Btrfs: improve the performance of the csums lookup
[firefly-linux-kernel-4.4.55.git] / fs / btrfs / file-item.c
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18
19 #include <linux/bio.h>
20 #include <linux/slab.h>
21 #include <linux/pagemap.h>
22 #include <linux/highmem.h>
23 #include "ctree.h"
24 #include "disk-io.h"
25 #include "transaction.h"
26 #include "print-tree.h"
27
28 #define __MAX_CSUM_ITEMS(r, size) ((unsigned long)(((BTRFS_LEAF_DATA_SIZE(r) - \
29                                    sizeof(struct btrfs_item) * 2) / \
30                                   size) - 1))
31
32 #define MAX_CSUM_ITEMS(r, size) (min_t(u32, __MAX_CSUM_ITEMS(r, size), \
33                                        PAGE_CACHE_SIZE))
34
35 #define MAX_ORDERED_SUM_BYTES(r) ((PAGE_SIZE - \
36                                    sizeof(struct btrfs_ordered_sum)) / \
37                                    sizeof(struct btrfs_sector_sum) * \
38                                    (r)->sectorsize - (r)->sectorsize)
39
40 int btrfs_insert_file_extent(struct btrfs_trans_handle *trans,
41                              struct btrfs_root *root,
42                              u64 objectid, u64 pos,
43                              u64 disk_offset, u64 disk_num_bytes,
44                              u64 num_bytes, u64 offset, u64 ram_bytes,
45                              u8 compression, u8 encryption, u16 other_encoding)
46 {
47         int ret = 0;
48         struct btrfs_file_extent_item *item;
49         struct btrfs_key file_key;
50         struct btrfs_path *path;
51         struct extent_buffer *leaf;
52
53         path = btrfs_alloc_path();
54         if (!path)
55                 return -ENOMEM;
56         file_key.objectid = objectid;
57         file_key.offset = pos;
58         btrfs_set_key_type(&file_key, BTRFS_EXTENT_DATA_KEY);
59
60         path->leave_spinning = 1;
61         ret = btrfs_insert_empty_item(trans, root, path, &file_key,
62                                       sizeof(*item));
63         if (ret < 0)
64                 goto out;
65         BUG_ON(ret); /* Can't happen */
66         leaf = path->nodes[0];
67         item = btrfs_item_ptr(leaf, path->slots[0],
68                               struct btrfs_file_extent_item);
69         btrfs_set_file_extent_disk_bytenr(leaf, item, disk_offset);
70         btrfs_set_file_extent_disk_num_bytes(leaf, item, disk_num_bytes);
71         btrfs_set_file_extent_offset(leaf, item, offset);
72         btrfs_set_file_extent_num_bytes(leaf, item, num_bytes);
73         btrfs_set_file_extent_ram_bytes(leaf, item, ram_bytes);
74         btrfs_set_file_extent_generation(leaf, item, trans->transid);
75         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
76         btrfs_set_file_extent_compression(leaf, item, compression);
77         btrfs_set_file_extent_encryption(leaf, item, encryption);
78         btrfs_set_file_extent_other_encoding(leaf, item, other_encoding);
79
80         btrfs_mark_buffer_dirty(leaf);
81 out:
82         btrfs_free_path(path);
83         return ret;
84 }
85
86 struct btrfs_csum_item *btrfs_lookup_csum(struct btrfs_trans_handle *trans,
87                                           struct btrfs_root *root,
88                                           struct btrfs_path *path,
89                                           u64 bytenr, int cow)
90 {
91         int ret;
92         struct btrfs_key file_key;
93         struct btrfs_key found_key;
94         struct btrfs_csum_item *item;
95         struct extent_buffer *leaf;
96         u64 csum_offset = 0;
97         u16 csum_size = btrfs_super_csum_size(root->fs_info->super_copy);
98         int csums_in_item;
99
100         file_key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
101         file_key.offset = bytenr;
102         btrfs_set_key_type(&file_key, BTRFS_EXTENT_CSUM_KEY);
103         ret = btrfs_search_slot(trans, root, &file_key, path, 0, cow);
104         if (ret < 0)
105                 goto fail;
106         leaf = path->nodes[0];
