ext4: Implement range_cyclic in ext4_da_writepages instead of write_cache_pages
[sfrench/cifs-2.6.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/jbd2.h>
29 #include <linux/highuid.h>
30 #include <linux/pagemap.h>
31 #include <linux/quotaops.h>
32 #include <linux/string.h>
33 #include <linux/buffer_head.h>
34 #include <linux/writeback.h>
35 #include <linux/pagevec.h>
36 #include <linux/mpage.h>
37 #include <linux/namei.h>
38 #include <linux/uio.h>
39 #include <linux/bio.h>
40 #include "ext4_jbd2.h"
41 #include "xattr.h"
42 #include "acl.h"
43 #include "ext4_extents.h"
44
45 #define MPAGE_DA_EXTENT_TAIL 0x01
46
47 static inline int ext4_begin_ordered_truncate(struct inode *inode,
48                                               loff_t new_size)
49 {
50         return jbd2_journal_begin_ordered_truncate(
51                                         EXT4_SB(inode->i_sb)->s_journal,
52                                         &EXT4_I(inode)->jinode,
53                                         new_size);
54 }
55
56 static void ext4_invalidatepage(struct page *page, unsigned long offset);
57
58 /*
59  * Test whether an inode is a fast symlink.
60  */
61 static int ext4_inode_is_fast_symlink(struct inode *inode)
62 {
63         int ea_blocks = EXT4_I(inode)->i_file_acl ?
64                 (inode->i_sb->s_blocksize >> 9) : 0;
65
66         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
67 }
68
69 /*
70  * The ext4 forget function must perform a revoke if we are freeing data
71  * which has been journaled.  Metadata (eg. indirect blocks) must be
72  * revoked in all cases.
73  *
74  * "bh" may be NULL: a metadata block may have been freed from memory
75  * but there may still be a record of it in the journal, and that record
76  * still needs to be revoked.
77  *
78  * If the handle isn't valid we're not journaling so there's nothing to do.
79  */
80 int ext4_forget(handle_t *handle, int is_metadata, struct inode *inode,
81                         struct buffer_head *bh, ext4_fsblk_t blocknr)
82 {
83         int err;
84
85         if (!ext4_handle_valid(handle))
86                 return 0;
87
88         might_sleep();
89
90         BUFFER_TRACE(bh, "enter");
91
92         jbd_debug(4, "forgetting bh %p: is_metadata = %d, mode %o, "
93                   "data mode %lx\n",
94                   bh, is_metadata, inode->i_mode,
95                   test_opt(inode->i_sb, DATA_FLAGS));
96
97         /* Never use the revoke function if we are doing full data
98          * journaling: there is no need to, and a V1 superblock won't
99          * support it.  Otherwise, only skip the revoke on un-journaled
100          * data blocks. */
101
102         if (test_opt(inode->i_sb, DATA_FLAGS) == EXT4_MOUNT_JOURNAL_DATA ||
103             (!is_metadata && !ext4_should_journal_data(inode))) {
104                 if (bh) {
105                         BUFFER_TRACE(bh, "call jbd2_journal_forget");
106                         return ext4_journal_forget(handle, bh);
107                 }
108                 return 0;
109         }
110
111         /*
112          * data!=journal && (is_metadata || should_journal_data(inode))
113          */
114         BUFFER_TRACE(bh, "call ext4_journal_revoke");
115         err = ext4_journal_revoke(handle, blocknr, bh);
116         if (err)
117                 ext4_abort(inode->i_sb, __func__,
118                            "error %d when attempting revoke", err);
119         BUFFER_TRACE(bh, "exit");
120         return err;
121 }
122
123 /*
124  * Work out how many blocks we need to proceed with the next chunk of a
125  * truncate transaction.
126  */
127 static unsigned long blocks_for_truncate(struct inode *inode)
128 {
129         ext4_lblk_t needed;
130
131         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
132
133         /* Give ourselves just enough room to cope with inodes in which
134          * i_blocks is corrupt: we've seen disk corruptions in the past
135          * which resulted in random data in an inode which looked enough
136          * like a regular file for ext4 to try to delete it.  Things
137          * will go a bit crazy if that happens, but at least we should
138          * try not to panic the whole kernel. */
139         if (needed < 2)
140                 needed = 2;
141
142         /* But we need to bound the transaction so we don't overflow the
143          * journal. */
144         if (needed > EXT4_MAX_TRANS_DATA)
145                 needed = EXT4_MAX_TRANS_DATA;
146
147         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
148 }
149
150 /*
151  * Truncate transactions can be complex and absolutely huge.  So we need to
152  * be able to restart the transaction at a conventient checkpoint to make
153  * sure we don't overflow the journal.
154  *
155  * start_transaction gets us a new handle for a truncate transaction,
156  * and extend_transaction tries to extend the existing one a bit.  If
157  * extend fails, we need to propagate the failure up and restart the
158  * transaction in the top-level truncate loop. --sct
159  */
160 static handle_t *start_transaction(struct inode *inode)
161 {
162         handle_t *result;
163
164         result = ext4_journal_start(inode, blocks_for_truncate(inode));
165         if (!IS_ERR(result))
166                 return result;
167
168         ext4_std_error(inode->i_sb, PTR_ERR(result));
169         return result;
170 }
171
172 /*
173  * Try to extend this transaction for the purposes of truncation.
174  *
175  * Returns 0 if we managed to create more room.  If we can't create more
176  * room, and the transaction must be restarted we return 1.
177  */
178 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
179 {
180         if (!ext4_handle_valid(handle))
181                 return 0;
182         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
183                 return 0;
184         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
185                 return 0;
186         return 1;
187 }
188
189 /*
190  * Restart the transaction associated with *handle.  This does a commit,
191  * so before we call here everything must be consistently dirtied against
192  * this transaction.
193  */
194 static int ext4_journal_test_restart(handle_t *handle, struct inode *inode)
195 {
196         BUG_ON(EXT4_JOURNAL(inode) == NULL);
197         jbd_debug(2, "restarting handle %p\n", handle);
198         return ext4_journal_restart(handle, blocks_for_truncate(inode));
199 }
200
201 /*
202  * Called at the last iput() if i_nlink is zero.
203  */
204 void ext4_delete_inode(struct inode *inode)
205 {
206         handle_t *handle;
207         int err;
208
209         if (ext4_should_order_data(inode))
210                 ext4_begin_ordered_truncate(inode, 0);
211         truncate_inode_pages(&inode->i_data, 0);
212
213         if (is_bad_inode(inode))
214                 goto no_delete;
215
216         handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
217         if (IS_ERR(handle)) {
218                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
219                 /*
220                  * If we're going to skip the normal cleanup, we still need to
221                  * make sure that the in-core orphan linked list is properly
222                  * cleaned up.
223                  */
224                 ext4_orphan_del(NULL, inode);
225                 goto no_delete;
226         }
227
228         if (IS_SYNC(inode))
229                 ext4_handle_sync(handle);
230         inode->i_size = 0;
231         err = ext4_mark_inode_dirty(handle, inode);
232         if (err) {
233                 ext4_warning(inode->i_sb, __func__,
234                              "couldn't mark inode dirty (err %d)", err);
235                 goto stop_handle;
236         }
237         if (inode->i_blocks)
238                 ext4_truncate(inode);
239
240         /*
241          * ext4_ext_truncate() doesn't reserve any slop when it
242          * restarts journal transactions; therefore there may not be
243          * enough credits left in the handle to remove the inode from
244          * the orphan list and set the dtime field.
245          */
246         if (!ext4_handle_has_enough_credits(handle, 3)) {
247                 err = ext4_journal_extend(handle, 3);
248                 if (err > 0)
249                         err = ext4_journal_restart(handle, 3);
250                 if (err != 0) {
251                         ext4_warning(inode->i_sb, __func__,
252                                      "couldn't extend journal (err %d)", err);
253                 stop_handle:
254                         ext4_journal_stop(handle);
255                         goto no_delete;
256                 }
257         }
258
259         /*
260          * Kill off the orphan record which ext4_truncate created.
261          * AKPM: I think this can be inside the above `if'.
262          * Note that ext4_orphan_del() has to be able to cope with the
263          * deletion of a non-existent orphan - this is because we don't
264          * know if ext4_truncate() actually created an orphan record.
265          * (Well, we could do this if we need to, but heck - it works)
266          */
267         ext4_orphan_del(handle, inode);
268         EXT4_I(inode)->i_dtime  = get_seconds();
269
270         /*
271          * One subtle ordering requirement: if anything has gone wrong
272          * (transaction abort, IO errors, whatever), then we can still
273          * do these next steps (the fs will already have been marked as
274          * having errors), but we can't free the inode if the mark_dirty
275          * fails.
276          */
277         if (ext4_mark_inode_dirty(handle, inode))
278                 /* If that failed, just do the required in-core inode clear. */
279                 clear_inode(inode);
280         else
281                 ext4_free_inode(handle, inode);
282         ext4_journal_stop(handle);
283         return;
284 no_delete:
285         clear_inode(inode);     /* We must guarantee clearing of inode... */
286 }
287
288 typedef struct {
289         __le32  *p;
290         __le32  key;
291         struct buffer_head *bh;
292 } Indirect;
293
294 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
295 {
296         p->key = *(p->p = v);
297         p->bh = bh;
298 }
299
300 /**
301  *      ext4_block_to_path - parse the block number into array of offsets
302  *      @inode: inode in question (we are only interested in its superblock)
303  *      @i_block: block number to be parsed
304  *      @offsets: array to store the offsets in
305  *      @boundary: set this non-zero if the referred-to block is likely to be
306  *             followed (on disk) by an indirect block.
307  *
308  *      To store the locations of file's data ext4 uses a data structure common
309  *      for UNIX filesystems - tree of pointers anchored in the inode, with
310  *      data blocks at leaves and indirect blocks in intermediate nodes.
311  *      This function translates the block number into path in that tree -
312  *      return value is the path length and @offsets[n] is the offset of
313  *      pointer to (n+1)th node in the nth one. If @block is out of range
314  *      (negative or too large) warning is printed and zero returned.
315  *
316  *      Note: function doesn't find node addresses, so no IO is needed. All
317  *      we need to know is the capacity of indirect blocks (taken from the
318  *      inode->i_sb).
319  */
320
321 /*
322  * Portability note: the last comparison (check that we fit into triple
323  * indirect block) is spelled differently, because otherwise on an
324  * architecture with 32-bit longs and 8Kb pages we might get into trouble
325  * if our filesystem had 8Kb blocks. We might use long long, but that would
326  * kill us on x86. Oh, well, at least the sign propagation does not matter -
327  * i_block would have to be negative in the very beginning, so we would not
328  * get there at all.
329  */
330
331 static int ext4_block_to_path(struct inode *inode,
332                         ext4_lblk_t i_block,
333                         ext4_lblk_t offsets[4], int *boundary)
334 {
335         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
336         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
337         const long direct_blocks = EXT4_NDIR_BLOCKS,
338                 indirect_blocks = ptrs,
339                 double_blocks = (1 << (ptrs_bits * 2));
340         int n = 0;
341         int final = 0;
342
343         if (i_block < 0) {
344                 ext4_warning(inode->i_sb, "ext4_block_to_path", "block < 0");
345         } else if (i_block < direct_blocks) {
346                 offsets[n++] = i_block;
347                 final = direct_blocks;
348         } else if ((i_block -= direct_blocks) < indirect_blocks) {
349                 offsets[n++] = EXT4_IND_BLOCK;
350                 offsets[n++] = i_block;
351                 final = ptrs;
352         } else if ((i_block -= indirect_blocks) < double_blocks) {
353                 offsets[n++] = EXT4_DIND_BLOCK;
354                 offsets[n++] = i_block >> ptrs_bits;
355                 offsets[n++] = i_block & (ptrs - 1);
356                 final = ptrs;
357         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
358                 offsets[n++] = EXT4_TIND_BLOCK;
359                 offsets[n++] = i_block >> (ptrs_bits * 2);
360                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
361                 offsets[n++] = i_block & (ptrs - 1);
362                 final = ptrs;
363         } else {
364                 ext4_warning(inode->i_sb, "ext4_block_to_path",
365                                 "block %lu > max in inode %lu",
366                                 i_block + direct_blocks +
367                                 indirect_blocks + double_blocks, inode->i_ino);
368         }
369         if (boundary)
370                 *boundary = final - 1 - (i_block & (ptrs - 1));
371         return n;
372 }
373
374 /**
375  *      ext4_get_branch - read the chain of indirect blocks leading to data
376  *      @inode: inode in question
377  *      @depth: depth of the chain (1 - direct pointer, etc.)
378  *      @offsets: offsets of pointers in inode/indirect blocks
379  *      @chain: place to store the result
380  *      @err: here we store the error value
381  *
382  *      Function fills the array of triples <key, p, bh> and returns %NULL
383  *      if everything went OK or the pointer to the last filled triple
384  *      (incomplete one) otherwise. Upon the return chain[i].key contains
385  *      the number of (i+1)-th block in the chain (as it is stored in memory,
386  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
387  *      number (it points into struct inode for i==0 and into the bh->b_data
388  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
389  *      block for i>0 and NULL for i==0. In other words, it holds the block
390  *      numbers of the chain, addresses they were taken from (and where we can
391  *      verify that chain did not change) and buffer_heads hosting these
392  *      numbers.
393  *
394  *      Function stops when it stumbles upon zero pointer (absent block)
395  *              (pointer to last triple returned, *@err == 0)
396  *      or when it gets an IO error reading an indirect block
397  *              (ditto, *@err == -EIO)
398  *      or when it reads all @depth-1 indirect blocks successfully and finds
399  *      the whole chain, all way to the data (returns %NULL, *err == 0).
