Btrfs: Fix page leak in compressed writeback path
[linux-2.6.git] / fs / btrfs / inode.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/kernel.h>
20 #include <linux/bio.h>
21 #include <linux/buffer_head.h>
22 #include <linux/file.h>
23 #include <linux/fs.h>
24 #include <linux/pagemap.h>
25 #include <linux/highmem.h>
26 #include <linux/time.h>
27 #include <linux/init.h>
28 #include <linux/string.h>
29 #include <linux/backing-dev.h>
30 #include <linux/mpage.h>
31 #include <linux/swap.h>
32 #include <linux/writeback.h>
33 #include <linux/statfs.h>
34 #include <linux/compat.h>
35 #include <linux/bit_spinlock.h>
36 #include <linux/xattr.h>
37 #include <linux/posix_acl.h>
38 #include <linux/falloc.h>
39 #include <linux/slab.h>
40 #include "compat.h"
41 #include "ctree.h"
42 #include "disk-io.h"
43 #include "transaction.h"
44 #include "btrfs_inode.h"
45 #include "ioctl.h"
46 #include "print-tree.h"
47 #include "volumes.h"
48 #include "ordered-data.h"
49 #include "xattr.h"
50 #include "tree-log.h"
51 #include "compression.h"
52 #include "locking.h"
53
54 struct btrfs_iget_args {
55         u64 ino;
56         struct btrfs_root *root;
57 };
58
59 static const struct inode_operations btrfs_dir_inode_operations;
60 static const struct inode_operations btrfs_symlink_inode_operations;
61 static const struct inode_operations btrfs_dir_ro_inode_operations;
62 static const struct inode_operations btrfs_special_inode_operations;
63 static const struct inode_operations btrfs_file_inode_operations;
64 static const struct address_space_operations btrfs_aops;
65 static const struct address_space_operations btrfs_symlink_aops;
66 static const struct file_operations btrfs_dir_file_operations;
67 static struct extent_io_ops btrfs_extent_io_ops;
68
69 static struct kmem_cache *btrfs_inode_cachep;
70 struct kmem_cache *btrfs_trans_handle_cachep;
71 struct kmem_cache *btrfs_transaction_cachep;
72 struct kmem_cache *btrfs_path_cachep;
73
74 #define S_SHIFT 12
75 static unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
76         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
77         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
78         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
79         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
80         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
81         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
82         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
83 };
84
85 static void btrfs_truncate(struct inode *inode);
86 static int btrfs_finish_ordered_io(struct inode *inode, u64 start, u64 end);
87 static noinline int cow_file_range(struct inode *inode,
88                                    struct page *locked_page,
89                                    u64 start, u64 end, int *page_started,
90                                    unsigned long *nr_written, int unlock);
91
92 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
93                                      struct inode *inode,  struct inode *dir)
94 {
95         int err;
96
97         err = btrfs_init_acl(trans, inode, dir);
98         if (!err)
99                 err = btrfs_xattr_security_init(trans, inode, dir);
100         return err;
101 }
102
103 /*
104  * this does all the hard work for inserting an inline extent into
105  * the btree.  The caller should have done a btrfs_drop_extents so that
106  * no overlapping inline items exist in the btree
107  */
108 static noinline int insert_inline_extent(struct btrfs_trans_handle *trans,
109                                 struct btrfs_root *root, struct inode *inode,
110                                 u64 start, size_t size, size_t compressed_size,
111                                 struct page **compressed_pages)
112 {
113         struct btrfs_key key;
114         struct btrfs_path *path;
115         struct extent_buffer *leaf;
116         struct page *page = NULL;
117         char *kaddr;
118         unsigned long ptr;
119         struct btrfs_file_extent_item *ei;
120         int err = 0;
121         int ret;
122         size_t cur_size = size;
123         size_t datasize;
124         unsigned long offset;
125         int use_compress = 0;
126
127         if (compressed_size && compressed_pages) {
128                 use_compress = 1;
129                 cur_size = compressed_size;
130         }
131
132         path = btrfs_alloc_path();
133         if (!path)
134                 return -ENOMEM;
135
136         path->leave_spinning = 1;
137         btrfs_set_trans_block_group(trans, inode);
138
139         key.objectid = inode->i_ino;
140         key.offset = start;
141         btrfs_set_key_type(&key, BTRFS_EXTENT_DATA_KEY);
142         datasize = btrfs_file_extent_calc_inline_size(cur_size);
143
144         inode_add_bytes(inode, size);
145         ret = btrfs_insert_empty_item(trans, root, path, &key,
146                                       datasize);
147         BUG_ON(ret);
148         if (ret) {
149                 err = ret;
150                 goto fail;
151         }
152         leaf = path->nodes[0];
153         ei = btrfs_item_ptr(leaf, path->slots[0],
154                             struct btrfs_file_extent_item);
155         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
156         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
157         btrfs_set_file_extent_encryption(leaf, ei, 0);
158         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
159         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
160         ptr = btrfs_file_extent_inline_start(ei);
161
162         if (use_compress) {
163                 struct page *cpage;
164                 int i = 0;
165                 while (compressed_size > 0) {
166                         cpage = compressed_pages[i];
167                         cur_size = min_t(unsigned long, compressed_size,
168                                        PAGE_CACHE_SIZE);
169
170                         kaddr = kmap_atomic(cpage, KM_USER0);
171                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
172                         kunmap_atomic(kaddr, KM_USER0);
173
174                         i++;
175                         ptr += cur_size;
176                         compressed_size -= cur_size;
177                 }
178                 btrfs_set_file_extent_compression(leaf, ei,
179                                                   BTRFS_COMPRESS_ZLIB);
180         } else {
181                 page = find_get_page(inode->i_mapping,
182                                      start >> PAGE_CACHE_SHIFT);
183                 btrfs_set_file_extent_compression(leaf, ei, 0);
184                 kaddr = kmap_atomic(page, KM_USER0);
185                 offset = start & (PAGE_CACHE_SIZE - 1);
186                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
187                 kunmap_atomic(kaddr, KM_USER0);
188                 page_cache_release(page);
189         }
190         btrfs_mark_buffer_dirty(leaf);
191         btrfs_free_path(path);
192
193         /*
194          * we're an inline extent, so nobody can
195          * extend the file past i_size without locking
196          * a page we already have locked.
197          *
198          * We must do any isize and inode updates
199          * before we unlock the pages.  Otherwise we
200          * could end up racing with unlink.
201          */
202         BTRFS_I(inode)->disk_i_size = inode->i_size;
203         btrfs_update_inode(trans, root, inode);
204
205         return 0;
206 fail:
207         btrfs_free_path(path);
208         return err;
209 }
210
211
212 /*
213  * conditionally insert an inline extent into the file.  This
214  * does the checks required to make sure the data is small enough
215  * to fit as an inline extent.
216  */
217 static noinline int cow_file_range_inline(struct btrfs_trans_handle *trans,
218                                  struct btrfs_root *root,
219                                  struct inode *inode, u64 start, u64 end,
220                                  size_t compressed_size,
221                                  struct page **compressed_pages)
222 {
223         u64 isize = i_size_read(inode);
224         u64 actual_end = min(end + 1, isize);
225         u64 inline_len = actual_end - start;
226         u64 aligned_end = (end + root->sectorsize - 1) &
227                         ~((u64)root->sectorsize - 1);
228         u64 hint_byte;
229         u64 data_len = inline_len;
230         int ret;
231
232         if (compressed_size)
233                 data_len = compressed_size;
234
235         if (start > 0 ||
236             actual_end >= PAGE_CACHE_SIZE ||
237             data_len >= BTRFS_MAX_INLINE_DATA_SIZE(root) ||
238             (!compressed_size &&
239             (actual_end & (root->sectorsize - 1)) == 0) ||
240             end + 1 < isize ||
241             data_len > root->fs_info->max_inline) {
242                 return 1;
243         }
244
245         ret = btrfs_drop_extents(trans, inode, start, aligned_end,
246                                  &hint_byte, 1);
247         BUG_ON(ret);
248
249         if (isize > actual_end)
250                 inline_len = min_t(u64, isize, actual_end);
251         ret = insert_inline_extent(trans, root, inode, start,
252                                    inline_len, compressed_size,
253                                    compressed_pages);
254         BUG_ON(ret);
255         btrfs_delalloc_release_metadata(inode, end + 1 - start);
256         btrfs_drop_extent_cache(inode, start, aligned_end - 1, 0);
257         return 0;
258 }
259
260 struct async_extent {
261         u64 start;
262         u64 ram_size;
263         u64 compressed_size;
264         struct page **pages;
265         unsigned long nr_pages;
266         struct list_head list;
267 };
268
269 struct async_cow {
270         struct inode *inode;
271         struct btrfs_root *root;
272         struct page *locked_page;
273         u64 start;
274         u64 end;
275         struct list_head extents;
276         struct btrfs_work work;
277 };
278
279 static noinline int add_async_extent(struct async_cow *cow,
280                                      u64 start, u64 ram_size,
281                                      u64 compressed_size,
282                                      struct page **pages,
283                                      unsigned long nr_pages)
284 {
285         struct async_extent *async_extent;
286
287         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
288         async_extent->start = start;
289         async_extent->ram_size = ram_size;
290         async_extent->compressed_size = compressed_size;
291         async_extent->pages = pages;
292         async_extent->nr_pages = nr_pages;
293         list_add_tail(&async_extent->list, &cow->extents);
294         return 0;
295 }
296
297 /*
298  * we create compressed extents in two phases.  The first
299  * phase compresses a range of pages that have already been
300  * locked (both pages and state bits are locked).
301  *
302  * This is done inside an ordered work queue, and the compression
303  * is spread across many cpus.  The actual IO submission is step
304  * two, and the ordered work queue takes care of making sure that
305  * happens in the same order things were put onto the queue by
306  * writepages and friends.
307  *
308  * If this code finds it can't get good compression, it puts an
309  * entry onto the work queue to write the uncompressed bytes.  This
310  * makes sure that both compressed inodes and uncompressed inodes
311  * are written in the same order that pdflush sent them down.
312  */
313 static noinline int compress_file_range(struct inode *inode,
314                                         struct page *locked_page,
315                                         u64 start, u64 end,
316                                         struct async_cow *async_cow,
317                                         int *num_added)
318 {
319         struct btrfs_root *root = BTRFS_I(inode)->root;
320         struct btrfs_trans_handle *trans;
321         u64 num_bytes;
322         u64 blocksize = root->sectorsize;
323         u64 actual_end;
324         u64 isize = i_size_read(inode);
325         int ret = 0;
326         struct page **pages = NULL;
327         unsigned long nr_pages;
328         unsigned long nr_pages_ret = 0;
329         unsigned long total_compressed = 0;
330         unsigned long total_in = 0;
331         unsigned long max_compressed = 128 * 1024;
332         unsigned long max_uncompressed = 128 * 1024;
333         int i;
334         int will_compress;
335
336         actual_end = min_t(u64, isize, end + 1);
337 again:
338         will_compress = 0;
339         nr_pages = (end >> PAGE_CACHE_SHIFT) - (start >> PAGE_CACHE_SHIFT) + 1;
340         nr_pages = min(nr_pages, (128 * 1024UL) / PAGE_CACHE_SIZE);
341
342         /*
343          * we don't want to send crud past the end of i_size through
344          * compression, that's just a waste of CPU time.  So, if the
345          * end of the file is before the start of our current
346          * requested range of bytes, we bail out to the uncompressed
347          * cleanup code that can deal with all of this.
348          *
349          * It isn't really the fastest way to fix things, but this is a
350          * very uncommon corner.
351          */
352         if (actual_end <= start)
353                 goto cleanup_and_bail_uncompressed;
354
355         total_compressed = actual_end - start;
356
357         /* we want to make sure that amount of ram required to uncompress
358          * an extent is reasonable, so we limit the total size in ram
359          * of a compressed extent to 128k.  This is a crucial number
360          * because it also controls how easily we can spread reads across
361          * cpus for decompression.
362          *
363          * We also want to make sure the amount of IO required to do
364          * a random read is reasonably small, so we limit the size of
365          * a compressed extent to 128k.
366          */
367         total_compressed = min(total_compressed, max_uncompressed);
368         num_bytes = (end - start + blocksize) & ~(blocksize - 1);
369         num_bytes = max(blocksize,  num_bytes);
370         total_in = 0;
371         ret = 0;
372
373         /*
374          * we do compression for mount -o compress and when the
375          * inode has not been flagged as nocompress.  This flag can
376          * change at any time if we discover bad compression ratios.
377          */
378         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS) &&
379             (btrfs_test_opt(root, COMPRESS) ||
380              (BTRFS_I(inode)->force_compress))) {
381                 WARN_ON(pages);
382                 pages = kzalloc(sizeof(struct page *) * nr_pages, GFP_NOFS);
383
384                 ret = btrfs_zlib_compress_pages(inode->i_mapping, start,
385                                                 total_compressed, pages,
386                                                 nr_pages, &nr_pages_ret,
387                                                 &total_in,
388                                                 &total_compressed,
389                                                 max_compressed);
390
391                 if (!ret) {
392                         unsigned long offset = total_compressed &
393                                 (PAGE_CACHE_SIZE - 1);
394                         struct page *page = pages[nr_pages_ret - 1];
395                         char *kaddr;
396
397                         /* zero the tail end of the last page, we might be
398                          * sending it down to disk
399                          */
400                         if (offset) {
401                                 kaddr = kmap_atomic(page, KM_USER0);
402                                 memset(kaddr + offset, 0,
403                                        PAGE_CACHE_SIZE - offset);
404                                 kunmap_atomic(kaddr, KM_USER0);
405                         }
406                         will_compress = 1;
407                 }
408         }
409         if (start == 0) {
410                 trans = btrfs_join_transaction(root, 1);
411                 BUG_ON(!trans);
412                 btrfs_set_trans_block_group(trans, inode);
413                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
414
415                 /* lets try to make an inline extent */
416                 if (ret || total_in < (actual_end - start)) {
417                         /* we didn't compress the entire range, try
418                          * to make an uncompressed inline extent.
419                          */
420                         ret = cow_file_range_inline(trans, root, inode,
421                                                     start, end, 0, NULL);
422                 } else {
423                         /* try making a compressed inline extent */
424                         ret = cow_file_range_inline(trans, root, inode,
425                                                     start, end,
426                                                     total_compressed, pages);
427                 }
428                 if (ret == 0) {
429                         /*
430                          * inline extent creation worked, we don't need
431                          * to create any more async work items.  Unlock
432                          * and free up our temp pages.
433                          */
434                         extent_clear_unlock_delalloc(inode,
435                              &BTRFS_I(inode)->io_tree,
436                              start, end, NULL,
437                              EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
438                              EXTENT_CLEAR_DELALLOC |
439                              EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK);
440
441                         btrfs_end_transaction(trans, root);
442                         goto free_pages_out;
443                 }
444                 btrfs_end_transaction(trans, root);
445         }
446
447         if (will_compress) {
448                 /*
449                  * we aren't doing an inline extent round the compressed size
450                  * up to a block size boundary so the allocator does sane
451                  * things
452                  */
453                 total_compressed = (total_compressed + blocksize - 1) &
454                         ~(blocksize - 1);
455
456                 /*
457                  * one last check to make sure the compression is really a
458                  * win, compare the page count read with the blocks on disk
459                  */
460                 total_in = (total_in + PAGE_CACHE_SIZE - 1) &
461                         ~(PAGE_CACHE_SIZE - 1);
462                 if (total_compressed >= total_in) {
463                         will_compress = 0;
464                 } else {
465                         num_bytes = total_in;
466                 }
467         }
468         if (!will_compress && pages) {
469                 /*
470                  * the compression code ran but failed to make things smaller,
471                  * free any pages it allocated and our page pointer array
472                  */
473                 for (i = 0; i < nr_pages_ret; i++) {
474                         WARN_ON(pages[i]->mapping);
475                         page_cache_release(pages[i]);
476                 }
477                 kfree(pages);
478                 pages = NULL;
479                 total_compressed = 0;
480                 nr_pages_ret = 0;
481
482                 /* flag the file so we don't compress in the future */
483                 if (!btrfs_test_opt(root, FORCE_COMPRESS) &&
484                     !(BTRFS_I(inode)->force_compress)) {
485                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
486                 }
487         }
488         if (will_compress) {
489                 *num_added += 1;
490
491                 /* the async work queues will take care of doing actual
492                  * allocation on disk for these compressed pages,
493                  * and will submit them to the elevator.
494                  */
495                 add_async_extent(async_cow, start, num_bytes,
496                                  total_compressed, pages, nr_pages_ret);
497
498                 if (start + num_bytes < end) {
499                         start += num_bytes;
500                         pages = NULL;
501                         cond_resched();
502                         goto again;
503                 }
504         } else {
505 cleanup_and_bail_uncompressed:
506                 /*
507                  * No compression, but we still need to write the pages in
508                  * the file we've been given so far.  redirty the locked
509                  * page if it corresponds to our extent and set things up
510                  * for the async work queue to run cow_file_range to do
511                  * the normal delalloc dance
512                  */
513                 if (page_offset(locked_page) >= start &&
514                     page_offset(locked_page) <= end) {
515                         __set_page_dirty_nobuffers(locked_page);
516                         /* unlocked later on in the async handlers */
517                 }
518                 add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0);
519                 *num_added += 1;
520         }
521
522 out:
523         return 0;
524
525 free_pages_out:
526         for (i = 0; i < nr_pages_ret; i++) {
527                 WARN_ON(pages[i]->mapping);
528                 page_cache_release(pages[i]);
529         }
530         kfree(pages);
531
532         goto out;
533 }
534
535 /*
536  * phase two of compressed writeback.  This is the ordered portion
537  * of the code, which only gets called in the order the work was
538  * queued.  We walk all the async extents created by compress_file_range
539  * and send them down to the disk.
540  */
541 static noinline int submit_compressed_extents(struct inode *inode,
542                                               struct async_cow *async_cow)
543 {
544         struct async_extent *async_extent;
545         u64 alloc_hint = 0;
546         struct btrfs_trans_handle *trans;
547         struct btrfs_key ins;
548         struct extent_map *em;
549         struct btrfs_root *root = BTRFS_I(inode)->root;
550         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
551         struct extent_io_tree *io_tree;
552         int ret = 0;
553
554         if (list_empty(&async_cow->extents))
555                 return 0;
556
557
558         while (!list_empty(&async_cow->extents)) {
559                 async_extent = list_entry(async_cow->extents.next,
560                                           struct async_extent, list);
561                 list_del(&async_extent->list);
562
563                 io_tree = &BTRFS_I(inode)->io_tree;
564
565 retry:
566                 /* did the compression code fall back to uncompressed IO? */
567                 if (!async_extent->pages) {
568                         int page_started = 0;
569                         unsigned long nr_written = 0;
570
571                         lock_extent(io_tree, async_extent->start,
572                                          async_extent->start +
573                                          async_extent->ram_size - 1, GFP_NOFS);
574
575                         /* allocate blocks */
576                         ret = cow_file_range(inode, async_cow->locked_page,
577                                              async_extent->start,
578                                              async_extent->start +
579                                              async_extent->ram_size - 1,
580                                              &page_started, &nr_written, 0);
581
582                         /*
583                          * if page_started, cow_file_range inserted an
584                          * inline extent and took care of all the unlocking
585                          * and IO for us.  Otherwise, we need to submit
586                          * all those pages down to the drive.
587                          */
588                         if (!page_started && !ret)
589                                 extent_write_locked_range(io_tree,
590                                                   inode, async_extent->start,
591                                                   async_extent->start +
592                                                   async_extent->ram_size - 1,
593                                                   btrfs_get_extent,
594                                                   WB_SYNC_ALL);
595                         kfree(async_extent);
596                         cond_resched();
597                         continue;
598                 }
599
600                 lock_extent(io_tree, async_extent->start,
601                             async_extent->start + async_extent->ram_size - 1,
602                             GFP_NOFS);
603
604                 trans = btrfs_join_transaction(root, 1);
605                 ret = btrfs_reserve_extent(trans, root,
606                                            async_extent->compressed_size,
607                                            async_extent->compressed_size,
608                                            0, alloc_hint,
609                                            (u64)-1, &ins, 1);
610                 btrfs_end_transaction(trans, root);
611
612                 if (ret) {
613                         int i;
614                         for (i = 0; i < async_extent->nr_pages; i++) {
615                                 WARN_ON(async_extent->pages[i]->mapping);
616                                 page_cache_release(async_extent->pages[i]);
617                         }
618                         kfree(async_extent->pages);
619                         async_extent->nr_pages = 0;
620                         async_extent->pages = NULL;
621                         unlock_extent(io_tree, async_extent->start,
622                                       async_extent->start +
623                                       async_extent->ram_size - 1, GFP_NOFS);
624                         goto retry;
625                 }
626
627                 /*
628                  * here we're doing allocation and writeback of the
629                  * compressed pages
630                  */
631                 btrfs_drop_extent_cache(inode, async_extent->start,
632                                         async_extent->start +
633                                         async_extent->ram_size - 1, 0);
634
635                 em = alloc_extent_map(GFP_NOFS);
636                 em->start = async_extent->start;
637                 em->len = async_extent->ram_size;
638                 em->orig_start = em->start;
639
640                 em->block_start = ins.objectid;
641                 em->block_len = ins.offset;
642                 em->bdev = root->fs_info->fs_devices->latest_bdev;
643                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
644                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
645
646                 while (1) {
647                         write_lock(&em_tree->lock);
648                         ret = add_extent_mapping(em_tree, em);
649                         write_unlock(&em_tree->lock);
650                         if (ret != -EEXIST) {
651                                 free_extent_map(em);
652                                 break;
653                         }
654                         btrfs_drop_extent_cache(inode, async_extent->start,
655                                                 async_extent->start +
656                                                 async_extent->ram_size - 1, 0);
657                 }
658
659                 ret = btrfs_add_ordered_extent(inode, async_extent->start,
660                                                ins.objectid,
661                                                async_extent->ram_size,
662                                                ins.offset,
663                                                BTRFS_ORDERED_COMPRESSED);
664                 BUG_ON(ret);
665
666                 /*
667                  * clear dirty, set writeback and unlock the pages.
668                  */
669                 extent_clear_unlock_delalloc(inode,
670                                 &BTRFS_I(inode)->io_tree,
671                                 async_extent->start,
672                                 async_extent->start +
673                                 async_extent->ram_size - 1,
674                                 NULL, EXTENT_CLEAR_UNLOCK_PAGE |
675                                 EXTENT_CLEAR_UNLOCK |
676                                 EXTENT_CLEAR_DELALLOC |
677                                 EXTENT_CLEAR_DIRTY | EXTENT_SET_WRITEBACK);
678
679                 ret = btrfs_submit_compressed_write(inode,
680                                     async_extent->start,
681                                     async_extent->ram_size,
682                                     ins.objectid,
683                                     ins.offset, async_extent->pages,
684                                     async_extent->nr_pages);
685
686                 BUG_ON(ret);
687                 alloc_hint = ins.objectid + ins.offset;
688                 kfree(async_extent);
689                 cond_resched();
690         }
691
692         return 0;
693 }
694
695 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
696                                       u64 num_bytes)
697 {
698         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
699         struct extent_map *em;
700         u64 alloc_hint = 0;
701
702         read_lock(&em_tree->lock);
703         em = search_extent_mapping(em_tree, start, num_bytes);
704         if (em) {
705                 /*
706                  * if block start isn't an actual block number then find the
707                  * first block in this inode and use that as a hint.  If that
708                  * block is also bogus then just don't worry about it.
709                  */
710                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
711                         free_extent_map(em);
712                         em = search_extent_mapping(em_tree, 0, 0);
713                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
714                                 alloc_hint = em->block_start;
715                         if (em)
716                                 free_extent_map(em);
717                 } else {
718                         alloc_hint = em->block_start;
719                         free_extent_map(em);
720                 }
721         }
722         read_unlock(&em_tree->lock);
723
724         return alloc_hint;
725 }
726
727 /*
728  * when extent_io.c finds a delayed allocation range in the file,
729  * the call backs end up in this code.  The basic idea is to
730  * allocate extents on disk for the range, and create ordered data structs
731  * in ram to track those extents.
732  *
733  * locked_page is the page that writepage had locked already.  We use
734  * it to make sure we don't do extra locks or unlocks.
735  *
736  * *page_started is set to one if we unlock locked_page and do everything
737  * required to start IO on it.  It may be clean and already done with
738  * IO when we return.
