handle;
1349         int retries = 0;
1350         struct page *page;
1351         pgoff_t index;
1352         unsigned from, to;
1353
1354         trace_mark(ext4_write_begin,
1355                    "dev %s ino %lu pos %llu len %u flags %u",
1356                    inode->i_sb->s_id, inode->i_ino,
1357                    (unsigned long long) pos, len, flags);
1358         index = pos >> PAGE_CACHE_SHIFT;
1359         from = pos & (PAGE_CACHE_SIZE - 1);
1360         to = from + len;
1361
1362 retry:
1363         handle = ext4_journal_start(inode, needed_blocks);
1364         if (IS_ERR(handle)) {
1365                 ret = PTR_ERR(handle);
1366                 goto out;
1367         }
1368
1369         page = grab_cache_page_write_begin(mapping, index, flags);
1370         if (!page) {
1371                 ext4_journal_stop(handle);
1372                 ret = -ENOMEM;
1373                 goto out;
1374         }
1375         *pagep = page;
1376
1377         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1378                                                         ext4_get_block);
1379
1380         if (!ret && ext4_should_journal_data(inode)) {
1381                 ret = walk_page_buffers(handle, page_buffers(page),
1382                                 from, to, NULL, do_journal_get_write_access);
1383         }
1384
1385         if (ret) {
1386                 unlock_page(page);
1387                 ext4_journal_stop(handle);
1388                 page_cache_release(page);
1389                 /*
1390                  * block_write_begin may have instantiated a few blocks
1391                  * outside i_size.  Trim these off again. Don't need
1392                  * i_size_read because we hold i_mutex.
1393                  */
1394                 if (pos + len > inode->i_size)
1395                         vmtruncate(inode, inode->i_size);
1396         }
1397
1398         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1399                 goto retry;
1400 out:
1401         return ret;
1402 }
1403
1404 /* For write_end() in data=journal mode */
1405 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1406 {
1407         if (!buffer_mapped(bh) || buffer_freed(bh))
1408                 return 0;
1409         set_buffer_uptodate(bh);
1410         return ext4_handle_dirty_metadata(handle, NULL, bh);
1411 }
1412
1413 /*
1414  * We need to pick up the new inode size which generic_commit_write gave us
1415  * `file' can be NULL - eg, when called from page_symlink().
1416  *
1417  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1418  * buffers are managed internally.
1419  */
1420 static int ext4_ordered_write_end(struct file *file,
1421                                 struct address_space *mapping,
1422                                 loff_t pos, unsigned len, unsigned copied,
1423                                 struct page *page, void *fsdata)
1424 {
1425         handle_t *handle = ext4_journal_current_handle();
1426         struct inode *inode = mapping->host;
1427         int ret = 0, ret2;
1428
1429         trace_mark(ext4_ordered_write_end,
1430                    "dev %s ino %lu pos %llu len %u copied %u",
1431                    inode->i_sb->s_id, inode->i_ino,
1432                    (unsigned long long) pos, len, copied);
1433         ret = ext4_jbd2_file_inode(handle, inode);
1434
1435         if (ret == 0) {
1436                 loff_t new_i_size;
1437
1438                 new_i_size = pos + copied;
1439                 if (new_i_size > EXT4_I(inode)->i_disksize) {
1440                         ext4_update_i_disksize(inode, new_i_size);
1441                         /* We need to mark inode dirty even if
1442                          * new_i_size is less that inode->i_size
1443                          * bu greater than i_disksize.(hint delalloc)
1444                          */
1445                         ext4_mark_inode_dirty(handle, inode);
1446                 }
1447
1448                 ret2 = generic_write_end(file, mapping, pos, len, copied,
1449                                                         page, fsdata);
1450                 copied = ret2;
1451                 if (ret2 < 0)
1452                         ret = ret2;
1453         }
1454         ret2 = ext4_journal_stop(handle);
1455         if (!ret)
1456                 ret = ret2;
1457
1458         return ret ? ret : copied;
1459 }
1460
1461 static int ext4_writeback_write_end(struct file *file,
1462                                 struct address_space *mapping,
1463                                 loff_t pos, unsigned len, unsigned copied,
1464                                 struct page *page, void *fsdata)
1465 {
1466         handle_t *handle = ext4_journal_current_handle();
1467         struct inode *inode = mapping->host;
1468         int ret = 0, ret2;
1469         loff_t new_i_size;
1470
1471         trace_mark(ext4_writeback_write_end,
1472                    "dev %s ino %lu pos %llu len %u copied %u",
1473                    inode->i_sb->s_id, inode->i_ino,
1474                    (unsigned long long) pos, len, copied);
1475         new_i_size = pos + copied;
1476         if (new_i_size > EXT4_I(inode)->i_disksize) {
1477                 ext4_update_i_disksize(inode, new_i_size);
1478                 /* We need to mark inode dirty even if
1479                  * new_i_size is less that inode->i_size
1480                  * bu greater than i_disksize.(hint delalloc)
1481                  */
1482                 ext4_mark_inode_dirty(handle, inode);
1483         }
1484
1485         ret2 = generic_write_end(file, mapping, pos, len, copied,
1486                                                         page, fsdata);
1487         copied = ret2;
1488         if (ret2 < 0)
1489                 ret = ret2;
1490
1491         ret2 = ext4_journal_stop(handle);
1492         if (!ret)
1493                 ret = ret2;
1494
1495         return ret ? ret : copied;
1496 }
1497
1498 static int ext4_journalled_write_end(struct file *file,
1499                                 struct address_space *mapping,
1500                                 loff_t pos, unsigned len, unsigned copied,
1501                                 struct page *page, void *fsdata)
1502 {
1503         handle_t *handle = ext4_journal_current_handle();
1504         struct inode *inode = mapping->host;
1505         int ret = 0, ret2;
1506         int partial = 0;
1507         unsigned from, to;
1508         loff_t new_i_size;
1509
1510         trace_mark(ext4_journalled_write_end,
1511                    "dev %s ino %lu pos %llu len %u copied %u",
1512                    inode->i_sb->s_id, inode->i_ino,
1513                    (unsigned long long) pos, len, copied);
1514         from = pos & (PAGE_CACHE_SIZE - 1);
1515         to = from + len;
1516
1517         if (copied < len) {
1518                 if (!PageUptodate(page))
1519                         copied = 0;
1520                 page_zero_new_buffers(page, from+copied, to);
1521         }
1522
1523         ret = walk_page_buffers(handle, page_buffers(page), from,
1524                                 to, &partial, write_end_fn);
1525         if (!partial)
1526                 SetPageUptodate(page);
1527         new_i_size = pos + copied;
1528         if (new_i_size > inode->i_size)
1529                 i_size_write(inode, pos+copied);
1530         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
1531         if (new_i_size > EXT4_I(inode)->i_disksize) {
1532                 ext4_update_i_disksize(inode, new_i_size);
1533                 ret2 = ext4_mark_inode_dirty(handle, inode);
1534                 if (!ret)
1535                         ret = ret2;
1536         }
1537
1538         unlock_page(page);
1539         ret2 = ext4_journal_stop(handle);
1540         if (!ret)
1541                 ret = ret2;
1542         page_cache_release(page);
1543
1544         return ret ? ret : copied;
1545 }
1546
1547 static int ext4_da_reserve_space(struct inode *inode, int nrblocks)
1548 {
1549         int retries = 0;
1550        struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1551        unsigned long md_needed, mdblocks, total = 0;
1552
1553         /*
1554          * recalculate the amount of metadata blocks to reserve
1555          * in order to allocate nrblocks
1556          * worse case is one extent per block
1557          */
1558 repeat:
1559         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1560         total = EXT4_I(inode)->i_reserved_data_blocks + nrblocks;
1561         mdblocks = ext4_calc_metadata_amount(inode, total);
1562         BUG_ON(mdblocks < EXT4_I(inode)->i_reserved_meta_blocks);
1563
1564         md_needed = mdblocks - EXT4_I(inode)->i_reserved_meta_blocks;
1565         total = md_needed + nrblocks;
1566
1567         if (ext4_claim_free_blocks(sbi, total)) {
1568                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1569                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1570                         yield();
1571                         goto repeat;
1572                 }
1573                 return -ENOSPC;
1574         }
1575         EXT4_I(inode)->i_reserved_data_blocks += nrblocks;
1576         EXT4_I(inode)->i_reserved_meta_blocks = mdblocks;
1577
1578         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1579         return 0;       /* success */
1580 }
1581
1582 static void ext4_da_release_space(struct inode *inode, int to_free)
1583 {
1584         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1585         int total, mdb, mdb_free, release;
1586
1587         if (!to_free)
1588                 return;         /* Nothing to release, exit */
1589
1590         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1591
1592         if (!EXT4_I(inode)->i_reserved_data_blocks) {
1593                 /*
1594                  * if there is no reserved blocks, but we try to free some
1595                  * then the counter is messed up somewhere.
1596                  * but since this function is called from invalidate
1597                  * page, it's harmless to return without any action
1598                  */
1599                 printk(KERN_INFO "ext4 delalloc try to release %d reserved "
1600                             "blocks for inode %lu, but there is no reserved "
1601                             "data blocks\n", to_free, inode->i_ino);
1602                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1603                 return;
1604         }
1605
1606         /* recalculate the number of metablocks still need to be reserved */
1607         total = EXT4_I(inode)->i_reserved_data_blocks - to_free;
1608         mdb = ext4_calc_metadata_amount(inode, total);
1609
1610         /* figure out how many metablocks to release */
1611         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1612         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1613
1614         release = to_free + mdb_free;
1615
1616         /* update fs dirty blocks counter for truncate case */
1617         percpu_counter_sub(&sbi->s_dirtyblocks_counter, release);
1618
1619         /* update per-inode reservations */
1620         BUG_ON(to_free > EXT4_I(inode)->i_reserved_data_blocks);
1621         EXT4_I(inode)->i_reserved_data_blocks -= to_free;
1622
1623         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1624         EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1625         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1626 }
1627
1628 static void ext4_da_page_release_reservation(struct page *page,
1629                                                 unsigned long offset)
1630 {
1631         int to_release = 0;
1632         struct buffer_head *head, *bh;
1633         unsigned int curr_off = 0;
1634
1635         head = page_buffers(page);
1636         bh = head;
1637         do {
1638                 unsigned int next_off = curr_off + bh->b_size;
1639
1640                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1641                         to_release++;
1642                         clear_buffer_delay(bh);
1643                 }
1644                 curr_off = next_off;
1645         } while ((bh = bh->b_this_page) != head);
1646         ext4_da_release_space(page->mapping->host, to_release);
1647 }
1648
1649 /*
1650  * Delayed allocation stuff
1651  */
1652
1653 struct mpage_da_data {
1654         struct inode *inode;
1655         struct buffer_head lbh;                 /* extent of blocks */
1656         unsigned long first_page, next_page;    /* extent of pages */
1657         get_block_t *get_block;
1658         struct writeback_control *wbc;
1659         int io_done;
1660         int pages_written;
1661         int retval;
1662 };
1663
1664 /*
1665  * mpage_da_submit_io - walks through extent of pages and try to write
1666  * them with writepage() call back
1667  *
1668  * @mpd->inode: inode
1669  * @mpd->first_page: first page of the extent
1670  * @mpd->next_page: page after the last page of the extent
1671  * @mpd->get_block: the filesystem's block mapper function
1672  *
1673  * By the time mpage_da_submit_io() is called we expect all blocks
1674  * to be allocated. this may be wrong if allocation failed.
1675  *
1676  * As pages are already locked by write_cache_pages(), we can't use it
1677  */
1678 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1679 {
1680         long pages_skipped;
1681         struct pagevec pvec;
1682         unsigned long index, end;
1683         int ret = 0, err, nr_pages, i;
1684         struct inode *inode = mpd->inode;
1685         struct address_space *mapping = inode->i_mapping;
1686
1687         BUG_ON(mpd->next_page <= mpd->first_page);
1688         /*
1689          * We need to start from the first_page to the next_page - 1
1690          * to make sure we also write the mapped dirty buffer_heads.
1691          * If we look at mpd->lbh.b_blocknr we would only be looking
1692          * at the currently mapped buffer_heads.
1693          */
1694         index = mpd->first_page;
1695         end = mpd->next_page - 1;
1696
1697         pagevec_init(&pvec, 0);
1698         while (index <= end) {
1699                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1700                 if (nr_pages == 0)
1701                         break;
1702                 for (i = 0; i < nr_pages; i++) {
1703                         struct page *page = pvec.pages[i];
1704
1705                         index = page->index;
1706                         if (index > end)
1707                                 break;
1708                         index++;
1709
1710                         BUG_ON(!PageLocked(page));
1711                         BUG_ON(PageWriteback(page));
1712
1713                         pages_skipped = mpd->wbc->pages_skipped;
1714                         err = mapping->a_ops->writepage(page, mpd->wbc);
1715                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
1716                                 /*
1717                                  * have successfully written the page
1718                                  * without skipping the same
1719                                  */
1720                                 mpd->pages_written++;
1721                         /*
1722                          * In error case, we have to continue because
1723                          * remaining pages are still locked
1724                          * XXX: unlock and re-dirty them?
1725                          */
1726                         if (ret == 0)
1727                                 ret = err;
1728                 }
1729                 pagevec_release(&pvec);
1730         }
1731         return ret;
1732 }
1733
1734 /*
1735  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
1736  *
1737  * @mpd->inode - inode to walk through
1738  * @exbh->b_blocknr - first block on a disk
1739  * @exbh->b_size - amount of space in bytes
1740  * @logical - first logical block to start assignment with
1741  *
1742  * the function goes through all passed space and put actual disk
1743  * block numbers into buffer heads, dropping BH_Delay
1744  */
1745 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
1746                                  struct buffer_head *exbh)
1747 {
1748         struct inode *inode = mpd->inode;
1749         struct address_space *mapping = inode->i_mapping;
1750         int blocks = exbh->b_size >> inode->i_blkbits;
1751         sector_t pblock = exbh->b_blocknr, cur_logical;
1752         struct buffer_head *head, *bh;
1753         pgoff_t index, end;
1754         struct pagevec pvec;
1755         int nr_pages, i;
1756
1757         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1758         end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1759         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1760
1761         pagevec_init(&pvec, 0);
1762
1763         while (index <= end) {
1764                 /* XXX: optimize tail */
1765                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1766                 if (nr_pages == 0)
1767                         break;
1768                 for (i = 0; i < nr_pages; i++) {
1769                         struct page *page = pvec.pages[i];
1770
1771                         index = page->index;
1772                         if (index > end)
1773                                 break;
1774                         index++;
1775
1776                         BUG_ON(!PageLocked(page));
1777                         BUG_ON(PageWriteback(page));
1778                         BUG_ON(!page_has_buffers(page));
1779
1780                         bh = page_buffers(page);
1781                         head = bh;
1782
1783                         /* skip blocks out of the range */
1784                         do {
1785                                 if (cur_logical >= logical)
1786                                         break;
1787                                 cur_logical++;
1788                         } while ((bh = bh->b_this_page) != head);
1789
1790                         do {
1791                                 if (cur_logical >= logical + blocks)
1792                                         break;
1793                                 if (buffer_delay(bh)) {
1794                                         bh->b_blocknr = pblock;
1795                                         clear_buffer_delay(bh);
1796                                         bh->b_bdev = inode->i_sb->s_bdev;
1797                                 } else if (buffer_unwritten(bh)) {
1798                                         bh->b_blocknr = pblock;
1799                                         clear_buffer_unwritten(bh);
1800                                         set_buffer_mapped(bh);
1801                                         set_buffer_new(bh);
1802                                         bh->b_bdev = inode->i_sb->s_bdev;
1803                                 } else if (buffer_mapped(bh))
1804                                         BUG_ON(bh->b_blocknr != pblock);
1805
1806                                 cur_logical++;
1807                                 pblock++;
1808                         } while ((bh = bh->b_this_page) != head);
1809                 }
1810                 pagevec_release(&pvec);
1811         }
1812 }
1813
1814
1815 /*
1816  * __unmap_underlying_blocks - just a helper function to unmap
1817  * set of blocks described by @bh
1818  */
1819 static inline void __unmap_underlying_blocks(struct inode *inode,
1820                                              struct buffer_head *bh)
1821 {
1822         struct block_device *bdev = inode->i_sb->s_bdev;
1823         int blocks, i;
1824
1825         blocks = bh->b_size >> inode->i_blkbits;
1826         for (i = 0; i < blocks; i++)
1827                 unmap_underlying_metadata(bdev, bh->b_blocknr + i);
1828 }
1829
1830 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
1831                                         sector_t logical, long blk_cnt)
1832 {
1833         int nr_pages, i;
1834         pgoff_t index, end;
1835         struct pagevec pvec;
1836         struct inode *inode = mpd->inode;
1837         struct address_space *mapping = inode->i_mapping;
1838
1839         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1840         end   = (logical + blk_cnt - 1) >>
1841                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
1842         while (index <= end) {
1843                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1844                 if (nr_pages == 0)
1845                         break;
1846                 for (i = 0; i < nr_pages; i++) {
1847                         struct page *page = pvec.pages[i];
1848                         index = page->index;
1849                         if (index > end)
1850                                 break;
1851                         index++;
1852
1853                         BUG_ON(!PageLocked(page));
1854                         BUG_ON(PageWriteback(page));
1855                         block_invalidatepage(page, 0);
1856                         ClearPageUptodate(page);
1857                         unlock_page(page);
1858                 }
1859         }
1860         return;
1861 }
1862
1863 static void ext4_print_free_blocks(struct inode *inode)
1864 {
1865         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1866         printk(KERN_EMERG "Total free blocks count %lld\n",
1867                         ext4_count_free_blocks(inode->i_sb));
1868         printk(KERN_EMERG "Free/Dirty block details\n");
1869         printk(KERN_EMERG "free_blocks=%lld\n",
1870                         (long long)percpu_counter_sum(&sbi->s_freeblocks_counter));
1871         printk(KERN_EMERG "dirty_blocks=%lld\n",
1872                         (long long)percpu_counter_sum(&sbi->s_dirtyblocks_counter));
1873         printk(KERN_EMERG "Block reservation details\n");
1874         printk(KERN_EMERG "i_reserved_data_blocks=%u\n",
1875                         EXT4_I(inode)->i_reserved_data_blocks);
1876         printk(KERN_EMERG "i_reserved_meta_blocks=%u\n",
1877                         EXT4_I(inode)->i_reserved_meta_blocks);
1878         return;
1879 }
1880
1881 /*
1882  * mpage_da_map_blocks - go through given space
1883  *
1884  * @mpd->lbh - bh describing space
1885  * @mpd->get_block - the filesystem's block mapper function
1886  *
1887  * The function skips space we know is already mapped to disk blocks.
