Merge tag 'for-5.8-rc4-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/kdave...
[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 <linux/swap.h>
31 #include <linux/migrate.h>
32 #include <linux/sched/mm.h>
33 #include <asm/unaligned.h>
34 #include "misc.h"
35 #include "ctree.h"
36 #include "disk-io.h"
37 #include "transaction.h"
38 #include "btrfs_inode.h"
39 #include "print-tree.h"
40 #include "ordered-data.h"
41 #include "xattr.h"
42 #include "tree-log.h"
43 #include "volumes.h"
44 #include "compression.h"
45 #include "locking.h"
46 #include "free-space-cache.h"
47 #include "inode-map.h"
48 #include "props.h"
49 #include "qgroup.h"
50 #include "delalloc-space.h"
51 #include "block-group.h"
52 #include "space-info.h"
53
54 struct btrfs_iget_args {
55         u64 ino;
56         struct btrfs_root *root;
57 };
58
59 struct btrfs_dio_data {
60         u64 reserve;
61         u64 unsubmitted_oe_range_start;
62         u64 unsubmitted_oe_range_end;
63         int overwrite;
64 };
65
66 static const struct inode_operations btrfs_dir_inode_operations;
67 static const struct inode_operations btrfs_symlink_inode_operations;
68 static const struct inode_operations btrfs_special_inode_operations;
69 static const struct inode_operations btrfs_file_inode_operations;
70 static const struct address_space_operations btrfs_aops;
71 static const struct file_operations btrfs_dir_file_operations;
72 static const struct extent_io_ops btrfs_extent_io_ops;
73
74 static struct kmem_cache *btrfs_inode_cachep;
75 struct kmem_cache *btrfs_trans_handle_cachep;
76 struct kmem_cache *btrfs_path_cachep;
77 struct kmem_cache *btrfs_free_space_cachep;
78 struct kmem_cache *btrfs_free_space_bitmap_cachep;
79
80 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
81 static int btrfs_truncate(struct inode *inode, bool skip_writeback);
82 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
83 static noinline int cow_file_range(struct inode *inode,
84                                    struct page *locked_page,
85                                    u64 start, u64 end, int *page_started,
86                                    unsigned long *nr_written, int unlock);
87 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
88                                        u64 orig_start, u64 block_start,
89                                        u64 block_len, u64 orig_block_len,
90                                        u64 ram_bytes, int compress_type,
91                                        int type);
92
93 static void __endio_write_update_ordered(struct inode *inode,
94                                          const u64 offset, const u64 bytes,
95                                          const bool uptodate);
96
97 /*
98  * Cleanup all submitted ordered extents in specified range to handle errors
99  * from the btrfs_run_delalloc_range() callback.
100  *
101  * NOTE: caller must ensure that when an error happens, it can not call
102  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
103  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
104  * to be released, which we want to happen only when finishing the ordered
105  * extent (btrfs_finish_ordered_io()).
106  */
107 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
108                                                  struct page *locked_page,
109                                                  u64 offset, u64 bytes)
110 {
111         unsigned long index = offset >> PAGE_SHIFT;
112         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
113         u64 page_start = page_offset(locked_page);
114         u64 page_end = page_start + PAGE_SIZE - 1;
115
116         struct page *page;
117
118         while (index <= end_index) {
119                 page = find_get_page(inode->i_mapping, index);
120                 index++;
121                 if (!page)
122                         continue;
123                 ClearPagePrivate2(page);
124                 put_page(page);
125         }
126
127         /*
128          * In case this page belongs to the delalloc range being instantiated
129          * then skip it, since the first page of a range is going to be
130          * properly cleaned up by the caller of run_delalloc_range
131          */
132         if (page_start >= offset && page_end <= (offset + bytes - 1)) {
133                 offset += PAGE_SIZE;
134                 bytes -= PAGE_SIZE;
135         }
136
137         return __endio_write_update_ordered(inode, offset, bytes, false);
138 }
139
140 static int btrfs_dirty_inode(struct inode *inode);
141
142 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
143 void btrfs_test_inode_set_ops(struct inode *inode)
144 {
145         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
146 }
147 #endif
148
149 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
150                                      struct inode *inode,  struct inode *dir,
151                                      const struct qstr *qstr)
152 {
153         int err;
154
155         err = btrfs_init_acl(trans, inode, dir);
156         if (!err)
157                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
158         return err;
159 }
160
161 /*
162  * this does all the hard work for inserting an inline extent into
163  * the btree.  The caller should have done a btrfs_drop_extents so that
164  * no overlapping inline items exist in the btree
165  */
166 static int insert_inline_extent(struct btrfs_trans_handle *trans,
167                                 struct btrfs_path *path, int extent_inserted,
168                                 struct btrfs_root *root, struct inode *inode,
169                                 u64 start, size_t size, size_t compressed_size,
170                                 int compress_type,
171                                 struct page **compressed_pages)
172 {
173         struct extent_buffer *leaf;
174         struct page *page = NULL;
175         char *kaddr;
176         unsigned long ptr;
177         struct btrfs_file_extent_item *ei;
178         int ret;
179         size_t cur_size = size;
180         unsigned long offset;
181
182         ASSERT((compressed_size > 0 && compressed_pages) ||
183                (compressed_size == 0 && !compressed_pages));
184
185         if (compressed_size && compressed_pages)
186                 cur_size = compressed_size;
187
188         inode_add_bytes(inode, size);
189
190         if (!extent_inserted) {
191                 struct btrfs_key key;
192                 size_t datasize;
193
194                 key.objectid = btrfs_ino(BTRFS_I(inode));
195                 key.offset = start;
196                 key.type = BTRFS_EXTENT_DATA_KEY;
197
198                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
199                 path->leave_spinning = 1;
200                 ret = btrfs_insert_empty_item(trans, root, path, &key,
201                                               datasize);
202                 if (ret)
203                         goto fail;
204         }
205         leaf = path->nodes[0];
206         ei = btrfs_item_ptr(leaf, path->slots[0],
207                             struct btrfs_file_extent_item);
208         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
209         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
210         btrfs_set_file_extent_encryption(leaf, ei, 0);
211         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
212         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
213         ptr = btrfs_file_extent_inline_start(ei);
214
215         if (compress_type != BTRFS_COMPRESS_NONE) {
216                 struct page *cpage;
217                 int i = 0;
218                 while (compressed_size > 0) {
219                         cpage = compressed_pages[i];
220                         cur_size = min_t(unsigned long, compressed_size,
221                                        PAGE_SIZE);
222
223                         kaddr = kmap_atomic(cpage);
224                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
225                         kunmap_atomic(kaddr);
226
227                         i++;
228                         ptr += cur_size;
229                         compressed_size -= cur_size;
230                 }
231                 btrfs_set_file_extent_compression(leaf, ei,
232                                                   compress_type);
233         } else {
234                 page = find_get_page(inode->i_mapping,
235                                      start >> PAGE_SHIFT);
236                 btrfs_set_file_extent_compression(leaf, ei, 0);
237                 kaddr = kmap_atomic(page);
238                 offset = offset_in_page(start);
239                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
240                 kunmap_atomic(kaddr);
241                 put_page(page);
242         }
243         btrfs_mark_buffer_dirty(leaf);
244         btrfs_release_path(path);
245
246         /*
247          * We align size to sectorsize for inline extents just for simplicity
248          * sake.
249          */
250         size = ALIGN(size, root->fs_info->sectorsize);
251         ret = btrfs_inode_set_file_extent_range(BTRFS_I(inode), start, size);
252         if (ret)
253                 goto fail;
254
255         /*
256          * we're an inline extent, so nobody can
257          * extend the file past i_size without locking
258          * a page we already have locked.
259          *
260          * We must do any isize and inode updates
261          * before we unlock the pages.  Otherwise we
262          * could end up racing with unlink.
263          */
264         BTRFS_I(inode)->disk_i_size = inode->i_size;
265         ret = btrfs_update_inode(trans, root, inode);
266
267 fail:
268         return ret;
269 }
270
271
272 /*
273  * conditionally insert an inline extent into the file.  This
274  * does the checks required to make sure the data is small enough
275  * to fit as an inline extent.
276  */
277 static noinline int cow_file_range_inline(struct inode *inode, u64 start,
278                                           u64 end, size_t compressed_size,
279                                           int compress_type,
280                                           struct page **compressed_pages)
281 {
282         struct btrfs_root *root = BTRFS_I(inode)->root;
283         struct btrfs_fs_info *fs_info = root->fs_info;
284         struct btrfs_trans_handle *trans;
285         u64 isize = i_size_read(inode);
286         u64 actual_end = min(end + 1, isize);
287         u64 inline_len = actual_end - start;
288         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
289         u64 data_len = inline_len;
290         int ret;
291         struct btrfs_path *path;
292         int extent_inserted = 0;
293         u32 extent_item_size;
294
295         if (compressed_size)
296                 data_len = compressed_size;
297
298         if (start > 0 ||
299             actual_end > fs_info->sectorsize ||
300             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
301             (!compressed_size &&
302             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
303             end + 1 < isize ||
304             data_len > fs_info->max_inline) {
305                 return 1;
306         }
307
308         path = btrfs_alloc_path();
309         if (!path)
310                 return -ENOMEM;
311
312         trans = btrfs_join_transaction(root);
313         if (IS_ERR(trans)) {
314                 btrfs_free_path(path);
315                 return PTR_ERR(trans);
316         }
317         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
318
319         if (compressed_size && compressed_pages)
320                 extent_item_size = btrfs_file_extent_calc_inline_size(
321                    compressed_size);
322         else
323                 extent_item_size = btrfs_file_extent_calc_inline_size(
324                     inline_len);
325
326         ret = __btrfs_drop_extents(trans, root, inode, path,
327                                    start, aligned_end, NULL,
328                                    1, 1, extent_item_size, &extent_inserted);
329         if (ret) {
330                 btrfs_abort_transaction(trans, ret);
331                 goto out;
332         }
333
334         if (isize > actual_end)
335                 inline_len = min_t(u64, isize, actual_end);
336         ret = insert_inline_extent(trans, path, extent_inserted,
337                                    root, inode, start,
338                                    inline_len, compressed_size,
339                                    compress_type, compressed_pages);
340         if (ret && ret != -ENOSPC) {
341                 btrfs_abort_transaction(trans, ret);
342                 goto out;
343         } else if (ret == -ENOSPC) {
344                 ret = 1;
345                 goto out;
346         }
347
348         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
349         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
350 out:
351         /*
352          * Don't forget to free the reserved space, as for inlined extent
353          * it won't count as data extent, free them directly here.
354          * And at reserve time, it's always aligned to page size, so
355          * just free one page here.
356          */
357         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
358         btrfs_free_path(path);
359         btrfs_end_transaction(trans);
360         return ret;
361 }
362
363 struct async_extent {
364         u64 start;
365         u64 ram_size;
366         u64 compressed_size;
367         struct page **pages;
368         unsigned long nr_pages;
369         int compress_type;
370         struct list_head list;
371 };
372
373 struct async_chunk {
374         struct inode *inode;
375         struct page *locked_page;
376         u64 start;
377         u64 end;
378         unsigned int write_flags;
379         struct list_head extents;
380         struct cgroup_subsys_state *blkcg_css;
381         struct btrfs_work work;
382         atomic_t *pending;
383 };
384
385 struct async_cow {
386         /* Number of chunks in flight; must be first in the structure */
387         atomic_t num_chunks;
388         struct async_chunk chunks[];
389 };
390
391 static noinline int add_async_extent(struct async_chunk *cow,
392                                      u64 start, u64 ram_size,
393                                      u64 compressed_size,
394                                      struct page **pages,
395                                      unsigned long nr_pages,
396                                      int compress_type)
397 {
398         struct async_extent *async_extent;
399
400         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
401         BUG_ON(!async_extent); /* -ENOMEM */
402         async_extent->start = start;
403         async_extent->ram_size = ram_size;
404         async_extent->compressed_size = compressed_size;
405         async_extent->pages = pages;
406         async_extent->nr_pages = nr_pages;
407         async_extent->compress_type = compress_type;
408         list_add_tail(&async_extent->list, &cow->extents);
409         return 0;
410 }
411
412 /*
413  * Check if the inode has flags compatible with compression
414  */
415 static inline bool inode_can_compress(struct inode *inode)
416 {
417         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW ||
418             BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
419                 return false;
420         return true;
421 }
422
423 /*
424  * Check if the inode needs to be submitted to compression, based on mount
425  * options, defragmentation, properties or heuristics.
426  */
427 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
428 {
429         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
430
431         if (!inode_can_compress(inode)) {
432                 WARN(IS_ENABLED(CONFIG_BTRFS_DEBUG),
433                         KERN_ERR "BTRFS: unexpected compression for ino %llu\n",
434                         btrfs_ino(BTRFS_I(inode)));
435                 return 0;
436         }
437         /* force compress */
438         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
439                 return 1;
440         /* defrag ioctl */
441         if (BTRFS_I(inode)->defrag_compress)
442                 return 1;
443         /* bad compression ratios */
444         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
445                 return 0;
446         if (btrfs_test_opt(fs_info, COMPRESS) ||
447             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
448             BTRFS_I(inode)->prop_compress)
449                 return btrfs_compress_heuristic(inode, start, end);
450         return 0;
451 }
452
453 static inline void inode_should_defrag(struct btrfs_inode *inode,
454                 u64 start, u64 end, u64 num_bytes, u64 small_write)
455 {
456         /* If this is a small write inside eof, kick off a defrag */
457         if (num_bytes < small_write &&
458             (start > 0 || end + 1 < inode->disk_i_size))
459                 btrfs_add_inode_defrag(NULL, inode);
460 }
461
462 /*
463  * we create compressed extents in two phases.  The first
464  * phase compresses a range of pages that have already been
465  * locked (both pages and state bits are locked).
466  *
467  * This is done inside an ordered work queue, and the compression
468  * is spread across many cpus.  The actual IO submission is step
469  * two, and the ordered work queue takes care of making sure that
470  * happens in the same order things were put onto the queue by
471  * writepages and friends.
472  *
473  * If this code finds it can't get good compression, it puts an
474  * entry onto the work queue to write the uncompressed bytes.  This
475  * makes sure that both compressed inodes and uncompressed inodes
476  * are written in the same order that the flusher thread sent them
477  * down.
478  */
479 static noinline int compress_file_range(struct async_chunk *async_chunk)
480 {
481         struct inode *inode = async_chunk->inode;
482         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
483         u64 blocksize = fs_info->sectorsize;
484         u64 start = async_chunk->start;
485         u64 end = async_chunk->end;
486         u64 actual_end;
487         u64 i_size;
488         int ret = 0;
489         struct page **pages = NULL;
490         unsigned long nr_pages;
491         unsigned long total_compressed = 0;
492         unsigned long total_in = 0;
493         int i;
494         int will_compress;
495         int compress_type = fs_info->compress_type;
496         int compressed_extents = 0;
497         int redirty = 0;
498
499         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
500                         SZ_16K);
501
502         /*
503          * We need to save i_size before now because it could change in between
504          * us evaluating the size and assigning it.  This is because we lock and
505          * unlock the page in truncate and fallocate, and then modify the i_size
506          * later on.
507          *
508          * The barriers are to emulate READ_ONCE, remove that once i_size_read
509          * does that for us.
510          */
511         barrier();
512         i_size = i_size_read(inode);
513         barrier();
514         actual_end = min_t(u64, i_size, end + 1);
515 again:
516         will_compress = 0;
517         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
518         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
519         nr_pages = min_t(unsigned long, nr_pages,
520                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
521
522         /*
523          * we don't want to send crud past the end of i_size through
524          * compression, that's just a waste of CPU time.  So, if the
525          * end of the file is before the start of our current
526          * requested range of bytes, we bail out to the uncompressed
527          * cleanup code that can deal with all of this.
528          *
529          * It isn't really the fastest way to fix things, but this is a
530          * very uncommon corner.
531          */
532         if (actual_end <= start)
533                 goto cleanup_and_bail_uncompressed;
534
535         total_compressed = actual_end - start;
536
537         /*
538          * skip compression for a small file range(<=blocksize) that
539          * isn't an inline extent, since it doesn't save disk space at all.
540          */
541         if (total_compressed <= blocksize &&
542            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
543                 goto cleanup_and_bail_uncompressed;
544
545         total_compressed = min_t(unsigned long, total_compressed,
546                         BTRFS_MAX_UNCOMPRESSED);
547         total_in = 0;
548         ret = 0;
549
550         /*
551          * we do compression for mount -o compress and when the
552          * inode has not been flagged as nocompress.  This flag can
553          * change at any time if we discover bad compression ratios.
554          */
555         if (inode_need_compress(inode, start, end)) {
556                 WARN_ON(pages);
557                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
558                 if (!pages) {
559                         /* just bail out to the uncompressed code */
560                         nr_pages = 0;
561                         goto cont;
562                 }
563
564                 if (BTRFS_I(inode)->defrag_compress)
565                         compress_type = BTRFS_I(inode)->defrag_compress;
566                 else if (BTRFS_I(inode)->prop_compress)
567                         compress_type = BTRFS_I(inode)->prop_compress;
568
569                 /*
570                  * we need to call clear_page_dirty_for_io on each
571                  * page in the range.  Otherwise applications with the file
572                  * mmap'd can wander in and change the page contents while
573                  * we are compressing them.
574                  *
575                  * If the compression fails for any reason, we set the pages
576                  * dirty again later on.
577                  *
578                  * Note that the remaining part is redirtied, the start pointer
579                  * has moved, the end is the original one.
580                  */
581                 if (!redirty) {
582                         extent_range_clear_dirty_for_io(inode, start, end);
583                         redirty = 1;
584                 }
585
586                 /* Compression level is applied here and only here */
587                 ret = btrfs_compress_pages(
588                         compress_type | (fs_info->compress_level << 4),
589                                            inode->i_mapping, start,
590                                            pages,
591                                            &nr_pages,
592                                            &total_in,
593                                            &total_compressed);
594
595                 if (!ret) {
596                         unsigned long offset = offset_in_page(total_compressed);
597                         struct page *page = pages[nr_pages - 1];
598                         char *kaddr;
599
600                         /* zero the tail end of the last page, we might be
601                          * sending it down to disk
602                          */
603                         if (offset) {
604                                 kaddr = kmap_atomic(page);
605                                 memset(kaddr + offset, 0,
606                                        PAGE_SIZE - offset);
607                                 kunmap_atomic(kaddr);
608                         }
609                         will_compress = 1;
610                 }
611         }
612 cont:
613         if (start == 0) {
614                 /* lets try to make an inline extent */
615                 if (ret || total_in < actual_end) {
616                         /* we didn't compress the entire range, try
617                          * to make an uncompressed inline extent.
618                          */
619                         ret = cow_file_range_inline(inode, start, end, 0,
620                                                     BTRFS_COMPRESS_NONE, NULL);
621                 } else {
622                         /* try making a compressed inline extent */
623                         ret = cow_file_range_inline(inode, start, end,
624                                                     total_compressed,
625                                                     compress_type, pages);
626                 }
627                 if (ret <= 0) {
628                         unsigned long clear_flags = EXTENT_DELALLOC |
629                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
630                                 EXTENT_DO_ACCOUNTING;
631                         unsigned long page_error_op;
632
633                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
634
635                         /*
636                          * inline extent creation worked or returned error,
637                          * we don't need to create any more async work items.
638                          * Unlock and free up our temp pages.
639                          *
640                          * We use DO_ACCOUNTING here because we need the
641                          * delalloc_release_metadata to be done _after_ we drop
642                          * our outstanding extent for clearing delalloc for this
643                          * range.
644                          */
645                         extent_clear_unlock_delalloc(inode, start, end, NULL,
646                                                      clear_flags,
647                                                      PAGE_UNLOCK |
648                                                      PAGE_CLEAR_DIRTY |
649                                                      PAGE_SET_WRITEBACK |
650                                                      page_error_op |
651                                                      PAGE_END_WRITEBACK);
652
653                         for (i = 0; i < nr_pages; i++) {
654                                 WARN_ON(pages[i]->mapping);
655                                 put_page(pages[i]);
656                         }
657                         kfree(pages);
658
659                         return 0;
660                 }
661         }
662
663         if (will_compress) {
664                 /*
665                  * we aren't doing an inline extent round the compressed size
666                  * up to a block size boundary so the allocator does sane
667                  * things
668                  */
669                 total_compressed = ALIGN(total_compressed, blocksize);
670
671                 /*
672                  * one last check to make sure the compression is really a
673                  * win, compare the page count read with the blocks on disk,
674                  * compression must free at least one sector size
675                  */
676                 total_in = ALIGN(total_in, PAGE_SIZE);
677                 if (total_compressed + blocksize <= total_in) {
678                         compressed_extents++;
679
680                         /*
681                          * The async work queues will take care of doing actual
682                          * allocation on disk for these compressed pages, and
683                          * will submit them to the elevator.
684                          */
685                         add_async_extent(async_chunk, start, total_in,
686                                         total_compressed, pages, nr_pages,
687                                         compress_type);
688
689                         if (start + total_in < end) {
690                                 start += total_in;
691                                 pages = NULL;
692                                 cond_resched();
693                                 goto again;
694                         }
695                         return compressed_extents;
696                 }
697         }
698         if (pages) {
699                 /*
700                  * the compression code ran but failed to make things smaller,
701                  * free any pages it allocated and our page pointer array
702                  */
703                 for (i = 0; i < nr_pages; i++) {
704                         WARN_ON(pages[i]->mapping);
705                         put_page(pages[i]);
706                 }
707                 kfree(pages);
708                 pages = NULL;
709                 total_compressed = 0;
710                 nr_pages = 0;
711
712                 /* flag the file so we don't compress in the future */
713                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
714                     !(BTRFS_I(inode)->prop_compress)) {
715                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
716                 }
717         }
718 cleanup_and_bail_uncompressed:
719         /*
720          * No compression, but we still need to write the pages in the file
721          * we've been given so far.  redirty the locked page if it corresponds
722          * to our extent and set things up for the async work queue to run
723          * cow_file_range to do the normal delalloc dance.
724          */
725         if (async_chunk->locked_page &&
726             (page_offset(async_chunk->locked_page) >= start &&
727              page_offset(async_chunk->locked_page)) <= end) {
728                 __set_page_dirty_nobuffers(async_chunk->locked_page);
729                 /* unlocked later on in the async handlers */
730         }
731
732         if (redirty)
733                 extent_range_redirty_for_io(inode, start, end);
734         add_async_extent(async_chunk, start, end - start + 1, 0, NULL, 0,
735                          BTRFS_COMPRESS_NONE);
736         compressed_extents++;
737
738         return compressed_extents;
739 }
740
741 static void free_async_extent_pages(struct async_extent *async_extent)
742 {
743         int i;
744
745         if (!async_extent->pages)
746                 return;
747
748         for (i = 0; i < async_extent->nr_pages; i++) {
749                 WARN_ON(async_extent->pages[i]->mapping);
750                 put_page(async_extent->pages[i]);
751         }
752         kfree(async_extent->pages);
753         async_extent->nr_pages = 0;
754         async_extent->pages = NULL;
755 }
756
757 /*
758  * phase two of compressed writeback.  This is the ordered portion
759  * of the code, which only gets called in the order the work was
760  * queued.  We walk all the async extents created by compress_file_range
761  * and send them down to the disk.
762  */
763 static noinline void submit_compressed_extents(struct async_chunk *async_chunk)
764 {
765         struct inode *inode = async_chunk->inode;
766         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
767         struct async_extent *async_extent;
768         u64 alloc_hint = 0;
769         struct btrfs_key ins;
770         struct extent_map *em;
771         struct btrfs_root *root = BTRFS_I(inode)->root;
772         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
773         int ret = 0;
774
775 again:
776         while (!list_empty(&async_chunk->extents)) {
777                 async_extent = list_entry(async_chunk->extents.next,
778                                           struct async_extent, list);
779                 list_del(&async_extent->list);
780
781 retry:
782                 lock_extent(io_tree, async_extent->start,
783                             async_extent->start + async_extent->ram_size - 1);
784                 /* did the compression code fall back to uncompressed IO? */
785                 if (!async_extent->pages) {
786                         int page_started = 0;
787                         unsigned long nr_written = 0;
788
789                         /* allocate blocks */
790                         ret = cow_file_range(inode, async_chunk->locked_page,
791                                              async_extent->start,
792                                              async_extent->start +
793                                              async_extent->ram_size - 1,
794                                              &page_started, &nr_written, 0);
795
796                         /* JDM XXX */
797
798                         /*
799                          * if page_started, cow_file_range inserted an
800                          * inline extent and took care of all the unlocking
801                          * and IO for us.  Otherwise, we need to submit
802                          * all those pages down to the drive.
803                          */
804                         if (!page_started && !ret)
805                                 extent_write_locked_range(inode,
806                                                   async_extent->start,
807                                                   async_extent->start +
808                                                   async_extent->ram_size - 1,
809                                                   WB_SYNC_ALL);
810                         else if (ret && async_chunk->locked_page)
811                                 unlock_page(async_chunk->locked_page);
812                         kfree(async_extent);
813                         cond_resched();
814                         continue;
815                 }
816
817                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
818                                            async_extent->compressed_size,
819                                            async_extent->compressed_size,
820                                            0, alloc_hint, &ins, 1, 1);
821                 if (ret) {
822                         free_async_extent_pages(async_extent);
823
824                         if (ret == -ENOSPC) {
825                                 unlock_extent(io_tree, async_extent->start,
826                                               async_extent->start +
827                                               async_extent->ram_size - 1);
828
829                                 /*
830                                  * we need to redirty the pages if we decide to
831                                  * fallback to uncompressed IO, otherwise we
832                                  * will not submit these pages down to lower
833                                  * layers.
834                                  */
835                                 extent_range_redirty_for_io(inode,
836                                                 async_extent->start,
837                                                 async_extent->start +
838                                                 async_extent->ram_size - 1);
839
840                                 goto retry;
841                         }
842                         goto out_free;
843                 }
844                 /*
845                  * here we're doing allocation and writeback of the
846                  * compressed pages
847                  */
848                 em = create_io_em(inode, async_extent->start,
849                                   async_extent->ram_size, /* len */
850                                   async_extent->start, /* orig_start */
851                                   ins.objectid, /* block_start */
852                                   ins.offset, /* block_len */
853                                   ins.offset, /* orig_block_len */
854                                   async_extent->ram_size, /* ram_bytes */
855                                   async_extent->compress_type,
856                                   BTRFS_ORDERED_COMPRESSED);
857                 if (IS_ERR(em))
858                         /* ret value is not necessary due to void function */
859                         goto out_free_reserve;
860                 free_extent_map(em);
861
862                 ret = btrfs_add_ordered_extent_compress(inode,
863                                                 async_extent->start,
864                                                 ins.objectid,
865                                                 async_extent->ram_size,
866                                                 ins.offset,
867                                                 BTRFS_ORDERED_COMPRESSED,
868                                                 async_extent->compress_type);
869                 if (ret) {
870                         btrfs_drop_extent_cache(BTRFS_I(inode),
871                                                 async_extent->start,
872                                                 async_extent->start +
873                                                 async_extent->ram_size - 1, 0);
874                         goto out_free_reserve;
875                 }
876                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
877
878                 /*
879                  * clear dirty, set writeback and unlock the pages.
880                  */
881                 extent_clear_unlock_delalloc(inode, async_extent->start,
882                                 async_extent->start +
883                                 async_extent->ram_size - 1,
884                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
885                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
886                                 PAGE_SET_WRITEBACK);
887                 if (btrfs_submit_compressed_write(inode,
888                                     async_extent->start,
889                                     async_extent->ram_size,
890                                     ins.objectid,
891                                     ins.offset, async_extent->pages,
892                                     async_extent->nr_pages,
893                                     async_chunk->write_flags,
894                                     async_chunk->blkcg_css)) {
895                         struct page *p = async_extent->pages[0];
896                         const u64 start = async_extent->start;
897                         const u64 end = start + async_extent->ram_size - 1;
898
899                         p->mapping = inode->i_mapping;
900                         btrfs_writepage_endio_finish_ordered(p, start, end, 0);
901
902                         p->mapping = NULL;
903                         extent_clear_unlock_delalloc(inode, start, end,
904                                                      NULL, 0,
905                                                      PAGE_END_WRITEBACK |
906                                                      PAGE_SET_ERROR);
907                         free_async_extent_pages(async_extent);
908                 }
909                 alloc_hint = ins.objectid + ins.offset;
910                 kfree(async_extent);
911                 cond_resched();
912         }
913         return;
914 out_free_reserve:
915         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
916         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
917 out_free:
918         extent_clear_unlock_delalloc(inode, async_extent->start,
919                                      async_extent->start +
920                                      async_extent->ram_size - 1,
921                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
922                                      EXTENT_DELALLOC_NEW |
923                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
924                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
925                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
926                                      PAGE_SET_ERROR);
927         free_async_extent_pages(async_extent);
928         kfree(async_extent);
929         goto again;
930 }
931
932 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
933                                       u64 num_bytes)
934 {
935         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
936         struct extent_map *em;
937         u64 alloc_hint = 0;
938
939         read_lock(&em_tree->lock);
940         em = search_extent_mapping(em_tree, start, num_bytes);
941         if (em) {
942                 /*
943                  * if block start isn't an actual block number then find the
944                  * first block in this inode and use that as a hint.  If that
945                  * block is also bogus then just don't worry about it.
946                  */
947                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
948                         free_extent_map(em);
949                         em = search_extent_mapping(em_tree, 0, 0);
950                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
951                                 alloc_hint = em->block_start;
952                         if (em)
953                                 free_extent_map(em);
954                 } else {
955                         alloc_hint = em->block_start;
956                         free_extent_map(em);
957                 }
958         }
959         read_unlock(&em_tree->lock);
960
961         return alloc_hint;
962 }
963
964 /*
965  * when extent_io.c finds a delayed allocation range in the file,
966  * the call backs end up in this code.  The basic idea is to
967  * allocate extents on disk for the range, and create ordered data structs
968  * in ram to track those extents.
969  *
970  * locked_page is the page that writepage had locked already.  We use
971  * it to make sure we don't do extra locks or unlocks.
972  *
973  * *page_started is set to one if we unlock locked_page and do everything
974  * required to start IO on it.  It may be clean and already done with
975  * IO when we return.
976  */
977 static noinline int cow_file_range(struct inode *inode,
978                                    struct page *locked_page,
979                                    u64 start, u64 end, int *page_started,
980                                    unsigned long *nr_written, int unlock)
981 {
982         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
983         struct btrfs_root *root = BTRFS_I(inode)->root;
984         u64 alloc_hint = 0;
985         u64 num_bytes;
986         unsigned long ram_size;
987         u64 cur_alloc_size = 0;
988         u64 min_alloc_size;
989         u64 blocksize = fs_info->sectorsize;
990         struct btrfs_key ins;
991         struct extent_map *em;
992         unsigned clear_bits;
993         unsigned long page_ops;
994         bool extent_reserved = false;
995         int ret = 0;
996
997         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
998                 WARN_ON_ONCE(1);
999                 ret = -EINVAL;
1000                 goto out_unlock;
1001         }
1002
1003         num_bytes = ALIGN(end - start + 1, blocksize);
1004         num_bytes = max(blocksize,  num_bytes);
1005         ASSERT(num_bytes <= btrfs_super_total_bytes(fs_info->super_copy));
1006
1007         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
1008
1009         if (start == 0) {
1010                 /* lets try to make an inline extent */
1011                 ret = cow_file_range_inline(inode, start, end, 0,
1012                                             BTRFS_COMPRESS_NONE, NULL);
1013                 if (ret == 0) {
1014                         /*
1015                          * We use DO_ACCOUNTING here because we need the
1016                          * delalloc_release_metadata to be run _after_ we drop
1017                          * our outstanding extent for clearing delalloc for this
1018                          * range.
1019                          */
1020                         extent_clear_unlock_delalloc(inode, start, end, NULL,
1021                                      EXTENT_LOCKED | EXTENT_DELALLOC |
1022                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
1023                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1024                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1025                                      PAGE_END_WRITEBACK);
1026                         *nr_written = *nr_written +
1027                              (end - start + PAGE_SIZE) / PAGE_SIZE;
1028                         *page_started = 1;
1029                         goto out;
1030                 } else if (ret < 0) {
1031                         goto out_unlock;
1032                 }
1033         }
1034
1035         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
1036         btrfs_drop_extent_cache(BTRFS_I(inode), start,
1037                         start + num_bytes - 1, 0);
1038
1039         /*
1040          * Relocation relies on the relocated extents to have exactly the same
1041          * size as the original extents. Normally writeback for relocation data
1042          * extents follows a NOCOW path because relocation preallocates the
1043          * extents. However, due to an operation such as scrub turning a block
1044          * group to RO mode, it may fallback to COW mode, so we must make sure
1045          * an extent allocated during COW has exactly the requested size and can
1046          * not be split into smaller extents, otherwise relocation breaks and
1047          * fails during the stage where it updates the bytenr of file extent
1048          * items.
1049          */
1050         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1051                 min_alloc_size = num_bytes;
1052         else
1053                 min_alloc_size = fs_info->sectorsize;
1054
1055         while (num_bytes > 0) {
1056                 cur_alloc_size = num_bytes;
1057                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1058                                            min_alloc_size, 0, alloc_hint,
1059                                            &ins, 1, 1);
1060                 if (ret < 0)
1061                         goto out_unlock;
1062                 cur_alloc_size = ins.offset;
1063                 extent_reserved = true;
1064
1065                 ram_size = ins.offset;
1066                 em = create_io_em(inode, start, ins.offset, /* len */
1067                                   start, /* orig_start */
1068                                   ins.objectid, /* block_start */
1069                                   ins.offset, /* block_len */
1070                                   ins.offset, /* orig_block_len */
1071                                   ram_size, /* ram_bytes */
1072                                   BTRFS_COMPRESS_NONE, /* compress_type */
1073                                   BTRFS_ORDERED_REGULAR /* type */);
1074                 if (IS_ERR(em)) {
1075                         ret = PTR_ERR(em);
1076                         goto out_reserve;
1077                 }
1078                 free_extent_map(em);
1079
1080                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1081                                                ram_size, cur_alloc_size, 0);
1082                 if (ret)
1083                         goto out_drop_extent_cache;
1084
1085                 if (root->root_key.objectid ==
1086                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1087                         ret = btrfs_reloc_clone_csums(inode, start,
1088                                                       cur_alloc_size);
1089                         /*
1090                          * Only drop cache here, and process as normal.
1091                          *
1092                          * We must not allow extent_clear_unlock_delalloc()
1093                          * at out_unlock label to free meta of this ordered
1094                          * extent, as its meta should be freed by
1095                          * btrfs_finish_ordered_io().
1096                          *
1097                          * So we must continue until @start is increased to
1098                          * skip current ordered extent.
1099                          */
1100                         if (ret)
1101                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1102                                                 start + ram_size - 1, 0);
1103                 }
1104
1105                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1106
1107                 /* we're not doing compressed IO, don't unlock the first
1108                  * page (which the caller expects to stay locked), don't
1109                  * clear any dirty bits and don't set any writeback bits
1110                  *
1111                  * Do set the Private2 bit so we know this page was properly
1112                  * setup for writepage
1113                  */
1114                 page_ops = unlock ? PAGE_UNLOCK : 0;
1115                 page_ops |= PAGE_SET_PRIVATE2;
1116
1117                 extent_clear_unlock_delalloc(inode, start,
1118                                              start + ram_size - 1,
1119                                              locked_page,
1120                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1121                                              page_ops);
1122                 if (num_bytes < cur_alloc_size)
1123                         num_bytes = 0;
1124                 else
1125                         num_bytes -= cur_alloc_size;
1126                 alloc_hint = ins.objectid + ins.offset;
1127                 start += cur_alloc_size;
1128                 extent_reserved = false;
1129
1130                 /*
1131                  * btrfs_reloc_clone_csums() error, since start is increased
1132                  * extent_clear_unlock_delalloc() at out_unlock label won't
1133                  * free metadata of current ordered extent, we're OK to exit.
1134                  */
1135                 if (ret)
1136                         goto out_unlock;
1137         }
1138 out:
1139         return ret;
1140
1141 out_drop_extent_cache:
1142         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1143 out_reserve:
1144         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1145         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1146 out_unlock:
1147         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1148                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1149         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1150                 PAGE_END_WRITEBACK;
1151         /*
1152          * If we reserved an extent for our delalloc range (or a subrange) and
1153          * failed to create the respective ordered extent, then it means that
1154          * when we reserved the extent we decremented the extent's size from
1155          * the data space_info's bytes_may_use counter and incremented the
1156          * space_info's bytes_reserved counter by the same amount. We must make
1157          * sure extent_clear_unlock_delalloc() does not try to decrement again
1158          * the data space_info's bytes_may_use counter, therefore we do not pass
1159          * it the flag EXTENT_CLEAR_DATA_RESV.
1160          */
1161         if (extent_reserved) {
1162                 extent_clear_unlock_delalloc(inode, start,
1163                                              start + cur_alloc_size - 1,
1164                                              locked_page,
1165                                              clear_bits,
1166                                              page_ops);
1167                 start += cur_alloc_size;
1168                 if (start >= end)
1169                         goto out;
1170         }
1171         extent_clear_unlock_delalloc(inode, start, end, locked_page,
1172                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1173                                      page_ops);
1174         goto out;
1175 }
1176
1177 /*
1178  * work queue call back to started compression on a file and pages
1179  */
1180 static noinline void async_cow_start(struct btrfs_work *work)
1181 {
1182         struct async_chunk *async_chunk;
1183         int compressed_extents;
1184
1185         async_chunk = container_of(work, struct async_chunk, work);
1186
1187         compressed_extents = compress_file_range(async_chunk);
1188         if (compressed_extents == 0) {
1189                 btrfs_add_delayed_iput(async_chunk->inode);
1190                 async_chunk->inode = NULL;
1191         }
1192 }
1193
1194 /*
1195  * work queue call back to submit previously compressed pages
1196  */
1197 static noinline void async_cow_submit(struct btrfs_work *work)
1198 {
1199         struct async_chunk *async_chunk = container_of(work, struct async_chunk,
1200                                                      work);
1201         struct btrfs_fs_info *fs_info = btrfs_work_owner(work);
1202         unsigned long nr_pages;
1203
1204         nr_pages = (async_chunk->end - async_chunk->start + PAGE_SIZE) >>
1205                 PAGE_SHIFT;
1206
1207         /* atomic_sub_return implies a barrier */
1208         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1209             5 * SZ_1M)
1210                 cond_wake_up_nomb(&fs_info->async_submit_wait);
1211
1212         /*
1213          * ->inode could be NULL if async_chunk_start has failed to compress,
1214          * in which case we don't have anything to submit, yet we need to
1215          * always adjust ->async_delalloc_pages as its paired with the init
1216          * happening in cow_file_range_async
1217          */
1218         if (async_chunk->inode)
1219                 submit_compressed_extents(async_chunk);
1220 }
1221
1222 static noinline void async_cow_free(struct btrfs_work *work)
1223 {
1224         struct async_chunk *async_chunk;
1225
1226         async_chunk = container_of(work, struct async_chunk, work);
1227         if (async_chunk->inode)
1228                 btrfs_add_delayed_iput(async_chunk->inode);
1229         if (async_chunk->blkcg_css)
1230                 css_put(async_chunk->blkcg_css);
1231         /*
1232          * Since the pointer to 'pending' is at the beginning of the array of
1233          * async_chunk's, freeing it ensures the whole array has been freed.
1234          */
1235         if (atomic_dec_and_test(async_chunk->pending))
1236                 kvfree(async_chunk->pending);
1237 }
1238
1239 static int cow_file_range_async(struct inode *inode,
1240                                 struct writeback_control *wbc,
1241                                 struct page *locked_page,
1242                                 u64 start, u64 end, int *page_started,
1243                                 unsigned long *nr_written)
1244 {
1245         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1246         struct cgroup_subsys_state *blkcg_css = wbc_blkcg_css(wbc);
1247         struct async_cow *ctx;
1248         struct async_chunk *async_chunk;
1249         unsigned long nr_pages;
1250         u64 cur_end;
1251         u64 num_chunks = DIV_ROUND_UP(end - start, SZ_512K);
1252         int i;
1253         bool should_compress;
1254         unsigned nofs_flag;
1255         const unsigned int write_flags = wbc_to_write_flags(wbc);
1256
1257         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
1258
1259         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1260             !btrfs_test_opt(fs_info, FORCE_COMPRESS)) {
1261                 num_chunks = 1;
1262                 should_compress = false;
1263         } else {
1264                 should_compress = true;
1265         }
1266
1267         nofs_flag = memalloc_nofs_save();
1268         ctx = kvmalloc(struct_size(ctx, chunks, num_chunks), GFP_KERNEL);
1269         memalloc_nofs_restore(nofs_flag);
1270
1271         if (!ctx) {
1272                 unsigned clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC |
1273                         EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
1274                         EXTENT_DO_ACCOUNTING;
1275                 unsigned long page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1276                         PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
1277                         PAGE_SET_ERROR;
1278
1279                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1280                                              clear_bits, page_ops);
1281                 return -ENOMEM;
1282         }
1283
1284         async_chunk = ctx->chunks;
1285         atomic_set(&ctx->num_chunks, num_chunks);
1286
1287         for (i = 0; i < num_chunks; i++) {
1288                 if (should_compress)
1289                         cur_end = min(end, start + SZ_512K - 1);
1290                 else
1291                         cur_end = end;
1292
1293                 /*
1294                  * igrab is called higher up in the call chain, take only the
1295                  * lightweight reference for the callback lifetime
1296                  */
1297                 ihold(inode);
1298                 async_chunk[i].pending = &ctx->num_chunks;
1299                 async_chunk[i].inode = inode;
1300                 async_chunk[i].start = start;
1301                 async_chunk[i].end = cur_end;
1302                 async_chunk[i].write_flags = write_flags;
1303                 INIT_LIST_HEAD(&async_chunk[i].extents);
1304
1305                 /*
1306                  * The locked_page comes all the way from writepage and its
1307                  * the original page we were actually given.  As we spread
1308                  * this large delalloc region across multiple async_chunk
1309                  * structs, only the first struct needs a pointer to locked_page
1310                  *
1311                  * This way we don't need racey decisions about who is supposed
1312                  * to unlock it.
1313                  */
1314                 if (locked_page) {
1315                         /*
1316                          * Depending on the compressibility, the pages might or
1317                          * might not go through async.  We want all of them to
1318                          * be accounted against wbc once.  Let's do it here
1319                          * before the paths diverge.  wbc accounting is used
1320                          * only for foreign writeback detection and doesn't
1321                          * need full accuracy.  Just account the whole thing
1322                          * against the first page.
