btrfs: drop min_size from evict_refill_and_join
[platform/kernel/linux-starfive.git] / fs / btrfs / inode.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2007 Oracle.  All rights reserved.
4  */
5
6 #include <linux/kernel.h>
7 #include <linux/bio.h>
8 #include <linux/buffer_head.h>
9 #include <linux/file.h>
10 #include <linux/fs.h>
11 #include <linux/pagemap.h>
12 #include <linux/highmem.h>
13 #include <linux/time.h>
14 #include <linux/init.h>
15 #include <linux/string.h>
16 #include <linux/backing-dev.h>
17 #include <linux/writeback.h>
18 #include <linux/compat.h>
19 #include <linux/xattr.h>
20 #include <linux/posix_acl.h>
21 #include <linux/falloc.h>
22 #include <linux/slab.h>
23 #include <linux/ratelimit.h>
24 #include <linux/btrfs.h>
25 #include <linux/blkdev.h>
26 #include <linux/posix_acl_xattr.h>
27 #include <linux/uio.h>
28 #include <linux/magic.h>
29 #include <linux/iversion.h>
30 #include <asm/unaligned.h>
31 #include "ctree.h"
32 #include "disk-io.h"
33 #include "transaction.h"
34 #include "btrfs_inode.h"
35 #include "print-tree.h"
36 #include "ordered-data.h"
37 #include "xattr.h"
38 #include "tree-log.h"
39 #include "volumes.h"
40 #include "compression.h"
41 #include "locking.h"
42 #include "free-space-cache.h"
43 #include "inode-map.h"
44 #include "backref.h"
45 #include "props.h"
46 #include "qgroup.h"
47 #include "dedupe.h"
48
49 struct btrfs_iget_args {
50         struct btrfs_key *location;
51         struct btrfs_root *root;
52 };
53
54 struct btrfs_dio_data {
55         u64 reserve;
56         u64 unsubmitted_oe_range_start;
57         u64 unsubmitted_oe_range_end;
58         int overwrite;
59 };
60
61 static const struct inode_operations btrfs_dir_inode_operations;
62 static const struct inode_operations btrfs_symlink_inode_operations;
63 static const struct inode_operations btrfs_dir_ro_inode_operations;
64 static const struct inode_operations btrfs_special_inode_operations;
65 static const struct inode_operations btrfs_file_inode_operations;
66 static const struct address_space_operations btrfs_aops;
67 static const struct file_operations btrfs_dir_file_operations;
68 static const struct extent_io_ops btrfs_extent_io_ops;
69
70 static struct kmem_cache *btrfs_inode_cachep;
71 struct kmem_cache *btrfs_trans_handle_cachep;
72 struct kmem_cache *btrfs_path_cachep;
73 struct kmem_cache *btrfs_free_space_cachep;
74
75 #define S_SHIFT 12
76 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
77         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
78         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
79         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
80         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
81         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
82         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
83         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
84 };
85
86 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
87 static int btrfs_truncate(struct inode *inode, bool skip_writeback);
88 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
89 static noinline int cow_file_range(struct inode *inode,
90                                    struct page *locked_page,
91                                    u64 start, u64 end, u64 delalloc_end,
92                                    int *page_started, unsigned long *nr_written,
93                                    int unlock, struct btrfs_dedupe_hash *hash);
94 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
95                                        u64 orig_start, u64 block_start,
96                                        u64 block_len, u64 orig_block_len,
97                                        u64 ram_bytes, int compress_type,
98                                        int type);
99
100 static void __endio_write_update_ordered(struct inode *inode,
101                                          const u64 offset, const u64 bytes,
102                                          const bool uptodate);
103
104 /*
105  * Cleanup all submitted ordered extents in specified range to handle errors
106  * from the fill_dellaloc() callback.
107  *
108  * NOTE: caller must ensure that when an error happens, it can not call
109  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
110  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
111  * to be released, which we want to happen only when finishing the ordered
112  * extent (btrfs_finish_ordered_io()). Also note that the caller of the
113  * fill_delalloc() callback already does proper cleanup for the first page of
114  * the range, that is, it invokes the callback writepage_end_io_hook() for the
115  * range of the first page.
116  */
117 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
118                                                  const u64 offset,
119                                                  const u64 bytes)
120 {
121         unsigned long index = offset >> PAGE_SHIFT;
122         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
123         struct page *page;
124
125         while (index <= end_index) {
126                 page = find_get_page(inode->i_mapping, index);
127                 index++;
128                 if (!page)
129                         continue;
130                 ClearPagePrivate2(page);
131                 put_page(page);
132         }
133         return __endio_write_update_ordered(inode, offset + PAGE_SIZE,
134                                             bytes - PAGE_SIZE, false);
135 }
136
137 static int btrfs_dirty_inode(struct inode *inode);
138
139 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
140 void btrfs_test_inode_set_ops(struct inode *inode)
141 {
142         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
143 }
144 #endif
145
146 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
147                                      struct inode *inode,  struct inode *dir,
148                                      const struct qstr *qstr)
149 {
150         int err;
151
152         err = btrfs_init_acl(trans, inode, dir);
153         if (!err)
154                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
155         return err;
156 }
157
158 /*
159  * this does all the hard work for inserting an inline extent into
160  * the btree.  The caller should have done a btrfs_drop_extents so that
161  * no overlapping inline items exist in the btree
162  */
163 static int insert_inline_extent(struct btrfs_trans_handle *trans,
164                                 struct btrfs_path *path, int extent_inserted,
165                                 struct btrfs_root *root, struct inode *inode,
166                                 u64 start, size_t size, size_t compressed_size,
167                                 int compress_type,
168                                 struct page **compressed_pages)
169 {
170         struct extent_buffer *leaf;
171         struct page *page = NULL;
172         char *kaddr;
173         unsigned long ptr;
174         struct btrfs_file_extent_item *ei;
175         int ret;
176         size_t cur_size = size;
177         unsigned long offset;
178
179         if (compressed_size && compressed_pages)
180                 cur_size = compressed_size;
181
182         inode_add_bytes(inode, size);
183
184         if (!extent_inserted) {
185                 struct btrfs_key key;
186                 size_t datasize;
187
188                 key.objectid = btrfs_ino(BTRFS_I(inode));
189                 key.offset = start;
190                 key.type = BTRFS_EXTENT_DATA_KEY;
191
192                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
193                 path->leave_spinning = 1;
194                 ret = btrfs_insert_empty_item(trans, root, path, &key,
195                                               datasize);
196                 if (ret)
197                         goto fail;
198         }
199         leaf = path->nodes[0];
200         ei = btrfs_item_ptr(leaf, path->slots[0],
201                             struct btrfs_file_extent_item);
202         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
203         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
204         btrfs_set_file_extent_encryption(leaf, ei, 0);
205         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
206         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
207         ptr = btrfs_file_extent_inline_start(ei);
208
209         if (compress_type != BTRFS_COMPRESS_NONE) {
210                 struct page *cpage;
211                 int i = 0;
212                 while (compressed_size > 0) {
213                         cpage = compressed_pages[i];
214                         cur_size = min_t(unsigned long, compressed_size,
215                                        PAGE_SIZE);
216
217                         kaddr = kmap_atomic(cpage);
218                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
219                         kunmap_atomic(kaddr);
220
221                         i++;
222                         ptr += cur_size;
223                         compressed_size -= cur_size;
224                 }
225                 btrfs_set_file_extent_compression(leaf, ei,
226                                                   compress_type);
227         } else {
228                 page = find_get_page(inode->i_mapping,
229                                      start >> PAGE_SHIFT);
230                 btrfs_set_file_extent_compression(leaf, ei, 0);
231                 kaddr = kmap_atomic(page);
232                 offset = start & (PAGE_SIZE - 1);
233                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
234                 kunmap_atomic(kaddr);
235                 put_page(page);
236         }
237         btrfs_mark_buffer_dirty(leaf);
238         btrfs_release_path(path);
239
240         /*
241          * we're an inline extent, so nobody can
242          * extend the file past i_size without locking
243          * a page we already have locked.
244          *
245          * We must do any isize and inode updates
246          * before we unlock the pages.  Otherwise we
247          * could end up racing with unlink.
248          */
249         BTRFS_I(inode)->disk_i_size = inode->i_size;
250         ret = btrfs_update_inode(trans, root, inode);
251
252 fail:
253         return ret;
254 }
255
256
257 /*
258  * conditionally insert an inline extent into the file.  This
259  * does the checks required to make sure the data is small enough
260  * to fit as an inline extent.
261  */
262 static noinline int cow_file_range_inline(struct inode *inode, u64 start,
263                                           u64 end, size_t compressed_size,
264                                           int compress_type,
265                                           struct page **compressed_pages)
266 {
267         struct btrfs_root *root = BTRFS_I(inode)->root;
268         struct btrfs_fs_info *fs_info = root->fs_info;
269         struct btrfs_trans_handle *trans;
270         u64 isize = i_size_read(inode);
271         u64 actual_end = min(end + 1, isize);
272         u64 inline_len = actual_end - start;
273         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
274         u64 data_len = inline_len;
275         int ret;
276         struct btrfs_path *path;
277         int extent_inserted = 0;
278         u32 extent_item_size;
279
280         if (compressed_size)
281                 data_len = compressed_size;
282
283         if (start > 0 ||
284             actual_end > fs_info->sectorsize ||
285             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
286             (!compressed_size &&
287             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
288             end + 1 < isize ||
289             data_len > fs_info->max_inline) {
290                 return 1;
291         }
292
293         path = btrfs_alloc_path();
294         if (!path)
295                 return -ENOMEM;
296
297         trans = btrfs_join_transaction(root);
298         if (IS_ERR(trans)) {
299                 btrfs_free_path(path);
300                 return PTR_ERR(trans);
301         }
302         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
303
304         if (compressed_size && compressed_pages)
305                 extent_item_size = btrfs_file_extent_calc_inline_size(
306                    compressed_size);
307         else
308                 extent_item_size = btrfs_file_extent_calc_inline_size(
309                     inline_len);
310
311         ret = __btrfs_drop_extents(trans, root, inode, path,
312                                    start, aligned_end, NULL,
313                                    1, 1, extent_item_size, &extent_inserted);
314         if (ret) {
315                 btrfs_abort_transaction(trans, ret);
316                 goto out;
317         }
318
319         if (isize > actual_end)
320                 inline_len = min_t(u64, isize, actual_end);
321         ret = insert_inline_extent(trans, path, extent_inserted,
322                                    root, inode, start,
323                                    inline_len, compressed_size,
324                                    compress_type, compressed_pages);
325         if (ret && ret != -ENOSPC) {
326                 btrfs_abort_transaction(trans, ret);
327                 goto out;
328         } else if (ret == -ENOSPC) {
329                 ret = 1;
330                 goto out;
331         }
332
333         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
334         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
335 out:
336         /*
337          * Don't forget to free the reserved space, as for inlined extent
338          * it won't count as data extent, free them directly here.
339          * And at reserve time, it's always aligned to page size, so
340          * just free one page here.
341          */
342         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
343         btrfs_free_path(path);
344         btrfs_end_transaction(trans);
345         return ret;
346 }
347
348 struct async_extent {
349         u64 start;
350         u64 ram_size;
351         u64 compressed_size;
352         struct page **pages;
353         unsigned long nr_pages;
354         int compress_type;
355         struct list_head list;
356 };
357
358 struct async_cow {
359         struct inode *inode;
360         struct btrfs_root *root;
361         struct page *locked_page;
362         u64 start;
363         u64 end;
364         unsigned int write_flags;
365         struct list_head extents;
366         struct btrfs_work work;
367 };
368
369 static noinline int add_async_extent(struct async_cow *cow,
370                                      u64 start, u64 ram_size,
371                                      u64 compressed_size,
372                                      struct page **pages,
373                                      unsigned long nr_pages,
374                                      int compress_type)
375 {
376         struct async_extent *async_extent;
377
378         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
379         BUG_ON(!async_extent); /* -ENOMEM */
380         async_extent->start = start;
381         async_extent->ram_size = ram_size;
382         async_extent->compressed_size = compressed_size;
383         async_extent->pages = pages;
384         async_extent->nr_pages = nr_pages;
385         async_extent->compress_type = compress_type;
386         list_add_tail(&async_extent->list, &cow->extents);
387         return 0;
388 }
389
390 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
391 {
392         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
393
394         /* force compress */
395         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
396                 return 1;
397         /* defrag ioctl */
398         if (BTRFS_I(inode)->defrag_compress)
399                 return 1;
400         /* bad compression ratios */
401         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
402                 return 0;
403         if (btrfs_test_opt(fs_info, COMPRESS) ||
404             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
405             BTRFS_I(inode)->prop_compress)
406                 return btrfs_compress_heuristic(inode, start, end);
407         return 0;
408 }
409
410 static inline void inode_should_defrag(struct btrfs_inode *inode,
411                 u64 start, u64 end, u64 num_bytes, u64 small_write)
412 {
413         /* If this is a small write inside eof, kick off a defrag */
414         if (num_bytes < small_write &&
415             (start > 0 || end + 1 < inode->disk_i_size))
416                 btrfs_add_inode_defrag(NULL, inode);
417 }
418
419 /*
420  * we create compressed extents in two phases.  The first
421  * phase compresses a range of pages that have already been
422  * locked (both pages and state bits are locked).
423  *
424  * This is done inside an ordered work queue, and the compression
425  * is spread across many cpus.  The actual IO submission is step
426  * two, and the ordered work queue takes care of making sure that
427  * happens in the same order things were put onto the queue by
428  * writepages and friends.
429  *
430  * If this code finds it can't get good compression, it puts an
431  * entry onto the work queue to write the uncompressed bytes.  This
432  * makes sure that both compressed inodes and uncompressed inodes
433  * are written in the same order that the flusher thread sent them
434  * down.
435  */
436 static noinline void compress_file_range(struct inode *inode,
437                                         struct page *locked_page,
438                                         u64 start, u64 end,
439                                         struct async_cow *async_cow,
440                                         int *num_added)
441 {
442         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
443         u64 blocksize = fs_info->sectorsize;
444         u64 actual_end;
445         u64 isize = i_size_read(inode);
446         int ret = 0;
447         struct page **pages = NULL;
448         unsigned long nr_pages;
449         unsigned long total_compressed = 0;
450         unsigned long total_in = 0;
451         int i;
452         int will_compress;
453         int compress_type = fs_info->compress_type;
454         int redirty = 0;
455
456         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
457                         SZ_16K);
458
459         actual_end = min_t(u64, isize, end + 1);
460 again:
461         will_compress = 0;
462         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
463         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
464         nr_pages = min_t(unsigned long, nr_pages,
465                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
466
467         /*
468          * we don't want to send crud past the end of i_size through
469          * compression, that's just a waste of CPU time.  So, if the
470          * end of the file is before the start of our current
471          * requested range of bytes, we bail out to the uncompressed
472          * cleanup code that can deal with all of this.
473          *
474          * It isn't really the fastest way to fix things, but this is a
475          * very uncommon corner.
476          */
477         if (actual_end <= start)
478                 goto cleanup_and_bail_uncompressed;
479
480         total_compressed = actual_end - start;
481
482         /*
483          * skip compression for a small file range(<=blocksize) that
484          * isn't an inline extent, since it doesn't save disk space at all.
485          */
486         if (total_compressed <= blocksize &&
487            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
488                 goto cleanup_and_bail_uncompressed;
489
490         total_compressed = min_t(unsigned long, total_compressed,
491                         BTRFS_MAX_UNCOMPRESSED);
492         total_in = 0;
493         ret = 0;
494
495         /*
496          * we do compression for mount -o compress and when the
497          * inode has not been flagged as nocompress.  This flag can
498          * change at any time if we discover bad compression ratios.
499          */
500         if (inode_need_compress(inode, start, end)) {
501                 WARN_ON(pages);
502                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
503                 if (!pages) {
504                         /* just bail out to the uncompressed code */
505                         goto cont;
506                 }
507
508                 if (BTRFS_I(inode)->defrag_compress)
509                         compress_type = BTRFS_I(inode)->defrag_compress;
510                 else if (BTRFS_I(inode)->prop_compress)
511                         compress_type = BTRFS_I(inode)->prop_compress;
512
513                 /*
514                  * we need to call clear_page_dirty_for_io on each
515                  * page in the range.  Otherwise applications with the file
516                  * mmap'd can wander in and change the page contents while
517                  * we are compressing them.
518                  *
519                  * If the compression fails for any reason, we set the pages
520                  * dirty again later on.
521                  *
522                  * Note that the remaining part is redirtied, the start pointer
523                  * has moved, the end is the original one.
524                  */
525                 if (!redirty) {
526                         extent_range_clear_dirty_for_io(inode, start, end);
527                         redirty = 1;
528                 }
529
530                 /* Compression level is applied here and only here */
531                 ret = btrfs_compress_pages(
532                         compress_type | (fs_info->compress_level << 4),
533                                            inode->i_mapping, start,
534                                            pages,
535                                            &nr_pages,
536                                            &total_in,
537                                            &total_compressed);
538
539                 if (!ret) {
540                         unsigned long offset = total_compressed &
541                                 (PAGE_SIZE - 1);
542                         struct page *page = pages[nr_pages - 1];
543                         char *kaddr;
544
545                         /* zero the tail end of the last page, we might be
546                          * sending it down to disk
547                          */
548                         if (offset) {
549                                 kaddr = kmap_atomic(page);
550                                 memset(kaddr + offset, 0,
551                                        PAGE_SIZE - offset);
552                                 kunmap_atomic(kaddr);
553                         }
554                         will_compress = 1;
555                 }
556         }
557 cont:
558         if (start == 0) {
559                 /* lets try to make an inline extent */
560                 if (ret || total_in < actual_end) {
561                         /* we didn't compress the entire range, try
562                          * to make an uncompressed inline extent.
563                          */
564                         ret = cow_file_range_inline(inode, start, end, 0,
565                                                     BTRFS_COMPRESS_NONE, NULL);
566                 } else {
567                         /* try making a compressed inline extent */
568                         ret = cow_file_range_inline(inode, start, end,
569                                                     total_compressed,
570                                                     compress_type, pages);
571                 }
572                 if (ret <= 0) {
573                         unsigned long clear_flags = EXTENT_DELALLOC |
574                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
575                                 EXTENT_DO_ACCOUNTING;
576                         unsigned long page_error_op;
577
578                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
579
580                         /*
581                          * inline extent creation worked or returned error,
582                          * we don't need to create any more async work items.
583                          * Unlock and free up our temp pages.
584                          *
585                          * We use DO_ACCOUNTING here because we need the
586                          * delalloc_release_metadata to be done _after_ we drop
587                          * our outstanding extent for clearing delalloc for this
588                          * range.
589                          */
590                         extent_clear_unlock_delalloc(inode, start, end, end,
591                                                      NULL, clear_flags,
592                                                      PAGE_UNLOCK |
593                                                      PAGE_CLEAR_DIRTY |
594                                                      PAGE_SET_WRITEBACK |
595                                                      page_error_op |
596                                                      PAGE_END_WRITEBACK);
597                         goto free_pages_out;
598                 }
599         }
600
601         if (will_compress) {
602                 /*
603                  * we aren't doing an inline extent round the compressed size
604                  * up to a block size boundary so the allocator does sane
605                  * things
606                  */
607                 total_compressed = ALIGN(total_compressed, blocksize);
608
609                 /*
610                  * one last check to make sure the compression is really a
611                  * win, compare the page count read with the blocks on disk,
612                  * compression must free at least one sector size
613                  */
614                 total_in = ALIGN(total_in, PAGE_SIZE);
615                 if (total_compressed + blocksize <= total_in) {
616                         *num_added += 1;
617
618                         /*
619                          * The async work queues will take care of doing actual
620                          * allocation on disk for these compressed pages, and
621                          * will submit them to the elevator.
622                          */
623                         add_async_extent(async_cow, start, total_in,
624                                         total_compressed, pages, nr_pages,
625                                         compress_type);
626
627                         if (start + total_in < end) {
628                                 start += total_in;
629                                 pages = NULL;
630                                 cond_resched();
631                                 goto again;
632                         }
633                         return;
634                 }
635         }
636         if (pages) {
637                 /*
638                  * the compression code ran but failed to make things smaller,
639                  * free any pages it allocated and our page pointer array
640                  */
641                 for (i = 0; i < nr_pages; i++) {
642                         WARN_ON(pages[i]->mapping);
643                         put_page(pages[i]);
644                 }
645                 kfree(pages);
646                 pages = NULL;
647                 total_compressed = 0;
648                 nr_pages = 0;
649
650                 /* flag the file so we don't compress in the future */
651                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
652                     !(BTRFS_I(inode)->prop_compress)) {
653                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
654                 }
655         }
656 cleanup_and_bail_uncompressed:
657         /*
658          * No compression, but we still need to write the pages in the file
659          * we've been given so far.  redirty the locked page if it corresponds
660          * to our extent and set things up for the async work queue to run
661          * cow_file_range to do the normal delalloc dance.
662          */
663         if (page_offset(locked_page) >= start &&
664             page_offset(locked_page) <= end)
665                 __set_page_dirty_nobuffers(locked_page);
666                 /* unlocked later on in the async handlers */
667
668         if (redirty)
669                 extent_range_redirty_for_io(inode, start, end);
670         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
671                          BTRFS_COMPRESS_NONE);
672         *num_added += 1;
673
674         return;
675
676 free_pages_out:
677         for (i = 0; i < nr_pages; i++) {
678                 WARN_ON(pages[i]->mapping);
679                 put_page(pages[i]);
680         }
681         kfree(pages);
682 }
683
684 static void free_async_extent_pages(struct async_extent *async_extent)
685 {
686         int i;
687
688         if (!async_extent->pages)
689                 return;
690
691         for (i = 0; i < async_extent->nr_pages; i++) {
692                 WARN_ON(async_extent->pages[i]->mapping);
693                 put_page(async_extent->pages[i]);
694         }
695         kfree(async_extent->pages);
696         async_extent->nr_pages = 0;
697         async_extent->pages = NULL;
698 }
699
700 /*
701  * phase two of compressed writeback.  This is the ordered portion
702  * of the code, which only gets called in the order the work was
703  * queued.  We walk all the async extents created by compress_file_range
704  * and send them down to the disk.
705  */
706 static noinline void submit_compressed_extents(struct inode *inode,
707                                               struct async_cow *async_cow)
708 {
709         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
710         struct async_extent *async_extent;
711         u64 alloc_hint = 0;
712         struct btrfs_key ins;
713         struct extent_map *em;
714         struct btrfs_root *root = BTRFS_I(inode)->root;
715         struct extent_io_tree *io_tree;
716         int ret = 0;
717
718 again:
719         while (!list_empty(&async_cow->extents)) {
720                 async_extent = list_entry(async_cow->extents.next,
721                                           struct async_extent, list);
722                 list_del(&async_extent->list);
723
724                 io_tree = &BTRFS_I(inode)->io_tree;
725
726 retry:
727                 /* did the compression code fall back to uncompressed IO? */
728                 if (!async_extent->pages) {
729                         int page_started = 0;
730                         unsigned long nr_written = 0;
731
732                         lock_extent(io_tree, async_extent->start,
733                                          async_extent->start +
734                                          async_extent->ram_size - 1);
735
736                         /* allocate blocks */
737                         ret = cow_file_range(inode, async_cow->locked_page,
738                                              async_extent->start,
739                                              async_extent->start +
740                                              async_extent->ram_size - 1,
741                                              async_extent->start +
742                                              async_extent->ram_size - 1,
743                                              &page_started, &nr_written, 0,
744                                              NULL);
745
746                         /* JDM XXX */
747
748                         /*
749                          * if page_started, cow_file_range inserted an
750                          * inline extent and took care of all the unlocking
751                          * and IO for us.  Otherwise, we need to submit
752                          * all those pages down to the drive.
753                          */
754                         if (!page_started && !ret)
755                                 extent_write_locked_range(inode,
756                                                   async_extent->start,
757                                                   async_extent->start +
758                                                   async_extent->ram_size - 1,
759                                                   WB_SYNC_ALL);
760                         else if (ret)
761                                 unlock_page(async_cow->locked_page);
762                         kfree(async_extent);
763                         cond_resched();
764                         continue;
765                 }
766
767                 lock_extent(io_tree, async_extent->start,
768                             async_extent->start + async_extent->ram_size - 1);
769
770                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
771                                            async_extent->compressed_size,
772                                            async_extent->compressed_size,
773                                            0, alloc_hint, &ins, 1, 1);
774                 if (ret) {
775                         free_async_extent_pages(async_extent);
776
777                         if (ret == -ENOSPC) {
778                                 unlock_extent(io_tree, async_extent->start,
779                                               async_extent->start +
780                                               async_extent->ram_size - 1);
781
782                                 /*
783                                  * we need to redirty the pages if we decide to
784                                  * fallback to uncompressed IO, otherwise we
785                                  * will not submit these pages down to lower
786                                  * layers.
787                                  */
788                                 extent_range_redirty_for_io(inode,
789                                                 async_extent->start,
790                                                 async_extent->start +
791                                                 async_extent->ram_size - 1);
792
793                                 goto retry;
794                         }
795                         goto out_free;
796                 }
797                 /*
798                  * here we're doing allocation and writeback of the
799                  * compressed pages
800                  */
801                 em = create_io_em(inode, async_extent->start,
802                                   async_extent->ram_size, /* len */
803                                   async_extent->start, /* orig_start */
804                                   ins.objectid, /* block_start */
805                                   ins.offset, /* block_len */
806                                   ins.offset, /* orig_block_len */
807                                   async_extent->ram_size, /* ram_bytes */
808                                   async_extent->compress_type,
809                                   BTRFS_ORDERED_COMPRESSED);
810                 if (IS_ERR(em))
811                         /* ret value is not necessary due to void function */
812                         goto out_free_reserve;
813                 free_extent_map(em);
814
815                 ret = btrfs_add_ordered_extent_compress(inode,
816                                                 async_extent->start,
817                                                 ins.objectid,
818                                                 async_extent->ram_size,
819                                                 ins.offset,
820                                                 BTRFS_ORDERED_COMPRESSED,
821                                                 async_extent->compress_type);
822                 if (ret) {
823                         btrfs_drop_extent_cache(BTRFS_I(inode),
824                                                 async_extent->start,
825                                                 async_extent->start +
826                                                 async_extent->ram_size - 1, 0);
827                         goto out_free_reserve;
828                 }
829                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
830
831                 /*
832                  * clear dirty, set writeback and unlock the pages.
833                  */
834                 extent_clear_unlock_delalloc(inode, async_extent->start,
835                                 async_extent->start +
836                                 async_extent->ram_size - 1,
837                                 async_extent->start +
838                                 async_extent->ram_size - 1,
839                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
840                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
841                                 PAGE_SET_WRITEBACK);
842                 if (btrfs_submit_compressed_write(inode,
843                                     async_extent->start,
844                                     async_extent->ram_size,
845                                     ins.objectid,
846                                     ins.offset, async_extent->pages,
847                                     async_extent->nr_pages,
848                                     async_cow->write_flags)) {
849                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
850                         struct page *p = async_extent->pages[0];
851                         const u64 start = async_extent->start;
852                         const u64 end = start + async_extent->ram_size - 1;
853
854                         p->mapping = inode->i_mapping;
855                         tree->ops->writepage_end_io_hook(p, start, end,
856                                                          NULL, 0);
857                         p->mapping = NULL;
858                         extent_clear_unlock_delalloc(inode, start, end, end,
859                                                      NULL, 0,
860                                                      PAGE_END_WRITEBACK |
861                                                      PAGE_SET_ERROR);
862                         free_async_extent_pages(async_extent);
863                 }
864                 alloc_hint = ins.objectid + ins.offset;
865                 kfree(async_extent);
866                 cond_resched();
867         }
868         return;
869 out_free_reserve:
870         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
871         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
872 out_free:
873         extent_clear_unlock_delalloc(inode, async_extent->start,
874                                      async_extent->start +
875                                      async_extent->ram_size - 1,
876                                      async_extent->start +
877                                      async_extent->ram_size - 1,
878                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
879                                      EXTENT_DELALLOC_NEW |
880                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
881                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
882                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
883                                      PAGE_SET_ERROR);
884         free_async_extent_pages(async_extent);
885         kfree(async_extent);
886         goto again;
887 }
888
889 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
890                                       u64 num_bytes)
891 {
892         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
893         struct extent_map *em;
894         u64 alloc_hint = 0;
895
896         read_lock(&em_tree->lock);
897         em = search_extent_mapping(em_tree, start, num_bytes);
898         if (em) {
899                 /*
900                  * if block start isn't an actual block number then find the
901                  * first block in this inode and use that as a hint.  If that
902                  * block is also bogus then just don't worry about it.
903                  */
904                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
905                         free_extent_map(em);
906                         em = search_extent_mapping(em_tree, 0, 0);
907                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
908                                 alloc_hint = em->block_start;
909                         if (em)
910                                 free_extent_map(em);
911                 } else {
912                         alloc_hint = em->block_start;
913                         free_extent_map(em);
914                 }
915         }
916         read_unlock(&em_tree->lock);
917
918         return alloc_hint;
919 }
920
921 /*
922  * when extent_io.c finds a delayed allocation range in the file,
923  * the call backs end up in this code.  The basic idea is to
924  * allocate extents on disk for the range, and create ordered data structs
925  * in ram to track those extents.
926  *
927  * locked_page is the page that writepage had locked already.  We use
928  * it to make sure we don't do extra locks or unlocks.
929  *
930  * *page_started is set to one if we unlock locked_page and do everything
931  * required to start IO on it.  It may be clean and already done with
932  * IO when we return.
933  */
934 static noinline int cow_file_range(struct inode *inode,
935                                    struct page *locked_page,
936                                    u64 start, u64 end, u64 delalloc_end,
937                                    int *page_started, unsigned long *nr_written,
938                                    int unlock, struct btrfs_dedupe_hash *hash)
939 {
940         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
941         struct btrfs_root *root = BTRFS_I(inode)->root;
942         u64 alloc_hint = 0;
943         u64 num_bytes;
944         unsigned long ram_size;
945         u64 cur_alloc_size = 0;
946         u64 blocksize = fs_info->sectorsize;
947         struct btrfs_key ins;
948         struct extent_map *em;
949         unsigned clear_bits;
950         unsigned long page_ops;
951         bool extent_reserved = false;
952         int ret = 0;
953
954         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
955                 WARN_ON_ONCE(1);
956                 ret = -EINVAL;
957                 goto out_unlock;
958         }
959
960         num_bytes = ALIGN(end - start + 1, blocksize);
961         num_bytes = max(blocksize,  num_bytes);
962         ASSERT(num_bytes <= btrfs_super_total_bytes(fs_info->super_copy));
963
964         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
965
966         if (start == 0) {
967                 /* lets try to make an inline extent */
968                 ret = cow_file_range_inline(inode, start, end, 0,
969                                             BTRFS_COMPRESS_NONE, NULL);
970                 if (ret == 0) {
971                         /*
972                          * We use DO_ACCOUNTING here because we need the
973                          * delalloc_release_metadata to be run _after_ we drop
974                          * our outstanding extent for clearing delalloc for this
975                          * range.
976                          */
977                         extent_clear_unlock_delalloc(inode, start, end,
978                                      delalloc_end, NULL,
979                                      EXTENT_LOCKED | EXTENT_DELALLOC |
980                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
981                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
982                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
983                                      PAGE_END_WRITEBACK);
984                         *nr_written = *nr_written +
985                              (end - start + PAGE_SIZE) / PAGE_SIZE;
986                         *page_started = 1;
987                         goto out;
988                 } else if (ret < 0) {
989                         goto out_unlock;
990                 }
991         }
992
993         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
994         btrfs_drop_extent_cache(BTRFS_I(inode), start,
995                         start + num_bytes - 1, 0);
996
997         while (num_bytes > 0) {
998                 cur_alloc_size = num_bytes;
999                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1000                                            fs_info->sectorsize, 0, alloc_hint,
1001                                            &ins, 1, 1);
1002                 if (ret < 0)
1003                         goto out_unlock;
1004                 cur_alloc_size = ins.offset;
1005                 extent_reserved = true;
1006
1007                 ram_size = ins.offset;
1008                 em = create_io_em(inode, start, ins.offset, /* len */
1009                                   start, /* orig_start */
1010                                   ins.objectid, /* block_start */
1011                                   ins.offset, /* block_len */
1012                                   ins.offset, /* orig_block_len */
1013                                   ram_size, /* ram_bytes */
1014                                   BTRFS_COMPRESS_NONE, /* compress_type */
1015                                   BTRFS_ORDERED_REGULAR /* type */);
1016                 if (IS_ERR(em)) {
1017                         ret = PTR_ERR(em);
1018                         goto out_reserve;
1019                 }
1020                 free_extent_map(em);
1021
1022                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1023                                                ram_size, cur_alloc_size, 0);
1024                 if (ret)
1025                         goto out_drop_extent_cache;
1026
1027                 if (root->root_key.objectid ==
1028                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1029                         ret = btrfs_reloc_clone_csums(inode, start,
1030                                                       cur_alloc_size);
1031                         /*
1032                          * Only drop cache here, and process as normal.
1033                          *
1034                          * We must not allow extent_clear_unlock_delalloc()
1035                          * at out_unlock label to free meta of this ordered
1036                          * extent, as its meta should be freed by
1037                          * btrfs_finish_ordered_io().
1038                          *
1039                          * So we must continue until @start is increased to
1040                          * skip current ordered extent.
1041                          */
1042                         if (ret)
1043                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1044                                                 start + ram_size - 1, 0);
1045                 }
1046
1047                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1048
1049                 /* we're not doing compressed IO, don't unlock the first
1050                  * page (which the caller expects to stay locked), don't
1051                  * clear any dirty bits and don't set any writeback bits
1052                  *
1053                  * Do set the Private2 bit so we know this page was properly
1054                  * setup for writepage
1055                  */
1056                 page_ops = unlock ? PAGE_UNLOCK : 0;
1057                 page_ops |= PAGE_SET_PRIVATE2;
1058
1059                 extent_clear_unlock_delalloc(inode, start,
1060                                              start + ram_size - 1,
1061                                              delalloc_end, locked_page,
1062                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1063                                              page_ops);
1064                 if (num_bytes < cur_alloc_size)
1065                         num_bytes = 0;
1066                 else
1067                         num_bytes -= cur_alloc_size;
1068                 alloc_hint = ins.objectid + ins.offset;
1069                 start += cur_alloc_size;
1070                 extent_reserved = false;
1071
1072                 /*
1073                  * btrfs_reloc_clone_csums() error, since start is increased
1074                  * extent_clear_unlock_delalloc() at out_unlock label won't
1075                  * free metadata of current ordered extent, we're OK to exit.
1076                  */
1077                 if (ret)
1078                         goto out_unlock;
1079         }
1080 out:
1081         return ret;
1082
1083 out_drop_extent_cache:
1084         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1085 out_reserve:
1086         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1087         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1088 out_unlock:
1089         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1090                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1091         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1092                 PAGE_END_WRITEBACK;
1093         /*
1094          * If we reserved an extent for our delalloc range (or a subrange) and
1095          * failed to create the respective ordered extent, then it means that
1096          * when we reserved the extent we decremented the extent's size from
1097          * the data space_info's bytes_may_use counter and incremented the
1098          * space_info's bytes_reserved counter by the same amount. We must make
1099          * sure extent_clear_unlock_delalloc() does not try to decrement again
1100          * the data space_info's bytes_may_use counter, therefore we do not pass
1101          * it the flag EXTENT_CLEAR_DATA_RESV.
1102          */
1103         if (extent_reserved) {
1104                 extent_clear_unlock_delalloc(inode, start,
1105                                              start + cur_alloc_size,
1106                                              start + cur_alloc_size,
1107                                              locked_page,
1108                                              clear_bits,
1109                                              page_ops);
1110                 start += cur_alloc_size;
1111                 if (start >= end)
1112                         goto out;
1113         }
1114         extent_clear_unlock_delalloc(inode, start, end, delalloc_end,
1115                                      locked_page,
1116                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1117                                      page_ops);
1118         goto out;
1119 }
1120
1121 /*
1122  * work queue call back to started compression on a file and pages
1123  */
1124 static noinline void async_cow_start(struct btrfs_work *work)
1125 {
1126         struct async_cow *async_cow;
1127         int num_added = 0;
1128         async_cow = container_of(work, struct async_cow, work);
1129
1130         compress_file_range(async_cow->inode, async_cow->locked_page,
1131                             async_cow->start, async_cow->end, async_cow,
1132                             &num_added);
1133         if (num_added == 0) {
1134                 btrfs_add_delayed_iput(async_cow->inode);
1135                 async_cow->inode = NULL;
1136         }
1137 }
1138
1139 /*
1140  * work queue call back to submit previously compressed pages
1141  */
1142 static noinline void async_cow_submit(struct btrfs_work *work)
1143 {
1144         struct btrfs_fs_info *fs_info;
1145         struct async_cow *async_cow;
1146         struct btrfs_root *root;
1147         unsigned long nr_pages;
1148
1149         async_cow = container_of(work, struct async_cow, work);
1150
1151         root = async_cow->root;
1152         fs_info = root->fs_info;
1153         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1154                 PAGE_SHIFT;
1155
1156         /* atomic_sub_return implies a barrier */
1157         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1158             5 * SZ_1M)
1159                 cond_wake_up_nomb(&fs_info->async_submit_wait);
1160
1161         if (async_cow->inode)
1162                 submit_compressed_extents(async_cow->inode, async_cow);
1163 }
1164
1165 static noinline void async_cow_free(struct btrfs_work *work)
1166 {
1167         struct async_cow *async_cow;
1168         async_cow = container_of(work, struct async_cow, work);
1169         if (async_cow->inode)
1170                 btrfs_add_delayed_iput(async_cow->inode);
1171         kfree(async_cow);
1172 }
1173
1174 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1175                                 u64 start, u64 end, int *page_started,
1176                                 unsigned long *nr_written,
1177                                 unsigned int write_flags)
1178 {
1179         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1180         struct async_cow *async_cow;
1181         struct btrfs_root *root = BTRFS_I(inode)->root;
1182         unsigned long nr_pages;
1183         u64 cur_end;
1184
1185         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1186                          1, 0, NULL);
1187         while (start < end) {
1188                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1189                 BUG_ON(!async_cow); /* -ENOMEM */
1190                 async_cow->inode = igrab(inode);
1191                 async_cow->root = root;
1192                 async_cow->locked_page = locked_page;
1193                 async_cow->start = start;
1194                 async_cow->write_flags = write_flags;
1195
1196                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1197                     !btrfs_test_opt(fs_info, FORCE_COMPRESS))
1198                         cur_end = end;
1199                 else
1200                         cur_end = min(end, start + SZ_512K - 1);
1201
1202                 async_cow->end = cur_end;
1203                 INIT_LIST_HEAD(&async_cow->extents);
1204
1205                 btrfs_init_work(&async_cow->work,
1206                                 btrfs_delalloc_helper,
1207                                 async_cow_start, async_cow_submit,
1208                                 async_cow_free);
1209
1210                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1211                         PAGE_SHIFT;
1212                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1213
1214                 btrfs_queue_work(fs_info->delalloc_workers, &async_cow->work);
1215
1216                 *nr_written += nr_pages;
1217                 start = cur_end + 1;
1218         }
1219         *page_started = 1;
1220         return 0;
1221 }
1222
1223 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1224                                         u64 bytenr, u64 num_bytes)
1225 {
1226         int ret;
1227         struct btrfs_ordered_sum *sums;
1228         LIST_HEAD(list);
1229
1230         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1231                                        bytenr + num_bytes - 1, &list, 0);
1232         if (ret == 0 && list_empty(&list))
1233                 return 0;
1234
1235         while (!list_empty(&list)) {
1236                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1237                 list_del(&sums->list);
1238                 kfree(sums);
1239         }
1240         if (ret < 0)
1241                 return ret;
1242         return 1;
1243 }
1244
1245 /*
1246  * when nowcow writeback call back.  This checks for snapshots or COW copies
1247  * of the extents that exist in the file, and COWs the file as required.
1248  *
1249  * If no cow copies or snapshots exist, we write directly to the existing
1250  * blocks on disk
1251  */
1252 static noinline int run_delalloc_nocow(struct inode *inode,
1253                                        struct page *locked_page,
1254                               u64 start, u64 end, int *page_started, int force,
1255                               unsigned long *nr_written)
1256 {
1257         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1258         struct btrfs_root *root = BTRFS_I(inode)->root;
1259         struct extent_buffer *leaf;
1260         struct btrfs_path *path;
1261         struct btrfs_file_extent_item *fi;
1262         struct btrfs_key found_key;
1263         struct extent_map *em;
1264         u64 cow_start;
1265         u64 cur_offset;
1266         u64 extent_end;
1267         u64 extent_offset;
1268         u64 disk_bytenr;
1269         u64 num_bytes;
1270         u64 disk_num_bytes;
1271         u64 ram_bytes;
1272         int extent_type;
1273         int ret;
1274         int type;
1275         int nocow;
1276         int check_prev = 1;
1277         bool nolock;
1278         u64 ino = btrfs_ino(BTRFS_I(inode));
1279
1280         path = btrfs_alloc_path();
1281         if (!path) {
1282                 extent_clear_unlock_delalloc(inode, start, end, end,
1283                                              locked_page,
1284                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1285                                              EXTENT_DO_ACCOUNTING |
1286                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1287                                              PAGE_CLEAR_DIRTY |
1288                                              PAGE_SET_WRITEBACK |
1289                                              PAGE_END_WRITEBACK);
1290                 return -ENOMEM;
1291         }
1292
1293         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
1294
1295         cow_start = (u64)-1;
1296         cur_offset = start;
1297         while (1) {
1298                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1299                                                cur_offset, 0);
1300                 if (ret < 0)
1301                         goto error;
1302                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1303                         leaf = path->nodes[0];
1304                         btrfs_item_key_to_cpu(leaf, &found_key,
1305                                               path->slots[0] - 1);
1306                         if (found_key.objectid == ino &&
1307                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1308                                 path->slots[0]--;
1309                 }
1310                 check_prev = 0;
1311 next_slot:
1312                 leaf = path->nodes[0];
1313                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1314                         ret = btrfs_next_leaf(root, path);
1315                         if (ret < 0) {
1316                                 if (cow_start != (u64)-1)
1317                                         cur_offset = cow_start;
1318                                 goto error;
1319                         }
1320                         if (ret > 0)
1321                                 break;
1322                         leaf = path->nodes[0];
1323                 }
1324
1325                 nocow = 0;
1326                 disk_bytenr = 0;
1327                 num_bytes = 0;
1328                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1329
1330                 if (found_key.objectid > ino)
1331                         break;
1332                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1333                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1334                         path->slots[0]++;
1335                         goto next_slot;
1336                 }
1337                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1338                     found_key.offset > end)
1339                         break;
1340
1341                 if (found_key.offset > cur_offset) {
1342                         extent_end = found_key.offset;
1343                         extent_type = 0;
1344                         goto out_check;
1345                 }
1346
1347                 fi = btrfs_item_ptr(leaf, path->slots[0],
1348                                     struct btrfs_file_extent_item);
1349                 extent_type = btrfs_file_extent_type(leaf, fi);
1350
1351                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1352                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1353                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1354                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1355                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1356                         extent_end = found_key.offset +
1357                                 btrfs_file_extent_num_bytes(leaf, fi);
1358                         disk_num_bytes =
1359                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1360                         if (extent_end <= start) {
1361                                 path->slots[0]++;
1362                                 goto next_slot;
1363                         }
1364                         if (disk_bytenr == 0)
1365                                 goto out_check;
1366                         if (btrfs_file_extent_compression(leaf, fi) ||
1367                             btrfs_file_extent_encryption(leaf, fi) ||
1368                             btrfs_file_extent_other_encoding(leaf, fi))
1369                                 goto out_check;
1370                         /*
1371                          * Do the same check as in btrfs_cross_ref_exist but
1372                          * without the unnecessary search.