739  */
740 static noinline int cow_file_range(struct inode *inode,
741                                    struct page *locked_page,
742                                    u64 start, u64 end, int *page_started,
743                                    unsigned long *nr_written,
744                                    int unlock)
745 {
746         struct btrfs_root *root = BTRFS_I(inode)->root;
747         struct btrfs_trans_handle *trans;
748         u64 alloc_hint = 0;
749         u64 num_bytes;
750         unsigned long ram_size;
751         u64 disk_num_bytes;
752         u64 cur_alloc_size;
753         u64 blocksize = root->sectorsize;
754         struct btrfs_key ins;
755         struct extent_map *em;
756         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
757         int ret = 0;
758
759         BUG_ON(root == root->fs_info->tree_root);
760         trans = btrfs_join_transaction(root, 1);
761         BUG_ON(!trans);
762         btrfs_set_trans_block_group(trans, inode);
763         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
764
765         num_bytes = (end - start + blocksize) & ~(blocksize - 1);
766         num_bytes = max(blocksize,  num_bytes);
767         disk_num_bytes = num_bytes;
768         ret = 0;
769
770         if (start == 0) {
771                 /* lets try to make an inline extent */
772                 ret = cow_file_range_inline(trans, root, inode,
773                                             start, end, 0, NULL);
774                 if (ret == 0) {
775                         extent_clear_unlock_delalloc(inode,
776                                      &BTRFS_I(inode)->io_tree,
777                                      start, end, NULL,
778                                      EXTENT_CLEAR_UNLOCK_PAGE |
779                                      EXTENT_CLEAR_UNLOCK |
780                                      EXTENT_CLEAR_DELALLOC |
781                                      EXTENT_CLEAR_DIRTY |
782                                      EXTENT_SET_WRITEBACK |
783                                      EXTENT_END_WRITEBACK);
784
785                         *nr_written = *nr_written +
786                              (end - start + PAGE_CACHE_SIZE) / PAGE_CACHE_SIZE;
787                         *page_started = 1;
788                         ret = 0;
789                         goto out;
790                 }
791         }
792
793         BUG_ON(disk_num_bytes >
794                btrfs_super_total_bytes(&root->fs_info->super_copy));
795
796         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
797         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
798
799         while (disk_num_bytes > 0) {
800                 unsigned long op;
801
802                 cur_alloc_size = disk_num_bytes;
803                 ret = btrfs_reserve_extent(trans, root, cur_alloc_size,
804                                            root->sectorsize, 0, alloc_hint,
805                                            (u64)-1, &ins, 1);
806                 BUG_ON(ret);
807
808                 em = alloc_extent_map(GFP_NOFS);
809                 em->start = start;
810                 em->orig_start = em->start;
811                 ram_size = ins.offset;
812                 em->len = ins.offset;
813
814                 em->block_start = ins.objectid;
815                 em->block_len = ins.offset;
816                 em->bdev = root->fs_info->fs_devices->latest_bdev;
817                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
818
819                 while (1) {
820                         write_lock(&em_tree->lock);
821                         ret = add_extent_mapping(em_tree, em);
822                         write_unlock(&em_tree->lock);
823                         if (ret != -EEXIST) {
824                                 free_extent_map(em);
825                                 break;
826                         }
827                         btrfs_drop_extent_cache(inode, start,
828                                                 start + ram_size - 1, 0);
829                 }
830
831                 cur_alloc_size = ins.offset;
832                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
833                                                ram_size, cur_alloc_size, 0);
834                 BUG_ON(ret);
835
836                 if (root->root_key.objectid ==
837                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
838                         ret = btrfs_reloc_clone_csums(inode, start,
839                                                       cur_alloc_size);
840                         BUG_ON(ret);
841                 }
842
843                 if (disk_num_bytes < cur_alloc_size)
844                         break;
845
846                 /* we're not doing compressed IO, don't unlock the first
847                  * page (which the caller expects to stay locked), don't
848                  * clear any dirty bits and don't set any writeback bits
849                  *
850                  * Do set the Private2 bit so we know this page was properly
851                  * setup for writepage
852                  */
853                 op = unlock ? EXTENT_CLEAR_UNLOCK_PAGE : 0;
854                 op |= EXTENT_CLEAR_UNLOCK | EXTENT_CLEAR_DELALLOC |
855                         EXTENT_SET_PRIVATE2;
856
857                 extent_clear_unlock_delalloc(inode, &BTRFS_I(inode)->io_tree,
858                                              start, start + ram_size - 1,
859                                              locked_page, op);
860                 disk_num_bytes -= cur_alloc_size;
861                 num_bytes -= cur_alloc_size;
862                 alloc_hint = ins.objectid + ins.offset;
863                 start += cur_alloc_size;
864         }
865 out:
866         ret = 0;
867         btrfs_end_transaction(trans, root);
868
869         return ret;
870 }
871
872 /*
873  * work queue call back to started compression on a file and pages
874  */
875 static noinline void async_cow_start(struct btrfs_work *work)
876 {
877         struct async_cow *async_cow;
878         int num_added = 0;
879         async_cow = container_of(work, struct async_cow, work);
880
881         compress_file_range(async_cow->inode, async_cow->locked_page,
882                             async_cow->start, async_cow->end, async_cow,
883                             &num_added);
884         if (num_added == 0)
885                 async_cow->inode = NULL;
886 }
887
888 /*
889  * work queue call back to submit previously compressed pages
890  */
891 static noinline void async_cow_submit(struct btrfs_work *work)
892 {
893         struct async_cow *async_cow;
894         struct btrfs_root *root;
895         unsigned long nr_pages;
896
897         async_cow = container_of(work, struct async_cow, work);
898
899         root = async_cow->root;
900         nr_pages = (async_cow->end - async_cow->start + PAGE_CACHE_SIZE) >>
901                 PAGE_CACHE_SHIFT;
902
903         atomic_sub(nr_pages, &root->fs_info->async_delalloc_pages);
904
905         if (atomic_read(&root->fs_info->async_delalloc_pages) <
906             5 * 1042 * 1024 &&
907             waitqueue_active(&root->fs_info->async_submit_wait))
908                 wake_up(&root->fs_info->async_submit_wait);
909
910         if (async_cow->inode)
911                 submit_compressed_extents(async_cow->inode, async_cow);
912 }
913
914 static noinline void async_cow_free(struct btrfs_work *work)
915 {
916         struct async_cow *async_cow;
917         async_cow = container_of(work, struct async_cow, work);
918         kfree(async_cow);
919 }
920
921 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
922                                 u64 start, u64 end, int *page_started,
923                                 unsigned long *nr_written)
924 {
925         struct async_cow *async_cow;
926         struct btrfs_root *root = BTRFS_I(inode)->root;
927         unsigned long nr_pages;
928         u64 cur_end;
929         int limit = 10 * 1024 * 1042;
930
931         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
932                          1, 0, NULL, GFP_NOFS);
933         while (start < end) {
934                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
935                 async_cow->inode = inode;
936                 async_cow->root = root;
937                 async_cow->locked_page = locked_page;
938                 async_cow->start = start;
939
940                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
941                         cur_end = end;
942                 else
943                         cur_end = min(end, start + 512 * 1024 - 1);
944
945                 async_cow->end = cur_end;
946                 INIT_LIST_HEAD(&async_cow->extents);
947
948                 async_cow->work.func = async_cow_start;
949                 async_cow->work.ordered_func = async_cow_submit;
950                 async_cow->work.ordered_free = async_cow_free;
951                 async_cow->work.flags = 0;
952
953                 nr_pages = (cur_end - start + PAGE_CACHE_SIZE) >>
954                         PAGE_CACHE_SHIFT;
955                 atomic_add(nr_pages, &root->fs_info->async_delalloc_pages);
956
957                 btrfs_queue_worker(&root->fs_info->delalloc_workers,
958                                    &async_cow->work);
959
960                 if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
961                         wait_event(root->fs_info->async_submit_wait,
962                            (atomic_read(&root->fs_info->async_delalloc_pages) <
963                             limit));
964                 }
965
966                 while (atomic_read(&root->fs_info->async_submit_draining) &&
967                       atomic_read(&root->fs_info->async_delalloc_pages)) {
968                         wait_event(root->fs_info->async_submit_wait,
969                           (atomic_read(&root->fs_info->async_delalloc_pages) ==
970                            0));
971                 }
972
973                 *nr_written += nr_pages;
974                 start = cur_end + 1;
975         }
976         *page_started = 1;
977         return 0;
978 }
979
980 static noinline int csum_exist_in_range(struct btrfs_root *root,
981                                         u64 bytenr, u64 num_bytes)
982 {
983         int ret;
984         struct btrfs_ordered_sum *sums;
985         LIST_HEAD(list);
986
987         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, bytenr,
988                                        bytenr + num_bytes - 1, &list);
989         if (ret == 0 && list_empty(&list))
990                 return 0;
991
992         while (!list_empty(&list)) {
993                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
994                 list_del(&sums->list);
995                 kfree(sums);
996         }
997         return 1;
998 }
999
1000 /*
1001  * when nowcow writeback call back.  This checks for snapshots or COW copies
1002  * of the extents that exist in the file, and COWs the file as required.
1003  *
1004  * If no cow copies or snapshots exist, we write directly to the existing
1005  * blocks on disk
1006  */
1007 static noinline int run_delalloc_nocow(struct inode *inode,
1008                                        struct page *locked_page,
1009                               u64 start, u64 end, int *page_started, int force,
1010                               unsigned long *nr_written)
1011 {
1012         struct btrfs_root *root = BTRFS_I(inode)->root;
1013         struct btrfs_trans_handle *trans;
1014         struct extent_buffer *leaf;
1015         struct btrfs_path *path;
1016         struct btrfs_file_extent_item *fi;
1017         struct btrfs_key found_key;
1018         u64 cow_start;
1019         u64 cur_offset;
1020         u64 extent_end;
1021         u64 extent_offset;
1022         u64 disk_bytenr;
1023         u64 num_bytes;
1024         int extent_type;
1025         int ret;
1026         int type;
1027         int nocow;
1028         int check_prev = 1;
1029         bool nolock = false;
1030
1031         path = btrfs_alloc_path();
1032         BUG_ON(!path);
1033         if (root == root->fs_info->tree_root) {
1034                 nolock = true;
1035                 trans = btrfs_join_transaction_nolock(root, 1);
1036         } else {
1037                 trans = btrfs_join_transaction(root, 1);
1038         }
1039         BUG_ON(!trans);
1040
1041         cow_start = (u64)-1;
1042         cur_offset = start;
1043         while (1) {
1044                 ret = btrfs_lookup_file_extent(trans, root, path, inode->i_ino,
1045                                                cur_offset, 0);
1046                 BUG_ON(ret < 0);
1047                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1048                         leaf = path->nodes[0];
1049                         btrfs_item_key_to_cpu(leaf, &found_key,
1050                                               path->slots[0] - 1);
1051                         if (found_key.objectid == inode->i_ino &&
1052                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1053                                 path->slots[0]--;
1054                 }
1055                 check_prev = 0;
1056 next_slot:
1057                 leaf = path->nodes[0];
1058                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1059                         ret = btrfs_next_leaf(root, path);
1060                         if (ret < 0)
1061                                 BUG_ON(1);
1062                         if (ret > 0)
1063                                 break;
1064                         leaf = path->nodes[0];
1065                 }
1066
1067                 nocow = 0;
1068                 disk_bytenr = 0;
1069                 num_bytes = 0;
1070                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1071
1072                 if (found_key.objectid > inode->i_ino ||
1073                     found_key.type > BTRFS_EXTENT_DATA_KEY ||
1074                     found_key.offset > end)
1075                         break;
1076
1077                 if (found_key.offset > cur_offset) {
1078                         extent_end = found_key.offset;
1079                         extent_type = 0;
1080                         goto out_check;
1081                 }
1082
1083                 fi = btrfs_item_ptr(leaf, path->slots[0],
1084                                     struct btrfs_file_extent_item);
1085                 extent_type = btrfs_file_extent_type(leaf, fi);
1086
1087                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1088                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1089                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1090                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1091                         extent_end = found_key.offset +
1092                                 btrfs_file_extent_num_bytes(leaf, fi);
1093                         if (extent_end <= start) {
1094                                 path->slots[0]++;
1095                                 goto next_slot;
1096                         }
1097                         if (disk_bytenr == 0)
1098                                 goto out_check;
1099                         if (btrfs_file_extent_compression(leaf, fi) ||
1100                             btrfs_file_extent_encryption(leaf, fi) ||
1101                             btrfs_file_extent_other_encoding(leaf, fi))
1102                                 goto out_check;
1103                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1104                                 goto out_check;
1105                         if (btrfs_extent_readonly(root, disk_bytenr))
1106                                 goto out_check;
1107                         if (btrfs_cross_ref_exist(trans, root, inode->i_ino,
1108                                                   found_key.offset -
1109                                                   extent_offset, disk_bytenr))
1110                                 goto out_check;
1111                         disk_bytenr += extent_offset;
1112                         disk_bytenr += cur_offset - found_key.offset;
1113                         num_bytes = min(end + 1, extent_end) - cur_offset;
1114                         /*
1115                          * force cow if csum exists in the range.
1116                          * this ensure that csum for a given extent are
1117                          * either valid or do not exist.
1118                          */
1119                         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
1120                                 goto out_check;
1121                         nocow = 1;
1122                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1123                         extent_end = found_key.offset +
1124                                 btrfs_file_extent_inline_len(leaf, fi);
1125                         extent_end = ALIGN(extent_end, root->sectorsize);
1126                 } else {
1127                         BUG_ON(1);
1128                 }
1129 out_check:
1130                 if (extent_end <= start) {
1131                         path->slots[0]++;
1132                         goto next_slot;
1133                 }
1134                 if (!nocow) {
1135                         if (cow_start == (u64)-1)
1136                                 cow_start = cur_offset;
1137                         cur_offset = extent_end;
1138                         if (cur_offset > end)
1139                                 break;
1140                         path->slots[0]++;
1141                         goto next_slot;
1142                 }
1143
1144                 btrfs_release_path(root, path);
1145                 if (cow_start != (u64)-1) {
1146                         ret = cow_file_range(inode, locked_page, cow_start,
1147                                         found_key.offset - 1, page_started,
1148                                         nr_written, 1);
1149                         BUG_ON(ret);
1150                         cow_start = (u64)-1;
1151                 }
1152
1153                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1154                         struct extent_map *em;
1155                         struct extent_map_tree *em_tree;
1156                         em_tree = &BTRFS_I(inode)->extent_tree;
1157                         em = alloc_extent_map(GFP_NOFS);
1158                         em->start = cur_offset;
1159                         em->orig_start = em->start;
1160                         em->len = num_bytes;
1161                         em->block_len = num_bytes;
1162                         em->block_start = disk_bytenr;
1163                         em->bdev = root->fs_info->fs_devices->latest_bdev;
1164                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1165                         while (1) {
1166                                 write_lock(&em_tree->lock);
1167                                 ret = add_extent_mapping(em_tree, em);
1168                                 write_unlock(&em_tree->lock);
1169                                 if (ret != -EEXIST) {
1170                                         free_extent_map(em);
1171                                         break;
1172                                 }
1173                                 btrfs_drop_extent_cache(inode, em->start,
1174                                                 em->start + em->len - 1, 0);
1175                         }
1176                         type = BTRFS_ORDERED_PREALLOC;
1177                 } else {
1178                         type = BTRFS_ORDERED_NOCOW;
1179                 }
1180
1181                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1182                                                num_bytes, num_bytes, type);
1183                 BUG_ON(ret);
1184
1185                 if (root->root_key.objectid ==
1186                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1187                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1188                                                       num_bytes);
1189                         BUG_ON(ret);
1190                 }
1191
1192                 extent_clear_unlock_delalloc(inode, &BTRFS_I(inode)->io_tree,
1193                                 cur_offset, cur_offset + num_bytes - 1,
1194                                 locked_page, EXTENT_CLEAR_UNLOCK_PAGE |
1195                                 EXTENT_CLEAR_UNLOCK | EXTENT_CLEAR_DELALLOC |
1196                                 EXTENT_SET_PRIVATE2);
1197                 cur_offset = extent_end;
1198                 if (cur_offset > end)
1199                         break;
1200         }
1201         btrfs_release_path(root, path);
1202
1203         if (cur_offset <= end && cow_start == (u64)-1)
1204                 cow_start = cur_offset;
1205         if (cow_start != (u64)-1) {
1206                 ret = cow_file_range(inode, locked_page, cow_start, end,
1207                                      page_started, nr_written, 1);
1208                 BUG_ON(ret);
1209         }
1210
1211         if (nolock) {
1212                 ret = btrfs_end_transaction_nolock(trans, root);
1213                 BUG_ON(ret);
1214         } else {
1215                 ret = btrfs_end_transaction(trans, root);
1216                 BUG_ON(ret);
1217         }
1218         btrfs_free_path(path);
1219         return 0;
1220 }
1221
1222 /*
1223  * extent_io.c call back to do delayed allocation processing
1224  */
1225 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1226                               u64 start, u64 end, int *page_started,
1227                               unsigned long *nr_written)
1228 {
1229         int ret;
1230         struct btrfs_root *root = BTRFS_I(inode)->root;
1231
1232         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW)
1233                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1234                                          page_started, 1, nr_written);
1235         else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)
1236                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1237                                          page_started, 0, nr_written);
1238         else if (!btrfs_test_opt(root, COMPRESS) &&
1239                  !(BTRFS_I(inode)->force_compress))
1240                 ret = cow_file_range(inode, locked_page, start, end,
1241                                       page_started, nr_written, 1);
1242         else
1243                 ret = cow_file_range_async(inode, locked_page, start, end,
1244                                            page_started, nr_written);
1245         return ret;
1246 }
1247
1248 static int btrfs_split_extent_hook(struct inode *inode,
1249                                    struct extent_state *orig, u64 split)
1250 {
1251         /* not delalloc, ignore it */
1252         if (!(orig->state & EXTENT_DELALLOC))
1253                 return 0;
1254
1255         atomic_inc(&BTRFS_I(inode)->outstanding_extents);
1256         return 0;
1257 }
1258
1259 /*
1260  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1261  * extents so we can keep track of new extents that are just merged onto old
1262  * extents, such as when we are doing sequential writes, so we can properly
1263  * account for the metadata space we'll need.
1264  */
1265 static int btrfs_merge_extent_hook(struct inode *inode,
1266                                    struct extent_state *new,
1267                                    struct extent_state *other)
1268 {
1269         /* not delalloc, ignore it */
1270         if (!(other->state & EXTENT_DELALLOC))
1271                 return 0;
1272
1273         atomic_dec(&BTRFS_I(inode)->outstanding_extents);
1274         return 0;
1275 }
1276
1277 /*
1278  * extent_io.c set_bit_hook, used to track delayed allocation
1279  * bytes in this file, and to maintain the list of inodes that
1280  * have pending delalloc work to be done.
1281  */
1282 static int btrfs_set_bit_hook(struct inode *inode,
1283                               struct extent_state *state, int *bits)
1284 {
1285
1286         /*
1287          * set_bit and clear bit hooks normally require _irqsave/restore
1288          * but in this case, we are only testeing for the DELALLOC
1289          * bit, which is only set or cleared with irqs on
1290          */
1291         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1292                 struct btrfs_root *root = BTRFS_I(inode)->root;
1293                 u64 len = state->end + 1 - state->start;
1294                 int do_list = (root->root_key.objectid !=
1295                                BTRFS_ROOT_TREE_OBJECTID);
1296
1297                 if (*bits & EXTENT_FIRST_DELALLOC)
1298                         *bits &= ~EXTENT_FIRST_DELALLOC;
1299                 else
1300                         atomic_inc(&BTRFS_I(inode)->outstanding_extents);
1301
1302                 spin_lock(&root->fs_info->delalloc_lock);
1303                 BTRFS_I(inode)->delalloc_bytes += len;
1304                 root->fs_info->delalloc_bytes += len;
1305                 if (do_list && list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1306                         list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1307                                       &root->fs_info->delalloc_inodes);
1308                 }
1309                 spin_unlock(&root->fs_info->delalloc_lock);
1310         }
1311         return 0;
1312 }
1313
1314 /*
1315  * extent_io.c clear_bit_hook, see set_bit_hook for why
1316  */
1317 static int btrfs_clear_bit_hook(struct inode *inode,
1318                                 struct extent_state *state, int *bits)
1319 {
1320         /*
1321          * set_bit and clear bit hooks normally require _irqsave/restore
1322          * but in this case, we are only testeing for the DELALLOC
1323          * bit, which is only set or cleared with irqs on
1324          */
1325         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1326                 struct btrfs_root *root = BTRFS_I(inode)->root;
1327                 u64 len = state->end + 1 - state->start;
1328                 int do_list = (root->root_key.objectid !=
1329                                BTRFS_ROOT_TREE_OBJECTID);
1330
1331                 if (*bits & EXTENT_FIRST_DELALLOC)
1332                         *bits &= ~EXTENT_FIRST_DELALLOC;
1333                 else if (!(*bits & EXTENT_DO_ACCOUNTING))
1334                         atomic_dec(&BTRFS_I(inode)->outstanding_extents);
1335
1336                 if (*bits & EXTENT_DO_ACCOUNTING)
1337                         btrfs_delalloc_release_metadata(inode, len);
1338
1339                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1340                     && do_list)
1341                         btrfs_free_reserved_data_space(inode, len);
1342
1343                 spin_lock(&root->fs_info->delalloc_lock);
1344                 root->fs_info->delalloc_bytes -= len;
1345                 BTRFS_I(inode)->delalloc_bytes -= len;
1346
1347                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1348                     !list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1349                         list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1350                 }
1351                 spin_unlock(&root->fs_info->delalloc_lock);
1352         }
1353         return 0;
1354 }
1355
1356 /*
1357  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1358  * we don't create bios that span stripes or chunks
1359  */
1360 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1361                          size_t size, struct bio *bio,
1362                          unsigned long bio_flags)
1363 {
1364         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
1365         struct btrfs_mapping_tree *map_tree;
1366         u64 logical = (u64)bio->bi_sector << 9;
1367         u64 length = 0;
1368         u64 map_length;
1369         int ret;
1370
1371         if (bio_flags & EXTENT_BIO_COMPRESSED)
1372                 return 0;
1373
1374         length = bio->bi_size;
1375         map_tree = &root->fs_info->mapping_tree;
1376         map_length = length;
1377         ret = btrfs_map_block(map_tree, READ, logical,
1378                               &map_length, NULL, 0);
1379
1380         if (map_length < length + size)
1381                 return 1;
1382         return ret;
1383 }
1384
1385 /*
1386  * in order to insert checksums into the metadata in large chunks,
1387  * we wait until bio submission time.   All the pages in the bio are
1388  * checksummed and sums are attached onto the ordered extent record.
1389  *
1390  * At IO completion time the cums attached on the ordered extent record
1391  * are inserted into the btree
1392  */
1393 static int __btrfs_submit_bio_start(struct inode *inode, int rw,
1394                                     struct bio *bio, int mirror_num,
1395                                     unsigned long bio_flags,
1396                                     u64 bio_offset)
1397 {
1398         struct btrfs_root *root = BTRFS_I(inode)->root;
1399         int ret = 0;
1400
1401         ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1402         BUG_ON(ret);
1403         return 0;
1404 }
1405
1406 /*
1407  * in order to insert checksums into the metadata in large chunks,
1408  * we wait until bio submission time.   All the pages in the bio are
1409  * checksummed and sums are attached onto the ordered extent record.