1323                          */
1324                         wbc_account_cgroup_owner(wbc, locked_page,
1325                                                  cur_end - start);
1326                         async_chunk[i].locked_page = locked_page;
1327                         locked_page = NULL;
1328                 } else {
1329                         async_chunk[i].locked_page = NULL;
1330                 }
1331
1332                 if (blkcg_css != blkcg_root_css) {
1333                         css_get(blkcg_css);
1334                         async_chunk[i].blkcg_css = blkcg_css;
1335                 } else {
1336                         async_chunk[i].blkcg_css = NULL;
1337                 }
1338
1339                 btrfs_init_work(&async_chunk[i].work, async_cow_start,
1340                                 async_cow_submit, async_cow_free);
1341
1342                 nr_pages = DIV_ROUND_UP(cur_end - start, PAGE_SIZE);
1343                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1344
1345                 btrfs_queue_work(fs_info->delalloc_workers, &async_chunk[i].work);
1346
1347                 *nr_written += nr_pages;
1348                 start = cur_end + 1;
1349         }
1350         *page_started = 1;
1351         return 0;
1352 }
1353
1354 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1355                                         u64 bytenr, u64 num_bytes)
1356 {
1357         int ret;
1358         struct btrfs_ordered_sum *sums;
1359         LIST_HEAD(list);
1360
1361         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1362                                        bytenr + num_bytes - 1, &list, 0);
1363         if (ret == 0 && list_empty(&list))
1364                 return 0;
1365
1366         while (!list_empty(&list)) {
1367                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1368                 list_del(&sums->list);
1369                 kfree(sums);
1370         }
1371         if (ret < 0)
1372                 return ret;
1373         return 1;
1374 }
1375
1376 static int fallback_to_cow(struct inode *inode, struct page *locked_page,
1377                            const u64 start, const u64 end,
1378                            int *page_started, unsigned long *nr_written)
1379 {
1380         const bool is_space_ino = btrfs_is_free_space_inode(BTRFS_I(inode));
1381         const bool is_reloc_ino = (BTRFS_I(inode)->root->root_key.objectid ==
1382                                    BTRFS_DATA_RELOC_TREE_OBJECTID);
1383         const u64 range_bytes = end + 1 - start;
1384         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
1385         u64 range_start = start;
1386         u64 count;
1387
1388         /*
1389          * If EXTENT_NORESERVE is set it means that when the buffered write was
1390          * made we had not enough available data space and therefore we did not
1391          * reserve data space for it, since we though we could do NOCOW for the
1392          * respective file range (either there is prealloc extent or the inode
1393          * has the NOCOW bit set).
1394          *
1395          * However when we need to fallback to COW mode (because for example the
1396          * block group for the corresponding extent was turned to RO mode by a
1397          * scrub or relocation) we need to do the following:
1398          *
1399          * 1) We increment the bytes_may_use counter of the data space info.
1400          *    If COW succeeds, it allocates a new data extent and after doing
1401          *    that it decrements the space info's bytes_may_use counter and
1402          *    increments its bytes_reserved counter by the same amount (we do
1403          *    this at btrfs_add_reserved_bytes()). So we need to increment the
1404          *    bytes_may_use counter to compensate (when space is reserved at
1405          *    buffered write time, the bytes_may_use counter is incremented);
1406          *
1407          * 2) We clear the EXTENT_NORESERVE bit from the range. We do this so
1408          *    that if the COW path fails for any reason, it decrements (through
1409          *    extent_clear_unlock_delalloc()) the bytes_may_use counter of the
1410          *    data space info, which we incremented in the step above.
1411          *
1412          * If we need to fallback to cow and the inode corresponds to a free
1413          * space cache inode or an inode of the data relocation tree, we must
1414          * also increment bytes_may_use of the data space_info for the same
1415          * reason. Space caches and relocated data extents always get a prealloc
1416          * extent for them, however scrub or balance may have set the block
1417          * group that contains that extent to RO mode and therefore force COW
1418          * when starting writeback.
1419          */
1420         count = count_range_bits(io_tree, &range_start, end, range_bytes,
1421                                  EXTENT_NORESERVE, 0);
1422         if (count > 0 || is_space_ino || is_reloc_ino) {
1423                 u64 bytes = count;
1424                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
1425                 struct btrfs_space_info *sinfo = fs_info->data_sinfo;
1426
1427                 if (is_space_ino || is_reloc_ino)
1428                         bytes = range_bytes;
1429
1430                 spin_lock(&sinfo->lock);
1431                 btrfs_space_info_update_bytes_may_use(fs_info, sinfo, bytes);
1432                 spin_unlock(&sinfo->lock);
1433
1434                 if (count > 0)
1435                         clear_extent_bit(io_tree, start, end, EXTENT_NORESERVE,
1436                                          0, 0, NULL);
1437         }
1438
1439         return cow_file_range(inode, locked_page, start, end, page_started,
1440                               nr_written, 1);
1441 }
1442
1443 /*
1444  * when nowcow writeback call back.  This checks for snapshots or COW copies
1445  * of the extents that exist in the file, and COWs the file as required.
1446  *
1447  * If no cow copies or snapshots exist, we write directly to the existing
1448  * blocks on disk
1449  */
1450 static noinline int run_delalloc_nocow(struct inode *inode,
1451                                        struct page *locked_page,
1452                                        const u64 start, const u64 end,
1453                                        int *page_started, int force,
1454                                        unsigned long *nr_written)
1455 {
1456         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1457         struct btrfs_root *root = BTRFS_I(inode)->root;
1458         struct btrfs_path *path;
1459         u64 cow_start = (u64)-1;
1460         u64 cur_offset = start;
1461         int ret;
1462         bool check_prev = true;
1463         const bool freespace_inode = btrfs_is_free_space_inode(BTRFS_I(inode));
1464         u64 ino = btrfs_ino(BTRFS_I(inode));
1465         bool nocow = false;
1466         u64 disk_bytenr = 0;
1467
1468         path = btrfs_alloc_path();
1469         if (!path) {
1470                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1471                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1472                                              EXTENT_DO_ACCOUNTING |
1473                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1474                                              PAGE_CLEAR_DIRTY |
1475                                              PAGE_SET_WRITEBACK |
1476                                              PAGE_END_WRITEBACK);
1477                 return -ENOMEM;
1478         }
1479
1480         while (1) {
1481                 struct btrfs_key found_key;
1482                 struct btrfs_file_extent_item *fi;
1483                 struct extent_buffer *leaf;
1484                 u64 extent_end;
1485                 u64 extent_offset;
1486                 u64 num_bytes = 0;
1487                 u64 disk_num_bytes;
1488                 u64 ram_bytes;
1489                 int extent_type;
1490
1491                 nocow = false;
1492
1493                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1494                                                cur_offset, 0);
1495                 if (ret < 0)
1496                         goto error;
1497
1498                 /*
1499                  * If there is no extent for our range when doing the initial
1500                  * search, then go back to the previous slot as it will be the
1501                  * one containing the search offset
1502                  */
1503                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1504                         leaf = path->nodes[0];
1505                         btrfs_item_key_to_cpu(leaf, &found_key,
1506                                               path->slots[0] - 1);
1507                         if (found_key.objectid == ino &&
1508                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1509                                 path->slots[0]--;
1510                 }
1511                 check_prev = false;
1512 next_slot:
1513                 /* Go to next leaf if we have exhausted the current one */
1514                 leaf = path->nodes[0];
1515                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1516                         ret = btrfs_next_leaf(root, path);
1517                         if (ret < 0) {
1518                                 if (cow_start != (u64)-1)
1519                                         cur_offset = cow_start;
1520                                 goto error;
1521                         }
1522                         if (ret > 0)
1523                                 break;
1524                         leaf = path->nodes[0];
1525                 }
1526
1527                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1528
1529                 /* Didn't find anything for our INO */
1530                 if (found_key.objectid > ino)
1531                         break;
1532                 /*
1533                  * Keep searching until we find an EXTENT_ITEM or there are no
1534                  * more extents for this inode
1535                  */
1536                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1537                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1538                         path->slots[0]++;
1539                         goto next_slot;
1540                 }
1541
1542                 /* Found key is not EXTENT_DATA_KEY or starts after req range */
1543                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1544                     found_key.offset > end)
1545                         break;
1546
1547                 /*
1548                  * If the found extent starts after requested offset, then
1549                  * adjust extent_end to be right before this extent begins
1550                  */
1551                 if (found_key.offset > cur_offset) {
1552                         extent_end = found_key.offset;
1553                         extent_type = 0;
1554                         goto out_check;
1555                 }
1556
1557                 /*
1558                  * Found extent which begins before our range and potentially
1559                  * intersect it
1560                  */
1561                 fi = btrfs_item_ptr(leaf, path->slots[0],
1562                                     struct btrfs_file_extent_item);
1563                 extent_type = btrfs_file_extent_type(leaf, fi);
1564
1565                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1566                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1567                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1568                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1569                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1570                         extent_end = found_key.offset +
1571                                 btrfs_file_extent_num_bytes(leaf, fi);
1572                         disk_num_bytes =
1573                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1574                         /*
1575                          * If the extent we got ends before our current offset,
1576                          * skip to the next extent.
1577                          */
1578                         if (extent_end <= cur_offset) {
1579                                 path->slots[0]++;
1580                                 goto next_slot;
1581                         }
1582                         /* Skip holes */
1583                         if (disk_bytenr == 0)
1584                                 goto out_check;
1585                         /* Skip compressed/encrypted/encoded extents */
1586                         if (btrfs_file_extent_compression(leaf, fi) ||
1587                             btrfs_file_extent_encryption(leaf, fi) ||
1588                             btrfs_file_extent_other_encoding(leaf, fi))
1589                                 goto out_check;
1590                         /*
1591                          * If extent is created before the last volume's snapshot
1592                          * this implies the extent is shared, hence we can't do
1593                          * nocow. This is the same check as in
1594                          * btrfs_cross_ref_exist but without calling
1595                          * btrfs_search_slot.
1596                          */
1597                         if (!freespace_inode &&
1598                             btrfs_file_extent_generation(leaf, fi) <=
1599                             btrfs_root_last_snapshot(&root->root_item))
1600                                 goto out_check;
1601                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1602                                 goto out_check;
1603                         /* If extent is RO, we must COW it */
1604                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1605                                 goto out_check;
1606                         ret = btrfs_cross_ref_exist(root, ino,
1607                                                     found_key.offset -
1608                                                     extent_offset, disk_bytenr);
1609                         if (ret) {
1610                                 /*
1611                                  * ret could be -EIO if the above fails to read
1612                                  * metadata.
1613                                  */
1614                                 if (ret < 0) {
1615                                         if (cow_start != (u64)-1)
1616                                                 cur_offset = cow_start;
1617                                         goto error;
1618                                 }
1619
1620                                 WARN_ON_ONCE(freespace_inode);
1621                                 goto out_check;
1622                         }
1623                         disk_bytenr += extent_offset;
1624                         disk_bytenr += cur_offset - found_key.offset;
1625                         num_bytes = min(end + 1, extent_end) - cur_offset;
1626                         /*
1627                          * If there are pending snapshots for this root, we
1628                          * fall into common COW way
1629                          */
1630                         if (!freespace_inode && atomic_read(&root->snapshot_force_cow))
1631                                 goto out_check;
1632                         /*
1633                          * force cow if csum exists in the range.
1634                          * this ensure that csum for a given extent are
1635                          * either valid or do not exist.
1636                          */
1637                         ret = csum_exist_in_range(fs_info, disk_bytenr,
1638                                                   num_bytes);
1639                         if (ret) {
1640                                 /*
1641                                  * ret could be -EIO if the above fails to read
1642                                  * metadata.
1643                                  */
1644                                 if (ret < 0) {
1645                                         if (cow_start != (u64)-1)
1646                                                 cur_offset = cow_start;
1647                                         goto error;
1648                                 }
1649                                 WARN_ON_ONCE(freespace_inode);
1650                                 goto out_check;
1651                         }
1652                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr))
1653                                 goto out_check;
1654                         nocow = true;
1655                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1656                         extent_end = found_key.offset + ram_bytes;
1657                         extent_end = ALIGN(extent_end, fs_info->sectorsize);
1658                         /* Skip extents outside of our requested range */
1659                         if (extent_end <= start) {
1660                                 path->slots[0]++;
1661                                 goto next_slot;
1662                         }
1663                 } else {
1664                         /* If this triggers then we have a memory corruption */
1665                         BUG();
1666                 }
1667 out_check:
1668                 /*
1669                  * If nocow is false then record the beginning of the range
1670                  * that needs to be COWed
1671                  */
1672                 if (!nocow) {
1673                         if (cow_start == (u64)-1)
1674                                 cow_start = cur_offset;
1675                         cur_offset = extent_end;
1676                         if (cur_offset > end)
1677                                 break;
1678                         path->slots[0]++;
1679                         goto next_slot;
1680                 }
1681
1682                 btrfs_release_path(path);
1683
1684                 /*
1685                  * COW range from cow_start to found_key.offset - 1. As the key
1686                  * will contain the beginning of the first extent that can be
1687                  * NOCOW, following one which needs to be COW'ed
1688                  */
1689                 if (cow_start != (u64)-1) {
1690                         ret = fallback_to_cow(inode, locked_page, cow_start,
1691                                               found_key.offset - 1,
1692                                               page_started, nr_written);
1693                         if (ret)
1694                                 goto error;
1695                         cow_start = (u64)-1;
1696                 }
1697
1698                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1699                         u64 orig_start = found_key.offset - extent_offset;
1700                         struct extent_map *em;
1701
1702                         em = create_io_em(inode, cur_offset, num_bytes,
1703                                           orig_start,
1704                                           disk_bytenr, /* block_start */
1705                                           num_bytes, /* block_len */
1706                                           disk_num_bytes, /* orig_block_len */
1707                                           ram_bytes, BTRFS_COMPRESS_NONE,
1708                                           BTRFS_ORDERED_PREALLOC);
1709                         if (IS_ERR(em)) {
1710                                 ret = PTR_ERR(em);
1711                                 goto error;
1712                         }
1713                         free_extent_map(em);
1714                         ret = btrfs_add_ordered_extent(inode, cur_offset,
1715                                                        disk_bytenr, num_bytes,
1716                                                        num_bytes,
1717                                                        BTRFS_ORDERED_PREALLOC);
1718                         if (ret) {
1719                                 btrfs_drop_extent_cache(BTRFS_I(inode),
1720                                                         cur_offset,
1721                                                         cur_offset + num_bytes - 1,
1722                                                         0);
1723                                 goto error;
1724                         }
1725                 } else {
1726                         ret = btrfs_add_ordered_extent(inode, cur_offset,
1727                                                        disk_bytenr, num_bytes,
1728                                                        num_bytes,
1729                                                        BTRFS_ORDERED_NOCOW);
1730                         if (ret)
1731                                 goto error;
1732                 }
1733
1734                 if (nocow)
1735                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1736                 nocow = false;
1737
1738                 if (root->root_key.objectid ==
1739                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1740                         /*
1741                          * Error handled later, as we must prevent
1742                          * extent_clear_unlock_delalloc() in error handler
1743                          * from freeing metadata of created ordered extent.
1744                          */
1745                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1746                                                       num_bytes);
1747
1748                 extent_clear_unlock_delalloc(inode, cur_offset,
1749                                              cur_offset + num_bytes - 1,
1750                                              locked_page, EXTENT_LOCKED |
1751                                              EXTENT_DELALLOC |
1752                                              EXTENT_CLEAR_DATA_RESV,
1753                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1754
1755                 cur_offset = extent_end;
1756
1757                 /*
1758                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1759                  * handler, as metadata for created ordered extent will only
1760                  * be freed by btrfs_finish_ordered_io().
1761                  */
1762                 if (ret)
1763                         goto error;
1764                 if (cur_offset > end)
1765                         break;
1766         }
1767         btrfs_release_path(path);
1768
1769         if (cur_offset <= end && cow_start == (u64)-1)
1770                 cow_start = cur_offset;
1771
1772         if (cow_start != (u64)-1) {
1773                 cur_offset = end;
1774                 ret = fallback_to_cow(inode, locked_page, cow_start, end,
1775                                       page_started, nr_written);
1776                 if (ret)
1777                         goto error;
1778         }
1779
1780 error:
1781         if (nocow)
1782                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1783
1784         if (ret && cur_offset < end)
1785                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1786                                              locked_page, EXTENT_LOCKED |
1787                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1788                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1789                                              PAGE_CLEAR_DIRTY |
1790                                              PAGE_SET_WRITEBACK |
1791                                              PAGE_END_WRITEBACK);
1792         btrfs_free_path(path);
1793         return ret;
1794 }
1795
1796 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1797 {
1798
1799         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1800             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1801                 return 0;
1802
1803         /*
1804          * @defrag_bytes is a hint value, no spinlock held here,
1805          * if is not zero, it means the file is defragging.
1806          * Force cow if given extent needs to be defragged.
1807          */
1808         if (BTRFS_I(inode)->defrag_bytes &&
1809             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1810                            EXTENT_DEFRAG, 0, NULL))
1811                 return 1;
1812
1813         return 0;
1814 }
1815
1816 /*
1817  * Function to process delayed allocation (create CoW) for ranges which are
1818  * being touched for the first time.
1819  */
1820 int btrfs_run_delalloc_range(struct inode *inode, struct page *locked_page,
1821                 u64 start, u64 end, int *page_started, unsigned long *nr_written,
1822                 struct writeback_control *wbc)
1823 {
1824         int ret;
1825         int force_cow = need_force_cow(inode, start, end);
1826
1827         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1828                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1829                                          page_started, 1, nr_written);
1830         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1831                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1832                                          page_started, 0, nr_written);
1833         } else if (!inode_can_compress(inode) ||
1834                    !inode_need_compress(inode, start, end)) {
1835                 ret = cow_file_range(inode, locked_page, start, end,
1836                                       page_started, nr_written, 1);
1837         } else {
1838                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1839                         &BTRFS_I(inode)->runtime_flags);
1840                 ret = cow_file_range_async(inode, wbc, locked_page, start, end,
1841                                            page_started, nr_written);
1842         }
1843         if (ret)
1844                 btrfs_cleanup_ordered_extents(inode, locked_page, start,
1845                                               end - start + 1);
1846         return ret;
1847 }
1848
1849 void btrfs_split_delalloc_extent(struct inode *inode,
1850                                  struct extent_state *orig, u64 split)
1851 {
1852         u64 size;
1853
1854         /* not delalloc, ignore it */
1855         if (!(orig->state & EXTENT_DELALLOC))
1856                 return;
1857
1858         size = orig->end - orig->start + 1;
1859         if (size > BTRFS_MAX_EXTENT_SIZE) {
1860                 u32 num_extents;
1861                 u64 new_size;
1862
1863                 /*
1864                  * See the explanation in btrfs_merge_delalloc_extent, the same
1865                  * applies here, just in reverse.
1866                  */
1867                 new_size = orig->end - split + 1;
1868                 num_extents = count_max_extents(new_size);
1869                 new_size = split - orig->start;
1870                 num_extents += count_max_extents(new_size);
1871                 if (count_max_extents(size) >= num_extents)
1872                         return;
1873         }
1874
1875         spin_lock(&BTRFS_I(inode)->lock);
1876         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1877         spin_unlock(&BTRFS_I(inode)->lock);
1878 }
1879
1880 /*
1881  * Handle merged delayed allocation extents so we can keep track of new extents
1882  * that are just merged onto old extents, such as when we are doing sequential
1883  * writes, so we can properly account for the metadata space we'll need.
1884  */
1885 void btrfs_merge_delalloc_extent(struct inode *inode, struct extent_state *new,
1886                                  struct extent_state *other)
1887 {
1888         u64 new_size, old_size;
1889         u32 num_extents;
1890
1891         /* not delalloc, ignore it */
1892         if (!(other->state & EXTENT_DELALLOC))
1893                 return;
1894
1895         if (new->start > other->start)
1896                 new_size = new->end - other->start + 1;
1897         else
1898                 new_size = other->end - new->start + 1;
1899
1900         /* we're not bigger than the max, unreserve the space and go */
1901         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1902                 spin_lock(&BTRFS_I(inode)->lock);
1903                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1904                 spin_unlock(&BTRFS_I(inode)->lock);
1905                 return;
1906         }
1907
1908         /*
1909          * We have to add up either side to figure out how many extents were
1910          * accounted for before we merged into one big extent.  If the number of
1911          * extents we accounted for is <= the amount we need for the new range
1912          * then we can return, otherwise drop.  Think of it like this
1913          *
1914          * [ 4k][MAX_SIZE]
1915          *
1916          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1917          * need 2 outstanding extents, on one side we have 1 and the other side
1918          * we have 1 so they are == and we can return.  But in this case
1919          *
1920          * [MAX_SIZE+4k][MAX_SIZE+4k]
1921          *
1922          * Each range on their own accounts for 2 extents, but merged together
1923          * they are only 3 extents worth of accounting, so we need to drop in
1924          * this case.
1925          */
1926         old_size = other->end - other->start + 1;
1927         num_extents = count_max_extents(old_size);
1928         old_size = new->end - new->start + 1;
1929         num_extents += count_max_extents(old_size);
1930         if (count_max_extents(new_size) >= num_extents)
1931                 return;
1932
1933         spin_lock(&BTRFS_I(inode)->lock);
1934         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1935         spin_unlock(&BTRFS_I(inode)->lock);
1936 }
1937
1938 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1939                                       struct inode *inode)
1940 {
1941         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1942
1943         spin_lock(&root->delalloc_lock);
1944         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1945                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1946                               &root->delalloc_inodes);
1947                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1948                         &BTRFS_I(inode)->runtime_flags);
1949                 root->nr_delalloc_inodes++;
1950                 if (root->nr_delalloc_inodes == 1) {
1951                         spin_lock(&fs_info->delalloc_root_lock);
1952                         BUG_ON(!list_empty(&root->delalloc_root));
1953                         list_add_tail(&root->delalloc_root,
1954                                       &fs_info->delalloc_roots);
1955                         spin_unlock(&fs_info->delalloc_root_lock);
1956                 }
1957         }
1958         spin_unlock(&root->delalloc_lock);
1959 }
1960
1961
1962 void __btrfs_del_delalloc_inode(struct btrfs_root *root,
1963                                 struct btrfs_inode *inode)
1964 {
1965         struct btrfs_fs_info *fs_info = root->fs_info;
1966
1967         if (!list_empty(&inode->delalloc_inodes)) {
1968                 list_del_init(&inode->delalloc_inodes);
1969                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1970                           &inode->runtime_flags);
1971                 root->nr_delalloc_inodes--;
1972                 if (!root->nr_delalloc_inodes) {
1973                         ASSERT(list_empty(&root->delalloc_inodes));
1974                         spin_lock(&fs_info->delalloc_root_lock);
1975                         BUG_ON(list_empty(&root->delalloc_root));
1976                         list_del_init(&root->delalloc_root);
1977                         spin_unlock(&fs_info->delalloc_root_lock);
1978                 }
1979         }
1980 }
1981
1982 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1983                                      struct btrfs_inode *inode)
1984 {
1985         spin_lock(&root->delalloc_lock);
1986         __btrfs_del_delalloc_inode(root, inode);
1987         spin_unlock(&root->delalloc_lock);
1988 }
1989
1990 /*
1991  * Properly track delayed allocation bytes in the inode and to maintain the
1992  * list of inodes that have pending delalloc work to be done.
1993  */
1994 void btrfs_set_delalloc_extent(struct inode *inode, struct extent_state *state,
1995                                unsigned *bits)
1996 {
1997         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1998
1999         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
2000                 WARN_ON(1);
2001         /*
2002          * set_bit and clear bit hooks normally require _irqsave/restore
2003          * but in this case, we are only testing for the DELALLOC
2004          * bit, which is only set or cleared with irqs on
2005          */
2006         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
2007                 struct btrfs_root *root = BTRFS_I(inode)->root;
2008                 u64 len = state->end + 1 - state->start;
2009                 u32 num_extents = count_max_extents(len);
2010                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
2011
2012                 spin_lock(&BTRFS_I(inode)->lock);
2013                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
2014                 spin_unlock(&BTRFS_I(inode)->lock);
2015
2016                 /* For sanity tests */
2017                 if (btrfs_is_testing(fs_info))
2018                         return;
2019
2020                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
2021                                          fs_info->delalloc_batch);
2022                 spin_lock(&BTRFS_I(inode)->lock);
2023                 BTRFS_I(inode)->delalloc_bytes += len;
2024                 if (*bits & EXTENT_DEFRAG)
2025                         BTRFS_I(inode)->defrag_bytes += len;
2026                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
2027                                          &BTRFS_I(inode)->runtime_flags))
2028                         btrfs_add_delalloc_inodes(root, inode);
2029                 spin_unlock(&BTRFS_I(inode)->lock);
2030         }
2031
2032         if (!(state->state & EXTENT_DELALLOC_NEW) &&
2033             (*bits & EXTENT_DELALLOC_NEW)) {
2034                 spin_lock(&BTRFS_I(inode)->lock);
2035                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
2036                         state->start;
2037                 spin_unlock(&BTRFS_I(inode)->lock);
2038         }
2039 }
2040
2041 /*
2042  * Once a range is no longer delalloc this function ensures that proper
2043  * accounting happens.
2044  */
2045 void btrfs_clear_delalloc_extent(struct inode *vfs_inode,
2046                                  struct extent_state *state, unsigned *bits)
2047 {
2048         struct btrfs_inode *inode = BTRFS_I(vfs_inode);
2049         struct btrfs_fs_info *fs_info = btrfs_sb(vfs_inode->i_sb);
2050         u64 len = state->end + 1 - state->start;
2051         u32 num_extents = count_max_extents(len);
2052
2053         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
2054                 spin_lock(&inode->lock);
2055                 inode->defrag_bytes -= len;
2056                 spin_unlock(&inode->lock);
2057         }
2058
2059         /*
2060          * set_bit and clear bit hooks normally require _irqsave/restore
2061          * but in this case, we are only testing for the DELALLOC
2062          * bit, which is only set or cleared with irqs on
2063          */
2064         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
2065                 struct btrfs_root *root = inode->root;
2066                 bool do_list = !btrfs_is_free_space_inode(inode);
2067
2068                 spin_lock(&inode->lock);
2069                 btrfs_mod_outstanding_extents(inode, -num_extents);
2070                 spin_unlock(&inode->lock);
2071
2072                 /*
2073                  * We don't reserve metadata space for space cache inodes so we
2074                  * don't need to call delalloc_release_metadata if there is an
2075                  * error.
2076                  */
2077                 if (*bits & EXTENT_CLEAR_META_RESV &&
2078                     root != fs_info->tree_root)
2079                         btrfs_delalloc_release_metadata(inode, len, false);
2080
2081                 /* For sanity tests. */
2082                 if (btrfs_is_testing(fs_info))
2083                         return;
2084
2085                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
2086                     do_list && !(state->state & EXTENT_NORESERVE) &&
2087                     (*bits & EXTENT_CLEAR_DATA_RESV))
2088                         btrfs_free_reserved_data_space_noquota(
2089                                         &inode->vfs_inode,
2090                                         state->start, len);
2091
2092                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
2093                                          fs_info->delalloc_batch);
2094                 spin_lock(&inode->lock);
2095                 inode->delalloc_bytes -= len;
2096                 if (do_list && inode->delalloc_bytes == 0 &&
2097                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
2098                                         &inode->runtime_flags))
2099                         btrfs_del_delalloc_inode(root, inode);
2100                 spin_unlock(&inode->lock);
2101         }
2102
2103         if ((state->state & EXTENT_DELALLOC_NEW) &&
2104             (*bits & EXTENT_DELALLOC_NEW)) {
2105                 spin_lock(&inode->lock);
2106                 ASSERT(inode->new_delalloc_bytes >= len);
2107                 inode->new_delalloc_bytes -= len;
2108                 spin_unlock(&inode->lock);
2109         }
2110 }
2111
2112 /*
2113  * btrfs_bio_fits_in_stripe - Checks whether the size of the given bio will fit
2114  * in a chunk's stripe. This function ensures that bios do not span a
2115  * stripe/chunk
2116  *
2117  * @page - The page we are about to add to the bio
2118  * @size - size we want to add to the bio
2119  * @bio - bio we want to ensure is smaller than a stripe
2120  * @bio_flags - flags of the bio
2121  *
2122  * return 1 if page cannot be added to the bio
2123  * return 0 if page can be added to the bio
2124  * return error otherwise
2125  */
2126 int btrfs_bio_fits_in_stripe(struct page *page, size_t size, struct bio *bio,
2127                              unsigned long bio_flags)
2128 {
2129         struct inode *inode = page->mapping->host;
2130         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2131         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
2132         u64 length = 0;
2133         u64 map_length;
2134         int ret;
2135         struct btrfs_io_geometry geom;
2136
2137         if (bio_flags & EXTENT_BIO_COMPRESSED)
2138                 return 0;
2139
2140         length = bio->bi_iter.bi_size;
2141         map_length = length;
2142         ret = btrfs_get_io_geometry(fs_info, btrfs_op(bio), logical, map_length,
2143                                     &geom);
2144         if (ret < 0)
2145                 return ret;
2146
2147         if (geom.len < length + size)
2148                 return 1;
2149         return 0;
2150 }
2151
2152 /*
2153  * in order to insert checksums into the metadata in large chunks,
2154  * we wait until bio submission time.   All the pages in the bio are
2155  * checksummed and sums are attached onto the ordered extent record.
2156  *
2157  * At IO completion time the cums attached on the ordered extent record
2158  * are inserted into the btree
2159  */
2160 static blk_status_t btrfs_submit_bio_start(void *private_data, struct bio *bio,
2161                                     u64 bio_offset)
2162 {
2163         struct inode *inode = private_data;
2164         blk_status_t ret = 0;
2165
2166         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2167         BUG_ON(ret); /* -ENOMEM */
2168         return 0;
2169 }
2170
2171 /*
2172  * extent_io.c submission hook. This does the right thing for csum calculation
2173  * on write, or reading the csums from the tree before a read.
2174  *
2175  * Rules about async/sync submit,
2176  * a) read:                             sync submit
2177  *
2178  * b) write without checksum:           sync submit
2179  *
2180  * c) write with checksum:
2181  *    c-1) if bio is issued by fsync:   sync submit
2182  *         (sync_writers != 0)
2183  *
2184  *    c-2) if root is reloc root:       sync submit
2185  *         (only in case of buffered IO)
2186  *
2187  *    c-3) otherwise:                   async submit
2188  */
2189 static blk_status_t btrfs_submit_bio_hook(struct inode *inode, struct bio *bio,
2190                                           int mirror_num,
2191                                           unsigned long bio_flags)
2192
2193 {
2194         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2195         struct btrfs_root *root = BTRFS_I(inode)->root;
2196         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
2197         blk_status_t ret = 0;
2198         int skip_sum;
2199         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
2200
2201         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
2202
2203         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
2204                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
2205
2206         if (bio_op(bio) != REQ_OP_WRITE) {
2207                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
2208                 if (ret)
2209                         goto out;
2210
2211                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
2212                         ret = btrfs_submit_compressed_read(inode, bio,
2213                                                            mirror_num,
2214                                                            bio_flags);
2215                         goto out;
2216                 } else if (!skip_sum) {
2217                         ret = btrfs_lookup_bio_sums(inode, bio, (u64)-1, NULL);
2218                         if (ret)
2219                                 goto out;
2220                 }
2221                 goto mapit;
2222         } else if (async && !skip_sum) {
2223                 /* csum items have already been cloned */
2224                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2225                         goto mapit;
2226                 /* we're doing a write, do the async checksumming */
2227                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
2228                                           0, inode, btrfs_submit_bio_start);
2229                 goto out;
2230         } else if (!skip_sum) {
2231                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2232                 if (ret)
2233                         goto out;
2234         }
2235
2236 mapit:
2237         ret = btrfs_map_bio(fs_info, bio, mirror_num);
2238
2239 out:
2240         if (ret) {
2241                 bio->bi_status = ret;
2242                 bio_endio(bio);
2243         }
2244         return ret;
2245 }
2246
2247 /*
2248  * given a list of ordered sums record them in the inode.  This happens
2249  * at IO completion time based on sums calculated at bio submission time.
2250  */
2251 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2252                              struct inode *inode, struct list_head *list)
2253 {
2254         struct btrfs_ordered_sum *sum;
2255         int ret;
2256
2257         list_for_each_entry(sum, list, list) {
2258                 trans->adding_csums = true;
2259                 ret = btrfs_csum_file_blocks(trans,
2260                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2261                 trans->adding_csums = false;
2262                 if (ret)
2263                         return ret;
2264         }
2265         return 0;
2266 }
2267
2268 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2269                               unsigned int extra_bits,
2270                               struct extent_state **cached_state)
2271 {
2272         WARN_ON(PAGE_ALIGNED(end));
2273         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2274                                    extra_bits, cached_state);
2275 }
2276
2277 /* see btrfs_writepage_start_hook for details on why this is required */
2278 struct btrfs_writepage_fixup {
2279         struct page *page;
2280         struct inode *inode;
2281         struct btrfs_work work;
2282 };
2283
2284 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2285 {
2286         struct btrfs_writepage_fixup *fixup;
2287         struct btrfs_ordered_extent *ordered;
2288         struct extent_state *cached_state = NULL;
2289         struct extent_changeset *data_reserved = NULL;
2290         struct page *page;
2291         struct inode *inode;
2292         u64 page_start;
2293         u64 page_end;
2294         int ret = 0;
2295         bool free_delalloc_space = true;
2296
2297         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2298         page = fixup->page;
2299         inode = fixup->inode;
2300         page_start = page_offset(page);
2301         page_end = page_offset(page) + PAGE_SIZE - 1;
2302
2303         /*
2304          * This is similar to page_mkwrite, we need to reserve the space before
2305          * we take the page lock.
2306          */
2307         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2308                                            PAGE_SIZE);
2309 again:
2310         lock_page(page);
2311
2312         /*
2313          * Before we queued this fixup, we took a reference on the page.
2314          * page->mapping may go NULL, but it shouldn't be moved to a different
2315          * address space.
2316          */
2317         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2318                 /*
2319                  * Unfortunately this is a little tricky, either
2320                  *
2321                  * 1) We got here and our page had already been dealt with and
2322                  *    we reserved our space, thus ret == 0, so we need to just
2323                  *    drop our space reservation and bail.  This can happen the
2324                  *    first time we come into the fixup worker, or could happen
2325                  *    while waiting for the ordered extent.
2326                  * 2) Our page was already dealt with, but we happened to get an
2327                  *    ENOSPC above from the btrfs_delalloc_reserve_space.  In
2328                  *    this case we obviously don't have anything to release, but
2329                  *    because the page was already dealt with we don't want to
2330                  *    mark the page with an error, so make sure we're resetting
2331                  *    ret to 0.  This is why we have this check _before_ the ret
2332                  *    check, because we do not want to have a surprise ENOSPC
2333                  *    when the page was already properly dealt with.
2334                  */
2335                 if (!ret) {
2336                         btrfs_delalloc_release_extents(BTRFS_I(inode),
2337                                                        PAGE_SIZE);
2338                         btrfs_delalloc_release_space(inode, data_reserved,
2339                                                      page_start, PAGE_SIZE,
2340                                                      true);
2341                 }
2342                 ret = 0;
2343                 goto out_page;
2344         }
2345
2346         /*
2347          * We can't mess with the page state unless it is locked, so now that
2348          * it is locked bail if we failed to make our space reservation.
2349          */
2350         if (ret)
2351                 goto out_page;
2352
2353         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2354                          &cached_state);
2355
2356         /* already ordered? We're done */
2357         if (PagePrivate2(page))
2358                 goto out_reserved;
2359
2360         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2361                                         PAGE_SIZE);
2362         if (ordered) {
2363                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2364                                      page_end, &cached_state);
2365                 unlock_page(page);
2366                 btrfs_start_ordered_extent(inode, ordered, 1);
2367                 btrfs_put_ordered_extent(ordered);
2368                 goto again;
2369         }
2370
2371         ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
2372                                         &cached_state);
2373         if (ret)
2374                 goto out_reserved;
2375
2376         /*
2377          * Everything went as planned, we're now the owner of a dirty page with
2378          * delayed allocation bits set and space reserved for our COW
2379          * destination.
2380          *
2381          * The page was dirty when we started, nothing should have cleaned it.
2382          */
2383         BUG_ON(!PageDirty(page));
2384         free_delalloc_space = false;
2385 out_reserved:
2386         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
2387         if (free_delalloc_space)
2388                 btrfs_delalloc_release_space(inode, data_reserved, page_start,
2389                                              PAGE_SIZE, true);
2390         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2391                              &cached_state);
2392 out_page:
2393         if (ret) {
2394                 /*
2395                  * We hit ENOSPC or other errors.  Update the mapping and page
2396                  * to reflect the errors and clean the page.
2397                  */
2398                 mapping_set_error(page->mapping, ret);
2399                 end_extent_writepage(page, ret, page_start, page_end);
2400                 clear_page_dirty_for_io(page);
2401                 SetPageError(page);
2402         }
2403         ClearPageChecked(page);
2404         unlock_page(page);
2405         put_page(page);
2406         kfree(fixup);
2407         extent_changeset_free(data_reserved);
2408         /*
2409          * As a precaution, do a delayed iput in case it would be the last iput
2410          * that could need flushing space. Recursing back to fixup worker would
2411          * deadlock.
2412          */
2413         btrfs_add_delayed_iput(inode);
2414 }
2415
2416 /*
2417  * There are a few paths in the higher layers of the kernel that directly
2418  * set the page dirty bit without asking the filesystem if it is a
2419  * good idea.  This causes problems because we want to make sure COW
2420  * properly happens and the data=ordered rules are followed.
2421  *
2422  * In our case any range that doesn't have the ORDERED bit set
2423  * hasn't been properly setup for IO.  We kick off an async process
2424  * to fix it up.  The async helper will wait for ordered extents, set
2425  * the delalloc bit and make it safe to write the page.
2426  */
2427 int btrfs_writepage_cow_fixup(struct page *page, u64 start, u64 end)
2428 {
2429         struct inode *inode = page->mapping->host;
2430         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2431         struct btrfs_writepage_fixup *fixup;
2432
2433         /* this page is properly in the ordered list */
2434         if (TestClearPagePrivate2(page))
2435                 return 0;
2436
2437         /*
2438          * PageChecked is set below when we create a fixup worker for this page,
2439          * don't try to create another one if we're already PageChecked()
2440          *
2441          * The extent_io writepage code will redirty the page if we send back
2442          * EAGAIN.
2443          */
2444         if (PageChecked(page))
2445                 return -EAGAIN;
2446
2447         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2448         if (!fixup)
2449                 return -EAGAIN;
2450
2451         /*
2452          * We are already holding a reference to this inode from
2453          * write_cache_pages.  We need to hold it because the space reservation
2454          * takes place outside of the page lock, and we can't trust
2455          * page->mapping outside of the page lock.
2456          */
2457         ihold(inode);
2458         SetPageChecked(page);
2459         get_page(page);
2460         btrfs_init_work(&fixup->work, btrfs_writepage_fixup_worker, NULL, NULL);
2461         fixup->page = page;
2462         fixup->inode = inode;
2463         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2464
2465         return -EAGAIN;
2466 }
2467
2468 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2469                                        struct inode *inode, u64 file_pos,
2470                                        u64 disk_bytenr, u64 disk_num_bytes,
2471                                        u64 num_bytes, u64 ram_bytes,
2472                                        u8 compression, u8 encryption,
2473                                        u16 other_encoding, int extent_type)
2474 {
2475         struct btrfs_root *root = BTRFS_I(inode)->root;
2476         struct btrfs_file_extent_item *fi;
2477         struct btrfs_path *path;
2478         struct extent_buffer *leaf;
2479         struct btrfs_key ins;
2480         u64 qg_released;
2481         int extent_inserted = 0;
2482         int ret;
2483
2484         path = btrfs_alloc_path();
2485         if (!path)
2486                 return -ENOMEM;
2487
2488         /*
2489          * we may be replacing one extent in the tree with another.
2490          * The new extent is pinned in the extent map, and we don't want
2491          * to drop it from the cache until it is completely in the btree.
2492          *
2493          * So, tell btrfs_drop_extents to leave this extent in the cache.
2494          * the caller is expected to unpin it and allow it to be merged
2495          * with the others.
2496          */
2497         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2498                                    file_pos + num_bytes, NULL, 0,
2499                                    1, sizeof(*fi), &extent_inserted);
2500         if (ret)
2501                 goto out;
2502
2503         if (!extent_inserted) {
2504                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2505                 ins.offset = file_pos;
2506                 ins.type = BTRFS_EXTENT_DATA_KEY;
2507
2508                 path->leave_spinning = 1;
2509                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2510                                               sizeof(*fi));
2511                 if (ret)
2512                         goto out;
2513         }
2514         leaf = path->nodes[0];
2515         fi = btrfs_item_ptr(leaf, path->slots[0],
2516                             struct btrfs_file_extent_item);
2517         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2518         btrfs_set_file_extent_type(leaf, fi, extent_type);
2519         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2520         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2521         btrfs_set_file_extent_offset(leaf, fi, 0);
2522         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2523         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2524         btrfs_set_file_extent_compression(leaf, fi, compression);
2525         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2526         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2527
2528         btrfs_mark_buffer_dirty(leaf);
2529         btrfs_release_path(path);
2530
2531         inode_add_bytes(inode, num_bytes);
2532
2533         ins.objectid = disk_bytenr;
2534         ins.offset = disk_num_bytes;
2535         ins.type = BTRFS_EXTENT_ITEM_KEY;
2536
2537         ret = btrfs_inode_set_file_extent_range(BTRFS_I(inode), file_pos,
2538                                                 ram_bytes);
2539         if (ret)
2540                 goto out;
2541
2542         /*
2543          * Release the reserved range from inode dirty range map, as it is
2544          * already moved into delayed_ref_head
2545          */
2546         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2547         if (ret < 0)
2548                 goto out;
2549         qg_released = ret;
2550         ret = btrfs_alloc_reserved_file_extent(trans, root,
2551                                                btrfs_ino(BTRFS_I(inode)),
2552                                                file_pos, qg_released, &ins);
2553 out:
2554         btrfs_free_path(path);
2555
2556         return ret;
2557 }
2558
2559 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2560                                          u64 start, u64 len)
2561 {
2562         struct btrfs_block_group *cache;
2563
2564         cache = btrfs_lookup_block_group(fs_info, start);
2565         ASSERT(cache);
2566
2567         spin_lock(&cache->lock);
2568         cache->delalloc_bytes -= len;
2569         spin_unlock(&cache->lock);
2570
2571         btrfs_put_block_group(cache);
2572 }
2573
2574 /* as ordered data IO finishes, this gets called so we can finish
2575  * an ordered extent if the range of bytes in the file it covers are
2576  * fully written.
