Merge git://git.kernel.org/pub/scm/linux/kernel/git/mason/btrfs-unstable
[~shefty/rdma-dev.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 compress_type = BTRFS_COMPRESS_NONE;
126
127         if (compressed_size && compressed_pages) {
128                 compress_type = root->fs_info->compress_type;
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 (compress_type != BTRFS_COMPRESS_NONE) {
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                                                   compress_type);
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         int compress_type;
267         struct list_head list;
268 };
269
270 struct async_cow {
271         struct inode *inode;
272         struct btrfs_root *root;
273         struct page *locked_page;
274         u64 start;
275         u64 end;
276         struct list_head extents;
277         struct btrfs_work work;
278 };
279
280 static noinline int add_async_extent(struct async_cow *cow,
281                                      u64 start, u64 ram_size,
282                                      u64 compressed_size,
283                                      struct page **pages,
284                                      unsigned long nr_pages,
285                                      int compress_type)
286 {
287         struct async_extent *async_extent;
288
289         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
290         async_extent->start = start;
291         async_extent->ram_size = ram_size;
292         async_extent->compressed_size = compressed_size;
293         async_extent->pages = pages;
294         async_extent->nr_pages = nr_pages;
295         async_extent->compress_type = compress_type;
296         list_add_tail(&async_extent->list, &cow->extents);
297         return 0;
298 }
299
300 /*
301  * we create compressed extents in two phases.  The first
302  * phase compresses a range of pages that have already been
303  * locked (both pages and state bits are locked).
304  *
305  * This is done inside an ordered work queue, and the compression
306  * is spread across many cpus.  The actual IO submission is step
307  * two, and the ordered work queue takes care of making sure that
308  * happens in the same order things were put onto the queue by
309  * writepages and friends.
310  *
311  * If this code finds it can't get good compression, it puts an
312  * entry onto the work queue to write the uncompressed bytes.  This
313  * makes sure that both compressed inodes and uncompressed inodes
314  * are written in the same order that pdflush sent them down.
315  */
316 static noinline int compress_file_range(struct inode *inode,
317                                         struct page *locked_page,
318                                         u64 start, u64 end,
319                                         struct async_cow *async_cow,
320                                         int *num_added)
321 {
322         struct btrfs_root *root = BTRFS_I(inode)->root;
323         struct btrfs_trans_handle *trans;
324         u64 num_bytes;
325         u64 blocksize = root->sectorsize;
326         u64 actual_end;
327         u64 isize = i_size_read(inode);
328         int ret = 0;
329         struct page **pages = NULL;
330         unsigned long nr_pages;
331         unsigned long nr_pages_ret = 0;
332         unsigned long total_compressed = 0;
333         unsigned long total_in = 0;
334         unsigned long max_compressed = 128 * 1024;
335         unsigned long max_uncompressed = 128 * 1024;
336         int i;
337         int will_compress;
338         int compress_type = root->fs_info->compress_type;
339
340         actual_end = min_t(u64, isize, end + 1);
341 again:
342         will_compress = 0;
343         nr_pages = (end >> PAGE_CACHE_SHIFT) - (start >> PAGE_CACHE_SHIFT) + 1;
344         nr_pages = min(nr_pages, (128 * 1024UL) / PAGE_CACHE_SIZE);
345
346         /*
347          * we don't want to send crud past the end of i_size through
348          * compression, that's just a waste of CPU time.  So, if the
349          * end of the file is before the start of our current
350          * requested range of bytes, we bail out to the uncompressed
351          * cleanup code that can deal with all of this.
352          *
353          * It isn't really the fastest way to fix things, but this is a
354          * very uncommon corner.
355          */
356         if (actual_end <= start)
357                 goto cleanup_and_bail_uncompressed;
358
359         total_compressed = actual_end - start;
360
361         /* we want to make sure that amount of ram required to uncompress
362          * an extent is reasonable, so we limit the total size in ram
363          * of a compressed extent to 128k.  This is a crucial number
364          * because it also controls how easily we can spread reads across
365          * cpus for decompression.
366          *
367          * We also want to make sure the amount of IO required to do
368          * a random read is reasonably small, so we limit the size of
369          * a compressed extent to 128k.
370          */
371         total_compressed = min(total_compressed, max_uncompressed);
372         num_bytes = (end - start + blocksize) & ~(blocksize - 1);
373         num_bytes = max(blocksize,  num_bytes);
374         total_in = 0;
375         ret = 0;
376
377         /*
378          * we do compression for mount -o compress and when the
379          * inode has not been flagged as nocompress.  This flag can
380          * change at any time if we discover bad compression ratios.
381          */
382         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS) &&
383             (btrfs_test_opt(root, COMPRESS) ||
384              (BTRFS_I(inode)->force_compress))) {
385                 WARN_ON(pages);
386                 pages = kzalloc(sizeof(struct page *) * nr_pages, GFP_NOFS);
387
388                 if (BTRFS_I(inode)->force_compress)
389                         compress_type = BTRFS_I(inode)->force_compress;
390
391                 ret = btrfs_compress_pages(compress_type,
392                                            inode->i_mapping, start,
393                                            total_compressed, pages,
394                                            nr_pages, &nr_pages_ret,
395                                            &total_in,
396                                            &total_compressed,
397                                            max_compressed);
398
399                 if (!ret) {
400                         unsigned long offset = total_compressed &
401                                 (PAGE_CACHE_SIZE - 1);
402                         struct page *page = pages[nr_pages_ret - 1];
403                         char *kaddr;
404
405                         /* zero the tail end of the last page, we might be
406                          * sending it down to disk
407                          */
408                         if (offset) {
409                                 kaddr = kmap_atomic(page, KM_USER0);
410                                 memset(kaddr + offset, 0,
411                                        PAGE_CACHE_SIZE - offset);
412                                 kunmap_atomic(kaddr, KM_USER0);
413                         }
414                         will_compress = 1;
415                 }
416         }
417         if (start == 0) {
418                 trans = btrfs_join_transaction(root, 1);
419                 BUG_ON(IS_ERR(trans));
420                 btrfs_set_trans_block_group(trans, inode);
421                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
422
423                 /* lets try to make an inline extent */
424                 if (ret || total_in < (actual_end - start)) {
425                         /* we didn't compress the entire range, try
426                          * to make an uncompressed inline extent.
427                          */
428                         ret = cow_file_range_inline(trans, root, inode,
429                                                     start, end, 0, NULL);
430                 } else {
431                         /* try making a compressed inline extent */
432                         ret = cow_file_range_inline(trans, root, inode,
433                                                     start, end,
434                                                     total_compressed, pages);
435                 }
436                 if (ret == 0) {
437                         /*
438                          * inline extent creation worked, we don't need
439                          * to create any more async work items.  Unlock
440                          * and free up our temp pages.
441                          */
442                         extent_clear_unlock_delalloc(inode,
443                              &BTRFS_I(inode)->io_tree,
444                              start, end, NULL,
445                              EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
446                              EXTENT_CLEAR_DELALLOC |
447                              EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK);
448
449                         btrfs_end_transaction(trans, root);
450                         goto free_pages_out;
451                 }
452                 btrfs_end_transaction(trans, root);
453         }
454
455         if (will_compress) {
456                 /*
457                  * we aren't doing an inline extent round the compressed size
458                  * up to a block size boundary so the allocator does sane
459                  * things
460                  */
461                 total_compressed = (total_compressed + blocksize - 1) &
462                         ~(blocksize - 1);
463
464                 /*
465                  * one last check to make sure the compression is really a
466                  * win, compare the page count read with the blocks on disk
467                  */
468                 total_in = (total_in + PAGE_CACHE_SIZE - 1) &
469                         ~(PAGE_CACHE_SIZE - 1);
470                 if (total_compressed >= total_in) {
471                         will_compress = 0;
472                 } else {
473                         num_bytes = total_in;
474                 }
475         }
476         if (!will_compress && pages) {
477                 /*
478                  * the compression code ran but failed to make things smaller,
479                  * free any pages it allocated and our page pointer array
480                  */
481                 for (i = 0; i < nr_pages_ret; i++) {
482                         WARN_ON(pages[i]->mapping);
483                         page_cache_release(pages[i]);
484                 }
485                 kfree(pages);
486                 pages = NULL;
487                 total_compressed = 0;
488                 nr_pages_ret = 0;
489
490                 /* flag the file so we don't compress in the future */
491                 if (!btrfs_test_opt(root, FORCE_COMPRESS) &&
492                     !(BTRFS_I(inode)->force_compress)) {
493                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
494                 }
495         }
496         if (will_compress) {
497                 *num_added += 1;
498
499                 /* the async work queues will take care of doing actual
500                  * allocation on disk for these compressed pages,
501                  * and will submit them to the elevator.
502                  */
503                 add_async_extent(async_cow, start, num_bytes,
504                                  total_compressed, pages, nr_pages_ret,
505                                  compress_type);
506
507                 if (start + num_bytes < end) {
508                         start += num_bytes;
509                         pages = NULL;
510                         cond_resched();
511                         goto again;
512                 }
513         } else {
514 cleanup_and_bail_uncompressed:
515                 /*
516                  * No compression, but we still need to write the pages in
517                  * the file we've been given so far.  redirty the locked
518                  * page if it corresponds to our extent and set things up
519                  * for the async work queue to run cow_file_range to do
520                  * the normal delalloc dance
521                  */
522                 if (page_offset(locked_page) >= start &&
523                     page_offset(locked_page) <= end) {
524                         __set_page_dirty_nobuffers(locked_page);
525                         /* unlocked later on in the async handlers */
526                 }
527                 add_async_extent(async_cow, start, end - start + 1,
528                                  0, NULL, 0, BTRFS_COMPRESS_NONE);
529                 *num_added += 1;
530         }
531
532 out:
533         return 0;
534
535 free_pages_out:
536         for (i = 0; i < nr_pages_ret; i++) {
537                 WARN_ON(pages[i]->mapping);
538                 page_cache_release(pages[i]);
539         }
540         kfree(pages);
541
542         goto out;
543 }
544
545 /*
546  * phase two of compressed writeback.  This is the ordered portion
547  * of the code, which only gets called in the order the work was
548  * queued.  We walk all the async extents created by compress_file_range
549  * and send them down to the disk.
550  */
551 static noinline int submit_compressed_extents(struct inode *inode,
552                                               struct async_cow *async_cow)
553 {
554         struct async_extent *async_extent;
555         u64 alloc_hint = 0;
556         struct btrfs_trans_handle *trans;
557         struct btrfs_key ins;
558         struct extent_map *em;
559         struct btrfs_root *root = BTRFS_I(inode)->root;
560         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
561         struct extent_io_tree *io_tree;
562         int ret = 0;
563
564         if (list_empty(&async_cow->extents))
565                 return 0;
566
567
568         while (!list_empty(&async_cow->extents)) {
569                 async_extent = list_entry(async_cow->extents.next,
570                                           struct async_extent, list);
571                 list_del(&async_extent->list);
572
573                 io_tree = &BTRFS_I(inode)->io_tree;
574
575 retry:
576                 /* did the compression code fall back to uncompressed IO? */
577                 if (!async_extent->pages) {
578                         int page_started = 0;
579                         unsigned long nr_written = 0;
580
581                         lock_extent(io_tree, async_extent->start,
582                                          async_extent->start +
583                                          async_extent->ram_size - 1, GFP_NOFS);
584
585                         /* allocate blocks */
586                         ret = cow_file_range(inode, async_cow->locked_page,
587                                              async_extent->start,
588                                              async_extent->start +
589                                              async_extent->ram_size - 1,
590                                              &page_started, &nr_written, 0);
591
592                         /*
593                          * if page_started, cow_file_range inserted an
594                          * inline extent and took care of all the unlocking
595                          * and IO for us.  Otherwise, we need to submit
596                          * all those pages down to the drive.