1888  *
1889  */
1890 static int  mpage_da_map_blocks(struct mpage_da_data *mpd)
1891 {
1892         int err = 0;
1893         struct buffer_head new;
1894         struct buffer_head *lbh = &mpd->lbh;
1895         sector_t next;
1896
1897         /*
1898          * We consider only non-mapped and non-allocated blocks
1899          */
1900         if (buffer_mapped(lbh) && !buffer_delay(lbh))
1901                 return 0;
1902         new.b_state = lbh->b_state;
1903         new.b_blocknr = 0;
1904         new.b_size = lbh->b_size;
1905         next = lbh->b_blocknr;
1906         /*
1907          * If we didn't accumulate anything
1908          * to write simply return
1909          */
1910         if (!new.b_size)
1911                 return 0;
1912         err = mpd->get_block(mpd->inode, next, &new, 1);
1913         if (err) {
1914
1915                 /* If get block returns with error
1916                  * we simply return. Later writepage
1917                  * will redirty the page and writepages
1918                  * will find the dirty page again
1919                  */
1920                 if (err == -EAGAIN)
1921                         return 0;
1922
1923                 if (err == -ENOSPC &&
1924                                 ext4_count_free_blocks(mpd->inode->i_sb)) {
1925                         mpd->retval = err;
1926                         return 0;
1927                 }
1928
1929                 /*
1930                  * get block failure will cause us
1931                  * to loop in writepages. Because
1932                  * a_ops->writepage won't be able to
1933                  * make progress. The page will be redirtied
1934                  * by writepage and writepages will again
1935                  * try to write the same.
1936                  */
1937                 printk(KERN_EMERG "%s block allocation failed for inode %lu "
1938                                   "at logical offset %llu with max blocks "
1939                                   "%zd with error %d\n",
1940                                   __func__, mpd->inode->i_ino,
1941                                   (unsigned long long)next,
1942                                   lbh->b_size >> mpd->inode->i_blkbits, err);
1943                 printk(KERN_EMERG "This should not happen.!! "
1944                                         "Data will be lost\n");
1945                 if (err == -ENOSPC) {
1946                         ext4_print_free_blocks(mpd->inode);
1947                 }
1948                 /* invlaidate all the pages */
1949                 ext4_da_block_invalidatepages(mpd, next,
1950                                 lbh->b_size >> mpd->inode->i_blkbits);
1951                 return err;
1952         }
1953         BUG_ON(new.b_size == 0);
1954
1955         if (buffer_new(&new))
1956                 __unmap_underlying_blocks(mpd->inode, &new);
1957
1958         /*
1959          * If blocks are delayed marked, we need to
1960          * put actual blocknr and drop delayed bit
1961          */
1962         if (buffer_delay(lbh) || buffer_unwritten(lbh))
1963                 mpage_put_bnr_to_bhs(mpd, next, &new);
1964
1965         return 0;
1966 }
1967
1968 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
1969                 (1 << BH_Delay) | (1 << BH_Unwritten))
1970
1971 /*
1972  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
1973  *
1974  * @mpd->lbh - extent of blocks
1975  * @logical - logical number of the block in the file
1976  * @bh - bh of the block (used to access block's state)
1977  *
1978  * the function is used to collect contig. blocks in same state
1979  */
1980 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
1981                                    sector_t logical, struct buffer_head *bh)
1982 {
1983         sector_t next;
1984         size_t b_size = bh->b_size;
1985         struct buffer_head *lbh = &mpd->lbh;
1986         int nrblocks = lbh->b_size >> mpd->inode->i_blkbits;
1987
1988         /* check if thereserved journal credits might overflow */
1989         if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
1990                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
1991                         /*
1992                          * With non-extent format we are limited by the journal
1993                          * credit available.  Total credit needed to insert
1994                          * nrblocks contiguous blocks is dependent on the
1995                          * nrblocks.  So limit nrblocks.
1996                          */
1997                         goto flush_it;
1998                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
1999                                 EXT4_MAX_TRANS_DATA) {
2000                         /*
2001                          * Adding the new buffer_head would make it cross the
2002                          * allowed limit for which we have journal credit
2003                          * reserved. So limit the new bh->b_size
2004                          */
2005                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2006                                                 mpd->inode->i_blkbits;
2007                         /* we will do mpage_da_submit_io in the next loop */
2008                 }
2009         }
2010         /*
2011          * First block in the extent
2012          */
2013         if (lbh->b_size == 0) {
2014                 lbh->b_blocknr = logical;
2015                 lbh->b_size = b_size;
2016                 lbh->b_state = bh->b_state & BH_FLAGS;
2017                 return;
2018         }
2019
2020         next = lbh->b_blocknr + nrblocks;
2021         /*
2022          * Can we merge the block to our big extent?
2023          */
2024         if (logical == next && (bh->b_state & BH_FLAGS) == lbh->b_state) {
2025                 lbh->b_size += b_size;
2026                 return;
2027         }
2028
2029 flush_it:
2030         /*
2031          * We couldn't merge the block to our extent, so we
2032          * need to flush current  extent and start new one
2033          */
2034         if (mpage_da_map_blocks(mpd) == 0)
2035                 mpage_da_submit_io(mpd);
2036         mpd->io_done = 1;
2037         return;
2038 }
2039
2040 /*
2041  * __mpage_da_writepage - finds extent of pages and blocks
2042  *
2043  * @page: page to consider
2044  * @wbc: not used, we just follow rules
2045  * @data: context
2046  *
2047  * The function finds extents of pages and scan them for all blocks.
2048  */
2049 static int __mpage_da_writepage(struct page *page,
2050                                 struct writeback_control *wbc, void *data)
2051 {
2052         struct mpage_da_data *mpd = data;
2053         struct inode *inode = mpd->inode;
2054         struct buffer_head *bh, *head, fake;
2055         sector_t logical;
2056
2057         if (mpd->io_done) {
2058                 /*
2059                  * Rest of the page in the page_vec
2060                  * redirty then and skip then. We will
2061                  * try to to write them again after
2062                  * starting a new transaction
2063                  */
2064                 redirty_page_for_writepage(wbc, page);
2065                 unlock_page(page);
2066                 return MPAGE_DA_EXTENT_TAIL;
2067         }
2068         /*
2069          * Can we merge this page to current extent?
2070          */
2071         if (mpd->next_page != page->index) {
2072                 /*
2073                  * Nope, we can't. So, we map non-allocated blocks
2074                  * and start IO on them using writepage()
2075                  */
2076                 if (mpd->next_page != mpd->first_page) {
2077                         if (mpage_da_map_blocks(mpd) == 0)
2078                                 mpage_da_submit_io(mpd);
2079                         /*
2080                          * skip rest of the page in the page_vec
2081                          */
2082                         mpd->io_done = 1;
2083                         redirty_page_for_writepage(wbc, page);
2084                         unlock_page(page);
2085                         return MPAGE_DA_EXTENT_TAIL;
2086                 }
2087
2088                 /*
2089                  * Start next extent of pages ...
2090                  */
2091                 mpd->first_page = page->index;
2092
2093                 /*
2094                  * ... and blocks
2095                  */
2096                 mpd->lbh.b_size = 0;
2097                 mpd->lbh.b_state = 0;
2098                 mpd->lbh.b_blocknr = 0;
2099         }
2100
2101         mpd->next_page = page->index + 1;
2102         logical = (sector_t) page->index <<
2103                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2104
2105         if (!page_has_buffers(page)) {
2106                 /*
2107                  * There is no attached buffer heads yet (mmap?)
2108                  * we treat the page asfull of dirty blocks
2109                  */
2110                 bh = &fake;
2111                 bh->b_size = PAGE_CACHE_SIZE;
2112                 bh->b_state = 0;
2113                 set_buffer_dirty(bh);
2114                 set_buffer_uptodate(bh);
2115                 mpage_add_bh_to_extent(mpd, logical, bh);
2116                 if (mpd->io_done)
2117                         return MPAGE_DA_EXTENT_TAIL;
2118         } else {
2119                 /*
2120                  * Page with regular buffer heads, just add all dirty ones
2121                  */
2122                 head = page_buffers(page);
2123                 bh = head;
2124                 do {
2125                         BUG_ON(buffer_locked(bh));
2126                         /*
2127                          * We need to try to allocate
2128                          * unmapped blocks in the same page.
2129                          * Otherwise we won't make progress
2130                          * with the page in ext4_da_writepage
2131                          */
2132                         if (buffer_dirty(bh) &&
2133                                 (!buffer_mapped(bh) || buffer_delay(bh))) {
2134                                 mpage_add_bh_to_extent(mpd, logical, bh);
2135                                 if (mpd->io_done)
2136                                         return MPAGE_DA_EXTENT_TAIL;
2137                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2138                                 /*
2139                                  * mapped dirty buffer. We need to update
2140                                  * the b_state because we look at
2141                                  * b_state in mpage_da_map_blocks. We don't
2142                                  * update b_size because if we find an
2143                                  * unmapped buffer_head later we need to
2144                                  * use the b_state flag of that buffer_head.
2145                                  */
2146                                 if (mpd->lbh.b_size == 0)
2147                                         mpd->lbh.b_state =
2148                                                 bh->b_state & BH_FLAGS;
2149                         }
2150                         logical++;
2151                 } while ((bh = bh->b_this_page) != head);
2152         }
2153
2154         return 0;
2155 }
2156
2157 /*
2158  * mpage_da_writepages - walk the list of dirty pages of the given
2159  * address space, allocates non-allocated blocks, maps newly-allocated
2160  * blocks to existing bhs and issue IO them
2161  *
2162  * @mapping: address space structure to write
2163  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2164  * @get_block: the filesystem's block mapper function.
2165  *
2166  * This is a library function, which implements the writepages()
2167  * address_space_operation.
2168  */
2169 static int mpage_da_writepages(struct address_space *mapping,
2170                                struct writeback_control *wbc,
2171                                struct mpage_da_data *mpd)
2172 {
2173         int ret;
2174
2175         if (!mpd->get_block)
2176                 return generic_writepages(mapping, wbc);
2177
2178         mpd->lbh.b_size = 0;
2179         mpd->lbh.b_state = 0;
2180         mpd->lbh.b_blocknr = 0;
2181         mpd->first_page = 0;
2182         mpd->next_page = 0;
2183         mpd->io_done = 0;
2184         mpd->pages_written = 0;
2185         mpd->retval = 0;
2186
2187         ret = write_cache_pages(mapping, wbc, __mpage_da_writepage, mpd);
2188         /*
2189          * Handle last extent of pages
2190          */
2191         if (!mpd->io_done && mpd->next_page != mpd->first_page) {
2192                 if (mpage_da_map_blocks(mpd) == 0)
2193                         mpage_da_submit_io(mpd);
2194
2195                 mpd->io_done = 1;
2196                 ret = MPAGE_DA_EXTENT_TAIL;
2197         }
2198         wbc->nr_to_write -= mpd->pages_written;
2199         return ret;
2200 }
2201
2202 /*
2203  * this is a special callback for ->write_begin() only
2204  * it's intention is to return mapped block or reserve space
2205  */
2206 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2207                                   struct buffer_head *bh_result, int create)
2208 {
2209         int ret = 0;
2210
2211         BUG_ON(create == 0);
2212         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2213
2214         /*
2215          * first, we need to know whether the block is allocated already
2216          * preallocated blocks are unmapped but should treated
2217          * the same as allocated blocks.
2218          */
2219         ret = ext4_get_blocks_wrap(NULL, inode, iblock, 1,  bh_result, 0, 0, 0);
2220         if ((ret == 0) && !buffer_delay(bh_result)) {
2221                 /* the block isn't (pre)allocated yet, let's reserve space */
2222                 /*
2223                  * XXX: __block_prepare_write() unmaps passed block,
2224                  * is it OK?
2225                  */
2226                 ret = ext4_da_reserve_space(inode, 1);
2227                 if (ret)
2228                         /* not enough space to reserve */
2229                         return ret;
2230
2231                 map_bh(bh_result, inode->i_sb, 0);
2232                 set_buffer_new(bh_result);
2233                 set_buffer_delay(bh_result);
2234         } else if (ret > 0) {
2235                 bh_result->b_size = (ret << inode->i_blkbits);
2236                 ret = 0;
2237         }
2238
2239         return ret;
2240 }
2241 #define         EXT4_DELALLOC_RSVED     1
2242 static int ext4_da_get_block_write(struct inode *inode, sector_t iblock,
2243                                    struct buffer_head *bh_result, int create)
2244 {
2245         int ret;
2246         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2247         loff_t disksize = EXT4_I(inode)->i_disksize;
2248         handle_t *handle = NULL;
2249
2250         handle = ext4_journal_current_handle();
2251         BUG_ON(!handle);
2252         ret = ext4_get_blocks_wrap(handle, inode, iblock, max_blocks,
2253                         bh_result, create, 0, EXT4_DELALLOC_RSVED);
2254         if (ret > 0) {
2255
2256                 bh_result->b_size = (ret << inode->i_blkbits);
2257
2258                 if (ext4_should_order_data(inode)) {
2259                         int retval;
2260                         retval = ext4_jbd2_file_inode(handle, inode);
2261                         if (retval)
2262                                 /*
2263                                  * Failed to add inode for ordered
2264                                  * mode. Don't update file size
2265                                  */
2266                                 return retval;
2267                 }
2268
2269                 /*
2270                  * Update on-disk size along with block allocation
2271                  * we don't use 'extend_disksize' as size may change
2272                  * within already allocated block -bzzz
2273                  */
2274                 disksize = ((loff_t) iblock + ret) << inode->i_blkbits;
2275                 if (disksize > i_size_read(inode))
2276                         disksize = i_size_read(inode);
2277                 if (disksize > EXT4_I(inode)->i_disksize) {
2278                         ext4_update_i_disksize(inode, disksize);
2279                         ret = ext4_mark_inode_dirty(handle, inode);
2280                         return ret;
2281                 }
2282                 ret = 0;
2283         }
2284         return ret;
2285 }
2286
2287 static int ext4_bh_unmapped_or_delay(handle_t *handle, struct buffer_head *bh)
2288 {
2289         /*
2290          * unmapped buffer is possible for holes.
