?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge branch 'linus' of git://git.kernel.org/pub/scm/linux/kernel/git/herbert/crypto-2.6
[linux/fpc-iii.git] / fs / btrfs / scrub.c
blob46d847f66e4b870dcd54dc18d023f8ca321f40fb
1 /*
2 * Copyright (C) 2011, 2012 STRATO. 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/blkdev.h>
20 #include <linux/ratelimit.h>
21 #include "ctree.h"
22 #include "volumes.h"
23 #include "disk-io.h"
24 #include "ordered-data.h"
25 #include "transaction.h"
26 #include "backref.h"
27 #include "extent_io.h"
28 #include "dev-replace.h"
29 #include "check-integrity.h"
30 #include "rcu-string.h"
31 #include "raid56.h"
32
33 /*
34 * This is only the first step towards a full-features scrub. It reads all
35 * extent and super block and verifies the checksums. In case a bad checksum
36 * is found or the extent cannot be read, good data will be written back if
37 * any can be found.
38 *
39 * Future enhancements:
40 * - In case an unrepairable extent is encountered, track which files are
41 * affected and report them
42 * - track and record media errors, throw out bad devices
43 * - add a mode to also read unallocated space
44 */
45
46 struct scrub_block;
47 struct scrub_ctx;
48
49 /*
50 * the following three values only influence the performance.
51 * The last one configures the number of parallel and outstanding I/O
52 * operations. The first two values configure an upper limit for the number
53 * of (dynamically allocated) pages that are added to a bio.
54 */
55 #define SCRUB_PAGES_PER_RD_BIO 32 /* 128k per bio */
56 #define SCRUB_PAGES_PER_WR_BIO 32 /* 128k per bio */
57 #define SCRUB_BIOS_PER_SCTX 64 /* 8MB per device in flight */
58
59 /*
60 * the following value times PAGE_SIZE needs to be large enough to match the
61 * largest node/leaf/sector size that shall be supported.
62 * Values larger than BTRFS_STRIPE_LEN are not supported.
63 */
64 #define SCRUB_MAX_PAGES_PER_BLOCK 16 /* 64k per node/leaf/sector */
65
66 struct scrub_recover {
67 atomic_t refs;
68 struct btrfs_bio *bbio;
69 u64 map_length;
70 };
71
72 struct scrub_page {
73 struct scrub_block *sblock;
74 struct page *page;
75 struct btrfs_device *dev;
76 struct list_head list;
77 u64 flags; /* extent flags */
78 u64 generation;
79 u64 logical;
80 u64 physical;
81 u64 physical_for_dev_replace;
82 atomic_t refs;
83 struct {
84 unsigned int mirror_num:8;
85 unsigned int have_csum:1;
86 unsigned int io_error:1;
87 };
88 u8 csum[BTRFS_CSUM_SIZE];
89
90 struct scrub_recover *recover;
91 };
92
93 struct scrub_bio {
94 int index;
95 struct scrub_ctx *sctx;
96 struct btrfs_device *dev;
97 struct bio *bio;
98 int err;
99 u64 logical;
100 u64 physical;
101 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
102 struct scrub_page *pagev[SCRUB_PAGES_PER_WR_BIO];
103 #else
104 struct scrub_page *pagev[SCRUB_PAGES_PER_RD_BIO];
105 #endif
106 int page_count;
107 int next_free;
108 struct btrfs_work work;
109 };
110
111 struct scrub_block {
112 struct scrub_page *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
113 int page_count;
114 atomic_t outstanding_pages;
115 atomic_t refs; /* free mem on transition to zero */
116 struct scrub_ctx *sctx;
117 struct scrub_parity *sparity;
118 struct {
119 unsigned int header_error:1;
120 unsigned int checksum_error:1;
121 unsigned int no_io_error_seen:1;
122 unsigned int generation_error:1; /* also sets header_error */
123
124 /* The following is for the data used to check parity */
125 /* It is for the data with checksum */
126 unsigned int data_corrected:1;
127 };
128 struct btrfs_work work;
129 };
130
131 /* Used for the chunks with parity stripe such RAID5/6 */
132 struct scrub_parity {
133 struct scrub_ctx *sctx;
134
135 struct btrfs_device *scrub_dev;
136
137 u64 logic_start;
138
139 u64 logic_end;
140
141 int nsectors;
142
143 int stripe_len;
144
145 atomic_t refs;
146
147 struct list_head spages;
