Btrfs: reduce CPU usage in the extent_state tree
[linux-2.6.git] / fs / btrfs / file.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/fs.h>
20 #include <linux/pagemap.h>
21 #include <linux/highmem.h>
22 #include <linux/time.h>
23 #include <linux/init.h>
24 #include <linux/string.h>
25 #include <linux/smp_lock.h>
26 #include <linux/backing-dev.h>
27 #include <linux/mpage.h>
28 #include <linux/swap.h>
29 #include <linux/writeback.h>
30 #include <linux/statfs.h>
31 #include <linux/compat.h>
32 #include "ctree.h"
33 #include "disk-io.h"
34 #include "transaction.h"
35 #include "btrfs_inode.h"
36 #include "ioctl.h"
37 #include "print-tree.h"
38 #include "tree-log.h"
39 #include "locking.h"
40 #include "compat.h"
41
42
43 /* simple helper to fault in pages and copy.  This should go away
44  * and be replaced with calls into generic code.
45  */
46 static noinline int btrfs_copy_from_user(loff_t pos, int num_pages,
47                                          int write_bytes,
48                                          struct page **prepared_pages,
49                                          const char __user *buf)
50 {
51         long page_fault = 0;
52         int i;
53         int offset = pos & (PAGE_CACHE_SIZE - 1);
54
55         for (i = 0; i < num_pages && write_bytes > 0; i++, offset = 0) {
56                 size_t count = min_t(size_t,
57                                      PAGE_CACHE_SIZE - offset, write_bytes);
58                 struct page *page = prepared_pages[i];
59                 fault_in_pages_readable(buf, count);
60
61                 /* Copy data from userspace to the current page */
62                 kmap(page);
63                 page_fault = __copy_from_user(page_address(page) + offset,
64                                               buf, count);
65                 /* Flush processor's dcache for this page */
66                 flush_dcache_page(page);
67                 kunmap(page);
68                 buf += count;
69                 write_bytes -= count;
70
71                 if (page_fault)
72                         break;
73         }
74         return page_fault ? -EFAULT : 0;
75 }
76
77 /*
78  * unlocks pages after btrfs_file_write is done with them
79  */
80 static noinline void btrfs_drop_pages(struct page **pages, size_t num_pages)
81 {
82         size_t i;
83         for (i = 0; i < num_pages; i++) {
84                 if (!pages[i])
85                         break;
86                 /* page checked is some magic around finding pages that
87                  * have been modified without going through btrfs_set_page_dirty
88                  * clear it here
89                  */
90                 ClearPageChecked(pages[i]);
91                 unlock_page(pages[i]);
92                 mark_page_accessed(pages[i]);
93                 page_cache_release(pages[i]);
94         }
95 }
96
97 /*
98  * after copy_from_user, pages need to be dirtied and we need to make
99  * sure holes are created between the current EOF and the start of
100  * any next extents (if required).
101  *
102  * this also makes the decision about creating an inline extent vs
103  * doing real data extents, marking pages dirty and delalloc as required.
104  */
105 static noinline int dirty_and_release_pages(struct btrfs_trans_handle *trans,
106                                    struct btrfs_root *root,
107                                    struct file *file,
108                                    struct page **pages,
109                                    size_t num_pages,
110                                    loff_t pos,
111                                    size_t write_bytes)
112 {
113         int err = 0;
114         int i;
115         struct inode *inode = fdentry(file)->d_inode;
116         u64 num_bytes;
117         u64 start_pos;
118         u64 end_of_last_block;
119         u64 end_pos = pos + write_bytes;
120         loff_t isize = i_size_read(inode);
121
122         start_pos = pos & ~((u64)root->sectorsize - 1);
123         num_bytes = (write_bytes + pos - start_pos +
124                     root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
125
126         end_of_last_block = start_pos + num_bytes - 1;
127         btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block);
128         for (i = 0; i < num_pages; i++) {
129                 struct page *p = pages[i];
130                 SetPageUptodate(p);
131                 ClearPageChecked(p);
132                 set_page_dirty(p);
133         }
134         if (end_pos > isize) {
135                 i_size_write(inode, end_pos);
136                 /* we've only changed i_size in ram, and we haven't updated
137                  * the disk i_size.  There is no need to log the inode
138                  * at this time.
139                  */
140         }
141         return err;
142 }
143
144 /*
145  * this drops all the extents in the cache that intersect the range
146  * [start, end].  Existing extents are split as required.