597                          */
598                         if (!page_started && !ret)
599                                 extent_write_locked_range(io_tree,
600                                                   inode, async_extent->start,
601                                                   async_extent->start +
602                                                   async_extent->ram_size - 1,
603                                                   btrfs_get_extent,
604                                                   WB_SYNC_ALL);
605                         kfree(async_extent);
606                         cond_resched();
607                         continue;
608                 }
609
610                 lock_extent(io_tree, async_extent->start,
611                             async_extent->start + async_extent->ram_size - 1,
612                             GFP_NOFS);
613
614                 trans = btrfs_join_transaction(root, 1);
615                 BUG_ON(IS_ERR(trans));
616                 ret = btrfs_reserve_extent(trans, root,
617                                            async_extent->compressed_size,
618                                            async_extent->compressed_size,
619                                            0, alloc_hint,
620                                            (u64)-1, &ins, 1);
621                 btrfs_end_transaction(trans, root);
622
623                 if (ret) {
624                         int i;
625                         for (i = 0; i < async_extent->nr_pages; i++) {
626                                 WARN_ON(async_extent->pages[i]->mapping);
627                                 page_cache_release(async_extent->pages[i]);
628                         }
629                         kfree(async_extent->pages);
630                         async_extent->nr_pages = 0;
631                         async_extent->pages = NULL;
632                         unlock_extent(io_tree, async_extent->start,
633                                       async_extent->start +
634                                       async_extent->ram_size - 1, GFP_NOFS);
635                         goto retry;
636                 }
637
638                 /*
639                  * here we're doing allocation and writeback of the
640                  * compressed pages
641                  */
642                 btrfs_drop_extent_cache(inode, async_extent->start,
643                                         async_extent->start +
644                                         async_extent->ram_size - 1, 0);
645
646                 em = alloc_extent_map(GFP_NOFS);
647                 BUG_ON(!em);
648                 em->start = async_extent->start;
649                 em->len = async_extent->ram_size;
650                 em->orig_start = em->start;
651
652                 em->block_start = ins.objectid;
653                 em->block_len = ins.offset;
654                 em->bdev = root->fs_info->fs_devices->latest_bdev;
655                 em->compress_type = async_extent->compress_type;
656                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
657                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
658
659                 while (1) {
660                         write_lock(&em_tree->lock);
661                         ret = add_extent_mapping(em_tree, em);
662                         write_unlock(&em_tree->lock);
663                         if (ret != -EEXIST) {
664                                 free_extent_map(em);
665                                 break;
666                         }
667                         btrfs_drop_extent_cache(inode, async_extent->start,
668                                                 async_extent->start +
669                                                 async_extent->ram_size - 1, 0);
670                 }
671
672                 ret = btrfs_add_ordered_extent_compress(inode,
673                                                 async_extent->start,
674                                                 ins.objectid,
675                                                 async_extent->ram_size,
676                                                 ins.offset,
677                                                 BTRFS_ORDERED_COMPRESSED,
678                                                 async_extent->compress_type);
679                 BUG_ON(ret);
680
681                 /*
682                  * clear dirty, set writeback and unlock the pages.
683                  */
684                 extent_clear_unlock_delalloc(inode,
685                                 &BTRFS_I(inode)->io_tree,
686                                 async_extent->start,
687                                 async_extent->start +
688                                 async_extent->ram_size - 1,
689                                 NULL, EXTENT_CLEAR_UNLOCK_PAGE |
690                                 EXTENT_CLEAR_UNLOCK |
691                                 EXTENT_CLEAR_DELALLOC |
692                                 EXTENT_CLEAR_DIRTY | EXTENT_SET_WRITEBACK);
693
694                 ret = btrfs_submit_compressed_write(inode,
695                                     async_extent->start,
696                                     async_extent->ram_size,
697                                     ins.objectid,
698                                     ins.offset, async_extent->pages,
699                                     async_extent->nr_pages);
700
701                 BUG_ON(ret);
702                 alloc_hint = ins.objectid + ins.offset;
703                 kfree(async_extent);
704                 cond_resched();
705         }
706
707         return 0;
708 }
709
710 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
711                                       u64 num_bytes)
712 {
713         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
714         struct extent_map *em;
715         u64 alloc_hint = 0;
716
717         read_lock(&em_tree->lock);
718         em = search_extent_mapping(em_tree, start, num_bytes);
719         if (em) {
720                 /*
721                  * if block start isn't an actual block number then find the
722                  * first block in this inode and use that as a hint.  If that
723                  * block is also bogus then just don't worry about it.
724                  */
725                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
726                         free_extent_map(em);
727                         em = search_extent_mapping(em_tree, 0, 0);
728                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
729                                 alloc_hint = em->block_start;
730                         if (em)
731                                 free_extent_map(em);
732                 } else {
733                         alloc_hint = em->block_start;
734                         free_extent_map(em);
735                 }
736         }
737         read_unlock(&em_tree->lock);
738
739         return alloc_hint;
740 }
741
742 /*
743  * when extent_io.c finds a delayed allocation range in the file,
744  * the call backs end up in this code.  The basic idea is to
745  * allocate extents on disk for the range, and create ordered data structs
746  * in ram to track those extents.
747  *
748  * locked_page is the page that writepage had locked already.  We use
749  * it to make sure we don't do extra locks or unlocks.
750  *
751  * *page_started is set to one if we unlock locked_page and do everything
752  * required to start IO on it.  It may be clean and already done with
753  * IO when we return.
754  */
755 static noinline int cow_file_range(struct inode *inode,
756                                    struct page *locked_page,
757                                    u64 start, u64 end, int *page_started,
758                                    unsigned long *nr_written,
759                                    int unlock)
760 {
761         struct btrfs_root *root = BTRFS_I(inode)->root;
762         struct btrfs_trans_handle *trans;
763         u64 alloc_hint = 0;
764         u64 num_bytes;
765         unsigned long ram_size;
766         u64 disk_num_bytes;
767         u64 cur_alloc_size;
768         u64 blocksize = root->sectorsize;
769         struct btrfs_key ins;
770         struct extent_map *em;
771         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
772         int ret = 0;
773
774         BUG_ON(root == root->fs_info->tree_root);
775         trans = btrfs_join_transaction(root, 1);
776         BUG_ON(IS_ERR(trans));
777         btrfs_set_trans_block_group(trans, inode);
778         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
779
780         num_bytes = (end - start + blocksize) & ~(blocksize - 1);
781         num_bytes = max(blocksize,  num_bytes);
782         disk_num_bytes = num_bytes;
783         ret = 0;
784
785         if (start == 0) {
786                 /* lets try to make an inline extent */
787                 ret = cow_file_range_inline(trans, root, inode,
788                                             start, end, 0, NULL);
789                 if (ret == 0) {
790                         extent_clear_unlock_delalloc(inode,
791                                      &BTRFS_I(inode)->io_tree,
792                                      start, end, NULL,
793                                      EXTENT_CLEAR_UNLOCK_PAGE |
794                                      EXTENT_CLEAR_UNLOCK |
795                                      EXTENT_CLEAR_DELALLOC |
796                                      EXTENT_CLEAR_DIRTY |
797                                      EXTENT_SET_WRITEBACK |
798                                      EXTENT_END_WRITEBACK);
799
800                         *nr_written = *nr_written +
801                              (end - start + PAGE_CACHE_SIZE) / PAGE_CACHE_SIZE;
802                         *page_started = 1;
803                         ret = 0;
804                         goto out;
805                 }
806         }
807
808         BUG_ON(disk_num_bytes >
809                btrfs_super_total_bytes(&root->fs_info->super_copy));
810
811         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
812         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
813
814         while (disk_num_bytes > 0) {
815                 unsigned long op;
816
817                 cur_alloc_size = disk_num_bytes;
818                 ret = btrfs_reserve_extent(trans, root, cur_alloc_size,
819                                            root->sectorsize, 0, alloc_hint,
820                                            (u64)-1, &ins, 1);
821                 BUG_ON(ret);
822
823                 em = alloc_extent_map(GFP_NOFS);
824                 BUG_ON(!em);
825                 em->start = start;
826                 em->orig_start = em->start;
827                 ram_size = ins.offset;
828                 em->len = ins.offset;
829
830                 em->block_start = ins.objectid;
831                 em->block_len = ins.offset;
832                 em->bdev = root->fs_info->fs_devices->latest_bdev;
833                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
834
835                 while (1) {
836                         write_lock(&em_tree->lock);
837                         ret = add_extent_mapping(em_tree, em);
838                         write_unlock(&em_tree->lock);
839                         if (ret != -EEXIST) {
840                                 free_extent_map(em);
841                                 break;
842                         }
843                         btrfs_drop_extent_cache(inode, start,
844                                                 start + ram_size - 1, 0);
845                 }
846
847                 cur_alloc_size = ins.offset;
848                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
849                                                ram_size, cur_alloc_size, 0);
850                 BUG_ON(ret);
851
852                 if (root->root_key.objectid ==
853                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
854                         ret = btrfs_reloc_clone_csums(inode, start,
855                                                       cur_alloc_size);
856                         BUG_ON(ret);
857                 }
858
859                 if (disk_num_bytes < cur_alloc_size)
860                         break;
861
862                 /* we're not doing compressed IO, don't unlock the first
863                  * page (which the caller expects to stay locked), don't
864                  * clear any dirty bits and don't set any writeback bits
865                  *
866                  * Do set the Private2 bit so we know this page was properly
867                  * setup for writepage
868                  */
869                 op = unlock ? EXTENT_CLEAR_UNLOCK_PAGE : 0;
870                 op |= EXTENT_CLEAR_UNLOCK | EXTENT_CLEAR_DELALLOC |
871                         EXTENT_SET_PRIVATE2;
872
873                 extent_clear_unlock_delalloc(inode, &BTRFS_I(inode)->io_tree,
874                                              start, start + ram_size - 1,
875                                              locked_page, op);
876                 disk_num_bytes -= cur_alloc_size;
877                 num_bytes -= cur_alloc_size;
878                 alloc_hint = ins.objectid + ins.offset;
879                 start += cur_alloc_size;
880         }
881 out:
882         ret = 0;
883         btrfs_end_transaction(trans, root);
884
885         return ret;
886 }
887
888 /*
889  * work queue call back to started compression on a file and pages
890  */
891 static noinline void async_cow_start(struct btrfs_work *work)
892 {
893         struct async_cow *async_cow;
894         int num_added = 0;
895         async_cow = container_of(work, struct async_cow, work);
896
897         compress_file_range(async_cow->inode, async_cow->locked_page,
898                             async_cow->start, async_cow->end, async_cow,
899                             &num_added);
900         if (num_added == 0)
901                 async_cow->inode = NULL;
902 }
903
904 /*
905  * work queue call back to submit previously compressed pages
906  */
907 static noinline void async_cow_submit(struct btrfs_work *work)
908 {
909         struct async_cow *async_cow;
910         struct btrfs_root *root;
911         unsigned long nr_pages;
912
913         async_cow = container_of(work, struct async_cow, work);
914
915         root = async_cow->root;
916         nr_pages = (async_cow->end - async_cow->start + PAGE_CACHE_SIZE) >>
917                 PAGE_CACHE_SHIFT;
918
919         atomic_sub(nr_pages, &root->fs_info->async_delalloc_pages);
920
921         if (atomic_read(&root->fs_info->async_delalloc_pages) <
922             5 * 1042 * 1024 &&
923             waitqueue_active(&root->fs_info->async_submit_wait))
924                 wake_up(&root->fs_info->async_submit_wait);
925
926         if (async_cow->inode)
927                 submit_compressed_extents(async_cow->inode, async_cow);
928 }
929
930 static noinline void async_cow_free(struct btrfs_work *work)
931 {
932         struct async_cow *async_cow;
933         async_cow = container_of(work, struct async_cow, work);
934         kfree(async_cow);
935 }
936
937 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
938                                 u64 start, u64 end, int *page_started,
939                                 unsigned long *nr_written)
940 {
941         struct async_cow *async_cow;
942         struct btrfs_root *root = BTRFS_I(inode)->root;
943         unsigned long nr_pages;
944         u64 cur_end;
945         int limit = 10 * 1024 * 1042;
946
947         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
948                          1, 0, NULL, GFP_NOFS);
949         while (start < end) {
950                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
951                 async_cow->inode = inode;
952                 async_cow->root = root;
953                 async_cow->locked_page = locked_page;
954                 async_cow->start = start;
955
956                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
957                         cur_end = end;
958                 else
959                         cur_end = min(end, start + 512 * 1024 - 1);
960
961                 async_cow->end = cur_end;
962                 INIT_LIST_HEAD(&async_cow->extents);
963
964                 async_cow->work.func = async_cow_start;
965                 async_cow->work.ordered_func = async_cow_submit;
966                 async_cow->work.ordered_free = async_cow_free;
967                 async_cow->work.flags = 0;
968
969                 nr_pages = (cur_end - start + PAGE_CACHE_SIZE) >>
970                         PAGE_CACHE_SHIFT;
971                 atomic_add(nr_pages, &root->fs_info->async_delalloc_pages);
972
973                 btrfs_queue_worker(&root->fs_info->delalloc_workers,
974                                    &async_cow->work);
975
976                 if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
977                         wait_event(root->fs_info->async_submit_wait,
978                            (atomic_read(&root->fs_info->async_delalloc_pages) <
979                             limit));
980                 }
981
982                 while (atomic_read(&root->fs_info->async_submit_draining) &&
983                       atomic_read(&root->fs_info->async_delalloc_pages)) {
984                         wait_event(root->fs_info->async_submit_wait,
985                           (atomic_read(&root->fs_info->async_delalloc_pages) ==
986                            0));
987                 }
988
989                 *nr_written += nr_pages;
990                 start = cur_end + 1;
991         }
992         *page_started = 1;
993         return 0;
994 }
995
996 static noinline int csum_exist_in_range(struct btrfs_root *root,
997                                         u64 bytenr, u64 num_bytes)
998 {
999         int ret;
1000         struct btrfs_ordered_sum *sums;
1001         LIST_HEAD(list);
1002
1003         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, bytenr,
1004                                        bytenr + num_bytes - 1, &list);
1005         if (ret == 0 && list_empty(&list))
1006                 return 0;
1007
1008         while (!list_empty(&list)) {
1009                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1010                 list_del(&sums->list);
1011                 kfree(sums);
1012         }
1013         return 1;
1014 }
1015
1016 /*
1017  * when nowcow writeback call back.  This checks for snapshots or COW copies
1018  * of the extents that exist in the file, and COWs the file as required.