2291          * delay buffer is possible with delayed allocation
2292          */
2293         return ((!buffer_mapped(bh) || buffer_delay(bh)) && buffer_dirty(bh));
2294 }
2295
2296 static int ext4_normal_get_block_write(struct inode *inode, sector_t iblock,
2297                                    struct buffer_head *bh_result, int create)
2298 {
2299         int ret = 0;
2300         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2301
2302         /*
2303          * we don't want to do block allocation in writepage
2304          * so call get_block_wrap with create = 0
2305          */
2306         ret = ext4_get_blocks_wrap(NULL, inode, iblock, max_blocks,
2307                                    bh_result, 0, 0, 0);
2308         if (ret > 0) {
2309                 bh_result->b_size = (ret << inode->i_blkbits);
2310                 ret = 0;
2311         }
2312         return ret;
2313 }
2314
2315 /*
2316  * get called vi ext4_da_writepages after taking page lock (have journal handle)
2317  * get called via journal_submit_inode_data_buffers (no journal handle)
2318  * get called via shrink_page_list via pdflush (no journal handle)
2319  * or grab_page_cache when doing write_begin (have journal handle)
2320  */
2321 static int ext4_da_writepage(struct page *page,
2322                                 struct writeback_control *wbc)
2323 {
2324         int ret = 0;
2325         loff_t size;
2326         unsigned int len;
2327         struct buffer_head *page_bufs;
2328         struct inode *inode = page->mapping->host;
2329
2330         trace_mark(ext4_da_writepage,
2331                    "dev %s ino %lu page_index %lu",
2332                    inode->i_sb->s_id, inode->i_ino, page->index);
2333         size = i_size_read(inode);
2334         if (page->index == size >> PAGE_CACHE_SHIFT)
2335                 len = size & ~PAGE_CACHE_MASK;
2336         else
2337                 len = PAGE_CACHE_SIZE;
2338
2339         if (page_has_buffers(page)) {
2340                 page_bufs = page_buffers(page);
2341                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2342                                         ext4_bh_unmapped_or_delay)) {
2343                         /*
2344                          * We don't want to do  block allocation
2345                          * So redirty the page and return
2346                          * We may reach here when we do a journal commit
2347                          * via journal_submit_inode_data_buffers.
2348                          * If we don't have mapping block we just ignore
2349                          * them. We can also reach here via shrink_page_list
2350                          */
2351                         redirty_page_for_writepage(wbc, page);
2352                         unlock_page(page);
2353                         return 0;
2354                 }
2355         } else {
2356                 /*
2357                  * The test for page_has_buffers() is subtle:
2358                  * We know the page is dirty but it lost buffers. That means
2359                  * that at some moment in time after write_begin()/write_end()
2360                  * has been called all buffers have been clean and thus they
2361                  * must have been written at least once. So they are all
2362                  * mapped and we can happily proceed with mapping them
2363                  * and writing the page.
2364                  *
2365                  * Try to initialize the buffer_heads and check whether
2366                  * all are mapped and non delay. We don't want to
2367                  * do block allocation here.
2368                  */
2369                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
2370                                                 ext4_normal_get_block_write);
2371                 if (!ret) {
2372                         page_bufs = page_buffers(page);
2373                         /* check whether all are mapped and non delay */
2374                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2375                                                 ext4_bh_unmapped_or_delay)) {
2376                                 redirty_page_for_writepage(wbc, page);
2377                                 unlock_page(page);
2378                                 return 0;
2379                         }
2380                 } else {
2381                         /*
2382                          * We can't do block allocation here
2383                          * so just redity the page and unlock
2384                          * and return
2385                          */
2386                         redirty_page_for_writepage(wbc, page);
2387                         unlock_page(page);
2388                         return 0;
2389                 }
2390                 /* now mark the buffer_heads as dirty and uptodate */
2391                 block_commit_write(page, 0, PAGE_CACHE_SIZE);
2392         }
2393
2394         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2395                 ret = nobh_writepage(page, ext4_normal_get_block_write, wbc);
2396         else
2397                 ret = block_write_full_page(page,
2398                                                 ext4_normal_get_block_write,
2399                                                 wbc);
2400
2401         return ret;
2402 }
2403
2404 /*
2405  * This is called via ext4_da_writepages() to
2406  * calulate the total number of credits to reserve to fit
2407  * a single extent allocation into a single transaction,
2408  * ext4_da_writpeages() will loop calling this before
2409  * the block allocation.
2410  */
2411
2412 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2413 {
2414         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2415
2416         /*
2417          * With non-extent format the journal credit needed to
2418          * insert nrblocks contiguous block is dependent on
2419          * number of contiguous block. So we will limit
2420          * number of contiguous block to a sane value
2421          */
2422         if (!(inode->i_flags & EXT4_EXTENTS_FL) &&
2423             (max_blocks > EXT4_MAX_TRANS_DATA))
2424                 max_blocks = EXT4_MAX_TRANS_DATA;
2425
2426         return ext4_chunk_trans_blocks(inode, max_blocks);
2427 }
2428
2429 static int ext4_da_writepages(struct address_space *mapping,
2430                               struct writeback_control *wbc)
2431 {
2432         pgoff_t index;
2433         int range_whole = 0;
2434         handle_t *handle = NULL;
2435         struct mpage_da_data mpd;
2436         struct inode *inode = mapping->host;
2437         int no_nrwrite_index_update;
2438         int pages_written = 0;
2439         long pages_skipped;
2440         int needed_blocks, ret = 0, nr_to_writebump = 0;
2441         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2442
2443         trace_mark(ext4_da_writepages,
2444                    "dev %s ino %lu nr_t_write %ld "
2445                    "pages_skipped %ld range_start %llu "
2446                    "range_end %llu nonblocking %d "
2447                    "for_kupdate %d for_reclaim %d "
2448                    "for_writepages %d range_cyclic %d",
2449                    inode->i_sb->s_id, inode->i_ino,
2450                    wbc->nr_to_write, wbc->pages_skipped,
2451                    (unsigned long long) wbc->range_start,
2452                    (unsigned long long) wbc->range_end,
2453                    wbc->nonblocking, wbc->for_kupdate,
2454                    wbc->for_reclaim, wbc->for_writepages,
2455                    wbc->range_cyclic);
2456
2457         /*
2458          * No pages to write? This is mainly a kludge to avoid starting
2459          * a transaction for special inodes like journal inode on last iput()
2460          * because that could violate lock ordering on umount
2461          */
2462         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2463                 return 0;
2464
2465         /*
2466          * If the filesystem has aborted, it is read-only, so return
2467          * right away instead of dumping stack traces later on that
2468          * will obscure the real source of the problem.  We test
2469          * EXT4_MOUNT_ABORT instead of sb->s_flag's MS_RDONLY because
2470          * the latter could be true if the filesystem is mounted
2471          * read-only, and in that case, ext4_da_writepages should
2472          * *never* be called, so if that ever happens, we would want
2473          * the stack trace.
2474          */
2475         if (unlikely(sbi->s_mount_opt & EXT4_MOUNT_ABORT))
2476                 return -EROFS;
2477
2478         /*
2479          * Make sure nr_to_write is >= sbi->s_mb_stream_request
2480          * This make sure small files blocks are allocated in
2481          * single attempt. This ensure that small files
2482          * get less fragmented.
2483          */
2484         if (wbc->nr_to_write < sbi->s_mb_stream_request) {
2485                 nr_to_writebump = sbi->s_mb_stream_request - wbc->nr_to_write;
2486                 wbc->nr_to_write = sbi->s_mb_stream_request;
2487         }
2488         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2489                 range_whole = 1;
2490
2491         if (wbc->range_cyclic)
2492                 index = mapping->writeback_index;
2493         else
2494                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2495
2496         mpd.wbc = wbc;
2497         mpd.inode = mapping->host;
2498
2499         /*
2500          * we don't want write_cache_pages to update
2501          * nr_to_write and writeback_index
2502          */
2503         no_nrwrite_index_update = wbc->no_nrwrite_index_update;
2504         wbc->no_nrwrite_index_update = 1;
2505         pages_skipped = wbc->pages_skipped;
2506
2507         while (!ret && wbc->nr_to_write > 0) {
2508
2509                 /*
2510                  * we  insert one extent at a time. So we need
2511                  * credit needed for single extent allocation.
2512                  * journalled mode is currently not supported
2513                  * by delalloc
2514                  */
2515                 BUG_ON(ext4_should_journal_data(inode));
2516                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2517
2518                 /* start a new transaction*/
2519                 handle = ext4_journal_start(inode, needed_blocks);
2520                 if (IS_ERR(handle)) {
2521                         ret = PTR_ERR(handle);
2522                         printk(KERN_CRIT "%s: jbd2_start: "
2523                                "%ld pages, ino %lu; err %d\n", __func__,
2524                                 wbc->nr_to_write, inode->i_ino, ret);
2525                         dump_stack();
2526                         goto out_writepages;
2527                 }
2528                 mpd.get_block = ext4_da_get_block_write;
2529                 ret = mpage_da_writepages(mapping, wbc, &mpd);
2530
2531                 ext4_journal_stop(handle);
2532
2533                 if (mpd.retval == -ENOSPC) {
2534                         /* commit the transaction which would
2535                          * free blocks released in the transaction
2536                          * and try again
2537                          */
2538                         jbd2_journal_force_commit_nested(sbi->s_journal);
2539                         wbc->pages_skipped = pages_skipped;
2540                         ret = 0;
2541                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2542                         /*
2543                          * got one extent now try with
2544                          * rest of the pages
2545                          */
2546                         pages_written += mpd.pages_written;
2547                         wbc->pages_skipped = pages_skipped;
2548                         ret = 0;
2549                 } else if (wbc->nr_to_write)
2550                         /*
2551                          * There is no more writeout needed
2552                          * or we requested for a noblocking writeout
2553                          * and we found the device congested
2554                          */
2555                         break;
2556         }
2557         if (pages_skipped != wbc->pages_skipped)
2558                 printk(KERN_EMERG "This should not happen leaving %s "
2559                                 "with nr_to_write = %ld ret = %d\n",
2560                                 __func__, wbc->nr_to_write, ret);
2561
2562         /* Update index */
2563         index += pages_written;
2564         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2565                 /*
2566                  * set the writeback_index so that range_cyclic
2567                  * mode will write it back later
2568                  */
2569                 mapping->writeback_index = index;
2570
2571 out_writepages:
2572         if (!no_nrwrite_index_update)
2573                 wbc->no_nrwrite_index_update = 0;
2574         wbc->nr_to_write -= nr_to_writebump;
2575         trace_mark(ext4_da_writepage_result,
2576                    "dev %s ino %lu ret %d pages_written %d "
2577                    "pages_skipped %ld congestion %d "
2578                    "more_io %d no_nrwrite_index_update %d",
2579                    inode->i_sb->s_id, inode->i_ino, ret,
2580                    pages_written, wbc->pages_skipped,
2581                    wbc->encountered_congestion, wbc->more_io,
2582                    wbc->no_nrwrite_index_update);
2583         return ret;
2584 }
2585
2586 #define FALL_BACK_TO_NONDELALLOC 1
2587 static int ext4_nonda_switch(struct super_block *sb)
2588 {
2589         s64 free_blocks, dirty_blocks;
2590         struct ext4_sb_info *sbi = EXT4_SB(sb);
2591
2592         /*
2593          * switch to non delalloc mode if we are running low
2594          * on free block. The free block accounting via percpu
2595          * counters can get slightly wrong with percpu_counter_batch getting
2596          * accumulated on each CPU without updating global counters
2597          * Delalloc need an accurate free block accounting. So switch
2598          * to non delalloc when we are near to error range.
2599          */
2600         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
2601         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
2602         if (2 * free_blocks < 3 * dirty_blocks ||
2603                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
2604                 /*
2605                  * free block count is less that 150% of dirty blocks
2606                  * or free blocks is less that watermark
2607                  */
2608                 return 1;
2609         }
2610         return 0;
2611 }
2612
2613 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2614                                 loff_t pos, unsigned len, unsigned flags,
2615                                 struct page **pagep, void **fsdata)
2616 {
2617         int ret, retries = 0;
2618         struct page *page;
2619         pgoff_t index;
2620         unsigned from, to;
2621         struct inode *inode = mapping->host;
2622         handle_t *handle;
2623
2624         index = pos >> PAGE_CACHE_SHIFT;
2625         from = pos & (PAGE_CACHE_SIZE - 1);
2626         to = from + len;
2627
2628         if (ext4_nonda_switch(inode->i_sb)) {
2629                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2630                 return ext4_write_begin(file, mapping, pos,
2631                                         len, flags, pagep, fsdata);
2632         }
2633         *fsdata = (void *)0;
2634
2635         trace_mark(ext4_da_write_begin,
2636                    "dev %s ino %lu pos %llu len %u flags %u",
2637                    inode->i_sb->s_id, inode->i_ino,
2638                    (unsigned long long) pos, len, flags);
2639 retry:
2640         /*
2641          * With delayed allocation, we don't log the i_disksize update
2642          * if there is delayed block allocation. But we still need
2643          * to journalling the i_disksize update if writes to the end
2644          * of file which has an already mapped buffer.
2645          */
2646         handle = ext4_journal_start(inode, 1);
2647         if (IS_ERR(handle)) {
2648                 ret = PTR_ERR(handle);
2649                 goto out;
2650         }
2651
2652         page = grab_cache_page_write_begin(mapping, index, flags);
2653         if (!page) {
2654                 ext4_journal_stop(handle);
2655                 ret = -ENOMEM;
2656                 goto out;
2657         }
2658         *pagep = page;
2659
2660         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
2661                                                         ext4_da_get_block_prep);
2662         if (ret < 0) {
2663                 unlock_page(page);
2664                 ext4_journal_stop(handle);
2665                 page_cache_release(page);
2666                 /*
2667                  * block_write_begin may have instantiated a few blocks
2668                  * outside i_size.  Trim these off again. Don't need
2669                  * i_size_read because we hold i_mutex.
2670                  */
2671                 if (pos + len > inode->i_size)
2672                         vmtruncate(inode, inode->i_size);
2673         }
2674
2675         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
2676                 goto retry;
2677 out:
2678         return ret;
2679 }
2680
2681 /*
2682  * Check if we should update i_disksize
2683  * when write to the end of file but not require block allocation
2684  */
2685 static int ext4_da_should_update_i_disksize(struct page *page,
2686                                          unsigned long offset)
2687 {
2688         struct buffer_head *bh;
2689         struct inode *inode = page->mapping->host;
2690         unsigned int idx;
2691         int i;
2692
2693         bh = page_buffers(page);
2694         idx = offset >> inode->i_blkbits;
2695
2696         for (i = 0; i < idx; i++)
2697                 bh = bh->b_this_page;
2698
2699         if (!buffer_mapped(bh) || (buffer_delay(bh)))
2700                 return 0;
2701         return 1;
2702 }
2703
2704 static int ext4_da_write_end(struct file *file,
2705                                 struct address_space *mapping,
2706                                 loff_t pos, unsigned len, unsigned copied,
2707                                 struct page *page, void *fsdata)
2708 {
2709         struct inode *inode = mapping->host;
2710         int ret = 0, ret2;
2711         handle_t *handle = ext4_journal_current_handle();
2712         loff_t new_i_size;
2713         unsigned long start, end;
2714         int write_mode = (int)(unsigned long)fsdata;
2715
2716         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
2717                 if (ext4_should_order_data(inode)) {
2718                         return ext4_ordered_write_end(file, mapping, pos,
2719                                         len, copied, page, fsdata);
2720                 } else if (ext4_should_writeback_data(inode)) {
2721                         return ext4_writeback_write_end(file, mapping, pos,
2722                                         len, copied, page, fsdata);
2723                 } else {
2724                         BUG();
2725                 }
2726         }
2727
2728         trace_mark(ext4_da_write_end,
2729                    "dev %s ino %lu pos %llu len %u copied %u",
2730                    inode->i_sb->s_id, inode->i_ino,
2731                    (unsigned long long) pos, len, copied);
2732         start = pos & (PAGE_CACHE_SIZE - 1);
2733         end = start + copied - 1;
2734
2735         /*
2736          * generic_write_end() will run mark_inode_dirty() if i_size
2737          * changes.  So let's piggyback the i_disksize mark_inode_dirty
2738          * into that.
2739          */
2740
2741         new_i_size = pos + copied;
2742         if (new_i_size > EXT4_I(inode)->i_disksize) {
2743                 if (ext4_da_should_update_i_disksize(page, end)) {
2744                         down_write(&EXT4_I(inode)->i_data_sem);
2745                         if (new_i_size > EXT4_I(inode)->i_disksize) {
2746                                 /*
2747                                  * Updating i_disksize when extending file
2748                                  * without needing block allocation
2749                                  */
2750                                 if (ext4_should_order_data(inode))
2751                                         ret = ext4_jbd2_file_inode(handle,
2752                                                                    inode);
2753
2754                                 EXT4_I(inode)->i_disksize = new_i_size;
2755                         }
2756                         up_write(&EXT4_I(inode)->i_data_sem);
2757                         /* We need to mark inode dirty even if
2758                          * new_i_size is less that inode->i_size
2759                          * bu greater than i_disksize.(hint delalloc)
2760                          */
2761                         ext4_mark_inode_dirty(handle, inode);
2762                 }
2763         }
2764         ret2 = generic_write_end(file, mapping, pos, len, copied,
2765                                                         page, fsdata);
2766         copied = ret2;
2767         if (ret2 < 0)
2768                 ret = ret2;
2769         ret2 = ext4_journal_stop(handle);
2770         if (!ret)
2771                 ret = ret2;
2772
2773         return ret ? ret : copied;
2774 }
2775
2776 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
2777 {
2778         /*
2779          * Drop reserved blocks
2780          */
2781         BUG_ON(!PageLocked(page));
2782         if (!page_has_buffers(page))
2783                 goto out;
2784
2785         ext4_da_page_release_reservation(page, offset);
2786
2787 out:
2788         ext4_invalidatepage(page, offset);
2789
2790         return;
2791 }
2792
2793
2794 /*
2795  * bmap() is special.  It gets used by applications such as lilo and by
2796  * the swapper to find the on-disk block of a specific piece of data.