1373                          */
1374                         if (btrfs_file_extent_generation(leaf, fi) <=
1375                             btrfs_root_last_snapshot(&root->root_item))
1376                                 goto out_check;
1377                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1378                                 goto out_check;
1379                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1380                                 goto out_check;
1381                         ret = btrfs_cross_ref_exist(root, ino,
1382                                                     found_key.offset -
1383                                                     extent_offset, disk_bytenr);
1384                         if (ret) {
1385                                 /*
1386                                  * ret could be -EIO if the above fails to read
1387                                  * metadata.
1388                                  */
1389                                 if (ret < 0) {
1390                                         if (cow_start != (u64)-1)
1391                                                 cur_offset = cow_start;
1392                                         goto error;
1393                                 }
1394
1395                                 WARN_ON_ONCE(nolock);
1396                                 goto out_check;
1397                         }
1398                         disk_bytenr += extent_offset;
1399                         disk_bytenr += cur_offset - found_key.offset;
1400                         num_bytes = min(end + 1, extent_end) - cur_offset;
1401                         /*
1402                          * if there are pending snapshots for this root,
1403                          * we fall into common COW way.
1404                          */
1405                         if (!nolock && atomic_read(&root->snapshot_force_cow))
1406                                 goto out_check;
1407                         /*
1408                          * force cow if csum exists in the range.
1409                          * this ensure that csum for a given extent are
1410                          * either valid or do not exist.
1411                          */
1412                         ret = csum_exist_in_range(fs_info, disk_bytenr,
1413                                                   num_bytes);
1414                         if (ret) {
1415                                 /*
1416                                  * ret could be -EIO if the above fails to read
1417                                  * metadata.
1418                                  */
1419                                 if (ret < 0) {
1420                                         if (cow_start != (u64)-1)
1421                                                 cur_offset = cow_start;
1422                                         goto error;
1423                                 }
1424                                 WARN_ON_ONCE(nolock);
1425                                 goto out_check;
1426                         }
1427                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr))
1428                                 goto out_check;
1429                         nocow = 1;
1430                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1431                         extent_end = found_key.offset +
1432                                 btrfs_file_extent_ram_bytes(leaf, fi);
1433                         extent_end = ALIGN(extent_end,
1434                                            fs_info->sectorsize);
1435                 } else {
1436                         BUG_ON(1);
1437                 }
1438 out_check:
1439                 if (extent_end <= start) {
1440                         path->slots[0]++;
1441                         if (nocow)
1442                                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1443                         goto next_slot;
1444                 }
1445                 if (!nocow) {
1446                         if (cow_start == (u64)-1)
1447                                 cow_start = cur_offset;
1448                         cur_offset = extent_end;
1449                         if (cur_offset > end)
1450                                 break;
1451                         path->slots[0]++;
1452                         goto next_slot;
1453                 }
1454
1455                 btrfs_release_path(path);
1456                 if (cow_start != (u64)-1) {
1457                         ret = cow_file_range(inode, locked_page,
1458                                              cow_start, found_key.offset - 1,
1459                                              end, page_started, nr_written, 1,
1460                                              NULL);
1461                         if (ret) {
1462                                 if (nocow)
1463                                         btrfs_dec_nocow_writers(fs_info,
1464                                                                 disk_bytenr);
1465                                 goto error;
1466                         }
1467                         cow_start = (u64)-1;
1468                 }
1469
1470                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1471                         u64 orig_start = found_key.offset - extent_offset;
1472
1473                         em = create_io_em(inode, cur_offset, num_bytes,
1474                                           orig_start,
1475                                           disk_bytenr, /* block_start */
1476                                           num_bytes, /* block_len */
1477                                           disk_num_bytes, /* orig_block_len */
1478                                           ram_bytes, BTRFS_COMPRESS_NONE,
1479                                           BTRFS_ORDERED_PREALLOC);
1480                         if (IS_ERR(em)) {
1481                                 if (nocow)
1482                                         btrfs_dec_nocow_writers(fs_info,
1483                                                                 disk_bytenr);
1484                                 ret = PTR_ERR(em);
1485                                 goto error;
1486                         }
1487                         free_extent_map(em);
1488                 }
1489
1490                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1491                         type = BTRFS_ORDERED_PREALLOC;
1492                 } else {
1493                         type = BTRFS_ORDERED_NOCOW;
1494                 }
1495
1496                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1497                                                num_bytes, num_bytes, type);
1498                 if (nocow)
1499                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1500                 BUG_ON(ret); /* -ENOMEM */
1501
1502                 if (root->root_key.objectid ==
1503                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1504                         /*
1505                          * Error handled later, as we must prevent
1506                          * extent_clear_unlock_delalloc() in error handler
1507                          * from freeing metadata of created ordered extent.
1508                          */
1509                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1510                                                       num_bytes);
1511
1512                 extent_clear_unlock_delalloc(inode, cur_offset,
1513                                              cur_offset + num_bytes - 1, end,
1514                                              locked_page, EXTENT_LOCKED |
1515                                              EXTENT_DELALLOC |
1516                                              EXTENT_CLEAR_DATA_RESV,
1517                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1518
1519                 cur_offset = extent_end;
1520
1521                 /*
1522                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1523                  * handler, as metadata for created ordered extent will only
1524                  * be freed by btrfs_finish_ordered_io().
1525                  */
1526                 if (ret)
1527                         goto error;
1528                 if (cur_offset > end)
1529                         break;
1530         }
1531         btrfs_release_path(path);
1532
1533         if (cur_offset <= end && cow_start == (u64)-1) {
1534                 cow_start = cur_offset;
1535                 cur_offset = end;
1536         }
1537
1538         if (cow_start != (u64)-1) {
1539                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1540                                      page_started, nr_written, 1, NULL);
1541                 if (ret)
1542                         goto error;
1543         }
1544
1545 error:
1546         if (ret && cur_offset < end)
1547                 extent_clear_unlock_delalloc(inode, cur_offset, end, end,
1548                                              locked_page, EXTENT_LOCKED |
1549                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1550                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1551                                              PAGE_CLEAR_DIRTY |
1552                                              PAGE_SET_WRITEBACK |
1553                                              PAGE_END_WRITEBACK);
1554         btrfs_free_path(path);
1555         return ret;
1556 }
1557
1558 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1559 {
1560
1561         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1562             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1563                 return 0;
1564
1565         /*
1566          * @defrag_bytes is a hint value, no spinlock held here,
1567          * if is not zero, it means the file is defragging.
1568          * Force cow if given extent needs to be defragged.
1569          */
1570         if (BTRFS_I(inode)->defrag_bytes &&
1571             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1572                            EXTENT_DEFRAG, 0, NULL))
1573                 return 1;
1574
1575         return 0;
1576 }
1577
1578 /*
1579  * extent_io.c call back to do delayed allocation processing
1580  */
1581 static int run_delalloc_range(void *private_data, struct page *locked_page,
1582                               u64 start, u64 end, int *page_started,
1583                               unsigned long *nr_written,
1584                               struct writeback_control *wbc)
1585 {
1586         struct inode *inode = private_data;
1587         int ret;
1588         int force_cow = need_force_cow(inode, start, end);
1589         unsigned int write_flags = wbc_to_write_flags(wbc);
1590
1591         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1592                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1593                                          page_started, 1, nr_written);
1594         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1595                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1596                                          page_started, 0, nr_written);
1597         } else if (!inode_need_compress(inode, start, end)) {
1598                 ret = cow_file_range(inode, locked_page, start, end, end,
1599                                       page_started, nr_written, 1, NULL);
1600         } else {
1601                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1602                         &BTRFS_I(inode)->runtime_flags);
1603                 ret = cow_file_range_async(inode, locked_page, start, end,
1604                                            page_started, nr_written,
1605                                            write_flags);
1606         }
1607         if (ret)
1608                 btrfs_cleanup_ordered_extents(inode, start, end - start + 1);
1609         return ret;
1610 }
1611
1612 static void btrfs_split_extent_hook(void *private_data,
1613                                     struct extent_state *orig, u64 split)
1614 {
1615         struct inode *inode = private_data;
1616         u64 size;
1617
1618         /* not delalloc, ignore it */
1619         if (!(orig->state & EXTENT_DELALLOC))
1620                 return;
1621
1622         size = orig->end - orig->start + 1;
1623         if (size > BTRFS_MAX_EXTENT_SIZE) {
1624                 u32 num_extents;
1625                 u64 new_size;
1626
1627                 /*
1628                  * See the explanation in btrfs_merge_extent_hook, the same
1629                  * applies here, just in reverse.
1630                  */
1631                 new_size = orig->end - split + 1;
1632                 num_extents = count_max_extents(new_size);
1633                 new_size = split - orig->start;
1634                 num_extents += count_max_extents(new_size);
1635                 if (count_max_extents(size) >= num_extents)
1636                         return;
1637         }
1638
1639         spin_lock(&BTRFS_I(inode)->lock);
1640         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1641         spin_unlock(&BTRFS_I(inode)->lock);
1642 }
1643
1644 /*
1645  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1646  * extents so we can keep track of new extents that are just merged onto old
1647  * extents, such as when we are doing sequential writes, so we can properly
1648  * account for the metadata space we'll need.
1649  */
1650 static void btrfs_merge_extent_hook(void *private_data,
1651                                     struct extent_state *new,
1652                                     struct extent_state *other)
1653 {
1654         struct inode *inode = private_data;
1655         u64 new_size, old_size;
1656         u32 num_extents;
1657
1658         /* not delalloc, ignore it */
1659         if (!(other->state & EXTENT_DELALLOC))
1660                 return;
1661
1662         if (new->start > other->start)
1663                 new_size = new->end - other->start + 1;
1664         else
1665                 new_size = other->end - new->start + 1;
1666
1667         /* we're not bigger than the max, unreserve the space and go */
1668         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1669                 spin_lock(&BTRFS_I(inode)->lock);
1670                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1671                 spin_unlock(&BTRFS_I(inode)->lock);
1672                 return;
1673         }
1674
1675         /*
1676          * We have to add up either side to figure out how many extents were
1677          * accounted for before we merged into one big extent.  If the number of
1678          * extents we accounted for is <= the amount we need for the new range
1679          * then we can return, otherwise drop.  Think of it like this
1680          *
1681          * [ 4k][MAX_SIZE]
1682          *
1683          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1684          * need 2 outstanding extents, on one side we have 1 and the other side
1685          * we have 1 so they are == and we can return.  But in this case
1686          *
1687          * [MAX_SIZE+4k][MAX_SIZE+4k]
1688          *
1689          * Each range on their own accounts for 2 extents, but merged together
1690          * they are only 3 extents worth of accounting, so we need to drop in
1691          * this case.
1692          */
1693         old_size = other->end - other->start + 1;
1694         num_extents = count_max_extents(old_size);
1695         old_size = new->end - new->start + 1;
1696         num_extents += count_max_extents(old_size);
1697         if (count_max_extents(new_size) >= num_extents)
1698                 return;
1699
1700         spin_lock(&BTRFS_I(inode)->lock);
1701         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1702         spin_unlock(&BTRFS_I(inode)->lock);
1703 }
1704
1705 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1706                                       struct inode *inode)
1707 {
1708         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1709
1710         spin_lock(&root->delalloc_lock);
1711         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1712                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1713                               &root->delalloc_inodes);
1714                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1715                         &BTRFS_I(inode)->runtime_flags);
1716                 root->nr_delalloc_inodes++;
1717                 if (root->nr_delalloc_inodes == 1) {
1718                         spin_lock(&fs_info->delalloc_root_lock);
1719                         BUG_ON(!list_empty(&root->delalloc_root));
1720                         list_add_tail(&root->delalloc_root,
1721                                       &fs_info->delalloc_roots);
1722                         spin_unlock(&fs_info->delalloc_root_lock);
1723                 }
1724         }
1725         spin_unlock(&root->delalloc_lock);
1726 }
1727
1728
1729 void __btrfs_del_delalloc_inode(struct btrfs_root *root,
1730                                 struct btrfs_inode *inode)
1731 {
1732         struct btrfs_fs_info *fs_info = root->fs_info;
1733
1734         if (!list_empty(&inode->delalloc_inodes)) {
1735                 list_del_init(&inode->delalloc_inodes);
1736                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1737                           &inode->runtime_flags);
1738                 root->nr_delalloc_inodes--;
1739                 if (!root->nr_delalloc_inodes) {
1740                         ASSERT(list_empty(&root->delalloc_inodes));
1741                         spin_lock(&fs_info->delalloc_root_lock);
1742                         BUG_ON(list_empty(&root->delalloc_root));
1743                         list_del_init(&root->delalloc_root);
1744                         spin_unlock(&fs_info->delalloc_root_lock);
1745                 }
1746         }
1747 }
1748
1749 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1750                                      struct btrfs_inode *inode)
1751 {
1752         spin_lock(&root->delalloc_lock);
1753         __btrfs_del_delalloc_inode(root, inode);
1754         spin_unlock(&root->delalloc_lock);
1755 }
1756
1757 /*
1758  * extent_io.c set_bit_hook, used to track delayed allocation
1759  * bytes in this file, and to maintain the list of inodes that
1760  * have pending delalloc work to be done.
1761  */
1762 static void btrfs_set_bit_hook(void *private_data,
1763                                struct extent_state *state, unsigned *bits)
1764 {
1765         struct inode *inode = private_data;
1766
1767         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1768
1769         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1770                 WARN_ON(1);
1771         /*
1772          * set_bit and clear bit hooks normally require _irqsave/restore
1773          * but in this case, we are only testing for the DELALLOC
1774          * bit, which is only set or cleared with irqs on
1775          */
1776         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1777                 struct btrfs_root *root = BTRFS_I(inode)->root;
1778                 u64 len = state->end + 1 - state->start;
1779                 u32 num_extents = count_max_extents(len);
1780                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1781
1782                 spin_lock(&BTRFS_I(inode)->lock);
1783                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1784                 spin_unlock(&BTRFS_I(inode)->lock);
1785
1786                 /* For sanity tests */
1787                 if (btrfs_is_testing(fs_info))
1788                         return;
1789
1790                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1791                                          fs_info->delalloc_batch);
1792                 spin_lock(&BTRFS_I(inode)->lock);
1793                 BTRFS_I(inode)->delalloc_bytes += len;
1794                 if (*bits & EXTENT_DEFRAG)
1795                         BTRFS_I(inode)->defrag_bytes += len;
1796                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1797                                          &BTRFS_I(inode)->runtime_flags))
1798                         btrfs_add_delalloc_inodes(root, inode);
1799                 spin_unlock(&BTRFS_I(inode)->lock);
1800         }
1801
1802         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1803             (*bits & EXTENT_DELALLOC_NEW)) {
1804                 spin_lock(&BTRFS_I(inode)->lock);
1805                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1806                         state->start;
1807                 spin_unlock(&BTRFS_I(inode)->lock);
1808         }
1809 }
1810
1811 /*
1812  * extent_io.c clear_bit_hook, see set_bit_hook for why
1813  */
1814 static void btrfs_clear_bit_hook(void *private_data,
1815                                  struct extent_state *state,
1816                                  unsigned *bits)
1817 {
1818         struct btrfs_inode *inode = BTRFS_I((struct inode *)private_data);
1819         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1820         u64 len = state->end + 1 - state->start;
1821         u32 num_extents = count_max_extents(len);
1822
1823         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1824                 spin_lock(&inode->lock);
1825                 inode->defrag_bytes -= len;
1826                 spin_unlock(&inode->lock);
1827         }
1828
1829         /*
1830          * set_bit and clear bit hooks normally require _irqsave/restore
1831          * but in this case, we are only testing for the DELALLOC
1832          * bit, which is only set or cleared with irqs on
1833          */
1834         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1835                 struct btrfs_root *root = inode->root;
1836                 bool do_list = !btrfs_is_free_space_inode(inode);
1837
1838                 spin_lock(&inode->lock);
1839                 btrfs_mod_outstanding_extents(inode, -num_extents);
1840                 spin_unlock(&inode->lock);
1841
1842                 /*
1843                  * We don't reserve metadata space for space cache inodes so we
1844                  * don't need to call dellalloc_release_metadata if there is an
1845                  * error.
1846                  */
1847                 if (*bits & EXTENT_CLEAR_META_RESV &&
1848                     root != fs_info->tree_root)
1849                         btrfs_delalloc_release_metadata(inode, len, false);
1850
1851                 /* For sanity tests. */
1852                 if (btrfs_is_testing(fs_info))
1853                         return;
1854
1855                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
1856                     do_list && !(state->state & EXTENT_NORESERVE) &&
1857                     (*bits & EXTENT_CLEAR_DATA_RESV))
1858                         btrfs_free_reserved_data_space_noquota(
1859                                         &inode->vfs_inode,
1860                                         state->start, len);
1861
1862                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
1863                                          fs_info->delalloc_batch);
1864                 spin_lock(&inode->lock);
1865                 inode->delalloc_bytes -= len;
1866                 if (do_list && inode->delalloc_bytes == 0 &&
1867                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1868                                         &inode->runtime_flags))
1869                         btrfs_del_delalloc_inode(root, inode);
1870                 spin_unlock(&inode->lock);
1871         }
1872
1873         if ((state->state & EXTENT_DELALLOC_NEW) &&
1874             (*bits & EXTENT_DELALLOC_NEW)) {
1875                 spin_lock(&inode->lock);
1876                 ASSERT(inode->new_delalloc_bytes >= len);
1877                 inode->new_delalloc_bytes -= len;
1878                 spin_unlock(&inode->lock);
1879         }
1880 }
1881
1882 /*
1883  * Merge bio hook, this must check the chunk tree to make sure we don't create
1884  * bios that span stripes or chunks
1885  *
1886  * return 1 if page cannot be merged to bio
1887  * return 0 if page can be merged to bio
1888  * return error otherwise
1889  */
1890 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1891                          size_t size, struct bio *bio,
1892                          unsigned long bio_flags)
1893 {
1894         struct inode *inode = page->mapping->host;
1895         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1896         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1897         u64 length = 0;
1898         u64 map_length;
1899         int ret;
1900
1901         if (bio_flags & EXTENT_BIO_COMPRESSED)
1902                 return 0;
1903
1904         length = bio->bi_iter.bi_size;
1905         map_length = length;
1906         ret = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
1907                               NULL, 0);
1908         if (ret < 0)
1909                 return ret;
1910         if (map_length < length + size)
1911                 return 1;
1912         return 0;
1913 }
1914
1915 /*
1916  * in order to insert checksums into the metadata in large chunks,
1917  * we wait until bio submission time.   All the pages in the bio are
1918  * checksummed and sums are attached onto the ordered extent record.
1919  *
1920  * At IO completion time the cums attached on the ordered extent record
1921  * are inserted into the btree
1922  */
1923 static blk_status_t btrfs_submit_bio_start(void *private_data, struct bio *bio,
1924                                     u64 bio_offset)
1925 {
1926         struct inode *inode = private_data;
1927         blk_status_t ret = 0;
1928
1929         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1930         BUG_ON(ret); /* -ENOMEM */
1931         return 0;
1932 }
1933
1934 /*
1935  * in order to insert checksums into the metadata in large chunks,
1936  * we wait until bio submission time.   All the pages in the bio are
1937  * checksummed and sums are attached onto the ordered extent record.
1938  *
1939  * At IO completion time the cums attached on the ordered extent record
1940  * are inserted into the btree
1941  */
1942 blk_status_t btrfs_submit_bio_done(void *private_data, struct bio *bio,
1943                           int mirror_num)
1944 {
1945         struct inode *inode = private_data;
1946         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1947         blk_status_t ret;
1948
1949         ret = btrfs_map_bio(fs_info, bio, mirror_num, 1);
1950         if (ret) {
1951                 bio->bi_status = ret;
1952                 bio_endio(bio);
1953         }
1954         return ret;
1955 }
1956
1957 /*
1958  * extent_io.c submission hook. This does the right thing for csum calculation
1959  * on write, or reading the csums from the tree before a read.
1960  *
1961  * Rules about async/sync submit,
1962  * a) read:                             sync submit
1963  *
1964  * b) write without checksum:           sync submit
1965  *
1966  * c) write with checksum:
1967  *    c-1) if bio is issued by fsync:   sync submit
1968  *         (sync_writers != 0)
1969  *
1970  *    c-2) if root is reloc root:       sync submit
1971  *         (only in case of buffered IO)
1972  *
1973  *    c-3) otherwise:                   async submit
1974  */
1975 static blk_status_t btrfs_submit_bio_hook(void *private_data, struct bio *bio,
1976                                  int mirror_num, unsigned long bio_flags,
1977                                  u64 bio_offset)
1978 {
1979         struct inode *inode = private_data;
1980         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1981         struct btrfs_root *root = BTRFS_I(inode)->root;
1982         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1983         blk_status_t ret = 0;
1984         int skip_sum;
1985         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1986
1987         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1988
1989         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
1990                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1991
1992         if (bio_op(bio) != REQ_OP_WRITE) {
1993                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
1994                 if (ret)
1995                         goto out;
1996
1997                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1998                         ret = btrfs_submit_compressed_read(inode, bio,
1999                                                            mirror_num,
2000                                                            bio_flags);
2001                         goto out;
2002                 } else if (!skip_sum) {
2003                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
2004                         if (ret)
2005                                 goto out;
2006                 }
2007                 goto mapit;
2008         } else if (async && !skip_sum) {
2009                 /* csum items have already been cloned */
2010                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2011                         goto mapit;
2012                 /* we're doing a write, do the async checksumming */
2013                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
2014                                           bio_offset, inode,
2015                                           btrfs_submit_bio_start);
2016                 goto out;
2017         } else if (!skip_sum) {
2018                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2019                 if (ret)
2020                         goto out;
2021         }
2022
2023 mapit:
2024         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
2025
2026 out:
2027         if (ret) {
2028                 bio->bi_status = ret;
2029                 bio_endio(bio);
2030         }
2031         return ret;
2032 }
2033
2034 /*
2035  * given a list of ordered sums record them in the inode.  This happens
2036  * at IO completion time based on sums calculated at bio submission time.
2037  */
2038 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2039                              struct inode *inode, struct list_head *list)
2040 {
2041         struct btrfs_ordered_sum *sum;
2042         int ret;
2043
2044         list_for_each_entry(sum, list, list) {
2045                 trans->adding_csums = true;
2046                 ret = btrfs_csum_file_blocks(trans,
2047                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2048                 trans->adding_csums = false;
2049                 if (ret)
2050                         return ret;
2051         }
2052         return 0;
2053 }
2054
2055 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2056                               unsigned int extra_bits,
2057                               struct extent_state **cached_state, int dedupe)
2058 {
2059         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
2060         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2061                                    extra_bits, cached_state);
2062 }
2063
2064 /* see btrfs_writepage_start_hook for details on why this is required */
2065 struct btrfs_writepage_fixup {
2066         struct page *page;
2067         struct btrfs_work work;
2068 };
2069
2070 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2071 {
2072         struct btrfs_writepage_fixup *fixup;
2073         struct btrfs_ordered_extent *ordered;
2074         struct extent_state *cached_state = NULL;
2075         struct extent_changeset *data_reserved = NULL;
2076         struct page *page;
2077         struct inode *inode;
2078         u64 page_start;
2079         u64 page_end;
2080         int ret;
2081
2082         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2083         page = fixup->page;
2084 again:
2085         lock_page(page);
2086         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2087                 ClearPageChecked(page);
2088                 goto out_page;
2089         }
2090
2091         inode = page->mapping->host;
2092         page_start = page_offset(page);
2093         page_end = page_offset(page) + PAGE_SIZE - 1;
2094
2095         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2096                          &cached_state);
2097
2098         /* already ordered? We're done */
2099         if (PagePrivate2(page))
2100                 goto out;
2101
2102         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2103                                         PAGE_SIZE);
2104         if (ordered) {
2105                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2106                                      page_end, &cached_state);
2107                 unlock_page(page);
2108                 btrfs_start_ordered_extent(inode, ordered, 1);
2109                 btrfs_put_ordered_extent(ordered);
2110                 goto again;
2111         }
2112
2113         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2114                                            PAGE_SIZE);
2115         if (ret) {
2116                 mapping_set_error(page->mapping, ret);
2117                 end_extent_writepage(page, ret, page_start, page_end);
2118                 ClearPageChecked(page);
2119                 goto out;
2120          }
2121
2122         ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
2123                                         &cached_state, 0);
2124         if (ret) {
2125                 mapping_set_error(page->mapping, ret);
2126                 end_extent_writepage(page, ret, page_start, page_end);
2127                 ClearPageChecked(page);
2128                 goto out;
2129         }
2130
2131         ClearPageChecked(page);
2132         set_page_dirty(page);
2133         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, false);
2134 out:
2135         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2136                              &cached_state);
2137 out_page:
2138         unlock_page(page);
2139         put_page(page);
2140         kfree(fixup);
2141         extent_changeset_free(data_reserved);
2142 }
2143
2144 /*
2145  * There are a few paths in the higher layers of the kernel that directly
2146  * set the page dirty bit without asking the filesystem if it is a
2147  * good idea.  This causes problems because we want to make sure COW
2148  * properly happens and the data=ordered rules are followed.
2149  *
2150  * In our case any range that doesn't have the ORDERED bit set
2151  * hasn't been properly setup for IO.  We kick off an async process
2152  * to fix it up.  The async helper will wait for ordered extents, set
2153  * the delalloc bit and make it safe to write the page.
2154  */
2155 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2156 {
2157         struct inode *inode = page->mapping->host;
2158         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2159         struct btrfs_writepage_fixup *fixup;
2160
2161         /* this page is properly in the ordered list */
2162         if (TestClearPagePrivate2(page))
2163                 return 0;
2164
2165         if (PageChecked(page))
2166                 return -EAGAIN;
2167
2168         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2169         if (!fixup)
2170                 return -EAGAIN;
2171
2172         SetPageChecked(page);
2173         get_page(page);
2174         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2175                         btrfs_writepage_fixup_worker, NULL, NULL);
2176         fixup->page = page;
2177         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2178         return -EBUSY;
2179 }
2180
2181 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2182                                        struct inode *inode, u64 file_pos,
2183                                        u64 disk_bytenr, u64 disk_num_bytes,
2184                                        u64 num_bytes, u64 ram_bytes,
2185                                        u8 compression, u8 encryption,
2186                                        u16 other_encoding, int extent_type)
2187 {
2188         struct btrfs_root *root = BTRFS_I(inode)->root;
2189         struct btrfs_file_extent_item *fi;
2190         struct btrfs_path *path;
2191         struct extent_buffer *leaf;
2192         struct btrfs_key ins;
2193         u64 qg_released;
2194         int extent_inserted = 0;
2195         int ret;
2196
2197         path = btrfs_alloc_path();
2198         if (!path)
2199                 return -ENOMEM;
2200
2201         /*
2202          * we may be replacing one extent in the tree with another.
2203          * The new extent is pinned in the extent map, and we don't want
2204          * to drop it from the cache until it is completely in the btree.
2205          *
2206          * So, tell btrfs_drop_extents to leave this extent in the cache.
2207          * the caller is expected to unpin it and allow it to be merged
2208          * with the others.
2209          */
2210         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2211                                    file_pos + num_bytes, NULL, 0,
2212                                    1, sizeof(*fi), &extent_inserted);
2213         if (ret)
2214                 goto out;
2215
2216         if (!extent_inserted) {
2217                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2218                 ins.offset = file_pos;
2219                 ins.type = BTRFS_EXTENT_DATA_KEY;
2220
2221                 path->leave_spinning = 1;
2222                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2223                                               sizeof(*fi));
2224                 if (ret)
2225                         goto out;
2226         }
2227         leaf = path->nodes[0];
2228         fi = btrfs_item_ptr(leaf, path->slots[0],
2229                             struct btrfs_file_extent_item);
2230         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2231         btrfs_set_file_extent_type(leaf, fi, extent_type);
2232         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2233         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2234         btrfs_set_file_extent_offset(leaf, fi, 0);
2235         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2236         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2237         btrfs_set_file_extent_compression(leaf, fi, compression);
2238         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2239         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2240
2241         btrfs_mark_buffer_dirty(leaf);
2242         btrfs_release_path(path);
2243
2244         inode_add_bytes(inode, num_bytes);
2245
2246         ins.objectid = disk_bytenr;
2247         ins.offset = disk_num_bytes;
2248         ins.type = BTRFS_EXTENT_ITEM_KEY;
2249
2250         /*
2251          * Release the reserved range from inode dirty range map, as it is
2252          * already moved into delayed_ref_head
2253          */
2254         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2255         if (ret < 0)
2256                 goto out;
2257         qg_released = ret;
2258         ret = btrfs_alloc_reserved_file_extent(trans, root,
2259                                                btrfs_ino(BTRFS_I(inode)),
2260                                                file_pos, qg_released, &ins);
2261 out:
2262         btrfs_free_path(path);
2263
2264         return ret;
2265 }
2266
2267 /* snapshot-aware defrag */
2268 struct sa_defrag_extent_backref {
2269         struct rb_node node;
2270         struct old_sa_defrag_extent *old;
2271         u64 root_id;
2272         u64 inum;
2273         u64 file_pos;
2274         u64 extent_offset;
2275         u64 num_bytes;
2276         u64 generation;
2277 };
2278
2279 struct old_sa_defrag_extent {
2280         struct list_head list;
2281         struct new_sa_defrag_extent *new;
2282
2283         u64 extent_offset;
2284         u64 bytenr;
2285         u64 offset;
2286         u64 len;
2287         int count;
2288 };
2289
2290 struct new_sa_defrag_extent {
2291         struct rb_root root;
2292         struct list_head head;
2293         struct btrfs_path *path;
2294         struct inode *inode;
2295         u64 file_pos;
2296         u64 len;
2297         u64 bytenr;
2298         u64 disk_len;
2299         u8 compress_type;
2300 };
2301
2302 static int backref_comp(struct sa_defrag_extent_backref *b1,
2303                         struct sa_defrag_extent_backref *b2)
2304 {
2305         if (b1->root_id < b2->root_id)
2306                 return -1;
2307         else if (b1->root_id > b2->root_id)
2308                 return 1;
2309
2310         if (b1->inum < b2->inum)
2311                 return -1;
2312         else if (b1->inum > b2->inum)
2313                 return 1;
2314
2315         if (b1->file_pos < b2->file_pos)
2316                 return -1;
2317         else if (b1->file_pos > b2->file_pos)
2318                 return 1;
2319
2320         /*
2321          * [------------------------------] ===> (a range of space)
2322          *     |<--->|   |<---->| =============> (fs/file tree A)
2323          * |<---------------------------->| ===> (fs/file tree B)
2324          *
2325          * A range of space can refer to two file extents in one tree while
2326          * refer to only one file extent in another tree.
2327          *
2328          * So we may process a disk offset more than one time(two extents in A)
2329          * and locate at the same extent(one extent in B), then insert two same
2330          * backrefs(both refer to the extent in B).
2331          */
2332         return 0;
2333 }
2334
2335 static void backref_insert(struct rb_root *root,
2336                            struct sa_defrag_extent_backref *backref)
2337 {
2338         struct rb_node **p = &root->rb_node;
2339         struct rb_node *parent = NULL;
2340         struct sa_defrag_extent_backref *entry;
2341         int ret;
2342
2343         while (*p) {
2344                 parent = *p;
2345                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2346
2347                 ret = backref_comp(backref, entry);
2348                 if (ret < 0)
2349                         p = &(*p)->rb_left;
2350                 else
2351                         p = &(*p)->rb_right;
2352         }
2353
2354         rb_link_node(&backref->node, parent, p);
2355         rb_insert_color(&backref->node, root);
2356 }
2357
2358 /*
2359  * Note the backref might has changed, and in this case we just return 0.
2360  */
2361 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2362                                        void *ctx)
2363 {
2364         struct btrfs_file_extent_item *extent;
2365         struct old_sa_defrag_extent *old = ctx;
2366         struct new_sa_defrag_extent *new = old->new;
2367         struct btrfs_path *path = new->path;
2368         struct btrfs_key key;
2369         struct btrfs_root *root;
2370         struct sa_defrag_extent_backref *backref;
2371         struct extent_buffer *leaf;
2372         struct inode *inode = new->inode;
2373         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2374         int slot;
2375         int ret;
2376         u64 extent_offset;
2377         u64 num_bytes;
2378
2379         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2380             inum == btrfs_ino(BTRFS_I(inode)))
2381                 return 0;
2382
2383         key.objectid = root_id;
2384         key.type = BTRFS_ROOT_ITEM_KEY;
2385         key.offset = (u64)-1;
2386
2387         root = btrfs_read_fs_root_no_name(fs_info, &key);
2388         if (IS_ERR(root)) {
2389                 if (PTR_ERR(root) == -ENOENT)
2390                         return 0;
2391                 WARN_ON(1);
2392                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2393                          inum, offset, root_id);
2394                 return PTR_ERR(root);
2395         }
2396
2397         key.objectid = inum;
2398         key.type = BTRFS_EXTENT_DATA_KEY;
2399         if (offset > (u64)-1 << 32)
2400                 key.offset = 0;
2401         else
2402                 key.offset = offset;
2403
2404         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2405         if (WARN_ON(ret < 0))
2406                 return ret;
2407         ret = 0;
2408
2409         while (1) {
2410                 cond_resched();
2411
2412                 leaf = path->nodes[0];
2413                 slot = path->slots[0];
2414
2415                 if (slot >= btrfs_header_nritems(leaf)) {
2416                         ret = btrfs_next_leaf(root, path);
2417                         if (ret < 0) {
2418                                 goto out;
2419                         } else if (ret > 0) {
2420                                 ret = 0;
2421                                 goto out;
2422                         }
2423                         continue;
2424                 }
2425
2426                 path->slots[0]++;
2427
2428                 btrfs_item_key_to_cpu(leaf, &key, slot);
2429
2430                 if (key.objectid > inum)
2431                         goto out;
2432
2433                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2434                         continue;
2435
2436                 extent = btrfs_item_ptr(leaf, slot,
2437                                         struct btrfs_file_extent_item);
2438
2439                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2440                         continue;
2441
2442                 /*
2443                  * 'offset' refers to the exact key.offset,
2444                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2445                  * (key.offset - extent_offset).
2446                  */
2447                 if (key.offset != offset)
2448                         continue;
2449
2450                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2451                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2452
2453                 if (extent_offset >= old->extent_offset + old->offset +
2454                     old->len || extent_offset + num_bytes <=
2455                     old->extent_offset + old->offset)
2456                         continue;
2457                 break;
2458         }
2459
2460         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2461         if (!backref) {
2462                 ret = -ENOENT;
2463                 goto out;
2464         }
2465
2466         backref->root_id = root_id;
2467         backref->inum = inum;
2468         backref->file_pos = offset;
2469         backref->num_bytes = num_bytes;
2470         backref->extent_offset = extent_offset;
2471         backref->generation = btrfs_file_extent_generation(leaf, extent);
2472         backref->old = old;
2473         backref_insert(&new->root, backref);
2474         old->count++;
2475 out:
2476         btrfs_release_path(path);
2477         WARN_ON(ret);
2478         return ret;
2479 }
2480
2481 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2482                                    struct new_sa_defrag_extent *new)
2483 {
2484         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2485         struct old_sa_defrag_extent *old, *tmp;
2486         int ret;
2487
2488         new->path = path;
2489
2490         list_for_each_entry_safe(old, tmp, &new->head, list) {
2491                 ret = iterate_inodes_from_logical(old->bytenr +
2492                                                   old->extent_offset, fs_info,
2493                                                   path, record_one_backref,
2494                                                   old, false);
2495                 if (ret < 0 && ret != -ENOENT)
2496                         return false;
2497
2498                 /* no backref to be processed for this extent */
2499                 if (!old->count) {
2500                         list_del(&old->list);
2501                         kfree(old);
2502                 }
2503         }
2504
2505         if (list_empty(&new->head))
2506                 return false;
2507
2508         return true;
2509 }
2510
2511 static int relink_is_mergable(struct extent_buffer *leaf,
2512                               struct btrfs_file_extent_item *fi,
2513                               struct new_sa_defrag_extent *new)
2514 {
2515         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2516                 return 0;
2517
2518         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2519                 return 0;
2520
2521         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2522                 return 0;
2523
2524         if (btrfs_file_extent_encryption(leaf, fi) ||
2525             btrfs_file_extent_other_encoding(leaf, fi))
2526                 return 0;
2527
2528         return 1;
2529 }
2530
2531 /*
2532  * Note the backref might has changed, and in this case we just return 0.