148
149 /* Work of parity check and repair */
150 struct btrfs_work work;
151
152 /* Mark the parity blocks which have data */
153 unsigned long *dbitmap;
154
155 /*
156 * Mark the parity blocks which have data, but errors happen when
157 * read data or check data
158 */
159 unsigned long *ebitmap;
160
161 unsigned long bitmap[0];
162 };
163
164 struct scrub_wr_ctx {
165 struct scrub_bio *wr_curr_bio;
166 struct btrfs_device *tgtdev;
167 int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
168 atomic_t flush_all_writes;
169 struct mutex wr_lock;
170 };
171
172 struct scrub_ctx {
173 struct scrub_bio *bios[SCRUB_BIOS_PER_SCTX];
174 struct btrfs_root *dev_root;
175 int first_free;
176 int curr;
177 atomic_t bios_in_flight;
178 atomic_t workers_pending;
179 spinlock_t list_lock;
180 wait_queue_head_t list_wait;
181 u16 csum_size;
182 struct list_head csum_list;
183 atomic_t cancel_req;
184 int readonly;
185 int pages_per_rd_bio;
186 u32 sectorsize;
187 u32 nodesize;
188
189 int is_dev_replace;
190 struct scrub_wr_ctx wr_ctx;
191
192 /*
193 * statistics
194 */
195 struct btrfs_scrub_progress stat;
196 spinlock_t stat_lock;
197
198 /*
199 * Use a ref counter to avoid use-after-free issues. Scrub workers
200 * decrement bios_in_flight and workers_pending and then do a wakeup
201 * on the list_wait wait queue. We must ensure the main scrub task
202 * doesn't free the scrub context before or while the workers are
203 * doing the wakeup() call.
204 */
205 atomic_t refs;
206 };
207
208 struct scrub_fixup_nodatasum {
209 struct scrub_ctx *sctx;
210 struct btrfs_device *dev;
211 u64 logical;
212 struct btrfs_root *root;
213 struct btrfs_work work;
214 int mirror_num;
215 };
216
217 struct scrub_nocow_inode {
218 u64 inum;
219 u64 offset;
220 u64 root;
221 struct list_head list;
222 };
223
224 struct scrub_copy_nocow_ctx {
225 struct scrub_ctx *sctx;
226 u64 logical;
227 u64 len;
228 int mirror_num;
229 u64 physical_for_dev_replace;
230 struct list_head inodes;
231 struct btrfs_work work;
232 };
233
234 struct scrub_warning {
235 struct btrfs_path *path;
236 u64 extent_item_size;
237 const char *errstr;
238 sector_t sector;
239 u64 logical;
240 struct btrfs_device *dev;
241 };
242
243 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
244 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
245 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
246 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
247 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
248 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
249 struct scrub_block *sblocks_for_recheck);
250 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
251 struct scrub_block *sblock,
252 int retry_failed_mirror);
253 static void scrub_recheck_block_checksum(struct scrub_block *sblock);
254 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
255 struct scrub_block *sblock_good);
256 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
257 struct scrub_block *sblock_good,
258 int page_num, int force_write);
259 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
260 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
261 int page_num);
262 static int scrub_checksum_data(struct scrub_block *sblock);
263 static int scrub_checksum_tree_block(struct scrub_block *sblock);
264 static int scrub_checksum_super(struct scrub_block *sblock);
265 static void scrub_block_get(struct scrub_block *sblock);
266 static void scrub_block_put(struct scrub_block *sblock);
267 static void scrub_page_get(struct scrub_page *spage);
268 static void scrub_page_put(struct scrub_page *spage);
269 static void scrub_parity_get(struct scrub_parity *sparity);
270 static void scrub_parity_put(struct scrub_parity *sparity);
271 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
272 struct scrub_page *spage);
273 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