2577  */
2578 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2579 {
2580         struct inode *inode = ordered_extent->inode;
2581         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2582         struct btrfs_root *root = BTRFS_I(inode)->root;
2583         struct btrfs_trans_handle *trans = NULL;
2584         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2585         struct extent_state *cached_state = NULL;
2586         u64 start, end;
2587         int compress_type = 0;
2588         int ret = 0;
2589         u64 logical_len = ordered_extent->num_bytes;
2590         bool freespace_inode;
2591         bool truncated = false;
2592         bool range_locked = false;
2593         bool clear_new_delalloc_bytes = false;
2594         bool clear_reserved_extent = true;
2595         unsigned int clear_bits;
2596
2597         start = ordered_extent->file_offset;
2598         end = start + ordered_extent->num_bytes - 1;
2599
2600         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2601             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
2602             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
2603                 clear_new_delalloc_bytes = true;
2604
2605         freespace_inode = btrfs_is_free_space_inode(BTRFS_I(inode));
2606
2607         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2608                 ret = -EIO;
2609                 goto out;
2610         }
2611
2612         btrfs_free_io_failure_record(BTRFS_I(inode), start, end);
2613
2614         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2615                 truncated = true;
2616                 logical_len = ordered_extent->truncated_len;
2617                 /* Truncated the entire extent, don't bother adding */
2618                 if (!logical_len)
2619                         goto out;
2620         }
2621
2622         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2623                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2624
2625                 /*
2626                  * For mwrite(mmap + memset to write) case, we still reserve
2627                  * space for NOCOW range.
2628                  * As NOCOW won't cause a new delayed ref, just free the space
2629                  */
2630                 btrfs_qgroup_free_data(inode, NULL, start,
2631                                        ordered_extent->num_bytes);
2632                 btrfs_inode_safe_disk_i_size_write(inode, 0);
2633                 if (freespace_inode)
2634                         trans = btrfs_join_transaction_spacecache(root);
2635                 else
2636                         trans = btrfs_join_transaction(root);
2637                 if (IS_ERR(trans)) {
2638                         ret = PTR_ERR(trans);
2639                         trans = NULL;
2640                         goto out;
2641                 }
2642                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
2643                 ret = btrfs_update_inode_fallback(trans, root, inode);
2644                 if (ret) /* -ENOMEM or corruption */
2645                         btrfs_abort_transaction(trans, ret);
2646                 goto out;
2647         }
2648
2649         range_locked = true;
2650         lock_extent_bits(io_tree, start, end, &cached_state);
2651
2652         if (freespace_inode)
2653                 trans = btrfs_join_transaction_spacecache(root);
2654         else
2655                 trans = btrfs_join_transaction(root);
2656         if (IS_ERR(trans)) {
2657                 ret = PTR_ERR(trans);
2658                 trans = NULL;
2659                 goto out;
2660         }
2661
2662         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
2663
2664         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2665                 compress_type = ordered_extent->compress_type;
2666         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2667                 BUG_ON(compress_type);
2668                 btrfs_qgroup_free_data(inode, NULL, start,
2669                                        ordered_extent->num_bytes);
2670                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
2671                                                 ordered_extent->file_offset,
2672                                                 ordered_extent->file_offset +
2673                                                 logical_len);
2674         } else {
2675                 BUG_ON(root == fs_info->tree_root);
2676                 ret = insert_reserved_file_extent(trans, inode, start,
2677                                                 ordered_extent->disk_bytenr,
2678                                                 ordered_extent->disk_num_bytes,
2679                                                 logical_len, logical_len,
2680                                                 compress_type, 0, 0,
2681                                                 BTRFS_FILE_EXTENT_REG);
2682                 if (!ret) {
2683                         clear_reserved_extent = false;
2684                         btrfs_release_delalloc_bytes(fs_info,
2685                                                 ordered_extent->disk_bytenr,
2686                                                 ordered_extent->disk_num_bytes);
2687                 }
2688         }
2689         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2690                            ordered_extent->file_offset,
2691                            ordered_extent->num_bytes, trans->transid);
2692         if (ret < 0) {
2693                 btrfs_abort_transaction(trans, ret);
2694                 goto out;
2695         }
2696
2697         ret = add_pending_csums(trans, inode, &ordered_extent->list);
2698         if (ret) {
2699                 btrfs_abort_transaction(trans, ret);
2700                 goto out;
2701         }
2702
2703         btrfs_inode_safe_disk_i_size_write(inode, 0);
2704         ret = btrfs_update_inode_fallback(trans, root, inode);
2705         if (ret) { /* -ENOMEM or corruption */
2706                 btrfs_abort_transaction(trans, ret);
2707                 goto out;
2708         }
2709         ret = 0;
2710 out:
2711         clear_bits = EXTENT_DEFRAG;
2712         if (range_locked)
2713                 clear_bits |= EXTENT_LOCKED;
2714         if (clear_new_delalloc_bytes)
2715                 clear_bits |= EXTENT_DELALLOC_NEW;
2716         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, clear_bits,
2717                          (clear_bits & EXTENT_LOCKED) ? 1 : 0, 0,
2718                          &cached_state);
2719
2720         if (trans)
2721                 btrfs_end_transaction(trans);
2722
2723         if (ret || truncated) {
2724                 u64 unwritten_start = start;
2725
2726                 if (truncated)
2727                         unwritten_start += logical_len;
2728                 clear_extent_uptodate(io_tree, unwritten_start, end, NULL);
2729
2730                 /* Drop the cache for the part of the extent we didn't write. */
2731                 btrfs_drop_extent_cache(BTRFS_I(inode), unwritten_start, end, 0);
2732
2733                 /*
2734                  * If the ordered extent had an IOERR or something else went
2735                  * wrong we need to return the space for this ordered extent
2736                  * back to the allocator.  We only free the extent in the
2737                  * truncated case if we didn't write out the extent at all.
2738                  *
2739                  * If we made it past insert_reserved_file_extent before we
2740                  * errored out then we don't need to do this as the accounting
2741                  * has already been done.
2742                  */
2743                 if ((ret || !logical_len) &&
2744                     clear_reserved_extent &&
2745                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2746                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2747                         /*
2748                          * Discard the range before returning it back to the
2749                          * free space pool
2750                          */
2751                         if (ret && btrfs_test_opt(fs_info, DISCARD_SYNC))
2752                                 btrfs_discard_extent(fs_info,
2753                                                 ordered_extent->disk_bytenr,
2754                                                 ordered_extent->disk_num_bytes,
2755                                                 NULL);
2756                         btrfs_free_reserved_extent(fs_info,
2757                                         ordered_extent->disk_bytenr,
2758                                         ordered_extent->disk_num_bytes, 1);
2759                 }
2760         }
2761
2762         /*
2763          * This needs to be done to make sure anybody waiting knows we are done
2764          * updating everything for this ordered extent.
2765          */
2766         btrfs_remove_ordered_extent(inode, ordered_extent);
2767
2768         /* once for us */
2769         btrfs_put_ordered_extent(ordered_extent);
2770         /* once for the tree */
2771         btrfs_put_ordered_extent(ordered_extent);
2772
2773         return ret;
2774 }
2775
2776 static void finish_ordered_fn(struct btrfs_work *work)
2777 {
2778         struct btrfs_ordered_extent *ordered_extent;
2779         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
2780         btrfs_finish_ordered_io(ordered_extent);
2781 }
2782
2783 void btrfs_writepage_endio_finish_ordered(struct page *page, u64 start,
2784                                           u64 end, int uptodate)
2785 {
2786         struct inode *inode = page->mapping->host;
2787         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2788         struct btrfs_ordered_extent *ordered_extent = NULL;
2789         struct btrfs_workqueue *wq;
2790
2791         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
2792
2793         ClearPagePrivate2(page);
2794         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
2795                                             end - start + 1, uptodate))
2796                 return;
2797
2798         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
2799                 wq = fs_info->endio_freespace_worker;
2800         else
2801                 wq = fs_info->endio_write_workers;
2802
2803         btrfs_init_work(&ordered_extent->work, finish_ordered_fn, NULL, NULL);
2804         btrfs_queue_work(wq, &ordered_extent->work);
2805 }
2806
2807 static int check_data_csum(struct inode *inode, struct btrfs_io_bio *io_bio,
2808                            int icsum, struct page *page, int pgoff, u64 start,
2809                            size_t len)
2810 {
2811         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2812         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
2813         char *kaddr;
2814         u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2815         u8 *csum_expected;
2816         u8 csum[BTRFS_CSUM_SIZE];
2817
2818         csum_expected = ((u8 *)io_bio->csum) + icsum * csum_size;
2819
2820         kaddr = kmap_atomic(page);
2821         shash->tfm = fs_info->csum_shash;
2822
2823         crypto_shash_digest(shash, kaddr + pgoff, len, csum);
2824
2825         if (memcmp(csum, csum_expected, csum_size))
2826                 goto zeroit;
2827
2828         kunmap_atomic(kaddr);
2829         return 0;
2830 zeroit:
2831         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
2832                                     io_bio->mirror_num);
2833         memset(kaddr + pgoff, 1, len);
2834         flush_dcache_page(page);
2835         kunmap_atomic(kaddr);
2836         return -EIO;
2837 }
2838
2839 /*
2840  * when reads are done, we need to check csums to verify the data is correct
2841  * if there's a match, we allow the bio to finish.  If not, the code in
2842  * extent_io.c will try to find good copies for us.
2843  */
2844 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
2845                                       u64 phy_offset, struct page *page,
2846                                       u64 start, u64 end, int mirror)
2847 {
2848         size_t offset = start - page_offset(page);
2849         struct inode *inode = page->mapping->host;
2850         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2851         struct btrfs_root *root = BTRFS_I(inode)->root;
2852
2853         if (PageChecked(page)) {
2854                 ClearPageChecked(page);
2855                 return 0;
2856         }
2857
2858         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
2859                 return 0;
2860
2861         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
2862             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
2863                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
2864                 return 0;
2865         }
2866
2867         phy_offset >>= inode->i_sb->s_blocksize_bits;
2868         return check_data_csum(inode, io_bio, phy_offset, page, offset, start,
2869                                (size_t)(end - start + 1));
2870 }
2871
2872 /*
2873  * btrfs_add_delayed_iput - perform a delayed iput on @inode
2874  *
2875  * @inode: The inode we want to perform iput on
2876  *
2877  * This function uses the generic vfs_inode::i_count to track whether we should
2878  * just decrement it (in case it's > 1) or if this is the last iput then link
2879  * the inode to the delayed iput machinery. Delayed iputs are processed at
2880  * transaction commit time/superblock commit/cleaner kthread.
2881  */
2882 void btrfs_add_delayed_iput(struct inode *inode)
2883 {
2884         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2885         struct btrfs_inode *binode = BTRFS_I(inode);
2886
2887         if (atomic_add_unless(&inode->i_count, -1, 1))
2888                 return;
2889
2890         atomic_inc(&fs_info->nr_delayed_iputs);
2891         spin_lock(&fs_info->delayed_iput_lock);
2892         ASSERT(list_empty(&binode->delayed_iput));
2893         list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
2894         spin_unlock(&fs_info->delayed_iput_lock);
2895         if (!test_bit(BTRFS_FS_CLEANER_RUNNING, &fs_info->flags))
2896                 wake_up_process(fs_info->cleaner_kthread);
2897 }
2898
2899 static void run_delayed_iput_locked(struct btrfs_fs_info *fs_info,
2900                                     struct btrfs_inode *inode)
2901 {
2902         list_del_init(&inode->delayed_iput);
2903         spin_unlock(&fs_info->delayed_iput_lock);
2904         iput(&inode->vfs_inode);
2905         if (atomic_dec_and_test(&fs_info->nr_delayed_iputs))
2906                 wake_up(&fs_info->delayed_iputs_wait);
2907         spin_lock(&fs_info->delayed_iput_lock);
2908 }
2909
2910 static void btrfs_run_delayed_iput(struct btrfs_fs_info *fs_info,
2911                                    struct btrfs_inode *inode)
2912 {
2913         if (!list_empty(&inode->delayed_iput)) {
2914                 spin_lock(&fs_info->delayed_iput_lock);
2915                 if (!list_empty(&inode->delayed_iput))
2916                         run_delayed_iput_locked(fs_info, inode);
2917                 spin_unlock(&fs_info->delayed_iput_lock);
2918         }
2919 }
2920
2921 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
2922 {
2923
2924         spin_lock(&fs_info->delayed_iput_lock);
2925         while (!list_empty(&fs_info->delayed_iputs)) {
2926                 struct btrfs_inode *inode;
2927
2928                 inode = list_first_entry(&fs_info->delayed_iputs,
2929                                 struct btrfs_inode, delayed_iput);
2930                 run_delayed_iput_locked(fs_info, inode);
2931         }
2932         spin_unlock(&fs_info->delayed_iput_lock);
2933 }
2934
2935 /**
2936  * btrfs_wait_on_delayed_iputs - wait on the delayed iputs to be done running
2937  * @fs_info - the fs_info for this fs
2938  * @return - EINTR if we were killed, 0 if nothing's pending
2939  *
2940  * This will wait on any delayed iputs that are currently running with KILLABLE
2941  * set.  Once they are all done running we will return, unless we are killed in
2942  * which case we return EINTR. This helps in user operations like fallocate etc
2943  * that might get blocked on the iputs.
2944  */
2945 int btrfs_wait_on_delayed_iputs(struct btrfs_fs_info *fs_info)
2946 {
2947         int ret = wait_event_killable(fs_info->delayed_iputs_wait,
2948                         atomic_read(&fs_info->nr_delayed_iputs) == 0);
2949         if (ret)
2950                 return -EINTR;
2951         return 0;
2952 }
2953
2954 /*
2955  * This creates an orphan entry for the given inode in case something goes wrong
2956  * in the middle of an unlink.
2957  */
2958 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
2959                      struct btrfs_inode *inode)
2960 {
2961         int ret;
2962
2963         ret = btrfs_insert_orphan_item(trans, inode->root, btrfs_ino(inode));
2964         if (ret && ret != -EEXIST) {
2965                 btrfs_abort_transaction(trans, ret);
2966                 return ret;
2967         }
2968
2969         return 0;
2970 }
2971
2972 /*
2973  * We have done the delete so we can go ahead and remove the orphan item for
2974  * this particular inode.
2975  */
2976 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
2977                             struct btrfs_inode *inode)
2978 {
2979         return btrfs_del_orphan_item(trans, inode->root, btrfs_ino(inode));
2980 }
2981
2982 /*
2983  * this cleans up any orphans that may be left on the list from the last use
2984  * of this root.
2985  */
2986 int btrfs_orphan_cleanup(struct btrfs_root *root)
2987 {
2988         struct btrfs_fs_info *fs_info = root->fs_info;
2989         struct btrfs_path *path;
2990         struct extent_buffer *leaf;
2991         struct btrfs_key key, found_key;
2992         struct btrfs_trans_handle *trans;
2993         struct inode *inode;
2994         u64 last_objectid = 0;
2995         int ret = 0, nr_unlink = 0;
2996
2997         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
2998                 return 0;
2999
3000         path = btrfs_alloc_path();
3001         if (!path) {
3002                 ret = -ENOMEM;
3003                 goto out;
3004         }
3005         path->reada = READA_BACK;
3006
3007         key.objectid = BTRFS_ORPHAN_OBJECTID;
3008         key.type = BTRFS_ORPHAN_ITEM_KEY;
3009         key.offset = (u64)-1;
3010
3011         while (1) {
3012                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3013                 if (ret < 0)
3014                         goto out;
3015
3016                 /*
3017                  * if ret == 0 means we found what we were searching for, which
3018                  * is weird, but possible, so only screw with path if we didn't
3019                  * find the key and see if we have stuff that matches
3020                  */
3021                 if (ret > 0) {
3022                         ret = 0;
3023                         if (path->slots[0] == 0)
3024                                 break;
3025                         path->slots[0]--;
3026                 }
3027
3028                 /* pull out the item */
3029                 leaf = path->nodes[0];
3030                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3031
3032                 /* make sure the item matches what we want */
3033                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3034                         break;
3035                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3036                         break;
3037
3038                 /* release the path since we're done with it */
3039                 btrfs_release_path(path);
3040
3041                 /*
3042                  * this is where we are basically btrfs_lookup, without the
3043                  * crossing root thing.  we store the inode number in the
3044                  * offset of the orphan item.
3045                  */
3046
3047                 if (found_key.offset == last_objectid) {
3048                         btrfs_err(fs_info,
3049                                   "Error removing orphan entry, stopping orphan cleanup");
3050                         ret = -EINVAL;
3051                         goto out;
3052                 }
3053
3054                 last_objectid = found_key.offset;
3055
3056                 found_key.objectid = found_key.offset;
3057                 found_key.type = BTRFS_INODE_ITEM_KEY;
3058                 found_key.offset = 0;
3059                 inode = btrfs_iget(fs_info->sb, last_objectid, root);
3060                 ret = PTR_ERR_OR_ZERO(inode);
3061                 if (ret && ret != -ENOENT)
3062                         goto out;
3063
3064                 if (ret == -ENOENT && root == fs_info->tree_root) {
3065                         struct btrfs_root *dead_root;
3066                         struct btrfs_fs_info *fs_info = root->fs_info;
3067                         int is_dead_root = 0;
3068
3069                         /*
3070                          * this is an orphan in the tree root. Currently these
3071                          * could come from 2 sources:
3072                          *  a) a snapshot deletion in progress
3073                          *  b) a free space cache inode
3074                          * We need to distinguish those two, as the snapshot
3075                          * orphan must not get deleted.
3076                          * find_dead_roots already ran before us, so if this
3077                          * is a snapshot deletion, we should find the root
3078                          * in the fs_roots radix tree.
3079                          */
3080
3081                         spin_lock(&fs_info->fs_roots_radix_lock);
3082                         dead_root = radix_tree_lookup(&fs_info->fs_roots_radix,
3083                                                          (unsigned long)found_key.objectid);
3084                         if (dead_root && btrfs_root_refs(&dead_root->root_item) == 0)
3085                                 is_dead_root = 1;
3086                         spin_unlock(&fs_info->fs_roots_radix_lock);
3087
3088                         if (is_dead_root) {
3089                                 /* prevent this orphan from being found again */
3090                                 key.offset = found_key.objectid - 1;
3091                                 continue;
3092                         }
3093
3094                 }
3095
3096                 /*
3097                  * If we have an inode with links, there are a couple of
3098                  * possibilities. Old kernels (before v3.12) used to create an
3099                  * orphan item for truncate indicating that there were possibly
3100                  * extent items past i_size that needed to be deleted. In v3.12,
3101                  * truncate was changed to update i_size in sync with the extent
3102                  * items, but the (useless) orphan item was still created. Since
3103                  * v4.18, we don't create the orphan item for truncate at all.
3104                  *
3105                  * So, this item could mean that we need to do a truncate, but
3106                  * only if this filesystem was last used on a pre-v3.12 kernel
3107                  * and was not cleanly unmounted. The odds of that are quite
3108                  * slim, and it's a pain to do the truncate now, so just delete
3109                  * the orphan item.
3110                  *
3111                  * It's also possible that this orphan item was supposed to be
3112                  * deleted but wasn't. The inode number may have been reused,
3113                  * but either way, we can delete the orphan item.
3114                  */
3115                 if (ret == -ENOENT || inode->i_nlink) {
3116                         if (!ret)
3117                                 iput(inode);
3118                         trans = btrfs_start_transaction(root, 1);
3119                         if (IS_ERR(trans)) {
3120                                 ret = PTR_ERR(trans);
3121                                 goto out;
3122                         }
3123                         btrfs_debug(fs_info, "auto deleting %Lu",
3124                                     found_key.objectid);
3125                         ret = btrfs_del_orphan_item(trans, root,
3126                                                     found_key.objectid);
3127                         btrfs_end_transaction(trans);
3128                         if (ret)
3129                                 goto out;
3130                         continue;
3131                 }
3132
3133                 nr_unlink++;
3134
3135                 /* this will do delete_inode and everything for us */
3136                 iput(inode);
3137         }
3138         /* release the path since we're done with it */
3139         btrfs_release_path(path);
3140
3141         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3142
3143         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3144                 trans = btrfs_join_transaction(root);
3145                 if (!IS_ERR(trans))
3146                         btrfs_end_transaction(trans);
3147         }
3148
3149         if (nr_unlink)
3150                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3151
3152 out:
3153         if (ret)
3154                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3155         btrfs_free_path(path);
3156         return ret;
3157 }
3158
3159 /*
3160  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3161  * don't find any xattrs, we know there can't be any acls.
3162  *
3163  * slot is the slot the inode is in, objectid is the objectid of the inode
3164  */
3165 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3166                                           int slot, u64 objectid,
3167                                           int *first_xattr_slot)
3168 {
3169         u32 nritems = btrfs_header_nritems(leaf);
3170         struct btrfs_key found_key;
3171         static u64 xattr_access = 0;
3172         static u64 xattr_default = 0;
3173         int scanned = 0;
3174
3175         if (!xattr_access) {
3176                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3177                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3178                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3179                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3180         }
3181
3182         slot++;
3183         *first_xattr_slot = -1;
3184         while (slot < nritems) {
3185                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3186
3187                 /* we found a different objectid, there must not be acls */
3188                 if (found_key.objectid != objectid)
3189                         return 0;
3190
3191                 /* we found an xattr, assume we've got an acl */
3192                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3193                         if (*first_xattr_slot == -1)
3194                                 *first_xattr_slot = slot;
3195                         if (found_key.offset == xattr_access ||
3196                             found_key.offset == xattr_default)
3197                                 return 1;
3198                 }
3199
3200                 /*
3201                  * we found a key greater than an xattr key, there can't
3202                  * be any acls later on
3203                  */
3204                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3205                         return 0;
3206
3207                 slot++;
3208                 scanned++;
3209
3210                 /*
3211                  * it goes inode, inode backrefs, xattrs, extents,
3212                  * so if there are a ton of hard links to an inode there can
3213                  * be a lot of backrefs.  Don't waste time searching too hard,
3214                  * this is just an optimization
3215                  */
3216                 if (scanned >= 8)
3217                         break;
3218         }
3219         /* we hit the end of the leaf before we found an xattr or
3220          * something larger than an xattr.  We have to assume the inode
3221          * has acls
3222          */
3223         if (*first_xattr_slot == -1)
3224                 *first_xattr_slot = slot;
3225         return 1;
3226 }
3227
3228 /*
3229  * read an inode from the btree into the in-memory inode
3230  */
3231 static int btrfs_read_locked_inode(struct inode *inode,
3232                                    struct btrfs_path *in_path)
3233 {
3234         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3235         struct btrfs_path *path = in_path;
3236         struct extent_buffer *leaf;
3237         struct btrfs_inode_item *inode_item;
3238         struct btrfs_root *root = BTRFS_I(inode)->root;
3239         struct btrfs_key location;
3240         unsigned long ptr;
3241         int maybe_acls;
3242         u32 rdev;
3243         int ret;
3244         bool filled = false;
3245         int first_xattr_slot;
3246
3247         ret = btrfs_fill_inode(inode, &rdev);
3248         if (!ret)
3249                 filled = true;
3250
3251         if (!path) {
3252                 path = btrfs_alloc_path();
3253                 if (!path)
3254                         return -ENOMEM;
3255         }
3256
3257         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3258
3259         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3260         if (ret) {
3261                 if (path != in_path)
3262                         btrfs_free_path(path);
3263                 return ret;
3264         }
3265
3266         leaf = path->nodes[0];
3267
3268         if (filled)
3269                 goto cache_index;
3270
3271         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3272                                     struct btrfs_inode_item);
3273         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3274         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3275         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3276         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3277         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3278         btrfs_inode_set_file_extent_range(BTRFS_I(inode), 0,
3279                         round_up(i_size_read(inode), fs_info->sectorsize));
3280
3281         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3282         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3283
3284         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3285         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3286
3287         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3288         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3289
3290         BTRFS_I(inode)->i_otime.tv_sec =
3291                 btrfs_timespec_sec(leaf, &inode_item->otime);
3292         BTRFS_I(inode)->i_otime.tv_nsec =
3293                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3294
3295         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3296         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3297         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3298
3299         inode_set_iversion_queried(inode,
3300                                    btrfs_inode_sequence(leaf, inode_item));
3301         inode->i_generation = BTRFS_I(inode)->generation;
3302         inode->i_rdev = 0;
3303         rdev = btrfs_inode_rdev(leaf, inode_item);
3304
3305         BTRFS_I(inode)->index_cnt = (u64)-1;
3306         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3307
3308 cache_index:
3309         /*
3310          * If we were modified in the current generation and evicted from memory
3311          * and then re-read we need to do a full sync since we don't have any
3312          * idea about which extents were modified before we were evicted from
3313          * cache.
3314          *
3315          * This is required for both inode re-read from disk and delayed inode
3316          * in delayed_nodes_tree.
3317          */
3318         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3319                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3320                         &BTRFS_I(inode)->runtime_flags);
3321
3322         /*
3323          * We don't persist the id of the transaction where an unlink operation
3324          * against the inode was last made. So here we assume the inode might
3325          * have been evicted, and therefore the exact value of last_unlink_trans
3326          * lost, and set it to last_trans to avoid metadata inconsistencies
3327          * between the inode and its parent if the inode is fsync'ed and the log
3328          * replayed. For example, in the scenario:
3329          *
3330          * touch mydir/foo
3331          * ln mydir/foo mydir/bar
3332          * sync
3333          * unlink mydir/bar
3334          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3335          * xfs_io -c fsync mydir/foo
3336          * <power failure>
3337          * mount fs, triggers fsync log replay
3338          *
3339          * We must make sure that when we fsync our inode foo we also log its
3340          * parent inode, otherwise after log replay the parent still has the
3341          * dentry with the "bar" name but our inode foo has a link count of 1
3342          * and doesn't have an inode ref with the name "bar" anymore.
3343          *
3344          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3345          * but it guarantees correctness at the expense of occasional full
3346          * transaction commits on fsync if our inode is a directory, or if our
3347          * inode is not a directory, logging its parent unnecessarily.
3348          */
3349         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3350
3351         path->slots[0]++;
3352         if (inode->i_nlink != 1 ||
3353             path->slots[0] >= btrfs_header_nritems(leaf))
3354                 goto cache_acl;
3355
3356         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3357         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3358                 goto cache_acl;
3359
3360         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3361         if (location.type == BTRFS_INODE_REF_KEY) {
3362                 struct btrfs_inode_ref *ref;
3363
3364                 ref = (struct btrfs_inode_ref *)ptr;
3365                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3366         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3367                 struct btrfs_inode_extref *extref;
3368
3369                 extref = (struct btrfs_inode_extref *)ptr;
3370                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3371                                                                      extref);
3372         }
3373 cache_acl:
3374         /*
3375          * try to precache a NULL acl entry for files that don't have
3376          * any xattrs or acls
3377          */
3378         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3379                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3380         if (first_xattr_slot != -1) {
3381                 path->slots[0] = first_xattr_slot;
3382                 ret = btrfs_load_inode_props(inode, path);
3383                 if (ret)
3384                         btrfs_err(fs_info,
3385                                   "error loading props for ino %llu (root %llu): %d",
3386                                   btrfs_ino(BTRFS_I(inode)),
3387                                   root->root_key.objectid, ret);
3388         }
3389         if (path != in_path)
3390                 btrfs_free_path(path);
3391
3392         if (!maybe_acls)
3393                 cache_no_acl(inode);
3394
3395         switch (inode->i_mode & S_IFMT) {
3396         case S_IFREG:
3397                 inode->i_mapping->a_ops = &btrfs_aops;
3398                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3399                 inode->i_fop = &btrfs_file_operations;
3400                 inode->i_op = &btrfs_file_inode_operations;
3401                 break;
3402         case S_IFDIR:
3403                 inode->i_fop = &btrfs_dir_file_operations;
3404                 inode->i_op = &btrfs_dir_inode_operations;
3405                 break;
3406         case S_IFLNK:
3407                 inode->i_op = &btrfs_symlink_inode_operations;
3408                 inode_nohighmem(inode);
3409                 inode->i_mapping->a_ops = &btrfs_aops;
3410                 break;
3411         default:
3412                 inode->i_op = &btrfs_special_inode_operations;
3413                 init_special_inode(inode, inode->i_mode, rdev);
3414                 break;
3415         }
3416
3417         btrfs_sync_inode_flags_to_i_flags(inode);
3418         return 0;
3419 }
3420
3421 /*
3422  * given a leaf and an inode, copy the inode fields into the leaf
3423  */
3424 static void fill_inode_item(struct btrfs_trans_handle *trans,
3425                             struct extent_buffer *leaf,
3426                             struct btrfs_inode_item *item,
3427                             struct inode *inode)
3428 {
3429         struct btrfs_map_token token;
3430
3431         btrfs_init_map_token(&token, leaf);
3432
3433         btrfs_set_token_inode_uid(&token, item, i_uid_read(inode));
3434         btrfs_set_token_inode_gid(&token, item, i_gid_read(inode));
3435         btrfs_set_token_inode_size(&token, item, BTRFS_I(inode)->disk_i_size);
3436         btrfs_set_token_inode_mode(&token, item, inode->i_mode);
3437         btrfs_set_token_inode_nlink(&token, item, inode->i_nlink);
3438
3439         btrfs_set_token_timespec_sec(&token, &item->atime,
3440                                      inode->i_atime.tv_sec);
3441         btrfs_set_token_timespec_nsec(&token, &item->atime,
3442                                       inode->i_atime.tv_nsec);
3443
3444         btrfs_set_token_timespec_sec(&token, &item->mtime,
3445                                      inode->i_mtime.tv_sec);
3446         btrfs_set_token_timespec_nsec(&token, &item->mtime,
3447                                       inode->i_mtime.tv_nsec);
3448
3449         btrfs_set_token_timespec_sec(&token, &item->ctime,
3450                                      inode->i_ctime.tv_sec);
3451         btrfs_set_token_timespec_nsec(&token, &item->ctime,
3452                                       inode->i_ctime.tv_nsec);
3453
3454         btrfs_set_token_timespec_sec(&token, &item->otime,
3455                                      BTRFS_I(inode)->i_otime.tv_sec);
3456         btrfs_set_token_timespec_nsec(&token, &item->otime,
3457                                       BTRFS_I(inode)->i_otime.tv_nsec);
3458
3459         btrfs_set_token_inode_nbytes(&token, item, inode_get_bytes(inode));
3460         btrfs_set_token_inode_generation(&token, item,
3461                                          BTRFS_I(inode)->generation);
3462         btrfs_set_token_inode_sequence(&token, item, inode_peek_iversion(inode));
3463         btrfs_set_token_inode_transid(&token, item, trans->transid);
3464         btrfs_set_token_inode_rdev(&token, item, inode->i_rdev);
3465         btrfs_set_token_inode_flags(&token, item, BTRFS_I(inode)->flags);
3466         btrfs_set_token_inode_block_group(&token, item, 0);
3467 }
3468
3469 /*
3470  * copy everything in the in-memory inode into the btree.
3471  */
3472 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3473                                 struct btrfs_root *root, struct inode *inode)
3474 {
3475         struct btrfs_inode_item *inode_item;
3476         struct btrfs_path *path;
3477         struct extent_buffer *leaf;
3478         int ret;
3479
3480         path = btrfs_alloc_path();
3481         if (!path)
3482                 return -ENOMEM;
3483
3484         path->leave_spinning = 1;
3485         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3486                                  1);
3487         if (ret) {
3488                 if (ret > 0)
3489                         ret = -ENOENT;
3490                 goto failed;
3491         }
3492
3493         leaf = path->nodes[0];
3494         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3495                                     struct btrfs_inode_item);
3496
3497         fill_inode_item(trans, leaf, inode_item, inode);
3498         btrfs_mark_buffer_dirty(leaf);
3499         btrfs_set_inode_last_trans(trans, inode);
3500         ret = 0;
3501 failed:
3502         btrfs_free_path(path);
3503         return ret;
3504 }
3505
3506 /*
3507  * copy everything in the in-memory inode into the btree.
3508  */
3509 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3510                                 struct btrfs_root *root, struct inode *inode)
3511 {
3512         struct btrfs_fs_info *fs_info = root->fs_info;
3513         int ret;
3514
3515         /*
3516          * If the inode is a free space inode, we can deadlock during commit
3517          * if we put it into the delayed code.
3518          *
3519          * The data relocation inode should also be directly updated
3520          * without delay
3521          */
3522         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
3523             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3524             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
3525                 btrfs_update_root_times(trans, root);
3526
3527                 ret = btrfs_delayed_update_inode(trans, root, inode);
3528                 if (!ret)
3529                         btrfs_set_inode_last_trans(trans, inode);
3530                 return ret;
3531         }
3532
3533         return btrfs_update_inode_item(trans, root, inode);
3534 }
3535
3536 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3537                                          struct btrfs_root *root,
3538                                          struct inode *inode)
3539 {
3540         int ret;
3541
3542         ret = btrfs_update_inode(trans, root, inode);
3543         if (ret == -ENOSPC)
3544                 return btrfs_update_inode_item(trans, root, inode);
3545         return ret;
3546 }
3547
3548 /*
3549  * unlink helper that gets used here in inode.c and in the tree logging
3550  * recovery code.  It remove a link in a directory with a given name, and
3551  * also drops the back refs in the inode to the directory
3552  */
3553 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3554                                 struct btrfs_root *root,
3555                                 struct btrfs_inode *dir,
3556                                 struct btrfs_inode *inode,
3557                                 const char *name, int name_len)
3558 {
3559         struct btrfs_fs_info *fs_info = root->fs_info;
3560         struct btrfs_path *path;
3561         int ret = 0;
3562         struct btrfs_dir_item *di;
3563         u64 index;
3564         u64 ino = btrfs_ino(inode);
3565         u64 dir_ino = btrfs_ino(dir);
3566
3567         path = btrfs_alloc_path();
3568         if (!path) {
3569                 ret = -ENOMEM;
3570                 goto out;
3571         }
3572
3573         path->leave_spinning = 1;
3574         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3575                                     name, name_len, -1);
3576         if (IS_ERR_OR_NULL(di)) {
3577                 ret = di ? PTR_ERR(di) : -ENOENT;
3578                 goto err;
3579         }
3580         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3581         if (ret)
3582                 goto err;
3583         btrfs_release_path(path);
3584
3585         /*
3586          * If we don't have dir index, we have to get it by looking up
3587          * the inode ref, since we get the inode ref, remove it directly,
3588          * it is unnecessary to do delayed deletion.
3589          *
3590          * But if we have dir index, needn't search inode ref to get it.
3591          * Since the inode ref is close to the inode item, it is better
3592          * that we delay to delete it, and just do this deletion when
3593          * we update the inode item.
3594          */
3595         if (inode->dir_index) {
3596                 ret = btrfs_delayed_delete_inode_ref(inode);
3597                 if (!ret) {
3598                         index = inode->dir_index;
3599                         goto skip_backref;
3600                 }
3601         }
3602
3603         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
3604                                   dir_ino, &index);
3605         if (ret) {
3606                 btrfs_info(fs_info,
3607                         "failed to delete reference to %.*s, inode %llu parent %llu",
3608                         name_len, name, ino, dir_ino);
3609                 btrfs_abort_transaction(trans, ret);
3610                 goto err;
3611         }
3612 skip_backref:
3613         ret = btrfs_delete_delayed_dir_index(trans, dir, index);
3614         if (ret) {
3615                 btrfs_abort_transaction(trans, ret);
3616                 goto err;
3617         }
3618
3619         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
3620                         dir_ino);
3621         if (ret != 0 && ret != -ENOENT) {
3622                 btrfs_abort_transaction(trans, ret);
3623                 goto err;
3624         }
3625
3626         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
3627                         index);
3628         if (ret == -ENOENT)
3629                 ret = 0;
3630         else if (ret)
3631                 btrfs_abort_transaction(trans, ret);
3632
3633         /*
3634          * If we have a pending delayed iput we could end up with the final iput
3635          * being run in btrfs-cleaner context.  If we have enough of these built
3636          * up we can end up burning a lot of time in btrfs-cleaner without any
3637          * way to throttle the unlinks.  Since we're currently holding a ref on
3638          * the inode we can run the delayed iput here without any issues as the
3639          * final iput won't be done until after we drop the ref we're currently
3640          * holding.
3641          */
3642         btrfs_run_delayed_iput(fs_info, inode);
3643 err:
3644         btrfs_free_path(path);
3645         if (ret)
3646                 goto out;
3647
3648         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
3649         inode_inc_iversion(&inode->vfs_inode);
3650         inode_inc_iversion(&dir->vfs_inode);
3651         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
3652                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
3653         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
3654 out:
3655         return ret;
3656 }
3657
3658 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3659                        struct btrfs_root *root,
3660                        struct btrfs_inode *dir, struct btrfs_inode *inode,
3661                        const char *name, int name_len)
3662 {
3663         int ret;
3664         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
3665         if (!ret) {
3666                 drop_nlink(&inode->vfs_inode);
3667                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
3668         }
3669         return ret;
3670 }
3671
3672 /*
3673  * helper to start transaction for unlink and rmdir.
3674  *
3675  * unlink and rmdir are special in btrfs, they do not always free space, so
3676  * if we cannot make our reservations the normal way try and see if there is
3677  * plenty of slack room in the global reserve to migrate, otherwise we cannot
3678  * allow the unlink to occur.
3679  */
3680 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
3681 {
3682         struct btrfs_root *root = BTRFS_I(dir)->root;
3683
3684         /*
3685          * 1 for the possible orphan item
3686          * 1 for the dir item
3687          * 1 for the dir index
3688          * 1 for the inode ref
3689          * 1 for the inode
3690          */
3691         return btrfs_start_transaction_fallback_global_rsv(root, 5);
3692 }
3693
3694 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
3695 {
3696         struct btrfs_root *root = BTRFS_I(dir)->root;
3697         struct btrfs_trans_handle *trans;
3698         struct inode *inode = d_inode(dentry);
3699         int ret;
3700
3701         trans = __unlink_start_trans(dir);
3702         if (IS_ERR(trans))
3703                 return PTR_ERR(trans);
3704
3705         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
3706                         0);
3707
3708         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
3709                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
3710                         dentry->d_name.len);
3711         if (ret)
3712                 goto out;
3713
3714         if (inode->i_nlink == 0) {
3715                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
3716                 if (ret)
3717                         goto out;
3718         }
3719
3720 out:
3721         btrfs_end_transaction(trans);
3722         btrfs_btree_balance_dirty(root->fs_info);
3723         return ret;
3724 }
3725
3726 static int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
3727                                struct inode *dir, struct dentry *dentry)
3728 {
3729         struct btrfs_root *root = BTRFS_I(dir)->root;
3730         struct btrfs_inode *inode = BTRFS_I(d_inode(dentry));
3731         struct btrfs_path *path;
3732         struct extent_buffer *leaf;
3733         struct btrfs_dir_item *di;
3734         struct btrfs_key key;
3735         const char *name = dentry->d_name.name;
3736         int name_len = dentry->d_name.len;
3737         u64 index;
3738         int ret;
3739         u64 objectid;
3740         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
3741
3742         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID) {
3743                 objectid = inode->root->root_key.objectid;
3744         } else if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID) {
3745                 objectid = inode->location.objectid;
3746         } else {
3747                 WARN_ON(1);
3748                 return -EINVAL;
3749         }
3750
3751         path = btrfs_alloc_path();
3752         if (!path)
3753                 return -ENOMEM;
3754
3755         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3756                                    name, name_len, -1);
3757         if (IS_ERR_OR_NULL(di)) {
3758                 ret = di ? PTR_ERR(di) : -ENOENT;
3759                 goto out;
3760         }
3761
3762         leaf = path->nodes[0];
3763         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3764         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
3765         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3766         if (ret) {
3767                 btrfs_abort_transaction(trans, ret);
3768                 goto out;
3769         }
3770         btrfs_release_path(path);
3771
3772         /*
3773          * This is a placeholder inode for a subvolume we didn't have a
3774          * reference to at the time of the snapshot creation.  In the meantime
3775          * we could have renamed the real subvol link into our snapshot, so
3776          * depending on btrfs_del_root_ref to return -ENOENT here is incorret.
3777          * Instead simply lookup the dir_index_item for this entry so we can
3778          * remove it.  Otherwise we know we have a ref to the root and we can
3779          * call btrfs_del_root_ref, and it _shouldn't_ fail.
3780          */
3781         if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID) {
3782                 di = btrfs_search_dir_index_item(root, path, dir_ino,
3783                                                  name, name_len);
3784                 if (IS_ERR_OR_NULL(di)) {
3785                         if (!di)
3786                                 ret = -ENOENT;
3787                         else
3788                                 ret = PTR_ERR(di);
3789                         btrfs_abort_transaction(trans, ret);
3790                         goto out;
3791                 }
3792
3793                 leaf = path->nodes[0];
3794                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3795                 index = key.offset;
3796                 btrfs_release_path(path);
3797         } else {
3798                 ret = btrfs_del_root_ref(trans, objectid,
3799                                          root->root_key.objectid, dir_ino,
3800                                          &index, name, name_len);
3801                 if (ret) {
3802                         btrfs_abort_transaction(trans, ret);
3803                         goto out;
3804                 }
3805         }
3806
3807         ret = btrfs_delete_delayed_dir_index(trans, BTRFS_I(dir), index);
3808         if (ret) {
3809                 btrfs_abort_transaction(trans, ret);
3810                 goto out;
3811         }
3812
3813         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
3814         inode_inc_iversion(dir);
3815         dir->i_mtime = dir->i_ctime = current_time(dir);
3816         ret = btrfs_update_inode_fallback(trans, root, dir);
3817         if (ret)
3818                 btrfs_abort_transaction(trans, ret);
3819 out:
3820         btrfs_free_path(path);
3821         return ret;
3822 }
3823
3824 /*
3825  * Helper to check if the subvolume references other subvolumes or if it's
3826  * default.