2797  *
2798  * Naturally, this is dangerous if the block concerned is still in the
2799  * journal.  If somebody makes a swapfile on an ext4 data-journaling
2800  * filesystem and enables swap, then they may get a nasty shock when the
2801  * data getting swapped to that swapfile suddenly gets overwritten by
2802  * the original zero's written out previously to the journal and
2803  * awaiting writeback in the kernel's buffer cache.
2804  *
2805  * So, if we see any bmap calls here on a modified, data-journaled file,
2806  * take extra steps to flush any blocks which might be in the cache.
2807  */
2808 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
2809 {
2810         struct inode *inode = mapping->host;
2811         journal_t *journal;
2812         int err;
2813
2814         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
2815                         test_opt(inode->i_sb, DELALLOC)) {
2816                 /*
2817                  * With delalloc we want to sync the file
2818                  * so that we can make sure we allocate
2819                  * blocks for file
2820                  */
2821                 filemap_write_and_wait(mapping);
2822         }
2823
2824         BUG_ON(!EXT4_JOURNAL(inode) &&
2825                EXT4_I(inode)->i_state & EXT4_STATE_JDATA);
2826
2827         if (EXT4_JOURNAL(inode) && EXT4_I(inode)->i_state & EXT4_STATE_JDATA) {
2828                 /*
2829                  * This is a REALLY heavyweight approach, but the use of
2830                  * bmap on dirty files is expected to be extremely rare:
2831                  * only if we run lilo or swapon on a freshly made file
2832                  * do we expect this to happen.
2833                  *
2834                  * (bmap requires CAP_SYS_RAWIO so this does not
2835                  * represent an unprivileged user DOS attack --- we'd be
2836                  * in trouble if mortal users could trigger this path at
2837                  * will.)
2838                  *
2839                  * NB. EXT4_STATE_JDATA is not set on files other than
2840                  * regular files.  If somebody wants to bmap a directory
2841                  * or symlink and gets confused because the buffer
2842                  * hasn't yet been flushed to disk, they deserve
2843                  * everything they get.
2844                  */
2845
2846                 EXT4_I(inode)->i_state &= ~EXT4_STATE_JDATA;
2847                 journal = EXT4_JOURNAL(inode);
2848                 jbd2_journal_lock_updates(journal);
2849                 err = jbd2_journal_flush(journal);
2850                 jbd2_journal_unlock_updates(journal);
2851
2852                 if (err)
2853                         return 0;
2854         }
2855
2856         return generic_block_bmap(mapping, block, ext4_get_block);
2857 }
2858
2859 static int bget_one(handle_t *handle, struct buffer_head *bh)
2860 {
2861         get_bh(bh);
2862         return 0;
2863 }
2864
2865 static int bput_one(handle_t *handle, struct buffer_head *bh)
2866 {
2867         put_bh(bh);
2868         return 0;
2869 }
2870
2871 /*
2872  * Note that we don't need to start a transaction unless we're journaling data
2873  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2874  * need to file the inode to the transaction's list in ordered mode because if
2875  * we are writing back data added by write(), the inode is already there and if
2876  * we are writing back data modified via mmap(), noone guarantees in which
2877  * transaction the data will hit the disk. In case we are journaling data, we
2878  * cannot start transaction directly because transaction start ranks above page
2879  * lock so we have to do some magic.
2880  *
2881  * In all journaling modes block_write_full_page() will start the I/O.
2882  *
2883  * Problem:
2884  *
2885  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2886  *              ext4_writepage()
2887  *
2888  * Similar for:
2889  *
2890  *      ext4_file_write() -> generic_file_write() -> __alloc_pages() -> ...
2891  *
2892  * Same applies to ext4_get_block().  We will deadlock on various things like
2893  * lock_journal and i_data_sem
2894  *
2895  * Setting PF_MEMALLOC here doesn't work - too many internal memory
2896  * allocations fail.
2897  *
2898  * 16May01: If we're reentered then journal_current_handle() will be
2899  *          non-zero. We simply *return*.
2900  *
2901  * 1 July 2001: @@@ FIXME:
2902  *   In journalled data mode, a data buffer may be metadata against the
2903  *   current transaction.  But the same file is part of a shared mapping
2904  *   and someone does a writepage() on it.
2905  *
2906  *   We will move the buffer onto the async_data list, but *after* it has
2907  *   been dirtied. So there's a small window where we have dirty data on
2908  *   BJ_Metadata.
2909  *
2910  *   Note that this only applies to the last partial page in the file.  The
2911  *   bit which block_write_full_page() uses prepare/commit for.  (That's
2912  *   broken code anyway: it's wrong for msync()).
2913  *
2914  *   It's a rare case: affects the final partial page, for journalled data
2915  *   where the file is subject to bith write() and writepage() in the same
2916  *   transction.  To fix it we'll need a custom block_write_full_page().
2917  *   We'll probably need that anyway for journalling writepage() output.
2918  *
2919  * We don't honour synchronous mounts for writepage().  That would be
2920  * disastrous.  Any write() or metadata operation will sync the fs for
2921  * us.
2922  *
2923  */
2924 static int __ext4_normal_writepage(struct page *page,
2925                                 struct writeback_control *wbc)
2926 {
2927         struct inode *inode = page->mapping->host;
2928
2929         if (test_opt(inode->i_sb, NOBH))
2930                 return nobh_writepage(page,
2931                                         ext4_normal_get_block_write, wbc);
2932         else
2933                 return block_write_full_page(page,
2934                                                 ext4_normal_get_block_write,
2935                                                 wbc);
2936 }
2937
2938 static int ext4_normal_writepage(struct page *page,
2939                                 struct writeback_control *wbc)
2940 {
2941         struct inode *inode = page->mapping->host;
2942         loff_t size = i_size_read(inode);
2943         loff_t len;
2944
2945         trace_mark(ext4_normal_writepage,
2946                    "dev %s ino %lu page_index %lu",
2947                    inode->i_sb->s_id, inode->i_ino, page->index);
2948         J_ASSERT(PageLocked(page));
2949         if (page->index == size >> PAGE_CACHE_SHIFT)
2950                 len = size & ~PAGE_CACHE_MASK;
2951         else
2952                 len = PAGE_CACHE_SIZE;
2953
2954         if (page_has_buffers(page)) {
2955                 /* if page has buffers it should all be mapped
2956                  * and allocated. If there are not buffers attached
2957                  * to the page we know the page is dirty but it lost
2958                  * buffers. That means that at some moment in time
2959                  * after write_begin() / write_end() has been called
2960                  * all buffers have been clean and thus they must have been
2961                  * written at least once. So they are all mapped and we can
2962                  * happily proceed with mapping them and writing the page.
2963                  */
2964                 BUG_ON(walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
2965                                         ext4_bh_unmapped_or_delay));
2966         }
2967
2968         if (!ext4_journal_current_handle())
2969                 return __ext4_normal_writepage(page, wbc);
2970
2971         redirty_page_for_writepage(wbc, page);
2972         unlock_page(page);
2973         return 0;
2974 }
2975
2976 static int __ext4_journalled_writepage(struct page *page,
2977                                 struct writeback_control *wbc)
2978 {
2979         struct address_space *mapping = page->mapping;
2980         struct inode *inode = mapping->host;
2981         struct buffer_head *page_bufs;
2982         handle_t *handle = NULL;
2983         int ret = 0;
2984         int err;
2985
2986         ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
2987                                         ext4_normal_get_block_write);
2988         if (ret != 0)
2989                 goto out_unlock;
2990
2991         page_bufs = page_buffers(page);
2992         walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE, NULL,
2993                                                                 bget_one);
2994         /* As soon as we unlock the page, it can go away, but we have
2995          * references to buffers so we are safe */
2996         unlock_page(page);
2997
2998         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2999         if (IS_ERR(handle)) {
3000                 ret = PTR_ERR(handle);
3001                 goto out;
3002         }
3003
3004         ret = walk_page_buffers(handle, page_bufs, 0,
3005                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
3006
3007         err = walk_page_buffers(handle, page_bufs, 0,
3008                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
3009         if (ret == 0)
3010                 ret = err;
3011         err = ext4_journal_stop(handle);
3012         if (!ret)
3013                 ret = err;
3014
3015         walk_page_buffers(handle, page_bufs, 0,
3016                                 PAGE_CACHE_SIZE, NULL, bput_one);
3017         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
3018         goto out;
3019
3020 out_unlock:
3021         unlock_page(page);
3022 out:
3023         return ret;
3024 }
3025
3026 static int ext4_journalled_writepage(struct page *page,
3027                                 struct writeback_control *wbc)
3028 {
3029         struct inode *inode = page->mapping->host;
3030         loff_t size = i_size_read(inode);
3031         loff_t len;
3032
3033         trace_mark(ext4_journalled_writepage,
3034                    "dev %s ino %lu page_index %lu",
3035                    inode->i_sb->s_id, inode->i_ino, page->index);
3036         J_ASSERT(PageLocked(page));
3037         if (page->index == size >> PAGE_CACHE_SHIFT)
3038                 len = size & ~PAGE_CACHE_MASK;
3039         else
3040                 len = PAGE_CACHE_SIZE;
3041
3042         if (page_has_buffers(page)) {
3043                 /* if page has buffers it should all be mapped
3044                  * and allocated. If there are not buffers attached
3045                  * to the page we know the page is dirty but it lost
3046                  * buffers. That means that at some moment in time
3047                  * after write_begin() / write_end() has been called
3048                  * all buffers have been clean and thus they must have been
3049                  * written at least once. So they are all mapped and we can
3050                  * happily proceed with mapping them and writing the page.
3051                  */
3052                 BUG_ON(walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
3053                                         ext4_bh_unmapped_or_delay));
3054         }
3055
3056         if (ext4_journal_current_handle())
3057                 goto no_write;
3058
3059         if (PageChecked(page)) {
3060                 /*
3061                  * It's mmapped pagecache.  Add buffers and journal it.  There
3062                  * doesn't seem much point in redirtying the page here.
3063                  */
3064                 ClearPageChecked(page);
3065                 return __ext4_journalled_writepage(page, wbc);
3066         } else {
3067                 /*
3068                  * It may be a page full of checkpoint-mode buffers.  We don't
3069                  * really know unless we go poke around in the buffer_heads.
3070                  * But block_write_full_page will do the right thing.
3071                  */
3072                 return block_write_full_page(page,
3073                                                 ext4_normal_get_block_write,
3074                                                 wbc);
3075         }
3076 no_write:
3077         redirty_page_for_writepage(wbc, page);
3078         unlock_page(page);
3079         return 0;
3080 }
3081
3082 static int ext4_readpage(struct file *file, struct page *page)
3083 {
3084         return mpage_readpage(page, ext4_get_block);
3085 }
3086
3087 static int
3088 ext4_readpages(struct file *file, struct address_space *mapping,
3089                 struct list_head *pages, unsigned nr_pages)
3090 {
3091         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3092 }
3093
3094 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3095 {
3096         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3097
3098         /*
3099          * If it's a full truncate we just forget about the pending dirtying
3100          */
3101         if (offset == 0)
3102                 ClearPageChecked(page);
3103
3104         if (journal)
3105                 jbd2_journal_invalidatepage(journal, page, offset);
3106         else
3107                 block_invalidatepage(page, offset);
3108 }
3109
3110 static int ext4_releasepage(struct page *page, gfp_t wait)
3111 {
3112         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3113
3114         WARN_ON(PageChecked(page));
3115         if (!page_has_buffers(page))
3116                 return 0;
3117         if (journal)
3118                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3119         else
3120                 return try_to_free_buffers(page);
3121 }
3122
3123 /*
3124  * If the O_DIRECT write will extend the file then add this inode to the
3125  * orphan list.  So recovery will truncate it back to the original size
3126  * if the machine crashes during the write.
3127  *
3128  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3129  * crashes then stale disk data _may_ be exposed inside the file. But current
3130  * VFS code falls back into buffered path in that case so we are safe.
3131  */
3132 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3133                         const struct iovec *iov, loff_t offset,
3134                         unsigned long nr_segs)
3135 {
3136         struct file *file = iocb->ki_filp;
3137         struct inode *inode = file->f_mapping->host;
3138         struct ext4_inode_info *ei = EXT4_I(inode);
3139         handle_t *handle;
3140         ssize_t ret;
3141         int orphan = 0;
3142         size_t count = iov_length(iov, nr_segs);
3143
3144         if (rw == WRITE) {
3145                 loff_t final_size = offset + count;
3146
3147                 if (final_size > inode->i_size) {
3148                         /* Credits for sb + inode write */
3149                         handle = ext4_journal_start(inode, 2);
3150                         if (IS_ERR(handle)) {
3151                                 ret = PTR_ERR(handle);
3152                                 goto out;
3153                         }
3154                         ret = ext4_orphan_add(handle, inode);
3155                         if (ret) {
3156                                 ext4_journal_stop(handle);
3157                                 goto out;
3158                         }
3159                         orphan = 1;
3160                         ei->i_disksize = inode->i_size;
3161                         ext4_journal_stop(handle);
3162                 }
3163         }
3164
3165         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
3166                                  offset, nr_segs,
3167                                  ext4_get_block, NULL);
3168
3169         if (orphan) {
3170                 int err;
3171
3172                 /* Credits for sb + inode write */
3173                 handle = ext4_journal_start(inode, 2);
3174                 if (IS_ERR(handle)) {
3175                         /* This is really bad luck. We've written the data
3176                          * but cannot extend i_size. Bail out and pretend
3177                          * the write failed... */
3178                         ret = PTR_ERR(handle);
3179                         goto out;
3180                 }
3181                 if (inode->i_nlink)
3182                         ext4_orphan_del(handle, inode);
3183                 if (ret > 0) {
3184                         loff_t end = offset + ret;
3185                         if (end > inode->i_size) {
3186                                 ei->i_disksize = end;
3187                                 i_size_write(inode, end);
3188                                 /*
3189                                  * We're going to return a positive `ret'
3190                                  * here due to non-zero-length I/O, so there's
3191                                  * no way of reporting error returns from
3192                                  * ext4_mark_inode_dirty() to userspace.  So
3193                                  * ignore it.
3194                                  */
3195                                 ext4_mark_inode_dirty(handle, inode);
3196                         }
3197                 }
3198                 err = ext4_journal_stop(handle);
3199                 if (ret == 0)
3200                         ret = err;
3201         }
3202 out:
3203         return ret;
3204 }
3205
3206 /*
3207  * Pages can be marked dirty completely asynchronously from ext4's journalling
3208  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3209  * much here because ->set_page_dirty is called under VFS locks.  The page is
3210  * not necessarily locked.
3211  *
3212  * We cannot just dirty the page and leave attached buffers clean, because the
3213  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3214  * or jbddirty because all the journalling code will explode.
3215  *
3216  * So what we do is to mark the page "pending dirty" and next time writepage
3217  * is called, propagate that into the buffers appropriately.