400  *
401  *      Need to be called with
402  *      down_read(&EXT4_I(inode)->i_data_sem)
403  */
404 static Indirect *ext4_get_branch(struct inode *inode, int depth,
405                                  ext4_lblk_t  *offsets,
406                                  Indirect chain[4], int *err)
407 {
408         struct super_block *sb = inode->i_sb;
409         Indirect *p = chain;
410         struct buffer_head *bh;
411
412         *err = 0;
413         /* i_data is not going away, no lock needed */
414         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
415         if (!p->key)
416                 goto no_block;
417         while (--depth) {
418                 bh = sb_bread(sb, le32_to_cpu(p->key));
419                 if (!bh)
420                         goto failure;
421                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
422                 /* Reader: end */
423                 if (!p->key)
424                         goto no_block;
425         }
426         return NULL;
427
428 failure:
429         *err = -EIO;
430 no_block:
431         return p;
432 }
433
434 /**
435  *      ext4_find_near - find a place for allocation with sufficient locality
436  *      @inode: owner
437  *      @ind: descriptor of indirect block.
438  *
439  *      This function returns the preferred place for block allocation.
440  *      It is used when heuristic for sequential allocation fails.
441  *      Rules are:
442  *        + if there is a block to the left of our position - allocate near it.
443  *        + if pointer will live in indirect block - allocate near that block.
444  *        + if pointer will live in inode - allocate in the same
445  *          cylinder group.
446  *
447  * In the latter case we colour the starting block by the callers PID to
448  * prevent it from clashing with concurrent allocations for a different inode
449  * in the same block group.   The PID is used here so that functionally related
450  * files will be close-by on-disk.
451  *
452  *      Caller must make sure that @ind is valid and will stay that way.
453  */
454 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
455 {
456         struct ext4_inode_info *ei = EXT4_I(inode);
457         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
458         __le32 *p;
459         ext4_fsblk_t bg_start;
460         ext4_fsblk_t last_block;
461         ext4_grpblk_t colour;
462
463         /* Try to find previous block */
464         for (p = ind->p - 1; p >= start; p--) {
465                 if (*p)
466                         return le32_to_cpu(*p);
467         }
468
469         /* No such thing, so let's try location of indirect block */
470         if (ind->bh)
471                 return ind->bh->b_blocknr;
472
473         /*
474          * It is going to be referred to from the inode itself? OK, just put it
475          * into the same cylinder group then.
476          */
477         bg_start = ext4_group_first_block_no(inode->i_sb, ei->i_block_group);
478         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
479
480         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
481                 colour = (current->pid % 16) *
482                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
483         else
484                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
485         return bg_start + colour;
486 }
487
488 /**
489  *      ext4_find_goal - find a preferred place for allocation.
490  *      @inode: owner
491  *      @block:  block we want
492  *      @partial: pointer to the last triple within a chain
493  *
494  *      Normally this function find the preferred place for block allocation,
495  *      returns it.
496  */
497 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
498                 Indirect *partial)
499 {
500         /*
501          * XXX need to get goal block from mballoc's data structures
502          */
503
504         return ext4_find_near(inode, partial);
505 }
506
507 /**
508  *      ext4_blks_to_allocate: Look up the block map and count the number
509  *      of direct blocks need to be allocated for the given branch.
510  *
511  *      @branch: chain of indirect blocks
512  *      @k: number of blocks need for indirect blocks
513  *      @blks: number of data blocks to be mapped.
514  *      @blocks_to_boundary:  the offset in the indirect block
515  *
516  *      return the total number of blocks to be allocate, including the
517  *      direct and indirect blocks.
518  */
519 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
520                 int blocks_to_boundary)
521 {
522         unsigned int count = 0;
523
524         /*
525          * Simple case, [t,d]Indirect block(s) has not allocated yet
526          * then it's clear blocks on that path have not allocated
527          */
528         if (k > 0) {
529                 /* right now we don't handle cross boundary allocation */
530                 if (blks < blocks_to_boundary + 1)
531                         count += blks;
532                 else
533                         count += blocks_to_boundary + 1;
534                 return count;
535         }
536
537         count++;
538         while (count < blks && count <= blocks_to_boundary &&
539                 le32_to_cpu(*(branch[0].p + count)) == 0) {
540                 count++;
541         }
542         return count;
543 }
544
545 /**
546  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
547  *      @indirect_blks: the number of blocks need to allocate for indirect
548  *                      blocks
549  *
550  *      @new_blocks: on return it will store the new block numbers for
551  *      the indirect blocks(if needed) and the first direct block,
552  *      @blks:  on return it will store the total number of allocated
553  *              direct blocks
554  */
555 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
556                                 ext4_lblk_t iblock, ext4_fsblk_t goal,
557                                 int indirect_blks, int blks,
558                                 ext4_fsblk_t new_blocks[4], int *err)
559 {
560         struct ext4_allocation_request ar;
561         int target, i;
562         unsigned long count = 0, blk_allocated = 0;
563         int index = 0;
564         ext4_fsblk_t current_block = 0;
565         int ret = 0;
566
567         /*
568          * Here we try to allocate the requested multiple blocks at once,
569          * on a best-effort basis.
570          * To build a branch, we should allocate blocks for
571          * the indirect blocks(if not allocated yet), and at least
572          * the first direct block of this branch.  That's the
573          * minimum number of blocks need to allocate(required)
574          */
575         /* first we try to allocate the indirect blocks */
576         target = indirect_blks;
577         while (target > 0) {
578                 count = target;
579                 /* allocating blocks for indirect blocks and direct blocks */
580                 current_block = ext4_new_meta_blocks(handle, inode,
581                                                         goal, &count, err);
582                 if (*err)
583                         goto failed_out;
584
585                 target -= count;
586                 /* allocate blocks for indirect blocks */
587                 while (index < indirect_blks && count) {
588                         new_blocks[index++] = current_block++;
589                         count--;
590                 }
591                 if (count > 0) {
592                         /*
593                          * save the new block number
594                          * for the first direct block
595                          */
596                         new_blocks[index] = current_block;
597                         printk(KERN_INFO "%s returned more blocks than "
598                                                 "requested\n", __func__);
599                         WARN_ON(1);
600                         break;
601                 }
602         }
603
604         target = blks - count ;
605         blk_allocated = count;
606         if (!target)
607                 goto allocated;
608         /* Now allocate data blocks */
609         memset(&ar, 0, sizeof(ar));
610         ar.inode = inode;
611         ar.goal = goal;
612         ar.len = target;
613         ar.logical = iblock;
614         if (S_ISREG(inode->i_mode))
615                 /* enable in-core preallocation only for regular files */
616                 ar.flags = EXT4_MB_HINT_DATA;
617
618         current_block = ext4_mb_new_blocks(handle, &ar, err);
619
620         if (*err && (target == blks)) {
621                 /*
622                  * if the allocation failed and we didn't allocate
623                  * any blocks before
624                  */
625                 goto failed_out;
626         }
627         if (!*err) {
628                 if (target == blks) {
629                 /*
630                  * save the new block number
631                  * for the first direct block
632                  */
633                         new_blocks[index] = current_block;
634                 }
635                 blk_allocated += ar.len;
636         }
637 allocated:
638         /* total number of blocks allocated for direct blocks */
639         ret = blk_allocated;
640         *err = 0;
641         return ret;
642 failed_out:
643         for (i = 0; i < index; i++)
644                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
645         return ret;
646 }
647
648 /**
649  *      ext4_alloc_branch - allocate and set up a chain of blocks.
650  *      @inode: owner
651  *      @indirect_blks: number of allocated indirect blocks
652  *      @blks: number of allocated direct blocks
653  *      @offsets: offsets (in the blocks) to store the pointers to next.
654  *      @branch: place to store the chain in.
655  *
656  *      This function allocates blocks, zeroes out all but the last one,
657  *      links them into chain and (if we are synchronous) writes them to disk.
658  *      In other words, it prepares a branch that can be spliced onto the
659  *      inode. It stores the information about that chain in the branch[], in
660  *      the same format as ext4_get_branch() would do. We are calling it after
661  *      we had read the existing part of chain and partial points to the last
662  *      triple of that (one with zero ->key). Upon the exit we have the same
663  *      picture as after the successful ext4_get_block(), except that in one
664  *      place chain is disconnected - *branch->p is still zero (we did not
665  *      set the last link), but branch->key contains the number that should
666  *      be placed into *branch->p to fill that gap.
667  *
668  *      If allocation fails we free all blocks we've allocated (and forget
669  *      their buffer_heads) and return the error value the from failed
670  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
671  *      as described above and return 0.
672  */
673 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
674                                 ext4_lblk_t iblock, int indirect_blks,
675                                 int *blks, ext4_fsblk_t goal,
676                                 ext4_lblk_t *offsets, Indirect *branch)
677 {
678         int blocksize = inode->i_sb->s_blocksize;
679         int i, n = 0;
680         int err = 0;
681         struct buffer_head *bh;
682         int num;
683         ext4_fsblk_t new_blocks[4];
684         ext4_fsblk_t current_block;
685
686         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
687                                 *blks, new_blocks, &err);
688         if (err)
689                 return err;
690
691         branch[0].key = cpu_to_le32(new_blocks[0]);
692         /*
693          * metadata blocks and data blocks are allocated.
694          */
695         for (n = 1; n <= indirect_blks;  n++) {
696                 /*
697                  * Get buffer_head for parent block, zero it out
698                  * and set the pointer to new one, then send
699                  * parent to disk.
700                  */
701                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
702                 branch[n].bh = bh;
703                 lock_buffer(bh);
704                 BUFFER_TRACE(bh, "call get_create_access");
705                 err = ext4_journal_get_create_access(handle, bh);
706                 if (err) {
707                         unlock_buffer(bh);
708                         brelse(bh);
709                         goto failed;
710                 }
711
712                 memset(bh->b_data, 0, blocksize);
713                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
714                 branch[n].key = cpu_to_le32(new_blocks[n]);
715                 *branch[n].p = branch[n].key;
716                 if (n == indirect_blks) {
717                         current_block = new_blocks[n];
718                         /*
719                          * End of chain, update the last new metablock of
720                          * the chain to point to the new allocated
721                          * data blocks numbers
722                          */
723                         for (i=1; i < num; i++)
724                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
725                 }
726                 BUFFER_TRACE(bh, "marking uptodate");
727                 set_buffer_uptodate(bh);
728                 unlock_buffer(bh);
729
730                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
731                 err = ext4_handle_dirty_metadata(handle, inode, bh);
732                 if (err)
733                         goto failed;
734         }
735         *blks = num;
736         return err;
737 failed:
738         /* Allocation failed, free what we already allocated */
739         for (i = 1; i <= n ; i++) {
740                 BUFFER_TRACE(branch[i].bh, "call jbd2_journal_forget");
741                 ext4_journal_forget(handle, branch[i].bh);
742         }
743         for (i = 0; i < indirect_blks; i++)
744                 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
745
746         ext4_free_blocks(handle, inode, new_blocks[i], num, 0);
747
748         return err;
749 }
750
751 /**
752  * ext4_splice_branch - splice the allocated branch onto inode.
753  * @inode: owner
754  * @block: (logical) number of block we are adding
755  * @chain: chain of indirect blocks (with a missing link - see
756  *      ext4_alloc_branch)
757  * @where: location of missing link
758  * @num:   number of indirect blocks we are adding
759  * @blks:  number of direct blocks we are adding
760  *
761  * This function fills the missing link and does all housekeeping needed in
762  * inode (->i_blocks, etc.). In case of success we end up with the full
763  * chain to new block and return 0.
764  */
765 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
766                         ext4_lblk_t block, Indirect *where, int num, int blks)
767 {
768         int i;
769         int err = 0;
770         ext4_fsblk_t current_block;
771
772         /*
773          * If we're splicing into a [td]indirect block (as opposed to the
774          * inode) then we need to get write access to the [td]indirect block
775          * before the splice.
776          */
777         if (where->bh) {
778                 BUFFER_TRACE(where->bh, "get_write_access");
779                 err = ext4_journal_get_write_access(handle, where->bh);
780                 if (err)
781                         goto err_out;
782         }
783         /* That's it */
784
785         *where->p = where->key;
786
787         /*
788          * Update the host buffer_head or inode to point to more just allocated
789          * direct blocks blocks
790          */
791         if (num == 0 && blks > 1) {
792                 current_block = le32_to_cpu(where->key) + 1;
793                 for (i = 1; i < blks; i++)
794                         *(where->p + i) = cpu_to_le32(current_block++);
795         }
796
797         /* We are done with atomic stuff, now do the rest of housekeeping */
798
799         inode->i_ctime = ext4_current_time(inode);
800         ext4_mark_inode_dirty(handle, inode);
801
802         /* had we spliced it onto indirect block? */
803         if (where->bh) {
804                 /*
805                  * If we spliced it onto an indirect block, we haven't
806                  * altered the inode.  Note however that if it is being spliced
807                  * onto an indirect block at the very end of the file (the
808                  * file is growing) then we *will* alter the inode to reflect
809                  * the new i_size.  But that is not done here - it is done in
810                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
811                  */
812                 jbd_debug(5, "splicing indirect only\n");
813                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
814                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
815                 if (err)
816                         goto err_out;
817         } else {
818                 /*
819                  * OK, we spliced it into the inode itself on a direct block.
820                  * Inode was dirtied above.
821                  */
822                 jbd_debug(5, "splicing direct\n");
823         }
824         return err;
825
826 err_out:
827         for (i = 1; i <= num; i++) {
828                 BUFFER_TRACE(where[i].bh, "call jbd2_journal_forget");
829                 ext4_journal_forget(handle, where[i].bh);
830                 ext4_free_blocks(handle, inode,
831                                         le32_to_cpu(where[i-1].key), 1, 0);
832         }
833         ext4_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks, 0);
834
835         return err;
836 }
837
838 /*
839  * Allocation strategy is simple: if we have to allocate something, we will
840  * have to go the whole way to leaf. So let's do it before attaching anything
841  * to tree, set linkage between the newborn blocks, write them if sync is
842  * required, recheck the path, free and repeat if check fails, otherwise
843  * set the last missing link (that will protect us from any truncate-generated
844  * removals - all blocks on the path are immune now) and possibly force the
845  * write on the parent block.
