]> git.karo-electronics.de Git - karo-tx-linux.git/blob - fs/btrfs/ordered-data.c
Btrfs: don't mix the ordered extents of all files together during logging the inodes
[karo-tx-linux.git] / fs / btrfs / ordered-data.c
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18
19 #include <linux/slab.h>
20 #include <linux/blkdev.h>
21 #include <linux/writeback.h>
22 #include <linux/pagevec.h>
23 #include "ctree.h"
24 #include "transaction.h"
25 #include "btrfs_inode.h"
26 #include "extent_io.h"
27 #include "disk-io.h"
28
29 static struct kmem_cache *btrfs_ordered_extent_cache;
30
31 static u64 entry_end(struct btrfs_ordered_extent *entry)
32 {
33         if (entry->file_offset + entry->len < entry->file_offset)
34                 return (u64)-1;
35         return entry->file_offset + entry->len;
36 }
37
38 /* returns NULL if the insertion worked, or it returns the node it did find
39  * in the tree
40  */
41 static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
42                                    struct rb_node *node)
43 {
44         struct rb_node **p = &root->rb_node;
45         struct rb_node *parent = NULL;
46         struct btrfs_ordered_extent *entry;
47
48         while (*p) {
49                 parent = *p;
50                 entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
51
52                 if (file_offset < entry->file_offset)
53                         p = &(*p)->rb_left;
54                 else if (file_offset >= entry_end(entry))
55                         p = &(*p)->rb_right;
56                 else
57                         return parent;
58         }
59
60         rb_link_node(node, parent, p);
61         rb_insert_color(node, root);
62         return NULL;
63 }
64
65 static void ordered_data_tree_panic(struct inode *inode, int errno,
66                                                u64 offset)
67 {
68         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
69         btrfs_panic(fs_info, errno, "Inconsistency in ordered tree at offset "
70                     "%llu\n", offset);
71 }
72
73 /*
74  * look for a given offset in the tree, and if it can't be found return the
75  * first lesser offset
76  */
77 static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
78                                      struct rb_node **prev_ret)
79 {
80         struct rb_node *n = root->rb_node;
81         struct rb_node *prev = NULL;
82         struct rb_node *test;
83         struct btrfs_ordered_extent *entry;
84         struct btrfs_ordered_extent *prev_entry = NULL;
85
86         while (n) {
87                 entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
88                 prev = n;
89                 prev_entry = entry;
90
91                 if (file_offset < entry->file_offset)
92                         n = n->rb_left;
93                 else if (file_offset >= entry_end(entry))
94                         n = n->rb_right;
95                 else
96                         return n;
97         }
98         if (!prev_ret)
99                 return NULL;
100
101         while (prev && file_offset >= entry_end(prev_entry)) {
102                 test = rb_next(prev);
103                 if (!test)
104                         break;
105                 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
106                                       rb_node);
107                 if (file_offset < entry_end(prev_entry))
108                         break;
109
110                 prev = test;
111         }
112         if (prev)
113                 prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
114                                       rb_node);
115         while (prev && file_offset < entry_end(prev_entry)) {
116                 test = rb_prev(prev);
117                 if (!test)
118                         break;
119                 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
120                                       rb_node);
121                 prev = test;
122         }
123         *prev_ret = prev;
124         return NULL;
125 }
126
127 /*
128  * helper to check if a given offset is inside a given entry
129  */
130 static int offset_in_entry(struct btrfs_ordered_extent *entry, u64 file_offset)
131 {
132         if (file_offset < entry->file_offset ||
133             entry->file_offset + entry->len <= file_offset)
134                 return 0;
135         return 1;
136 }
137
138 static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
139                           u64 len)
140 {
141         if (file_offset + len <= entry->file_offset ||
142             entry->file_offset + entry->len <= file_offset)
143                 return 0;
144         return 1;
145 }
146
147 /*
148  * look find the first ordered struct that has this offset, otherwise
149  * the first one less than this offset
150  */
151 static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
152                                           u64 file_offset)
153 {
154         struct rb_root *root = &tree->tree;
155         struct rb_node *prev = NULL;
156         struct rb_node *ret;
157         struct btrfs_ordered_extent *entry;
158
159         if (tree->last) {
160                 entry = rb_entry(tree->last, struct btrfs_ordered_extent,
161                                  rb_node);
162                 if (offset_in_entry(entry, file_offset))
163                         return tree->last;
164         }
165         ret = __tree_search(root, file_offset, &prev);
166         if (!ret)
167                 ret = prev;
168         if (ret)
169                 tree->last = ret;
170         return ret;
171 }
172
173 /* allocate and add a new ordered_extent into the per-inode tree.
174  * file_offset is the logical offset in the file
175  *
176  * start is the disk block number of an extent already reserved in the
177  * extent allocation tree
178  *
179  * len is the length of the extent
180  *
181  * The tree is given a single reference on the ordered extent that was
182  * inserted.