3218  */
3219 static int ext4_journalled_set_page_dirty(struct page *page)
3220 {
3221         SetPageChecked(page);
3222         return __set_page_dirty_nobuffers(page);
3223 }
3224
3225 static const struct address_space_operations ext4_ordered_aops = {
3226         .readpage               = ext4_readpage,
3227         .readpages              = ext4_readpages,
3228         .writepage              = ext4_normal_writepage,
3229         .sync_page              = block_sync_page,
3230         .write_begin            = ext4_write_begin,
3231         .write_end              = ext4_ordered_write_end,
3232         .bmap                   = ext4_bmap,
3233         .invalidatepage         = ext4_invalidatepage,
3234         .releasepage            = ext4_releasepage,
3235         .direct_IO              = ext4_direct_IO,
3236         .migratepage            = buffer_migrate_page,
3237         .is_partially_uptodate  = block_is_partially_uptodate,
3238 };
3239
3240 static const struct address_space_operations ext4_writeback_aops = {
3241         .readpage               = ext4_readpage,
3242         .readpages              = ext4_readpages,
3243         .writepage              = ext4_normal_writepage,
3244         .sync_page              = block_sync_page,
3245         .write_begin            = ext4_write_begin,
3246         .write_end              = ext4_writeback_write_end,
3247         .bmap                   = ext4_bmap,
3248         .invalidatepage         = ext4_invalidatepage,
3249         .releasepage            = ext4_releasepage,
3250         .direct_IO              = ext4_direct_IO,
3251         .migratepage            = buffer_migrate_page,
3252         .is_partially_uptodate  = block_is_partially_uptodate,
3253 };
3254
3255 static const struct address_space_operations ext4_journalled_aops = {
3256         .readpage               = ext4_readpage,
3257         .readpages              = ext4_readpages,
3258         .writepage              = ext4_journalled_writepage,
3259         .sync_page              = block_sync_page,
3260         .write_begin            = ext4_write_begin,
3261         .write_end              = ext4_journalled_write_end,
3262         .set_page_dirty         = ext4_journalled_set_page_dirty,
3263         .bmap                   = ext4_bmap,
3264         .invalidatepage         = ext4_invalidatepage,
3265         .releasepage            = ext4_releasepage,
3266         .is_partially_uptodate  = block_is_partially_uptodate,
3267 };
3268
3269 static const struct address_space_operations ext4_da_aops = {
3270         .readpage               = ext4_readpage,
3271         .readpages              = ext4_readpages,
3272         .writepage              = ext4_da_writepage,
3273         .writepages             = ext4_da_writepages,
3274         .sync_page              = block_sync_page,
3275         .write_begin            = ext4_da_write_begin,
3276         .write_end              = ext4_da_write_end,
3277         .bmap                   = ext4_bmap,
3278         .invalidatepage         = ext4_da_invalidatepage,
3279         .releasepage            = ext4_releasepage,
3280         .direct_IO              = ext4_direct_IO,
3281         .migratepage            = buffer_migrate_page,
3282         .is_partially_uptodate  = block_is_partially_uptodate,
3283 };
3284
3285 void ext4_set_aops(struct inode *inode)
3286 {
3287         if (ext4_should_order_data(inode) &&
3288                 test_opt(inode->i_sb, DELALLOC))
3289                 inode->i_mapping->a_ops = &ext4_da_aops;
3290         else if (ext4_should_order_data(inode))
3291                 inode->i_mapping->a_ops = &ext4_ordered_aops;
3292         else if (ext4_should_writeback_data(inode) &&
3293                  test_opt(inode->i_sb, DELALLOC))
3294                 inode->i_mapping->a_ops = &ext4_da_aops;
3295         else if (ext4_should_writeback_data(inode))
3296                 inode->i_mapping->a_ops = &ext4_writeback_aops;
3297         else
3298                 inode->i_mapping->a_ops = &ext4_journalled_aops;
3299 }
3300
3301 /*
3302  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3303  * up to the end of the block which corresponds to `from'.
3304  * This required during truncate. We need to physically zero the tail end
3305  * of that block so it doesn't yield old data if the file is later grown.
3306  */
3307 int ext4_block_truncate_page(handle_t *handle,
3308                 struct address_space *mapping, loff_t from)
3309 {
3310         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3311         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3312         unsigned blocksize, length, pos;
3313         ext4_lblk_t iblock;
3314         struct inode *inode = mapping->host;
3315         struct buffer_head *bh;
3316         struct page *page;
3317         int err = 0;
3318
3319         page = grab_cache_page(mapping, from >> PAGE_CACHE_SHIFT);
3320         if (!page)
3321                 return -EINVAL;
3322
3323         blocksize = inode->i_sb->s_blocksize;
3324         length = blocksize - (offset & (blocksize - 1));
3325         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3326
3327         /*
3328          * For "nobh" option,  we can only work if we don't need to
3329          * read-in the page - otherwise we create buffers to do the IO.
3330          */
3331         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
3332              ext4_should_writeback_data(inode) && PageUptodate(page)) {
3333                 zero_user(page, offset, length);
3334                 set_page_dirty(page);
3335                 goto unlock;
3336         }
3337
3338         if (!page_has_buffers(page))
3339                 create_empty_buffers(page, blocksize, 0);
3340
3341         /* Find the buffer that contains "offset" */
3342         bh = page_buffers(page);
3343         pos = blocksize;
3344         while (offset >= pos) {
3345                 bh = bh->b_this_page;
3346                 iblock++;
3347                 pos += blocksize;
3348         }
3349
3350         err = 0;
3351         if (buffer_freed(bh)) {
3352                 BUFFER_TRACE(bh, "freed: skip");
3353                 goto unlock;
3354         }
3355
3356         if (!buffer_mapped(bh)) {
3357                 BUFFER_TRACE(bh, "unmapped");
3358                 ext4_get_block(inode, iblock, bh, 0);
3359                 /* unmapped? It's a hole - nothing to do */
3360                 if (!buffer_mapped(bh)) {
3361                         BUFFER_TRACE(bh, "still unmapped");
3362                         goto unlock;
3363                 }
3364         }
3365
3366         /* Ok, it's mapped. Make sure it's up-to-date */
3367         if (PageUptodate(page))
3368                 set_buffer_uptodate(bh);
3369
3370         if (!buffer_uptodate(bh)) {
3371                 err = -EIO;
3372                 ll_rw_block(READ, 1, &bh);
3373                 wait_on_buffer(bh);
3374                 /* Uhhuh. Read error. Complain and punt. */
3375                 if (!buffer_uptodate(bh))
3376                         goto unlock;
3377         }
3378
3379         if (ext4_should_journal_data(inode)) {
3380                 BUFFER_TRACE(bh, "get write access");
3381                 err = ext4_journal_get_write_access(handle, bh);
3382                 if (err)
3383                         goto unlock;
3384         }
3385
3386         zero_user(page, offset, length);
3387
3388         BUFFER_TRACE(bh, "zeroed end of block");
3389
3390         err = 0;
3391         if (ext4_should_journal_data(inode)) {
3392                 err = ext4_handle_dirty_metadata(handle, inode, bh);
3393         } else {
3394                 if (ext4_should_order_data(inode))
3395                         err = ext4_jbd2_file_inode(handle, inode);
3396                 mark_buffer_dirty(bh);
3397         }
3398
3399 unlock:
3400         unlock_page(page);
3401         page_cache_release(page);
3402         return err;
3403 }
3404
3405 /*
3406  * Probably it should be a library function... search for first non-zero word
3407  * or memcmp with zero_page, whatever is better for particular architecture.
3408  * Linus?
3409  */
3410 static inline int all_zeroes(__le32 *p, __le32 *q)
3411 {
3412         while (p < q)
3413                 if (*p++)
3414                         return 0;
3415         return 1;
3416 }
3417
3418 /**
3419  *      ext4_find_shared - find the indirect blocks for partial truncation.
3420  *      @inode:   inode in question
3421  *      @depth:   depth of the affected branch
3422  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
3423  *      @chain:   place to store the pointers to partial indirect blocks
3424  *      @top:     place to the (detached) top of branch
3425  *
3426  *      This is a helper function used by ext4_truncate().
3427  *
3428  *      When we do truncate() we may have to clean the ends of several
3429  *      indirect blocks but leave the blocks themselves alive. Block is
3430  *      partially truncated if some data below the new i_size is refered
3431  *      from it (and it is on the path to the first completely truncated
3432  *      data block, indeed).  We have to free the top of that path along
3433  *      with everything to the right of the path. Since no allocation
3434  *      past the truncation point is possible until ext4_truncate()
3435  *      finishes, we may safely do the latter, but top of branch may
3436  *      require special attention - pageout below the truncation point
3437  *      might try to populate it.
3438  *
3439  *      We atomically detach the top of branch from the tree, store the
3440  *      block number of its root in *@top, pointers to buffer_heads of
3441  *      partially truncated blocks - in @chain[].bh and pointers to
3442  *      their last elements that should not be removed - in
3443  *      @chain[].p. Return value is the pointer to last filled element
3444  *      of @chain.
3445  *
3446  *      The work left to caller to do the actual freeing of subtrees:
3447  *              a) free the subtree starting from *@top
3448  *              b) free the subtrees whose roots are stored in
3449  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
3450  *              c) free the subtrees growing from the inode past the @chain[0].
3451  *                      (no partially truncated stuff there).  */
3452
3453 static Indirect *ext4_find_shared(struct inode *inode, int depth,
3454                         ext4_lblk_t offsets[4], Indirect chain[4], __le32 *top)
3455 {
3456         Indirect *partial, *p;
3457         int k, err;
3458
3459         *top = 0;
3460         /* Make k index the deepest non-null offest + 1 */
3461         for (k = depth; k > 1 && !offsets[k-1]; k--)
3462                 ;
3463         partial = ext4_get_branch(inode, k, offsets, chain, &err);
3464         /* Writer: pointers */
3465         if (!partial)
3466                 partial = chain + k-1;
3467         /*
3468          * If the branch acquired continuation since we've looked at it -
3469          * fine, it should all survive and (new) top doesn't belong to us.
3470          */
3471         if (!partial->key && *partial->p)
3472                 /* Writer: end */
3473                 goto no_top;
3474         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
3475                 ;
3476         /*
3477          * OK, we've found the last block that must survive. The rest of our
3478          * branch should be detached before unlocking. However, if that rest
3479          * of branch is all ours and does not grow immediately from the inode
3480          * it's easier to cheat and just decrement partial->p.
3481          */
3482         if (p == chain + k - 1 && p > chain) {
3483                 p->p--;
3484         } else {
3485                 *top = *p->p;
3486                 /* Nope, don't do this in ext4.  Must leave the tree intact */
3487 #if 0
3488                 *p->p = 0;
3489 #endif
3490         }
3491         /* Writer: end */
3492
3493         while (partial > p) {
3494                 brelse(partial->bh);
3495                 partial--;
3496         }
3497 no_top:
3498         return partial;
3499 }
3500
3501 /*
3502  * Zero a number of block pointers in either an inode or an indirect block.
3503  * If we restart the transaction we must again get write access to the
3504  * indirect block for further modification.
3505  *
3506  * We release `count' blocks on disk, but (last - first) may be greater
3507  * than `count' because there can be holes in there.
3508  */
3509 static void ext4_clear_blocks(handle_t *handle, struct inode *inode,
3510                 struct buffer_head *bh, ext4_fsblk_t block_to_free,
3511                 unsigned long count, __le32 *first, __le32 *last)
3512 {
3513         __le32 *p;
3514         if (try_to_extend_transaction(handle, inode)) {
3515                 if (bh) {
3516                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
3517                         ext4_handle_dirty_metadata(handle, inode, bh);
3518                 }
3519                 ext4_mark_inode_dirty(handle, inode);
3520                 ext4_journal_test_restart(handle, inode);
3521                 if (bh) {
3522                         BUFFER_TRACE(bh, "retaking write access");
3523                         ext4_journal_get_write_access(handle, bh);
3524                 }
3525         }
3526
3527         /*
3528          * Any buffers which are on the journal will be in memory. We find
3529          * them on the hash table so jbd2_journal_revoke() will run jbd2_journal_forget()
3530          * on them.  We've already detached each block from the file, so
3531          * bforget() in jbd2_journal_forget() should be safe.
3532          *
3533          * AKPM: turn on bforget in jbd2_journal_forget()!!!
3534          */
3535         for (p = first; p < last; p++) {
3536                 u32 nr = le32_to_cpu(*p);
3537                 if (nr) {
3538                         struct buffer_head *tbh;
3539
3540                         *p = 0;
3541                         tbh = sb_find_get_block(inode->i_sb, nr);
3542                         ext4_forget(handle, 0, inode, tbh, nr);
3543                 }
3544         }
3545
3546         ext4_free_blocks(handle, inode, block_to_free, count, 0);
3547 }
3548
3549 /**
3550  * ext4_free_data - free a list of data blocks
3551  * @handle:     handle for this transaction
3552  * @inode:      inode we are dealing with
3553  * @this_bh:    indirect buffer_head which contains *@first and *@last
3554  * @first:      array of block numbers
3555  * @last:       points immediately past the end of array
3556  *
3557  * We are freeing all blocks refered from that array (numbers are stored as
3558  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
3559  *
3560  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
3561  * blocks are contiguous then releasing them at one time will only affect one
3562  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
3563  * actually use a lot of journal space.
3564  *
3565  * @this_bh will be %NULL if @first and @last point into the inode's direct
3566  * block pointers.
3567  */
3568 static void ext4_free_data(handle_t *handle, struct inode *inode,
3569                            struct buffer_head *this_bh,
3570                            __le32 *first, __le32 *last)
3571 {
3572         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
3573         unsigned long count = 0;            /* Number of blocks in the run */
3574         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
3575                                                corresponding to
3576                                                block_to_free */
3577         ext4_fsblk_t nr;                    /* Current block # */
3578         __le32 *p;                          /* Pointer into inode/ind
3579                                                for current block */
3580         int err;
3581
3582         if (this_bh) {                          /* For indirect block */
3583                 BUFFER_TRACE(this_bh, "get_write_access");
3584                 err = ext4_journal_get_write_access(handle, this_bh);
3585                 /* Important: if we can't update the indirect pointers
3586                  * to the blocks, we can't free them. */
3587                 if (err)
3588                         return;
3589         }
3590
3591         for (p = first; p < last; p++) {
3592                 nr = le32_to_cpu(*p);
3593                 if (nr) {
3594                         /* accumulate blocks to free if they're contiguous */
3595                         if (count == 0) {
3596                                 block_to_free = nr;
3597                                 block_to_free_p = p;
3598                                 count = 1;
3599                         } else if (nr == block_to_free + count) {
3600                                 count++;
3601                         } else {
3602                                 ext4_clear_blocks(handle, inode, this_bh,
3603                                                   block_to_free,
3604                                                   count, block_to_free_p, p);
3605                                 block_to_free = nr;
3606                                 block_to_free_p = p;
3607                                 count = 1;
3608                         }
3609                 }
3610         }
3611
3612         if (count > 0)
3613                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
3614                                   count, block_to_free_p, p);
3615
3616         if (this_bh) {
3617                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
3618
3619                 /*
3620                  * The buffer head should have an attached journal head at this
3621                  * point. However, if the data is corrupted and an indirect
3622                  * block pointed to itself, it would have been detached when
3623                  * the block was cleared. Check for this instead of OOPSing.
3624                  */
3625                 if (bh2jh(this_bh))
3626                         ext4_handle_dirty_metadata(handle, inode, this_bh);
3627                 else
3628                         ext4_error(inode->i_sb, __func__,
3629                                    "circular indirect block detected, "
3630                                    "inode=%lu, block=%llu",
3631                                    inode->i_ino,
3632                                    (unsigned long long) this_bh->b_blocknr);
3633         }
3634 }
3635
3636 /**
3637  *      ext4_free_branches - free an array of branches
3638  *      @handle: JBD handle for this transaction
3639  *      @inode: inode we are dealing with
3640  *      @parent_bh: the buffer_head which contains *@first and *@last
3641  *      @first: array of block numbers
3642  *      @last:  pointer immediately past the end of array
3643  *      @depth: depth of the branches to free
3644  *
3645  *      We are freeing all blocks refered from these branches (numbers are
3646  *      stored as little-endian 32-bit) and updating @inode->i_blocks
3647  *      appropriately.
3648  */
3649 static void ext4_free_branches(handle_t *handle, struct inode *inode,
3650                                struct buffer_head *parent_bh,
3651                                __le32 *first, __le32 *last, int depth)
3652 {
3653         ext4_fsblk_t nr;
3654         __le32 *p;
3655
3656         if (ext4_handle_is_aborted(handle))
3657                 return;
3658
3659         if (depth--) {
3660                 struct buffer_head *bh;
3661                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
3662                 p = last;
3663                 while (--p >= first) {
3664                         nr = le32_to_cpu(*p);
3665                         if (!nr)
3666                                 continue;               /* A hole */
3667
3668                         /* Go read the buffer for the next level down */
3669                         bh = sb_bread(inode->i_sb, nr);
3670
3671                         /*
3672                          * A read failure? Report error and clear slot
3673                          * (should be rare).
3674                          */
3675                         if (!bh) {
3676                                 ext4_error(inode->i_sb, "ext4_free_branches",
3677                                            "Read failure, inode=%lu, block=%llu",
3678                                            inode->i_ino, nr);
3679                                 continue;
3680                         }
3681
3682                         /* This zaps the entire block.  Bottom up. */
3683                         BUFFER_TRACE(bh, "free child branches");
3684                         ext4_free_branches(handle, inode, bh,
3685                                         (__le32 *) bh->b_data,
3686                                         (__le32 *) bh->b_data + addr_per_block,
3687                                         depth);
3688
3689                         /*
3690                          * We've probably journalled the indirect block several
3691                          * times during the truncate.  But it's no longer
3692                          * needed and we now drop it from the transaction via
3693                          * jbd2_journal_revoke().
3694                          *
3695                          * That's easy if it's exclusively part of this
3696                          * transaction.  But if it's part of the committing
3697                          * transaction then jbd2_journal_forget() will simply
3698                          * brelse() it.  That means that if the underlying
3699                          * block is reallocated in ext4_get_block(),
3700                          * unmap_underlying_metadata() will find this block
3701                          * and will try to get rid of it.  damn, damn.
