Btrfs: stop using highmem for extent_buffers
[pandora-kernel.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 <linux/ratelimit.h>
41 #include "compat.h"
42 #include "ctree.h"
43 #include "disk-io.h"
44 #include "transaction.h"
45 #include "btrfs_inode.h"
46 #include "ioctl.h"
47 #include "print-tree.h"
48 #include "volumes.h"
49 #include "ordered-data.h"
50 #include "xattr.h"
51 #include "tree-log.h"
52 #include "compression.h"
53 #include "locking.h"
54 #include "free-space-cache.h"
55 #include "inode-map.h"
56
57 struct btrfs_iget_args {
58         u64 ino;
59         struct btrfs_root *root;
60 };
61
62 static const struct inode_operations btrfs_dir_inode_operations;
63 static const struct inode_operations btrfs_symlink_inode_operations;
64 static const struct inode_operations btrfs_dir_ro_inode_operations;
65 static const struct inode_operations btrfs_special_inode_operations;
66 static const struct inode_operations btrfs_file_inode_operations;
67 static const struct address_space_operations btrfs_aops;
68 static const struct address_space_operations btrfs_symlink_aops;
69 static const struct file_operations btrfs_dir_file_operations;
70 static struct extent_io_ops btrfs_extent_io_ops;
71
72 static struct kmem_cache *btrfs_inode_cachep;
73 struct kmem_cache *btrfs_trans_handle_cachep;
74 struct kmem_cache *btrfs_transaction_cachep;
75 struct kmem_cache *btrfs_path_cachep;
76 struct kmem_cache *btrfs_free_space_cachep;
77
78 #define S_SHIFT 12
79 static unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
80         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
81         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
82         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
83         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
84         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
85         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
86         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
87 };
88
89 static int btrfs_setsize(struct inode *inode, loff_t newsize);
90 static int btrfs_truncate(struct inode *inode);
91 static int btrfs_finish_ordered_io(struct inode *inode, u64 start, u64 end);
92 static noinline int cow_file_range(struct inode *inode,
93                                    struct page *locked_page,
94                                    u64 start, u64 end, int *page_started,
95                                    unsigned long *nr_written, int unlock);
96
97 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
98                                      struct inode *inode,  struct inode *dir,
99                                      const struct qstr *qstr)
100 {
101         int err;
102
103         err = btrfs_init_acl(trans, inode, dir);
104         if (!err)
105                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
106         return err;
107 }
108
109 /*
110  * this does all the hard work for inserting an inline extent into
111  * the btree.  The caller should have done a btrfs_drop_extents so that
112  * no overlapping inline items exist in the btree
113  */
114 static noinline int insert_inline_extent(struct btrfs_trans_handle *trans,
115                                 struct btrfs_root *root, struct inode *inode,
116                                 u64 start, size_t size, size_t compressed_size,
117                                 int compress_type,
118                                 struct page **compressed_pages)
119 {
120         struct btrfs_key key;
121         struct btrfs_path *path;
122         struct extent_buffer *leaf;
123         struct page *page = NULL;
124         char *kaddr;
125         unsigned long ptr;
126         struct btrfs_file_extent_item *ei;
127         int err = 0;
128         int ret;
129         size_t cur_size = size;
130         size_t datasize;
131         unsigned long offset;
132
133         if (compressed_size && compressed_pages)
134                 cur_size = compressed_size;
135
136         path = btrfs_alloc_path();
137         if (!path)
138                 return -ENOMEM;
139
140         path->leave_spinning = 1;
141
142         key.objectid = btrfs_ino(inode);
143         key.offset = start;
144         btrfs_set_key_type(&key, BTRFS_EXTENT_DATA_KEY);
145         datasize = btrfs_file_extent_calc_inline_size(cur_size);
146
147         inode_add_bytes(inode, size);
148         ret = btrfs_insert_empty_item(trans, root, path, &key,
149                                       datasize);
150         BUG_ON(ret);
151         if (ret) {
152                 err = ret;
153                 goto fail;
154         }
155         leaf = path->nodes[0];
156         ei = btrfs_item_ptr(leaf, path->slots[0],
157                             struct btrfs_file_extent_item);
158         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
159         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
160         btrfs_set_file_extent_encryption(leaf, ei, 0);
161         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
162         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
163         ptr = btrfs_file_extent_inline_start(ei);
164
165         if (compress_type != BTRFS_COMPRESS_NONE) {
166                 struct page *cpage;
167                 int i = 0;
168                 while (compressed_size > 0) {
169                         cpage = compressed_pages[i];
170                         cur_size = min_t(unsigned long, compressed_size,
171                                        PAGE_CACHE_SIZE);
172
173                         kaddr = kmap_atomic(cpage, KM_USER0);
174                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
175                         kunmap_atomic(kaddr, KM_USER0);
176
177                         i++;
178                         ptr += cur_size;
179                         compressed_size -= cur_size;
180                 }
181                 btrfs_set_file_extent_compression(leaf, ei,
182                                                   compress_type);
183         } else {
184                 page = find_get_page(inode->i_mapping,
185                                      start >> PAGE_CACHE_SHIFT);
186                 btrfs_set_file_extent_compression(leaf, ei, 0);
187                 kaddr = kmap_atomic(page, KM_USER0);
188                 offset = start & (PAGE_CACHE_SIZE - 1);
189                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
190                 kunmap_atomic(kaddr, KM_USER0);
191                 page_cache_release(page);
192         }
193         btrfs_mark_buffer_dirty(leaf);
194         btrfs_free_path(path);
195
196         /*
197          * we're an inline extent, so nobody can
198          * extend the file past i_size without locking
199          * a page we already have locked.
200          *
201          * We must do any isize and inode updates
202          * before we unlock the pages.  Otherwise we
203          * could end up racing with unlink.
204          */
205         BTRFS_I(inode)->disk_i_size = inode->i_size;
206         btrfs_update_inode(trans, root, inode);
207
208         return 0;
209 fail:
210         btrfs_free_path(path);
211         return err;
212 }
213
214
215 /*
216  * conditionally insert an inline extent into the file.  This
217  * does the checks required to make sure the data is small enough
218  * to fit as an inline extent.
219  */
220 static noinline int cow_file_range_inline(struct btrfs_trans_handle *trans,
221                                  struct btrfs_root *root,
222                                  struct inode *inode, u64 start, u64 end,
223                                  size_t compressed_size, int compress_type,
224                                  struct page **compressed_pages)
225 {
226         u64 isize = i_size_read(inode);
227         u64 actual_end = min(end + 1, isize);
228         u64 inline_len = actual_end - start;
229         u64 aligned_end = (end + root->sectorsize - 1) &
230                         ~((u64)root->sectorsize - 1);
231         u64 hint_byte;
232         u64 data_len = inline_len;
233         int ret;
234
235         if (compressed_size)
236                 data_len = compressed_size;
237
238         if (start > 0 ||
239             actual_end >= PAGE_CACHE_SIZE ||
240             data_len >= BTRFS_MAX_INLINE_DATA_SIZE(root) ||
241             (!compressed_size &&
242             (actual_end & (root->sectorsize - 1)) == 0) ||
243             end + 1 < isize ||
244             data_len > root->fs_info->max_inline) {
245                 return 1;
246         }
247
248         ret = btrfs_drop_extents(trans, inode, start, aligned_end,
249                                  &hint_byte, 1);
250         BUG_ON(ret);
251
252         if (isize > actual_end)
253                 inline_len = min_t(u64, isize, actual_end);
254         ret = insert_inline_extent(trans, root, inode, start,
255                                    inline_len, compressed_size,
256                                    compress_type, compressed_pages);
257         BUG_ON(ret);
258         btrfs_delalloc_release_metadata(inode, end + 1 - start);
259         btrfs_drop_extent_cache(inode, start, aligned_end - 1, 0);
260         return 0;
261 }
262
263 struct async_extent {
264         u64 start;
265         u64 ram_size;
266         u64 compressed_size;
267         struct page **pages;
268         unsigned long nr_pages;
269         int compress_type;
270         struct list_head list;
271 };
272
273 struct async_cow {
274         struct inode *inode;
275         struct btrfs_root *root;
276         struct page *locked_page;
277         u64 start;
278         u64 end;
279         struct list_head extents;
280         struct btrfs_work work;
281 };
282
283 static noinline int add_async_extent(struct async_cow *cow,
284                                      u64 start, u64 ram_size,
285                                      u64 compressed_size,
286                                      struct page **pages,
287                                      unsigned long nr_pages,
288                                      int compress_type)
289 {
290         struct async_extent *async_extent;
291
292         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
293         BUG_ON(!async_extent);
294         async_extent->start = start;
295         async_extent->ram_size = ram_size;
296         async_extent->compressed_size = compressed_size;
297         async_extent->pages = pages;
298         async_extent->nr_pages = nr_pages;
299         async_extent->compress_type = compress_type;
300         list_add_tail(&async_extent->list, &cow->extents);
301         return 0;
302 }
303
304 /*
305  * we create compressed extents in two phases.  The first
306  * phase compresses a range of pages that have already been
307  * locked (both pages and state bits are locked).
308  *
309  * This is done inside an ordered work queue, and the compression
310  * is spread across many cpus.  The actual IO submission is step
311  * two, and the ordered work queue takes care of making sure that
312  * happens in the same order things were put onto the queue by
313  * writepages and friends.
314  *
315  * If this code finds it can't get good compression, it puts an
316  * entry onto the work queue to write the uncompressed bytes.  This
317  * makes sure that both compressed inodes and uncompressed inodes
318  * are written in the same order that pdflush sent them down.
319  */
320 static noinline int compress_file_range(struct inode *inode,
321                                         struct page *locked_page,
322                                         u64 start, u64 end,
323                                         struct async_cow *async_cow,
324                                         int *num_added)
325 {
326         struct btrfs_root *root = BTRFS_I(inode)->root;
327         struct btrfs_trans_handle *trans;
328         u64 num_bytes;
329         u64 blocksize = root->sectorsize;
330         u64 actual_end;
331         u64 isize = i_size_read(inode);
332         int ret = 0;
333         struct page **pages = NULL;
334         unsigned long nr_pages;
335         unsigned long nr_pages_ret = 0;
336         unsigned long total_compressed = 0;
337         unsigned long total_in = 0;
338         unsigned long max_compressed = 128 * 1024;
339         unsigned long max_uncompressed = 128 * 1024;
340         int i;
341         int will_compress;
342         int compress_type = root->fs_info->compress_type;
343
344         /* if this is a small write inside eof, kick off a defragbot */
345         if (end <= BTRFS_I(inode)->disk_i_size && (end - start + 1) < 16 * 1024)
346                 btrfs_add_inode_defrag(NULL, inode);
347
348         actual_end = min_t(u64, isize, end + 1);
349 again:
350         will_compress = 0;
351         nr_pages = (end >> PAGE_CACHE_SHIFT) - (start >> PAGE_CACHE_SHIFT) + 1;
352         nr_pages = min(nr_pages, (128 * 1024UL) / PAGE_CACHE_SIZE);
353
354         /*
355          * we don't want to send crud past the end of i_size through
356          * compression, that's just a waste of CPU time.  So, if the
357          * end of the file is before the start of our current
358          * requested range of bytes, we bail out to the uncompressed
359          * cleanup code that can deal with all of this.
360          *
361          * It isn't really the fastest way to fix things, but this is a
362          * very uncommon corner.
363          */
364         if (actual_end <= start)
365                 goto cleanup_and_bail_uncompressed;
366
367         total_compressed = actual_end - start;
368
369         /* we want to make sure that amount of ram required to uncompress
370          * an extent is reasonable, so we limit the total size in ram
371          * of a compressed extent to 128k.  This is a crucial number
372          * because it also controls how easily we can spread reads across
373          * cpus for decompression.
374          *
375          * We also want to make sure the amount of IO required to do
376          * a random read is reasonably small, so we limit the size of
377          * a compressed extent to 128k.
378          */
379         total_compressed = min(total_compressed, max_uncompressed);
380         num_bytes = (end - start + blocksize) & ~(blocksize - 1);
381         num_bytes = max(blocksize,  num_bytes);
382         total_in = 0;
383         ret = 0;
384
385         /*
386          * we do compression for mount -o compress and when the
387          * inode has not been flagged as nocompress.  This flag can
388          * change at any time if we discover bad compression ratios.
389          */
390         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS) &&
391             (btrfs_test_opt(root, COMPRESS) ||
392              (BTRFS_I(inode)->force_compress) ||
393              (BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS))) {
394                 WARN_ON(pages);
395                 pages = kzalloc(sizeof(struct page *) * nr_pages, GFP_NOFS);
396                 BUG_ON(!pages);
397
398                 if (BTRFS_I(inode)->force_compress)
399                         compress_type = BTRFS_I(inode)->force_compress;
400
401                 ret = btrfs_compress_pages(compress_type,
402                                            inode->i_mapping, start,
403                                            total_compressed, pages,
404                                            nr_pages, &nr_pages_ret,
405                                            &total_in,
406                                            &total_compressed,
407                                            max_compressed);
408
409                 if (!ret) {
410                         unsigned long offset = total_compressed &
411                                 (PAGE_CACHE_SIZE - 1);
412                         struct page *page = pages[nr_pages_ret - 1];
413                         char *kaddr;
414
415                         /* zero the tail end of the last page, we might be
416                          * sending it down to disk
417                          */
418                         if (offset) {
419                                 kaddr = kmap_atomic(page, KM_USER0);
420                                 memset(kaddr + offset, 0,
421                                        PAGE_CACHE_SIZE - offset);
422                                 kunmap_atomic(kaddr, KM_USER0);
423                         }
424                         will_compress = 1;
425                 }
426         }
427         if (start == 0) {
428                 trans = btrfs_join_transaction(root);
429                 BUG_ON(IS_ERR(trans));
430                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
431
432                 /* lets try to make an inline extent */
433                 if (ret || total_in < (actual_end - start)) {
434                         /* we didn't compress the entire range, try
435                          * to make an uncompressed inline extent.