183  */
184 static int __btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
185                                       u64 start, u64 len, u64 disk_len,
186                                       int type, int dio, int compress_type)
187 {
188         struct btrfs_root *root = BTRFS_I(inode)->root;
189         struct btrfs_ordered_inode_tree *tree;
190         struct rb_node *node;
191         struct btrfs_ordered_extent *entry;
192
193         tree = &BTRFS_I(inode)->ordered_tree;
194         entry = kmem_cache_zalloc(btrfs_ordered_extent_cache, GFP_NOFS);
195         if (!entry)
196                 return -ENOMEM;
197
198         entry->file_offset = file_offset;
199         entry->start = start;
200         entry->len = len;
201         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM) &&
202             !(type == BTRFS_ORDERED_NOCOW))
203                 entry->csum_bytes_left = disk_len;
204         entry->disk_len = disk_len;
205         entry->bytes_left = len;
206         entry->inode = igrab(inode);
207         entry->compress_type = compress_type;
208         entry->truncated_len = (u64)-1;
209         if (type != BTRFS_ORDERED_IO_DONE && type != BTRFS_ORDERED_COMPLETE)
210                 set_bit(type, &entry->flags);
211
212         if (dio)
213                 set_bit(BTRFS_ORDERED_DIRECT, &entry->flags);
214
215         /* one ref for the tree */
216         atomic_set(&entry->refs, 1);
217         init_waitqueue_head(&entry->wait);
218         INIT_LIST_HEAD(&entry->list);
219         INIT_LIST_HEAD(&entry->root_extent_list);
220         INIT_LIST_HEAD(&entry->work_list);
221         init_completion(&entry->completion);
222         INIT_LIST_HEAD(&entry->log_list);
223
224         trace_btrfs_ordered_extent_add(inode, entry);
225
226         spin_lock_irq(&tree->lock);
227         node = tree_insert(&tree->tree, file_offset,
228                            &entry->rb_node);
229         if (node)
230                 ordered_data_tree_panic(inode, -EEXIST, file_offset);
231         spin_unlock_irq(&tree->lock);
232
233         spin_lock(&root->ordered_extent_lock);
234         list_add_tail(&entry->root_extent_list,
235                       &root->ordered_extents);
236         root->nr_ordered_extents++;
237         if (root->nr_ordered_extents == 1) {
238                 spin_lock(&root->fs_info->ordered_root_lock);
239                 BUG_ON(!list_empty(&root->ordered_root));
240                 list_add_tail(&root->ordered_root,
241                               &root->fs_info->ordered_roots);
242                 spin_unlock(&root->fs_info->ordered_root_lock);
243         }
244         spin_unlock(&root->ordered_extent_lock);
245
246         return 0;
247 }
248
249 int btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
250                              u64 start, u64 len, u64 disk_len, int type)
251 {
252         return __btrfs_add_ordered_extent(inode, file_offset, start, len,
253                                           disk_len, type, 0,
254                                           BTRFS_COMPRESS_NONE);
255 }
256
257 int btrfs_add_ordered_extent_dio(struct inode *inode, u64 file_offset,
258                                  u64 start, u64 len, u64 disk_len, int type)
259 {
260         return __btrfs_add_ordered_extent(inode, file_offset, start, len,
261                                           disk_len, type, 1,
262                                           BTRFS_COMPRESS_NONE);
263 }
264
265 int btrfs_add_ordered_extent_compress(struct inode *inode, u64 file_offset,
266                                       u64 start, u64 len, u64 disk_len,
267                                       int type, int compress_type)
268 {
269         return __btrfs_add_ordered_extent(inode, file_offset, start, len,
270                                           disk_len, type, 0,
271                                           compress_type);
272 }
273
274 /*
275  * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
276  * when an ordered extent is finished.  If the list covers more than one
277  * ordered extent, it is split across multiples.
278  */
279 void btrfs_add_ordered_sum(struct inode *inode,
280                            struct btrfs_ordered_extent *entry,
281                            struct btrfs_ordered_sum *sum)
282 {
283         struct btrfs_ordered_inode_tree *tree;
284
285         tree = &BTRFS_I(inode)->ordered_tree;
286         spin_lock_irq(&tree->lock);
287         list_add_tail(&sum->list, &entry->list);
288         WARN_ON(entry->csum_bytes_left < sum->len);
289         entry->csum_bytes_left -= sum->len;
290         if (entry->csum_bytes_left == 0)
291                 wake_up(&entry->wait);
292         spin_unlock_irq(&tree->lock);
293 }
294
295 /*
296  * this is used to account for finished IO across a given range
297  * of the file.  The IO may span ordered extents.  If
298  * a given ordered_extent is completely done, 1 is returned, otherwise
299  * 0.
300  *
301  * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
302  * to make sure this function only returns 1 once for a given ordered extent.
303  *
304  * file_offset is updated to one byte past the range that is recorded as
305  * complete.  This allows you to walk forward in the file.