147  */
148 int btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
149                             int skip_pinned)
150 {
151         struct extent_map *em;
152         struct extent_map *split = NULL;
153         struct extent_map *split2 = NULL;
154         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
155         u64 len = end - start + 1;
156         int ret;
157         int testend = 1;
158         unsigned long flags;
159         int compressed = 0;
160
161         WARN_ON(end < start);
162         if (end == (u64)-1) {
163                 len = (u64)-1;
164                 testend = 0;
165         }
166         while (1) {
167                 if (!split)
168                         split = alloc_extent_map(GFP_NOFS);
169                 if (!split2)
170                         split2 = alloc_extent_map(GFP_NOFS);
171
172                 write_lock(&em_tree->lock);
173                 em = lookup_extent_mapping(em_tree, start, len);
174                 if (!em) {
175                         write_unlock(&em_tree->lock);
176                         break;
177                 }
178                 flags = em->flags;
179                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
180                         write_unlock(&em_tree->lock);
181                         if (em->start <= start &&
182                             (!testend || em->start + em->len >= start + len)) {
183                                 free_extent_map(em);
184                                 break;
185                         }
186                         if (start < em->start) {
187                                 len = em->start - start;
188                         } else {
189                                 len = start + len - (em->start + em->len);
190                                 start = em->start + em->len;
191                         }
192                         free_extent_map(em);
193                         continue;
194                 }
195                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
196                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
197                 remove_extent_mapping(em_tree, em);
198
199                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
200                     em->start < start) {
201                         split->start = em->start;
202                         split->len = start - em->start;
203                         split->orig_start = em->orig_start;
204                         split->block_start = em->block_start;
205
206                         if (compressed)
207                                 split->block_len = em->block_len;
208                         else
209                                 split->block_len = split->len;
210
211                         split->bdev = em->bdev;
212                         split->flags = flags;
213                         ret = add_extent_mapping(em_tree, split);
214                         BUG_ON(ret);
215                         free_extent_map(split);
216                         split = split2;
217                         split2 = NULL;
218                 }
219                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
220                     testend && em->start + em->len > start + len) {
221                         u64 diff = start + len - em->start;
222
223                         split->start = start + len;
224                         split->len = em->start + em->len - (start + len);
225                         split->bdev = em->bdev;
226                         split->flags = flags;
227
228                         if (compressed) {
229                                 split->block_len = em->block_len;
230                                 split->block_start = em->block_start;
231                                 split->orig_start = em->orig_start;
232                         } else {
233                                 split->block_len = split->len;
234                                 split->block_start = em->block_start + diff;
235                                 split->orig_start = split->start;
236                         }
237
238                         ret = add_extent_mapping(em_tree, split);
239                         BUG_ON(ret);
240                         free_extent_map(split);
241                         split = NULL;
242                 }
243                 write_unlock(&em_tree->lock);
244
245                 /* once for us */
246                 free_extent_map(em);
247                 /* once for the tree*/
248                 free_extent_map(em);
249         }
250         if (split)
251                 free_extent_map(split);
252         if (split2)
253                 free_extent_map(split2);
254         return 0;
255 }
256
257 /*
258  * this is very complex, but the basic idea is to drop all extents
259  * in the range start - end.  hint_block is filled in with a block number
260  * that would be a good hint to the block allocator for this file.
261  *
262  * If an extent intersects the range but is not entirely inside the range
263  * it is either truncated or split.  Anything entirely inside the range
264  * is deleted from the tree.
265  *
266  * inline_limit is used to tell this code which offsets in the file to keep
267  * if they contain inline extents.
268  */
269 noinline int btrfs_drop_extents(struct btrfs_trans_handle *trans,
270                        struct btrfs_root *root, struct inode *inode,
271                        u64 start, u64 end, u64 locked_end,
272                        u64 inline_limit, u64 *hint_byte)
273 {
274         u64 extent_end = 0;
275         u64 search_start = start;
276         u64 ram_bytes = 0;
277         u64 disk_bytenr = 0;
278         u64 orig_locked_end = locked_end;
279         u8 compression;
280         u8 encryption;
281         u16 other_encoding = 0;
282         struct extent_buffer *leaf;
283         struct btrfs_file_extent_item *extent;
284         struct btrfs_path *path;
285         struct btrfs_key key;
286         struct btrfs_file_extent_item old;
287         int keep;
288         int slot;
289         int bookend;
290         int found_type = 0;
291         int found_extent;
292         int found_inline;
293         int recow;
294         int ret;
295
296         inline_limit = 0;
297         btrfs_drop_extent_cache(inode, start, end - 1, 0);
298
299         path = btrfs_alloc_path();
300         if (!path)
301                 return -ENOMEM;
302         while (1) {
303                 recow = 0;
304                 btrfs_release_path(root, path);
305                 ret = btrfs_lookup_file_extent(trans, root, path, inode->i_ino,
306                                                search_start, -1);
307                 if (ret < 0)
308                         goto out;
309                 if (ret > 0) {
310                         if (path->slots[0] == 0) {
311                                 ret = 0;
312                                 goto out;
313                         }
314                         path->slots[0]--;
315                 }
316 next_slot:
317                 keep = 0;
318                 bookend = 0;
319                 found_extent = 0;
320                 found_inline = 0;
321                 compression = 0;
322                 encryption = 0;
323                 extent = NULL;
324                 leaf = path->nodes[0];
325                 slot = path->slots[0];
326                 ret = 0;
327                 btrfs_item_key_to_cpu(leaf, &key, slot);
328                 if (btrfs_key_type(&key) == BTRFS_EXTENT_DATA_KEY &&
329                     key.offset >= end) {