107         if (ret > 0) {
108                 ret = 1;
109                 if (path->slots[0] == 0)
110                         goto fail;
111                 path->slots[0]--;
112                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
113                 if (btrfs_key_type(&found_key) != BTRFS_EXTENT_CSUM_KEY)
114                         goto fail;
115
116                 csum_offset = (bytenr - found_key.offset) >>
117                                 root->fs_info->sb->s_blocksize_bits;
118                 csums_in_item = btrfs_item_size_nr(leaf, path->slots[0]);
119                 csums_in_item /= csum_size;
120
121                 if (csum_offset == csums_in_item) {
122                         ret = -EFBIG;
123                         goto fail;
124                 } else if (csum_offset > csums_in_item) {
125                         goto fail;
126                 }
127         }
128         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_csum_item);
129         item = (struct btrfs_csum_item *)((unsigned char *)item +
130                                           csum_offset * csum_size);
131         return item;
132 fail:
133         if (ret > 0)
134                 ret = -ENOENT;
135         return ERR_PTR(ret);
136 }
137
138 int btrfs_lookup_file_extent(struct btrfs_trans_handle *trans,
139                              struct btrfs_root *root,
140                              struct btrfs_path *path, u64 objectid,
141                              u64 offset, int mod)
142 {
143         int ret;
144         struct btrfs_key file_key;
145         int ins_len = mod < 0 ? -1 : 0;
146         int cow = mod != 0;
147
148         file_key.objectid = objectid;
149         file_key.offset = offset;
150         btrfs_set_key_type(&file_key, BTRFS_EXTENT_DATA_KEY);
151         ret = btrfs_search_slot(trans, root, &file_key, path, ins_len, cow);
152         return ret;
153 }
154
155 u64 btrfs_file_extent_length(struct btrfs_path *path)
156 {
157         int extent_type;
158         struct btrfs_file_extent_item *fi;
159         u64 len;
160
161         fi = btrfs_item_ptr(path->nodes[0], path->slots[0],
162                             struct btrfs_file_extent_item);
163         extent_type = btrfs_file_extent_type(path->nodes[0], fi);
164
165         if (extent_type == BTRFS_FILE_EXTENT_REG ||
166             extent_type == BTRFS_FILE_EXTENT_PREALLOC)
167                 len = btrfs_file_extent_num_bytes(path->nodes[0], fi);
168         else if (extent_type == BTRFS_FILE_EXTENT_INLINE)
169                 len = btrfs_file_extent_inline_len(path->nodes[0], fi);
170         else
171                 BUG();
172
173         return len;
174 }
175
176 static int __btrfs_lookup_bio_sums(struct btrfs_root *root,
177                                    struct inode *inode, struct bio *bio,
178                                    u64 logical_offset, u32 *dst, int dio)
179 {
180         u32 sum[16];
181         int len;
182         struct bio_vec *bvec = bio->bi_io_vec;
183         int bio_index = 0;
184         u64 offset = 0;
185         u64 item_start_offset = 0;
186         u64 item_last_offset = 0;
187         u64 disk_bytenr;
188         u32 diff;
189         u16 csum_size = btrfs_super_csum_size(root->fs_info->super_copy);
190         int count;
191         struct btrfs_path *path;
192         struct btrfs_csum_item *item = NULL;
193         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
194
195         path = btrfs_alloc_path();
196         if (!path)
197                 return -ENOMEM;
198         if (bio->bi_size > PAGE_CACHE_SIZE * 8)
199                 path->reada = 2;
200
201         WARN_ON(bio->bi_vcnt <= 0);
202
203         /*
204          * the free space stuff is only read when it hasn't been
205          * updated in the current transaction.  So, we can safely
206          * read from the commit root and sidestep a nasty deadlock
207          * between reading the free space cache and updating the csum tree.