2533  */
2534 static noinline int relink_extent_backref(struct btrfs_path *path,
2535                                  struct sa_defrag_extent_backref *prev,
2536                                  struct sa_defrag_extent_backref *backref)
2537 {
2538         struct btrfs_file_extent_item *extent;
2539         struct btrfs_file_extent_item *item;
2540         struct btrfs_ordered_extent *ordered;
2541         struct btrfs_trans_handle *trans;
2542         struct btrfs_root *root;
2543         struct btrfs_key key;
2544         struct extent_buffer *leaf;
2545         struct old_sa_defrag_extent *old = backref->old;
2546         struct new_sa_defrag_extent *new = old->new;
2547         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2548         struct inode *inode;
2549         struct extent_state *cached = NULL;
2550         int ret = 0;
2551         u64 start;
2552         u64 len;
2553         u64 lock_start;
2554         u64 lock_end;
2555         bool merge = false;
2556         int index;
2557
2558         if (prev && prev->root_id == backref->root_id &&
2559             prev->inum == backref->inum &&
2560             prev->file_pos + prev->num_bytes == backref->file_pos)
2561                 merge = true;
2562
2563         /* step 1: get root */
2564         key.objectid = backref->root_id;
2565         key.type = BTRFS_ROOT_ITEM_KEY;
2566         key.offset = (u64)-1;
2567
2568         index = srcu_read_lock(&fs_info->subvol_srcu);
2569
2570         root = btrfs_read_fs_root_no_name(fs_info, &key);
2571         if (IS_ERR(root)) {
2572                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2573                 if (PTR_ERR(root) == -ENOENT)
2574                         return 0;
2575                 return PTR_ERR(root);
2576         }
2577
2578         if (btrfs_root_readonly(root)) {
2579                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2580                 return 0;
2581         }
2582
2583         /* step 2: get inode */
2584         key.objectid = backref->inum;
2585         key.type = BTRFS_INODE_ITEM_KEY;
2586         key.offset = 0;
2587
2588         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2589         if (IS_ERR(inode)) {
2590                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2591                 return 0;
2592         }
2593
2594         srcu_read_unlock(&fs_info->subvol_srcu, index);
2595
2596         /* step 3: relink backref */
2597         lock_start = backref->file_pos;
2598         lock_end = backref->file_pos + backref->num_bytes - 1;
2599         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2600                          &cached);
2601
2602         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2603         if (ordered) {
2604                 btrfs_put_ordered_extent(ordered);
2605                 goto out_unlock;
2606         }
2607
2608         trans = btrfs_join_transaction(root);
2609         if (IS_ERR(trans)) {
2610                 ret = PTR_ERR(trans);
2611                 goto out_unlock;
2612         }
2613
2614         key.objectid = backref->inum;
2615         key.type = BTRFS_EXTENT_DATA_KEY;
2616         key.offset = backref->file_pos;
2617
2618         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2619         if (ret < 0) {
2620                 goto out_free_path;
2621         } else if (ret > 0) {
2622                 ret = 0;
2623                 goto out_free_path;
2624         }
2625
2626         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2627                                 struct btrfs_file_extent_item);
2628
2629         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2630             backref->generation)
2631                 goto out_free_path;
2632
2633         btrfs_release_path(path);
2634
2635         start = backref->file_pos;
2636         if (backref->extent_offset < old->extent_offset + old->offset)
2637                 start += old->extent_offset + old->offset -
2638                          backref->extent_offset;
2639
2640         len = min(backref->extent_offset + backref->num_bytes,
2641                   old->extent_offset + old->offset + old->len);
2642         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2643
2644         ret = btrfs_drop_extents(trans, root, inode, start,
2645                                  start + len, 1);
2646         if (ret)
2647                 goto out_free_path;
2648 again:
2649         key.objectid = btrfs_ino(BTRFS_I(inode));
2650         key.type = BTRFS_EXTENT_DATA_KEY;
2651         key.offset = start;
2652
2653         path->leave_spinning = 1;
2654         if (merge) {
2655                 struct btrfs_file_extent_item *fi;
2656                 u64 extent_len;
2657                 struct btrfs_key found_key;
2658
2659                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2660                 if (ret < 0)
2661                         goto out_free_path;
2662
2663                 path->slots[0]--;
2664                 leaf = path->nodes[0];
2665                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2666
2667                 fi = btrfs_item_ptr(leaf, path->slots[0],
2668                                     struct btrfs_file_extent_item);
2669                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2670
2671                 if (extent_len + found_key.offset == start &&
2672                     relink_is_mergable(leaf, fi, new)) {
2673                         btrfs_set_file_extent_num_bytes(leaf, fi,
2674                                                         extent_len + len);
2675                         btrfs_mark_buffer_dirty(leaf);
2676                         inode_add_bytes(inode, len);
2677
2678                         ret = 1;
2679                         goto out_free_path;
2680                 } else {
2681                         merge = false;
2682                         btrfs_release_path(path);
2683                         goto again;
2684                 }
2685         }
2686
2687         ret = btrfs_insert_empty_item(trans, root, path, &key,
2688                                         sizeof(*extent));
2689         if (ret) {
2690                 btrfs_abort_transaction(trans, ret);
2691                 goto out_free_path;
2692         }
2693
2694         leaf = path->nodes[0];
2695         item = btrfs_item_ptr(leaf, path->slots[0],
2696                                 struct btrfs_file_extent_item);
2697         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2698         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2699         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2700         btrfs_set_file_extent_num_bytes(leaf, item, len);
2701         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2702         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2703         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2704         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2705         btrfs_set_file_extent_encryption(leaf, item, 0);
2706         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2707
2708         btrfs_mark_buffer_dirty(leaf);
2709         inode_add_bytes(inode, len);
2710         btrfs_release_path(path);
2711
2712         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2713                         new->disk_len, 0,
2714                         backref->root_id, backref->inum,
2715                         new->file_pos); /* start - extent_offset */
2716         if (ret) {
2717                 btrfs_abort_transaction(trans, ret);
2718                 goto out_free_path;
2719         }
2720
2721         ret = 1;
2722 out_free_path:
2723         btrfs_release_path(path);
2724         path->leave_spinning = 0;
2725         btrfs_end_transaction(trans);
2726 out_unlock:
2727         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2728                              &cached);
2729         iput(inode);
2730         return ret;
2731 }
2732
2733 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2734 {
2735         struct old_sa_defrag_extent *old, *tmp;
2736
2737         if (!new)
2738                 return;
2739
2740         list_for_each_entry_safe(old, tmp, &new->head, list) {
2741                 kfree(old);
2742         }
2743         kfree(new);
2744 }
2745
2746 static void relink_file_extents(struct new_sa_defrag_extent *new)
2747 {
2748         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2749         struct btrfs_path *path;
2750         struct sa_defrag_extent_backref *backref;
2751         struct sa_defrag_extent_backref *prev = NULL;
2752         struct rb_node *node;
2753         int ret;
2754
2755         path = btrfs_alloc_path();
2756         if (!path)
2757                 return;
2758
2759         if (!record_extent_backrefs(path, new)) {
2760                 btrfs_free_path(path);
2761                 goto out;
2762         }
2763         btrfs_release_path(path);
2764
2765         while (1) {
2766                 node = rb_first(&new->root);
2767                 if (!node)
2768                         break;
2769                 rb_erase(node, &new->root);
2770
2771                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2772
2773                 ret = relink_extent_backref(path, prev, backref);
2774                 WARN_ON(ret < 0);
2775
2776                 kfree(prev);
2777
2778                 if (ret == 1)
2779                         prev = backref;
2780                 else
2781                         prev = NULL;
2782                 cond_resched();
2783         }
2784         kfree(prev);
2785
2786         btrfs_free_path(path);
2787 out:
2788         free_sa_defrag_extent(new);
2789
2790         atomic_dec(&fs_info->defrag_running);
2791         wake_up(&fs_info->transaction_wait);
2792 }
2793
2794 static struct new_sa_defrag_extent *
2795 record_old_file_extents(struct inode *inode,
2796                         struct btrfs_ordered_extent *ordered)
2797 {
2798         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2799         struct btrfs_root *root = BTRFS_I(inode)->root;
2800         struct btrfs_path *path;
2801         struct btrfs_key key;
2802         struct old_sa_defrag_extent *old;
2803         struct new_sa_defrag_extent *new;
2804         int ret;
2805
2806         new = kmalloc(sizeof(*new), GFP_NOFS);
2807         if (!new)
2808                 return NULL;
2809
2810         new->inode = inode;
2811         new->file_pos = ordered->file_offset;
2812         new->len = ordered->len;
2813         new->bytenr = ordered->start;
2814         new->disk_len = ordered->disk_len;
2815         new->compress_type = ordered->compress_type;
2816         new->root = RB_ROOT;
2817         INIT_LIST_HEAD(&new->head);
2818
2819         path = btrfs_alloc_path();
2820         if (!path)
2821                 goto out_kfree;
2822
2823         key.objectid = btrfs_ino(BTRFS_I(inode));
2824         key.type = BTRFS_EXTENT_DATA_KEY;
2825         key.offset = new->file_pos;
2826
2827         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2828         if (ret < 0)
2829                 goto out_free_path;
2830         if (ret > 0 && path->slots[0] > 0)
2831                 path->slots[0]--;
2832
2833         /* find out all the old extents for the file range */
2834         while (1) {
2835                 struct btrfs_file_extent_item *extent;
2836                 struct extent_buffer *l;
2837                 int slot;
2838                 u64 num_bytes;
2839                 u64 offset;
2840                 u64 end;
2841                 u64 disk_bytenr;
2842                 u64 extent_offset;
2843
2844                 l = path->nodes[0];
2845                 slot = path->slots[0];
2846
2847                 if (slot >= btrfs_header_nritems(l)) {
2848                         ret = btrfs_next_leaf(root, path);
2849                         if (ret < 0)
2850                                 goto out_free_path;
2851                         else if (ret > 0)
2852                                 break;
2853                         continue;
2854                 }
2855
2856                 btrfs_item_key_to_cpu(l, &key, slot);
2857
2858                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2859                         break;
2860                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2861                         break;
2862                 if (key.offset >= new->file_pos + new->len)
2863                         break;
2864
2865                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2866
2867                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2868                 if (key.offset + num_bytes < new->file_pos)
2869                         goto next;
2870
2871                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2872                 if (!disk_bytenr)
2873                         goto next;
2874
2875                 extent_offset = btrfs_file_extent_offset(l, extent);
2876
2877                 old = kmalloc(sizeof(*old), GFP_NOFS);
2878                 if (!old)
2879                         goto out_free_path;
2880
2881                 offset = max(new->file_pos, key.offset);
2882                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2883
2884                 old->bytenr = disk_bytenr;
2885                 old->extent_offset = extent_offset;
2886                 old->offset = offset - key.offset;
2887                 old->len = end - offset;
2888                 old->new = new;
2889                 old->count = 0;
2890                 list_add_tail(&old->list, &new->head);
2891 next:
2892                 path->slots[0]++;
2893                 cond_resched();
2894         }
2895
2896         btrfs_free_path(path);
2897         atomic_inc(&fs_info->defrag_running);
2898
2899         return new;
2900
2901 out_free_path:
2902         btrfs_free_path(path);
2903 out_kfree:
2904         free_sa_defrag_extent(new);
2905         return NULL;
2906 }
2907
2908 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2909                                          u64 start, u64 len)
2910 {
2911         struct btrfs_block_group_cache *cache;
2912
2913         cache = btrfs_lookup_block_group(fs_info, start);
2914         ASSERT(cache);
2915
2916         spin_lock(&cache->lock);
2917         cache->delalloc_bytes -= len;
2918         spin_unlock(&cache->lock);
2919
2920         btrfs_put_block_group(cache);
2921 }
2922
2923 /* as ordered data IO finishes, this gets called so we can finish
2924  * an ordered extent if the range of bytes in the file it covers are
2925  * fully written.
2926  */
2927 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2928 {
2929         struct inode *inode = ordered_extent->inode;
2930         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2931         struct btrfs_root *root = BTRFS_I(inode)->root;
2932         struct btrfs_trans_handle *trans = NULL;
2933         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2934         struct extent_state *cached_state = NULL;
2935         struct new_sa_defrag_extent *new = NULL;
2936         int compress_type = 0;
2937         int ret = 0;
2938         u64 logical_len = ordered_extent->len;
2939         bool nolock;
2940         bool truncated = false;
2941         bool range_locked = false;
2942         bool clear_new_delalloc_bytes = false;
2943
2944         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2945             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
2946             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
2947                 clear_new_delalloc_bytes = true;
2948
2949         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
2950
2951         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2952                 ret = -EIO;
2953                 goto out;
2954         }
2955
2956         btrfs_free_io_failure_record(BTRFS_I(inode),
2957                         ordered_extent->file_offset,
2958                         ordered_extent->file_offset +
2959                         ordered_extent->len - 1);
2960
2961         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2962                 truncated = true;
2963                 logical_len = ordered_extent->truncated_len;
2964                 /* Truncated the entire extent, don't bother adding */
2965                 if (!logical_len)
2966                         goto out;
2967         }
2968
2969         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2970                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2971
2972                 /*
2973                  * For mwrite(mmap + memset to write) case, we still reserve
2974                  * space for NOCOW range.
2975                  * As NOCOW won't cause a new delayed ref, just free the space
2976                  */
2977                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
2978                                        ordered_extent->len);
2979                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2980                 if (nolock)
2981                         trans = btrfs_join_transaction_nolock(root);
2982                 else
2983                         trans = btrfs_join_transaction(root);
2984                 if (IS_ERR(trans)) {
2985                         ret = PTR_ERR(trans);
2986                         trans = NULL;
2987                         goto out;
2988                 }
2989                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
2990                 ret = btrfs_update_inode_fallback(trans, root, inode);
2991                 if (ret) /* -ENOMEM or corruption */
2992                         btrfs_abort_transaction(trans, ret);
2993                 goto out;
2994         }
2995
2996         range_locked = true;
2997         lock_extent_bits(io_tree, ordered_extent->file_offset,
2998                          ordered_extent->file_offset + ordered_extent->len - 1,
2999                          &cached_state);
3000
3001         ret = test_range_bit(io_tree, ordered_extent->file_offset,
3002                         ordered_extent->file_offset + ordered_extent->len - 1,
3003                         EXTENT_DEFRAG, 0, cached_state);
3004         if (ret) {
3005                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
3006                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
3007                         /* the inode is shared */
3008                         new = record_old_file_extents(inode, ordered_extent);
3009
3010                 clear_extent_bit(io_tree, ordered_extent->file_offset,
3011                         ordered_extent->file_offset + ordered_extent->len - 1,
3012                         EXTENT_DEFRAG, 0, 0, &cached_state);
3013         }
3014
3015         if (nolock)
3016                 trans = btrfs_join_transaction_nolock(root);
3017         else
3018                 trans = btrfs_join_transaction(root);
3019         if (IS_ERR(trans)) {
3020                 ret = PTR_ERR(trans);
3021                 trans = NULL;
3022                 goto out;
3023         }
3024
3025         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3026
3027         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
3028                 compress_type = ordered_extent->compress_type;
3029         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
3030                 BUG_ON(compress_type);
3031                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3032                                        ordered_extent->len);
3033                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
3034                                                 ordered_extent->file_offset,
3035                                                 ordered_extent->file_offset +
3036                                                 logical_len);
3037         } else {
3038                 BUG_ON(root == fs_info->tree_root);
3039                 ret = insert_reserved_file_extent(trans, inode,
3040                                                 ordered_extent->file_offset,
3041                                                 ordered_extent->start,
3042                                                 ordered_extent->disk_len,
3043                                                 logical_len, logical_len,
3044                                                 compress_type, 0, 0,
3045                                                 BTRFS_FILE_EXTENT_REG);
3046                 if (!ret)
3047                         btrfs_release_delalloc_bytes(fs_info,
3048                                                      ordered_extent->start,
3049                                                      ordered_extent->disk_len);
3050         }
3051         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3052                            ordered_extent->file_offset, ordered_extent->len,
3053                            trans->transid);
3054         if (ret < 0) {
3055                 btrfs_abort_transaction(trans, ret);
3056                 goto out;
3057         }
3058
3059         ret = add_pending_csums(trans, inode, &ordered_extent->list);
3060         if (ret) {
3061                 btrfs_abort_transaction(trans, ret);
3062                 goto out;
3063         }
3064
3065         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3066         ret = btrfs_update_inode_fallback(trans, root, inode);
3067         if (ret) { /* -ENOMEM or corruption */
3068                 btrfs_abort_transaction(trans, ret);
3069                 goto out;
3070         }
3071         ret = 0;
3072 out:
3073         if (range_locked || clear_new_delalloc_bytes) {
3074                 unsigned int clear_bits = 0;
3075
3076                 if (range_locked)
3077                         clear_bits |= EXTENT_LOCKED;
3078                 if (clear_new_delalloc_bytes)
3079                         clear_bits |= EXTENT_DELALLOC_NEW;
3080                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3081                                  ordered_extent->file_offset,
3082                                  ordered_extent->file_offset +
3083                                  ordered_extent->len - 1,
3084                                  clear_bits,
3085                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3086                                  0, &cached_state);
3087         }
3088
3089         if (trans)
3090                 btrfs_end_transaction(trans);
3091
3092         if (ret || truncated) {
3093                 u64 start, end;
3094
3095                 if (truncated)
3096                         start = ordered_extent->file_offset + logical_len;
3097                 else
3098                         start = ordered_extent->file_offset;
3099                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3100                 clear_extent_uptodate(io_tree, start, end, NULL);
3101
3102                 /* Drop the cache for the part of the extent we didn't write. */
3103                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3104
3105                 /*
3106                  * If the ordered extent had an IOERR or something else went
3107                  * wrong we need to return the space for this ordered extent
3108                  * back to the allocator.  We only free the extent in the
3109                  * truncated case if we didn't write out the extent at all.
3110                  */
3111                 if ((ret || !logical_len) &&
3112                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3113                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3114                         btrfs_free_reserved_extent(fs_info,
3115                                                    ordered_extent->start,
3116                                                    ordered_extent->disk_len, 1);
3117         }
3118
3119
3120         /*
3121          * This needs to be done to make sure anybody waiting knows we are done
3122          * updating everything for this ordered extent.
3123          */
3124         btrfs_remove_ordered_extent(inode, ordered_extent);
3125
3126         /* for snapshot-aware defrag */
3127         if (new) {
3128                 if (ret) {
3129                         free_sa_defrag_extent(new);
3130                         atomic_dec(&fs_info->defrag_running);
3131                 } else {
3132                         relink_file_extents(new);
3133                 }
3134         }
3135
3136         /* once for us */
3137         btrfs_put_ordered_extent(ordered_extent);
3138         /* once for the tree */
3139         btrfs_put_ordered_extent(ordered_extent);
3140
3141         /* Try to release some metadata so we don't get an OOM but don't wait */
3142         btrfs_btree_balance_dirty_nodelay(fs_info);
3143
3144         return ret;
3145 }
3146
3147 static void finish_ordered_fn(struct btrfs_work *work)
3148 {
3149         struct btrfs_ordered_extent *ordered_extent;
3150         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3151         btrfs_finish_ordered_io(ordered_extent);
3152 }
3153
3154 static void btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3155                                 struct extent_state *state, int uptodate)
3156 {
3157         struct inode *inode = page->mapping->host;
3158         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3159         struct btrfs_ordered_extent *ordered_extent = NULL;
3160         struct btrfs_workqueue *wq;
3161         btrfs_work_func_t func;
3162
3163         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3164
3165         ClearPagePrivate2(page);
3166         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3167                                             end - start + 1, uptodate))
3168                 return;
3169
3170         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
3171                 wq = fs_info->endio_freespace_worker;
3172                 func = btrfs_freespace_write_helper;
3173         } else {
3174                 wq = fs_info->endio_write_workers;
3175                 func = btrfs_endio_write_helper;
3176         }
3177
3178         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3179                         NULL);
3180         btrfs_queue_work(wq, &ordered_extent->work);
3181 }
3182
3183 static int __readpage_endio_check(struct inode *inode,
3184                                   struct btrfs_io_bio *io_bio,
3185                                   int icsum, struct page *page,
3186                                   int pgoff, u64 start, size_t len)
3187 {
3188         char *kaddr;
3189         u32 csum_expected;
3190         u32 csum = ~(u32)0;
3191
3192         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3193
3194         kaddr = kmap_atomic(page);
3195         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3196         btrfs_csum_final(csum, (u8 *)&csum);
3197         if (csum != csum_expected)
3198                 goto zeroit;
3199
3200         kunmap_atomic(kaddr);
3201         return 0;
3202 zeroit:
3203         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3204                                     io_bio->mirror_num);
3205         memset(kaddr + pgoff, 1, len);
3206         flush_dcache_page(page);
3207         kunmap_atomic(kaddr);
3208         return -EIO;
3209 }
3210
3211 /*
3212  * when reads are done, we need to check csums to verify the data is correct
3213  * if there's a match, we allow the bio to finish.  If not, the code in
3214  * extent_io.c will try to find good copies for us.
3215  */
3216 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3217                                       u64 phy_offset, struct page *page,
3218                                       u64 start, u64 end, int mirror)
3219 {
3220         size_t offset = start - page_offset(page);
3221         struct inode *inode = page->mapping->host;
3222         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3223         struct btrfs_root *root = BTRFS_I(inode)->root;
3224
3225         if (PageChecked(page)) {
3226                 ClearPageChecked(page);
3227                 return 0;
3228         }
3229
3230         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3231                 return 0;
3232
3233         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3234             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3235                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3236                 return 0;
3237         }
3238
3239         phy_offset >>= inode->i_sb->s_blocksize_bits;
3240         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3241                                       start, (size_t)(end - start + 1));
3242 }
3243
3244 /*
3245  * btrfs_add_delayed_iput - perform a delayed iput on @inode
3246  *
3247  * @inode: The inode we want to perform iput on
3248  *
3249  * This function uses the generic vfs_inode::i_count to track whether we should
3250  * just decrement it (in case it's > 1) or if this is the last iput then link
3251  * the inode to the delayed iput machinery. Delayed iputs are processed at
3252  * transaction commit time/superblock commit/cleaner kthread.
3253  */
3254 void btrfs_add_delayed_iput(struct inode *inode)
3255 {
3256         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3257         struct btrfs_inode *binode = BTRFS_I(inode);
3258
3259         if (atomic_add_unless(&inode->i_count, -1, 1))
3260                 return;
3261
3262         spin_lock(&fs_info->delayed_iput_lock);
3263         ASSERT(list_empty(&binode->delayed_iput));
3264         list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3265         spin_unlock(&fs_info->delayed_iput_lock);
3266 }
3267
3268 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3269 {
3270
3271         spin_lock(&fs_info->delayed_iput_lock);
3272         while (!list_empty(&fs_info->delayed_iputs)) {
3273                 struct btrfs_inode *inode;
3274
3275                 inode = list_first_entry(&fs_info->delayed_iputs,
3276                                 struct btrfs_inode, delayed_iput);
3277                 list_del_init(&inode->delayed_iput);
3278                 spin_unlock(&fs_info->delayed_iput_lock);
3279                 iput(&inode->vfs_inode);
3280                 spin_lock(&fs_info->delayed_iput_lock);
3281         }
3282         spin_unlock(&fs_info->delayed_iput_lock);
3283 }
3284
3285 /*
3286  * This creates an orphan entry for the given inode in case something goes wrong
3287  * in the middle of an unlink.
3288  */
3289 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3290                      struct btrfs_inode *inode)
3291 {
3292         int ret;
3293
3294         ret = btrfs_insert_orphan_item(trans, inode->root, btrfs_ino(inode));
3295         if (ret && ret != -EEXIST) {
3296                 btrfs_abort_transaction(trans, ret);
3297                 return ret;
3298         }
3299
3300         return 0;
3301 }
3302
3303 /*
3304  * We have done the delete so we can go ahead and remove the orphan item for
3305  * this particular inode.
3306  */
3307 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3308                             struct btrfs_inode *inode)
3309 {
3310         return btrfs_del_orphan_item(trans, inode->root, btrfs_ino(inode));
3311 }
3312
3313 /*
3314  * this cleans up any orphans that may be left on the list from the last use
3315  * of this root.
3316  */
3317 int btrfs_orphan_cleanup(struct btrfs_root *root)
3318 {
3319         struct btrfs_fs_info *fs_info = root->fs_info;
3320         struct btrfs_path *path;
3321         struct extent_buffer *leaf;
3322         struct btrfs_key key, found_key;
3323         struct btrfs_trans_handle *trans;
3324         struct inode *inode;
3325         u64 last_objectid = 0;
3326         int ret = 0, nr_unlink = 0;
3327
3328         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3329                 return 0;
3330
3331         path = btrfs_alloc_path();
3332         if (!path) {
3333                 ret = -ENOMEM;
3334                 goto out;
3335         }
3336         path->reada = READA_BACK;
3337
3338         key.objectid = BTRFS_ORPHAN_OBJECTID;
3339         key.type = BTRFS_ORPHAN_ITEM_KEY;
3340         key.offset = (u64)-1;
3341
3342         while (1) {
3343                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3344                 if (ret < 0)
3345                         goto out;
3346
3347                 /*
3348                  * if ret == 0 means we found what we were searching for, which
3349                  * is weird, but possible, so only screw with path if we didn't
3350                  * find the key and see if we have stuff that matches
3351                  */
3352                 if (ret > 0) {
3353                         ret = 0;
3354                         if (path->slots[0] == 0)
3355                                 break;
3356                         path->slots[0]--;
3357                 }
3358
3359                 /* pull out the item */
3360                 leaf = path->nodes[0];
3361                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3362
3363                 /* make sure the item matches what we want */
3364                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3365                         break;
3366                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3367                         break;
3368
3369                 /* release the path since we're done with it */
3370                 btrfs_release_path(path);
3371
3372                 /*
3373                  * this is where we are basically btrfs_lookup, without the
3374                  * crossing root thing.  we store the inode number in the
3375                  * offset of the orphan item.
3376                  */
3377
3378                 if (found_key.offset == last_objectid) {
3379                         btrfs_err(fs_info,
3380                                   "Error removing orphan entry, stopping orphan cleanup");
3381                         ret = -EINVAL;
3382                         goto out;
3383                 }
3384
3385                 last_objectid = found_key.offset;
3386
3387                 found_key.objectid = found_key.offset;
3388                 found_key.type = BTRFS_INODE_ITEM_KEY;
3389                 found_key.offset = 0;
3390                 inode = btrfs_iget(fs_info->sb, &found_key, root, NULL);
3391                 ret = PTR_ERR_OR_ZERO(inode);
3392                 if (ret && ret != -ENOENT)
3393                         goto out;
3394
3395                 if (ret == -ENOENT && root == fs_info->tree_root) {
3396                         struct btrfs_root *dead_root;
3397                         struct btrfs_fs_info *fs_info = root->fs_info;
3398                         int is_dead_root = 0;
3399
3400                         /*
3401                          * this is an orphan in the tree root. Currently these
3402                          * could come from 2 sources:
3403                          *  a) a snapshot deletion in progress
3404                          *  b) a free space cache inode
3405                          * We need to distinguish those two, as the snapshot
3406                          * orphan must not get deleted.
3407                          * find_dead_roots already ran before us, so if this
3408                          * is a snapshot deletion, we should find the root
3409                          * in the dead_roots list
3410                          */
3411                         spin_lock(&fs_info->trans_lock);
3412                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3413                                             root_list) {
3414                                 if (dead_root->root_key.objectid ==
3415                                     found_key.objectid) {
3416                                         is_dead_root = 1;
3417                                         break;
3418                                 }
3419                         }
3420                         spin_unlock(&fs_info->trans_lock);
3421                         if (is_dead_root) {
3422                                 /* prevent this orphan from being found again */
3423                                 key.offset = found_key.objectid - 1;
3424                                 continue;
3425                         }
3426
3427                 }
3428
3429                 /*
3430                  * If we have an inode with links, there are a couple of
3431                  * possibilities. Old kernels (before v3.12) used to create an
3432                  * orphan item for truncate indicating that there were possibly
3433                  * extent items past i_size that needed to be deleted. In v3.12,
3434                  * truncate was changed to update i_size in sync with the extent
3435                  * items, but the (useless) orphan item was still created. Since
3436                  * v4.18, we don't create the orphan item for truncate at all.
3437                  *
3438                  * So, this item could mean that we need to do a truncate, but
3439                  * only if this filesystem was last used on a pre-v3.12 kernel
3440                  * and was not cleanly unmounted. The odds of that are quite
3441                  * slim, and it's a pain to do the truncate now, so just delete
3442                  * the orphan item.
3443                  *
3444                  * It's also possible that this orphan item was supposed to be
3445                  * deleted but wasn't. The inode number may have been reused,
3446                  * but either way, we can delete the orphan item.
3447                  */
3448                 if (ret == -ENOENT || inode->i_nlink) {
3449                         if (!ret)
3450                                 iput(inode);
3451                         trans = btrfs_start_transaction(root, 1);
3452                         if (IS_ERR(trans)) {
3453                                 ret = PTR_ERR(trans);
3454                                 goto out;
3455                         }
3456                         btrfs_debug(fs_info, "auto deleting %Lu",
3457                                     found_key.objectid);
3458                         ret = btrfs_del_orphan_item(trans, root,
3459                                                     found_key.objectid);
3460                         btrfs_end_transaction(trans);
3461                         if (ret)
3462                                 goto out;
3463                         continue;
3464                 }
3465
3466                 nr_unlink++;
3467
3468                 /* this will do delete_inode and everything for us */
3469                 iput(inode);
3470         }
3471         /* release the path since we're done with it */
3472         btrfs_release_path(path);
3473
3474         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3475
3476         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3477                 trans = btrfs_join_transaction(root);
3478                 if (!IS_ERR(trans))
3479                         btrfs_end_transaction(trans);
3480         }
3481
3482         if (nr_unlink)
3483                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3484
3485 out:
3486         if (ret)
3487                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3488         btrfs_free_path(path);
3489         return ret;
3490 }
3491
3492 /*
3493  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3494  * don't find any xattrs, we know there can't be any acls.
3495  *
3496  * slot is the slot the inode is in, objectid is the objectid of the inode
3497  */
3498 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3499                                           int slot, u64 objectid,
3500                                           int *first_xattr_slot)
3501 {
3502         u32 nritems = btrfs_header_nritems(leaf);
3503         struct btrfs_key found_key;
3504         static u64 xattr_access = 0;
3505         static u64 xattr_default = 0;
3506         int scanned = 0;
3507
3508         if (!xattr_access) {
3509                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3510                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3511                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3512                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3513         }
3514
3515         slot++;
3516         *first_xattr_slot = -1;
3517         while (slot < nritems) {
3518                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3519
3520                 /* we found a different objectid, there must not be acls */
3521                 if (found_key.objectid != objectid)
3522                         return 0;
3523
3524                 /* we found an xattr, assume we've got an acl */
3525                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3526                         if (*first_xattr_slot == -1)
3527                                 *first_xattr_slot = slot;
3528                         if (found_key.offset == xattr_access ||
3529                             found_key.offset == xattr_default)
3530                                 return 1;
3531                 }
3532
3533                 /*
3534                  * we found a key greater than an xattr key, there can't
3535                  * be any acls later on
3536                  */
3537                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3538                         return 0;
3539
3540                 slot++;
3541                 scanned++;
3542
3543                 /*
3544                  * it goes inode, inode backrefs, xattrs, extents,
3545                  * so if there are a ton of hard links to an inode there can
3546                  * be a lot of backrefs.  Don't waste time searching too hard,
3547                  * this is just an optimization
3548                  */
3549                 if (scanned >= 8)
3550                         break;
3551         }
3552         /* we hit the end of the leaf before we found an xattr or
3553          * something larger than an xattr.  We have to assume the inode
3554          * has acls
3555          */
3556         if (*first_xattr_slot == -1)
3557                 *first_xattr_slot = slot;
3558         return 1;
3559 }
3560
3561 /*
3562  * read an inode from the btree into the in-memory inode
3563  */
3564 static int btrfs_read_locked_inode(struct inode *inode)
3565 {
3566         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3567         struct btrfs_path *path;
3568         struct extent_buffer *leaf;
3569         struct btrfs_inode_item *inode_item;
3570         struct btrfs_root *root = BTRFS_I(inode)->root;
3571         struct btrfs_key location;
3572         unsigned long ptr;
3573         int maybe_acls;
3574         u32 rdev;
3575         int ret;
3576         bool filled = false;
3577         int first_xattr_slot;
3578
3579         ret = btrfs_fill_inode(inode, &rdev);
3580         if (!ret)
3581                 filled = true;
3582
3583         path = btrfs_alloc_path();
3584         if (!path)
3585                 return -ENOMEM;
3586
3587         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3588
3589         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3590         if (ret) {
3591                 btrfs_free_path(path);
3592                 return ret;
3593         }
3594
3595         leaf = path->nodes[0];
3596
3597         if (filled)
3598                 goto cache_index;
3599
3600         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3601                                     struct btrfs_inode_item);
3602         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3603         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3604         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3605         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3606         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3607
3608         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3609         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3610
3611         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3612         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3613
3614         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3615         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3616
3617         BTRFS_I(inode)->i_otime.tv_sec =
3618                 btrfs_timespec_sec(leaf, &inode_item->otime);
3619         BTRFS_I(inode)->i_otime.tv_nsec =
3620                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3621
3622         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3623         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3624         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3625
3626         inode_set_iversion_queried(inode,
3627                                    btrfs_inode_sequence(leaf, inode_item));
3628         inode->i_generation = BTRFS_I(inode)->generation;
3629         inode->i_rdev = 0;
3630         rdev = btrfs_inode_rdev(leaf, inode_item);
3631
3632         BTRFS_I(inode)->index_cnt = (u64)-1;
3633         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3634
3635 cache_index:
3636         /*
3637          * If we were modified in the current generation and evicted from memory
3638          * and then re-read we need to do a full sync since we don't have any
3639          * idea about which extents were modified before we were evicted from
3640          * cache.
3641          *
3642          * This is required for both inode re-read from disk and delayed inode
3643          * in delayed_nodes_tree.
3644          */
3645         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3646                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3647                         &BTRFS_I(inode)->runtime_flags);
3648
3649         /*
3650          * We don't persist the id of the transaction where an unlink operation
3651          * against the inode was last made. So here we assume the inode might
3652          * have been evicted, and therefore the exact value of last_unlink_trans
3653          * lost, and set it to last_trans to avoid metadata inconsistencies
3654          * between the inode and its parent if the inode is fsync'ed and the log
3655          * replayed. For example, in the scenario:
3656          *
3657          * touch mydir/foo
3658          * ln mydir/foo mydir/bar
3659          * sync
3660          * unlink mydir/bar
3661          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3662          * xfs_io -c fsync mydir/foo
3663          * <power failure>
3664          * mount fs, triggers fsync log replay
3665          *
3666          * We must make sure that when we fsync our inode foo we also log its
3667          * parent inode, otherwise after log replay the parent still has the
3668          * dentry with the "bar" name but our inode foo has a link count of 1
3669          * and doesn't have an inode ref with the name "bar" anymore.
3670          *
3671          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3672          * but it guarantees correctness at the expense of occasional full
3673          * transaction commits on fsync if our inode is a directory, or if our
3674          * inode is not a directory, logging its parent unnecessarily.
3675          */
3676         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3677
3678         path->slots[0]++;
3679         if (inode->i_nlink != 1 ||
3680             path->slots[0] >= btrfs_header_nritems(leaf))
3681                 goto cache_acl;
3682
3683         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3684         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3685                 goto cache_acl;
3686
3687         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3688         if (location.type == BTRFS_INODE_REF_KEY) {
3689                 struct btrfs_inode_ref *ref;
3690
3691                 ref = (struct btrfs_inode_ref *)ptr;
3692                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3693         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3694                 struct btrfs_inode_extref *extref;
3695
3696                 extref = (struct btrfs_inode_extref *)ptr;
3697                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3698                                                                      extref);
3699         }
3700 cache_acl:
3701         /*
3702          * try to precache a NULL acl entry for files that don't have
3703          * any xattrs or acls
3704          */
3705         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3706                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3707         if (first_xattr_slot != -1) {
3708                 path->slots[0] = first_xattr_slot;
3709                 ret = btrfs_load_inode_props(inode, path);
3710                 if (ret)
3711                         btrfs_err(fs_info,
3712                                   "error loading props for ino %llu (root %llu): %d",
3713                                   btrfs_ino(BTRFS_I(inode)),
3714                                   root->root_key.objectid, ret);
3715         }
3716         btrfs_free_path(path);
3717
3718         if (!maybe_acls)
3719                 cache_no_acl(inode);
3720
3721         switch (inode->i_mode & S_IFMT) {
3722         case S_IFREG:
3723                 inode->i_mapping->a_ops = &btrfs_aops;
3724                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3725                 inode->i_fop = &btrfs_file_operations;
3726                 inode->i_op = &btrfs_file_inode_operations;
3727                 break;
3728         case S_IFDIR:
3729                 inode->i_fop = &btrfs_dir_file_operations;
3730                 inode->i_op = &btrfs_dir_inode_operations;
3731                 break;
3732         case S_IFLNK:
3733                 inode->i_op = &btrfs_symlink_inode_operations;
3734                 inode_nohighmem(inode);
3735                 inode->i_mapping->a_ops = &btrfs_aops;
3736                 break;
3737         default:
3738                 inode->i_op = &btrfs_special_inode_operations;
3739                 init_special_inode(inode, inode->i_mode, rdev);
3740                 break;
3741         }
3742
3743         btrfs_sync_inode_flags_to_i_flags(inode);
3744         return 0;
3745 }
3746
3747 /*
3748  * given a leaf and an inode, copy the inode fields into the leaf
3749  */
3750 static void fill_inode_item(struct btrfs_trans_handle *trans,
3751                             struct extent_buffer *leaf,
3752                             struct btrfs_inode_item *item,
3753                             struct inode *inode)
3754 {
3755         struct btrfs_map_token token;
3756
3757         btrfs_init_map_token(&token);
3758
3759         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3760         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3761         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3762                                    &token);
3763         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3764         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3765
3766         btrfs_set_token_timespec_sec(leaf, &item->atime,
3767                                      inode->i_atime.tv_sec, &token);
3768         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3769                                       inode->i_atime.tv_nsec, &token);
3770
3771         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3772                                      inode->i_mtime.tv_sec, &token);
3773         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3774                                       inode->i_mtime.tv_nsec, &token);
3775
3776         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3777                                      inode->i_ctime.tv_sec, &token);
3778         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3779                                       inode->i_ctime.tv_nsec, &token);
3780
3781         btrfs_set_token_timespec_sec(leaf, &item->otime,
3782                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3783         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3784                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3785
3786         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3787                                      &token);
3788         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3789                                          &token);
3790         btrfs_set_token_inode_sequence(leaf, item, inode_peek_iversion(inode),
3791                                        &token);
3792         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3793         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3794         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3795         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3796 }
3797
3798 /*
3799  * copy everything in the in-memory inode into the btree.
3800  */
3801 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3802                                 struct btrfs_root *root, struct inode *inode)
3803 {
3804         struct btrfs_inode_item *inode_item;
3805         struct btrfs_path *path;
3806         struct extent_buffer *leaf;
3807         int ret;
3808
3809         path = btrfs_alloc_path();
3810         if (!path)
3811                 return -ENOMEM;
3812
3813         path->leave_spinning = 1;
3814         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3815                                  1);
3816         if (ret) {
3817                 if (ret > 0)
3818                         ret = -ENOENT;
3819                 goto failed;
3820         }
3821
3822         leaf = path->nodes[0];
3823         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3824                                     struct btrfs_inode_item);
3825
3826         fill_inode_item(trans, leaf, inode_item, inode);
3827         btrfs_mark_buffer_dirty(leaf);
3828         btrfs_set_inode_last_trans(trans, inode);
3829         ret = 0;
3830 failed:
3831         btrfs_free_path(path);
3832         return ret;
3833 }
3834
3835 /*
3836  * copy everything in the in-memory inode into the btree.
3837  */
3838 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3839                                 struct btrfs_root *root, struct inode *inode)
3840 {
3841         struct btrfs_fs_info *fs_info = root->fs_info;
3842         int ret;
3843
3844         /*
3845          * If the inode is a free space inode, we can deadlock during commit
3846          * if we put it into the delayed code.
3847          *
3848          * The data relocation inode should also be directly updated
3849          * without delay
3850          */
3851         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
3852             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3853             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
3854                 btrfs_update_root_times(trans, root);
3855
3856                 ret = btrfs_delayed_update_inode(trans, root, inode);
3857                 if (!ret)
3858                         btrfs_set_inode_last_trans(trans, inode);
3859                 return ret;
3860         }
3861
3862         return btrfs_update_inode_item(trans, root, inode);
3863 }
3864
3865 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3866                                          struct btrfs_root *root,
3867                                          struct inode *inode)
3868 {
3869         int ret;
3870
3871         ret = btrfs_update_inode(trans, root, inode);
3872         if (ret == -ENOSPC)
3873                 return btrfs_update_inode_item(trans, root, inode);
3874         return ret;
3875 }
3876
3877 /*
3878  * unlink helper that gets used here in inode.c and in the tree logging
3879  * recovery code.  It remove a link in a directory with a given name, and
3880  * also drops the back refs in the inode to the directory
3881  */
3882 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3883                                 struct btrfs_root *root,
3884                                 struct btrfs_inode *dir,
3885                                 struct btrfs_inode *inode,
3886                                 const char *name, int name_len)
3887 {
3888         struct btrfs_fs_info *fs_info = root->fs_info;
3889         struct btrfs_path *path;
3890         int ret = 0;
3891         struct extent_buffer *leaf;
3892         struct btrfs_dir_item *di;
3893         struct btrfs_key key;
3894         u64 index;
3895         u64 ino = btrfs_ino(inode);
3896         u64 dir_ino = btrfs_ino(dir);
3897
3898         path = btrfs_alloc_path();
3899         if (!path) {
3900                 ret = -ENOMEM;
3901                 goto out;
3902         }
3903
3904         path->leave_spinning = 1;
3905         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3906                                     name, name_len, -1);
3907         if (IS_ERR_OR_NULL(di)) {
3908                 ret = di ? PTR_ERR(di) : -ENOENT;
3909                 goto err;
3910         }
3911         leaf = path->nodes[0];
3912         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3913         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3914         if (ret)
3915                 goto err;
3916         btrfs_release_path(path);
3917
3918         /*
3919          * If we don't have dir index, we have to get it by looking up
3920          * the inode ref, since we get the inode ref, remove it directly,
3921          * it is unnecessary to do delayed deletion.
3922          *
3923          * But if we have dir index, needn't search inode ref to get it.
3924          * Since the inode ref is close to the inode item, it is better
3925          * that we delay to delete it, and just do this deletion when
3926          * we update the inode item.
3927          */
3928         if (inode->dir_index) {
3929                 ret = btrfs_delayed_delete_inode_ref(inode);
3930                 if (!ret) {
3931                         index = inode->dir_index;
3932                         goto skip_backref;
3933                 }
3934         }
3935
3936         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
3937                                   dir_ino, &index);
3938         if (ret) {
3939                 btrfs_info(fs_info,
3940                         "failed to delete reference to %.*s, inode %llu parent %llu",
3941                         name_len, name, ino, dir_ino);
3942                 btrfs_abort_transaction(trans, ret);
3943                 goto err;
3944         }
3945 skip_backref:
3946         ret = btrfs_delete_delayed_dir_index(trans, dir, index);
3947         if (ret) {
3948                 btrfs_abort_transaction(trans, ret);
3949                 goto err;
3950         }
3951
3952         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
3953                         dir_ino);
3954         if (ret != 0 && ret != -ENOENT) {
3955                 btrfs_abort_transaction(trans, ret);
3956                 goto err;
3957         }
3958
3959         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
3960                         index);
3961         if (ret == -ENOENT)
3962                 ret = 0;
3963         else if (ret)
3964                 btrfs_abort_transaction(trans, ret);
3965 err:
3966         btrfs_free_path(path);
3967         if (ret)
3968                 goto out;
3969
3970         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
3971         inode_inc_iversion(&inode->vfs_inode);
3972         inode_inc_iversion(&dir->vfs_inode);
3973         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
3974                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
3975         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
3976 out:
3977         return ret;
3978 }
3979
3980 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3981                        struct btrfs_root *root,
3982                        struct btrfs_inode *dir, struct btrfs_inode *inode,
3983                        const char *name, int name_len)
3984 {
3985         int ret;
3986         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
3987         if (!ret) {
3988                 drop_nlink(&inode->vfs_inode);
3989                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
3990         }
3991         return ret;
3992 }
3993
3994 /*
3995  * helper to start transaction for unlink and rmdir.
3996  *
3997  * unlink and rmdir are special in btrfs, they do not always free space, so
3998  * if we cannot make our reservations the normal way try and see if there is
3999  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4000  * allow the unlink to occur.