330                         goto out;
331                 }
332                 if (btrfs_key_type(&key) > BTRFS_EXTENT_DATA_KEY ||
333                     key.objectid != inode->i_ino) {
334                         goto out;
335                 }
336                 if (recow) {
337                         search_start = max(key.offset, start);
338                         continue;
339                 }
340                 if (btrfs_key_type(&key) == BTRFS_EXTENT_DATA_KEY) {
341                         extent = btrfs_item_ptr(leaf, slot,
342                                                 struct btrfs_file_extent_item);
343                         found_type = btrfs_file_extent_type(leaf, extent);
344                         compression = btrfs_file_extent_compression(leaf,
345                                                                     extent);
346                         encryption = btrfs_file_extent_encryption(leaf,
347                                                                   extent);
348                         other_encoding = btrfs_file_extent_other_encoding(leaf,
349                                                                   extent);
350                         if (found_type == BTRFS_FILE_EXTENT_REG ||
351                             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
352                                 extent_end =
353                                      btrfs_file_extent_disk_bytenr(leaf,
354                                                                    extent);
355                                 if (extent_end)
356                                         *hint_byte = extent_end;
357
358                                 extent_end = key.offset +
359                                      btrfs_file_extent_num_bytes(leaf, extent);
360                                 ram_bytes = btrfs_file_extent_ram_bytes(leaf,
361                                                                 extent);
362                                 found_extent = 1;
363                         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
364                                 found_inline = 1;
365                                 extent_end = key.offset +
366                                      btrfs_file_extent_inline_len(leaf, extent);
367                         }
368                 } else {
369                         extent_end = search_start;
370                 }
371
372                 /* we found nothing we can drop */
373                 if ((!found_extent && !found_inline) ||
374                     search_start >= extent_end) {
375                         int nextret;
376                         u32 nritems;
377                         nritems = btrfs_header_nritems(leaf);
378                         if (slot >= nritems - 1) {
379                                 nextret = btrfs_next_leaf(root, path);
380                                 if (nextret)
381                                         goto out;
382                                 recow = 1;
383                         } else {
384                                 path->slots[0]++;
385                         }
386                         goto next_slot;
387                 }
388
389                 if (end <= extent_end && start >= key.offset && found_inline)
390                         *hint_byte = EXTENT_MAP_INLINE;
391
392                 if (found_extent) {
393                         read_extent_buffer(leaf, &old, (unsigned long)extent,
394                                            sizeof(old));
395                 }
396
397                 if (end < extent_end && end >= key.offset) {
398                         bookend = 1;
399                         if (found_inline && start <= key.offset)
400                                 keep = 1;
401                 }
402
403                 if (bookend && found_extent) {
404                         if (locked_end < extent_end) {
405                                 ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
406                                                 locked_end, extent_end - 1,
407                                                 GFP_NOFS);
408                                 if (!ret) {
409                                         btrfs_release_path(root, path);
410                                         lock_extent(&BTRFS_I(inode)->io_tree,
411                                                 locked_end, extent_end - 1,
412                                                 GFP_NOFS);
413                                         locked_end = extent_end;
414                                         continue;
415                                 }
416                                 locked_end = extent_end;
417                         }
418                         disk_bytenr = le64_to_cpu(old.disk_bytenr);
419                         if (disk_bytenr != 0) {
420                                 ret = btrfs_inc_extent_ref(trans, root,
421                                            disk_bytenr,
422                                            le64_to_cpu(old.disk_num_bytes), 0,
423                                            root->root_key.objectid,
424                                            key.objectid, key.offset -
425                                            le64_to_cpu(old.offset));
426                                 BUG_ON(ret);
427                         }
428                 }
429
430                 if (found_inline) {
431                         u64 mask = root->sectorsize - 1;
432                         search_start = (extent_end + mask) & ~mask;
433                 } else
434                         search_start = extent_end;
435
436                 /* truncate existing extent */
437                 if (start > key.offset) {
438                         u64 new_num;
439                         u64 old_num;
440                         keep = 1;
441                         WARN_ON(start & (root->sectorsize - 1));
442                         if (found_extent) {
443                                 new_num = start - key.offset;
444                                 old_num = btrfs_file_extent_num_bytes(leaf,
445                                                                       extent);
446                                 *hint_byte =
447                                         btrfs_file_extent_disk_bytenr(leaf,
448                                                                       extent);
449                                 if (btrfs_file_extent_disk_bytenr(leaf,
450                                                                   extent)) {
451                                         inode_sub_bytes(inode, old_num -
452                                                         new_num);
453                                 }
454                                 btrfs_set_file_extent_num_bytes(leaf,
455                                                         extent, new_num);
456                                 btrfs_mark_buffer_dirty(leaf);
457                         } else if (key.offset < inline_limit &&
458                                    (end > extent_end) &&
459                                    (inline_limit < extent_end)) {
460                                 u32 new_size;
461                                 new_size = btrfs_file_extent_calc_inline_size(
462                                                    inline_limit - key.offset);
463                                 inode_sub_bytes(inode, extent_end -
464                                                 inline_limit);
465                                 btrfs_set_file_extent_ram_bytes(leaf, extent,