3702                          *
3703                          * If this block has already been committed to the
3704                          * journal, a revoke record will be written.  And
3705                          * revoke records must be emitted *before* clearing
3706                          * this block's bit in the bitmaps.
3707                          */
3708                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
3709
3710                         /*
3711                          * Everything below this this pointer has been
3712                          * released.  Now let this top-of-subtree go.
3713                          *
3714                          * We want the freeing of this indirect block to be
3715                          * atomic in the journal with the updating of the
3716                          * bitmap block which owns it.  So make some room in
3717                          * the journal.
3718                          *
3719                          * We zero the parent pointer *after* freeing its
3720                          * pointee in the bitmaps, so if extend_transaction()
3721                          * for some reason fails to put the bitmap changes and
3722                          * the release into the same transaction, recovery
3723                          * will merely complain about releasing a free block,
3724                          * rather than leaking blocks.
3725                          */
3726                         if (ext4_handle_is_aborted(handle))
3727                                 return;
3728                         if (try_to_extend_transaction(handle, inode)) {
3729                                 ext4_mark_inode_dirty(handle, inode);
3730                                 ext4_journal_test_restart(handle, inode);
3731                         }
3732
3733                         ext4_free_blocks(handle, inode, nr, 1, 1);
3734
3735                         if (parent_bh) {
3736                                 /*
3737                                  * The block which we have just freed is
3738                                  * pointed to by an indirect block: journal it
3739                                  */
3740                                 BUFFER_TRACE(parent_bh, "get_write_access");
3741                                 if (!ext4_journal_get_write_access(handle,
3742                                                                    parent_bh)){
3743                                         *p = 0;
3744                                         BUFFER_TRACE(parent_bh,
3745                                         "call ext4_handle_dirty_metadata");
3746                                         ext4_handle_dirty_metadata(handle,
3747                                                                    inode,
3748                                                                    parent_bh);
3749                                 }
3750                         }
3751                 }
3752         } else {
3753                 /* We have reached the bottom of the tree. */
3754                 BUFFER_TRACE(parent_bh, "free data blocks");
3755                 ext4_free_data(handle, inode, parent_bh, first, last);
3756         }
3757 }
3758
3759 int ext4_can_truncate(struct inode *inode)
3760 {
3761         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
3762                 return 0;
3763         if (S_ISREG(inode->i_mode))
3764                 return 1;
3765         if (S_ISDIR(inode->i_mode))
3766                 return 1;
3767         if (S_ISLNK(inode->i_mode))
3768                 return !ext4_inode_is_fast_symlink(inode);
3769         return 0;
3770 }
3771
3772 /*
3773  * ext4_truncate()
3774  *
3775  * We block out ext4_get_block() block instantiations across the entire
3776  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
3777  * simultaneously on behalf of the same inode.
3778  *
3779  * As we work through the truncate and commmit bits of it to the journal there
3780  * is one core, guiding principle: the file's tree must always be consistent on
3781  * disk.  We must be able to restart the truncate after a crash.
3782  *
3783  * The file's tree may be transiently inconsistent in memory (although it
3784  * probably isn't), but whenever we close off and commit a journal transaction,
3785  * the contents of (the filesystem + the journal) must be consistent and
3786  * restartable.  It's pretty simple, really: bottom up, right to left (although
3787  * left-to-right works OK too).
3788  *
3789  * Note that at recovery time, journal replay occurs *before* the restart of
3790  * truncate against the orphan inode list.
3791  *
3792  * The committed inode has the new, desired i_size (which is the same as
3793  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
3794  * that this inode's truncate did not complete and it will again call
3795  * ext4_truncate() to have another go.  So there will be instantiated blocks
3796  * to the right of the truncation point in a crashed ext4 filesystem.  But
3797  * that's fine - as long as they are linked from the inode, the post-crash
3798  * ext4_truncate() run will find them and release them.
3799  */
3800 void ext4_truncate(struct inode *inode)
3801 {
3802         handle_t *handle;
3803         struct ext4_inode_info *ei = EXT4_I(inode);
3804         __le32 *i_data = ei->i_data;
3805         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
3806         struct address_space *mapping = inode->i_mapping;
3807         ext4_lblk_t offsets[4];
3808         Indirect chain[4];
3809         Indirect *partial;
3810         __le32 nr = 0;
3811         int n;
3812         ext4_lblk_t last_block;
3813         unsigned blocksize = inode->i_sb->s_blocksize;
3814
3815         if (!ext4_can_truncate(inode))
3816                 return;
3817
3818         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
3819                 ext4_ext_truncate(inode);
3820                 return;
3821         }
3822
3823         handle = start_transaction(inode);
3824         if (IS_ERR(handle))
3825                 return;         /* AKPM: return what? */
3826
3827         last_block = (inode->i_size + blocksize-1)
3828                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
3829
3830         if (inode->i_size & (blocksize - 1))
3831                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
3832                         goto out_stop;
3833
3834         n = ext4_block_to_path(inode, last_block, offsets, NULL);
3835         if (n == 0)
3836                 goto out_stop;  /* error */
3837
3838         /*
3839          * OK.  This truncate is going to happen.  We add the inode to the
3840          * orphan list, so that if this truncate spans multiple transactions,
3841          * and we crash, we will resume the truncate when the filesystem
3842          * recovers.  It also marks the inode dirty, to catch the new size.
3843          *
3844          * Implication: the file must always be in a sane, consistent
3845          * truncatable state while each transaction commits.
3846          */
3847         if (ext4_orphan_add(handle, inode))
3848                 goto out_stop;
3849
3850         /*
3851          * From here we block out all ext4_get_block() callers who want to
3852          * modify the block allocation tree.
3853          */
3854         down_write(&ei->i_data_sem);
3855
3856         ext4_discard_preallocations(inode);
3857
3858         /*
3859          * The orphan list entry will now protect us from any crash which
3860          * occurs before the truncate completes, so it is now safe to propagate
3861          * the new, shorter inode size (held for now in i_size) into the
3862          * on-disk inode. We do this via i_disksize, which is the value which
3863          * ext4 *really* writes onto the disk inode.
3864          */
3865         ei->i_disksize = inode->i_size;
3866
3867         if (n == 1) {           /* direct blocks */
3868                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
3869                                i_data + EXT4_NDIR_BLOCKS);
3870                 goto do_indirects;
3871         }
3872
3873         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
3874         /* Kill the top of shared branch (not detached) */
3875         if (nr) {
3876                 if (partial == chain) {
3877                         /* Shared branch grows from the inode */
3878                         ext4_free_branches(handle, inode, NULL,
3879                                            &nr, &nr+1, (chain+n-1) - partial);
3880                         *partial->p = 0;
3881                         /*
3882                          * We mark the inode dirty prior to restart,
3883                          * and prior to stop.  No need for it here.
3884                          */
3885                 } else {
3886                         /* Shared branch grows from an indirect block */
3887                         BUFFER_TRACE(partial->bh, "get_write_access");
3888                         ext4_free_branches(handle, inode, partial->bh,
3889                                         partial->p,
3890                                         partial->p+1, (chain+n-1) - partial);
3891                 }
3892         }
3893         /* Clear the ends of indirect blocks on the shared branch */
3894         while (partial > chain) {
3895                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
3896                                    (__le32*)partial->bh->b_data+addr_per_block,
3897                                    (chain+n-1) - partial);
3898                 BUFFER_TRACE(partial->bh, "call brelse");
3899                 brelse (partial->bh);
3900                 partial--;
3901         }
3902 do_indirects:
3903         /* Kill the remaining (whole) subtrees */
3904         switch (offsets[0]) {
3905         default:
3906                 nr = i_data[EXT4_IND_BLOCK];
3907                 if (nr) {
3908                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
3909                         i_data[EXT4_IND_BLOCK] = 0;
3910                 }
3911         case EXT4_IND_BLOCK:
3912                 nr = i_data[EXT4_DIND_BLOCK];
3913                 if (nr) {
3914                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
3915                         i_data[EXT4_DIND_BLOCK] = 0;
3916                 }
3917         case EXT4_DIND_BLOCK:
3918                 nr = i_data[EXT4_TIND_BLOCK];
3919                 if (nr) {
3920                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
3921                         i_data[EXT4_TIND_BLOCK] = 0;
3922                 }
3923         case EXT4_TIND_BLOCK:
3924                 ;
3925         }
3926
3927         up_write(&ei->i_data_sem);
3928         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
3929         ext4_mark_inode_dirty(handle, inode);
3930
3931         /*
3932          * In a multi-transaction truncate, we only make the final transaction
3933          * synchronous
3934          */
3935         if (IS_SYNC(inode))
3936                 ext4_handle_sync(handle);
3937 out_stop:
3938         /*
3939          * If this was a simple ftruncate(), and the file will remain alive
3940          * then we need to clear up the orphan record which we created above.
3941          * However, if this was a real unlink then we were called by
3942          * ext4_delete_inode(), and we allow that function to clean up the
3943          * orphan info for us.
3944          */
3945         if (inode->i_nlink)
3946                 ext4_orphan_del(handle, inode);
3947
3948         ext4_journal_stop(handle);
3949 }
3950
3951 /*
3952  * ext4_get_inode_loc returns with an extra refcount against the inode's
3953  * underlying buffer_head on success. If 'in_mem' is true, we have all
3954  * data in memory that is needed to recreate the on-disk version of this
3955  * inode.
3956  */
3957 static int __ext4_get_inode_loc(struct inode *inode,
3958                                 struct ext4_iloc *iloc, int in_mem)
3959 {
3960         struct ext4_group_desc  *gdp;
3961         struct buffer_head      *bh;
3962         struct super_block      *sb = inode->i_sb;
3963         ext4_fsblk_t            block;
3964         int                     inodes_per_block, inode_offset;
3965
3966         iloc->bh = NULL;
3967         if (!ext4_valid_inum(sb, inode->i_ino))
3968                 return -EIO;
3969
3970         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
3971         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
3972         if (!gdp)
3973                 return -EIO;
3974
3975         /*
3976          * Figure out the offset within the block group inode table
3977          */
3978         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
3979         inode_offset = ((inode->i_ino - 1) %
3980                         EXT4_INODES_PER_GROUP(sb));
3981         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
3982         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
3983
3984         bh = sb_getblk(sb, block);
3985         if (!bh) {
3986                 ext4_error(sb, "ext4_get_inode_loc", "unable to read "
3987                            "inode block - inode=%lu, block=%llu",
3988                            inode->i_ino, block);
3989                 return -EIO;
3990         }
3991         if (!buffer_uptodate(bh)) {
3992                 lock_buffer(bh);
3993
3994                 /*
3995                  * If the buffer has the write error flag, we have failed
3996                  * to write out another inode in the same block.  In this
3997                  * case, we don't have to read the block because we may
3998                  * read the old inode data successfully.
3999                  */
4000                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4001                         set_buffer_uptodate(bh);
4002
4003                 if (buffer_uptodate(bh)) {
4004                         /* someone brought it uptodate while we waited */
4005                         unlock_buffer(bh);
4006                         goto has_buffer;
4007                 }
4008
4009                 /*
4010                  * If we have all information of the inode in memory and this
4011                  * is the only valid inode in the block, we need not read the
4012                  * block.
4013                  */
4014                 if (in_mem) {
4015                         struct buffer_head *bitmap_bh;
4016                         int i, start;
4017
4018                         start = inode_offset & ~(inodes_per_block - 1);
4019
4020                         /* Is the inode bitmap in cache? */
4021                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4022                         if (!bitmap_bh)
4023                                 goto make_io;
4024
4025                         /*
4026                          * If the inode bitmap isn't in cache then the
4027                          * optimisation may end up performing two reads instead
4028                          * of one, so skip it.
4029                          */
4030                         if (!buffer_uptodate(bitmap_bh)) {
4031                                 brelse(bitmap_bh);
4032                                 goto make_io;
4033                         }
4034                         for (i = start; i < start + inodes_per_block; i++) {
4035                                 if (i == inode_offset)
4036                                         continue;
4037                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4038                                         break;
4039                         }
4040                         brelse(bitmap_bh);
4041                         if (i == start + inodes_per_block) {
4042                                 /* all other inodes are free, so skip I/O */
4043                                 memset(bh->b_data, 0, bh->b_size);
4044                                 set_buffer_uptodate(bh);
4045                                 unlock_buffer(bh);
4046                                 goto has_buffer;
4047                         }
4048                 }
4049
4050 make_io:
4051                 /*
4052                  * If we need to do any I/O, try to pre-readahead extra
4053                  * blocks from the inode table.
4054                  */
4055                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4056                         ext4_fsblk_t b, end, table;
4057                         unsigned num;
4058
4059                         table = ext4_inode_table(sb, gdp);
4060                         /* Make sure s_inode_readahead_blks is a power of 2 */
4061                         while (EXT4_SB(sb)->s_inode_readahead_blks &
4062                                (EXT4_SB(sb)->s_inode_readahead_blks-1))
4063                                 EXT4_SB(sb)->s_inode_readahead_blks = 
4064                                    (EXT4_SB(sb)->s_inode_readahead_blks &
4065                                     (EXT4_SB(sb)->s_inode_readahead_blks-1));
4066                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4067                         if (table > b)
4068                                 b = table;
4069                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4070                         num = EXT4_INODES_PER_GROUP(sb);
4071                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4072                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4073                                 num -= ext4_itable_unused_count(sb, gdp);
4074                         table += num / inodes_per_block;
4075                         if (end > table)
4076                                 end = table;
4077                         while (b <= end)
4078                                 sb_breadahead(sb, b++);
4079                 }
4080
4081                 /*
4082                  * There are other valid inodes in the buffer, this inode
4083                  * has in-inode xattrs, or we don't have this inode in memory.
4084                  * Read the block from disk.
4085                  */
4086                 get_bh(bh);
4087                 bh->b_end_io = end_buffer_read_sync;
4088                 submit_bh(READ_META, bh);
4089                 wait_on_buffer(bh);
4090                 if (!buffer_uptodate(bh)) {
4091                         ext4_error(sb, __func__,
4092                                    "unable to read inode block - inode=%lu, "
4093                                    "block=%llu", inode->i_ino, block);
4094                         brelse(bh);
4095                         return -EIO;
4096                 }
4097         }
4098 has_buffer:
4099         iloc->bh = bh;
4100         return 0;
4101 }
4102
4103 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4104 {
4105         /* We have all inode data except xattrs in memory here. */
4106         return __ext4_get_inode_loc(inode, iloc,
4107                 !(EXT4_I(inode)->i_state & EXT4_STATE_XATTR));
4108 }
4109
4110 void ext4_set_inode_flags(struct inode *inode)
4111 {
4112         unsigned int flags = EXT4_I(inode)->i_flags;
4113
4114         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4115         if (flags & EXT4_SYNC_FL)
4116                 inode->i_flags |= S_SYNC;
4117         if (flags & EXT4_APPEND_FL)
4118                 inode->i_flags |= S_APPEND;
4119         if (flags & EXT4_IMMUTABLE_FL)
4120                 inode->i_flags |= S_IMMUTABLE;
4121         if (flags & EXT4_NOATIME_FL)
4122                 inode->i_flags |= S_NOATIME;
4123         if (flags & EXT4_DIRSYNC_FL)
4124                 inode->i_flags |= S_DIRSYNC;
4125 }
4126
4127 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4128 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4129 {
4130         unsigned int flags = ei->vfs_inode.i_flags;
4131
4132         ei->i_flags &= ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4133                         EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|EXT4_DIRSYNC_FL);
4134         if (flags & S_SYNC)
4135                 ei->i_flags |= EXT4_SYNC_FL;
4136         if (flags & S_APPEND)
4137                 ei->i_flags |= EXT4_APPEND_FL;
4138         if (flags & S_IMMUTABLE)
4139                 ei->i_flags |= EXT4_IMMUTABLE_FL;
4140         if (flags & S_NOATIME)
4141                 ei->i_flags |= EXT4_NOATIME_FL;
4142         if (flags & S_DIRSYNC)
4143                 ei->i_flags |= EXT4_DIRSYNC_FL;
4144 }
4145 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4146                                         struct ext4_inode_info *ei)
4147 {
4148         blkcnt_t i_blocks ;
4149         struct inode *inode = &(ei->vfs_inode);
4150         struct super_block *sb = inode->i_sb;
4151
4152         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4153                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4154                 /* we are using combined 48 bit field */
4155                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4156                                         le32_to_cpu(raw_inode->i_blocks_lo);
4157                 if (ei->i_flags & EXT4_HUGE_FILE_FL) {
4158                         /* i_blocks represent file system block size */
4159                         return i_blocks  << (inode->i_blkbits - 9);
4160                 } else {
4161                         return i_blocks;
4162                 }
4163         } else {
4164                 return le32_to_cpu(raw_inode->i_blocks_lo);
4165         }
4166 }
4167
4168 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4169 {
4170         struct ext4_iloc iloc;
4171         struct ext4_inode *raw_inode;
4172         struct ext4_inode_info *ei;
4173         struct buffer_head *bh;
4174         struct inode *inode;
4175         long ret;
4176         int block;
4177
4178         inode = iget_locked(sb, ino);
4179         if (!inode)
4180                 return ERR_PTR(-ENOMEM);
4181         if (!(inode->i_state & I_NEW))
4182                 return inode;
4183
4184         ei = EXT4_I(inode);
4185 #ifdef CONFIG_EXT4_FS_POSIX_ACL
4186         ei->i_acl = EXT4_ACL_NOT_CACHED;
4187         ei->i_default_acl = EXT4_ACL_NOT_CACHED;
4188 #endif
4189
4190         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4191         if (ret < 0)
4192                 goto bad_inode;
4193         bh = iloc.bh;
4194         raw_inode = ext4_raw_inode(&iloc);
4195         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4196         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4197         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4198         if (!(test_opt(inode->i_sb, NO_UID32))) {
4199                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4200                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4201         }
4202         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
4203
4204         ei->i_state = 0;
4205         ei->i_dir_start_lookup = 0;
4206         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4207         /* We now have enough fields to check if the inode was active or not.