1019  *
1020  * If no cow copies or snapshots exist, we write directly to the existing
1021  * blocks on disk
1022  */
1023 static noinline int run_delalloc_nocow(struct inode *inode,
1024                                        struct page *locked_page,
1025                               u64 start, u64 end, int *page_started, int force,
1026                               unsigned long *nr_written)
1027 {
1028         struct btrfs_root *root = BTRFS_I(inode)->root;
1029         struct btrfs_trans_handle *trans;
1030         struct extent_buffer *leaf;
1031         struct btrfs_path *path;
1032         struct btrfs_file_extent_item *fi;
1033         struct btrfs_key found_key;
1034         u64 cow_start;
1035         u64 cur_offset;
1036         u64 extent_end;
1037         u64 extent_offset;
1038         u64 disk_bytenr;
1039         u64 num_bytes;
1040         int extent_type;
1041         int ret;
1042         int type;
1043         int nocow;
1044         int check_prev = 1;
1045         bool nolock = false;
1046
1047         path = btrfs_alloc_path();
1048         BUG_ON(!path);
1049         if (root == root->fs_info->tree_root) {
1050                 nolock = true;
1051                 trans = btrfs_join_transaction_nolock(root, 1);
1052         } else {
1053                 trans = btrfs_join_transaction(root, 1);
1054         }
1055         BUG_ON(IS_ERR(trans));
1056
1057         cow_start = (u64)-1;
1058         cur_offset = start;
1059         while (1) {
1060                 ret = btrfs_lookup_file_extent(trans, root, path, inode->i_ino,
1061                                                cur_offset, 0);
1062                 BUG_ON(ret < 0);
1063                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1064                         leaf = path->nodes[0];
1065                         btrfs_item_key_to_cpu(leaf, &found_key,
1066                                               path->slots[0] - 1);
1067                         if (found_key.objectid == inode->i_ino &&
1068                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1069                                 path->slots[0]--;
1070                 }
1071                 check_prev = 0;
1072 next_slot:
1073                 leaf = path->nodes[0];
1074                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1075                         ret = btrfs_next_leaf(root, path);
1076                         if (ret < 0)
1077                                 BUG_ON(1);
1078                         if (ret > 0)
1079                                 break;
1080                         leaf = path->nodes[0];
1081                 }
1082
1083                 nocow = 0;
1084                 disk_bytenr = 0;
1085                 num_bytes = 0;
1086                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1087
1088                 if (found_key.objectid > inode->i_ino ||
1089                     found_key.type > BTRFS_EXTENT_DATA_KEY ||
1090                     found_key.offset > end)
1091                         break;
1092
1093                 if (found_key.offset > cur_offset) {
1094                         extent_end = found_key.offset;
1095                         extent_type = 0;
1096                         goto out_check;
1097                 }
1098
1099                 fi = btrfs_item_ptr(leaf, path->slots[0],
1100                                     struct btrfs_file_extent_item);
1101                 extent_type = btrfs_file_extent_type(leaf, fi);
1102
1103                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1104                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1105                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1106                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1107                         extent_end = found_key.offset +
1108                                 btrfs_file_extent_num_bytes(leaf, fi);
1109                         if (extent_end <= start) {
1110                                 path->slots[0]++;
1111                                 goto next_slot;
1112                         }
1113                         if (disk_bytenr == 0)
1114                                 goto out_check;
1115                         if (btrfs_file_extent_compression(leaf, fi) ||
1116                             btrfs_file_extent_encryption(leaf, fi) ||
1117                             btrfs_file_extent_other_encoding(leaf, fi))
1118                                 goto out_check;
1119                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1120                                 goto out_check;
1121                         if (btrfs_extent_readonly(root, disk_bytenr))
1122                                 goto out_check;
1123                         if (btrfs_cross_ref_exist(trans, root, inode->i_ino,
1124                                                   found_key.offset -
1125                                                   extent_offset, disk_bytenr))
1126                                 goto out_check;
1127                         disk_bytenr += extent_offset;
1128                         disk_bytenr += cur_offset - found_key.offset;
1129                         num_bytes = min(end + 1, extent_end) - cur_offset;
1130                         /*
1131                          * force cow if csum exists in the range.
1132                          * this ensure that csum for a given extent are
1133                          * either valid or do not exist.
1134                          */
1135                         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
1136                                 goto out_check;
1137                         nocow = 1;
1138                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1139                         extent_end = found_key.offset +
1140                                 btrfs_file_extent_inline_len(leaf, fi);
1141                         extent_end = ALIGN(extent_end, root->sectorsize);
1142                 } else {
1143                         BUG_ON(1);
1144                 }
1145 out_check:
1146                 if (extent_end <= start) {
1147                         path->slots[0]++;
1148                         goto next_slot;
1149                 }
1150                 if (!nocow) {
1151                         if (cow_start == (u64)-1)
1152                                 cow_start = cur_offset;
1153                         cur_offset = extent_end;
1154                         if (cur_offset > end)
1155                                 break;
1156                         path->slots[0]++;
1157                         goto next_slot;
1158                 }
1159
1160                 btrfs_release_path(root, path);
1161                 if (cow_start != (u64)-1) {
1162                         ret = cow_file_range(inode, locked_page, cow_start,
1163                                         found_key.offset - 1, page_started,
1164                                         nr_written, 1);
1165                         BUG_ON(ret);
1166                         cow_start = (u64)-1;
1167                 }
1168
1169                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1170                         struct extent_map *em;
1171                         struct extent_map_tree *em_tree;
1172                         em_tree = &BTRFS_I(inode)->extent_tree;
1173                         em = alloc_extent_map(GFP_NOFS);
1174                         BUG_ON(!em);
1175                         em->start = cur_offset;
1176                         em->orig_start = em->start;
1177                         em->len = num_bytes;
1178                         em->block_len = num_bytes;
1179                         em->block_start = disk_bytenr;
1180                         em->bdev = root->fs_info->fs_devices->latest_bdev;
1181                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1182                         while (1) {
1183                                 write_lock(&em_tree->lock);
1184                                 ret = add_extent_mapping(em_tree, em);
1185                                 write_unlock(&em_tree->lock);
1186                                 if (ret != -EEXIST) {
1187                                         free_extent_map(em);
1188                                         break;
1189                                 }
1190                                 btrfs_drop_extent_cache(inode, em->start,
1191                                                 em->start + em->len - 1, 0);
1192                         }
1193                         type = BTRFS_ORDERED_PREALLOC;
1194                 } else {
1195                         type = BTRFS_ORDERED_NOCOW;
1196                 }
1197
1198                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1199                                                num_bytes, num_bytes, type);
1200                 BUG_ON(ret);
1201
1202                 if (root->root_key.objectid ==
1203                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1204                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1205                                                       num_bytes);
1206                         BUG_ON(ret);
1207                 }
1208
1209                 extent_clear_unlock_delalloc(inode, &BTRFS_I(inode)->io_tree,
1210                                 cur_offset, cur_offset + num_bytes - 1,
1211                                 locked_page, EXTENT_CLEAR_UNLOCK_PAGE |
1212                                 EXTENT_CLEAR_UNLOCK | EXTENT_CLEAR_DELALLOC |
1213                                 EXTENT_SET_PRIVATE2);
1214                 cur_offset = extent_end;
1215                 if (cur_offset > end)
1216                         break;
1217         }
1218         btrfs_release_path(root, path);
1219
1220         if (cur_offset <= end && cow_start == (u64)-1)
1221                 cow_start = cur_offset;
1222         if (cow_start != (u64)-1) {
1223                 ret = cow_file_range(inode, locked_page, cow_start, end,
1224                                      page_started, nr_written, 1);
1225                 BUG_ON(ret);
1226         }
1227
1228         if (nolock) {
1229                 ret = btrfs_end_transaction_nolock(trans, root);
1230                 BUG_ON(ret);
1231         } else {
1232                 ret = btrfs_end_transaction(trans, root);
1233                 BUG_ON(ret);
1234         }
1235         btrfs_free_path(path);
1236         return 0;
1237 }
1238
1239 /*
1240  * extent_io.c call back to do delayed allocation processing
1241  */
1242 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1243                               u64 start, u64 end, int *page_started,
1244                               unsigned long *nr_written)
1245 {
1246         int ret;
1247         struct btrfs_root *root = BTRFS_I(inode)->root;
1248
1249         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW)
1250                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1251                                          page_started, 1, nr_written);
1252         else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)
1253                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1254                                          page_started, 0, nr_written);
1255         else if (!btrfs_test_opt(root, COMPRESS) &&
1256                  !(BTRFS_I(inode)->force_compress))
1257                 ret = cow_file_range(inode, locked_page, start, end,
1258                                       page_started, nr_written, 1);
1259         else
1260                 ret = cow_file_range_async(inode, locked_page, start, end,
1261                                            page_started, nr_written);
1262         return ret;
1263 }
1264
1265 static int btrfs_split_extent_hook(struct inode *inode,
1266                                    struct extent_state *orig, u64 split)
1267 {
1268         /* not delalloc, ignore it */
1269         if (!(orig->state & EXTENT_DELALLOC))
1270                 return 0;
1271
1272         atomic_inc(&BTRFS_I(inode)->outstanding_extents);
1273         return 0;
1274 }
1275
1276 /*
1277  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1278  * extents so we can keep track of new extents that are just merged onto old
1279  * extents, such as when we are doing sequential writes, so we can properly
1280  * account for the metadata space we'll need.
1281  */
1282 static int btrfs_merge_extent_hook(struct inode *inode,
1283                                    struct extent_state *new,
1284                                    struct extent_state *other)
1285 {
1286         /* not delalloc, ignore it */
1287         if (!(other->state & EXTENT_DELALLOC))
1288                 return 0;
1289
1290         atomic_dec(&BTRFS_I(inode)->outstanding_extents);
1291         return 0;
1292 }
1293
1294 /*
1295  * extent_io.c set_bit_hook, used to track delayed allocation
1296  * bytes in this file, and to maintain the list of inodes that
1297  * have pending delalloc work to be done.
1298  */
1299 static int btrfs_set_bit_hook(struct inode *inode,
1300                               struct extent_state *state, int *bits)
1301 {
1302
1303         /*
1304          * set_bit and clear bit hooks normally require _irqsave/restore
1305          * but in this case, we are only testeing for the DELALLOC
1306          * bit, which is only set or cleared with irqs on
1307          */
1308         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1309                 struct btrfs_root *root = BTRFS_I(inode)->root;
1310                 u64 len = state->end + 1 - state->start;
1311                 int do_list = (root->root_key.objectid !=
1312                                BTRFS_ROOT_TREE_OBJECTID);
1313
1314                 if (*bits & EXTENT_FIRST_DELALLOC)
1315                         *bits &= ~EXTENT_FIRST_DELALLOC;
1316                 else
1317                         atomic_inc(&BTRFS_I(inode)->outstanding_extents);
1318
1319                 spin_lock(&root->fs_info->delalloc_lock);
1320                 BTRFS_I(inode)->delalloc_bytes += len;
1321                 root->fs_info->delalloc_bytes += len;
1322                 if (do_list && list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1323                         list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1324                                       &root->fs_info->delalloc_inodes);
1325                 }
1326                 spin_unlock(&root->fs_info->delalloc_lock);
1327         }
1328         return 0;
1329 }
1330
1331 /*
1332  * extent_io.c clear_bit_hook, see set_bit_hook for why
1333  */
1334 static int btrfs_clear_bit_hook(struct inode *inode,
1335                                 struct extent_state *state, int *bits)
1336 {
1337         /*
1338          * set_bit and clear bit hooks normally require _irqsave/restore
1339          * but in this case, we are only testeing for the DELALLOC
1340          * bit, which is only set or cleared with irqs on
1341          */
1342         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1343                 struct btrfs_root *root = BTRFS_I(inode)->root;
1344                 u64 len = state->end + 1 - state->start;
1345                 int do_list = (root->root_key.objectid !=
1346                                BTRFS_ROOT_TREE_OBJECTID);
1347
1348                 if (*bits & EXTENT_FIRST_DELALLOC)
1349                         *bits &= ~EXTENT_FIRST_DELALLOC;
1350                 else if (!(*bits & EXTENT_DO_ACCOUNTING))
1351                         atomic_dec(&BTRFS_I(inode)->outstanding_extents);
1352
1353                 if (*bits & EXTENT_DO_ACCOUNTING)
1354                         btrfs_delalloc_release_metadata(inode, len);
1355
1356                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1357                     && do_list)
1358                         btrfs_free_reserved_data_space(inode, len);
1359
1360                 spin_lock(&root->fs_info->delalloc_lock);
1361                 root->fs_info->delalloc_bytes -= len;
1362                 BTRFS_I(inode)->delalloc_bytes -= len;
1363
1364                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1365                     !list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1366                         list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1367                 }
1368                 spin_unlock(&root->fs_info->delalloc_lock);
1369         }
1370         return 0;
1371 }
1372
1373 /*
1374  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1375  * we don't create bios that span stripes or chunks
1376  */
1377 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1378                          size_t size, struct bio *bio,
1379                          unsigned long bio_flags)
1380 {
1381         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
1382         struct btrfs_mapping_tree *map_tree;
1383         u64 logical = (u64)bio->bi_sector << 9;
1384         u64 length = 0;
1385         u64 map_length;
1386         int ret;
1387
1388         if (bio_flags & EXTENT_BIO_COMPRESSED)
1389                 return 0;
1390
1391         length = bio->bi_size;
1392         map_tree = &root->fs_info->mapping_tree;
1393         map_length = length;
1394         ret = btrfs_map_block(map_tree, READ, logical,
1395                               &map_length, NULL, 0);
1396
1397         if (map_length < length + size)
1398                 return 1;
1399         return ret;
1400 }
1401
1402 /*
1403  * in order to insert checksums into the metadata in large chunks,
1404  * we wait until bio submission time.   All the pages in the bio are
1405  * checksummed and sums are attached onto the ordered extent record.