466                                                         new_size);
467                                 if (!compression && !encryption) {
468                                         btrfs_truncate_item(trans, root, path,
469                                                             new_size, 1);
470                                 }
471                         }
472                 }
473                 /* delete the entire extent */
474                 if (!keep) {
475                         if (found_inline)
476                                 inode_sub_bytes(inode, extent_end -
477                                                 key.offset);
478                         ret = btrfs_del_item(trans, root, path);
479                         /* TODO update progress marker and return */
480                         BUG_ON(ret);
481                         extent = NULL;
482                         btrfs_release_path(root, path);
483                         /* the extent will be freed later */
484                 }
485                 if (bookend && found_inline && start <= key.offset) {
486                         u32 new_size;
487                         new_size = btrfs_file_extent_calc_inline_size(
488                                                    extent_end - end);
489                         inode_sub_bytes(inode, end - key.offset);
490                         btrfs_set_file_extent_ram_bytes(leaf, extent,
491                                                         new_size);
492                         if (!compression && !encryption)
493                                 ret = btrfs_truncate_item(trans, root, path,
494                                                           new_size, 0);
495                         BUG_ON(ret);
496                 }
497                 /* create bookend, splitting the extent in two */
498                 if (bookend && found_extent) {
499                         struct btrfs_key ins;
500                         ins.objectid = inode->i_ino;
501                         ins.offset = end;
502                         btrfs_set_key_type(&ins, BTRFS_EXTENT_DATA_KEY);
503
504                         btrfs_release_path(root, path);
505                         path->leave_spinning = 1;
506                         ret = btrfs_insert_empty_item(trans, root, path, &ins,
507                                                       sizeof(*extent));
508                         BUG_ON(ret);
509
510                         leaf = path->nodes[0];
511                         extent = btrfs_item_ptr(leaf, path->slots[0],
512                                                 struct btrfs_file_extent_item);
513                         write_extent_buffer(leaf, &old,
514                                             (unsigned long)extent, sizeof(old));
515
516                         btrfs_set_file_extent_compression(leaf, extent,
517                                                           compression);
518                         btrfs_set_file_extent_encryption(leaf, extent,
519                                                          encryption);
520                         btrfs_set_file_extent_other_encoding(leaf, extent,
521                                                              other_encoding);
522                         btrfs_set_file_extent_offset(leaf, extent,
523                                     le64_to_cpu(old.offset) + end - key.offset);
524                         WARN_ON(le64_to_cpu(old.num_bytes) <
525                                 (extent_end - end));
526                         btrfs_set_file_extent_num_bytes(leaf, extent,
527                                                         extent_end - end);
528
529                         /*
530                          * set the ram bytes to the size of the full extent
531                          * before splitting.  This is a worst case flag,
532                          * but its the best we can do because we don't know
533                          * how splitting affects compression
534                          */
535                         btrfs_set_file_extent_ram_bytes(leaf, extent,
536                                                         ram_bytes);
537                         btrfs_set_file_extent_type(leaf, extent, found_type);
538
539                         btrfs_unlock_up_safe(path, 1);
540                         btrfs_mark_buffer_dirty(path->nodes[0]);
541                         btrfs_set_lock_blocking(path->nodes[0]);
542
543                         path->leave_spinning = 0;
544                         btrfs_release_path(root, path);
545                         if (disk_bytenr != 0)
546                                 inode_add_bytes(inode, extent_end - end);
547                 }
548
549                 if (found_extent && !keep) {
550                         u64 old_disk_bytenr = le64_to_cpu(old.disk_bytenr);
551
552                         if (old_disk_bytenr != 0) {
553                                 inode_sub_bytes(inode,
554                                                 le64_to_cpu(old.num_bytes));
555                                 ret = btrfs_free_extent(trans, root,
556                                                 old_disk_bytenr,
557                                                 le64_to_cpu(old.disk_num_bytes),
558                                                 0, root->root_key.objectid,
559                                                 key.objectid, key.offset -
560                                                 le64_to_cpu(old.offset));
561                                 BUG_ON(ret);
562                                 *hint_byte = old_disk_bytenr;
563                         }
564                 }
565
566                 if (search_start >= end) {
567                         ret = 0;
568                         goto out;
569                 }
570         }
571 out:
572         btrfs_free_path(path);
573         if (locked_end > orig_locked_end) {
574                 unlock_extent(&BTRFS_I(inode)->io_tree, orig_locked_end,
575                               locked_end - 1, GFP_NOFS);
576         }
577         return ret;
578 }
579
580 static int extent_mergeable(struct extent_buffer *leaf, int slot,
581                             u64 objectid, u64 bytenr, u64 *start, u64 *end)
582 {
583         struct btrfs_file_extent_item *fi;
584         struct btrfs_key key;
585         u64 extent_end;
586
587         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
588                 return 0;
589
590         btrfs_item_key_to_cpu(leaf, &key, slot);
591         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
592                 return 0;
593
594         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
595         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
596             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
597             btrfs_file_extent_compression(leaf, fi) ||
598             btrfs_file_extent_encryption(leaf, fi) ||
599             btrfs_file_extent_other_encoding(leaf, fi))
600                 return 0;
601
602         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
603         if ((*start && *start != key.offset) || (*end && *end != extent_end))
604                 return 0;
605
606         *start = key.offset;
607         *end = extent_end;
608         return 1;
609 }
610
611 /*
612  * Mark extent in the range start - end as written.
613  *
614  * This changes extent type from 'pre-allocated' to 'regular'. If only
615  * part of extent is marked as written, the extent will be split into
616  * two or three.