3827  */
3828 static noinline int may_destroy_subvol(struct btrfs_root *root)
3829 {
3830         struct btrfs_fs_info *fs_info = root->fs_info;
3831         struct btrfs_path *path;
3832         struct btrfs_dir_item *di;
3833         struct btrfs_key key;
3834         u64 dir_id;
3835         int ret;
3836
3837         path = btrfs_alloc_path();
3838         if (!path)
3839                 return -ENOMEM;
3840
3841         /* Make sure this root isn't set as the default subvol */
3842         dir_id = btrfs_super_root_dir(fs_info->super_copy);
3843         di = btrfs_lookup_dir_item(NULL, fs_info->tree_root, path,
3844                                    dir_id, "default", 7, 0);
3845         if (di && !IS_ERR(di)) {
3846                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
3847                 if (key.objectid == root->root_key.objectid) {
3848                         ret = -EPERM;
3849                         btrfs_err(fs_info,
3850                                   "deleting default subvolume %llu is not allowed",
3851                                   key.objectid);
3852                         goto out;
3853                 }
3854                 btrfs_release_path(path);
3855         }
3856
3857         key.objectid = root->root_key.objectid;
3858         key.type = BTRFS_ROOT_REF_KEY;
3859         key.offset = (u64)-1;
3860
3861         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
3862         if (ret < 0)
3863                 goto out;
3864         BUG_ON(ret == 0);
3865
3866         ret = 0;
3867         if (path->slots[0] > 0) {
3868                 path->slots[0]--;
3869                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3870                 if (key.objectid == root->root_key.objectid &&
3871                     key.type == BTRFS_ROOT_REF_KEY)
3872                         ret = -ENOTEMPTY;
3873         }
3874 out:
3875         btrfs_free_path(path);
3876         return ret;
3877 }
3878
3879 /* Delete all dentries for inodes belonging to the root */
3880 static void btrfs_prune_dentries(struct btrfs_root *root)
3881 {
3882         struct btrfs_fs_info *fs_info = root->fs_info;
3883         struct rb_node *node;
3884         struct rb_node *prev;
3885         struct btrfs_inode *entry;
3886         struct inode *inode;
3887         u64 objectid = 0;
3888
3889         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
3890                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
3891
3892         spin_lock(&root->inode_lock);
3893 again:
3894         node = root->inode_tree.rb_node;
3895         prev = NULL;
3896         while (node) {
3897                 prev = node;
3898                 entry = rb_entry(node, struct btrfs_inode, rb_node);
3899
3900                 if (objectid < btrfs_ino(entry))
3901                         node = node->rb_left;
3902                 else if (objectid > btrfs_ino(entry))
3903                         node = node->rb_right;
3904                 else
3905                         break;
3906         }
3907         if (!node) {
3908                 while (prev) {
3909                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
3910                         if (objectid <= btrfs_ino(entry)) {
3911                                 node = prev;
3912                                 break;
3913                         }
3914                         prev = rb_next(prev);
3915                 }
3916         }
3917         while (node) {
3918                 entry = rb_entry(node, struct btrfs_inode, rb_node);
3919                 objectid = btrfs_ino(entry) + 1;
3920                 inode = igrab(&entry->vfs_inode);
3921                 if (inode) {
3922                         spin_unlock(&root->inode_lock);
3923                         if (atomic_read(&inode->i_count) > 1)
3924                                 d_prune_aliases(inode);
3925                         /*
3926                          * btrfs_drop_inode will have it removed from the inode
3927                          * cache when its usage count hits zero.
3928                          */
3929                         iput(inode);
3930                         cond_resched();
3931                         spin_lock(&root->inode_lock);
3932                         goto again;
3933                 }
3934
3935                 if (cond_resched_lock(&root->inode_lock))
3936                         goto again;
3937
3938                 node = rb_next(node);
3939         }
3940         spin_unlock(&root->inode_lock);
3941 }
3942
3943 int btrfs_delete_subvolume(struct inode *dir, struct dentry *dentry)
3944 {
3945         struct btrfs_fs_info *fs_info = btrfs_sb(dentry->d_sb);
3946         struct btrfs_root *root = BTRFS_I(dir)->root;
3947         struct inode *inode = d_inode(dentry);
3948         struct btrfs_root *dest = BTRFS_I(inode)->root;
3949         struct btrfs_trans_handle *trans;
3950         struct btrfs_block_rsv block_rsv;
3951         u64 root_flags;
3952         int ret;
3953         int err;
3954
3955         /*
3956          * Don't allow to delete a subvolume with send in progress. This is
3957          * inside the inode lock so the error handling that has to drop the bit
3958          * again is not run concurrently.
3959          */
3960         spin_lock(&dest->root_item_lock);
3961         if (dest->send_in_progress) {
3962                 spin_unlock(&dest->root_item_lock);
3963                 btrfs_warn(fs_info,
3964                            "attempt to delete subvolume %llu during send",
3965                            dest->root_key.objectid);
3966                 return -EPERM;
3967         }
3968         root_flags = btrfs_root_flags(&dest->root_item);
3969         btrfs_set_root_flags(&dest->root_item,
3970                              root_flags | BTRFS_ROOT_SUBVOL_DEAD);
3971         spin_unlock(&dest->root_item_lock);
3972
3973         down_write(&fs_info->subvol_sem);
3974
3975         err = may_destroy_subvol(dest);
3976         if (err)
3977                 goto out_up_write;
3978
3979         btrfs_init_block_rsv(&block_rsv, BTRFS_BLOCK_RSV_TEMP);
3980         /*
3981          * One for dir inode,
3982          * two for dir entries,
3983          * two for root ref/backref.
3984          */
3985         err = btrfs_subvolume_reserve_metadata(root, &block_rsv, 5, true);
3986         if (err)
3987                 goto out_up_write;
3988
3989         trans = btrfs_start_transaction(root, 0);
3990         if (IS_ERR(trans)) {
3991                 err = PTR_ERR(trans);
3992                 goto out_release;
3993         }
3994         trans->block_rsv = &block_rsv;
3995         trans->bytes_reserved = block_rsv.size;
3996
3997         btrfs_record_snapshot_destroy(trans, BTRFS_I(dir));
3998
3999         ret = btrfs_unlink_subvol(trans, dir, dentry);
4000         if (ret) {
4001                 err = ret;
4002                 btrfs_abort_transaction(trans, ret);
4003                 goto out_end_trans;
4004         }
4005
4006         btrfs_record_root_in_trans(trans, dest);
4007
4008         memset(&dest->root_item.drop_progress, 0,
4009                 sizeof(dest->root_item.drop_progress));
4010         dest->root_item.drop_level = 0;
4011         btrfs_set_root_refs(&dest->root_item, 0);
4012
4013         if (!test_and_set_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &dest->state)) {
4014                 ret = btrfs_insert_orphan_item(trans,
4015                                         fs_info->tree_root,
4016                                         dest->root_key.objectid);
4017                 if (ret) {
4018                         btrfs_abort_transaction(trans, ret);
4019                         err = ret;
4020                         goto out_end_trans;
4021                 }
4022         }
4023
4024         ret = btrfs_uuid_tree_remove(trans, dest->root_item.uuid,
4025                                   BTRFS_UUID_KEY_SUBVOL,
4026                                   dest->root_key.objectid);
4027         if (ret && ret != -ENOENT) {
4028                 btrfs_abort_transaction(trans, ret);
4029                 err = ret;
4030                 goto out_end_trans;
4031         }
4032         if (!btrfs_is_empty_uuid(dest->root_item.received_uuid)) {
4033                 ret = btrfs_uuid_tree_remove(trans,
4034                                           dest->root_item.received_uuid,
4035                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
4036                                           dest->root_key.objectid);
4037                 if (ret && ret != -ENOENT) {
4038                         btrfs_abort_transaction(trans, ret);
4039                         err = ret;
4040                         goto out_end_trans;
4041                 }
4042         }
4043
4044 out_end_trans:
4045         trans->block_rsv = NULL;
4046         trans->bytes_reserved = 0;
4047         ret = btrfs_end_transaction(trans);
4048         if (ret && !err)
4049                 err = ret;
4050         inode->i_flags |= S_DEAD;
4051 out_release:
4052         btrfs_subvolume_release_metadata(fs_info, &block_rsv);
4053 out_up_write:
4054         up_write(&fs_info->subvol_sem);
4055         if (err) {
4056                 spin_lock(&dest->root_item_lock);
4057                 root_flags = btrfs_root_flags(&dest->root_item);
4058                 btrfs_set_root_flags(&dest->root_item,
4059                                 root_flags & ~BTRFS_ROOT_SUBVOL_DEAD);
4060                 spin_unlock(&dest->root_item_lock);
4061         } else {
4062                 d_invalidate(dentry);
4063                 btrfs_prune_dentries(dest);
4064                 ASSERT(dest->send_in_progress == 0);
4065
4066                 /* the last ref */
4067                 if (dest->ino_cache_inode) {
4068                         iput(dest->ino_cache_inode);
4069                         dest->ino_cache_inode = NULL;
4070                 }
4071         }
4072
4073         return err;
4074 }
4075
4076 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4077 {
4078         struct inode *inode = d_inode(dentry);
4079         int err = 0;
4080         struct btrfs_root *root = BTRFS_I(dir)->root;
4081         struct btrfs_trans_handle *trans;
4082         u64 last_unlink_trans;
4083
4084         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4085                 return -ENOTEMPTY;
4086         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4087                 return btrfs_delete_subvolume(dir, dentry);
4088
4089         trans = __unlink_start_trans(dir);
4090         if (IS_ERR(trans))
4091                 return PTR_ERR(trans);
4092
4093         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4094                 err = btrfs_unlink_subvol(trans, dir, dentry);
4095                 goto out;
4096         }
4097
4098         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4099         if (err)
4100                 goto out;
4101
4102         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4103
4104         /* now the directory is empty */
4105         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4106                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4107                         dentry->d_name.len);
4108         if (!err) {
4109                 btrfs_i_size_write(BTRFS_I(inode), 0);
4110                 /*
4111                  * Propagate the last_unlink_trans value of the deleted dir to
4112                  * its parent directory. This is to prevent an unrecoverable
4113                  * log tree in the case we do something like this:
4114                  * 1) create dir foo
4115                  * 2) create snapshot under dir foo
4116                  * 3) delete the snapshot
4117                  * 4) rmdir foo
4118                  * 5) mkdir foo
4119                  * 6) fsync foo or some file inside foo
4120                  */
4121                 if (last_unlink_trans >= trans->transid)
4122                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4123         }
4124 out:
4125         btrfs_end_transaction(trans);
4126         btrfs_btree_balance_dirty(root->fs_info);
4127
4128         return err;
4129 }
4130
4131 /*
4132  * Return this if we need to call truncate_block for the last bit of the
4133  * truncate.
4134  */
4135 #define NEED_TRUNCATE_BLOCK 1
4136
4137 /*
4138  * this can truncate away extent items, csum items and directory items.
4139  * It starts at a high offset and removes keys until it can't find
4140  * any higher than new_size
4141  *
4142  * csum items that cross the new i_size are truncated to the new size
4143  * as well.
4144  *
4145  * min_type is the minimum key type to truncate down to.  If set to 0, this
4146  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4147  */
4148 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4149                                struct btrfs_root *root,
4150                                struct inode *inode,
4151                                u64 new_size, u32 min_type)
4152 {
4153         struct btrfs_fs_info *fs_info = root->fs_info;
4154         struct btrfs_path *path;
4155         struct extent_buffer *leaf;
4156         struct btrfs_file_extent_item *fi;
4157         struct btrfs_key key;
4158         struct btrfs_key found_key;
4159         u64 extent_start = 0;
4160         u64 extent_num_bytes = 0;
4161         u64 extent_offset = 0;
4162         u64 item_end = 0;
4163         u64 last_size = new_size;
4164         u32 found_type = (u8)-1;
4165         int found_extent;
4166         int del_item;
4167         int pending_del_nr = 0;
4168         int pending_del_slot = 0;
4169         int extent_type = -1;
4170         int ret;
4171         u64 ino = btrfs_ino(BTRFS_I(inode));
4172         u64 bytes_deleted = 0;
4173         bool be_nice = false;
4174         bool should_throttle = false;
4175         const u64 lock_start = ALIGN_DOWN(new_size, fs_info->sectorsize);
4176         struct extent_state *cached_state = NULL;
4177
4178         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4179
4180         /*
4181          * For non-free space inodes and non-shareable roots, we want to back
4182          * off from time to time.  This means all inodes in subvolume roots,
4183          * reloc roots, and data reloc roots.
4184          */
4185         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4186             test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
4187                 be_nice = true;
4188
4189         path = btrfs_alloc_path();
4190         if (!path)
4191                 return -ENOMEM;
4192         path->reada = READA_BACK;
4193
4194         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4195                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, (u64)-1,
4196                                  &cached_state);
4197
4198                 /*
4199                  * We want to drop from the next block forward in case this
4200                  * new size is not block aligned since we will be keeping the
4201                  * last block of the extent just the way it is.
4202                  */
4203                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4204                                         fs_info->sectorsize),
4205                                         (u64)-1, 0);
4206         }
4207
4208         /*
4209          * This function is also used to drop the items in the log tree before
4210          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4211          * it is used to drop the logged items. So we shouldn't kill the delayed
4212          * items.
4213          */
4214         if (min_type == 0 && root == BTRFS_I(inode)->root)
4215                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4216
4217         key.objectid = ino;
4218         key.offset = (u64)-1;
4219         key.type = (u8)-1;
4220
4221 search_again:
4222         /*
4223          * with a 16K leaf size and 128MB extents, you can actually queue
4224          * up a huge file in a single leaf.  Most of the time that
4225          * bytes_deleted is > 0, it will be huge by the time we get here
4226          */
4227         if (be_nice && bytes_deleted > SZ_32M &&
4228             btrfs_should_end_transaction(trans)) {
4229                 ret = -EAGAIN;
4230                 goto out;
4231         }
4232
4233         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4234         if (ret < 0)
4235                 goto out;
4236
4237         if (ret > 0) {
4238                 ret = 0;
4239                 /* there are no items in the tree for us to truncate, we're
4240                  * done
4241                  */
4242                 if (path->slots[0] == 0)
4243                         goto out;
4244                 path->slots[0]--;
4245         }
4246
4247         while (1) {
4248                 u64 clear_start = 0, clear_len = 0;
4249
4250                 fi = NULL;
4251                 leaf = path->nodes[0];
4252                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4253                 found_type = found_key.type;
4254
4255                 if (found_key.objectid != ino)
4256                         break;
4257
4258                 if (found_type < min_type)
4259                         break;
4260
4261                 item_end = found_key.offset;
4262                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4263                         fi = btrfs_item_ptr(leaf, path->slots[0],
4264                                             struct btrfs_file_extent_item);
4265                         extent_type = btrfs_file_extent_type(leaf, fi);
4266                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4267                                 item_end +=
4268                                     btrfs_file_extent_num_bytes(leaf, fi);
4269
4270                                 trace_btrfs_truncate_show_fi_regular(
4271                                         BTRFS_I(inode), leaf, fi,
4272                                         found_key.offset);
4273                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4274                                 item_end += btrfs_file_extent_ram_bytes(leaf,
4275                                                                         fi);
4276
4277                                 trace_btrfs_truncate_show_fi_inline(
4278                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4279                                         found_key.offset);
4280                         }
4281                         item_end--;
4282                 }
4283                 if (found_type > min_type) {
4284                         del_item = 1;
4285                 } else {
4286                         if (item_end < new_size)
4287                                 break;
4288                         if (found_key.offset >= new_size)
4289                                 del_item = 1;
4290                         else
4291                                 del_item = 0;
4292                 }
4293                 found_extent = 0;
4294                 /* FIXME, shrink the extent if the ref count is only 1 */
4295                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4296                         goto delete;
4297
4298                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4299                         u64 num_dec;
4300
4301                         clear_start = found_key.offset;
4302                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4303                         if (!del_item) {
4304                                 u64 orig_num_bytes =
4305                                         btrfs_file_extent_num_bytes(leaf, fi);
4306                                 extent_num_bytes = ALIGN(new_size -
4307                                                 found_key.offset,
4308                                                 fs_info->sectorsize);
4309                                 clear_start = ALIGN(new_size, fs_info->sectorsize);
4310                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4311                                                          extent_num_bytes);
4312                                 num_dec = (orig_num_bytes -
4313                                            extent_num_bytes);
4314                                 if (test_bit(BTRFS_ROOT_SHAREABLE,
4315                                              &root->state) &&
4316                                     extent_start != 0)
4317                                         inode_sub_bytes(inode, num_dec);
4318                                 btrfs_mark_buffer_dirty(leaf);
4319                         } else {
4320                                 extent_num_bytes =
4321                                         btrfs_file_extent_disk_num_bytes(leaf,
4322                                                                          fi);
4323                                 extent_offset = found_key.offset -
4324                                         btrfs_file_extent_offset(leaf, fi);
4325
4326                                 /* FIXME blocksize != 4096 */
4327                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4328                                 if (extent_start != 0) {
4329                                         found_extent = 1;
4330                                         if (test_bit(BTRFS_ROOT_SHAREABLE,
4331                                                      &root->state))
4332                                                 inode_sub_bytes(inode, num_dec);
4333                                 }
4334                         }
4335                         clear_len = num_dec;
4336                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4337                         /*
4338                          * we can't truncate inline items that have had
4339                          * special encodings
4340                          */
4341                         if (!del_item &&
4342                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4343                             btrfs_file_extent_other_encoding(leaf, fi) == 0 &&
4344                             btrfs_file_extent_compression(leaf, fi) == 0) {
4345                                 u32 size = (u32)(new_size - found_key.offset);
4346
4347                                 btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4348                                 size = btrfs_file_extent_calc_inline_size(size);
4349                                 btrfs_truncate_item(path, size, 1);
4350                         } else if (!del_item) {
4351                                 /*
4352                                  * We have to bail so the last_size is set to
4353                                  * just before this extent.
4354                                  */
4355                                 ret = NEED_TRUNCATE_BLOCK;
4356                                 break;
4357                         } else {
4358                                 /*
4359                                  * Inline extents are special, we just treat
4360                                  * them as a full sector worth in the file
4361                                  * extent tree just for simplicity sake.
4362                                  */
4363                                 clear_len = fs_info->sectorsize;
4364                         }
4365
4366                         if (test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
4367                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4368                 }
4369 delete:
4370                 /*
4371                  * We use btrfs_truncate_inode_items() to clean up log trees for
4372                  * multiple fsyncs, and in this case we don't want to clear the
4373                  * file extent range because it's just the log.
4374                  */
4375                 if (root == BTRFS_I(inode)->root) {
4376                         ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
4377                                                   clear_start, clear_len);
4378                         if (ret) {
4379                                 btrfs_abort_transaction(trans, ret);
4380                                 break;
4381                         }
4382                 }
4383
4384                 if (del_item)
4385                         last_size = found_key.offset;
4386                 else
4387                         last_size = new_size;
4388                 if (del_item) {
4389                         if (!pending_del_nr) {
4390                                 /* no pending yet, add ourselves */
4391                                 pending_del_slot = path->slots[0];
4392                                 pending_del_nr = 1;
4393                         } else if (pending_del_nr &&
4394                                    path->slots[0] + 1 == pending_del_slot) {
4395                                 /* hop on the pending chunk */
4396                                 pending_del_nr++;
4397                                 pending_del_slot = path->slots[0];
4398                         } else {
4399                                 BUG();
4400                         }
4401                 } else {
4402                         break;
4403                 }
4404                 should_throttle = false;
4405
4406                 if (found_extent &&
4407                     root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4408                         struct btrfs_ref ref = { 0 };
4409
4410                         bytes_deleted += extent_num_bytes;
4411
4412                         btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF,
4413                                         extent_start, extent_num_bytes, 0);
4414                         ref.real_root = root->root_key.objectid;
4415                         btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
4416                                         ino, extent_offset);
4417                         ret = btrfs_free_extent(trans, &ref);
4418                         if (ret) {
4419                                 btrfs_abort_transaction(trans, ret);
4420                                 break;
4421                         }
4422                         if (be_nice) {
4423                                 if (btrfs_should_throttle_delayed_refs(trans))
4424                                         should_throttle = true;
4425                         }
4426                 }
4427
4428                 if (found_type == BTRFS_INODE_ITEM_KEY)
4429                         break;
4430
4431                 if (path->slots[0] == 0 ||
4432                     path->slots[0] != pending_del_slot ||
4433                     should_throttle) {
4434                         if (pending_del_nr) {
4435                                 ret = btrfs_del_items(trans, root, path,
4436                                                 pending_del_slot,
4437                                                 pending_del_nr);
4438                                 if (ret) {
4439                                         btrfs_abort_transaction(trans, ret);
4440                                         break;
4441                                 }
4442                                 pending_del_nr = 0;
4443                         }
4444                         btrfs_release_path(path);
4445
4446                         /*
4447                          * We can generate a lot of delayed refs, so we need to
4448                          * throttle every once and a while and make sure we're
4449                          * adding enough space to keep up with the work we are
4450                          * generating.  Since we hold a transaction here we
4451                          * can't flush, and we don't want to FLUSH_LIMIT because
4452                          * we could have generated too many delayed refs to
4453                          * actually allocate, so just bail if we're short and
4454                          * let the normal reservation dance happen higher up.
4455                          */
4456                         if (should_throttle) {
4457                                 ret = btrfs_delayed_refs_rsv_refill(fs_info,
4458                                                         BTRFS_RESERVE_NO_FLUSH);
4459                                 if (ret) {
4460                                         ret = -EAGAIN;
4461                                         break;
4462                                 }
4463                         }
4464                         goto search_again;
4465                 } else {
4466                         path->slots[0]--;
4467                 }
4468         }
4469 out:
4470         if (ret >= 0 && pending_del_nr) {
4471                 int err;
4472
4473                 err = btrfs_del_items(trans, root, path, pending_del_slot,
4474                                       pending_del_nr);
4475                 if (err) {
4476                         btrfs_abort_transaction(trans, err);
4477                         ret = err;
4478                 }
4479         }
4480         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4481                 ASSERT(last_size >= new_size);
4482                 if (!ret && last_size > new_size)
4483                         last_size = new_size;
4484                 btrfs_inode_safe_disk_i_size_write(inode, last_size);
4485                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start,
4486                                      (u64)-1, &cached_state);
4487         }
4488
4489         btrfs_free_path(path);
4490         return ret;
4491 }
4492
4493 /*
4494  * btrfs_truncate_block - read, zero a chunk and write a block
4495  * @inode - inode that we're zeroing
4496  * @from - the offset to start zeroing
4497  * @len - the length to zero, 0 to zero the entire range respective to the
4498  *      offset
4499  * @front - zero up to the offset instead of from the offset on
4500  *
4501  * This will find the block for the "from" offset and cow the block and zero the
4502  * part we want to zero.  This is used with truncate and hole punching.
4503  */
4504 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4505                         int front)
4506 {
4507         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4508         struct address_space *mapping = inode->i_mapping;
4509         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4510         struct btrfs_ordered_extent *ordered;
4511         struct extent_state *cached_state = NULL;
4512         struct extent_changeset *data_reserved = NULL;
4513         char *kaddr;
4514         u32 blocksize = fs_info->sectorsize;
4515         pgoff_t index = from >> PAGE_SHIFT;
4516         unsigned offset = from & (blocksize - 1);
4517         struct page *page;
4518         gfp_t mask = btrfs_alloc_write_mask(mapping);
4519         int ret = 0;
4520         u64 block_start;
4521         u64 block_end;
4522
4523         if (IS_ALIGNED(offset, blocksize) &&
4524             (!len || IS_ALIGNED(len, blocksize)))
4525                 goto out;
4526
4527         block_start = round_down(from, blocksize);
4528         block_end = block_start + blocksize - 1;
4529
4530         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4531                                            block_start, blocksize);
4532         if (ret)
4533                 goto out;
4534
4535 again:
4536         page = find_or_create_page(mapping, index, mask);
4537         if (!page) {
4538                 btrfs_delalloc_release_space(inode, data_reserved,
4539                                              block_start, blocksize, true);
4540                 btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4541                 ret = -ENOMEM;
4542                 goto out;
4543         }
4544
4545         if (!PageUptodate(page)) {
4546                 ret = btrfs_readpage(NULL, page);
4547                 lock_page(page);
4548                 if (page->mapping != mapping) {
4549                         unlock_page(page);
4550                         put_page(page);
4551                         goto again;
4552                 }
4553                 if (!PageUptodate(page)) {
4554                         ret = -EIO;
4555                         goto out_unlock;
4556                 }
4557         }
4558         wait_on_page_writeback(page);
4559
4560         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4561         set_page_extent_mapped(page);
4562
4563         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4564         if (ordered) {
4565                 unlock_extent_cached(io_tree, block_start, block_end,
4566                                      &cached_state);
4567                 unlock_page(page);
4568                 put_page(page);
4569                 btrfs_start_ordered_extent(inode, ordered, 1);
4570                 btrfs_put_ordered_extent(ordered);
4571                 goto again;
4572         }
4573
4574         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4575                          EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4576                          0, 0, &cached_state);
4577
4578         ret = btrfs_set_extent_delalloc(inode, block_start, block_end, 0,
4579                                         &cached_state);
4580         if (ret) {
4581                 unlock_extent_cached(io_tree, block_start, block_end,
4582                                      &cached_state);
4583                 goto out_unlock;
4584         }
4585
4586         if (offset != blocksize) {
4587                 if (!len)
4588                         len = blocksize - offset;
4589                 kaddr = kmap(page);
4590                 if (front)
4591                         memset(kaddr + (block_start - page_offset(page)),
4592                                 0, offset);
4593                 else
4594                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4595                                 0, len);
4596                 flush_dcache_page(page);
4597                 kunmap(page);
4598         }
4599         ClearPageChecked(page);
4600         set_page_dirty(page);
4601         unlock_extent_cached(io_tree, block_start, block_end, &cached_state);
4602
4603 out_unlock:
4604         if (ret)
4605                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
4606                                              blocksize, true);
4607         btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4608         unlock_page(page);
4609         put_page(page);
4610 out:
4611         extent_changeset_free(data_reserved);
4612         return ret;
4613 }
4614
4615 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4616                              u64 offset, u64 len)
4617 {
4618         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4619         struct btrfs_trans_handle *trans;
4620         int ret;
4621
4622         /*
4623          * Still need to make sure the inode looks like it's been updated so
4624          * that any holes get logged if we fsync.
4625          */
4626         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
4627                 BTRFS_I(inode)->last_trans = fs_info->generation;
4628                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4629                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4630                 return 0;
4631         }
4632
4633         /*
4634          * 1 - for the one we're dropping
4635          * 1 - for the one we're adding
4636          * 1 - for updating the inode.
4637          */
4638         trans = btrfs_start_transaction(root, 3);
4639         if (IS_ERR(trans))
4640                 return PTR_ERR(trans);
4641
4642         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4643         if (ret) {
4644                 btrfs_abort_transaction(trans, ret);
4645                 btrfs_end_transaction(trans);
4646                 return ret;
4647         }
4648
4649         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
4650                         offset, 0, 0, len, 0, len, 0, 0, 0);
4651         if (ret)
4652                 btrfs_abort_transaction(trans, ret);
4653         else
4654                 btrfs_update_inode(trans, root, inode);
4655         btrfs_end_transaction(trans);
4656         return ret;
4657 }
4658
4659 /*
4660  * This function puts in dummy file extents for the area we're creating a hole
4661  * for.  So if we are truncating this file to a larger size we need to insert
4662  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4663  * the range between oldsize and size
4664  */
4665 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4666 {
4667         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4668         struct btrfs_root *root = BTRFS_I(inode)->root;
4669         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4670         struct extent_map *em = NULL;
4671         struct extent_state *cached_state = NULL;
4672         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4673         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
4674         u64 block_end = ALIGN(size, fs_info->sectorsize);
4675         u64 last_byte;
4676         u64 cur_offset;
4677         u64 hole_size;
4678         int err = 0;
4679
4680         /*
4681          * If our size started in the middle of a block we need to zero out the
4682          * rest of the block before we expand the i_size, otherwise we could
4683          * expose stale data.
4684          */
4685         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4686         if (err)
4687                 return err;
4688
4689         if (size <= hole_start)
4690                 return 0;
4691
4692         btrfs_lock_and_flush_ordered_range(BTRFS_I(inode), hole_start,
4693                                            block_end - 1, &cached_state);
4694         cur_offset = hole_start;
4695         while (1) {
4696                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
4697                                       block_end - cur_offset);
4698                 if (IS_ERR(em)) {
4699                         err = PTR_ERR(em);
4700                         em = NULL;
4701                         break;
4702                 }
4703                 last_byte = min(extent_map_end(em), block_end);
4704                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
4705                 hole_size = last_byte - cur_offset;
4706
4707                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4708                         struct extent_map *hole_em;
4709
4710                         err = maybe_insert_hole(root, inode, cur_offset,
4711                                                 hole_size);
4712                         if (err)
4713                                 break;
4714
4715                         err = btrfs_inode_set_file_extent_range(BTRFS_I(inode),
4716                                                         cur_offset, hole_size);
4717                         if (err)
4718                                 break;
4719
4720                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
4721                                                 cur_offset + hole_size - 1, 0);
4722                         hole_em = alloc_extent_map();
4723                         if (!hole_em) {
4724                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4725                                         &BTRFS_I(inode)->runtime_flags);
4726                                 goto next;
4727                         }
4728                         hole_em->start = cur_offset;
4729                         hole_em->len = hole_size;
4730                         hole_em->orig_start = cur_offset;
4731
4732                         hole_em->block_start = EXTENT_MAP_HOLE;
4733                         hole_em->block_len = 0;
4734                         hole_em->orig_block_len = 0;
4735                         hole_em->ram_bytes = hole_size;
4736                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4737                         hole_em->generation = fs_info->generation;
4738
4739                         while (1) {
4740                                 write_lock(&em_tree->lock);
4741                                 err = add_extent_mapping(em_tree, hole_em, 1);
4742                                 write_unlock(&em_tree->lock);
4743                                 if (err != -EEXIST)
4744                                         break;
4745                                 btrfs_drop_extent_cache(BTRFS_I(inode),
4746                                                         cur_offset,
4747                                                         cur_offset +
4748                                                         hole_size - 1, 0);
4749                         }
4750                         free_extent_map(hole_em);
4751                 } else {
4752                         err = btrfs_inode_set_file_extent_range(BTRFS_I(inode),
4753                                                         cur_offset, hole_size);
4754                         if (err)
4755                                 break;
4756                 }
4757 next:
4758                 free_extent_map(em);
4759                 em = NULL;
4760                 cur_offset = last_byte;
4761                 if (cur_offset >= block_end)
4762                         break;
4763         }
4764         free_extent_map(em);
4765         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state);
4766         return err;
4767 }
4768
4769 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4770 {
4771         struct btrfs_root *root = BTRFS_I(inode)->root;
4772         struct btrfs_trans_handle *trans;
4773         loff_t oldsize = i_size_read(inode);
4774         loff_t newsize = attr->ia_size;
4775         int mask = attr->ia_valid;
4776         int ret;
4777
4778         /*
4779          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4780          * special case where we need to update the times despite not having
4781          * these flags set.  For all other operations the VFS set these flags
4782          * explicitly if it wants a timestamp update.
4783          */
4784         if (newsize != oldsize) {
4785                 inode_inc_iversion(inode);
4786                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
4787                         inode->i_ctime = inode->i_mtime =
4788                                 current_time(inode);
4789         }
4790
4791         if (newsize > oldsize) {
4792                 /*
4793                  * Don't do an expanding truncate while snapshotting is ongoing.
4794                  * This is to ensure the snapshot captures a fully consistent
4795                  * state of this file - if the snapshot captures this expanding
4796                  * truncation, it must capture all writes that happened before
4797                  * this truncation.
4798                  */
4799                 btrfs_drew_write_lock(&root->snapshot_lock);
4800                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4801                 if (ret) {
4802                         btrfs_drew_write_unlock(&root->snapshot_lock);
4803                         return ret;
4804                 }
4805
4806                 trans = btrfs_start_transaction(root, 1);
4807                 if (IS_ERR(trans)) {
4808                         btrfs_drew_write_unlock(&root->snapshot_lock);
4809                         return PTR_ERR(trans);
4810                 }
4811
4812                 i_size_write(inode, newsize);
4813                 btrfs_inode_safe_disk_i_size_write(inode, 0);
4814                 pagecache_isize_extended(inode, oldsize, newsize);
4815                 ret = btrfs_update_inode(trans, root, inode);
4816                 btrfs_drew_write_unlock(&root->snapshot_lock);
4817                 btrfs_end_transaction(trans);
4818         } else {
4819
4820                 /*
4821                  * We're truncating a file that used to have good data down to
4822                  * zero. Make sure it gets into the ordered flush list so that
4823                  * any new writes get down to disk quickly.
4824                  */
4825                 if (newsize == 0)
4826                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
4827                                 &BTRFS_I(inode)->runtime_flags);
4828
4829                 truncate_setsize(inode, newsize);
4830
4831                 /* Disable nonlocked read DIO to avoid the endless truncate */
4832                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
4833                 inode_dio_wait(inode);
4834                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
4835
4836                 ret = btrfs_truncate(inode, newsize == oldsize);
4837                 if (ret && inode->i_nlink) {
4838                         int err;
4839
4840                         /*
4841                          * Truncate failed, so fix up the in-memory size. We
4842                          * adjusted disk_i_size down as we removed extents, so
4843                          * wait for disk_i_size to be stable and then update the
4844                          * in-memory size to match.
4845                          */
4846                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
4847                         if (err)
4848                                 return err;
4849                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
4850                 }
4851         }
4852
4853         return ret;
4854 }
4855
4856 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
4857 {
4858         struct inode *inode = d_inode(dentry);
4859         struct btrfs_root *root = BTRFS_I(inode)->root;
4860         int err;
4861
4862         if (btrfs_root_readonly(root))
4863                 return -EROFS;
4864
4865         err = setattr_prepare(dentry, attr);
4866         if (err)
4867                 return err;
4868
4869         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
4870                 err = btrfs_setsize(inode, attr);
4871                 if (err)
4872                         return err;
4873         }
4874
4875         if (attr->ia_valid) {
4876                 setattr_copy(inode, attr);
4877                 inode_inc_iversion(inode);
4878                 err = btrfs_dirty_inode(inode);
4879
4880                 if (!err && attr->ia_valid & ATTR_MODE)
4881                         err = posix_acl_chmod(inode, inode->i_mode);
4882         }
4883
4884         return err;
4885 }
4886
4887 /*
4888  * While truncating the inode pages during eviction, we get the VFS calling
4889  * btrfs_invalidatepage() against each page of the inode. This is slow because
4890  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
4891  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
4892  * extent_state structures over and over, wasting lots of time.
4893  *
4894  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
4895  * those expensive operations on a per page basis and do only the ordered io
4896  * finishing, while we release here the extent_map and extent_state structures,
4897  * without the excessive merging and splitting.
4898  */
4899 static void evict_inode_truncate_pages(struct inode *inode)
4900 {
4901         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4902         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
4903         struct rb_node *node;
4904
4905         ASSERT(inode->i_state & I_FREEING);
4906         truncate_inode_pages_final(&inode->i_data);
4907
4908         write_lock(&map_tree->lock);
4909         while (!RB_EMPTY_ROOT(&map_tree->map.rb_root)) {
4910                 struct extent_map *em;
4911
4912                 node = rb_first_cached(&map_tree->map);
4913                 em = rb_entry(node, struct extent_map, rb_node);
4914                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
4915                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
4916                 remove_extent_mapping(map_tree, em);
4917                 free_extent_map(em);
4918                 if (need_resched()) {
4919                         write_unlock(&map_tree->lock);
4920                         cond_resched();
4921                         write_lock(&map_tree->lock);
4922                 }
4923         }
4924         write_unlock(&map_tree->lock);
4925
4926         /*
4927          * Keep looping until we have no more ranges in the io tree.
4928          * We can have ongoing bios started by readahead that have
4929          * their endio callback (extent_io.c:end_bio_extent_readpage)
4930          * still in progress (unlocked the pages in the bio but did not yet
4931          * unlocked the ranges in the io tree). Therefore this means some
4932          * ranges can still be locked and eviction started because before
4933          * submitting those bios, which are executed by a separate task (work
4934          * queue kthread), inode references (inode->i_count) were not taken
4935          * (which would be dropped in the end io callback of each bio).
4936          * Therefore here we effectively end up waiting for those bios and
4937          * anyone else holding locked ranges without having bumped the inode's
4938          * reference count - if we don't do it, when they access the inode's
4939          * io_tree to unlock a range it may be too late, leading to an
4940          * use-after-free issue.
4941          */
4942         spin_lock(&io_tree->lock);
4943         while (!RB_EMPTY_ROOT(&io_tree->state)) {
4944                 struct extent_state *state;
4945                 struct extent_state *cached_state = NULL;
4946                 u64 start;
4947                 u64 end;
4948                 unsigned state_flags;
4949
4950                 node = rb_first(&io_tree->state);
4951                 state = rb_entry(node, struct extent_state, rb_node);
4952                 start = state->start;
4953                 end = state->end;
4954                 state_flags = state->state;
4955                 spin_unlock(&io_tree->lock);
4956
4957                 lock_extent_bits(io_tree, start, end, &cached_state);
4958
4959                 /*
4960                  * If still has DELALLOC flag, the extent didn't reach disk,
4961                  * and its reserved space won't be freed by delayed_ref.
4962                  * So we need to free its reserved space here.
4963                  * (Refer to comment in btrfs_invalidatepage, case 2)
4964                  *
4965                  * Note, end is the bytenr of last byte, so we need + 1 here.
4966                  */
4967                 if (state_flags & EXTENT_DELALLOC)
4968                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
4969
4970                 clear_extent_bit(io_tree, start, end,
4971                                  EXTENT_LOCKED | EXTENT_DELALLOC |
4972                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
4973                                  &cached_state);
4974
4975                 cond_resched();
4976                 spin_lock(&io_tree->lock);
4977         }
4978         spin_unlock(&io_tree->lock);
4979 }
4980
4981 static struct btrfs_trans_handle *evict_refill_and_join(struct btrfs_root *root,
4982                                                         struct btrfs_block_rsv *rsv)
4983 {
4984         struct btrfs_fs_info *fs_info = root->fs_info;
4985         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4986         struct btrfs_trans_handle *trans;
4987         u64 delayed_refs_extra = btrfs_calc_insert_metadata_size(fs_info, 1);
4988         int ret;
4989
4990         /*
4991          * Eviction should be taking place at some place safe because of our
4992          * delayed iputs.  However the normal flushing code will run delayed
4993          * iputs, so we cannot use FLUSH_ALL otherwise we'll deadlock.
4994          *
4995          * We reserve the delayed_refs_extra here again because we can't use
4996          * btrfs_start_transaction(root, 0) for the same deadlocky reason as
4997          * above.  We reserve our extra bit here because we generate a ton of
4998          * delayed refs activity by truncating.
4999          *
5000          * If we cannot make our reservation we'll attempt to steal from the
5001          * global reserve, because we really want to be able to free up space.
5002          */
5003         ret = btrfs_block_rsv_refill(root, rsv, rsv->size + delayed_refs_extra,
5004                                      BTRFS_RESERVE_FLUSH_EVICT);
5005         if (ret) {
5006                 /*
5007                  * Try to steal from the global reserve if there is space for
5008                  * it.
5009                  */
5010                 if (btrfs_check_space_for_delayed_refs(fs_info) ||
5011                     btrfs_block_rsv_migrate(global_rsv, rsv, rsv->size, 0)) {
5012                         btrfs_warn(fs_info,
5013                                    "could not allocate space for delete; will truncate on mount");
5014                         return ERR_PTR(-ENOSPC);
5015                 }
5016                 delayed_refs_extra = 0;
5017         }
5018
5019         trans = btrfs_join_transaction(root);
5020         if (IS_ERR(trans))
5021                 return trans;
5022
5023         if (delayed_refs_extra) {
5024                 trans->block_rsv = &fs_info->trans_block_rsv;
5025                 trans->bytes_reserved = delayed_refs_extra;
5026                 btrfs_block_rsv_migrate(rsv, trans->block_rsv,
5027                                         delayed_refs_extra, 1);
5028         }
5029         return trans;
5030 }
5031
5032 void btrfs_evict_inode(struct inode *inode)
5033 {
5034         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5035         struct btrfs_trans_handle *trans;
5036         struct btrfs_root *root = BTRFS_I(inode)->root;
5037         struct btrfs_block_rsv *rsv;
5038         int ret;
5039
5040         trace_btrfs_inode_evict(inode);
5041
5042         if (!root) {
5043                 clear_inode(inode);
5044                 return;
5045         }
5046
5047         evict_inode_truncate_pages(inode);
5048
5049         if (inode->i_nlink &&
5050             ((btrfs_root_refs(&root->root_item) != 0 &&
5051               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5052              btrfs_is_free_space_inode(BTRFS_I(inode))))
5053                 goto no_delete;
5054
5055         if (is_bad_inode(inode))
5056                 goto no_delete;
5057
5058         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5059
5060         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
5061                 goto no_delete;
5062
5063         if (inode->i_nlink > 0) {
5064                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5065                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5066                 goto no_delete;
5067         }
5068
5069         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5070         if (ret)
5071                 goto no_delete;
5072
5073         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5074         if (!rsv)
5075                 goto no_delete;
5076         rsv->size = btrfs_calc_metadata_size(fs_info, 1);
5077         rsv->failfast = 1;
5078
5079         btrfs_i_size_write(BTRFS_I(inode), 0);
5080
5081         while (1) {
5082                 trans = evict_refill_and_join(root, rsv);
5083                 if (IS_ERR(trans))
5084                         goto free_rsv;
5085
5086                 trans->block_rsv = rsv;
5087
5088                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5089                 trans->block_rsv = &fs_info->trans_block_rsv;
5090                 btrfs_end_transaction(trans);
5091                 btrfs_btree_balance_dirty(fs_info);
5092                 if (ret && ret != -ENOSPC && ret != -EAGAIN)
5093                         goto free_rsv;
5094                 else if (!ret)
5095                         break;
5096         }
5097
5098         /*
5099          * Errors here aren't a big deal, it just means we leave orphan items in
5100          * the tree. They will be cleaned up on the next mount. If the inode
5101          * number gets reused, cleanup deletes the orphan item without doing
5102          * anything, and unlink reuses the existing orphan item.
5103          *
5104          * If it turns out that we are dropping too many of these, we might want
5105          * to add a mechanism for retrying these after a commit.
5106          */
5107         trans = evict_refill_and_join(root, rsv);
5108         if (!IS_ERR(trans)) {
5109                 trans->block_rsv = rsv;
5110                 btrfs_orphan_del(trans, BTRFS_I(inode));
5111                 trans->block_rsv = &fs_info->trans_block_rsv;
5112                 btrfs_end_transaction(trans);
5113         }
5114
5115         if (!(root == fs_info->tree_root ||
5116               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5117                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5118
5119 free_rsv:
5120         btrfs_free_block_rsv(fs_info, rsv);
5121 no_delete:
5122         /*
5123          * If we didn't successfully delete, the orphan item will still be in
5124          * the tree and we'll retry on the next mount. Again, we might also want
5125          * to retry these periodically in the future.
5126          */
5127         btrfs_remove_delayed_node(BTRFS_I(inode));
5128         clear_inode(inode);
5129 }
5130
5131 /*
5132  * Return the key found in the dir entry in the location pointer, fill @type
5133  * with BTRFS_FT_*, and return 0.
5134  *
5135  * If no dir entries were found, returns -ENOENT.
5136  * If found a corrupted location in dir entry, returns -EUCLEAN.