1410  *
1411  * At IO completion time the cums attached on the ordered extent record
1412  * are inserted into the btree
1413  */
1414 static int __btrfs_submit_bio_done(struct inode *inode, int rw, struct bio *bio,
1415                           int mirror_num, unsigned long bio_flags,
1416                           u64 bio_offset)
1417 {
1418         struct btrfs_root *root = BTRFS_I(inode)->root;
1419         return btrfs_map_bio(root, rw, bio, mirror_num, 1);
1420 }
1421
1422 /*
1423  * extent_io.c submission hook. This does the right thing for csum calculation
1424  * on write, or reading the csums from the tree before a read
1425  */
1426 static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
1427                           int mirror_num, unsigned long bio_flags,
1428                           u64 bio_offset)
1429 {
1430         struct btrfs_root *root = BTRFS_I(inode)->root;
1431         int ret = 0;
1432         int skip_sum;
1433
1434         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1435
1436         if (root == root->fs_info->tree_root)
1437                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, 2);
1438         else
1439                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
1440         BUG_ON(ret);
1441
1442         if (!(rw & REQ_WRITE)) {
1443                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1444                         return btrfs_submit_compressed_read(inode, bio,
1445                                                     mirror_num, bio_flags);
1446                 } else if (!skip_sum)
1447                         btrfs_lookup_bio_sums(root, inode, bio, NULL);
1448                 goto mapit;
1449         } else if (!skip_sum) {
1450                 /* csum items have already been cloned */
1451                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1452                         goto mapit;
1453                 /* we're doing a write, do the async checksumming */
1454                 return btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
1455                                    inode, rw, bio, mirror_num,
1456                                    bio_flags, bio_offset,
1457                                    __btrfs_submit_bio_start,
1458                                    __btrfs_submit_bio_done);
1459         }
1460
1461 mapit:
1462         return btrfs_map_bio(root, rw, bio, mirror_num, 0);
1463 }
1464
1465 /*
1466  * given a list of ordered sums record them in the inode.  This happens
1467  * at IO completion time based on sums calculated at bio submission time.
1468  */
1469 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1470                              struct inode *inode, u64 file_offset,
1471                              struct list_head *list)
1472 {
1473         struct btrfs_ordered_sum *sum;
1474
1475         btrfs_set_trans_block_group(trans, inode);
1476
1477         list_for_each_entry(sum, list, list) {
1478                 btrfs_csum_file_blocks(trans,
1479                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1480         }
1481         return 0;
1482 }
1483
1484 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
1485                               struct extent_state **cached_state)
1486 {
1487         if ((end & (PAGE_CACHE_SIZE - 1)) == 0)
1488                 WARN_ON(1);
1489         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
1490                                    cached_state, GFP_NOFS);
1491 }
1492
1493 /* see btrfs_writepage_start_hook for details on why this is required */
1494 struct btrfs_writepage_fixup {
1495         struct page *page;
1496         struct btrfs_work work;
1497 };
1498
1499 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
1500 {
1501         struct btrfs_writepage_fixup *fixup;
1502         struct btrfs_ordered_extent *ordered;
1503         struct extent_state *cached_state = NULL;
1504         struct page *page;
1505         struct inode *inode;
1506         u64 page_start;
1507         u64 page_end;
1508
1509         fixup = container_of(work, struct btrfs_writepage_fixup, work);
1510         page = fixup->page;
1511 again:
1512         lock_page(page);
1513         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
1514                 ClearPageChecked(page);
1515                 goto out_page;
1516         }
1517
1518         inode = page->mapping->host;
1519         page_start = page_offset(page);
1520         page_end = page_offset(page) + PAGE_CACHE_SIZE - 1;
1521
1522         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end, 0,
1523                          &cached_state, GFP_NOFS);
1524
1525         /* already ordered? We're done */
1526         if (PagePrivate2(page))
1527                 goto out;
1528
1529         ordered = btrfs_lookup_ordered_extent(inode, page_start);
1530         if (ordered) {
1531                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
1532                                      page_end, &cached_state, GFP_NOFS);
1533                 unlock_page(page);
1534                 btrfs_start_ordered_extent(inode, ordered, 1);
1535                 goto again;
1536         }
1537
1538         BUG();
1539         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state);
1540         ClearPageChecked(page);
1541 out:
1542         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
1543                              &cached_state, GFP_NOFS);
1544 out_page:
1545         unlock_page(page);
1546         page_cache_release(page);
1547 }
1548
1549 /*
1550  * There are a few paths in the higher layers of the kernel that directly
1551  * set the page dirty bit without asking the filesystem if it is a
1552  * good idea.  This causes problems because we want to make sure COW
1553  * properly happens and the data=ordered rules are followed.
1554  *
1555  * In our case any range that doesn't have the ORDERED bit set
1556  * hasn't been properly setup for IO.  We kick off an async process
1557  * to fix it up.  The async helper will wait for ordered extents, set
1558  * the delalloc bit and make it safe to write the page.
1559  */
1560 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
1561 {
1562         struct inode *inode = page->mapping->host;
1563         struct btrfs_writepage_fixup *fixup;
1564         struct btrfs_root *root = BTRFS_I(inode)->root;
1565
1566         /* this page is properly in the ordered list */
1567         if (TestClearPagePrivate2(page))
1568                 return 0;
1569
1570         if (PageChecked(page))
1571                 return -EAGAIN;
1572
1573         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
1574         if (!fixup)
1575                 return -EAGAIN;
1576
1577         SetPageChecked(page);
1578         page_cache_get(page);
1579         fixup->work.func = btrfs_writepage_fixup_worker;
1580         fixup->page = page;
1581         btrfs_queue_worker(&root->fs_info->fixup_workers, &fixup->work);
1582         return -EAGAIN;
1583 }
1584
1585 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
1586                                        struct inode *inode, u64 file_pos,
1587                                        u64 disk_bytenr, u64 disk_num_bytes,
1588                                        u64 num_bytes, u64 ram_bytes,
1589                                        u8 compression, u8 encryption,
1590                                        u16 other_encoding, int extent_type)
1591 {
1592         struct btrfs_root *root = BTRFS_I(inode)->root;
1593         struct btrfs_file_extent_item *fi;
1594         struct btrfs_path *path;
1595         struct extent_buffer *leaf;
1596         struct btrfs_key ins;
1597         u64 hint;
1598         int ret;
1599
1600         path = btrfs_alloc_path();
1601         BUG_ON(!path);
1602
1603         path->leave_spinning = 1;
1604
1605         /*
1606          * we may be replacing one extent in the tree with another.
1607          * The new extent is pinned in the extent map, and we don't want
1608          * to drop it from the cache until it is completely in the btree.
1609          *
1610          * So, tell btrfs_drop_extents to leave this extent in the cache.
1611          * the caller is expected to unpin it and allow it to be merged
1612          * with the others.
1613          */
1614         ret = btrfs_drop_extents(trans, inode, file_pos, file_pos + num_bytes,
1615                                  &hint, 0);
1616         BUG_ON(ret);
1617
1618         ins.objectid = inode->i_ino;
1619         ins.offset = file_pos;
1620         ins.type = BTRFS_EXTENT_DATA_KEY;
1621         ret = btrfs_insert_empty_item(trans, root, path, &ins, sizeof(*fi));
1622         BUG_ON(ret);
1623         leaf = path->nodes[0];
1624         fi = btrfs_item_ptr(leaf, path->slots[0],
1625                             struct btrfs_file_extent_item);
1626         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1627         btrfs_set_file_extent_type(leaf, fi, extent_type);
1628         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
1629         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
1630         btrfs_set_file_extent_offset(leaf, fi, 0);
1631         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
1632         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
1633         btrfs_set_file_extent_compression(leaf, fi, compression);
1634         btrfs_set_file_extent_encryption(leaf, fi, encryption);
1635         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
1636
1637         btrfs_unlock_up_safe(path, 1);
1638         btrfs_set_lock_blocking(leaf);
1639
1640         btrfs_mark_buffer_dirty(leaf);
1641
1642         inode_add_bytes(inode, num_bytes);
1643
1644         ins.objectid = disk_bytenr;
1645         ins.offset = disk_num_bytes;
1646         ins.type = BTRFS_EXTENT_ITEM_KEY;
1647         ret = btrfs_alloc_reserved_file_extent(trans, root,
1648                                         root->root_key.objectid,
1649                                         inode->i_ino, file_pos, &ins);
1650         BUG_ON(ret);
1651         btrfs_free_path(path);
1652
1653         return 0;
1654 }
1655
1656 /*
1657  * helper function for btrfs_finish_ordered_io, this
1658  * just reads in some of the csum leaves to prime them into ram
1659  * before we start the transaction.  It limits the amount of btree
1660  * reads required while inside the transaction.
1661  */
1662 /* as ordered data IO finishes, this gets called so we can finish
1663  * an ordered extent if the range of bytes in the file it covers are
1664  * fully written.
1665  */
1666 static int btrfs_finish_ordered_io(struct inode *inode, u64 start, u64 end)
1667 {
1668         struct btrfs_root *root = BTRFS_I(inode)->root;
1669         struct btrfs_trans_handle *trans = NULL;
1670         struct btrfs_ordered_extent *ordered_extent = NULL;
1671         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
1672         struct extent_state *cached_state = NULL;
1673         int compressed = 0;
1674         int ret;
1675         bool nolock = false;
1676
1677         ret = btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
1678                                              end - start + 1);
1679         if (!ret)
1680                 return 0;
1681         BUG_ON(!ordered_extent);
1682
1683         nolock = (root == root->fs_info->tree_root);
1684
1685         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
1686                 BUG_ON(!list_empty(&ordered_extent->list));
1687                 ret = btrfs_ordered_update_i_size(inode, 0, ordered_extent);
1688                 if (!ret) {
1689                         if (nolock)
1690                                 trans = btrfs_join_transaction_nolock(root, 1);
1691                         else
1692                                 trans = btrfs_join_transaction(root, 1);
1693                         BUG_ON(!trans);
1694                         btrfs_set_trans_block_group(trans, inode);
1695                         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1696                         ret = btrfs_update_inode(trans, root, inode);
1697                         BUG_ON(ret);
1698                 }
1699                 goto out;
1700         }
1701
1702         lock_extent_bits(io_tree, ordered_extent->file_offset,
1703                          ordered_extent->file_offset + ordered_extent->len - 1,
1704                          0, &cached_state, GFP_NOFS);
1705
1706         if (nolock)
1707                 trans = btrfs_join_transaction_nolock(root, 1);
1708         else
1709                 trans = btrfs_join_transaction(root, 1);
1710         btrfs_set_trans_block_group(trans, inode);
1711         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1712
1713         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
1714                 compressed = 1;
1715         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
1716                 BUG_ON(compressed);
1717                 ret = btrfs_mark_extent_written(trans, inode,
1718                                                 ordered_extent->file_offset,
1719                                                 ordered_extent->file_offset +
1720                                                 ordered_extent->len);
1721                 BUG_ON(ret);
1722         } else {
1723                 BUG_ON(root == root->fs_info->tree_root);
1724                 ret = insert_reserved_file_extent(trans, inode,
1725                                                 ordered_extent->file_offset,
1726                                                 ordered_extent->start,
1727                                                 ordered_extent->disk_len,
1728                                                 ordered_extent->len,
1729                                                 ordered_extent->len,
1730                                                 compressed, 0, 0,
1731                                                 BTRFS_FILE_EXTENT_REG);
1732                 unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
1733                                    ordered_extent->file_offset,
1734                                    ordered_extent->len);
1735                 BUG_ON(ret);
1736         }
1737         unlock_extent_cached(io_tree, ordered_extent->file_offset,
1738                              ordered_extent->file_offset +
1739                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
1740
1741         add_pending_csums(trans, inode, ordered_extent->file_offset,
1742                           &ordered_extent->list);
1743
1744         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
1745         ret = btrfs_update_inode(trans, root, inode);
1746         BUG_ON(ret);
1747 out:
1748         if (nolock) {
1749                 if (trans)
1750                         btrfs_end_transaction_nolock(trans, root);
1751         } else {
1752                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
1753                 if (trans)
1754                         btrfs_end_transaction(trans, root);
1755         }
1756
1757         /* once for us */
1758         btrfs_put_ordered_extent(ordered_extent);
1759         /* once for the tree */
1760         btrfs_put_ordered_extent(ordered_extent);
1761
1762         return 0;
1763 }
1764
1765 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
1766                                 struct extent_state *state, int uptodate)
1767 {
1768         ClearPagePrivate2(page);
1769         return btrfs_finish_ordered_io(page->mapping->host, start, end);
1770 }
1771
1772 /*
1773  * When IO fails, either with EIO or csum verification fails, we
1774  * try other mirrors that might have a good copy of the data.  This
1775  * io_failure_record is used to record state as we go through all the
1776  * mirrors.  If another mirror has good data, the page is set up to date
1777  * and things continue.  If a good mirror can't be found, the original
1778  * bio end_io callback is called to indicate things have failed.
1779  */
1780 struct io_failure_record {
1781         struct page *page;
1782         u64 start;
1783         u64 len;
1784         u64 logical;
1785         unsigned long bio_flags;
1786         int last_mirror;
1787 };
1788
1789 static int btrfs_io_failed_hook(struct bio *failed_bio,
1790                          struct page *page, u64 start, u64 end,
1791                          struct extent_state *state)
1792 {
1793         struct io_failure_record *failrec = NULL;
1794         u64 private;
1795         struct extent_map *em;
1796         struct inode *inode = page->mapping->host;
1797         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1798         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
1799         struct bio *bio;
1800         int num_copies;
1801         int ret;
1802         int rw;
1803         u64 logical;
1804
1805         ret = get_state_private(failure_tree, start, &private);
1806         if (ret) {
1807                 failrec = kmalloc(sizeof(*failrec), GFP_NOFS);
1808                 if (!failrec)
1809                         return -ENOMEM;
1810                 failrec->start = start;
1811                 failrec->len = end - start + 1;
1812                 failrec->last_mirror = 0;
1813                 failrec->bio_flags = 0;
1814
1815                 read_lock(&em_tree->lock);
1816                 em = lookup_extent_mapping(em_tree, start, failrec->len);
1817                 if (em->start > start || em->start + em->len < start) {
1818                         free_extent_map(em);
1819                         em = NULL;
1820                 }
1821                 read_unlock(&em_tree->lock);
1822
1823                 if (!em || IS_ERR(em)) {
1824                         kfree(failrec);
1825                         return -EIO;
1826                 }
1827                 logical = start - em->start;
1828                 logical = em->block_start + logical;
1829                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
1830                         logical = em->block_start;
1831                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
1832                 }
1833                 failrec->logical = logical;
1834                 free_extent_map(em);
1835                 set_extent_bits(failure_tree, start, end, EXTENT_LOCKED |
1836                                 EXTENT_DIRTY, GFP_NOFS);
1837                 set_state_private(failure_tree, start,
1838                                  (u64)(unsigned long)failrec);
1839         } else {
1840                 failrec = (struct io_failure_record *)(unsigned long)private;
1841         }
1842         num_copies = btrfs_num_copies(
1843                               &BTRFS_I(inode)->root->fs_info->mapping_tree,
1844                               failrec->logical, failrec->len);
1845         failrec->last_mirror++;
1846         if (!state) {
1847                 spin_lock(&BTRFS_I(inode)->io_tree.lock);
1848                 state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
1849                                                     failrec->start,
1850                                                     EXTENT_LOCKED);
1851                 if (state && state->start != failrec->start)
1852                         state = NULL;
1853                 spin_unlock(&BTRFS_I(inode)->io_tree.lock);
1854         }
1855         if (!state || failrec->last_mirror > num_copies) {
1856                 set_state_private(failure_tree, failrec->start, 0);
1857                 clear_extent_bits(failure_tree, failrec->start,
1858                                   failrec->start + failrec->len - 1,
1859                                   EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1860                 kfree(failrec);
1861                 return -EIO;
1862         }
1863         bio = bio_alloc(GFP_NOFS, 1);
1864         bio->bi_private = state;
1865         bio->bi_end_io = failed_bio->bi_end_io;
1866         bio->bi_sector = failrec->logical >> 9;
1867         bio->bi_bdev = failed_bio->bi_bdev;
1868         bio->bi_size = 0;
1869
1870         bio_add_page(bio, page, failrec->len, start - page_offset(page));
1871         if (failed_bio->bi_rw & REQ_WRITE)
1872                 rw = WRITE;
1873         else
1874                 rw = READ;
1875
1876         BTRFS_I(inode)->io_tree.ops->submit_bio_hook(inode, rw, bio,
1877                                                       failrec->last_mirror,
1878                                                       failrec->bio_flags, 0);
1879         return 0;
1880 }
1881
1882 /*
1883  * each time an IO finishes, we do a fast check in the IO failure tree
1884  * to see if we need to process or clean up an io_failure_record
1885  */
1886 static int btrfs_clean_io_failures(struct inode *inode, u64 start)
1887 {
1888         u64 private;
1889         u64 private_failure;
1890         struct io_failure_record *failure;
1891         int ret;
1892
1893         private = 0;
1894         if (count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
1895                              (u64)-1, 1, EXTENT_DIRTY)) {
1896                 ret = get_state_private(&BTRFS_I(inode)->io_failure_tree,
1897                                         start, &private_failure);
1898                 if (ret == 0) {
1899                         failure = (struct io_failure_record *)(unsigned long)
1900                                    private_failure;
1901                         set_state_private(&BTRFS_I(inode)->io_failure_tree,
1902                                           failure->start, 0);
1903                         clear_extent_bits(&BTRFS_I(inode)->io_failure_tree,
1904                                           failure->start,
1905                                           failure->start + failure->len - 1,
1906                                           EXTENT_DIRTY | EXTENT_LOCKED,
1907                                           GFP_NOFS);
1908                         kfree(failure);
1909                 }
1910         }
1911         return 0;
1912 }
1913
1914 /*
1915  * when reads are done, we need to check csums to verify the data is correct
1916  * if there's a match, we allow the bio to finish.  If not, we go through
1917  * the io_failure_record routines to find good copies
1918  */
1919 static int btrfs_readpage_end_io_hook(struct page *page, u64 start, u64 end,
1920                                struct extent_state *state)
1921 {
1922         size_t offset = start - ((u64)page->index << PAGE_CACHE_SHIFT);
1923         struct inode *inode = page->mapping->host;
1924         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
1925         char *kaddr;
1926         u64 private = ~(u32)0;
1927         int ret;
1928         struct btrfs_root *root = BTRFS_I(inode)->root;
1929         u32 csum = ~(u32)0;
1930
1931         if (PageChecked(page)) {
1932                 ClearPageChecked(page);
1933                 goto good;
1934         }
1935
1936         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
1937                 return 0;
1938
1939         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
1940             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
1941                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM,
1942                                   GFP_NOFS);
1943                 return 0;
1944         }
1945
1946         if (state && state->start == start) {
1947                 private = state->private;
1948                 ret = 0;
1949         } else {
1950                 ret = get_state_private(io_tree, start, &private);
1951         }
1952         kaddr = kmap_atomic(page, KM_USER0);
1953         if (ret)
1954                 goto zeroit;
1955
1956         csum = btrfs_csum_data(root, kaddr + offset, csum,  end - start + 1);
1957         btrfs_csum_final(csum, (char *)&csum);
1958         if (csum != private)
1959                 goto zeroit;
1960
1961         kunmap_atomic(kaddr, KM_USER0);
1962 good:
1963         /* if the io failure tree for this inode is non-empty,
1964          * check to see if we've recovered from a failed IO
1965          */
1966         btrfs_clean_io_failures(inode, start);
1967         return 0;
1968
1969 zeroit:
1970         if (printk_ratelimit()) {
1971                 printk(KERN_INFO "btrfs csum failed ino %lu off %llu csum %u "
1972                        "private %llu\n", page->mapping->host->i_ino,
1973                        (unsigned long long)start, csum,
1974                        (unsigned long long)private);
1975         }
1976         memset(kaddr + offset, 1, end - start + 1);
1977         flush_dcache_page(page);
1978         kunmap_atomic(kaddr, KM_USER0);
1979         if (private == 0)
1980                 return 0;
1981         return -EIO;
1982 }
1983
1984 struct delayed_iput {
1985         struct list_head list;
1986         struct inode *inode;
1987 };
1988
1989 void btrfs_add_delayed_iput(struct inode *inode)
1990 {
1991         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
1992         struct delayed_iput *delayed;
1993
1994         if (atomic_add_unless(&inode->i_count, -1, 1))
1995                 return;
1996
1997         delayed = kmalloc(sizeof(*delayed), GFP_NOFS | __GFP_NOFAIL);
1998         delayed->inode = inode;
1999
2000         spin_lock(&fs_info->delayed_iput_lock);
2001         list_add_tail(&delayed->list, &fs_info->delayed_iputs);
2002         spin_unlock(&fs_info->delayed_iput_lock);
2003 }
2004
2005 void btrfs_run_delayed_iputs(struct btrfs_root *root)
2006 {
2007         LIST_HEAD(list);
2008         struct btrfs_fs_info *fs_info = root->fs_info;
2009         struct delayed_iput *delayed;
2010         int empty;
2011
2012         spin_lock(&fs_info->delayed_iput_lock);
2013         empty = list_empty(&fs_info->delayed_iputs);
2014         spin_unlock(&fs_info->delayed_iput_lock);
2015         if (empty)
2016                 return;
2017
2018         down_read(&root->fs_info->cleanup_work_sem);
2019         spin_lock(&fs_info->delayed_iput_lock);
2020         list_splice_init(&fs_info->delayed_iputs, &list);
2021         spin_unlock(&fs_info->delayed_iput_lock);
2022
2023         while (!list_empty(&list)) {
2024                 delayed = list_entry(list.next, struct delayed_iput, list);
2025                 list_del(&delayed->list);
2026                 iput(delayed->inode);
2027                 kfree(delayed);
2028         }
2029         up_read(&root->fs_info->cleanup_work_sem);
2030 }
2031
2032 /*
2033  * calculate extra metadata reservation when snapshotting a subvolume
2034  * contains orphan files.
2035  */
2036 void btrfs_orphan_pre_snapshot(struct btrfs_trans_handle *trans,
2037                                 struct btrfs_pending_snapshot *pending,
2038                                 u64 *bytes_to_reserve)
2039 {
2040         struct btrfs_root *root;
2041         struct btrfs_block_rsv *block_rsv;
2042         u64 num_bytes;
2043         int index;
2044
2045         root = pending->root;
2046         if (!root->orphan_block_rsv || list_empty(&root->orphan_list))
2047                 return;
2048
2049         block_rsv = root->orphan_block_rsv;
2050
2051         /* orphan block reservation for the snapshot */
2052         num_bytes = block_rsv->size;
2053
2054         /*
2055          * after the snapshot is created, COWing tree blocks may use more
2056          * space than it frees. So we should make sure there is enough
2057          * reserved space.
2058          */
2059         index = trans->transid & 0x1;
2060         if (block_rsv->reserved + block_rsv->freed[index] < block_rsv->size) {
2061                 num_bytes += block_rsv->size -
2062                              (block_rsv->reserved + block_rsv->freed[index]);
2063         }
2064
2065         *bytes_to_reserve += num_bytes;
2066 }
2067
2068 void btrfs_orphan_post_snapshot(struct btrfs_trans_handle *trans,
2069                                 struct btrfs_pending_snapshot *pending)
2070 {
2071         struct btrfs_root *root = pending->root;
2072         struct btrfs_root *snap = pending->snap;
2073         struct btrfs_block_rsv *block_rsv;
2074         u64 num_bytes;
2075         int index;
2076         int ret;
2077
2078         if (!root->orphan_block_rsv || list_empty(&root->orphan_list))
2079                 return;
2080
2081         /* refill source subvolume's orphan block reservation */
2082         block_rsv = root->orphan_block_rsv;
2083         index = trans->transid & 0x1;
2084         if (block_rsv->reserved + block_rsv->freed[index] < block_rsv->size) {
2085                 num_bytes = block_rsv->size -
2086                             (block_rsv->reserved + block_rsv->freed[index]);
2087                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
2088                                               root->orphan_block_rsv,
2089                                               num_bytes);
2090                 BUG_ON(ret);
2091         }
2092
2093         /* setup orphan block reservation for the snapshot */
2094         block_rsv = btrfs_alloc_block_rsv(snap);
2095         BUG_ON(!block_rsv);
2096
2097         btrfs_add_durable_block_rsv(root->fs_info, block_rsv);
2098         snap->orphan_block_rsv = block_rsv;
2099
2100         num_bytes = root->orphan_block_rsv->size;
2101         ret = btrfs_block_rsv_migrate(&pending->block_rsv,
2102                                       block_rsv, num_bytes);
2103         BUG_ON(ret);
2104
2105 #if 0
2106         /* insert orphan item for the snapshot */
2107         WARN_ON(!root->orphan_item_inserted);
2108         ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
2109                                        snap->root_key.objectid);
2110         BUG_ON(ret);
2111         snap->orphan_item_inserted = 1;
2112 #endif
2113 }
2114
2115 enum btrfs_orphan_cleanup_state {
2116         ORPHAN_CLEANUP_STARTED  = 1,
2117         ORPHAN_CLEANUP_DONE     = 2,
2118 };
2119
2120 /*
2121  * This is called in transaction commmit time. If there are no orphan
2122  * files in the subvolume, it removes orphan item and frees block_rsv
2123  * structure.