617  */
618 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
619                               struct btrfs_root *root,
620                               struct inode *inode, u64 start, u64 end)
621 {
622         struct extent_buffer *leaf;
623         struct btrfs_path *path;
624         struct btrfs_file_extent_item *fi;
625         struct btrfs_key key;
626         u64 bytenr;
627         u64 num_bytes;
628         u64 extent_end;
629         u64 orig_offset;
630         u64 other_start;
631         u64 other_end;
632         u64 split = start;
633         u64 locked_end = end;
634         int extent_type;
635         int split_end = 1;
636         int ret;
637
638         btrfs_drop_extent_cache(inode, start, end - 1, 0);
639
640         path = btrfs_alloc_path();
641         BUG_ON(!path);
642 again:
643         key.objectid = inode->i_ino;
644         key.type = BTRFS_EXTENT_DATA_KEY;
645         if (split == start)
646                 key.offset = split;
647         else
648                 key.offset = split - 1;
649
650         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
651         if (ret > 0 && path->slots[0] > 0)
652                 path->slots[0]--;
653
654         leaf = path->nodes[0];
655         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
656         BUG_ON(key.objectid != inode->i_ino ||
657                key.type != BTRFS_EXTENT_DATA_KEY);
658         fi = btrfs_item_ptr(leaf, path->slots[0],
659                             struct btrfs_file_extent_item);
660         extent_type = btrfs_file_extent_type(leaf, fi);
661         BUG_ON(extent_type != BTRFS_FILE_EXTENT_PREALLOC);
662         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
663         BUG_ON(key.offset > start || extent_end < end);
664
665         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
666         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
667         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
668
669         if (key.offset == start)
670                 split = end;
671
672         if (key.offset == start && extent_end == end) {
673                 int del_nr = 0;
674                 int del_slot = 0;
675                 other_start = end;
676                 other_end = 0;
677                 if (extent_mergeable(leaf, path->slots[0] + 1, inode->i_ino,
678                                      bytenr, &other_start, &other_end)) {
679                         extent_end = other_end;
680                         del_slot = path->slots[0] + 1;
681                         del_nr++;
682                         ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
683                                                 0, root->root_key.objectid,
684                                                 inode->i_ino, orig_offset);
685                         BUG_ON(ret);
686                 }
687                 other_start = 0;
688                 other_end = start;
689                 if (extent_mergeable(leaf, path->slots[0] - 1, inode->i_ino,
690                                      bytenr, &other_start, &other_end)) {
691                         key.offset = other_start;
692                         del_slot = path->slots[0];
693                         del_nr++;
694                         ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
695                                                 0, root->root_key.objectid,
696                                                 inode->i_ino, orig_offset);
697                         BUG_ON(ret);
698                 }
699                 split_end = 0;
700                 if (del_nr == 0) {
701                         btrfs_set_file_extent_type(leaf, fi,
702                                                    BTRFS_FILE_EXTENT_REG);
703                         goto done;
704                 }
705
706                 fi = btrfs_item_ptr(leaf, del_slot - 1,
707                                     struct btrfs_file_extent_item);
708                 btrfs_set_file_extent_type(leaf, fi, BTRFS_FILE_EXTENT_REG);
709                 btrfs_set_file_extent_num_bytes(leaf, fi,
710                                                 extent_end - key.offset);
711                 btrfs_mark_buffer_dirty(leaf);
712
713                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
714                 BUG_ON(ret);
715                 goto release;
716         } else if (split == start) {
717                 if (locked_end < extent_end) {
718                         ret = try_lock_extent(&BTRFS_I(inode)->io_tree,
719                                         locked_end, extent_end - 1, GFP_NOFS);
720                         if (!ret) {
721                                 btrfs_release_path(root, path);
722                                 lock_extent(&BTRFS_I(inode)->io_tree,
723                                         locked_end, extent_end - 1, GFP_NOFS);
724                                 locked_end = extent_end;
725                                 goto again;
726                         }
727                         locked_end = extent_end;
728                 }
729                 btrfs_set_file_extent_num_bytes(leaf, fi, split - key.offset);
730         } else  {
731                 BUG_ON(key.offset != start);
732                 key.offset = split;
733                 btrfs_set_file_extent_offset(leaf, fi, key.offset -
734                                              orig_offset);
735                 btrfs_set_file_extent_num_bytes(leaf, fi, extent_end - split);
736                 btrfs_set_item_key_safe(trans, root, path, &key);
737                 extent_end = split;
738         }
739
740         if (extent_end == end) {
741                 split_end = 0;
742                 extent_type = BTRFS_FILE_EXTENT_REG;
743         }
744         if (extent_end == end && split == start) {
745                 other_start = end;
746                 other_end = 0;
747                 if (extent_mergeable(leaf, path->slots[0] + 1, inode->i_ino,
748                                      bytenr, &other_start, &other_end)) {
749                         path->slots[0]++;
750                         fi = btrfs_item_ptr(leaf, path->slots[0],
751                                             struct btrfs_file_extent_item);
752                         key.offset = split;
753                         btrfs_set_item_key_safe(trans, root, path, &key);
754                         btrfs_set_file_extent_offset(leaf, fi, key.offset -
755                                                      orig_offset);
756                         btrfs_set_file_extent_num_bytes(leaf, fi,
757                                                         other_end - split);
758                         goto done;
759                 }
760         }
761         if (extent_end == end && split == end) {
762                 other_start = 0;
763                 other_end = start;
764                 if (extent_mergeable(leaf, path->slots[0] - 1 , inode->i_ino,
765                                      bytenr, &other_start, &other_end)) {
766                         path->slots[0]--;
767                         fi = btrfs_item_ptr(leaf, path->slots[0],
768                                             struct btrfs_file_extent_item);
769                         btrfs_set_file_extent_num_bytes(leaf, fi, extent_end -
770                                                         other_start);
771                         goto done;
772                 }
773         }
774
775         btrfs_mark_buffer_dirty(leaf);
776
777         ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
778                                    root->root_key.objectid,
779                                    inode->i_ino, orig_offset);
780         BUG_ON(ret);
781         btrfs_release_path(root, path);
782
783         key.offset = start;
784         ret = btrfs_insert_empty_item(trans, root, path, &key, sizeof(*fi));
785         BUG_ON(ret);
786
787         leaf = path->nodes[0];
788         fi = btrfs_item_ptr(leaf, path->slots[0],
789                             struct btrfs_file_extent_item);
790         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
791         btrfs_set_file_extent_type(leaf, fi, extent_type);
792         btrfs_set_file_extent_disk_bytenr(leaf, fi, bytenr);
793         btrfs_set_file_extent_disk_num_bytes(leaf, fi, num_bytes);
794         btrfs_set_file_extent_offset(leaf, fi, key.offset - orig_offset);
795         btrfs_set_file_extent_num_bytes(leaf, fi, extent_end - key.offset);
796         btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
797         btrfs_set_file_extent_compression(leaf, fi, 0);
798         btrfs_set_file_extent_encryption(leaf, fi, 0);
799         btrfs_set_file_extent_other_encoding(leaf, fi, 0);
800 done:
801         btrfs_mark_buffer_dirty(leaf);
802
803 release:
804         btrfs_release_path(root, path);
805         if (split_end && split == start) {
806                 split = end;
807                 goto again;
808         }
809         if (locked_end > end) {
810                 unlock_extent(&BTRFS_I(inode)->io_tree, end, locked_end - 1,
811                               GFP_NOFS);
812         }
813         btrfs_free_path(path);
814         return 0;
815 }
816
817 /*
818  * this gets pages into the page cache and locks them down, it also properly
819  * waits for data=ordered extents to finish before allowing the pages to be
820  * modified.