846  * That has a nice additional property: no special recovery from the failed
847  * allocations is needed - we simply release blocks and do not touch anything
848  * reachable from inode.
849  *
850  * `handle' can be NULL if create == 0.
851  *
852  * return > 0, # of blocks mapped or allocated.
853  * return = 0, if plain lookup failed.
854  * return < 0, error case.
855  *
856  *
857  * Need to be called with
858  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
859  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
860  */
861 static int ext4_get_blocks_handle(handle_t *handle, struct inode *inode,
862                                   ext4_lblk_t iblock, unsigned int maxblocks,
863                                   struct buffer_head *bh_result,
864                                   int create, int extend_disksize)
865 {
866         int err = -EIO;
867         ext4_lblk_t offsets[4];
868         Indirect chain[4];
869         Indirect *partial;
870         ext4_fsblk_t goal;
871         int indirect_blks;
872         int blocks_to_boundary = 0;
873         int depth;
874         struct ext4_inode_info *ei = EXT4_I(inode);
875         int count = 0;
876         ext4_fsblk_t first_block = 0;
877         loff_t disksize;
878
879
880         J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
881         J_ASSERT(handle != NULL || create == 0);
882         depth = ext4_block_to_path(inode, iblock, offsets,
883                                         &blocks_to_boundary);
884
885         if (depth == 0)
886                 goto out;
887
888         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
889
890         /* Simplest case - block found, no allocation needed */
891         if (!partial) {
892                 first_block = le32_to_cpu(chain[depth - 1].key);
893                 clear_buffer_new(bh_result);
894                 count++;
895                 /*map more blocks*/
896                 while (count < maxblocks && count <= blocks_to_boundary) {
897                         ext4_fsblk_t blk;
898
899                         blk = le32_to_cpu(*(chain[depth-1].p + count));
900
901                         if (blk == first_block + count)
902                                 count++;
903                         else
904                                 break;
905                 }
906                 goto got_it;
907         }
908
909         /* Next simple case - plain lookup or failed read of indirect block */
910         if (!create || err == -EIO)
911                 goto cleanup;
912
913         /*
914          * Okay, we need to do block allocation.
915         */
916         goal = ext4_find_goal(inode, iblock, partial);
917
918         /* the number of blocks need to allocate for [d,t]indirect blocks */
919         indirect_blks = (chain + depth) - partial - 1;
920
921         /*
922          * Next look up the indirect map to count the totoal number of
923          * direct blocks to allocate for this branch.
924          */
925         count = ext4_blks_to_allocate(partial, indirect_blks,
926                                         maxblocks, blocks_to_boundary);
927         /*
928          * Block out ext4_truncate while we alter the tree
929          */
930         err = ext4_alloc_branch(handle, inode, iblock, indirect_blks,
931                                         &count, goal,
932                                         offsets + (partial - chain), partial);
933
934         /*
935          * The ext4_splice_branch call will free and forget any buffers
936          * on the new chain if there is a failure, but that risks using
937          * up transaction credits, especially for bitmaps where the
938          * credits cannot be returned.  Can we handle this somehow?  We
939          * may need to return -EAGAIN upwards in the worst case.  --sct
940          */
941         if (!err)
942                 err = ext4_splice_branch(handle, inode, iblock,
943                                         partial, indirect_blks, count);
944         /*
945          * i_disksize growing is protected by i_data_sem.  Don't forget to
946          * protect it if you're about to implement concurrent
947          * ext4_get_block() -bzzz
948         */
949         if (!err && extend_disksize) {
950                 disksize = ((loff_t) iblock + count) << inode->i_blkbits;
951                 if (disksize > i_size_read(inode))
952                         disksize = i_size_read(inode);
953                 if (disksize > ei->i_disksize)
954                         ei->i_disksize = disksize;
955         }
956         if (err)
957                 goto cleanup;
958
959         set_buffer_new(bh_result);
960 got_it:
961         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
962         if (count > blocks_to_boundary)
963                 set_buffer_boundary(bh_result);
964         err = count;
965         /* Clean up and exit */
966         partial = chain + depth - 1;    /* the whole chain */
967 cleanup:
968         while (partial > chain) {
969                 BUFFER_TRACE(partial->bh, "call brelse");
970                 brelse(partial->bh);
971                 partial--;
972         }
973         BUFFER_TRACE(bh_result, "returned");
974 out:
975         return err;
976 }
977
978 /*
979  * Calculate the number of metadata blocks need to reserve
980  * to allocate @blocks for non extent file based file
981  */
982 static int ext4_indirect_calc_metadata_amount(struct inode *inode, int blocks)
983 {
984         int icap = EXT4_ADDR_PER_BLOCK(inode->i_sb);
985         int ind_blks, dind_blks, tind_blks;
986
987         /* number of new indirect blocks needed */
988         ind_blks = (blocks + icap - 1) / icap;
989
990         dind_blks = (ind_blks + icap - 1) / icap;
991
992         tind_blks = 1;
993
994         return ind_blks + dind_blks + tind_blks;
995 }
996
997 /*
998  * Calculate the number of metadata blocks need to reserve
999  * to allocate given number of blocks
1000  */
1001 static int ext4_calc_metadata_amount(struct inode *inode, int blocks)
1002 {
1003         if (!blocks)
1004                 return 0;
1005
1006         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
1007                 return ext4_ext_calc_metadata_amount(inode, blocks);
1008
1009         return ext4_indirect_calc_metadata_amount(inode, blocks);
1010 }
1011
1012 static void ext4_da_update_reserve_space(struct inode *inode, int used)
1013 {
1014         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1015         int total, mdb, mdb_free;
1016
1017         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1018         /* recalculate the number of metablocks still need to be reserved */
1019         total = EXT4_I(inode)->i_reserved_data_blocks - used;
1020         mdb = ext4_calc_metadata_amount(inode, total);
1021
1022         /* figure out how many metablocks to release */
1023         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1024         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1025
1026         if (mdb_free) {
1027                 /* Account for allocated meta_blocks */
1028                 mdb_free -= EXT4_I(inode)->i_allocated_meta_blocks;
1029
1030                 /* update fs dirty blocks counter */
1031                 percpu_counter_sub(&sbi->s_dirtyblocks_counter, mdb_free);
1032                 EXT4_I(inode)->i_allocated_meta_blocks = 0;
1033                 EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1034         }
1035
1036         /* update per-inode reservations */
1037         BUG_ON(used  > EXT4_I(inode)->i_reserved_data_blocks);
1038         EXT4_I(inode)->i_reserved_data_blocks -= used;
1039
1040         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1041 }
1042
1043 /*
1044  * The ext4_get_blocks_wrap() function try to look up the requested blocks,
1045  * and returns if the blocks are already mapped.
1046  *
1047  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1048  * and store the allocated blocks in the result buffer head and mark it
1049  * mapped.
1050  *
1051  * If file type is extents based, it will call ext4_ext_get_blocks(),
1052  * Otherwise, call with ext4_get_blocks_handle() to handle indirect mapping
1053  * based files
1054  *
1055  * On success, it returns the number of blocks being mapped or allocate.
1056  * if create==0 and the blocks are pre-allocated and uninitialized block,
1057  * the result buffer head is unmapped. If the create ==1, it will make sure
1058  * the buffer head is mapped.
1059  *
1060  * It returns 0 if plain look up failed (blocks have not been allocated), in
1061  * that casem, buffer head is unmapped
1062  *
1063  * It returns the error in case of allocation failure.
1064  */
1065 int ext4_get_blocks_wrap(handle_t *handle, struct inode *inode, sector_t block,
1066                         unsigned int max_blocks, struct buffer_head *bh,
1067                         int create, int extend_disksize, int flag)
1068 {
1069         int retval;
1070
1071         clear_buffer_mapped(bh);
1072
1073         /*
1074          * Try to see if we can get  the block without requesting
1075          * for new file system block.
1076          */
1077         down_read((&EXT4_I(inode)->i_data_sem));
1078         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1079                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1080                                 bh, 0, 0);
1081         } else {
1082                 retval = ext4_get_blocks_handle(handle,
1083                                 inode, block, max_blocks, bh, 0, 0);
1084         }
1085         up_read((&EXT4_I(inode)->i_data_sem));
1086
1087         /* If it is only a block(s) look up */
1088         if (!create)
1089                 return retval;
1090
1091         /*
1092          * Returns if the blocks have already allocated
1093          *
1094          * Note that if blocks have been preallocated
1095          * ext4_ext_get_block() returns th create = 0
1096          * with buffer head unmapped.
1097          */
1098         if (retval > 0 && buffer_mapped(bh))
1099                 return retval;
1100
1101         /*
1102          * New blocks allocate and/or writing to uninitialized extent
1103          * will possibly result in updating i_data, so we take
1104          * the write lock of i_data_sem, and call get_blocks()
1105          * with create == 1 flag.
1106          */
1107         down_write((&EXT4_I(inode)->i_data_sem));
1108
1109         /*
1110          * if the caller is from delayed allocation writeout path
1111          * we have already reserved fs blocks for allocation
1112          * let the underlying get_block() function know to
1113          * avoid double accounting
1114          */
1115         if (flag)
1116                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1117         /*
1118          * We need to check for EXT4 here because migrate
1119          * could have changed the inode type in between
1120          */
1121         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1122                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1123                                 bh, create, extend_disksize);
1124         } else {
1125                 retval = ext4_get_blocks_handle(handle, inode, block,
1126                                 max_blocks, bh, create, extend_disksize);
1127
1128                 if (retval > 0 && buffer_new(bh)) {
1129                         /*
1130                          * We allocated new blocks which will result in
1131                          * i_data's format changing.  Force the migrate
1132                          * to fail by clearing migrate flags
1133                          */
1134                         EXT4_I(inode)->i_flags = EXT4_I(inode)->i_flags &
1135                                                         ~EXT4_EXT_MIGRATE;
1136                 }
1137         }
1138
1139         if (flag) {
1140                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1141                 /*
1142                  * Update reserved blocks/metadata blocks
1143                  * after successful block allocation
1144                  * which were deferred till now
1145                  */
1146                 if ((retval > 0) && buffer_delay(bh))
1147                         ext4_da_update_reserve_space(inode, retval);
1148         }
1149
1150         up_write((&EXT4_I(inode)->i_data_sem));
1151         return retval;
1152 }
1153
1154 /* Maximum number of blocks we map for direct IO at once. */
1155 #define DIO_MAX_BLOCKS 4096
1156
1157 int ext4_get_block(struct inode *inode, sector_t iblock,
1158                    struct buffer_head *bh_result, int create)
1159 {
1160         handle_t *handle = ext4_journal_current_handle();
1161         int ret = 0, started = 0;
1162         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1163         int dio_credits;
1164
1165         if (create && !handle) {
1166                 /* Direct IO write... */
1167                 if (max_blocks > DIO_MAX_BLOCKS)
1168                         max_blocks = DIO_MAX_BLOCKS;
1169                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
1170                 handle = ext4_journal_start(inode, dio_credits);
1171                 if (IS_ERR(handle)) {
1172                         ret = PTR_ERR(handle);
1173                         goto out;
1174                 }
1175                 started = 1;
1176         }
1177
1178         ret = ext4_get_blocks_wrap(handle, inode, iblock,
1179                                         max_blocks, bh_result, create, 0, 0);
1180         if (ret > 0) {
1181                 bh_result->b_size = (ret << inode->i_blkbits);
1182                 ret = 0;
1183         }
1184         if (started)
1185                 ext4_journal_stop(handle);
1186 out:
1187         return ret;
1188 }
1189
1190 /*
1191  * `handle' can be NULL if create is zero
1192  */
1193 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1194                                 ext4_lblk_t block, int create, int *errp)
1195 {
1196         struct buffer_head dummy;
1197         int fatal = 0, err;
1198
1199         J_ASSERT(handle != NULL || create == 0);
1200
1201         dummy.b_state = 0;
1202         dummy.b_blocknr = -1000;
1203         buffer_trace_init(&dummy.b_history);
1204         err = ext4_get_blocks_wrap(handle, inode, block, 1,
1205                                         &dummy, create, 1, 0);
1206         /*
1207          * ext4_get_blocks_handle() returns number of blocks
1208          * mapped. 0 in case of a HOLE.
1209          */
1210         if (err > 0) {
1211                 if (err > 1)
1212                         WARN_ON(1);
1213                 err = 0;
1214         }
1215         *errp = err;
1216         if (!err && buffer_mapped(&dummy)) {
1217                 struct buffer_head *bh;
1218                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1219                 if (!bh) {
1220                         *errp = -EIO;
1221                         goto err;
1222                 }
1223                 if (buffer_new(&dummy)) {
1224                         J_ASSERT(create != 0);
1225                         J_ASSERT(handle != NULL);
1226
1227                         /*
1228                          * Now that we do not always journal data, we should
1229                          * keep in mind whether this should always journal the
1230                          * new buffer as metadata.  For now, regular file
1231                          * writes use ext4_get_block instead, so it's not a
1232                          * problem.