306  */
307 int btrfs_dec_test_first_ordered_pending(struct inode *inode,
308                                    struct btrfs_ordered_extent **cached,
309                                    u64 *file_offset, u64 io_size, int uptodate)
310 {
311         struct btrfs_ordered_inode_tree *tree;
312         struct rb_node *node;
313         struct btrfs_ordered_extent *entry = NULL;
314         int ret;
315         unsigned long flags;
316         u64 dec_end;
317         u64 dec_start;
318         u64 to_dec;
319
320         tree = &BTRFS_I(inode)->ordered_tree;
321         spin_lock_irqsave(&tree->lock, flags);
322         node = tree_search(tree, *file_offset);
323         if (!node) {
324                 ret = 1;
325                 goto out;
326         }
327
328         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
329         if (!offset_in_entry(entry, *file_offset)) {
330                 ret = 1;
331                 goto out;
332         }
333
334         dec_start = max(*file_offset, entry->file_offset);
335         dec_end = min(*file_offset + io_size, entry->file_offset +
336                       entry->len);
337         *file_offset = dec_end;
338         if (dec_start > dec_end) {
339                 btrfs_crit(BTRFS_I(inode)->root->fs_info,
340                         "bad ordering dec_start %llu end %llu", dec_start, dec_end);
341         }
342         to_dec = dec_end - dec_start;
343         if (to_dec > entry->bytes_left) {
344                 btrfs_crit(BTRFS_I(inode)->root->fs_info,
345                         "bad ordered accounting left %llu size %llu",
346                         entry->bytes_left, to_dec);
347         }
348         entry->bytes_left -= to_dec;
349         if (!uptodate)
350                 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
351
352         if (entry->bytes_left == 0)
353                 ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
354         else
355                 ret = 1;
356 out:
357         if (!ret && cached && entry) {
358                 *cached = entry;
359                 atomic_inc(&entry->refs);
360         }
361         spin_unlock_irqrestore(&tree->lock, flags);
362         return ret == 0;
363 }
364
365 /*
366  * this is used to account for finished IO across a given range
367  * of the file.  The IO should not span ordered extents.  If
368  * a given ordered_extent is completely done, 1 is returned, otherwise
369  * 0.
370  *
371  * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
372  * to make sure this function only returns 1 once for a given ordered extent.
373  */
374 int btrfs_dec_test_ordered_pending(struct inode *inode,
375                                    struct btrfs_ordered_extent **cached,
376                                    u64 file_offset, u64 io_size, int uptodate)
377 {
378         struct btrfs_ordered_inode_tree *tree;
379         struct rb_node *node;
380         struct btrfs_ordered_extent *entry = NULL;
381         unsigned long flags;
382         int ret;
383
384         tree = &BTRFS_I(inode)->ordered_tree;
385         spin_lock_irqsave(&tree->lock, flags);
386         if (cached && *cached) {
387                 entry = *cached;
388                 goto have_entry;
389         }
390
391         node = tree_search(tree, file_offset);
392         if (!node) {
393                 ret = 1;
394                 goto out;
395         }
396
397         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
398 have_entry:
399         if (!offset_in_entry(entry, file_offset)) {
400                 ret = 1;
401                 goto out;
402         }
403
404         if (io_size > entry->bytes_left) {
405                 btrfs_crit(BTRFS_I(inode)->root->fs_info,
406                            "bad ordered accounting left %llu size %llu",
407                        entry->bytes_left, io_size);
408         }
409         entry->bytes_left -= io_size;
410         if (!uptodate)
411                 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
412
413         if (entry->bytes_left == 0)
414                 ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
415         else
416                 ret = 1;
417 out:
418         if (!ret && cached && entry) {
419                 *cached = entry;
420                 atomic_inc(&entry->refs);
421         }
422         spin_unlock_irqrestore(&tree->lock, flags);
423         return ret == 0;
424 }
425
426 /* Needs to either be called under a log transaction or the log_mutex */
427 void btrfs_get_logged_extents(struct inode *inode,
428                               struct list_head *logged_list)
429 {
430         struct btrfs_ordered_inode_tree *tree;
431         struct btrfs_ordered_extent *ordered;
432         struct rb_node *n;
433
434         tree = &BTRFS_I(inode)->ordered_tree;
435         spin_lock_irq(&tree->lock);
436         for (n = rb_first(&tree->tree); n; n = rb_next(n)) {
437                 ordered = rb_entry(n, struct btrfs_ordered_extent, rb_node);
438                 if (!list_empty(&ordered->log_list))
439                         continue;
440                 list_add_tail(&ordered->log_list, logged_list);
441                 atomic_inc(&ordered->refs);
442         }
443         spin_unlock_irq(&tree->lock);
444 }
445
446 void btrfs_put_logged_extents(struct list_head *logged_list)
447 {
448         struct btrfs_ordered_extent *ordered;
449
450         while (!list_empty(logged_list)) {
451                 ordered = list_first_entry(logged_list,
452                                            struct btrfs_ordered_extent,