436                          */
437                         ret = cow_file_range_inline(trans, root, inode,
438                                                     start, end, 0, 0, NULL);
439                 } else {
440                         /* try making a compressed inline extent */
441                         ret = cow_file_range_inline(trans, root, inode,
442                                                     start, end,
443                                                     total_compressed,
444                                                     compress_type, pages);
445                 }
446                 if (ret == 0) {
447                         /*
448                          * inline extent creation worked, we don't need
449                          * to create any more async work items.  Unlock
450                          * and free up our temp pages.
451                          */
452                         extent_clear_unlock_delalloc(inode,
453                              &BTRFS_I(inode)->io_tree,
454                              start, end, NULL,
455                              EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
456                              EXTENT_CLEAR_DELALLOC |
457                              EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK);
458
459                         btrfs_end_transaction(trans, root);
460                         goto free_pages_out;
461                 }
462                 btrfs_end_transaction(trans, root);
463         }
464
465         if (will_compress) {
466                 /*
467                  * we aren't doing an inline extent round the compressed size
468                  * up to a block size boundary so the allocator does sane
469                  * things
470                  */
471                 total_compressed = (total_compressed + blocksize - 1) &
472                         ~(blocksize - 1);
473
474                 /*
475                  * one last check to make sure the compression is really a
476                  * win, compare the page count read with the blocks on disk
477                  */
478                 total_in = (total_in + PAGE_CACHE_SIZE - 1) &
479                         ~(PAGE_CACHE_SIZE - 1);
480                 if (total_compressed >= total_in) {
481                         will_compress = 0;
482                 } else {
483                         num_bytes = total_in;
484                 }
485         }
486         if (!will_compress && pages) {
487                 /*
488                  * the compression code ran but failed to make things smaller,
489                  * free any pages it allocated and our page pointer array
490                  */
491                 for (i = 0; i < nr_pages_ret; i++) {
492                         WARN_ON(pages[i]->mapping);
493                         page_cache_release(pages[i]);
494                 }
495                 kfree(pages);
496                 pages = NULL;
497                 total_compressed = 0;
498                 nr_pages_ret = 0;
499
500                 /* flag the file so we don't compress in the future */
501                 if (!btrfs_test_opt(root, FORCE_COMPRESS) &&
502                     !(BTRFS_I(inode)->force_compress)) {
503                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
504                 }
505         }
506         if (will_compress) {
507                 *num_added += 1;
508
509                 /* the async work queues will take care of doing actual
510                  * allocation on disk for these compressed pages,
511                  * and will submit them to the elevator.
512                  */
513                 add_async_extent(async_cow, start, num_bytes,
514                                  total_compressed, pages, nr_pages_ret,
515                                  compress_type);
516
517                 if (start + num_bytes < end) {
518                         start += num_bytes;
519                         pages = NULL;
520                         cond_resched();
521                         goto again;
522                 }
523         } else {
524 cleanup_and_bail_uncompressed:
525                 /*
526                  * No compression, but we still need to write the pages in
527                  * the file we've been given so far.  redirty the locked
528                  * page if it corresponds to our extent and set things up
529                  * for the async work queue to run cow_file_range to do
530                  * the normal delalloc dance
531                  */
532                 if (page_offset(locked_page) >= start &&
533                     page_offset(locked_page) <= end) {
534                         __set_page_dirty_nobuffers(locked_page);
535                         /* unlocked later on in the async handlers */
536                 }
537                 add_async_extent(async_cow, start, end - start + 1,
538                                  0, NULL, 0, BTRFS_COMPRESS_NONE);
539                 *num_added += 1;
540         }
541
542 out:
543         return 0;
544
545 free_pages_out:
546         for (i = 0; i < nr_pages_ret; i++) {
547                 WARN_ON(pages[i]->mapping);
548                 page_cache_release(pages[i]);
549         }
550         kfree(pages);
551
552         goto out;
553 }
554
555 /*
556  * phase two of compressed writeback.  This is the ordered portion
557  * of the code, which only gets called in the order the work was
558  * queued.  We walk all the async extents created by compress_file_range
559  * and send them down to the disk.
560  */
561 static noinline int submit_compressed_extents(struct inode *inode,
562                                               struct async_cow *async_cow)
563 {
564         struct async_extent *async_extent;
565         u64 alloc_hint = 0;
566         struct btrfs_trans_handle *trans;
567         struct btrfs_key ins;
568         struct extent_map *em;
569         struct btrfs_root *root = BTRFS_I(inode)->root;
570         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
571         struct extent_io_tree *io_tree;
572         int ret = 0;
573
574         if (list_empty(&async_cow->extents))
575                 return 0;
576
577
578         while (!list_empty(&async_cow->extents)) {
579                 async_extent = list_entry(async_cow->extents.next,
580                                           struct async_extent, list);
581                 list_del(&async_extent->list);
582
583                 io_tree = &BTRFS_I(inode)->io_tree;
584
585 retry:
586                 /* did the compression code fall back to uncompressed IO? */
587                 if (!async_extent->pages) {
588                         int page_started = 0;
589                         unsigned long nr_written = 0;
590
591                         lock_extent(io_tree, async_extent->start,
592                                          async_extent->start +
593                                          async_extent->ram_size - 1, GFP_NOFS);
594
595                         /* allocate blocks */
596                         ret = cow_file_range(inode, async_cow->locked_page,
597                                              async_extent->start,
598                                              async_extent->start +
599                                              async_extent->ram_size - 1,
600                                              &page_started, &nr_written, 0);
601
602                         /*
603                          * if page_started, cow_file_range inserted an
604                          * inline extent and took care of all the unlocking
605                          * and IO for us.  Otherwise, we need to submit
606                          * all those pages down to the drive.
607                          */
608                         if (!page_started && !ret)
609                                 extent_write_locked_range(io_tree,
610                                                   inode, async_extent->start,
611                                                   async_extent->start +
612                                                   async_extent->ram_size - 1,
613                                                   btrfs_get_extent,
614                                                   WB_SYNC_ALL);
615                         kfree(async_extent);
616                         cond_resched();
617                         continue;
618                 }
619
620                 lock_extent(io_tree, async_extent->start,
621                             async_extent->start + async_extent->ram_size - 1,
622                             GFP_NOFS);
623
624                 trans = btrfs_join_transaction(root);
625                 BUG_ON(IS_ERR(trans));
626                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
627                 ret = btrfs_reserve_extent(trans, root,
628                                            async_extent->compressed_size,
629                                            async_extent->compressed_size,
630                                            0, alloc_hint,
631                                            (u64)-1, &ins, 1);
632                 btrfs_end_transaction(trans, root);
633
634                 if (ret) {
635                         int i;
636                         for (i = 0; i < async_extent->nr_pages; i++) {
637                                 WARN_ON(async_extent->pages[i]->mapping);
638                                 page_cache_release(async_extent->pages[i]);
639                         }
640                         kfree(async_extent->pages);
641                         async_extent->nr_pages = 0;
642                         async_extent->pages = NULL;
643                         unlock_extent(io_tree, async_extent->start,
644                                       async_extent->start +
645                                       async_extent->ram_size - 1, GFP_NOFS);
646                         goto retry;
647                 }
648
649                 /*
650                  * here we're doing allocation and writeback of the
651                  * compressed pages
652                  */
653                 btrfs_drop_extent_cache(inode, async_extent->start,
654                                         async_extent->start +
655                                         async_extent->ram_size - 1, 0);
656
657                 em = alloc_extent_map();
658                 BUG_ON(!em);
659                 em->start = async_extent->start;
660                 em->len = async_extent->ram_size;
661                 em->orig_start = em->start;
662
663                 em->block_start = ins.objectid;
664                 em->block_len = ins.offset;
665                 em->bdev = root->fs_info->fs_devices->latest_bdev;
666                 em->compress_type = async_extent->compress_type;
667                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
668                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
669
670                 while (1) {
671                         write_lock(&em_tree->lock);
672                         ret = add_extent_mapping(em_tree, em);
673                         write_unlock(&em_tree->lock);
674                         if (ret != -EEXIST) {
675                                 free_extent_map(em);
676                                 break;
677                         }
678                         btrfs_drop_extent_cache(inode, async_extent->start,
679                                                 async_extent->start +
680                                                 async_extent->ram_size - 1, 0);
681                 }
682
683                 ret = btrfs_add_ordered_extent_compress(inode,
684                                                 async_extent->start,
685                                                 ins.objectid,
686                                                 async_extent->ram_size,
687                                                 ins.offset,
688                                                 BTRFS_ORDERED_COMPRESSED,
689                                                 async_extent->compress_type);
690                 BUG_ON(ret);
691
692                 /*
693                  * clear dirty, set writeback and unlock the pages.
694                  */
695                 extent_clear_unlock_delalloc(inode,
696                                 &BTRFS_I(inode)->io_tree,
697                                 async_extent->start,
698                                 async_extent->start +
699                                 async_extent->ram_size - 1,
700                                 NULL, EXTENT_CLEAR_UNLOCK_PAGE |
701                                 EXTENT_CLEAR_UNLOCK |
702                                 EXTENT_CLEAR_DELALLOC |
703                                 EXTENT_CLEAR_DIRTY | EXTENT_SET_WRITEBACK);
704
705                 ret = btrfs_submit_compressed_write(inode,
706                                     async_extent->start,
707                                     async_extent->ram_size,
708                                     ins.objectid,
709                                     ins.offset, async_extent->pages,
710                                     async_extent->nr_pages);
711
712                 BUG_ON(ret);
713                 alloc_hint = ins.objectid + ins.offset;
714                 kfree(async_extent);
715                 cond_resched();
716         }
717
718         return 0;
719 }
720
721 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
722                                       u64 num_bytes)
723 {
724         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
725         struct extent_map *em;
726         u64 alloc_hint = 0;
727
728         read_lock(&em_tree->lock);
729         em = search_extent_mapping(em_tree, start, num_bytes);
730         if (em) {
731                 /*
732                  * if block start isn't an actual block number then find the
733                  * first block in this inode and use that as a hint.  If that
734                  * block is also bogus then just don't worry about it.
735                  */
736                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
737                         free_extent_map(em);
738                         em = search_extent_mapping(em_tree, 0, 0);
739                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
740                                 alloc_hint = em->block_start;
741                         if (em)
742                                 free_extent_map(em);
743                 } else {
744                         alloc_hint = em->block_start;
745                         free_extent_map(em);
746                 }
747         }
748         read_unlock(&em_tree->lock);
749
750         return alloc_hint;
751 }
752
753 static inline bool is_free_space_inode(struct btrfs_root *root,
754                                        struct inode *inode)
755 {
756         if (root == root->fs_info->tree_root ||
757             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID)
758                 return true;
759         return false;
760 }
761
762 /*
763  * when extent_io.c finds a delayed allocation range in the file,
764  * the call backs end up in this code.  The basic idea is to
765  * allocate extents on disk for the range, and create ordered data structs
766  * in ram to track those extents.
767  *
768  * locked_page is the page that writepage had locked already.  We use
769  * it to make sure we don't do extra locks or unlocks.
770  *
771  * *page_started is set to one if we unlock locked_page and do everything
772  * required to start IO on it.  It may be clean and already done with
773  * IO when we return.