1233                          */
1234                         lock_buffer(bh);
1235                         BUFFER_TRACE(bh, "call get_create_access");
1236                         fatal = ext4_journal_get_create_access(handle, bh);
1237                         if (!fatal && !buffer_uptodate(bh)) {
1238                                 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1239                                 set_buffer_uptodate(bh);
1240                         }
1241                         unlock_buffer(bh);
1242                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1243                         err = ext4_handle_dirty_metadata(handle, inode, bh);
1244                         if (!fatal)
1245                                 fatal = err;
1246                 } else {
1247                         BUFFER_TRACE(bh, "not a new buffer");
1248                 }
1249                 if (fatal) {
1250                         *errp = fatal;
1251                         brelse(bh);
1252                         bh = NULL;
1253                 }
1254                 return bh;
1255         }
1256 err:
1257         return NULL;
1258 }
1259
1260 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1261                                ext4_lblk_t block, int create, int *err)
1262 {
1263         struct buffer_head *bh;
1264
1265         bh = ext4_getblk(handle, inode, block, create, err);
1266         if (!bh)
1267                 return bh;
1268         if (buffer_uptodate(bh))
1269                 return bh;
1270         ll_rw_block(READ_META, 1, &bh);
1271         wait_on_buffer(bh);
1272         if (buffer_uptodate(bh))
1273                 return bh;
1274         put_bh(bh);
1275         *err = -EIO;
1276         return NULL;
1277 }
1278
1279 static int walk_page_buffers(handle_t *handle,
1280                              struct buffer_head *head,
1281                              unsigned from,
1282                              unsigned to,
1283                              int *partial,
1284                              int (*fn)(handle_t *handle,
1285                                        struct buffer_head *bh))
1286 {
1287         struct buffer_head *bh;
1288         unsigned block_start, block_end;
1289         unsigned blocksize = head->b_size;
1290         int err, ret = 0;
1291         struct buffer_head *next;
1292
1293         for (bh = head, block_start = 0;
1294              ret == 0 && (bh != head || !block_start);
1295              block_start = block_end, bh = next)
1296         {
1297                 next = bh->b_this_page;
1298                 block_end = block_start + blocksize;
1299                 if (block_end <= from || block_start >= to) {
1300                         if (partial && !buffer_uptodate(bh))
1301                                 *partial = 1;
1302                         continue;
1303                 }
1304                 err = (*fn)(handle, bh);
1305                 if (!ret)
1306                         ret = err;
1307         }
1308         return ret;
1309 }
1310
1311 /*
1312  * To preserve ordering, it is essential that the hole instantiation and
1313  * the data write be encapsulated in a single transaction.  We cannot
1314  * close off a transaction and start a new one between the ext4_get_block()
1315  * and the commit_write().  So doing the jbd2_journal_start at the start of
1316  * prepare_write() is the right place.
1317  *
1318  * Also, this function can nest inside ext4_writepage() ->
1319  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1320  * has generated enough buffer credits to do the whole page.  So we won't
1321  * block on the journal in that case, which is good, because the caller may
1322  * be PF_MEMALLOC.
1323  *
1324  * By accident, ext4 can be reentered when a transaction is open via
1325  * quota file writes.  If we were to commit the transaction while thus
1326  * reentered, there can be a deadlock - we would be holding a quota
1327  * lock, and the commit would never complete if another thread had a
1328  * transaction open and was blocking on the quota lock - a ranking
1329  * violation.
1330  *
1331  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1332  * will _not_ run commit under these circumstances because handle->h_ref
1333  * is elevated.  We'll still have enough credits for the tiny quotafile
1334  * write.
1335  */
1336 static int do_journal_get_write_access(handle_t *handle,
1337                                         struct buffer_head *bh)
1338 {
1339         if (!buffer_mapped(bh) || buffer_freed(bh))
1340                 return 0;
1341         return ext4_journal_get_write_access(handle, bh);
1342 }
1343
1344 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1345                                 loff_t pos, unsigned len, unsigned flags,
1346                                 struct page **pagep, void **fsdata)
1347 {
1348         struct inode *inode = mapping->host;
1349         int ret, needed_blocks = ext4_writepage_trans_blocks(inode);
1350         handle_t *handle;
1351         int retries = 0;
1352         struct page *page;
1353         pgoff_t index;
1354         unsigned from, to;
1355
1356         trace_mark(ext4_write_begin,
1357                    "dev %s ino %lu pos %llu len %u flags %u",
1358                    inode->i_sb->s_id, inode->i_ino,
1359                    (unsigned long long) pos, len, flags);
1360         index = pos >> PAGE_CACHE_SHIFT;
1361         from = pos & (PAGE_CACHE_SIZE - 1);
1362         to = from + len;
1363
1364 retry:
1365         handle = ext4_journal_start(inode, needed_blocks);
1366         if (IS_ERR(handle)) {
1367                 ret = PTR_ERR(handle);
1368                 goto out;
1369         }
1370
1371         page = grab_cache_page_write_begin(mapping, index, flags);
1372         if (!page) {
1373                 ext4_journal_stop(handle);
1374                 ret = -ENOMEM;
1375                 goto out;
1376         }
1377         *pagep = page;
1378
1379         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1380                                                         ext4_get_block);
1381
1382         if (!ret && ext4_should_journal_data(inode)) {
1383                 ret = walk_page_buffers(handle, page_buffers(page),
1384                                 from, to, NULL, do_journal_get_write_access);
1385         }
1386
1387         if (ret) {
1388                 unlock_page(page);
1389                 ext4_journal_stop(handle);
1390                 page_cache_release(page);
1391                 /*
1392                  * block_write_begin may have instantiated a few blocks
1393                  * outside i_size.  Trim these off again. Don't need
1394                  * i_size_read because we hold i_mutex.
1395                  */
1396                 if (pos + len > inode->i_size)
1397                         vmtruncate(inode, inode->i_size);
1398         }
1399
1400         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1401                 goto retry;
1402 out:
1403         return ret;
1404 }
1405
1406 /* For write_end() in data=journal mode */
1407 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1408 {
1409         if (!buffer_mapped(bh) || buffer_freed(bh))
1410                 return 0;
1411         set_buffer_uptodate(bh);
1412         return ext4_handle_dirty_metadata(handle, NULL, bh);
1413 }
1414
1415 /*
1416  * We need to pick up the new inode size which generic_commit_write gave us
1417  * `file' can be NULL - eg, when called from page_symlink().
1418  *
1419  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1420  * buffers are managed internally.
1421  */
1422 static int ext4_ordered_write_end(struct file *file,
1423                                 struct address_space *mapping,
1424                                 loff_t pos, unsigned len, unsigned copied,
1425                                 struct page *page, void *fsdata)
1426 {
1427         handle_t *handle = ext4_journal_current_handle();
1428         struct inode *inode = mapping->host;
1429         int ret = 0, ret2;
1430
1431         trace_mark(ext4_ordered_write_end,
1432                    "dev %s ino %lu pos %llu len %u copied %u",
1433                    inode->i_sb->s_id, inode->i_ino,
1434                    (unsigned long long) pos, len, copied);
1435         ret = ext4_jbd2_file_inode(handle, inode);
1436
1437         if (ret == 0) {
1438                 loff_t new_i_size;
1439
1440                 new_i_size = pos + copied;
1441                 if (new_i_size > EXT4_I(inode)->i_disksize) {
1442                         ext4_update_i_disksize(inode, new_i_size);
1443                         /* We need to mark inode dirty even if
1444                          * new_i_size is less that inode->i_size
1445                          * bu greater than i_disksize.(hint delalloc)
1446                          */
1447                         ext4_mark_inode_dirty(handle, inode);
1448                 }
1449
1450                 ret2 = generic_write_end(file, mapping, pos, len, copied,
1451                                                         page, fsdata);
1452                 copied = ret2;
1453                 if (ret2 < 0)
1454                         ret = ret2;
1455         }
1456         ret2 = ext4_journal_stop(handle);
1457         if (!ret)
1458                 ret = ret2;
1459
1460         return ret ? ret : copied;
1461 }
1462
1463 static int ext4_writeback_write_end(struct file *file,
1464                                 struct address_space *mapping,
1465                                 loff_t pos, unsigned len, unsigned copied,
1466                                 struct page *page, void *fsdata)
1467 {
1468         handle_t *handle = ext4_journal_current_handle();
1469         struct inode *inode = mapping->host;
1470         int ret = 0, ret2;
1471         loff_t new_i_size;
1472
1473         trace_mark(ext4_writeback_write_end,
1474                    "dev %s ino %lu pos %llu len %u copied %u",
1475                    inode->i_sb->s_id, inode->i_ino,
1476                    (unsigned long long) pos, len, copied);
1477         new_i_size = pos + copied;
1478         if (new_i_size > EXT4_I(inode)->i_disksize) {
1479                 ext4_update_i_disksize(inode, new_i_size);
1480                 /* We need to mark inode dirty even if
1481                  * new_i_size is less that inode->i_size
1482                  * bu greater than i_disksize.(hint delalloc)
1483                  */
1484                 ext4_mark_inode_dirty(handle, inode);
1485         }
1486
1487         ret2 = generic_write_end(file, mapping, pos, len, copied,
1488                                                         page, fsdata);
1489         copied = ret2;
1490         if (ret2 < 0)
1491                 ret = ret2;
1492
1493         ret2 = ext4_journal_stop(handle);
1494         if (!ret)
1495                 ret = ret2;
1496
1497         return ret ? ret : copied;
1498 }
1499
1500 static int ext4_journalled_write_end(struct file *file,
1501                                 struct address_space *mapping,
1502                                 loff_t pos, unsigned len, unsigned copied,
1503                                 struct page *page, void *fsdata)
1504 {
1505         handle_t *handle = ext4_journal_current_handle();
1506         struct inode *inode = mapping->host;
1507         int ret = 0, ret2;
1508         int partial = 0;
1509         unsigned from, to;
1510         loff_t new_i_size;
1511
1512         trace_mark(ext4_journalled_write_end,
1513                    "dev %s ino %lu pos %llu len %u copied %u",
1514                    inode->i_sb->s_id, inode->i_ino,
1515                    (unsigned long long) pos, len, copied);
1516         from = pos & (PAGE_CACHE_SIZE - 1);
1517         to = from + len;
1518
1519         if (copied < len) {
1520                 if (!PageUptodate(page))
1521                         copied = 0;
1522                 page_zero_new_buffers(page, from+copied, to);
1523         }
1524
1525         ret = walk_page_buffers(handle, page_buffers(page), from,
1526                                 to, &partial, write_end_fn);
1527         if (!partial)
1528                 SetPageUptodate(page);
1529         new_i_size = pos + copied;
1530         if (new_i_size > inode->i_size)
1531                 i_size_write(inode, pos+copied);
1532         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
1533         if (new_i_size > EXT4_I(inode)->i_disksize) {
1534                 ext4_update_i_disksize(inode, new_i_size);
1535                 ret2 = ext4_mark_inode_dirty(handle, inode);
1536                 if (!ret)
1537                         ret = ret2;
1538         }
1539
1540         unlock_page(page);
1541         ret2 = ext4_journal_stop(handle);
1542         if (!ret)
1543                 ret = ret2;
1544         page_cache_release(page);
1545
1546         return ret ? ret : copied;
1547 }
1548
1549 static int ext4_da_reserve_space(struct inode *inode, int nrblocks)
1550 {
1551         int retries = 0;
1552        struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1553        unsigned long md_needed, mdblocks, total = 0;
1554
1555         /*
1556          * recalculate the amount of metadata blocks to reserve
1557          * in order to allocate nrblocks
1558          * worse case is one extent per block
1559          */
1560 repeat:
1561         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1562         total = EXT4_I(inode)->i_reserved_data_blocks + nrblocks;
1563         mdblocks = ext4_calc_metadata_amount(inode, total);
1564         BUG_ON(mdblocks < EXT4_I(inode)->i_reserved_meta_blocks);
1565
1566         md_needed = mdblocks - EXT4_I(inode)->i_reserved_meta_blocks;
1567         total = md_needed + nrblocks;
1568
1569         if (ext4_claim_free_blocks(sbi, total)) {
1570                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1571                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1572                         yield();
1573                         goto repeat;
1574                 }
1575                 return -ENOSPC;
1576         }
1577         EXT4_I(inode)->i_reserved_data_blocks += nrblocks;
1578         EXT4_I(inode)->i_reserved_meta_blocks = mdblocks;
1579
1580         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1581         return 0;       /* success */
1582 }
1583
1584 static void ext4_da_release_space(struct inode *inode, int to_free)
1585 {
1586         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1587         int total, mdb, mdb_free, release;
1588
1589         if (!to_free)
1590                 return;         /* Nothing to release, exit */
1591
1592         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1593
1594         if (!EXT4_I(inode)->i_reserved_data_blocks) {
1595                 /*
1596                  * if there is no reserved blocks, but we try to free some
1597                  * then the counter is messed up somewhere.