453                                            log_list);
454                 list_del_init(&ordered->log_list);
455                 btrfs_put_ordered_extent(ordered);
456         }
457 }
458
459 void btrfs_submit_logged_extents(struct list_head *logged_list,
460                                  struct btrfs_root *log)
461 {
462         int index = log->log_transid % 2;
463
464         spin_lock_irq(&log->log_extents_lock[index]);
465         list_splice_tail(logged_list, &log->logged_list[index]);
466         spin_unlock_irq(&log->log_extents_lock[index]);
467 }
468
469 void btrfs_wait_logged_extents(struct btrfs_root *log, u64 transid)
470 {
471         struct btrfs_ordered_extent *ordered;
472         int index = transid % 2;
473
474         spin_lock_irq(&log->log_extents_lock[index]);
475         while (!list_empty(&log->logged_list[index])) {
476                 ordered = list_first_entry(&log->logged_list[index],
477                                            struct btrfs_ordered_extent,
478                                            log_list);
479                 list_del_init(&ordered->log_list);
480                 spin_unlock_irq(&log->log_extents_lock[index]);
481                 wait_event(ordered->wait, test_bit(BTRFS_ORDERED_IO_DONE,
482                                                    &ordered->flags));
483                 btrfs_put_ordered_extent(ordered);
484                 spin_lock_irq(&log->log_extents_lock[index]);
485         }
486         spin_unlock_irq(&log->log_extents_lock[index]);
487 }
488
489 void btrfs_free_logged_extents(struct btrfs_root *log, u64 transid)
490 {
491         struct btrfs_ordered_extent *ordered;
492         int index = transid % 2;
493
494         spin_lock_irq(&log->log_extents_lock[index]);
495         while (!list_empty(&log->logged_list[index])) {
496                 ordered = list_first_entry(&log->logged_list[index],
497                                            struct btrfs_ordered_extent,
498                                            log_list);
499                 list_del_init(&ordered->log_list);
500                 spin_unlock_irq(&log->log_extents_lock[index]);
501                 btrfs_put_ordered_extent(ordered);
502                 spin_lock_irq(&log->log_extents_lock[index]);
503         }
504         spin_unlock_irq(&log->log_extents_lock[index]);
505 }
506
507 /*
508  * used to drop a reference on an ordered extent.  This will free
509  * the extent if the last reference is dropped
510  */
511 void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
512 {
513         struct list_head *cur;
514         struct btrfs_ordered_sum *sum;
515
516         trace_btrfs_ordered_extent_put(entry->inode, entry);
517
518         if (atomic_dec_and_test(&entry->refs)) {
519                 if (entry->inode)
520                         btrfs_add_delayed_iput(entry->inode);
521                 while (!list_empty(&entry->list)) {
522                         cur = entry->list.next;
523                         sum = list_entry(cur, struct btrfs_ordered_sum, list);
524                         list_del(&sum->list);
525                         kfree(sum);
526                 }
527                 kmem_cache_free(btrfs_ordered_extent_cache, entry);
528         }
529 }
530
531 /*
532  * remove an ordered extent from the tree.  No references are dropped
533  * and waiters are woken up.
534  */
535 void btrfs_remove_ordered_extent(struct inode *inode,
536                                  struct btrfs_ordered_extent *entry)
537 {
538         struct btrfs_ordered_inode_tree *tree;
539         struct btrfs_root *root = BTRFS_I(inode)->root;
540         struct rb_node *node;
541
542         tree = &BTRFS_I(inode)->ordered_tree;
543         spin_lock_irq(&tree->lock);
544         node = &entry->rb_node;
545         rb_erase(node, &tree->tree);
546         if (tree->last == node)
547                 tree->last = NULL;
548         set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
549         spin_unlock_irq(&tree->lock);
550
551         spin_lock(&root->ordered_extent_lock);
552         list_del_init(&entry->root_extent_list);
553         root->nr_ordered_extents--;
554
555         trace_btrfs_ordered_extent_remove(inode, entry);
556
557         /*
558          * we have no more ordered extents for this inode and
559          * no dirty pages.  We can safely remove it from the
560          * list of ordered extents
561          */
562         if (RB_EMPTY_ROOT(&tree->tree) &&
563             !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
564                 spin_lock(&root->fs_info->ordered_root_lock);
565                 list_del_init(&BTRFS_I(inode)->ordered_operations);
566                 spin_unlock(&root->fs_info->ordered_root_lock);
567         }
568
569         if (!root->nr_ordered_extents) {
570                 spin_lock(&root->fs_info->ordered_root_lock);
571                 BUG_ON(list_empty(&root->ordered_root));
572                 list_del_init(&root->ordered_root);
573                 spin_unlock(&root->fs_info->ordered_root_lock);
574         }
575         spin_unlock(&root->ordered_extent_lock);
576         wake_up(&entry->wait);
577 }
578
579 static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
580 {
581         struct btrfs_ordered_extent *ordered;
582
583         ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
584         btrfs_start_ordered_extent(ordered->inode, ordered, 1);
585         complete(&ordered->completion);
586 }
587
588 /*
589  * wait for all the ordered extents in a root.  This is done when balancing
590  * space between drives.