774  */
775 static noinline int cow_file_range(struct inode *inode,
776                                    struct page *locked_page,
777                                    u64 start, u64 end, int *page_started,
778                                    unsigned long *nr_written,
779                                    int unlock)
780 {
781         struct btrfs_root *root = BTRFS_I(inode)->root;
782         struct btrfs_trans_handle *trans;
783         u64 alloc_hint = 0;
784         u64 num_bytes;
785         unsigned long ram_size;
786         u64 disk_num_bytes;
787         u64 cur_alloc_size;
788         u64 blocksize = root->sectorsize;
789         struct btrfs_key ins;
790         struct extent_map *em;
791         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
792         int ret = 0;
793
794         BUG_ON(is_free_space_inode(root, inode));
795         trans = btrfs_join_transaction(root);
796         BUG_ON(IS_ERR(trans));
797         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
798
799         num_bytes = (end - start + blocksize) & ~(blocksize - 1);
800         num_bytes = max(blocksize,  num_bytes);
801         disk_num_bytes = num_bytes;
802         ret = 0;
803
804         /* if this is a small write inside eof, kick off defrag */
805         if (end <= BTRFS_I(inode)->disk_i_size && num_bytes < 64 * 1024)
806                 btrfs_add_inode_defrag(trans, inode);
807
808         if (start == 0) {
809                 /* lets try to make an inline extent */
810                 ret = cow_file_range_inline(trans, root, inode,
811                                             start, end, 0, 0, NULL);
812                 if (ret == 0) {
813                         extent_clear_unlock_delalloc(inode,
814                                      &BTRFS_I(inode)->io_tree,
815                                      start, end, NULL,
816                                      EXTENT_CLEAR_UNLOCK_PAGE |
817                                      EXTENT_CLEAR_UNLOCK |
818                                      EXTENT_CLEAR_DELALLOC |
819                                      EXTENT_CLEAR_DIRTY |
820                                      EXTENT_SET_WRITEBACK |
821                                      EXTENT_END_WRITEBACK);
822
823                         *nr_written = *nr_written +
824                              (end - start + PAGE_CACHE_SIZE) / PAGE_CACHE_SIZE;
825                         *page_started = 1;
826                         ret = 0;
827                         goto out;
828                 }
829         }
830
831         BUG_ON(disk_num_bytes >
832                btrfs_super_total_bytes(&root->fs_info->super_copy));
833
834         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
835         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
836
837         while (disk_num_bytes > 0) {
838                 unsigned long op;
839
840                 cur_alloc_size = disk_num_bytes;
841                 ret = btrfs_reserve_extent(trans, root, cur_alloc_size,
842                                            root->sectorsize, 0, alloc_hint,
843                                            (u64)-1, &ins, 1);
844                 BUG_ON(ret);
845
846                 em = alloc_extent_map();
847                 BUG_ON(!em);
848                 em->start = start;
849                 em->orig_start = em->start;
850                 ram_size = ins.offset;
851                 em->len = ins.offset;
852
853                 em->block_start = ins.objectid;
854                 em->block_len = ins.offset;
855                 em->bdev = root->fs_info->fs_devices->latest_bdev;
856                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
857
858                 while (1) {
859                         write_lock(&em_tree->lock);
860                         ret = add_extent_mapping(em_tree, em);
861                         write_unlock(&em_tree->lock);
862                         if (ret != -EEXIST) {
863                                 free_extent_map(em);
864                                 break;
865                         }
866                         btrfs_drop_extent_cache(inode, start,
867                                                 start + ram_size - 1, 0);
868                 }
869
870                 cur_alloc_size = ins.offset;
871                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
872                                                ram_size, cur_alloc_size, 0);
873                 BUG_ON(ret);
874
875                 if (root->root_key.objectid ==
876                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
877                         ret = btrfs_reloc_clone_csums(inode, start,
878                                                       cur_alloc_size);
879                         BUG_ON(ret);
880                 }
881
882                 if (disk_num_bytes < cur_alloc_size)
883                         break;
884
885                 /* we're not doing compressed IO, don't unlock the first
886                  * page (which the caller expects to stay locked), don't
887                  * clear any dirty bits and don't set any writeback bits
888                  *
889                  * Do set the Private2 bit so we know this page was properly
890                  * setup for writepage
891                  */
892                 op = unlock ? EXTENT_CLEAR_UNLOCK_PAGE : 0;
893                 op |= EXTENT_CLEAR_UNLOCK | EXTENT_CLEAR_DELALLOC |
894                         EXTENT_SET_PRIVATE2;
895
896                 extent_clear_unlock_delalloc(inode, &BTRFS_I(inode)->io_tree,
897                                              start, start + ram_size - 1,
898                                              locked_page, op);
899                 disk_num_bytes -= cur_alloc_size;
900                 num_bytes -= cur_alloc_size;
901                 alloc_hint = ins.objectid + ins.offset;
902                 start += cur_alloc_size;
903         }
904 out:
905         ret = 0;
906         btrfs_end_transaction(trans, root);
907
908         return ret;
909 }
910
911 /*
912  * work queue call back to started compression on a file and pages
913  */
914 static noinline void async_cow_start(struct btrfs_work *work)
915 {
916         struct async_cow *async_cow;
917         int num_added = 0;
918         async_cow = container_of(work, struct async_cow, work);
919
920         compress_file_range(async_cow->inode, async_cow->locked_page,
921                             async_cow->start, async_cow->end, async_cow,
922                             &num_added);
923         if (num_added == 0)
924                 async_cow->inode = NULL;
925 }
926
927 /*
928  * work queue call back to submit previously compressed pages
929  */
930 static noinline void async_cow_submit(struct btrfs_work *work)
931 {
932         struct async_cow *async_cow;
933         struct btrfs_root *root;
934         unsigned long nr_pages;
935
936         async_cow = container_of(work, struct async_cow, work);
937
938         root = async_cow->root;
939         nr_pages = (async_cow->end - async_cow->start + PAGE_CACHE_SIZE) >>
940                 PAGE_CACHE_SHIFT;
941
942         atomic_sub(nr_pages, &root->fs_info->async_delalloc_pages);
943
944         if (atomic_read(&root->fs_info->async_delalloc_pages) <
945             5 * 1042 * 1024 &&
946             waitqueue_active(&root->fs_info->async_submit_wait))
947                 wake_up(&root->fs_info->async_submit_wait);
948
949         if (async_cow->inode)
950                 submit_compressed_extents(async_cow->inode, async_cow);
951 }
952
953 static noinline void async_cow_free(struct btrfs_work *work)
954 {
955         struct async_cow *async_cow;
956         async_cow = container_of(work, struct async_cow, work);
957         kfree(async_cow);
958 }
959
960 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
961                                 u64 start, u64 end, int *page_started,
962                                 unsigned long *nr_written)
963 {
964         struct async_cow *async_cow;
965         struct btrfs_root *root = BTRFS_I(inode)->root;
966         unsigned long nr_pages;
967         u64 cur_end;
968         int limit = 10 * 1024 * 1042;
969
970         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
971                          1, 0, NULL, GFP_NOFS);
972         while (start < end) {
973                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
974                 BUG_ON(!async_cow);
975                 async_cow->inode = inode;
976                 async_cow->root = root;
977                 async_cow->locked_page = locked_page;
978                 async_cow->start = start;
979
980                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
981                         cur_end = end;
982                 else
983                         cur_end = min(end, start + 512 * 1024 - 1);
984
985                 async_cow->end = cur_end;
986                 INIT_LIST_HEAD(&async_cow->extents);
987
988                 async_cow->work.func = async_cow_start;
989                 async_cow->work.ordered_func = async_cow_submit;
990                 async_cow->work.ordered_free = async_cow_free;
991                 async_cow->work.flags = 0;
992
993                 nr_pages = (cur_end - start + PAGE_CACHE_SIZE) >>
994                         PAGE_CACHE_SHIFT;
995                 atomic_add(nr_pages, &root->fs_info->async_delalloc_pages);
996
997                 btrfs_queue_worker(&root->fs_info->delalloc_workers,
998                                    &async_cow->work);
999
1000                 if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
1001                         wait_event(root->fs_info->async_submit_wait,
1002                            (atomic_read(&root->fs_info->async_delalloc_pages) <
1003                             limit));
1004                 }
1005
1006                 while (atomic_read(&root->fs_info->async_submit_draining) &&
1007                       atomic_read(&root->fs_info->async_delalloc_pages)) {
1008                         wait_event(root->fs_info->async_submit_wait,
1009                           (atomic_read(&root->fs_info->async_delalloc_pages) ==
1010                            0));
1011                 }
1012
1013                 *nr_written += nr_pages;
1014                 start = cur_end + 1;
1015         }
1016         *page_started = 1;
1017         return 0;
1018 }
1019
1020 static noinline int csum_exist_in_range(struct btrfs_root *root,
1021                                         u64 bytenr, u64 num_bytes)
1022 {
1023         int ret;
1024         struct btrfs_ordered_sum *sums;
1025         LIST_HEAD(list);
1026
1027         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, bytenr,
1028                                        bytenr + num_bytes - 1, &list, 0);
1029         if (ret == 0 && list_empty(&list))
1030                 return 0;
1031
1032         while (!list_empty(&list)) {
1033                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1034                 list_del(&sums->list);
1035                 kfree(sums);
1036         }
1037         return 1;
1038 }
1039
1040 /*
1041  * when nowcow writeback call back.  This checks for snapshots or COW copies
1042  * of the extents that exist in the file, and COWs the file as required.
1043  *
1044  * If no cow copies or snapshots exist, we write directly to the existing
1045  * blocks on disk
1046  */
1047 static noinline int run_delalloc_nocow(struct inode *inode,
1048                                        struct page *locked_page,
1049                               u64 start, u64 end, int *page_started, int force,
1050                               unsigned long *nr_written)
1051 {
1052         struct btrfs_root *root = BTRFS_I(inode)->root;
1053         struct btrfs_trans_handle *trans;
1054         struct extent_buffer *leaf;
1055         struct btrfs_path *path;
1056         struct btrfs_file_extent_item *fi;
1057         struct btrfs_key found_key;
1058         u64 cow_start;
1059         u64 cur_offset;
1060         u64 extent_end;
1061         u64 extent_offset;
1062         u64 disk_bytenr;
1063         u64 num_bytes;
1064         int extent_type;
1065         int ret;
1066         int type;
1067         int nocow;
1068         int check_prev = 1;
1069         bool nolock;
1070         u64 ino = btrfs_ino(inode);
1071
1072         path = btrfs_alloc_path();
1073         BUG_ON(!path);
1074
1075         nolock = is_free_space_inode(root, inode);
1076
1077         if (nolock)
1078                 trans = btrfs_join_transaction_nolock(root);
1079         else
1080                 trans = btrfs_join_transaction(root);
1081
1082         BUG_ON(IS_ERR(trans));
1083         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1084
1085         cow_start = (u64)-1;
1086         cur_offset = start;
1087         while (1) {
1088                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
1089                                                cur_offset, 0);
1090                 BUG_ON(ret < 0);
1091                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1092                         leaf = path->nodes[0];
1093                         btrfs_item_key_to_cpu(leaf, &found_key,
1094                                               path->slots[0] - 1);
1095                         if (found_key.objectid == ino &&
1096                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1097                                 path->slots[0]--;
1098                 }
1099                 check_prev = 0;
1100 next_slot:
1101                 leaf = path->nodes[0];
1102                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1103                         ret = btrfs_next_leaf(root, path);
1104                         if (ret < 0)
1105                                 BUG_ON(1);
1106                         if (ret > 0)
1107                                 break;
1108                         leaf = path->nodes[0];
1109                 }
1110
1111                 nocow = 0;
1112                 disk_bytenr = 0;
1113                 num_bytes = 0;
1114                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1115
1116                 if (found_key.objectid > ino ||
1117                     found_key.type > BTRFS_EXTENT_DATA_KEY ||
1118                     found_key.offset > end)
1119                         break;
1120
1121                 if (found_key.offset > cur_offset) {
1122                         extent_end = found_key.offset;
1123                         extent_type = 0;
1124                         goto out_check;
1125                 }
1126
1127                 fi = btrfs_item_ptr(leaf, path->slots[0],
1128                                     struct btrfs_file_extent_item);
1129                 extent_type = btrfs_file_extent_type(leaf, fi);
1130
1131                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1132                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1133                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1134                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1135                         extent_end = found_key.offset +
1136                                 btrfs_file_extent_num_bytes(leaf, fi);
1137                         if (extent_end <= start) {
1138                                 path->slots[0]++;
1139                                 goto next_slot;
1140                         }
1141                         if (disk_bytenr == 0)
1142                                 goto out_check;
1143                         if (btrfs_file_extent_compression(leaf, fi) ||
1144                             btrfs_file_extent_encryption(leaf, fi) ||
1145                             btrfs_file_extent_other_encoding(leaf, fi))
1146                                 goto out_check;
1147                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1148                                 goto out_check;
1149                         if (btrfs_extent_readonly(root, disk_bytenr))
1150                                 goto out_check;
1151                         if (btrfs_cross_ref_exist(trans, root, ino,
1152                                                   found_key.offset -
1153                                                   extent_offset, disk_bytenr))
1154                                 goto out_check;
1155                         disk_bytenr += extent_offset;
1156                         disk_bytenr += cur_offset - found_key.offset;
1157                         num_bytes = min(end + 1, extent_end) - cur_offset;
1158                         /*
1159                          * force cow if csum exists in the range.
1160                          * this ensure that csum for a given extent are
1161                          * either valid or do not exist.