4001  */
4002 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4003 {
4004         struct btrfs_root *root = BTRFS_I(dir)->root;
4005
4006         /*
4007          * 1 for the possible orphan item
4008          * 1 for the dir item
4009          * 1 for the dir index
4010          * 1 for the inode ref
4011          * 1 for the inode
4012          */
4013         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4014 }
4015
4016 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4017 {
4018         struct btrfs_root *root = BTRFS_I(dir)->root;
4019         struct btrfs_trans_handle *trans;
4020         struct inode *inode = d_inode(dentry);
4021         int ret;
4022
4023         trans = __unlink_start_trans(dir);
4024         if (IS_ERR(trans))
4025                 return PTR_ERR(trans);
4026
4027         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4028                         0);
4029
4030         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4031                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4032                         dentry->d_name.len);
4033         if (ret)
4034                 goto out;
4035
4036         if (inode->i_nlink == 0) {
4037                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4038                 if (ret)
4039                         goto out;
4040         }
4041
4042 out:
4043         btrfs_end_transaction(trans);
4044         btrfs_btree_balance_dirty(root->fs_info);
4045         return ret;
4046 }
4047
4048 static int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4049                                struct inode *dir, u64 objectid,
4050                                const char *name, int name_len)
4051 {
4052         struct btrfs_root *root = BTRFS_I(dir)->root;
4053         struct btrfs_path *path;
4054         struct extent_buffer *leaf;
4055         struct btrfs_dir_item *di;
4056         struct btrfs_key key;
4057         u64 index;
4058         int ret;
4059         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4060
4061         path = btrfs_alloc_path();
4062         if (!path)
4063                 return -ENOMEM;
4064
4065         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4066                                    name, name_len, -1);
4067         if (IS_ERR_OR_NULL(di)) {
4068                 ret = di ? PTR_ERR(di) : -ENOENT;
4069                 goto out;
4070         }
4071
4072         leaf = path->nodes[0];
4073         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4074         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4075         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4076         if (ret) {
4077                 btrfs_abort_transaction(trans, ret);
4078                 goto out;
4079         }
4080         btrfs_release_path(path);
4081
4082         ret = btrfs_del_root_ref(trans, objectid, root->root_key.objectid,
4083                                  dir_ino, &index, name, name_len);
4084         if (ret < 0) {
4085                 if (ret != -ENOENT) {
4086                         btrfs_abort_transaction(trans, ret);
4087                         goto out;
4088                 }
4089                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4090                                                  name, name_len);
4091                 if (IS_ERR_OR_NULL(di)) {
4092                         if (!di)
4093                                 ret = -ENOENT;
4094                         else
4095                                 ret = PTR_ERR(di);
4096                         btrfs_abort_transaction(trans, ret);
4097                         goto out;
4098                 }
4099
4100                 leaf = path->nodes[0];
4101                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4102                 index = key.offset;
4103         }
4104         btrfs_release_path(path);
4105
4106         ret = btrfs_delete_delayed_dir_index(trans, BTRFS_I(dir), index);
4107         if (ret) {
4108                 btrfs_abort_transaction(trans, ret);
4109                 goto out;
4110         }
4111
4112         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4113         inode_inc_iversion(dir);
4114         dir->i_mtime = dir->i_ctime = current_time(dir);
4115         ret = btrfs_update_inode_fallback(trans, root, dir);
4116         if (ret)
4117                 btrfs_abort_transaction(trans, ret);
4118 out:
4119         btrfs_free_path(path);
4120         return ret;
4121 }
4122
4123 /*
4124  * Helper to check if the subvolume references other subvolumes or if it's
4125  * default.
4126  */
4127 static noinline int may_destroy_subvol(struct btrfs_root *root)
4128 {
4129         struct btrfs_fs_info *fs_info = root->fs_info;
4130         struct btrfs_path *path;
4131         struct btrfs_dir_item *di;
4132         struct btrfs_key key;
4133         u64 dir_id;
4134         int ret;
4135
4136         path = btrfs_alloc_path();
4137         if (!path)
4138                 return -ENOMEM;
4139
4140         /* Make sure this root isn't set as the default subvol */
4141         dir_id = btrfs_super_root_dir(fs_info->super_copy);
4142         di = btrfs_lookup_dir_item(NULL, fs_info->tree_root, path,
4143                                    dir_id, "default", 7, 0);
4144         if (di && !IS_ERR(di)) {
4145                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
4146                 if (key.objectid == root->root_key.objectid) {
4147                         ret = -EPERM;
4148                         btrfs_err(fs_info,
4149                                   "deleting default subvolume %llu is not allowed",
4150                                   key.objectid);
4151                         goto out;
4152                 }
4153                 btrfs_release_path(path);
4154         }
4155
4156         key.objectid = root->root_key.objectid;
4157         key.type = BTRFS_ROOT_REF_KEY;
4158         key.offset = (u64)-1;
4159
4160         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
4161         if (ret < 0)
4162                 goto out;
4163         BUG_ON(ret == 0);
4164
4165         ret = 0;
4166         if (path->slots[0] > 0) {
4167                 path->slots[0]--;
4168                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
4169                 if (key.objectid == root->root_key.objectid &&
4170                     key.type == BTRFS_ROOT_REF_KEY)
4171                         ret = -ENOTEMPTY;
4172         }
4173 out:
4174         btrfs_free_path(path);
4175         return ret;
4176 }
4177
4178 /* Delete all dentries for inodes belonging to the root */
4179 static void btrfs_prune_dentries(struct btrfs_root *root)
4180 {
4181         struct btrfs_fs_info *fs_info = root->fs_info;
4182         struct rb_node *node;
4183         struct rb_node *prev;
4184         struct btrfs_inode *entry;
4185         struct inode *inode;
4186         u64 objectid = 0;
4187
4188         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
4189                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
4190
4191         spin_lock(&root->inode_lock);
4192 again:
4193         node = root->inode_tree.rb_node;
4194         prev = NULL;
4195         while (node) {
4196                 prev = node;
4197                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4198
4199                 if (objectid < btrfs_ino(entry))
4200                         node = node->rb_left;
4201                 else if (objectid > btrfs_ino(entry))
4202                         node = node->rb_right;
4203                 else
4204                         break;
4205         }
4206         if (!node) {
4207                 while (prev) {
4208                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
4209                         if (objectid <= btrfs_ino(entry)) {
4210                                 node = prev;
4211                                 break;
4212                         }
4213                         prev = rb_next(prev);
4214                 }
4215         }
4216         while (node) {
4217                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4218                 objectid = btrfs_ino(entry) + 1;
4219                 inode = igrab(&entry->vfs_inode);
4220                 if (inode) {
4221                         spin_unlock(&root->inode_lock);
4222                         if (atomic_read(&inode->i_count) > 1)
4223                                 d_prune_aliases(inode);
4224                         /*
4225                          * btrfs_drop_inode will have it removed from the inode
4226                          * cache when its usage count hits zero.
4227                          */
4228                         iput(inode);
4229                         cond_resched();
4230                         spin_lock(&root->inode_lock);
4231                         goto again;
4232                 }
4233
4234                 if (cond_resched_lock(&root->inode_lock))
4235                         goto again;
4236
4237                 node = rb_next(node);
4238         }
4239         spin_unlock(&root->inode_lock);
4240 }
4241
4242 int btrfs_delete_subvolume(struct inode *dir, struct dentry *dentry)
4243 {
4244         struct btrfs_fs_info *fs_info = btrfs_sb(dentry->d_sb);
4245         struct btrfs_root *root = BTRFS_I(dir)->root;
4246         struct inode *inode = d_inode(dentry);
4247         struct btrfs_root *dest = BTRFS_I(inode)->root;
4248         struct btrfs_trans_handle *trans;
4249         struct btrfs_block_rsv block_rsv;
4250         u64 root_flags;
4251         int ret;
4252         int err;
4253
4254         /*
4255          * Don't allow to delete a subvolume with send in progress. This is
4256          * inside the inode lock so the error handling that has to drop the bit
4257          * again is not run concurrently.
4258          */
4259         spin_lock(&dest->root_item_lock);
4260         if (dest->send_in_progress) {
4261                 spin_unlock(&dest->root_item_lock);
4262                 btrfs_warn(fs_info,
4263                            "attempt to delete subvolume %llu during send",
4264                            dest->root_key.objectid);
4265                 return -EPERM;
4266         }
4267         root_flags = btrfs_root_flags(&dest->root_item);
4268         btrfs_set_root_flags(&dest->root_item,
4269                              root_flags | BTRFS_ROOT_SUBVOL_DEAD);
4270         spin_unlock(&dest->root_item_lock);
4271
4272         down_write(&fs_info->subvol_sem);
4273
4274         err = may_destroy_subvol(dest);
4275         if (err)
4276                 goto out_up_write;
4277
4278         btrfs_init_block_rsv(&block_rsv, BTRFS_BLOCK_RSV_TEMP);
4279         /*
4280          * One for dir inode,
4281          * two for dir entries,
4282          * two for root ref/backref.
4283          */
4284         err = btrfs_subvolume_reserve_metadata(root, &block_rsv, 5, true);
4285         if (err)
4286                 goto out_up_write;
4287
4288         trans = btrfs_start_transaction(root, 0);
4289         if (IS_ERR(trans)) {
4290                 err = PTR_ERR(trans);
4291                 goto out_release;
4292         }
4293         trans->block_rsv = &block_rsv;
4294         trans->bytes_reserved = block_rsv.size;
4295
4296         btrfs_record_snapshot_destroy(trans, BTRFS_I(dir));
4297
4298         ret = btrfs_unlink_subvol(trans, dir, dest->root_key.objectid,
4299                                   dentry->d_name.name, dentry->d_name.len);
4300         if (ret) {
4301                 err = ret;
4302                 btrfs_abort_transaction(trans, ret);
4303                 goto out_end_trans;
4304         }
4305
4306         btrfs_record_root_in_trans(trans, dest);
4307
4308         memset(&dest->root_item.drop_progress, 0,
4309                 sizeof(dest->root_item.drop_progress));
4310         dest->root_item.drop_level = 0;
4311         btrfs_set_root_refs(&dest->root_item, 0);
4312
4313         if (!test_and_set_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &dest->state)) {
4314                 ret = btrfs_insert_orphan_item(trans,
4315                                         fs_info->tree_root,
4316                                         dest->root_key.objectid);
4317                 if (ret) {
4318                         btrfs_abort_transaction(trans, ret);
4319                         err = ret;
4320                         goto out_end_trans;
4321                 }
4322         }
4323
4324         ret = btrfs_uuid_tree_remove(trans, dest->root_item.uuid,
4325                                   BTRFS_UUID_KEY_SUBVOL,
4326                                   dest->root_key.objectid);
4327         if (ret && ret != -ENOENT) {
4328                 btrfs_abort_transaction(trans, ret);
4329                 err = ret;
4330                 goto out_end_trans;
4331         }
4332         if (!btrfs_is_empty_uuid(dest->root_item.received_uuid)) {
4333                 ret = btrfs_uuid_tree_remove(trans,
4334                                           dest->root_item.received_uuid,
4335                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
4336                                           dest->root_key.objectid);
4337                 if (ret && ret != -ENOENT) {
4338                         btrfs_abort_transaction(trans, ret);
4339                         err = ret;
4340                         goto out_end_trans;
4341                 }
4342         }
4343
4344 out_end_trans:
4345         trans->block_rsv = NULL;
4346         trans->bytes_reserved = 0;
4347         ret = btrfs_end_transaction(trans);
4348         if (ret && !err)
4349                 err = ret;
4350         inode->i_flags |= S_DEAD;
4351 out_release:
4352         btrfs_subvolume_release_metadata(fs_info, &block_rsv);
4353 out_up_write:
4354         up_write(&fs_info->subvol_sem);
4355         if (err) {
4356                 spin_lock(&dest->root_item_lock);
4357                 root_flags = btrfs_root_flags(&dest->root_item);
4358                 btrfs_set_root_flags(&dest->root_item,
4359                                 root_flags & ~BTRFS_ROOT_SUBVOL_DEAD);
4360                 spin_unlock(&dest->root_item_lock);
4361         } else {
4362                 d_invalidate(dentry);
4363                 btrfs_prune_dentries(dest);
4364                 ASSERT(dest->send_in_progress == 0);
4365
4366                 /* the last ref */
4367                 if (dest->ino_cache_inode) {
4368                         iput(dest->ino_cache_inode);
4369                         dest->ino_cache_inode = NULL;
4370                 }
4371         }
4372
4373         return err;
4374 }
4375
4376 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4377 {
4378         struct inode *inode = d_inode(dentry);
4379         int err = 0;
4380         struct btrfs_root *root = BTRFS_I(dir)->root;
4381         struct btrfs_trans_handle *trans;
4382         u64 last_unlink_trans;
4383
4384         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4385                 return -ENOTEMPTY;
4386         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4387                 return btrfs_delete_subvolume(dir, dentry);
4388
4389         trans = __unlink_start_trans(dir);
4390         if (IS_ERR(trans))
4391                 return PTR_ERR(trans);
4392
4393         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4394                 err = btrfs_unlink_subvol(trans, dir,
4395                                           BTRFS_I(inode)->location.objectid,
4396                                           dentry->d_name.name,
4397                                           dentry->d_name.len);
4398                 goto out;
4399         }
4400
4401         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4402         if (err)
4403                 goto out;
4404
4405         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4406
4407         /* now the directory is empty */
4408         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4409                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4410                         dentry->d_name.len);
4411         if (!err) {
4412                 btrfs_i_size_write(BTRFS_I(inode), 0);
4413                 /*
4414                  * Propagate the last_unlink_trans value of the deleted dir to
4415                  * its parent directory. This is to prevent an unrecoverable
4416                  * log tree in the case we do something like this:
4417                  * 1) create dir foo
4418                  * 2) create snapshot under dir foo
4419                  * 3) delete the snapshot
4420                  * 4) rmdir foo
4421                  * 5) mkdir foo
4422                  * 6) fsync foo or some file inside foo
4423                  */
4424                 if (last_unlink_trans >= trans->transid)
4425                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4426         }
4427 out:
4428         btrfs_end_transaction(trans);
4429         btrfs_btree_balance_dirty(root->fs_info);
4430
4431         return err;
4432 }
4433
4434 static int truncate_space_check(struct btrfs_trans_handle *trans,
4435                                 struct btrfs_root *root,
4436                                 u64 bytes_deleted)
4437 {
4438         struct btrfs_fs_info *fs_info = root->fs_info;
4439         int ret;
4440
4441         /*
4442          * This is only used to apply pressure to the enospc system, we don't
4443          * intend to use this reservation at all.
4444          */
4445         bytes_deleted = btrfs_csum_bytes_to_leaves(fs_info, bytes_deleted);
4446         bytes_deleted *= fs_info->nodesize;
4447         ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
4448                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4449         if (!ret) {
4450                 trace_btrfs_space_reservation(fs_info, "transaction",
4451                                               trans->transid,
4452                                               bytes_deleted, 1);
4453                 trans->bytes_reserved += bytes_deleted;
4454         }
4455         return ret;
4456
4457 }
4458
4459 /*
4460  * Return this if we need to call truncate_block for the last bit of the
4461  * truncate.
4462  */
4463 #define NEED_TRUNCATE_BLOCK 1
4464
4465 /*
4466  * this can truncate away extent items, csum items and directory items.
4467  * It starts at a high offset and removes keys until it can't find
4468  * any higher than new_size
4469  *
4470  * csum items that cross the new i_size are truncated to the new size
4471  * as well.
4472  *
4473  * min_type is the minimum key type to truncate down to.  If set to 0, this
4474  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4475  */
4476 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4477                                struct btrfs_root *root,
4478                                struct inode *inode,
4479                                u64 new_size, u32 min_type)
4480 {
4481         struct btrfs_fs_info *fs_info = root->fs_info;
4482         struct btrfs_path *path;
4483         struct extent_buffer *leaf;
4484         struct btrfs_file_extent_item *fi;
4485         struct btrfs_key key;
4486         struct btrfs_key found_key;
4487         u64 extent_start = 0;
4488         u64 extent_num_bytes = 0;
4489         u64 extent_offset = 0;
4490         u64 item_end = 0;
4491         u64 last_size = new_size;
4492         u32 found_type = (u8)-1;
4493         int found_extent;
4494         int del_item;
4495         int pending_del_nr = 0;
4496         int pending_del_slot = 0;
4497         int extent_type = -1;
4498         int ret;
4499         u64 ino = btrfs_ino(BTRFS_I(inode));
4500         u64 bytes_deleted = 0;
4501         bool be_nice = false;
4502         bool should_throttle = false;
4503         bool should_end = false;
4504
4505         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4506
4507         /*
4508          * for non-free space inodes and ref cows, we want to back off from
4509          * time to time
4510          */
4511         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4512             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4513                 be_nice = true;
4514
4515         path = btrfs_alloc_path();
4516         if (!path)
4517                 return -ENOMEM;
4518         path->reada = READA_BACK;
4519
4520         /*
4521          * We want to drop from the next block forward in case this new size is
4522          * not block aligned since we will be keeping the last block of the
4523          * extent just the way it is.
4524          */
4525         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4526             root == fs_info->tree_root)
4527                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4528                                         fs_info->sectorsize),
4529                                         (u64)-1, 0);
4530
4531         /*
4532          * This function is also used to drop the items in the log tree before
4533          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4534          * it is used to drop the loged items. So we shouldn't kill the delayed
4535          * items.
4536          */
4537         if (min_type == 0 && root == BTRFS_I(inode)->root)
4538                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4539
4540         key.objectid = ino;
4541         key.offset = (u64)-1;
4542         key.type = (u8)-1;
4543
4544 search_again:
4545         /*
4546          * with a 16K leaf size and 128MB extents, you can actually queue
4547          * up a huge file in a single leaf.  Most of the time that
4548          * bytes_deleted is > 0, it will be huge by the time we get here
4549          */
4550         if (be_nice && bytes_deleted > SZ_32M &&
4551             btrfs_should_end_transaction(trans)) {
4552                 ret = -EAGAIN;
4553                 goto out;
4554         }
4555
4556         path->leave_spinning = 1;
4557         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4558         if (ret < 0)
4559                 goto out;
4560
4561         if (ret > 0) {
4562                 ret = 0;
4563                 /* there are no items in the tree for us to truncate, we're
4564                  * done
4565                  */
4566                 if (path->slots[0] == 0)
4567                         goto out;
4568                 path->slots[0]--;
4569         }
4570
4571         while (1) {
4572                 fi = NULL;
4573                 leaf = path->nodes[0];
4574                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4575                 found_type = found_key.type;
4576
4577                 if (found_key.objectid != ino)
4578                         break;
4579
4580                 if (found_type < min_type)
4581                         break;
4582
4583                 item_end = found_key.offset;
4584                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4585                         fi = btrfs_item_ptr(leaf, path->slots[0],
4586                                             struct btrfs_file_extent_item);
4587                         extent_type = btrfs_file_extent_type(leaf, fi);
4588                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4589                                 item_end +=
4590                                     btrfs_file_extent_num_bytes(leaf, fi);
4591
4592                                 trace_btrfs_truncate_show_fi_regular(
4593                                         BTRFS_I(inode), leaf, fi,
4594                                         found_key.offset);
4595                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4596                                 item_end += btrfs_file_extent_ram_bytes(leaf,
4597                                                                         fi);
4598
4599                                 trace_btrfs_truncate_show_fi_inline(
4600                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4601                                         found_key.offset);
4602                         }
4603                         item_end--;
4604                 }
4605                 if (found_type > min_type) {
4606                         del_item = 1;
4607                 } else {
4608                         if (item_end < new_size)
4609                                 break;
4610                         if (found_key.offset >= new_size)
4611                                 del_item = 1;
4612                         else
4613                                 del_item = 0;
4614                 }
4615                 found_extent = 0;
4616                 /* FIXME, shrink the extent if the ref count is only 1 */
4617                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4618                         goto delete;
4619
4620                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4621                         u64 num_dec;
4622                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4623                         if (!del_item) {
4624                                 u64 orig_num_bytes =
4625                                         btrfs_file_extent_num_bytes(leaf, fi);
4626                                 extent_num_bytes = ALIGN(new_size -
4627                                                 found_key.offset,
4628                                                 fs_info->sectorsize);
4629                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4630                                                          extent_num_bytes);
4631                                 num_dec = (orig_num_bytes -
4632                                            extent_num_bytes);
4633                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4634                                              &root->state) &&
4635                                     extent_start != 0)
4636                                         inode_sub_bytes(inode, num_dec);
4637                                 btrfs_mark_buffer_dirty(leaf);
4638                         } else {
4639                                 extent_num_bytes =
4640                                         btrfs_file_extent_disk_num_bytes(leaf,
4641                                                                          fi);
4642                                 extent_offset = found_key.offset -
4643                                         btrfs_file_extent_offset(leaf, fi);
4644
4645                                 /* FIXME blocksize != 4096 */
4646                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4647                                 if (extent_start != 0) {
4648                                         found_extent = 1;
4649                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4650                                                      &root->state))
4651                                                 inode_sub_bytes(inode, num_dec);
4652                                 }
4653                         }
4654                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4655                         /*
4656                          * we can't truncate inline items that have had
4657                          * special encodings
4658                          */
4659                         if (!del_item &&
4660                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4661                             btrfs_file_extent_other_encoding(leaf, fi) == 0 &&
4662                             btrfs_file_extent_compression(leaf, fi) == 0) {
4663                                 u32 size = (u32)(new_size - found_key.offset);
4664
4665                                 btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4666                                 size = btrfs_file_extent_calc_inline_size(size);
4667                                 btrfs_truncate_item(root->fs_info, path, size, 1);
4668                         } else if (!del_item) {
4669                                 /*
4670                                  * We have to bail so the last_size is set to
4671                                  * just before this extent.
4672                                  */
4673                                 ret = NEED_TRUNCATE_BLOCK;
4674                                 break;
4675                         }
4676
4677                         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4678                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4679                 }
4680 delete:
4681                 if (del_item)
4682                         last_size = found_key.offset;
4683                 else
4684                         last_size = new_size;
4685                 if (del_item) {
4686                         if (!pending_del_nr) {
4687                                 /* no pending yet, add ourselves */
4688                                 pending_del_slot = path->slots[0];
4689                                 pending_del_nr = 1;
4690                         } else if (pending_del_nr &&
4691                                    path->slots[0] + 1 == pending_del_slot) {
4692                                 /* hop on the pending chunk */
4693                                 pending_del_nr++;
4694                                 pending_del_slot = path->slots[0];
4695                         } else {
4696                                 BUG();
4697                         }
4698                 } else {
4699                         break;
4700                 }
4701                 should_throttle = false;
4702
4703                 if (found_extent &&
4704                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4705                      root == fs_info->tree_root)) {
4706                         btrfs_set_path_blocking(path);
4707                         bytes_deleted += extent_num_bytes;
4708                         ret = btrfs_free_extent(trans, root, extent_start,
4709                                                 extent_num_bytes, 0,
4710                                                 btrfs_header_owner(leaf),
4711                                                 ino, extent_offset);
4712                         if (ret) {
4713                                 btrfs_abort_transaction(trans, ret);
4714                                 break;
4715                         }
4716                         if (btrfs_should_throttle_delayed_refs(trans, fs_info))
4717                                 btrfs_async_run_delayed_refs(fs_info,
4718                                         trans->delayed_ref_updates * 2,
4719                                         trans->transid, 0);
4720                         if (be_nice) {
4721                                 if (truncate_space_check(trans, root,
4722                                                          extent_num_bytes)) {
4723                                         should_end = true;
4724                                 }
4725                                 if (btrfs_should_throttle_delayed_refs(trans,
4726                                                                        fs_info))
4727                                         should_throttle = true;
4728                         }
4729                 }
4730
4731                 if (found_type == BTRFS_INODE_ITEM_KEY)
4732                         break;
4733
4734                 if (path->slots[0] == 0 ||
4735                     path->slots[0] != pending_del_slot ||
4736                     should_throttle || should_end) {
4737                         if (pending_del_nr) {
4738                                 ret = btrfs_del_items(trans, root, path,
4739                                                 pending_del_slot,
4740                                                 pending_del_nr);
4741                                 if (ret) {
4742                                         btrfs_abort_transaction(trans, ret);
4743                                         break;
4744                                 }
4745                                 pending_del_nr = 0;
4746                         }
4747                         btrfs_release_path(path);
4748                         if (should_throttle) {
4749                                 unsigned long updates = trans->delayed_ref_updates;
4750                                 if (updates) {
4751                                         trans->delayed_ref_updates = 0;
4752                                         ret = btrfs_run_delayed_refs(trans,
4753                                                                    updates * 2);
4754                                         if (ret)
4755                                                 break;
4756                                 }
4757                         }
4758                         /*
4759                          * if we failed to refill our space rsv, bail out
4760                          * and let the transaction restart
4761                          */
4762                         if (should_end) {
4763                                 ret = -EAGAIN;
4764                                 break;
4765                         }
4766                         goto search_again;
4767                 } else {
4768                         path->slots[0]--;
4769                 }
4770         }
4771 out:
4772         if (ret >= 0 && pending_del_nr) {
4773                 int err;
4774
4775                 err = btrfs_del_items(trans, root, path, pending_del_slot,
4776                                       pending_del_nr);
4777                 if (err) {
4778                         btrfs_abort_transaction(trans, err);
4779                         ret = err;
4780                 }
4781         }
4782         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4783                 ASSERT(last_size >= new_size);
4784                 if (!ret && last_size > new_size)
4785                         last_size = new_size;
4786                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4787         }
4788
4789         btrfs_free_path(path);
4790
4791         if (be_nice && bytes_deleted > SZ_32M && (ret >= 0 || ret == -EAGAIN)) {
4792                 unsigned long updates = trans->delayed_ref_updates;
4793                 int err;
4794
4795                 if (updates) {
4796                         trans->delayed_ref_updates = 0;
4797                         err = btrfs_run_delayed_refs(trans, updates * 2);
4798                         if (err)
4799                                 ret = err;
4800                 }
4801         }
4802         return ret;
4803 }
4804
4805 /*
4806  * btrfs_truncate_block - read, zero a chunk and write a block
4807  * @inode - inode that we're zeroing
4808  * @from - the offset to start zeroing
4809  * @len - the length to zero, 0 to zero the entire range respective to the
4810  *      offset
4811  * @front - zero up to the offset instead of from the offset on
4812  *
4813  * This will find the block for the "from" offset and cow the block and zero the
4814  * part we want to zero.  This is used with truncate and hole punching.
4815  */
4816 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4817                         int front)
4818 {
4819         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4820         struct address_space *mapping = inode->i_mapping;
4821         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4822         struct btrfs_ordered_extent *ordered;
4823         struct extent_state *cached_state = NULL;
4824         struct extent_changeset *data_reserved = NULL;
4825         char *kaddr;
4826         u32 blocksize = fs_info->sectorsize;
4827         pgoff_t index = from >> PAGE_SHIFT;
4828         unsigned offset = from & (blocksize - 1);
4829         struct page *page;
4830         gfp_t mask = btrfs_alloc_write_mask(mapping);
4831         int ret = 0;
4832         u64 block_start;
4833         u64 block_end;
4834
4835         if (IS_ALIGNED(offset, blocksize) &&
4836             (!len || IS_ALIGNED(len, blocksize)))
4837                 goto out;
4838
4839         block_start = round_down(from, blocksize);
4840         block_end = block_start + blocksize - 1;
4841
4842         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4843                                            block_start, blocksize);
4844         if (ret)
4845                 goto out;
4846
4847 again:
4848         page = find_or_create_page(mapping, index, mask);
4849         if (!page) {
4850                 btrfs_delalloc_release_space(inode, data_reserved,
4851                                              block_start, blocksize, true);
4852                 btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize, true);
4853                 ret = -ENOMEM;
4854                 goto out;
4855         }
4856
4857         if (!PageUptodate(page)) {
4858                 ret = btrfs_readpage(NULL, page);
4859                 lock_page(page);
4860                 if (page->mapping != mapping) {
4861                         unlock_page(page);
4862                         put_page(page);
4863                         goto again;
4864                 }
4865                 if (!PageUptodate(page)) {
4866                         ret = -EIO;
4867                         goto out_unlock;
4868                 }
4869         }
4870         wait_on_page_writeback(page);
4871
4872         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4873         set_page_extent_mapped(page);
4874
4875         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4876         if (ordered) {
4877                 unlock_extent_cached(io_tree, block_start, block_end,
4878                                      &cached_state);
4879                 unlock_page(page);
4880                 put_page(page);
4881                 btrfs_start_ordered_extent(inode, ordered, 1);
4882                 btrfs_put_ordered_extent(ordered);
4883                 goto again;
4884         }
4885
4886         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4887                           EXTENT_DIRTY | EXTENT_DELALLOC |
4888                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4889                           0, 0, &cached_state);
4890
4891         ret = btrfs_set_extent_delalloc(inode, block_start, block_end, 0,
4892                                         &cached_state, 0);
4893         if (ret) {
4894                 unlock_extent_cached(io_tree, block_start, block_end,
4895                                      &cached_state);
4896                 goto out_unlock;
4897         }
4898
4899         if (offset != blocksize) {
4900                 if (!len)
4901                         len = blocksize - offset;
4902                 kaddr = kmap(page);
4903                 if (front)
4904                         memset(kaddr + (block_start - page_offset(page)),
4905                                 0, offset);
4906                 else
4907                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4908                                 0, len);
4909                 flush_dcache_page(page);
4910                 kunmap(page);
4911         }
4912         ClearPageChecked(page);
4913         set_page_dirty(page);
4914         unlock_extent_cached(io_tree, block_start, block_end, &cached_state);
4915
4916 out_unlock:
4917         if (ret)
4918                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
4919                                              blocksize, true);
4920         btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize, (ret != 0));
4921         unlock_page(page);
4922         put_page(page);
4923 out:
4924         extent_changeset_free(data_reserved);
4925         return ret;
4926 }
4927
4928 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4929                              u64 offset, u64 len)
4930 {
4931         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4932         struct btrfs_trans_handle *trans;
4933         int ret;
4934
4935         /*
4936          * Still need to make sure the inode looks like it's been updated so
4937          * that any holes get logged if we fsync.
4938          */
4939         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
4940                 BTRFS_I(inode)->last_trans = fs_info->generation;
4941                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4942                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4943                 return 0;
4944         }
4945
4946         /*
4947          * 1 - for the one we're dropping
4948          * 1 - for the one we're adding
4949          * 1 - for updating the inode.
4950          */
4951         trans = btrfs_start_transaction(root, 3);
4952         if (IS_ERR(trans))
4953                 return PTR_ERR(trans);
4954
4955         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4956         if (ret) {
4957                 btrfs_abort_transaction(trans, ret);
4958                 btrfs_end_transaction(trans);
4959                 return ret;
4960         }
4961
4962         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
4963                         offset, 0, 0, len, 0, len, 0, 0, 0);
4964         if (ret)
4965                 btrfs_abort_transaction(trans, ret);
4966         else
4967                 btrfs_update_inode(trans, root, inode);
4968         btrfs_end_transaction(trans);
4969         return ret;
4970 }
4971
4972 /*
4973  * This function puts in dummy file extents for the area we're creating a hole
4974  * for.  So if we are truncating this file to a larger size we need to insert
4975  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4976  * the range between oldsize and size
4977  */
4978 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4979 {
4980         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4981         struct btrfs_root *root = BTRFS_I(inode)->root;
4982         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4983         struct extent_map *em = NULL;
4984         struct extent_state *cached_state = NULL;
4985         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4986         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
4987         u64 block_end = ALIGN(size, fs_info->sectorsize);
4988         u64 last_byte;
4989         u64 cur_offset;
4990         u64 hole_size;
4991         int err = 0;
4992
4993         /*
4994          * If our size started in the middle of a block we need to zero out the
4995          * rest of the block before we expand the i_size, otherwise we could
4996          * expose stale data.
4997          */
4998         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4999         if (err)
5000                 return err;
5001
5002         if (size <= hole_start)
5003                 return 0;
5004
5005         while (1) {
5006                 struct btrfs_ordered_extent *ordered;
5007
5008                 lock_extent_bits(io_tree, hole_start, block_end - 1,
5009                                  &cached_state);
5010                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), hole_start,
5011                                                      block_end - hole_start);
5012                 if (!ordered)
5013                         break;
5014                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
5015                                      &cached_state);
5016                 btrfs_start_ordered_extent(inode, ordered, 1);
5017                 btrfs_put_ordered_extent(ordered);
5018         }
5019
5020         cur_offset = hole_start;
5021         while (1) {
5022                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
5023                                 block_end - cur_offset, 0);
5024                 if (IS_ERR(em)) {
5025                         err = PTR_ERR(em);
5026                         em = NULL;
5027                         break;
5028                 }
5029                 last_byte = min(extent_map_end(em), block_end);
5030                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
5031                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
5032                         struct extent_map *hole_em;
5033                         hole_size = last_byte - cur_offset;
5034
5035                         err = maybe_insert_hole(root, inode, cur_offset,
5036                                                 hole_size);
5037                         if (err)
5038                                 break;
5039                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
5040                                                 cur_offset + hole_size - 1, 0);
5041                         hole_em = alloc_extent_map();
5042                         if (!hole_em) {
5043                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
5044                                         &BTRFS_I(inode)->runtime_flags);
5045                                 goto next;
5046                         }
5047                         hole_em->start = cur_offset;
5048                         hole_em->len = hole_size;
5049                         hole_em->orig_start = cur_offset;
5050
5051                         hole_em->block_start = EXTENT_MAP_HOLE;
5052                         hole_em->block_len = 0;
5053                         hole_em->orig_block_len = 0;
5054                         hole_em->ram_bytes = hole_size;
5055                         hole_em->bdev = fs_info->fs_devices->latest_bdev;
5056                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
5057                         hole_em->generation = fs_info->generation;
5058
5059                         while (1) {
5060                                 write_lock(&em_tree->lock);
5061                                 err = add_extent_mapping(em_tree, hole_em, 1);
5062                                 write_unlock(&em_tree->lock);
5063                                 if (err != -EEXIST)
5064                                         break;
5065                                 btrfs_drop_extent_cache(BTRFS_I(inode),
5066                                                         cur_offset,
5067                                                         cur_offset +
5068                                                         hole_size - 1, 0);
5069                         }
5070                         free_extent_map(hole_em);
5071                 }
5072 next:
5073                 free_extent_map(em);
5074                 em = NULL;
5075                 cur_offset = last_byte;
5076                 if (cur_offset >= block_end)
5077                         break;
5078         }
5079         free_extent_map(em);
5080         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state);
5081         return err;
5082 }
5083
5084 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
5085 {
5086         struct btrfs_root *root = BTRFS_I(inode)->root;
5087         struct btrfs_trans_handle *trans;
5088         loff_t oldsize = i_size_read(inode);
5089         loff_t newsize = attr->ia_size;
5090         int mask = attr->ia_valid;
5091         int ret;
5092
5093         /*
5094          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5095          * special case where we need to update the times despite not having
5096          * these flags set.  For all other operations the VFS set these flags
5097          * explicitly if it wants a timestamp update.
5098          */
5099         if (newsize != oldsize) {
5100                 inode_inc_iversion(inode);
5101                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5102                         inode->i_ctime = inode->i_mtime =
5103                                 current_time(inode);
5104         }
5105
5106         if (newsize > oldsize) {
5107                 /*
5108                  * Don't do an expanding truncate while snapshotting is ongoing.
5109                  * This is to ensure the snapshot captures a fully consistent
5110                  * state of this file - if the snapshot captures this expanding
5111                  * truncation, it must capture all writes that happened before
5112                  * this truncation.
5113                  */
5114                 btrfs_wait_for_snapshot_creation(root);
5115                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5116                 if (ret) {
5117                         btrfs_end_write_no_snapshotting(root);
5118                         return ret;
5119                 }
5120
5121                 trans = btrfs_start_transaction(root, 1);
5122                 if (IS_ERR(trans)) {
5123                         btrfs_end_write_no_snapshotting(root);
5124                         return PTR_ERR(trans);
5125                 }
5126
5127                 i_size_write(inode, newsize);
5128                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5129                 pagecache_isize_extended(inode, oldsize, newsize);
5130                 ret = btrfs_update_inode(trans, root, inode);
5131                 btrfs_end_write_no_snapshotting(root);
5132                 btrfs_end_transaction(trans);
5133         } else {
5134
5135                 /*
5136                  * We're truncating a file that used to have good data down to
5137                  * zero. Make sure it gets into the ordered flush list so that
5138                  * any new writes get down to disk quickly.
5139                  */
5140                 if (newsize == 0)
5141                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5142                                 &BTRFS_I(inode)->runtime_flags);
5143
5144                 truncate_setsize(inode, newsize);
5145
5146                 /* Disable nonlocked read DIO to avoid the end less truncate */
5147                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
5148                 inode_dio_wait(inode);
5149                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
5150
5151                 ret = btrfs_truncate(inode, newsize == oldsize);
5152                 if (ret && inode->i_nlink) {
5153                         int err;
5154
5155                         /*
5156                          * Truncate failed, so fix up the in-memory size. We
5157                          * adjusted disk_i_size down as we removed extents, so
5158                          * wait for disk_i_size to be stable and then update the
5159                          * in-memory size to match.
5160                          */
5161                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5162                         if (err)
5163                                 return err;
5164                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5165                 }
5166         }
5167
5168         return ret;
5169 }
5170
5171 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5172 {
5173         struct inode *inode = d_inode(dentry);
5174         struct btrfs_root *root = BTRFS_I(inode)->root;
5175         int err;
5176
5177         if (btrfs_root_readonly(root))
5178                 return -EROFS;
5179
5180         err = setattr_prepare(dentry, attr);
5181         if (err)
5182                 return err;
5183
5184         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5185                 err = btrfs_setsize(inode, attr);
5186                 if (err)
5187                         return err;
5188         }
5189
5190         if (attr->ia_valid) {
5191                 setattr_copy(inode, attr);
5192                 inode_inc_iversion(inode);
5193                 err = btrfs_dirty_inode(inode);
5194
5195                 if (!err && attr->ia_valid & ATTR_MODE)
5196                         err = posix_acl_chmod(inode, inode->i_mode);
5197         }
5198
5199         return err;
5200 }
5201
5202 /*
5203  * While truncating the inode pages during eviction, we get the VFS calling
5204  * btrfs_invalidatepage() against each page of the inode. This is slow because
5205  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5206  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5207  * extent_state structures over and over, wasting lots of time.
5208  *
5209  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5210  * those expensive operations on a per page basis and do only the ordered io
5211  * finishing, while we release here the extent_map and extent_state structures,
5212  * without the excessive merging and splitting.
5213  */
5214 static void evict_inode_truncate_pages(struct inode *inode)
5215 {
5216         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5217         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5218         struct rb_node *node;
5219
5220         ASSERT(inode->i_state & I_FREEING);
5221         truncate_inode_pages_final(&inode->i_data);
5222
5223         write_lock(&map_tree->lock);
5224         while (!RB_EMPTY_ROOT(&map_tree->map.rb_root)) {
5225                 struct extent_map *em;
5226
5227                 node = rb_first_cached(&map_tree->map);
5228                 em = rb_entry(node, struct extent_map, rb_node);
5229                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5230                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5231                 remove_extent_mapping(map_tree, em);
5232                 free_extent_map(em);
5233                 if (need_resched()) {
5234                         write_unlock(&map_tree->lock);
5235                         cond_resched();
5236                         write_lock(&map_tree->lock);
5237                 }
5238         }
5239         write_unlock(&map_tree->lock);
5240
5241         /*
5242          * Keep looping until we have no more ranges in the io tree.
5243          * We can have ongoing bios started by readpages (called from readahead)
5244          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5245          * still in progress (unlocked the pages in the bio but did not yet
5246          * unlocked the ranges in the io tree). Therefore this means some
5247          * ranges can still be locked and eviction started because before
5248          * submitting those bios, which are executed by a separate task (work
5249          * queue kthread), inode references (inode->i_count) were not taken
5250          * (which would be dropped in the end io callback of each bio).
5251          * Therefore here we effectively end up waiting for those bios and
5252          * anyone else holding locked ranges without having bumped the inode's
5253          * reference count - if we don't do it, when they access the inode's
5254          * io_tree to unlock a range it may be too late, leading to an
5255          * use-after-free issue.
5256          */
5257         spin_lock(&io_tree->lock);
5258         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5259                 struct extent_state *state;
5260                 struct extent_state *cached_state = NULL;
5261                 u64 start;
5262                 u64 end;
5263
5264                 node = rb_first(&io_tree->state);
5265                 state = rb_entry(node, struct extent_state, rb_node);
5266                 start = state->start;
5267                 end = state->end;
5268                 spin_unlock(&io_tree->lock);
5269
5270                 lock_extent_bits(io_tree, start, end, &cached_state);
5271
5272                 /*
5273                  * If still has DELALLOC flag, the extent didn't reach disk,
5274                  * and its reserved space won't be freed by delayed_ref.
5275                  * So we need to free its reserved space here.
5276                  * (Refer to comment in btrfs_invalidatepage, case 2)
5277                  *
5278                  * Note, end is the bytenr of last byte, so we need + 1 here.
5279                  */
5280                 if (state->state & EXTENT_DELALLOC)
5281                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
5282
5283                 clear_extent_bit(io_tree, start, end,
5284                                  EXTENT_LOCKED | EXTENT_DIRTY |
5285                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5286                                  EXTENT_DEFRAG, 1, 1, &cached_state);
5287
5288                 cond_resched();
5289                 spin_lock(&io_tree->lock);
5290         }
5291         spin_unlock(&io_tree->lock);
5292 }
5293
5294 static struct btrfs_trans_handle *evict_refill_and_join(struct btrfs_root *root,
5295                                                         struct btrfs_block_rsv *rsv)
5296 {
5297         struct btrfs_fs_info *fs_info = root->fs_info;
5298         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5299         int failures = 0;
5300
5301         for (;;) {
5302                 struct btrfs_trans_handle *trans;
5303                 int ret;
5304
5305                 ret = btrfs_block_rsv_refill(root, rsv, rsv->size,
5306                                              BTRFS_RESERVE_FLUSH_LIMIT);
5307
5308                 if (ret && ++failures > 2) {
5309                         btrfs_warn(fs_info,
5310                                    "could not allocate space for a delete; will truncate on mount");
5311                         return ERR_PTR(-ENOSPC);
5312                 }
5313
5314                 trans = btrfs_join_transaction(root);
5315                 if (IS_ERR(trans) || !ret)
5316                         return trans;
5317
5318                 /*
5319                  * Try to steal from the global reserve if there is space for
5320                  * it.