274 u64 physical, struct btrfs_device *dev, u64 flags,
275 u64 gen, int mirror_num, u8 *csum, int force,
276 u64 physical_for_dev_replace);
277 static void scrub_bio_end_io(struct bio *bio);
278 static void scrub_bio_end_io_worker(struct btrfs_work *work);
279 static void scrub_block_complete(struct scrub_block *sblock);
280 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
281 u64 extent_logical, u64 extent_len,
282 u64 *extent_physical,
283 struct btrfs_device **extent_dev,
284 int *extent_mirror_num);
285 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
286 struct scrub_wr_ctx *wr_ctx,
287 struct btrfs_fs_info *fs_info,
288 struct btrfs_device *dev,
289 int is_dev_replace);
290 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx);
291 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
292 struct scrub_page *spage);
293 static void scrub_wr_submit(struct scrub_ctx *sctx);
294 static void scrub_wr_bio_end_io(struct bio *bio);
295 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
296 static int write_page_nocow(struct scrub_ctx *sctx,
297 u64 physical_for_dev_replace, struct page *page);
298 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
299 struct scrub_copy_nocow_ctx *ctx);
300 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
301 int mirror_num, u64 physical_for_dev_replace);
302 static void copy_nocow_pages_worker(struct btrfs_work *work);
303 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
304 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
305 static void scrub_put_ctx(struct scrub_ctx *sctx);
306
307
308 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
309 {
310 atomic_inc(&sctx->refs);
311 atomic_inc(&sctx->bios_in_flight);
312 }
313
314 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
315 {
316 atomic_dec(&sctx->bios_in_flight);
317 wake_up(&sctx->list_wait);
318 scrub_put_ctx(sctx);
319 }
320
321 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
322 {
323 while (atomic_read(&fs_info->scrub_pause_req)) {
324 mutex_unlock(&fs_info->scrub_lock);
325 wait_event(fs_info->scrub_pause_wait,
326 atomic_read(&fs_info->scrub_pause_req) == 0);
327 mutex_lock(&fs_info->scrub_lock);
328 }
329 }
330
331 static void scrub_pause_on(struct btrfs_fs_info *fs_info)
332 {
333 atomic_inc(&fs_info->scrubs_paused);
334 wake_up(&fs_info->scrub_pause_wait);
335 }
336
337 static void scrub_pause_off(struct btrfs_fs_info *fs_info)
338 {
339 mutex_lock(&fs_info->scrub_lock);
340 __scrub_blocked_if_needed(fs_info);
341 atomic_dec(&fs_info->scrubs_paused);
342 mutex_unlock(&fs_info->scrub_lock);
343
344 wake_up(&fs_info->scrub_pause_wait);
345 }
346
347 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
348 {
349 scrub_pause_on(fs_info);
350 scrub_pause_off(fs_info);
351 }
352
353 /*
354 * used for workers that require transaction commits (i.e., for the
355 * NOCOW case)
356 */
357 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
358 {
359 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
360
361 atomic_inc(&sctx->refs);
362 /*
363 * increment scrubs_running to prevent cancel requests from
364 * completing as long as a worker is running. we must also
365 * increment scrubs_paused to prevent deadlocking on pause
366 * requests used for transactions commits (as the worker uses a
367 * transaction context). it is safe to regard the worker
368 * as paused for all matters practical. effectively, we only
369 * avoid cancellation requests from completing.
370 */
371 mutex_lock(&fs_info->scrub_lock);
372 atomic_inc(&fs_info->scrubs_running);
373 atomic_inc(&fs_info->scrubs_paused);
374 mutex_unlock(&fs_info->scrub_lock);
375
376 /*
377 * check if @scrubs_running=@scrubs_paused condition
378 * inside wait_event() is not an atomic operation.
379 * which means we may inc/dec @scrub_running/paused
380 * at any time. Let's wake up @scrub_pause_wait as
381 * much as we can to let commit transaction blocked less.