821  */
822 static noinline int prepare_pages(struct btrfs_root *root, struct file *file,
823                          struct page **pages, size_t num_pages,
824                          loff_t pos, unsigned long first_index,
825                          unsigned long last_index, size_t write_bytes)
826 {
827         int i;
828         unsigned long index = pos >> PAGE_CACHE_SHIFT;
829         struct inode *inode = fdentry(file)->d_inode;
830         int err = 0;
831         u64 start_pos;
832         u64 last_pos;
833
834         start_pos = pos & ~((u64)root->sectorsize - 1);
835         last_pos = ((u64)index + num_pages) << PAGE_CACHE_SHIFT;
836
837         if (start_pos > inode->i_size) {
838                 err = btrfs_cont_expand(inode, start_pos);
839                 if (err)
840                         return err;
841         }
842
843         memset(pages, 0, num_pages * sizeof(struct page *));
844 again:
845         for (i = 0; i < num_pages; i++) {
846                 pages[i] = grab_cache_page(inode->i_mapping, index + i);
847                 if (!pages[i]) {
848                         err = -ENOMEM;
849                         BUG_ON(1);
850                 }
851                 wait_on_page_writeback(pages[i]);
852         }
853         if (start_pos < inode->i_size) {
854                 struct btrfs_ordered_extent *ordered;
855                 lock_extent(&BTRFS_I(inode)->io_tree,
856                             start_pos, last_pos - 1, GFP_NOFS);
857                 ordered = btrfs_lookup_first_ordered_extent(inode,
858                                                             last_pos - 1);
859                 if (ordered &&
860                     ordered->file_offset + ordered->len > start_pos &&
861                     ordered->file_offset < last_pos) {
862                         btrfs_put_ordered_extent(ordered);
863                         unlock_extent(&BTRFS_I(inode)->io_tree,
864                                       start_pos, last_pos - 1, GFP_NOFS);
865                         for (i = 0; i < num_pages; i++) {
866                                 unlock_page(pages[i]);
867                                 page_cache_release(pages[i]);
868                         }
869                         btrfs_wait_ordered_range(inode, start_pos,
870                                                  last_pos - start_pos);
871                         goto again;
872                 }
873                 if (ordered)
874                         btrfs_put_ordered_extent(ordered);
875
876                 clear_extent_bits(&BTRFS_I(inode)->io_tree, start_pos,
877                                   last_pos - 1, EXTENT_DIRTY | EXTENT_DELALLOC,
878                                   GFP_NOFS);
879                 unlock_extent(&BTRFS_I(inode)->io_tree,
880                               start_pos, last_pos - 1, GFP_NOFS);
881         }
882         for (i = 0; i < num_pages; i++) {
883                 clear_page_dirty_for_io(pages[i]);
884                 set_page_extent_mapped(pages[i]);
885                 WARN_ON(!PageLocked(pages[i]));
886         }
887         return 0;
888 }
889
890 static ssize_t btrfs_file_write(struct file *file, const char __user *buf,
891                                 size_t count, loff_t *ppos)
892 {
893         loff_t pos;
894         loff_t start_pos;
895         ssize_t num_written = 0;
896         ssize_t err = 0;
897         int ret = 0;
898         struct inode *inode = fdentry(file)->d_inode;
899         struct btrfs_root *root = BTRFS_I(inode)->root;
900         struct page **pages = NULL;
901         int nrptrs;
902         struct page *pinned[2];
903         unsigned long first_index;
904         unsigned long last_index;
905         int will_write;
906
907         will_write = ((file->f_flags & O_SYNC) || IS_SYNC(inode) ||
908                       (file->f_flags & O_DIRECT));
909
910         nrptrs = min((count + PAGE_CACHE_SIZE - 1) / PAGE_CACHE_SIZE,
911                      PAGE_CACHE_SIZE / (sizeof(struct page *)));
912         pinned[0] = NULL;
913         pinned[1] = NULL;
914
915         pos = *ppos;
916         start_pos = pos;
917
918         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
919         current->backing_dev_info = inode->i_mapping->backing_dev_info;
920         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
921         if (err)
922                 goto out_nolock;
923         if (count == 0)
924                 goto out_nolock;
925
926         err = file_remove_suid(file);
927         if (err)
928                 goto out_nolock;
929         file_update_time(file);
930
931         pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
932
933         mutex_lock(&inode->i_mutex);
934         BTRFS_I(inode)->sequence++;
935         first_index = pos >> PAGE_CACHE_SHIFT;
936         last_index = (pos + count) >> PAGE_CACHE_SHIFT;
937
938         /*
939          * there are lots of better ways to do this, but this code
940          * makes sure the first and last page in the file range are
941          * up to date and ready for cow
942          */
943         if ((pos & (PAGE_CACHE_SIZE - 1))) {
944                 pinned[0] = grab_cache_page(inode->i_mapping, first_index);
945                 if (!PageUptodate(pinned[0])) {
946                         ret = btrfs_readpage(NULL, pinned[0]);
947                         BUG_ON(ret);
948                         wait_on_page_locked(pinned[0]);
949                 } else {
950                         unlock_page(pinned[0]);
951                 }
952         }
953         if ((pos + count) & (PAGE_CACHE_SIZE - 1)) {
954                 pinned[1] = grab_cache_page(inode->i_mapping, last_index);
955                 if (!PageUptodate(pinned[1])) {
956                         ret = btrfs_readpage(NULL, pinned[1]);
957                         BUG_ON(ret);
958                         wait_on_page_locked(pinned[1]);
959                 } else {