5321                  */
5322                 if (!btrfs_check_space_for_delayed_refs(trans, fs_info) &&
5323                     !btrfs_block_rsv_migrate(global_rsv, rsv, rsv->size, false))
5324                         return trans;
5325
5326                 /* If not, commit and try again. */
5327                 ret = btrfs_commit_transaction(trans);
5328                 if (ret)
5329                         return ERR_PTR(ret);
5330         }
5331 }
5332
5333 void btrfs_evict_inode(struct inode *inode)
5334 {
5335         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5336         struct btrfs_trans_handle *trans;
5337         struct btrfs_root *root = BTRFS_I(inode)->root;
5338         struct btrfs_block_rsv *rsv;
5339         int ret;
5340
5341         trace_btrfs_inode_evict(inode);
5342
5343         if (!root) {
5344                 clear_inode(inode);
5345                 return;
5346         }
5347
5348         evict_inode_truncate_pages(inode);
5349
5350         if (inode->i_nlink &&
5351             ((btrfs_root_refs(&root->root_item) != 0 &&
5352               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5353              btrfs_is_free_space_inode(BTRFS_I(inode))))
5354                 goto no_delete;
5355
5356         if (is_bad_inode(inode))
5357                 goto no_delete;
5358
5359         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5360
5361         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
5362                 goto no_delete;
5363
5364         if (inode->i_nlink > 0) {
5365                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5366                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5367                 goto no_delete;
5368         }
5369
5370         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5371         if (ret)
5372                 goto no_delete;
5373
5374         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5375         if (!rsv)
5376                 goto no_delete;
5377         rsv->size = btrfs_calc_trunc_metadata_size(fs_info, 1);
5378         rsv->failfast = 1;
5379
5380         btrfs_i_size_write(BTRFS_I(inode), 0);
5381
5382         while (1) {
5383                 trans = evict_refill_and_join(root, rsv);
5384                 if (IS_ERR(trans))
5385                         goto free_rsv;
5386
5387                 trans->block_rsv = rsv;
5388
5389                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5390                 trans->block_rsv = &fs_info->trans_block_rsv;
5391                 btrfs_end_transaction(trans);
5392                 btrfs_btree_balance_dirty(fs_info);
5393                 if (ret && ret != -ENOSPC && ret != -EAGAIN)
5394                         goto free_rsv;
5395                 else if (!ret)
5396                         break;
5397         }
5398
5399         /*
5400          * Errors here aren't a big deal, it just means we leave orphan items in
5401          * the tree. They will be cleaned up on the next mount. If the inode
5402          * number gets reused, cleanup deletes the orphan item without doing
5403          * anything, and unlink reuses the existing orphan item.
5404          *
5405          * If it turns out that we are dropping too many of these, we might want
5406          * to add a mechanism for retrying these after a commit.
5407          */
5408         trans = evict_refill_and_join(root, rsv);
5409         if (!IS_ERR(trans)) {
5410                 trans->block_rsv = rsv;
5411                 btrfs_orphan_del(trans, BTRFS_I(inode));
5412                 trans->block_rsv = &fs_info->trans_block_rsv;
5413                 btrfs_end_transaction(trans);
5414         }
5415
5416         if (!(root == fs_info->tree_root ||
5417               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5418                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5419
5420 free_rsv:
5421         btrfs_free_block_rsv(fs_info, rsv);
5422 no_delete:
5423         /*
5424          * If we didn't successfully delete, the orphan item will still be in
5425          * the tree and we'll retry on the next mount. Again, we might also want
5426          * to retry these periodically in the future.
5427          */
5428         btrfs_remove_delayed_node(BTRFS_I(inode));
5429         clear_inode(inode);
5430 }
5431
5432 /*
5433  * this returns the key found in the dir entry in the location pointer.
5434  * If no dir entries were found, returns -ENOENT.
5435  * If found a corrupted location in dir entry, returns -EUCLEAN.
5436  */
5437 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5438                                struct btrfs_key *location)
5439 {
5440         const char *name = dentry->d_name.name;
5441         int namelen = dentry->d_name.len;
5442         struct btrfs_dir_item *di;
5443         struct btrfs_path *path;
5444         struct btrfs_root *root = BTRFS_I(dir)->root;
5445         int ret = 0;
5446
5447         path = btrfs_alloc_path();
5448         if (!path)
5449                 return -ENOMEM;
5450
5451         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5452                         name, namelen, 0);
5453         if (IS_ERR_OR_NULL(di)) {
5454                 ret = di ? PTR_ERR(di) : -ENOENT;
5455                 goto out;
5456         }
5457
5458         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5459         if (location->type != BTRFS_INODE_ITEM_KEY &&
5460             location->type != BTRFS_ROOT_ITEM_KEY) {
5461                 ret = -EUCLEAN;
5462                 btrfs_warn(root->fs_info,
5463 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5464                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5465                            location->objectid, location->type, location->offset);
5466         }
5467 out:
5468         btrfs_free_path(path);
5469         return ret;
5470 }
5471
5472 /*
5473  * when we hit a tree root in a directory, the btrfs part of the inode
5474  * needs to be changed to reflect the root directory of the tree root.  This
5475  * is kind of like crossing a mount point.
5476  */
5477 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5478                                     struct inode *dir,
5479                                     struct dentry *dentry,
5480                                     struct btrfs_key *location,
5481                                     struct btrfs_root **sub_root)
5482 {
5483         struct btrfs_path *path;
5484         struct btrfs_root *new_root;
5485         struct btrfs_root_ref *ref;
5486         struct extent_buffer *leaf;
5487         struct btrfs_key key;
5488         int ret;
5489         int err = 0;
5490
5491         path = btrfs_alloc_path();
5492         if (!path) {
5493                 err = -ENOMEM;
5494                 goto out;
5495         }
5496
5497         err = -ENOENT;
5498         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5499         key.type = BTRFS_ROOT_REF_KEY;
5500         key.offset = location->objectid;
5501
5502         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5503         if (ret) {
5504                 if (ret < 0)
5505                         err = ret;
5506                 goto out;
5507         }
5508
5509         leaf = path->nodes[0];
5510         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5511         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5512             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5513                 goto out;
5514
5515         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5516                                    (unsigned long)(ref + 1),
5517                                    dentry->d_name.len);
5518         if (ret)
5519                 goto out;
5520
5521         btrfs_release_path(path);
5522
5523         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5524         if (IS_ERR(new_root)) {
5525                 err = PTR_ERR(new_root);
5526                 goto out;
5527         }
5528
5529         *sub_root = new_root;
5530         location->objectid = btrfs_root_dirid(&new_root->root_item);
5531         location->type = BTRFS_INODE_ITEM_KEY;
5532         location->offset = 0;
5533         err = 0;
5534 out:
5535         btrfs_free_path(path);
5536         return err;
5537 }
5538
5539 static void inode_tree_add(struct inode *inode)
5540 {
5541         struct btrfs_root *root = BTRFS_I(inode)->root;
5542         struct btrfs_inode *entry;
5543         struct rb_node **p;
5544         struct rb_node *parent;
5545         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5546         u64 ino = btrfs_ino(BTRFS_I(inode));
5547
5548         if (inode_unhashed(inode))
5549                 return;
5550         parent = NULL;
5551         spin_lock(&root->inode_lock);
5552         p = &root->inode_tree.rb_node;
5553         while (*p) {
5554                 parent = *p;
5555                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5556
5557                 if (ino < btrfs_ino(entry))
5558                         p = &parent->rb_left;
5559                 else if (ino > btrfs_ino(entry))
5560                         p = &parent->rb_right;
5561                 else {
5562                         WARN_ON(!(entry->vfs_inode.i_state &
5563                                   (I_WILL_FREE | I_FREEING)));
5564                         rb_replace_node(parent, new, &root->inode_tree);
5565                         RB_CLEAR_NODE(parent);
5566                         spin_unlock(&root->inode_lock);
5567                         return;
5568                 }
5569         }
5570         rb_link_node(new, parent, p);
5571         rb_insert_color(new, &root->inode_tree);
5572         spin_unlock(&root->inode_lock);
5573 }
5574
5575 static void inode_tree_del(struct inode *inode)
5576 {
5577         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5578         struct btrfs_root *root = BTRFS_I(inode)->root;
5579         int empty = 0;
5580
5581         spin_lock(&root->inode_lock);
5582         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5583                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5584                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5585                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5586         }
5587         spin_unlock(&root->inode_lock);
5588
5589         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5590                 synchronize_srcu(&fs_info->subvol_srcu);
5591                 spin_lock(&root->inode_lock);
5592                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5593                 spin_unlock(&root->inode_lock);
5594                 if (empty)
5595                         btrfs_add_dead_root(root);
5596         }
5597 }
5598
5599
5600 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5601 {
5602         struct btrfs_iget_args *args = p;
5603         inode->i_ino = args->location->objectid;
5604         memcpy(&BTRFS_I(inode)->location, args->location,
5605                sizeof(*args->location));
5606         BTRFS_I(inode)->root = args->root;
5607         return 0;
5608 }
5609
5610 static int btrfs_find_actor(struct inode *inode, void *opaque)
5611 {
5612         struct btrfs_iget_args *args = opaque;
5613         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5614                 args->root == BTRFS_I(inode)->root;
5615 }
5616
5617 static struct inode *btrfs_iget_locked(struct super_block *s,
5618                                        struct btrfs_key *location,
5619                                        struct btrfs_root *root)
5620 {
5621         struct inode *inode;
5622         struct btrfs_iget_args args;
5623         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5624
5625         args.location = location;
5626         args.root = root;
5627
5628         inode = iget5_locked(s, hashval, btrfs_find_actor,
5629                              btrfs_init_locked_inode,
5630                              (void *)&args);
5631         return inode;
5632 }
5633
5634 /* Get an inode object given its location and corresponding root.
5635  * Returns in *is_new if the inode was read from disk
5636  */
5637 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5638                          struct btrfs_root *root, int *new)
5639 {
5640         struct inode *inode;
5641
5642         inode = btrfs_iget_locked(s, location, root);
5643         if (!inode)
5644                 return ERR_PTR(-ENOMEM);
5645
5646         if (inode->i_state & I_NEW) {
5647                 int ret;
5648
5649                 ret = btrfs_read_locked_inode(inode);
5650                 if (!ret) {
5651                         inode_tree_add(inode);
5652                         unlock_new_inode(inode);
5653                         if (new)
5654                                 *new = 1;
5655                 } else {
5656                         iget_failed(inode);
5657                         /*
5658                          * ret > 0 can come from btrfs_search_slot called by
5659                          * btrfs_read_locked_inode, this means the inode item
5660                          * was not found.
5661                          */
5662                         if (ret > 0)
5663                                 ret = -ENOENT;
5664                         inode = ERR_PTR(ret);
5665                 }
5666         }
5667
5668         return inode;
5669 }
5670
5671 static struct inode *new_simple_dir(struct super_block *s,
5672                                     struct btrfs_key *key,
5673                                     struct btrfs_root *root)
5674 {
5675         struct inode *inode = new_inode(s);
5676
5677         if (!inode)
5678                 return ERR_PTR(-ENOMEM);
5679
5680         BTRFS_I(inode)->root = root;
5681         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5682         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5683
5684         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5685         inode->i_op = &btrfs_dir_ro_inode_operations;
5686         inode->i_opflags &= ~IOP_XATTR;
5687         inode->i_fop = &simple_dir_operations;
5688         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5689         inode->i_mtime = current_time(inode);
5690         inode->i_atime = inode->i_mtime;
5691         inode->i_ctime = inode->i_mtime;
5692         BTRFS_I(inode)->i_otime = inode->i_mtime;
5693
5694         return inode;
5695 }
5696
5697 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5698 {
5699         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5700         struct inode *inode;
5701         struct btrfs_root *root = BTRFS_I(dir)->root;
5702         struct btrfs_root *sub_root = root;
5703         struct btrfs_key location;
5704         int index;
5705         int ret = 0;
5706
5707         if (dentry->d_name.len > BTRFS_NAME_LEN)
5708                 return ERR_PTR(-ENAMETOOLONG);
5709
5710         ret = btrfs_inode_by_name(dir, dentry, &location);
5711         if (ret < 0)
5712                 return ERR_PTR(ret);
5713
5714         if (location.type == BTRFS_INODE_ITEM_KEY) {
5715                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5716                 return inode;
5717         }
5718
5719         index = srcu_read_lock(&fs_info->subvol_srcu);
5720         ret = fixup_tree_root_location(fs_info, dir, dentry,
5721                                        &location, &sub_root);
5722         if (ret < 0) {
5723                 if (ret != -ENOENT)
5724                         inode = ERR_PTR(ret);
5725                 else
5726                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5727         } else {
5728                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5729         }
5730         srcu_read_unlock(&fs_info->subvol_srcu, index);
5731
5732         if (!IS_ERR(inode) && root != sub_root) {
5733                 down_read(&fs_info->cleanup_work_sem);
5734                 if (!sb_rdonly(inode->i_sb))
5735                         ret = btrfs_orphan_cleanup(sub_root);
5736                 up_read(&fs_info->cleanup_work_sem);
5737                 if (ret) {
5738                         iput(inode);
5739                         inode = ERR_PTR(ret);
5740                 }
5741         }
5742
5743         return inode;
5744 }
5745
5746 static int btrfs_dentry_delete(const struct dentry *dentry)
5747 {
5748         struct btrfs_root *root;
5749         struct inode *inode = d_inode(dentry);
5750
5751         if (!inode && !IS_ROOT(dentry))
5752                 inode = d_inode(dentry->d_parent);
5753
5754         if (inode) {
5755                 root = BTRFS_I(inode)->root;
5756                 if (btrfs_root_refs(&root->root_item) == 0)
5757                         return 1;
5758
5759                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5760                         return 1;
5761         }
5762         return 0;
5763 }
5764
5765 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5766                                    unsigned int flags)
5767 {
5768         struct inode *inode;
5769
5770         inode = btrfs_lookup_dentry(dir, dentry);
5771         if (IS_ERR(inode)) {
5772                 if (PTR_ERR(inode) == -ENOENT)
5773                         inode = NULL;
5774                 else
5775                         return ERR_CAST(inode);
5776         }
5777
5778         return d_splice_alias(inode, dentry);
5779 }
5780
5781 unsigned char btrfs_filetype_table[] = {
5782         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5783 };
5784
5785 /*
5786  * All this infrastructure exists because dir_emit can fault, and we are holding
5787  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5788  * our information into that, and then dir_emit from the buffer.  This is
5789  * similar to what NFS does, only we don't keep the buffer around in pagecache
5790  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5791  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5792  * tree lock.
5793  */
5794 static int btrfs_opendir(struct inode *inode, struct file *file)
5795 {
5796         struct btrfs_file_private *private;
5797
5798         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5799         if (!private)
5800                 return -ENOMEM;
5801         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5802         if (!private->filldir_buf) {
5803                 kfree(private);
5804                 return -ENOMEM;
5805         }
5806         file->private_data = private;
5807         return 0;
5808 }
5809
5810 struct dir_entry {
5811         u64 ino;
5812         u64 offset;
5813         unsigned type;
5814         int name_len;
5815 };
5816
5817 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
5818 {
5819         while (entries--) {
5820                 struct dir_entry *entry = addr;
5821                 char *name = (char *)(entry + 1);
5822
5823                 ctx->pos = get_unaligned(&entry->offset);
5824                 if (!dir_emit(ctx, name, get_unaligned(&entry->name_len),
5825                                          get_unaligned(&entry->ino),
5826                                          get_unaligned(&entry->type)))
5827                         return 1;
5828                 addr += sizeof(struct dir_entry) +
5829                         get_unaligned(&entry->name_len);
5830                 ctx->pos++;
5831         }
5832         return 0;
5833 }
5834
5835 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5836 {
5837         struct inode *inode = file_inode(file);
5838         struct btrfs_root *root = BTRFS_I(inode)->root;
5839         struct btrfs_file_private *private = file->private_data;
5840         struct btrfs_dir_item *di;
5841         struct btrfs_key key;
5842         struct btrfs_key found_key;
5843         struct btrfs_path *path;
5844         void *addr;
5845         struct list_head ins_list;
5846         struct list_head del_list;
5847         int ret;
5848         struct extent_buffer *leaf;
5849         int slot;
5850         char *name_ptr;
5851         int name_len;
5852         int entries = 0;
5853         int total_len = 0;
5854         bool put = false;
5855         struct btrfs_key location;
5856
5857         if (!dir_emit_dots(file, ctx))
5858                 return 0;
5859
5860         path = btrfs_alloc_path();
5861         if (!path)
5862                 return -ENOMEM;
5863
5864         addr = private->filldir_buf;
5865         path->reada = READA_FORWARD;
5866
5867         INIT_LIST_HEAD(&ins_list);
5868         INIT_LIST_HEAD(&del_list);
5869         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
5870
5871 again:
5872         key.type = BTRFS_DIR_INDEX_KEY;
5873         key.offset = ctx->pos;
5874         key.objectid = btrfs_ino(BTRFS_I(inode));
5875
5876         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5877         if (ret < 0)
5878                 goto err;
5879
5880         while (1) {
5881                 struct dir_entry *entry;
5882
5883                 leaf = path->nodes[0];
5884                 slot = path->slots[0];
5885                 if (slot >= btrfs_header_nritems(leaf)) {
5886                         ret = btrfs_next_leaf(root, path);
5887                         if (ret < 0)
5888                                 goto err;
5889                         else if (ret > 0)
5890                                 break;
5891                         continue;
5892                 }
5893
5894                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5895
5896                 if (found_key.objectid != key.objectid)
5897                         break;
5898                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
5899                         break;
5900                 if (found_key.offset < ctx->pos)
5901                         goto next;
5902                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
5903                         goto next;
5904                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5905                 name_len = btrfs_dir_name_len(leaf, di);
5906                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
5907                     PAGE_SIZE) {
5908                         btrfs_release_path(path);
5909                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5910                         if (ret)
5911                                 goto nopos;
5912                         addr = private->filldir_buf;
5913                         entries = 0;
5914                         total_len = 0;
5915                         goto again;
5916                 }
5917
5918                 entry = addr;
5919                 put_unaligned(name_len, &entry->name_len);
5920                 name_ptr = (char *)(entry + 1);
5921                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
5922                                    name_len);
5923                 put_unaligned(btrfs_filetype_table[btrfs_dir_type(leaf, di)],
5924                                 &entry->type);
5925                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
5926                 put_unaligned(location.objectid, &entry->ino);
5927                 put_unaligned(found_key.offset, &entry->offset);
5928                 entries++;
5929                 addr += sizeof(struct dir_entry) + name_len;
5930                 total_len += sizeof(struct dir_entry) + name_len;
5931 next:
5932                 path->slots[0]++;
5933         }
5934         btrfs_release_path(path);
5935
5936         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5937         if (ret)
5938                 goto nopos;
5939
5940         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
5941         if (ret)
5942                 goto nopos;
5943
5944         /*
5945          * Stop new entries from being returned after we return the last
5946          * entry.
5947          *
5948          * New directory entries are assigned a strictly increasing
5949          * offset.  This means that new entries created during readdir
5950          * are *guaranteed* to be seen in the future by that readdir.
5951          * This has broken buggy programs which operate on names as
5952          * they're returned by readdir.  Until we re-use freed offsets
5953          * we have this hack to stop new entries from being returned
5954          * under the assumption that they'll never reach this huge
5955          * offset.
5956          *
5957          * This is being careful not to overflow 32bit loff_t unless the
5958          * last entry requires it because doing so has broken 32bit apps
5959          * in the past.
5960          */
5961         if (ctx->pos >= INT_MAX)
5962                 ctx->pos = LLONG_MAX;
5963         else
5964                 ctx->pos = INT_MAX;
5965 nopos:
5966         ret = 0;
5967 err:
5968         if (put)
5969                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
5970         btrfs_free_path(path);
5971         return ret;
5972 }
5973
5974 /*
5975  * This is somewhat expensive, updating the tree every time the
5976  * inode changes.  But, it is most likely to find the inode in cache.
5977  * FIXME, needs more benchmarking...there are no reasons other than performance
5978  * to keep or drop this code.
5979  */
5980 static int btrfs_dirty_inode(struct inode *inode)
5981 {
5982         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5983         struct btrfs_root *root = BTRFS_I(inode)->root;
5984         struct btrfs_trans_handle *trans;
5985         int ret;
5986
5987         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5988                 return 0;
5989
5990         trans = btrfs_join_transaction(root);
5991         if (IS_ERR(trans))
5992                 return PTR_ERR(trans);
5993
5994         ret = btrfs_update_inode(trans, root, inode);
5995         if (ret && ret == -ENOSPC) {
5996                 /* whoops, lets try again with the full transaction */
5997                 btrfs_end_transaction(trans);
5998                 trans = btrfs_start_transaction(root, 1);
5999                 if (IS_ERR(trans))
6000                         return PTR_ERR(trans);
6001
6002                 ret = btrfs_update_inode(trans, root, inode);
6003         }
6004         btrfs_end_transaction(trans);
6005         if (BTRFS_I(inode)->delayed_node)
6006                 btrfs_balance_delayed_items(fs_info);
6007
6008         return ret;
6009 }
6010
6011 /*
6012  * This is a copy of file_update_time.  We need this so we can return error on
6013  * ENOSPC for updating the inode in the case of file write and mmap writes.
6014  */
6015 static int btrfs_update_time(struct inode *inode, struct timespec64 *now,
6016                              int flags)
6017 {
6018         struct btrfs_root *root = BTRFS_I(inode)->root;
6019         bool dirty = flags & ~S_VERSION;
6020
6021         if (btrfs_root_readonly(root))
6022                 return -EROFS;
6023
6024         if (flags & S_VERSION)
6025                 dirty |= inode_maybe_inc_iversion(inode, dirty);
6026         if (flags & S_CTIME)
6027                 inode->i_ctime = *now;
6028         if (flags & S_MTIME)
6029                 inode->i_mtime = *now;
6030         if (flags & S_ATIME)
6031                 inode->i_atime = *now;
6032         return dirty ? btrfs_dirty_inode(inode) : 0;
6033 }
6034
6035 /*
6036  * find the highest existing sequence number in a directory
6037  * and then set the in-memory index_cnt variable to reflect
6038  * free sequence numbers
6039  */
6040 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
6041 {
6042         struct btrfs_root *root = inode->root;
6043         struct btrfs_key key, found_key;
6044         struct btrfs_path *path;
6045         struct extent_buffer *leaf;
6046         int ret;
6047
6048         key.objectid = btrfs_ino(inode);
6049         key.type = BTRFS_DIR_INDEX_KEY;
6050         key.offset = (u64)-1;
6051
6052         path = btrfs_alloc_path();
6053         if (!path)
6054                 return -ENOMEM;
6055
6056         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6057         if (ret < 0)
6058                 goto out;
6059         /* FIXME: we should be able to handle this */
6060         if (ret == 0)
6061                 goto out;
6062         ret = 0;
6063
6064         /*
6065          * MAGIC NUMBER EXPLANATION:
6066          * since we search a directory based on f_pos we have to start at 2
6067          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6068          * else has to start at 2
6069          */
6070         if (path->slots[0] == 0) {
6071                 inode->index_cnt = 2;
6072                 goto out;
6073         }
6074
6075         path->slots[0]--;
6076
6077         leaf = path->nodes[0];
6078         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6079
6080         if (found_key.objectid != btrfs_ino(inode) ||
6081             found_key.type != BTRFS_DIR_INDEX_KEY) {
6082                 inode->index_cnt = 2;
6083                 goto out;
6084         }
6085
6086         inode->index_cnt = found_key.offset + 1;
6087 out:
6088         btrfs_free_path(path);
6089         return ret;
6090 }
6091
6092 /*
6093  * helper to find a free sequence number in a given directory.  This current
6094  * code is very simple, later versions will do smarter things in the btree
6095  */
6096 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6097 {
6098         int ret = 0;
6099
6100         if (dir->index_cnt == (u64)-1) {
6101                 ret = btrfs_inode_delayed_dir_index_count(dir);
6102                 if (ret) {
6103                         ret = btrfs_set_inode_index_count(dir);
6104                         if (ret)
6105                                 return ret;
6106                 }
6107         }
6108
6109         *index = dir->index_cnt;
6110         dir->index_cnt++;
6111
6112         return ret;
6113 }
6114
6115 static int btrfs_insert_inode_locked(struct inode *inode)
6116 {
6117         struct btrfs_iget_args args;
6118         args.location = &BTRFS_I(inode)->location;
6119         args.root = BTRFS_I(inode)->root;
6120
6121         return insert_inode_locked4(inode,
6122                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6123                    btrfs_find_actor, &args);
6124 }
6125
6126 /*
6127  * Inherit flags from the parent inode.
6128  *
6129  * Currently only the compression flags and the cow flags are inherited.
6130  */
6131 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
6132 {
6133         unsigned int flags;
6134
6135         if (!dir)
6136                 return;
6137
6138         flags = BTRFS_I(dir)->flags;
6139
6140         if (flags & BTRFS_INODE_NOCOMPRESS) {
6141                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
6142                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
6143         } else if (flags & BTRFS_INODE_COMPRESS) {
6144                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
6145                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
6146         }
6147
6148         if (flags & BTRFS_INODE_NODATACOW) {
6149                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
6150                 if (S_ISREG(inode->i_mode))
6151                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6152         }
6153
6154         btrfs_sync_inode_flags_to_i_flags(inode);
6155 }
6156
6157 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6158                                      struct btrfs_root *root,
6159                                      struct inode *dir,
6160                                      const char *name, int name_len,
6161                                      u64 ref_objectid, u64 objectid,
6162                                      umode_t mode, u64 *index)
6163 {
6164         struct btrfs_fs_info *fs_info = root->fs_info;
6165         struct inode *inode;
6166         struct btrfs_inode_item *inode_item;
6167         struct btrfs_key *location;
6168         struct btrfs_path *path;
6169         struct btrfs_inode_ref *ref;
6170         struct btrfs_key key[2];
6171         u32 sizes[2];
6172         int nitems = name ? 2 : 1;
6173         unsigned long ptr;
6174         int ret;
6175
6176         path = btrfs_alloc_path();
6177         if (!path)
6178                 return ERR_PTR(-ENOMEM);
6179
6180         inode = new_inode(fs_info->sb);
6181         if (!inode) {
6182                 btrfs_free_path(path);
6183                 return ERR_PTR(-ENOMEM);
6184         }
6185
6186         /*
6187          * O_TMPFILE, set link count to 0, so that after this point,
6188          * we fill in an inode item with the correct link count.
6189          */
6190         if (!name)
6191                 set_nlink(inode, 0);
6192
6193         /*
6194          * we have to initialize this early, so we can reclaim the inode
6195          * number if we fail afterwards in this function.
6196          */
6197         inode->i_ino = objectid;
6198
6199         if (dir && name) {
6200                 trace_btrfs_inode_request(dir);
6201
6202                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6203                 if (ret) {
6204                         btrfs_free_path(path);
6205                         iput(inode);
6206                         return ERR_PTR(ret);
6207                 }
6208         } else if (dir) {
6209                 *index = 0;
6210         }
6211         /*
6212          * index_cnt is ignored for everything but a dir,
6213          * btrfs_set_inode_index_count has an explanation for the magic
6214          * number
6215          */
6216         BTRFS_I(inode)->index_cnt = 2;
6217         BTRFS_I(inode)->dir_index = *index;
6218         BTRFS_I(inode)->root = root;
6219         BTRFS_I(inode)->generation = trans->transid;
6220         inode->i_generation = BTRFS_I(inode)->generation;
6221
6222         /*
6223          * We could have gotten an inode number from somebody who was fsynced
6224          * and then removed in this same transaction, so let's just set full
6225          * sync since it will be a full sync anyway and this will blow away the
6226          * old info in the log.
6227          */
6228         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6229
6230         key[0].objectid = objectid;
6231         key[0].type = BTRFS_INODE_ITEM_KEY;
6232         key[0].offset = 0;
6233
6234         sizes[0] = sizeof(struct btrfs_inode_item);
6235
6236         if (name) {
6237                 /*
6238                  * Start new inodes with an inode_ref. This is slightly more
6239                  * efficient for small numbers of hard links since they will
6240                  * be packed into one item. Extended refs will kick in if we
6241                  * add more hard links than can fit in the ref item.
6242                  */
6243                 key[1].objectid = objectid;
6244                 key[1].type = BTRFS_INODE_REF_KEY;
6245                 key[1].offset = ref_objectid;
6246
6247                 sizes[1] = name_len + sizeof(*ref);
6248         }
6249
6250         location = &BTRFS_I(inode)->location;
6251         location->objectid = objectid;
6252         location->offset = 0;
6253         location->type = BTRFS_INODE_ITEM_KEY;
6254
6255         ret = btrfs_insert_inode_locked(inode);
6256         if (ret < 0) {
6257                 iput(inode);
6258                 goto fail;
6259         }
6260
6261         path->leave_spinning = 1;
6262         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6263         if (ret != 0)
6264                 goto fail_unlock;
6265
6266         inode_init_owner(inode, dir, mode);
6267         inode_set_bytes(inode, 0);
6268
6269         inode->i_mtime = current_time(inode);
6270         inode->i_atime = inode->i_mtime;
6271         inode->i_ctime = inode->i_mtime;
6272         BTRFS_I(inode)->i_otime = inode->i_mtime;
6273
6274         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6275                                   struct btrfs_inode_item);
6276         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6277                              sizeof(*inode_item));
6278         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6279
6280         if (name) {
6281                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6282                                      struct btrfs_inode_ref);
6283                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6284                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6285                 ptr = (unsigned long)(ref + 1);
6286                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6287         }
6288
6289         btrfs_mark_buffer_dirty(path->nodes[0]);
6290         btrfs_free_path(path);
6291
6292         btrfs_inherit_iflags(inode, dir);
6293
6294         if (S_ISREG(mode)) {
6295                 if (btrfs_test_opt(fs_info, NODATASUM))
6296                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6297                 if (btrfs_test_opt(fs_info, NODATACOW))
6298                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6299                                 BTRFS_INODE_NODATASUM;
6300         }
6301
6302         inode_tree_add(inode);
6303
6304         trace_btrfs_inode_new(inode);
6305         btrfs_set_inode_last_trans(trans, inode);
6306
6307         btrfs_update_root_times(trans, root);
6308
6309         ret = btrfs_inode_inherit_props(trans, inode, dir);
6310         if (ret)
6311                 btrfs_err(fs_info,
6312                           "error inheriting props for ino %llu (root %llu): %d",
6313                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6314
6315         return inode;
6316
6317 fail_unlock:
6318         discard_new_inode(inode);
6319 fail:
6320         if (dir && name)
6321                 BTRFS_I(dir)->index_cnt--;
6322         btrfs_free_path(path);
6323         return ERR_PTR(ret);
6324 }
6325
6326 static inline u8 btrfs_inode_type(struct inode *inode)
6327 {
6328         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6329 }
6330
6331 /*
6332  * utility function to add 'inode' into 'parent_inode' with
6333  * a give name and a given sequence number.
6334  * if 'add_backref' is true, also insert a backref from the
6335  * inode to the parent directory.
6336  */
6337 int btrfs_add_link(struct btrfs_trans_handle *trans,
6338                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6339                    const char *name, int name_len, int add_backref, u64 index)
6340 {
6341         int ret = 0;
6342         struct btrfs_key key;
6343         struct btrfs_root *root = parent_inode->root;
6344         u64 ino = btrfs_ino(inode);
6345         u64 parent_ino = btrfs_ino(parent_inode);
6346
6347         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6348                 memcpy(&key, &inode->root->root_key, sizeof(key));
6349         } else {
6350                 key.objectid = ino;
6351                 key.type = BTRFS_INODE_ITEM_KEY;
6352                 key.offset = 0;
6353         }
6354
6355         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6356                 ret = btrfs_add_root_ref(trans, key.objectid,
6357                                          root->root_key.objectid, parent_ino,
6358                                          index, name, name_len);
6359         } else if (add_backref) {
6360                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6361                                              parent_ino, index);
6362         }
6363
6364         /* Nothing to clean up yet */
6365         if (ret)
6366                 return ret;
6367
6368         ret = btrfs_insert_dir_item(trans, name, name_len, parent_inode, &key,
6369                                     btrfs_inode_type(&inode->vfs_inode), index);
6370         if (ret == -EEXIST || ret == -EOVERFLOW)
6371                 goto fail_dir_item;
6372         else if (ret) {
6373                 btrfs_abort_transaction(trans, ret);
6374                 return ret;
6375         }
6376
6377         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6378                            name_len * 2);
6379         inode_inc_iversion(&parent_inode->vfs_inode);
6380         parent_inode->vfs_inode.i_mtime = parent_inode->vfs_inode.i_ctime =
6381                 current_time(&parent_inode->vfs_inode);
6382         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6383         if (ret)
6384                 btrfs_abort_transaction(trans, ret);
6385         return ret;
6386
6387 fail_dir_item:
6388         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6389                 u64 local_index;
6390                 int err;
6391                 err = btrfs_del_root_ref(trans, key.objectid,
6392                                          root->root_key.objectid, parent_ino,
6393                                          &local_index, name, name_len);
6394
6395         } else if (add_backref) {
6396                 u64 local_index;
6397                 int err;
6398
6399                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6400                                           ino, parent_ino, &local_index);
6401         }
6402         return ret;
6403 }
6404
6405 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6406                             struct btrfs_inode *dir, struct dentry *dentry,
6407                             struct btrfs_inode *inode, int backref, u64 index)
6408 {
6409         int err = btrfs_add_link(trans, dir, inode,
6410                                  dentry->d_name.name, dentry->d_name.len,
6411                                  backref, index);
6412         if (err > 0)
6413                 err = -EEXIST;
6414         return err;
6415 }
6416
6417 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6418                         umode_t mode, dev_t rdev)
6419 {
6420         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6421         struct btrfs_trans_handle *trans;
6422         struct btrfs_root *root = BTRFS_I(dir)->root;
6423         struct inode *inode = NULL;
6424         int err;
6425         u64 objectid;
6426         u64 index = 0;
6427
6428         /*
6429          * 2 for inode item and ref
6430          * 2 for dir items
6431          * 1 for xattr if selinux is on
6432          */
6433         trans = btrfs_start_transaction(root, 5);
6434         if (IS_ERR(trans))
6435                 return PTR_ERR(trans);
6436
6437         err = btrfs_find_free_ino(root, &objectid);
6438         if (err)
6439                 goto out_unlock;
6440
6441         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6442                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6443                         mode, &index);
6444         if (IS_ERR(inode)) {
6445                 err = PTR_ERR(inode);
6446                 inode = NULL;
6447                 goto out_unlock;
6448         }
6449
6450         /*
6451         * If the active LSM wants to access the inode during
6452         * d_instantiate it needs these. Smack checks to see
6453         * if the filesystem supports xattrs by looking at the
6454         * ops vector.
6455         */
6456         inode->i_op = &btrfs_special_inode_operations;
6457         init_special_inode(inode, inode->i_mode, rdev);
6458
6459         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6460         if (err)
6461                 goto out_unlock;
6462
6463         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6464                         0, index);
6465         if (err)
6466                 goto out_unlock;
6467
6468         btrfs_update_inode(trans, root, inode);
6469         d_instantiate_new(dentry, inode);
6470
6471 out_unlock:
6472         btrfs_end_transaction(trans);
6473         btrfs_btree_balance_dirty(fs_info);
6474         if (err && inode) {
6475                 inode_dec_link_count(inode);
6476                 discard_new_inode(inode);
6477         }
6478         return err;
6479 }
6480
6481 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6482                         umode_t mode, bool excl)
6483 {
6484         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6485         struct btrfs_trans_handle *trans;
6486         struct btrfs_root *root = BTRFS_I(dir)->root;
6487         struct inode *inode = NULL;
6488         int err;
6489         u64 objectid;
6490         u64 index = 0;
6491
6492         /*
6493          * 2 for inode item and ref
6494          * 2 for dir items
6495          * 1 for xattr if selinux is on
6496          */
6497         trans = btrfs_start_transaction(root, 5);
6498         if (IS_ERR(trans))
6499                 return PTR_ERR(trans);
6500
6501         err = btrfs_find_free_ino(root, &objectid);
6502         if (err)
6503                 goto out_unlock;
6504
6505         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6506                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6507                         mode, &index);
6508         if (IS_ERR(inode)) {
6509                 err = PTR_ERR(inode);
6510                 inode = NULL;
6511                 goto out_unlock;
6512         }
6513         /*
6514         * If the active LSM wants to access the inode during
6515         * d_instantiate it needs these. Smack checks to see
6516         * if the filesystem supports xattrs by looking at the
6517         * ops vector.
6518         */
6519         inode->i_fop = &btrfs_file_operations;
6520         inode->i_op = &btrfs_file_inode_operations;
6521         inode->i_mapping->a_ops = &btrfs_aops;
6522
6523         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6524         if (err)
6525                 goto out_unlock;
6526
6527         err = btrfs_update_inode(trans, root, inode);
6528         if (err)
6529                 goto out_unlock;
6530
6531         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6532                         0, index);
6533         if (err)
6534                 goto out_unlock;
6535
6536         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6537         d_instantiate_new(dentry, inode);
6538
6539 out_unlock:
6540         btrfs_end_transaction(trans);
6541         if (err && inode) {
6542                 inode_dec_link_count(inode);
6543                 discard_new_inode(inode);
6544         }
6545         btrfs_btree_balance_dirty(fs_info);
6546         return err;
6547 }
6548
6549 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6550                       struct dentry *dentry)
6551 {
6552         struct btrfs_trans_handle *trans = NULL;
6553         struct btrfs_root *root = BTRFS_I(dir)->root;
6554         struct inode *inode = d_inode(old_dentry);
6555         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6556         u64 index;
6557         int err;
6558         int drop_inode = 0;
6559
6560         /* do not allow sys_link's with other subvols of the same device */
6561         if (root->root_key.objectid != BTRFS_I(inode)->root->root_key.objectid)
6562                 return -EXDEV;
6563
6564         if (inode->i_nlink >= BTRFS_LINK_MAX)
6565                 return -EMLINK;
6566
6567         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6568         if (err)
6569                 goto fail;
6570
6571         /*
6572          * 2 items for inode and inode ref
6573          * 2 items for dir items
6574          * 1 item for parent inode
6575          * 1 item for orphan item deletion if O_TMPFILE
6576          */
6577         trans = btrfs_start_transaction(root, inode->i_nlink ? 5 : 6);
6578         if (IS_ERR(trans)) {
6579                 err = PTR_ERR(trans);
6580                 trans = NULL;
6581                 goto fail;
6582         }
6583
6584         /* There are several dir indexes for this inode, clear the cache. */
6585         BTRFS_I(inode)->dir_index = 0ULL;
6586         inc_nlink(inode);
6587         inode_inc_iversion(inode);
6588         inode->i_ctime = current_time(inode);
6589         ihold(inode);
6590         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6591
6592         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6593                         1, index);
6594
6595         if (err) {
6596                 drop_inode = 1;
6597         } else {
6598                 struct dentry *parent = dentry->d_parent;
6599                 int ret;
6600
6601                 err = btrfs_update_inode(trans, root, inode);
6602                 if (err)
6603                         goto fail;
6604                 if (inode->i_nlink == 1) {
6605                         /*
6606                          * If new hard link count is 1, it's a file created
6607                          * with open(2) O_TMPFILE flag.
6608                          */
6609                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6610                         if (err)
6611                                 goto fail;
6612                 }
6613                 d_instantiate(dentry, inode);
6614                 ret = btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent,
6615                                          true, NULL);
6616                 if (ret == BTRFS_NEED_TRANS_COMMIT) {
6617                         err = btrfs_commit_transaction(trans);
6618                         trans = NULL;
6619                 }
6620         }
6621
6622 fail:
6623         if (trans)
6624                 btrfs_end_transaction(trans);
6625         if (drop_inode) {
6626                 inode_dec_link_count(inode);
6627                 iput(inode);
6628         }
6629         btrfs_btree_balance_dirty(fs_info);
6630         return err;
6631 }
6632
6633 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6634 {
6635         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6636         struct inode *inode = NULL;
6637         struct btrfs_trans_handle *trans;
6638         struct btrfs_root *root = BTRFS_I(dir)->root;
6639         int err = 0;
6640         int drop_on_err = 0;
6641         u64 objectid = 0;
6642         u64 index = 0;
6643
6644         /*
6645          * 2 items for inode and ref
6646          * 2 items for dir items
6647          * 1 for xattr if selinux is on
6648          */
6649         trans = btrfs_start_transaction(root, 5);
6650         if (IS_ERR(trans))
6651                 return PTR_ERR(trans);
6652
6653         err = btrfs_find_free_ino(root, &objectid);
6654         if (err)
6655                 goto out_fail;
6656
6657         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6658                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6659                         S_IFDIR | mode, &index);
6660         if (IS_ERR(inode)) {
6661                 err = PTR_ERR(inode);
6662                 inode = NULL;
6663                 goto out_fail;
6664         }
6665
6666         drop_on_err = 1;
6667         /* these must be set before we unlock the inode */
6668         inode->i_op = &btrfs_dir_inode_operations;
6669         inode->i_fop = &btrfs_dir_file_operations;
6670
6671         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6672         if (err)
6673                 goto out_fail;
6674
6675         btrfs_i_size_write(BTRFS_I(inode), 0);
6676         err = btrfs_update_inode(trans, root, inode);
6677         if (err)
6678                 goto out_fail;
6679
6680         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6681                         dentry->d_name.name,
6682                         dentry->d_name.len, 0, index);
6683         if (err)
6684                 goto out_fail;
6685
6686         d_instantiate_new(dentry, inode);
6687         drop_on_err = 0;
6688
6689 out_fail:
6690         btrfs_end_transaction(trans);
6691         if (err && inode) {
6692                 inode_dec_link_count(inode);
6693                 discard_new_inode(inode);
6694         }
6695         btrfs_btree_balance_dirty(fs_info);
6696         return err;
6697 }
6698
6699 static noinline int uncompress_inline(struct btrfs_path *path,
6700                                       struct page *page,
6701                                       size_t pg_offset, u64 extent_offset,
6702                                       struct btrfs_file_extent_item *item)
6703 {
6704         int ret;
6705         struct extent_buffer *leaf = path->nodes[0];
6706         char *tmp;
6707         size_t max_size;
6708         unsigned long inline_size;
6709         unsigned long ptr;
6710         int compress_type;
6711
6712         WARN_ON(pg_offset != 0);
6713         compress_type = btrfs_file_extent_compression(leaf, item);
6714         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6715         inline_size = btrfs_file_extent_inline_item_len(leaf,
6716                                         btrfs_item_nr(path->slots[0]));
6717         tmp = kmalloc(inline_size, GFP_NOFS);
6718         if (!tmp)
6719                 return -ENOMEM;
6720         ptr = btrfs_file_extent_inline_start(item);
6721
6722         read_extent_buffer(leaf, tmp, ptr, inline_size);
6723
6724         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6725         ret = btrfs_decompress(compress_type, tmp, page,
6726                                extent_offset, inline_size, max_size);
6727
6728         /*
6729          * decompression code contains a memset to fill in any space between the end
6730          * of the uncompressed data and the end of max_size in case the decompressed
6731          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6732          * the end of an inline extent and the beginning of the next block, so we
6733          * cover that region here.