382 */
383 wake_up(&fs_info->scrub_pause_wait);
384
385 atomic_inc(&sctx->workers_pending);
386 }
387
388 /* used for workers that require transaction commits */
389 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
390 {
391 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
392
393 /*
394 * see scrub_pending_trans_workers_inc() why we're pretending
395 * to be paused in the scrub counters
396 */
397 mutex_lock(&fs_info->scrub_lock);
398 atomic_dec(&fs_info->scrubs_running);
399 atomic_dec(&fs_info->scrubs_paused);
400 mutex_unlock(&fs_info->scrub_lock);
401 atomic_dec(&sctx->workers_pending);
402 wake_up(&fs_info->scrub_pause_wait);
403 wake_up(&sctx->list_wait);
404 scrub_put_ctx(sctx);
405 }
406
407 static void scrub_free_csums(struct scrub_ctx *sctx)
408 {
409 while (!list_empty(&sctx->csum_list)) {
410 struct btrfs_ordered_sum *sum;
411 sum = list_first_entry(&sctx->csum_list,
412 struct btrfs_ordered_sum, list);
413 list_del(&sum->list);
414 kfree(sum);
415 }
416 }
417
418 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
419 {
420 int i;
421
422 if (!sctx)
423 return;
424
425 scrub_free_wr_ctx(&sctx->wr_ctx);
426
427 /* this can happen when scrub is cancelled */
428 if (sctx->curr != -1) {
429 struct scrub_bio *sbio = sctx->bios[sctx->curr];
430
431 for (i = 0; i < sbio->page_count; i++) {
432 WARN_ON(!sbio->pagev[i]->page);
433 scrub_block_put(sbio->pagev[i]->sblock);
434 }
435 bio_put(sbio->bio);
436 }
437
438 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
439 struct scrub_bio *sbio = sctx->bios[i];
440
441 if (!sbio)
442 break;
443 kfree(sbio);
444 }
445
446 scrub_free_csums(sctx);
447 kfree(sctx);
448 }
449
450 static void scrub_put_ctx(struct scrub_ctx *sctx)
451 {
452 if (atomic_dec_and_test(&sctx->refs))
453 scrub_free_ctx(sctx);
454 }
455
456 static noinline_for_stack
457 struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
458 {
459 struct scrub_ctx *sctx;
460 int i;
461 struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
462 int ret;
463
464 sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
465 if (!sctx)
466 goto nomem;
467 atomic_set(&sctx->refs, 1);
468 sctx->is_dev_replace = is_dev_replace;
469 sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
470 sctx->curr = -1;
471 sctx->dev_root = dev->dev_root;
472 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
473 struct scrub_bio *sbio;
474
475 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
476 if (!sbio)
477 goto nomem;
478 sctx->bios[i] = sbio;
479
480 sbio->index = i;
481 sbio->sctx = sctx;
482 sbio->page_count = 0;
483 btrfs_init_work(&sbio->work, btrfs_scrub_helper,
484 scrub_bio_end_io_worker, NULL, NULL);
485
486 if (i != SCRUB_BIOS_PER_SCTX - 1)
487 sctx->bios[i]->next_free = i + 1;
488 else
489 sctx->bios[i]->next_free = -1;
490 }
491 sctx->first_free = 0;
492 sctx->nodesize = dev->dev_root->nodesize;
493 sctx->sectorsize = dev->dev_root->sectorsize;
494 atomic_set(&sctx->bios_in_flight, 0);
495 atomic_set(&sctx->workers_pending, 0);
496 atomic_set(&sctx->cancel_req, 0);
497 sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
498 INIT_LIST_HEAD(&sctx->csum_list);
499
500 spin_lock_init(&sctx->list_lock);
501 spin_lock_init(&sctx->stat_lock);
502 init_waitqueue_head(&sctx->list_wait);
503
504 ret = scrub_setup_wr_ctx(sctx, &sctx->wr_ctx, fs_info,
505 fs_info->dev_replace.tgtdev, is_dev_replace);
506 if (ret) {
507 scrub_free_ctx(sctx);
508 return ERR_PTR(ret);
509 }
510 return sctx;
511
512 nomem:
513 scrub_free_ctx(sctx);
514 return ERR_PTR(-ENOMEM);
515 }
516
517 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
518 void *warn_ctx)
519 {
520 u64 isize;
521 u32 nlink;
522 int ret;
523 int i;
524 struct extent_buffer *eb;
525 struct btrfs_inode_item *inode_item;
526 struct scrub_warning *swarn = warn_ctx;
527 struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
528 struct inode_fs_paths *ipath = NULL;
529 struct btrfs_root *local_root;
530 struct btrfs_key root_key;
531 struct btrfs_key key;