1406  *
1407  * At IO completion time the cums attached on the ordered extent record
1408  * are inserted into the btree
1409  */
1410 static int __btrfs_submit_bio_start(struct inode *inode, int rw,
1411                                     struct bio *bio, int mirror_num,
1412                                     unsigned long bio_flags,
1413                                     u64 bio_offset)
1414 {
1415         struct btrfs_root *root = BTRFS_I(inode)->root;
1416         int ret = 0;
1417
1418         ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1419         BUG_ON(ret);
1420         return 0;
1421 }
1422
1423 /*
1424  * in order to insert checksums into the metadata in large chunks,
1425  * we wait until bio submission time.   All the pages in the bio are
1426  * checksummed and sums are attached onto the ordered extent record.
1427  *
1428  * At IO completion time the cums attached on the ordered extent record
1429  * are inserted into the btree
1430  */
1431 static int __btrfs_submit_bio_done(struct inode *inode, int rw, struct bio *bio,
1432                           int mirror_num, unsigned long bio_flags,
1433                           u64 bio_offset)
1434 {
1435         struct btrfs_root *root = BTRFS_I(inode)->root;
1436         return btrfs_map_bio(root, rw, bio, mirror_num, 1);
1437 }
1438
1439 /*
1440  * extent_io.c submission hook. This does the right thing for csum calculation
1441  * on write, or reading the csums from the tree before a read
1442  */
1443 static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
1444                           int mirror_num, unsigned long bio_flags,
1445                           u64 bio_offset)
1446 {
1447         struct btrfs_root *root = BTRFS_I(inode)->root;
1448         int ret = 0;
1449         int skip_sum;
1450
1451         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1452
1453         if (root == root->fs_info->tree_root)
1454                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, 2);
1455         else
1456                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
1457         BUG_ON(ret);
1458
1459         if (!(rw & REQ_WRITE)) {
1460                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1461                         return btrfs_submit_compressed_read(inode, bio,
1462                                                     mirror_num, bio_flags);
1463                 } else if (!skip_sum)
1464                         btrfs_lookup_bio_sums(root, inode, bio, NULL);
1465                 goto mapit;
1466         } else if (!skip_sum) {
1467                 /* csum items have already been cloned */
1468                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1469                         goto mapit;
1470                 /* we're doing a write, do the async checksumming */
1471                 return btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
1472                                    inode, rw, bio, mirror_num,
1473                                    bio_flags, bio_offset,
1474                                    __btrfs_submit_bio_start,
1475                                    __btrfs_submit_bio_done);
1476         }
1477
1478 mapit:
1479         return btrfs_map_bio(root, rw, bio, mirror_num, 0);
1480 }
1481
1482 /*
1483  * given a list of ordered sums record them in the inode.  This happens
1484  * at IO completion time based on sums calculated at bio submission time.
1485  */
1486 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1487                              struct inode *inode, u64 file_offset,
1488                              struct list_head *list)
1489 {
1490         struct btrfs_ordered_sum *sum;
1491
1492         btrfs_set_trans_block_group(trans, inode);
1493
1494         list_for_each_entry(sum, list, list) {
1495                 btrfs_csum_file_blocks(trans,
1496                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1497         }
1498         return 0;
1499 }
1500
1501 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
1502                               struct extent_state **cached_state)
1503 {
1504         if ((end & (PAGE_CACHE_SIZE - 1)) == 0)
1505                 WARN_ON(1);
1506         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
1507                                    cached_state, GFP_NOFS);
1508 }
1509
1510 /* see btrfs_writepage_start_hook for details on why this is required */
1511 struct btrfs_writepage_fixup {
1512         struct page *page;
1513         struct btrfs_work work;
1514 };
1515
1516 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
1517 {
1518         struct btrfs_writepage_fixup *fixup;
1519         struct btrfs_ordered_extent *ordered;
1520         struct extent_state *cached_state = NULL;
1521         struct page *page;
1522         struct inode *inode;
1523         u64 page_start;
1524         u64 page_end;
1525
1526         fixup = container_of(work, struct btrfs_writepage_fixup, work);
1527         page = fixup->page;
1528 again:
1529         lock_page(page);
1530         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
1531                 ClearPageChecked(page);
1532                 goto out_page;
1533         }
1534
1535         inode = page->mapping->host;
1536         page_start = page_offset(page);
1537         page_end = page_offset(page) + PAGE_CACHE_SIZE - 1;
1538
1539         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end, 0,
1540                          &cached_state, GFP_NOFS);
1541
1542         /* already ordered? We're done */
1543         if (PagePrivate2(page))
1544                 goto out;
1545
1546         ordered = btrfs_lookup_ordered_extent(inode, page_start);
1547         if (ordered) {
1548                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
1549                                      page_end, &cached_state, GFP_NOFS);
1550                 unlock_page(page);
1551                 btrfs_start_ordered_extent(inode, ordered, 1);
1552                 goto again;
1553         }
1554
1555         BUG();
1556         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state);
1557         ClearPageChecked(page);
1558 out:
1559         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
1560                              &cached_state, GFP_NOFS);
1561 out_page:
1562         unlock_page(page);
1563         page_cache_release(page);
1564         kfree(fixup);
1565 }
1566
1567 /*
1568  * There are a few paths in the higher layers of the kernel that directly
1569  * set the page dirty bit without asking the filesystem if it is a
1570  * good idea.  This causes problems because we want to make sure COW
1571  * properly happens and the data=ordered rules are followed.
1572  *
1573  * In our case any range that doesn't have the ORDERED bit set
1574  * hasn't been properly setup for IO.  We kick off an async process
1575  * to fix it up.  The async helper will wait for ordered extents, set
1576  * the delalloc bit and make it safe to write the page.
1577  */
1578 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
1579 {
1580         struct inode *inode = page->mapping->host;
1581         struct btrfs_writepage_fixup *fixup;
1582         struct btrfs_root *root = BTRFS_I(inode)->root;
1583
1584         /* this page is properly in the ordered list */
1585         if (TestClearPagePrivate2(page))
1586                 return 0;
1587
1588         if (PageChecked(page))
1589                 return -EAGAIN;
1590
1591         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
1592         if (!fixup)
1593                 return -EAGAIN;
1594
1595         SetPageChecked(page);
1596         page_cache_get(page);
1597         fixup->work.func = btrfs_writepage_fixup_worker;
1598         fixup->page = page;
1599         btrfs_queue_worker(&root->fs_info->fixup_workers, &fixup->work);
1600         return -EAGAIN;
1601 }
1602
1603 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
1604                                        struct inode *inode, u64 file_pos,
1605                                        u64 disk_bytenr, u64 disk_num_bytes,
1606                                        u64 num_bytes, u64 ram_bytes,
1607                                        u8 compression, u8 encryption,
1608                                        u16 other_encoding, int extent_type)
1609 {
1610         struct btrfs_root *root = BTRFS_I(inode)->root;
1611         struct btrfs_file_extent_item *fi;
1612         struct btrfs_path *path;
1613         struct extent_buffer *leaf;
1614         struct btrfs_key ins;
1615         u64 hint;
1616         int ret;
1617
1618         path = btrfs_alloc_path();
1619         BUG_ON(!path);
1620
1621         path->leave_spinning = 1;
1622
1623         /*
1624          * we may be replacing one extent in the tree with another.
1625          * The new extent is pinned in the extent map, and we don't want
1626          * to drop it from the cache until it is completely in the btree.
1627          *
1628          * So, tell btrfs_drop_extents to leave this extent in the cache.
1629          * the caller is expected to unpin it and allow it to be merged
1630          * with the others.
1631          */
1632         ret = btrfs_drop_extents(trans, inode, file_pos, file_pos + num_bytes,
1633                                  &hint, 0);
1634         BUG_ON(ret);
1635
1636         ins.objectid = inode->i_ino;
1637         ins.offset = file_pos;
1638         ins.type = BTRFS_EXTENT_DATA_KEY;
1639         ret = btrfs_insert_empty_item(trans, root, path, &ins, sizeof(*fi));
1640         BUG_ON(ret);
1641         leaf = path->nodes[0];
1642         fi = btrfs_item_ptr(leaf, path->slots[0],
1643                             struct btrfs_file_extent_item);
1644         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1645         btrfs_set_file_extent_type(leaf, fi, extent_type);
1646         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
1647         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
1648         btrfs_set_file_extent_offset(leaf, fi, 0);
1649         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
1650         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
1651         btrfs_set_file_extent_compression(leaf, fi, compression);
1652         btrfs_set_file_extent_encryption(leaf, fi, encryption);
1653         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
1654
1655         btrfs_unlock_up_safe(path, 1);
1656         btrfs_set_lock_blocking(leaf);
1657
1658         btrfs_mark_buffer_dirty(leaf);
1659
1660         inode_add_bytes(inode, num_bytes);
1661
1662         ins.objectid = disk_bytenr;
1663         ins.offset = disk_num_bytes;
1664         ins.type = BTRFS_EXTENT_ITEM_KEY;
1665         ret = btrfs_alloc_reserved_file_extent(trans, root,
1666                                         root->root_key.objectid,
1667                                         inode->i_ino, file_pos, &ins);
1668         BUG_ON(ret);
1669         btrfs_free_path(path);
1670
1671         return 0;
1672 }
1673
1674 /*
1675  * helper function for btrfs_finish_ordered_io, this
1676  * just reads in some of the csum leaves to prime them into ram
1677  * before we start the transaction.  It limits the amount of btree
1678  * reads required while inside the transaction.
1679  */
1680 /* as ordered data IO finishes, this gets called so we can finish
1681  * an ordered extent if the range of bytes in the file it covers are
1682  * fully written.
1683  */
1684 static int btrfs_finish_ordered_io(struct inode *inode, u64 start, u64 end)
1685 {
1686         struct btrfs_root *root = BTRFS_I(inode)->root;
1687         struct btrfs_trans_handle *trans = NULL;
1688         struct btrfs_ordered_extent *ordered_extent = NULL;
1689         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
1690         struct extent_state *cached_state = NULL;
1691         int compress_type = 0;
1692         int ret;
1693         bool nolock = false;
1694
1695         ret = btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
1696                                              end - start + 1);
1697         if (!ret)
1698                 return 0;
1699         BUG_ON(!ordered_extent);
1700
1701         nolock = (root == root->fs_info->tree_root);
1702
1703         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
1704                 BUG_ON(!list_empty(&ordered_extent->list));
1705                 ret = btrfs_ordered_update_i_size(inode, 0, ordered_extent);
1706                 if (!ret) {
1707                         if (nolock)
1708                                 trans = btrfs_join_transaction_nolock(root, 1);
1709                         else
1710                                 trans = btrfs_join_transaction(root, 1);
1711                         BUG_ON(IS_ERR(trans));
1712                         btrfs_set_trans_block_group(trans, inode);
1713                         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1714                         ret = btrfs_update_inode(trans, root, inode);
1715                         BUG_ON(ret);
1716                 }
1717                 goto out;
1718         }
1719
1720         lock_extent_bits(io_tree, ordered_extent->file_offset,
1721                          ordered_extent->file_offset + ordered_extent->len - 1,
1722                          0, &cached_state, GFP_NOFS);
1723
1724         if (nolock)
1725                 trans = btrfs_join_transaction_nolock(root, 1);
1726         else
1727                 trans = btrfs_join_transaction(root, 1);
1728         BUG_ON(IS_ERR(trans));
1729         btrfs_set_trans_block_group(trans, inode);
1730         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1731
1732         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
1733                 compress_type = ordered_extent->compress_type;
1734         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
1735                 BUG_ON(compress_type);
1736                 ret = btrfs_mark_extent_written(trans, inode,
1737                                                 ordered_extent->file_offset,
1738                                                 ordered_extent->file_offset +
1739                                                 ordered_extent->len);
1740                 BUG_ON(ret);
1741         } else {
1742                 BUG_ON(root == root->fs_info->tree_root);
1743                 ret = insert_reserved_file_extent(trans, inode,
1744                                                 ordered_extent->file_offset,
1745                                                 ordered_extent->start,
1746                                                 ordered_extent->disk_len,
1747                                                 ordered_extent->len,
1748                                                 ordered_extent->len,
1749                                                 compress_type, 0, 0,
1750                                                 BTRFS_FILE_EXTENT_REG);
1751                 unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
1752                                    ordered_extent->file_offset,
1753                                    ordered_extent->len);
1754                 BUG_ON(ret);
1755         }
1756         unlock_extent_cached(io_tree, ordered_extent->file_offset,
1757                              ordered_extent->file_offset +
1758                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
1759
1760         add_pending_csums(trans, inode, ordered_extent->file_offset,
1761                           &ordered_extent->list);
1762
1763         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
1764         ret = btrfs_update_inode(trans, root, inode);
1765         BUG_ON(ret);
1766 out:
1767         if (nolock) {
1768                 if (trans)
1769                         btrfs_end_transaction_nolock(trans, root);
1770         } else {
1771                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
1772                 if (trans)
1773                         btrfs_end_transaction(trans, root);
1774         }
1775
1776         /* once for us */
1777         btrfs_put_ordered_extent(ordered_extent);
1778         /* once for the tree */
1779         btrfs_put_ordered_extent(ordered_extent);
1780
1781         return 0;
1782 }
1783
1784 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
1785                                 struct extent_state *state, int uptodate)
1786 {
1787         ClearPagePrivate2(page);
1788         return btrfs_finish_ordered_io(page->mapping->host, start, end);
1789 }
1790
1791 /*
1792  * When IO fails, either with EIO or csum verification fails, we
1793  * try other mirrors that might have a good copy of the data.  This
1794  * io_failure_record is used to record state as we go through all the
1795  * mirrors.  If another mirror has good data, the page is set up to date
1796  * and things continue.  If a good mirror can't be found, the original
1797  * bio end_io callback is called to indicate things have failed.