1162                          */
1163                         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
1164                                 goto out_check;
1165                         nocow = 1;
1166                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1167                         extent_end = found_key.offset +
1168                                 btrfs_file_extent_inline_len(leaf, fi);
1169                         extent_end = ALIGN(extent_end, root->sectorsize);
1170                 } else {
1171                         BUG_ON(1);
1172                 }
1173 out_check:
1174                 if (extent_end <= start) {
1175                         path->slots[0]++;
1176                         goto next_slot;
1177                 }
1178                 if (!nocow) {
1179                         if (cow_start == (u64)-1)
1180                                 cow_start = cur_offset;
1181                         cur_offset = extent_end;
1182                         if (cur_offset > end)
1183                                 break;
1184                         path->slots[0]++;
1185                         goto next_slot;
1186                 }
1187
1188                 btrfs_release_path(path);
1189                 if (cow_start != (u64)-1) {
1190                         ret = cow_file_range(inode, locked_page, cow_start,
1191                                         found_key.offset - 1, page_started,
1192                                         nr_written, 1);
1193                         BUG_ON(ret);
1194                         cow_start = (u64)-1;
1195                 }
1196
1197                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1198                         struct extent_map *em;
1199                         struct extent_map_tree *em_tree;
1200                         em_tree = &BTRFS_I(inode)->extent_tree;
1201                         em = alloc_extent_map();
1202                         BUG_ON(!em);
1203                         em->start = cur_offset;
1204                         em->orig_start = em->start;
1205                         em->len = num_bytes;
1206                         em->block_len = num_bytes;
1207                         em->block_start = disk_bytenr;
1208                         em->bdev = root->fs_info->fs_devices->latest_bdev;
1209                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1210                         while (1) {
1211                                 write_lock(&em_tree->lock);
1212                                 ret = add_extent_mapping(em_tree, em);
1213                                 write_unlock(&em_tree->lock);
1214                                 if (ret != -EEXIST) {
1215                                         free_extent_map(em);
1216                                         break;
1217                                 }
1218                                 btrfs_drop_extent_cache(inode, em->start,
1219                                                 em->start + em->len - 1, 0);
1220                         }
1221                         type = BTRFS_ORDERED_PREALLOC;
1222                 } else {
1223                         type = BTRFS_ORDERED_NOCOW;
1224                 }
1225
1226                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1227                                                num_bytes, num_bytes, type);
1228                 BUG_ON(ret);
1229
1230                 if (root->root_key.objectid ==
1231                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1232                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1233                                                       num_bytes);
1234                         BUG_ON(ret);
1235                 }
1236
1237                 extent_clear_unlock_delalloc(inode, &BTRFS_I(inode)->io_tree,
1238                                 cur_offset, cur_offset + num_bytes - 1,
1239                                 locked_page, EXTENT_CLEAR_UNLOCK_PAGE |
1240                                 EXTENT_CLEAR_UNLOCK | EXTENT_CLEAR_DELALLOC |
1241                                 EXTENT_SET_PRIVATE2);
1242                 cur_offset = extent_end;
1243                 if (cur_offset > end)
1244                         break;
1245         }
1246         btrfs_release_path(path);
1247
1248         if (cur_offset <= end && cow_start == (u64)-1)
1249                 cow_start = cur_offset;
1250         if (cow_start != (u64)-1) {
1251                 ret = cow_file_range(inode, locked_page, cow_start, end,
1252                                      page_started, nr_written, 1);
1253                 BUG_ON(ret);
1254         }
1255
1256         if (nolock) {
1257                 ret = btrfs_end_transaction_nolock(trans, root);
1258                 BUG_ON(ret);
1259         } else {
1260                 ret = btrfs_end_transaction(trans, root);
1261                 BUG_ON(ret);
1262         }
1263         btrfs_free_path(path);
1264         return 0;
1265 }
1266
1267 /*
1268  * extent_io.c call back to do delayed allocation processing
1269  */
1270 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1271                               u64 start, u64 end, int *page_started,
1272                               unsigned long *nr_written)
1273 {
1274         int ret;
1275         struct btrfs_root *root = BTRFS_I(inode)->root;
1276
1277         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW)
1278                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1279                                          page_started, 1, nr_written);
1280         else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)
1281                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1282                                          page_started, 0, nr_written);
1283         else if (!btrfs_test_opt(root, COMPRESS) &&
1284                  !(BTRFS_I(inode)->force_compress) &&
1285                  !(BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS))
1286                 ret = cow_file_range(inode, locked_page, start, end,
1287                                       page_started, nr_written, 1);
1288         else
1289                 ret = cow_file_range_async(inode, locked_page, start, end,
1290                                            page_started, nr_written);
1291         return ret;
1292 }
1293
1294 static int btrfs_split_extent_hook(struct inode *inode,
1295                                    struct extent_state *orig, u64 split)
1296 {
1297         /* not delalloc, ignore it */
1298         if (!(orig->state & EXTENT_DELALLOC))
1299                 return 0;
1300
1301         spin_lock(&BTRFS_I(inode)->lock);
1302         BTRFS_I(inode)->outstanding_extents++;
1303         spin_unlock(&BTRFS_I(inode)->lock);
1304         return 0;
1305 }
1306
1307 /*
1308  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1309  * extents so we can keep track of new extents that are just merged onto old
1310  * extents, such as when we are doing sequential writes, so we can properly
1311  * account for the metadata space we'll need.
1312  */
1313 static int btrfs_merge_extent_hook(struct inode *inode,
1314                                    struct extent_state *new,
1315                                    struct extent_state *other)
1316 {
1317         /* not delalloc, ignore it */
1318         if (!(other->state & EXTENT_DELALLOC))
1319                 return 0;
1320
1321         spin_lock(&BTRFS_I(inode)->lock);
1322         BTRFS_I(inode)->outstanding_extents--;
1323         spin_unlock(&BTRFS_I(inode)->lock);
1324         return 0;
1325 }
1326
1327 /*
1328  * extent_io.c set_bit_hook, used to track delayed allocation
1329  * bytes in this file, and to maintain the list of inodes that
1330  * have pending delalloc work to be done.
1331  */
1332 static int btrfs_set_bit_hook(struct inode *inode,
1333                               struct extent_state *state, int *bits)
1334 {
1335
1336         /*
1337          * set_bit and clear bit hooks normally require _irqsave/restore
1338          * but in this case, we are only testing for the DELALLOC
1339          * bit, which is only set or cleared with irqs on
1340          */
1341         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1342                 struct btrfs_root *root = BTRFS_I(inode)->root;
1343                 u64 len = state->end + 1 - state->start;
1344                 bool do_list = !is_free_space_inode(root, inode);
1345
1346                 if (*bits & EXTENT_FIRST_DELALLOC) {
1347                         *bits &= ~EXTENT_FIRST_DELALLOC;
1348                 } else {
1349                         spin_lock(&BTRFS_I(inode)->lock);
1350                         BTRFS_I(inode)->outstanding_extents++;
1351                         spin_unlock(&BTRFS_I(inode)->lock);
1352                 }
1353
1354                 spin_lock(&root->fs_info->delalloc_lock);
1355                 BTRFS_I(inode)->delalloc_bytes += len;
1356                 root->fs_info->delalloc_bytes += len;
1357                 if (do_list && list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1358                         list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1359                                       &root->fs_info->delalloc_inodes);
1360                 }
1361                 spin_unlock(&root->fs_info->delalloc_lock);
1362         }
1363         return 0;
1364 }
1365
1366 /*
1367  * extent_io.c clear_bit_hook, see set_bit_hook for why
1368  */
1369 static int btrfs_clear_bit_hook(struct inode *inode,
1370                                 struct extent_state *state, int *bits)
1371 {
1372         /*
1373          * set_bit and clear bit hooks normally require _irqsave/restore
1374          * but in this case, we are only testing for the DELALLOC
1375          * bit, which is only set or cleared with irqs on
1376          */
1377         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1378                 struct btrfs_root *root = BTRFS_I(inode)->root;
1379                 u64 len = state->end + 1 - state->start;
1380                 bool do_list = !is_free_space_inode(root, inode);
1381
1382                 if (*bits & EXTENT_FIRST_DELALLOC) {
1383                         *bits &= ~EXTENT_FIRST_DELALLOC;
1384                 } else if (!(*bits & EXTENT_DO_ACCOUNTING)) {
1385                         spin_lock(&BTRFS_I(inode)->lock);
1386                         BTRFS_I(inode)->outstanding_extents--;
1387                         spin_unlock(&BTRFS_I(inode)->lock);
1388                 }
1389
1390                 if (*bits & EXTENT_DO_ACCOUNTING)
1391                         btrfs_delalloc_release_metadata(inode, len);
1392
1393                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1394                     && do_list)
1395                         btrfs_free_reserved_data_space(inode, len);
1396
1397                 spin_lock(&root->fs_info->delalloc_lock);
1398                 root->fs_info->delalloc_bytes -= len;
1399                 BTRFS_I(inode)->delalloc_bytes -= len;
1400
1401                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1402                     !list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1403                         list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1404                 }
1405                 spin_unlock(&root->fs_info->delalloc_lock);
1406         }
1407         return 0;
1408 }
1409
1410 /*
1411  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1412  * we don't create bios that span stripes or chunks
1413  */
1414 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1415                          size_t size, struct bio *bio,
1416                          unsigned long bio_flags)
1417 {
1418         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
1419         struct btrfs_mapping_tree *map_tree;
1420         u64 logical = (u64)bio->bi_sector << 9;
1421         u64 length = 0;
1422         u64 map_length;
1423         int ret;
1424
1425         if (bio_flags & EXTENT_BIO_COMPRESSED)
1426                 return 0;
1427
1428         length = bio->bi_size;
1429         map_tree = &root->fs_info->mapping_tree;
1430         map_length = length;
1431         ret = btrfs_map_block(map_tree, READ, logical,
1432                               &map_length, NULL, 0);
1433
1434         if (map_length < length + size)
1435                 return 1;
1436         return ret;
1437 }
1438
1439 /*
1440  * in order to insert checksums into the metadata in large chunks,
1441  * we wait until bio submission time.   All the pages in the bio are
1442  * checksummed and sums are attached onto the ordered extent record.
1443  *
1444  * At IO completion time the cums attached on the ordered extent record
1445  * are inserted into the btree
1446  */
1447 static int __btrfs_submit_bio_start(struct inode *inode, int rw,
1448                                     struct bio *bio, int mirror_num,
1449                                     unsigned long bio_flags,
1450                                     u64 bio_offset)
1451 {
1452         struct btrfs_root *root = BTRFS_I(inode)->root;
1453         int ret = 0;
1454
1455         ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1456         BUG_ON(ret);
1457         return 0;
1458 }
1459
1460 /*
1461  * in order to insert checksums into the metadata in large chunks,
1462  * we wait until bio submission time.   All the pages in the bio are
1463  * checksummed and sums are attached onto the ordered extent record.
1464  *
1465  * At IO completion time the cums attached on the ordered extent record
1466  * are inserted into the btree
1467  */
1468 static int __btrfs_submit_bio_done(struct inode *inode, int rw, struct bio *bio,
1469                           int mirror_num, unsigned long bio_flags,
1470                           u64 bio_offset)
1471 {
1472         struct btrfs_root *root = BTRFS_I(inode)->root;
1473         return btrfs_map_bio(root, rw, bio, mirror_num, 1);
1474 }
1475
1476 /*
1477  * extent_io.c submission hook. This does the right thing for csum calculation
1478  * on write, or reading the csums from the tree before a read
1479  */
1480 static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
1481                           int mirror_num, unsigned long bio_flags,
1482                           u64 bio_offset)
1483 {
1484         struct btrfs_root *root = BTRFS_I(inode)->root;
1485         int ret = 0;
1486         int skip_sum;
1487
1488         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1489
1490         if (is_free_space_inode(root, inode))
1491                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, 2);
1492         else
1493                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
1494         BUG_ON(ret);
1495
1496         if (!(rw & REQ_WRITE)) {
1497                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1498                         return btrfs_submit_compressed_read(inode, bio,
1499                                                     mirror_num, bio_flags);
1500                 } else if (!skip_sum) {
1501                         ret = btrfs_lookup_bio_sums(root, inode, bio, NULL);
1502                         if (ret)
1503                                 return ret;
1504                 }
1505                 goto mapit;
1506         } else if (!skip_sum) {
1507                 /* csum items have already been cloned */
1508                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1509                         goto mapit;
1510                 /* we're doing a write, do the async checksumming */
1511                 return btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
1512                                    inode, rw, bio, mirror_num,
1513                                    bio_flags, bio_offset,
1514                                    __btrfs_submit_bio_start,
1515                                    __btrfs_submit_bio_done);
1516         }
1517
1518 mapit:
1519         return btrfs_map_bio(root, rw, bio, mirror_num, 0);
1520 }
1521
1522 /*
1523  * given a list of ordered sums record them in the inode.  This happens
1524  * at IO completion time based on sums calculated at bio submission time.