5137  */
5138 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5139                                struct btrfs_key *location, u8 *type)
5140 {
5141         const char *name = dentry->d_name.name;
5142         int namelen = dentry->d_name.len;
5143         struct btrfs_dir_item *di;
5144         struct btrfs_path *path;
5145         struct btrfs_root *root = BTRFS_I(dir)->root;
5146         int ret = 0;
5147
5148         path = btrfs_alloc_path();
5149         if (!path)
5150                 return -ENOMEM;
5151
5152         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5153                         name, namelen, 0);
5154         if (IS_ERR_OR_NULL(di)) {
5155                 ret = di ? PTR_ERR(di) : -ENOENT;
5156                 goto out;
5157         }
5158
5159         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5160         if (location->type != BTRFS_INODE_ITEM_KEY &&
5161             location->type != BTRFS_ROOT_ITEM_KEY) {
5162                 ret = -EUCLEAN;
5163                 btrfs_warn(root->fs_info,
5164 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5165                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5166                            location->objectid, location->type, location->offset);
5167         }
5168         if (!ret)
5169                 *type = btrfs_dir_type(path->nodes[0], di);
5170 out:
5171         btrfs_free_path(path);
5172         return ret;
5173 }
5174
5175 /*
5176  * when we hit a tree root in a directory, the btrfs part of the inode
5177  * needs to be changed to reflect the root directory of the tree root.  This
5178  * is kind of like crossing a mount point.
5179  */
5180 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5181                                     struct inode *dir,
5182                                     struct dentry *dentry,
5183                                     struct btrfs_key *location,
5184                                     struct btrfs_root **sub_root)
5185 {
5186         struct btrfs_path *path;
5187         struct btrfs_root *new_root;
5188         struct btrfs_root_ref *ref;
5189         struct extent_buffer *leaf;
5190         struct btrfs_key key;
5191         int ret;
5192         int err = 0;
5193
5194         path = btrfs_alloc_path();
5195         if (!path) {
5196                 err = -ENOMEM;
5197                 goto out;
5198         }
5199
5200         err = -ENOENT;
5201         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5202         key.type = BTRFS_ROOT_REF_KEY;
5203         key.offset = location->objectid;
5204
5205         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5206         if (ret) {
5207                 if (ret < 0)
5208                         err = ret;
5209                 goto out;
5210         }
5211
5212         leaf = path->nodes[0];
5213         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5214         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5215             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5216                 goto out;
5217
5218         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5219                                    (unsigned long)(ref + 1),
5220                                    dentry->d_name.len);
5221         if (ret)
5222                 goto out;
5223
5224         btrfs_release_path(path);
5225
5226         new_root = btrfs_get_fs_root(fs_info, location->objectid, true);
5227         if (IS_ERR(new_root)) {
5228                 err = PTR_ERR(new_root);
5229                 goto out;
5230         }
5231
5232         *sub_root = new_root;
5233         location->objectid = btrfs_root_dirid(&new_root->root_item);
5234         location->type = BTRFS_INODE_ITEM_KEY;
5235         location->offset = 0;
5236         err = 0;
5237 out:
5238         btrfs_free_path(path);
5239         return err;
5240 }
5241
5242 static void inode_tree_add(struct inode *inode)
5243 {
5244         struct btrfs_root *root = BTRFS_I(inode)->root;
5245         struct btrfs_inode *entry;
5246         struct rb_node **p;
5247         struct rb_node *parent;
5248         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5249         u64 ino = btrfs_ino(BTRFS_I(inode));
5250
5251         if (inode_unhashed(inode))
5252                 return;
5253         parent = NULL;
5254         spin_lock(&root->inode_lock);
5255         p = &root->inode_tree.rb_node;
5256         while (*p) {
5257                 parent = *p;
5258                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5259
5260                 if (ino < btrfs_ino(entry))
5261                         p = &parent->rb_left;
5262                 else if (ino > btrfs_ino(entry))
5263                         p = &parent->rb_right;
5264                 else {
5265                         WARN_ON(!(entry->vfs_inode.i_state &
5266                                   (I_WILL_FREE | I_FREEING)));
5267                         rb_replace_node(parent, new, &root->inode_tree);
5268                         RB_CLEAR_NODE(parent);
5269                         spin_unlock(&root->inode_lock);
5270                         return;
5271                 }
5272         }
5273         rb_link_node(new, parent, p);
5274         rb_insert_color(new, &root->inode_tree);
5275         spin_unlock(&root->inode_lock);
5276 }
5277
5278 static void inode_tree_del(struct inode *inode)
5279 {
5280         struct btrfs_root *root = BTRFS_I(inode)->root;
5281         int empty = 0;
5282
5283         spin_lock(&root->inode_lock);
5284         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5285                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5286                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5287                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5288         }
5289         spin_unlock(&root->inode_lock);
5290
5291         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5292                 spin_lock(&root->inode_lock);
5293                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5294                 spin_unlock(&root->inode_lock);
5295                 if (empty)
5296                         btrfs_add_dead_root(root);
5297         }
5298 }
5299
5300
5301 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5302 {
5303         struct btrfs_iget_args *args = p;
5304
5305         inode->i_ino = args->ino;
5306         BTRFS_I(inode)->location.objectid = args->ino;
5307         BTRFS_I(inode)->location.type = BTRFS_INODE_ITEM_KEY;
5308         BTRFS_I(inode)->location.offset = 0;
5309         BTRFS_I(inode)->root = btrfs_grab_root(args->root);
5310         BUG_ON(args->root && !BTRFS_I(inode)->root);
5311         return 0;
5312 }
5313
5314 static int btrfs_find_actor(struct inode *inode, void *opaque)
5315 {
5316         struct btrfs_iget_args *args = opaque;
5317
5318         return args->ino == BTRFS_I(inode)->location.objectid &&
5319                 args->root == BTRFS_I(inode)->root;
5320 }
5321
5322 static struct inode *btrfs_iget_locked(struct super_block *s, u64 ino,
5323                                        struct btrfs_root *root)
5324 {
5325         struct inode *inode;
5326         struct btrfs_iget_args args;
5327         unsigned long hashval = btrfs_inode_hash(ino, root);
5328
5329         args.ino = ino;
5330         args.root = root;
5331
5332         inode = iget5_locked(s, hashval, btrfs_find_actor,
5333                              btrfs_init_locked_inode,
5334                              (void *)&args);
5335         return inode;
5336 }
5337
5338 /*
5339  * Get an inode object given its inode number and corresponding root.
5340  * Path can be preallocated to prevent recursing back to iget through
5341  * allocator. NULL is also valid but may require an additional allocation
5342  * later.
5343  */
5344 struct inode *btrfs_iget_path(struct super_block *s, u64 ino,
5345                               struct btrfs_root *root, struct btrfs_path *path)
5346 {
5347         struct inode *inode;
5348
5349         inode = btrfs_iget_locked(s, ino, root);
5350         if (!inode)
5351                 return ERR_PTR(-ENOMEM);
5352
5353         if (inode->i_state & I_NEW) {
5354                 int ret;
5355
5356                 ret = btrfs_read_locked_inode(inode, path);
5357                 if (!ret) {
5358                         inode_tree_add(inode);
5359                         unlock_new_inode(inode);
5360                 } else {
5361                         iget_failed(inode);
5362                         /*
5363                          * ret > 0 can come from btrfs_search_slot called by
5364                          * btrfs_read_locked_inode, this means the inode item
5365                          * was not found.
5366                          */
5367                         if (ret > 0)
5368                                 ret = -ENOENT;
5369                         inode = ERR_PTR(ret);
5370                 }
5371         }
5372
5373         return inode;
5374 }
5375
5376 struct inode *btrfs_iget(struct super_block *s, u64 ino, struct btrfs_root *root)
5377 {
5378         return btrfs_iget_path(s, ino, root, NULL);
5379 }
5380
5381 static struct inode *new_simple_dir(struct super_block *s,
5382                                     struct btrfs_key *key,
5383                                     struct btrfs_root *root)
5384 {
5385         struct inode *inode = new_inode(s);
5386
5387         if (!inode)
5388                 return ERR_PTR(-ENOMEM);
5389
5390         BTRFS_I(inode)->root = btrfs_grab_root(root);
5391         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5392         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5393
5394         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5395         /*
5396          * We only need lookup, the rest is read-only and there's no inode
5397          * associated with the dentry
5398          */
5399         inode->i_op = &simple_dir_inode_operations;
5400         inode->i_opflags &= ~IOP_XATTR;
5401         inode->i_fop = &simple_dir_operations;
5402         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5403         inode->i_mtime = current_time(inode);
5404         inode->i_atime = inode->i_mtime;
5405         inode->i_ctime = inode->i_mtime;
5406         BTRFS_I(inode)->i_otime = inode->i_mtime;
5407
5408         return inode;
5409 }
5410
5411 static inline u8 btrfs_inode_type(struct inode *inode)
5412 {
5413         /*
5414          * Compile-time asserts that generic FT_* types still match
5415          * BTRFS_FT_* types
5416          */
5417         BUILD_BUG_ON(BTRFS_FT_UNKNOWN != FT_UNKNOWN);
5418         BUILD_BUG_ON(BTRFS_FT_REG_FILE != FT_REG_FILE);
5419         BUILD_BUG_ON(BTRFS_FT_DIR != FT_DIR);
5420         BUILD_BUG_ON(BTRFS_FT_CHRDEV != FT_CHRDEV);
5421         BUILD_BUG_ON(BTRFS_FT_BLKDEV != FT_BLKDEV);
5422         BUILD_BUG_ON(BTRFS_FT_FIFO != FT_FIFO);
5423         BUILD_BUG_ON(BTRFS_FT_SOCK != FT_SOCK);
5424         BUILD_BUG_ON(BTRFS_FT_SYMLINK != FT_SYMLINK);
5425
5426         return fs_umode_to_ftype(inode->i_mode);
5427 }
5428
5429 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5430 {
5431         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5432         struct inode *inode;
5433         struct btrfs_root *root = BTRFS_I(dir)->root;
5434         struct btrfs_root *sub_root = root;
5435         struct btrfs_key location;
5436         u8 di_type = 0;
5437         int ret = 0;
5438
5439         if (dentry->d_name.len > BTRFS_NAME_LEN)
5440                 return ERR_PTR(-ENAMETOOLONG);
5441
5442         ret = btrfs_inode_by_name(dir, dentry, &location, &di_type);
5443         if (ret < 0)
5444                 return ERR_PTR(ret);
5445
5446         if (location.type == BTRFS_INODE_ITEM_KEY) {
5447                 inode = btrfs_iget(dir->i_sb, location.objectid, root);
5448                 if (IS_ERR(inode))
5449                         return inode;
5450
5451                 /* Do extra check against inode mode with di_type */
5452                 if (btrfs_inode_type(inode) != di_type) {
5453                         btrfs_crit(fs_info,
5454 "inode mode mismatch with dir: inode mode=0%o btrfs type=%u dir type=%u",
5455                                   inode->i_mode, btrfs_inode_type(inode),
5456                                   di_type);
5457                         iput(inode);
5458                         return ERR_PTR(-EUCLEAN);
5459                 }
5460                 return inode;
5461         }
5462
5463         ret = fixup_tree_root_location(fs_info, dir, dentry,
5464                                        &location, &sub_root);
5465         if (ret < 0) {
5466                 if (ret != -ENOENT)
5467                         inode = ERR_PTR(ret);
5468                 else
5469                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5470         } else {
5471                 inode = btrfs_iget(dir->i_sb, location.objectid, sub_root);
5472         }
5473         if (root != sub_root)
5474                 btrfs_put_root(sub_root);
5475
5476         if (!IS_ERR(inode) && root != sub_root) {
5477                 down_read(&fs_info->cleanup_work_sem);
5478                 if (!sb_rdonly(inode->i_sb))
5479                         ret = btrfs_orphan_cleanup(sub_root);
5480                 up_read(&fs_info->cleanup_work_sem);
5481                 if (ret) {
5482                         iput(inode);
5483                         inode = ERR_PTR(ret);
5484                 }
5485         }
5486
5487         return inode;
5488 }
5489
5490 static int btrfs_dentry_delete(const struct dentry *dentry)
5491 {
5492         struct btrfs_root *root;
5493         struct inode *inode = d_inode(dentry);
5494
5495         if (!inode && !IS_ROOT(dentry))
5496                 inode = d_inode(dentry->d_parent);
5497
5498         if (inode) {
5499                 root = BTRFS_I(inode)->root;
5500                 if (btrfs_root_refs(&root->root_item) == 0)
5501                         return 1;
5502
5503                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5504                         return 1;
5505         }
5506         return 0;
5507 }
5508
5509 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5510                                    unsigned int flags)
5511 {
5512         struct inode *inode = btrfs_lookup_dentry(dir, dentry);
5513
5514         if (inode == ERR_PTR(-ENOENT))
5515                 inode = NULL;
5516         return d_splice_alias(inode, dentry);
5517 }
5518
5519 /*
5520  * All this infrastructure exists because dir_emit can fault, and we are holding
5521  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5522  * our information into that, and then dir_emit from the buffer.  This is
5523  * similar to what NFS does, only we don't keep the buffer around in pagecache
5524  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5525  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5526  * tree lock.
5527  */
5528 static int btrfs_opendir(struct inode *inode, struct file *file)
5529 {
5530         struct btrfs_file_private *private;
5531
5532         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5533         if (!private)
5534                 return -ENOMEM;
5535         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5536         if (!private->filldir_buf) {
5537                 kfree(private);
5538                 return -ENOMEM;
5539         }
5540         file->private_data = private;
5541         return 0;
5542 }
5543
5544 struct dir_entry {
5545         u64 ino;
5546         u64 offset;
5547         unsigned type;
5548         int name_len;
5549 };
5550
5551 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
5552 {
5553         while (entries--) {
5554                 struct dir_entry *entry = addr;
5555                 char *name = (char *)(entry + 1);
5556
5557                 ctx->pos = get_unaligned(&entry->offset);
5558                 if (!dir_emit(ctx, name, get_unaligned(&entry->name_len),
5559                                          get_unaligned(&entry->ino),
5560                                          get_unaligned(&entry->type)))
5561                         return 1;
5562                 addr += sizeof(struct dir_entry) +
5563                         get_unaligned(&entry->name_len);
5564                 ctx->pos++;
5565         }
5566         return 0;
5567 }
5568
5569 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5570 {
5571         struct inode *inode = file_inode(file);
5572         struct btrfs_root *root = BTRFS_I(inode)->root;
5573         struct btrfs_file_private *private = file->private_data;
5574         struct btrfs_dir_item *di;
5575         struct btrfs_key key;
5576         struct btrfs_key found_key;
5577         struct btrfs_path *path;
5578         void *addr;
5579         struct list_head ins_list;
5580         struct list_head del_list;
5581         int ret;
5582         struct extent_buffer *leaf;
5583         int slot;
5584         char *name_ptr;
5585         int name_len;
5586         int entries = 0;
5587         int total_len = 0;
5588         bool put = false;
5589         struct btrfs_key location;
5590
5591         if (!dir_emit_dots(file, ctx))
5592                 return 0;
5593
5594         path = btrfs_alloc_path();
5595         if (!path)
5596                 return -ENOMEM;
5597
5598         addr = private->filldir_buf;
5599         path->reada = READA_FORWARD;
5600
5601         INIT_LIST_HEAD(&ins_list);
5602         INIT_LIST_HEAD(&del_list);
5603         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
5604
5605 again:
5606         key.type = BTRFS_DIR_INDEX_KEY;
5607         key.offset = ctx->pos;
5608         key.objectid = btrfs_ino(BTRFS_I(inode));
5609
5610         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5611         if (ret < 0)
5612                 goto err;
5613
5614         while (1) {
5615                 struct dir_entry *entry;
5616
5617                 leaf = path->nodes[0];
5618                 slot = path->slots[0];
5619                 if (slot >= btrfs_header_nritems(leaf)) {
5620                         ret = btrfs_next_leaf(root, path);
5621                         if (ret < 0)
5622                                 goto err;
5623                         else if (ret > 0)
5624                                 break;
5625                         continue;
5626                 }
5627
5628                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5629
5630                 if (found_key.objectid != key.objectid)
5631                         break;
5632                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
5633                         break;
5634                 if (found_key.offset < ctx->pos)
5635                         goto next;
5636                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
5637                         goto next;
5638                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5639                 name_len = btrfs_dir_name_len(leaf, di);
5640                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
5641                     PAGE_SIZE) {
5642                         btrfs_release_path(path);
5643                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5644                         if (ret)
5645                                 goto nopos;
5646                         addr = private->filldir_buf;
5647                         entries = 0;
5648                         total_len = 0;
5649                         goto again;
5650                 }
5651
5652                 entry = addr;
5653                 put_unaligned(name_len, &entry->name_len);
5654                 name_ptr = (char *)(entry + 1);
5655                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
5656                                    name_len);
5657                 put_unaligned(fs_ftype_to_dtype(btrfs_dir_type(leaf, di)),
5658                                 &entry->type);
5659                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
5660                 put_unaligned(location.objectid, &entry->ino);
5661                 put_unaligned(found_key.offset, &entry->offset);
5662                 entries++;
5663                 addr += sizeof(struct dir_entry) + name_len;
5664                 total_len += sizeof(struct dir_entry) + name_len;
5665 next:
5666                 path->slots[0]++;
5667         }
5668         btrfs_release_path(path);
5669
5670         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5671         if (ret)
5672                 goto nopos;
5673
5674         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
5675         if (ret)
5676                 goto nopos;
5677
5678         /*
5679          * Stop new entries from being returned after we return the last
5680          * entry.
5681          *
5682          * New directory entries are assigned a strictly increasing
5683          * offset.  This means that new entries created during readdir
5684          * are *guaranteed* to be seen in the future by that readdir.
5685          * This has broken buggy programs which operate on names as
5686          * they're returned by readdir.  Until we re-use freed offsets
5687          * we have this hack to stop new entries from being returned
5688          * under the assumption that they'll never reach this huge
5689          * offset.
5690          *
5691          * This is being careful not to overflow 32bit loff_t unless the
5692          * last entry requires it because doing so has broken 32bit apps
5693          * in the past.
5694          */
5695         if (ctx->pos >= INT_MAX)
5696                 ctx->pos = LLONG_MAX;
5697         else
5698                 ctx->pos = INT_MAX;
5699 nopos:
5700         ret = 0;
5701 err:
5702         if (put)
5703                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
5704         btrfs_free_path(path);
5705         return ret;
5706 }
5707
5708 /*
5709  * This is somewhat expensive, updating the tree every time the
5710  * inode changes.  But, it is most likely to find the inode in cache.
5711  * FIXME, needs more benchmarking...there are no reasons other than performance
5712  * to keep or drop this code.
5713  */
5714 static int btrfs_dirty_inode(struct inode *inode)
5715 {
5716         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5717         struct btrfs_root *root = BTRFS_I(inode)->root;
5718         struct btrfs_trans_handle *trans;
5719         int ret;
5720
5721         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5722                 return 0;
5723
5724         trans = btrfs_join_transaction(root);
5725         if (IS_ERR(trans))
5726                 return PTR_ERR(trans);
5727
5728         ret = btrfs_update_inode(trans, root, inode);
5729         if (ret && ret == -ENOSPC) {
5730                 /* whoops, lets try again with the full transaction */
5731                 btrfs_end_transaction(trans);
5732                 trans = btrfs_start_transaction(root, 1);
5733                 if (IS_ERR(trans))
5734                         return PTR_ERR(trans);
5735
5736                 ret = btrfs_update_inode(trans, root, inode);
5737         }
5738         btrfs_end_transaction(trans);
5739         if (BTRFS_I(inode)->delayed_node)
5740                 btrfs_balance_delayed_items(fs_info);
5741
5742         return ret;
5743 }
5744
5745 /*
5746  * This is a copy of file_update_time.  We need this so we can return error on
5747  * ENOSPC for updating the inode in the case of file write and mmap writes.
5748  */
5749 static int btrfs_update_time(struct inode *inode, struct timespec64 *now,
5750                              int flags)
5751 {
5752         struct btrfs_root *root = BTRFS_I(inode)->root;
5753         bool dirty = flags & ~S_VERSION;
5754
5755         if (btrfs_root_readonly(root))
5756                 return -EROFS;
5757
5758         if (flags & S_VERSION)
5759                 dirty |= inode_maybe_inc_iversion(inode, dirty);
5760         if (flags & S_CTIME)
5761                 inode->i_ctime = *now;
5762         if (flags & S_MTIME)
5763                 inode->i_mtime = *now;
5764         if (flags & S_ATIME)
5765                 inode->i_atime = *now;
5766         return dirty ? btrfs_dirty_inode(inode) : 0;
5767 }
5768
5769 /*
5770  * find the highest existing sequence number in a directory
5771  * and then set the in-memory index_cnt variable to reflect
5772  * free sequence numbers
5773  */
5774 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
5775 {
5776         struct btrfs_root *root = inode->root;
5777         struct btrfs_key key, found_key;
5778         struct btrfs_path *path;
5779         struct extent_buffer *leaf;
5780         int ret;
5781
5782         key.objectid = btrfs_ino(inode);
5783         key.type = BTRFS_DIR_INDEX_KEY;
5784         key.offset = (u64)-1;
5785
5786         path = btrfs_alloc_path();
5787         if (!path)
5788                 return -ENOMEM;
5789
5790         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5791         if (ret < 0)
5792                 goto out;
5793         /* FIXME: we should be able to handle this */
5794         if (ret == 0)
5795                 goto out;
5796         ret = 0;
5797
5798         /*
5799          * MAGIC NUMBER EXPLANATION:
5800          * since we search a directory based on f_pos we have to start at 2
5801          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
5802          * else has to start at 2
5803          */
5804         if (path->slots[0] == 0) {
5805                 inode->index_cnt = 2;
5806                 goto out;
5807         }
5808
5809         path->slots[0]--;
5810
5811         leaf = path->nodes[0];
5812         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5813
5814         if (found_key.objectid != btrfs_ino(inode) ||
5815             found_key.type != BTRFS_DIR_INDEX_KEY) {
5816                 inode->index_cnt = 2;
5817                 goto out;
5818         }
5819
5820         inode->index_cnt = found_key.offset + 1;
5821 out:
5822         btrfs_free_path(path);
5823         return ret;
5824 }
5825
5826 /*
5827  * helper to find a free sequence number in a given directory.  This current
5828  * code is very simple, later versions will do smarter things in the btree
5829  */
5830 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
5831 {
5832         int ret = 0;
5833
5834         if (dir->index_cnt == (u64)-1) {
5835                 ret = btrfs_inode_delayed_dir_index_count(dir);
5836                 if (ret) {
5837                         ret = btrfs_set_inode_index_count(dir);
5838                         if (ret)
5839                                 return ret;
5840                 }
5841         }
5842
5843         *index = dir->index_cnt;
5844         dir->index_cnt++;
5845
5846         return ret;
5847 }
5848
5849 static int btrfs_insert_inode_locked(struct inode *inode)
5850 {
5851         struct btrfs_iget_args args;
5852
5853         args.ino = BTRFS_I(inode)->location.objectid;
5854         args.root = BTRFS_I(inode)->root;
5855
5856         return insert_inode_locked4(inode,
5857                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
5858                    btrfs_find_actor, &args);
5859 }
5860
5861 /*
5862  * Inherit flags from the parent inode.
5863  *
5864  * Currently only the compression flags and the cow flags are inherited.
5865  */
5866 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
5867 {
5868         unsigned int flags;
5869
5870         if (!dir)
5871                 return;
5872
5873         flags = BTRFS_I(dir)->flags;
5874
5875         if (flags & BTRFS_INODE_NOCOMPRESS) {
5876                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
5877                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
5878         } else if (flags & BTRFS_INODE_COMPRESS) {
5879                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
5880                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
5881         }
5882
5883         if (flags & BTRFS_INODE_NODATACOW) {
5884                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
5885                 if (S_ISREG(inode->i_mode))
5886                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
5887         }
5888
5889         btrfs_sync_inode_flags_to_i_flags(inode);
5890 }
5891
5892 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
5893                                      struct btrfs_root *root,
5894                                      struct inode *dir,
5895                                      const char *name, int name_len,
5896                                      u64 ref_objectid, u64 objectid,
5897                                      umode_t mode, u64 *index)
5898 {
5899         struct btrfs_fs_info *fs_info = root->fs_info;
5900         struct inode *inode;
5901         struct btrfs_inode_item *inode_item;
5902         struct btrfs_key *location;
5903         struct btrfs_path *path;
5904         struct btrfs_inode_ref *ref;
5905         struct btrfs_key key[2];
5906         u32 sizes[2];
5907         int nitems = name ? 2 : 1;
5908         unsigned long ptr;
5909         unsigned int nofs_flag;
5910         int ret;
5911
5912         path = btrfs_alloc_path();
5913         if (!path)
5914                 return ERR_PTR(-ENOMEM);
5915
5916         nofs_flag = memalloc_nofs_save();
5917         inode = new_inode(fs_info->sb);
5918         memalloc_nofs_restore(nofs_flag);
5919         if (!inode) {
5920                 btrfs_free_path(path);
5921                 return ERR_PTR(-ENOMEM);
5922         }
5923
5924         /*
5925          * O_TMPFILE, set link count to 0, so that after this point,
5926          * we fill in an inode item with the correct link count.
5927          */
5928         if (!name)
5929                 set_nlink(inode, 0);
5930
5931         /*
5932          * we have to initialize this early, so we can reclaim the inode
5933          * number if we fail afterwards in this function.
5934          */
5935         inode->i_ino = objectid;
5936
5937         if (dir && name) {
5938                 trace_btrfs_inode_request(dir);
5939
5940                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
5941                 if (ret) {
5942                         btrfs_free_path(path);
5943                         iput(inode);
5944                         return ERR_PTR(ret);
5945                 }
5946         } else if (dir) {
5947                 *index = 0;
5948         }
5949         /*
5950          * index_cnt is ignored for everything but a dir,
5951          * btrfs_set_inode_index_count has an explanation for the magic
5952          * number
5953          */
5954         BTRFS_I(inode)->index_cnt = 2;
5955         BTRFS_I(inode)->dir_index = *index;
5956         BTRFS_I(inode)->root = btrfs_grab_root(root);
5957         BTRFS_I(inode)->generation = trans->transid;
5958         inode->i_generation = BTRFS_I(inode)->generation;
5959
5960         /*
5961          * We could have gotten an inode number from somebody who was fsynced
5962          * and then removed in this same transaction, so let's just set full
5963          * sync since it will be a full sync anyway and this will blow away the
5964          * old info in the log.
5965          */
5966         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
5967
5968         key[0].objectid = objectid;
5969         key[0].type = BTRFS_INODE_ITEM_KEY;
5970         key[0].offset = 0;
5971
5972         sizes[0] = sizeof(struct btrfs_inode_item);
5973
5974         if (name) {
5975                 /*
5976                  * Start new inodes with an inode_ref. This is slightly more
5977                  * efficient for small numbers of hard links since they will
5978                  * be packed into one item. Extended refs will kick in if we
5979                  * add more hard links than can fit in the ref item.
5980                  */
5981                 key[1].objectid = objectid;
5982                 key[1].type = BTRFS_INODE_REF_KEY;
5983                 key[1].offset = ref_objectid;
5984
5985                 sizes[1] = name_len + sizeof(*ref);
5986         }
5987
5988         location = &BTRFS_I(inode)->location;
5989         location->objectid = objectid;
5990         location->offset = 0;
5991         location->type = BTRFS_INODE_ITEM_KEY;
5992
5993         ret = btrfs_insert_inode_locked(inode);
5994         if (ret < 0) {
5995                 iput(inode);
5996                 goto fail;
5997         }
5998
5999         path->leave_spinning = 1;
6000         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6001         if (ret != 0)
6002                 goto fail_unlock;
6003
6004         inode_init_owner(inode, dir, mode);
6005         inode_set_bytes(inode, 0);
6006
6007         inode->i_mtime = current_time(inode);
6008         inode->i_atime = inode->i_mtime;
6009         inode->i_ctime = inode->i_mtime;
6010         BTRFS_I(inode)->i_otime = inode->i_mtime;
6011
6012         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6013                                   struct btrfs_inode_item);
6014         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6015                              sizeof(*inode_item));
6016         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6017
6018         if (name) {
6019                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6020                                      struct btrfs_inode_ref);
6021                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6022                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6023                 ptr = (unsigned long)(ref + 1);
6024                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6025         }
6026
6027         btrfs_mark_buffer_dirty(path->nodes[0]);
6028         btrfs_free_path(path);
6029
6030         btrfs_inherit_iflags(inode, dir);
6031
6032         if (S_ISREG(mode)) {
6033                 if (btrfs_test_opt(fs_info, NODATASUM))
6034                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6035                 if (btrfs_test_opt(fs_info, NODATACOW))
6036                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6037                                 BTRFS_INODE_NODATASUM;
6038         }
6039
6040         inode_tree_add(inode);
6041
6042         trace_btrfs_inode_new(inode);
6043         btrfs_set_inode_last_trans(trans, inode);
6044
6045         btrfs_update_root_times(trans, root);
6046
6047         ret = btrfs_inode_inherit_props(trans, inode, dir);
6048         if (ret)
6049                 btrfs_err(fs_info,
6050                           "error inheriting props for ino %llu (root %llu): %d",
6051                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6052
6053         return inode;
6054
6055 fail_unlock:
6056         discard_new_inode(inode);
6057 fail:
6058         if (dir && name)
6059                 BTRFS_I(dir)->index_cnt--;
6060         btrfs_free_path(path);
6061         return ERR_PTR(ret);
6062 }
6063
6064 /*
6065  * utility function to add 'inode' into 'parent_inode' with
6066  * a give name and a given sequence number.
6067  * if 'add_backref' is true, also insert a backref from the
6068  * inode to the parent directory.
6069  */
6070 int btrfs_add_link(struct btrfs_trans_handle *trans,
6071                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6072                    const char *name, int name_len, int add_backref, u64 index)
6073 {
6074         int ret = 0;
6075         struct btrfs_key key;
6076         struct btrfs_root *root = parent_inode->root;
6077         u64 ino = btrfs_ino(inode);
6078         u64 parent_ino = btrfs_ino(parent_inode);
6079
6080         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6081                 memcpy(&key, &inode->root->root_key, sizeof(key));
6082         } else {
6083                 key.objectid = ino;
6084                 key.type = BTRFS_INODE_ITEM_KEY;
6085                 key.offset = 0;
6086         }
6087
6088         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6089                 ret = btrfs_add_root_ref(trans, key.objectid,
6090                                          root->root_key.objectid, parent_ino,
6091                                          index, name, name_len);
6092         } else if (add_backref) {
6093                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6094                                              parent_ino, index);
6095         }
6096
6097         /* Nothing to clean up yet */
6098         if (ret)
6099                 return ret;
6100
6101         ret = btrfs_insert_dir_item(trans, name, name_len, parent_inode, &key,
6102                                     btrfs_inode_type(&inode->vfs_inode), index);
6103         if (ret == -EEXIST || ret == -EOVERFLOW)
6104                 goto fail_dir_item;
6105         else if (ret) {
6106                 btrfs_abort_transaction(trans, ret);
6107                 return ret;
6108         }
6109
6110         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6111                            name_len * 2);
6112         inode_inc_iversion(&parent_inode->vfs_inode);
6113         /*
6114          * If we are replaying a log tree, we do not want to update the mtime
6115          * and ctime of the parent directory with the current time, since the
6116          * log replay procedure is responsible for setting them to their correct
6117          * values (the ones it had when the fsync was done).
6118          */
6119         if (!test_bit(BTRFS_FS_LOG_RECOVERING, &root->fs_info->flags)) {
6120                 struct timespec64 now = current_time(&parent_inode->vfs_inode);
6121
6122                 parent_inode->vfs_inode.i_mtime = now;
6123                 parent_inode->vfs_inode.i_ctime = now;
6124         }
6125         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6126         if (ret)
6127                 btrfs_abort_transaction(trans, ret);
6128         return ret;
6129
6130 fail_dir_item:
6131         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6132                 u64 local_index;
6133                 int err;
6134                 err = btrfs_del_root_ref(trans, key.objectid,
6135                                          root->root_key.objectid, parent_ino,
6136                                          &local_index, name, name_len);
6137                 if (err)
6138                         btrfs_abort_transaction(trans, err);
6139         } else if (add_backref) {
6140                 u64 local_index;
6141                 int err;
6142
6143                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6144                                           ino, parent_ino, &local_index);
6145                 if (err)
6146                         btrfs_abort_transaction(trans, err);
6147         }
6148
6149         /* Return the original error code */
6150         return ret;
6151 }
6152
6153 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6154                             struct btrfs_inode *dir, struct dentry *dentry,
6155                             struct btrfs_inode *inode, int backref, u64 index)
6156 {
6157         int err = btrfs_add_link(trans, dir, inode,
6158                                  dentry->d_name.name, dentry->d_name.len,
6159                                  backref, index);
6160         if (err > 0)
6161                 err = -EEXIST;
6162         return err;
6163 }
6164
6165 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6166                         umode_t mode, dev_t rdev)
6167 {
6168         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6169         struct btrfs_trans_handle *trans;
6170         struct btrfs_root *root = BTRFS_I(dir)->root;
6171         struct inode *inode = NULL;
6172         int err;
6173         u64 objectid;
6174         u64 index = 0;
6175
6176         /*
6177          * 2 for inode item and ref
6178          * 2 for dir items
6179          * 1 for xattr if selinux is on
6180          */
6181         trans = btrfs_start_transaction(root, 5);
6182         if (IS_ERR(trans))
6183                 return PTR_ERR(trans);
6184
6185         err = btrfs_find_free_ino(root, &objectid);
6186         if (err)
6187                 goto out_unlock;
6188
6189         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6190                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6191                         mode, &index);
6192         if (IS_ERR(inode)) {
6193                 err = PTR_ERR(inode);
6194                 inode = NULL;
6195                 goto out_unlock;
6196         }
6197
6198         /*
6199         * If the active LSM wants to access the inode during
6200         * d_instantiate it needs these. Smack checks to see
6201         * if the filesystem supports xattrs by looking at the
6202         * ops vector.
6203         */
6204         inode->i_op = &btrfs_special_inode_operations;
6205         init_special_inode(inode, inode->i_mode, rdev);
6206
6207         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6208         if (err)
6209                 goto out_unlock;
6210
6211         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6212                         0, index);
6213         if (err)
6214                 goto out_unlock;
6215
6216         btrfs_update_inode(trans, root, inode);
6217         d_instantiate_new(dentry, inode);
6218
6219 out_unlock:
6220         btrfs_end_transaction(trans);
6221         btrfs_btree_balance_dirty(fs_info);
6222         if (err && inode) {
6223                 inode_dec_link_count(inode);
6224                 discard_new_inode(inode);
6225         }
6226         return err;
6227 }
6228
6229 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6230                         umode_t mode, bool excl)
6231 {
6232         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6233         struct btrfs_trans_handle *trans;
6234         struct btrfs_root *root = BTRFS_I(dir)->root;
6235         struct inode *inode = NULL;
6236         int err;
6237         u64 objectid;
6238         u64 index = 0;
6239
6240         /*
6241          * 2 for inode item and ref
6242          * 2 for dir items
6243          * 1 for xattr if selinux is on
6244          */
6245         trans = btrfs_start_transaction(root, 5);
6246         if (IS_ERR(trans))
6247                 return PTR_ERR(trans);
6248
6249         err = btrfs_find_free_ino(root, &objectid);
6250         if (err)
6251                 goto out_unlock;
6252
6253         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6254                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6255                         mode, &index);
6256         if (IS_ERR(inode)) {
6257                 err = PTR_ERR(inode);
6258                 inode = NULL;
6259                 goto out_unlock;
6260         }
6261         /*
6262         * If the active LSM wants to access the inode during
6263         * d_instantiate it needs these. Smack checks to see
6264         * if the filesystem supports xattrs by looking at the
6265         * ops vector.
6266         */
6267         inode->i_fop = &btrfs_file_operations;
6268         inode->i_op = &btrfs_file_inode_operations;
6269         inode->i_mapping->a_ops = &btrfs_aops;
6270
6271         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6272         if (err)
6273                 goto out_unlock;
6274
6275         err = btrfs_update_inode(trans, root, inode);
6276         if (err)
6277                 goto out_unlock;
6278
6279         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6280                         0, index);
6281         if (err)
6282                 goto out_unlock;
6283
6284         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6285         d_instantiate_new(dentry, inode);
6286
6287 out_unlock:
6288         btrfs_end_transaction(trans);
6289         if (err && inode) {
6290                 inode_dec_link_count(inode);
6291                 discard_new_inode(inode);
6292         }
6293         btrfs_btree_balance_dirty(fs_info);
6294         return err;
6295 }
6296
6297 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6298                       struct dentry *dentry)
6299 {
6300         struct btrfs_trans_handle *trans = NULL;
6301         struct btrfs_root *root = BTRFS_I(dir)->root;
6302         struct inode *inode = d_inode(old_dentry);
6303         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6304         u64 index;
6305         int err;
6306         int drop_inode = 0;
6307
6308         /* do not allow sys_link's with other subvols of the same device */
6309         if (root->root_key.objectid != BTRFS_I(inode)->root->root_key.objectid)
6310                 return -EXDEV;
6311
6312         if (inode->i_nlink >= BTRFS_LINK_MAX)
6313                 return -EMLINK;
6314
6315         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6316         if (err)
6317                 goto fail;
6318
6319         /*
6320          * 2 items for inode and inode ref
6321          * 2 items for dir items
6322          * 1 item for parent inode
6323          * 1 item for orphan item deletion if O_TMPFILE
6324          */
6325         trans = btrfs_start_transaction(root, inode->i_nlink ? 5 : 6);
6326         if (IS_ERR(trans)) {
6327                 err = PTR_ERR(trans);
6328                 trans = NULL;
6329                 goto fail;
6330         }
6331
6332         /* There are several dir indexes for this inode, clear the cache. */
6333         BTRFS_I(inode)->dir_index = 0ULL;
6334         inc_nlink(inode);
6335         inode_inc_iversion(inode);
6336         inode->i_ctime = current_time(inode);
6337         ihold(inode);
6338         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6339
6340         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6341                         1, index);
6342
6343         if (err) {
6344                 drop_inode = 1;
6345         } else {
6346                 struct dentry *parent = dentry->d_parent;
6347                 int ret;
6348
6349                 err = btrfs_update_inode(trans, root, inode);
6350                 if (err)
6351                         goto fail;
6352                 if (inode->i_nlink == 1) {
6353                         /*
6354                          * If new hard link count is 1, it's a file created
6355                          * with open(2) O_TMPFILE flag.
6356                          */
6357                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6358                         if (err)
6359                                 goto fail;
6360                 }
6361                 d_instantiate(dentry, inode);
6362                 ret = btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent,
6363                                          true, NULL);
6364                 if (ret == BTRFS_NEED_TRANS_COMMIT) {
6365                         err = btrfs_commit_transaction(trans);
6366                         trans = NULL;
6367                 }
6368         }
6369
6370 fail:
6371         if (trans)
6372                 btrfs_end_transaction(trans);
6373         if (drop_inode) {
6374                 inode_dec_link_count(inode);
6375                 iput(inode);
6376         }
6377         btrfs_btree_balance_dirty(fs_info);
6378         return err;
6379 }
6380
6381 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6382 {
6383         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6384         struct inode *inode = NULL;
6385         struct btrfs_trans_handle *trans;
6386         struct btrfs_root *root = BTRFS_I(dir)->root;
6387         int err = 0;
6388         u64 objectid = 0;
6389         u64 index = 0;
6390
6391         /*
6392          * 2 items for inode and ref
6393          * 2 items for dir items
6394          * 1 for xattr if selinux is on
6395          */
6396         trans = btrfs_start_transaction(root, 5);
6397         if (IS_ERR(trans))
6398                 return PTR_ERR(trans);
6399
6400         err = btrfs_find_free_ino(root, &objectid);
6401         if (err)
6402                 goto out_fail;
6403
6404         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6405                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6406                         S_IFDIR | mode, &index);
6407         if (IS_ERR(inode)) {
6408                 err = PTR_ERR(inode);
6409                 inode = NULL;
6410                 goto out_fail;
6411         }
6412
6413         /* these must be set before we unlock the inode */
6414         inode->i_op = &btrfs_dir_inode_operations;
6415         inode->i_fop = &btrfs_dir_file_operations;
6416
6417         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6418         if (err)
6419                 goto out_fail;
6420
6421         btrfs_i_size_write(BTRFS_I(inode), 0);
6422         err = btrfs_update_inode(trans, root, inode);
6423         if (err)
6424                 goto out_fail;
6425
6426         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6427                         dentry->d_name.name,
6428                         dentry->d_name.len, 0, index);
6429         if (err)
6430                 goto out_fail;
6431
6432         d_instantiate_new(dentry, inode);
6433
6434 out_fail:
6435         btrfs_end_transaction(trans);
6436         if (err && inode) {
6437                 inode_dec_link_count(inode);
6438                 discard_new_inode(inode);
6439         }
6440         btrfs_btree_balance_dirty(fs_info);
6441         return err;
6442 }
6443
6444 static noinline int uncompress_inline(struct btrfs_path *path,
6445                                       struct page *page,
6446                                       size_t pg_offset, u64 extent_offset,
6447                                       struct btrfs_file_extent_item *item)
6448 {
6449         int ret;
6450         struct extent_buffer *leaf = path->nodes[0];
6451         char *tmp;
6452         size_t max_size;
6453         unsigned long inline_size;
6454         unsigned long ptr;
6455         int compress_type;
6456
6457         WARN_ON(pg_offset != 0);
6458         compress_type = btrfs_file_extent_compression(leaf, item);
6459         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6460         inline_size = btrfs_file_extent_inline_item_len(leaf,
6461                                         btrfs_item_nr(path->slots[0]));
6462         tmp = kmalloc(inline_size, GFP_NOFS);
6463         if (!tmp)
6464                 return -ENOMEM;
6465         ptr = btrfs_file_extent_inline_start(item);
6466
6467         read_extent_buffer(leaf, tmp, ptr, inline_size);
6468
6469         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6470         ret = btrfs_decompress(compress_type, tmp, page,
6471                                extent_offset, inline_size, max_size);
6472
6473         /*
6474          * decompression code contains a memset to fill in any space between the end
6475          * of the uncompressed data and the end of max_size in case the decompressed
6476          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6477          * the end of an inline extent and the beginning of the next block, so we
6478          * cover that region here.
6479          */
6480
6481         if (max_size + pg_offset < PAGE_SIZE) {
6482                 char *map = kmap(page);
6483                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6484                 kunmap(page);
6485         }
6486         kfree(tmp);
6487         return ret;
6488 }
6489
6490 /**
6491  * btrfs_get_extent - Lookup the first extent overlapping a range in a file.
6492  * @inode:      file to search in
6493  * @page:       page to read extent data into if the extent is inline
6494  * @pg_offset:  offset into @page to copy to
6495  * @start:      file offset
6496  * @len:        length of range starting at @start
6497  *
6498  * This returns the first &struct extent_map which overlaps with the given
6499  * range, reading it from the B-tree and caching it if necessary. Note that
6500  * there may be more extents which overlap the given range after the returned
6501  * extent_map.