591  */
592 int btrfs_wait_ordered_extents(struct btrfs_root *root, int nr)
593 {
594         struct list_head splice, works;
595         struct btrfs_ordered_extent *ordered, *next;
596         int count = 0;
597
598         INIT_LIST_HEAD(&splice);
599         INIT_LIST_HEAD(&works);
600
601         mutex_lock(&root->fs_info->ordered_operations_mutex);
602         spin_lock(&root->ordered_extent_lock);
603         list_splice_init(&root->ordered_extents, &splice);
604         while (!list_empty(&splice) && nr) {
605                 ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
606                                            root_extent_list);
607                 list_move_tail(&ordered->root_extent_list,
608                                &root->ordered_extents);
609                 atomic_inc(&ordered->refs);
610                 spin_unlock(&root->ordered_extent_lock);
611
612                 ordered->flush_work.func = btrfs_run_ordered_extent_work;
613                 list_add_tail(&ordered->work_list, &works);
614                 btrfs_queue_worker(&root->fs_info->flush_workers,
615                                    &ordered->flush_work);
616
617                 cond_resched();
618                 spin_lock(&root->ordered_extent_lock);
619                 if (nr != -1)
620                         nr--;
621                 count++;
622         }
623         list_splice_tail(&splice, &root->ordered_extents);
624         spin_unlock(&root->ordered_extent_lock);
625
626         list_for_each_entry_safe(ordered, next, &works, work_list) {
627                 list_del_init(&ordered->work_list);
628                 wait_for_completion(&ordered->completion);
629                 btrfs_put_ordered_extent(ordered);
630                 cond_resched();
631         }
632         mutex_unlock(&root->fs_info->ordered_operations_mutex);
633
634         return count;
635 }
636
637 void btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, int nr)
638 {
639         struct btrfs_root *root;
640         struct list_head splice;
641         int done;
642
643         INIT_LIST_HEAD(&splice);
644
645         spin_lock(&fs_info->ordered_root_lock);
646         list_splice_init(&fs_info->ordered_roots, &splice);
647         while (!list_empty(&splice) && nr) {
648                 root = list_first_entry(&splice, struct btrfs_root,
649                                         ordered_root);
650                 root = btrfs_grab_fs_root(root);
651                 BUG_ON(!root);
652                 list_move_tail(&root->ordered_root,
653                                &fs_info->ordered_roots);
654                 spin_unlock(&fs_info->ordered_root_lock);
655
656                 done = btrfs_wait_ordered_extents(root, nr);
657                 btrfs_put_fs_root(root);
658
659                 spin_lock(&fs_info->ordered_root_lock);
660                 if (nr != -1) {
661                         nr -= done;
662                         WARN_ON(nr < 0);
663                 }
664         }
665         list_splice_tail(&splice, &fs_info->ordered_roots);
666         spin_unlock(&fs_info->ordered_root_lock);
667 }
668
669 /*
670  * this is used during transaction commit to write all the inodes
671  * added to the ordered operation list.  These files must be fully on
672  * disk before the transaction commits.
673  *
674  * we have two modes here, one is to just start the IO via filemap_flush
675  * and the other is to wait for all the io.  When we wait, we have an
676  * extra check to make sure the ordered operation list really is empty
677  * before we return
678  */
679 int btrfs_run_ordered_operations(struct btrfs_trans_handle *trans,
680                                  struct btrfs_root *root, int wait)
681 {
682         struct btrfs_inode *btrfs_inode;
683         struct inode *inode;
684         struct btrfs_transaction *cur_trans = trans->transaction;
685         struct list_head splice;
686         struct list_head works;
687         struct btrfs_delalloc_work *work, *next;
688         int ret = 0;
689
690         INIT_LIST_HEAD(&splice);
691         INIT_LIST_HEAD(&works);
692
693         mutex_lock(&root->fs_info->ordered_extent_flush_mutex);
694         spin_lock(&root->fs_info->ordered_root_lock);
695         list_splice_init(&cur_trans->ordered_operations, &splice);
696         while (!list_empty(&splice)) {
697                 btrfs_inode = list_entry(splice.next, struct btrfs_inode,
698                                    ordered_operations);
699                 inode = &btrfs_inode->vfs_inode;
700
701                 list_del_init(&btrfs_inode->ordered_operations);
702
703                 /*
704                  * the inode may be getting freed (in sys_unlink path).
705                  */
706                 inode = igrab(inode);
707                 if (!inode)
708                         continue;
709
710                 if (!wait)
711                         list_add_tail(&BTRFS_I(inode)->ordered_operations,
712                                       &cur_trans->ordered_operations);
713                 spin_unlock(&root->fs_info->ordered_root_lock);
714
715                 work = btrfs_alloc_delalloc_work(inode, wait, 1);
716                 if (!work) {
717                         spin_lock(&root->fs_info->ordered_root_lock);
718                         if (list_empty(&BTRFS_I(inode)->ordered_operations))
719                                 list_add_tail(&btrfs_inode->ordered_operations,
720                                               &splice);
721                         list_splice_tail(&splice,
722                                          &cur_trans->ordered_operations);
723                         spin_unlock(&root->fs_info->ordered_root_lock);
724                         ret = -ENOMEM;
725                         goto out;
726                 }
727                 list_add_tail(&work->list, &works);
728                 btrfs_queue_worker(&root->fs_info->flush_workers,
729                                    &work->work);
730
731                 cond_resched();
732                 spin_lock(&root->fs_info->ordered_root_lock);
733         }
734         spin_unlock(&root->fs_info->ordered_root_lock);
735 out:
736         list_for_each_entry_safe(work, next, &works, list) {
737                 list_del_init(&work->list);
738                 btrfs_wait_and_free_delalloc_work(work);
739         }
740         mutex_unlock(&root->fs_info->ordered_extent_flush_mutex);
741         return ret;
742 }
743
744 /*
745  * Used to start IO or wait for a given ordered extent to finish.