1798  */
1799 struct io_failure_record {
1800         struct page *page;
1801         u64 start;
1802         u64 len;
1803         u64 logical;
1804         unsigned long bio_flags;
1805         int last_mirror;
1806 };
1807
1808 static int btrfs_io_failed_hook(struct bio *failed_bio,
1809                          struct page *page, u64 start, u64 end,
1810                          struct extent_state *state)
1811 {
1812         struct io_failure_record *failrec = NULL;
1813         u64 private;
1814         struct extent_map *em;
1815         struct inode *inode = page->mapping->host;
1816         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1817         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
1818         struct bio *bio;
1819         int num_copies;
1820         int ret;
1821         int rw;
1822         u64 logical;
1823
1824         ret = get_state_private(failure_tree, start, &private);
1825         if (ret) {
1826                 failrec = kmalloc(sizeof(*failrec), GFP_NOFS);
1827                 if (!failrec)
1828                         return -ENOMEM;
1829                 failrec->start = start;
1830                 failrec->len = end - start + 1;
1831                 failrec->last_mirror = 0;
1832                 failrec->bio_flags = 0;
1833
1834                 read_lock(&em_tree->lock);
1835                 em = lookup_extent_mapping(em_tree, start, failrec->len);
1836                 if (em->start > start || em->start + em->len < start) {
1837                         free_extent_map(em);
1838                         em = NULL;
1839                 }
1840                 read_unlock(&em_tree->lock);
1841
1842                 if (!em || IS_ERR(em)) {
1843                         kfree(failrec);
1844                         return -EIO;
1845                 }
1846                 logical = start - em->start;
1847                 logical = em->block_start + logical;
1848                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
1849                         logical = em->block_start;
1850                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
1851                         extent_set_compress_type(&failrec->bio_flags,
1852                                                  em->compress_type);
1853                 }
1854                 failrec->logical = logical;
1855                 free_extent_map(em);
1856                 set_extent_bits(failure_tree, start, end, EXTENT_LOCKED |
1857                                 EXTENT_DIRTY, GFP_NOFS);
1858                 set_state_private(failure_tree, start,
1859                                  (u64)(unsigned long)failrec);
1860         } else {
1861                 failrec = (struct io_failure_record *)(unsigned long)private;
1862         }
1863         num_copies = btrfs_num_copies(
1864                               &BTRFS_I(inode)->root->fs_info->mapping_tree,
1865                               failrec->logical, failrec->len);
1866         failrec->last_mirror++;
1867         if (!state) {
1868                 spin_lock(&BTRFS_I(inode)->io_tree.lock);
1869                 state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
1870                                                     failrec->start,
1871                                                     EXTENT_LOCKED);
1872                 if (state && state->start != failrec->start)
1873                         state = NULL;
1874                 spin_unlock(&BTRFS_I(inode)->io_tree.lock);
1875         }
1876         if (!state || failrec->last_mirror > num_copies) {
1877                 set_state_private(failure_tree, failrec->start, 0);
1878                 clear_extent_bits(failure_tree, failrec->start,
1879                                   failrec->start + failrec->len - 1,
1880                                   EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1881                 kfree(failrec);
1882                 return -EIO;
1883         }
1884         bio = bio_alloc(GFP_NOFS, 1);
1885         bio->bi_private = state;
1886         bio->bi_end_io = failed_bio->bi_end_io;
1887         bio->bi_sector = failrec->logical >> 9;
1888         bio->bi_bdev = failed_bio->bi_bdev;
1889         bio->bi_size = 0;
1890
1891         bio_add_page(bio, page, failrec->len, start - page_offset(page));
1892         if (failed_bio->bi_rw & REQ_WRITE)
1893                 rw = WRITE;
1894         else
1895                 rw = READ;
1896
1897         BTRFS_I(inode)->io_tree.ops->submit_bio_hook(inode, rw, bio,
1898                                                       failrec->last_mirror,
1899                                                       failrec->bio_flags, 0);
1900         return 0;
1901 }
1902
1903 /*
1904  * each time an IO finishes, we do a fast check in the IO failure tree
1905  * to see if we need to process or clean up an io_failure_record
1906  */
1907 static int btrfs_clean_io_failures(struct inode *inode, u64 start)
1908 {
1909         u64 private;
1910         u64 private_failure;
1911         struct io_failure_record *failure;
1912         int ret;
1913
1914         private = 0;
1915         if (count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
1916                              (u64)-1, 1, EXTENT_DIRTY, 0)) {
1917                 ret = get_state_private(&BTRFS_I(inode)->io_failure_tree,
1918                                         start, &private_failure);
1919                 if (ret == 0) {
1920                         failure = (struct io_failure_record *)(unsigned long)
1921                                    private_failure;
1922                         set_state_private(&BTRFS_I(inode)->io_failure_tree,
1923                                           failure->start, 0);
1924                         clear_extent_bits(&BTRFS_I(inode)->io_failure_tree,
1925                                           failure->start,
1926                                           failure->start + failure->len - 1,
1927                                           EXTENT_DIRTY | EXTENT_LOCKED,
1928                                           GFP_NOFS);
1929                         kfree(failure);
1930                 }
1931         }
1932         return 0;
1933 }
1934
1935 /*
1936  * when reads are done, we need to check csums to verify the data is correct
1937  * if there's a match, we allow the bio to finish.  If not, we go through
1938  * the io_failure_record routines to find good copies
1939  */
1940 static int btrfs_readpage_end_io_hook(struct page *page, u64 start, u64 end,
1941                                struct extent_state *state)
1942 {
1943         size_t offset = start - ((u64)page->index << PAGE_CACHE_SHIFT);
1944         struct inode *inode = page->mapping->host;
1945         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
1946         char *kaddr;
1947         u64 private = ~(u32)0;
1948         int ret;
1949         struct btrfs_root *root = BTRFS_I(inode)->root;
1950         u32 csum = ~(u32)0;
1951
1952         if (PageChecked(page)) {
1953                 ClearPageChecked(page);
1954                 goto good;
1955         }
1956
1957         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
1958                 return 0;
1959
1960         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
1961             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
1962                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM,
1963                                   GFP_NOFS);
1964                 return 0;
1965         }
1966
1967         if (state && state->start == start) {
1968                 private = state->private;
1969                 ret = 0;
1970         } else {
1971                 ret = get_state_private(io_tree, start, &private);
1972         }
1973         kaddr = kmap_atomic(page, KM_USER0);
1974         if (ret)
1975                 goto zeroit;
1976
1977         csum = btrfs_csum_data(root, kaddr + offset, csum,  end - start + 1);
1978         btrfs_csum_final(csum, (char *)&csum);
1979         if (csum != private)
1980                 goto zeroit;
1981
1982         kunmap_atomic(kaddr, KM_USER0);
1983 good:
1984         /* if the io failure tree for this inode is non-empty,
1985          * check to see if we've recovered from a failed IO
1986          */
1987         btrfs_clean_io_failures(inode, start);
1988         return 0;
1989
1990 zeroit:
1991         if (printk_ratelimit()) {
1992                 printk(KERN_INFO "btrfs csum failed ino %lu off %llu csum %u "
1993                        "private %llu\n", page->mapping->host->i_ino,
1994                        (unsigned long long)start, csum,
1995                        (unsigned long long)private);
1996         }
1997         memset(kaddr + offset, 1, end - start + 1);
1998         flush_dcache_page(page);
1999         kunmap_atomic(kaddr, KM_USER0);
2000         if (private == 0)
2001                 return 0;
2002         return -EIO;
2003 }
2004
2005 struct delayed_iput {
2006         struct list_head list;
2007         struct inode *inode;
2008 };
2009
2010 void btrfs_add_delayed_iput(struct inode *inode)
2011 {
2012         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2013         struct delayed_iput *delayed;
2014
2015         if (atomic_add_unless(&inode->i_count, -1, 1))
2016                 return;
2017
2018         delayed = kmalloc(sizeof(*delayed), GFP_NOFS | __GFP_NOFAIL);
2019         delayed->inode = inode;
2020
2021         spin_lock(&fs_info->delayed_iput_lock);
2022         list_add_tail(&delayed->list, &fs_info->delayed_iputs);
2023         spin_unlock(&fs_info->delayed_iput_lock);
2024 }
2025
2026 void btrfs_run_delayed_iputs(struct btrfs_root *root)
2027 {
2028         LIST_HEAD(list);
2029         struct btrfs_fs_info *fs_info = root->fs_info;
2030         struct delayed_iput *delayed;
2031         int empty;
2032
2033         spin_lock(&fs_info->delayed_iput_lock);
2034         empty = list_empty(&fs_info->delayed_iputs);
2035         spin_unlock(&fs_info->delayed_iput_lock);
2036         if (empty)
2037                 return;
2038
2039         down_read(&root->fs_info->cleanup_work_sem);
2040         spin_lock(&fs_info->delayed_iput_lock);
2041         list_splice_init(&fs_info->delayed_iputs, &list);
2042         spin_unlock(&fs_info->delayed_iput_lock);
2043
2044         while (!list_empty(&list)) {
2045                 delayed = list_entry(list.next, struct delayed_iput, list);
2046                 list_del(&delayed->list);
2047                 iput(delayed->inode);
2048                 kfree(delayed);
2049         }
2050         up_read(&root->fs_info->cleanup_work_sem);
2051 }
2052
2053 /*
2054  * calculate extra metadata reservation when snapshotting a subvolume
2055  * contains orphan files.
2056  */
2057 void btrfs_orphan_pre_snapshot(struct btrfs_trans_handle *trans,
2058                                 struct btrfs_pending_snapshot *pending,
2059                                 u64 *bytes_to_reserve)
2060 {
2061         struct btrfs_root *root;
2062         struct btrfs_block_rsv *block_rsv;
2063         u64 num_bytes;
2064         int index;
2065
2066         root = pending->root;
2067         if (!root->orphan_block_rsv || list_empty(&root->orphan_list))
2068                 return;
2069
2070         block_rsv = root->orphan_block_rsv;
2071
2072         /* orphan block reservation for the snapshot */
2073         num_bytes = block_rsv->size;
2074
2075         /*
2076          * after the snapshot is created, COWing tree blocks may use more
2077          * space than it frees. So we should make sure there is enough
2078          * reserved space.
2079          */
2080         index = trans->transid & 0x1;
2081         if (block_rsv->reserved + block_rsv->freed[index] < block_rsv->size) {
2082                 num_bytes += block_rsv->size -
2083                              (block_rsv->reserved + block_rsv->freed[index]);
2084         }
2085
2086         *bytes_to_reserve += num_bytes;
2087 }
2088
2089 void btrfs_orphan_post_snapshot(struct btrfs_trans_handle *trans,
2090                                 struct btrfs_pending_snapshot *pending)
2091 {
2092         struct btrfs_root *root = pending->root;
2093         struct btrfs_root *snap = pending->snap;
2094         struct btrfs_block_rsv *block_rsv;
2095         u64 num_bytes;
2096         int index;
2097         int ret;
2098
2099         if (!root->orphan_block_rsv || list_empty(&root->orphan_list))
2100                 return;
2101
2102         /* refill source subvolume's orphan block reservation */
2103         block_rsv = root->orphan_block_rsv;
2104         index = trans->transid & 0x1;
2105         if (block_rsv->reserved + block_rsv->freed[index] < block_rsv->size) {
2106                 num_bytes = block_rsv->size -
2107                             (block_rsv->reserved + block_rsv->freed[index]);
2108                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
2109                                               root->orphan_block_rsv,
2110                                               num_bytes);
2111                 BUG_ON(ret);
2112         }
2113
2114         /* setup orphan block reservation for the snapshot */
2115         block_rsv = btrfs_alloc_block_rsv(snap);
2116         BUG_ON(!block_rsv);
2117
2118         btrfs_add_durable_block_rsv(root->fs_info, block_rsv);
2119         snap->orphan_block_rsv = block_rsv;
2120
2121         num_bytes = root->orphan_block_rsv->size;
2122         ret = btrfs_block_rsv_migrate(&pending->block_rsv,
2123                                       block_rsv, num_bytes);
2124         BUG_ON(ret);
2125
2126 #if 0
2127         /* insert orphan item for the snapshot */
2128         WARN_ON(!root->orphan_item_inserted);
2129         ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
2130                                        snap->root_key.objectid);
2131         BUG_ON(ret);
2132         snap->orphan_item_inserted = 1;
2133 #endif
2134 }
2135
2136 enum btrfs_orphan_cleanup_state {
2137         ORPHAN_CLEANUP_STARTED  = 1,
2138         ORPHAN_CLEANUP_DONE     = 2,
2139 };
2140
2141 /*
2142  * This is called in transaction commmit time. If there are no orphan
2143  * files in the subvolume, it removes orphan item and frees block_rsv
2144  * structure.