960                         unlock_page(pinned[1]);
961                 }
962         }
963
964         while (count > 0) {
965                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
966                 size_t write_bytes = min(count, nrptrs *
967                                         (size_t)PAGE_CACHE_SIZE -
968                                          offset);
969                 size_t num_pages = (write_bytes + PAGE_CACHE_SIZE - 1) >>
970                                         PAGE_CACHE_SHIFT;
971
972                 WARN_ON(num_pages > nrptrs);
973                 memset(pages, 0, sizeof(struct page *) * nrptrs);
974
975                 ret = btrfs_check_data_free_space(root, inode, write_bytes);
976                 if (ret)
977                         goto out;
978
979                 ret = prepare_pages(root, file, pages, num_pages,
980                                     pos, first_index, last_index,
981                                     write_bytes);
982                 if (ret) {
983                         btrfs_free_reserved_data_space(root, inode,
984                                                        write_bytes);
985                         goto out;
986                 }
987
988                 ret = btrfs_copy_from_user(pos, num_pages,
989                                            write_bytes, pages, buf);
990                 if (ret) {
991                         btrfs_free_reserved_data_space(root, inode,
992                                                        write_bytes);
993                         btrfs_drop_pages(pages, num_pages);
994                         goto out;
995                 }
996
997                 ret = dirty_and_release_pages(NULL, root, file, pages,
998                                               num_pages, pos, write_bytes);
999                 btrfs_drop_pages(pages, num_pages);
1000                 if (ret) {
1001                         btrfs_free_reserved_data_space(root, inode,
1002                                                        write_bytes);
1003                         goto out;
1004                 }
1005
1006                 if (will_write) {
1007                         btrfs_fdatawrite_range(inode->i_mapping, pos,
1008                                                pos + write_bytes - 1,
1009                                                WB_SYNC_ALL);
1010                 } else {
1011                         balance_dirty_pages_ratelimited_nr(inode->i_mapping,
1012                                                            num_pages);
1013                         if (num_pages <
1014                             (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
1015                                 btrfs_btree_balance_dirty(root, 1);
1016                         btrfs_throttle(root);
1017                 }
1018
1019                 buf += write_bytes;
1020                 count -= write_bytes;
1021                 pos += write_bytes;
1022                 num_written += write_bytes;
1023
1024                 cond_resched();
1025         }
1026 out:
1027         mutex_unlock(&inode->i_mutex);
1028         if (ret)
1029                 err = ret;
1030
1031 out_nolock:
1032         kfree(pages);
1033         if (pinned[0])
1034                 page_cache_release(pinned[0]);
1035         if (pinned[1])
1036                 page_cache_release(pinned[1]);
1037         *ppos = pos;
1038
1039         /*
1040          * we want to make sure fsync finds this change
1041          * but we haven't joined a transaction running right now.
1042          *
1043          * Later on, someone is sure to update the inode and get the
1044          * real transid recorded.
1045          *
1046          * We set last_trans now to the fs_info generation + 1,
1047          * this will either be one more than the running transaction
1048          * or the generation used for the next transaction if there isn't
1049          * one running right now.
1050          */
1051         BTRFS_I(inode)->last_trans = root->fs_info->generation + 1;
1052
1053         if (num_written > 0 && will_write) {
1054                 struct btrfs_trans_handle *trans;
1055
1056                 err = btrfs_wait_ordered_range(inode, start_pos, num_written);
1057                 if (err)
1058                         num_written = err;
1059
1060                 if ((file->f_flags & O_SYNC) || IS_SYNC(inode)) {
1061                         trans = btrfs_start_transaction(root, 1);
1062                         ret = btrfs_log_dentry_safe(trans, root,
1063                                                     file->f_dentry);
1064                         if (ret == 0) {
1065                                 ret = btrfs_sync_log(trans, root);
1066                                 if (ret == 0)
1067                                         btrfs_end_transaction(trans, root);
1068                                 else
1069                                         btrfs_commit_transaction(trans, root);
1070                         } else {
1071                                 btrfs_commit_transaction(trans, root);
1072                         }
1073                 }
1074                 if (file->f_flags & O_DIRECT) {
1075                         invalidate_mapping_pages(inode->i_mapping,
1076                               start_pos >> PAGE_CACHE_SHIFT,
1077                              (start_pos + num_written - 1) >> PAGE_CACHE_SHIFT);
1078                 }
1079         }
1080         current->backing_dev_info = NULL;
1081         return num_written ? num_written : err;
1082 }
1083
1084 int btrfs_release_file(struct inode *inode, struct file *filp)
1085 {
1086         /*
1087          * ordered_data_close is set by settattr when we are about to truncate
1088          * a file from a non-zero size to a zero size.  This tries to
1089          * flush down new bytes that may have been written if the
1090          * application were using truncate to replace a file in place.