4208          * This is needed because nfsd might try to access dead inodes
4209          * the test is that same one that e2fsck uses
4210          * NeilBrown 1999oct15
4211          */
4212         if (inode->i_nlink == 0) {
4213                 if (inode->i_mode == 0 ||
4214                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
4215                         /* this inode is deleted */
4216                         brelse(bh);
4217                         ret = -ESTALE;
4218                         goto bad_inode;
4219                 }
4220                 /* The only unlinked inodes we let through here have
4221                  * valid i_mode and are being read by the orphan
4222                  * recovery code: that's fine, we're about to complete
4223                  * the process of deleting those. */
4224         }
4225         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4226         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4227         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4228         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
4229             cpu_to_le32(EXT4_OS_HURD)) {
4230                 ei->i_file_acl |=
4231                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4232         }
4233         inode->i_size = ext4_isize(raw_inode);
4234         ei->i_disksize = inode->i_size;
4235         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4236         ei->i_block_group = iloc.block_group;
4237         /*
4238          * NOTE! The in-memory inode i_data array is in little-endian order
4239          * even on big-endian machines: we do NOT byteswap the block numbers!
4240          */
4241         for (block = 0; block < EXT4_N_BLOCKS; block++)
4242                 ei->i_data[block] = raw_inode->i_block[block];
4243         INIT_LIST_HEAD(&ei->i_orphan);
4244
4245         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4246                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4247                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4248                     EXT4_INODE_SIZE(inode->i_sb)) {
4249                         brelse(bh);
4250                         ret = -EIO;
4251                         goto bad_inode;
4252                 }
4253                 if (ei->i_extra_isize == 0) {
4254                         /* The extra space is currently unused. Use it. */
4255                         ei->i_extra_isize = sizeof(struct ext4_inode) -
4256                                             EXT4_GOOD_OLD_INODE_SIZE;
4257                 } else {
4258                         __le32 *magic = (void *)raw_inode +
4259                                         EXT4_GOOD_OLD_INODE_SIZE +
4260                                         ei->i_extra_isize;
4261                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
4262                                  ei->i_state |= EXT4_STATE_XATTR;
4263                 }
4264         } else
4265                 ei->i_extra_isize = 0;
4266
4267         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4268         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4269         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4270         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4271
4272         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4273         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4274                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4275                         inode->i_version |=
4276                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4277         }
4278
4279         if (S_ISREG(inode->i_mode)) {
4280                 inode->i_op = &ext4_file_inode_operations;
4281                 inode->i_fop = &ext4_file_operations;
4282                 ext4_set_aops(inode);
4283         } else if (S_ISDIR(inode->i_mode)) {
4284                 inode->i_op = &ext4_dir_inode_operations;
4285                 inode->i_fop = &ext4_dir_operations;
4286         } else if (S_ISLNK(inode->i_mode)) {
4287                 if (ext4_inode_is_fast_symlink(inode)) {
4288                         inode->i_op = &ext4_fast_symlink_inode_operations;
4289                         nd_terminate_link(ei->i_data, inode->i_size,
4290                                 sizeof(ei->i_data) - 1);
4291                 } else {
4292                         inode->i_op = &ext4_symlink_inode_operations;
4293                         ext4_set_aops(inode);
4294                 }
4295         } else {
4296                 inode->i_op = &ext4_special_inode_operations;
4297                 if (raw_inode->i_block[0])
4298                         init_special_inode(inode, inode->i_mode,
4299                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4300                 else
4301                         init_special_inode(inode, inode->i_mode,
4302                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4303         }
4304         brelse(iloc.bh);
4305         ext4_set_inode_flags(inode);
4306         unlock_new_inode(inode);
4307         return inode;
4308
4309 bad_inode:
4310         iget_failed(inode);
4311         return ERR_PTR(ret);
4312 }
4313
4314 static int ext4_inode_blocks_set(handle_t *handle,
4315                                 struct ext4_inode *raw_inode,
4316                                 struct ext4_inode_info *ei)
4317 {
4318         struct inode *inode = &(ei->vfs_inode);
4319         u64 i_blocks = inode->i_blocks;
4320         struct super_block *sb = inode->i_sb;
4321
4322         if (i_blocks <= ~0U) {
4323                 /*
4324                  * i_blocks can be represnted in a 32 bit variable
4325                  * as multiple of 512 bytes
4326                  */
4327                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4328                 raw_inode->i_blocks_high = 0;
4329                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4330                 return 0;
4331         }
4332         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
4333                 return -EFBIG;
4334
4335         if (i_blocks <= 0xffffffffffffULL) {
4336                 /*
4337                  * i_blocks can be represented in a 48 bit variable
4338                  * as multiple of 512 bytes
4339                  */
4340                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4341                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4342                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4343         } else {
4344                 ei->i_flags |= EXT4_HUGE_FILE_FL;
4345                 /* i_block is stored in file system block size */
4346                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4347                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4348                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4349         }
4350         return 0;
4351 }
4352
4353 /*
4354  * Post the struct inode info into an on-disk inode location in the
4355  * buffer-cache.  This gobbles the caller's reference to the
4356  * buffer_head in the inode location struct.
4357  *
4358  * The caller must have write access to iloc->bh.
4359  */
4360 static int ext4_do_update_inode(handle_t *handle,
4361                                 struct inode *inode,
4362                                 struct ext4_iloc *iloc)
4363 {
4364         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
4365         struct ext4_inode_info *ei = EXT4_I(inode);
4366         struct buffer_head *bh = iloc->bh;
4367         int err = 0, rc, block;
4368
4369         /* For fields not not tracking in the in-memory inode,
4370          * initialise them to zero for new inodes. */
4371         if (ei->i_state & EXT4_STATE_NEW)
4372                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
4373
4374         ext4_get_inode_flags(ei);
4375         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
4376         if (!(test_opt(inode->i_sb, NO_UID32))) {
4377                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
4378                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
4379 /*
4380  * Fix up interoperability with old kernels. Otherwise, old inodes get
4381  * re-used with the upper 16 bits of the uid/gid intact
4382  */
4383                 if (!ei->i_dtime) {
4384                         raw_inode->i_uid_high =
4385                                 cpu_to_le16(high_16_bits(inode->i_uid));
4386                         raw_inode->i_gid_high =
4387                                 cpu_to_le16(high_16_bits(inode->i_gid));
4388                 } else {
4389                         raw_inode->i_uid_high = 0;
4390                         raw_inode->i_gid_high = 0;
4391                 }
4392         } else {
4393                 raw_inode->i_uid_low =
4394                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
4395                 raw_inode->i_gid_low =
4396                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
4397                 raw_inode->i_uid_high = 0;
4398                 raw_inode->i_gid_high = 0;
4399         }
4400         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
4401
4402         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
4403         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
4404         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
4405         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
4406
4407         if (ext4_inode_blocks_set(handle, raw_inode, ei))
4408                 goto out_brelse;
4409         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
4410         /* clear the migrate flag in the raw_inode */
4411         raw_inode->i_flags = cpu_to_le32(ei->i_flags & ~EXT4_EXT_MIGRATE);
4412         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
4413             cpu_to_le32(EXT4_OS_HURD))
4414                 raw_inode->i_file_acl_high =
4415                         cpu_to_le16(ei->i_file_acl >> 32);
4416         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
4417         ext4_isize_set(raw_inode, ei->i_disksize);
4418         if (ei->i_disksize > 0x7fffffffULL) {
4419                 struct super_block *sb = inode->i_sb;
4420                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
4421                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
4422                                 EXT4_SB(sb)->s_es->s_rev_level ==
4423                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
4424                         /* If this is the first large file
4425                          * created, add a flag to the superblock.
4426                          */
4427                         err = ext4_journal_get_write_access(handle,
4428                                         EXT4_SB(sb)->s_sbh);
4429                         if (err)
4430                                 goto out_brelse;
4431                         ext4_update_dynamic_rev(sb);
4432                         EXT4_SET_RO_COMPAT_FEATURE(sb,
4433                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
4434                         sb->s_dirt = 1;
4435                         ext4_handle_sync(handle);
4436                         err = ext4_handle_dirty_metadata(handle, inode,
4437                                         EXT4_SB(sb)->s_sbh);
4438                 }
4439         }
4440         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
4441         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
4442                 if (old_valid_dev(inode->i_rdev)) {
4443                         raw_inode->i_block[0] =
4444                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
4445                         raw_inode->i_block[1] = 0;
4446                 } else {
4447                         raw_inode->i_block[0] = 0;
4448                         raw_inode->i_block[1] =
4449                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
4450                         raw_inode->i_block[2] = 0;
4451                 }
4452         } else for (block = 0; block < EXT4_N_BLOCKS; block++)
4453                 raw_inode->i_block[block] = ei->i_data[block];
4454
4455         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
4456         if (ei->i_extra_isize) {
4457                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4458                         raw_inode->i_version_hi =
4459                         cpu_to_le32(inode->i_version >> 32);
4460                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
4461         }
4462
4463         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4464         rc = ext4_handle_dirty_metadata(handle, inode, bh);
4465         if (!err)
4466                 err = rc;
4467         ei->i_state &= ~EXT4_STATE_NEW;
4468
4469 out_brelse:
4470         brelse(bh);
4471         ext4_std_error(inode->i_sb, err);
4472         return err;
4473 }
4474
4475 /*
4476  * ext4_write_inode()
4477  *
4478  * We are called from a few places:
4479  *
4480  * - Within generic_file_write() for O_SYNC files.
4481  *   Here, there will be no transaction running. We wait for any running
4482  *   trasnaction to commit.
4483  *
4484  * - Within sys_sync(), kupdate and such.
4485  *   We wait on commit, if tol to.
4486  *
4487  * - Within prune_icache() (PF_MEMALLOC == true)
4488  *   Here we simply return.  We can't afford to block kswapd on the
4489  *   journal commit.
4490  *
4491  * In all cases it is actually safe for us to return without doing anything,
4492  * because the inode has been copied into a raw inode buffer in
4493  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
4494  * knfsd.
4495  *
4496  * Note that we are absolutely dependent upon all inode dirtiers doing the
4497  * right thing: they *must* call mark_inode_dirty() after dirtying info in
4498  * which we are interested.
4499  *
4500  * It would be a bug for them to not do this.  The code:
4501  *
4502  *      mark_inode_dirty(inode)
4503  *      stuff();
4504  *      inode->i_size = expr;
4505  *
4506  * is in error because a kswapd-driven write_inode() could occur while
4507  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
4508  * will no longer be on the superblock's dirty inode list.
4509  */
4510 int ext4_write_inode(struct inode *inode, int wait)
4511 {
4512         if (current->flags & PF_MEMALLOC)
4513                 return 0;
4514
4515         if (ext4_journal_current_handle()) {
4516                 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
4517                 dump_stack();
4518                 return -EIO;
4519         }
4520
4521         if (!wait)
4522                 return 0;
4523
4524         return ext4_force_commit(inode->i_sb);
4525 }
4526
4527 int __ext4_write_dirty_metadata(struct inode *inode, struct buffer_head *bh)
4528 {
4529         int err = 0;
4530
4531         mark_buffer_dirty(bh);
4532         if (inode && inode_needs_sync(inode)) {
4533                 sync_dirty_buffer(bh);
4534                 if (buffer_req(bh) && !buffer_uptodate(bh)) {
4535                         ext4_error(inode->i_sb, __func__,
4536                                    "IO error syncing inode, "
4537                                    "inode=%lu, block=%llu",
4538                                    inode->i_ino,
4539                                    (unsigned long long)bh->b_blocknr);
4540                         err = -EIO;
4541                 }
4542         }
4543         return err;
4544 }
4545
4546 /*
4547  * ext4_setattr()
4548  *
4549  * Called from notify_change.
4550  *
4551  * We want to trap VFS attempts to truncate the file as soon as
4552  * possible.  In particular, we want to make sure that when the VFS
4553  * shrinks i_size, we put the inode on the orphan list and modify
4554  * i_disksize immediately, so that during the subsequent flushing of
4555  * dirty pages and freeing of disk blocks, we can guarantee that any
4556  * commit will leave the blocks being flushed in an unused state on
4557  * disk.  (On recovery, the inode will get truncated and the blocks will
4558  * be freed, so we have a strong guarantee that no future commit will
4559  * leave these blocks visible to the user.)
4560  *
4561  * Another thing we have to assure is that if we are in ordered mode
4562  * and inode is still attached to the committing transaction, we must
4563  * we start writeout of all the dirty pages which are being truncated.
4564  * This way we are sure that all the data written in the previous
4565  * transaction are already on disk (truncate waits for pages under
4566  * writeback).
4567  *
4568  * Called with inode->i_mutex down.
4569  */
4570 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
4571 {
4572         struct inode *inode = dentry->d_inode;
4573         int error, rc = 0;
4574         const unsigned int ia_valid = attr->ia_valid;
4575
4576         error = inode_change_ok(inode, attr);
4577         if (error)
4578                 return error;
4579
4580         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
4581                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
4582                 handle_t *handle;
4583
4584                 /* (user+group)*(old+new) structure, inode write (sb,
4585                  * inode block, ? - but truncate inode update has it) */
4586                 handle = ext4_journal_start(inode, 2*(EXT4_QUOTA_INIT_BLOCKS(inode->i_sb)+
4587                                         EXT4_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
4588                 if (IS_ERR(handle)) {
4589                         error = PTR_ERR(handle);
4590                         goto err_out;
4591                 }
4592                 error = DQUOT_TRANSFER(inode, attr) ? -EDQUOT : 0;
4593                 if (error) {
4594                         ext4_journal_stop(handle);
4595                         return error;
4596                 }
4597                 /* Update corresponding info in inode so that everything is in
4598                  * one transaction */
4599                 if (attr->ia_valid & ATTR_UID)
4600                         inode->i_uid = attr->ia_uid;
4601                 if (attr->ia_valid & ATTR_GID)
4602                         inode->i_gid = attr->ia_gid;
4603                 error = ext4_mark_inode_dirty(handle, inode);
4604                 ext4_journal_stop(handle);
4605         }
4606
4607         if (attr->ia_valid & ATTR_SIZE) {
4608                 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)) {
4609                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4610
4611                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
4612                                 error = -EFBIG;
4613                                 goto err_out;
4614                         }
4615                 }
4616         }
4617
4618         if (S_ISREG(inode->i_mode) &&
4619             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
4620                 handle_t *handle;
4621
4622                 handle = ext4_journal_start(inode, 3);
4623                 if (IS_ERR(handle)) {
4624                         error = PTR_ERR(handle);
4625                         goto err_out;
4626                 }
4627
4628                 error = ext4_orphan_add(handle, inode);
4629                 EXT4_I(inode)->i_disksize = attr->ia_size;
4630                 rc = ext4_mark_inode_dirty(handle, inode);
4631                 if (!error)
4632                         error = rc;
4633                 ext4_journal_stop(handle);
4634
4635                 if (ext4_should_order_data(inode)) {
4636                         error = ext4_begin_ordered_truncate(inode,
4637                                                             attr->ia_size);
4638                         if (error) {
4639                                 /* Do as much error cleanup as possible */
4640                                 handle = ext4_journal_start(inode, 3);
4641                                 if (IS_ERR(handle)) {
4642                                         ext4_orphan_del(NULL, inode);
4643                                         goto err_out;
4644                                 }
4645                                 ext4_orphan_del(handle, inode);
4646                                 ext4_journal_stop(handle);
4647                                 goto err_out;
4648                         }
4649                 }
4650         }
4651
4652         rc = inode_setattr(inode, attr);
4653
4654         /* If inode_setattr's call to ext4_truncate failed to get a
4655          * transaction handle at all, we need to clean up the in-core
4656          * orphan list manually. */
4657         if (inode->i_nlink)
4658                 ext4_orphan_del(NULL, inode);
4659
4660         if (!rc && (ia_valid & ATTR_MODE))
4661                 rc = ext4_acl_chmod(inode);
4662
4663 err_out:
4664         ext4_std_error(inode->i_sb, error);
4665         if (!error)
4666                 error = rc;
4667         return error;
4668 }
4669
4670 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
4671                  struct kstat *stat)
4672 {
4673         struct inode *inode;
4674         unsigned long delalloc_blocks;
4675
4676         inode = dentry->d_inode;
4677         generic_fillattr(inode, stat);
4678
4679         /*
4680          * We can't update i_blocks if the block allocation is delayed
4681          * otherwise in the case of system crash before the real block
4682          * allocation is done, we will have i_blocks inconsistent with
4683          * on-disk file blocks.