208          */
209         if (btrfs_is_free_space_inode(inode)) {
210                 path->search_commit_root = 1;
211                 path->skip_locking = 1;
212         }
213
214         disk_bytenr = (u64)bio->bi_sector << 9;
215         if (dio)
216                 offset = logical_offset;
217         while (bio_index < bio->bi_vcnt) {
218                 len = min_t(int, ARRAY_SIZE(sum), bio->bi_vcnt - bio_index);
219                 if (!dio)
220                         offset = page_offset(bvec->bv_page) + bvec->bv_offset;
221                 count = btrfs_find_ordered_sum(inode, offset, disk_bytenr, sum,
222                                                len);
223                 if (count)
224                         goto found;
225
226                 if (!item || disk_bytenr < item_start_offset ||
227                     disk_bytenr >= item_last_offset) {
228                         struct btrfs_key found_key;
229                         u32 item_size;
230
231                         if (item)
232                                 btrfs_release_path(path);
233                         item = btrfs_lookup_csum(NULL, root->fs_info->csum_root,
234                                                  path, disk_bytenr, 0);
235                         if (IS_ERR(item)) {
236                                 count = 1;
237                                 sum[0] = 0;
238                                 if (BTRFS_I(inode)->root->root_key.objectid ==
239                                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
240                                         set_extent_bits(io_tree, offset,
241                                                 offset + bvec->bv_len - 1,
242                                                 EXTENT_NODATASUM, GFP_NOFS);
243                                 } else {
244                                         printk(KERN_INFO "btrfs no csum found "
245                                                "for inode %llu start %llu\n",
246                                                (unsigned long long)
247                                                btrfs_ino(inode),
248                                                (unsigned long long)offset);
249                                 }
250                                 item = NULL;
251                                 btrfs_release_path(path);
252                                 goto found;
253                         }
254                         btrfs_item_key_to_cpu(path->nodes[0], &found_key,
255                                               path->slots[0]);
256
257                         item_start_offset = found_key.offset;
258                         item_size = btrfs_item_size_nr(path->nodes[0],
259                                                        path->slots[0]);
260                         item_last_offset = item_start_offset +
261                                 (item_size / csum_size) *
262                                 root->sectorsize;
263                         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
264                                               struct btrfs_csum_item);
265                 }
266                 /*
267                  * this byte range must be able to fit inside
268                  * a single leaf so it will also fit inside a u32
269                  */
270                 diff = disk_bytenr - item_start_offset;
271                 diff = diff / root->sectorsize;
272                 diff = diff * csum_size;
273                 count = min_t(int, len, (item_last_offset - disk_bytenr) >>
274                                         inode->i_sb->s_blocksize_bits);
275                 read_extent_buffer(path->nodes[0], sum,
276                                    ((unsigned long)item) + diff,
277                                    csum_size * count);
278 found:
279                 if (dst) {
280                         memcpy(dst, sum, count * csum_size);
281                         dst += count;
282                 } else {
283                         if (dio)
284                                 extent_cache_csums_dio(io_tree, offset, sum,
285                                                        count);
286                         else
287                                 extent_cache_csums(io_tree, bio, bio_index, sum,
288                                             count);
289                 }
290                 while (count--) {
291                         disk_bytenr += bvec->bv_len;
292                         offset += bvec->bv_len;
293                         bio_index++;
294                         bvec++;
295                 }
296         }
297         btrfs_free_path(path);
298         return 0;
299 }
300
301 int btrfs_lookup_bio_sums(struct btrfs_root *root, struct inode *inode,
302                           struct bio *bio, u32 *dst)
303 {
304         return __btrfs_lookup_bio_sums(root, inode, bio, 0, dst, 0);
305 }
306
307 int btrfs_lookup_bio_sums_dio(struct btrfs_root *root, struct inode *inode,
308                               struct bio *bio, u64 offset)
309 {
310         return __btrfs_lookup_bio_sums(root, inode, bio, offset, NULL, 1);
311 }
312
313 int btrfs_lookup_csums_range(struct btrfs_root *root, u64 start, u64 end,
314                              struct list_head *list, int search_commit)
315 {
316         struct btrfs_key key;
317         struct btrfs_path *path;
318         struct extent_buffer *leaf;