2124  */
2125 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
2126                               struct btrfs_root *root)
2127 {
2128         int ret;
2129
2130         if (!list_empty(&root->orphan_list) ||
2131             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
2132                 return;
2133
2134         if (root->orphan_item_inserted &&
2135             btrfs_root_refs(&root->root_item) > 0) {
2136                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
2137                                             root->root_key.objectid);
2138                 BUG_ON(ret);
2139                 root->orphan_item_inserted = 0;
2140         }
2141
2142         if (root->orphan_block_rsv) {
2143                 WARN_ON(root->orphan_block_rsv->size > 0);
2144                 btrfs_free_block_rsv(root, root->orphan_block_rsv);
2145                 root->orphan_block_rsv = NULL;
2146         }
2147 }
2148
2149 /*
2150  * This creates an orphan entry for the given inode in case something goes
2151  * wrong in the middle of an unlink/truncate.
2152  *
2153  * NOTE: caller of this function should reserve 5 units of metadata for
2154  *       this function.
2155  */
2156 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
2157 {
2158         struct btrfs_root *root = BTRFS_I(inode)->root;
2159         struct btrfs_block_rsv *block_rsv = NULL;
2160         int reserve = 0;
2161         int insert = 0;
2162         int ret;
2163
2164         if (!root->orphan_block_rsv) {
2165                 block_rsv = btrfs_alloc_block_rsv(root);
2166                 BUG_ON(!block_rsv);
2167         }
2168
2169         spin_lock(&root->orphan_lock);
2170         if (!root->orphan_block_rsv) {
2171                 root->orphan_block_rsv = block_rsv;
2172         } else if (block_rsv) {
2173                 btrfs_free_block_rsv(root, block_rsv);
2174                 block_rsv = NULL;
2175         }
2176
2177         if (list_empty(&BTRFS_I(inode)->i_orphan)) {
2178                 list_add(&BTRFS_I(inode)->i_orphan, &root->orphan_list);
2179 #if 0
2180                 /*
2181                  * For proper ENOSPC handling, we should do orphan
2182                  * cleanup when mounting. But this introduces backward
2183                  * compatibility issue.
2184                  */
2185                 if (!xchg(&root->orphan_item_inserted, 1))
2186                         insert = 2;
2187                 else
2188                         insert = 1;
2189 #endif
2190                 insert = 1;
2191         } else {
2192                 WARN_ON(!BTRFS_I(inode)->orphan_meta_reserved);
2193         }
2194
2195         if (!BTRFS_I(inode)->orphan_meta_reserved) {
2196                 BTRFS_I(inode)->orphan_meta_reserved = 1;
2197                 reserve = 1;
2198         }
2199         spin_unlock(&root->orphan_lock);
2200
2201         if (block_rsv)
2202                 btrfs_add_durable_block_rsv(root->fs_info, block_rsv);
2203
2204         /* grab metadata reservation from transaction handle */
2205         if (reserve) {
2206                 ret = btrfs_orphan_reserve_metadata(trans, inode);
2207                 BUG_ON(ret);
2208         }
2209
2210         /* insert an orphan item to track this unlinked/truncated file */
2211         if (insert >= 1) {
2212                 ret = btrfs_insert_orphan_item(trans, root, inode->i_ino);
2213                 BUG_ON(ret);
2214         }
2215
2216         /* insert an orphan item to track subvolume contains orphan files */
2217         if (insert >= 2) {
2218                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
2219                                                root->root_key.objectid);
2220                 BUG_ON(ret);
2221         }
2222         return 0;
2223 }
2224
2225 /*
2226  * We have done the truncate/delete so we can go ahead and remove the orphan
2227  * item for this particular inode.
2228  */
2229 int btrfs_orphan_del(struct btrfs_trans_handle *trans, struct inode *inode)
2230 {
2231         struct btrfs_root *root = BTRFS_I(inode)->root;
2232         int delete_item = 0;
2233         int release_rsv = 0;
2234         int ret = 0;
2235
2236         spin_lock(&root->orphan_lock);
2237         if (!list_empty(&BTRFS_I(inode)->i_orphan)) {
2238                 list_del_init(&BTRFS_I(inode)->i_orphan);
2239                 delete_item = 1;
2240         }
2241
2242         if (BTRFS_I(inode)->orphan_meta_reserved) {
2243                 BTRFS_I(inode)->orphan_meta_reserved = 0;
2244                 release_rsv = 1;
2245         }
2246         spin_unlock(&root->orphan_lock);
2247
2248         if (trans && delete_item) {
2249                 ret = btrfs_del_orphan_item(trans, root, inode->i_ino);
2250                 BUG_ON(ret);
2251         }
2252
2253         if (release_rsv)
2254                 btrfs_orphan_release_metadata(inode);
2255
2256         return 0;
2257 }
2258
2259 /*
2260  * this cleans up any orphans that may be left on the list from the last use
2261  * of this root.
2262  */
2263 void btrfs_orphan_cleanup(struct btrfs_root *root)
2264 {
2265         struct btrfs_path *path;
2266         struct extent_buffer *leaf;
2267         struct btrfs_key key, found_key;
2268         struct btrfs_trans_handle *trans;
2269         struct inode *inode;
2270         int ret = 0, nr_unlink = 0, nr_truncate = 0;
2271
2272         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
2273                 return;
2274
2275         path = btrfs_alloc_path();
2276         BUG_ON(!path);
2277         path->reada = -1;
2278
2279         key.objectid = BTRFS_ORPHAN_OBJECTID;
2280         btrfs_set_key_type(&key, BTRFS_ORPHAN_ITEM_KEY);
2281         key.offset = (u64)-1;
2282
2283         while (1) {
2284                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2285                 if (ret < 0) {
2286                         printk(KERN_ERR "Error searching slot for orphan: %d"
2287                                "\n", ret);
2288                         break;
2289                 }
2290
2291                 /*
2292                  * if ret == 0 means we found what we were searching for, which
2293                  * is weird, but possible, so only screw with path if we didnt
2294                  * find the key and see if we have stuff that matches
2295                  */
2296                 if (ret > 0) {
2297                         if (path->slots[0] == 0)
2298                                 break;
2299                         path->slots[0]--;
2300                 }
2301
2302                 /* pull out the item */
2303                 leaf = path->nodes[0];
2304                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2305
2306                 /* make sure the item matches what we want */
2307                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
2308                         break;
2309                 if (btrfs_key_type(&found_key) != BTRFS_ORPHAN_ITEM_KEY)
2310                         break;
2311
2312                 /* release the path since we're done with it */
2313                 btrfs_release_path(root, path);
2314
2315                 /*
2316                  * this is where we are basically btrfs_lookup, without the
2317                  * crossing root thing.  we store the inode number in the
2318                  * offset of the orphan item.
2319                  */
2320                 found_key.objectid = found_key.offset;
2321                 found_key.type = BTRFS_INODE_ITEM_KEY;
2322                 found_key.offset = 0;
2323                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
2324                 BUG_ON(IS_ERR(inode));
2325
2326                 /*
2327                  * add this inode to the orphan list so btrfs_orphan_del does
2328                  * the proper thing when we hit it
2329                  */
2330                 spin_lock(&root->orphan_lock);
2331                 list_add(&BTRFS_I(inode)->i_orphan, &root->orphan_list);
2332                 spin_unlock(&root->orphan_lock);
2333
2334                 /*
2335                  * if this is a bad inode, means we actually succeeded in
2336                  * removing the inode, but not the orphan record, which means
2337                  * we need to manually delete the orphan since iput will just
2338                  * do a destroy_inode
2339                  */
2340                 if (is_bad_inode(inode)) {
2341                         trans = btrfs_start_transaction(root, 0);
2342                         btrfs_orphan_del(trans, inode);
2343                         btrfs_end_transaction(trans, root);
2344                         iput(inode);
2345                         continue;
2346                 }
2347
2348                 /* if we have links, this was a truncate, lets do that */
2349                 if (inode->i_nlink) {
2350                         nr_truncate++;
2351                         btrfs_truncate(inode);
2352                 } else {
2353                         nr_unlink++;
2354                 }
2355
2356                 /* this will do delete_inode and everything for us */
2357                 iput(inode);
2358         }
2359         btrfs_free_path(path);
2360
2361         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
2362
2363         if (root->orphan_block_rsv)
2364                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
2365                                         (u64)-1);
2366
2367         if (root->orphan_block_rsv || root->orphan_item_inserted) {
2368                 trans = btrfs_join_transaction(root, 1);
2369                 btrfs_end_transaction(trans, root);
2370         }
2371
2372         if (nr_unlink)
2373                 printk(KERN_INFO "btrfs: unlinked %d orphans\n", nr_unlink);
2374         if (nr_truncate)
2375                 printk(KERN_INFO "btrfs: truncated %d orphans\n", nr_truncate);
2376 }
2377
2378 /*
2379  * very simple check to peek ahead in the leaf looking for xattrs.  If we
2380  * don't find any xattrs, we know there can't be any acls.
2381  *
2382  * slot is the slot the inode is in, objectid is the objectid of the inode
2383  */
2384 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
2385                                           int slot, u64 objectid)
2386 {
2387         u32 nritems = btrfs_header_nritems(leaf);
2388         struct btrfs_key found_key;
2389         int scanned = 0;
2390
2391         slot++;
2392         while (slot < nritems) {
2393                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
2394
2395                 /* we found a different objectid, there must not be acls */
2396                 if (found_key.objectid != objectid)
2397                         return 0;
2398
2399                 /* we found an xattr, assume we've got an acl */
2400                 if (found_key.type == BTRFS_XATTR_ITEM_KEY)
2401                         return 1;
2402
2403                 /*
2404                  * we found a key greater than an xattr key, there can't
2405                  * be any acls later on
2406                  */
2407                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
2408                         return 0;
2409
2410                 slot++;
2411                 scanned++;
2412
2413                 /*
2414                  * it goes inode, inode backrefs, xattrs, extents,
2415                  * so if there are a ton of hard links to an inode there can
2416                  * be a lot of backrefs.  Don't waste time searching too hard,
2417                  * this is just an optimization
2418                  */
2419                 if (scanned >= 8)
2420                         break;
2421         }
2422         /* we hit the end of the leaf before we found an xattr or
2423          * something larger than an xattr.  We have to assume the inode
2424          * has acls
2425          */
2426         return 1;
2427 }
2428
2429 /*
2430  * read an inode from the btree into the in-memory inode
2431  */
2432 static void btrfs_read_locked_inode(struct inode *inode)
2433 {
2434         struct btrfs_path *path;
2435         struct extent_buffer *leaf;
2436         struct btrfs_inode_item *inode_item;
2437         struct btrfs_timespec *tspec;
2438         struct btrfs_root *root = BTRFS_I(inode)->root;
2439         struct btrfs_key location;
2440         int maybe_acls;
2441         u64 alloc_group_block;
2442         u32 rdev;
2443         int ret;
2444
2445         path = btrfs_alloc_path();
2446         BUG_ON(!path);
2447         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
2448
2449         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
2450         if (ret)
2451                 goto make_bad;
2452
2453         leaf = path->nodes[0];
2454         inode_item = btrfs_item_ptr(leaf, path->slots[0],
2455                                     struct btrfs_inode_item);
2456
2457         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
2458         inode->i_nlink = btrfs_inode_nlink(leaf, inode_item);
2459         inode->i_uid = btrfs_inode_uid(leaf, inode_item);
2460         inode->i_gid = btrfs_inode_gid(leaf, inode_item);
2461         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
2462
2463         tspec = btrfs_inode_atime(inode_item);
2464         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, tspec);
2465         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
2466
2467         tspec = btrfs_inode_mtime(inode_item);
2468         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, tspec);
2469         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
2470
2471         tspec = btrfs_inode_ctime(inode_item);
2472         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, tspec);
2473         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
2474
2475         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
2476         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
2477         BTRFS_I(inode)->sequence = btrfs_inode_sequence(leaf, inode_item);
2478         inode->i_generation = BTRFS_I(inode)->generation;
2479         inode->i_rdev = 0;
2480         rdev = btrfs_inode_rdev(leaf, inode_item);
2481
2482         BTRFS_I(inode)->index_cnt = (u64)-1;
2483         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
2484
2485         alloc_group_block = btrfs_inode_block_group(leaf, inode_item);
2486
2487         /*
2488          * try to precache a NULL acl entry for files that don't have
2489          * any xattrs or acls
2490          */
2491         maybe_acls = acls_after_inode_item(leaf, path->slots[0], inode->i_ino);
2492         if (!maybe_acls)
2493                 cache_no_acl(inode);
2494
2495         BTRFS_I(inode)->block_group = btrfs_find_block_group(root, 0,
2496                                                 alloc_group_block, 0);
2497         btrfs_free_path(path);
2498         inode_item = NULL;
2499
2500         switch (inode->i_mode & S_IFMT) {
2501         case S_IFREG:
2502                 inode->i_mapping->a_ops = &btrfs_aops;
2503                 inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
2504                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
2505                 inode->i_fop = &btrfs_file_operations;
2506                 inode->i_op = &btrfs_file_inode_operations;
2507                 break;
2508         case S_IFDIR:
2509                 inode->i_fop = &btrfs_dir_file_operations;
2510                 if (root == root->fs_info->tree_root)
2511                         inode->i_op = &btrfs_dir_ro_inode_operations;
2512                 else
2513                         inode->i_op = &btrfs_dir_inode_operations;
2514                 break;
2515         case S_IFLNK:
2516                 inode->i_op = &btrfs_symlink_inode_operations;
2517                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
2518                 inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
2519                 break;
2520         default:
2521                 inode->i_op = &btrfs_special_inode_operations;
2522                 init_special_inode(inode, inode->i_mode, rdev);
2523                 break;
2524         }
2525
2526         btrfs_update_iflags(inode);
2527         return;
2528
2529 make_bad:
2530         btrfs_free_path(path);
2531         make_bad_inode(inode);
2532 }
2533
2534 /*
2535  * given a leaf and an inode, copy the inode fields into the leaf
2536  */
2537 static void fill_inode_item(struct btrfs_trans_handle *trans,
2538                             struct extent_buffer *leaf,
2539                             struct btrfs_inode_item *item,
2540                             struct inode *inode)
2541 {
2542         btrfs_set_inode_uid(leaf, item, inode->i_uid);
2543         btrfs_set_inode_gid(leaf, item, inode->i_gid);
2544         btrfs_set_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size);
2545         btrfs_set_inode_mode(leaf, item, inode->i_mode);
2546         btrfs_set_inode_nlink(leaf, item, inode->i_nlink);
2547
2548         btrfs_set_timespec_sec(leaf, btrfs_inode_atime(item),
2549                                inode->i_atime.tv_sec);
2550         btrfs_set_timespec_nsec(leaf, btrfs_inode_atime(item),
2551                                 inode->i_atime.tv_nsec);
2552
2553         btrfs_set_timespec_sec(leaf, btrfs_inode_mtime(item),
2554                                inode->i_mtime.tv_sec);
2555         btrfs_set_timespec_nsec(leaf, btrfs_inode_mtime(item),
2556                                 inode->i_mtime.tv_nsec);
2557
2558         btrfs_set_timespec_sec(leaf, btrfs_inode_ctime(item),
2559                                inode->i_ctime.tv_sec);
2560         btrfs_set_timespec_nsec(leaf, btrfs_inode_ctime(item),
2561                                 inode->i_ctime.tv_nsec);
2562
2563         btrfs_set_inode_nbytes(leaf, item, inode_get_bytes(inode));
2564         btrfs_set_inode_generation(leaf, item, BTRFS_I(inode)->generation);
2565         btrfs_set_inode_sequence(leaf, item, BTRFS_I(inode)->sequence);
2566         btrfs_set_inode_transid(leaf, item, trans->transid);
2567         btrfs_set_inode_rdev(leaf, item, inode->i_rdev);
2568         btrfs_set_inode_flags(leaf, item, BTRFS_I(inode)->flags);
2569         btrfs_set_inode_block_group(leaf, item, BTRFS_I(inode)->block_group);
2570 }
2571
2572 /*
2573  * copy everything in the in-memory inode into the btree.
2574  */
2575 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
2576                                 struct btrfs_root *root, struct inode *inode)
2577 {
2578         struct btrfs_inode_item *inode_item;
2579         struct btrfs_path *path;
2580         struct extent_buffer *leaf;
2581         int ret;
2582
2583         path = btrfs_alloc_path();
2584         BUG_ON(!path);
2585         path->leave_spinning = 1;
2586         ret = btrfs_lookup_inode(trans, root, path,
2587                                  &BTRFS_I(inode)->location, 1);
2588         if (ret) {
2589                 if (ret > 0)
2590                         ret = -ENOENT;
2591                 goto failed;
2592         }
2593
2594         btrfs_unlock_up_safe(path, 1);
2595         leaf = path->nodes[0];
2596         inode_item = btrfs_item_ptr(leaf, path->slots[0],
2597                                   struct btrfs_inode_item);
2598
2599         fill_inode_item(trans, leaf, inode_item, inode);
2600         btrfs_mark_buffer_dirty(leaf);
2601         btrfs_set_inode_last_trans(trans, inode);
2602         ret = 0;
2603 failed:
2604         btrfs_free_path(path);
2605         return ret;
2606 }
2607
2608
2609 /*
2610  * unlink helper that gets used here in inode.c and in the tree logging
2611  * recovery code.  It remove a link in a directory with a given name, and
2612  * also drops the back refs in the inode to the directory
2613  */
2614 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
2615                        struct btrfs_root *root,
2616                        struct inode *dir, struct inode *inode,
2617                        const char *name, int name_len)
2618 {
2619         struct btrfs_path *path;
2620         int ret = 0;
2621         struct extent_buffer *leaf;
2622         struct btrfs_dir_item *di;
2623         struct btrfs_key key;
2624         u64 index;
2625
2626         path = btrfs_alloc_path();
2627         if (!path) {
2628                 ret = -ENOMEM;
2629                 goto err;
2630         }
2631
2632         path->leave_spinning = 1;
2633         di = btrfs_lookup_dir_item(trans, root, path, dir->i_ino,
2634                                     name, name_len, -1);
2635         if (IS_ERR(di)) {
2636                 ret = PTR_ERR(di);
2637                 goto err;
2638         }
2639         if (!di) {
2640                 ret = -ENOENT;
2641                 goto err;
2642         }
2643         leaf = path->nodes[0];
2644         btrfs_dir_item_key_to_cpu(leaf, di, &key);
2645         ret = btrfs_delete_one_dir_name(trans, root, path, di);
2646         if (ret)
2647                 goto err;
2648         btrfs_release_path(root, path);
2649
2650         ret = btrfs_del_inode_ref(trans, root, name, name_len,
2651                                   inode->i_ino,
2652                                   dir->i_ino, &index);
2653         if (ret) {
2654                 printk(KERN_INFO "btrfs failed to delete reference to %.*s, "
2655                        "inode %lu parent %lu\n", name_len, name,
2656                        inode->i_ino, dir->i_ino);
2657                 goto err;
2658         }
2659
2660         di = btrfs_lookup_dir_index_item(trans, root, path, dir->i_ino,
2661                                          index, name, name_len, -1);
2662         if (IS_ERR(di)) {
2663                 ret = PTR_ERR(di);
2664                 goto err;
2665         }
2666         if (!di) {
2667                 ret = -ENOENT;
2668                 goto err;
2669         }
2670         ret = btrfs_delete_one_dir_name(trans, root, path, di);
2671         btrfs_release_path(root, path);
2672
2673         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
2674                                          inode, dir->i_ino);
2675         BUG_ON(ret != 0 && ret != -ENOENT);
2676
2677         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
2678                                            dir, index);
2679         if (ret == -ENOENT)
2680                 ret = 0;
2681 err:
2682         btrfs_free_path(path);
2683         if (ret)
2684                 goto out;
2685
2686         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
2687         inode->i_ctime = dir->i_mtime = dir->i_ctime = CURRENT_TIME;
2688         btrfs_update_inode(trans, root, dir);
2689         btrfs_drop_nlink(inode);
2690         ret = btrfs_update_inode(trans, root, inode);
2691 out:
2692         return ret;
2693 }
2694
2695 /* helper to check if there is any shared block in the path */
2696 static int check_path_shared(struct btrfs_root *root,
2697                              struct btrfs_path *path)
2698 {
2699         struct extent_buffer *eb;
2700         int level;
2701         u64 refs = 1;
2702         int uninitialized_var(ret);
2703
2704         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2705                 if (!path->nodes[level])
2706                         break;
2707                 eb = path->nodes[level];
2708                 if (!btrfs_block_can_be_shared(root, eb))
2709                         continue;
2710                 ret = btrfs_lookup_extent_info(NULL, root, eb->start, eb->len,
2711                                                &refs, NULL);
2712                 if (refs > 1)
2713                         return 1;
2714         }
2715         return ret; /* XXX callers? */
2716 }
2717
2718 /*
2719  * helper to start transaction for unlink and rmdir.
2720  *
2721  * unlink and rmdir are special in btrfs, they do not always free space.
2722  * so in enospc case, we should make sure they will free space before
2723  * allowing them to use the global metadata reservation.