1525  */
1526 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1527                              struct inode *inode, u64 file_offset,
1528                              struct list_head *list)
1529 {
1530         struct btrfs_ordered_sum *sum;
1531
1532         list_for_each_entry(sum, list, list) {
1533                 btrfs_csum_file_blocks(trans,
1534                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1535         }
1536         return 0;
1537 }
1538
1539 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
1540                               struct extent_state **cached_state)
1541 {
1542         if ((end & (PAGE_CACHE_SIZE - 1)) == 0)
1543                 WARN_ON(1);
1544         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
1545                                    cached_state, GFP_NOFS);
1546 }
1547
1548 /* see btrfs_writepage_start_hook for details on why this is required */
1549 struct btrfs_writepage_fixup {
1550         struct page *page;
1551         struct btrfs_work work;
1552 };
1553
1554 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
1555 {
1556         struct btrfs_writepage_fixup *fixup;
1557         struct btrfs_ordered_extent *ordered;
1558         struct extent_state *cached_state = NULL;
1559         struct page *page;
1560         struct inode *inode;
1561         u64 page_start;
1562         u64 page_end;
1563
1564         fixup = container_of(work, struct btrfs_writepage_fixup, work);
1565         page = fixup->page;
1566 again:
1567         lock_page(page);
1568         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
1569                 ClearPageChecked(page);
1570                 goto out_page;
1571         }
1572
1573         inode = page->mapping->host;
1574         page_start = page_offset(page);
1575         page_end = page_offset(page) + PAGE_CACHE_SIZE - 1;
1576
1577         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end, 0,
1578                          &cached_state, GFP_NOFS);
1579
1580         /* already ordered? We're done */
1581         if (PagePrivate2(page))
1582                 goto out;
1583
1584         ordered = btrfs_lookup_ordered_extent(inode, page_start);
1585         if (ordered) {
1586                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
1587                                      page_end, &cached_state, GFP_NOFS);
1588                 unlock_page(page);
1589                 btrfs_start_ordered_extent(inode, ordered, 1);
1590                 goto again;
1591         }
1592
1593         BUG();
1594         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state);
1595         ClearPageChecked(page);
1596 out:
1597         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
1598                              &cached_state, GFP_NOFS);
1599 out_page:
1600         unlock_page(page);
1601         page_cache_release(page);
1602         kfree(fixup);
1603 }
1604
1605 /*
1606  * There are a few paths in the higher layers of the kernel that directly
1607  * set the page dirty bit without asking the filesystem if it is a
1608  * good idea.  This causes problems because we want to make sure COW
1609  * properly happens and the data=ordered rules are followed.
1610  *
1611  * In our case any range that doesn't have the ORDERED bit set
1612  * hasn't been properly setup for IO.  We kick off an async process
1613  * to fix it up.  The async helper will wait for ordered extents, set
1614  * the delalloc bit and make it safe to write the page.
1615  */
1616 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
1617 {
1618         struct inode *inode = page->mapping->host;
1619         struct btrfs_writepage_fixup *fixup;
1620         struct btrfs_root *root = BTRFS_I(inode)->root;
1621
1622         /* this page is properly in the ordered list */
1623         if (TestClearPagePrivate2(page))
1624                 return 0;
1625
1626         if (PageChecked(page))
1627                 return -EAGAIN;
1628
1629         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
1630         if (!fixup)
1631                 return -EAGAIN;
1632
1633         SetPageChecked(page);
1634         page_cache_get(page);
1635         fixup->work.func = btrfs_writepage_fixup_worker;
1636         fixup->page = page;
1637         btrfs_queue_worker(&root->fs_info->fixup_workers, &fixup->work);
1638         return -EAGAIN;
1639 }
1640
1641 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
1642                                        struct inode *inode, u64 file_pos,
1643                                        u64 disk_bytenr, u64 disk_num_bytes,
1644                                        u64 num_bytes, u64 ram_bytes,
1645                                        u8 compression, u8 encryption,
1646                                        u16 other_encoding, int extent_type)
1647 {
1648         struct btrfs_root *root = BTRFS_I(inode)->root;
1649         struct btrfs_file_extent_item *fi;
1650         struct btrfs_path *path;
1651         struct extent_buffer *leaf;
1652         struct btrfs_key ins;
1653         u64 hint;
1654         int ret;
1655
1656         path = btrfs_alloc_path();
1657         BUG_ON(!path);
1658
1659         path->leave_spinning = 1;
1660
1661         /*
1662          * we may be replacing one extent in the tree with another.
1663          * The new extent is pinned in the extent map, and we don't want
1664          * to drop it from the cache until it is completely in the btree.
1665          *
1666          * So, tell btrfs_drop_extents to leave this extent in the cache.
1667          * the caller is expected to unpin it and allow it to be merged
1668          * with the others.
1669          */
1670         ret = btrfs_drop_extents(trans, inode, file_pos, file_pos + num_bytes,
1671                                  &hint, 0);
1672         BUG_ON(ret);
1673
1674         ins.objectid = btrfs_ino(inode);
1675         ins.offset = file_pos;
1676         ins.type = BTRFS_EXTENT_DATA_KEY;
1677         ret = btrfs_insert_empty_item(trans, root, path, &ins, sizeof(*fi));
1678         BUG_ON(ret);
1679         leaf = path->nodes[0];
1680         fi = btrfs_item_ptr(leaf, path->slots[0],
1681                             struct btrfs_file_extent_item);
1682         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1683         btrfs_set_file_extent_type(leaf, fi, extent_type);
1684         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
1685         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
1686         btrfs_set_file_extent_offset(leaf, fi, 0);
1687         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
1688         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
1689         btrfs_set_file_extent_compression(leaf, fi, compression);
1690         btrfs_set_file_extent_encryption(leaf, fi, encryption);
1691         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
1692
1693         btrfs_unlock_up_safe(path, 1);
1694         btrfs_set_lock_blocking(leaf);
1695
1696         btrfs_mark_buffer_dirty(leaf);
1697
1698         inode_add_bytes(inode, num_bytes);
1699
1700         ins.objectid = disk_bytenr;
1701         ins.offset = disk_num_bytes;
1702         ins.type = BTRFS_EXTENT_ITEM_KEY;
1703         ret = btrfs_alloc_reserved_file_extent(trans, root,
1704                                         root->root_key.objectid,
1705                                         btrfs_ino(inode), file_pos, &ins);
1706         BUG_ON(ret);
1707         btrfs_free_path(path);
1708
1709         return 0;
1710 }
1711
1712 /*
1713  * helper function for btrfs_finish_ordered_io, this
1714  * just reads in some of the csum leaves to prime them into ram
1715  * before we start the transaction.  It limits the amount of btree
1716  * reads required while inside the transaction.
1717  */
1718 /* as ordered data IO finishes, this gets called so we can finish
1719  * an ordered extent if the range of bytes in the file it covers are
1720  * fully written.
1721  */
1722 static int btrfs_finish_ordered_io(struct inode *inode, u64 start, u64 end)
1723 {
1724         struct btrfs_root *root = BTRFS_I(inode)->root;
1725         struct btrfs_trans_handle *trans = NULL;
1726         struct btrfs_ordered_extent *ordered_extent = NULL;
1727         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
1728         struct extent_state *cached_state = NULL;
1729         int compress_type = 0;
1730         int ret;
1731         bool nolock;
1732
1733         ret = btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
1734                                              end - start + 1);
1735         if (!ret)
1736                 return 0;
1737         BUG_ON(!ordered_extent);
1738
1739         nolock = is_free_space_inode(root, inode);
1740
1741         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
1742                 BUG_ON(!list_empty(&ordered_extent->list));
1743                 ret = btrfs_ordered_update_i_size(inode, 0, ordered_extent);
1744                 if (!ret) {
1745                         if (nolock)
1746                                 trans = btrfs_join_transaction_nolock(root);
1747                         else
1748                                 trans = btrfs_join_transaction(root);
1749                         BUG_ON(IS_ERR(trans));
1750                         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1751                         ret = btrfs_update_inode(trans, root, inode);
1752                         BUG_ON(ret);
1753                 }
1754                 goto out;
1755         }
1756
1757         lock_extent_bits(io_tree, ordered_extent->file_offset,
1758                          ordered_extent->file_offset + ordered_extent->len - 1,
1759                          0, &cached_state, GFP_NOFS);
1760
1761         if (nolock)
1762                 trans = btrfs_join_transaction_nolock(root);
1763         else
1764                 trans = btrfs_join_transaction(root);
1765         BUG_ON(IS_ERR(trans));
1766         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1767
1768         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
1769                 compress_type = ordered_extent->compress_type;
1770         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
1771                 BUG_ON(compress_type);
1772                 ret = btrfs_mark_extent_written(trans, inode,
1773                                                 ordered_extent->file_offset,
1774                                                 ordered_extent->file_offset +
1775                                                 ordered_extent->len);
1776                 BUG_ON(ret);
1777         } else {
1778                 BUG_ON(root == root->fs_info->tree_root);
1779                 ret = insert_reserved_file_extent(trans, inode,
1780                                                 ordered_extent->file_offset,
1781                                                 ordered_extent->start,
1782                                                 ordered_extent->disk_len,
1783                                                 ordered_extent->len,
1784                                                 ordered_extent->len,
1785                                                 compress_type, 0, 0,
1786                                                 BTRFS_FILE_EXTENT_REG);
1787                 unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
1788                                    ordered_extent->file_offset,
1789                                    ordered_extent->len);
1790                 BUG_ON(ret);
1791         }
1792         unlock_extent_cached(io_tree, ordered_extent->file_offset,
1793                              ordered_extent->file_offset +
1794                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
1795
1796         add_pending_csums(trans, inode, ordered_extent->file_offset,
1797                           &ordered_extent->list);
1798
1799         ret = btrfs_ordered_update_i_size(inode, 0, ordered_extent);
1800         if (!ret) {
1801                 ret = btrfs_update_inode(trans, root, inode);
1802                 BUG_ON(ret);
1803         }
1804         ret = 0;
1805 out:
1806         if (nolock) {
1807                 if (trans)
1808                         btrfs_end_transaction_nolock(trans, root);
1809         } else {
1810                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
1811                 if (trans)
1812                         btrfs_end_transaction(trans, root);
1813         }
1814
1815         /* once for us */
1816         btrfs_put_ordered_extent(ordered_extent);
1817         /* once for the tree */
1818         btrfs_put_ordered_extent(ordered_extent);
1819
1820         return 0;
1821 }
1822
1823 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
1824                                 struct extent_state *state, int uptodate)
1825 {
1826         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
1827
1828         ClearPagePrivate2(page);
1829         return btrfs_finish_ordered_io(page->mapping->host, start, end);
1830 }
1831
1832 /*
1833  * When IO fails, either with EIO or csum verification fails, we
1834  * try other mirrors that might have a good copy of the data.  This
1835  * io_failure_record is used to record state as we go through all the
1836  * mirrors.  If another mirror has good data, the page is set up to date
1837  * and things continue.  If a good mirror can't be found, the original
1838  * bio end_io callback is called to indicate things have failed.