319         struct btrfs_ordered_sum *sums;
320         struct btrfs_sector_sum *sector_sum;
321         struct btrfs_csum_item *item;
322         LIST_HEAD(tmplist);
323         unsigned long offset;
324         int ret;
325         size_t size;
326         u64 csum_end;
327         u16 csum_size = btrfs_super_csum_size(root->fs_info->super_copy);
328
329         path = btrfs_alloc_path();
330         if (!path)
331                 return -ENOMEM;
332
333         if (search_commit) {
334                 path->skip_locking = 1;
335                 path->reada = 2;
336                 path->search_commit_root = 1;
337         }
338
339         key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
340         key.offset = start;
341         key.type = BTRFS_EXTENT_CSUM_KEY;
342
343         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
344         if (ret < 0)
345                 goto fail;
346         if (ret > 0 && path->slots[0] > 0) {
347                 leaf = path->nodes[0];
348                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
349                 if (key.objectid == BTRFS_EXTENT_CSUM_OBJECTID &&
350                     key.type == BTRFS_EXTENT_CSUM_KEY) {
351                         offset = (start - key.offset) >>
352                                  root->fs_info->sb->s_blocksize_bits;
353                         if (offset * csum_size <
354                             btrfs_item_size_nr(leaf, path->slots[0] - 1))
355                                 path->slots[0]--;
356                 }
357         }
358
359         while (start <= end) {
360                 leaf = path->nodes[0];
361                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
362                         ret = btrfs_next_leaf(root, path);
363                         if (ret < 0)
364                                 goto fail;
365                         if (ret > 0)
366                                 break;
367                         leaf = path->nodes[0];
368                 }
369
370                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
371                 if (key.objectid != BTRFS_EXTENT_CSUM_OBJECTID ||
372                     key.type != BTRFS_EXTENT_CSUM_KEY ||
373                     key.offset > end)
374                         break;
375
376                 if (key.offset > start)
377                         start = key.offset;
378
379                 size = btrfs_item_size_nr(leaf, path->slots[0]);
380                 csum_end = key.offset + (size / csum_size) * root->sectorsize;
381                 if (csum_end <= start) {
382                         path->slots[0]++;
383                         continue;
384                 }
385
386                 csum_end = min(csum_end, end + 1);
387                 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
388                                       struct btrfs_csum_item);
389                 while (start < csum_end) {
390                         size = min_t(size_t, csum_end - start,
391                                         MAX_ORDERED_SUM_BYTES(root));
392                         sums = kzalloc(btrfs_ordered_sum_size(root, size),
393                                         GFP_NOFS);
394                         if (!sums) {
395                                 ret = -ENOMEM;
396                                 goto fail;
397                         }
398
399                         sector_sum = sums->sums;
400                         sums->bytenr = start;
401                         sums->len = size;
402
403                         offset = (start - key.offset) >>
404                                 root->fs_info->sb->s_blocksize_bits;
405                         offset *= csum_size;
406
407                         while (size > 0) {
408                                 read_extent_buffer(path->nodes[0],
409                                                 &sector_sum->sum,
410                                                 ((unsigned long)item) +
411                                                 offset, csum_size);
412                                 sector_sum->bytenr = start;
413
414                                 size -= root->sectorsize;
415                                 start += root->sectorsize;
416                                 offset += csum_size;
417                                 sector_sum++;
418                         }
419                         list_add_tail(&sums->list, &tmplist);
420                 }
421                 path->slots[0]++;
422         }
423         ret = 0;
424 fail:
425         while (ret < 0 && !list_empty(&tmplist)) {
426                 sums = list_entry(&tmplist, struct btrfs_ordered_sum, list);
427                 list_del(&sums->list);
428                 kfree(sums);
429         }
430         list_splice_tail(&tmplist, list);
431
432         btrfs_free_path(path);
433         return ret;
434 }
435
436 int btrfs_csum_one_bio(struct btrfs_root *root, struct inode *inode,
437                        struct bio *bio, u64 file_start, int contig)
438 {
439         struct btrfs_ordered_sum *sums;
440         struct btrfs_sector_sum *sector_sum;
441         struct btrfs_ordered_extent *ordered;
442         char *data;
443         struct bio_vec *bvec = bio->bi_io_vec;
444         int bio_index = 0;
445         unsigned long total_bytes = 0;
446         unsigned long this_sum_bytes = 0;
447         u64 offset;
448         u64 disk_bytenr;
449
450         WARN_ON(bio->bi_vcnt <= 0);