2145  */
2146 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
2147                               struct btrfs_root *root)
2148 {
2149         int ret;
2150
2151         if (!list_empty(&root->orphan_list) ||
2152             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
2153                 return;
2154
2155         if (root->orphan_item_inserted &&
2156             btrfs_root_refs(&root->root_item) > 0) {
2157                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
2158                                             root->root_key.objectid);
2159                 BUG_ON(ret);
2160                 root->orphan_item_inserted = 0;
2161         }
2162
2163         if (root->orphan_block_rsv) {
2164                 WARN_ON(root->orphan_block_rsv->size > 0);
2165                 btrfs_free_block_rsv(root, root->orphan_block_rsv);
2166                 root->orphan_block_rsv = NULL;
2167         }
2168 }
2169
2170 /*
2171  * This creates an orphan entry for the given inode in case something goes
2172  * wrong in the middle of an unlink/truncate.
2173  *
2174  * NOTE: caller of this function should reserve 5 units of metadata for
2175  *       this function.
2176  */
2177 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
2178 {
2179         struct btrfs_root *root = BTRFS_I(inode)->root;
2180         struct btrfs_block_rsv *block_rsv = NULL;
2181         int reserve = 0;
2182         int insert = 0;
2183         int ret;
2184
2185         if (!root->orphan_block_rsv) {
2186                 block_rsv = btrfs_alloc_block_rsv(root);
2187                 BUG_ON(!block_rsv);
2188         }
2189
2190         spin_lock(&root->orphan_lock);
2191         if (!root->orphan_block_rsv) {
2192                 root->orphan_block_rsv = block_rsv;
2193         } else if (block_rsv) {
2194                 btrfs_free_block_rsv(root, block_rsv);
2195                 block_rsv = NULL;
2196         }
2197
2198         if (list_empty(&BTRFS_I(inode)->i_orphan)) {
2199                 list_add(&BTRFS_I(inode)->i_orphan, &root->orphan_list);
2200 #if 0
2201                 /*
2202                  * For proper ENOSPC handling, we should do orphan
2203                  * cleanup when mounting. But this introduces backward
2204                  * compatibility issue.
2205                  */
2206                 if (!xchg(&root->orphan_item_inserted, 1))
2207                         insert = 2;
2208                 else
2209                         insert = 1;
2210 #endif
2211                 insert = 1;
2212         } else {
2213                 WARN_ON(!BTRFS_I(inode)->orphan_meta_reserved);
2214         }
2215
2216         if (!BTRFS_I(inode)->orphan_meta_reserved) {
2217                 BTRFS_I(inode)->orphan_meta_reserved = 1;
2218                 reserve = 1;
2219         }
2220         spin_unlock(&root->orphan_lock);
2221
2222         if (block_rsv)
2223                 btrfs_add_durable_block_rsv(root->fs_info, block_rsv);
2224
2225         /* grab metadata reservation from transaction handle */
2226         if (reserve) {
2227                 ret = btrfs_orphan_reserve_metadata(trans, inode);
2228                 BUG_ON(ret);
2229         }
2230
2231         /* insert an orphan item to track this unlinked/truncated file */
2232         if (insert >= 1) {
2233                 ret = btrfs_insert_orphan_item(trans, root, inode->i_ino);
2234                 BUG_ON(ret);
2235         }
2236
2237         /* insert an orphan item to track subvolume contains orphan files */
2238         if (insert >= 2) {
2239                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
2240                                                root->root_key.objectid);
2241                 BUG_ON(ret);
2242         }
2243         return 0;
2244 }
2245
2246 /*
2247  * We have done the truncate/delete so we can go ahead and remove the orphan
2248  * item for this particular inode.
2249  */
2250 int btrfs_orphan_del(struct btrfs_trans_handle *trans, struct inode *inode)
2251 {
2252         struct btrfs_root *root = BTRFS_I(inode)->root;
2253         int delete_item = 0;
2254         int release_rsv = 0;
2255         int ret = 0;
2256
2257         spin_lock(&root->orphan_lock);
2258         if (!list_empty(&BTRFS_I(inode)->i_orphan)) {
2259                 list_del_init(&BTRFS_I(inode)->i_orphan);
2260                 delete_item = 1;
2261         }
2262
2263         if (BTRFS_I(inode)->orphan_meta_reserved) {
2264                 BTRFS_I(inode)->orphan_meta_reserved = 0;
2265                 release_rsv = 1;
2266         }
2267         spin_unlock(&root->orphan_lock);
2268
2269         if (trans && delete_item) {
2270                 ret = btrfs_del_orphan_item(trans, root, inode->i_ino);
2271                 BUG_ON(ret);
2272         }
2273
2274         if (release_rsv)
2275                 btrfs_orphan_release_metadata(inode);
2276
2277         return 0;
2278 }
2279
2280 /*
2281  * this cleans up any orphans that may be left on the list from the last use
2282  * of this root.
2283  */
2284 void btrfs_orphan_cleanup(struct btrfs_root *root)
2285 {
2286         struct btrfs_path *path;
2287         struct extent_buffer *leaf;
2288         struct btrfs_key key, found_key;
2289         struct btrfs_trans_handle *trans;
2290         struct inode *inode;
2291         int ret = 0, nr_unlink = 0, nr_truncate = 0;
2292
2293         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
2294                 return;
2295
2296         path = btrfs_alloc_path();
2297         BUG_ON(!path);
2298         path->reada = -1;
2299
2300         key.objectid = BTRFS_ORPHAN_OBJECTID;
2301         btrfs_set_key_type(&key, BTRFS_ORPHAN_ITEM_KEY);
2302         key.offset = (u64)-1;
2303
2304         while (1) {
2305                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2306                 if (ret < 0) {
2307                         printk(KERN_ERR "Error searching slot for orphan: %d"
2308                                "\n", ret);
2309                         break;
2310                 }
2311
2312                 /*
2313                  * if ret == 0 means we found what we were searching for, which
2314                  * is weird, but possible, so only screw with path if we didnt
2315                  * find the key and see if we have stuff that matches
2316                  */
2317                 if (ret > 0) {
2318                         if (path->slots[0] == 0)
2319                                 break;
2320                         path->slots[0]--;
2321                 }
2322
2323                 /* pull out the item */
2324                 leaf = path->nodes[0];
2325                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2326
2327                 /* make sure the item matches what we want */
2328                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
2329                         break;
2330                 if (btrfs_key_type(&found_key) != BTRFS_ORPHAN_ITEM_KEY)
2331                         break;
2332
2333                 /* release the path since we're done with it */
2334                 btrfs_release_path(root, path);
2335
2336                 /*
2337                  * this is where we are basically btrfs_lookup, without the
2338                  * crossing root thing.  we store the inode number in the
2339                  * offset of the orphan item.
2340                  */
2341                 found_key.objectid = found_key.offset;
2342                 found_key.type = BTRFS_INODE_ITEM_KEY;
2343                 found_key.offset = 0;
2344                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
2345                 BUG_ON(IS_ERR(inode));
2346
2347                 /*
2348                  * add this inode to the orphan list so btrfs_orphan_del does
2349                  * the proper thing when we hit it
2350                  */
2351                 spin_lock(&root->orphan_lock);
2352                 list_add(&BTRFS_I(inode)->i_orphan, &root->orphan_list);
2353                 spin_unlock(&root->orphan_lock);
2354
2355                 /*
2356                  * if this is a bad inode, means we actually succeeded in
2357                  * removing the inode, but not the orphan record, which means
2358                  * we need to manually delete the orphan since iput will just
2359                  * do a destroy_inode
2360                  */
2361                 if (is_bad_inode(inode)) {
2362                         trans = btrfs_start_transaction(root, 0);
2363                         BUG_ON(IS_ERR(trans));
2364                         btrfs_orphan_del(trans, inode);
2365                         btrfs_end_transaction(trans, root);
2366                         iput(inode);
2367                         continue;
2368                 }
2369
2370                 /* if we have links, this was a truncate, lets do that */
2371                 if (inode->i_nlink) {
2372                         nr_truncate++;
2373                         btrfs_truncate(inode);
2374                 } else {
2375                         nr_unlink++;
2376                 }
2377
2378                 /* this will do delete_inode and everything for us */
2379                 iput(inode);
2380         }
2381         btrfs_free_path(path);
2382
2383         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
2384
2385         if (root->orphan_block_rsv)
2386                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
2387                                         (u64)-1);
2388
2389         if (root->orphan_block_rsv || root->orphan_item_inserted) {
2390                 trans = btrfs_join_transaction(root, 1);
2391                 BUG_ON(IS_ERR(trans));
2392                 btrfs_end_transaction(trans, root);
2393         }
2394
2395         if (nr_unlink)
2396                 printk(KERN_INFO "btrfs: unlinked %d orphans\n", nr_unlink);
2397         if (nr_truncate)
2398                 printk(KERN_INFO "btrfs: truncated %d orphans\n", nr_truncate);
2399 }
2400
2401 /*
2402  * very simple check to peek ahead in the leaf looking for xattrs.  If we
2403  * don't find any xattrs, we know there can't be any acls.