6734          */
6735
6736         if (max_size + pg_offset < PAGE_SIZE) {
6737                 char *map = kmap(page);
6738                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6739                 kunmap(page);
6740         }
6741         kfree(tmp);
6742         return ret;
6743 }
6744
6745 /*
6746  * a bit scary, this does extent mapping from logical file offset to the disk.
6747  * the ugly parts come from merging extents from the disk with the in-ram
6748  * representation.  This gets more complex because of the data=ordered code,
6749  * where the in-ram extents might be locked pending data=ordered completion.
6750  *
6751  * This also copies inline extents directly into the page.
6752  */
6753 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6754                                     struct page *page,
6755                                     size_t pg_offset, u64 start, u64 len,
6756                                     int create)
6757 {
6758         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6759         int ret;
6760         int err = 0;
6761         u64 extent_start = 0;
6762         u64 extent_end = 0;
6763         u64 objectid = btrfs_ino(inode);
6764         u32 found_type;
6765         struct btrfs_path *path = NULL;
6766         struct btrfs_root *root = inode->root;
6767         struct btrfs_file_extent_item *item;
6768         struct extent_buffer *leaf;
6769         struct btrfs_key found_key;
6770         struct extent_map *em = NULL;
6771         struct extent_map_tree *em_tree = &inode->extent_tree;
6772         struct extent_io_tree *io_tree = &inode->io_tree;
6773         const bool new_inline = !page || create;
6774
6775         read_lock(&em_tree->lock);
6776         em = lookup_extent_mapping(em_tree, start, len);
6777         if (em)
6778                 em->bdev = fs_info->fs_devices->latest_bdev;
6779         read_unlock(&em_tree->lock);
6780
6781         if (em) {
6782                 if (em->start > start || em->start + em->len <= start)
6783                         free_extent_map(em);
6784                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6785                         free_extent_map(em);
6786                 else
6787                         goto out;
6788         }
6789         em = alloc_extent_map();
6790         if (!em) {
6791                 err = -ENOMEM;
6792                 goto out;
6793         }
6794         em->bdev = fs_info->fs_devices->latest_bdev;
6795         em->start = EXTENT_MAP_HOLE;
6796         em->orig_start = EXTENT_MAP_HOLE;
6797         em->len = (u64)-1;
6798         em->block_len = (u64)-1;
6799
6800         path = btrfs_alloc_path();
6801         if (!path) {
6802                 err = -ENOMEM;
6803                 goto out;
6804         }
6805
6806         /* Chances are we'll be called again, so go ahead and do readahead */
6807         path->reada = READA_FORWARD;
6808
6809         /*
6810          * Unless we're going to uncompress the inline extent, no sleep would
6811          * happen.
6812          */
6813         path->leave_spinning = 1;
6814
6815         ret = btrfs_lookup_file_extent(NULL, root, path, objectid, start, 0);
6816         if (ret < 0) {
6817                 err = ret;
6818                 goto out;
6819         }
6820
6821         if (ret != 0) {
6822                 if (path->slots[0] == 0)
6823                         goto not_found;
6824                 path->slots[0]--;
6825         }
6826
6827         leaf = path->nodes[0];
6828         item = btrfs_item_ptr(leaf, path->slots[0],
6829                               struct btrfs_file_extent_item);
6830         /* are we inside the extent that was found? */
6831         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6832         found_type = found_key.type;
6833         if (found_key.objectid != objectid ||
6834             found_type != BTRFS_EXTENT_DATA_KEY) {
6835                 /*
6836                  * If we backup past the first extent we want to move forward
6837                  * and see if there is an extent in front of us, otherwise we'll
6838                  * say there is a hole for our whole search range which can
6839                  * cause problems.
6840                  */
6841                 extent_end = start;
6842                 goto next;
6843         }
6844
6845         found_type = btrfs_file_extent_type(leaf, item);
6846         extent_start = found_key.offset;
6847         if (found_type == BTRFS_FILE_EXTENT_REG ||
6848             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6849                 extent_end = extent_start +
6850                        btrfs_file_extent_num_bytes(leaf, item);
6851
6852                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
6853                                                        extent_start);
6854         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6855                 size_t size;
6856
6857                 size = btrfs_file_extent_ram_bytes(leaf, item);
6858                 extent_end = ALIGN(extent_start + size,
6859                                    fs_info->sectorsize);
6860
6861                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
6862                                                       path->slots[0],
6863                                                       extent_start);
6864         }
6865 next:
6866         if (start >= extent_end) {
6867                 path->slots[0]++;
6868                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6869                         ret = btrfs_next_leaf(root, path);
6870                         if (ret < 0) {
6871                                 err = ret;
6872                                 goto out;
6873                         }
6874                         if (ret > 0)
6875                                 goto not_found;
6876                         leaf = path->nodes[0];
6877                 }
6878                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6879                 if (found_key.objectid != objectid ||
6880                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6881                         goto not_found;
6882                 if (start + len <= found_key.offset)
6883                         goto not_found;
6884                 if (start > found_key.offset)
6885                         goto next;
6886                 em->start = start;
6887                 em->orig_start = start;
6888                 em->len = found_key.offset - start;
6889                 goto not_found_em;
6890         }
6891
6892         btrfs_extent_item_to_extent_map(inode, path, item,
6893                         new_inline, em);
6894
6895         if (found_type == BTRFS_FILE_EXTENT_REG ||
6896             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6897                 goto insert;
6898         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6899                 unsigned long ptr;
6900                 char *map;
6901                 size_t size;
6902                 size_t extent_offset;
6903                 size_t copy_size;
6904
6905                 if (new_inline)
6906                         goto out;
6907
6908                 size = btrfs_file_extent_ram_bytes(leaf, item);
6909                 extent_offset = page_offset(page) + pg_offset - extent_start;
6910                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
6911                                   size - extent_offset);
6912                 em->start = extent_start + extent_offset;
6913                 em->len = ALIGN(copy_size, fs_info->sectorsize);
6914                 em->orig_block_len = em->len;
6915                 em->orig_start = em->start;
6916                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6917
6918                 btrfs_set_path_blocking(path);
6919                 if (!PageUptodate(page)) {
6920                         if (btrfs_file_extent_compression(leaf, item) !=
6921                             BTRFS_COMPRESS_NONE) {
6922                                 ret = uncompress_inline(path, page, pg_offset,
6923                                                         extent_offset, item);
6924                                 if (ret) {
6925                                         err = ret;
6926                                         goto out;
6927                                 }
6928                         } else {
6929                                 map = kmap(page);
6930                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6931                                                    copy_size);
6932                                 if (pg_offset + copy_size < PAGE_SIZE) {
6933                                         memset(map + pg_offset + copy_size, 0,
6934                                                PAGE_SIZE - pg_offset -
6935                                                copy_size);
6936                                 }
6937                                 kunmap(page);
6938                         }
6939                         flush_dcache_page(page);
6940                 }
6941                 set_extent_uptodate(io_tree, em->start,
6942                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6943                 goto insert;
6944         }
6945 not_found:
6946         em->start = start;
6947         em->orig_start = start;
6948         em->len = len;
6949 not_found_em:
6950         em->block_start = EXTENT_MAP_HOLE;
6951 insert:
6952         btrfs_release_path(path);
6953         if (em->start > start || extent_map_end(em) <= start) {
6954                 btrfs_err(fs_info,
6955                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
6956                           em->start, em->len, start, len);
6957                 err = -EIO;
6958                 goto out;
6959         }
6960
6961         err = 0;
6962         write_lock(&em_tree->lock);
6963         err = btrfs_add_extent_mapping(fs_info, em_tree, &em, start, len);
6964         write_unlock(&em_tree->lock);
6965 out:
6966         btrfs_free_path(path);
6967
6968         trace_btrfs_get_extent(root, inode, em);
6969
6970         if (err) {
6971                 free_extent_map(em);
6972                 return ERR_PTR(err);
6973         }
6974         BUG_ON(!em); /* Error is always set */
6975         return em;
6976 }
6977
6978 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
6979                 struct page *page,
6980                 size_t pg_offset, u64 start, u64 len,
6981                 int create)
6982 {
6983         struct extent_map *em;
6984         struct extent_map *hole_em = NULL;
6985         u64 range_start = start;
6986         u64 end;
6987         u64 found;
6988         u64 found_end;
6989         int err = 0;
6990
6991         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
6992         if (IS_ERR(em))
6993                 return em;
6994         /*
6995          * If our em maps to:
6996          * - a hole or
6997          * - a pre-alloc extent,
6998          * there might actually be delalloc bytes behind it.
6999          */
7000         if (em->block_start != EXTENT_MAP_HOLE &&
7001             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7002                 return em;
7003         else
7004                 hole_em = em;
7005
7006         /* check to see if we've wrapped (len == -1 or similar) */
7007         end = start + len;
7008         if (end < start)
7009                 end = (u64)-1;
7010         else
7011                 end -= 1;
7012
7013         em = NULL;
7014
7015         /* ok, we didn't find anything, lets look for delalloc */
7016         found = count_range_bits(&inode->io_tree, &range_start,
7017                                  end, len, EXTENT_DELALLOC, 1);
7018         found_end = range_start + found;
7019         if (found_end < range_start)
7020                 found_end = (u64)-1;
7021
7022         /*
7023          * we didn't find anything useful, return
7024          * the original results from get_extent()
7025          */
7026         if (range_start > end || found_end <= start) {
7027                 em = hole_em;
7028                 hole_em = NULL;
7029                 goto out;
7030         }
7031
7032         /* adjust the range_start to make sure it doesn't
7033          * go backwards from the start they passed in
7034          */
7035         range_start = max(start, range_start);
7036         found = found_end - range_start;
7037
7038         if (found > 0) {
7039                 u64 hole_start = start;
7040                 u64 hole_len = len;
7041
7042                 em = alloc_extent_map();
7043                 if (!em) {
7044                         err = -ENOMEM;
7045                         goto out;
7046                 }
7047                 /*
7048                  * when btrfs_get_extent can't find anything it
7049                  * returns one huge hole
7050                  *
7051                  * make sure what it found really fits our range, and
7052                  * adjust to make sure it is based on the start from
7053                  * the caller
7054                  */
7055                 if (hole_em) {
7056                         u64 calc_end = extent_map_end(hole_em);
7057
7058                         if (calc_end <= start || (hole_em->start > end)) {
7059                                 free_extent_map(hole_em);
7060                                 hole_em = NULL;
7061                         } else {
7062                                 hole_start = max(hole_em->start, start);
7063                                 hole_len = calc_end - hole_start;
7064                         }
7065                 }
7066                 em->bdev = NULL;
7067                 if (hole_em && range_start > hole_start) {
7068                         /* our hole starts before our delalloc, so we
7069                          * have to return just the parts of the hole
7070                          * that go until  the delalloc starts
7071                          */
7072                         em->len = min(hole_len,
7073                                       range_start - hole_start);
7074                         em->start = hole_start;
7075                         em->orig_start = hole_start;
7076                         /*
7077                          * don't adjust block start at all,
7078                          * it is fixed at EXTENT_MAP_HOLE
7079                          */
7080                         em->block_start = hole_em->block_start;
7081                         em->block_len = hole_len;
7082                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7083                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7084                 } else {
7085                         em->start = range_start;
7086                         em->len = found;
7087                         em->orig_start = range_start;
7088                         em->block_start = EXTENT_MAP_DELALLOC;
7089                         em->block_len = found;
7090                 }
7091         } else {
7092                 return hole_em;
7093         }
7094 out:
7095
7096         free_extent_map(hole_em);
7097         if (err) {
7098                 free_extent_map(em);
7099                 return ERR_PTR(err);
7100         }
7101         return em;
7102 }
7103
7104 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7105                                                   const u64 start,
7106                                                   const u64 len,
7107                                                   const u64 orig_start,
7108                                                   const u64 block_start,
7109                                                   const u64 block_len,
7110                                                   const u64 orig_block_len,
7111                                                   const u64 ram_bytes,
7112                                                   const int type)
7113 {
7114         struct extent_map *em = NULL;
7115         int ret;
7116
7117         if (type != BTRFS_ORDERED_NOCOW) {
7118                 em = create_io_em(inode, start, len, orig_start,
7119                                   block_start, block_len, orig_block_len,
7120                                   ram_bytes,
7121                                   BTRFS_COMPRESS_NONE, /* compress_type */
7122                                   type);
7123                 if (IS_ERR(em))
7124                         goto out;
7125         }
7126         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7127                                            len, block_len, type);
7128         if (ret) {
7129                 if (em) {
7130                         free_extent_map(em);
7131                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7132                                                 start + len - 1, 0);
7133                 }
7134                 em = ERR_PTR(ret);
7135         }
7136  out:
7137
7138         return em;
7139 }
7140
7141 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7142                                                   u64 start, u64 len)
7143 {
7144         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7145         struct btrfs_root *root = BTRFS_I(inode)->root;
7146         struct extent_map *em;
7147         struct btrfs_key ins;
7148         u64 alloc_hint;
7149         int ret;
7150
7151         alloc_hint = get_extent_allocation_hint(inode, start, len);
7152         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7153                                    0, alloc_hint, &ins, 1, 1);
7154         if (ret)
7155                 return ERR_PTR(ret);
7156
7157         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7158                                      ins.objectid, ins.offset, ins.offset,
7159                                      ins.offset, BTRFS_ORDERED_REGULAR);
7160         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7161         if (IS_ERR(em))
7162                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7163                                            ins.offset, 1);
7164
7165         return em;
7166 }
7167
7168 /*
7169  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7170  * block must be cow'd
7171  */
7172 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7173                               u64 *orig_start, u64 *orig_block_len,
7174                               u64 *ram_bytes)
7175 {
7176         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7177         struct btrfs_path *path;
7178         int ret;
7179         struct extent_buffer *leaf;
7180         struct btrfs_root *root = BTRFS_I(inode)->root;
7181         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7182         struct btrfs_file_extent_item *fi;
7183         struct btrfs_key key;
7184         u64 disk_bytenr;
7185         u64 backref_offset;
7186         u64 extent_end;
7187         u64 num_bytes;
7188         int slot;
7189         int found_type;
7190         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7191
7192         path = btrfs_alloc_path();
7193         if (!path)
7194                 return -ENOMEM;
7195
7196         ret = btrfs_lookup_file_extent(NULL, root, path,
7197                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7198         if (ret < 0)
7199                 goto out;
7200
7201         slot = path->slots[0];
7202         if (ret == 1) {
7203                 if (slot == 0) {
7204                         /* can't find the item, must cow */
7205                         ret = 0;
7206                         goto out;
7207                 }
7208                 slot--;
7209         }
7210         ret = 0;
7211         leaf = path->nodes[0];
7212         btrfs_item_key_to_cpu(leaf, &key, slot);
7213         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7214             key.type != BTRFS_EXTENT_DATA_KEY) {
7215                 /* not our file or wrong item type, must cow */
7216                 goto out;
7217         }
7218
7219         if (key.offset > offset) {
7220                 /* Wrong offset, must cow */
7221                 goto out;
7222         }
7223
7224         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7225         found_type = btrfs_file_extent_type(leaf, fi);
7226         if (found_type != BTRFS_FILE_EXTENT_REG &&
7227             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7228                 /* not a regular extent, must cow */
7229                 goto out;
7230         }
7231
7232         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7233                 goto out;
7234
7235         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7236         if (extent_end <= offset)
7237                 goto out;
7238
7239         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7240         if (disk_bytenr == 0)
7241                 goto out;
7242
7243         if (btrfs_file_extent_compression(leaf, fi) ||
7244             btrfs_file_extent_encryption(leaf, fi) ||
7245             btrfs_file_extent_other_encoding(leaf, fi))
7246                 goto out;
7247
7248         /*
7249          * Do the same check as in btrfs_cross_ref_exist but without the
7250          * unnecessary search.
7251          */
7252         if (btrfs_file_extent_generation(leaf, fi) <=
7253             btrfs_root_last_snapshot(&root->root_item))
7254                 goto out;
7255
7256         backref_offset = btrfs_file_extent_offset(leaf, fi);
7257
7258         if (orig_start) {
7259                 *orig_start = key.offset - backref_offset;
7260                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7261                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7262         }
7263
7264         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7265                 goto out;
7266
7267         num_bytes = min(offset + *len, extent_end) - offset;
7268         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7269                 u64 range_end;
7270
7271                 range_end = round_up(offset + num_bytes,
7272                                      root->fs_info->sectorsize) - 1;
7273                 ret = test_range_bit(io_tree, offset, range_end,
7274                                      EXTENT_DELALLOC, 0, NULL);
7275                 if (ret) {
7276                         ret = -EAGAIN;
7277                         goto out;
7278                 }
7279         }
7280
7281         btrfs_release_path(path);
7282
7283         /*
7284          * look for other files referencing this extent, if we
7285          * find any we must cow
7286          */
7287
7288         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7289                                     key.offset - backref_offset, disk_bytenr);
7290         if (ret) {
7291                 ret = 0;
7292                 goto out;
7293         }
7294
7295         /*
7296          * adjust disk_bytenr and num_bytes to cover just the bytes
7297          * in this extent we are about to write.  If there
7298          * are any csums in that range we have to cow in order
7299          * to keep the csums correct
7300          */
7301         disk_bytenr += backref_offset;
7302         disk_bytenr += offset - key.offset;
7303         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7304                 goto out;
7305         /*
7306          * all of the above have passed, it is safe to overwrite this extent
7307          * without cow
7308          */
7309         *len = num_bytes;
7310         ret = 1;
7311 out:
7312         btrfs_free_path(path);
7313         return ret;
7314 }
7315
7316 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7317                               struct extent_state **cached_state, int writing)
7318 {
7319         struct btrfs_ordered_extent *ordered;
7320         int ret = 0;
7321
7322         while (1) {
7323                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7324                                  cached_state);
7325                 /*
7326                  * We're concerned with the entire range that we're going to be
7327                  * doing DIO to, so we need to make sure there's no ordered
7328                  * extents in this range.
7329                  */
7330                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7331                                                      lockend - lockstart + 1);
7332
7333                 /*
7334                  * We need to make sure there are no buffered pages in this
7335                  * range either, we could have raced between the invalidate in
7336                  * generic_file_direct_write and locking the extent.  The
7337                  * invalidate needs to happen so that reads after a write do not
7338                  * get stale data.
7339                  */
7340                 if (!ordered &&
7341                     (!writing || !filemap_range_has_page(inode->i_mapping,
7342                                                          lockstart, lockend)))
7343                         break;
7344
7345                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7346                                      cached_state);
7347
7348                 if (ordered) {
7349                         /*
7350                          * If we are doing a DIO read and the ordered extent we
7351                          * found is for a buffered write, we can not wait for it
7352                          * to complete and retry, because if we do so we can
7353                          * deadlock with concurrent buffered writes on page
7354                          * locks. This happens only if our DIO read covers more
7355                          * than one extent map, if at this point has already
7356                          * created an ordered extent for a previous extent map
7357                          * and locked its range in the inode's io tree, and a
7358                          * concurrent write against that previous extent map's
7359                          * range and this range started (we unlock the ranges
7360                          * in the io tree only when the bios complete and
7361                          * buffered writes always lock pages before attempting
7362                          * to lock range in the io tree).
7363                          */
7364                         if (writing ||
7365                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7366                                 btrfs_start_ordered_extent(inode, ordered, 1);
7367                         else
7368                                 ret = -ENOTBLK;
7369                         btrfs_put_ordered_extent(ordered);
7370                 } else {
7371                         /*
7372                          * We could trigger writeback for this range (and wait
7373                          * for it to complete) and then invalidate the pages for
7374                          * this range (through invalidate_inode_pages2_range()),
7375                          * but that can lead us to a deadlock with a concurrent
7376                          * call to readpages() (a buffered read or a defrag call
7377                          * triggered a readahead) on a page lock due to an
7378                          * ordered dio extent we created before but did not have
7379                          * yet a corresponding bio submitted (whence it can not
7380                          * complete), which makes readpages() wait for that
7381                          * ordered extent to complete while holding a lock on
7382                          * that page.
7383                          */
7384                         ret = -ENOTBLK;
7385                 }
7386
7387                 if (ret)
7388                         break;
7389
7390                 cond_resched();
7391         }
7392
7393         return ret;
7394 }
7395
7396 /* The callers of this must take lock_extent() */
7397 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7398                                        u64 orig_start, u64 block_start,
7399                                        u64 block_len, u64 orig_block_len,
7400                                        u64 ram_bytes, int compress_type,
7401                                        int type)
7402 {
7403         struct extent_map_tree *em_tree;
7404         struct extent_map *em;
7405         struct btrfs_root *root = BTRFS_I(inode)->root;
7406         int ret;
7407
7408         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7409                type == BTRFS_ORDERED_COMPRESSED ||
7410                type == BTRFS_ORDERED_NOCOW ||
7411                type == BTRFS_ORDERED_REGULAR);
7412
7413         em_tree = &BTRFS_I(inode)->extent_tree;
7414         em = alloc_extent_map();
7415         if (!em)
7416                 return ERR_PTR(-ENOMEM);
7417
7418         em->start = start;
7419         em->orig_start = orig_start;
7420         em->len = len;
7421         em->block_len = block_len;
7422         em->block_start = block_start;
7423         em->bdev = root->fs_info->fs_devices->latest_bdev;
7424         em->orig_block_len = orig_block_len;
7425         em->ram_bytes = ram_bytes;
7426         em->generation = -1;
7427         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7428         if (type == BTRFS_ORDERED_PREALLOC) {
7429                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7430         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7431                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7432                 em->compress_type = compress_type;
7433         }
7434
7435         do {
7436                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7437                                 em->start + em->len - 1, 0);
7438                 write_lock(&em_tree->lock);
7439                 ret = add_extent_mapping(em_tree, em, 1);
7440                 write_unlock(&em_tree->lock);
7441                 /*
7442                  * The caller has taken lock_extent(), who could race with us
7443                  * to add em?
7444                  */
7445         } while (ret == -EEXIST);
7446
7447         if (ret) {
7448                 free_extent_map(em);
7449                 return ERR_PTR(ret);
7450         }
7451
7452         /* em got 2 refs now, callers needs to do free_extent_map once. */
7453         return em;
7454 }
7455
7456
7457 static int btrfs_get_blocks_direct_read(struct extent_map *em,
7458                                         struct buffer_head *bh_result,
7459                                         struct inode *inode,
7460                                         u64 start, u64 len)
7461 {
7462         if (em->block_start == EXTENT_MAP_HOLE ||
7463                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7464                 return -ENOENT;
7465
7466         len = min(len, em->len - (start - em->start));
7467
7468         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7469                 inode->i_blkbits;
7470         bh_result->b_size = len;
7471         bh_result->b_bdev = em->bdev;
7472         set_buffer_mapped(bh_result);
7473
7474         return 0;
7475 }
7476
7477 static int btrfs_get_blocks_direct_write(struct extent_map **map,
7478                                          struct buffer_head *bh_result,
7479                                          struct inode *inode,
7480                                          struct btrfs_dio_data *dio_data,
7481                                          u64 start, u64 len)
7482 {
7483         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7484         struct extent_map *em = *map;
7485         int ret = 0;
7486
7487         /*
7488          * We don't allocate a new extent in the following cases
7489          *
7490          * 1) The inode is marked as NODATACOW. In this case we'll just use the
7491          * existing extent.
7492          * 2) The extent is marked as PREALLOC. We're good to go here and can
7493          * just use the extent.
7494          *
7495          */
7496         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7497             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7498              em->block_start != EXTENT_MAP_HOLE)) {
7499                 int type;
7500                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7501
7502                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7503                         type = BTRFS_ORDERED_PREALLOC;
7504                 else
7505                         type = BTRFS_ORDERED_NOCOW;
7506                 len = min(len, em->len - (start - em->start));
7507                 block_start = em->block_start + (start - em->start);
7508
7509                 if (can_nocow_extent(inode, start, &len, &orig_start,
7510                                      &orig_block_len, &ram_bytes) == 1 &&
7511                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7512                         struct extent_map *em2;
7513
7514                         em2 = btrfs_create_dio_extent(inode, start, len,
7515                                                       orig_start, block_start,
7516                                                       len, orig_block_len,
7517                                                       ram_bytes, type);
7518                         btrfs_dec_nocow_writers(fs_info, block_start);
7519                         if (type == BTRFS_ORDERED_PREALLOC) {
7520                                 free_extent_map(em);
7521                                 *map = em = em2;
7522                         }
7523
7524                         if (em2 && IS_ERR(em2)) {
7525                                 ret = PTR_ERR(em2);
7526                                 goto out;
7527                         }
7528                         /*
7529                          * For inode marked NODATACOW or extent marked PREALLOC,
7530                          * use the existing or preallocated extent, so does not
7531                          * need to adjust btrfs_space_info's bytes_may_use.
7532                          */
7533                         btrfs_free_reserved_data_space_noquota(inode, start,
7534                                                                len);
7535                         goto skip_cow;
7536                 }
7537         }
7538
7539         /* this will cow the extent */
7540         len = bh_result->b_size;
7541         free_extent_map(em);
7542         *map = em = btrfs_new_extent_direct(inode, start, len);
7543         if (IS_ERR(em)) {
7544                 ret = PTR_ERR(em);
7545                 goto out;
7546         }
7547
7548         len = min(len, em->len - (start - em->start));
7549
7550 skip_cow:
7551         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7552                 inode->i_blkbits;
7553         bh_result->b_size = len;
7554         bh_result->b_bdev = em->bdev;
7555         set_buffer_mapped(bh_result);
7556
7557         if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7558                 set_buffer_new(bh_result);
7559
7560         /*
7561          * Need to update the i_size under the extent lock so buffered
7562          * readers will get the updated i_size when we unlock.
7563          */
7564         if (!dio_data->overwrite && start + len > i_size_read(inode))
7565                 i_size_write(inode, start + len);
7566
7567         WARN_ON(dio_data->reserve < len);
7568         dio_data->reserve -= len;
7569         dio_data->unsubmitted_oe_range_end = start + len;
7570         current->journal_info = dio_data;
7571 out:
7572         return ret;
7573 }
7574
7575 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7576                                    struct buffer_head *bh_result, int create)
7577 {
7578         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7579         struct extent_map *em;
7580         struct extent_state *cached_state = NULL;
7581         struct btrfs_dio_data *dio_data = NULL;
7582         u64 start = iblock << inode->i_blkbits;
7583         u64 lockstart, lockend;
7584         u64 len = bh_result->b_size;
7585         int unlock_bits = EXTENT_LOCKED;
7586         int ret = 0;
7587
7588         if (create)
7589                 unlock_bits |= EXTENT_DIRTY;
7590         else
7591                 len = min_t(u64, len, fs_info->sectorsize);
7592
7593         lockstart = start;
7594         lockend = start + len - 1;
7595
7596         if (current->journal_info) {
7597                 /*
7598                  * Need to pull our outstanding extents and set journal_info to NULL so
7599                  * that anything that needs to check if there's a transaction doesn't get
7600                  * confused.
7601                  */
7602                 dio_data = current->journal_info;
7603                 current->journal_info = NULL;
7604         }
7605
7606         /*
7607          * If this errors out it's because we couldn't invalidate pagecache for
7608          * this range and we need to fallback to buffered.
7609          */
7610         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7611                                create)) {
7612                 ret = -ENOTBLK;
7613                 goto err;
7614         }
7615
7616         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7617         if (IS_ERR(em)) {
7618                 ret = PTR_ERR(em);
7619                 goto unlock_err;
7620         }
7621
7622         /*
7623          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7624          * io.  INLINE is special, and we could probably kludge it in here, but
7625          * it's still buffered so for safety lets just fall back to the generic
7626          * buffered path.
7627          *
7628          * For COMPRESSED we _have_ to read the entire extent in so we can
7629          * decompress it, so there will be buffering required no matter what we
7630          * do, so go ahead and fallback to buffered.
7631          *
7632          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7633          * to buffered IO.  Don't blame me, this is the price we pay for using
7634          * the generic code.
7635          */
7636         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7637             em->block_start == EXTENT_MAP_INLINE) {
7638                 free_extent_map(em);
7639                 ret = -ENOTBLK;
7640                 goto unlock_err;
7641         }
7642
7643         if (create) {
7644                 ret = btrfs_get_blocks_direct_write(&em, bh_result, inode,
7645                                                     dio_data, start, len);
7646                 if (ret < 0)
7647                         goto unlock_err;
7648
7649                 /* clear and unlock the entire range */
7650                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7651                                  unlock_bits, 1, 0, &cached_state);
7652         } else {
7653                 ret = btrfs_get_blocks_direct_read(em, bh_result, inode,
7654                                                    start, len);
7655                 /* Can be negative only if we read from a hole */
7656                 if (ret < 0) {
7657                         ret = 0;
7658                         free_extent_map(em);
7659                         goto unlock_err;
7660                 }
7661                 /*
7662                  * We need to unlock only the end area that we aren't using.
7663                  * The rest is going to be unlocked by the endio routine.
7664                  */
7665                 lockstart = start + bh_result->b_size;
7666                 if (lockstart < lockend) {
7667                         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7668                                          lockend, unlock_bits, 1, 0,
7669                                          &cached_state);
7670                 } else {
7671                         free_extent_state(cached_state);
7672                 }
7673         }
7674
7675         free_extent_map(em);
7676
7677         return 0;
7678
7679 unlock_err:
7680         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7681                          unlock_bits, 1, 0, &cached_state);
7682 err:
7683         if (dio_data)
7684                 current->journal_info = dio_data;
7685         return ret;
7686 }
7687
7688 static inline blk_status_t submit_dio_repair_bio(struct inode *inode,
7689                                                  struct bio *bio,
7690                                                  int mirror_num)
7691 {
7692         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7693         blk_status_t ret;
7694
7695         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7696
7697         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7698         if (ret)
7699                 return ret;
7700
7701         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
7702
7703         return ret;
7704 }
7705
7706 static int btrfs_check_dio_repairable(struct inode *inode,
7707                                       struct bio *failed_bio,
7708                                       struct io_failure_record *failrec,
7709                                       int failed_mirror)
7710 {
7711         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7712         int num_copies;
7713
7714         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7715         if (num_copies == 1) {
7716                 /*
7717                  * we only have a single copy of the data, so don't bother with
7718                  * all the retry and error correction code that follows. no
7719                  * matter what the error is, it is very likely to persist.
7720                  */
7721                 btrfs_debug(fs_info,
7722                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7723                         num_copies, failrec->this_mirror, failed_mirror);
7724                 return 0;
7725         }
7726
7727         failrec->failed_mirror = failed_mirror;
7728         failrec->this_mirror++;
7729         if (failrec->this_mirror == failed_mirror)
7730                 failrec->this_mirror++;
7731
7732         if (failrec->this_mirror > num_copies) {
7733                 btrfs_debug(fs_info,
7734                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7735                         num_copies, failrec->this_mirror, failed_mirror);
7736                 return 0;
7737         }
7738
7739         return 1;
7740 }
7741
7742 static blk_status_t dio_read_error(struct inode *inode, struct bio *failed_bio,
7743                                    struct page *page, unsigned int pgoff,
7744                                    u64 start, u64 end, int failed_mirror,
7745                                    bio_end_io_t *repair_endio, void *repair_arg)
7746 {
7747         struct io_failure_record *failrec;
7748         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7749         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
7750         struct bio *bio;
7751         int isector;
7752         unsigned int read_mode = 0;
7753         int segs;
7754         int ret;
7755         blk_status_t status;
7756         struct bio_vec bvec;
7757
7758         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
7759
7760         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7761         if (ret)
7762                 return errno_to_blk_status(ret);
7763
7764         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7765                                          failed_mirror);
7766         if (!ret) {
7767                 free_io_failure(failure_tree, io_tree, failrec);
7768                 return BLK_STS_IOERR;
7769         }
7770
7771         segs = bio_segments(failed_bio);
7772         bio_get_first_bvec(failed_bio, &bvec);
7773         if (segs > 1 ||
7774             (bvec.bv_len > btrfs_inode_sectorsize(inode)))
7775                 read_mode |= REQ_FAILFAST_DEV;
7776
7777         isector = start - btrfs_io_bio(failed_bio)->logical;
7778         isector >>= inode->i_sb->s_blocksize_bits;
7779         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7780                                 pgoff, isector, repair_endio, repair_arg);
7781         bio->bi_opf = REQ_OP_READ | read_mode;
7782
7783         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7784                     "repair DIO read error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d",
7785                     read_mode, failrec->this_mirror, failrec->in_validation);
7786
7787         status = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
7788         if (status) {
7789                 free_io_failure(failure_tree, io_tree, failrec);
7790                 bio_put(bio);
7791         }
7792
7793         return status;
7794 }
7795
7796 struct btrfs_retry_complete {
7797         struct completion done;
7798         struct inode *inode;
7799         u64 start;
7800         int uptodate;
7801 };
7802
7803 static void btrfs_retry_endio_nocsum(struct bio *bio)
7804 {
7805         struct btrfs_retry_complete *done = bio->bi_private;
7806         struct inode *inode = done->inode;
7807         struct bio_vec *bvec;
7808         struct extent_io_tree *io_tree, *failure_tree;
7809         int i;
7810
7811         if (bio->bi_status)
7812                 goto end;
7813
7814         ASSERT(bio->bi_vcnt == 1);
7815         io_tree = &BTRFS_I(inode)->io_tree;
7816         failure_tree = &BTRFS_I(inode)->io_failure_tree;
7817         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(inode));
7818
7819         done->uptodate = 1;
7820         ASSERT(!bio_flagged(bio, BIO_CLONED));
7821         bio_for_each_segment_all(bvec, bio, i)
7822                 clean_io_failure(BTRFS_I(inode)->root->fs_info, failure_tree,
7823                                  io_tree, done->start, bvec->bv_page,
7824                                  btrfs_ino(BTRFS_I(inode)), 0);
7825 end:
7826         complete(&done->done);
7827         bio_put(bio);
7828 }
7829
7830 static blk_status_t __btrfs_correct_data_nocsum(struct inode *inode,
7831                                                 struct btrfs_io_bio *io_bio)
7832 {
7833         struct btrfs_fs_info *fs_info;
7834         struct bio_vec bvec;
7835         struct bvec_iter iter;
7836         struct btrfs_retry_complete done;
7837         u64 start;
7838         unsigned int pgoff;
7839         u32 sectorsize;
7840         int nr_sectors;
7841         blk_status_t ret;
7842         blk_status_t err = BLK_STS_OK;
7843
7844         fs_info = BTRFS_I(inode)->root->fs_info;
7845         sectorsize = fs_info->sectorsize;
7846
7847         start = io_bio->logical;
7848         done.inode = inode;
7849         io_bio->bio.bi_iter = io_bio->iter;
7850
7851         bio_for_each_segment(bvec, &io_bio->bio, iter) {
7852                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
7853                 pgoff = bvec.bv_offset;
7854
7855 next_block_or_try_again:
7856                 done.uptodate = 0;
7857                 done.start = start;
7858                 init_completion(&done.done);
7859
7860                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
7861                                 pgoff, start, start + sectorsize - 1,
7862                                 io_bio->mirror_num,
7863                                 btrfs_retry_endio_nocsum, &done);
7864                 if (ret) {
7865                         err = ret;
7866                         goto next;
7867                 }
7868
7869                 wait_for_completion_io(&done.done);
7870
7871                 if (!done.uptodate) {
7872                         /* We might have another mirror, so try again */
7873                         goto next_block_or_try_again;
7874                 }
7875
7876 next:
7877                 start += sectorsize;
7878
7879                 nr_sectors--;
7880                 if (nr_sectors) {
7881                         pgoff += sectorsize;
7882                         ASSERT(pgoff < PAGE_SIZE);
7883                         goto next_block_or_try_again;
7884                 }
7885         }
7886
7887         return err;
7888 }
7889
7890 static void btrfs_retry_endio(struct bio *bio)
7891 {
7892         struct btrfs_retry_complete *done = bio->bi_private;
7893         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7894         struct extent_io_tree *io_tree, *failure_tree;
7895         struct inode *inode = done->inode;
7896         struct bio_vec *bvec;
7897         int uptodate;
7898         int ret;
7899         int i;
7900
7901         if (bio->bi_status)
7902                 goto end;
7903
7904         uptodate = 1;
7905
7906         ASSERT(bio->bi_vcnt == 1);
7907         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(done->inode));
7908
7909         io_tree = &BTRFS_I(inode)->io_tree;
7910         failure_tree = &BTRFS_I(inode)->io_failure_tree;
7911
7912         ASSERT(!bio_flagged(bio, BIO_CLONED));
7913         bio_for_each_segment_all(bvec, bio, i) {
7914                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
7915                                              bvec->bv_offset, done->start,
7916                                              bvec->bv_len);
7917                 if (!ret)
7918                         clean_io_failure(BTRFS_I(inode)->root->fs_info,
7919                                          failure_tree, io_tree, done->start,
7920                                          bvec->bv_page,
7921                                          btrfs_ino(BTRFS_I(inode)),
7922                                          bvec->bv_offset);
7923                 else
7924                         uptodate = 0;
7925         }
7926
7927         done->uptodate = uptodate;
7928 end:
7929         complete(&done->done);
7930         bio_put(bio);
7931 }
7932
7933 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
7934                 struct btrfs_io_bio *io_bio, blk_status_t err)
7935 {
7936         struct btrfs_fs_info *fs_info;
7937         struct bio_vec bvec;
7938         struct bvec_iter iter;
7939         struct btrfs_retry_complete done;
7940         u64 start;
7941         u64 offset = 0;
7942         u32 sectorsize;
7943         int nr_sectors;
7944         unsigned int pgoff;
7945         int csum_pos;
7946         bool uptodate = (err == 0);
7947         int ret;
7948         blk_status_t status;
7949
7950         fs_info = BTRFS_I(inode)->root->fs_info;
7951         sectorsize = fs_info->sectorsize;
7952
7953         err = BLK_STS_OK;
7954         start = io_bio->logical;
7955         done.inode = inode;
7956         io_bio->bio.bi_iter = io_bio->iter;
7957
7958         bio_for_each_segment(bvec, &io_bio->bio, iter) {
7959                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
7960
7961                 pgoff = bvec.bv_offset;
7962 next_block:
7963                 if (uptodate) {
7964                         csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
7965                         ret = __readpage_endio_check(inode, io_bio, csum_pos,
7966                                         bvec.bv_page, pgoff, start, sectorsize);
7967                         if (likely(!ret))
7968                                 goto next;
7969                 }
7970 try_again:
7971                 done.uptodate = 0;
7972                 done.start = start;
7973                 init_completion(&done.done);
7974
7975                 status = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
7976                                         pgoff, start, start + sectorsize - 1,
7977                                         io_bio->mirror_num, btrfs_retry_endio,
7978                                         &done);
7979                 if (status) {
7980                         err = status;
7981                         goto next;
7982                 }
7983
7984                 wait_for_completion_io(&done.done);
7985
7986                 if (!done.uptodate) {
7987                         /* We might have another mirror, so try again */
7988                         goto try_again;
7989                 }
7990 next:
7991                 offset += sectorsize;
7992                 start += sectorsize;
7993
7994                 ASSERT(nr_sectors);
7995
7996                 nr_sectors--;
7997                 if (nr_sectors) {
7998                         pgoff += sectorsize;
7999                         ASSERT(pgoff < PAGE_SIZE);
8000                         goto next_block;
8001                 }
8002         }
8003
8004         return err;
8005 }
8006
8007 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
8008                 struct btrfs_io_bio *io_bio, blk_status_t err)
8009 {
8010         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8011
8012         if (skip_csum) {
8013                 if (unlikely(err))
8014                         return __btrfs_correct_data_nocsum(inode, io_bio);
8015                 else
8016                         return BLK_STS_OK;
8017         } else {
8018                 return __btrfs_subio_endio_read(inode, io_bio, err);
8019         }
8020 }
8021
8022 static void btrfs_endio_direct_read(struct bio *bio)
8023 {
8024         struct btrfs_dio_private *dip = bio->bi_private;
8025         struct inode *inode = dip->inode;
8026         struct bio *dio_bio;
8027         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8028         blk_status_t err = bio->bi_status;
8029
8030         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8031                 err = btrfs_subio_endio_read(inode, io_bio, err);
8032
8033         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8034                       dip->logical_offset + dip->bytes - 1);
8035         dio_bio = dip->dio_bio;
8036
8037         kfree(dip);
8038
8039         dio_bio->bi_status = err;
8040         dio_end_io(dio_bio);
8041
8042         if (io_bio->end_io)
8043                 io_bio->end_io(io_bio, blk_status_to_errno(err));
8044         bio_put(bio);
8045 }
8046
8047 static void __endio_write_update_ordered(struct inode *inode,
8048                                          const u64 offset, const u64 bytes,
8049                                          const bool uptodate)
8050 {
8051         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8052         struct btrfs_ordered_extent *ordered = NULL;
8053         struct btrfs_workqueue *wq;
8054         btrfs_work_func_t func;
8055         u64 ordered_offset = offset;
8056         u64 ordered_bytes = bytes;
8057         u64 last_offset;
8058
8059         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
8060                 wq = fs_info->endio_freespace_worker;
8061                 func = btrfs_freespace_write_helper;
8062         } else {
8063                 wq = fs_info->endio_write_workers;
8064                 func = btrfs_endio_write_helper;
8065         }
8066
8067         while (ordered_offset < offset + bytes) {
8068                 last_offset = ordered_offset;
8069                 if (btrfs_dec_test_first_ordered_pending(inode, &ordered,
8070                                                            &ordered_offset,
8071                                                            ordered_bytes,
8072                                                            uptodate)) {
8073                         btrfs_init_work(&ordered->work, func,
8074                                         finish_ordered_fn,
8075                                         NULL, NULL);
8076                         btrfs_queue_work(wq, &ordered->work);
8077                 }
8078                 /*
8079                  * If btrfs_dec_test_ordered_pending does not find any ordered
8080                  * extent in the range, we can exit.