1598                  * but since this function is called from invalidate
1599                  * page, it's harmless to return without any action
1600                  */
1601                 printk(KERN_INFO "ext4 delalloc try to release %d reserved "
1602                             "blocks for inode %lu, but there is no reserved "
1603                             "data blocks\n", to_free, inode->i_ino);
1604                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1605                 return;
1606         }
1607
1608         /* recalculate the number of metablocks still need to be reserved */
1609         total = EXT4_I(inode)->i_reserved_data_blocks - to_free;
1610         mdb = ext4_calc_metadata_amount(inode, total);
1611
1612         /* figure out how many metablocks to release */
1613         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1614         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1615
1616         release = to_free + mdb_free;
1617
1618         /* update fs dirty blocks counter for truncate case */
1619         percpu_counter_sub(&sbi->s_dirtyblocks_counter, release);
1620
1621         /* update per-inode reservations */
1622         BUG_ON(to_free > EXT4_I(inode)->i_reserved_data_blocks);
1623         EXT4_I(inode)->i_reserved_data_blocks -= to_free;
1624
1625         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1626         EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1627         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1628 }
1629
1630 static void ext4_da_page_release_reservation(struct page *page,
1631                                                 unsigned long offset)
1632 {
1633         int to_release = 0;
1634         struct buffer_head *head, *bh;
1635         unsigned int curr_off = 0;
1636
1637         head = page_buffers(page);
1638         bh = head;
1639         do {
1640                 unsigned int next_off = curr_off + bh->b_size;
1641
1642                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1643                         to_release++;
1644                         clear_buffer_delay(bh);
1645                 }
1646                 curr_off = next_off;
1647         } while ((bh = bh->b_this_page) != head);
1648         ext4_da_release_space(page->mapping->host, to_release);
1649 }
1650
1651 /*
1652  * Delayed allocation stuff
1653  */
1654
1655 struct mpage_da_data {
1656         struct inode *inode;
1657         struct buffer_head lbh;                 /* extent of blocks */
1658         unsigned long first_page, next_page;    /* extent of pages */
1659         get_block_t *get_block;
1660         struct writeback_control *wbc;
1661         int io_done;
1662         int pages_written;
1663         int retval;
1664 };
1665
1666 /*
1667  * mpage_da_submit_io - walks through extent of pages and try to write
1668  * them with writepage() call back
1669  *
1670  * @mpd->inode: inode
1671  * @mpd->first_page: first page of the extent
1672  * @mpd->next_page: page after the last page of the extent
1673  * @mpd->get_block: the filesystem's block mapper function
1674  *
1675  * By the time mpage_da_submit_io() is called we expect all blocks
1676  * to be allocated. this may be wrong if allocation failed.
1677  *
1678  * As pages are already locked by write_cache_pages(), we can't use it
1679  */
1680 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1681 {
1682         long pages_skipped;
1683         struct pagevec pvec;
1684         unsigned long index, end;
1685         int ret = 0, err, nr_pages, i;
1686         struct inode *inode = mpd->inode;
1687         struct address_space *mapping = inode->i_mapping;
1688
1689         BUG_ON(mpd->next_page <= mpd->first_page);
1690         /*
1691          * We need to start from the first_page to the next_page - 1
1692          * to make sure we also write the mapped dirty buffer_heads.
1693          * If we look at mpd->lbh.b_blocknr we would only be looking
1694          * at the currently mapped buffer_heads.
1695          */
1696         index = mpd->first_page;
1697         end = mpd->next_page - 1;
1698
1699         pagevec_init(&pvec, 0);
1700         while (index <= end) {
1701                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1702                 if (nr_pages == 0)
1703                         break;
1704                 for (i = 0; i < nr_pages; i++) {
1705                         struct page *page = pvec.pages[i];
1706
1707                         index = page->index;
1708                         if (index > end)
1709                                 break;
1710                         index++;
1711
1712                         BUG_ON(!PageLocked(page));
1713                         BUG_ON(PageWriteback(page));
1714
1715                         pages_skipped = mpd->wbc->pages_skipped;
1716                         err = mapping->a_ops->writepage(page, mpd->wbc);
1717                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
1718                                 /*
1719                                  * have successfully written the page
1720                                  * without skipping the same
1721                                  */
1722                                 mpd->pages_written++;
1723                         /*
1724                          * In error case, we have to continue because
1725                          * remaining pages are still locked
1726                          * XXX: unlock and re-dirty them?
1727                          */
1728                         if (ret == 0)
1729                                 ret = err;
1730                 }
1731                 pagevec_release(&pvec);
1732         }
1733         return ret;
1734 }
1735
1736 /*
1737  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
1738  *
1739  * @mpd->inode - inode to walk through
1740  * @exbh->b_blocknr - first block on a disk
1741  * @exbh->b_size - amount of space in bytes
1742  * @logical - first logical block to start assignment with
1743  *
1744  * the function goes through all passed space and put actual disk
1745  * block numbers into buffer heads, dropping BH_Delay
1746  */
1747 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
1748                                  struct buffer_head *exbh)
1749 {
1750         struct inode *inode = mpd->inode;
1751         struct address_space *mapping = inode->i_mapping;
1752         int blocks = exbh->b_size >> inode->i_blkbits;
1753         sector_t pblock = exbh->b_blocknr, cur_logical;
1754         struct buffer_head *head, *bh;
1755         pgoff_t index, end;
1756         struct pagevec pvec;
1757         int nr_pages, i;
1758
1759         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1760         end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1761         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1762
1763         pagevec_init(&pvec, 0);
1764
1765         while (index <= end) {
1766                 /* XXX: optimize tail */
1767                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1768                 if (nr_pages == 0)
1769                         break;
1770                 for (i = 0; i < nr_pages; i++) {
1771                         struct page *page = pvec.pages[i];
1772
1773                         index = page->index;
1774                         if (index > end)
1775                                 break;
1776                         index++;
1777
1778                         BUG_ON(!PageLocked(page));
1779                         BUG_ON(PageWriteback(page));
1780                         BUG_ON(!page_has_buffers(page));
1781
1782                         bh = page_buffers(page);
1783                         head = bh;
1784
1785                         /* skip blocks out of the range */
1786                         do {
1787                                 if (cur_logical >= logical)
1788                                         break;
1789                                 cur_logical++;
1790                         } while ((bh = bh->b_this_page) != head);
1791
1792                         do {
1793                                 if (cur_logical >= logical + blocks)
1794                                         break;
1795                                 if (buffer_delay(bh)) {
1796                                         bh->b_blocknr = pblock;
1797                                         clear_buffer_delay(bh);
1798                                         bh->b_bdev = inode->i_sb->s_bdev;
1799                                 } else if (buffer_unwritten(bh)) {
1800                                         bh->b_blocknr = pblock;
1801                                         clear_buffer_unwritten(bh);
1802                                         set_buffer_mapped(bh);
1803                                         set_buffer_new(bh);
1804                                         bh->b_bdev = inode->i_sb->s_bdev;
1805                                 } else if (buffer_mapped(bh))
1806                                         BUG_ON(bh->b_blocknr != pblock);
1807
1808                                 cur_logical++;
1809                                 pblock++;
1810                         } while ((bh = bh->b_this_page) != head);
1811                 }
1812                 pagevec_release(&pvec);
1813         }
1814 }
1815
1816
1817 /*
1818  * __unmap_underlying_blocks - just a helper function to unmap
1819  * set of blocks described by @bh
1820  */
1821 static inline void __unmap_underlying_blocks(struct inode *inode,
1822                                              struct buffer_head *bh)
1823 {
1824         struct block_device *bdev = inode->i_sb->s_bdev;
1825         int blocks, i;
1826
1827         blocks = bh->b_size >> inode->i_blkbits;
1828         for (i = 0; i < blocks; i++)
1829                 unmap_underlying_metadata(bdev, bh->b_blocknr + i);
1830 }
1831
1832 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
1833                                         sector_t logical, long blk_cnt)
1834 {
1835         int nr_pages, i;
1836         pgoff_t index, end;
1837         struct pagevec pvec;
1838         struct inode *inode = mpd->inode;
1839         struct address_space *mapping = inode->i_mapping;
1840
1841         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1842         end   = (logical + blk_cnt - 1) >>
1843                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
1844         while (index <= end) {
1845                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1846                 if (nr_pages == 0)
1847                         break;
1848                 for (i = 0; i < nr_pages; i++) {
1849                         struct page *page = pvec.pages[i];
1850                         index = page->index;
1851                         if (index > end)
1852                                 break;
1853                         index++;
1854
1855                         BUG_ON(!PageLocked(page));
1856                         BUG_ON(PageWriteback(page));
1857                         block_invalidatepage(page, 0);
1858                         ClearPageUptodate(page);
1859                         unlock_page(page);
1860                 }
1861         }
1862         return;
1863 }
1864
1865 static void ext4_print_free_blocks(struct inode *inode)
1866 {
1867         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1868         printk(KERN_EMERG "Total free blocks count %lld\n",
1869                         ext4_count_free_blocks(inode->i_sb));
1870         printk(KERN_EMERG "Free/Dirty block details\n");
1871         printk(KERN_EMERG "free_blocks=%lld\n",
1872                         (long long)percpu_counter_sum(&sbi->s_freeblocks_counter));
1873         printk(KERN_EMERG "dirty_blocks=%lld\n",
1874                         (long long)percpu_counter_sum(&sbi->s_dirtyblocks_counter));
1875         printk(KERN_EMERG "Block reservation details\n");
1876         printk(KERN_EMERG "i_reserved_data_blocks=%u\n",
1877                         EXT4_I(inode)->i_reserved_data_blocks);
1878         printk(KERN_EMERG "i_reserved_meta_blocks=%u\n",
1879                         EXT4_I(inode)->i_reserved_meta_blocks);
1880         return;
1881 }
1882
1883 /*
1884  * mpage_da_map_blocks - go through given space
1885  *
1886  * @mpd->lbh - bh describing space
1887  * @mpd->get_block - the filesystem's block mapper function
1888  *
1889  * The function skips space we know is already mapped to disk blocks.
1890  *
1891  */
1892 static int  mpage_da_map_blocks(struct mpage_da_data *mpd)
1893 {
1894         int err = 0;
1895         struct buffer_head new;
1896         struct buffer_head *lbh = &mpd->lbh;
1897         sector_t next;
1898
1899         /*
1900          * We consider only non-mapped and non-allocated blocks
1901          */
1902         if (buffer_mapped(lbh) && !buffer_delay(lbh))
1903                 return 0;
1904         new.b_state = lbh->b_state;
1905         new.b_blocknr = 0;
1906         new.b_size = lbh->b_size;
1907         next = lbh->b_blocknr;
1908         /*
1909          * If we didn't accumulate anything
1910          * to write simply return
1911          */
1912         if (!new.b_size)
1913                 return 0;
1914         err = mpd->get_block(mpd->inode, next, &new, 1);
1915         if (err) {
1916
1917                 /* If get block returns with error
1918                  * we simply return. Later writepage
1919                  * will redirty the page and writepages
1920                  * will find the dirty page again
1921                  */
1922                 if (err == -EAGAIN)
1923                         return 0;
1924
1925                 if (err == -ENOSPC &&
1926                                 ext4_count_free_blocks(mpd->inode->i_sb)) {
1927                         mpd->retval = err;
1928                         return 0;
1929                 }
1930
1931                 /*
1932                  * get block failure will cause us
1933                  * to loop in writepages. Because
1934                  * a_ops->writepage won't be able to
1935                  * make progress. The page will be redirtied
1936                  * by writepage and writepages will again
1937                  * try to write the same.
1938                  */
1939                 printk(KERN_EMERG "%s block allocation failed for inode %lu "
1940                                   "at logical offset %llu with max blocks "
1941                                   "%zd with error %d\n",
1942                                   __func__, mpd->inode->i_ino,
1943                                   (unsigned long long)next,
1944                                   lbh->b_size >> mpd->inode->i_blkbits, err);
1945                 printk(KERN_EMERG "This should not happen.!! "
1946                                         "Data will be lost\n");
1947                 if (err == -ENOSPC) {
1948                         ext4_print_free_blocks(mpd->inode);
1949                 }
1950                 /* invlaidate all the pages */
1951                 ext4_da_block_invalidatepages(mpd, next,
1952                                 lbh->b_size >> mpd->inode->i_blkbits);
1953                 return err;
1954         }
1955         BUG_ON(new.b_size == 0);
1956
1957         if (buffer_new(&new))
1958                 __unmap_underlying_blocks(mpd->inode, &new);
1959
1960         /*
1961          * If blocks are delayed marked, we need to
1962          * put actual blocknr and drop delayed bit
1963          */
1964         if (buffer_delay(lbh) || buffer_unwritten(lbh))
1965                 mpage_put_bnr_to_bhs(mpd, next, &new);
1966
1967         return 0;
1968 }
1969
1970 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
1971                 (1 << BH_Delay) | (1 << BH_Unwritten))
1972
1973 /*
1974  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
1975  *
1976  * @mpd->lbh - extent of blocks
1977  * @logical - logical number of the block in the file
1978  * @bh - bh of the block (used to access block's state)
1979  *
1980  * the function is used to collect contig. blocks in same state
1981  */
1982 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
1983                                    sector_t logical, struct buffer_head *bh)
1984 {
1985         sector_t next;
1986         size_t b_size = bh->b_size;
1987         struct buffer_head *lbh = &mpd->lbh;
1988         int nrblocks = lbh->b_size >> mpd->inode->i_blkbits;
1989
1990         /* check if thereserved journal credits might overflow */
1991         if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
1992                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
1993                         /*
1994                          * With non-extent format we are limited by the journal
1995                          * credit available.  Total credit needed to insert
1996                          * nrblocks contiguous blocks is dependent on the
1997                          * nrblocks.  So limit nrblocks.
1998                          */
1999                         goto flush_it;
2000                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2001                                 EXT4_MAX_TRANS_DATA) {
2002                         /*
2003                          * Adding the new buffer_head would make it cross the
2004                          * allowed limit for which we have journal credit
2005                          * reserved. So limit the new bh->b_size
2006                          */
2007                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2008                                                 mpd->inode->i_blkbits;
2009                         /* we will do mpage_da_submit_io in the next loop */
2010                 }
2011         }
2012         /*
2013          * First block in the extent
2014          */
2015         if (lbh->b_size == 0) {
2016                 lbh->b_blocknr = logical;
2017                 lbh->b_size = b_size;
2018                 lbh->b_state = bh->b_state & BH_FLAGS;
2019                 return;
2020         }
2021
2022         next = lbh->b_blocknr + nrblocks;
2023         /*
2024          * Can we merge the block to our big extent?