532
533 root_key.objectid = root;
534 root_key.type = BTRFS_ROOT_ITEM_KEY;
535 root_key.offset = (u64)-1;
536 local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
537 if (IS_ERR(local_root)) {
538 ret = PTR_ERR(local_root);
539 goto err;
540 }
541
542 /*
543 * this makes the path point to (inum INODE_ITEM ioff)
544 */
545 key.objectid = inum;
546 key.type = BTRFS_INODE_ITEM_KEY;
547 key.offset = 0;
548
549 ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
550 if (ret) {
551 btrfs_release_path(swarn->path);
552 goto err;
553 }
554
555 eb = swarn->path->nodes[0];
556 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
557 struct btrfs_inode_item);
558 isize = btrfs_inode_size(eb, inode_item);
559 nlink = btrfs_inode_nlink(eb, inode_item);
560 btrfs_release_path(swarn->path);
561
562 ipath = init_ipath(4096, local_root, swarn->path);
563 if (IS_ERR(ipath)) {
564 ret = PTR_ERR(ipath);
565 ipath = NULL;
566 goto err;
567 }
568 ret = paths_from_inode(inum, ipath);
569
570 if (ret < 0)
571 goto err;
572
573 /*
574 * we deliberately ignore the bit ipath might have been too small to
575 * hold all of the paths here
576 */
577 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
578 btrfs_warn_in_rcu(fs_info, "%s at logical %llu on dev "
579 "%s, sector %llu, root %llu, inode %llu, offset %llu, "
580 "length %llu, links %u (path: %s)", swarn->errstr,
581 swarn->logical, rcu_str_deref(swarn->dev->name),
582 (unsigned long long)swarn->sector, root, inum, offset,
583 min(isize - offset, (u64)PAGE_SIZE), nlink,
584 (char *)(unsigned long)ipath->fspath->val[i]);
585
586 free_ipath(ipath);
587 return 0;
588
589 err:
590 btrfs_warn_in_rcu(fs_info, "%s at logical %llu on dev "
591 "%s, sector %llu, root %llu, inode %llu, offset %llu: path "
592 "resolving failed with ret=%d", swarn->errstr,
593 swarn->logical, rcu_str_deref(swarn->dev->name),
594 (unsigned long long)swarn->sector, root, inum, offset, ret);
595
596 free_ipath(ipath);
597 return 0;
598 }
599
600 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
601 {
602 struct btrfs_device *dev;
603 struct btrfs_fs_info *fs_info;
604 struct btrfs_path *path;
605 struct btrfs_key found_key;
606 struct extent_buffer *eb;
607 struct btrfs_extent_item *ei;
608 struct scrub_warning swarn;
609 unsigned long ptr = 0;
610 u64 extent_item_pos;
611 u64 flags = 0;
612 u64 ref_root;
613 u32 item_size;
614 u8 ref_level = 0;
615 int ret;
616
617 WARN_ON(sblock->page_count < 1);
618 dev = sblock->pagev[0]->dev;
619 fs_info = sblock->sctx->dev_root->fs_info;
620
621 path = btrfs_alloc_path();
622 if (!path)
623 return;
624
625 swarn.sector = (sblock->pagev[0]->physical) >> 9;
626 swarn.logical = sblock->pagev[0]->logical;
627 swarn.errstr = errstr;
628 swarn.dev = NULL;
629
630 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
631 &flags);
632 if (ret < 0)
633 goto out;
634
635 extent_item_pos = swarn.logical - found_key.objectid;
636 swarn.extent_item_size = found_key.offset;
637
638 eb = path->nodes[0];
639 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
640 item_size = btrfs_item_size_nr(eb, path->slots[0]);
641
642 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
643 do {
644 ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
645 item_size, &ref_root,
646 &ref_level);
647 btrfs_warn_in_rcu(fs_info,
648 "%s at logical %llu on dev %s, "
649 "sector %llu: metadata %s (level %d) in tree "
650 "%llu", errstr, swarn.logical,
651 rcu_str_deref(dev->name),
652 (unsigned long long)swarn.sector,
653 ref_level ? "node" : "leaf",
654 ret < 0 ? -1 : ref_level,
655 ret < 0 ? -1 : ref_root);
656 } while (ret != 1);
657 btrfs_release_path(path);
658 } else {
659 btrfs_release_path(path);
660 swarn.path = path;
661 swarn.dev = dev;
662 iterate_extent_inodes(fs_info, found_key.objectid,
663 extent_item_pos, 1,
664 scrub_print_warning_inode, &swarn);
665 }
666
667 out:
668 btrfs_free_path(path);
669 }
670
671 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
672 {
673 struct page *page = NULL;