2724  */
2725 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir,
2726                                                        struct dentry *dentry)
2727 {
2728         struct btrfs_trans_handle *trans;
2729         struct btrfs_root *root = BTRFS_I(dir)->root;
2730         struct btrfs_path *path;
2731         struct btrfs_inode_ref *ref;
2732         struct btrfs_dir_item *di;
2733         struct inode *inode = dentry->d_inode;
2734         u64 index;
2735         int check_link = 1;
2736         int err = -ENOSPC;
2737         int ret;
2738
2739         trans = btrfs_start_transaction(root, 10);
2740         if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
2741                 return trans;
2742
2743         if (inode->i_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
2744                 return ERR_PTR(-ENOSPC);
2745
2746         /* check if there is someone else holds reference */
2747         if (S_ISDIR(inode->i_mode) && atomic_read(&inode->i_count) > 1)
2748                 return ERR_PTR(-ENOSPC);
2749
2750         if (atomic_read(&inode->i_count) > 2)
2751                 return ERR_PTR(-ENOSPC);
2752
2753         if (xchg(&root->fs_info->enospc_unlink, 1))
2754                 return ERR_PTR(-ENOSPC);
2755
2756         path = btrfs_alloc_path();
2757         if (!path) {
2758                 root->fs_info->enospc_unlink = 0;
2759                 return ERR_PTR(-ENOMEM);
2760         }
2761
2762         trans = btrfs_start_transaction(root, 0);
2763         if (IS_ERR(trans)) {
2764                 btrfs_free_path(path);
2765                 root->fs_info->enospc_unlink = 0;
2766                 return trans;
2767         }
2768
2769         path->skip_locking = 1;
2770         path->search_commit_root = 1;
2771
2772         ret = btrfs_lookup_inode(trans, root, path,
2773                                 &BTRFS_I(dir)->location, 0);
2774         if (ret < 0) {
2775                 err = ret;
2776                 goto out;
2777         }
2778         if (ret == 0) {
2779                 if (check_path_shared(root, path))
2780                         goto out;
2781         } else {
2782                 check_link = 0;
2783         }
2784         btrfs_release_path(root, path);
2785
2786         ret = btrfs_lookup_inode(trans, root, path,
2787                                 &BTRFS_I(inode)->location, 0);
2788         if (ret < 0) {
2789                 err = ret;
2790                 goto out;
2791         }
2792         if (ret == 0) {
2793                 if (check_path_shared(root, path))
2794                         goto out;
2795         } else {
2796                 check_link = 0;
2797         }
2798         btrfs_release_path(root, path);
2799
2800         if (ret == 0 && S_ISREG(inode->i_mode)) {
2801                 ret = btrfs_lookup_file_extent(trans, root, path,
2802                                                inode->i_ino, (u64)-1, 0);
2803                 if (ret < 0) {
2804                         err = ret;
2805                         goto out;
2806                 }
2807                 BUG_ON(ret == 0);
2808                 if (check_path_shared(root, path))
2809                         goto out;
2810                 btrfs_release_path(root, path);
2811         }
2812
2813         if (!check_link) {
2814                 err = 0;
2815                 goto out;
2816         }
2817
2818         di = btrfs_lookup_dir_item(trans, root, path, dir->i_ino,
2819                                 dentry->d_name.name, dentry->d_name.len, 0);
2820         if (IS_ERR(di)) {
2821                 err = PTR_ERR(di);
2822                 goto out;
2823         }
2824         if (di) {
2825                 if (check_path_shared(root, path))
2826                         goto out;
2827         } else {
2828                 err = 0;
2829                 goto out;
2830         }
2831         btrfs_release_path(root, path);
2832
2833         ref = btrfs_lookup_inode_ref(trans, root, path,
2834                                 dentry->d_name.name, dentry->d_name.len,
2835                                 inode->i_ino, dir->i_ino, 0);
2836         if (IS_ERR(ref)) {
2837                 err = PTR_ERR(ref);
2838                 goto out;
2839         }
2840         BUG_ON(!ref);
2841         if (check_path_shared(root, path))
2842                 goto out;
2843         index = btrfs_inode_ref_index(path->nodes[0], ref);
2844         btrfs_release_path(root, path);
2845
2846         di = btrfs_lookup_dir_index_item(trans, root, path, dir->i_ino, index,
2847                                 dentry->d_name.name, dentry->d_name.len, 0);
2848         if (IS_ERR(di)) {
2849                 err = PTR_ERR(di);
2850                 goto out;
2851         }
2852         BUG_ON(ret == -ENOENT);
2853         if (check_path_shared(root, path))
2854                 goto out;
2855
2856         err = 0;
2857 out:
2858         btrfs_free_path(path);
2859         if (err) {
2860                 btrfs_end_transaction(trans, root);
2861                 root->fs_info->enospc_unlink = 0;
2862                 return ERR_PTR(err);
2863         }
2864
2865         trans->block_rsv = &root->fs_info->global_block_rsv;
2866         return trans;
2867 }
2868
2869 static void __unlink_end_trans(struct btrfs_trans_handle *trans,
2870                                struct btrfs_root *root)
2871 {
2872         if (trans->block_rsv == &root->fs_info->global_block_rsv) {
2873                 BUG_ON(!root->fs_info->enospc_unlink);
2874                 root->fs_info->enospc_unlink = 0;
2875         }
2876         btrfs_end_transaction_throttle(trans, root);
2877 }
2878
2879 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
2880 {
2881         struct btrfs_root *root = BTRFS_I(dir)->root;
2882         struct btrfs_trans_handle *trans;
2883         struct inode *inode = dentry->d_inode;
2884         int ret;
2885         unsigned long nr = 0;
2886
2887         trans = __unlink_start_trans(dir, dentry);
2888         if (IS_ERR(trans))
2889                 return PTR_ERR(trans);
2890
2891         btrfs_set_trans_block_group(trans, dir);
2892
2893         btrfs_record_unlink_dir(trans, dir, dentry->d_inode, 0);
2894
2895         ret = btrfs_unlink_inode(trans, root, dir, dentry->d_inode,
2896                                  dentry->d_name.name, dentry->d_name.len);
2897         BUG_ON(ret);
2898
2899         if (inode->i_nlink == 0) {
2900                 ret = btrfs_orphan_add(trans, inode);
2901                 BUG_ON(ret);
2902         }
2903
2904         nr = trans->blocks_used;
2905         __unlink_end_trans(trans, root);
2906         btrfs_btree_balance_dirty(root, nr);
2907         return ret;
2908 }
2909
2910 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
2911                         struct btrfs_root *root,
2912                         struct inode *dir, u64 objectid,
2913                         const char *name, int name_len)
2914 {
2915         struct btrfs_path *path;
2916         struct extent_buffer *leaf;
2917         struct btrfs_dir_item *di;
2918         struct btrfs_key key;
2919         u64 index;
2920         int ret;
2921
2922         path = btrfs_alloc_path();
2923         if (!path)
2924                 return -ENOMEM;
2925
2926         di = btrfs_lookup_dir_item(trans, root, path, dir->i_ino,
2927                                    name, name_len, -1);
2928         BUG_ON(!di || IS_ERR(di));
2929
2930         leaf = path->nodes[0];
2931         btrfs_dir_item_key_to_cpu(leaf, di, &key);
2932         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
2933         ret = btrfs_delete_one_dir_name(trans, root, path, di);
2934         BUG_ON(ret);
2935         btrfs_release_path(root, path);
2936
2937         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
2938                                  objectid, root->root_key.objectid,
2939                                  dir->i_ino, &index, name, name_len);
2940         if (ret < 0) {
2941                 BUG_ON(ret != -ENOENT);
2942                 di = btrfs_search_dir_index_item(root, path, dir->i_ino,
2943                                                  name, name_len);
2944                 BUG_ON(!di || IS_ERR(di));
2945
2946                 leaf = path->nodes[0];
2947                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2948                 btrfs_release_path(root, path);
2949                 index = key.offset;
2950         }
2951
2952         di = btrfs_lookup_dir_index_item(trans, root, path, dir->i_ino,
2953                                          index, name, name_len, -1);
2954         BUG_ON(!di || IS_ERR(di));
2955
2956         leaf = path->nodes[0];
2957         btrfs_dir_item_key_to_cpu(leaf, di, &key);
2958         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
2959         ret = btrfs_delete_one_dir_name(trans, root, path, di);
2960         BUG_ON(ret);
2961         btrfs_release_path(root, path);
2962
2963         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
2964         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
2965         ret = btrfs_update_inode(trans, root, dir);
2966         BUG_ON(ret);
2967
2968         btrfs_free_path(path);
2969         return 0;
2970 }
2971
2972 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
2973 {
2974         struct inode *inode = dentry->d_inode;
2975         int err = 0;
2976         struct btrfs_root *root = BTRFS_I(dir)->root;
2977         struct btrfs_trans_handle *trans;
2978         unsigned long nr = 0;
2979
2980         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE ||
2981             inode->i_ino == BTRFS_FIRST_FREE_OBJECTID)
2982                 return -ENOTEMPTY;
2983
2984         trans = __unlink_start_trans(dir, dentry);
2985         if (IS_ERR(trans))
2986                 return PTR_ERR(trans);
2987
2988         btrfs_set_trans_block_group(trans, dir);
2989
2990         if (unlikely(inode->i_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
2991                 err = btrfs_unlink_subvol(trans, root, dir,
2992                                           BTRFS_I(inode)->location.objectid,
2993                                           dentry->d_name.name,
2994                                           dentry->d_name.len);
2995                 goto out;
2996         }
2997
2998         err = btrfs_orphan_add(trans, inode);
2999         if (err)
3000                 goto out;
3001
3002         /* now the directory is empty */
3003         err = btrfs_unlink_inode(trans, root, dir, dentry->d_inode,
3004                                  dentry->d_name.name, dentry->d_name.len);
3005         if (!err)
3006                 btrfs_i_size_write(inode, 0);
3007 out:
3008         nr = trans->blocks_used;
3009         __unlink_end_trans(trans, root);
3010         btrfs_btree_balance_dirty(root, nr);
3011
3012         return err;
3013 }
3014
3015 #if 0
3016 /*
3017  * when truncating bytes in a file, it is possible to avoid reading
3018  * the leaves that contain only checksum items.  This can be the
3019  * majority of the IO required to delete a large file, but it must
3020  * be done carefully.
3021  *
3022  * The keys in the level just above the leaves are checked to make sure
3023  * the lowest key in a given leaf is a csum key, and starts at an offset
3024  * after the new  size.
3025  *
3026  * Then the key for the next leaf is checked to make sure it also has
3027  * a checksum item for the same file.  If it does, we know our target leaf
3028  * contains only checksum items, and it can be safely freed without reading
3029  * it.
3030  *
3031  * This is just an optimization targeted at large files.  It may do
3032  * nothing.  It will return 0 unless things went badly.
3033  */
3034 static noinline int drop_csum_leaves(struct btrfs_trans_handle *trans,
3035                                      struct btrfs_root *root,
3036                                      struct btrfs_path *path,
3037                                      struct inode *inode, u64 new_size)
3038 {
3039         struct btrfs_key key;
3040         int ret;
3041         int nritems;
3042         struct btrfs_key found_key;
3043         struct btrfs_key other_key;
3044         struct btrfs_leaf_ref *ref;
3045         u64 leaf_gen;
3046         u64 leaf_start;
3047
3048         path->lowest_level = 1;
3049         key.objectid = inode->i_ino;
3050         key.type = BTRFS_CSUM_ITEM_KEY;
3051         key.offset = new_size;
3052 again:
3053         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
3054         if (ret < 0)
3055                 goto out;
3056
3057         if (path->nodes[1] == NULL) {
3058                 ret = 0;
3059                 goto out;
3060         }
3061         ret = 0;
3062         btrfs_node_key_to_cpu(path->nodes[1], &found_key, path->slots[1]);
3063         nritems = btrfs_header_nritems(path->nodes[1]);
3064
3065         if (!nritems)
3066                 goto out;
3067
3068         if (path->slots[1] >= nritems)
3069                 goto next_node;
3070
3071         /* did we find a key greater than anything we want to delete? */
3072         if (found_key.objectid > inode->i_ino ||
3073            (found_key.objectid == inode->i_ino && found_key.type > key.type))
3074                 goto out;
3075
3076         /* we check the next key in the node to make sure the leave contains
3077          * only checksum items.  This comparison doesn't work if our
3078          * leaf is the last one in the node
3079          */
3080         if (path->slots[1] + 1 >= nritems) {
3081 next_node:
3082                 /* search forward from the last key in the node, this
3083                  * will bring us into the next node in the tree
3084                  */
3085                 btrfs_node_key_to_cpu(path->nodes[1], &found_key, nritems - 1);
3086
3087                 /* unlikely, but we inc below, so check to be safe */
3088                 if (found_key.offset == (u64)-1)
3089                         goto out;
3090
3091                 /* search_forward needs a path with locks held, do the
3092                  * search again for the original key.  It is possible
3093                  * this will race with a balance and return a path that
3094                  * we could modify, but this drop is just an optimization
3095                  * and is allowed to miss some leaves.
3096                  */
3097                 btrfs_release_path(root, path);
3098                 found_key.offset++;
3099
3100                 /* setup a max key for search_forward */
3101                 other_key.offset = (u64)-1;
3102                 other_key.type = key.type;
3103                 other_key.objectid = key.objectid;
3104
3105                 path->keep_locks = 1;
3106                 ret = btrfs_search_forward(root, &found_key, &other_key,
3107                                            path, 0, 0);
3108                 path->keep_locks = 0;
3109                 if (ret || found_key.objectid != key.objectid ||
3110                     found_key.type != key.type) {
3111                         ret = 0;
3112                         goto out;
3113                 }
3114
3115                 key.offset = found_key.offset;
3116                 btrfs_release_path(root, path);
3117                 cond_resched();
3118                 goto again;
3119         }
3120
3121         /* we know there's one more slot after us in the tree,
3122          * read that key so we can verify it is also a checksum item
3123          */
3124         btrfs_node_key_to_cpu(path->nodes[1], &other_key, path->slots[1] + 1);
3125
3126         if (found_key.objectid < inode->i_ino)
3127                 goto next_key;
3128
3129         if (found_key.type != key.type || found_key.offset < new_size)
3130                 goto next_key;
3131
3132         /*
3133          * if the key for the next leaf isn't a csum key from this objectid,
3134          * we can't be sure there aren't good items inside this leaf.
3135          * Bail out
3136          */
3137         if (other_key.objectid != inode->i_ino || other_key.type != key.type)
3138                 goto out;
3139
3140         leaf_start = btrfs_node_blockptr(path->nodes[1], path->slots[1]);
3141         leaf_gen = btrfs_node_ptr_generation(path->nodes[1], path->slots[1]);
3142         /*
3143          * it is safe to delete this leaf, it contains only
3144          * csum items from this inode at an offset >= new_size
3145          */
3146         ret = btrfs_del_leaf(trans, root, path, leaf_start);
3147         BUG_ON(ret);
3148
3149         if (root->ref_cows && leaf_gen < trans->transid) {
3150                 ref = btrfs_alloc_leaf_ref(root, 0);
3151                 if (ref) {
3152                         ref->root_gen = root->root_key.offset;
3153                         ref->bytenr = leaf_start;
3154                         ref->owner = 0;
3155                         ref->generation = leaf_gen;
3156                         ref->nritems = 0;
3157
3158                         btrfs_sort_leaf_ref(ref);
3159
3160                         ret = btrfs_add_leaf_ref(root, ref, 0);
3161                         WARN_ON(ret);
3162                         btrfs_free_leaf_ref(root, ref);
3163                 } else {
3164                         WARN_ON(1);
3165                 }
3166         }
3167 next_key:
3168         btrfs_release_path(root, path);
3169
3170         if (other_key.objectid == inode->i_ino &&
3171             other_key.type == key.type && other_key.offset > key.offset) {
3172                 key.offset = other_key.offset;
3173                 cond_resched();
3174                 goto again;
3175         }
3176         ret = 0;
3177 out:
3178         /* fixup any changes we've made to the path */
3179         path->lowest_level = 0;
3180         path->keep_locks = 0;
3181         btrfs_release_path(root, path);
3182         return ret;
3183 }
3184
3185 #endif
3186
3187 /*
3188  * this can truncate away extent items, csum items and directory items.
3189  * It starts at a high offset and removes keys until it can't find
3190  * any higher than new_size
3191  *
3192  * csum items that cross the new i_size are truncated to the new size
3193  * as well.
3194  *
3195  * min_type is the minimum key type to truncate down to.  If set to 0, this
3196  * will kill all the items on this inode, including the INODE_ITEM_KEY.
3197  */
3198 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
3199                                struct btrfs_root *root,
3200                                struct inode *inode,
3201                                u64 new_size, u32 min_type)
3202 {
3203         struct btrfs_path *path;
3204         struct extent_buffer *leaf;
3205         struct btrfs_file_extent_item *fi;
3206         struct btrfs_key key;
3207         struct btrfs_key found_key;
3208         u64 extent_start = 0;
3209         u64 extent_num_bytes = 0;
3210         u64 extent_offset = 0;
3211         u64 item_end = 0;
3212         u64 mask = root->sectorsize - 1;
3213         u32 found_type = (u8)-1;
3214         int found_extent;
3215         int del_item;
3216         int pending_del_nr = 0;
3217         int pending_del_slot = 0;
3218         int extent_type = -1;
3219         int encoding;
3220         int ret;
3221         int err = 0;
3222
3223         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
3224
3225         if (root->ref_cows || root == root->fs_info->tree_root)
3226                 btrfs_drop_extent_cache(inode, new_size & (~mask), (u64)-1, 0);
3227
3228         path = btrfs_alloc_path();
3229         BUG_ON(!path);
3230         path->reada = -1;
3231
3232         key.objectid = inode->i_ino;
3233         key.offset = (u64)-1;
3234         key.type = (u8)-1;
3235
3236 search_again:
3237         path->leave_spinning = 1;
3238         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
3239         if (ret < 0) {
3240                 err = ret;
3241                 goto out;
3242         }
3243
3244         if (ret > 0) {
3245                 /* there are no items in the tree for us to truncate, we're
3246                  * done
3247                  */
3248                 if (path->slots[0] == 0)
3249                         goto out;
3250                 path->slots[0]--;
3251         }
3252
3253         while (1) {
3254                 fi = NULL;
3255                 leaf = path->nodes[0];
3256                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3257                 found_type = btrfs_key_type(&found_key);
3258                 encoding = 0;
3259
3260                 if (found_key.objectid != inode->i_ino)
3261                         break;
3262
3263                 if (found_type < min_type)
3264                         break;
3265
3266                 item_end = found_key.offset;
3267                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
3268                         fi = btrfs_item_ptr(leaf, path->slots[0],
3269                                             struct btrfs_file_extent_item);
3270                         extent_type = btrfs_file_extent_type(leaf, fi);
3271                         encoding = btrfs_file_extent_compression(leaf, fi);
3272                         encoding |= btrfs_file_extent_encryption(leaf, fi);
3273                         encoding |= btrfs_file_extent_other_encoding(leaf, fi);
3274
3275                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
3276                                 item_end +=
3277                                     btrfs_file_extent_num_bytes(leaf, fi);
3278                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
3279                                 item_end += btrfs_file_extent_inline_len(leaf,
3280                                                                          fi);
3281                         }
3282                         item_end--;
3283                 }
3284                 if (found_type > min_type) {
3285                         del_item = 1;
3286                 } else {
3287                         if (item_end < new_size)
3288                                 break;
3289                         if (found_key.offset >= new_size)
3290                                 del_item = 1;
3291                         else
3292                                 del_item = 0;
3293                 }
3294                 found_extent = 0;
3295                 /* FIXME, shrink the extent if the ref count is only 1 */
3296                 if (found_type != BTRFS_EXTENT_DATA_KEY)
3297                         goto delete;
3298
3299                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
3300                         u64 num_dec;
3301                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
3302                         if (!del_item && !encoding) {
3303                                 u64 orig_num_bytes =
3304                                         btrfs_file_extent_num_bytes(leaf, fi);
3305                                 extent_num_bytes = new_size -
3306                                         found_key.offset + root->sectorsize - 1;
3307                                 extent_num_bytes = extent_num_bytes &
3308                                         ~((u64)root->sectorsize - 1);
3309                                 btrfs_set_file_extent_num_bytes(leaf, fi,
3310                                                          extent_num_bytes);
3311                                 num_dec = (orig_num_bytes -
3312                                            extent_num_bytes);
3313                                 if (root->ref_cows && extent_start != 0)
3314                                         inode_sub_bytes(inode, num_dec);
3315                                 btrfs_mark_buffer_dirty(leaf);
3316                         } else {
3317                                 extent_num_bytes =
3318                                         btrfs_file_extent_disk_num_bytes(leaf,
3319                                                                          fi);
3320                                 extent_offset = found_key.offset -
3321                                         btrfs_file_extent_offset(leaf, fi);
3322
3323                                 /* FIXME blocksize != 4096 */
3324                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
3325                                 if (extent_start != 0) {
3326                                         found_extent = 1;
3327                                         if (root->ref_cows)
3328                                                 inode_sub_bytes(inode, num_dec);
3329                                 }
3330                         }
3331                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
3332                         /*
3333                          * we can't truncate inline items that have had
3334                          * special encodings
3335                          */
3336                         if (!del_item &&
3337                             btrfs_file_extent_compression(leaf, fi) == 0 &&
3338                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
3339                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
3340                                 u32 size = new_size - found_key.offset;
3341
3342                                 if (root->ref_cows) {
3343                                         inode_sub_bytes(inode, item_end + 1 -
3344                                                         new_size);
3345                                 }
3346                                 size =
3347                                     btrfs_file_extent_calc_inline_size(size);
3348                                 ret = btrfs_truncate_item(trans, root, path,
3349                                                           size, 1);
3350                                 BUG_ON(ret);
3351                         } else if (root->ref_cows) {
3352                                 inode_sub_bytes(inode, item_end + 1 -
3353                                                 found_key.offset);
3354                         }
3355                 }
3356 delete:
3357                 if (del_item) {
3358                         if (!pending_del_nr) {
3359                                 /* no pending yet, add ourselves */
3360                                 pending_del_slot = path->slots[0];
3361                                 pending_del_nr = 1;
3362                         } else if (pending_del_nr &&
3363                                    path->slots[0] + 1 == pending_del_slot) {
3364                                 /* hop on the pending chunk */
3365                                 pending_del_nr++;
3366                                 pending_del_slot = path->slots[0];
3367                         } else {
3368                                 BUG();
3369                         }
3370                 } else {
3371                         break;
3372                 }
3373                 if (found_extent && (root->ref_cows ||
3374                                      root == root->fs_info->tree_root)) {
3375                         btrfs_set_path_blocking(path);
3376                         ret = btrfs_free_extent(trans, root, extent_start,
3377                                                 extent_num_bytes, 0,
3378                                                 btrfs_header_owner(leaf),
3379                                                 inode->i_ino, extent_offset);
3380                         BUG_ON(ret);
3381                 }
3382
3383                 if (found_type == BTRFS_INODE_ITEM_KEY)
3384                         break;
3385
3386                 if (path->slots[0] == 0 ||
3387                     path->slots[0] != pending_del_slot) {
3388                         if (root->ref_cows) {
3389                                 err = -EAGAIN;
3390                                 goto out;
3391                         }
3392                         if (pending_del_nr) {
3393                                 ret = btrfs_del_items(trans, root, path,
3394                                                 pending_del_slot,
3395                                                 pending_del_nr);
3396                                 BUG_ON(ret);
3397                                 pending_del_nr = 0;
3398                         }
3399                         btrfs_release_path(root, path);
3400                         goto search_again;
3401                 } else {
3402                         path->slots[0]--;
3403                 }
3404         }
3405 out:
3406         if (pending_del_nr) {
3407                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
3408                                       pending_del_nr);
3409                 BUG_ON(ret);
3410         }
3411         btrfs_free_path(path);
3412         return err;
3413 }
3414
3415 /*
3416  * taken from block_truncate_page, but does cow as it zeros out
3417  * any bytes left in the last page in the file.