1839  */
1840 struct io_failure_record {
1841         struct page *page;
1842         u64 start;
1843         u64 len;
1844         u64 logical;
1845         unsigned long bio_flags;
1846         int last_mirror;
1847 };
1848
1849 static int btrfs_io_failed_hook(struct bio *failed_bio,
1850                          struct page *page, u64 start, u64 end,
1851                          struct extent_state *state)
1852 {
1853         struct io_failure_record *failrec = NULL;
1854         u64 private;
1855         struct extent_map *em;
1856         struct inode *inode = page->mapping->host;
1857         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1858         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
1859         struct bio *bio;
1860         int num_copies;
1861         int ret;
1862         int rw;
1863         u64 logical;
1864
1865         ret = get_state_private(failure_tree, start, &private);
1866         if (ret) {
1867                 failrec = kmalloc(sizeof(*failrec), GFP_NOFS);
1868                 if (!failrec)
1869                         return -ENOMEM;
1870                 failrec->start = start;
1871                 failrec->len = end - start + 1;
1872                 failrec->last_mirror = 0;
1873                 failrec->bio_flags = 0;
1874
1875                 read_lock(&em_tree->lock);
1876                 em = lookup_extent_mapping(em_tree, start, failrec->len);
1877                 if (em->start > start || em->start + em->len < start) {
1878                         free_extent_map(em);
1879                         em = NULL;
1880                 }
1881                 read_unlock(&em_tree->lock);
1882
1883                 if (IS_ERR_OR_NULL(em)) {
1884                         kfree(failrec);
1885                         return -EIO;
1886                 }
1887                 logical = start - em->start;
1888                 logical = em->block_start + logical;
1889                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
1890                         logical = em->block_start;
1891                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
1892                         extent_set_compress_type(&failrec->bio_flags,
1893                                                  em->compress_type);
1894                 }
1895                 failrec->logical = logical;
1896                 free_extent_map(em);
1897                 set_extent_bits(failure_tree, start, end, EXTENT_LOCKED |
1898                                 EXTENT_DIRTY, GFP_NOFS);
1899                 set_state_private(failure_tree, start,
1900                                  (u64)(unsigned long)failrec);
1901         } else {
1902                 failrec = (struct io_failure_record *)(unsigned long)private;
1903         }
1904         num_copies = btrfs_num_copies(
1905                               &BTRFS_I(inode)->root->fs_info->mapping_tree,
1906                               failrec->logical, failrec->len);
1907         failrec->last_mirror++;
1908         if (!state) {
1909                 spin_lock(&BTRFS_I(inode)->io_tree.lock);
1910                 state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
1911                                                     failrec->start,
1912                                                     EXTENT_LOCKED);
1913                 if (state && state->start != failrec->start)
1914                         state = NULL;
1915                 spin_unlock(&BTRFS_I(inode)->io_tree.lock);
1916         }
1917         if (!state || failrec->last_mirror > num_copies) {
1918                 set_state_private(failure_tree, failrec->start, 0);
1919                 clear_extent_bits(failure_tree, failrec->start,
1920                                   failrec->start + failrec->len - 1,
1921                                   EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1922                 kfree(failrec);
1923                 return -EIO;
1924         }
1925         bio = bio_alloc(GFP_NOFS, 1);
1926         bio->bi_private = state;
1927         bio->bi_end_io = failed_bio->bi_end_io;
1928         bio->bi_sector = failrec->logical >> 9;
1929         bio->bi_bdev = failed_bio->bi_bdev;
1930         bio->bi_size = 0;
1931
1932         bio_add_page(bio, page, failrec->len, start - page_offset(page));
1933         if (failed_bio->bi_rw & REQ_WRITE)
1934                 rw = WRITE;
1935         else
1936                 rw = READ;
1937
1938         ret = BTRFS_I(inode)->io_tree.ops->submit_bio_hook(inode, rw, bio,
1939                                                       failrec->last_mirror,
1940                                                       failrec->bio_flags, 0);
1941         return ret;
1942 }
1943
1944 /*
1945  * each time an IO finishes, we do a fast check in the IO failure tree
1946  * to see if we need to process or clean up an io_failure_record
1947  */
1948 static int btrfs_clean_io_failures(struct inode *inode, u64 start)
1949 {
1950         u64 private;
1951         u64 private_failure;
1952         struct io_failure_record *failure;
1953         int ret;
1954
1955         private = 0;
1956         if (count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
1957                              (u64)-1, 1, EXTENT_DIRTY, 0)) {
1958                 ret = get_state_private(&BTRFS_I(inode)->io_failure_tree,
1959                                         start, &private_failure);
1960                 if (ret == 0) {
1961                         failure = (struct io_failure_record *)(unsigned long)
1962                                    private_failure;
1963                         set_state_private(&BTRFS_I(inode)->io_failure_tree,
1964                                           failure->start, 0);
1965                         clear_extent_bits(&BTRFS_I(inode)->io_failure_tree,
1966                                           failure->start,
1967                                           failure->start + failure->len - 1,
1968                                           EXTENT_DIRTY | EXTENT_LOCKED,
1969                                           GFP_NOFS);
1970                         kfree(failure);
1971                 }
1972         }
1973         return 0;
1974 }
1975
1976 /*
1977  * when reads are done, we need to check csums to verify the data is correct
1978  * if there's a match, we allow the bio to finish.  If not, we go through
1979  * the io_failure_record routines to find good copies
1980  */
1981 static int btrfs_readpage_end_io_hook(struct page *page, u64 start, u64 end,
1982                                struct extent_state *state)
1983 {
1984         size_t offset = start - ((u64)page->index << PAGE_CACHE_SHIFT);
1985         struct inode *inode = page->mapping->host;
1986         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
1987         char *kaddr;
1988         u64 private = ~(u32)0;
1989         int ret;
1990         struct btrfs_root *root = BTRFS_I(inode)->root;
1991         u32 csum = ~(u32)0;
1992
1993         if (PageChecked(page)) {
1994                 ClearPageChecked(page);
1995                 goto good;
1996         }
1997
1998         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
1999                 goto good;
2000
2001         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
2002             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
2003                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM,
2004                                   GFP_NOFS);
2005                 return 0;
2006         }
2007
2008         if (state && state->start == start) {
2009                 private = state->private;
2010                 ret = 0;
2011         } else {
2012                 ret = get_state_private(io_tree, start, &private);
2013         }
2014         kaddr = kmap_atomic(page, KM_USER0);
2015         if (ret)
2016                 goto zeroit;
2017
2018         csum = btrfs_csum_data(root, kaddr + offset, csum,  end - start + 1);
2019         btrfs_csum_final(csum, (char *)&csum);
2020         if (csum != private)
2021                 goto zeroit;
2022
2023         kunmap_atomic(kaddr, KM_USER0);
2024 good:
2025         /* if the io failure tree for this inode is non-empty,
2026          * check to see if we've recovered from a failed IO
2027          */
2028         btrfs_clean_io_failures(inode, start);
2029         return 0;
2030
2031 zeroit:
2032         printk_ratelimited(KERN_INFO "btrfs csum failed ino %llu off %llu csum %u "
2033                        "private %llu\n",
2034                        (unsigned long long)btrfs_ino(page->mapping->host),
2035                        (unsigned long long)start, csum,
2036                        (unsigned long long)private);
2037         memset(kaddr + offset, 1, end - start + 1);
2038         flush_dcache_page(page);
2039         kunmap_atomic(kaddr, KM_USER0);
2040         if (private == 0)
2041                 return 0;
2042         return -EIO;
2043 }
2044
2045 struct delayed_iput {
2046         struct list_head list;
2047         struct inode *inode;
2048 };
2049
2050 void btrfs_add_delayed_iput(struct inode *inode)
2051 {
2052         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2053         struct delayed_iput *delayed;
2054
2055         if (atomic_add_unless(&inode->i_count, -1, 1))
2056                 return;
2057
2058         delayed = kmalloc(sizeof(*delayed), GFP_NOFS | __GFP_NOFAIL);
2059         delayed->inode = inode;
2060
2061         spin_lock(&fs_info->delayed_iput_lock);
2062         list_add_tail(&delayed->list, &fs_info->delayed_iputs);
2063         spin_unlock(&fs_info->delayed_iput_lock);
2064 }
2065
2066 void btrfs_run_delayed_iputs(struct btrfs_root *root)
2067 {
2068         LIST_HEAD(list);
2069         struct btrfs_fs_info *fs_info = root->fs_info;
2070         struct delayed_iput *delayed;
2071         int empty;
2072
2073         spin_lock(&fs_info->delayed_iput_lock);
2074         empty = list_empty(&fs_info->delayed_iputs);
2075         spin_unlock(&fs_info->delayed_iput_lock);
2076         if (empty)
2077                 return;
2078
2079         down_read(&root->fs_info->cleanup_work_sem);
2080         spin_lock(&fs_info->delayed_iput_lock);
2081         list_splice_init(&fs_info->delayed_iputs, &list);
2082         spin_unlock(&fs_info->delayed_iput_lock);
2083
2084         while (!list_empty(&list)) {
2085                 delayed = list_entry(list.next, struct delayed_iput, list);
2086                 list_del(&delayed->list);
2087                 iput(delayed->inode);
2088                 kfree(delayed);
2089         }
2090         up_read(&root->fs_info->cleanup_work_sem);
2091 }
2092
2093 /*
2094  * calculate extra metadata reservation when snapshotting a subvolume
2095  * contains orphan files.
2096  */
2097 void btrfs_orphan_pre_snapshot(struct btrfs_trans_handle *trans,
2098                                 struct btrfs_pending_snapshot *pending,
2099                                 u64 *bytes_to_reserve)
2100 {
2101         struct btrfs_root *root;
2102         struct btrfs_block_rsv *block_rsv;
2103         u64 num_bytes;
2104         int index;
2105
2106         root = pending->root;
2107         if (!root->orphan_block_rsv || list_empty(&root->orphan_list))
2108                 return;
2109
2110         block_rsv = root->orphan_block_rsv;
2111
2112         /* orphan block reservation for the snapshot */
2113         num_bytes = block_rsv->size;
2114
2115         /*
2116          * after the snapshot is created, COWing tree blocks may use more
2117          * space than it frees. So we should make sure there is enough
2118          * reserved space.
2119          */
2120         index = trans->transid & 0x1;
2121         if (block_rsv->reserved + block_rsv->freed[index] < block_rsv->size) {
2122                 num_bytes += block_rsv->size -
2123                              (block_rsv->reserved + block_rsv->freed[index]);
2124         }
2125
2126         *bytes_to_reserve += num_bytes;
2127 }
2128
2129 void btrfs_orphan_post_snapshot(struct btrfs_trans_handle *trans,
2130                                 struct btrfs_pending_snapshot *pending)
2131 {
2132         struct btrfs_root *root = pending->root;
2133         struct btrfs_root *snap = pending->snap;
2134         struct btrfs_block_rsv *block_rsv;
2135         u64 num_bytes;
2136         int index;
2137         int ret;
2138
2139         if (!root->orphan_block_rsv || list_empty(&root->orphan_list))
2140                 return;
2141
2142         /* refill source subvolume's orphan block reservation */
2143         block_rsv = root->orphan_block_rsv;
2144         index = trans->transid & 0x1;
2145         if (block_rsv->reserved + block_rsv->freed[index] < block_rsv->size) {
2146                 num_bytes = block_rsv->size -
2147                             (block_rsv->reserved + block_rsv->freed[index]);
2148                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
2149                                               root->orphan_block_rsv,
2150                                               num_bytes);
2151                 BUG_ON(ret);
2152         }
2153
2154         /* setup orphan block reservation for the snapshot */
2155         block_rsv = btrfs_alloc_block_rsv(snap);
2156         BUG_ON(!block_rsv);
2157
2158         btrfs_add_durable_block_rsv(root->fs_info, block_rsv);
2159         snap->orphan_block_rsv = block_rsv;
2160
2161         num_bytes = root->orphan_block_rsv->size;
2162         ret = btrfs_block_rsv_migrate(&pending->block_rsv,
2163                                       block_rsv, num_bytes);
2164         BUG_ON(ret);
2165
2166 #if 0
2167         /* insert orphan item for the snapshot */
2168         WARN_ON(!root->orphan_item_inserted);
2169         ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
2170                                        snap->root_key.objectid);
2171         BUG_ON(ret);
2172         snap->orphan_item_inserted = 1;
2173 #endif
2174 }
2175
2176 enum btrfs_orphan_cleanup_state {
2177         ORPHAN_CLEANUP_STARTED  = 1,
2178         ORPHAN_CLEANUP_DONE     = 2,
2179 };
2180
2181 /*
2182  * This is called in transaction commmit time. If there are no orphan
2183  * files in the subvolume, it removes orphan item and frees block_rsv
2184  * structure.
2185  */
2186 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
2187                               struct btrfs_root *root)
2188 {
2189         int ret;
2190
2191         if (!list_empty(&root->orphan_list) ||
2192             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
2193                 return;
2194
2195         if (root->orphan_item_inserted &&
2196             btrfs_root_refs(&root->root_item) > 0) {
2197                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
2198                                             root->root_key.objectid);
2199                 BUG_ON(ret);
2200                 root->orphan_item_inserted = 0;
2201         }
2202
2203         if (root->orphan_block_rsv) {
2204                 WARN_ON(root->orphan_block_rsv->size > 0);
2205                 btrfs_free_block_rsv(root, root->orphan_block_rsv);
2206                 root->orphan_block_rsv = NULL;
2207         }
2208 }
2209
2210 /*
2211  * This creates an orphan entry for the given inode in case something goes
2212  * wrong in the middle of an unlink/truncate.
2213  *
2214  * NOTE: caller of this function should reserve 5 units of metadata for
2215  *       this function.
2216  */
2217 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
2218 {
2219         struct btrfs_root *root = BTRFS_I(inode)->root;
2220         struct btrfs_block_rsv *block_rsv = NULL;
2221         int reserve = 0;
2222         int insert = 0;
2223         int ret;
2224
2225         if (!root->orphan_block_rsv) {
2226                 block_rsv = btrfs_alloc_block_rsv(root);
2227                 BUG_ON(!block_rsv);
2228         }
2229
2230         spin_lock(&root->orphan_lock);
2231         if (!root->orphan_block_rsv) {
2232                 root->orphan_block_rsv = block_rsv;
2233         } else if (block_rsv) {
2234                 btrfs_free_block_rsv(root, block_rsv);
2235                 block_rsv = NULL;
2236         }
2237
2238         if (list_empty(&BTRFS_I(inode)->i_orphan)) {
2239                 list_add(&BTRFS_I(inode)->i_orphan, &root->orphan_list);
2240 #if 0
2241                 /*
2242                  * For proper ENOSPC handling, we should do orphan
2243                  * cleanup when mounting. But this introduces backward
2244                  * compatibility issue.
2245                  */
2246                 if (!xchg(&root->orphan_item_inserted, 1))
2247                         insert = 2;
2248                 else
2249                         insert = 1;
2250 #endif
2251                 insert = 1;
2252         }
2253
2254         if (!BTRFS_I(inode)->orphan_meta_reserved) {
2255                 BTRFS_I(inode)->orphan_meta_reserved = 1;
2256                 reserve = 1;
2257         }
2258         spin_unlock(&root->orphan_lock);
2259
2260         if (block_rsv)
2261                 btrfs_add_durable_block_rsv(root->fs_info, block_rsv);
2262
2263         /* grab metadata reservation from transaction handle */
2264         if (reserve) {
2265                 ret = btrfs_orphan_reserve_metadata(trans, inode);
2266                 BUG_ON(ret);
2267         }
2268
2269         /* insert an orphan item to track this unlinked/truncated file */
2270         if (insert >= 1) {
2271                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
2272                 BUG_ON(ret);
2273         }
2274
2275         /* insert an orphan item to track subvolume contains orphan files */
2276         if (insert >= 2) {
2277                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
2278                                                root->root_key.objectid);
2279                 BUG_ON(ret);
2280         }
2281         return 0;
2282 }
2283
2284 /*
2285  * We have done the truncate/delete so we can go ahead and remove the orphan
2286  * item for this particular inode.