2404  *
2405  * slot is the slot the inode is in, objectid is the objectid of the inode
2406  */
2407 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
2408                                           int slot, u64 objectid)
2409 {
2410         u32 nritems = btrfs_header_nritems(leaf);
2411         struct btrfs_key found_key;
2412         int scanned = 0;
2413
2414         slot++;
2415         while (slot < nritems) {
2416                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
2417
2418                 /* we found a different objectid, there must not be acls */
2419                 if (found_key.objectid != objectid)
2420                         return 0;
2421
2422                 /* we found an xattr, assume we've got an acl */
2423                 if (found_key.type == BTRFS_XATTR_ITEM_KEY)
2424                         return 1;
2425
2426                 /*
2427                  * we found a key greater than an xattr key, there can't
2428                  * be any acls later on
2429                  */
2430                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
2431                         return 0;
2432
2433                 slot++;
2434                 scanned++;
2435
2436                 /*
2437                  * it goes inode, inode backrefs, xattrs, extents,
2438                  * so if there are a ton of hard links to an inode there can
2439                  * be a lot of backrefs.  Don't waste time searching too hard,
2440                  * this is just an optimization
2441                  */
2442                 if (scanned >= 8)
2443                         break;
2444         }
2445         /* we hit the end of the leaf before we found an xattr or
2446          * something larger than an xattr.  We have to assume the inode
2447          * has acls
2448          */
2449         return 1;
2450 }
2451
2452 /*
2453  * read an inode from the btree into the in-memory inode
2454  */
2455 static void btrfs_read_locked_inode(struct inode *inode)
2456 {
2457         struct btrfs_path *path;
2458         struct extent_buffer *leaf;
2459         struct btrfs_inode_item *inode_item;
2460         struct btrfs_timespec *tspec;
2461         struct btrfs_root *root = BTRFS_I(inode)->root;
2462         struct btrfs_key location;
2463         int maybe_acls;
2464         u64 alloc_group_block;
2465         u32 rdev;
2466         int ret;
2467
2468         path = btrfs_alloc_path();
2469         BUG_ON(!path);
2470         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
2471
2472         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
2473         if (ret)
2474                 goto make_bad;
2475
2476         leaf = path->nodes[0];
2477         inode_item = btrfs_item_ptr(leaf, path->slots[0],
2478                                     struct btrfs_inode_item);
2479
2480         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
2481         inode->i_nlink = btrfs_inode_nlink(leaf, inode_item);
2482         inode->i_uid = btrfs_inode_uid(leaf, inode_item);
2483         inode->i_gid = btrfs_inode_gid(leaf, inode_item);
2484         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
2485
2486         tspec = btrfs_inode_atime(inode_item);
2487         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, tspec);
2488         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
2489
2490         tspec = btrfs_inode_mtime(inode_item);
2491         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, tspec);
2492         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
2493
2494         tspec = btrfs_inode_ctime(inode_item);
2495         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, tspec);
2496         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
2497
2498         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
2499         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
2500         BTRFS_I(inode)->sequence = btrfs_inode_sequence(leaf, inode_item);
2501         inode->i_generation = BTRFS_I(inode)->generation;
2502         inode->i_rdev = 0;
2503         rdev = btrfs_inode_rdev(leaf, inode_item);
2504
2505         BTRFS_I(inode)->index_cnt = (u64)-1;
2506         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
2507
2508         alloc_group_block = btrfs_inode_block_group(leaf, inode_item);
2509
2510         /*
2511          * try to precache a NULL acl entry for files that don't have
2512          * any xattrs or acls
2513          */
2514         maybe_acls = acls_after_inode_item(leaf, path->slots[0], inode->i_ino);
2515         if (!maybe_acls)
2516                 cache_no_acl(inode);
2517
2518         BTRFS_I(inode)->block_group = btrfs_find_block_group(root, 0,
2519                                                 alloc_group_block, 0);
2520         btrfs_free_path(path);
2521         inode_item = NULL;
2522
2523         switch (inode->i_mode & S_IFMT) {
2524         case S_IFREG:
2525                 inode->i_mapping->a_ops = &btrfs_aops;
2526                 inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
2527                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
2528                 inode->i_fop = &btrfs_file_operations;
2529                 inode->i_op = &btrfs_file_inode_operations;
2530                 break;
2531         case S_IFDIR:
2532                 inode->i_fop = &btrfs_dir_file_operations;
2533                 if (root == root->fs_info->tree_root)
2534                         inode->i_op = &btrfs_dir_ro_inode_operations;
2535                 else
2536                         inode->i_op = &btrfs_dir_inode_operations;
2537                 break;
2538         case S_IFLNK:
2539                 inode->i_op = &btrfs_symlink_inode_operations;
2540                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
2541                 inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
2542                 break;
2543         default:
2544                 inode->i_op = &btrfs_special_inode_operations;
2545                 init_special_inode(inode, inode->i_mode, rdev);
2546                 break;
2547         }
2548
2549         btrfs_update_iflags(inode);
2550         return;
2551
2552 make_bad:
2553         btrfs_free_path(path);
2554         make_bad_inode(inode);
2555 }
2556
2557 /*
2558  * given a leaf and an inode, copy the inode fields into the leaf
2559  */
2560 static void fill_inode_item(struct btrfs_trans_handle *trans,
2561                             struct extent_buffer *leaf,
2562                             struct btrfs_inode_item *item,
2563                             struct inode *inode)
2564 {
2565         btrfs_set_inode_uid(leaf, item, inode->i_uid);
2566         btrfs_set_inode_gid(leaf, item, inode->i_gid);
2567         btrfs_set_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size);
2568         btrfs_set_inode_mode(leaf, item, inode->i_mode);
2569         btrfs_set_inode_nlink(leaf, item, inode->i_nlink);
2570
2571         btrfs_set_timespec_sec(leaf, btrfs_inode_atime(item),
2572                                inode->i_atime.tv_sec);
2573         btrfs_set_timespec_nsec(leaf, btrfs_inode_atime(item),
2574                                 inode->i_atime.tv_nsec);
2575
2576         btrfs_set_timespec_sec(leaf, btrfs_inode_mtime(item),
2577                                inode->i_mtime.tv_sec);
2578         btrfs_set_timespec_nsec(leaf, btrfs_inode_mtime(item),
2579                                 inode->i_mtime.tv_nsec);
2580
2581         btrfs_set_timespec_sec(leaf, btrfs_inode_ctime(item),
2582                                inode->i_ctime.tv_sec);
2583         btrfs_set_timespec_nsec(leaf, btrfs_inode_ctime(item),
2584                                 inode->i_ctime.tv_nsec);
2585
2586         btrfs_set_inode_nbytes(leaf, item, inode_get_bytes(inode));
2587         btrfs_set_inode_generation(leaf, item, BTRFS_I(inode)->generation);
2588         btrfs_set_inode_sequence(leaf, item, BTRFS_I(inode)->sequence);
2589         btrfs_set_inode_transid(leaf, item, trans->transid);
2590         btrfs_set_inode_rdev(leaf, item, inode->i_rdev);
2591         btrfs_set_inode_flags(leaf, item, BTRFS_I(inode)->flags);
2592         btrfs_set_inode_block_group(leaf, item, BTRFS_I(inode)->block_group);
2593 }
2594
2595 /*
2596  * copy everything in the in-memory inode into the btree.
2597  */
2598 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
2599                                 struct btrfs_root *root, struct inode *inode)
2600 {
2601         struct btrfs_inode_item *inode_item;
2602         struct btrfs_path *path;
2603         struct extent_buffer *leaf;
2604         int ret;
2605
2606         path = btrfs_alloc_path();
2607         BUG_ON(!path);
2608         path->leave_spinning = 1;
2609         ret = btrfs_lookup_inode(trans, root, path,
2610                                  &BTRFS_I(inode)->location, 1);
2611         if (ret) {
2612                 if (ret > 0)
2613                         ret = -ENOENT;
2614                 goto failed;
2615         }
2616
2617         btrfs_unlock_up_safe(path, 1);
2618         leaf = path->nodes[0];
2619         inode_item = btrfs_item_ptr(leaf, path->slots[0],
2620                                   struct btrfs_inode_item);
2621
2622         fill_inode_item(trans, leaf, inode_item, inode);
2623         btrfs_mark_buffer_dirty(leaf);
2624         btrfs_set_inode_last_trans(trans, inode);
2625         ret = 0;
2626 failed:
2627         btrfs_free_path(path);
2628         return ret;
2629 }
2630
2631
2632 /*
2633  * unlink helper that gets used here in inode.c and in the tree logging
2634  * recovery code.  It remove a link in a directory with a given name, and
2635  * also drops the back refs in the inode to the directory
2636  */
2637 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
2638                        struct btrfs_root *root,
2639                        struct inode *dir, struct inode *inode,
2640                        const char *name, int name_len)
2641 {
2642         struct btrfs_path *path;
2643         int ret = 0;
2644         struct extent_buffer *leaf;
2645         struct btrfs_dir_item *di;
2646         struct btrfs_key key;
2647         u64 index;
2648
2649         path = btrfs_alloc_path();
2650         if (!path) {
2651                 ret = -ENOMEM;
2652                 goto out;
2653         }
2654
2655         path->leave_spinning = 1;
2656         di = btrfs_lookup_dir_item(trans, root, path, dir->i_ino,
2657                                     name, name_len, -1);
2658         if (IS_ERR(di)) {
2659                 ret = PTR_ERR(di);
2660                 goto err;
2661         }
2662         if (!di) {
2663                 ret = -ENOENT;
2664                 goto err;
2665         }
2666         leaf = path->nodes[0];
2667         btrfs_dir_item_key_to_cpu(leaf, di, &key);
2668         ret = btrfs_delete_one_dir_name(trans, root, path, di);
2669         if (ret)
2670                 goto err;
2671         btrfs_release_path(root, path);
2672
2673         ret = btrfs_del_inode_ref(trans, root, name, name_len,
2674                                   inode->i_ino,
2675                                   dir->i_ino, &index);
2676         if (ret) {
2677                 printk(KERN_INFO "btrfs failed to delete reference to %.*s, "
2678                        "inode %lu parent %lu\n", name_len, name,
2679                        inode->i_ino, dir->i_ino);
2680                 goto err;
2681         }
2682
2683         di = btrfs_lookup_dir_index_item(trans, root, path, dir->i_ino,
2684                                          index, name, name_len, -1);
2685         if (IS_ERR(di)) {
2686                 ret = PTR_ERR(di);
2687                 goto err;
2688         }
2689         if (!di) {
2690                 ret = -ENOENT;
2691                 goto err;
2692         }
2693         ret = btrfs_delete_one_dir_name(trans, root, path, di);
2694         btrfs_release_path(root, path);
2695
2696         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
2697                                          inode, dir->i_ino);
2698         BUG_ON(ret != 0 && ret != -ENOENT);
2699
2700         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
2701                                            dir, index);
2702         if (ret == -ENOENT)
2703                 ret = 0;
2704 err:
2705         btrfs_free_path(path);
2706         if (ret)
2707                 goto out;
2708
2709         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
2710         inode->i_ctime = dir->i_mtime = dir->i_ctime = CURRENT_TIME;
2711         btrfs_update_inode(trans, root, dir);
2712         btrfs_drop_nlink(inode);
2713         ret = btrfs_update_inode(trans, root, inode);
2714 out:
2715         return ret;
2716 }
2717
2718 /* helper to check if there is any shared block in the path */
2719 static int check_path_shared(struct btrfs_root *root,
2720                              struct btrfs_path *path)
2721 {
2722         struct extent_buffer *eb;
2723         int level;
2724         u64 refs = 1;
2725
2726         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2727                 int ret;
2728
2729                 if (!path->nodes[level])
2730                         break;
2731                 eb = path->nodes[level];
2732                 if (!btrfs_block_can_be_shared(root, eb))
2733                         continue;
2734                 ret = btrfs_lookup_extent_info(NULL, root, eb->start, eb->len,
2735                                                &refs, NULL);
2736                 if (refs > 1)
2737                         return 1;
2738         }
2739         return 0;
2740 }
2741
2742 /*
2743  * helper to start transaction for unlink and rmdir.
2744  *
2745  * unlink and rmdir are special in btrfs, they do not always free space.
2746  * so in enospc case, we should make sure they will free space before
2747  * allowing them to use the global metadata reservation.
2748  */
2749 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir,
2750                                                        struct dentry *dentry)
2751 {
2752         struct btrfs_trans_handle *trans;
2753         struct btrfs_root *root = BTRFS_I(dir)->root;
2754         struct btrfs_path *path;
2755         struct btrfs_inode_ref *ref;
2756         struct btrfs_dir_item *di;
2757         struct inode *inode = dentry->d_inode;
2758         u64 index;
2759         int check_link = 1;
2760         int err = -ENOSPC;
2761         int ret;
2762
2763         trans = btrfs_start_transaction(root, 10);
2764         if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
2765                 return trans;
2766
2767         if (inode->i_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
2768                 return ERR_PTR(-ENOSPC);
2769
2770         /* check if there is someone else holds reference */
2771         if (S_ISDIR(inode->i_mode) && atomic_read(&inode->i_count) > 1)
2772                 return ERR_PTR(-ENOSPC);
2773
2774         if (atomic_read(&inode->i_count) > 2)
2775                 return ERR_PTR(-ENOSPC);
2776
2777         if (xchg(&root->fs_info->enospc_unlink, 1))
2778                 return ERR_PTR(-ENOSPC);
2779
2780         path = btrfs_alloc_path();
2781         if (!path) {
2782                 root->fs_info->enospc_unlink = 0;
2783                 return ERR_PTR(-ENOMEM);
2784         }
2785
2786         trans = btrfs_start_transaction(root, 0);
2787         if (IS_ERR(trans)) {
2788                 btrfs_free_path(path);
2789                 root->fs_info->enospc_unlink = 0;
2790                 return trans;
2791         }
2792
2793         path->skip_locking = 1;
2794         path->search_commit_root = 1;
2795
2796         ret = btrfs_lookup_inode(trans, root, path,
2797                                 &BTRFS_I(dir)->location, 0);
2798         if (ret < 0) {
2799                 err = ret;
2800                 goto out;
2801         }
2802         if (ret == 0) {
2803                 if (check_path_shared(root, path))
2804                         goto out;
2805         } else {
2806                 check_link = 0;
2807         }
2808         btrfs_release_path(root, path);
2809
2810         ret = btrfs_lookup_inode(trans, root, path,
2811                                 &BTRFS_I(inode)->location, 0);
2812         if (ret < 0) {
2813                 err = ret;
2814                 goto out;
2815         }
2816         if (ret == 0) {
2817                 if (check_path_shared(root, path))
2818                         goto out;
2819         } else {
2820                 check_link = 0;
2821         }
2822         btrfs_release_path(root, path);
2823
2824         if (ret == 0 && S_ISREG(inode->i_mode)) {