451         sums = kzalloc(btrfs_ordered_sum_size(root, bio->bi_size), GFP_NOFS);
452         if (!sums)
453                 return -ENOMEM;
454
455         sector_sum = sums->sums;
456         disk_bytenr = (u64)bio->bi_sector << 9;
457         sums->len = bio->bi_size;
458         INIT_LIST_HEAD(&sums->list);
459
460         if (contig)
461                 offset = file_start;
462         else
463                 offset = page_offset(bvec->bv_page) + bvec->bv_offset;
464
465         ordered = btrfs_lookup_ordered_extent(inode, offset);
466         BUG_ON(!ordered); /* Logic error */
467         sums->bytenr = ordered->start;
468
469         while (bio_index < bio->bi_vcnt) {
470                 if (!contig)
471                         offset = page_offset(bvec->bv_page) + bvec->bv_offset;
472
473                 if (offset >= ordered->file_offset + ordered->len ||
474                     offset < ordered->file_offset) {
475                         unsigned long bytes_left;
476                         sums->len = this_sum_bytes;
477                         this_sum_bytes = 0;
478                         btrfs_add_ordered_sum(inode, ordered, sums);
479                         btrfs_put_ordered_extent(ordered);
480
481                         bytes_left = bio->bi_size - total_bytes;
482
483                         sums = kzalloc(btrfs_ordered_sum_size(root, bytes_left),
484                                        GFP_NOFS);
485                         BUG_ON(!sums); /* -ENOMEM */
486                         sector_sum = sums->sums;
487                         sums->len = bytes_left;
488                         ordered = btrfs_lookup_ordered_extent(inode, offset);
489                         BUG_ON(!ordered); /* Logic error */
490                         sums->bytenr = ordered->start;
491                 }
492
493                 data = kmap_atomic(bvec->bv_page);
494                 sector_sum->sum = ~(u32)0;
495                 sector_sum->sum = btrfs_csum_data(data + bvec->bv_offset,
496                                                   sector_sum->sum,
497                                                   bvec->bv_len);
498                 kunmap_atomic(data);
499                 btrfs_csum_final(sector_sum->sum,
500                                  (char *)&sector_sum->sum);
501                 sector_sum->bytenr = disk_bytenr;
502
503                 sector_sum++;
504                 bio_index++;
505                 total_bytes += bvec->bv_len;
506                 this_sum_bytes += bvec->bv_len;
507                 disk_bytenr += bvec->bv_len;
508                 offset += bvec->bv_len;
509                 bvec++;
510         }
511         this_sum_bytes = 0;
512         btrfs_add_ordered_sum(inode, ordered, sums);
513         btrfs_put_ordered_extent(ordered);
514         return 0;
515 }
516
517 /*
518  * helper function for csum removal, this expects the
519  * key to describe the csum pointed to by the path, and it expects
520  * the csum to overlap the range [bytenr, len]
521  *
522  * The csum should not be entirely contained in the range and the
523  * range should not be entirely contained in the csum.
524  *
525  * This calls btrfs_truncate_item with the correct args based on the
526  * overlap, and fixes up the key as required.
527  */
528 static noinline void truncate_one_csum(struct btrfs_trans_handle *trans,
529                                        struct btrfs_root *root,
530                                        struct btrfs_path *path,
531                                        struct btrfs_key *key,
532                                        u64 bytenr, u64 len)
533 {
534         struct extent_buffer *leaf;
535         u16 csum_size = btrfs_super_csum_size(root->fs_info->super_copy);
536         u64 csum_end;
537         u64 end_byte = bytenr + len;
538         u32 blocksize_bits = root->fs_info->sb->s_blocksize_bits;
539
540         leaf = path->nodes[0];
541         csum_end = btrfs_item_size_nr(leaf, path->slots[0]) / csum_size;
542         csum_end <<= root->fs_info->sb->s_blocksize_bits;
543         csum_end += key->offset;
544
545         if (key->offset < bytenr && csum_end <= end_byte) {
546                 /*
547                  *         [ bytenr - len ]
548                  *         [   ]
549                  *   [csum     ]
550                  *   A simple truncate off the end of the item
551                  */
552                 u32 new_size = (bytenr - key->offset) >> blocksize_bits;
553                 new_size *= csum_size;
554                 btrfs_truncate_item(trans, root, path, new_size, 1);
555         } else if (key->offset >= bytenr && csum_end > end_byte &&
556                    end_byte > key->offset) {
557                 /*
558                  *         [ bytenr - len ]
559                  *                 [ ]
560                  *                 [csum     ]
561                  * we need to truncate from the beginning of the csum
562                  */
563                 u32 new_size = (csum_end - end_byte) >> blocksize_bits;
564                 new_size *= csum_size;
565
566                 btrfs_truncate_item(trans, root, path, new_size, 0);
567
568                 key->offset = end_byte;
569                 btrfs_set_item_key_safe(trans, root, path, key);
570         } else {
571                 BUG();
572         }
573 }
574
575 /*
576  * deletes the csum items from the csum tree for a given
577  * range of bytes.