3418  */
3419 static int btrfs_truncate_page(struct address_space *mapping, loff_t from)
3420 {
3421         struct inode *inode = mapping->host;
3422         struct btrfs_root *root = BTRFS_I(inode)->root;
3423         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3424         struct btrfs_ordered_extent *ordered;
3425         struct extent_state *cached_state = NULL;
3426         char *kaddr;
3427         u32 blocksize = root->sectorsize;
3428         pgoff_t index = from >> PAGE_CACHE_SHIFT;
3429         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3430         struct page *page;
3431         int ret = 0;
3432         u64 page_start;
3433         u64 page_end;
3434
3435         if ((offset & (blocksize - 1)) == 0)
3436                 goto out;
3437         ret = btrfs_delalloc_reserve_space(inode, PAGE_CACHE_SIZE);
3438         if (ret)
3439                 goto out;
3440
3441         ret = -ENOMEM;
3442 again:
3443         page = grab_cache_page(mapping, index);
3444         if (!page) {
3445                 btrfs_delalloc_release_space(inode, PAGE_CACHE_SIZE);
3446                 goto out;
3447         }
3448
3449         page_start = page_offset(page);
3450         page_end = page_start + PAGE_CACHE_SIZE - 1;
3451
3452         if (!PageUptodate(page)) {
3453                 ret = btrfs_readpage(NULL, page);
3454                 lock_page(page);
3455                 if (page->mapping != mapping) {
3456                         unlock_page(page);
3457                         page_cache_release(page);
3458                         goto again;
3459                 }
3460                 if (!PageUptodate(page)) {
3461                         ret = -EIO;
3462                         goto out_unlock;
3463                 }
3464         }
3465         wait_on_page_writeback(page);
3466
3467         lock_extent_bits(io_tree, page_start, page_end, 0, &cached_state,
3468                          GFP_NOFS);
3469         set_page_extent_mapped(page);
3470
3471         ordered = btrfs_lookup_ordered_extent(inode, page_start);
3472         if (ordered) {
3473                 unlock_extent_cached(io_tree, page_start, page_end,
3474                                      &cached_state, GFP_NOFS);
3475                 unlock_page(page);
3476                 page_cache_release(page);
3477                 btrfs_start_ordered_extent(inode, ordered, 1);
3478                 btrfs_put_ordered_extent(ordered);
3479                 goto again;
3480         }
3481
3482         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, page_end,
3483                           EXTENT_DIRTY | EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING,
3484                           0, 0, &cached_state, GFP_NOFS);
3485
3486         ret = btrfs_set_extent_delalloc(inode, page_start, page_end,
3487                                         &cached_state);
3488         if (ret) {
3489                 unlock_extent_cached(io_tree, page_start, page_end,
3490                                      &cached_state, GFP_NOFS);
3491                 goto out_unlock;
3492         }
3493
3494         ret = 0;
3495         if (offset != PAGE_CACHE_SIZE) {
3496                 kaddr = kmap(page);
3497                 memset(kaddr + offset, 0, PAGE_CACHE_SIZE - offset);
3498                 flush_dcache_page(page);
3499                 kunmap(page);
3500         }
3501         ClearPageChecked(page);
3502         set_page_dirty(page);
3503         unlock_extent_cached(io_tree, page_start, page_end, &cached_state,
3504                              GFP_NOFS);
3505
3506 out_unlock:
3507         if (ret)
3508                 btrfs_delalloc_release_space(inode, PAGE_CACHE_SIZE);
3509         unlock_page(page);
3510         page_cache_release(page);
3511 out:
3512         return ret;
3513 }
3514
3515 int btrfs_cont_expand(struct inode *inode, loff_t size)
3516 {
3517         struct btrfs_trans_handle *trans;
3518         struct btrfs_root *root = BTRFS_I(inode)->root;
3519         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3520         struct extent_map *em = NULL;
3521         struct extent_state *cached_state = NULL;
3522         u64 mask = root->sectorsize - 1;
3523         u64 hole_start = (inode->i_size + mask) & ~mask;
3524         u64 block_end = (size + mask) & ~mask;
3525         u64 last_byte;
3526         u64 cur_offset;
3527         u64 hole_size;
3528         int err = 0;
3529
3530         if (size <= hole_start)
3531                 return 0;
3532
3533         while (1) {
3534                 struct btrfs_ordered_extent *ordered;
3535                 btrfs_wait_ordered_range(inode, hole_start,
3536                                          block_end - hole_start);
3537                 lock_extent_bits(io_tree, hole_start, block_end - 1, 0,
3538                                  &cached_state, GFP_NOFS);
3539                 ordered = btrfs_lookup_ordered_extent(inode, hole_start);
3540                 if (!ordered)
3541                         break;
3542                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
3543                                      &cached_state, GFP_NOFS);
3544                 btrfs_put_ordered_extent(ordered);
3545         }
3546
3547         cur_offset = hole_start;
3548         while (1) {
3549                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
3550                                 block_end - cur_offset, 0);
3551                 BUG_ON(IS_ERR(em) || !em);
3552                 last_byte = min(extent_map_end(em), block_end);
3553                 last_byte = (last_byte + mask) & ~mask;
3554                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3555                         u64 hint_byte = 0;
3556                         hole_size = last_byte - cur_offset;
3557
3558                         trans = btrfs_start_transaction(root, 2);
3559                         if (IS_ERR(trans)) {
3560                                 err = PTR_ERR(trans);
3561                                 break;
3562                         }
3563                         btrfs_set_trans_block_group(trans, inode);
3564
3565                         err = btrfs_drop_extents(trans, inode, cur_offset,
3566                                                  cur_offset + hole_size,
3567                                                  &hint_byte, 1);
3568                         BUG_ON(err);
3569
3570                         err = btrfs_insert_file_extent(trans, root,
3571                                         inode->i_ino, cur_offset, 0,
3572                                         0, hole_size, 0, hole_size,
3573                                         0, 0, 0);
3574                         BUG_ON(err);
3575
3576                         btrfs_drop_extent_cache(inode, hole_start,
3577                                         last_byte - 1, 0);
3578
3579                         btrfs_end_transaction(trans, root);
3580                 }
3581                 free_extent_map(em);
3582                 em = NULL;
3583                 cur_offset = last_byte;
3584                 if (cur_offset >= block_end)
3585                         break;
3586         }
3587
3588         free_extent_map(em);
3589         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
3590                              GFP_NOFS);
3591         return err;
3592 }
3593
3594 static int btrfs_setattr_size(struct inode *inode, struct iattr *attr)
3595 {
3596         struct btrfs_root *root = BTRFS_I(inode)->root;
3597         struct btrfs_trans_handle *trans;
3598         unsigned long nr;
3599         int ret;
3600
3601         if (attr->ia_size == inode->i_size)
3602                 return 0;
3603
3604         if (attr->ia_size > inode->i_size) {
3605                 unsigned long limit;
3606                 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
3607                 if (attr->ia_size > inode->i_sb->s_maxbytes)
3608                         return -EFBIG;
3609                 if (limit != RLIM_INFINITY && attr->ia_size > limit) {
3610                         send_sig(SIGXFSZ, current, 0);
3611                         return -EFBIG;
3612                 }
3613         }
3614
3615         trans = btrfs_start_transaction(root, 5);
3616         if (IS_ERR(trans))
3617                 return PTR_ERR(trans);
3618
3619         btrfs_set_trans_block_group(trans, inode);
3620
3621         ret = btrfs_orphan_add(trans, inode);
3622         BUG_ON(ret);
3623
3624         nr = trans->blocks_used;
3625         btrfs_end_transaction(trans, root);
3626         btrfs_btree_balance_dirty(root, nr);
3627
3628         if (attr->ia_size > inode->i_size) {
3629                 ret = btrfs_cont_expand(inode, attr->ia_size);
3630                 if (ret) {
3631                         btrfs_truncate(inode);
3632                         return ret;
3633                 }
3634
3635                 i_size_write(inode, attr->ia_size);
3636                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
3637
3638                 trans = btrfs_start_transaction(root, 0);
3639                 BUG_ON(IS_ERR(trans));
3640                 btrfs_set_trans_block_group(trans, inode);
3641                 trans->block_rsv = root->orphan_block_rsv;
3642                 BUG_ON(!trans->block_rsv);
3643
3644                 ret = btrfs_update_inode(trans, root, inode);
3645                 BUG_ON(ret);
3646                 if (inode->i_nlink > 0) {
3647                         ret = btrfs_orphan_del(trans, inode);
3648                         BUG_ON(ret);
3649                 }
3650                 nr = trans->blocks_used;
3651                 btrfs_end_transaction(trans, root);
3652                 btrfs_btree_balance_dirty(root, nr);
3653                 return 0;
3654         }
3655
3656         /*
3657          * We're truncating a file that used to have good data down to
3658          * zero. Make sure it gets into the ordered flush list so that
3659          * any new writes get down to disk quickly.
3660          */
3661         if (attr->ia_size == 0)
3662                 BTRFS_I(inode)->ordered_data_close = 1;
3663
3664         /* we don't support swapfiles, so vmtruncate shouldn't fail */
3665         ret = vmtruncate(inode, attr->ia_size);
3666         BUG_ON(ret);
3667
3668         return 0;
3669 }
3670
3671 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
3672 {
3673         struct inode *inode = dentry->d_inode;
3674         int err;
3675
3676         err = inode_change_ok(inode, attr);
3677         if (err)
3678                 return err;
3679
3680         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
3681                 err = btrfs_setattr_size(inode, attr);
3682                 if (err)
3683                         return err;
3684         }
3685
3686         if (attr->ia_valid) {
3687                 setattr_copy(inode, attr);
3688                 mark_inode_dirty(inode);
3689
3690                 if (attr->ia_valid & ATTR_MODE)
3691                         err = btrfs_acl_chmod(inode);
3692         }
3693
3694         return err;
3695 }
3696
3697 void btrfs_evict_inode(struct inode *inode)
3698 {
3699         struct btrfs_trans_handle *trans;
3700         struct btrfs_root *root = BTRFS_I(inode)->root;
3701         unsigned long nr;
3702         int ret;
3703
3704         truncate_inode_pages(&inode->i_data, 0);
3705         if (inode->i_nlink && (btrfs_root_refs(&root->root_item) != 0 ||
3706                                root == root->fs_info->tree_root))
3707                 goto no_delete;
3708
3709         if (is_bad_inode(inode)) {
3710                 btrfs_orphan_del(NULL, inode);
3711                 goto no_delete;
3712         }
3713         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
3714         btrfs_wait_ordered_range(inode, 0, (u64)-1);
3715
3716         if (root->fs_info->log_root_recovering) {
3717                 BUG_ON(!list_empty(&BTRFS_I(inode)->i_orphan));
3718                 goto no_delete;
3719         }
3720
3721         if (inode->i_nlink > 0) {
3722                 BUG_ON(btrfs_root_refs(&root->root_item) != 0);
3723                 goto no_delete;
3724         }
3725
3726         btrfs_i_size_write(inode, 0);
3727
3728         while (1) {
3729                 trans = btrfs_start_transaction(root, 0);
3730                 BUG_ON(IS_ERR(trans));
3731                 btrfs_set_trans_block_group(trans, inode);
3732                 trans->block_rsv = root->orphan_block_rsv;
3733
3734                 ret = btrfs_block_rsv_check(trans, root,
3735                                             root->orphan_block_rsv, 0, 5);
3736                 if (ret) {
3737                         BUG_ON(ret != -EAGAIN);
3738                         ret = btrfs_commit_transaction(trans, root);
3739                         BUG_ON(ret);
3740                         continue;
3741                 }
3742
3743                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
3744                 if (ret != -EAGAIN)
3745                         break;
3746
3747                 nr = trans->blocks_used;
3748                 btrfs_end_transaction(trans, root);
3749                 trans = NULL;
3750                 btrfs_btree_balance_dirty(root, nr);
3751
3752         }
3753
3754         if (ret == 0) {
3755                 ret = btrfs_orphan_del(trans, inode);
3756                 BUG_ON(ret);
3757         }
3758
3759         nr = trans->blocks_used;
3760         btrfs_end_transaction(trans, root);
3761         btrfs_btree_balance_dirty(root, nr);
3762 no_delete:
3763         end_writeback(inode);
3764         return;
3765 }
3766
3767 /*
3768  * this returns the key found in the dir entry in the location pointer.
3769  * If no dir entries were found, location->objectid is 0.
3770  */
3771 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
3772                                struct btrfs_key *location)
3773 {
3774         const char *name = dentry->d_name.name;
3775         int namelen = dentry->d_name.len;
3776         struct btrfs_dir_item *di;
3777         struct btrfs_path *path;
3778         struct btrfs_root *root = BTRFS_I(dir)->root;
3779         int ret = 0;
3780
3781         path = btrfs_alloc_path();
3782         BUG_ON(!path);
3783
3784         di = btrfs_lookup_dir_item(NULL, root, path, dir->i_ino, name,
3785                                     namelen, 0);
3786         if (IS_ERR(di))
3787                 ret = PTR_ERR(di);
3788
3789         if (!di || IS_ERR(di))
3790                 goto out_err;
3791
3792         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
3793 out:
3794         btrfs_free_path(path);
3795         return ret;
3796 out_err:
3797         location->objectid = 0;
3798         goto out;
3799 }
3800
3801 /*
3802  * when we hit a tree root in a directory, the btrfs part of the inode
3803  * needs to be changed to reflect the root directory of the tree root.  This
3804  * is kind of like crossing a mount point.
3805  */
3806 static int fixup_tree_root_location(struct btrfs_root *root,
3807                                     struct inode *dir,
3808                                     struct dentry *dentry,
3809                                     struct btrfs_key *location,
3810                                     struct btrfs_root **sub_root)
3811 {
3812         struct btrfs_path *path;
3813         struct btrfs_root *new_root;
3814         struct btrfs_root_ref *ref;
3815         struct extent_buffer *leaf;
3816         int ret;
3817         int err = 0;
3818
3819         path = btrfs_alloc_path();
3820         if (!path) {
3821                 err = -ENOMEM;
3822                 goto out;
3823         }
3824
3825         err = -ENOENT;
3826         ret = btrfs_find_root_ref(root->fs_info->tree_root, path,
3827                                   BTRFS_I(dir)->root->root_key.objectid,
3828                                   location->objectid);
3829         if (ret) {
3830                 if (ret < 0)
3831                         err = ret;
3832                 goto out;
3833         }
3834
3835         leaf = path->nodes[0];
3836         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
3837         if (btrfs_root_ref_dirid(leaf, ref) != dir->i_ino ||
3838             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
3839                 goto out;
3840
3841         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
3842                                    (unsigned long)(ref + 1),
3843                                    dentry->d_name.len);
3844         if (ret)
3845                 goto out;
3846
3847         btrfs_release_path(root->fs_info->tree_root, path);
3848
3849         new_root = btrfs_read_fs_root_no_name(root->fs_info, location);
3850         if (IS_ERR(new_root)) {
3851                 err = PTR_ERR(new_root);
3852                 goto out;
3853         }
3854
3855         if (btrfs_root_refs(&new_root->root_item) == 0) {
3856                 err = -ENOENT;
3857                 goto out;
3858         }
3859
3860         *sub_root = new_root;
3861         location->objectid = btrfs_root_dirid(&new_root->root_item);
3862         location->type = BTRFS_INODE_ITEM_KEY;
3863         location->offset = 0;
3864         err = 0;
3865 out:
3866         btrfs_free_path(path);
3867         return err;
3868 }
3869
3870 static void inode_tree_add(struct inode *inode)
3871 {
3872         struct btrfs_root *root = BTRFS_I(inode)->root;
3873         struct btrfs_inode *entry;
3874         struct rb_node **p;
3875         struct rb_node *parent;
3876 again:
3877         p = &root->inode_tree.rb_node;
3878         parent = NULL;
3879
3880         if (hlist_unhashed(&inode->i_hash))
3881                 return;
3882
3883         spin_lock(&root->inode_lock);
3884         while (*p) {
3885                 parent = *p;
3886                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
3887
3888                 if (inode->i_ino < entry->vfs_inode.i_ino)
3889                         p = &parent->rb_left;
3890                 else if (inode->i_ino > entry->vfs_inode.i_ino)
3891                         p = &parent->rb_right;
3892                 else {
3893                         WARN_ON(!(entry->vfs_inode.i_state &
3894                                   (I_WILL_FREE | I_FREEING)));
3895                         rb_erase(parent, &root->inode_tree);
3896                         RB_CLEAR_NODE(parent);
3897                         spin_unlock(&root->inode_lock);
3898                         goto again;
3899                 }
3900         }
3901         rb_link_node(&BTRFS_I(inode)->rb_node, parent, p);
3902         rb_insert_color(&BTRFS_I(inode)->rb_node, &root->inode_tree);
3903         spin_unlock(&root->inode_lock);
3904 }
3905
3906 static void inode_tree_del(struct inode *inode)
3907 {
3908         struct btrfs_root *root = BTRFS_I(inode)->root;
3909         int empty = 0;
3910
3911         spin_lock(&root->inode_lock);
3912         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
3913                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
3914                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
3915                 empty = RB_EMPTY_ROOT(&root->inode_tree);
3916         }
3917         spin_unlock(&root->inode_lock);
3918
3919         /*
3920          * Free space cache has inodes in the tree root, but the tree root has a
3921          * root_refs of 0, so this could end up dropping the tree root as a
3922          * snapshot, so we need the extra !root->fs_info->tree_root check to
3923          * make sure we don't drop it.
3924          */
3925         if (empty && btrfs_root_refs(&root->root_item) == 0 &&
3926             root != root->fs_info->tree_root) {
3927                 synchronize_srcu(&root->fs_info->subvol_srcu);
3928                 spin_lock(&root->inode_lock);
3929                 empty = RB_EMPTY_ROOT(&root->inode_tree);
3930                 spin_unlock(&root->inode_lock);
3931                 if (empty)
3932                         btrfs_add_dead_root(root);
3933         }
3934 }
3935
3936 int btrfs_invalidate_inodes(struct btrfs_root *root)
3937 {
3938         struct rb_node *node;
3939         struct rb_node *prev;
3940         struct btrfs_inode *entry;
3941         struct inode *inode;
3942         u64 objectid = 0;
3943
3944         WARN_ON(btrfs_root_refs(&root->root_item) != 0);
3945
3946         spin_lock(&root->inode_lock);
3947 again:
3948         node = root->inode_tree.rb_node;
3949         prev = NULL;
3950         while (node) {
3951                 prev = node;
3952                 entry = rb_entry(node, struct btrfs_inode, rb_node);
3953
3954                 if (objectid < entry->vfs_inode.i_ino)
3955                         node = node->rb_left;
3956                 else if (objectid > entry->vfs_inode.i_ino)
3957                         node = node->rb_right;
3958                 else
3959                         break;
3960         }
3961         if (!node) {
3962                 while (prev) {
3963                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
3964                         if (objectid <= entry->vfs_inode.i_ino) {
3965                                 node = prev;
3966                                 break;
3967                         }
3968                         prev = rb_next(prev);
3969                 }
3970         }
3971         while (node) {
3972                 entry = rb_entry(node, struct btrfs_inode, rb_node);
3973                 objectid = entry->vfs_inode.i_ino + 1;
3974                 inode = igrab(&entry->vfs_inode);
3975                 if (inode) {
3976                         spin_unlock(&root->inode_lock);
3977                         if (atomic_read(&inode->i_count) > 1)
3978                                 d_prune_aliases(inode);
3979                         /*
3980                          * btrfs_drop_inode will have it removed from
3981                          * the inode cache when its usage count
3982                          * hits zero.
3983                          */
3984                         iput(inode);
3985                         cond_resched();
3986                         spin_lock(&root->inode_lock);
3987                         goto again;
3988                 }
3989
3990                 if (cond_resched_lock(&root->inode_lock))
3991                         goto again;
3992
3993                 node = rb_next(node);
3994         }
3995         spin_unlock(&root->inode_lock);
3996         return 0;
3997 }
3998
3999 static int btrfs_init_locked_inode(struct inode *inode, void *p)
4000 {
4001         struct btrfs_iget_args *args = p;
4002         inode->i_ino = args->ino;
4003         BTRFS_I(inode)->root = args->root;
4004         btrfs_set_inode_space_info(args->root, inode);
4005         return 0;
4006 }
4007
4008 static int btrfs_find_actor(struct inode *inode, void *opaque)
4009 {
4010         struct btrfs_iget_args *args = opaque;
4011         return args->ino == inode->i_ino &&
4012                 args->root == BTRFS_I(inode)->root;
4013 }
4014
4015 static struct inode *btrfs_iget_locked(struct super_block *s,
4016                                        u64 objectid,
4017                                        struct btrfs_root *root)
4018 {
4019         struct inode *inode;
4020         struct btrfs_iget_args args;
4021         args.ino = objectid;
4022         args.root = root;
4023
4024         inode = iget5_locked(s, objectid, btrfs_find_actor,
4025                              btrfs_init_locked_inode,
4026                              (void *)&args);
4027         return inode;
4028 }
4029
4030 /* Get an inode object given its location and corresponding root.
4031  * Returns in *is_new if the inode was read from disk
4032  */
4033 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
4034                          struct btrfs_root *root, int *new)
4035 {
4036         struct inode *inode;
4037
4038         inode = btrfs_iget_locked(s, location->objectid, root);
4039         if (!inode)
4040                 return ERR_PTR(-ENOMEM);
4041
4042         if (inode->i_state & I_NEW) {
4043                 BTRFS_I(inode)->root = root;
4044                 memcpy(&BTRFS_I(inode)->location, location, sizeof(*location));
4045                 btrfs_read_locked_inode(inode);
4046
4047                 inode_tree_add(inode);
4048                 unlock_new_inode(inode);
4049                 if (new)
4050                         *new = 1;
4051         }
4052
4053         return inode;
4054 }
4055
4056 static struct inode *new_simple_dir(struct super_block *s,
4057                                     struct btrfs_key *key,
4058                                     struct btrfs_root *root)
4059 {
4060         struct inode *inode = new_inode(s);
4061
4062         if (!inode)
4063                 return ERR_PTR(-ENOMEM);
4064
4065         BTRFS_I(inode)->root = root;
4066         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
4067         BTRFS_I(inode)->dummy_inode = 1;
4068
4069         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
4070         inode->i_op = &simple_dir_inode_operations;
4071         inode->i_fop = &simple_dir_operations;
4072         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
4073         inode->i_mtime = inode->i_atime = inode->i_ctime = CURRENT_TIME;
4074
4075         return inode;
4076 }
4077
4078 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
4079 {
4080         struct inode *inode;
4081         struct btrfs_root *root = BTRFS_I(dir)->root;
4082         struct btrfs_root *sub_root = root;
4083         struct btrfs_key location;
4084         int index;
4085         int ret;
4086
4087         dentry->d_op = &btrfs_dentry_operations;
4088
4089         if (dentry->d_name.len > BTRFS_NAME_LEN)
4090                 return ERR_PTR(-ENAMETOOLONG);
4091
4092         ret = btrfs_inode_by_name(dir, dentry, &location);
4093
4094         if (ret < 0)
4095                 return ERR_PTR(ret);
4096
4097         if (location.objectid == 0)
4098                 return NULL;
4099
4100         if (location.type == BTRFS_INODE_ITEM_KEY) {
4101                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
4102                 return inode;
4103         }
4104
4105         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
4106
4107         index = srcu_read_lock(&root->fs_info->subvol_srcu);
4108         ret = fixup_tree_root_location(root, dir, dentry,
4109                                        &location, &sub_root);
4110         if (ret < 0) {
4111                 if (ret != -ENOENT)
4112                         inode = ERR_PTR(ret);
4113                 else
4114                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
4115         } else {
4116                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
4117         }
4118         srcu_read_unlock(&root->fs_info->subvol_srcu, index);
4119
4120         if (root != sub_root) {
4121                 down_read(&root->fs_info->cleanup_work_sem);
4122                 if (!(inode->i_sb->s_flags & MS_RDONLY))
4123                         btrfs_orphan_cleanup(sub_root);
4124                 up_read(&root->fs_info->cleanup_work_sem);
4125         }
4126
4127         return inode;
4128 }
4129
4130 static int btrfs_dentry_delete(struct dentry *dentry)
4131 {
4132         struct btrfs_root *root;
4133
4134         if (!dentry->d_inode && !IS_ROOT(dentry))
4135                 dentry = dentry->d_parent;
4136
4137         if (dentry->d_inode) {
4138                 root = BTRFS_I(dentry->d_inode)->root;
4139                 if (btrfs_root_refs(&root->root_item) == 0)
4140                         return 1;
4141         }
4142         return 0;
4143 }
4144
4145 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
4146                                    struct nameidata *nd)
4147 {
4148         struct inode *inode;
4149
4150         inode = btrfs_lookup_dentry(dir, dentry);
4151         if (IS_ERR(inode))
4152                 return ERR_CAST(inode);
4153
4154         return d_splice_alias(inode, dentry);
4155 }
4156
4157 static unsigned char btrfs_filetype_table[] = {
4158         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
4159 };
4160
4161 static int btrfs_real_readdir(struct file *filp, void *dirent,
4162                               filldir_t filldir)
4163 {
4164         struct inode *inode = filp->f_dentry->d_inode;
4165         struct btrfs_root *root = BTRFS_I(inode)->root;
4166         struct btrfs_item *item;
4167         struct btrfs_dir_item *di;
4168         struct btrfs_key key;
4169         struct btrfs_key found_key;
4170         struct btrfs_path *path;
4171         int ret;
4172         u32 nritems;
4173         struct extent_buffer *leaf;
4174         int slot;
4175         int advance;
4176         unsigned char d_type;
4177         int over = 0;
4178         u32 di_cur;
4179         u32 di_total;
4180         u32 di_len;
4181         int key_type = BTRFS_DIR_INDEX_KEY;
4182         char tmp_name[32];
4183         char *name_ptr;
4184         int name_len;
4185
4186         /* FIXME, use a real flag for deciding about the key type */
4187         if (root->fs_info->tree_root == root)
4188                 key_type = BTRFS_DIR_ITEM_KEY;
4189
4190         /* special case for "." */
4191         if (filp->f_pos == 0) {
4192                 over = filldir(dirent, ".", 1,
4193                                1, inode->i_ino,
4194                                DT_DIR);
4195                 if (over)
4196                         return 0;
4197                 filp->f_pos = 1;
4198         }
4199         /* special case for .., just use the back ref */
4200         if (filp->f_pos == 1) {
4201                 u64 pino = parent_ino(filp->f_path.dentry);
4202                 over = filldir(dirent, "..", 2,
4203                                2, pino, DT_DIR);
4204                 if (over)
4205                         return 0;
4206                 filp->f_pos = 2;
4207         }
4208         path = btrfs_alloc_path();
4209         path->reada = 2;
4210
4211         btrfs_set_key_type(&key, key_type);
4212         key.offset = filp->f_pos;
4213         key.objectid = inode->i_ino;
4214
4215         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4216         if (ret < 0)
4217                 goto err;
4218         advance = 0;
4219
4220         while (1) {
4221                 leaf = path->nodes[0];
4222                 nritems = btrfs_header_nritems(leaf);
4223                 slot = path->slots[0];
4224                 if (advance || slot >= nritems) {
4225                         if (slot >= nritems - 1) {
4226                                 ret = btrfs_next_leaf(root, path);
4227                                 if (ret)
4228                                         break;
4229                                 leaf = path->nodes[0];
4230                                 nritems = btrfs_header_nritems(leaf);
4231                                 slot = path->slots[0];
4232                         } else {
4233                                 slot++;
4234                                 path->slots[0]++;
4235                         }
4236                 }
4237
4238                 advance = 1;
4239                 item = btrfs_item_nr(leaf, slot);
4240                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
4241
4242                 if (found_key.objectid != key.objectid)
4243                         break;
4244                 if (btrfs_key_type(&found_key) != key_type)
4245                         break;
4246                 if (found_key.offset < filp->f_pos)
4247                         continue;
4248
4249                 filp->f_pos = found_key.offset;
4250
4251                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
4252                 di_cur = 0;
4253                 di_total = btrfs_item_size(leaf, item);
4254
4255                 while (di_cur < di_total) {
4256                         struct btrfs_key location;
4257
4258                         name_len = btrfs_dir_name_len(leaf, di);
4259                         if (name_len <= sizeof(tmp_name)) {
4260                                 name_ptr = tmp_name;
4261                         } else {
4262                                 name_ptr = kmalloc(name_len, GFP_NOFS);
4263                                 if (!name_ptr) {
4264                                         ret = -ENOMEM;
4265                                         goto err;
4266                                 }
4267                         }
4268                         read_extent_buffer(leaf, name_ptr,
4269                                            (unsigned long)(di + 1), name_len);
4270
4271                         d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
4272                         btrfs_dir_item_key_to_cpu(leaf, di, &location);
4273
4274                         /* is this a reference to our own snapshot? If so
4275                          * skip it
4276                          */
4277                         if (location.type == BTRFS_ROOT_ITEM_KEY &&
4278                             location.objectid == root->root_key.objectid) {
4279                                 over = 0;
4280                                 goto skip;
4281                         }
4282                         over = filldir(dirent, name_ptr, name_len,
4283                                        found_key.offset, location.objectid,
4284                                        d_type);
4285
4286 skip:
4287                         if (name_ptr != tmp_name)
4288                                 kfree(name_ptr);
4289
4290                         if (over)
4291                                 goto nopos;
4292                         di_len = btrfs_dir_name_len(leaf, di) +
4293                                  btrfs_dir_data_len(leaf, di) + sizeof(*di);
4294                         di_cur += di_len;
4295                         di = (struct btrfs_dir_item *)((char *)di + di_len);
4296                 }
4297         }
4298
4299         /* Reached end of directory/root. Bump pos past the last item. */
4300         if (key_type == BTRFS_DIR_INDEX_KEY)
4301                 /*
4302                  * 32-bit glibc will use getdents64, but then strtol -
4303                  * so the last number we can serve is this.