2825                 ret = btrfs_lookup_file_extent(trans, root, path,
2826                                                inode->i_ino, (u64)-1, 0);
2827                 if (ret < 0) {
2828                         err = ret;
2829                         goto out;
2830                 }
2831                 BUG_ON(ret == 0);
2832                 if (check_path_shared(root, path))
2833                         goto out;
2834                 btrfs_release_path(root, path);
2835         }
2836
2837         if (!check_link) {
2838                 err = 0;
2839                 goto out;
2840         }
2841
2842         di = btrfs_lookup_dir_item(trans, root, path, dir->i_ino,
2843                                 dentry->d_name.name, dentry->d_name.len, 0);
2844         if (IS_ERR(di)) {
2845                 err = PTR_ERR(di);
2846                 goto out;
2847         }
2848         if (di) {
2849                 if (check_path_shared(root, path))
2850                         goto out;
2851         } else {
2852                 err = 0;
2853                 goto out;
2854         }
2855         btrfs_release_path(root, path);
2856
2857         ref = btrfs_lookup_inode_ref(trans, root, path,
2858                                 dentry->d_name.name, dentry->d_name.len,
2859                                 inode->i_ino, dir->i_ino, 0);
2860         if (IS_ERR(ref)) {
2861                 err = PTR_ERR(ref);
2862                 goto out;
2863         }
2864         BUG_ON(!ref);
2865         if (check_path_shared(root, path))
2866                 goto out;
2867         index = btrfs_inode_ref_index(path->nodes[0], ref);
2868         btrfs_release_path(root, path);
2869
2870         di = btrfs_lookup_dir_index_item(trans, root, path, dir->i_ino, index,
2871                                 dentry->d_name.name, dentry->d_name.len, 0);
2872         if (IS_ERR(di)) {
2873                 err = PTR_ERR(di);
2874                 goto out;
2875         }
2876         BUG_ON(ret == -ENOENT);
2877         if (check_path_shared(root, path))
2878                 goto out;
2879
2880         err = 0;
2881 out:
2882         btrfs_free_path(path);
2883         if (err) {
2884                 btrfs_end_transaction(trans, root);
2885                 root->fs_info->enospc_unlink = 0;
2886                 return ERR_PTR(err);
2887         }
2888
2889         trans->block_rsv = &root->fs_info->global_block_rsv;
2890         return trans;
2891 }
2892
2893 static void __unlink_end_trans(struct btrfs_trans_handle *trans,
2894                                struct btrfs_root *root)
2895 {
2896         if (trans->block_rsv == &root->fs_info->global_block_rsv) {
2897                 BUG_ON(!root->fs_info->enospc_unlink);
2898                 root->fs_info->enospc_unlink = 0;
2899         }
2900         btrfs_end_transaction_throttle(trans, root);
2901 }
2902
2903 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
2904 {
2905         struct btrfs_root *root = BTRFS_I(dir)->root;
2906         struct btrfs_trans_handle *trans;
2907         struct inode *inode = dentry->d_inode;
2908         int ret;
2909         unsigned long nr = 0;
2910
2911         trans = __unlink_start_trans(dir, dentry);
2912         if (IS_ERR(trans))
2913                 return PTR_ERR(trans);
2914
2915         btrfs_set_trans_block_group(trans, dir);
2916
2917         btrfs_record_unlink_dir(trans, dir, dentry->d_inode, 0);
2918
2919         ret = btrfs_unlink_inode(trans, root, dir, dentry->d_inode,
2920                                  dentry->d_name.name, dentry->d_name.len);
2921         BUG_ON(ret);
2922
2923         if (inode->i_nlink == 0) {
2924                 ret = btrfs_orphan_add(trans, inode);
2925                 BUG_ON(ret);
2926         }
2927
2928         nr = trans->blocks_used;
2929         __unlink_end_trans(trans, root);
2930         btrfs_btree_balance_dirty(root, nr);
2931         return ret;
2932 }
2933
2934 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
2935                         struct btrfs_root *root,
2936                         struct inode *dir, u64 objectid,
2937                         const char *name, int name_len)
2938 {
2939         struct btrfs_path *path;
2940         struct extent_buffer *leaf;
2941         struct btrfs_dir_item *di;
2942         struct btrfs_key key;
2943         u64 index;
2944         int ret;
2945
2946         path = btrfs_alloc_path();
2947         if (!path)
2948                 return -ENOMEM;
2949
2950         di = btrfs_lookup_dir_item(trans, root, path, dir->i_ino,
2951                                    name, name_len, -1);
2952         BUG_ON(!di || IS_ERR(di));
2953
2954         leaf = path->nodes[0];
2955         btrfs_dir_item_key_to_cpu(leaf, di, &key);
2956         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
2957         ret = btrfs_delete_one_dir_name(trans, root, path, di);
2958         BUG_ON(ret);
2959         btrfs_release_path(root, path);
2960
2961         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
2962                                  objectid, root->root_key.objectid,
2963                                  dir->i_ino, &index, name, name_len);
2964         if (ret < 0) {
2965                 BUG_ON(ret != -ENOENT);
2966                 di = btrfs_search_dir_index_item(root, path, dir->i_ino,
2967                                                  name, name_len);
2968                 BUG_ON(!di || IS_ERR(di));
2969
2970                 leaf = path->nodes[0];
2971                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2972                 btrfs_release_path(root, path);
2973                 index = key.offset;
2974         }
2975
2976         di = btrfs_lookup_dir_index_item(trans, root, path, dir->i_ino,
2977                                          index, name, name_len, -1);
2978         BUG_ON(!di || IS_ERR(di));
2979
2980         leaf = path->nodes[0];
2981         btrfs_dir_item_key_to_cpu(leaf, di, &key);
2982         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
2983         ret = btrfs_delete_one_dir_name(trans, root, path, di);
2984         BUG_ON(ret);
2985         btrfs_release_path(root, path);
2986
2987         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
2988         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
2989         ret = btrfs_update_inode(trans, root, dir);
2990         BUG_ON(ret);
2991
2992         btrfs_free_path(path);
2993         return 0;
2994 }
2995
2996 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
2997 {
2998         struct inode *inode = dentry->d_inode;
2999         int err = 0;
3000         struct btrfs_root *root = BTRFS_I(dir)->root;
3001         struct btrfs_trans_handle *trans;
3002         unsigned long nr = 0;
3003
3004         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE ||
3005             inode->i_ino == BTRFS_FIRST_FREE_OBJECTID)
3006                 return -ENOTEMPTY;
3007
3008         trans = __unlink_start_trans(dir, dentry);
3009         if (IS_ERR(trans))
3010                 return PTR_ERR(trans);
3011
3012         btrfs_set_trans_block_group(trans, dir);
3013
3014         if (unlikely(inode->i_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
3015                 err = btrfs_unlink_subvol(trans, root, dir,
3016                                           BTRFS_I(inode)->location.objectid,
3017                                           dentry->d_name.name,
3018                                           dentry->d_name.len);
3019                 goto out;
3020         }
3021
3022         err = btrfs_orphan_add(trans, inode);
3023         if (err)
3024                 goto out;
3025
3026         /* now the directory is empty */
3027         err = btrfs_unlink_inode(trans, root, dir, dentry->d_inode,
3028                                  dentry->d_name.name, dentry->d_name.len);
3029         if (!err)
3030                 btrfs_i_size_write(inode, 0);
3031 out:
3032         nr = trans->blocks_used;
3033         __unlink_end_trans(trans, root);
3034         btrfs_btree_balance_dirty(root, nr);
3035
3036         return err;
3037 }
3038
3039 #if 0
3040 /*
3041  * when truncating bytes in a file, it is possible to avoid reading
3042  * the leaves that contain only checksum items.  This can be the
3043  * majority of the IO required to delete a large file, but it must
3044  * be done carefully.
3045  *
3046  * The keys in the level just above the leaves are checked to make sure
3047  * the lowest key in a given leaf is a csum key, and starts at an offset
3048  * after the new  size.
3049  *
3050  * Then the key for the next leaf is checked to make sure it also has
3051  * a checksum item for the same file.  If it does, we know our target leaf
3052  * contains only checksum items, and it can be safely freed without reading
3053  * it.
3054  *
3055  * This is just an optimization targeted at large files.  It may do
3056  * nothing.  It will return 0 unless things went badly.
3057  */
3058 static noinline int drop_csum_leaves(struct btrfs_trans_handle *trans,
3059                                      struct btrfs_root *root,
3060                                      struct btrfs_path *path,
3061                                      struct inode *inode, u64 new_size)
3062 {
3063         struct btrfs_key key;
3064         int ret;
3065         int nritems;
3066         struct btrfs_key found_key;
3067         struct btrfs_key other_key;
3068         struct btrfs_leaf_ref *ref;
3069         u64 leaf_gen;
3070         u64 leaf_start;
3071
3072         path->lowest_level = 1;
3073         key.objectid = inode->i_ino;
3074         key.type = BTRFS_CSUM_ITEM_KEY;
3075         key.offset = new_size;
3076 again:
3077         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
3078         if (ret < 0)
3079                 goto out;
3080
3081         if (path->nodes[1] == NULL) {
3082                 ret = 0;
3083                 goto out;
3084         }
3085         ret = 0;
3086         btrfs_node_key_to_cpu(path->nodes[1], &found_key, path->slots[1]);
3087         nritems = btrfs_header_nritems(path->nodes[1]);
3088
3089         if (!nritems)
3090                 goto out;
3091
3092         if (path->slots[1] >= nritems)
3093                 goto next_node;
3094
3095         /* did we find a key greater than anything we want to delete? */
3096         if (found_key.objectid > inode->i_ino ||
3097            (found_key.objectid == inode->i_ino && found_key.type > key.type))
3098                 goto out;
3099
3100         /* we check the next key in the node to make sure the leave contains
3101          * only checksum items.  This comparison doesn't work if our
3102          * leaf is the last one in the node
3103          */
3104         if (path->slots[1] + 1 >= nritems) {
3105 next_node:
3106                 /* search forward from the last key in the node, this
3107                  * will bring us into the next node in the tree
3108                  */
3109                 btrfs_node_key_to_cpu(path->nodes[1], &found_key, nritems - 1);
3110
3111                 /* unlikely, but we inc below, so check to be safe */
3112                 if (found_key.offset == (u64)-1)
3113                         goto out;
3114
3115                 /* search_forward needs a path with locks held, do the
3116                  * search again for the original key.  It is possible
3117                  * this will race with a balance and return a path that
3118                  * we could modify, but this drop is just an optimization
3119                  * and is allowed to miss some leaves.
3120                  */
3121                 btrfs_release_path(root, path);
3122                 found_key.offset++;
3123
3124                 /* setup a max key for search_forward */
3125                 other_key.offset = (u64)-1;
3126                 other_key.type = key.type;
3127                 other_key.objectid = key.objectid;
3128
3129                 path->keep_locks = 1;
3130                 ret = btrfs_search_forward(root, &found_key, &other_key,
3131                                            path, 0, 0);
3132                 path->keep_locks = 0;
3133                 if (ret || found_key.objectid != key.objectid ||
3134                     found_key.type != key.type) {
3135                         ret = 0;
3136                         goto out;
3137                 }
3138
3139                 key.offset = found_key.offset;
3140                 btrfs_release_path(root, path);
3141                 cond_resched();
3142                 goto again;
3143         }
3144
3145         /* we know there's one more slot after us in the tree,
3146          * read that key so we can verify it is also a checksum item
3147          */
3148         btrfs_node_key_to_cpu(path->nodes[1], &other_key, path->slots[1] + 1);
3149
3150         if (found_key.objectid < inode->i_ino)
3151                 goto next_key;
3152
3153         if (found_key.type != key.type || found_key.offset < new_size)
3154                 goto next_key;
3155
3156         /*
3157          * if the key for the next leaf isn't a csum key from this objectid,
3158          * we can't be sure there aren't good items inside this leaf.
3159          * Bail out
3160          */
3161         if (other_key.objectid != inode->i_ino || other_key.type != key.type)
3162                 goto out;
3163
3164         leaf_start = btrfs_node_blockptr(path->nodes[1], path->slots[1]);
3165         leaf_gen = btrfs_node_ptr_generation(path->nodes[1], path->slots[1]);
3166         /*
3167          * it is safe to delete this leaf, it contains only
3168          * csum items from this inode at an offset >= new_size
3169          */
3170         ret = btrfs_del_leaf(trans, root, path, leaf_start);
3171         BUG_ON(ret);
3172
3173         if (root->ref_cows && leaf_gen < trans->transid) {
3174                 ref = btrfs_alloc_leaf_ref(root, 0);
3175                 if (ref) {
3176                         ref->root_gen = root->root_key.offset;
3177                         ref->bytenr = leaf_start;
3178                         ref->owner = 0;
3179                         ref->generation = leaf_gen;
3180                         ref->nritems = 0;
3181
3182                         btrfs_sort_leaf_ref(ref);
3183
3184                         ret = btrfs_add_leaf_ref(root, ref, 0);
3185                         WARN_ON(ret);
3186                         btrfs_free_leaf_ref(root, ref);
3187                 } else {
3188                         WARN_ON(1);
3189                 }
3190         }
3191 next_key:
3192         btrfs_release_path(root, path);
3193
3194         if (other_key.objectid == inode->i_ino &&
3195             other_key.type == key.type && other_key.offset > key.offset) {
3196                 key.offset = other_key.offset;
3197                 cond_resched();
3198                 goto again;
3199         }
3200         ret = 0;
3201 out:
3202         /* fixup any changes we've made to the path */
3203         path->lowest_level = 0;
3204         path->keep_locks = 0;
3205         btrfs_release_path(root, path);
3206         return ret;
3207 }
3208
3209 #endif
3210
3211 /*
3212  * this can truncate away extent items, csum items and directory items.
3213  * It starts at a high offset and removes keys until it can't find
3214  * any higher than new_size
3215  *
3216  * csum items that cross the new i_size are truncated to the new size
3217  * as well.
3218  *
3219  * min_type is the minimum key type to truncate down to.  If set to 0, this
3220  * will kill all the items on this inode, including the INODE_ITEM_KEY.
3221  */
3222 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
3223                                struct btrfs_root *root,
3224                                struct inode *inode,
3225                                u64 new_size, u32 min_type)
3226 {
3227         struct btrfs_path *path;
3228         struct extent_buffer *leaf;
3229         struct btrfs_file_extent_item *fi;
3230         struct btrfs_key key;
3231         struct btrfs_key found_key;
3232         u64 extent_start = 0;
3233         u64 extent_num_bytes = 0;
3234         u64 extent_offset = 0;
3235         u64 item_end = 0;
3236         u64 mask = root->sectorsize - 1;
3237         u32 found_type = (u8)-1;
3238         int found_extent;
3239         int del_item;
3240         int pending_del_nr = 0;
3241         int pending_del_slot = 0;
3242         int extent_type = -1;
3243         int encoding;
3244         int ret;
3245         int err = 0;
3246
3247         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
3248
3249         if (root->ref_cows || root == root->fs_info->tree_root)
3250                 btrfs_drop_extent_cache(inode, new_size & (~mask), (u64)-1, 0);
3251
3252        &nb