1091          */
1092         if (BTRFS_I(inode)->ordered_data_close) {
1093                 BTRFS_I(inode)->ordered_data_close = 0;
1094                 btrfs_add_ordered_operation(NULL, BTRFS_I(inode)->root, inode);
1095                 if (inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
1096                         filemap_flush(inode->i_mapping);
1097         }
1098         if (filp->private_data)
1099                 btrfs_ioctl_trans_end(filp);
1100         return 0;
1101 }
1102
1103 /*
1104  * fsync call for both files and directories.  This logs the inode into
1105  * the tree log instead of forcing full commits whenever possible.
1106  *
1107  * It needs to call filemap_fdatawait so that all ordered extent updates are
1108  * in the metadata btree are up to date for copying to the log.
1109  *
1110  * It drops the inode mutex before doing the tree log commit.  This is an
1111  * important optimization for directories because holding the mutex prevents
1112  * new operations on the dir while we write to disk.
1113  */
1114 int btrfs_sync_file(struct file *file, struct dentry *dentry, int datasync)
1115 {
1116         struct inode *inode = dentry->d_inode;
1117         struct btrfs_root *root = BTRFS_I(inode)->root;
1118         int ret = 0;
1119         struct btrfs_trans_handle *trans;
1120
1121         /*
1122          * check the transaction that last modified this inode
1123          * and see if its already been committed
1124          */
1125         if (!BTRFS_I(inode)->last_trans)
1126                 goto out;
1127
1128         mutex_lock(&root->fs_info->trans_mutex);
1129         if (BTRFS_I(inode)->last_trans <=
1130             root->fs_info->last_trans_committed) {
1131                 BTRFS_I(inode)->last_trans = 0;
1132                 mutex_unlock(&root->fs_info->trans_mutex);
1133                 goto out;
1134         }
1135         mutex_unlock(&root->fs_info->trans_mutex);
1136
1137         root->log_batch++;
1138         filemap_fdatawrite(inode->i_mapping);
1139         btrfs_wait_ordered_range(inode, 0, (u64)-1);
1140         root->log_batch++;
1141
1142         if (datasync && !(inode->i_state & I_DIRTY_PAGES))
1143                 goto out;
1144         /*
1145          * ok we haven't committed the transaction yet, lets do a commit
1146          */
1147         if (file && file->private_data)
1148                 btrfs_ioctl_trans_end(file);
1149
1150         trans = btrfs_start_transaction(root, 1);
1151         if (!trans) {
1152                 ret = -ENOMEM;
1153                 goto out;
1154         }
1155
1156         ret = btrfs_log_dentry_safe(trans, root, dentry);
1157         if (ret < 0)
1158                 goto out;
1159
1160         /* we've logged all the items and now have a consistent
1161          * version of the file in the log.  It is possible that
1162          * someone will come in and modify the file, but that's
1163          * fine because the log is consistent on disk, and we
1164          * have references to all of the file's extents
1165          *
1166          * It is possible that someone will come in and log the
1167          * file again, but that will end up using the synchronization
1168          * inside btrfs_sync_log to keep things safe.
1169          */
1170         mutex_unlock(&dentry->d_inode->i_mutex);
1171
1172         if (ret > 0) {
1173                 ret = btrfs_commit_transaction(trans, root);
1174         } else {
1175                 ret = btrfs_sync_log(trans, root);
1176                 if (ret == 0)
1177                         ret = btrfs_end_transaction(trans, root);
1178                 else
1179                         ret = btrfs_commit_transaction(trans, root);
1180         }
1181         mutex_lock(&dentry->d_inode->i_mutex);
1182 out:
1183         return ret > 0 ? EIO : ret;
1184 }
1185
1186 static struct vm_operations_struct btrfs_file_vm_ops = {
1187         .fault          = filemap_fault,
1188         .page_mkwrite   = btrfs_page_mkwrite,
1189 };
1190
1191 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
1192 {
1193         vma->vm_ops = &btrfs_file_vm_ops;
1194         file_accessed(filp);
1195         return 0;
1196 }
1197
1198 struct file_operations btrfs_file_operations = {
1199         .llseek         = generic_file_llseek,
1200         .read           = do_sync_read,
1201         .aio_read       = generic_file_aio_read,
1202         .splice_read    = generic_file_splice_read,
1203         .write          = btrfs_file_write,
1204         .mmap           = btrfs_file_mmap,
1205         .open           = generic_file_open,
1206         .release        = btrfs_release_file,
1207         .fsync          = btrfs_sync_file,
1208         .unlocked_ioctl = btrfs_ioctl,
1209 #ifdef CONFIG_COMPAT
1210         .compat_ioctl   = btrfs_ioctl,
1211 #endif
1212 };