8081                  */
8082                 if (ordered_offset == last_offset)
8083                         return;
8084                 /*
8085                  * Our bio might span multiple ordered extents. In this case
8086                  * we keep goin until we have accounted the whole dio.
8087                  */
8088                 if (ordered_offset < offset + bytes) {
8089                         ordered_bytes = offset + bytes - ordered_offset;
8090                         ordered = NULL;
8091                 }
8092         }
8093 }
8094
8095 static void btrfs_endio_direct_write(struct bio *bio)
8096 {
8097         struct btrfs_dio_private *dip = bio->bi_private;
8098         struct bio *dio_bio = dip->dio_bio;
8099
8100         __endio_write_update_ordered(dip->inode, dip->logical_offset,
8101                                      dip->bytes, !bio->bi_status);
8102
8103         kfree(dip);
8104
8105         dio_bio->bi_status = bio->bi_status;
8106         dio_end_io(dio_bio);
8107         bio_put(bio);
8108 }
8109
8110 static blk_status_t btrfs_submit_bio_start_direct_io(void *private_data,
8111                                     struct bio *bio, u64 offset)
8112 {
8113         struct inode *inode = private_data;
8114         blk_status_t ret;
8115         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8116         BUG_ON(ret); /* -ENOMEM */
8117         return 0;
8118 }
8119
8120 static void btrfs_end_dio_bio(struct bio *bio)
8121 {
8122         struct btrfs_dio_private *dip = bio->bi_private;
8123         blk_status_t err = bio->bi_status;
8124
8125         if (err)
8126                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8127                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8128                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8129                            bio->bi_opf,
8130                            (unsigned long long)bio->bi_iter.bi_sector,
8131                            bio->bi_iter.bi_size, err);
8132
8133         if (dip->subio_endio)
8134                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8135
8136         if (err) {
8137                 /*
8138                  * We want to perceive the errors flag being set before
8139                  * decrementing the reference count. We don't need a barrier
8140                  * since atomic operations with a return value are fully
8141                  * ordered as per atomic_t.txt
8142                  */
8143                 dip->errors = 1;
8144         }
8145
8146         /* if there are more bios still pending for this dio, just exit */
8147         if (!atomic_dec_and_test(&dip->pending_bios))
8148                 goto out;
8149
8150         if (dip->errors) {
8151                 bio_io_error(dip->orig_bio);
8152         } else {
8153                 dip->dio_bio->bi_status = BLK_STS_OK;
8154                 bio_endio(dip->orig_bio);
8155         }
8156 out:
8157         bio_put(bio);
8158 }
8159
8160 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8161                                                  struct btrfs_dio_private *dip,
8162                                                  struct bio *bio,
8163                                                  u64 file_offset)
8164 {
8165         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8166         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8167         blk_status_t ret;
8168
8169         /*
8170          * We load all the csum data we need when we submit
8171          * the first bio to reduce the csum tree search and
8172          * contention.
8173          */
8174         if (dip->logical_offset == file_offset) {
8175                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8176                                                 file_offset);
8177                 if (ret)
8178                         return ret;
8179         }
8180
8181         if (bio == dip->orig_bio)
8182                 return 0;
8183
8184         file_offset -= dip->logical_offset;
8185         file_offset >>= inode->i_sb->s_blocksize_bits;
8186         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8187
8188         return 0;
8189 }
8190
8191 static inline blk_status_t btrfs_submit_dio_bio(struct bio *bio,
8192                 struct inode *inode, u64 file_offset, int async_submit)
8193 {
8194         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8195         struct btrfs_dio_private *dip = bio->bi_private;
8196         bool write = bio_op(bio) == REQ_OP_WRITE;
8197         blk_status_t ret;
8198
8199         /* Check btrfs_submit_bio_hook() for rules about async submit. */
8200         if (async_submit)
8201                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8202
8203         if (!write) {
8204                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8205                 if (ret)
8206                         goto err;
8207         }
8208
8209         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
8210                 goto map;
8211
8212         if (write && async_submit) {
8213                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
8214                                           file_offset, inode,
8215                                           btrfs_submit_bio_start_direct_io);
8216                 goto err;
8217         } else if (write) {
8218                 /*
8219                  * If we aren't doing async submit, calculate the csum of the
8220                  * bio now.
8221                  */
8222                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8223                 if (ret)
8224                         goto err;
8225         } else {
8226                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8227                                                      file_offset);
8228                 if (ret)
8229                         goto err;
8230         }
8231 map:
8232         ret = btrfs_map_bio(fs_info, bio, 0, 0);
8233 err:
8234         return ret;
8235 }
8236
8237 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
8238 {
8239         struct inode *inode = dip->inode;
8240         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8241         struct bio *bio;
8242         struct bio *orig_bio = dip->orig_bio;
8243         u64 start_sector = orig_bio->bi_iter.bi_sector;
8244         u64 file_offset = dip->logical_offset;
8245         u64 map_length;
8246         int async_submit = 0;
8247         u64 submit_len;
8248         int clone_offset = 0;
8249         int clone_len;
8250         int ret;
8251         blk_status_t status;
8252
8253         map_length = orig_bio->bi_iter.bi_size;
8254         submit_len = map_length;
8255         ret = btrfs_map_block(fs_info, btrfs_op(orig_bio), start_sector << 9,
8256                               &map_length, NULL, 0);
8257         if (ret)
8258                 return -EIO;
8259
8260         if (map_length >= submit_len) {
8261                 bio = orig_bio;
8262                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8263                 goto submit;
8264         }
8265
8266         /* async crcs make it difficult to collect full stripe writes. */
8267         if (btrfs_data_alloc_profile(fs_info) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8268                 async_submit = 0;
8269         else
8270                 async_submit = 1;
8271
8272         /* bio split */
8273         ASSERT(map_length <= INT_MAX);
8274         atomic_inc(&dip->pending_bios);
8275         do {
8276                 clone_len = min_t(int, submit_len, map_length);
8277
8278                 /*
8279                  * This will never fail as it's passing GPF_NOFS and
8280                  * the allocation is backed by btrfs_bioset.
8281                  */
8282                 bio = btrfs_bio_clone_partial(orig_bio, clone_offset,
8283                                               clone_len);
8284                 bio->bi_private = dip;
8285                 bio->bi_end_io = btrfs_end_dio_bio;
8286                 btrfs_io_bio(bio)->logical = file_offset;
8287
8288                 ASSERT(submit_len >= clone_len);
8289                 submit_len -= clone_len;
8290                 if (submit_len == 0)
8291                         break;
8292
8293                 /*
8294                  * Increase the count before we submit the bio so we know
8295                  * the end IO handler won't happen before we increase the
8296                  * count. Otherwise, the dip might get freed before we're
8297                  * done setting it up.
8298                  */
8299                 atomic_inc(&dip->pending_bios);
8300
8301                 status = btrfs_submit_dio_bio(bio, inode, file_offset,
8302                                                 async_submit);
8303                 if (status) {
8304                         bio_put(bio);
8305                         atomic_dec(&dip->pending_bios);
8306                         goto out_err;
8307                 }
8308
8309                 clone_offset += clone_len;
8310                 start_sector += clone_len >> 9;
8311                 file_offset += clone_len;
8312
8313                 map_length = submit_len;
8314                 ret = btrfs_map_block(fs_info, btrfs_op(orig_bio),
8315                                       start_sector << 9, &map_length, NULL, 0);
8316                 if (ret)
8317                         goto out_err;
8318         } while (submit_len > 0);
8319
8320 submit:
8321         status = btrfs_submit_dio_bio(bio, inode, file_offset, async_submit);
8322         if (!status)
8323                 return 0;
8324
8325         bio_put(bio);
8326 out_err:
8327         dip->errors = 1;
8328         /*
8329          * Before atomic variable goto zero, we must  make sure dip->errors is
8330          * perceived to be set. This ordering is ensured by the fact that an
8331          * atomic operations with a return value are fully ordered as per
8332          * atomic_t.txt
8333          */
8334         if (atomic_dec_and_test(&dip->pending_bios))
8335                 bio_io_error(dip->orig_bio);
8336
8337         /* bio_end_io() will handle error, so we needn't return it */
8338         return 0;
8339 }
8340
8341 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8342                                 loff_t file_offset)
8343 {
8344         struct btrfs_dio_private *dip = NULL;
8345         struct bio *bio = NULL;
8346         struct btrfs_io_bio *io_bio;
8347         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8348         int ret = 0;
8349
8350         bio = btrfs_bio_clone(dio_bio);
8351
8352         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8353         if (!dip) {
8354                 ret = -ENOMEM;
8355                 goto free_ordered;
8356         }
8357
8358         dip->private = dio_bio->bi_private;
8359         dip->inode = inode;
8360         dip->logical_offset = file_offset;
8361         dip->bytes = dio_bio->bi_iter.bi_size;
8362         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8363         bio->bi_private = dip;
8364         dip->orig_bio = bio;
8365         dip->dio_bio = dio_bio;
8366         atomic_set(&dip->pending_bios, 0);
8367         io_bio = btrfs_io_bio(bio);
8368         io_bio->logical = file_offset;
8369
8370         if (write) {
8371                 bio->bi_end_io = btrfs_endio_direct_write;
8372         } else {
8373                 bio->bi_end_io = btrfs_endio_direct_read;
8374                 dip->subio_endio = btrfs_subio_endio_read;
8375         }
8376
8377         /*
8378          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8379          * even if we fail to submit a bio, because in such case we do the
8380          * corresponding error handling below and it must not be done a second
8381          * time by btrfs_direct_IO().
8382          */
8383         if (write) {
8384                 struct btrfs_dio_data *dio_data = current->journal_info;
8385
8386                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8387                         dip->bytes;
8388                 dio_data->unsubmitted_oe_range_start =
8389                         dio_data->unsubmitted_oe_range_end;
8390         }
8391
8392         ret = btrfs_submit_direct_hook(dip);
8393         if (!ret)
8394                 return;
8395
8396         if (io_bio->end_io)
8397                 io_bio->end_io(io_bio, ret);
8398
8399 free_ordered:
8400         /*
8401          * If we arrived here it means either we failed to submit the dip
8402          * or we either failed to clone the dio_bio or failed to allocate the
8403          * dip. If we cloned the dio_bio and allocated the dip, we can just
8404          * call bio_endio against our io_bio so that we get proper resource
8405          * cleanup if we fail to submit the dip, otherwise, we must do the
8406          * same as btrfs_endio_direct_[write|read] because we can't call these
8407          * callbacks - they require an allocated dip and a clone of dio_bio.
8408          */
8409         if (bio && dip) {
8410                 bio_io_error(bio);
8411                 /*
8412                  * The end io callbacks free our dip, do the final put on bio
8413                  * and all the cleanup and final put for dio_bio (through
8414                  * dio_end_io()).
8415                  */
8416                 dip = NULL;
8417                 bio = NULL;
8418         } else {
8419                 if (write)
8420                         __endio_write_update_ordered(inode,
8421                                                 file_offset,
8422                                                 dio_bio->bi_iter.bi_size,
8423                                                 false);
8424                 else
8425                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8426                               file_offset + dio_bio->bi_iter.bi_size - 1);
8427
8428                 dio_bio->bi_status = BLK_STS_IOERR;
8429                 /*
8430                  * Releases and cleans up our dio_bio, no need to bio_put()
8431                  * nor bio_endio()/bio_io_error() against dio_bio.
8432                  */
8433                 dio_end_io(dio_bio);
8434         }
8435         if (bio)
8436                 bio_put(bio);
8437         kfree(dip);
8438 }
8439
8440 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8441                                const struct iov_iter *iter, loff_t offset)
8442 {
8443         int seg;
8444         int i;
8445         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8446         ssize_t retval = -EINVAL;
8447
8448         if (offset & blocksize_mask)
8449                 goto out;
8450
8451         if (iov_iter_alignment(iter) & blocksize_mask)
8452                 goto out;
8453
8454         /* If this is a write we don't need to check anymore */
8455         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8456                 return 0;
8457         /*
8458          * Check to make sure we don't have duplicate iov_base's in this
8459          * iovec, if so return EINVAL, otherwise we'll get csum errors
8460          * when reading back.
8461          */
8462         for (seg = 0; seg < iter->nr_segs; seg++) {
8463                 for (i = seg + 1; i < iter->nr_segs; i++) {
8464                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8465                                 goto out;
8466                 }
8467         }
8468         retval = 0;
8469 out:
8470         return retval;
8471 }
8472
8473 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8474 {
8475         struct file *file = iocb->ki_filp;
8476         struct inode *inode = file->f_mapping->host;
8477         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8478         struct btrfs_dio_data dio_data = { 0 };
8479         struct extent_changeset *data_reserved = NULL;
8480         loff_t offset = iocb->ki_pos;
8481         size_t count = 0;
8482         int flags = 0;
8483         bool wakeup = true;
8484         bool relock = false;
8485         ssize_t ret;
8486
8487         if (check_direct_IO(fs_info, iter, offset))
8488                 return 0;
8489
8490         inode_dio_begin(inode);
8491
8492         /*
8493          * The generic stuff only does filemap_write_and_wait_range, which
8494          * isn't enough if we've written compressed pages to this area, so
8495          * we need to flush the dirty pages again to make absolutely sure
8496          * that any outstanding dirty pages are on disk.
8497          */
8498         count = iov_iter_count(iter);
8499         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8500                      &BTRFS_I(inode)->runtime_flags))
8501                 filemap_fdatawrite_range(inode->i_mapping, offset,
8502                                          offset + count - 1);
8503
8504         if (iov_iter_rw(iter) == WRITE) {
8505                 /*
8506                  * If the write DIO is beyond the EOF, we need update
8507                  * the isize, but it is protected by i_mutex. So we can
8508                  * not unlock the i_mutex at this case.
8509                  */
8510                 if (offset + count <= inode->i_size) {
8511                         dio_data.overwrite = 1;
8512                         inode_unlock(inode);
8513                         relock = true;
8514                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8515                         ret = -EAGAIN;
8516                         goto out;
8517                 }
8518                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
8519                                                    offset, count);
8520                 if (ret)
8521                         goto out;
8522
8523                 /*
8524                  * We need to know how many extents we reserved so that we can
8525                  * do the accounting properly if we go over the number we
8526                  * originally calculated.  Abuse current->journal_info for this.
8527                  */
8528                 dio_data.reserve = round_up(count,
8529                                             fs_info->sectorsize);
8530                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8531                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8532                 current->journal_info = &dio_data;
8533                 down_read(&BTRFS_I(inode)->dio_sem);
8534         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8535                                      &BTRFS_I(inode)->runtime_flags)) {
8536                 inode_dio_end(inode);
8537                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8538                 wakeup = false;
8539         }
8540
8541         ret = __blockdev_direct_IO(iocb, inode,
8542                                    fs_info->fs_devices->latest_bdev,
8543                                    iter, btrfs_get_blocks_direct, NULL,
8544                                    btrfs_submit_direct, flags);
8545         if (iov_iter_rw(iter) == WRITE) {
8546                 up_read(&BTRFS_I(inode)->dio_sem);
8547                 current->journal_info = NULL;
8548                 if (ret < 0 && ret != -EIOCBQUEUED) {
8549                         if (dio_data.reserve)
8550                                 btrfs_delalloc_release_space(inode, data_reserved,
8551                                         offset, dio_data.reserve, true);
8552                         /*
8553                          * On error we might have left some ordered extents
8554                          * without submitting corresponding bios for them, so
8555                          * cleanup them up to avoid other tasks getting them
8556                          * and waiting for them to complete forever.
8557                          */
8558                         if (dio_data.unsubmitted_oe_range_start <
8559                             dio_data.unsubmitted_oe_range_end)
8560                                 __endio_write_update_ordered(inode,
8561                                         dio_data.unsubmitted_oe_range_start,
8562                                         dio_data.unsubmitted_oe_range_end -
8563                                         dio_data.unsubmitted_oe_range_start,
8564                                         false);
8565                 } else if (ret >= 0 && (size_t)ret < count)
8566                         btrfs_delalloc_release_space(inode, data_reserved,
8567                                         offset, count - (size_t)ret, true);
8568                 btrfs_delalloc_release_extents(BTRFS_I(inode), count, false);
8569         }
8570 out:
8571         if (wakeup)
8572                 inode_dio_end(inode);
8573         if (relock)
8574                 inode_lock(inode);
8575
8576         extent_changeset_free(data_reserved);
8577         return ret;
8578 }
8579
8580 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8581
8582 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8583                 __u64 start, __u64 len)
8584 {
8585         int     ret;
8586
8587         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8588         if (ret)
8589                 return ret;
8590
8591         return extent_fiemap(inode, fieinfo, start, len);
8592 }
8593
8594 int btrfs_readpage(struct file *file, struct page *page)
8595 {
8596         struct extent_io_tree *tree;
8597         tree = &BTRFS_I(page->mapping->host)->io_tree;
8598         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8599 }
8600
8601 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8602 {
8603         struct inode *inode = page->mapping->host;
8604         int ret;
8605
8606         if (current->flags & PF_MEMALLOC) {
8607                 redirty_page_for_writepage(wbc, page);
8608                 unlock_page(page);
8609                 return 0;
8610         }
8611
8612         /*
8613          * If we are under memory pressure we will call this directly from the
8614          * VM, we need to make sure we have the inode referenced for the ordered
8615          * extent.  If not just return like we didn't do anything.
8616          */
8617         if (!igrab(inode)) {
8618                 redirty_page_for_writepage(wbc, page);
8619                 return AOP_WRITEPAGE_ACTIVATE;
8620         }
8621         ret = extent_write_full_page(page, wbc);
8622         btrfs_add_delayed_iput(inode);
8623         return ret;
8624 }
8625
8626 static int btrfs_writepages(struct address_space *mapping,
8627                             struct writeback_control *wbc)
8628 {
8629         return extent_writepages(mapping, wbc);
8630 }
8631
8632 static int
8633 btrfs_readpages(struct file *file, struct address_space *mapping,
8634                 struct list_head *pages, unsigned nr_pages)
8635 {
8636         return extent_readpages(mapping, pages, nr_pages);
8637 }
8638
8639 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8640 {
8641         int ret = try_release_extent_mapping(page, gfp_flags);
8642         if (ret == 1) {
8643                 ClearPagePrivate(page);
8644                 set_page_private(page, 0);
8645                 put_page(page);
8646         }
8647         return ret;
8648 }
8649
8650 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8651 {
8652         if (PageWriteback(page) || PageDirty(page))
8653                 return 0;
8654         return __btrfs_releasepage(page, gfp_flags);
8655 }
8656
8657 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8658                                  unsigned int length)
8659 {
8660         struct inode *inode = page->mapping->host;
8661         struct extent_io_tree *tree;
8662         struct btrfs_ordered_extent *ordered;
8663         struct extent_state *cached_state = NULL;
8664         u64 page_start = page_offset(page);
8665         u64 page_end = page_start + PAGE_SIZE - 1;
8666         u64 start;
8667         u64 end;
8668         int inode_evicting = inode->i_state & I_FREEING;
8669
8670         /*
8671          * we have the page locked, so new writeback can't start,
8672          * and the dirty bit won't be cleared while we are here.
8673          *
8674          * Wait for IO on this page so that we can safely clear
8675          * the PagePrivate2 bit and do ordered accounting
8676          */
8677         wait_on_page_writeback(page);
8678
8679         tree = &BTRFS_I(inode)->io_tree;
8680         if (offset) {
8681                 btrfs_releasepage(page, GFP_NOFS);
8682                 return;
8683         }
8684
8685         if (!inode_evicting)
8686                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8687 again:
8688         start = page_start;
8689         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8690                                         page_end - start + 1);
8691         if (ordered) {
8692                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8693                 /*
8694                  * IO on this page will never be started, so we need
8695                  * to account for any ordered extents now
8696                  */
8697                 if (!inode_evicting)
8698                         clear_extent_bit(tree, start, end,
8699                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8700                                          EXTENT_DELALLOC_NEW |
8701                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8702                                          EXTENT_DEFRAG, 1, 0, &cached_state);
8703                 /*
8704                  * whoever cleared the private bit is responsible
8705                  * for the finish_ordered_io
8706                  */
8707                 if (TestClearPagePrivate2(page)) {
8708                         struct btrfs_ordered_inode_tree *tree;
8709                         u64 new_len;
8710
8711                         tree = &BTRFS_I(inode)->ordered_tree;
8712
8713                         spin_lock_irq(&tree->lock);
8714                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8715                         new_len = start - ordered->file_offset;
8716                         if (new_len < ordered->truncated_len)
8717                                 ordered->truncated_len = new_len;
8718                         spin_unlock_irq(&tree->lock);
8719
8720                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8721                                                            start,
8722                                                            end - start + 1, 1))
8723                                 btrfs_finish_ordered_io(ordered);
8724                 }
8725                 btrfs_put_ordered_extent(ordered);
8726                 if (!inode_evicting) {
8727                         cached_state = NULL;
8728                         lock_extent_bits(tree, start, end,
8729                                          &cached_state);
8730                 }
8731
8732                 start = end + 1;
8733                 if (start < page_end)
8734                         goto again;
8735         }
8736
8737         /*
8738          * Qgroup reserved space handler
8739          * Page here will be either
8740          * 1) Already written to disk
8741          *    In this case, its reserved space is released from data rsv map
8742          *    and will be freed by delayed_ref handler finally.
8743          *    So even we call qgroup_free_data(), it won't decrease reserved
8744          *    space.
8745          * 2) Not written to disk
8746          *    This means the reserved space should be freed here. However,
8747          *    if a truncate invalidates the page (by clearing PageDirty)
8748          *    and the page is accounted for while allocating extent
8749          *    in btrfs_check_data_free_space() we let delayed_ref to
8750          *    free the entire extent.
8751          */
8752         if (PageDirty(page))
8753                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
8754         if (!inode_evicting) {
8755                 clear_extent_bit(tree, page_start, page_end,
8756                                  EXTENT_LOCKED | EXTENT_DIRTY |
8757                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8758                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
8759                                  &cached_state);
8760
8761                 __btrfs_releasepage(page, GFP_NOFS);
8762         }
8763
8764         ClearPageChecked(page);
8765         if (PagePrivate(page)) {
8766                 ClearPagePrivate(page);
8767                 set_page_private(page, 0);
8768                 put_page(page);
8769         }
8770 }
8771
8772 /*
8773  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8774  * called from a page fault handler when a page is first dirtied. Hence we must
8775  * be careful to check for EOF conditions here. We set the page up correctly
8776  * for a written page which means we get ENOSPC checking when writing into
8777  * holes and correct delalloc and unwritten extent mapping on filesystems that
8778  * support these features.
8779  *
8780  * We are not allowed to take the i_mutex here so we have to play games to
8781  * protect against truncate races as the page could now be beyond EOF.  Because
8782  * truncate_setsize() writes the inode size before removing pages, once we have
8783  * the page lock we can determine safely if the page is beyond EOF. If it is not
8784  * beyond EOF, then the page is guaranteed safe against truncation until we
8785  * unlock the page.
8786  */
8787 vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
8788 {
8789         struct page *page = vmf->page;
8790         struct inode *inode = file_inode(vmf->vma->vm_file);
8791         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8792         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8793         struct btrfs_ordered_extent *ordered;
8794         struct extent_state *cached_state = NULL;
8795         struct extent_changeset *data_reserved = NULL;
8796         char *kaddr;
8797         unsigned long zero_start;
8798         loff_t size;
8799         vm_fault_t ret;
8800         int ret2;
8801         int reserved = 0;
8802         u64 reserved_space;
8803         u64 page_start;
8804         u64 page_end;
8805         u64 end;
8806
8807         reserved_space = PAGE_SIZE;
8808
8809         sb_start_pagefault(inode->i_sb);
8810         page_start = page_offset(page);
8811         page_end = page_start + PAGE_SIZE - 1;
8812         end = page_end;
8813
8814         /*
8815          * Reserving delalloc space after obtaining the page lock can lead to
8816          * deadlock. For example, if a dirty page is locked by this function
8817          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8818          * dirty page write out, then the btrfs_writepage() function could
8819          * end up waiting indefinitely to get a lock on the page currently
8820          * being processed by btrfs_page_mkwrite() function.
8821          */
8822         ret2 = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
8823                                            reserved_space);
8824         if (!ret2) {
8825                 ret2 = file_update_time(vmf->vma->vm_file);
8826                 reserved = 1;
8827         }
8828         if (ret2) {
8829                 ret = vmf_error(ret2);
8830                 if (reserved)
8831                         goto out;
8832                 goto out_noreserve;
8833         }
8834
8835         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8836 again:
8837         lock_page(page);
8838         size = i_size_read(inode);
8839
8840         if ((page->mapping != inode->i_mapping) ||
8841             (page_start >= size)) {
8842                 /* page got truncated out from underneath us */
8843                 goto out_unlock;
8844         }
8845         wait_on_page_writeback(page);
8846
8847         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8848         set_page_extent_mapped(page);
8849
8850         /*
8851          * we can't set the delalloc bits if there are pending ordered
8852          * extents.  Drop our locks and wait for them to finish
8853          */
8854         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
8855                         PAGE_SIZE);
8856         if (ordered) {
8857                 unlock_extent_cached(io_tree, page_start, page_end,
8858                                      &cached_state);
8859                 unlock_page(page);
8860                 btrfs_start_ordered_extent(inode, ordered, 1);
8861                 btrfs_put_ordered_extent(ordered);
8862                 goto again;
8863         }
8864
8865         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
8866                 reserved_space = round_up(size - page_start,
8867                                           fs_info->sectorsize);
8868                 if (reserved_space < PAGE_SIZE) {
8869                         end = page_start + reserved_space - 1;
8870                         btrfs_delalloc_release_space(inode, data_reserved,
8871                                         page_start, PAGE_SIZE - reserved_space,
8872                                         true);
8873                 }
8874         }
8875
8876         /*
8877          * page_mkwrite gets called when the page is firstly dirtied after it's
8878          * faulted in, but write(2) could also dirty a page and set delalloc
8879          * bits, thus in this case for space account reason, we still need to
8880          * clear any delalloc bits within this page range since we have to
8881          * reserve data&meta space before lock_page() (see above comments).
8882          */
8883         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
8884                           EXTENT_DIRTY | EXTENT_DELALLOC |
8885                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
8886                           0, 0, &cached_state);
8887
8888         ret2 = btrfs_set_extent_delalloc(inode, page_start, end, 0,
8889                                         &cached_state, 0);
8890         if (ret2) {
8891                 unlock_extent_cached(io_tree, page_start, page_end,
8892                                      &cached_state);
8893                 ret = VM_FAULT_SIGBUS;
8894                 goto out_unlock;
8895         }
8896         ret2 = 0;
8897
8898         /* page is wholly or partially inside EOF */
8899         if (page_start + PAGE_SIZE > size)
8900                 zero_start = size & ~PAGE_MASK;
8901         else
8902                 zero_start = PAGE_SIZE;
8903
8904         if (zero_start != PAGE_SIZE) {
8905                 kaddr = kmap(page);
8906                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
8907                 flush_dcache_page(page);
8908                 kunmap(page);
8909         }
8910         ClearPageChecked(page);
8911         set_page_dirty(page);
8912         SetPageUptodate(page);
8913
8914         BTRFS_I(inode)->last_trans = fs_info->generation;
8915         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
8916         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
8917
8918         unlock_extent_cached(io_tree, page_start, page_end, &cached_state);
8919
8920         if (!ret2) {
8921                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, true);
8922                 sb_end_pagefault(inode->i_sb);
8923                 extent_changeset_free(data_reserved);
8924                 return VM_FAULT_LOCKED;
8925         }
8926
8927 out_unlock:
8928         unlock_page(page);
8929 out:
8930         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, (ret != 0));
8931         btrfs_delalloc_release_space(inode, data_reserved, page_start,
8932                                      reserved_space, (ret != 0));
8933 out_noreserve:
8934         sb_end_pagefault(inode->i_sb);
8935         extent_changeset_free(data_reserved);
8936         return ret;
8937 }
8938
8939 static int btrfs_truncate(struct inode *inode, bool skip_writeback)
8940 {
8941         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8942         struct btrfs_root *root = BTRFS_I(inode)->root;
8943         struct btrfs_block_rsv *rsv;
8944         int ret;
8945         struct btrfs_trans_handle *trans;
8946         u64 mask = fs_info->sectorsize - 1;
8947         u64 min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
8948
8949         if (!skip_writeback) {
8950                 ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
8951                                                (u64)-1);
8952                 if (ret)
8953                         return ret;
8954         }
8955
8956         /*
8957          * Yes ladies and gentlemen, this is indeed ugly.  We have a couple of
8958          * things going on here:
8959          *
8960          * 1) We need to reserve space to update our inode.
8961          *
8962          * 2) We need to have something to cache all the space that is going to
8963          * be free'd up by the truncate operation, but also have some slack
8964          * space reserved in case it uses space during the truncate (thank you
8965          * very much snapshotting).
8966          *
8967          * And we need these to be separate.  The fact is we can use a lot of
8968          * space doing the truncate, and we have no earthly idea how much space
8969          * we will use, so we need the truncate reservation to be separate so it
8970          * doesn't end up using space reserved for updating the inode.  We also
8971          * need to be able to stop the transaction and start a new one, which
8972          * means we need to be able to update the inode several times, and we
8973          * have no idea of knowing how many times that will be, so we can't just
8974          * reserve 1 item for the entirety of the operation, so that has to be
8975          * done separately as well.
8976          *
8977          * So that leaves us with
8978          *
8979          * 1) rsv - for the truncate reservation, which we will steal from the
8980          * transaction reservation.
8981          * 2) fs_info->trans_block_rsv - this will have 1 items worth left for
8982          * updating the inode.
8983          */
8984         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
8985         if (!rsv)
8986                 return -ENOMEM;
8987         rsv->size = min_size;
8988         rsv->failfast = 1;
8989
8990         /*
8991          * 1 for the truncate slack space
8992          * 1 for updating the inode.
8993          */
8994         trans = btrfs_start_transaction(root, 2);
8995         if (IS_ERR(trans)) {
8996                 ret = PTR_ERR(trans);
8997                 goto out;
8998         }
8999
9000         /* Migrate the slack space for the truncate to our reserve */
9001         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9002                                       min_size, false);
9003         BUG_ON(ret);
9004
9005         /*
9006          * So if we truncate and then write and fsync we normally would just
9007          * write the extents that changed, which is a problem if we need to
9008          * first truncate that entire inode.  So set this flag so we write out
9009          * all of the extents in the inode to the sync log so we're completely
9010          * safe.
9011          */
9012         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9013         trans->block_rsv = rsv;
9014
9015         while (1) {
9016                 ret = btrfs_truncate_inode_items(trans, root, inode,
9017                                                  inode->i_size,
9018                                                  BTRFS_EXTENT_DATA_KEY);
9019                 trans->block_rsv = &fs_info->trans_block_rsv;
9020                 if (ret != -ENOSPC && ret != -EAGAIN)
9021                         break;
9022
9023                 ret = btrfs_update_inode(trans, root, inode);
9024                 if (ret)
9025                         break;
9026
9027                 btrfs_end_transaction(trans);
9028                 btrfs_btree_balance_dirty(fs_info);
9029
9030                 trans = btrfs_start_transaction(root, 2);
9031                 if (IS_ERR(trans)) {
9032                         ret = PTR_ERR(trans);
9033                         trans = NULL;
9034                         break;
9035                 }
9036
9037                 btrfs_block_rsv_release(fs_info, rsv, -1);
9038                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9039                                               rsv, min_size, false);
9040                 BUG_ON(ret);    /* shouldn't happen */
9041                 trans->block_rsv = rsv;
9042         }
9043
9044         /*
9045          * We can't call btrfs_truncate_block inside a trans handle as we could
9046          * deadlock with freeze, if we got NEED_TRUNCATE_BLOCK then we know
9047          * we've truncated everything except the last little bit, and can do
9048          * btrfs_truncate_block and then update the disk_i_size.
9049          */
9050         if (ret == NEED_TRUNCATE_BLOCK) {
9051                 btrfs_end_transaction(trans);
9052                 btrfs_btree_balance_dirty(fs_info);
9053
9054                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
9055                 if (ret)
9056                         goto out;
9057                 trans = btrfs_start_transaction(root, 1);
9058                 if (IS_ERR(trans)) {
9059                         ret = PTR_ERR(trans);
9060                         goto out;
9061                 }
9062                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
9063         }
9064
9065         if (trans) {
9066                 int ret2;
9067
9068                 trans->block_rsv = &fs_info->trans_block_rsv;
9069                 ret2 = btrfs_update_inode(trans, root, inode);
9070                 if (ret2 && !ret)
9071                         ret = ret2;
9072
9073                 ret2 = btrfs_end_transaction(trans);
9074                 if (ret2 && !ret)
9075                         ret = ret2;
9076                 btrfs_btree_balance_dirty(fs_info);
9077         }
9078 out:
9079         btrfs_free_block_rsv(fs_info, rsv);
9080
9081         return ret;
9082 }
9083
9084 /*
9085  * create a new subvolume directory/inode (helper for the ioctl).
9086  */
9087 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9088                              struct btrfs_root *new_root,
9089                              struct btrfs_root *parent_root,
9090                              u64 new_dirid)
9091 {
9092         struct inode *inode;
9093         int err;
9094         u64 index = 0;
9095
9096         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9097                                 new_dirid, new_dirid,
9098                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9099                                 &index);
9100         if (IS_ERR(inode))
9101                 return PTR_ERR(inode);
9102         inode->i_op = &btrfs_dir_inode_operations;
9103         inode->i_fop = &btrfs_dir_file_operations;
9104
9105         set_nlink(inode, 1);
9106         btrfs_i_size_write(BTRFS_I(inode), 0);
9107         unlock_new_inode(inode);
9108
9109         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9110         if (err)
9111                 btrfs_err(new_root->fs_info,
9112                           "error inheriting subvolume %llu properties: %d",
9113                           new_root->root_key.objectid, err);
9114
9115         err = btrfs_update_inode(trans, new_root, inode);
9116
9117         iput(inode);
9118         return err;
9119 }
9120
9121 struct inode *btrfs_alloc_inode(struct super_block *sb)
9122 {
9123         struct btrfs_fs_info *fs_info = btrfs_sb(sb);
9124         struct btrfs_inode *ei;
9125         struct inode *inode;
9126
9127         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_KERNEL);
9128         if (!ei)
9129                 return NULL;
9130
9131         ei->root = NULL;
9132         ei->generation = 0;
9133         ei->last_trans = 0;
9134         ei->last_sub_trans = 0;
9135         ei->logged_trans = 0;
9136         ei->delalloc_bytes = 0;
9137         ei->new_delalloc_bytes = 0;
9138         ei->defrag_bytes = 0;
9139         ei->disk_i_size = 0;
9140         ei->flags = 0;
9141         ei->csum_bytes = 0;
9142         ei->index_cnt = (u64)-1;
9143         ei->dir_index = 0;
9144         ei->last_unlink_trans = 0;
9145         ei->last_log_commit = 0;
9146
9147         spin_lock_init(&ei->lock);
9148         ei->outstanding_extents = 0;
9149         if (sb->s_magic != BTRFS_TEST_MAGIC)
9150                 btrfs_init_metadata_block_rsv(fs_info, &ei->block_rsv,
9151                                               BTRFS_BLOCK_RSV_DELALLOC);
9152         ei->runtime_flags = 0;
9153         ei->prop_compress = BTRFS_COMPRESS_NONE;
9154         ei->defrag_compress = BTRFS_COMPRESS_NONE;
9155
9156         ei->delayed_node = NULL;
9157
9158         ei->i_otime.tv_sec = 0;
9159         ei->i_otime.tv_nsec = 0;
9160
9161         inode = &ei->vfs_inode;
9162         extent_map_tree_init(&ei->extent_tree);
9163         extent_io_tree_init(&ei->io_tree, inode);
9164         extent_io_tree_init(&ei->io_failure_tree, inode);
9165         ei->io_tree.track_uptodate = 1;
9166         ei->io_failure_tree.track_uptodate = 1;
9167         atomic_set(&ei->sync_writers, 0);
9168         mutex_init(&ei->log_mutex);
9169         mutex_init(&ei->delalloc_mutex);
9170         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9171         INIT_LIST_HEAD(&ei->delalloc_inodes);
9172         INIT_LIST_HEAD(&ei->delayed_iput);
9173         RB_CLEAR_NODE(&ei->rb_node);
9174         init_rwsem(&ei->dio_sem);
9175
9176         return inode;
9177 }
9178
9179 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9180 void btrfs_test_destroy_inode(struct inode *inode)
9181 {
9182         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9183         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9184 }
9185 #endif
9186
9187 static void btrfs_i_callback(struct rcu_head *head)
9188 {
9189         struct inode *inode = container_of(head, struct inode, i_rcu);
9190         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9191 }
9192
9193 void btrfs_destroy_inode(struct inode *inode)
9194 {
9195         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9196         struct btrfs_ordered_extent *ordered;
9197         struct btrfs_root *root = BTRFS_I(inode)->root;
9198
9199         WARN_ON(!hlist_empty(&inode->i_dentry));
9200         WARN_ON(inode->i_data.nrpages);
9201         WARN_ON(BTRFS_I(inode)->block_rsv.reserved);
9202         WARN_ON(BTRFS_I(inode)->block_rsv.size);
9203         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9204         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9205         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
9206         WARN_ON(BTRFS_I(inode)->csum_bytes);
9207         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9208
9209         /*
9210          * This can happen where we create an inode, but somebody else also
9211          * created the same inode and we need to destroy the one we already
9212          * created.
9213          */
9214         if (!root)
9215                 goto free;
9216
9217         while (1) {
9218                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9219                 if (!ordered)
9220                         break;
9221                 else {
9222                         btrfs_err(fs_info,
9223                                   "found ordered extent %llu %llu on inode cleanup",
9224                                   ordered->file_offset, ordered->len);
9225                         btrfs_remove_ordered_extent(inode, ordered);
9226                         btrfs_put_ordered_extent(ordered);
9227                         btrfs_put_ordered_extent(ordered);
9228                 }
9229         }
9230         btrfs_qgroup_check_reserved_leak(inode);
9231         inode_tree_del(inode);
9232         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9233 free:
9234         call_rcu(&inode->i_rcu, btrfs_i_callback);
9235 }
9236
9237 int btrfs_drop_inode(struct inode *inode)
9238 {
9239         struct btrfs_root *root = BTRFS_I(inode)->root;
9240
9241         if (root == NULL)
9242                 return 1;
9243
9244         /* the snap/subvol tree is on deleting */
9245         if (btrfs_root_refs(&root->root_item) == 0)
9246                 return 1;
9247         else
9248                 return generic_drop_inode(inode);
9249 }
9250
9251 static void init_once(void *foo)
9252 {
9253         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9254
9255         inode_init_once(&ei->vfs_inode);
9256 }
9257
9258 void __cold btrfs_destroy_cachep(void)
9259 {
9260         /*
9261          * Make sure all delayed rcu free inodes are flushed before we
9262          * destroy cache.
9263          */
9264         rcu_barrier();
9265         kmem_cache_destroy(btrfs_inode_cachep);
9266         kmem_cache_destroy(btrfs_trans_handle_cachep);
9267         kmem_cache_destroy(btrfs_path_cachep);
9268         kmem_cache_destroy(btrfs_free_space_cachep);
9269 }
9270
9271 int __init btrfs_init_cachep(void)
9272 {
9273         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9274                         sizeof(struct btrfs_inode), 0,
9275                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9276                         init_once);
9277         if (!btrfs_inode_cachep)
9278                 goto fail;
9279
9280         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9281                         sizeof(struct btrfs_trans_handle), 0,
9282                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9283         if (!btrfs_trans_handle_cachep)
9284                 goto fail;
9285
9286         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9287                         sizeof(struct btrfs_path), 0,
9288                         SLAB_MEM_SPREAD, NULL);
9289         if (!btrfs_path_cachep)
9290                 goto fail;
9291
9292         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9293                         sizeof(struct btrfs_free_space), 0,
9294                         SLAB_MEM_SPREAD, NULL);
9295         if (!btrfs_free_space_cachep)
9296                 goto fail;
9297
9298         return 0;
9299 fail:
9300         btrfs_destroy_cachep();
9301         return -ENOMEM;
9302 }
9303
9304 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9305                          u32 request_mask, unsigned int flags)
9306 {
9307         u64 delalloc_bytes;
9308         struct inode *inode = d_inode(path->dentry);
9309         u32 blocksize = inode->i_sb->s_blocksize;
9310         u32 bi_flags = BTRFS_I(inode)->flags;
9311
9312         stat->result_mask |= STATX_BTIME;
9313         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
9314         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
9315         if (bi_flags & BTRFS_INODE_APPEND)
9316                 stat->attributes |= STATX_ATTR_APPEND;
9317         if (bi_flags & BTRFS_INODE_COMPRESS)
9318                 stat->attributes |= STATX_ATTR_COMPRESSED;
9319         if (bi_flags & BTRFS_INODE_IMMUTABLE)
9320                 stat->attributes |= STATX_ATTR_IMMUTABLE;
9321         if (bi_flags & BTRFS_INODE_NODUMP)
9322                 stat->attributes |= STATX_ATTR_NODUMP;
9323
9324         stat->attributes_mask |= (STATX_ATTR_APPEND |
9325                                   STATX_ATTR_COMPRESSED |
9326                                   STATX_ATTR_IMMUTABLE |
9327                                   STATX_ATTR_NODUMP);
9328
9329         generic_fillattr(inode, stat);
9330         stat->dev = BTRFS_I(inode)->root->anon_dev;
9331
9332         spin_lock(&BTRFS_I(inode)->lock);
9333         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9334         spin_unlock(&BTRFS_I(inode)->lock);
9335         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9336                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9337         return 0;
9338 }
9339
9340 static int btrfs_rename_exchange(struct inode *old_dir,
9341                               struct dentry *old_dentry,
9342                               struct inode *new_dir,
9343                               struct dentry *new_dentry)
9344 {
9345         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9346         struct btrfs_trans_handle *trans;
9347         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9348         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9349         struct inode *new_inode = new_dentry->d_inode;
9350         struct inode *old_inode = old_dentry->d_inode;
9351         struct timespec64 ctime = current_time(old_inode);
9352         struct dentry *parent;
9353         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9354         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9355         u64 old_idx = 0;
9356         u64 new_idx = 0;
9357         u64 root_objectid;
9358         int ret;
9359         bool root_log_pinned = false;
9360         bool dest_log_pinned = false;
9361         struct btrfs_log_ctx ctx_root;
9362         struct btrfs_log_ctx ctx_dest;
9363         bool sync_log_root = false;
9364         bool sync_log_dest = false;
9365         bool commit_transaction = false;
9366
9367         /* we only allow rename subvolume link between subvolumes */
9368         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9369                 return -EXDEV;
9370
9371         btrfs_init_log_ctx(&ctx_root, old_inode);
9372         btrfs_init_log_ctx(&ctx_dest, new_inode);
9373
9374         /* close the race window with snapshot create/destroy ioctl */
9375         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9376                 down_read(&fs_info->subvol_sem);
9377         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9378                 down_read(&fs_info->subvol_sem);
9379
9380         /*
9381          * We want to reserve the absolute worst case amount of items.  So if
9382          * both inodes are subvols and we need to unlink them then that would
9383          * require 4 item modifications, but if they are both normal inodes it
9384          * would require 5 item modifications, so we'll assume their normal
9385          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9386          * should cover the worst case number of items we'll modify.