2287  */
2288 int btrfs_orphan_del(struct btrfs_trans_handle *trans, struct inode *inode)
2289 {
2290         struct btrfs_root *root = BTRFS_I(inode)->root;
2291         int delete_item = 0;
2292         int release_rsv = 0;
2293         int ret = 0;
2294
2295         spin_lock(&root->orphan_lock);
2296         if (!list_empty(&BTRFS_I(inode)->i_orphan)) {
2297                 list_del_init(&BTRFS_I(inode)->i_orphan);
2298                 delete_item = 1;
2299         }
2300
2301         if (BTRFS_I(inode)->orphan_meta_reserved) {
2302                 BTRFS_I(inode)->orphan_meta_reserved = 0;
2303                 release_rsv = 1;
2304         }
2305         spin_unlock(&root->orphan_lock);
2306
2307         if (trans && delete_item) {
2308                 ret = btrfs_del_orphan_item(trans, root, btrfs_ino(inode));
2309                 BUG_ON(ret);
2310         }
2311
2312         if (release_rsv)
2313                 btrfs_orphan_release_metadata(inode);
2314
2315         return 0;
2316 }
2317
2318 /*
2319  * this cleans up any orphans that may be left on the list from the last use
2320  * of this root.
2321  */
2322 int btrfs_orphan_cleanup(struct btrfs_root *root)
2323 {
2324         struct btrfs_path *path;
2325         struct extent_buffer *leaf;
2326         struct btrfs_key key, found_key;
2327         struct btrfs_trans_handle *trans;
2328         struct inode *inode;
2329         int ret = 0, nr_unlink = 0, nr_truncate = 0;
2330
2331         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
2332                 return 0;
2333
2334         path = btrfs_alloc_path();
2335         if (!path) {
2336                 ret = -ENOMEM;
2337                 goto out;
2338         }
2339         path->reada = -1;
2340
2341         key.objectid = BTRFS_ORPHAN_OBJECTID;
2342         btrfs_set_key_type(&key, BTRFS_ORPHAN_ITEM_KEY);
2343         key.offset = (u64)-1;
2344
2345         while (1) {
2346                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2347                 if (ret < 0)
2348                         goto out;
2349
2350                 /*
2351                  * if ret == 0 means we found what we were searching for, which
2352                  * is weird, but possible, so only screw with path if we didn't
2353                  * find the key and see if we have stuff that matches
2354                  */
2355                 if (ret > 0) {
2356                         ret = 0;
2357                         if (path->slots[0] == 0)
2358                                 break;
2359                         path->slots[0]--;
2360                 }
2361
2362                 /* pull out the item */
2363                 leaf = path->nodes[0];
2364                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2365
2366                 /* make sure the item matches what we want */
2367                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
2368                         break;
2369                 if (btrfs_key_type(&found_key) != BTRFS_ORPHAN_ITEM_KEY)
2370                         break;
2371
2372                 /* release the path since we're done with it */
2373                 btrfs_release_path(path);
2374
2375                 /*
2376                  * this is where we are basically btrfs_lookup, without the
2377                  * crossing root thing.  we store the inode number in the
2378                  * offset of the orphan item.
2379                  */
2380                 found_key.objectid = found_key.offset;
2381                 found_key.type = BTRFS_INODE_ITEM_KEY;
2382                 found_key.offset = 0;
2383                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
2384                 if (IS_ERR(inode)) {
2385                         ret = PTR_ERR(inode);
2386                         goto out;
2387                 }
2388
2389                 /*
2390                  * add this inode to the orphan list so btrfs_orphan_del does
2391                  * the proper thing when we hit it
2392                  */
2393                 spin_lock(&root->orphan_lock);
2394                 list_add(&BTRFS_I(inode)->i_orphan, &root->orphan_list);
2395                 spin_unlock(&root->orphan_lock);
2396
2397                 /*
2398                  * if this is a bad inode, means we actually succeeded in
2399                  * removing the inode, but not the orphan record, which means
2400                  * we need to manually delete the orphan since iput will just
2401                  * do a destroy_inode
2402                  */
2403                 if (is_bad_inode(inode)) {
2404                         trans = btrfs_start_transaction(root, 0);
2405                         if (IS_ERR(trans)) {
2406                                 ret = PTR_ERR(trans);
2407                                 goto out;
2408                         }
2409                         btrfs_orphan_del(trans, inode);
2410                         btrfs_end_transaction(trans, root);
2411                         iput(inode);
2412                         continue;
2413                 }
2414
2415                 /* if we have links, this was a truncate, lets do that */
2416                 if (inode->i_nlink) {
2417                         if (!S_ISREG(inode->i_mode)) {
2418                                 WARN_ON(1);
2419                                 iput(inode);
2420                                 continue;
2421                         }
2422                         nr_truncate++;
2423                         ret = btrfs_truncate(inode);
2424                 } else {
2425                         nr_unlink++;
2426                 }
2427
2428                 /* this will do delete_inode and everything for us */
2429                 iput(inode);
2430                 if (ret)
2431                         goto out;
2432         }
2433         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
2434
2435         if (root->orphan_block_rsv)
2436                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
2437                                         (u64)-1);
2438
2439         if (root->orphan_block_rsv || root->orphan_item_inserted) {
2440                 trans = btrfs_join_transaction(root);
2441                 if (!IS_ERR(trans))
2442                         btrfs_end_transaction(trans, root);
2443         }
2444
2445         if (nr_unlink)
2446                 printk(KERN_INFO "btrfs: unlinked %d orphans\n", nr_unlink);
2447         if (nr_truncate)
2448                 printk(KERN_INFO "btrfs: truncated %d orphans\n", nr_truncate);
2449
2450 out:
2451         if (ret)
2452                 printk(KERN_CRIT "btrfs: could not do orphan cleanup %d\n", ret);
2453         btrfs_free_path(path);
2454         return ret;
2455 }
2456
2457 /*
2458  * very simple check to peek ahead in the leaf looking for xattrs.  If we
2459  * don't find any xattrs, we know there can't be any acls.
2460  *
2461  * slot is the slot the inode is in, objectid is the objectid of the inode
2462  */
2463 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
2464                                           int slot, u64 objectid)
2465 {
2466         u32 nritems = btrfs_header_nritems(leaf);
2467         struct btrfs_key found_key;
2468         int scanned = 0;
2469
2470         slot++;
2471         while (slot < nritems) {
2472                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
2473
2474                 /* we found a different objectid, there must not be acls */
2475                 if (found_key.objectid != objectid)
2476                         return 0;
2477
2478                 /* we found an xattr, assume we've got an acl */
2479                 if (found_key.type == BTRFS_XATTR_ITEM_KEY)
2480                         return 1;
2481
2482                 /*
2483                  * we found a key greater than an xattr key, there can't
2484                  * be any acls later on
2485                  */
2486                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
2487                         return 0;
2488
2489                 slot++;
2490                 scanned++;
2491
2492                 /*
2493                  * it goes inode, inode backrefs, xattrs, extents,
2494                  * so if there are a ton of hard links to an inode there can
2495                  * be a lot of backrefs.  Don't waste time searching too hard,
2496                  * this is just an optimization
2497                  */
2498                 if (scanned >= 8)
2499                         break;
2500         }
2501         /* we hit the end of the leaf before we found an xattr or
2502          * something larger than an xattr.  We have to assume the inode
2503          * has acls
2504          */
2505         return 1;
2506 }
2507
2508 /*
2509  * read an inode from the btree into the in-memory inode
2510  */
2511 static void btrfs_read_locked_inode(struct inode *inode)
2512 {
2513         struct btrfs_path *path;
2514         struct extent_buffer *leaf;
2515         struct btrfs_inode_item *inode_item;
2516         struct btrfs_timespec *tspec;
2517         struct btrfs_root *root = BTRFS_I(inode)->root;
2518         struct btrfs_key location;
2519         int maybe_acls;
2520         u32 rdev;
2521         int ret;
2522         bool filled = false;
2523
2524         ret = btrfs_fill_inode(inode, &rdev);
2525         if (!ret)
2526                 filled = true;
2527
2528         path = btrfs_alloc_path();
2529         BUG_ON(!path);
2530         path->leave_spinning = 1;
2531         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
2532
2533         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
2534         if (ret)
2535                 goto make_bad;
2536
2537         leaf = path->nodes[0];
2538
2539         if (filled)
2540                 goto cache_acl;
2541
2542         inode_item = btrfs_item_ptr(leaf, path->slots[0],
2543                                     struct btrfs_inode_item);
2544         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
2545         inode->i_nlink = btrfs_inode_nlink(leaf, inode_item);
2546         inode->i_uid = btrfs_inode_uid(leaf, inode_item);
2547         inode->i_gid = btrfs_inode_gid(leaf, inode_item);
2548         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
2549
2550         tspec = btrfs_inode_atime(inode_item);
2551         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, tspec);
2552         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
2553
2554         tspec = btrfs_inode_mtime(inode_item);
2555         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, tspec);
2556         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
2557
2558         tspec = btrfs_inode_ctime(inode_item);
2559         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, tspec);
2560         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
2561
2562         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
2563         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
2564         BTRFS_I(inode)->sequence = btrfs_inode_sequence(leaf, inode_item);
2565         inode->i_generation = BTRFS_I(inode)->generation;
2566         inode->i_rdev = 0;
2567         rdev = btrfs_inode_rdev(leaf, inode_item);
2568
2569         BTRFS_I(inode)->index_cnt = (u64)-1;
2570         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
2571 cache_acl:
2572         /*
2573          * try to precache a NULL acl entry for files that don't have
2574          * any xattrs or acls
2575          */
2576         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
2577                                            btrfs_ino(inode));
2578         if (!maybe_acls)
2579                 cache_no_acl(inode);
2580
2581         btrfs_free_path(path);
2582
2583         switch (inode->i_mode & S_IFMT) {
2584         case S_IFREG:
2585                 inode->i_mapping->a_ops = &btrfs_aops;
2586                 inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
2587                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
2588                 inode->i_fop = &btrfs_file_operations;
2589                 inode->i_op = &btrfs_file_inode_operations;
2590                 break;
2591         case S_IFDIR:
2592                 inode->i_fop = &btrfs_dir_file_operations;
2593                 if (root == root->fs_info->tree_root)
2594                         inode->i_op = &btrfs_dir_ro_inode_operations;
2595                 else
2596                         inode->i_op = &btrfs_dir_inode_operations;
2597                 break;
2598         case S_IFLNK:
2599                 inode->i_op = &btrfs_symlink_inode_operations;
2600                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
2601                 inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
2602                 break;
2603         default:
2604                 inode->i_op = &btrfs_special_inode_operations;
2605                 init_special_inode(inode, inode->i_mode, rdev);
2606                 break;
2607         }
2608
2609         btrfs_update_iflags(inode);
2610         return;
2611
2612 make_bad:
2613         btrfs_free_path(path);
2614         make_bad_inode(inode);
2615 }
2616
2617 /*
2618  * given a leaf and an inode, copy the inode fields into the leaf
2619  */
2620 static void fill_inode_item(struct btrfs_trans_handle *trans,
2621                             struct extent_buffer *leaf,
2622                             struct btrfs_inode_item *item,
2623                             struct inode *inode)
2624 {
2625         btrfs_set_inode_uid(leaf, item, inode->i_uid);
2626         btrfs_set_inode_gid(leaf, item, inode->i_gid);
2627         btrfs_set_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size);
2628         btrfs_set_inode_mode(leaf, item, inode->i_mode);
2629         btrfs_set_inode_nlink(leaf, item, inode->i_nlink);
2630
2631         btrfs_set_timespec_sec(leaf, btrfs_inode_atime(item),
2632                                inode->i_atime.tv_sec);
2633         btrfs_set_timespec_nsec(leaf, btrfs_inode_atime(item),
2634                                 inode->i_atime.tv_nsec);
2635
2636         btrfs_set_timespec_sec(leaf, btrfs_inode_mtime(item),
2637                                inode->i_mtime.tv_sec);
2638         btrfs_set_timespec_nsec(leaf, btrfs_inode_mtime(item),
2639                                 inode->i_mtime.tv_nsec);
2640
2641         btrfs_set_timespec_sec(leaf, btrfs_inode_ctime(item),
2642                                inode->i_ctime.tv_sec);
2643         btrfs_set_timespec_nsec(leaf, btrfs_inode_ctime(item),
2644                                 inode->i_ctime.tv_nsec);
2645
2646         btrfs_set_inode_nbytes(leaf, item, inode_get_bytes(inode));
2647         btrfs_set_inode_generation(leaf, item, BTRFS_I(inode)->generation);
2648         btrfs_set_inode_sequence(leaf, item, BTRFS_I(inode)->sequence);
2649         btrfs_set_inode_transid(leaf, item, trans->transid);
2650         btrfs_set_inode_rdev(leaf, item, inode->i_rdev);
2651         btrfs_set_inode_flags(leaf, item, BTRFS_I(inode)->flags);
2652         btrfs_set_inode_block_group(leaf, item, 0);
2653 }
2654
2655 /*
2656  * copy everything in the in-memory inode into the btree.
2657  */
2658 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
2659                                 struct btrfs_root *root, struct inode *inode)
2660 {
2661         struct btrfs_inode_item *inode_item;
2662         struct btrfs_path *path;
2663         struct extent_buffer *leaf;
2664         int ret;
2665
2666         /*
2667          * If the inode is a free space inode, we can deadlock during commit
2668          * if we put it into the delayed code.