746  *
747  * If wait is one, this effectively waits on page writeback for all the pages
748  * in the extent, and it waits on the io completion code to insert
749  * metadata into the btree corresponding to the extent
750  */
751 void btrfs_start_ordered_extent(struct inode *inode,
752                                        struct btrfs_ordered_extent *entry,
753                                        int wait)
754 {
755         u64 start = entry->file_offset;
756         u64 end = start + entry->len - 1;
757
758         trace_btrfs_ordered_extent_start(inode, entry);
759
760         /*
761          * pages in the range can be dirty, clean or writeback.  We
762          * start IO on any dirty ones so the wait doesn't stall waiting
763          * for the flusher thread to find them
764          */
765         if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
766                 filemap_fdatawrite_range(inode->i_mapping, start, end);
767         if (wait) {
768                 wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
769                                                  &entry->flags));
770         }
771 }
772
773 /*
774  * Used to wait on ordered extents across a large range of bytes.
775  */
776 int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
777 {
778         int ret = 0;
779         u64 end;
780         u64 orig_end;
781         struct btrfs_ordered_extent *ordered;
782
783         if (start + len < start) {
784                 orig_end = INT_LIMIT(loff_t);
785         } else {
786                 orig_end = start + len - 1;
787                 if (orig_end > INT_LIMIT(loff_t))
788                         orig_end = INT_LIMIT(loff_t);
789         }
790
791         /* start IO across the range first to instantiate any delalloc
792          * extents
793          */
794         ret = filemap_fdatawrite_range(inode->i_mapping, start, orig_end);
795         if (ret)
796                 return ret;
797         /*
798          * So with compression we will find and lock a dirty page and clear the
799          * first one as dirty, setup an async extent, and immediately return
800          * with the entire range locked but with nobody actually marked with
801          * writeback.  So we can't just filemap_write_and_wait_range() and
802          * expect it to work since it will just kick off a thread to do the
803          * actual work.  So we need to call filemap_fdatawrite_range _again_
804          * since it will wait on the page lock, which won't be unlocked until
805          * after the pages have been marked as writeback and so we're good to go
806          * from there.  We have to do this otherwise we'll miss the ordered
807          * extents and that results in badness.  Please Josef, do not think you
808          * know better and pull this out at some point in the future, it is
809          * right and you are wrong.
810          */
811         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
812                      &BTRFS_I(inode)->runtime_flags)) {
813                 ret = filemap_fdatawrite_range(inode->i_mapping, start,
814                                                orig_end);
815                 if (ret)
816                         return ret;
817         }
818         ret = filemap_fdatawait_range(inode->i_mapping, start, orig_end);
819         if (ret)
820                 return ret;
821
822         end = orig_end;
823         while (1) {
824                 ordered = btrfs_lookup_first_ordered_extent(inode, end);
825                 if (!ordered)
826                         break;
827                 if (ordered->file_offset > orig_end) {
828                         btrfs_put_ordered_extent(ordered);
829                         break;
830                 }
831                 if (ordered->file_offset + ordered->len <= start) {
832                         btrfs_put_ordered_extent(ordered);
833                         break;
834                 }
835                 btrfs_start_ordered_extent(inode, ordered, 1);
836                 end = ordered->file_offset;
837                 if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
838                         ret = -EIO;
839                 btrfs_put_ordered_extent(ordered);
840                 if (ret || end == 0 || end == start)
841                         break;
842                 end--;
843         }
844         return ret;
845 }
846
847 /*
848  * find an ordered extent corresponding to file_offset.  return NULL if
849  * nothing is found, otherwise take a reference on the extent and return it
850  */
851 struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct inode *inode,
852                                                          u64 file_offset)
853 {
854         struct btrfs_ordered_inode_tree *tree;
855         struct rb_node *node;
856         struct btrfs_ordered_extent *entry = NULL;
857
858         tree = &BTRFS_I(inode)->ordered_tree;
859         spin_lock_irq(&tree->lock);
860         node = tree_search(tree, file_offset);
861         if (!node)
862                 goto out;
863
864         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
865         if (!offset_in_entry(entry, file_offset))
866                 entry = NULL;
867         if (entry)
868                 atomic_inc(&entry->refs);
869 out:
870         spin_unlock_irq(&tree->lock);
871         return entry;
872 }
873
874 /* Since the DIO code tries to lock a wide area we need to look for any ordered
875  * extents that exist in the range, rather than just the start of the range.