4304                  */
4305                 filp->f_pos = 0x7fffffff;
4306         else
4307                 filp->f_pos++;
4308 nopos:
4309         ret = 0;
4310 err:
4311         btrfs_free_path(path);
4312         return ret;
4313 }
4314
4315 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
4316 {
4317         struct btrfs_root *root = BTRFS_I(inode)->root;
4318         struct btrfs_trans_handle *trans;
4319         int ret = 0;
4320         bool nolock = false;
4321
4322         if (BTRFS_I(inode)->dummy_inode)
4323                 return 0;
4324
4325         smp_mb();
4326         nolock = (root->fs_info->closing && root == root->fs_info->tree_root);
4327
4328         if (wbc->sync_mode == WB_SYNC_ALL) {
4329                 if (nolock)
4330                         trans = btrfs_join_transaction_nolock(root, 1);
4331                 else
4332                         trans = btrfs_join_transaction(root, 1);
4333                 btrfs_set_trans_block_group(trans, inode);
4334                 if (nolock)
4335                         ret = btrfs_end_transaction_nolock(trans, root);
4336                 else
4337                         ret = btrfs_commit_transaction(trans, root);
4338         }
4339         return ret;
4340 }
4341
4342 /*
4343  * This is somewhat expensive, updating the tree every time the
4344  * inode changes.  But, it is most likely to find the inode in cache.
4345  * FIXME, needs more benchmarking...there are no reasons other than performance
4346  * to keep or drop this code.
4347  */
4348 void btrfs_dirty_inode(struct inode *inode)
4349 {
4350         struct btrfs_root *root = BTRFS_I(inode)->root;
4351         struct btrfs_trans_handle *trans;
4352         int ret;
4353
4354         if (BTRFS_I(inode)->dummy_inode)
4355                 return;
4356
4357         trans = btrfs_join_transaction(root, 1);
4358         btrfs_set_trans_block_group(trans, inode);
4359
4360         ret = btrfs_update_inode(trans, root, inode);
4361         if (ret && ret == -ENOSPC) {
4362                 /* whoops, lets try again with the full transaction */
4363                 btrfs_end_transaction(trans, root);
4364                 trans = btrfs_start_transaction(root, 1);
4365                 if (IS_ERR(trans)) {
4366                         if (printk_ratelimit()) {
4367                                 printk(KERN_ERR "btrfs: fail to "
4368                                        "dirty  inode %lu error %ld\n",
4369                                        inode->i_ino, PTR_ERR(trans));
4370                         }
4371                         return;
4372                 }
4373                 btrfs_set_trans_block_group(trans, inode);
4374
4375                 ret = btrfs_update_inode(trans, root, inode);
4376                 if (ret) {
4377                         if (printk_ratelimit()) {
4378                                 printk(KERN_ERR "btrfs: fail to "
4379                                        "dirty  inode %lu error %d\n",
4380                                        inode->i_ino, ret);
4381                         }
4382                 }
4383         }
4384         btrfs_end_transaction(trans, root);
4385 }
4386
4387 /*
4388  * find the highest existing sequence number in a directory
4389  * and then set the in-memory index_cnt variable to reflect
4390  * free sequence numbers
4391  */
4392 static int btrfs_set_inode_index_count(struct inode *inode)
4393 {
4394         struct btrfs_root *root = BTRFS_I(inode)->root;
4395         struct btrfs_key key, found_key;
4396         struct btrfs_path *path;
4397         struct extent_buffer *leaf;
4398         int ret;
4399
4400         key.objectid = inode->i_ino;
4401         btrfs_set_key_type(&key, BTRFS_DIR_INDEX_KEY);
4402         key.offset = (u64)-1;
4403
4404         path = btrfs_alloc_path();
4405         if (!path)
4406                 return -ENOMEM;
4407
4408         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4409         if (ret < 0)
4410                 goto out;
4411         /* FIXME: we should be able to handle this */
4412         if (ret == 0)
4413                 goto out;
4414         ret = 0;
4415
4416         /*
4417          * MAGIC NUMBER EXPLANATION:
4418          * since we search a directory based on f_pos we have to start at 2
4419          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
4420          * else has to start at 2
4421          */
4422         if (path->slots[0] == 0) {
4423                 BTRFS_I(inode)->index_cnt = 2;
4424                 goto out;
4425         }
4426
4427         path->slots[0]--;
4428
4429         leaf = path->nodes[0];
4430         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4431
4432         if (found_key.objectid != inode->i_ino ||
4433             btrfs_key_type(&found_key) != BTRFS_DIR_INDEX_KEY) {
4434                 BTRFS_I(inode)->index_cnt = 2;
4435                 goto out;
4436         }
4437
4438         BTRFS_I(inode)->index_cnt = found_key.offset + 1;
4439 out:
4440         btrfs_free_path(path);
4441         return ret;
4442 }
4443
4444 /*
4445  * helper to find a free sequence number in a given directory.  This current
4446  * code is very simple, later versions will do smarter things in the btree
4447  */
4448 int btrfs_set_inode_index(struct inode *dir, u64 *index)
4449 {
4450         int ret = 0;
4451
4452         if (BTRFS_I(dir)->index_cnt == (u64)-1) {
4453                 ret = btrfs_set_inode_index_count(dir);
4454                 if (ret)
4455                         return ret;
4456         }
4457
4458         *index = BTRFS_I(dir)->index_cnt;
4459         BTRFS_I(dir)->index_cnt++;
4460
4461         return ret;
4462 }
4463
4464 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
4465                                      struct btrfs_root *root,
4466                                      struct inode *dir,
4467                                      const char *name, int name_len,
4468                                      u64 ref_objectid, u64 objectid,
4469                                      u64 alloc_hint, int mode, u64 *index)
4470 {
4471         struct inode *inode;
4472         struct btrfs_inode_item *inode_item;
4473         struct btrfs_key *location;
4474         struct btrfs_path *path;
4475         struct btrfs_inode_ref *ref;
4476         struct btrfs_key key[2];
4477         u32 sizes[2];
4478         unsigned long ptr;
4479         int ret;
4480         int owner;
4481
4482         path = btrfs_alloc_path();
4483         BUG_ON(!path);
4484
4485         inode = new_inode(root->fs_info->sb);
4486         if (!inode)
4487                 return ERR_PTR(-ENOMEM);
4488
4489         if (dir) {
4490                 ret = btrfs_set_inode_index(dir, index);
4491                 if (ret) {
4492                         iput(inode);
4493                         return ERR_PTR(ret);
4494                 }
4495         }
4496         /*
4497          * index_cnt is ignored for everything but a dir,
4498          * btrfs_get_inode_index_count has an explanation for the magic
4499          * number
4500          */
4501         BTRFS_I(inode)->index_cnt = 2;
4502         BTRFS_I(inode)->root = root;
4503         BTRFS_I(inode)->generation = trans->transid;
4504         inode->i_generation = BTRFS_I(inode)->generation;
4505         btrfs_set_inode_space_info(root, inode);
4506
4507         if (mode & S_IFDIR)
4508                 owner = 0;
4509         else
4510                 owner = 1;
4511         BTRFS_I(inode)->block_group =
4512                         btrfs_find_block_group(root, 0, alloc_hint, owner);
4513
4514         key[0].objectid = objectid;
4515         btrfs_set_key_type(&key[0], BTRFS_INODE_ITEM_KEY);
4516         key[0].offset = 0;
4517
4518         key[1].objectid = objectid;
4519         btrfs_set_key_type(&key[1], BTRFS_INODE_REF_KEY);
4520         key[1].offset = ref_objectid;
4521
4522         sizes[0] = sizeof(struct btrfs_inode_item);
4523         sizes[1] = name_len + sizeof(*ref);
4524
4525         path->leave_spinning = 1;
4526         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, 2);
4527         if (ret != 0)
4528                 goto fail;
4529
4530         inode_init_owner(inode, dir, mode);
4531         inode->i_ino = objectid;
4532         inode_set_bytes(inode, 0);
4533         inode->i_mtime = inode->i_atime = inode->i_ctime = CURRENT_TIME;
4534         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
4535                                   struct btrfs_inode_item);
4536         fill_inode_item(trans, path->nodes[0], inode_item, inode);
4537
4538         ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
4539                              struct btrfs_inode_ref);
4540         btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
4541         btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
4542         ptr = (unsigned long)(ref + 1);
4543         write_extent_buffer(path->nodes[0], name, ptr, name_len);
4544
4545         btrfs_mark_buffer_dirty(path->nodes[0]);
4546         btrfs_free_path(path);
4547
4548         location = &BTRFS_I(inode)->location;
4549         location->objectid = objectid;
4550         location->offset = 0;
4551         btrfs_set_key_type(location, BTRFS_INODE_ITEM_KEY);
4552
4553         btrfs_inherit_iflags(inode, dir);
4554
4555         if ((mode & S_IFREG)) {
4556                 if (btrfs_test_opt(root, NODATASUM))
4557                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
4558                 if (btrfs_test_opt(root, NODATACOW))
4559                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
4560         }
4561
4562         insert_inode_hash(inode);
4563         inode_tree_add(inode);
4564         return inode;
4565 fail:
4566         if (dir)
4567                 BTRFS_I(dir)->index_cnt--;
4568         btrfs_free_path(path);
4569         iput(inode);
4570         return ERR_PTR(ret);
4571 }
4572
4573 static inline u8 btrfs_inode_type(struct inode *inode)
4574 {
4575         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
4576 }
4577
4578 /*
4579  * utility function to add 'inode' into 'parent_inode' with
4580  * a give name and a given sequence number.
4581  * if 'add_backref' is true, also insert a backref from the
4582  * inode to the parent directory.
4583  */
4584 int btrfs_add_link(struct btrfs_trans_handle *trans,
4585                    struct inode *parent_inode, struct inode *inode,
4586                    const char *name, int name_len, int add_backref, u64 index)
4587 {
4588         int ret = 0;
4589         struct btrfs_key key;
4590         struct btrfs_root *root = BTRFS_I(parent_inode)->root;
4591
4592         if (unlikely(inode->i_ino == BTRFS_FIRST_FREE_OBJECTID)) {
4593                 memcpy(&key, &BTRFS_I(inode)->root->root_key, sizeof(key));
4594         } else {
4595                 key.objectid = inode->i_ino;
4596                 btrfs_set_key_type(&key, BTRFS_INODE_ITEM_KEY);
4597                 key.offset = 0;
4598         }
4599
4600         if (unlikely(inode->i_ino == BTRFS_FIRST_FREE_OBJECTID)) {
4601                 ret = btrfs_add_root_ref(trans, root->fs_info->tree_root,
4602                                          key.objectid, root->root_key.objectid,
4603                                          parent_inode->i_ino,
4604                                          index, name, name_len);
4605         } else if (add_backref) {
4606                 ret = btrfs_insert_inode_ref(trans, root,
4607                                              name, name_len, inode->i_ino,
4608                                              parent_inode->i_ino, index);
4609         }
4610
4611         if (ret == 0) {
4612                 ret = btrfs_insert_dir_item(trans, root, name, name_len,
4613                                             parent_inode->i_ino, &key,
4614                                             btrfs_inode_type(inode), index);
4615                 BUG_ON(ret);
4616
4617                 btrfs_i_size_write(parent_inode, parent_inode->i_size +
4618                                    name_len * 2);
4619                 parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
4620                 ret = btrfs_update_inode(trans, root, parent_inode);
4621         }
4622         return ret;
4623 }
4624
4625 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
4626                             struct inode *dir, struct dentry *dentry,
4627                             struct inode *inode, int backref, u64 index)
4628 {
4629         int err = btrfs_add_link(trans, dir, inode,
4630                                  dentry->d_name.name, dentry->d_name.len,
4631                                  backref, index);
4632         if (!err) {
4633                 d_instantiate(dentry, inode);
4634                 return 0;
4635         }
4636         if (err > 0)
4637                 err = -EEXIST;
4638         return err;
4639 }
4640
4641 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
4642                         int mode, dev_t rdev)
4643 {
4644         struct btrfs_trans_handle *trans;
4645         struct btrfs_root *root = BTRFS_I(dir)->root;
4646         struct inode *inode = NULL;
4647         int err;
4648         int drop_inode = 0;
4649         u64 objectid;
4650         unsigned long nr = 0;
4651         u64 index = 0;
4652
4653         if (!new_valid_dev(rdev))
4654                 return -EINVAL;
4655
4656         err = btrfs_find_free_objectid(NULL, root, dir->i_ino, &objectid);
4657         if (err)
4658                 return err;
4659
4660         /*
4661          * 2 for inode item and ref
4662          * 2 for dir items
4663          * 1 for xattr if selinux is on
4664          */
4665         trans = btrfs_start_transaction(root, 5);
4666         if (IS_ERR(trans))
4667                 return PTR_ERR(trans);
4668
4669         btrfs_set_trans_block_group(trans, dir);
4670
4671         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
4672                                 dentry->d_name.len, dir->i_ino, objectid,
4673                                 BTRFS_I(dir)->block_group, mode, &index);
4674         err = PTR_ERR(inode);
4675         if (IS_ERR(inode))
4676                 goto out_unlock;
4677
4678         err = btrfs_init_inode_security(trans, inode, dir);
4679         if (err) {
4680                 drop_inode = 1;
4681                 goto out_unlock;
4682         }
4683
4684         btrfs_set_trans_block_group(trans, inode);
4685         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
4686         if (err)
4687                 drop_inode = 1;
4688         else {
4689                 inode->i_op = &btrfs_special_inode_operations;
4690                 init_special_inode(inode, inode->i_mode, rdev);
4691                 btrfs_update_inode(trans, root, inode);
4692         }
4693         btrfs_update_inode_block_group(trans, inode);
4694         btrfs_update_inode_block_group(trans, dir);
4695 out_unlock:
4696         nr = trans->blocks_used;
4697         btrfs_end_transaction_throttle(trans, root);
4698         btrfs_btree_balance_dirty(root, nr);
4699         if (drop_inode) {
4700                 inode_dec_link_count(inode);
4701                 iput(inode);
4702         }
4703         return err;
4704 }
4705
4706 static int btrfs_create(struct inode *dir, struct dentry *dentry,
4707                         int mode, struct nameidata *nd)
4708 {
4709         struct btrfs_trans_handle *trans;
4710         struct btrfs_root *root = BTRFS_I(dir)->root;
4711         struct inode *inode = NULL;
4712         int drop_inode = 0;
4713         int err;
4714         unsigned long nr = 0;
4715         u64 objectid;
4716         u64 index = 0;
4717
4718         err = btrfs_find_free_objectid(NULL, root, dir->i_ino, &objectid);
4719         if (err)
4720                 return err;
4721         /*
4722          * 2 for inode item and ref
4723          * 2 for dir items
4724          * 1 for xattr if selinux is on
4725          */
4726         trans = btrfs_start_transaction(root, 5);
4727         if (IS_ERR(trans))
4728                 return PTR_ERR(trans);
4729
4730         btrfs_set_trans_block_group(trans, dir);
4731
4732         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
4733                                 dentry->d_name.len, dir->i_ino, objectid,
4734                                 BTRFS_I(dir)->block_group, mode, &index);
4735         err = PTR_ERR(inode);
4736         if (IS_ERR(inode))
4737                 goto out_unlock;
4738
4739         err = btrfs_init_inode_security(trans, inode, dir);
4740         if (err) {
4741                 drop_inode = 1;
4742                 goto out_unlock;
4743         }
4744
4745         btrfs_set_trans_block_group(trans, inode);
4746         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
4747         if (err)
4748                 drop_inode = 1;
4749         else {
4750                 inode->i_mapping->a_ops = &btrfs_aops;
4751                 inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
4752                 inode->i_fop = &btrfs_file_operations;
4753                 inode->i_op = &btrfs_file_inode_operations;
4754                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
4755         }
4756         btrfs_update_inode_block_group(trans, inode);
4757         btrfs_update_inode_block_group(trans, dir);
4758 out_unlock:
4759         nr = trans->blocks_used;
4760         btrfs_end_transaction_throttle(trans, root);
4761         if (drop_inode) {
4762                 inode_dec_link_count(inode);
4763                 iput(inode);
4764         }
4765         btrfs_btree_balance_dirty(root, nr);
4766         return err;
4767 }
4768
4769 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
4770                       struct dentry *dentry)
4771 {
4772         struct btrfs_trans_handle *trans;
4773         struct btrfs_root *root = BTRFS_I(dir)->root;
4774         struct inode *inode = old_dentry->d_inode;
4775         u64 index;
4776         unsigned long nr = 0;
4777         int err;
4778         int drop_inode = 0;
4779
4780         if (inode->i_nlink == 0)
4781                 return -ENOENT;
4782
4783         /* do not allow sys_link's with other subvols of the same device */
4784         if (root->objectid != BTRFS_I(inode)->root->objectid)
4785                 return -EPERM;
4786
4787         btrfs_inc_nlink(inode);
4788         inode->i_ctime = CURRENT_TIME;
4789
4790         err = btrfs_set_inode_index(dir, &index);
4791         if (err)
4792                 goto fail;
4793
4794         /*
4795          * 1 item for inode ref
4796          * 2 items for dir items
4797          */
4798         trans = btrfs_start_transaction(root, 3);
4799         if (IS_ERR(trans)) {
4800                 err = PTR_ERR(trans);
4801                 goto fail;
4802         }
4803
4804         btrfs_set_trans_block_group(trans, dir);
4805         atomic_inc(&inode->i_count);
4806
4807         err = btrfs_add_nondir(trans, dir, dentry, inode, 1, index);
4808
4809         if (err) {
4810                 drop_inode = 1;
4811         } else {
4812                 struct dentry *parent = dget_parent(dentry);
4813                 btrfs_update_inode_block_group(trans, dir);
4814                 err = btrfs_update_inode(trans, root, inode);
4815                 BUG_ON(err);
4816                 btrfs_log_new_name(trans, inode, NULL, parent);
4817                 dput(parent);
4818         }
4819
4820         nr = trans->blocks_used;
4821         btrfs_end_transaction_throttle(trans, root);
4822 fail:
4823         if (drop_inode) {
4824                 inode_dec_link_count(inode);
4825                 iput(inode);
4826         }
4827         btrfs_btree_balance_dirty(root, nr);
4828         return err;
4829 }
4830
4831 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, int mode)
4832 {
4833         struct inode *inode = NULL;
4834         struct btrfs_trans_handle *trans;
4835         struct btrfs_root *root = BTRFS_I(dir)->root;
4836         int err = 0;
4837         int drop_on_err = 0;
4838         u64 objectid = 0;
4839         u64 index = 0;
4840         unsigned long nr = 1;
4841
4842         err = btrfs_find_free_objectid(NULL, root, dir->i_ino, &objectid);
4843         if (err)
4844                 return err;
4845
4846         /*
4847          * 2 items for inode and ref
4848          * 2 items for dir items
4849          * 1 for xattr if selinux is on
4850          */
4851         trans = btrfs_start_transaction(root, 5);
4852         if (IS_ERR(trans))
4853                 return PTR_ERR(trans);
4854         btrfs_set_trans_block_group(trans, dir);
4855
4856         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
4857                                 dentry->d_name.len, dir->i_ino, objectid,
4858                                 BTRFS_I(dir)->block_group, S_IFDIR | mode,
4859                                 &index);
4860         if (IS_ERR(inode)) {
4861                 err = PTR_ERR(inode);
4862                 goto out_fail;
4863         }
4864
4865         drop_on_err = 1;
4866
4867         err = btrfs_init_inode_security(trans, inode, dir);
4868         if (err)
4869                 goto out_fail;
4870
4871         inode->i_op = &btrfs_dir_inode_operations;
4872         inode->i_fop = &btrfs_dir_file_operations;
4873         btrfs_set_trans_block_group(trans, inode);
4874
4875         btrfs_i_size_write(inode, 0);
4876         err = btrfs_update_inode(trans, root, inode);
4877         if (err)
4878                 goto out_fail;
4879
4880         err = btrfs_add_link(trans, dir, inode, dentry->d_name.name,
4881                              dentry->d_name.len, 0, index);
4882         if (err)
4883                 goto out_fail;
4884
4885         d_instantiate(dentry, inode);
4886         drop_on_err = 0;
4887         btrfs_update_inode_block_group(trans, inode);
4888         btrfs_update_inode_block_group(trans, dir);
4889
4890 out_fail:
4891         nr = trans->blocks_used;
4892         btrfs_end_transaction_throttle(trans, root);
4893         if (drop_on_err)
4894                 iput(inode);
4895         btrfs_btree_balance_dirty(root, nr);
4896         return err;
4897 }
4898
4899 /* helper for btfs_get_extent.  Given an existing extent in the tree,
4900  * and an extent that you want to insert, deal with overlap and insert
4901  * the new extent into the tree.
4902  */
4903 static int merge_extent_mapping(struct extent_map_tree *em_tree,
4904                                 struct extent_map *existing,
4905                                 struct extent_map *em,
4906                                 u64 map_start, u64 map_len)
4907 {
4908         u64 start_diff;
4909
4910         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
4911         start_diff = map_start - em->start;
4912         em->start = map_start;
4913         em->len = map_len;
4914         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
4915             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
4916                 em->block_start += start_diff;
4917                 em->block_len -= start_diff;
4918         }
4919         return add_extent_mapping(em_tree, em);
4920 }
4921
4922 static noinline int uncompress_inline(struct btrfs_path *path,
4923                                       struct inode *inode, struct page *page,
4924                                       size_t pg_offset, u64 extent_offset,
4925                                       struct btrfs_file_extent_item *item)
4926 {
4927         int ret;
4928         struct extent_buffer *leaf = path->nodes[0];
4929         char *tmp;
4930         size_t max_size;
4931         unsigned long inline_size;
4932         unsigned long ptr;
4933
4934         WARN_ON(pg_offset != 0);
4935         max_size = btrfs_file_extent_ram_bytes(leaf, item);
4936         inline_size = btrfs_file_extent_inline_item_len(leaf,
4937                                         btrfs_item_nr(leaf, path->slots[0]));
4938         tmp = kmalloc(inline_size, GFP_NOFS);
4939         ptr = btrfs_file_extent_inline_start(item);
4940
4941         read_extent_buffer(leaf, tmp, ptr, inline_size);
4942
4943         max_size = min_t(unsigned long, PAGE_CACHE_SIZE, max_size);
4944         ret = btrfs_zlib_decompress(tmp, page, extent_offset,
4945                                     inline_size, max_size);
4946         if (ret) {
4947                 char *kaddr = kmap_atomic(page, KM_USER0);
4948                 unsigned long copy_size = min_t(u64,
4949                                   PAGE_CACHE_SIZE - pg_offset,
4950                                   max_size - extent_offset);
4951                 memset(kaddr + pg_offset, 0, copy_size);
4952                 kunmap_atomic(kaddr, KM_USER0);
4953         }
4954         kfree(tmp);
4955         return 0;
4956 }
4957
4958 /*
4959  * a bit scary, this does extent mapping from logical file offset to the disk.