9387          */
9388         trans = btrfs_start_transaction(root, 12);
9389         if (IS_ERR(trans)) {
9390                 ret = PTR_ERR(trans);
9391                 goto out_notrans;
9392         }
9393
9394         /*
9395          * We need to find a free sequence number both in the source and
9396          * in the destination directory for the exchange.
9397          */
9398         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9399         if (ret)
9400                 goto out_fail;
9401         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9402         if (ret)
9403                 goto out_fail;
9404
9405         BTRFS_I(old_inode)->dir_index = 0ULL;
9406         BTRFS_I(new_inode)->dir_index = 0ULL;
9407
9408         /* Reference for the source. */
9409         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9410                 /* force full log commit if subvolume involved. */
9411                 btrfs_set_log_full_commit(fs_info, trans);
9412         } else {
9413                 btrfs_pin_log_trans(root);
9414                 root_log_pinned = true;
9415                 ret = btrfs_insert_inode_ref(trans, dest,
9416                                              new_dentry->d_name.name,
9417                                              new_dentry->d_name.len,
9418                                              old_ino,
9419                                              btrfs_ino(BTRFS_I(new_dir)),
9420                                              old_idx);
9421                 if (ret)
9422                         goto out_fail;
9423         }
9424
9425         /* And now for the dest. */
9426         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9427                 /* force full log commit if subvolume involved. */
9428                 btrfs_set_log_full_commit(fs_info, trans);
9429         } else {
9430                 btrfs_pin_log_trans(dest);
9431                 dest_log_pinned = true;
9432                 ret = btrfs_insert_inode_ref(trans, root,
9433                                              old_dentry->d_name.name,
9434                                              old_dentry->d_name.len,
9435                                              new_ino,
9436                                              btrfs_ino(BTRFS_I(old_dir)),
9437                                              new_idx);
9438                 if (ret)
9439                         goto out_fail;
9440         }
9441
9442         /* Update inode version and ctime/mtime. */
9443         inode_inc_iversion(old_dir);
9444         inode_inc_iversion(new_dir);
9445         inode_inc_iversion(old_inode);
9446         inode_inc_iversion(new_inode);
9447         old_dir->i_ctime = old_dir->i_mtime = ctime;
9448         new_dir->i_ctime = new_dir->i_mtime = ctime;
9449         old_inode->i_ctime = ctime;
9450         new_inode->i_ctime = ctime;
9451
9452         if (old_dentry->d_parent != new_dentry->d_parent) {
9453                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9454                                 BTRFS_I(old_inode), 1);
9455                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9456                                 BTRFS_I(new_inode), 1);
9457         }
9458
9459         /* src is a subvolume */
9460         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9461                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9462                 ret = btrfs_unlink_subvol(trans, old_dir, root_objectid,
9463                                           old_dentry->d_name.name,
9464                                           old_dentry->d_name.len);
9465         } else { /* src is an inode */
9466                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9467                                            BTRFS_I(old_dentry->d_inode),
9468                                            old_dentry->d_name.name,
9469                                            old_dentry->d_name.len);
9470                 if (!ret)
9471                         ret = btrfs_update_inode(trans, root, old_inode);
9472         }
9473         if (ret) {
9474                 btrfs_abort_transaction(trans, ret);
9475                 goto out_fail;
9476         }
9477
9478         /* dest is a subvolume */
9479         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9480                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9481                 ret = btrfs_unlink_subvol(trans, new_dir, root_objectid,
9482                                           new_dentry->d_name.name,
9483                                           new_dentry->d_name.len);
9484         } else { /* dest is an inode */
9485                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9486                                            BTRFS_I(new_dentry->d_inode),
9487                                            new_dentry->d_name.name,
9488                                            new_dentry->d_name.len);
9489                 if (!ret)
9490                         ret = btrfs_update_inode(trans, dest, new_inode);
9491         }
9492         if (ret) {
9493                 btrfs_abort_transaction(trans, ret);
9494                 goto out_fail;
9495         }
9496
9497         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9498                              new_dentry->d_name.name,
9499                              new_dentry->d_name.len, 0, old_idx);
9500         if (ret) {
9501                 btrfs_abort_transaction(trans, ret);
9502                 goto out_fail;
9503         }
9504
9505         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9506                              old_dentry->d_name.name,
9507                              old_dentry->d_name.len, 0, new_idx);
9508         if (ret) {
9509                 btrfs_abort_transaction(trans, ret);
9510                 goto out_fail;
9511         }
9512
9513         if (old_inode->i_nlink == 1)
9514                 BTRFS_I(old_inode)->dir_index = old_idx;
9515         if (new_inode->i_nlink == 1)
9516                 BTRFS_I(new_inode)->dir_index = new_idx;
9517
9518         if (root_log_pinned) {
9519                 parent = new_dentry->d_parent;
9520                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
9521                                          BTRFS_I(old_dir), parent,
9522                                          false, &ctx_root);
9523                 if (ret == BTRFS_NEED_LOG_SYNC)
9524                         sync_log_root = true;
9525                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
9526                         commit_transaction = true;
9527                 ret = 0;
9528                 btrfs_end_log_trans(root);
9529                 root_log_pinned = false;
9530         }
9531         if (dest_log_pinned) {
9532                 if (!commit_transaction) {
9533                         parent = old_dentry->d_parent;
9534                         ret = btrfs_log_new_name(trans, BTRFS_I(new_inode),
9535                                                  BTRFS_I(new_dir), parent,
9536                                                  false, &ctx_dest);
9537                         if (ret == BTRFS_NEED_LOG_SYNC)
9538                                 sync_log_dest = true;
9539                         else if (ret == BTRFS_NEED_TRANS_COMMIT)
9540                                 commit_transaction = true;
9541                         ret = 0;
9542                 }
9543                 btrfs_end_log_trans(dest);
9544                 dest_log_pinned = false;
9545         }
9546 out_fail:
9547         /*
9548          * If we have pinned a log and an error happened, we unpin tasks
9549          * trying to sync the log and force them to fallback to a transaction
9550          * commit if the log currently contains any of the inodes involved in
9551          * this rename operation (to ensure we do not persist a log with an
9552          * inconsistent state for any of these inodes or leading to any
9553          * inconsistencies when replayed). If the transaction was aborted, the
9554          * abortion reason is propagated to userspace when attempting to commit
9555          * the transaction. If the log does not contain any of these inodes, we
9556          * allow the tasks to sync it.
9557          */
9558         if (ret && (root_log_pinned || dest_log_pinned)) {
9559                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9560                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9561                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9562                     (new_inode &&
9563                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9564                         btrfs_set_log_full_commit(fs_info, trans);
9565
9566                 if (root_log_pinned) {
9567                         btrfs_end_log_trans(root);
9568                         root_log_pinned = false;
9569                 }
9570                 if (dest_log_pinned) {
9571                         btrfs_end_log_trans(dest);
9572                         dest_log_pinned = false;
9573                 }
9574         }
9575         if (!ret && sync_log_root && !commit_transaction) {
9576                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root,
9577                                      &ctx_root);
9578                 if (ret)
9579                         commit_transaction = true;
9580         }
9581         if (!ret && sync_log_dest && !commit_transaction) {
9582                 ret = btrfs_sync_log(trans, BTRFS_I(new_inode)->root,
9583                                      &ctx_dest);
9584                 if (ret)
9585                         commit_transaction = true;
9586         }
9587         if (commit_transaction) {
9588                 ret = btrfs_commit_transaction(trans);
9589         } else {
9590                 int ret2;
9591
9592                 ret2 = btrfs_end_transaction(trans);
9593                 ret = ret ? ret : ret2;
9594         }
9595 out_notrans:
9596         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9597                 up_read(&fs_info->subvol_sem);
9598         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9599                 up_read(&fs_info->subvol_sem);
9600
9601         return ret;
9602 }
9603
9604 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9605                                      struct btrfs_root *root,
9606                                      struct inode *dir,
9607                                      struct dentry *dentry)
9608 {
9609         int ret;
9610         struct inode *inode;
9611         u64 objectid;
9612         u64 index;
9613
9614         ret = btrfs_find_free_ino(root, &objectid);
9615         if (ret)
9616                 return ret;
9617
9618         inode = btrfs_new_inode(trans, root, dir,
9619                                 dentry->d_name.name,
9620                                 dentry->d_name.len,
9621                                 btrfs_ino(BTRFS_I(dir)),
9622                                 objectid,
9623                                 S_IFCHR | WHITEOUT_MODE,
9624                                 &index);
9625
9626         if (IS_ERR(inode)) {
9627                 ret = PTR_ERR(inode);
9628                 return ret;
9629         }
9630
9631         inode->i_op = &btrfs_special_inode_operations;
9632         init_special_inode(inode, inode->i_mode,
9633                 WHITEOUT_DEV);
9634
9635         ret = btrfs_init_inode_security(trans, inode, dir,
9636                                 &dentry->d_name);
9637         if (ret)
9638                 goto out;
9639
9640         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9641                                 BTRFS_I(inode), 0, index);
9642         if (ret)
9643                 goto out;
9644
9645         ret = btrfs_update_inode(trans, root, inode);
9646 out:
9647         unlock_new_inode(inode);
9648         if (ret)
9649                 inode_dec_link_count(inode);
9650         iput(inode);
9651
9652         return ret;
9653 }
9654
9655 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9656                            struct inode *new_dir, struct dentry *new_dentry,
9657                            unsigned int flags)
9658 {
9659         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9660         struct btrfs_trans_handle *trans;
9661         unsigned int trans_num_items;
9662         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9663         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9664         struct inode *new_inode = d_inode(new_dentry);
9665         struct inode *old_inode = d_inode(old_dentry);
9666         u64 index = 0;
9667         u64 root_objectid;
9668         int ret;
9669         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9670         bool log_pinned = false;
9671         struct btrfs_log_ctx ctx;
9672         bool sync_log = false;
9673         bool commit_transaction = false;
9674
9675         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9676                 return -EPERM;
9677
9678         /* we only allow rename subvolume link between subvolumes */
9679         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9680                 return -EXDEV;
9681
9682         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9683             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9684                 return -ENOTEMPTY;
9685
9686         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9687             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9688                 return -ENOTEMPTY;
9689
9690
9691         /* check for collisions, even if the  name isn't there */
9692         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9693                              new_dentry->d_name.name,
9694                              new_dentry->d_name.len);
9695
9696         if (ret) {
9697                 if (ret == -EEXIST) {
9698                         /* we shouldn't get
9699                          * eexist without a new_inode */
9700                         if (WARN_ON(!new_inode)) {
9701                                 return ret;
9702                         }
9703                 } else {
9704                         /* maybe -EOVERFLOW */
9705                         return ret;
9706                 }
9707         }
9708         ret = 0;
9709
9710         /*
9711          * we're using rename to replace one file with another.  Start IO on it
9712          * now so  we don't add too much work to the end of the transaction
9713          */
9714         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9715                 filemap_flush(old_inode->i_mapping);
9716
9717         /* close the racy window with snapshot create/destroy ioctl */
9718         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9719                 down_read(&fs_info->subvol_sem);
9720         /*
9721          * We want to reserve the absolute worst case amount of items.  So if
9722          * both inodes are subvols and we need to unlink them then that would
9723          * require 4 item modifications, but if they are both normal inodes it
9724          * would require 5 item modifications, so we'll assume they are normal
9725          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9726          * should cover the worst case number of items we'll modify.
9727          * If our rename has the whiteout flag, we need more 5 units for the
9728          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9729          * when selinux is enabled).
9730          */
9731         trans_num_items = 11;
9732         if (flags & RENAME_WHITEOUT)
9733                 trans_num_items += 5;
9734         trans = btrfs_start_transaction(root, trans_num_items);
9735         if (IS_ERR(trans)) {
9736                 ret = PTR_ERR(trans);
9737                 goto out_notrans;
9738         }
9739
9740         if (dest != root)
9741                 btrfs_record_root_in_trans(trans, dest);
9742
9743         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
9744         if (ret)
9745                 goto out_fail;
9746
9747         BTRFS_I(old_inode)->dir_index = 0ULL;
9748         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9749                 /* force full log commit if subvolume involved. */
9750                 btrfs_set_log_full_commit(fs_info, trans);
9751         } else {
9752                 btrfs_pin_log_trans(root);
9753                 log_pinned = true;
9754                 ret = btrfs_insert_inode_ref(trans, dest,
9755                                              new_dentry->d_name.name,
9756                                              new_dentry->d_name.len,
9757                                              old_ino,
9758                                              btrfs_ino(BTRFS_I(new_dir)), index);
9759                 if (ret)
9760                         goto out_fail;
9761         }
9762
9763         inode_inc_iversion(old_dir);
9764         inode_inc_iversion(new_dir);
9765         inode_inc_iversion(old_inode);
9766         old_dir->i_ctime = old_dir->i_mtime =
9767         new_dir->i_ctime = new_dir->i_mtime =
9768         old_inode->i_ctime = current_time(old_dir);
9769
9770         if (old_dentry->d_parent != new_dentry->d_parent)
9771                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9772                                 BTRFS_I(old_inode), 1);
9773
9774         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9775                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9776                 ret = btrfs_unlink_subvol(trans, old_dir, root_objectid,
9777                                         old_dentry->d_name.name,
9778                                         old_dentry->d_name.len);
9779         } else {
9780                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9781                                         BTRFS_I(d_inode(old_dentry)),
9782                                         old_dentry->d_name.name,
9783                                         old_dentry->d_name.len);
9784                 if (!ret)
9785                         ret = btrfs_update_inode(trans, root, old_inode);
9786         }
9787         if (ret) {
9788                 btrfs_abort_transaction(trans, ret);
9789                 goto out_fail;
9790         }
9791
9792         if (new_inode) {
9793                 inode_inc_iversion(new_inode);
9794                 new_inode->i_ctime = current_time(new_inode);
9795                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
9796                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9797                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9798                         ret = btrfs_unlink_subvol(trans, new_dir, root_objectid,
9799                                                 new_dentry->d_name.name,
9800                                                 new_dentry->d_name.len);
9801                         BUG_ON(new_inode->i_nlink == 0);
9802                 } else {
9803                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9804                                                  BTRFS_I(d_inode(new_dentry)),
9805                                                  new_dentry->d_name.name,
9806                                                  new_dentry->d_name.len);
9807                 }
9808                 if (!ret && new_inode->i_nlink == 0)
9809                         ret = btrfs_orphan_add(trans,
9810                                         BTRFS_I(d_inode(new_dentry)));
9811                 if (ret) {
9812                         btrfs_abort_transaction(trans, ret);
9813                         goto out_fail;
9814                 }
9815         }
9816
9817         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9818                              new_dentry->d_name.name,
9819                              new_dentry->d_name.len, 0, index);
9820         if (ret) {
9821                 btrfs_abort_transaction(trans, ret);
9822                 goto out_fail;
9823         }
9824
9825         if (old_inode->i_nlink == 1)
9826                 BTRFS_I(old_inode)->dir_index = index;
9827
9828         if (log_pinned) {
9829                 struct dentry *parent = new_dentry->d_parent;
9830
9831                 btrfs_init_log_ctx(&ctx, old_inode);
9832                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
9833                                          BTRFS_I(old_dir), parent,
9834                                          false, &ctx);
9835                 if (ret == BTRFS_NEED_LOG_SYNC)
9836                         sync_log = true;
9837                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
9838                         commit_transaction = true;
9839                 ret = 0;
9840                 btrfs_end_log_trans(root);
9841                 log_pinned = false;
9842         }
9843
9844         if (flags & RENAME_WHITEOUT) {
9845                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9846                                                 old_dentry);
9847
9848                 if (ret) {
9849                         btrfs_abort_transaction(trans, ret);
9850                         goto out_fail;
9851                 }
9852         }
9853 out_fail:
9854         /*
9855          * If we have pinned the log and an error happened, we unpin tasks
9856          * trying to sync the log and force them to fallback to a transaction
9857          * commit if the log currently contains any of the inodes involved in
9858          * this rename operation (to ensure we do not persist a log with an
9859          * inconsistent state for any of these inodes or leading to any
9860          * inconsistencies when replayed). If the transaction was aborted, the
9861          * abortion reason is propagated to userspace when attempting to commit
9862          * the transaction. If the log does not contain any of these inodes, we
9863          * allow the tasks to sync it.
9864          */
9865         if (ret && log_pinned) {
9866                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9867                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9868                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9869                     (new_inode &&
9870                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9871                         btrfs_set_log_full_commit(fs_info, trans);
9872
9873                 btrfs_end_log_trans(root);
9874                 log_pinned = false;
9875         }
9876         if (!ret && sync_log) {
9877                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root, &ctx);
9878                 if (ret)
9879                         commit_transaction = true;
9880         }
9881         if (commit_transaction) {
9882                 ret = btrfs_commit_transaction(trans);
9883         } else {
9884                 int ret2;
9885
9886                 ret2 = btrfs_end_transaction(trans);
9887                 ret = ret ? ret : ret2;
9888         }
9889 out_notrans:
9890         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9891                 up_read(&fs_info->subvol_sem);
9892
9893         return ret;
9894 }
9895
9896 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9897                          struct inode *new_dir, struct dentry *new_dentry,
9898                          unsigned int flags)
9899 {
9900         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
9901                 return -EINVAL;
9902
9903         if (flags & RENAME_EXCHANGE)
9904                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
9905                                           new_dentry);
9906
9907         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
9908 }
9909
9910 struct btrfs_delalloc_work {
9911         struct inode *inode;
9912         struct completion completion;
9913         struct list_head list;
9914         struct btrfs_work work;
9915 };
9916
9917 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9918 {
9919         struct btrfs_delalloc_work *delalloc_work;
9920         struct inode *inode;
9921
9922         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9923                                      work);
9924         inode = delalloc_work->inode;
9925         filemap_flush(inode->i_mapping);
9926         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9927                                 &BTRFS_I(inode)->runtime_flags))
9928                 filemap_flush(inode->i_mapping);
9929
9930         iput(inode);
9931         complete(&delalloc_work->completion);
9932 }
9933
9934 static struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode)
9935 {
9936         struct btrfs_delalloc_work *work;
9937
9938         work = kmalloc(sizeof(*work), GFP_NOFS);
9939         if (!work)
9940                 return NULL;
9941
9942         init_completion(&work->completion);
9943         INIT_LIST_HEAD(&work->list);
9944         work->inode = inode;
9945         WARN_ON_ONCE(!inode);
9946         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
9947                         btrfs_run_delalloc_work, NULL, NULL);
9948
9949         return work;
9950 }
9951
9952 /*
9953  * some fairly slow code that needs optimization. This walks the list
9954  * of all the inodes with pending delalloc and forces them to disk.
9955  */
9956 static int start_delalloc_inodes(struct btrfs_root *root, int nr)
9957 {
9958         struct btrfs_inode *binode;
9959         struct inode *inode;
9960         struct btrfs_delalloc_work *work, *next;
9961         struct list_head works;
9962         struct list_head splice;
9963         int ret = 0;
9964
9965         INIT_LIST_HEAD(&works);
9966         INIT_LIST_HEAD(&splice);
9967
9968         mutex_lock(&root->delalloc_mutex);
9969         spin_lock(&root->delalloc_lock);
9970         list_splice_init(&root->delalloc_inodes, &splice);
9971         while (!list_empty(&splice)) {
9972                 binode = list_entry(splice.next, struct btrfs_inode,
9973                                     delalloc_inodes);
9974
9975                 list_move_tail(&binode->delalloc_inodes,
9976                                &root->delalloc_inodes);
9977                 inode = igrab(&binode->vfs_inode);
9978                 if (!inode) {
9979                         cond_resched_lock(&root->delalloc_lock);
9980                         continue;
9981                 }
9982                 spin_unlock(&root->delalloc_lock);
9983
9984                 work = btrfs_alloc_delalloc_work(inode);
9985                 if (!work) {
9986                         iput(inode);
9987                         ret = -ENOMEM;
9988                         goto out;
9989                 }
9990                 list_add_tail(&work->list, &works);
9991                 btrfs_queue_work(root->fs_info->flush_workers,
9992                                  &work->work);
9993                 ret++;
9994                 if (nr != -1 && ret >= nr)
9995                         goto out;
9996                 cond_resched();
9997                 spin_lock(&root->delalloc_lock);
9998         }
9999         spin_unlock(&root->delalloc_lock);
10000
10001 out:
10002         list_for_each_entry_safe(work, next, &works, list) {
10003                 list_del_init(&work->list);
10004                 wait_for_completion(&work->completion);
10005                 kfree(work);
10006         }
10007
10008         if (!list_empty(&splice)) {
10009                 spin_lock(&root->delalloc_lock);
10010                 list_splice_tail(&splice, &root->delalloc_inodes);
10011                 spin_unlock(&root->delalloc_lock);
10012         }
10013         mutex_unlock(&root->delalloc_mutex);
10014         return ret;
10015 }
10016
10017 int btrfs_start_delalloc_inodes(struct btrfs_root *root)
10018 {
10019         struct btrfs_fs_info *fs_info = root->fs_info;
10020         int ret;
10021
10022         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10023                 return -EROFS;
10024
10025         ret = start_delalloc_inodes(root, -1);
10026         if (ret > 0)
10027                 ret = 0;
10028         return ret;
10029 }
10030
10031 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int nr)
10032 {
10033         struct btrfs_root *root;
10034         struct list_head splice;
10035         int ret;
10036
10037         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10038                 return -EROFS;
10039
10040         INIT_LIST_HEAD(&splice);
10041
10042         mutex_lock(&fs_info->delalloc_root_mutex);
10043         spin_lock(&fs_info->delalloc_root_lock);
10044         list_splice_init(&fs_info->delalloc_roots, &splice);
10045         while (!list_empty(&splice) && nr) {
10046                 root = list_first_entry(&splice, struct btrfs_root,
10047                                         delalloc_root);
10048                 root = btrfs_grab_fs_root(root);
10049                 BUG_ON(!root);
10050                 list_move_tail(&root->delalloc_root,
10051                                &fs_info->delalloc_roots);
10052                 spin_unlock(&fs_info->delalloc_root_lock);
10053
10054                 ret = start_delalloc_inodes(root, nr);
10055                 btrfs_put_fs_root(root);
10056                 if (ret < 0)
10057                         goto out;
10058
10059                 if (nr != -1) {
10060                         nr -= ret;
10061                         WARN_ON(nr < 0);
10062                 }
10063                 spin_lock(&fs_info->delalloc_root_lock);
10064         }
10065         spin_unlock(&fs_info->delalloc_root_lock);
10066
10067         ret = 0;
10068 out:
10069         if (!list_empty(&splice)) {
10070                 spin_lock(&fs_info->delalloc_root_lock);
10071                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10072                 spin_unlock(&fs_info->delalloc_root_lock);
10073         }
10074         mutex_unlock(&fs_info->delalloc_root_mutex);
10075         return ret;
10076 }
10077
10078 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10079                          const char *symname)
10080 {
10081         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10082         struct btrfs_trans_handle *trans;
10083         struct btrfs_root *root = BTRFS_I(dir)->root;
10084         struct btrfs_path *path;
10085         struct btrfs_key key;
10086         struct inode *inode = NULL;
10087         int err;
10088         u64 objectid;
10089         u64 index = 0;
10090         int name_len;
10091         int datasize;
10092         unsigned long ptr;
10093         struct btrfs_file_extent_item *ei;
10094         struct extent_buffer *leaf;
10095
10096         name_len = strlen(symname);
10097         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10098                 return -ENAMETOOLONG;
10099
10100         /*
10101          * 2 items for inode item and ref
10102          * 2 items for dir items
10103          * 1 item for updating parent inode item
10104          * 1 item for the inline extent item
10105          * 1 item for xattr if selinux is on
10106          */
10107         trans = btrfs_start_transaction(root, 7);
10108         if (IS_ERR(trans))
10109                 return PTR_ERR(trans);
10110
10111         err = btrfs_find_free_ino(root, &objectid);
10112         if (err)
10113                 goto out_unlock;
10114
10115         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10116                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10117                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10118         if (IS_ERR(inode)) {
10119                 err = PTR_ERR(inode);
10120                 inode = NULL;
10121                 goto out_unlock;
10122         }
10123
10124         /*
10125         * If the active LSM wants to access the inode during
10126         * d_instantiate it needs these. Smack checks to see
10127         * if the filesystem supports xattrs by looking at the
10128         * ops vector.
10129         */
10130         inode->i_fop = &btrfs_file_operations;
10131         inode->i_op = &btrfs_file_inode_operations;
10132         inode->i_mapping->a_ops = &btrfs_aops;
10133         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10134
10135         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10136         if (err)
10137                 goto out_unlock;
10138
10139         path = btrfs_alloc_path();
10140         if (!path) {
10141                 err = -ENOMEM;
10142                 goto out_unlock;
10143         }
10144         key.objectid = btrfs_ino(BTRFS_I(inode));
10145         key.offset = 0;
10146         key.type = BTRFS_EXTENT_DATA_KEY;
10147         datasize = btrfs_file_extent_calc_inline_size(name_len);
10148         err = btrfs_insert_empty_item(trans, root, path, &key,
10149                                       datasize);
10150         if (err) {
10151                 btrfs_free_path(path);
10152                 goto out_unlock;
10153         }
10154         leaf = path->nodes[0];
10155         ei = btrfs_item_ptr(leaf, path->slots[0],
10156                             struct btrfs_file_extent_item);
10157         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10158         btrfs_set_file_extent_type(leaf, ei,
10159                                    BTRFS_FILE_EXTENT_INLINE);
10160         btrfs_set_file_extent_encryption(leaf, ei, 0);
10161         btrfs_set_file_extent_compression(leaf, ei, 0);
10162         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10163         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10164
10165         ptr = btrfs_file_extent_inline_start(ei);
10166         write_extent_buffer(leaf, symname, ptr, name_len);
10167         btrfs_mark_buffer_dirty(leaf);
10168         btrfs_free_path(path);
10169
10170         inode->i_op = &btrfs_symlink_inode_operations;
10171         inode_nohighmem(inode);
10172         inode->i_mapping->a_ops = &btrfs_aops;
10173         inode_set_bytes(inode, name_len);
10174         btrfs_i_size_write(BTRFS_I(inode), name_len);
10175         err = btrfs_update_inode(trans, root, inode);
10176         /*
10177          * Last step, add directory indexes for our symlink inode. This is the
10178          * last step to avoid extra cleanup of these indexes if an error happens
10179          * elsewhere above.
10180          */
10181         if (!err)
10182                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10183                                 BTRFS_I(inode), 0, index);
10184         if (err)
10185                 goto out_unlock;
10186
10187         d_instantiate_new(dentry, inode);
10188
10189 out_unlock:
10190         btrfs_end_transaction(trans);
10191         if (err && inode) {
10192                 inode_dec_link_count(inode);
10193                 discard_new_inode(inode);
10194         }
10195         btrfs_btree_balance_dirty(fs_info);
10196         return err;
10197 }
10198
10199 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10200                                        u64 start, u64 num_bytes, u64 min_size,
10201                                        loff_t actual_len, u64 *alloc_hint,
10202                                        struct btrfs_trans_handle *trans)
10203 {
10204         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10205         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10206         struct extent_map *em;
10207         struct btrfs_root *root = BTRFS_I(inode)->root;
10208         struct btrfs_key ins;
10209         u64 cur_offset = start;
10210         u64 i_size;
10211         u64 cur_bytes;
10212         u64 last_alloc = (u64)-1;
10213         int ret = 0;
10214         bool own_trans = true;
10215         u64 end = start + num_bytes - 1;
10216
10217         if (trans)
10218                 own_trans = false;
10219         while (num_bytes > 0) {
10220                 if (own_trans) {
10221                         trans = btrfs_start_transaction(root, 3);
10222                         if (IS_ERR(trans)) {
10223                                 ret = PTR_ERR(trans);
10224                                 break;
10225                         }
10226                 }
10227
10228                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10229                 cur_bytes = max(cur_bytes, min_size);
10230                 /*
10231                  * If we are severely fragmented we could end up with really
10232                  * small allocations, so if the allocator is returning small
10233                  * chunks lets make its job easier by only searching for those
10234                  * sized chunks.
10235                  */
10236                 cur_bytes = min(cur_bytes, last_alloc);
10237                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10238                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10239                 if (ret) {
10240                         if (own_trans)
10241                                 btrfs_end_transaction(trans);
10242                         break;
10243                 }
10244                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10245
10246                 last_alloc = ins.offset;
10247                 ret = insert_reserved_file_extent(trans, inode,
10248                                                   cur_offset, ins.objectid,
10249                                                   ins.offset, ins.offset,
10250                                                   ins.offset, 0, 0, 0,
10251                                                   BTRFS_FILE_EXTENT_PREALLOC);
10252                 if (ret) {
10253                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10254                                                    ins.offset, 0);
10255                         btrfs_abort_transaction(trans, ret);
10256                         if (own_trans)
10257                                 btrfs_end_transaction(trans);
10258                         break;
10259                 }
10260
10261                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10262                                         cur_offset + ins.offset -1, 0);
10263
10264                 em = alloc_extent_map();
10265                 if (!em) {
10266                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10267                                 &BTRFS_I(inode)->runtime_flags);
10268                         goto next;
10269                 }
10270
10271                 em->start = cur_offset;
10272                 em->orig_start = cur_offset;
10273                 em->len = ins.offset;
10274                 em->block_start = ins.objectid;
10275                 em->block_len = ins.offset;
10276                 em->orig_block_len = ins.offset;
10277                 em->ram_bytes = ins.offset;
10278                 em->bdev = fs_info->fs_devices->latest_bdev;
10279                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10280                 em->generation = trans->transid;
10281
10282                 while (1) {
10283                         write_lock(&em_tree->lock);
10284                         ret = add_extent_mapping(em_tree, em, 1);
10285                         write_unlock(&em_tree->lock);
10286                         if (ret != -EEXIST)
10287                                 break;
10288                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10289                                                 cur_offset + ins.offset - 1,
10290                                                 0);
10291                 }
10292                 free_extent_map(em);
10293 next:
10294                 num_bytes -= ins.offset;
10295                 cur_offset += ins.offset;
10296                 *alloc_hint = ins.objectid + ins.offset;
10297
10298                 inode_inc_iversion(inode);
10299                 inode->i_ctime = current_time(inode);
10300                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10301                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10302                     (actual_len > inode->i_size) &&
10303                     (cur_offset > inode->i_size)) {
10304                         if (cur_offset > actual_len)
10305                                 i_size = actual_len;
10306                         else
10307                                 i_size = cur_offset;
10308                         i_size_write(inode, i_size);
10309                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10310                 }
10311
10312                 ret = btrfs_update_inode(trans, root, inode);
10313
10314                 if (ret) {
10315                         btrfs_abort_transaction(trans, ret);
10316                         if (own_trans)
10317                                 btrfs_end_transaction(trans);
10318                         break;
10319                 }
10320
10321                 if (own_trans)
10322                         btrfs_end_transaction(trans);
10323         }
10324         if (cur_offset < end)
10325                 btrfs_free_reserved_data_space(inode, NULL, cur_offset,
10326                         end - cur_offset + 1);
10327         return ret;
10328 }
10329
10330 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10331                               u64 start, u64 num_bytes, u64 min_size,
10332                               loff_t actual_len, u64 *alloc_hint)
10333 {
10334         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10335                                            min_size, actual_len, alloc_hint,
10336                                            NULL);
10337 }
10338
10339 int btrfs_prealloc_file_range_trans(struct inode *inode,
10340                                     struct btrfs_trans_handle *trans, int mode,
10341                                     u64 start, u64 num_bytes, u64 min_size,
10342                                     loff_t actual_len, u64 *alloc_hint)
10343 {
10344         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10345                                            min_size, actual_len, alloc_hint, trans);
10346 }
10347
10348 static int btrfs_set_page_dirty(struct page *page)
10349 {
10350         return __set_page_dirty_nobuffers(page);
10351 }
10352
10353 static int btrfs_permission(struct inode *inode, int mask)
10354 {
10355         struct btrfs_root *root = BTRFS_I(inode)->root;
10356         umode_t mode = inode->i_mode;
10357
10358         if (mask & MAY_WRITE &&
10359             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10360                 if (btrfs_root_readonly(root))
10361                         return -EROFS;
10362                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10363                         return -EACCES;
10364         }
10365         return generic_permission(inode, mask);
10366 }
10367
10368 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10369 {
10370         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10371         struct btrfs_trans_handle *trans;
10372         struct btrfs_root *root = BTRFS_I(dir)->root;
10373         struct inode *inode = NULL;
10374         u64 objectid;
10375         u64 index;
10376         int ret = 0;
10377
10378         /*
10379          * 5 units required for adding orphan entry
10380          */
10381         trans = btrfs_start_transaction(root, 5);
10382         if (IS_ERR(trans))
10383                 return PTR_ERR(trans);
10384
10385         ret = btrfs_find_free_ino(root, &objectid);
10386         if (ret)
10387                 goto out;
10388
10389         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10390                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10391         if (IS_ERR(inode)) {
10392                 ret = PTR_ERR(inode);
10393                 inode = NULL;
10394                 goto out;
10395         }
10396
10397         inode->i_fop = &btrfs_file_operations;
10398         inode->i_op = &btrfs_file_inode_operations;
10399
10400         inode->i_mapping->a_ops = &btrfs_aops;
10401         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10402
10403         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10404         if (ret)
10405                 goto out;
10406
10407         ret = btrfs_update_inode(trans, root, inode);
10408         if (ret)
10409                 goto out;
10410         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10411         if (ret)
10412                 goto out;
10413
10414         /*
10415          * We set number of links to 0 in btrfs_new_inode(), and here we set
10416          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10417          * through:
10418          *
10419          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10420          */
10421         set_nlink(inode, 1);
10422         d_tmpfile(dentry, inode);
10423         unlock_new_inode(inode);
10424         mark_inode_dirty(inode);
10425 out:
10426         btrfs_end_transaction(trans);
10427         if (ret && inode)
10428                 discard_new_inode(inode);
10429         btrfs_btree_balance_dirty(fs_info);
10430         return ret;
10431 }
10432
10433 __attribute__((const))
10434 static int btrfs_readpage_io_failed_hook(struct page *page, int failed_mirror)
10435 {
10436         return -EAGAIN;
10437 }
10438
10439 static void btrfs_check_extent_io_range(void *private_data, const char *caller,
10440                                         u64 start, u64 end)
10441 {
10442         struct inode *inode = private_data;
10443         u64 isize;
10444
10445         isize = i_size_read(inode);
10446         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
10447                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
10448                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
10449                         caller, btrfs_ino(BTRFS_I(inode)), isize, start, end);
10450         }
10451 }
10452
10453 void btrfs_set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
10454 {
10455         struct inode *inode = tree->private_data;
10456         unsigned long index = start >> PAGE_SHIFT;
10457         unsigned long end_index = end >> PAGE_SHIFT;
10458         struct page *page;
10459
10460         while (index <= end_index) {
10461                 page = find_get_page(inode->i_mapping, index);
10462                 ASSERT(page); /* Pages should be in the extent_io_tree */
10463                 set_page_writeback(page);
10464                 put_page(page);
10465                 index++;
10466         }
10467 }
10468
10469 static const struct inode_operations btrfs_dir_inode_operations = {
10470         .getattr        = btrfs_getattr,
10471         .lookup         = btrfs_lookup,
10472         .create         = btrfs_create,
10473         .unlink         = btrfs_unlink,
10474         .link           = btrfs_link,
10475         .mkdir          = btrfs_mkdir,
10476         .rmdir          = btrfs_rmdir,
10477         .rename         = btrfs_rename2,
10478         .symlink        = btrfs_symlink,
10479         .setattr        = btrfs_setattr,
10480         .mknod          = btrfs_mknod,
10481         .listxattr      = btrfs_listxattr,
10482         .permission     = btrfs_permission,
10483         .get_acl        = btrfs_get_acl,
10484         .set_acl        = btrfs_set_acl,
10485         .update_time    = btrfs_update_time,
10486         .tmpfile        = btrfs_tmpfile,
10487 };
10488 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10489         .lookup         = btrfs_lookup,
10490         .permission     = btrfs_permission,
10491         .update_time    = btrfs_update_time,
10492 };
10493
10494 static const struct file_operations btrfs_dir_file_operations = {
10495         .llseek         = generic_file_llseek,
10496         .read           = generic_read_dir,
10497         .iterate_shared = btrfs_real_readdir,
10498         .open           = btrfs_opendir,
10499         .unlocked_ioctl = btrfs_ioctl,
10500 #ifdef CONFIG_COMPAT
10501         .compat_ioctl   = btrfs_compat_ioctl,
10502 #endif
10503         .release        = btrfs_release_file,
10504         .fsync          = btrfs_sync_file,
10505 };
10506
10507 static const struct extent_io_ops btrfs_extent_io_ops = {
10508         /* mandatory callbacks */
10509         .submit_bio_hook = btrfs_submit_bio_hook,
10510         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10511         .readpage_io_failed_hook = btrfs_readpage_io_failed_hook,
10512
10513         /* optional callbacks */
10514         .fill_delalloc = run_delalloc_range,
10515         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10516         .writepage_start_hook = btrfs_writepage_start_hook,
10517         .set_bit_hook = btrfs_set_bit_hook,
10518         .clear_bit_hook = btrfs_clear_bit_hook,
10519         .merge_extent_hook = btrfs_merge_extent_hook,
10520         .split_extent_hook = btrfs_split_extent_hook,
10521         .check_extent_io_range = btrfs_check_extent_io_range,
10522 };
10523
10524 /*
10525  * btrfs doesn't support the bmap operation because swapfiles
10526  * use bmap to make a mapping of extents in the file.  They assume
10527  * these extents won't change over the life of the file and they
10528  * use the bmap result to do IO directly to the drive.
10529  *
10530  * the btrfs bmap call would return logical addresses that aren't
10531  * suitable for IO and they also will change frequently as COW
10532  * operations happen.  So, swapfile + btrfs == corruption.
10533  *
10534  * For now we're avoiding this by dropping bmap.
10535  */
10536 static const struct address_space_operations btrfs_aops = {
10537         .readpage       = btrfs_readpage,
10538         .writepage      = btrfs_writepage,
10539         .writepages     = btrfs_writepages,
10540         .readpages      = btrfs_readpages,
10541         .direct_IO      = btrfs_direct_IO,
10542         .invalidatepage = btrfs_invalidatepage,
10543         .releasepage    = btrfs_releasepage,
10544         .set_page_dirty = btrfs_set_page_dirty,
10545         .error_remove_page = generic_error_remove_page,
10546 };
10547
10548 static const struct inode_operations btrfs_file_inode_operations = {
10549         .getattr        = btrfs_getattr,
10550         .setattr        = btrfs_setattr,
10551         .listxattr      = btrfs_listxattr,
10552         .permission     = btrfs_permission,
10553         .fiemap         = btrfs_fiemap,
10554         .get_acl        = btrfs_get_acl,
10555         .set_acl        = btrfs_set_acl,
10556         .update_time    = btrfs_update_time,
10557 };
10558 static const struct inode_operations btrfs_special_inode_operations = {
10559         .getattr        = btrfs_getattr,
10560         .setattr        = btrfs_setattr,
10561         .permission     = btrfs_permission,
10562         .listxattr      = btrfs_listxattr,
10563         .get_acl        = btrfs_get_acl,
10564         .set_acl        = btrfs_set_acl,
10565         .update_time    = btrfs_update_time,
10566 };
10567 static const struct inode_operations btrfs_symlink_inode_operations = {
10568         .get_link       = page_get_link,
10569         .getattr        = btrfs_getattr,
10570         .setattr        = btrfs_setattr,
10571         .permission     = btrfs_permission,
10572         .listxattr      = btrfs_listxattr,
10573         .update_time    = btrfs_update_time,
10574 };
10575
10576 const struct dentry_operations btrfs_dentry_operations = {
10577         .d_delete       = btrfs_dentry_delete,
10578 };