2025          */
2026         if (logical == next && (bh->b_state & BH_FLAGS) == lbh->b_state) {
2027                 lbh->b_size += b_size;
2028                 return;
2029         }
2030
2031 flush_it:
2032         /*
2033          * We couldn't merge the block to our extent, so we
2034          * need to flush current  extent and start new one
2035          */
2036         if (mpage_da_map_blocks(mpd) == 0)
2037                 mpage_da_submit_io(mpd);
2038         mpd->io_done = 1;
2039         return;
2040 }
2041
2042 /*
2043  * __mpage_da_writepage - finds extent of pages and blocks
2044  *
2045  * @page: page to consider
2046  * @wbc: not used, we just follow rules
2047  * @data: context
2048  *
2049  * The function finds extents of pages and scan them for all blocks.
2050  */
2051 static int __mpage_da_writepage(struct page *page,
2052                                 struct writeback_control *wbc, void *data)
2053 {
2054         struct mpage_da_data *mpd = data;
2055         struct inode *inode = mpd->inode;
2056         struct buffer_head *bh, *head, fake;
2057         sector_t logical;
2058
2059         if (mpd->io_done) {
2060                 /*
2061                  * Rest of the page in the page_vec
2062                  * redirty then and skip then. We will
2063                  * try to to write them again after
2064                  * starting a new transaction
2065                  */
2066                 redirty_page_for_writepage(wbc, page);
2067                 unlock_page(page);
2068                 return MPAGE_DA_EXTENT_TAIL;
2069         }
2070         /*
2071          * Can we merge this page to current extent?
2072          */
2073         if (mpd->next_page != page->index) {
2074                 /*
2075                  * Nope, we can't. So, we map non-allocated blocks
2076                  * and start IO on them using writepage()
2077                  */
2078                 if (mpd->next_page != mpd->first_page) {
2079                         if (mpage_da_map_blocks(mpd) == 0)
2080                                 mpage_da_submit_io(mpd);
2081                         /*
2082                          * skip rest of the page in the page_vec
2083                          */
2084                         mpd->io_done = 1;
2085                         redirty_page_for_writepage(wbc, page);
2086                         unlock_page(page);
2087                         return MPAGE_DA_EXTENT_TAIL;
2088                 }
2089
2090                 /*
2091                  * Start next extent of pages ...
2092                  */
2093                 mpd->first_page = page->index;
2094
2095                 /*
2096                  * ... and blocks
2097                  */
2098                 mpd->lbh.b_size = 0;
2099                 mpd->lbh.b_state = 0;
2100                 mpd->lbh.b_blocknr = 0;
2101         }
2102
2103         mpd->next_page = page->index + 1;
2104         logical = (sector_t) page->index <<
2105                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2106
2107         if (!page_has_buffers(page)) {
2108                 /*
2109                  * There is no attached buffer heads yet (mmap?)
2110                  * we treat the page asfull of dirty blocks
2111                  */
2112                 bh = &fake;
2113                 bh->b_size = PAGE_CACHE_SIZE;
2114                 bh->b_state = 0;
2115                 set_buffer_dirty(bh);
2116                 set_buffer_uptodate(bh);
2117                 mpage_add_bh_to_extent(mpd, logical, bh);
2118                 if (mpd->io_done)
2119                         return MPAGE_DA_EXTENT_TAIL;
2120         } else {
2121                 /*
2122                  * Page with regular buffer heads, just add all dirty ones
2123                  */
2124                 head = page_buffers(page);
2125                 bh = head;
2126                 do {
2127                         BUG_ON(buffer_locked(bh));
2128                         /*
2129                          * We need to try to allocate
2130                          * unmapped blocks in the same page.
2131                          * Otherwise we won't make progress
2132                          * with the page in ext4_da_writepage
2133                          */
2134                         if (buffer_dirty(bh) &&
2135                                 (!buffer_mapped(bh) || buffer_delay(bh))) {
2136                                 mpage_add_bh_to_extent(mpd, logical, bh);
2137                                 if (mpd->io_done)
2138                                         return MPAGE_DA_EXTENT_TAIL;
2139                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2140                                 /*
2141                                  * mapped dirty buffer. We need to update
2142                                  * the b_state because we look at
2143                                  * b_state in mpage_da_map_blocks. We don't
2144                                  * update b_size because if we find an
2145                                  * unmapped buffer_head later we need to
2146                                  * use the b_state flag of that buffer_head.
2147                                  */
2148                                 if (mpd->lbh.b_size == 0)
2149                                         mpd->lbh.b_state =
2150                                                 bh->b_state & BH_FLAGS;
2151                         }
2152                         logical++;
2153                 } while ((bh = bh->b_this_page) != head);
2154         }
2155
2156         return 0;
2157 }
2158
2159 /*
2160  * mpage_da_writepages - walk the list of dirty pages of the given
2161  * address space, allocates non-allocated blocks, maps newly-allocated
2162  * blocks to existing bhs and issue IO them
2163  *
2164  * @mapping: address space structure to write
2165  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2166  * @get_block: the filesystem's block mapper function.
2167  *
2168  * This is a library function, which implements the writepages()
2169  * address_space_operation.
2170  */
2171 static int mpage_da_writepages(struct address_space *mapping,
2172                                struct writeback_control *wbc,
2173                                struct mpage_da_data *mpd)
2174 {
2175         int ret;
2176
2177         if (!mpd->get_block)
2178                 return generic_writepages(mapping, wbc);
2179
2180         mpd->lbh.b_size = 0;
2181         mpd->lbh.b_state = 0;
2182         mpd->lbh.b_blocknr = 0;
2183         mpd->first_page = 0;
2184         mpd->next_page = 0;
2185         mpd->io_done = 0;
2186         mpd->pages_written = 0;
2187         mpd->retval = 0;
2188
2189         ret = write_cache_pages(mapping, wbc, __mpage_da_writepage, mpd);
2190         /*
2191          * Handle last extent of pages
2192          */
2193         if (!mpd->io_done && mpd->next_page != mpd->first_page) {
2194                 if (mpage_da_map_blocks(mpd) == 0)
2195                         mpage_da_submit_io(mpd);
2196
2197                 mpd->io_done = 1;
2198                 ret = MPAGE_DA_EXTENT_TAIL;
2199         }
2200         wbc->nr_to_write -= mpd->pages_written;
2201         return ret;
2202 }
2203
2204 /*
2205  * this is a special callback for ->write_begin() only
2206  * it's intention is to return mapped block or reserve space
2207  */
2208 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2209                                   struct buffer_head *bh_result, int create)
2210 {
2211         int ret = 0;
2212
2213         BUG_ON(create == 0);
2214         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2215
2216         /*
2217          * first, we need to know whether the block is allocated already
2218          * preallocated blocks are unmapped but should treated
2219          * the same as allocated blocks.
2220          */
2221         ret = ext4_get_blocks_wrap(NULL, inode, iblock, 1,  bh_result, 0, 0, 0);
2222         if ((ret == 0) && !buffer_delay(bh_result)) {
2223                 /* the block isn't (pre)allocated yet, let's reserve space */
2224                 /*
2225                  * XXX: __block_prepare_write() unmaps passed block,
2226                  * is it OK?
2227                  */
2228                 ret = ext4_da_reserve_space(inode, 1);
2229                 if (ret)
2230                         /* not enough space to reserve */
2231                         return ret;
2232
2233                 map_bh(bh_result, inode->i_sb, 0);
2234                 set_buffer_new(bh_result);
2235                 set_buffer_delay(bh_result);
2236         } else if (ret > 0) {
2237                 bh_result->b_size = (ret << inode->i_blkbits);
2238                 ret = 0;
2239         }
2240
2241         return ret;
2242 }
2243 #define         EXT4_DELALLOC_RSVED     1
2244 static int ext4_da_get_block_write(struct inode *inode, sector_t iblock,
2245                                    struct buffer_head *bh_result, int create)
2246 {
2247         int ret;
2248         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2249         loff_t disksize = EXT4_I(inode)->i_disksize;
2250         handle_t *handle = NULL;
2251
2252         handle = ext4_journal_current_handle();
2253         BUG_ON(!handle);
2254         ret = ext4_get_blocks_wrap(handle, inode, iblock, max_blocks,
2255                         bh_result, create, 0, EXT4_DELALLOC_RSVED);
2256         if (ret > 0) {
2257
2258                 bh_result->b_size = (ret << inode->i_blkbits);
2259
2260                 if (ext4_should_order_data(inode)) {
2261                         int retval;
2262                         retval = ext4_jbd2_file_inode(handle, inode);
2263                         if (retval)
2264                                 /*
2265                                  * Failed to add inode for ordered
2266                                  * mode. Don't update file size
2267                                  */
2268                                 return retval;
2269                 }
2270
2271                 /*
2272                  * Update on-disk size along with block allocation
2273                  * we don't use 'extend_disksize' as size may change
2274                  * within already allocated block -bzzz
2275                  */
2276                 disksize = ((loff_t) iblock + ret) << inode->i_blkbits;
2277                 if (disksize > i_size_read(inode))
2278                         disksize = i_size_read(inode);
2279                 if (disksize > EXT4_I(inode)->i_disksize) {
2280                         ext4_update_i_disksize(inode, disksize);
2281                         ret = ext4_mark_inode_dirty(handle, inode);
2282                         return ret;
2283                 }
2284                 ret = 0;
2285         }
2286         return ret;
2287 }
2288
2289 static int ext4_bh_unmapped_or_delay(handle_t *handle, struct buffer_head *bh)
2290 {
2291         /*
2292          * unmapped buffer is possible for holes.
2293          * delay buffer is possible with delayed allocation
2294          */
2295         return ((!buffer_mapped(bh) || buffer_delay(bh)) && buffer_dirty(bh));
2296 }
2297
2298 static int ext4_normal_get_block_write(struct inode *inode, sector_t iblock,
2299                                    struct buffer_head *bh_result, int create)
2300 {
2301         int ret = 0;
2302         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2303
2304         /*
2305          * we don't want to do block allocation in writepage
2306          * so call get_block_wrap with create = 0
2307          */
2308         ret = ext4_get_blocks_wrap(NULL, inode, iblock, max_blocks,
2309                                    bh_result, 0, 0, 0);
2310         if (ret > 0) {
2311                 bh_result->b_size = (ret << inode->i_blkbits);
2312                 ret = 0;
2313         }
2314         return ret;
2315 }
2316
2317 /*
2318  * get called vi ext4_da_writepages after taking page lock (have journal handle)
2319  * get called via journal_submit_inode_data_buffers (no journal handle)
2320  * get called via shrink_page_list via pdflush (no journal handle)
2321  * or grab_page_cache when doing write_begin (have journal handle)
2322  */
2323 static int ext4_da_writepage(struct page *page,
2324                                 struct writeback_control *wbc)
2325 {
2326         int ret = 0;
2327         loff_t size;
2328         unsigned int len;
2329         struct buffer_head *page_bufs;
2330         struct inode *inode = page->mapping->host;
2331
2332         trace_mark(ext4_da_writepage,
2333                    "dev %s ino %lu page_index %lu",
2334                    inode->i_sb->s_id, inode->i_ino, page->index);
2335         size = i_size_read(inode);
2336         if (page->index == size >> PAGE_CACHE_SHIFT)
2337                 len = size & ~PAGE_CACHE_MASK;
2338         else
2339                 len = PAGE_CACHE_SIZE;
2340
2341         if (page_has_buffers(page)) {
2342                 page_bufs = page_buffers(page);
2343                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2344                                         ext4_bh_unmapped_or_delay)) {
2345                         /*
2346                          * We don't want to do  block allocation
2347                          * So redirty the page and return
2348                          * We may reach here when we do a journal commit
2349                          * via journal_submit_inode_data_buffers.
2350                          * If we don't have mapping block we just ignore
2351                          * them. We can also reach here via shrink_page_list
2352                          */
2353                         redirty_page_for_writepage(wbc, page);
2354                         unlock_page(page);
2355                         return 0;
2356                 }
2357         } else {
2358                 /*
2359                  * The test for page_has_buffers() is subtle:
2360                  * We know the page is dirty but it lost buffers. That means
2361                  * that at some moment in time after write_begin()/write_end()
2362                  * has been called all buffers have been clean and thus they
2363                  * must have been written at least once. So they are all
2364                  * mapped and we can happily proceed with mapping them
2365                  * and writing the page.
2366                  *
2367                  * Try to initialize the buffer_heads and check whether
2368                  * all are mapped and non delay. We don't want to
2369                  * do block allocation here.
2370                  */
2371                 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
2372                                                 ext4_normal_get_block_write);
2373                 if (!ret) {
2374                         page_bufs = page_buffers(page);
2375                         /* check whether all are mapped and non delay */
2376                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2377                                                 ext4_bh_unmapped_or_delay)) {
2378                                 redirty_page_for_writepage(wbc, page);
2379                                 unlock_page(page);
2380                                 return 0;
2381                         }
2382                 } else {
2383                         /*
2384                          * We can't do block allocation here
2385                          * so just redity the page and unlock
2386                          * and return
2387                          */
2388                         redirty_page_for_writepage(wbc, page);
2389                         unlock_page(page);
2390                         return 0;
2391                 }
2392                 /* now mark the buffer_heads as dirty and uptodate */
2393                 block_commit_write(page, 0, PAGE_CACHE_SIZE);
2394         }
2395
2396         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2397                 ret = nobh_writepage(page, ext4_normal_get_block_write, wbc);
2398         else
2399                 ret = block_write_full_page(page,
2400                                                 ext4_normal_get_block_write,
2401                                                 wbc);
2402
2403         return ret;
2404 }
2405
2406 /*
2407  * This is called via ext4_da_writepages() to
2408  * calulate the total number of credits to reserve to fit
2409  * a single extent allocation into a single transaction,
2410  * ext4_da_writpeages() will loop calling this before
2411  * the block allocation.