876  */
877 struct btrfs_ordered_extent *btrfs_lookup_ordered_range(struct inode *inode,
878                                                         u64 file_offset,
879                                                         u64 len)
880 {
881         struct btrfs_ordered_inode_tree *tree;
882         struct rb_node *node;
883         struct btrfs_ordered_extent *entry = NULL;
884
885         tree = &BTRFS_I(inode)->ordered_tree;
886         spin_lock_irq(&tree->lock);
887         node = tree_search(tree, file_offset);
888         if (!node) {
889                 node = tree_search(tree, file_offset + len);
890                 if (!node)
891                         goto out;
892         }
893
894         while (1) {
895                 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
896                 if (range_overlaps(entry, file_offset, len))
897                         break;
898
899                 if (entry->file_offset >= file_offset + len) {
900                         entry = NULL;
901                         break;
902                 }
903                 entry = NULL;
904                 node = rb_next(node);
905                 if (!node)
906                         break;
907         }
908 out:
909         if (entry)
910                 atomic_inc(&entry->refs);
911         spin_unlock_irq(&tree->lock);
912         return entry;
913 }
914
915 /*
916  * lookup and return any extent before 'file_offset'.  NULL is returned
917  * if none is found
918  */
919 struct btrfs_ordered_extent *
920 btrfs_lookup_first_ordered_extent(struct inode *inode, u64 file_offset)
921 {
922         struct btrfs_ordered_inode_tree *tree;
923         struct rb_node *node;
924         struct btrfs_ordered_extent *entry = NULL;
925
926         tree = &BTRFS_I(inode)->ordered_tree;
927         spin_lock_irq(&tree->lock);
928         node = tree_search(tree, file_offset);
929         if (!node)
930                 goto out;
931
932         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
933         atomic_inc(&entry->refs);
934 out:
935         spin_unlock_irq(&tree->lock);
936         return entry;
937 }
938
939 /*
940  * After an extent is done, call this to conditionally update the on disk
941  * i_size.  i_size is updated to cover any fully written part of the file.
942  */
943 int btrfs_ordered_update_i_size(struct inode *inode, u64 offset,
944                                 struct btrfs_ordered_extent *ordered)
945 {
946         struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
947         u64 disk_i_size;
948         u64 new_i_size;
949         u64 i_size = i_size_read(inode);
950         struct rb_node *node;
951         struct rb_node *prev = NULL;
952         struct btrfs_ordered_extent *test;
953         int ret = 1;
954
955         spin_lock_irq(&tree->lock);
956         if (ordered) {
957                 offset = entry_end(ordered);
958                 if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags))
959                         offset = min(offset,
960                                      ordered->file_offset +
961                                      ordered->truncated_len);
962         } else {
963                 offset = ALIGN(offset, BTRFS_I(inode)->root->sectorsize);
964         }
965         disk_i_size = BTRFS_I(inode)->disk_i_size;
966
967         /* truncate file */
968         if (disk_i_size > i_size) {
969                 BTRFS_I(inode)->disk_i_size = i_size;
970                 ret = 0;
971                 goto out;
972         }
973
974         /*
975          * if the disk i_size is already at the inode->i_size, or
976          * this ordered extent is inside the disk i_size, we're done
977          */
978         if (disk_i_size == i_size)
979                 goto out;
980
981         /*
982          * We still need to update disk_i_size if outstanding_isize is greater
983          * than disk_i_size.
984          */
985         if (offset <= disk_i_size &&
986             (!ordered || ordered->outstanding_isize <= disk_i_size))
987                 goto out;
988
989         /*
990          * walk backward from this ordered extent to disk_i_size.
991          * if we find an ordered extent then we can't update disk i_size
992          * yet
993          */
994         if (ordered) {
995                 node = rb_prev(&ordered->rb_node);
996         } else {
997                 prev = tree_search(tree, offset);
998                 /*
999                  * we insert file extents without involving ordered struct,
1000                  * so there should be no ordered struct cover this offset
1001                  */
1002                 if (prev) {
1003                         test = rb_entry(prev, struct btrfs_ordered_extent,
1004                                         rb_node);
1005                         BUG_ON(offset_in_entry(test, offset));
1006                 }
1007                 node = prev;
1008         }
1009         for (; node; node = rb_prev(node)) {
1010                 test = rb_entry(node, struct btrfs_ordered_extent, rb_node);
1011
1012                 /* We treat this entry as if it doesnt exist */
1013                 if (test_bit(BTRFS_ORDERED_UPDATED_ISIZE, &test->flags))
1014                         continue;
1015                 if (test->file_offset + test->len <= disk_i_size)
1016                         break;
1017                 if (test->file_offset >= i_size)
1018                         break;
1019                 if (entry_end(test) > disk_i_size) {
1020                         /*
1021                          * we don't update disk_i_size now, so record this
1022                          * undealt i_size. Or we will not know the real
1023                          * i_size.