6502  *
6503  * If @page is not NULL and the extent is inline, this also reads the extent
6504  * data directly into the page and marks the extent up to date in the io_tree.
6505  *
6506  * Return: ERR_PTR on error, non-NULL extent_map on success.
6507  */
6508 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6509                                     struct page *page, size_t pg_offset,
6510                                     u64 start, u64 len)
6511 {
6512         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6513         int ret;
6514         int err = 0;
6515         u64 extent_start = 0;
6516         u64 extent_end = 0;
6517         u64 objectid = btrfs_ino(inode);
6518         int extent_type = -1;
6519         struct btrfs_path *path = NULL;
6520         struct btrfs_root *root = inode->root;
6521         struct btrfs_file_extent_item *item;
6522         struct extent_buffer *leaf;
6523         struct btrfs_key found_key;
6524         struct extent_map *em = NULL;
6525         struct extent_map_tree *em_tree = &inode->extent_tree;
6526         struct extent_io_tree *io_tree = &inode->io_tree;
6527
6528         read_lock(&em_tree->lock);
6529         em = lookup_extent_mapping(em_tree, start, len);
6530         read_unlock(&em_tree->lock);
6531
6532         if (em) {
6533                 if (em->start > start || em->start + em->len <= start)
6534                         free_extent_map(em);
6535                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6536                         free_extent_map(em);
6537                 else
6538                         goto out;
6539         }
6540         em = alloc_extent_map();
6541         if (!em) {
6542                 err = -ENOMEM;
6543                 goto out;
6544         }
6545         em->start = EXTENT_MAP_HOLE;
6546         em->orig_start = EXTENT_MAP_HOLE;
6547         em->len = (u64)-1;
6548         em->block_len = (u64)-1;
6549
6550         path = btrfs_alloc_path();
6551         if (!path) {
6552                 err = -ENOMEM;
6553                 goto out;
6554         }
6555
6556         /* Chances are we'll be called again, so go ahead and do readahead */
6557         path->reada = READA_FORWARD;
6558
6559         /*
6560          * Unless we're going to uncompress the inline extent, no sleep would
6561          * happen.
6562          */
6563         path->leave_spinning = 1;
6564
6565         ret = btrfs_lookup_file_extent(NULL, root, path, objectid, start, 0);
6566         if (ret < 0) {
6567                 err = ret;
6568                 goto out;
6569         } else if (ret > 0) {
6570                 if (path->slots[0] == 0)
6571                         goto not_found;
6572                 path->slots[0]--;
6573         }
6574
6575         leaf = path->nodes[0];
6576         item = btrfs_item_ptr(leaf, path->slots[0],
6577                               struct btrfs_file_extent_item);
6578         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6579         if (found_key.objectid != objectid ||
6580             found_key.type != BTRFS_EXTENT_DATA_KEY) {
6581                 /*
6582                  * If we backup past the first extent we want to move forward
6583                  * and see if there is an extent in front of us, otherwise we'll
6584                  * say there is a hole for our whole search range which can
6585                  * cause problems.
6586                  */
6587                 extent_end = start;
6588                 goto next;
6589         }
6590
6591         extent_type = btrfs_file_extent_type(leaf, item);
6592         extent_start = found_key.offset;
6593         extent_end = btrfs_file_extent_end(path);
6594         if (extent_type == BTRFS_FILE_EXTENT_REG ||
6595             extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
6596                 /* Only regular file could have regular/prealloc extent */
6597                 if (!S_ISREG(inode->vfs_inode.i_mode)) {
6598                         ret = -EUCLEAN;
6599                         btrfs_crit(fs_info,
6600                 "regular/prealloc extent found for non-regular inode %llu",
6601                                    btrfs_ino(inode));
6602                         goto out;
6603                 }
6604                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
6605                                                        extent_start);
6606         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
6607                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
6608                                                       path->slots[0],
6609                                                       extent_start);
6610         }
6611 next:
6612         if (start >= extent_end) {
6613                 path->slots[0]++;
6614                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6615                         ret = btrfs_next_leaf(root, path);
6616                         if (ret < 0) {
6617                                 err = ret;
6618                                 goto out;
6619                         } else if (ret > 0) {
6620                                 goto not_found;
6621                         }
6622                         leaf = path->nodes[0];
6623                 }
6624                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6625                 if (found_key.objectid != objectid ||
6626                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6627                         goto not_found;
6628                 if (start + len <= found_key.offset)
6629                         goto not_found;
6630                 if (start > found_key.offset)
6631                         goto next;
6632
6633                 /* New extent overlaps with existing one */
6634                 em->start = start;
6635                 em->orig_start = start;
6636                 em->len = found_key.offset - start;
6637                 em->block_start = EXTENT_MAP_HOLE;
6638                 goto insert;
6639         }
6640
6641         btrfs_extent_item_to_extent_map(inode, path, item, !page, em);
6642
6643         if (extent_type == BTRFS_FILE_EXTENT_REG ||
6644             extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
6645                 goto insert;
6646         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
6647                 unsigned long ptr;
6648                 char *map;
6649                 size_t size;
6650                 size_t extent_offset;
6651                 size_t copy_size;
6652
6653                 if (!page)
6654                         goto out;
6655
6656                 size = btrfs_file_extent_ram_bytes(leaf, item);
6657                 extent_offset = page_offset(page) + pg_offset - extent_start;
6658                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
6659                                   size - extent_offset);
6660                 em->start = extent_start + extent_offset;
6661                 em->len = ALIGN(copy_size, fs_info->sectorsize);
6662                 em->orig_block_len = em->len;
6663                 em->orig_start = em->start;
6664                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6665
6666                 btrfs_set_path_blocking(path);
6667                 if (!PageUptodate(page)) {
6668                         if (btrfs_file_extent_compression(leaf, item) !=
6669                             BTRFS_COMPRESS_NONE) {
6670                                 ret = uncompress_inline(path, page, pg_offset,
6671                                                         extent_offset, item);
6672                                 if (ret) {
6673                                         err = ret;
6674                                         goto out;
6675                                 }
6676                         } else {
6677                                 map = kmap(page);
6678                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6679                                                    copy_size);
6680                                 if (pg_offset + copy_size < PAGE_SIZE) {
6681                                         memset(map + pg_offset + copy_size, 0,
6682                                                PAGE_SIZE - pg_offset -
6683                                                copy_size);
6684                                 }
6685                                 kunmap(page);
6686                         }
6687                         flush_dcache_page(page);
6688                 }
6689                 set_extent_uptodate(io_tree, em->start,
6690                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6691                 goto insert;
6692         }
6693 not_found:
6694         em->start = start;
6695         em->orig_start = start;
6696         em->len = len;
6697         em->block_start = EXTENT_MAP_HOLE;
6698 insert:
6699         btrfs_release_path(path);
6700         if (em->start > start || extent_map_end(em) <= start) {
6701                 btrfs_err(fs_info,
6702                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
6703                           em->start, em->len, start, len);
6704                 err = -EIO;
6705                 goto out;
6706         }
6707
6708         err = 0;
6709         write_lock(&em_tree->lock);
6710         err = btrfs_add_extent_mapping(fs_info, em_tree, &em, start, len);
6711         write_unlock(&em_tree->lock);
6712 out:
6713         btrfs_free_path(path);
6714
6715         trace_btrfs_get_extent(root, inode, em);
6716
6717         if (err) {
6718                 free_extent_map(em);
6719                 return ERR_PTR(err);
6720         }
6721         BUG_ON(!em); /* Error is always set */
6722         return em;
6723 }
6724
6725 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
6726                                            u64 start, u64 len)
6727 {
6728         struct extent_map *em;
6729         struct extent_map *hole_em = NULL;
6730         u64 delalloc_start = start;
6731         u64 end;
6732         u64 delalloc_len;
6733         u64 delalloc_end;
6734         int err = 0;
6735
6736         em = btrfs_get_extent(inode, NULL, 0, start, len);
6737         if (IS_ERR(em))
6738                 return em;
6739         /*
6740          * If our em maps to:
6741          * - a hole or
6742          * - a pre-alloc extent,
6743          * there might actually be delalloc bytes behind it.
6744          */
6745         if (em->block_start != EXTENT_MAP_HOLE &&
6746             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
6747                 return em;
6748         else
6749                 hole_em = em;
6750
6751         /* check to see if we've wrapped (len == -1 or similar) */
6752         end = start + len;
6753         if (end < start)
6754                 end = (u64)-1;
6755         else
6756                 end -= 1;
6757
6758         em = NULL;
6759
6760         /* ok, we didn't find anything, lets look for delalloc */
6761         delalloc_len = count_range_bits(&inode->io_tree, &delalloc_start,
6762                                  end, len, EXTENT_DELALLOC, 1);
6763         delalloc_end = delalloc_start + delalloc_len;
6764         if (delalloc_end < delalloc_start)
6765                 delalloc_end = (u64)-1;
6766
6767         /*
6768          * We didn't find anything useful, return the original results from
6769          * get_extent()
6770          */
6771         if (delalloc_start > end || delalloc_end <= start) {
6772                 em = hole_em;
6773                 hole_em = NULL;
6774                 goto out;
6775         }
6776
6777         /*
6778          * Adjust the delalloc_start to make sure it doesn't go backwards from
6779          * the start they passed in
6780          */
6781         delalloc_start = max(start, delalloc_start);
6782         delalloc_len = delalloc_end - delalloc_start;
6783
6784         if (delalloc_len > 0) {
6785                 u64 hole_start;
6786                 u64 hole_len;
6787                 const u64 hole_end = extent_map_end(hole_em);
6788
6789                 em = alloc_extent_map();
6790                 if (!em) {
6791                         err = -ENOMEM;
6792                         goto out;
6793                 }
6794
6795                 ASSERT(hole_em);
6796                 /*
6797                  * When btrfs_get_extent can't find anything it returns one
6798                  * huge hole
6799                  *
6800                  * Make sure what it found really fits our range, and adjust to
6801                  * make sure it is based on the start from the caller
6802                  */
6803                 if (hole_end <= start || hole_em->start > end) {
6804                        free_extent_map(hole_em);
6805                        hole_em = NULL;
6806                 } else {
6807                        hole_start = max(hole_em->start, start);
6808                        hole_len = hole_end - hole_start;
6809                 }
6810
6811                 if (hole_em && delalloc_start > hole_start) {
6812                         /*
6813                          * Our hole starts before our delalloc, so we have to
6814                          * return just the parts of the hole that go until the
6815                          * delalloc starts
6816                          */
6817                         em->len = min(hole_len, delalloc_start - hole_start);
6818                         em->start = hole_start;
6819                         em->orig_start = hole_start;
6820                         /*
6821                          * Don't adjust block start at all, it is fixed at
6822                          * EXTENT_MAP_HOLE
6823                          */
6824                         em->block_start = hole_em->block_start;
6825                         em->block_len = hole_len;
6826                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
6827                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
6828                 } else {
6829                         /*
6830                          * Hole is out of passed range or it starts after
6831                          * delalloc range
6832                          */
6833                         em->start = delalloc_start;
6834                         em->len = delalloc_len;
6835                         em->orig_start = delalloc_start;
6836                         em->block_start = EXTENT_MAP_DELALLOC;
6837                         em->block_len = delalloc_len;
6838                 }
6839         } else {
6840                 return hole_em;
6841         }
6842 out:
6843
6844         free_extent_map(hole_em);
6845         if (err) {
6846                 free_extent_map(em);
6847                 return ERR_PTR(err);
6848         }
6849         return em;
6850 }
6851
6852 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
6853                                                   const u64 start,
6854                                                   const u64 len,
6855                                                   const u64 orig_start,
6856                                                   const u64 block_start,
6857                                                   const u64 block_len,
6858                                                   const u64 orig_block_len,
6859                                                   const u64 ram_bytes,
6860                                                   const int type)
6861 {
6862         struct extent_map *em = NULL;
6863         int ret;
6864
6865         if (type != BTRFS_ORDERED_NOCOW) {
6866                 em = create_io_em(inode, start, len, orig_start,
6867                                   block_start, block_len, orig_block_len,
6868                                   ram_bytes,
6869                                   BTRFS_COMPRESS_NONE, /* compress_type */
6870                                   type);
6871                 if (IS_ERR(em))
6872                         goto out;
6873         }
6874         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
6875                                            len, block_len, type);
6876         if (ret) {
6877                 if (em) {
6878                         free_extent_map(em);
6879                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
6880                                                 start + len - 1, 0);
6881                 }
6882                 em = ERR_PTR(ret);
6883         }
6884  out:
6885
6886         return em;
6887 }
6888
6889 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
6890                                                   u64 start, u64 len)
6891 {
6892         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6893         struct btrfs_root *root = BTRFS_I(inode)->root;
6894         struct extent_map *em;
6895         struct btrfs_key ins;
6896         u64 alloc_hint;
6897         int ret;
6898
6899         alloc_hint = get_extent_allocation_hint(inode, start, len);
6900         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
6901                                    0, alloc_hint, &ins, 1, 1);
6902         if (ret)
6903                 return ERR_PTR(ret);
6904
6905         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
6906                                      ins.objectid, ins.offset, ins.offset,
6907                                      ins.offset, BTRFS_ORDERED_REGULAR);
6908         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
6909         if (IS_ERR(em))
6910                 btrfs_free_reserved_extent(fs_info, ins.objectid,
6911                                            ins.offset, 1);
6912
6913         return em;
6914 }
6915
6916 /*
6917  * returns 1 when the nocow is safe, < 1 on error, 0 if the
6918  * block must be cow'd
6919  */
6920 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
6921                               u64 *orig_start, u64 *orig_block_len,
6922                               u64 *ram_bytes)
6923 {
6924         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6925         struct btrfs_path *path;
6926         int ret;
6927         struct extent_buffer *leaf;
6928         struct btrfs_root *root = BTRFS_I(inode)->root;
6929         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6930         struct btrfs_file_extent_item *fi;
6931         struct btrfs_key key;
6932         u64 disk_bytenr;
6933         u64 backref_offset;
6934         u64 extent_end;
6935         u64 num_bytes;
6936         int slot;
6937         int found_type;
6938         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
6939
6940         path = btrfs_alloc_path();
6941         if (!path)
6942                 return -ENOMEM;
6943
6944         ret = btrfs_lookup_file_extent(NULL, root, path,
6945                         btrfs_ino(BTRFS_I(inode)), offset, 0);
6946         if (ret < 0)
6947                 goto out;
6948
6949         slot = path->slots[0];
6950         if (ret == 1) {
6951                 if (slot == 0) {
6952                         /* can't find the item, must cow */
6953                         ret = 0;
6954                         goto out;
6955                 }
6956                 slot--;
6957         }
6958         ret = 0;
6959         leaf = path->nodes[0];
6960         btrfs_item_key_to_cpu(leaf, &key, slot);
6961         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
6962             key.type != BTRFS_EXTENT_DATA_KEY) {
6963                 /* not our file or wrong item type, must cow */
6964                 goto out;
6965         }
6966
6967         if (key.offset > offset) {
6968                 /* Wrong offset, must cow */
6969                 goto out;
6970         }
6971
6972         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
6973         found_type = btrfs_file_extent_type(leaf, fi);
6974         if (found_type != BTRFS_FILE_EXTENT_REG &&
6975             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
6976                 /* not a regular extent, must cow */
6977                 goto out;
6978         }
6979
6980         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
6981                 goto out;
6982
6983         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
6984         if (extent_end <= offset)
6985                 goto out;
6986
6987         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
6988         if (disk_bytenr == 0)
6989                 goto out;
6990
6991         if (btrfs_file_extent_compression(leaf, fi) ||
6992             btrfs_file_extent_encryption(leaf, fi) ||
6993             btrfs_file_extent_other_encoding(leaf, fi))
6994                 goto out;
6995
6996         /*
6997          * Do the same check as in btrfs_cross_ref_exist but without the
6998          * unnecessary search.
6999          */
7000         if (btrfs_file_extent_generation(leaf, fi) <=
7001             btrfs_root_last_snapshot(&root->root_item))
7002                 goto out;
7003
7004         backref_offset = btrfs_file_extent_offset(leaf, fi);
7005
7006         if (orig_start) {
7007                 *orig_start = key.offset - backref_offset;
7008                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7009                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7010         }
7011
7012         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7013                 goto out;
7014
7015         num_bytes = min(offset + *len, extent_end) - offset;
7016         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7017                 u64 range_end;
7018
7019                 range_end = round_up(offset + num_bytes,
7020                                      root->fs_info->sectorsize) - 1;
7021                 ret = test_range_bit(io_tree, offset, range_end,
7022                                      EXTENT_DELALLOC, 0, NULL);
7023                 if (ret) {
7024                         ret = -EAGAIN;
7025                         goto out;
7026                 }
7027         }
7028
7029         btrfs_release_path(path);
7030
7031         /*
7032          * look for other files referencing this extent, if we
7033          * find any we must cow
7034          */
7035
7036         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7037                                     key.offset - backref_offset, disk_bytenr);
7038         if (ret) {
7039                 ret = 0;
7040                 goto out;
7041         }
7042
7043         /*
7044          * adjust disk_bytenr and num_bytes to cover just the bytes
7045          * in this extent we are about to write.  If there
7046          * are any csums in that range we have to cow in order
7047          * to keep the csums correct
7048          */
7049         disk_bytenr += backref_offset;
7050         disk_bytenr += offset - key.offset;
7051         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7052                 goto out;
7053         /*
7054          * all of the above have passed, it is safe to overwrite this extent
7055          * without cow
7056          */
7057         *len = num_bytes;
7058         ret = 1;
7059 out:
7060         btrfs_free_path(path);
7061         return ret;
7062 }
7063
7064 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7065                               struct extent_state **cached_state, int writing)
7066 {
7067         struct btrfs_ordered_extent *ordered;
7068         int ret = 0;
7069
7070         while (1) {
7071                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7072                                  cached_state);
7073                 /*
7074                  * We're concerned with the entire range that we're going to be
7075                  * doing DIO to, so we need to make sure there's no ordered
7076                  * extents in this range.
7077                  */
7078                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7079                                                      lockend - lockstart + 1);
7080
7081                 /*
7082                  * We need to make sure there are no buffered pages in this
7083                  * range either, we could have raced between the invalidate in
7084                  * generic_file_direct_write and locking the extent.  The
7085                  * invalidate needs to happen so that reads after a write do not
7086                  * get stale data.
7087                  */
7088                 if (!ordered &&
7089                     (!writing || !filemap_range_has_page(inode->i_mapping,
7090                                                          lockstart, lockend)))
7091                         break;
7092
7093                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7094                                      cached_state);
7095
7096                 if (ordered) {
7097                         /*
7098                          * If we are doing a DIO read and the ordered extent we
7099                          * found is for a buffered write, we can not wait for it
7100                          * to complete and retry, because if we do so we can
7101                          * deadlock with concurrent buffered writes on page
7102                          * locks. This happens only if our DIO read covers more
7103                          * than one extent map, if at this point has already
7104                          * created an ordered extent for a previous extent map
7105                          * and locked its range in the inode's io tree, and a
7106                          * concurrent write against that previous extent map's
7107                          * range and this range started (we unlock the ranges
7108                          * in the io tree only when the bios complete and
7109                          * buffered writes always lock pages before attempting
7110                          * to lock range in the io tree).
7111                          */
7112                         if (writing ||
7113                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7114                                 btrfs_start_ordered_extent(inode, ordered, 1);
7115                         else
7116                                 ret = -ENOTBLK;
7117                         btrfs_put_ordered_extent(ordered);
7118                 } else {
7119                         /*
7120                          * We could trigger writeback for this range (and wait
7121                          * for it to complete) and then invalidate the pages for
7122                          * this range (through invalidate_inode_pages2_range()),
7123                          * but that can lead us to a deadlock with a concurrent
7124                          * call to readahead (a buffered read or a defrag call
7125                          * triggered a readahead) on a page lock due to an
7126                          * ordered dio extent we created before but did not have
7127                          * yet a corresponding bio submitted (whence it can not
7128                          * complete), which makes readahead wait for that
7129                          * ordered extent to complete while holding a lock on
7130                          * that page.
7131                          */
7132                         ret = -ENOTBLK;
7133                 }
7134
7135                 if (ret)
7136                         break;
7137
7138                 cond_resched();
7139         }
7140
7141         return ret;
7142 }
7143
7144 /* The callers of this must take lock_extent() */
7145 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7146                                        u64 orig_start, u64 block_start,
7147                                        u64 block_len, u64 orig_block_len,
7148                                        u64 ram_bytes, int compress_type,
7149                                        int type)
7150 {
7151         struct extent_map_tree *em_tree;
7152         struct extent_map *em;
7153         int ret;
7154
7155         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7156                type == BTRFS_ORDERED_COMPRESSED ||
7157                type == BTRFS_ORDERED_NOCOW ||
7158                type == BTRFS_ORDERED_REGULAR);
7159
7160         em_tree = &BTRFS_I(inode)->extent_tree;
7161         em = alloc_extent_map();
7162         if (!em)
7163                 return ERR_PTR(-ENOMEM);
7164
7165         em->start = start;
7166         em->orig_start = orig_start;
7167         em->len = len;
7168         em->block_len = block_len;
7169         em->block_start = block_start;
7170         em->orig_block_len = orig_block_len;
7171         em->ram_bytes = ram_bytes;
7172         em->generation = -1;
7173         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7174         if (type == BTRFS_ORDERED_PREALLOC) {
7175                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7176         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7177                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7178                 em->compress_type = compress_type;
7179         }
7180
7181         do {
7182                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7183                                 em->start + em->len - 1, 0);
7184                 write_lock(&em_tree->lock);
7185                 ret = add_extent_mapping(em_tree, em, 1);
7186                 write_unlock(&em_tree->lock);
7187                 /*
7188                  * The caller has taken lock_extent(), who could race with us
7189                  * to add em?
7190                  */
7191         } while (ret == -EEXIST);
7192
7193         if (ret) {
7194                 free_extent_map(em);
7195                 return ERR_PTR(ret);
7196         }
7197
7198         /* em got 2 refs now, callers needs to do free_extent_map once. */
7199         return em;
7200 }
7201
7202
7203 static int btrfs_get_blocks_direct_read(struct extent_map *em,
7204                                         struct buffer_head *bh_result,
7205                                         struct inode *inode,
7206                                         u64 start, u64 len)
7207 {
7208         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7209
7210         if (em->block_start == EXTENT_MAP_HOLE ||
7211                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7212                 return -ENOENT;
7213
7214         len = min(len, em->len - (start - em->start));
7215
7216         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7217                 inode->i_blkbits;
7218         bh_result->b_size = len;
7219         bh_result->b_bdev = fs_info->fs_devices->latest_bdev;
7220         set_buffer_mapped(bh_result);
7221
7222         return 0;
7223 }
7224
7225 static int btrfs_get_blocks_direct_write(struct extent_map **map,
7226                                          struct buffer_head *bh_result,
7227                                          struct inode *inode,
7228                                          struct btrfs_dio_data *dio_data,
7229                                          u64 start, u64 len)
7230 {
7231         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7232         struct extent_map *em = *map;
7233         int ret = 0;
7234
7235         /*
7236          * We don't allocate a new extent in the following cases
7237          *
7238          * 1) The inode is marked as NODATACOW. In this case we'll just use the
7239          * existing extent.
7240          * 2) The extent is marked as PREALLOC. We're good to go here and can
7241          * just use the extent.
7242          *
7243          */
7244         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7245             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7246              em->block_start != EXTENT_MAP_HOLE)) {
7247                 int type;
7248                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7249
7250                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7251                         type = BTRFS_ORDERED_PREALLOC;
7252                 else
7253                         type = BTRFS_ORDERED_NOCOW;
7254                 len = min(len, em->len - (start - em->start));
7255                 block_start = em->block_start + (start - em->start);
7256
7257                 if (can_nocow_extent(inode, start, &len, &orig_start,
7258                                      &orig_block_len, &ram_bytes) == 1 &&
7259                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7260                         struct extent_map *em2;
7261
7262                         em2 = btrfs_create_dio_extent(inode, start, len,
7263                                                       orig_start, block_start,
7264                                                       len, orig_block_len,
7265                                                       ram_bytes, type);
7266                         btrfs_dec_nocow_writers(fs_info, block_start);
7267                         if (type == BTRFS_ORDERED_PREALLOC) {
7268                                 free_extent_map(em);
7269                                 *map = em = em2;
7270                         }
7271
7272                         if (em2 && IS_ERR(em2)) {
7273                                 ret = PTR_ERR(em2);
7274                                 goto out;
7275                         }
7276                         /*
7277                          * For inode marked NODATACOW or extent marked PREALLOC,
7278                          * use the existing or preallocated extent, so does not
7279                          * need to adjust btrfs_space_info's bytes_may_use.
7280                          */
7281                         btrfs_free_reserved_data_space_noquota(inode, start,
7282                                                                len);
7283                         goto skip_cow;
7284                 }
7285         }
7286
7287         /* this will cow the extent */
7288         len = bh_result->b_size;
7289         free_extent_map(em);
7290         *map = em = btrfs_new_extent_direct(inode, start, len);
7291         if (IS_ERR(em)) {
7292                 ret = PTR_ERR(em);
7293                 goto out;
7294         }
7295
7296         len = min(len, em->len - (start - em->start));
7297
7298 skip_cow:
7299         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7300                 inode->i_blkbits;
7301         bh_result->b_size = len;
7302         bh_result->b_bdev = fs_info->fs_devices->latest_bdev;
7303         set_buffer_mapped(bh_result);
7304
7305         if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7306                 set_buffer_new(bh_result);
7307
7308         /*
7309          * Need to update the i_size under the extent lock so buffered
7310          * readers will get the updated i_size when we unlock.
7311          */
7312         if (!dio_data->overwrite && start + len > i_size_read(inode))
7313                 i_size_write(inode, start + len);
7314
7315         WARN_ON(dio_data->reserve < len);
7316         dio_data->reserve -= len;
7317         dio_data->unsubmitted_oe_range_end = start + len;
7318         current->journal_info = dio_data;
7319 out:
7320         return ret;
7321 }
7322
7323 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7324                                    struct buffer_head *bh_result, int create)
7325 {
7326         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7327         struct extent_map *em;
7328         struct extent_state *cached_state = NULL;
7329         struct btrfs_dio_data *dio_data = NULL;
7330         u64 start = iblock << inode->i_blkbits;
7331         u64 lockstart, lockend;
7332         u64 len = bh_result->b_size;
7333         int ret = 0;
7334
7335         if (!create)
7336                 len = min_t(u64, len, fs_info->sectorsize);
7337
7338         lockstart = start;
7339         lockend = start + len - 1;
7340
7341         if (current->journal_info) {
7342                 /*
7343                  * Need to pull our outstanding extents and set journal_info to NULL so
7344                  * that anything that needs to check if there's a transaction doesn't get
7345                  * confused.
7346                  */
7347                 dio_data = current->journal_info;
7348                 current->journal_info = NULL;
7349         }
7350
7351         /*
7352          * If this errors out it's because we couldn't invalidate pagecache for
7353          * this range and we need to fallback to buffered.
7354          */
7355         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7356                                create)) {
7357                 ret = -ENOTBLK;
7358                 goto err;
7359         }
7360
7361         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len);
7362         if (IS_ERR(em)) {
7363                 ret = PTR_ERR(em);
7364                 goto unlock_err;
7365         }
7366
7367         /*
7368          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7369          * io.  INLINE is special, and we could probably kludge it in here, but
7370          * it's still buffered so for safety lets just fall back to the generic
7371          * buffered path.
7372          *
7373          * For COMPRESSED we _have_ to read the entire extent in so we can
7374          * decompress it, so there will be buffering required no matter what we
7375          * do, so go ahead and fallback to buffered.
7376          *
7377          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7378          * to buffered IO.  Don't blame me, this is the price we pay for using
7379          * the generic code.
7380          */
7381         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7382             em->block_start == EXTENT_MAP_INLINE) {
7383                 free_extent_map(em);
7384                 ret = -ENOTBLK;
7385                 goto unlock_err;
7386         }
7387
7388         if (create) {
7389                 ret = btrfs_get_blocks_direct_write(&em, bh_result, inode,
7390                                                     dio_data, start, len);
7391                 if (ret < 0)
7392                         goto unlock_err;
7393
7394                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
7395                                      lockend, &cached_state);
7396         } else {
7397                 ret = btrfs_get_blocks_direct_read(em, bh_result, inode,
7398                                                    start, len);
7399                 /* Can be negative only if we read from a hole */
7400                 if (ret < 0) {
7401                         ret = 0;
7402                         free_extent_map(em);
7403                         goto unlock_err;
7404                 }
7405                 /*
7406                  * We need to unlock only the end area that we aren't using.
7407                  * The rest is going to be unlocked by the endio routine.
7408                  */
7409                 lockstart = start + bh_result->b_size;
7410                 if (lockstart < lockend) {
7411                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
7412                                              lockstart, lockend, &cached_state);
7413                 } else {
7414                         free_extent_state(cached_state);
7415                 }
7416         }
7417
7418         free_extent_map(em);
7419
7420         return 0;
7421
7422 unlock_err:
7423         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7424                              &cached_state);
7425 err:
7426         if (dio_data)
7427                 current->journal_info = dio_data;
7428         return ret;
7429 }
7430
7431 static void btrfs_dio_private_put(struct btrfs_dio_private *dip)
7432 {
7433         /*
7434          * This implies a barrier so that stores to dio_bio->bi_status before
7435          * this and loads of dio_bio->bi_status after this are fully ordered.
7436          */
7437         if (!refcount_dec_and_test(&dip->refs))
7438                 return;
7439
7440         if (bio_op(dip->dio_bio) == REQ_OP_WRITE) {
7441                 __endio_write_update_ordered(dip->inode, dip->logical_offset,
7442                                              dip->bytes,
7443                                              !dip->dio_bio->bi_status);
7444         } else {
7445                 unlock_extent(&BTRFS_I(dip->inode)->io_tree,
7446                               dip->logical_offset,
7447                               dip->logical_offset + dip->bytes - 1);
7448         }
7449
7450         dio_end_io(dip->dio_bio);
7451         kfree(dip);
7452 }
7453
7454 static blk_status_t submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7455                                           int mirror_num,
7456                                           unsigned long bio_flags)
7457 {
7458         struct btrfs_dio_private *dip = bio->bi_private;
7459         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7460         blk_status_t ret;
7461
7462         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7463
7464         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
7465         if (ret)
7466                 return ret;
7467
7468         refcount_inc(&dip->refs);
7469         ret = btrfs_map_bio(fs_info, bio, mirror_num);
7470         if (ret)
7471                 refcount_dec(&dip->refs);
7472         return ret;
7473 }
7474
7475 static blk_status_t btrfs_check_read_dio_bio(struct inode *inode,
7476                                              struct btrfs_io_bio *io_bio,
7477                                              const bool uptodate)
7478 {
7479         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
7480         const u32 sectorsize = fs_info->sectorsize;
7481         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
7482         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7483         const bool csum = !(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM);
7484         struct bio_vec bvec;
7485         struct bvec_iter iter;
7486         u64 start = io_bio->logical;
7487         int icsum = 0;
7488         blk_status_t err = BLK_STS_OK;
7489
7490         __bio_for_each_segment(bvec, &io_bio->bio, iter, io_bio->iter) {
7491                 unsigned int i, nr_sectors, pgoff;
7492
7493                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
7494                 pgoff = bvec.bv_offset;
7495                 for (i = 0; i < nr_sectors; i++) {
7496                         ASSERT(pgoff < PAGE_SIZE);
7497                         if (uptodate &&
7498                             (!csum || !check_data_csum(inode, io_bio, icsum,
7499                                                        bvec.bv_page, pgoff,
7500                                                        start, sectorsize))) {
7501                                 clean_io_failure(fs_info, failure_tree, io_tree,
7502                                                  start, bvec.bv_page,
7503                                                  btrfs_ino(BTRFS_I(inode)),
7504                                                  pgoff);
7505                         } else {
7506                                 blk_status_t status;
7507
7508                                 status = btrfs_submit_read_repair(inode,
7509                                                         &io_bio->bio,
7510                                                         start - io_bio->logical,
7511                                                         bvec.bv_page, pgoff,
7512                                                         start,
7513                                                         start + sectorsize - 1,
7514                                                         io_bio->mirror_num,
7515                                                         submit_dio_repair_bio);
7516                                 if (status)
7517                                         err = status;
7518                         }
7519                         start += sectorsize;
7520                         icsum++;
7521                         pgoff += sectorsize;
7522                 }
7523         }
7524         return err;
7525 }
7526
7527 static void __endio_write_update_ordered(struct inode *inode,
7528                                          const u64 offset, const u64 bytes,
7529                                          const bool uptodate)
7530 {
7531         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7532         struct btrfs_ordered_extent *ordered = NULL;
7533         struct btrfs_workqueue *wq;
7534         u64 ordered_offset = offset;
7535         u64 ordered_bytes = bytes;
7536         u64 last_offset;
7537
7538         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
7539                 wq = fs_info->endio_freespace_worker;
7540         else
7541                 wq = fs_info->endio_write_workers;
7542
7543         while (ordered_offset < offset + bytes) {
7544                 last_offset = ordered_offset;
7545                 if (btrfs_dec_test_first_ordered_pending(inode, &ordered,
7546                                                            &ordered_offset,
7547                                                            ordered_bytes,
7548                                                            uptodate)) {
7549                         btrfs_init_work(&ordered->work, finish_ordered_fn, NULL,
7550                                         NULL);
7551                         btrfs_queue_work(wq, &ordered->work);
7552                 }
7553                 /*
7554                  * If btrfs_dec_test_ordered_pending does not find any ordered
7555                  * extent in the range, we can exit.
7556                  */
7557                 if (ordered_offset == last_offset)
7558                         return;
7559                 /*
7560                  * Our bio might span multiple ordered extents. In this case
7561                  * we keep going until we have accounted the whole dio.
7562                  */
7563                 if (ordered_offset < offset + bytes) {
7564                         ordered_bytes = offset + bytes - ordered_offset;
7565                         ordered = NULL;
7566                 }
7567         }
7568 }
7569
7570 static blk_status_t btrfs_submit_bio_start_direct_io(void *private_data,
7571                                     struct bio *bio, u64 offset)
7572 {
7573         struct inode *inode = private_data;
7574         blk_status_t ret;
7575         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
7576         BUG_ON(ret); /* -ENOMEM */
7577         return 0;
7578 }
7579
7580 static void btrfs_end_dio_bio(struct bio *bio)
7581 {
7582         struct btrfs_dio_private *dip = bio->bi_private;
7583         blk_status_t err = bio->bi_status;
7584
7585         if (err)
7586                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
7587                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
7588                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
7589                            bio->bi_opf,
7590                            (unsigned long long)bio->bi_iter.bi_sector,
7591                            bio->bi_iter.bi_size, err);
7592
7593         if (bio_op(bio) == REQ_OP_READ) {
7594                 err = btrfs_check_read_dio_bio(dip->inode, btrfs_io_bio(bio),
7595                                                !err);
7596         }
7597
7598         if (err)
7599                 dip->dio_bio->bi_status = err;
7600
7601         bio_put(bio);
7602         btrfs_dio_private_put(dip);
7603 }
7604
7605 static inline blk_status_t btrfs_submit_dio_bio(struct bio *bio,
7606                 struct inode *inode, u64 file_offset, int async_submit)
7607 {
7608         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7609         struct btrfs_dio_private *dip = bio->bi_private;
7610         bool write = bio_op(bio) == REQ_OP_WRITE;
7611         blk_status_t ret;
7612
7613         /* Check btrfs_submit_bio_hook() for rules about async submit. */
7614         if (async_submit)
7615                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
7616
7617         if (!write) {
7618                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
7619                 if (ret)
7620                         goto err;
7621         }
7622
7623         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
7624                 goto map;
7625
7626         if (write && async_submit) {
7627                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
7628                                           file_offset, inode,
7629                                           btrfs_submit_bio_start_direct_io);
7630                 goto err;
7631         } else if (write) {
7632                 /*
7633                  * If we aren't doing async submit, calculate the csum of the
7634                  * bio now.
7635                  */
7636                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
7637                 if (ret)
7638                         goto err;
7639         } else {
7640                 u64 csum_offset;
7641
7642                 csum_offset = file_offset - dip->logical_offset;
7643                 csum_offset >>= inode->i_sb->s_blocksize_bits;
7644                 csum_offset *= btrfs_super_csum_size(fs_info->super_copy);
7645                 btrfs_io_bio(bio)->csum = dip->csums + csum_offset;
7646         }
7647 map:
7648         ret = btrfs_map_bio(fs_info, bio, 0);
7649 err:
7650         return ret;
7651 }
7652
7653 /*
7654  * If this succeeds, the btrfs_dio_private is responsible for cleaning up locked
7655  * or ordered extents whether or not we submit any bios.
7656  */
7657 static struct btrfs_dio_private *btrfs_create_dio_private(struct bio *dio_bio,
7658                                                           struct inode *inode,
7659                                                           loff_t file_offset)
7660 {
7661         const bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
7662         const bool csum = !(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM);
7663         size_t dip_size;
7664         struct btrfs_dio_private *dip;
7665
7666         dip_size = sizeof(*dip);
7667         if (!write && csum) {
7668                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7669                 const u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
7670                 size_t nblocks;
7671
7672                 nblocks = dio_bio->bi_iter.bi_size >> inode->i_sb->s_blocksize_bits;
7673                 dip_size += csum_size * nblocks;
7674         }
7675
7676         dip = kzalloc(dip_size, GFP_NOFS);
7677         if (!dip)
7678                 return NULL;
7679
7680         dip->inode = inode;
7681         dip->logical_offset = file_offset;
7682         dip->bytes = dio_bio->bi_iter.bi_size;
7683         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
7684         dip->dio_bio = dio_bio;
7685         refcount_set(&dip->refs, 1);
7686
7687         if (write) {
7688                 struct btrfs_dio_data *dio_data = current->journal_info;
7689
7690                 /*
7691                  * Setting range start and end to the same value means that
7692                  * no cleanup will happen in btrfs_direct_IO
7693                  */
7694                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
7695                         dip->bytes;
7696                 dio_data->unsubmitted_oe_range_start =
7697                         dio_data->unsubmitted_oe_range_end;
7698         }
7699         return dip;
7700 }
7701
7702 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
7703                                 loff_t file_offset)
7704 {
7705         const bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
7706         const bool csum = !(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM);
7707         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7708         const bool raid56 = (btrfs_data_alloc_profile(fs_info) &
7709                              BTRFS_BLOCK_GROUP_RAID56_MASK);
7710         struct btrfs_dio_private *dip;
7711         struct bio *bio;
7712         u64 start_sector;
7713         int async_submit = 0;
7714         u64 submit_len;
7715         int clone_offset = 0;
7716         int clone_len;
7717         int ret;
7718         blk_status_t status;
7719         struct btrfs_io_geometry geom;
7720
7721         dip = btrfs_create_dio_private(dio_bio, inode, file_offset);
7722         if (!dip) {
7723                 if (!write) {
7724                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
7725                                 file_offset + dio_bio->bi_iter.bi_size - 1);
7726                 }
7727                 dio_bio->bi_status = BLK_STS_RESOURCE;
7728                 dio_end_io(dio_bio);
7729                 return;
7730         }
7731
7732         if (!write && csum) {
7733                 /*
7734                  * Load the csums up front to reduce csum tree searches and
7735                  * contention when submitting bios.
7736                  */
7737                 status = btrfs_lookup_bio_sums(inode, dio_bio, file_offset,
7738                                                dip->csums);
7739                 if (status != BLK_STS_OK)
7740                         goto out_err;
7741         }
7742
7743         start_sector = dio_bio->bi_iter.bi_sector;
7744         submit_len = dio_bio->bi_iter.bi_size;
7745
7746         do {
7747                 ret = btrfs_get_io_geometry(fs_info, btrfs_op(dio_bio),
7748                                             start_sector << 9, submit_len,
7749                                             &geom);
7750                 if (ret) {
7751                         status = errno_to_blk_status(ret);
7752                         goto out_err;
7753                 }
7754                 ASSERT(geom.len <= INT_MAX);
7755
7756                 clone_len = min_t(int, submit_len, geom.len);
7757
7758                 /*
7759                  * This will never fail as it's passing GPF_NOFS and
7760                  * the allocation is backed by btrfs_bioset.
7761                  */
7762                 bio = btrfs_bio_clone_partial(dio_bio, clone_offset, clone_len);
7763                 bio->bi_private = dip;
7764                 bio->bi_end_io = btrfs_end_dio_bio;
7765                 btrfs_io_bio(bio)->logical = file_offset;
7766
7767                 ASSERT(submit_len >= clone_len);
7768                 submit_len -= clone_len;
7769
7770                 /*
7771                  * Increase the count before we submit the bio so we know
7772                  * the end IO handler won't happen before we increase the
7773                  * count. Otherwise, the dip might get freed before we're
7774                  * done setting it up.
7775                  *
7776                  * We transfer the initial reference to the last bio, so we
7777                  * don't need to increment the reference count for the last one.
7778                  */
7779                 if (submit_len > 0) {
7780                         refcount_inc(&dip->refs);
7781                         /*
7782                          * If we are submitting more than one bio, submit them
7783                          * all asynchronously. The exception is RAID 5 or 6, as
7784                          * asynchronous checksums make it difficult to collect
7785                          * full stripe writes.
7786                          */
7787                         if (!raid56)
7788                                 async_submit = 1;
7789                 }
7790
7791                 status = btrfs_submit_dio_bio(bio, inode, file_offset,
7792                                                 async_submit);
7793                 if (status) {
7794                         bio_put(bio);
7795                         if (submit_len > 0)
7796                                 refcount_dec(&dip->refs);
7797                         goto out_err;
7798                 }
7799
7800                 clone_offset += clone_len;
7801                 start_sector += clone_len >> 9;
7802                 file_offset += clone_len;
7803         } while (submit_len > 0);
7804         return;
7805
7806 out_err:
7807         dip->dio_bio->bi_status = status;
7808         btrfs_dio_private_put(dip);
7809 }
7810
7811 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
7812                                const struct iov_iter *iter, loff_t offset)
7813 {
7814         int seg;
7815         int i;
7816         unsigned int blocksize_mask = fs_info->sectorsize - 1;
7817         ssize_t retval = -EINVAL;
7818
7819         if (offset & blocksize_mask)
7820                 goto out;
7821
7822         if (iov_iter_alignment(iter) & blocksize_mask)
7823                 goto out;
7824
7825         /* If this is a write we don't need to check anymore */
7826         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
7827                 return 0;
7828         /*
7829          * Check to make sure we don't have duplicate iov_base's in this
7830          * iovec, if so return EINVAL, otherwise we'll get csum errors
7831          * when reading back.