2669          *
2670          * The data relocation inode should also be directly updated
2671          * without delay
2672          */
2673         if (!is_free_space_inode(root, inode)
2674             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID) {
2675                 ret = btrfs_delayed_update_inode(trans, root, inode);
2676                 if (!ret)
2677                         btrfs_set_inode_last_trans(trans, inode);
2678                 return ret;
2679         }
2680
2681         path = btrfs_alloc_path();
2682         if (!path)
2683                 return -ENOMEM;
2684
2685         path->leave_spinning = 1;
2686         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
2687                                  1);
2688         if (ret) {
2689                 if (ret > 0)
2690                         ret = -ENOENT;
2691                 goto failed;
2692         }
2693
2694         btrfs_unlock_up_safe(path, 1);
2695         leaf = path->nodes[0];
2696         inode_item = btrfs_item_ptr(leaf, path->slots[0],
2697                                     struct btrfs_inode_item);
2698
2699         fill_inode_item(trans, leaf, inode_item, inode);
2700         btrfs_mark_buffer_dirty(leaf);
2701         btrfs_set_inode_last_trans(trans, inode);
2702         ret = 0;
2703 failed:
2704         btrfs_free_path(path);
2705         return ret;
2706 }
2707
2708 /*
2709  * unlink helper that gets used here in inode.c and in the tree logging
2710  * recovery code.  It remove a link in a directory with a given name, and
2711  * also drops the back refs in the inode to the directory
2712  */
2713 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
2714                                 struct btrfs_root *root,
2715                                 struct inode *dir, struct inode *inode,
2716                                 const char *name, int name_len)
2717 {
2718         struct btrfs_path *path;
2719         int ret = 0;
2720         struct extent_buffer *leaf;
2721         struct btrfs_dir_item *di;
2722         struct btrfs_key key;
2723         u64 index;
2724         u64 ino = btrfs_ino(inode);
2725         u64 dir_ino = btrfs_ino(dir);
2726
2727         path = btrfs_alloc_path();
2728         if (!path) {
2729                 ret = -ENOMEM;
2730                 goto out;
2731         }
2732
2733         path->leave_spinning = 1;
2734         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
2735                                     name, name_len, -1);
2736         if (IS_ERR(di)) {
2737                 ret = PTR_ERR(di);
2738                 goto err;
2739         }
2740         if (!di) {
2741                 ret = -ENOENT;
2742                 goto err;
2743         }
2744         leaf = path->nodes[0];
2745         btrfs_dir_item_key_to_cpu(leaf, di, &key);
2746         ret = btrfs_delete_one_dir_name(trans, root, path, di);
2747         if (ret)
2748                 goto err;
2749         btrfs_release_path(path);
2750
2751         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
2752                                   dir_ino, &index);
2753         if (ret) {
2754                 printk(KERN_INFO "btrfs failed to delete reference to %.*s, "
2755                        "inode %llu parent %llu\n", name_len, name,
2756                        (unsigned long long)ino, (unsigned long long)dir_ino);
2757                 goto err;
2758         }
2759
2760         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
2761         if (ret)
2762                 goto err;
2763
2764         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
2765                                          inode, dir_ino);
2766         BUG_ON(ret != 0 && ret != -ENOENT);
2767
2768         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
2769                                            dir, index);
2770         if (ret == -ENOENT)
2771                 ret = 0;
2772 err:
2773         btrfs_free_path(path);
2774         if (ret)
2775                 goto out;
2776
2777         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
2778         inode->i_ctime = dir->i_mtime = dir->i_ctime = CURRENT_TIME;
2779         btrfs_update_inode(trans, root, dir);
2780 out:
2781         return ret;
2782 }
2783
2784 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
2785                        struct btrfs_root *root,
2786                        struct inode *dir, struct inode *inode,
2787                        const char *name, int name_len)
2788 {
2789         int ret;
2790         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
2791         if (!ret) {
2792                 btrfs_drop_nlink(inode);
2793                 ret = btrfs_update_inode(trans, root, inode);
2794         }
2795         return ret;
2796 }
2797                 
2798
2799 /* helper to check if there is any shared block in the path */
2800 static int check_path_shared(struct btrfs_root *root,
2801                              struct btrfs_path *path)
2802 {
2803         struct extent_buffer *eb;
2804         int level;
2805         u64 refs = 1;
2806
2807         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2808                 int ret;
2809
2810                 if (!path->nodes[level])
2811                         break;
2812                 eb = path->nodes[level];
2813                 if (!btrfs_block_can_be_shared(root, eb))
2814                         continue;
2815                 ret = btrfs_lookup_extent_info(NULL, root, eb->start, eb->len,
2816                                                &refs, NULL);
2817                 if (refs > 1)
2818                         return 1;
2819         }
2820         return 0;
2821 }
2822
2823 /*
2824  * helper to start transaction for unlink and rmdir.
2825  *
2826  * unlink and rmdir are special in btrfs, they do not always free space.
2827  * so in enospc case, we should make sure they will free space before
2828  * allowing them to use the global metadata reservation.
2829  */
2830 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir,
2831                                                        struct dentry *dentry)
2832 {
2833         struct btrfs_trans_handle *trans;
2834         struct btrfs_root *root = BTRFS_I(dir)->root;
2835         struct btrfs_path *path;
2836         struct btrfs_inode_ref *ref;
2837         struct btrfs_dir_item *di;
2838         struct inode *inode = dentry->d_inode;
2839         u64 index;
2840         int check_link = 1;
2841         int err = -ENOSPC;
2842         int ret;
2843         u64 ino = btrfs_ino(inode);
2844         u64 dir_ino = btrfs_ino(dir);
2845
2846         trans = btrfs_start_transaction(root, 10);
2847         if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
2848                 return trans;
2849
2850         if (ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
2851                 return ERR_PTR(-ENOSPC);
2852
2853         /* check if there is someone else holds reference */
2854         if (S_ISDIR(inode->i_mode) && atomic_read(&inode->i_count) > 1)
2855                 return ERR_PTR(-ENOSPC);
2856
2857         if (atomic_read(&inode->i_count) > 2)
2858                 return ERR_PTR(-ENOSPC);
2859
2860         if (xchg(&root->fs_info->enospc_unlink, 1))
2861                 return ERR_PTR(-ENOSPC);
2862
2863         path = btrfs_alloc_path();
2864         if (!path) {
2865                 root->fs_info->enospc_unlink = 0;
2866                 return ERR_PTR(-ENOMEM);
2867         }
2868
2869         trans = btrfs_start_transaction(root, 0);
2870         if (IS_ERR(trans)) {
2871                 btrfs_free_path(path);
2872                 root->fs_info->enospc_unlink = 0;
2873                 return trans;
2874         }
2875
2876         path->skip_locking = 1;
2877         path->search_commit_root = 1;
2878
2879         ret = btrfs_lookup_inode(trans, root, path,
2880                                 &BTRFS_I(dir)->location, 0);
2881         if (ret < 0) {
2882                 err = ret;
2883                 goto out;
2884         }
2885         if (ret == 0) {
2886                 if (check_path_shared(root, path))
2887                         goto out;
2888         } else {
2889                 check_link = 0;
2890         }
2891         btrfs_release_path(path);
2892
2893         ret = btrfs_lookup_inode(trans, root, path,
2894                                 &BTRFS_I(inode)->location, 0);
2895         if (ret < 0) {
2896                 err = ret;
2897                 goto out;
2898         }
2899         if (ret == 0) {
2900                 if (check_path_shared(root, path))
2901                         goto out;
2902         } else {
2903                 check_link = 0;
2904         }
2905         btrfs_release_path(path);
2906
2907         if (ret == 0 && S_ISREG(inode->i_mode)) {
2908                 ret = btrfs_lookup_file_extent(trans, root, path,
2909                                                ino, (u64)-1, 0);
2910                 if (ret < 0) {
2911                         err = ret;
2912                         goto out;
2913                 }
2914                 BUG_ON(ret == 0);
2915                 if (check_path_shared(root, path))
2916                         goto out;
2917                 btrfs_release_path(path);
2918         }
2919
2920         if (!check_link) {
2921                 err = 0;
2922                 goto out;
2923         }
2924
2925         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
2926                                 dentry->d_name.name, dentry->d_name.len, 0);
2927         if (IS_ERR(di)) {
2928                 err = PTR_ERR(di);
2929                 goto out;
2930         }
2931         if (di) {
2932                 if (check_path_shared(root, path))
2933                         goto out;
2934         } else {
2935                 err = 0;
2936                 goto out;
2937         }
2938         btrfs_release_path(path);
2939
2940         ref = btrfs_lookup_inode_ref(trans, root, path,
2941                                 dentry->d_name.name, dentry->d_name.len,
2942                                 ino, dir_ino, 0);
2943         if (IS_ERR(ref)) {
2944                 err = PTR_ERR(ref);
2945                 goto out;
2946         }
2947         BUG_ON(!ref);
2948         if (check_path_shared(root, path))
2949                 goto out;
2950         index = btrfs_inode_ref_index(path->nodes[0], ref);
2951         btrfs_release_path(path);
2952
2953         /*
2954          * This is a commit root search, if we can lookup inode item and other
2955          * relative items in the commit root, it means the transaction of
2956          * dir/file creation has been committed, and the dir index item that we
2957          * delay to insert has also been inserted into the commit root. So
2958          * we needn't worry about the delayed insertion of the dir index item
2959          * here.
2960          */
2961         di = btrfs_lookup_dir_index_item(trans, root, path, dir_ino, index,
2962                                 dentry->d_name.name, dentry->d_name.len, 0);
2963         if (IS_ERR(di)) {
2964                 err = PTR_ERR(di);
2965                 goto out;
2966         }
2967         BUG_ON(ret == -ENOENT);
2968         if (check_path_shared(root, path))
2969                 goto out;
2970
2971         err = 0;
2972 out:
2973         btrfs_free_path(path);
2974         if (err) {
2975                 btrfs_end_transaction(trans, root);
2976                 root->fs_info->enospc_unlink = 0;
2977                 return ERR_PTR(err);
2978         }
2979
2980         trans->block_rsv = &root->fs_info->global_block_rsv;
2981         return trans;
2982 }
2983
2984 static void __unlink_end_trans(struct btrfs_trans_handle *trans,
2985                                struct btrfs_root *root)
2986 {
2987         if (trans->block_rsv == &root->fs_info->global_block_rsv) {
2988                 BUG_ON(!root->fs_info->enospc_unlink);
2989                 root->fs_info->enospc_unlink = 0;
2990         }
2991         btrfs_end_transaction_throttle(trans, root);
2992 }
2993
2994 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
2995 {
2996         struct btrfs_root *root = BTRFS_I(dir)->root;
2997         struct btrfs_trans_handle *trans;
2998         struct inode *inode = dentry->d_inode;
2999         int ret;
3000         unsigned long nr = 0;
3001
3002         trans = __unlink_start_trans(dir, dentry);
3003         if (IS_ERR(trans))
3004                 return PTR_ERR(trans);
3005
3006         btrfs_record_unlink_dir(trans, dir, dentry->d_inode, 0);
3007
3008         ret = btrfs_unlink_inode(trans, root, dir, dentry->d_inode,
3009                                  dentry->d_name.name, dentry->d_name.len);
3010         BUG_ON(ret);
3011
3012         if (inode->i_nlink == 0) {
3013                 ret = btrfs_orphan_add(trans, inode);
3014                 BUG_ON(ret);
3015         }
3016
3017         nr = trans->blocks_used;
3018         __unlink_end_trans(trans, root);
3019         btrfs_btree_balance_dirty(root, nr);
3020         return ret;
3021 }
3022
3023 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
3024                         struct btrfs_root *root,
3025                         struct inode *dir, u64 objectid,
3026                         const char *name, int name_len)
3027 {
3028         struct btrfs_path *path;
3029         struct extent_buffer *leaf;
3030         struct btrfs_dir_item *di;
3031         struct btrfs_key key;
3032         u64 index;
3033         int ret;
3034         u64 dir_ino = btrfs_ino(dir);
3035
3036         path = btrfs_alloc_path();
3037         if (!path)
3038                 return -ENOMEM;
3039
3040         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3041                                    name, name_len, -1);
3042         BUG_ON(IS_ERR_OR_NULL(di));
3043
3044         leaf = path->nodes[0];
3045         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3046         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
3047         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3048         BUG_ON(ret);
3049         btrfs_release_path(path);
3050
3051         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
3052                                  objectid, root->root_key.objectid,
3053                                  dir_ino, &index, name, name_len);
3054         if (ret < 0) {
3055                 BUG_ON(ret != -ENOENT);
3056                 di = btrfs_search_dir_index_item(root, path, dir_ino,
3057                                                  name, name_len);
3058                 BUG_ON(IS_ERR_OR_NULL(di));
3059
3060                 leaf = path->nodes[0];
3061                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3062                 btrfs_release_path(path);
3063                 index = key.offset;
3064         }
3065         btrfs_release_path(path);
3066
3067         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
3068         BUG_ON(ret);
3069
3070         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
3071         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
3072         ret = btrfs_update_inode(trans, root, dir);
3073         BUG_ON(ret);
3074
3075         btrfs_free_path(path);
3076         return 0;
3077 }
3078
3079 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
3080 {
3081         struct inode *inode = dentry->d_inode;
3082         int err = 0;
3083         struct btrfs_root *root = BTRFS_I(dir)->root;
3084         struct btrfs_trans_handle *trans;
3085         unsigned long nr = 0;
3086
3087         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE ||
3088             btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
3089                 return -ENOTEMPTY;