578  */
579 int btrfs_del_csums(struct btrfs_trans_handle *trans,
580                     struct btrfs_root *root, u64 bytenr, u64 len)
581 {
582         struct btrfs_path *path;
583         struct btrfs_key key;
584         u64 end_byte = bytenr + len;
585         u64 csum_end;
586         struct extent_buffer *leaf;
587         int ret;
588         u16 csum_size = btrfs_super_csum_size(root->fs_info->super_copy);
589         int blocksize_bits = root->fs_info->sb->s_blocksize_bits;
590
591         root = root->fs_info->csum_root;
592
593         path = btrfs_alloc_path();
594         if (!path)
595                 return -ENOMEM;
596
597         while (1) {
598                 key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
599                 key.offset = end_byte - 1;
600                 key.type = BTRFS_EXTENT_CSUM_KEY;
601
602                 path->leave_spinning = 1;
603                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
604                 if (ret > 0) {
605                         if (path->slots[0] == 0)
606                                 break;
607                         path->slots[0]--;
608                 } else if (ret < 0) {
609                         break;
610                 }
611
612                 leaf = path->nodes[0];
613                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
614
615                 if (key.objectid != BTRFS_EXTENT_CSUM_OBJECTID ||
616                     key.type != BTRFS_EXTENT_CSUM_KEY) {
617                         break;
618                 }
619
620                 if (key.offset >= end_byte)
621                         break;
622
623                 csum_end = btrfs_item_size_nr(leaf, path->slots[0]) / csum_size;
624                 csum_end <<= blocksize_bits;
625                 csum_end += key.offset;
626
627                 /* this csum ends before we start, we're done */
628                 if (csum_end <= bytenr)
629                         break;
630
631                 /* delete the entire item, it is inside our range */
632                 if (key.offset >= bytenr && csum_end <= end_byte) {
633                         ret = btrfs_del_item(trans, root, path);
634                         if (ret)
635                                 goto out;
636                         if (key.offset == bytenr)
637                                 break;
638                 } else if (key.offset < bytenr && csum_end > end_byte) {
639                         unsigned long offset;
640                         unsigned long shift_len;
641                         unsigned long item_offset;
642                         /*
643                          *        [ bytenr - len ]
644                          *     [csum                ]
645                          *
646                          * Our bytes are in the middle of the csum,
647                          * we need to split this item and insert a new one.
648                          *
649                          * But we can't drop the path because the
650                          * csum could change, get removed, extended etc.
651                          *
652                          * The trick here is the max size of a csum item leaves
653                          * enough room in the tree block for a single
654                          * item header.  So, we split the item in place,
655                          * adding a new header pointing to the existing
656                          * bytes.  Then we loop around again and we have
657                          * a nicely formed csum item that we can neatly
658                          * truncate.