674 unsigned long index;
675 struct scrub_fixup_nodatasum *fixup = fixup_ctx;
676 int ret;
677 int corrected = 0;
678 struct btrfs_key key;
679 struct inode *inode = NULL;
680 struct btrfs_fs_info *fs_info;
681 u64 end = offset + PAGE_SIZE - 1;
682 struct btrfs_root *local_root;
683 int srcu_index;
684
685 key.objectid = root;
686 key.type = BTRFS_ROOT_ITEM_KEY;
687 key.offset = (u64)-1;
688
689 fs_info = fixup->root->fs_info;
690 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
691
692 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
693 if (IS_ERR(local_root)) {
694 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
695 return PTR_ERR(local_root);
696 }
697
698 key.type = BTRFS_INODE_ITEM_KEY;
699 key.objectid = inum;
700 key.offset = 0;
701 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
702 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
703 if (IS_ERR(inode))
704 return PTR_ERR(inode);
705
706 index = offset >> PAGE_SHIFT;
707
708 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
709 if (!page) {
710 ret = -ENOMEM;
711 goto out;
712 }
713
714 if (PageUptodate(page)) {
715 if (PageDirty(page)) {
716 /*
717 * we need to write the data to the defect sector. the
718 * data that was in that sector is not in memory,
719 * because the page was modified. we must not write the
720 * modified page to that sector.
721 *
722 * TODO: what could be done here: wait for the delalloc
723 * runner to write out that page (might involve
724 * COW) and see whether the sector is still
725 * referenced afterwards.
726 *
727 * For the meantime, we'll treat this error
728 * incorrectable, although there is a chance that a
729 * later scrub will find the bad sector again and that
730 * there's no dirty page in memory, then.
731 */
732 ret = -EIO;
733 goto out;
734 }
735 ret = repair_io_failure(inode, offset, PAGE_SIZE,
736 fixup->logical, page,
737 offset - page_offset(page),
738 fixup->mirror_num);
739 unlock_page(page);
740 corrected = !ret;
741 } else {
742 /*
743 * we need to get good data first. the general readpage path
744 * will call repair_io_failure for us, we just have to make
745 * sure we read the bad mirror.
746 */
747 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
748 EXTENT_DAMAGED);
749 if (ret) {
750 /* set_extent_bits should give proper error */
751 WARN_ON(ret > 0);
752 if (ret > 0)
753 ret = -EFAULT;
754 goto out;
755 }
756
757 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
758 btrfs_get_extent,
759 fixup->mirror_num);
760 wait_on_page_locked(page);
761
762 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
763 end, EXTENT_DAMAGED, 0, NULL);
764 if (!corrected)
765 clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
766 EXTENT_DAMAGED);
767 }
768
769 out:
770 if (page)
771 put_page(page);
772
773 iput(inode);
774
775 if (ret < 0)
776 return ret;
777
778 if (ret == 0 && corrected) {
779 /*
780 * we only need to call readpage for one of the inodes belonging
781 * to this extent. so make iterate_extent_inodes stop
782 */
783 return 1;
784 }
785
786 return -EIO;
787 }
788
789 static void scrub_fixup_nodatasum(struct btrfs_work *work)
790 {
791 int ret;
792 struct scrub_fixup_nodatasum *fixup;
793 struct scrub_ctx *sctx;
794 struct btrfs_trans_handle *trans = NULL;
795 struct btrfs_path *path;
796 int uncorrectable = 0;
797
798 fixup = container_of(work, struct scrub_fixup_nodatasum, work);
799 sctx = fixup->sctx;
800
801 path = btrfs_alloc_path();
802 if (!path) {
803 spin_lock(&sctx->stat_lock);
804 ++sctx->stat.malloc_errors;
805 spin_unlock(&sctx->stat_lock);
806 uncorrectable = 1;
807 goto out;
808 }
809
810 trans = btrfs_join_transaction(fixup->root);
811 if (IS_ERR(trans)) {
812 uncorrectable = 1;
813 goto out;
814 }
815
816 /*
817 * the idea is to trigger a regular read through the standard path. we
818 * read a page from the (failed) logical address by specifying the
819 * corresponding copynum of the failed sector. thus, that readpage is
820 * expected to fail.