4684          * We always keep i_blocks updated together with real
4685          * allocation. But to not confuse with user, stat
4686          * will return the blocks that include the delayed allocation
4687          * blocks for this file.
4688          */
4689         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
4690         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
4691         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
4692
4693         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
4694         return 0;
4695 }
4696
4697 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
4698                                       int chunk)
4699 {
4700         int indirects;
4701
4702         /* if nrblocks are contiguous */
4703         if (chunk) {
4704                 /*
4705                  * With N contiguous data blocks, it need at most
4706                  * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) indirect blocks
4707                  * 2 dindirect blocks
4708                  * 1 tindirect block
4709                  */
4710                 indirects = nrblocks / EXT4_ADDR_PER_BLOCK(inode->i_sb);
4711                 return indirects + 3;
4712         }
4713         /*
4714          * if nrblocks are not contiguous, worse case, each block touch
4715          * a indirect block, and each indirect block touch a double indirect
4716          * block, plus a triple indirect block
4717          */
4718         indirects = nrblocks * 2 + 1;
4719         return indirects;
4720 }
4721
4722 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4723 {
4724         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL))
4725                 return ext4_indirect_trans_blocks(inode, nrblocks, chunk);
4726         return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
4727 }
4728
4729 /*
4730  * Account for index blocks, block groups bitmaps and block group
4731  * descriptor blocks if modify datablocks and index blocks
4732  * worse case, the indexs blocks spread over different block groups
4733  *
4734  * If datablocks are discontiguous, they are possible to spread over
4735  * different block groups too. If they are contiugous, with flexbg,
4736  * they could still across block group boundary.
4737  *
4738  * Also account for superblock, inode, quota and xattr blocks
4739  */
4740 int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4741 {
4742         int groups, gdpblocks;
4743         int idxblocks;
4744         int ret = 0;
4745
4746         /*
4747          * How many index blocks need to touch to modify nrblocks?
4748          * The "Chunk" flag indicating whether the nrblocks is
4749          * physically contiguous on disk
4750          *
4751          * For Direct IO and fallocate, they calls get_block to allocate
4752          * one single extent at a time, so they could set the "Chunk" flag
4753          */
4754         idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
4755
4756         ret = idxblocks;
4757
4758         /*
4759          * Now let's see how many group bitmaps and group descriptors need
4760          * to account
4761          */
4762         groups = idxblocks;
4763         if (chunk)
4764                 groups += 1;
4765         else
4766                 groups += nrblocks;
4767
4768         gdpblocks = groups;
4769         if (groups > EXT4_SB(inode->i_sb)->s_groups_count)
4770                 groups = EXT4_SB(inode->i_sb)->s_groups_count;
4771         if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
4772                 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
4773
4774         /* bitmaps and block group descriptor blocks */
4775         ret += groups + gdpblocks;
4776
4777         /* Blocks for super block, inode, quota and xattr blocks */
4778         ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
4779
4780         return ret;
4781 }
4782
4783 /*
4784  * Calulate the total number of credits to reserve to fit
4785  * the modification of a single pages into a single transaction,
4786  * which may include multiple chunks of block allocations.
4787  *
4788  * This could be called via ext4_write_begin()
4789  *
4790  * We need to consider the worse case, when
4791  * one new block per extent.
4792  */
4793 int ext4_writepage_trans_blocks(struct inode *inode)
4794 {
4795         int bpp = ext4_journal_blocks_per_page(inode);
4796         int ret;
4797
4798         ret = ext4_meta_trans_blocks(inode, bpp, 0);
4799
4800         /* Account for data blocks for journalled mode */
4801         if (ext4_should_journal_data(inode))
4802                 ret += bpp;
4803         return ret;
4804 }
4805
4806 /*
4807  * Calculate the journal credits for a chunk of data modification.
4808  *
4809  * This is called from DIO, fallocate or whoever calling
4810  * ext4_get_blocks_wrap() to map/allocate a chunk of contigous disk blocks.
4811  *
4812  * journal buffers for data blocks are not included here, as DIO
4813  * and fallocate do no need to journal data buffers.
4814  */
4815 int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
4816 {
4817         return ext4_meta_trans_blocks(inode, nrblocks, 1);
4818 }
4819
4820 /*
4821  * The caller must have previously called ext4_reserve_inode_write().
4822  * Give this, we know that the caller already has write access to iloc->bh.
4823  */
4824 int ext4_mark_iloc_dirty(handle_t *handle,
4825                 struct inode *inode, struct ext4_iloc *iloc)
4826 {
4827         int err = 0;
4828
4829         if (test_opt(inode->i_sb, I_VERSION))
4830                 inode_inc_iversion(inode);
4831
4832         /* the do_update_inode consumes one bh->b_count */
4833         get_bh(iloc->bh);
4834
4835         /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
4836         err = ext4_do_update_inode(handle, inode, iloc);
4837         put_bh(iloc->bh);
4838         return err;
4839 }
4840
4841 /*
4842  * On success, We end up with an outstanding reference count against
4843  * iloc->bh.  This _must_ be cleaned up later.
4844  */
4845
4846 int
4847 ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
4848                          struct ext4_iloc *iloc)
4849 {
4850         int err;
4851
4852         err = ext4_get_inode_loc(inode, iloc);
4853         if (!err) {
4854                 BUFFER_TRACE(iloc->bh, "get_write_access");
4855                 err = ext4_journal_get_write_access(handle, iloc->bh);
4856                 if (err) {
4857                         brelse(iloc->bh);
4858                         iloc->bh = NULL;
4859                 }
4860         }
4861         ext4_std_error(inode->i_sb, err);
4862         return err;
4863 }
4864
4865 /*
4866  * Expand an inode by new_extra_isize bytes.
4867  * Returns 0 on success or negative error number on failure.
4868  */
4869 static int ext4_expand_extra_isize(struct inode *inode,
4870                                    unsigned int new_extra_isize,
4871                                    struct ext4_iloc iloc,
4872                                    handle_t *handle)
4873 {
4874         struct ext4_inode *raw_inode;
4875         struct ext4_xattr_ibody_header *header;
4876         struct ext4_xattr_entry *entry;
4877
4878         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
4879                 return 0;
4880
4881         raw_inode = ext4_raw_inode(&iloc);
4882
4883         header = IHDR(inode, raw_inode);
4884         entry = IFIRST(header);
4885
4886         /* No extended attributes present */
4887         if (!(EXT4_I(inode)->i_state & EXT4_STATE_XATTR) ||
4888                 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
4889                 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
4890                         new_extra_isize);
4891                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
4892                 return 0;
4893         }
4894
4895         /* try to expand with EAs present */
4896         return ext4_expand_extra_isize_ea(inode, new_extra_isize,
4897                                           raw_inode, handle);
4898 }
4899
4900 /*
4901  * What we do here is to mark the in-core inode as clean with respect to inode
4902  * dirtiness (it may still be data-dirty).
4903  * This means that the in-core inode may be reaped by prune_icache
4904  * without having to perform any I/O.  This is a very good thing,
4905  * because *any* task may call prune_icache - even ones which
4906  * have a transaction open against a different journal.
4907  *
4908  * Is this cheating?  Not really.  Sure, we haven't written the
4909  * inode out, but prune_icache isn't a user-visible syncing function.
4910  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
4911  * we start and wait on commits.
4912  *
4913  * Is this efficient/effective?  Well, we're being nice to the system
4914  * by cleaning up our inodes proactively so they can be reaped
4915  * without I/O.  But we are potentially leaving up to five seconds'
4916  * worth of inodes floating about which prune_icache wants us to
4917  * write out.  One way to fix that would be to get prune_icache()
4918  * to do a write_super() to free up some memory.  It has the desired
4919  * effect.
4920  */
4921 int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
4922 {
4923         struct ext4_iloc iloc;
4924         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4925         static unsigned int mnt_count;
4926         int err, ret;
4927
4928         might_sleep();
4929         err = ext4_reserve_inode_write(handle, inode, &iloc);
4930         if (ext4_handle_valid(handle) &&
4931             EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
4932             !(EXT4_I(inode)->i_state & EXT4_STATE_NO_EXPAND)) {
4933                 /*
4934                  * We need extra buffer credits since we may write into EA block
4935                  * with this same handle. If journal_extend fails, then it will
4936                  * only result in a minor loss of functionality for that inode.
4937                  * If this is felt to be critical, then e2fsck should be run to
4938                  * force a large enough s_min_extra_isize.
4939                  */
4940                 if ((jbd2_journal_extend(handle,
4941                              EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
4942                         ret = ext4_expand_extra_isize(inode,
4943                                                       sbi->s_want_extra_isize,
4944                                                       iloc, handle);
4945                         if (ret) {
4946                                 EXT4_I(inode)->i_state |= EXT4_STATE_NO_EXPAND;
4947                                 if (mnt_count !=
4948                                         le16_to_cpu(sbi->s_es->s_mnt_count)) {
4949                                         ext4_warning(inode->i_sb, __func__,
4950                                         "Unable to expand inode %lu. Delete"
4951                                         " some EAs or run e2fsck.",
4952                                         inode->i_ino);
4953                                         mnt_count =
4954                                           le16_to_cpu(sbi->s_es->s_mnt_count);
4955                                 }
4956                         }
4957                 }
4958         }
4959         if (!err)
4960                 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
4961         return err;
4962 }
4963
4964 /*
4965  * ext4_dirty_inode() is called from __mark_inode_dirty()
4966  *
4967  * We're really interested in the case where a file is being extended.
4968  * i_size has been changed by generic_commit_write() and we thus need
4969  * to include the updated inode in the current transaction.
4970  *
4971  * Also, DQUOT_ALLOC_SPACE() will always dirty the inode when blocks
4972  * are allocated to the file.
4973  *
4974  * If the inode is marked synchronous, we don't honour that here - doing
4975  * so would cause a commit on atime updates, which we don't bother doing.
4976  * We handle synchronous inodes at the highest possible level.
4977  */
4978 void ext4_dirty_inode(struct inode *inode)
4979 {
4980         handle_t *current_handle = ext4_journal_current_handle();
4981         handle_t *handle;
4982
4983         if (!ext4_handle_valid(current_handle)) {
4984                 ext4_mark_inode_dirty(current_handle, inode);
4985                 return;
4986         }
4987
4988         handle = ext4_journal_start(inode, 2);
4989         if (IS_ERR(handle))
4990                 goto out;
4991         if (current_handle &&
4992                 current_handle->h_transaction != handle->h_transaction) {
4993                 /* This task has a transaction open against a different fs */
4994                 printk(KERN_EMERG "%s: transactions do not match!\n",
4995                        __func__);
4996         } else {
4997                 jbd_debug(5, "marking dirty.  outer handle=%p\n",
4998                                 current_handle);
4999                 ext4_mark_inode_dirty(handle, inode);
5000         }
5001         ext4_journal_stop(handle);
5002 out:
5003         return;
5004 }
5005
5006 #if 0
5007 /*
5008  * Bind an inode's backing buffer_head into this transaction, to prevent
5009  * it from being flushed to disk early.  Unlike
5010  * ext4_reserve_inode_write, this leaves behind no bh reference and
5011  * returns no iloc structure, so the caller needs to repeat the iloc
5012  * lookup to mark the inode dirty later.
5013  */
5014 static int ext4_pin_inode(handle_t *handle, struct inode *inode)
5015 {
5016         struct ext4_iloc iloc;
5017
5018         int err = 0;
5019         if (handle) {
5020                 err = ext4_get_inode_loc(inode, &iloc);
5021                 if (!err) {
5022                         BUFFER_TRACE(iloc.bh, "get_write_access");
5023                         err = jbd2_journal_get_write_access(handle, iloc.bh);
5024                         if (!err)
5025                                 err = ext4_handle_dirty_metadata(handle,
5026                                                                  inode,
5027                                                                  iloc.bh);
5028                         brelse(iloc.bh);
5029                 }
5030         }
5031         ext4_std_error(inode->i_sb, err);
5032         return err;
5033 }
5034 #endif
5035
5036 int ext4_change_inode_journal_flag(struct inode *inode, int val)
5037 {
5038         journal_t *journal;
5039         handle_t *handle;
5040         int err;
5041
5042         /*
5043          * We have to be very careful here: changing a data block's
5044          * journaling status dynamically is dangerous.  If we write a
5045          * data block to the journal, change the status and then delete
5046          * that block, we risk forgetting to revoke the old log record
5047          * from the journal and so a subsequent replay can corrupt data.
5048          * So, first we make sure that the journal is empty and that
5049          * nobody is changing anything.
5050          */
5051
5052         journal = EXT4_JOURNAL(inode);
5053         if (!journal)
5054                 return 0;
5055         if (is_journal_aborted(journal))
5056                 return -EROFS;
5057
5058         jbd2_journal_lock_updates(journal);
5059         jbd2_journal_flush(journal);
5060
5061         /*
5062          * OK, there are no updates running now, and all cached data is
5063          * synced to disk.  We are now in a completely consistent state
5064          * which doesn't have anything in the journal, and we know that
5065          * no filesystem updates are running, so it is safe to modify
5066          * the inode's in-core data-journaling state flag now.
5067          */
5068
5069         if (val)
5070                 EXT4_I(inode)->i_flags |= EXT4_JOURNAL_DATA_FL;
5071         else
5072                 EXT4_I(inode)->i_flags &= ~EXT4_JOURNAL_DATA_FL;
5073         ext4_set_aops(inode);
5074
5075         jbd2_journal_unlock_updates(journal);
5076
5077         /* Finally we can mark the inode as dirty. */
5078
5079         handle = ext4_journal_start(inode, 1);
5080         if (IS_ERR(handle))
5081                 return PTR_ERR(handle);
5082
5083         err = ext4_mark_inode_dirty(handle, inode);
5084         ext4_handle_sync(handle);
5085         ext4_journal_stop(handle);
5086         ext4_std_error(inode->i_sb, err);
5087
5088         return err;
5089 }
5090
5091 static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
5092 {
5093         return !buffer_mapped(bh);
5094 }
5095
5096 int ext4_page_mkwrite(struct vm_area_struct *vma, struct page *page)
5097 {
5098         loff_t size;
5099         unsigned long len;
5100         int ret = -EINVAL;
5101         void *fsdata;
5102         struct file *file = vma->vm_file;
5103         struct inode *inode = file->f_path.dentry->d_inode;
5104         struct address_space *mapping = inode->i_mapping;
5105
5106         /*
5107          * Get i_alloc_sem to stop truncates messing with the inode. We cannot
5108          * get i_mutex because we are already holding mmap_sem.
5109          */
5110         down_read(&inode->i_alloc_sem);
5111         size = i_size_read(inode);
5112         if (page->mapping != mapping || size <= page_offset(page)
5113             || !PageUptodate(page)) {
5114                 /* page got truncated from under us? */
5115                 goto out_unlock;
5116         }
5117         ret = 0;
5118         if (PageMappedToDisk(page))
5119                 goto out_unlock;
5120
5121         if (page->index == size >> PAGE_CACHE_SHIFT)
5122                 len = size & ~PAGE_CACHE_MASK;
5123         else
5124                 len = PAGE_CACHE_SIZE;
5125
5126         if (page_has_buffers(page)) {
5127                 /* return if we have all the buffers mapped */
5128                 if (!walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
5129                                        ext4_bh_unmapped))
5130                         goto out_unlock;
5131         }
5132         /*
5133          * OK, we need to fill the hole... Do write_begin write_end
5134          * to do block allocation/reservation.We are not holding
5135          * inode.i__mutex here. That allow * parallel write_begin,
5136          * write_end call. lock_page prevent this from happening
5137          * on the same page though
5138          */
5139         ret = mapping->a_ops->write_begin(file, mapping, page_offset(page),
5140                         len, AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
5141         if (ret < 0)
5142                 goto out_unlock;
5143         ret = mapping->a_ops->write_end(file, mapping, page_offset(page),
5144                         len, len, page, fsdata);
5145         if (ret < 0)
5146                 goto out_unlock;
5147         ret = 0;
5148 out_unlock:
5149         up_read(&inode->i_alloc_sem);
5150         return ret;
5151 }