659                          */
660                         offset = (bytenr - key.offset) >> blocksize_bits;
661                         offset *= csum_size;
662
663                         shift_len = (len >> blocksize_bits) * csum_size;
664
665                         item_offset = btrfs_item_ptr_offset(leaf,
666                                                             path->slots[0]);
667
668                         memset_extent_buffer(leaf, 0, item_offset + offset,
669                                              shift_len);
670                         key.offset = bytenr;
671
672                         /*
673                          * btrfs_split_item returns -EAGAIN when the
674                          * item changed size or key
675                          */
676                         ret = btrfs_split_item(trans, root, path, &key, offset);
677                         if (ret && ret != -EAGAIN) {
678                                 btrfs_abort_transaction(trans, root, ret);
679                                 goto out;
680                         }
681
682                         key.offset = end_byte - 1;
683                 } else {
684                         truncate_one_csum(trans, root, path, &key, bytenr, len);
685                         if (key.offset < bytenr)
686                                 break;
687                 }
688                 btrfs_release_path(path);
689         }
690         ret = 0;
691 out:
692         btrfs_free_path(path);
693         return ret;
694 }
695
696 static u64 btrfs_sector_sum_left(struct btrfs_ordered_sum *sums,
697                                  struct btrfs_sector_sum *sector_sum,
698                                  u64 total_bytes, u64 sectorsize)
699 {
700         u64 tmp = sectorsize;
701         u64 next_sector = sector_sum->bytenr;
702         struct btrfs_sector_sum *next = sector_sum + 1;
703
704         while ((tmp + total_bytes) < sums->len) {
705                 if (next_sector + sectorsize != next->bytenr)
706                         break;
707                 tmp += sectorsize;
708                 next_sector = next->bytenr;
709                 next++;
710         }
711         return tmp;
712 }
713
714 int btrfs_csum_file_blocks(struct btrfs_trans_handle *trans,
715                            struct btrfs_root *root,
716                            struct btrfs_ordered_sum *sums)
717 {
718         u64 bytenr;
719         int ret;
720         struct btrfs_key file_key;
721         struct btrfs_key found_key;
722         u64 next_offset;
723         u64 total_bytes = 0;
724         int found_next;
725         struct btrfs_path *path;
726         struct btrfs_csum_item *item;
727         struct btrfs_csum_item *item_end;
728         struct extent_buffer *leaf = NULL;
729         u64 csum_offset;
730         struct btrfs_sector_sum *sector_sum;
731         u32 nritems;
732         u32 ins_size;
733         u16 csum_size = btrfs_super_csum_size(root->fs_info->super_copy);
734
735         path = btrfs_alloc_path();
736         if (!path)
737                 return -ENOMEM;
738
739         sector_sum = sums->sums;
740 again:
741         next_offset = (u64)-1;
742         found_next = 0;
743         file_key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
744         file_key.offset = sector_sum->bytenr;
745         bytenr = sector_sum->bytenr;
746         btrfs_set_key_type(&file_key, BTRFS_EXTENT_CSUM_KEY);
747
748         item = btrfs_lookup_csum(trans, root, path, sector_sum->bytenr, 1);
749         if (!IS_ERR(item)) {
750                 leaf = path->nodes[0];
751                 ret = 0;
752                 goto found;
753         }
754         ret = PTR_ERR(item);
755         if (ret != -EFBIG && ret != -ENOENT)
756                 goto fail_unlock;
757
758         if (ret == -EFBIG) {
759                 u32 item_size;
760                 /* we found one, but it isn't big enough yet */
761                 leaf = path->nodes[0];
762                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
763                 if ((item_size / csum_size) >=
764                     MAX_CSUM_ITEMS(root, csum_size)) {
765                         /* already at max size, make a new one */
766                         goto insert;
767                 }
768         } else {
769                 int slot = path->slots[0] + 1;
770                 /* we didn't find a csum item, insert one */
771                 nritems = btrfs_header_nritems(path->nodes[0]);
772                 if (path->slots[0] >= nritems - 1) {
773                         ret = btrfs_next_leaf(root, path);
774                         if (ret == 1)
775                                 found_next = 1;
776                         if (ret != 0)
777                                 goto insert;
778                         slot = 0;
779                 }
780                 btrfs_item_key_to_cpu(path->nodes[0], &found_key, slot);
781                 if (found_key.objectid != BTRFS_EXTENT_CSUM_OBJECTID ||
782                     found_key.type != BTRFS_EXTENT_CSUM_KEY) {
783                         found_next = 1;
784                         goto insert;