7832          */
7833         for (seg = 0; seg < iter->nr_segs; seg++) {
7834                 for (i = seg + 1; i < iter->nr_segs; i++) {
7835                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
7836                                 goto out;
7837                 }
7838         }
7839         retval = 0;
7840 out:
7841         return retval;
7842 }
7843
7844 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
7845 {
7846         struct file *file = iocb->ki_filp;
7847         struct inode *inode = file->f_mapping->host;
7848         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7849         struct btrfs_dio_data dio_data = { 0 };
7850         struct extent_changeset *data_reserved = NULL;
7851         loff_t offset = iocb->ki_pos;
7852         size_t count = 0;
7853         int flags = 0;
7854         bool wakeup = true;
7855         bool relock = false;
7856         ssize_t ret;
7857
7858         if (check_direct_IO(fs_info, iter, offset))
7859                 return 0;
7860
7861         inode_dio_begin(inode);
7862
7863         /*
7864          * The generic stuff only does filemap_write_and_wait_range, which
7865          * isn't enough if we've written compressed pages to this area, so
7866          * we need to flush the dirty pages again to make absolutely sure
7867          * that any outstanding dirty pages are on disk.
7868          */
7869         count = iov_iter_count(iter);
7870         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
7871                      &BTRFS_I(inode)->runtime_flags))
7872                 filemap_fdatawrite_range(inode->i_mapping, offset,
7873                                          offset + count - 1);
7874
7875         if (iov_iter_rw(iter) == WRITE) {
7876                 /*
7877                  * If the write DIO is beyond the EOF, we need update
7878                  * the isize, but it is protected by i_mutex. So we can
7879                  * not unlock the i_mutex at this case.
7880                  */
7881                 if (offset + count <= inode->i_size) {
7882                         dio_data.overwrite = 1;
7883                         inode_unlock(inode);
7884                         relock = true;
7885                 }
7886                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
7887                                                    offset, count);
7888                 if (ret)
7889                         goto out;
7890
7891                 /*
7892                  * We need to know how many extents we reserved so that we can
7893                  * do the accounting properly if we go over the number we
7894                  * originally calculated.  Abuse current->journal_info for this.
7895                  */
7896                 dio_data.reserve = round_up(count,
7897                                             fs_info->sectorsize);
7898                 dio_data.unsubmitted_oe_range_start = (u64)offset;
7899                 dio_data.unsubmitted_oe_range_end = (u64)offset;
7900                 current->journal_info = &dio_data;
7901                 down_read(&BTRFS_I(inode)->dio_sem);
7902         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
7903                                      &BTRFS_I(inode)->runtime_flags)) {
7904                 inode_dio_end(inode);
7905                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
7906                 wakeup = false;
7907         }
7908
7909         ret = __blockdev_direct_IO(iocb, inode,
7910                                    fs_info->fs_devices->latest_bdev,
7911                                    iter, btrfs_get_blocks_direct, NULL,
7912                                    btrfs_submit_direct, flags);
7913         if (iov_iter_rw(iter) == WRITE) {
7914                 up_read(&BTRFS_I(inode)->dio_sem);
7915                 current->journal_info = NULL;
7916                 if (ret < 0 && ret != -EIOCBQUEUED) {
7917                         if (dio_data.reserve)
7918                                 btrfs_delalloc_release_space(inode, data_reserved,
7919                                         offset, dio_data.reserve, true);
7920                         /*
7921                          * On error we might have left some ordered extents
7922                          * without submitting corresponding bios for them, so
7923                          * cleanup them up to avoid other tasks getting them
7924                          * and waiting for them to complete forever.
7925                          */
7926                         if (dio_data.unsubmitted_oe_range_start <
7927                             dio_data.unsubmitted_oe_range_end)
7928                                 __endio_write_update_ordered(inode,
7929                                         dio_data.unsubmitted_oe_range_start,
7930                                         dio_data.unsubmitted_oe_range_end -
7931                                         dio_data.unsubmitted_oe_range_start,
7932                                         false);
7933                 } else if (ret >= 0 && (size_t)ret < count)
7934                         btrfs_delalloc_release_space(inode, data_reserved,
7935                                         offset, count - (size_t)ret, true);
7936                 btrfs_delalloc_release_extents(BTRFS_I(inode), count);
7937         }
7938 out:
7939         if (wakeup)
7940                 inode_dio_end(inode);
7941         if (relock)
7942                 inode_lock(inode);
7943
7944         extent_changeset_free(data_reserved);
7945         return ret;
7946 }
7947
7948 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
7949                 __u64 start, __u64 len)
7950 {
7951         int     ret;
7952
7953         ret = fiemap_prep(inode, fieinfo, start, &len, 0);
7954         if (ret)
7955                 return ret;
7956
7957         return extent_fiemap(inode, fieinfo, start, len);
7958 }
7959
7960 int btrfs_readpage(struct file *file, struct page *page)
7961 {
7962         return extent_read_full_page(page, btrfs_get_extent, 0);
7963 }
7964
7965 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
7966 {
7967         struct inode *inode = page->mapping->host;
7968         int ret;
7969
7970         if (current->flags & PF_MEMALLOC) {
7971                 redirty_page_for_writepage(wbc, page);
7972                 unlock_page(page);
7973                 return 0;
7974         }
7975
7976         /*
7977          * If we are under memory pressure we will call this directly from the
7978          * VM, we need to make sure we have the inode referenced for the ordered
7979          * extent.  If not just return like we didn't do anything.
7980          */
7981         if (!igrab(inode)) {
7982                 redirty_page_for_writepage(wbc, page);
7983                 return AOP_WRITEPAGE_ACTIVATE;
7984         }
7985         ret = extent_write_full_page(page, wbc);
7986         btrfs_add_delayed_iput(inode);
7987         return ret;
7988 }
7989
7990 static int btrfs_writepages(struct address_space *mapping,
7991                             struct writeback_control *wbc)
7992 {
7993         return extent_writepages(mapping, wbc);
7994 }
7995
7996 static void btrfs_readahead(struct readahead_control *rac)
7997 {
7998         extent_readahead(rac);
7999 }
8000
8001 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8002 {
8003         int ret = try_release_extent_mapping(page, gfp_flags);
8004         if (ret == 1)
8005                 detach_page_private(page);
8006         return ret;
8007 }
8008
8009 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8010 {
8011         if (PageWriteback(page) || PageDirty(page))
8012                 return 0;
8013         return __btrfs_releasepage(page, gfp_flags);
8014 }
8015
8016 #ifdef CONFIG_MIGRATION
8017 static int btrfs_migratepage(struct address_space *mapping,
8018                              struct page *newpage, struct page *page,
8019                              enum migrate_mode mode)
8020 {
8021         int ret;
8022
8023         ret = migrate_page_move_mapping(mapping, newpage, page, 0);
8024         if (ret != MIGRATEPAGE_SUCCESS)
8025                 return ret;
8026
8027         if (page_has_private(page))
8028                 attach_page_private(newpage, detach_page_private(page));
8029
8030         if (PagePrivate2(page)) {
8031                 ClearPagePrivate2(page);
8032                 SetPagePrivate2(newpage);
8033         }
8034
8035         if (mode != MIGRATE_SYNC_NO_COPY)
8036                 migrate_page_copy(newpage, page);
8037         else
8038                 migrate_page_states(newpage, page);
8039         return MIGRATEPAGE_SUCCESS;
8040 }
8041 #endif
8042
8043 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8044                                  unsigned int length)
8045 {
8046         struct inode *inode = page->mapping->host;
8047         struct extent_io_tree *tree;
8048         struct btrfs_ordered_extent *ordered;
8049         struct extent_state *cached_state = NULL;
8050         u64 page_start = page_offset(page);
8051         u64 page_end = page_start + PAGE_SIZE - 1;
8052         u64 start;
8053         u64 end;
8054         int inode_evicting = inode->i_state & I_FREEING;
8055
8056         /*
8057          * we have the page locked, so new writeback can't start,
8058          * and the dirty bit won't be cleared while we are here.
8059          *
8060          * Wait for IO on this page so that we can safely clear
8061          * the PagePrivate2 bit and do ordered accounting
8062          */
8063         wait_on_page_writeback(page);
8064
8065         tree = &BTRFS_I(inode)->io_tree;
8066         if (offset) {
8067                 btrfs_releasepage(page, GFP_NOFS);
8068                 return;
8069         }
8070
8071         if (!inode_evicting)
8072                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8073 again:
8074         start = page_start;
8075         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8076                                         page_end - start + 1);
8077         if (ordered) {
8078                 end = min(page_end,
8079                           ordered->file_offset + ordered->num_bytes - 1);
8080                 /*
8081                  * IO on this page will never be started, so we need
8082                  * to account for any ordered extents now
8083                  */
8084                 if (!inode_evicting)
8085                         clear_extent_bit(tree, start, end,
8086                                          EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8087                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8088                                          EXTENT_DEFRAG, 1, 0, &cached_state);
8089                 /*
8090                  * whoever cleared the private bit is responsible
8091                  * for the finish_ordered_io
8092                  */
8093                 if (TestClearPagePrivate2(page)) {
8094                         struct btrfs_ordered_inode_tree *tree;
8095                         u64 new_len;
8096
8097                         tree = &BTRFS_I(inode)->ordered_tree;
8098
8099                         spin_lock_irq(&tree->lock);
8100                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8101                         new_len = start - ordered->file_offset;
8102                         if (new_len < ordered->truncated_len)
8103                                 ordered->truncated_len = new_len;
8104                         spin_unlock_irq(&tree->lock);
8105
8106                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8107                                                            start,
8108                                                            end - start + 1, 1))
8109                                 btrfs_finish_ordered_io(ordered);
8110                 }
8111                 btrfs_put_ordered_extent(ordered);
8112                 if (!inode_evicting) {
8113                         cached_state = NULL;
8114                         lock_extent_bits(tree, start, end,
8115                                          &cached_state);
8116                 }
8117
8118                 start = end + 1;
8119                 if (start < page_end)
8120                         goto again;
8121         }
8122
8123         /*
8124          * Qgroup reserved space handler
8125          * Page here will be either
8126          * 1) Already written to disk
8127          *    In this case, its reserved space is released from data rsv map
8128          *    and will be freed by delayed_ref handler finally.
8129          *    So even we call qgroup_free_data(), it won't decrease reserved
8130          *    space.
8131          * 2) Not written to disk
8132          *    This means the reserved space should be freed here. However,
8133          *    if a truncate invalidates the page (by clearing PageDirty)
8134          *    and the page is accounted for while allocating extent
8135          *    in btrfs_check_data_free_space() we let delayed_ref to
8136          *    free the entire extent.
8137          */
8138         if (PageDirty(page))
8139                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
8140         if (!inode_evicting) {
8141                 clear_extent_bit(tree, page_start, page_end, EXTENT_LOCKED |
8142                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8143                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
8144                                  &cached_state);
8145
8146                 __btrfs_releasepage(page, GFP_NOFS);
8147         }
8148
8149         ClearPageChecked(page);
8150         detach_page_private(page);
8151 }
8152
8153 /*
8154  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8155  * called from a page fault handler when a page is first dirtied. Hence we must
8156  * be careful to check for EOF conditions here. We set the page up correctly
8157  * for a written page which means we get ENOSPC checking when writing into
8158  * holes and correct delalloc and unwritten extent mapping on filesystems that
8159  * support these features.
8160  *
8161  * We are not allowed to take the i_mutex here so we have to play games to
8162  * protect against truncate races as the page could now be beyond EOF.  Because
8163  * truncate_setsize() writes the inode size before removing pages, once we have
8164  * the page lock we can determine safely if the page is beyond EOF. If it is not
8165  * beyond EOF, then the page is guaranteed safe against truncation until we
8166  * unlock the page.
8167  */
8168 vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
8169 {
8170         struct page *page = vmf->page;
8171         struct inode *inode = file_inode(vmf->vma->vm_file);
8172         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8173         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8174         struct btrfs_ordered_extent *ordered;
8175         struct extent_state *cached_state = NULL;
8176         struct extent_changeset *data_reserved = NULL;
8177         char *kaddr;
8178         unsigned long zero_start;
8179         loff_t size;
8180         vm_fault_t ret;
8181         int ret2;
8182         int reserved = 0;
8183         u64 reserved_space;
8184         u64 page_start;
8185         u64 page_end;
8186         u64 end;
8187
8188         reserved_space = PAGE_SIZE;
8189
8190         sb_start_pagefault(inode->i_sb);
8191         page_start = page_offset(page);
8192         page_end = page_start + PAGE_SIZE - 1;
8193         end = page_end;
8194
8195         /*
8196          * Reserving delalloc space after obtaining the page lock can lead to
8197          * deadlock. For example, if a dirty page is locked by this function
8198          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8199          * dirty page write out, then the btrfs_writepage() function could
8200          * end up waiting indefinitely to get a lock on the page currently
8201          * being processed by btrfs_page_mkwrite() function.
8202          */
8203         ret2 = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
8204                                            reserved_space);
8205         if (!ret2) {
8206                 ret2 = file_update_time(vmf->vma->vm_file);
8207                 reserved = 1;
8208         }
8209         if (ret2) {
8210                 ret = vmf_error(ret2);
8211                 if (reserved)
8212                         goto out;
8213                 goto out_noreserve;
8214         }
8215
8216         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8217 again:
8218         lock_page(page);
8219         size = i_size_read(inode);
8220
8221         if ((page->mapping != inode->i_mapping) ||
8222             (page_start >= size)) {
8223                 /* page got truncated out from underneath us */
8224                 goto out_unlock;
8225         }
8226         wait_on_page_writeback(page);
8227
8228         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8229         set_page_extent_mapped(page);
8230
8231         /*
8232          * we can't set the delalloc bits if there are pending ordered
8233          * extents.  Drop our locks and wait for them to finish
8234          */
8235         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
8236                         PAGE_SIZE);
8237         if (ordered) {
8238                 unlock_extent_cached(io_tree, page_start, page_end,
8239                                      &cached_state);
8240                 unlock_page(page);
8241                 btrfs_start_ordered_extent(inode, ordered, 1);
8242                 btrfs_put_ordered_extent(ordered);
8243                 goto again;
8244         }
8245
8246         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
8247                 reserved_space = round_up(size - page_start,
8248                                           fs_info->sectorsize);
8249                 if (reserved_space < PAGE_SIZE) {
8250                         end = page_start + reserved_space - 1;
8251                         btrfs_delalloc_release_space(inode, data_reserved,
8252                                         page_start, PAGE_SIZE - reserved_space,
8253                                         true);
8254                 }
8255         }
8256
8257         /*
8258          * page_mkwrite gets called when the page is firstly dirtied after it's
8259          * faulted in, but write(2) could also dirty a page and set delalloc
8260          * bits, thus in this case for space account reason, we still need to
8261          * clear any delalloc bits within this page range since we have to
8262          * reserve data&meta space before lock_page() (see above comments).
8263          */
8264         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
8265                           EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8266                           EXTENT_DEFRAG, 0, 0, &cached_state);
8267
8268         ret2 = btrfs_set_extent_delalloc(inode, page_start, end, 0,
8269                                         &cached_state);
8270         if (ret2) {
8271                 unlock_extent_cached(io_tree, page_start, page_end,
8272                                      &cached_state);
8273                 ret = VM_FAULT_SIGBUS;
8274                 goto out_unlock;
8275         }
8276
8277         /* page is wholly or partially inside EOF */
8278         if (page_start + PAGE_SIZE > size)
8279                 zero_start = offset_in_page(size);
8280         else
8281                 zero_start = PAGE_SIZE;
8282
8283         if (zero_start != PAGE_SIZE) {
8284                 kaddr = kmap(page);
8285                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
8286                 flush_dcache_page(page);
8287                 kunmap(page);
8288         }
8289         ClearPageChecked(page);
8290         set_page_dirty(page);
8291         SetPageUptodate(page);
8292
8293         BTRFS_I(inode)->last_trans = fs_info->generation;
8294         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
8295         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
8296
8297         unlock_extent_cached(io_tree, page_start, page_end, &cached_state);
8298
8299         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
8300         sb_end_pagefault(inode->i_sb);
8301         extent_changeset_free(data_reserved);
8302         return VM_FAULT_LOCKED;
8303
8304 out_unlock:
8305         unlock_page(page);
8306 out:
8307         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
8308         btrfs_delalloc_release_space(inode, data_reserved, page_start,
8309                                      reserved_space, (ret != 0));
8310 out_noreserve:
8311         sb_end_pagefault(inode->i_sb);
8312         extent_changeset_free(data_reserved);
8313         return ret;
8314 }
8315
8316 static int btrfs_truncate(struct inode *inode, bool skip_writeback)
8317 {
8318         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8319         struct btrfs_root *root = BTRFS_I(inode)->root;
8320         struct btrfs_block_rsv *rsv;
8321         int ret;
8322         struct btrfs_trans_handle *trans;
8323         u64 mask = fs_info->sectorsize - 1;
8324         u64 min_size = btrfs_calc_metadata_size(fs_info, 1);
8325
8326         if (!skip_writeback) {
8327                 ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
8328                                                (u64)-1);
8329                 if (ret)
8330                         return ret;
8331         }
8332
8333         /*
8334          * Yes ladies and gentlemen, this is indeed ugly.  We have a couple of
8335          * things going on here:
8336          *
8337          * 1) We need to reserve space to update our inode.
8338          *
8339          * 2) We need to have something to cache all the space that is going to
8340          * be free'd up by the truncate operation, but also have some slack
8341          * space reserved in case it uses space during the truncate (thank you
8342          * very much snapshotting).
8343          *
8344          * And we need these to be separate.  The fact is we can use a lot of
8345          * space doing the truncate, and we have no earthly idea how much space
8346          * we will use, so we need the truncate reservation to be separate so it
8347          * doesn't end up using space reserved for updating the inode.  We also
8348          * need to be able to stop the transaction and start a new one, which
8349          * means we need to be able to update the inode several times, and we
8350          * have no idea of knowing how many times that will be, so we can't just
8351          * reserve 1 item for the entirety of the operation, so that has to be
8352          * done separately as well.
8353          *
8354          * So that leaves us with
8355          *
8356          * 1) rsv - for the truncate reservation, which we will steal from the
8357          * transaction reservation.
8358          * 2) fs_info->trans_block_rsv - this will have 1 items worth left for
8359          * updating the inode.
8360          */
8361         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
8362         if (!rsv)
8363                 return -ENOMEM;
8364         rsv->size = min_size;
8365         rsv->failfast = 1;
8366
8367         /*
8368          * 1 for the truncate slack space
8369          * 1 for updating the inode.
8370          */
8371         trans = btrfs_start_transaction(root, 2);
8372         if (IS_ERR(trans)) {
8373                 ret = PTR_ERR(trans);
8374                 goto out;
8375         }
8376
8377         /* Migrate the slack space for the truncate to our reserve */
8378         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
8379                                       min_size, false);
8380         BUG_ON(ret);
8381
8382         /*
8383          * So if we truncate and then write and fsync we normally would just
8384          * write the extents that changed, which is a problem if we need to
8385          * first truncate that entire inode.  So set this flag so we write out
8386          * all of the extents in the inode to the sync log so we're completely
8387          * safe.
8388          */
8389         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
8390         trans->block_rsv = rsv;
8391
8392         while (1) {
8393                 ret = btrfs_truncate_inode_items(trans, root, inode,
8394                                                  inode->i_size,
8395                                                  BTRFS_EXTENT_DATA_KEY);
8396                 trans->block_rsv = &fs_info->trans_block_rsv;
8397                 if (ret != -ENOSPC && ret != -EAGAIN)
8398                         break;
8399
8400                 ret = btrfs_update_inode(trans, root, inode);
8401                 if (ret)
8402                         break;
8403
8404                 btrfs_end_transaction(trans);
8405                 btrfs_btree_balance_dirty(fs_info);
8406
8407                 trans = btrfs_start_transaction(root, 2);
8408                 if (IS_ERR(trans)) {
8409                         ret = PTR_ERR(trans);
8410                         trans = NULL;
8411                         break;
8412                 }
8413
8414                 btrfs_block_rsv_release(fs_info, rsv, -1, NULL);
8415                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
8416                                               rsv, min_size, false);
8417                 BUG_ON(ret);    /* shouldn't happen */
8418                 trans->block_rsv = rsv;
8419         }
8420
8421         /*
8422          * We can't call btrfs_truncate_block inside a trans handle as we could
8423          * deadlock with freeze, if we got NEED_TRUNCATE_BLOCK then we know
8424          * we've truncated everything except the last little bit, and can do
8425          * btrfs_truncate_block and then update the disk_i_size.
8426          */
8427         if (ret == NEED_TRUNCATE_BLOCK) {
8428                 btrfs_end_transaction(trans);
8429                 btrfs_btree_balance_dirty(fs_info);
8430
8431                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
8432                 if (ret)
8433                         goto out;
8434                 trans = btrfs_start_transaction(root, 1);
8435                 if (IS_ERR(trans)) {
8436                         ret = PTR_ERR(trans);
8437                         goto out;
8438                 }
8439                 btrfs_inode_safe_disk_i_size_write(inode, 0);
8440         }
8441
8442         if (trans) {
8443                 int ret2;
8444
8445                 trans->block_rsv = &fs_info->trans_block_rsv;
8446                 ret2 = btrfs_update_inode(trans, root, inode);
8447                 if (ret2 && !ret)
8448                         ret = ret2;
8449
8450                 ret2 = btrfs_end_transaction(trans);
8451                 if (ret2 && !ret)
8452                         ret = ret2;
8453                 btrfs_btree_balance_dirty(fs_info);
8454         }
8455 out:
8456         btrfs_free_block_rsv(fs_info, rsv);
8457
8458         return ret;
8459 }
8460
8461 /*
8462  * create a new subvolume directory/inode (helper for the ioctl).
8463  */
8464 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
8465                              struct btrfs_root *new_root,
8466                              struct btrfs_root *parent_root,
8467                              u64 new_dirid)
8468 {
8469         struct inode *inode;
8470         int err;
8471         u64 index = 0;
8472
8473         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
8474                                 new_dirid, new_dirid,
8475                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
8476                                 &index);
8477         if (IS_ERR(inode))
8478                 return PTR_ERR(inode);
8479         inode->i_op = &btrfs_dir_inode_operations;
8480         inode->i_fop = &btrfs_dir_file_operations;
8481
8482         set_nlink(inode, 1);
8483         btrfs_i_size_write(BTRFS_I(inode), 0);
8484         unlock_new_inode(inode);
8485
8486         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
8487         if (err)
8488                 btrfs_err(new_root->fs_info,
8489                           "error inheriting subvolume %llu properties: %d",
8490                           new_root->root_key.objectid, err);
8491
8492         err = btrfs_update_inode(trans, new_root, inode);
8493
8494         iput(inode);
8495         return err;
8496 }
8497
8498 struct inode *btrfs_alloc_inode(struct super_block *sb)
8499 {
8500         struct btrfs_fs_info *fs_info = btrfs_sb(sb);
8501         struct btrfs_inode *ei;
8502         struct inode *inode;
8503
8504         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_KERNEL);
8505         if (!ei)
8506                 return NULL;
8507
8508         ei->root = NULL;
8509         ei->generation = 0;
8510         ei->last_trans = 0;
8511         ei->last_sub_trans = 0;
8512         ei->logged_trans = 0;
8513         ei->delalloc_bytes = 0;
8514         ei->new_delalloc_bytes = 0;
8515         ei->defrag_bytes = 0;
8516         ei->disk_i_size = 0;
8517         ei->flags = 0;
8518         ei->csum_bytes = 0;
8519         ei->index_cnt = (u64)-1;
8520         ei->dir_index = 0;
8521         ei->last_unlink_trans = 0;
8522         ei->last_log_commit = 0;
8523
8524         spin_lock_init(&ei->lock);
8525         ei->outstanding_extents = 0;
8526         if (sb->s_magic != BTRFS_TEST_MAGIC)
8527                 btrfs_init_metadata_block_rsv(fs_info, &ei->block_rsv,
8528                                               BTRFS_BLOCK_RSV_DELALLOC);
8529         ei->runtime_flags = 0;
8530         ei->prop_compress = BTRFS_COMPRESS_NONE;
8531         ei->defrag_compress = BTRFS_COMPRESS_NONE;
8532
8533         ei->delayed_node = NULL;
8534
8535         ei->i_otime.tv_sec = 0;
8536         ei->i_otime.tv_nsec = 0;
8537
8538         inode = &ei->vfs_inode;
8539         extent_map_tree_init(&ei->extent_tree);
8540         extent_io_tree_init(fs_info, &ei->io_tree, IO_TREE_INODE_IO, inode);
8541         extent_io_tree_init(fs_info, &ei->io_failure_tree,
8542                             IO_TREE_INODE_IO_FAILURE, inode);
8543         extent_io_tree_init(fs_info, &ei->file_extent_tree,
8544                             IO_TREE_INODE_FILE_EXTENT, inode);
8545         ei->io_tree.track_uptodate = true;
8546         ei->io_failure_tree.track_uptodate = true;
8547         atomic_set(&ei->sync_writers, 0);
8548         mutex_init(&ei->log_mutex);
8549         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
8550         INIT_LIST_HEAD(&ei->delalloc_inodes);
8551         INIT_LIST_HEAD(&ei->delayed_iput);
8552         RB_CLEAR_NODE(&ei->rb_node);
8553         init_rwsem(&ei->dio_sem);
8554
8555         return inode;
8556 }
8557
8558 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8559 void btrfs_test_destroy_inode(struct inode *inode)
8560 {
8561         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
8562         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
8563 }
8564 #endif
8565
8566 void btrfs_free_inode(struct inode *inode)
8567 {
8568         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
8569 }
8570
8571 void btrfs_destroy_inode(struct inode *inode)
8572 {
8573         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8574         struct btrfs_ordered_extent *ordered;
8575         struct btrfs_root *root = BTRFS_I(inode)->root;
8576
8577         WARN_ON(!hlist_empty(&inode->i_dentry));
8578         WARN_ON(inode->i_data.nrpages);
8579         WARN_ON(BTRFS_I(inode)->block_rsv.reserved);
8580         WARN_ON(BTRFS_I(inode)->block_rsv.size);
8581         WARN_ON(BTRFS_I(inode)->outstanding_extents);
8582         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
8583         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
8584         WARN_ON(BTRFS_I(inode)->csum_bytes);
8585         WARN_ON(BTRFS_I(inode)->defrag_bytes);
8586
8587         /*
8588          * This can happen where we create an inode, but somebody else also
8589          * created the same inode and we need to destroy the one we already
8590          * created.
8591          */
8592         if (!root)
8593                 return;
8594
8595         while (1) {
8596                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
8597                 if (!ordered)
8598                         break;
8599                 else {
8600                         btrfs_err(fs_info,
8601                                   "found ordered extent %llu %llu on inode cleanup",
8602                                   ordered->file_offset, ordered->num_bytes);
8603                         btrfs_remove_ordered_extent(inode, ordered);
8604                         btrfs_put_ordered_extent(ordered);
8605                         btrfs_put_ordered_extent(ordered);
8606                 }
8607         }
8608         btrfs_qgroup_check_reserved_leak(inode);
8609         inode_tree_del(inode);
8610         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
8611         btrfs_inode_clear_file_extent_range(BTRFS_I(inode), 0, (u64)-1);
8612         btrfs_put_root(BTRFS_I(inode)->root);
8613 }
8614
8615 int btrfs_drop_inode(struct inode *inode)
8616 {
8617         struct btrfs_root *root = BTRFS_I(inode)->root;
8618
8619         if (root == NULL)
8620                 return 1;
8621
8622         /* the snap/subvol tree is on deleting */
8623         if (btrfs_root_refs(&root->root_item) == 0)
8624                 return 1;
8625         else
8626                 return generic_drop_inode(inode);
8627 }
8628
8629 static void init_once(void *foo)
8630 {
8631         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
8632
8633         inode_init_once(&ei->vfs_inode);
8634 }
8635
8636 void __cold btrfs_destroy_cachep(void)
8637 {
8638         /*
8639          * Make sure all delayed rcu free inodes are flushed before we
8640          * destroy cache.
8641          */
8642         rcu_barrier();
8643         kmem_cache_destroy(btrfs_inode_cachep);
8644         kmem_cache_destroy(btrfs_trans_handle_cachep);
8645         kmem_cache_destroy(btrfs_path_cachep);
8646         kmem_cache_destroy(btrfs_free_space_cachep);
8647         kmem_cache_destroy(btrfs_free_space_bitmap_cachep);
8648 }
8649
8650 int __init btrfs_init_cachep(void)
8651 {
8652         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
8653                         sizeof(struct btrfs_inode), 0,
8654                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
8655                         init_once);
8656         if (!btrfs_inode_cachep)
8657                 goto fail;
8658
8659         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
8660                         sizeof(struct btrfs_trans_handle), 0,
8661                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
8662         if (!btrfs_trans_handle_cachep)
8663                 goto fail;
8664
8665         btrfs_path_cachep = kmem_cache_create("btrfs_path",
8666                         sizeof(struct btrfs_path), 0,
8667                         SLAB_MEM_SPREAD, NULL);
8668         if (!btrfs_path_cachep)
8669                 goto fail;
8670
8671         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
8672                         sizeof(struct btrfs_free_space), 0,
8673                         SLAB_MEM_SPREAD, NULL);
8674         if (!btrfs_free_space_cachep)
8675                 goto fail;
8676
8677         btrfs_free_space_bitmap_cachep = kmem_cache_create("btrfs_free_space_bitmap",
8678                                                         PAGE_SIZE, PAGE_SIZE,
8679                                                         SLAB_RED_ZONE, NULL);
8680         if (!btrfs_free_space_bitmap_cachep)
8681                 goto fail;
8682
8683         return 0;
8684 fail:
8685         btrfs_destroy_cachep();
8686         return -ENOMEM;
8687 }
8688
8689 static int btrfs_getattr(const struct path *path, struct kstat *stat,
8690                          u32 request_mask, unsigned int flags)
8691 {
8692         u64 delalloc_bytes;
8693         struct inode *inode = d_inode(path->dentry);
8694         u32 blocksize = inode->i_sb->s_blocksize;
8695         u32 bi_flags = BTRFS_I(inode)->flags;
8696
8697         stat->result_mask |= STATX_BTIME;
8698         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
8699         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
8700         if (bi_flags & BTRFS_INODE_APPEND)
8701                 stat->attributes |= STATX_ATTR_APPEND;
8702         if (bi_flags & BTRFS_INODE_COMPRESS)
8703                 stat->attributes |= STATX_ATTR_COMPRESSED;
8704         if (bi_flags & BTRFS_INODE_IMMUTABLE)
8705                 stat->attributes |= STATX_ATTR_IMMUTABLE;
8706         if (bi_flags & BTRFS_INODE_NODUMP)
8707                 stat->attributes |= STATX_ATTR_NODUMP;
8708
8709         stat->attributes_mask |= (STATX_ATTR_APPEND |
8710                                   STATX_ATTR_COMPRESSED |
8711                                   STATX_ATTR_IMMUTABLE |
8712                                   STATX_ATTR_NODUMP);
8713
8714         generic_fillattr(inode, stat);
8715         stat->dev = BTRFS_I(inode)->root->anon_dev;
8716
8717         spin_lock(&BTRFS_I(inode)->lock);
8718         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
8719         spin_unlock(&BTRFS_I(inode)->lock);
8720         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
8721                         ALIGN(delalloc_bytes, blocksize)) >> 9;
8722         return 0;
8723 }
8724
8725 static int btrfs_rename_exchange(struct inode *old_dir,
8726                               struct dentry *old_dentry,
8727                               struct inode *new_dir,
8728                               struct dentry *new_dentry)
8729 {
8730         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
8731         struct btrfs_trans_handle *trans;
8732         struct btrfs_root *root = BTRFS_I(old_dir)->root;
8733         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
8734         struct inode *new_inode = new_dentry->d_inode;
8735         struct inode *old_inode = old_dentry->d_inode;
8736         struct timespec64 ctime = current_time(old_inode);
8737         struct dentry *parent;
8738         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
8739         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
8740         u64 old_idx = 0;
8741         u64 new_idx = 0;
8742         int ret;
8743         bool root_log_pinned = false;
8744         bool dest_log_pinned = false;
8745         struct btrfs_log_ctx ctx_root;
8746         struct btrfs_log_ctx ctx_dest;
8747         bool sync_log_root = false;
8748         bool sync_log_dest = false;
8749         bool commit_transaction = false;
8750
8751         /* we only allow rename subvolume link between subvolumes */
8752         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
8753                 return -EXDEV;
8754
8755         btrfs_init_log_ctx(&ctx_root, old_inode);
8756         btrfs_init_log_ctx(&ctx_dest, new_inode);
8757
8758         /* close the race window with snapshot create/destroy ioctl */
8759         if (old_ino == BTRFS_FIRST_FREE_OBJECTID ||
8760             new_ino == BTRFS_FIRST_FREE_OBJECTID)
8761                 down_read(&fs_info->subvol_sem);
8762
8763         /*
8764          * We want to reserve the absolute worst case amount of items.  So if
8765          * both inodes are subvols and we need to unlink them then that would
8766          * require 4 item modifications, but if they are both normal inodes it
8767          * would require 5 item modifications, so we'll assume their normal
8768          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
8769          * should cover the worst case number of items we'll modify.
8770          */
8771         trans = btrfs_start_transaction(root, 12);
8772         if (IS_ERR(trans)) {
8773                 ret = PTR_ERR(trans);
8774                 goto out_notrans;
8775         }
8776
8777         if (dest != root)
8778                 btrfs_record_root_in_trans(trans, dest);
8779
8780         /*
8781          * We need to find a free sequence number both in the source and
8782          * in the destination directory for the exchange.
8783          */
8784         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
8785         if (ret)
8786                 goto out_fail;
8787         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
8788         if (ret)
8789                 goto out_fail;
8790
8791         BTRFS_I(old_inode)->dir_index = 0ULL;
8792         BTRFS_I(new_inode)->dir_index = 0ULL;
8793
8794         /* Reference for the source. */
8795         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
8796                 /* force full log commit if subvolume involved. */
8797                 btrfs_set_log_full_commit(trans);
8798         } else {
8799                 btrfs_pin_log_trans(root);
8800                 root_log_pinned = true;
8801                 ret = btrfs_insert_inode_ref(trans, dest,
8802                                              new_dentry->d_name.name,
8803                                              new_dentry->d_name.len,
8804                                              old_ino,
8805                                              btrfs_ino(BTRFS_I(new_dir)),
8806                                              old_idx);
8807                 if (ret)
8808                         goto out_fail;
8809         }
8810
8811         /* And now for the dest. */
8812         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
8813                 /* force full log commit if subvolume involved. */
8814                 btrfs_set_log_full_commit(trans);
8815         } else {
8816                 btrfs_pin_log_trans(dest);
8817                 dest_log_pinned = true;
8818                 ret = btrfs_insert_inode_ref(trans, root,
8819                                              old_dentry->d_name.name,
8820                                              old_dentry->d_name.len,
8821                                              new_ino,
8822                                              btrfs_ino(BTRFS_I(old_dir)),
8823                                              new_idx);
8824                 if (ret)
8825                         goto out_fail;
8826         }
8827
8828         /* Update inode version and ctime/mtime. */
8829         inode_inc_iversion(old_dir);
8830         inode_inc_iversion(new_dir);
8831         inode_inc_iversion(old_inode);
8832         inode_inc_iversion(new_inode);
8833         old_dir->i_ctime = old_dir->i_mtime = ctime;
8834         new_dir->i_ctime = new_dir->i_mtime = ctime;
8835         old_inode->i_ctime = ctime;
8836         new_inode->i_ctime = ctime;
8837
8838         if (old_dentry->d_parent != new_dentry->d_parent) {
8839                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
8840                                 BTRFS_I(old_inode), 1);
8841                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
8842                                 BTRFS_I(new_inode), 1);
8843         }
8844
8845         /* src is a subvolume */
8846         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
8847                 ret = btrfs_unlink_subvol(trans, old_dir, old_dentry);
8848         } else { /* src is an inode */
8849                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
8850                                            BTRFS_I(old_dentry->d_inode),
8851                                            old_dentry->d_name.name,
8852                                            old_dentry->d_name.len);
8853                 if (!ret)
8854                         ret = btrfs_update_inode(trans, root, old_inode);
8855         }
8856         if (ret) {
8857                 btrfs_abort_transaction(trans, ret);
8858                 goto out_fail;
8859         }
8860
8861         /* dest is a subvolume */
8862         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
8863                 ret = btrfs_unlink_subvol(trans, new_dir, new_dentry);
8864         } else { /* dest is an inode */
8865                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
8866                                            BTRFS_I(new_dentry->d_inode),
8867                                            new_dentry->d_name.name,
8868                                            new_dentry->d_name.len);
8869                 if (!ret)
8870                         ret = btrfs_update_inode(trans, dest, new_inode);
8871         }
8872         if (ret) {
8873                 btrfs_abort_transaction(trans, ret);
8874                 goto out_fail;
8875         }
8876
8877         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
8878                              new_dentry->d_name.name,
8879                              new_dentry->d_name.len, 0, old_idx);
8880         if (ret) {
8881                 btrfs_abort_transaction(trans, ret);
8882                 goto out_fail;
8883         }
8884
8885         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
8886                              old_dentry->d_name.name,
8887                              old_dentry->d_name.len, 0, new_idx);
8888         if (ret) {
8889                 btrfs_abort_transaction(trans, ret);
8890                 goto out_fail;
8891         }
8892
8893         if (old_inode->i_nlink == 1)
8894                 BTRFS_I(old_inode)->dir_index = old_idx;
8895         if (new_inode->i_nlink == 1)
8896                 BTRFS_I(new_inode)->dir_index = new_idx;
8897
8898         if (root_log_pinned) {
8899                 parent = new_dentry->d_parent;
8900                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
8901                                          BTRFS_I(old_dir), parent,
8902                                          false, &ctx_root);
8903                 if (ret == BTRFS_NEED_LOG_SYNC)
8904                         sync_log_root = true;
8905                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
8906                         commit_transaction = true;
8907                 ret = 0;
8908                 btrfs_end_log_trans(root);
8909                 root_log_pinned = false;
8910         }
8911         if (dest_log_pinned) {
8912                 if (!commit_transaction) {
8913                         parent = old_dentry->d_parent;
8914                         ret = btrfs_log_new_name(trans, BTRFS_I(new_inode),
8915                                                  BTRFS_I(new_dir), parent,
8916                                                  false, &ctx_dest);
8917                         if (ret == BTRFS_NEED_LOG_SYNC)
8918                                 sync_log_dest = true;
8919                         else if (ret == BTRFS_NEED_TRANS_COMMIT)
8920                                 commit_transaction = true;
8921                         ret = 0;
8922                 }
8923                 btrfs_end_log_trans(dest);
8924                 dest_log_pinned = false;
8925         }
8926 out_fail:
8927         /*
8928          * If we have pinned a log and an error happened, we unpin tasks
8929          * trying to sync the log and force them to fallback to a transaction
8930          * commit if the log currently contains any of the inodes involved in
8931          * this rename operation (to ensure we do not persist a log with an
8932          * inconsistent state for any of these inodes or leading to any
8933          * inconsistencies when replayed). If the transaction was aborted, the
8934          * abortion reason is propagated to userspace when attempting to commit
8935          * the transaction. If the log does not contain any of these inodes, we
8936          * allow the tasks to sync it.
8937          */
8938         if (ret && (root_log_pinned || dest_log_pinned)) {
8939                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
8940                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
8941                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
8942                     (new_inode &&
8943                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
8944                         btrfs_set_log_full_commit(trans);
8945
8946                 if (root_log_pinned) {
8947                         btrfs_end_log_trans(root);
8948                         root_log_pinned = false;
8949                 }
8950                 if (dest_log_pinned) {
8951                         btrfs_end_log_trans(dest);
8952                         dest_log_pinned = false;
8953                 }
8954         }
8955         if (!ret && sync_log_root && !commit_transaction) {
8956                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root,
8957                                      &ctx_root);
8958                 if (ret)
8959                         commit_transaction = true;
8960         }
8961         if (!ret && sync_log_dest && !commit_transaction) {
8962                 ret = btrfs_sync_log(trans, BTRFS_I(new_inode)->root,
8963                                      &ctx_dest);
8964                 if (ret)
8965                         commit_transaction = true;
8966         }
8967         if (commit_transaction) {
8968                 /*
8969                  * We may have set commit_transaction when logging the new name
8970                  * in the destination root, in which case we left the source
8971                  * root context in the list of log contextes. So make sure we
8972                  * remove it to avoid invalid memory accesses, since the context
8973                  * was allocated in our stack frame.