2412  */
2413
2414 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2415 {
2416         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2417
2418         /*
2419          * With non-extent format the journal credit needed to
2420          * insert nrblocks contiguous block is dependent on
2421          * number of contiguous block. So we will limit
2422          * number of contiguous block to a sane value
2423          */
2424         if (!(inode->i_flags & EXT4_EXTENTS_FL) &&
2425             (max_blocks > EXT4_MAX_TRANS_DATA))
2426                 max_blocks = EXT4_MAX_TRANS_DATA;
2427
2428         return ext4_chunk_trans_blocks(inode, max_blocks);
2429 }
2430
2431 static int ext4_da_writepages(struct address_space *mapping,
2432                               struct writeback_control *wbc)
2433 {
2434         pgoff_t index;
2435         int range_whole = 0;
2436         handle_t *handle = NULL;
2437         struct mpage_da_data mpd;
2438         struct inode *inode = mapping->host;
2439         int no_nrwrite_index_update;
2440         int pages_written = 0;
2441         long pages_skipped;
2442         int range_cyclic, cycled = 1, io_done = 0;
2443         int needed_blocks, ret = 0, nr_to_writebump = 0;
2444         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2445
2446         trace_mark(ext4_da_writepages,
2447                    "dev %s ino %lu nr_t_write %ld "
2448                    "pages_skipped %ld range_start %llu "
2449                    "range_end %llu nonblocking %d "
2450                    "for_kupdate %d for_reclaim %d "
2451                    "for_writepages %d range_cyclic %d",
2452                    inode->i_sb->s_id, inode->i_ino,
2453                    wbc->nr_to_write, wbc->pages_skipped,
2454                    (unsigned long long) wbc->range_start,
2455                    (unsigned long long) wbc->range_end,
2456                    wbc->nonblocking, wbc->for_kupdate,
2457                    wbc->for_reclaim, wbc->for_writepages,
2458                    wbc->range_cyclic);
2459
2460         /*
2461          * No pages to write? This is mainly a kludge to avoid starting
2462          * a transaction for special inodes like journal inode on last iput()
2463          * because that could violate lock ordering on umount
2464          */
2465         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2466                 return 0;
2467
2468         /*
2469          * If the filesystem has aborted, it is read-only, so return
2470          * right away instead of dumping stack traces later on that
2471          * will obscure the real source of the problem.  We test
2472          * EXT4_MOUNT_ABORT instead of sb->s_flag's MS_RDONLY because
2473          * the latter could be true if the filesystem is mounted
2474          * read-only, and in that case, ext4_da_writepages should
2475          * *never* be called, so if that ever happens, we would want
2476          * the stack trace.
2477          */
2478         if (unlikely(sbi->s_mount_opt & EXT4_MOUNT_ABORT))
2479                 return -EROFS;
2480
2481         /*
2482          * Make sure nr_to_write is >= sbi->s_mb_stream_request
2483          * This make sure small files blocks are allocated in
2484          * single attempt. This ensure that small files
2485          * get less fragmented.
2486          */
2487         if (wbc->nr_to_write < sbi->s_mb_stream_request) {
2488                 nr_to_writebump = sbi->s_mb_stream_request - wbc->nr_to_write;
2489                 wbc->nr_to_write = sbi->s_mb_stream_request;
2490         }
2491         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2492                 range_whole = 1;
2493
2494         range_cyclic = wbc->range_cyclic;
2495         if (wbc->range_cyclic) {
2496                 index = mapping->writeback_index;
2497                 if (index)
2498                         cycled = 0;
2499                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2500                 wbc->range_end  = LLONG_MAX;
2501                 wbc->range_cyclic = 0;
2502         } else
2503                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2504
2505         mpd.wbc = wbc;
2506         mpd.inode = mapping->host;
2507
2508         /*
2509          * we don't want write_cache_pages to update
2510          * nr_to_write and writeback_index
2511          */
2512         no_nrwrite_index_update = wbc->no_nrwrite_index_update;
2513         wbc->no_nrwrite_index_update = 1;
2514         pages_skipped = wbc->pages_skipped;
2515
2516 retry:
2517         while (!ret && wbc->nr_to_write > 0) {
2518
2519                 /*
2520                  * we  insert one extent at a time. So we need
2521                  * credit needed for single extent allocation.
2522                  * journalled mode is currently not supported
2523                  * by delalloc
2524                  */
2525                 BUG_ON(ext4_should_journal_data(inode));
2526                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2527
2528                 /* start a new transaction*/
2529                 handle = ext4_journal_start(inode, needed_blocks);
2530                 if (IS_ERR(handle)) {
2531                         ret = PTR_ERR(handle);
2532                         printk(KERN_CRIT "%s: jbd2_start: "
2533                                "%ld pages, ino %lu; err %d\n", __func__,
2534                                 wbc->nr_to_write, inode->i_ino, ret);
2535                         dump_stack();
2536                         goto out_writepages;
2537                 }
2538                 mpd.get_block = ext4_da_get_block_write;
2539                 ret = mpage_da_writepages(mapping, wbc, &mpd);
2540
2541                 ext4_journal_stop(handle);
2542
2543                 if (mpd.retval == -ENOSPC) {
2544                         /* commit the transaction which would
2545                          * free blocks released in the transaction
2546                          * and try again
2547                          */
2548                         jbd2_journal_force_commit_nested(sbi->s_journal);
2549                         wbc->pages_skipped = pages_skipped;
2550                         ret = 0;
2551                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2552                         /*
2553                          * got one extent now try with
2554                          * rest of the pages
2555                          */
2556                         pages_written += mpd.pages_written;
2557                         wbc->pages_skipped = pages_skipped;
2558                         ret = 0;
2559                         io_done = 1;
2560                 } else if (wbc->nr_to_write)
2561                         /*
2562                          * There is no more writeout needed
2563                          * or we requested for a noblocking writeout
2564                          * and we found the device congested
2565                          */
2566                         break;
2567         }
2568         if (!io_done && !cycled) {
2569                 cycled = 1;
2570                 index = 0;
2571                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2572                 wbc->range_end  = mapping->writeback_index - 1;
2573                 goto retry;
2574         }
2575         if (pages_skipped != wbc->pages_skipped)
2576                 printk(KERN_EMERG "This should not happen leaving %s "
2577                                 "with nr_to_write = %ld ret = %d\n",
2578                                 __func__, wbc->nr_to_write, ret);
2579
2580         /* Update index */
2581         index += pages_written;
2582         wbc->range_cyclic = range_cyclic;
2583         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2584                 /*
2585                  * set the writeback_index so that range_cyclic
2586                  * mode will write it back later
2587                  */
2588                 mapping->writeback_index = index;
2589
2590 out_writepages:
2591         if (!no_nrwrite_index_update)
2592                 wbc->no_nrwrite_index_update = 0;
2593         wbc->nr_to_write -= nr_to_writebump;
2594         trace_mark(ext4_da_writepage_result,
2595                    "dev %s ino %lu ret %d pages_written %d "
2596                    "pages_skipped %ld congestion %d "
2597                    "more_io %d no_nrwrite_index_update %d",
2598                    inode->i_sb->s_id, inode->i_ino, ret,
2599                    pages_written, wbc->pages_skipped,
2600                    wbc->encountered_congestion, wbc->more_io,
2601                    wbc->no_nrwrite_index_update);
2602         return ret;
2603 }
2604
2605 #define FALL_BACK_TO_NONDELALLOC 1
2606 static int ext4_nonda_switch(struct super_block *sb)
2607 {
2608         s64 free_blocks, dirty_blocks;
2609         struct ext4_sb_info *sbi = EXT4_SB(sb);
2610
2611         /*
2612          * switch to non delalloc mode if we are running low
2613          * on free block. The free block accounting via percpu
2614          * counters can get slightly wrong with percpu_counter_batch getting
2615          * accumulated on each CPU without updating global counters
2616          * Delalloc need an accurate free block accounting. So switch
2617          * to non delalloc when we are near to error range.
2618          */
2619         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
2620         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
2621         if (2 * free_blocks < 3 * dirty_blocks ||
2622                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
2623                 /*
2624                  * free block count is less that 150% of dirty blocks
2625                  * or free blocks is less that watermark
2626                  */
2627                 return 1;
2628         }
2629         return 0;
2630 }
2631
2632 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2633                                 loff_t pos, unsigned len, unsigned flags,
2634                                 struct page **pagep, void **fsdata)
2635 {
2636         int ret, retries = 0;
2637         struct page *page;
2638         pgoff_t index;
2639         unsigned from, to;
2640         struct inode *inode = mapping->host;
2641         handle_t *handle;
2642
2643         index = pos >> PAGE_CACHE_SHIFT;
2644         from = pos & (PAGE_CACHE_SIZE - 1);
2645         to = from + len;
2646
2647         if (ext4_nonda_switch(inode->i_sb)) {
2648                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2649                 return ext4_write_begin(file, mapping, pos,
2650                                         len, flags, pagep, fsdata);
2651         }
2652         *fsdata = (void *)0;
2653
2654         trace_mark(ext4_da_write_begin,
2655                    "dev %s ino %lu pos %llu len %u flags %u",
2656                    inode->i_sb->s_id, inode->i_ino,
2657                    (unsigned long long) pos, len, flags);
2658 retry:
2659         /*
2660          * With delayed allocation, we don't log the i_disksize update
2661          * if there is delayed block allocation. But we still need
2662          * to journalling the i_disksize update if writes to the end
2663          * of file which has an already mapped buffer.
2664          */
2665         handle = ext4_journal_start(inode, 1);
2666         if (IS_ERR(handle)) {
2667                 ret = PTR_ERR(handle);
2668                 goto out;
2669         }
2670
2671         page = grab_cache_page_write_begin(mapping, index, flags);
2672         if (!page) {
2673                 ext4_journal_stop(handle);
2674                 ret = -ENOMEM;
2675                 goto out;
2676         }
2677         *pagep = page;
2678
2679         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
2680                                                         ext4_da_get_block_prep);
2681         if (ret < 0) {
2682                 unlock_page(page);
2683                 ext4_journal_stop(handle);
2684                 page_cache_release(page);
2685                 /*
2686                  * block_write_begin may have instantiated a few blocks
2687                  * outside i_size.  Trim these off again. Don't need
2688                  * i_size_read because we hold i_mutex.
2689                  */
2690                 if (pos + len > inode->i_size)
2691                         vmtruncate(inode, inode->i_size);
2692         }
2693
2694         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
2695                 goto retry;
2696 out:
2697         return ret;
2698 }
2699
2700 /*
2701  * Check if we should update i_disksize
2702  * when write to the end of file but not require block allocation
2703  */
2704 static int ext4_da_should_update_i_disksize(struct page *page,
2705                                          unsigned long offset)
2706 {
2707         struct buffer_head *bh;
2708         struct inode *inode = page->mapping->host;
2709         unsigned int idx;
2710         int i;
2711
2712         bh = page_buffers(page);
2713         idx = offset >> inode->i_blkbits;
2714
2715         for (i = 0; i < idx; i++)
2716                 bh = bh->b_this_page;
2717
2718         if (!buffer_mapped(bh) || (buffer_delay(bh)))
2719                 return 0;
2720         return 1;
2721 }
2722
2723 static int ext4_da_write_end(struct file *file,
2724                                 struct address_space *mapping,
2725                                 loff_t pos, unsigned len, unsigned copied,
2726                                 struct page *page, void *fsdata)
2727 {
2728         struct inode *inode = mapping->host;
2729         int ret = 0, ret2;
2730         handle_t *handle = ext4_journal_current_handle();
2731         loff_t new_i_size;
2732         unsigned long start, end;
2733         int write_mode = (int)(unsigned long)fsdata;
2734
2735         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
2736                 if (ext4_should_order_data(inode)) {
2737                         return ext4_ordered_write_end(file, mapping, pos,
2738                                         len, copied, page, fsdata);
2739                 } else if (ext4_should_writeback_data(inode)) {
2740                         return ext4_writeback_write_end(file, mapping, pos,
2741                                         len, copied, page, fsdata);
2742                 } else {
2743                         BUG();
2744                 }
2745         }
2746
2747         trace_mark(ext4_da_write_end,
2748                    "dev %s ino %lu pos %llu len %u copied %u",
2749                    inode->i_sb->s_id, inode->i_ino,
2750                    (unsigned long long) pos, len, copied);
2751         start = pos & (PAGE_CACHE_SIZE - 1);
2752         end = start + copied - 1;
2753
2754         /*
2755          * generic_write_end() will run mark_inode_dirty() if i_size
2756          * changes.  So let's piggyback the i_disksize mark_inode_dirty
2757          * into that.
2758          */
2759
2760         new_i_size = pos + copied;
2761         if (new_i_size > EXT4_I(inode)->i_disksize) {
2762                 if (ext4_da_should_update_i_disksize(page, end)) {
2763                         down_write(&EXT4_I(inode)->i_data_sem);
2764                         if (new_i_size > EXT4_I(inode)->i_disksize) {
2765                                 /*
2766                                  * Updating i_disksize when extending file
2767                                  * without needing block allocation
2768                                  */
2769                                 if (ext4_should_order_data(inode))
2770                                         ret = ext4_jbd2_file_inode(handle,
2771                                                                    inode);
2772
2773                                 EXT4_I(inode)->i_disksize = new_i_size;
2774                         }
2775                         up_write(&EXT4_I(inode)->i_data_sem);
2776                         /* We need to mark inode dirty even if
2777                          * new_i_size is less that inode->i_size
2778                          * bu greater than i_disksize.(hint delalloc)
2779                          */
2780                         ext4_mark_inode_dirty(handle, inode);
2781                 }
2782         }
2783         ret2 = generic_write_end(file, mapping, pos, len, copied,
2784                                                         page, fsdata);
2785         copied = ret2;
2786         if (ret2 < 0)
2787                 ret = ret2;
2788         ret2 = ext4_journal_stop(handle);
2789         if (!ret)
2790                 ret = ret2;
2791
2792         return ret ? ret : copied;
2793 }
2794
2795 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
2796 {
2797         /*
2798          * Drop reserved blocks
2799          */
2800         BUG_ON(!PageLocked(page));
2801         if (!page_has_buffers(page))
2802                 goto out;
2803
2804         ext4_da_page_release_reservation(page, offset);
2805
2806 out:
2807         ext4_invalidatepage(page, offset);
2808
2809         return;
2810 }
2811
2812
2813 /*
2814  * bmap() is special.  It gets used by applications such as lilo and by
2815  * the swapper to find the on-disk block of a specific piece of data.