785                 }
786                 next_offset = found_key.offset;
787                 found_next = 1;
788                 goto insert;
789         }
790
791         /*
792          * at this point, we know the tree has an item, but it isn't big
793          * enough yet to put our csum in.  Grow it
794          */
795         btrfs_release_path(path);
796         ret = btrfs_search_slot(trans, root, &file_key, path,
797                                 csum_size, 1);
798         if (ret < 0)
799                 goto fail_unlock;
800
801         if (ret > 0) {
802                 if (path->slots[0] == 0)
803                         goto insert;
804                 path->slots[0]--;
805         }
806
807         leaf = path->nodes[0];
808         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
809         csum_offset = (bytenr - found_key.offset) >>
810                         root->fs_info->sb->s_blocksize_bits;
811
812         if (btrfs_key_type(&found_key) != BTRFS_EXTENT_CSUM_KEY ||
813             found_key.objectid != BTRFS_EXTENT_CSUM_OBJECTID ||
814             csum_offset >= MAX_CSUM_ITEMS(root, csum_size)) {
815                 goto insert;
816         }
817
818         if (csum_offset == btrfs_item_size_nr(leaf, path->slots[0]) /
819             csum_size) {
820                 int extend_nr;
821                 u64 tmp;
822                 u32 diff;
823                 u32 free_space;
824
825                 if (btrfs_leaf_free_space(root, leaf) <
826                                  sizeof(struct btrfs_item) + csum_size * 2)
827                         goto insert;
828
829                 free_space = btrfs_leaf_free_space(root, leaf) -
830                                          sizeof(struct btrfs_item) - csum_size;
831                 tmp = btrfs_sector_sum_left(sums, sector_sum, total_bytes,
832                                             root->sectorsize);
833                 tmp >>= root->fs_info->sb->s_blocksize_bits;
834                 WARN_ON(tmp < 1);
835
836                 extend_nr = max_t(int, 1, (int)tmp);
837                 diff = (csum_offset + extend_nr) * csum_size;
838                 diff = min(diff, MAX_CSUM_ITEMS(root, csum_size) * csum_size);
839
840                 diff = diff - btrfs_item_size_nr(leaf, path->slots[0]);
841                 diff = min(free_space, diff);
842                 diff /= csum_size;
843                 diff *= csum_size;
844
845                 btrfs_extend_item(trans, root, path, diff);
846                 goto csum;
847         }
848
849 insert:
850         btrfs_release_path(path);
851         csum_offset = 0;
852         if (found_next) {
853                 u64 tmp;
854
855                 tmp = btrfs_sector_sum_left(sums, sector_sum, total_bytes,
856                                             root->sectorsize);
857                 tmp >>= root->fs_info->sb->s_blocksize_bits;
858                 tmp = min(tmp, (next_offset - file_key.offset) >>
859                                          root->fs_info->sb->s_blocksize_bits);
860
861                 tmp = max((u64)1, tmp);
862                 tmp = min(tmp, (u64)MAX_CSUM_ITEMS(root, csum_size));
863                 ins_size = csum_size * tmp;
864         } else {
865                 ins_size = csum_size;
866         }
867         path->leave_spinning = 1;
868         ret = btrfs_insert_empty_item(trans, root, path, &file_key,
869                                       ins_size);
870         path->leave_spinning = 0;
871         if (ret < 0)
872                 goto fail_unlock;
873         if (ret != 0) {
874                 WARN_ON(1);
875                 goto fail_unlock;
876         }
877 csum:
878         leaf = path->nodes[0];
879         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_csum_item);
880         ret = 0;
881         item = (struct btrfs_csum_item *)((unsigned char *)item +
882                                           csum_offset * csum_size);
883 found:
884         item_end = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_csum_item);
885         item_end = (struct btrfs_csum_item *)((unsigned char *)item_end +
886                                       btrfs_item_size_nr(leaf, path->slots[0]));
887 next_sector:
888
889         write_extent_buffer(leaf, &sector_sum->sum, (unsigned long)item, csum_size);
890
891         total_bytes += root->sectorsize;
892         sector_sum++;
893         if (total_bytes < sums->len) {
894                 item = (struct btrfs_csum_item *)((char *)item +
895                                                   csum_size);
896                 if (item < item_end && bytenr + PAGE_CACHE_SIZE ==
897                     sector_sum->bytenr) {
898                         bytenr = sector_sum->bytenr;
899                         goto next_sector;
900                 }
901         }
902
903         btrfs_mark_buffer_dirty(path->nodes[0]);
904         if (total_bytes < sums->len) {
905                 btrfs_release_path(path);
906                 cond_resched();
907                 goto again;
908         }
909 out:
910         btrfs_free_path(path);
911         return ret;
912
913 fail_unlock:
914         goto out;
915 }