4960  * the ugly parts come from merging extents from the disk with the in-ram
4961  * representation.  This gets more complex because of the data=ordered code,
4962  * where the in-ram extents might be locked pending data=ordered completion.
4963  *
4964  * This also copies inline extents directly into the page.
4965  */
4966
4967 struct extent_map *btrfs_get_extent(struct inode *inode, struct page *page,
4968                                     size_t pg_offset, u64 start, u64 len,
4969                                     int create)
4970 {
4971         int ret;
4972         int err = 0;
4973         u64 bytenr;
4974         u64 extent_start = 0;
4975         u64 extent_end = 0;
4976         u64 objectid = inode->i_ino;
4977         u32 found_type;
4978         struct btrfs_path *path = NULL;
4979         struct btrfs_root *root = BTRFS_I(inode)->root;
4980         struct btrfs_file_extent_item *item;
4981         struct extent_buffer *leaf;
4982         struct btrfs_key found_key;
4983         struct extent_map *em = NULL;
4984         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4985         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4986         struct btrfs_trans_handle *trans = NULL;
4987         int compressed;
4988
4989 again:
4990         read_lock(&em_tree->lock);
4991         em = lookup_extent_mapping(em_tree, start, len);
4992         if (em)
4993                 em->bdev = root->fs_info->fs_devices->latest_bdev;
4994         read_unlock(&em_tree->lock);
4995
4996         if (em) {
4997                 if (em->start > start || em->start + em->len <= start)
4998                         free_extent_map(em);
4999                 else if (em->block_start == EXTENT_MAP_INLINE && page)
5000                         free_extent_map(em);
5001                 else
5002                         goto out;
5003         }
5004         em = alloc_extent_map(GFP_NOFS);
5005         if (!em) {
5006                 err = -ENOMEM;
5007                 goto out;
5008         }
5009         em->bdev = root->fs_info->fs_devices->latest_bdev;
5010         em->start = EXTENT_MAP_HOLE;
5011         em->orig_start = EXTENT_MAP_HOLE;
5012         em->len = (u64)-1;
5013         em->block_len = (u64)-1;
5014
5015         if (!path) {
5016                 path = btrfs_alloc_path();
5017                 BUG_ON(!path);
5018         }
5019
5020         ret = btrfs_lookup_file_extent(trans, root, path,
5021                                        objectid, start, trans != NULL);
5022         if (ret < 0) {
5023                 err = ret;
5024                 goto out;
5025         }
5026
5027         if (ret != 0) {
5028                 if (path->slots[0] == 0)
5029                         goto not_found;
5030                 path->slots[0]--;
5031         }
5032
5033         leaf = path->nodes[0];
5034         item = btrfs_item_ptr(leaf, path->slots[0],
5035                               struct btrfs_file_extent_item);
5036         /* are we inside the extent that was found? */
5037         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5038         found_type = btrfs_key_type(&found_key);
5039         if (found_key.objectid != objectid ||
5040             found_type != BTRFS_EXTENT_DATA_KEY) {
5041                 goto not_found;
5042         }
5043
5044         found_type = btrfs_file_extent_type(leaf, item);
5045         extent_start = found_key.offset;
5046         compressed = btrfs_file_extent_compression(leaf, item);
5047         if (found_type == BTRFS_FILE_EXTENT_REG ||
5048             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
5049                 extent_end = extent_start +
5050                        btrfs_file_extent_num_bytes(leaf, item);
5051         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
5052                 size_t size;
5053                 size = btrfs_file_extent_inline_len(leaf, item);
5054                 extent_end = (extent_start + size + root->sectorsize - 1) &
5055                         ~((u64)root->sectorsize - 1);
5056         }
5057
5058         if (start >= extent_end) {
5059                 path->slots[0]++;
5060                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
5061                         ret = btrfs_next_leaf(root, path);
5062                         if (ret < 0) {
5063                                 err = ret;
5064                                 goto out;
5065                         }
5066                         if (ret > 0)
5067                                 goto not_found;
5068                         leaf = path->nodes[0];
5069                 }
5070                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5071                 if (found_key.objectid != objectid ||
5072                     found_key.type != BTRFS_EXTENT_DATA_KEY)
5073                         goto not_found;
5074                 if (start + len <= found_key.offset)
5075                         goto not_found;
5076                 em->start = start;
5077                 em->len = found_key.offset - start;
5078                 goto not_found_em;
5079         }
5080
5081         if (found_type == BTRFS_FILE_EXTENT_REG ||
5082             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
5083                 em->start = extent_start;
5084                 em->len = extent_end - extent_start;
5085                 em->orig_start = extent_start -
5086                                  btrfs_file_extent_offset(leaf, item);
5087                 bytenr = btrfs_file_extent_disk_bytenr(leaf, item);
5088                 if (bytenr == 0) {
5089                         em->block_start = EXTENT_MAP_HOLE;
5090                         goto insert;
5091                 }
5092                 if (compressed) {
5093                         set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
5094                         em->block_start = bytenr;
5095                         em->block_len = btrfs_file_extent_disk_num_bytes(leaf,
5096                                                                          item);
5097                 } else {
5098                         bytenr += btrfs_file_extent_offset(leaf, item);
5099                         em->block_start = bytenr;
5100                         em->block_len = em->len;
5101                         if (found_type == BTRFS_FILE_EXTENT_PREALLOC)
5102                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
5103                 }
5104                 goto insert;
5105         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
5106                 unsigned long ptr;
5107                 char *map;
5108                 size_t size;
5109                 size_t extent_offset;
5110                 size_t copy_size;
5111
5112                 em->block_start = EXTENT_MAP_INLINE;
5113                 if (!page || create) {
5114                         em->start = extent_start;
5115                         em->len = extent_end - extent_start;
5116                         goto out;
5117                 }
5118
5119                 size = btrfs_file_extent_inline_len(leaf, item);
5120                 extent_offset = page_offset(page) + pg_offset - extent_start;
5121                 copy_size = min_t(u64, PAGE_CACHE_SIZE - pg_offset,
5122                                 size - extent_offset);
5123                 em->start = extent_start + extent_offset;
5124                 em->len = (copy_size + root->sectorsize - 1) &
5125                         ~((u64)root->sectorsize - 1);
5126                 em->orig_start = EXTENT_MAP_INLINE;
5127                 if (compressed)
5128                         set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
5129                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
5130                 if (create == 0 && !PageUptodate(page)) {
5131                         if (btrfs_file_extent_compression(leaf, item) ==
5132                             BTRFS_COMPRESS_ZLIB) {
5133                                 ret = uncompress_inline(path, inode, page,
5134                                                         pg_offset,
5135                                                         extent_offset, item);
5136                                 BUG_ON(ret);
5137                         } else {
5138                                 map = kmap(page);
5139                                 read_extent_buffer(leaf, map + pg_offset, ptr,
5140                                                    copy_size);
5141                                 if (pg_offset + copy_size < PAGE_CACHE_SIZE) {
5142                                         memset(map + pg_offset + copy_size, 0,
5143                                                PAGE_CACHE_SIZE - pg_offset -
5144                                                copy_size);
5145                                 }
5146                                 kunmap(page);
5147                         }
5148                         flush_dcache_page(page);
5149                 } else if (create && PageUptodate(page)) {
5150                         WARN_ON(1);
5151                         if (!trans) {
5152                                 kunmap(page);
5153                                 free_extent_map(em);
5154                                 em = NULL;
5155                                 btrfs_release_path(root, path);
5156                                 trans = btrfs_join_transaction(root, 1);
5157                                 goto again;
5158                         }
5159                         map = kmap(page);
5160                         write_extent_buffer(leaf, map + pg_offset, ptr,
5161                                             copy_size);
5162                         kunmap(page);
5163                         btrfs_mark_buffer_dirty(leaf);
5164                 }
5165                 set_extent_uptodate(io_tree, em->start,
5166                                     extent_map_end(em) - 1, GFP_NOFS);
5167                 goto insert;
5168         } else {
5169                 printk(KERN_ERR "btrfs unknown found_type %d\n", found_type);
5170                 WARN_ON(1);
5171         }
5172 not_found:
5173         em->start = start;
5174         em->len = len;
5175 not_found_em:
5176         em->block_start = EXTENT_MAP_HOLE;
5177         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
5178 insert:
5179         btrfs_release_path(root, path);
5180         if (em->start > start || extent_map_end(em) <= start) {
5181                 printk(KERN_ERR "Btrfs: bad extent! em: [%llu %llu] passed "
5182                        "[%llu %llu]\n", (unsigned long long)em->start,
5183                        (unsigned long long)em->len,
5184                        (unsigned long long)start,
5185                        (unsigned long long)len);
5186                 err = -EIO;
5187                 goto out;
5188         }
5189
5190         err = 0;
5191         write_lock(&em_tree->lock);
5192         ret = add_extent_mapping(em_tree, em);
5193         /* it is possible that someone inserted the extent into the tree
5194          * while we had the lock dropped.  It is also possible that
5195          * an overlapping map exists in the tree
5196          */
5197         if (ret == -EEXIST) {
5198                 struct extent_map *existing;
5199
5200                 ret = 0;
5201
5202                 existing = lookup_extent_mapping(em_tree, start, len);
5203                 if (existing && (existing->start > start ||
5204                     existing->start + existing->len <= start)) {
5205                         free_extent_map(existing);
5206                         existing = NULL;
5207                 }
5208                 if (!existing) {
5209                         existing = lookup_extent_mapping(em_tree, em->start,
5210                                                          em->len);
5211                         if (existing) {
5212                                 err = merge_extent_mapping(em_tree, existing,
5213                                                            em, start,
5214                                                            root->sectorsize);
5215                                 free_extent_map(existing);
5216                                 if (err) {
5217                                         free_extent_map(em);
5218                                         em = NULL;
5219                                 }
5220                         } else {
5221                                 err = -EIO;
5222                                 free_extent_map(em);
5223                                 em = NULL;
5224                         }
5225                 } else {
5226                         free_extent_map(em);
5227                         em = existing;
5228                         err = 0;
5229                 }
5230         }
5231         write_unlock(&em_tree->lock);
5232 out:
5233         if (path)
5234                 btrfs_free_path(path);
5235         if (trans) {
5236                 ret = btrfs_end_transaction(trans, root);
5237                 if (!err)
5238                         err = ret;
5239         }
5240         if (err) {
5241                 free_extent_map(em);
5242                 return ERR_PTR(err);
5243         }
5244         return em;
5245 }
5246
5247 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
5248                                                   u64 start, u64 len)
5249 {
5250         struct btrfs_root *root = BTRFS_I(inode)->root;
5251         struct btrfs_trans_handle *trans;
5252         struct extent_map *em;
5253         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
5254         struct btrfs_key ins;
5255         u64 alloc_hint;
5256         int ret;
5257
5258         btrfs_drop_extent_cache(inode, start, start + len - 1, 0);
5259
5260         trans = btrfs_join_transaction(root, 0);
5261         if (!trans)
5262                 return ERR_PTR(-ENOMEM);
5263
5264         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
5265
5266         alloc_hint = get_extent_allocation_hint(inode, start, len);
5267         ret = btrfs_reserve_extent(trans, root, len, root->sectorsize, 0,
5268                                    alloc_hint, (u64)-1, &ins, 1);
5269         if (ret) {
5270                 em = ERR_PTR(ret);
5271                 goto out;
5272         }
5273
5274         em = alloc_extent_map(GFP_NOFS);
5275         if (!em) {
5276                 em = ERR_PTR(-ENOMEM);
5277                 goto out;
5278         }
5279
5280         em->start = start;
5281         em->orig_start = em->start;
5282         em->len = ins.offset;
5283
5284         em->block_start = ins.objectid;
5285         em->block_len = ins.offset;
5286         em->bdev = root->fs_info->fs_devices->latest_bdev;
5287         set_bit(EXTENT_FLAG_PINNED, &em->flags);
5288
5289         while (1) {
5290                 write_lock(&em_tree->lock);
5291                 ret = add_extent_mapping(em_tree, em);
5292                 write_unlock(&em_tree->lock);
5293                 if (ret != -EEXIST)
5294                         break;
5295                 btrfs_drop_extent_cache(inode, start, start + em->len - 1, 0);
5296         }
5297
5298         ret = btrfs_add_ordered_extent_dio(inode, start, ins.objectid,
5299                                            ins.offset, ins.offset, 0);
5300         if (ret) {
5301                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset);
5302                 em = ERR_PTR(ret);
5303         }
5304 out:
5305         btrfs_end_transaction(trans, root);
5306         return em;
5307 }
5308
5309 /*
5310  * returns 1 when the nocow is safe, < 1 on error, 0 if the
5311  * block must be cow'd
5312  */
5313 static noinline int can_nocow_odirect(struct btrfs_trans_handle *trans,
5314                                       struct inode *inode, u64 offset, u64 len)
5315 {
5316         struct btrfs_path *path;
5317         int ret;
5318         struct extent_buffer *leaf;
5319         struct btrfs_root *root = BTRFS_I(inode)->root;
5320         struct btrfs_file_extent_item *fi;
5321         struct btrfs_key key;
5322         u64 disk_bytenr;
5323         u64 backref_offset;
5324         u64 extent_end;
5325         u64 num_bytes;
5326         int slot;
5327         int found_type;
5328
5329         path = btrfs_alloc_path();
5330         if (!path)
5331                 return -ENOMEM;
5332
5333         ret = btrfs_lookup_file_extent(trans, root, path, inode->i_ino,
5334                                        offset, 0);
5335         if (ret < 0)
5336                 goto out;
5337
5338         slot = path->slots[0];
5339         if (ret == 1) {
5340                 if (slot == 0) {
5341                         /* can't find the item, must cow */
5342                         ret = 0;
5343                         goto out;
5344                 }
5345                 slot--;
5346         }
5347         ret = 0;
5348         leaf = path->nodes[0];
5349         btrfs_item_key_to_cpu(leaf, &key, slot);
5350         if (key.objectid != inode->i_ino ||
5351             key.type != BTRFS_EXTENT_DATA_KEY) {
5352                 /* not our file or wrong item type, must cow */
5353                 goto out;
5354         }
5355
5356         if (key.offset > offset) {
5357                 /* Wrong offset, must cow */
5358                 goto out;
5359         }
5360
5361         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
5362         found_type = btrfs_file_extent_type(leaf, fi);
5363         if (found_type != BTRFS_FILE_EXTENT_REG &&
5364             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
5365                 /* not a regular extent, must cow */
5366                 goto out;
5367         }
5368         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
5369         backref_offset = btrfs_file_extent_offset(leaf, fi);
5370
5371         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
5372         if (extent_end < offset + len) {
5373                 /* extent doesn't include our full range, must cow */
5374                 goto out;
5375         }
5376
5377         if (btrfs_extent_readonly(root, disk_bytenr))
5378                 goto out;
5379
5380         /*
5381          * look for other files referencing this extent, if we
5382          * find any we must cow
5383          */
5384         if (btrfs_cross_ref_exist(trans, root, inode->i_ino,
5385                                   key.offset - backref_offset, disk_bytenr))
5386                 goto out;
5387
5388         /*
5389          * adjust disk_bytenr and num_bytes to cover just the bytes
5390          * in this extent we are about to write.  If there
5391          * are any csums in that range we have to cow in order
5392          * to keep the csums correct
5393          */
5394         disk_bytenr += backref_offset;
5395         disk_bytenr += offset - key.offset;
5396         num_bytes = min(offset + len, extent_end) - offset;
5397         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
5398                                 goto out;
5399         /*
5400          * all of the above have passed, it is safe to overwrite this extent
5401          * without cow
5402          */
5403         ret = 1;
5404 out:
5405         btrfs_free_path(path);
5406         return ret;
5407 }
5408
5409 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
5410                                    struct buffer_head *bh_result, int create)
5411 {
5412         struct extent_map *em;
5413         struct btrfs_root *root = BTRFS_I(inode)->root;
5414         u64 start = iblock << inode->i_blkbits;
5415         u64 len = bh_result->b_size;
5416         struct btrfs_trans_handle *trans;
5417
5418         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
5419         if (IS_ERR(em))
5420                 return PTR_ERR(em);
5421
5422         /*
5423          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
5424          * io.  INLINE is special, and we could probably kludge it in here, but
5425          * it's still buffered so for safety lets just fall back to the generic
5426          * buffered path.
5427          *
5428          * For COMPRESSED we _have_ to read the entire extent in so we can
5429          * decompress it, so there will be buffering required no matter what we
5430          * do, so go ahead and fallback to buffered.
5431          *
5432          * We return -ENOTBLK because thats what makes DIO go ahead and go back
5433          * to buffered IO.  Don't blame me, this is the price we pay for using
5434          * the generic code.
5435          */
5436         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
5437             em->block_start == EXTENT_MAP_INLINE) {
5438                 free_extent_map(em);
5439                 return -ENOTBLK;
5440         }
5441
5442         /* Just a good old fashioned hole, return */
5443         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
5444                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
5445                 free_extent_map(em);
5446                 /* DIO will do one hole at a time, so just unlock a sector */
5447                 unlock_extent(&BTRFS_I(inode)->io_tree, start,
5448                               start + root->sectorsize - 1, GFP_NOFS);
5449                 return 0;
5450         }
5451
5452         /*
5453          * We don't allocate a new extent in the following cases
5454          *
5455          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
5456          * existing extent.
5457          * 2) The extent is marked as PREALLOC.  We're good to go here and can
5458          * just use the extent.
5459          *
5460          */
5461         if (!create) {
5462                 len = em->len - (start - em->start);
5463                 goto map;
5464         }
5465
5466         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
5467             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
5468              em->block_start != EXTENT_MAP_HOLE)) {
5469                 int type;
5470                 int ret;
5471                 u64 block_start;
5472
5473                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
5474                         type = BTRFS_ORDERED_PREALLOC;
5475                 else
5476                         type = BTRFS_ORDERED_NOCOW;
5477                 len = min(len, em->len - (start - em->start));
5478                 block_start = em->block_start + (start - em->start);
5479
5480                 /*
5481                  * we're not going to log anything, but we do need
5482                  * to make sure the current transaction stays open
5483                  * while we look for nocow cross refs
5484                  */
5485                 trans = btrfs_join_transaction(root, 0);
5486                 if (!trans)
5487                         goto must_cow;
5488
5489                 if (can_nocow_odirect(trans, inode, start, len) == 1) {
5490                         ret = btrfs_add_ordered_extent_dio(inode, start,
5491                                            block_start, len, len, type);
5492                         btrfs_end_transaction(trans, root);
5493                         if (ret) {
5494                                 free_extent_map(em);
5495                                 return ret;
5496                         }
5497                         goto unlock;
5498                 }
5499                 btrfs_end_transaction(trans, root);
5500         }
5501 must_cow:
5502         /*
5503          * this will cow the extent, reset the len in case we changed
5504          * it above
5505          */
5506         len = bh_result->b_size;
5507         free_extent_map(em);
5508         em = btrfs_new_extent_direct(inode, start, len);
5509         if (IS_ERR(em))
5510                 return PTR_ERR(em);
5511         len = min(len, em->len - (start - em->start));
5512 unlock:
5513         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, start + len - 1,
5514                           EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DIRTY, 1,
5515                           0, NULL, GFP_NOFS);
5516 map:
5517         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
5518                 inode->i_blkbits;
5519         bh_result->b_size = len;
5520         bh_result->b_bdev = em->bdev;
5521         set_buffer_mapped(bh_result);
5522         if (create && !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
5523                 set_buffer_new(bh_result);
5524
5525         free_extent_map(em);
5526
5527         return 0;
5528 }
5529
5530 struct btrfs_dio_private {
5531         struct inode *inode;
5532         u64 logical_offset;
5533         u64 disk_bytenr;
5534         u64 bytes;
5535         u32 *csums;
5536         void *private;
5537
5538         /* number of bios pending for this dio */
5539         atomic_t pending_bios;
5540
5541         /* IO errors */
5542         int errors;
5543
5544         struct bio *orig_bio;
5545 };
5546
5547 static void btrfs_endio_direct_read(struct bio *bio, int err)
5548 {
5549         struct btrfs_dio_private *dip = bio->bi_private;
5550         struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
5551         struct bio_vec *bvec = bio->bi_io_vec;
5552         struct inode *inode = dip->inode;
5553         struct btrfs_root *root = BTRFS_I(inode)->root;
5554         u64 start;
5555         u32 *private = dip->csums;
5556
5557         start = dip->logical_offset;
5558         do {
5559                 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
5560                         struct page *page = bvec->bv_page;
5561                         char *kaddr;
5562                         u32 csum = ~(u32)0;
5563                         unsigned long flags;
5564
5565                         local_irq_save(flags);
5566                         kaddr = kmap_atomic(page, KM_IRQ0);
5567                         csum = btrfs_csum_data(root, kaddr + bvec->bv_offset,
5568                                                csum, bvec->bv_len);
5569                         btrfs_csum_final(csum, (char *)&csum);
5570                         kunmap_atomic(kaddr, KM_IRQ0);
5571                         local_irq_restore(flags);
5572
5573                         flush_dcache_page(bvec->bv_page);
5574                         if (csum != *private) {
5575                                 printk(KERN_ERR "btrfs csum failed ino %lu off"
5576                                       " %llu csum %u private %u\n",
5577                                       inode->i_ino, (unsigned long long)start,
5578                                       csum, *private);
5579                                 err = -EIO;
5580                         }
5581                 }
5582
5583                 start += bvec->bv_len;
5584                 private++;
5585                 bvec++;
5586         } while (bvec <= bvec_end);
5587
5588         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
5589                       dip->logical_offset + dip->bytes - 1, GFP_NOFS);
5590         bio->bi_private = dip->private;
5591
5592         kfree(dip->csums);
5593         kfree(dip);
5594         dio_end_io(bio, err);
5595 }
5596
5597 static void btrfs_endio_direct_write(struct bio *bio, int err)
5598 {
5599         struct btrfs_dio_private *dip = bio->bi_private;
5600         struct inode *inode = dip->inode;
5601         struct btrfs_root *root = BTRFS_I(inode)->root;
5602         struct btrfs_trans_handle *trans;
5603         struct btrfs_ordered_extent *ordered = NULL;
5604         struct extent_state *cached_state = NULL;
5605         u64 ordered_offset = dip->logical_offset;
5606         u64 ordered_bytes = dip->bytes;
5607         int ret;
5608
5609         if (err)
5610                 goto out_done;
5611 again:
5612         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
5613                                                    &ordered_offset,
5614                                                    ordered_bytes);
5615         if (!ret)
5616                 goto out_test;
5617
5618         BUG_ON(!ordered);
5619
5620         trans = btrfs_join_transaction(root, 1);
5621         if (!trans) {
5622                 err = -ENOMEM;
5623                 goto out;
5624         }
5625         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
5626
5627         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered->flags)) {
5628                 ret = btrfs_ordered_update_i_size(inode, 0, ordered);
5629                 if (!ret)
5630                         ret = btrfs_update_inode(trans, root, inode);
5631                 err = ret;
5632                 goto out;
5633         }
5634
5635         lock_extent_bits(&BTRFS_I(inode)->io_tree, ordered->file_offset,
5636                          ordered->file_offset + ordered->len - 1, 0,
5637                          &cached_state, GFP_NOFS);
5638
5639         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered->flags)) {
5640                 ret = btrfs_mark_extent_written(trans, inode,
5641                                                 ordered->file_offset,
5642                                                 ordered->file_offset +
5643                                                 ordered->len);
5644                 if (ret) {
5645                         err = ret;
5646                         goto out_unlock;
5647                 }
5648         } else {
5649                 ret = insert_reserved_file_extent(trans, inode,
5650                                                   ordered->file_offset,
5651                                                   ordered->start,
5652                                                   ordered->disk_len,
5653                                                   ordered->len,
5654                                                   ordered->len,
5655                                                   0, 0, 0,
5656                                                   BTRFS_FILE_EXTENT_REG);
5657                 unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
5658                                    ordered->file_offset, ordered->len);
5659                 if (ret) {
5660             &nb