2816  *
2817  * Naturally, this is dangerous if the block concerned is still in the
2818  * journal.  If somebody makes a swapfile on an ext4 data-journaling
2819  * filesystem and enables swap, then they may get a nasty shock when the
2820  * data getting swapped to that swapfile suddenly gets overwritten by
2821  * the original zero's written out previously to the journal and
2822  * awaiting writeback in the kernel's buffer cache.
2823  *
2824  * So, if we see any bmap calls here on a modified, data-journaled file,
2825  * take extra steps to flush any blocks which might be in the cache.
2826  */
2827 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
2828 {
2829         struct inode *inode = mapping->host;
2830         journal_t *journal;
2831         int err;
2832
2833         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
2834                         test_opt(inode->i_sb, DELALLOC)) {
2835                 /*
2836                  * With delalloc we want to sync the file
2837                  * so that we can make sure we allocate
2838                  * blocks for file
2839                  */
2840                 filemap_write_and_wait(mapping);
2841         }
2842
2843         if (EXT4_JOURNAL(inode) && EXT4_I(inode)->i_state & EXT4_STATE_JDATA) {
2844                 /*
2845                  * This is a REALLY heavyweight approach, but the use of
2846                  * bmap on dirty files is expected to be extremely rare:
2847                  * only if we run lilo or swapon on a freshly made file
2848                  * do we expect this to happen.
2849                  *
2850                  * (bmap requires CAP_SYS_RAWIO so this does not
2851                  * represent an unprivileged user DOS attack --- we'd be
2852                  * in trouble if mortal users could trigger this path at
2853                  * will.)
2854                  *
2855                  * NB. EXT4_STATE_JDATA is not set on files other than
2856                  * regular files.  If somebody wants to bmap a directory
2857                  * or symlink and gets confused because the buffer
2858                  * hasn't yet been flushed to disk, they deserve
2859                  * everything they get.
2860                  */
2861
2862                 EXT4_I(inode)->i_state &= ~EXT4_STATE_JDATA;
2863                 journal = EXT4_JOURNAL(inode);
2864                 jbd2_journal_lock_updates(journal);
2865                 err = jbd2_journal_flush(journal);
2866                 jbd2_journal_unlock_updates(journal);
2867
2868                 if (err)
2869                         return 0;
2870         }
2871
2872         return generic_block_bmap(mapping, block, ext4_get_block);
2873 }
2874
2875 static int bget_one(handle_t *handle, struct buffer_head *bh)
2876 {
2877         get_bh(bh);
2878         return 0;
2879 }
2880
2881 static int bput_one(handle_t *handle, struct buffer_head *bh)
2882 {
2883         put_bh(bh);
2884         return 0;
2885 }
2886
2887 /*
2888  * Note that we don't need to start a transaction unless we're journaling data
2889  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2890  * need to file the inode to the transaction's list in ordered mode because if
2891  * we are writing back data added by write(), the inode is already there and if
2892  * we are writing back data modified via mmap(), noone guarantees in which
2893  * transaction the data will hit the disk. In case we are journaling data, we
2894  * cannot start transaction directly because transaction start ranks above page
2895  * lock so we have to do some magic.
2896  *
2897  * In all journaling modes block_write_full_page() will start the I/O.
2898  *
2899  * Problem:
2900  *
2901  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2902  *              ext4_writepage()
2903  *
2904  * Similar for:
2905  *
2906  *      ext4_file_write() -> generic_file_write() -> __alloc_pages() -> ...
2907  *
2908  * Same applies to ext4_get_block().  We will deadlock on various things like
2909  * lock_journal and i_data_sem
2910  *
2911  * Setting PF_MEMALLOC here doesn't work - too many internal memory
2912  * allocations fail.
2913  *
2914  * 16May01: If we're reentered then journal_current_handle() will be
2915  *          non-zero. We simply *return*.
2916  *
2917  * 1 July 2001: @@@ FIXME:
2918  *   In journalled data mode, a data buffer may be metadata against the
2919  *   current transaction.  But the same file is part of a shared mapping
2920  *   and someone does a writepage() on it.
2921  *
2922  *   We will move the buffer onto the async_data list, but *after* it has
2923  *   been dirtied. So there's a small window where we have dirty data on
2924  *   BJ_Metadata.
2925  *
2926  *   Note that this only applies to the last partial page in the file.  The
2927  *   bit which block_write_full_page() uses prepare/commit for.  (That's
2928  *   broken code anyway: it's wrong for msync()).
2929  *
2930  *   It's a rare case: affects the final partial page, for journalled data
2931  *   where the file is subject to bith write() and writepage() in the same
2932  *   transction.  To fix it we'll need a custom block_write_full_page().
2933  *   We'll probably need that anyway for journalling writepage() output.
2934  *
2935  * We don't honour synchronous mounts for writepage().  That would be
2936  * disastrous.  Any write() or metadata operation will sync the fs for
2937  * us.
2938  *
2939  */
2940 static int __ext4_normal_writepage(struct page *page,
2941                                 struct writeback_control *wbc)
2942 {
2943         struct inode *inode = page->mapping->host;
2944
2945         if (test_opt(inode->i_sb, NOBH))
2946                 return nobh_writepage(page,
2947                                         ext4_normal_get_block_write, wbc);
2948         else
2949                 return block_write_full_page(page,
2950                                                 ext4_normal_get_block_write,
2951                                                 wbc);
2952 }
2953
2954 static int ext4_normal_writepage(struct page *page,
2955                                 struct writeback_control *wbc)
2956 {
2957         struct inode *inode = page->mapping->host;
2958         loff_t size = i_size_read(inode);
2959         loff_t len;
2960
2961         trace_mark(ext4_normal_writepage,
2962                    "dev %s ino %lu page_index %lu",
2963                    inode->i_sb->s_id, inode->i_ino, page->index);
2964         J_ASSERT(PageLocked(page));
2965         if (page->index == size >> PAGE_CACHE_SHIFT)
2966                 len = size & ~PAGE_CACHE_MASK;
2967         else
2968                 len = PAGE_CACHE_SIZE;
2969
2970         if (page_has_buffers(page)) {
2971                 /* if page has buffers it should all be mapped
2972                  * and allocated. If there are not buffers attached
2973                  * to the page we know the page is dirty but it lost
2974                  * buffers. That means that at some moment in time
2975                  * after write_begin() / write_end() has been called
2976                  * all buffers have been clean and thus they must have been
2977                  * written at least once. So they are all mapped and we can
2978                  * happily proceed with mapping them and writing the page.
2979                  */
2980                 BUG_ON(walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
2981                                         ext4_bh_unmapped_or_delay));
2982         }
2983
2984         if (!ext4_journal_current_handle())
2985                 return __ext4_normal_writepage(page, wbc);
2986
2987         redirty_page_for_writepage(wbc, page);
2988         unlock_page(page);
2989         return 0;
2990 }
2991
2992 static int __ext4_journalled_writepage(struct page *page,
2993                                 struct writeback_control *wbc)
2994 {
2995         struct address_space *mapping = page->mapping;
2996         struct inode *inode = mapping->host;
2997         struct buffer_head *page_bufs;
2998         handle_t *handle = NULL;
2999         int ret = 0;
3000         int err;
3001
3002         ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
3003                                         ext4_normal_get_block_write);
3004         if (ret != 0)
3005                 goto out_unlock;
3006
3007         page_bufs = page_buffers(page);
3008         walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE, NULL,
3009                                                                 bget_one);
3010         /* As soon as we unlock the page, it can go away, but we have
3011          * references to buffers so we are safe */
3012         unlock_page(page);
3013
3014         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
3015         if (IS_ERR(handle)) {
3016                 ret = PTR_ERR(handle);
3017                 goto out;
3018         }
3019
3020         ret = walk_page_buffers(handle, page_bufs, 0,
3021                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
3022
3023         err = walk_page_buffers(handle, page_bufs, 0,
3024                                 PAGE_CACHE_SIZE, NULL, write_end_fn);
3025         if (ret == 0)
3026                 ret = err;
3027         err = ext4_journal_stop(handle);
3028         if (!ret)
3029                 ret = err;
3030
3031         walk_page_buffers(handle, page_bufs, 0,
3032                                 PAGE_CACHE_SIZE, NULL, bput_one);
3033         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
3034         goto out;
3035
3036 out_unlock:
3037         unlock_page(page);
3038 out:
3039         return ret;
3040 }
3041
3042 static int ext4_journalled_writepage(struct page *page,
3043                                 struct writeback_control *wbc)
3044 {
3045         struct inode *inode = page->mapping->host;
3046         loff_t size = i_size_read(inode);
3047         loff_t len;
3048
3049         trace_mark(ext4_journalled_writepage,
3050                    "dev %s ino %lu page_index %lu",
3051                    inode->i_sb->s_id, inode->i_ino, page->index);
3052         J_ASSERT(PageLocked(page));
3053         if (page->index == size >> PAGE_CACHE_SHIFT)
3054                 len = size & ~PAGE_CACHE_MASK;
3055         else
3056                 len = PAGE_CACHE_SIZE;
3057
3058         if (page_has_buffers(page)) {
3059                 /* if page has buffers it should all be mapped
3060                  * and allocated. If there are not buffers attached
3061                  * to the page we know the page is dirty but it lost
3062                  * buffers. That means that at some moment in time
3063                  * after write_begin() / write_end() has been called
3064                  * all buffers have been clean and thus they must have been
3065                  * written at least once. So they are all mapped and we can
3066                  * happily proceed with mapping them and writing the page.
3067                  */
3068                 BUG_ON(walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
3069                                         ext4_bh_unmapped_or_delay));
3070         }
3071
3072         if (ext4_journal_current_handle())
3073                 goto no_write;
3074
3075         if (PageChecked(page)) {
3076                 /*
3077                  * It's mmapped pagecache.  Add buffers and journal it.  There
3078                  * doesn't seem much point in redirtying the page here.
3079                  */
3080                 ClearPageChecked(page);
3081                 return __ext4_journalled_writepage(page, wbc);
3082         } else {
3083                 /*
3084                  * It may be a page full of checkpoint-mode buffers.  We don't
3085                  * really know unless we go poke around in the buffer_heads.
3086                  * But block_write_full_page will do the right thing.
3087                  */
3088                 return block_write_full_page(page,
3089                                                 ext4_normal_get_block_write,
3090                                                 wbc);
3091         }
3092 no_write:
3093         redirty_page_for_writepage(wbc, page);
3094         unlock_page(page);
3095         return 0;
3096 }
3097
3098 static int ext4_readpage(struct file *file, struct page *page)
3099 {
3100         return mpage_readpage(page, ext4_get_block);
3101 }
3102
3103 static int
3104 ext4_readpages(struct file *file, struct address_space *mapping,
3105                 struct list_head *pages, unsigned nr_pages)
3106 {
3107         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3108 }
3109
3110 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3111 {
3112         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3113
3114         /*
3115          * If it's a full truncate we just forget about the pending dirtying
3116          */
3117         if (offset == 0)
3118                 ClearPageChecked(page);
3119
3120         if (journal)
3121                 jbd2_journal_invalidatepage(journal, page, offset);
3122         else
3123                 block_invalidatepage(page, offset);
3124 }
3125
3126 static int ext4_releasepage(struct page *page, gfp_t wait)
3127 {
3128         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3129
3130         WARN_ON(PageChecked(page));
3131         if (!page_has_buffers(page))
3132                 return 0;
3133         if (journal)
3134                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3135         else
3136                 return try_to_free_buffers(page);
3137 }
3138
3139 /*
3140  * If the O_DIRECT write will extend the file then add this inode to the
3141  * orphan list.  So recovery will truncate it back to the original size
3142  * if the machine crashes during the write.
3143  *
3144  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3145  * crashes then stale disk data _may_ be exposed inside the file. But current
3146  * VFS code falls back into buffered path in that case so we are safe.
3147  */
3148 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3149                         const struct iovec *iov, loff_t offset,
3150                         unsigned long nr_segs)
3151 {
3152         struct file *file = iocb->ki_filp;
3153         struct inode *inode = file->f_mapping->host;
3154         struct ext4_inode_info *ei = EXT4_I(inode);
3155         handle_t *handle;
3156         ssize_t ret;
3157         int orphan = 0;
3158         size_t count = iov_length(iov, nr_segs);
3159
3160         if (rw == WRITE) {
3161                 loff_t final_size = offset + count;
3162
3163                 if (final_size > inode->i_size) {
3164                         /* Credits for sb + inode write */
3165                         handle = ext4_journal_start(inode, 2);
3166                         if (IS_ERR(handle)) {
3167                                 ret = PTR_ERR(handle);
3168                                 goto out;
3169                         }
3170                         ret = ext4_orphan_add(handle, inode);
3171                         if (ret) {
3172                                 ext4_journal_stop(handle);
3173                                 goto out;
3174                         }
3175                         orphan = 1;
3176                         ei->i_disksize = inode->i_size;
3177                         ext4_journal_stop(handle);
3178                 }
3179         }
3180
3181         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
3182                                  offset, nr_segs,
3183                                  ext4_get_block, NULL);
3184
3185         if (orphan) {
3186                 int err;
3187