8974                  */
8975                 if (sync_log_root) {
8976                         mutex_lock(&root->log_mutex);
8977                         list_del_init(&ctx_root.list);
8978                         mutex_unlock(&root->log_mutex);
8979                 }
8980                 ret = btrfs_commit_transaction(trans);
8981         } else {
8982                 int ret2;
8983
8984                 ret2 = btrfs_end_transaction(trans);
8985                 ret = ret ? ret : ret2;
8986         }
8987 out_notrans:
8988         if (new_ino == BTRFS_FIRST_FREE_OBJECTID ||
8989             old_ino == BTRFS_FIRST_FREE_OBJECTID)
8990                 up_read(&fs_info->subvol_sem);
8991
8992         ASSERT(list_empty(&ctx_root.list));
8993         ASSERT(list_empty(&ctx_dest.list));
8994
8995         return ret;
8996 }
8997
8998 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
8999                                      struct btrfs_root *root,
9000                                      struct inode *dir,
9001                                      struct dentry *dentry)
9002 {
9003         int ret;
9004         struct inode *inode;
9005         u64 objectid;
9006         u64 index;
9007
9008         ret = btrfs_find_free_ino(root, &objectid);
9009         if (ret)
9010                 return ret;
9011
9012         inode = btrfs_new_inode(trans, root, dir,
9013                                 dentry->d_name.name,
9014                                 dentry->d_name.len,
9015                                 btrfs_ino(BTRFS_I(dir)),
9016                                 objectid,
9017                                 S_IFCHR | WHITEOUT_MODE,
9018                                 &index);
9019
9020         if (IS_ERR(inode)) {
9021                 ret = PTR_ERR(inode);
9022                 return ret;
9023         }
9024
9025         inode->i_op = &btrfs_special_inode_operations;
9026         init_special_inode(inode, inode->i_mode,
9027                 WHITEOUT_DEV);
9028
9029         ret = btrfs_init_inode_security(trans, inode, dir,
9030                                 &dentry->d_name);
9031         if (ret)
9032                 goto out;
9033
9034         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9035                                 BTRFS_I(inode), 0, index);
9036         if (ret)
9037                 goto out;
9038
9039         ret = btrfs_update_inode(trans, root, inode);
9040 out:
9041         unlock_new_inode(inode);
9042         if (ret)
9043                 inode_dec_link_count(inode);
9044         iput(inode);
9045
9046         return ret;
9047 }
9048
9049 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9050                            struct inode *new_dir, struct dentry *new_dentry,
9051                            unsigned int flags)
9052 {
9053         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9054         struct btrfs_trans_handle *trans;
9055         unsigned int trans_num_items;
9056         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9057         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9058         struct inode *new_inode = d_inode(new_dentry);
9059         struct inode *old_inode = d_inode(old_dentry);
9060         u64 index = 0;
9061         int ret;
9062         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9063         bool log_pinned = false;
9064         struct btrfs_log_ctx ctx;
9065         bool sync_log = false;
9066         bool commit_transaction = false;
9067
9068         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9069                 return -EPERM;
9070
9071         /* we only allow rename subvolume link between subvolumes */
9072         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9073                 return -EXDEV;
9074
9075         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9076             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9077                 return -ENOTEMPTY;
9078
9079         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9080             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9081                 return -ENOTEMPTY;
9082
9083
9084         /* check for collisions, even if the  name isn't there */
9085         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9086                              new_dentry->d_name.name,
9087                              new_dentry->d_name.len);
9088
9089         if (ret) {
9090                 if (ret == -EEXIST) {
9091                         /* we shouldn't get
9092                          * eexist without a new_inode */
9093                         if (WARN_ON(!new_inode)) {
9094                                 return ret;
9095                         }
9096                 } else {
9097                         /* maybe -EOVERFLOW */
9098                         return ret;
9099                 }
9100         }
9101         ret = 0;
9102
9103         /*
9104          * we're using rename to replace one file with another.  Start IO on it
9105          * now so  we don't add too much work to the end of the transaction
9106          */
9107         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9108                 filemap_flush(old_inode->i_mapping);
9109
9110         /* close the racy window with snapshot create/destroy ioctl */
9111         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9112                 down_read(&fs_info->subvol_sem);
9113         /*
9114          * We want to reserve the absolute worst case amount of items.  So if
9115          * both inodes are subvols and we need to unlink them then that would
9116          * require 4 item modifications, but if they are both normal inodes it
9117          * would require 5 item modifications, so we'll assume they are normal
9118          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9119          * should cover the worst case number of items we'll modify.
9120          * If our rename has the whiteout flag, we need more 5 units for the
9121          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9122          * when selinux is enabled).
9123          */
9124         trans_num_items = 11;
9125         if (flags & RENAME_WHITEOUT)
9126                 trans_num_items += 5;
9127         trans = btrfs_start_transaction(root, trans_num_items);
9128         if (IS_ERR(trans)) {
9129                 ret = PTR_ERR(trans);
9130                 goto out_notrans;
9131         }
9132
9133         if (dest != root)
9134                 btrfs_record_root_in_trans(trans, dest);
9135
9136         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
9137         if (ret)
9138                 goto out_fail;
9139
9140         BTRFS_I(old_inode)->dir_index = 0ULL;
9141         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9142                 /* force full log commit if subvolume involved. */
9143                 btrfs_set_log_full_commit(trans);
9144         } else {
9145                 btrfs_pin_log_trans(root);
9146                 log_pinned = true;
9147                 ret = btrfs_insert_inode_ref(trans, dest,
9148                                              new_dentry->d_name.name,
9149                                              new_dentry->d_name.len,
9150                                              old_ino,
9151                                              btrfs_ino(BTRFS_I(new_dir)), index);
9152                 if (ret)
9153                         goto out_fail;
9154         }
9155
9156         inode_inc_iversion(old_dir);
9157         inode_inc_iversion(new_dir);
9158         inode_inc_iversion(old_inode);
9159         old_dir->i_ctime = old_dir->i_mtime =
9160         new_dir->i_ctime = new_dir->i_mtime =
9161         old_inode->i_ctime = current_time(old_dir);
9162
9163         if (old_dentry->d_parent != new_dentry->d_parent)
9164                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9165                                 BTRFS_I(old_inode), 1);
9166
9167         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9168                 ret = btrfs_unlink_subvol(trans, old_dir, old_dentry);
9169         } else {
9170                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9171                                         BTRFS_I(d_inode(old_dentry)),
9172                                         old_dentry->d_name.name,
9173                                         old_dentry->d_name.len);
9174                 if (!ret)
9175                         ret = btrfs_update_inode(trans, root, old_inode);
9176         }
9177         if (ret) {
9178                 btrfs_abort_transaction(trans, ret);
9179                 goto out_fail;
9180         }
9181
9182         if (new_inode) {
9183                 inode_inc_iversion(new_inode);
9184                 new_inode->i_ctime = current_time(new_inode);
9185                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
9186                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9187                         ret = btrfs_unlink_subvol(trans, new_dir, new_dentry);
9188                         BUG_ON(new_inode->i_nlink == 0);
9189                 } else {
9190                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9191                                                  BTRFS_I(d_inode(new_dentry)),
9192                                                  new_dentry->d_name.name,
9193                                                  new_dentry->d_name.len);
9194                 }
9195                 if (!ret && new_inode->i_nlink == 0)
9196                         ret = btrfs_orphan_add(trans,
9197                                         BTRFS_I(d_inode(new_dentry)));
9198                 if (ret) {
9199                         btrfs_abort_transaction(trans, ret);
9200                         goto out_fail;
9201                 }
9202         }
9203
9204         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9205                              new_dentry->d_name.name,
9206                              new_dentry->d_name.len, 0, index);
9207         if (ret) {
9208                 btrfs_abort_transaction(trans, ret);
9209                 goto out_fail;
9210         }
9211
9212         if (old_inode->i_nlink == 1)
9213                 BTRFS_I(old_inode)->dir_index = index;
9214
9215         if (log_pinned) {
9216                 struct dentry *parent = new_dentry->d_parent;
9217
9218                 btrfs_init_log_ctx(&ctx, old_inode);
9219                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
9220                                          BTRFS_I(old_dir), parent,
9221                                          false, &ctx);
9222                 if (ret == BTRFS_NEED_LOG_SYNC)
9223                         sync_log = true;
9224                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
9225                         commit_transaction = true;
9226                 ret = 0;
9227                 btrfs_end_log_trans(root);
9228                 log_pinned = false;
9229         }
9230
9231         if (flags & RENAME_WHITEOUT) {
9232                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9233                                                 old_dentry);
9234
9235                 if (ret) {
9236                         btrfs_abort_transaction(trans, ret);
9237                         goto out_fail;
9238                 }
9239         }
9240 out_fail:
9241         /*
9242          * If we have pinned the log and an error happened, we unpin tasks
9243          * trying to sync the log and force them to fallback to a transaction
9244          * commit if the log currently contains any of the inodes involved in
9245          * this rename operation (to ensure we do not persist a log with an
9246          * inconsistent state for any of these inodes or leading to any
9247          * inconsistencies when replayed). If the transaction was aborted, the
9248          * abortion reason is propagated to userspace when attempting to commit
9249          * the transaction. If the log does not contain any of these inodes, we
9250          * allow the tasks to sync it.
9251          */
9252         if (ret && log_pinned) {
9253                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9254                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9255                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9256                     (new_inode &&
9257                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9258                         btrfs_set_log_full_commit(trans);
9259
9260                 btrfs_end_log_trans(root);
9261                 log_pinned = false;
9262         }
9263         if (!ret && sync_log) {
9264                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root, &ctx);
9265                 if (ret)
9266                         commit_transaction = true;
9267         } else if (sync_log) {
9268                 mutex_lock(&root->log_mutex);
9269                 list_del(&ctx.list);
9270                 mutex_unlock(&root->log_mutex);
9271         }
9272         if (commit_transaction) {
9273                 ret = btrfs_commit_transaction(trans);
9274         } else {
9275                 int ret2;
9276
9277                 ret2 = btrfs_end_transaction(trans);
9278                 ret = ret ? ret : ret2;
9279         }
9280 out_notrans:
9281         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9282                 up_read(&fs_info->subvol_sem);
9283
9284         return ret;
9285 }
9286
9287 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9288                          struct inode *new_dir, struct dentry *new_dentry,
9289                          unsigned int flags)
9290 {
9291         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
9292                 return -EINVAL;
9293
9294         if (flags & RENAME_EXCHANGE)
9295                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
9296                                           new_dentry);
9297
9298         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
9299 }
9300
9301 struct btrfs_delalloc_work {
9302         struct inode *inode;
9303         struct completion completion;
9304         struct list_head list;
9305         struct btrfs_work work;
9306 };
9307
9308 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9309 {
9310         struct btrfs_delalloc_work *delalloc_work;
9311         struct inode *inode;
9312
9313         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9314                                      work);
9315         inode = delalloc_work->inode;
9316         filemap_flush(inode->i_mapping);
9317         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9318                                 &BTRFS_I(inode)->runtime_flags))
9319                 filemap_flush(inode->i_mapping);
9320
9321         iput(inode);
9322         complete(&delalloc_work->completion);
9323 }
9324
9325 static struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode)
9326 {
9327         struct btrfs_delalloc_work *work;
9328
9329         work = kmalloc(sizeof(*work), GFP_NOFS);
9330         if (!work)
9331                 return NULL;
9332
9333         init_completion(&work->completion);
9334         INIT_LIST_HEAD(&work->list);
9335         work->inode = inode;
9336         btrfs_init_work(&work->work, btrfs_run_delalloc_work, NULL, NULL);
9337
9338         return work;
9339 }
9340
9341 /*
9342  * some fairly slow code that needs optimization. This walks the list
9343  * of all the inodes with pending delalloc and forces them to disk.
9344  */
9345 static int start_delalloc_inodes(struct btrfs_root *root, int nr, bool snapshot)
9346 {
9347         struct btrfs_inode *binode;
9348         struct inode *inode;
9349         struct btrfs_delalloc_work *work, *next;
9350         struct list_head works;
9351         struct list_head splice;
9352         int ret = 0;
9353
9354         INIT_LIST_HEAD(&works);
9355         INIT_LIST_HEAD(&splice);
9356
9357         mutex_lock(&root->delalloc_mutex);
9358         spin_lock(&root->delalloc_lock);
9359         list_splice_init(&root->delalloc_inodes, &splice);
9360         while (!list_empty(&splice)) {
9361                 binode = list_entry(splice.next, struct btrfs_inode,
9362                                     delalloc_inodes);
9363
9364                 list_move_tail(&binode->delalloc_inodes,
9365                                &root->delalloc_inodes);
9366                 inode = igrab(&binode->vfs_inode);
9367                 if (!inode) {
9368                         cond_resched_lock(&root->delalloc_lock);
9369                         continue;
9370                 }
9371                 spin_unlock(&root->delalloc_lock);
9372
9373                 if (snapshot)
9374                         set_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
9375                                 &binode->runtime_flags);
9376                 work = btrfs_alloc_delalloc_work(inode);
9377                 if (!work) {
9378                         iput(inode);
9379                         ret = -ENOMEM;
9380                         goto out;
9381                 }
9382                 list_add_tail(&work->list, &works);
9383                 btrfs_queue_work(root->fs_info->flush_workers,
9384                                  &work->work);
9385                 ret++;
9386                 if (nr != -1 && ret >= nr)
9387                         goto out;
9388                 cond_resched();
9389                 spin_lock(&root->delalloc_lock);
9390         }
9391         spin_unlock(&root->delalloc_lock);
9392
9393 out:
9394         list_for_each_entry_safe(work, next, &works, list) {
9395                 list_del_init(&work->list);
9396                 wait_for_completion(&work->completion);
9397                 kfree(work);
9398         }
9399
9400         if (!list_empty(&splice)) {
9401                 spin_lock(&root->delalloc_lock);
9402                 list_splice_tail(&splice, &root->delalloc_inodes);
9403                 spin_unlock(&root->delalloc_lock);
9404         }
9405         mutex_unlock(&root->delalloc_mutex);
9406         return ret;
9407 }
9408
9409 int btrfs_start_delalloc_snapshot(struct btrfs_root *root)
9410 {
9411         struct btrfs_fs_info *fs_info = root->fs_info;
9412         int ret;
9413
9414         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
9415                 return -EROFS;
9416
9417         ret = start_delalloc_inodes(root, -1, true);
9418         if (ret > 0)
9419                 ret = 0;
9420         return ret;
9421 }
9422
9423 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int nr)
9424 {
9425         struct btrfs_root *root;
9426         struct list_head splice;
9427         int ret;
9428
9429         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
9430                 return -EROFS;
9431
9432         INIT_LIST_HEAD(&splice);
9433
9434         mutex_lock(&fs_info->delalloc_root_mutex);
9435         spin_lock(&fs_info->delalloc_root_lock);
9436         list_splice_init(&fs_info->delalloc_roots, &splice);
9437         while (!list_empty(&splice) && nr) {
9438                 root = list_first_entry(&splice, struct btrfs_root,
9439                                         delalloc_root);
9440                 root = btrfs_grab_root(root);
9441                 BUG_ON(!root);
9442                 list_move_tail(&root->delalloc_root,
9443                                &fs_info->delalloc_roots);
9444                 spin_unlock(&fs_info->delalloc_root_lock);
9445
9446                 ret = start_delalloc_inodes(root, nr, false);
9447                 btrfs_put_root(root);
9448                 if (ret < 0)
9449                         goto out;
9450
9451                 if (nr != -1) {
9452                         nr -= ret;
9453                         WARN_ON(nr < 0);
9454                 }
9455                 spin_lock(&fs_info->delalloc_root_lock);
9456         }
9457         spin_unlock(&fs_info->delalloc_root_lock);
9458
9459         ret = 0;
9460 out:
9461         if (!list_empty(&splice)) {
9462                 spin_lock(&fs_info->delalloc_root_lock);
9463                 list_splice_tail(&splice, &fs_info->delalloc_roots);
9464                 spin_unlock(&fs_info->delalloc_root_lock);
9465         }
9466         mutex_unlock(&fs_info->delalloc_root_mutex);
9467         return ret;
9468 }
9469
9470 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
9471                          const char *symname)
9472 {
9473         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
9474         struct btrfs_trans_handle *trans;
9475         struct btrfs_root *root = BTRFS_I(dir)->root;
9476         struct btrfs_path *path;
9477         struct btrfs_key key;
9478         struct inode *inode = NULL;
9479         int err;
9480         u64 objectid;
9481         u64 index = 0;
9482         int name_len;
9483         int datasize;
9484         unsigned long ptr;
9485         struct btrfs_file_extent_item *ei;
9486         struct extent_buffer *leaf;
9487
9488         name_len = strlen(symname);
9489         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
9490                 return -ENAMETOOLONG;
9491
9492         /*
9493          * 2 items for inode item and ref
9494          * 2 items for dir items
9495          * 1 item for updating parent inode item
9496          * 1 item for the inline extent item
9497          * 1 item for xattr if selinux is on
9498          */
9499         trans = btrfs_start_transaction(root, 7);
9500         if (IS_ERR(trans))
9501                 return PTR_ERR(trans);
9502
9503         err = btrfs_find_free_ino(root, &objectid);
9504         if (err)
9505                 goto out_unlock;
9506
9507         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
9508                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
9509                                 objectid, S_IFLNK|S_IRWXUGO, &index);
9510         if (IS_ERR(inode)) {
9511                 err = PTR_ERR(inode);
9512                 inode = NULL;
9513                 goto out_unlock;
9514         }
9515
9516         /*
9517         * If the active LSM wants to access the inode during
9518         * d_instantiate it needs these. Smack checks to see
9519         * if the filesystem supports xattrs by looking at the
9520         * ops vector.
9521         */
9522         inode->i_fop = &btrfs_file_operations;
9523         inode->i_op = &btrfs_file_inode_operations;
9524         inode->i_mapping->a_ops = &btrfs_aops;
9525         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
9526
9527         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
9528         if (err)
9529                 goto out_unlock;
9530
9531         path = btrfs_alloc_path();
9532         if (!path) {
9533                 err = -ENOMEM;
9534                 goto out_unlock;
9535         }
9536         key.objectid = btrfs_ino(BTRFS_I(inode));
9537         key.offset = 0;
9538         key.type = BTRFS_EXTENT_DATA_KEY;
9539         datasize = btrfs_file_extent_calc_inline_size(name_len);
9540         err = btrfs_insert_empty_item(trans, root, path, &key,
9541                                       datasize);
9542         if (err) {
9543                 btrfs_free_path(path);
9544                 goto out_unlock;
9545         }
9546         leaf = path->nodes[0];
9547         ei = btrfs_item_ptr(leaf, path->slots[0],
9548                             struct btrfs_file_extent_item);
9549         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
9550         btrfs_set_file_extent_type(leaf, ei,
9551                                    BTRFS_FILE_EXTENT_INLINE);
9552         btrfs_set_file_extent_encryption(leaf, ei, 0);
9553         btrfs_set_file_extent_compression(leaf, ei, 0);
9554         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
9555         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
9556
9557         ptr = btrfs_file_extent_inline_start(ei);
9558         write_extent_buffer(leaf, symname, ptr, name_len);
9559         btrfs_mark_buffer_dirty(leaf);
9560         btrfs_free_path(path);
9561
9562         inode->i_op = &btrfs_symlink_inode_operations;
9563         inode_nohighmem(inode);
9564         inode_set_bytes(inode, name_len);
9565         btrfs_i_size_write(BTRFS_I(inode), name_len);
9566         err = btrfs_update_inode(trans, root, inode);
9567         /*
9568          * Last step, add directory indexes for our symlink inode. This is the
9569          * last step to avoid extra cleanup of these indexes if an error happens
9570          * elsewhere above.
9571          */
9572         if (!err)
9573                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9574                                 BTRFS_I(inode), 0, index);
9575         if (err)
9576                 goto out_unlock;
9577
9578         d_instantiate_new(dentry, inode);
9579
9580 out_unlock:
9581         btrfs_end_transaction(trans);
9582         if (err && inode) {
9583                 inode_dec_link_count(inode);
9584                 discard_new_inode(inode);
9585         }
9586         btrfs_btree_balance_dirty(fs_info);
9587         return err;
9588 }
9589
9590 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
9591                                        u64 start, u64 num_bytes, u64 min_size,
9592                                        loff_t actual_len, u64 *alloc_hint,
9593                                        struct btrfs_trans_handle *trans)
9594 {
9595         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9596         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
9597         struct extent_map *em;
9598         struct btrfs_root *root = BTRFS_I(inode)->root;
9599         struct btrfs_key ins;
9600         u64 cur_offset = start;
9601         u64 clear_offset = start;
9602         u64 i_size;
9603         u64 cur_bytes;
9604         u64 last_alloc = (u64)-1;
9605         int ret = 0;
9606         bool own_trans = true;
9607         u64 end = start + num_bytes - 1;
9608
9609         if (trans)
9610                 own_trans = false;
9611         while (num_bytes > 0) {
9612                 if (own_trans) {
9613                         trans = btrfs_start_transaction(root, 3);
9614                         if (IS_ERR(trans)) {
9615                                 ret = PTR_ERR(trans);
9616                                 break;
9617                         }
9618                 }
9619
9620                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
9621                 cur_bytes = max(cur_bytes, min_size);
9622                 /*
9623                  * If we are severely fragmented we could end up with really
9624                  * small allocations, so if the allocator is returning small
9625                  * chunks lets make its job easier by only searching for those
9626                  * sized chunks.
9627                  */
9628                 cur_bytes = min(cur_bytes, last_alloc);
9629                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
9630                                 min_size, 0, *alloc_hint, &ins, 1, 0);
9631                 if (ret) {
9632                         if (own_trans)
9633                                 btrfs_end_transaction(trans);
9634                         break;
9635                 }
9636
9637                 /*
9638                  * We've reserved this space, and thus converted it from
9639                  * ->bytes_may_use to ->bytes_reserved.  Any error that happens
9640                  * from here on out we will only need to clear our reservation
9641                  * for the remaining unreserved area, so advance our
9642                  * clear_offset by our extent size.
9643                  */
9644                 clear_offset += ins.offset;
9645                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
9646
9647                 last_alloc = ins.offset;
9648                 ret = insert_reserved_file_extent(trans, inode,
9649                                                   cur_offset, ins.objectid,
9650                                                   ins.offset, ins.offset,
9651                                                   ins.offset, 0, 0, 0,
9652                                                   BTRFS_FILE_EXTENT_PREALLOC);
9653                 if (ret) {
9654                         btrfs_free_reserved_extent(fs_info, ins.objectid,
9655                                                    ins.offset, 0);
9656                         btrfs_abort_transaction(trans, ret);
9657                         if (own_trans)
9658                                 btrfs_end_transaction(trans);
9659                         break;
9660                 }
9661
9662                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
9663                                         cur_offset + ins.offset -1, 0);
9664
9665                 em = alloc_extent_map();
9666                 if (!em) {
9667                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
9668                                 &BTRFS_I(inode)->runtime_flags);
9669                         goto next;
9670                 }
9671
9672                 em->start = cur_offset;
9673                 em->orig_start = cur_offset;
9674                 em->len = ins.offset;
9675                 em->block_start = ins.objectid;
9676                 em->block_len = ins.offset;
9677                 em->orig_block_len = ins.offset;
9678                 em->ram_bytes = ins.offset;
9679                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
9680                 em->generation = trans->transid;
9681
9682                 while (1) {
9683                         write_lock(&em_tree->lock);
9684                         ret = add_extent_mapping(em_tree, em, 1);
9685                         write_unlock(&em_tree->lock);
9686                         if (ret != -EEXIST)
9687                                 break;
9688                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
9689                                                 cur_offset + ins.offset - 1,
9690                                                 0);
9691                 }
9692                 free_extent_map(em);
9693 next:
9694                 num_bytes -= ins.offset;
9695                 cur_offset += ins.offset;
9696                 *alloc_hint = ins.objectid + ins.offset;
9697
9698                 inode_inc_iversion(inode);
9699                 inode->i_ctime = current_time(inode);
9700                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
9701                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
9702                     (actual_len > inode->i_size) &&
9703                     (cur_offset > inode->i_size)) {
9704                         if (cur_offset > actual_len)
9705                                 i_size = actual_len;
9706                         else
9707                                 i_size = cur_offset;
9708                         i_size_write(inode, i_size);
9709                         btrfs_inode_safe_disk_i_size_write(inode, 0);
9710                 }
9711
9712                 ret = btrfs_update_inode(trans, root, inode);
9713
9714                 if (ret) {
9715                         btrfs_abort_transaction(trans, ret);
9716                         if (own_trans)
9717                                 btrfs_end_transaction(trans);
9718                         break;
9719                 }
9720
9721                 if (own_trans)
9722                         btrfs_end_transaction(trans);
9723         }
9724         if (clear_offset < end)
9725                 btrfs_free_reserved_data_space(inode, NULL, clear_offset,
9726                         end - clear_offset + 1);
9727         return ret;
9728 }
9729
9730 int btrfs_prealloc_file_range(struct inode *inode, int mode,
9731                               u64 start, u64 num_bytes, u64 min_size,
9732                               loff_t actual_len, u64 *alloc_hint)
9733 {
9734         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
9735                                            min_size, actual_len, alloc_hint,
9736                                            NULL);
9737 }
9738
9739 int btrfs_prealloc_file_range_trans(struct inode *inode,
9740                                     struct btrfs_trans_handle *trans, int mode,
9741                                     u64 start, u64 num_bytes, u64 min_size,
9742                                     loff_t actual_len, u64 *alloc_hint)
9743 {
9744         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
9745                                            min_size, actual_len, alloc_hint, trans);
9746 }
9747
9748 static int btrfs_set_page_dirty(struct page *page)
9749 {
9750         return __set_page_dirty_nobuffers(page);
9751 }
9752
9753 static int btrfs_permission(struct inode *inode, int mask)
9754 {
9755         struct btrfs_root *root = BTRFS_I(inode)->root;
9756         umode_t mode = inode->i_mode;
9757
9758         if (mask & MAY_WRITE &&
9759             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
9760                 if (btrfs_root_readonly(root))
9761                         return -EROFS;
9762                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
9763                         return -EACCES;
9764         }
9765         return generic_permission(inode, mask);
9766 }
9767
9768 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
9769 {
9770         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
9771         struct btrfs_trans_handle *trans;
9772         struct btrfs_root *root = BTRFS_I(dir)->root;
9773         struct inode *inode = NULL;
9774         u64 objectid;
9775         u64 index;
9776         int ret = 0;
9777
9778         /*
9779          * 5 units required for adding orphan entry
9780          */
9781         trans = btrfs_start_transaction(root, 5);
9782         if (IS_ERR(trans))
9783                 return PTR_ERR(trans);
9784
9785         ret = btrfs_find_free_ino(root, &objectid);
9786         if (ret)
9787                 goto out;
9788
9789         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
9790                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
9791         if (IS_ERR(inode)) {
9792                 ret = PTR_ERR(inode);
9793                 inode = NULL;
9794                 goto out;
9795         }
9796
9797         inode->i_fop = &btrfs_file_operations;
9798         inode->i_op = &btrfs_file_inode_operations;
9799
9800         inode->i_mapping->a_ops = &btrfs_aops;
9801         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
9802
9803         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
9804         if (ret)
9805                 goto out;
9806
9807         ret = btrfs_update_inode(trans, root, inode);
9808         if (ret)
9809                 goto out;
9810         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
9811         if (ret)
9812                 goto out;
9813
9814         /*
9815          * We set number of links to 0 in btrfs_new_inode(), and here we set
9816          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
9817          * through:
9818          *
9819          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
9820          */
9821         set_nlink(inode, 1);
9822         d_tmpfile(dentry, inode);
9823         unlock_new_inode(inode);
9824         mark_inode_dirty(inode);
9825 out:
9826         btrfs_end_transaction(trans);
9827         if (ret && inode)
9828                 discard_new_inode(inode);
9829         btrfs_btree_balance_dirty(fs_info);
9830         return ret;
9831 }
9832
9833 void btrfs_set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
9834 {
9835         struct inode *inode = tree->private_data;
9836         unsigned long index = start >> PAGE_SHIFT;
9837         unsigned long end_index = end >> PAGE_SHIFT;
9838         struct page *page;
9839
9840         while (index <= end_index) {
9841                 page = find_get_page(inode->i_mapping, index);
9842                 ASSERT(page); /* Pages should be in the extent_io_tree */
9843                 set_page_writeback(page);
9844                 put_page(page);
9845                 index++;
9846         }
9847 }
9848
9849 #ifdef CONFIG_SWAP
9850 /*
9851  * Add an entry indicating a block group or device which is pinned by a
9852  * swapfile. Returns 0 on success, 1 if there is already an entry for it, or a
9853  * negative errno on failure.
9854  */
9855 static int btrfs_add_swapfile_pin(struct inode *inode, void *ptr,
9856                                   bool is_block_group)
9857 {
9858         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
9859         struct btrfs_swapfile_pin *sp, *entry;
9860         struct rb_node **p;
9861         struct rb_node *parent = NULL;
9862
9863         sp = kmalloc(sizeof(*sp), GFP_NOFS);
9864         if (!sp)
9865                 return -ENOMEM;
9866         sp->ptr = ptr;
9867         sp->inode = inode;
9868         sp->is_block_group = is_block_group;
9869
9870         spin_lock(&fs_info->swapfile_pins_lock);
9871         p = &fs_info->swapfile_pins.rb_node;
9872         while (*p) {
9873                 parent = *p;
9874                 entry = rb_entry(parent, struct btrfs_swapfile_pin, node);
9875                 if (sp->ptr < entry->ptr ||
9876                     (sp->ptr == entry->ptr && sp->inode < entry->inode)) {
9877                         p = &(*p)->rb_left;
9878                 } else if (sp->ptr > entry->ptr ||
9879                            (sp->ptr == entry->ptr && sp->inode > entry->inode)) {
9880                         p = &(*p)->rb_right;
9881                 } else {
9882                         spin_unlock(&fs_info->swapfile_pins_lock);
9883                         kfree(sp);
9884                         return 1;
9885                 }
9886         }
9887         rb_link_node(&sp->node, parent, p);
9888         rb_insert_color(&sp->node, &fs_info->swapfile_pins);
9889         spin_unlock(&fs_info->swapfile_pins_lock);
9890         return 0;
9891 }
9892
9893 /* Free all of the entries pinned by this swapfile. */
9894 static void btrfs_free_swapfile_pins(struct inode *inode)
9895 {
9896         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
9897         struct btrfs_swapfile_pin *sp;
9898         struct rb_node *node, *next;
9899
9900         spin_lock(&fs_info->swapfile_pins_lock);
9901         node = rb_first(&fs_info->swapfile_pins);
9902         while (node) {
9903                 next = rb_next(node);
9904                 sp = rb_entry(node, struct btrfs_swapfile_pin, node);
9905                 if (sp->inode == inode) {
9906                         rb_erase(&sp->node, &fs_info->swapfile_pins);
9907                         if (sp->is_block_group)
9908                                 btrfs_put_block_group(sp->ptr);
9909                         kfree(sp);
9910                 }
9911                 node = next;
9912         }
9913         spin_unlock(&fs_info->swapfile_pins_lock);
9914 }
9915
9916 struct btrfs_swap_info {
9917         u64 start;
9918         u64 block_start;
9919         u64 block_len;
9920         u64 lowest_ppage;
9921         u64 highest_ppage;
9922         unsigned long nr_pages;
9923         int nr_extents;
9924 };
9925
9926 static int btrfs_add_swap_extent(struct swap_info_struct *sis,
9927                                  struct btrfs_swap_info *bsi)
9928 {
9929         unsigned long nr_pages;
9930         u64 first_ppage, first_ppage_reported, next_ppage;
9931         int ret;
9932
9933         first_ppage = ALIGN(bsi->block_start, PAGE_SIZE) >> PAGE_SHIFT;
9934         next_ppage = ALIGN_DOWN(bsi->block_start + bsi->block_len,
9935                                 PAGE_SIZE) >> PAGE_SHIFT;
9936
9937         if (first_ppage >= next_ppage)
9938                 return 0;
9939         nr_pages = next_ppage - first_ppage;
9940
9941         first_ppage_reported = first_ppage;
9942         if (bsi->start == 0)
9943                 first_ppage_reported++;
9944         if (bsi->lowest_ppage > first_ppage_reported)
9945                 bsi->lowest_ppage = first_ppage_reported;
9946         if (bsi->highest_ppage < (next_ppage - 1))
9947                 bsi->highest_ppage = next_ppage - 1;
9948
9949         ret = add_swap_extent(sis, bsi->nr_pages, nr_pages, first_ppage);
9950         if (ret < 0)
9951                 return ret;
9952         bsi->nr_extents += ret;
9953         bsi->nr_pages += nr_pages;
9954         return 0;
9955 }
9956
9957 static void btrfs_swap_deactivate(struct file *file)
9958 {
9959         struct inode *inode = file_inode(file);
9960
9961         btrfs_free_swapfile_pins(inode);
9962         atomic_dec(&BTRFS_I(inode)->root->nr_swapfiles);
9963 }
9964
9965 static int btrfs_swap_activate(struct swap_info_struct *sis, struct file *file,
9966                                sector_t *span)
9967 {
9968         struct inode *inode = file_inode(file);
9969         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
9970         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
9971         struct extent_state *cached_state = NULL;
9972         struct extent_map *em = NULL;
9973         struct btrfs_device *device = NULL;
9974         struct btrfs_swap_info bsi = {
9975                 .lowest_ppage = (sector_t)-1ULL,
9976         };
9977         int ret = 0;
9978         u64 isize;
9979         u64 start;
9980
9981         /*
9982          * If the swap file was just created, make sure delalloc is done. If the
9983          * file changes again after this, the user is doing something stupid and
9984          * we don't really care.
9985          */
9986         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
9987         if (ret)
9988                 return ret;
9989
9990         /*
9991          * The inode is locked, so these flags won't change after we check them.
9992          */
9993         if (BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS) {
9994                 btrfs_warn(fs_info, "swapfile must not be compressed");
9995                 return -EINVAL;
9996         }
9997         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW)) {
9998                 btrfs_warn(fs_info, "swapfile must not be copy-on-write");
9999                 return -EINVAL;
10000         }
10001         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
10002                 btrfs_warn(fs_info, "swapfile must not be checksummed");
10003                 return -EINVAL;
10004         }
10005
10006         /*
10007          * Balance or device remove/replace/resize can move stuff around from
10008          * under us. The EXCL_OP flag makes sure they aren't running/won't run
10009          * concurrently while we are mapping the swap extents, and
10010          * fs_info->swapfile_pins prevents them from running while the swap file
10011          * is active and moving the extents. Note that this also prevents a
10012          * concurrent device add which isn't actually necessary, but it's not
10013          * really worth the trouble to allow it.
10014          */
10015         if (test_and_set_bit(BTRFS_FS_EXCL_OP, &fs_info->flags)) {
10016                 btrfs_warn(fs_info,
10017            "cannot activate swapfile while exclusive operation is running");
10018                 return -EBUSY;
10019         }
10020         /*
10021          * Snapshots can create extents which require COW even if NODATACOW is
10022          * set. We use this counter to prevent snapshots. We must increment it
10023          * before walking the extents because we don't want a concurrent
10024          * snapshot to run after we've already checked the extents.
10025          */
10026         atomic_inc(&BTRFS_I(inode)->root->nr_swapfiles);
10027
10028         isize = ALIGN_DOWN(inode->i_size, fs_info->sectorsize);
10029
10030         lock_extent_bits(io_tree, 0, isize - 1, &cached_state);
10031         start = 0;
10032         while (start < isize) {
10033                 u64 logical_block_start, physical_block_start;
10034                 struct btrfs_block_group *bg;
10035                 u64 len = isize - start;
10036
10037                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len);
10038                 if (IS_ERR(em)) {
10039                         ret = PTR_ERR(em);
10040                         goto out;
10041                 }
10042
10043                 if (em->block_start == EXTENT_MAP_HOLE) {
10044                         btrfs_warn(fs_info, "swapfile must not have holes");
10045                         ret = -EINVAL;
10046                         goto out;
10047                 }
10048                 if (em->block_start == EXTENT_MAP_INLINE) {
10049                         /*
10050                          * It's unlikely we'll ever actually find ourselves
10051                          * here, as a file small enough to fit inline won't be
10052                          * big enough to store more than the swap header, but in
10053                          * case something changes in the future, let's catch it
10054                          * here rather than later.
10055                          */
10056                         btrfs_warn(fs_info, "swapfile must not be inline");
10057                         ret = -EINVAL;
10058                         goto out;
10059                 }
10060                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
10061                         btrfs_warn(fs_info, "swapfile must not be compressed");
10062                         ret = -EINVAL;
10063                         goto out;
10064                 }
10065
10066                 logical_block_start = em->block_start + (start - em->start);
10067                 len = min(len, em->len - (start - em->start));
10068                 free_extent_map(em);
10069                 em = NULL;
10070
10071                 ret = can_nocow_extent(inode, start, &len, NULL, NULL, NULL);
10072                 if (ret < 0) {
10073                         goto out;
10074                 } else if (ret) {
10075                         ret = 0;
10076                 } else {
10077                         btrfs_warn(fs_info,
10078                                    "swapfile must not be copy-on-write");
10079                         ret = -EINVAL;
10080                         goto out;
10081                 }
10082
10083                 em = btrfs_get_chunk_map(fs_info, logical_block_start, len);
10084                 if (IS_ERR(em)) {
10085                         ret = PTR_ERR(em);
10086                         goto out;
10087                 }
10088
10089                 if (em->map_lookup->type & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
10090                         btrfs_warn(fs_info,
10091                                    "swapfile must have single data profile");
10092                         ret = -EINVAL;
10093                         goto out;
10094                 }
10095
10096                 if (device == NULL) {
10097                         device = em->map_lookup->stripes[0].dev;
10098                         ret = btrfs_add_swapfile_pin(inode, device, false);
10099                         if (ret == 1)
10100                                 ret = 0;
10101                         else if (ret)
10102                                 goto out;
10103                 } else if (device != em->map_lookup->stripes[0].dev) {
10104                         btrfs_warn(fs_info, "swapfile must be on one device");
10105                         ret = -EINVAL;
10106                         goto out;
10107                 }
10108
10109                 physical_block_start = (em->map_lookup->stripes[0].physical +
10110                                         (logical_block_start - em->start));
10111                 len = min(len, em->len - (logical_block_start - em->start));
10112                 free_extent_map(em);
10113                 em = NULL;
10114
10115                 bg = btrfs_lookup_block_group(fs_info, logical_block_start);
10116                 if (!bg) {
10117                         btrfs_warn(fs_info,
10118                            "could not find block group containing swapfile");
10119                         ret = -EINVAL;
10120                         goto out;
10121                 }
10122
10123                 ret = btrfs_add_swapfile_pin(inode, bg, true);
10124                 if (ret) {
10125                         btrfs_put_block_group(bg);
10126                         if (ret == 1)
10127                                 ret = 0;
10128                         else
10129                                 goto out;
10130                 }
10131
10132                 if (bsi.block_len &&
10133                     bsi.block_start + bsi.block_len == physical_block_start) {
10134                         bsi.block_len += len;
10135                 } else {
10136                         if (bsi.block_len) {
10137                                 ret = btrfs_add_swap_extent(sis, &bsi);
10138                                 if (ret)
10139                                         goto out;
10140                         }
10141                         bsi.start = start;
10142                         bsi.block_start = physical_block_start;
10143                         bsi.block_len = len;
10144                 }
10145
10146                 start += len;
10147         }
10148
10149         if (bsi.block_len)
10150                 ret = btrfs_add_swap_extent(sis, &bsi);
10151
10152 out:
10153         if (!IS_ERR_OR_NULL(em))
10154                 free_extent_map(em);
10155
10156         unlock_extent_cached(io_tree, 0, isize - 1, &cached_state);
10157
10158         if (ret)
10159                 btrfs_swap_deactivate(file);
10160
10161         clear_bit(BTRFS_FS_EXCL_OP, &fs_info->flags);
10162
10163         if (ret)
10164                 return ret;
10165
10166         if (device)
10167                 sis->bdev = device->bdev;
10168         *span = bsi.highest_ppage - bsi.lowest_ppage + 1;
10169         sis->max = bsi.nr_pages;
10170         sis->pages = bsi.nr_pages - 1;
10171         sis->highest_bit = bsi.nr_pages - 1;
10172         return bsi.nr_extents;
10173 }
10174 #else
10175 static void btrfs_swap_deactivate(struct file *file)
10176 {
10177 }
10178
10179 static int btrfs_swap_activate(struct swap_info_struct *sis, struct file *file,
10180                                sector_t *span)
10181 {
10182         return -EOPNOTSUPP;
10183 }
10184 #endif
10185
10186 static const struct inode_operations btrfs_dir_inode_operations = {
10187         .getattr        = btrfs_getattr,
10188         .lookup         = btrfs_lookup,
10189         .create         = btrfs_create,
10190         .unlink         = btrfs_unlink,
10191         .link           = btrfs_link,
10192         .mkdir          = btrfs_mkdir,
10193         .rmdir          = btrfs_rmdir,
10194         .rename         = btrfs_rename2,
10195         .symlink        = btrfs_symlink,
10196         .setattr        = btrfs_setattr,
10197         .mknod          = btrfs_mknod,
10198         .listxattr      = btrfs_listxattr,
10199         .permission     = btrfs_permission,
10200         .get_acl        = btrfs_get_acl,
10201         .set_acl        = btrfs_set_acl,
10202         .update_time    = btrfs_update_time,
10203         .tmpfile        = btrfs_tmpfile,
10204 };
10205
10206 static const struct file_operations btrfs_dir_file_operations = {
10207         .llseek         = generic_file_llseek,
10208         .read           = generic_read_dir,
10209         .iterate_shared = btrfs_real_readdir,
10210         .open           = btrfs_opendir,
10211         .unlocked_ioctl = btrfs_ioctl,
10212 #ifdef CONFIG_COMPAT
10213         .compat_ioctl   = btrfs_compat_ioctl,
10214 #endif
10215         .release        = btrfs_release_file,
10216         .fsync          = btrfs_sync_file,
10217 };
10218
10219 static const struct extent_io_ops btrfs_extent_io_ops = {
10220         /* mandatory callbacks */
10221         .submit_bio_hook = btrfs_submit_bio_hook,
10222         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10223 };
10224
10225 /*
10226  * btrfs doesn't support the bmap operation because swapfiles
10227  * use bmap to make a mapping of extents in the file.  They assume
10228  * these extents won't change over the life of the file and they
10229  * use the bmap result to do IO directly to the drive.
10230  *
10231  * the btrfs bmap call would return logical addresses that aren't
10232  * suitable for IO and they also will change frequently as COW
10233  * operations happen.  So, swapfile + btrfs == corruption.
10234  *
10235  * For now we're avoiding this by dropping bmap.
10236  */
10237 static const struct address_space_operations btrfs_aops = {
10238         .readpage       = btrfs_readpage,
10239         .writepage      = btrfs_writepage,
10240         .writepages     = btrfs_writepages,
10241         .readahead      = btrfs_readahead,
10242         .direct_IO      = btrfs_direct_IO,
10243         .invalidatepage = btrfs_invalidatepage,
10244         .releasepage    = btrfs_releasepage,
10245 #ifdef CONFIG_MIGRATION
10246         .migratepage    = btrfs_migratepage,
10247 #endif
10248         .set_page_dirty = btrfs_set_page_dirty,
10249         .error_remove_page = generic_error_remove_page,
10250         .swap_activate  = btrfs_swap_activate,
10251         .swap_deactivate = btrfs_swap_deactivate,
10252 };
10253
10254 static const struct inode_operations btrfs_file_inode_operations = {
10255         .getattr        = btrfs_getattr,
10256         .setattr        = btrfs_setattr,
10257         .listxattr      = btrfs_listxattr,
10258         .permission     = btrfs_permission,
10259         .fiemap         = btrfs_fiemap,
10260         .get_acl        = btrfs_get_acl,
10261         .set_acl        = btrfs_set_acl,
10262         .update_time    = btrfs_update_time,
10263 };
10264 static const struct inode_operations btrfs_special_inode_operations = {
10265         .getattr        = btrfs_getattr,
10266         .setattr        = btrfs_setattr,
10267         .permission     = btrfs_permission,
10268         .listxattr      = btrfs_listxattr,
10269         .get_acl        = btrfs_get_acl,
10270         .set_acl        = btrfs_set_acl,
10271         .update_time    = btrfs_update_time,
10272 };
10273 static const struct inode_operations btrfs_symlink_inode_operations = {
10274         .get_link       = page_get_link,
10275         .getattr        = btrfs_getattr,
10276         .setattr        = btrfs_setattr,
10277         .permission     = btrfs_permission,
10278         .listxattr      = btrfs_listxattr,
10279         .update_time    = btrfs_update_time,
10280 };
10281
10282 const struct dentry_operations btrfs_dentry_operations = {
10283         .d_delete       = btrfs_dentry_delete,
10284 };