1024                          */
1025                         if (test->outstanding_isize < offset)
1026                                 test->outstanding_isize = offset;
1027                         if (ordered &&
1028                             ordered->outstanding_isize >
1029                             test->outstanding_isize)
1030                                 test->outstanding_isize =
1031                                                 ordered->outstanding_isize;
1032                         goto out;
1033                 }
1034         }
1035         new_i_size = min_t(u64, offset, i_size);
1036
1037         /*
1038          * Some ordered extents may completed before the current one, and
1039          * we hold the real i_size in ->outstanding_isize.
1040          */
1041         if (ordered && ordered->outstanding_isize > new_i_size)
1042                 new_i_size = min_t(u64, ordered->outstanding_isize, i_size);
1043         BTRFS_I(inode)->disk_i_size = new_i_size;
1044         ret = 0;
1045 out:
1046         /*
1047          * We need to do this because we can't remove ordered extents until
1048          * after the i_disk_size has been updated and then the inode has been
1049          * updated to reflect the change, so we need to tell anybody who finds
1050          * this ordered extent that we've already done all the real work, we
1051          * just haven't completed all the other work.
1052          */
1053         if (ordered)
1054                 set_bit(BTRFS_ORDERED_UPDATED_ISIZE, &ordered->flags);
1055         spin_unlock_irq(&tree->lock);
1056         return ret;
1057 }
1058
1059 /*
1060  * search the ordered extents for one corresponding to 'offset' and
1061  * try to find a checksum.  This is used because we allow pages to
1062  * be reclaimed before their checksum is actually put into the btree
1063  */
1064 int btrfs_find_ordered_sum(struct inode *inode, u64 offset, u64 disk_bytenr,
1065                            u32 *sum, int len)
1066 {
1067         struct btrfs_ordered_sum *ordered_sum;
1068         struct btrfs_ordered_extent *ordered;
1069         struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
1070         unsigned long num_sectors;
1071         unsigned long i;
1072         u32 sectorsize = BTRFS_I(inode)->root->sectorsize;
1073         int index = 0;
1074
1075         ordered = btrfs_lookup_ordered_extent(inode, offset);
1076         if (!ordered)
1077                 return 0;
1078
1079         spin_lock_irq(&tree->lock);
1080         list_for_each_entry_reverse(ordered_sum, &ordered->list, list) {
1081                 if (disk_bytenr >= ordered_sum->bytenr &&
1082                     disk_bytenr < ordered_sum->bytenr + ordered_sum->len) {
1083                         i = (disk_bytenr - ordered_sum->bytenr) >>
1084                             inode->i_sb->s_blocksize_bits;
1085                         num_sectors = ordered_sum->len >>
1086                                       inode->i_sb->s_blocksize_bits;
1087                         num_sectors = min_t(int, len - index, num_sectors - i);
1088                         memcpy(sum + index, ordered_sum->sums + i,
1089                                num_sectors);
1090
1091                         index += (int)num_sectors;
1092                         if (index == len)
1093                                 goto out;
1094                         disk_bytenr += num_sectors * sectorsize;
1095                 }
1096         }
1097 out:
1098         spin_unlock_irq(&tree->lock);
1099         btrfs_put_ordered_extent(ordered);
1100         return index;
1101 }
1102
1103
1104 /*
1105  * add a given inode to the list of inodes that must be fully on
1106  * disk before a transaction commit finishes.
1107  *
1108  * This basically gives us the ext3 style data=ordered mode, and it is mostly
1109  * used to make sure renamed files are fully on disk.
1110  *
1111  * It is a noop if the inode is already fully on disk.
1112  *
1113  * If trans is not null, we'll do a friendly check for a transaction that
1114  * is already flushing things and force the IO down ourselves.
1115  */
1116 void btrfs_add_ordered_operation(struct btrfs_trans_handle *trans,
1117                                  struct btrfs_root *root, struct inode *inode)
1118 {
1119         struct btrfs_transaction *cur_trans = trans->transaction;
1120         u64 last_mod;
1121
1122         last_mod = max(BTRFS_I(inode)->generation, BTRFS_I(inode)->last_trans);
1123
1124         /*
1125          * if this file hasn't been changed since the last transaction
1126          * commit, we can safely return without doing anything
1127          */
1128         if (last_mod <= root->fs_info->last_trans_committed)
1129                 return;
1130
1131         spin_lock(&root->fs_info->ordered_root_lock);
1132         if (list_empty(&BTRFS_I(inode)->ordered_operations)) {
1133                 list_add_tail(&BTRFS_I(inode)->ordered_operations,
1134                               &cur_trans->ordered_operations);
1135         }
1136         spin_unlock(&root->fs_info->ordered_root_lock);
1137 }
1138
1139 int __init ordered_data_init(void)
1140 {
1141         btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
1142                                      sizeof(struct btrfs_ordered_extent), 0,
1143                                      SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
1144                                      NULL);
1145         if (!btrfs_ordered_extent_cache)
1146                 return -ENOMEM;
1147
1148         return 0;
1149 }
1150
1151 void ordered_data_exit(void)
1152 {
1153         if (btrfs_ordered_extent_cache)
1154                 kmem_cache_destroy(btrfs_ordered_extent_cache);
1155 }