821 * that is the point where on-the-fly error correction will kick in
822 * (once it's finished) and rewrite the failed sector if a good copy
823 * can be found.
824 */
825 ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
826 path, scrub_fixup_readpage,
827 fixup);
828 if (ret < 0) {
829 uncorrectable = 1;
830 goto out;
831 }
832 WARN_ON(ret != 1);
833
834 spin_lock(&sctx->stat_lock);
835 ++sctx->stat.corrected_errors;
836 spin_unlock(&sctx->stat_lock);
837
838 out:
839 if (trans && !IS_ERR(trans))
840 btrfs_end_transaction(trans, fixup->root);
841 if (uncorrectable) {
842 spin_lock(&sctx->stat_lock);
843 ++sctx->stat.uncorrectable_errors;
844 spin_unlock(&sctx->stat_lock);
845 btrfs_dev_replace_stats_inc(
846 &sctx->dev_root->fs_info->dev_replace.
847 num_uncorrectable_read_errors);
848 btrfs_err_rl_in_rcu(sctx->dev_root->fs_info,
849 "unable to fixup (nodatasum) error at logical %llu on dev %s",
850 fixup->logical, rcu_str_deref(fixup->dev->name));
851 }
852
853 btrfs_free_path(path);
854 kfree(fixup);
855
856 scrub_pending_trans_workers_dec(sctx);
857 }
858
859 static inline void scrub_get_recover(struct scrub_recover *recover)
860 {
861 atomic_inc(&recover->refs);
862 }
863
864 static inline void scrub_put_recover(struct scrub_recover *recover)
865 {
866 if (atomic_dec_and_test(&recover->refs)) {
867 btrfs_put_bbio(recover->bbio);
868 kfree(recover);
869 }
870 }
871
872 /*
873 * scrub_handle_errored_block gets called when either verification of the
874 * pages failed or the bio failed to read, e.g. with EIO. In the latter
875 * case, this function handles all pages in the bio, even though only one
876 * may be bad.
877 * The goal of this function is to repair the errored block by using the
878 * contents of one of the mirrors.
879 */
880 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
881 {
882 struct scrub_ctx *sctx = sblock_to_check->sctx;
883 struct btrfs_device *dev;
884 struct btrfs_fs_info *fs_info;
885 u64 length;
886 u64 logical;
887 unsigned int failed_mirror_index;
888 unsigned int is_metadata;
889 unsigned int have_csum;
890 struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
891 struct scrub_block *sblock_bad;
892 int ret;
893 int mirror_index;
894 int page_num;
895 int success;
896 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
897 DEFAULT_RATELIMIT_BURST);
898
899 BUG_ON(sblock_to_check->page_count < 1);
900 fs_info = sctx->dev_root->fs_info;
901 if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
902 /*
903 * if we find an error in a super block, we just report it.
904 * They will get written with the next transaction commit
905 * anyway
906 */
907 spin_lock(&sctx->stat_lock);
908 ++sctx->stat.super_errors;
909 spin_unlock(&sctx->stat_lock);
910 return 0;
911 }
912 length = sblock_to_check->page_count * PAGE_SIZE;
913 logical = sblock_to_check->pagev[0]->logical;
914 BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
915 failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
916 is_metadata = !(sblock_to_check->pagev[0]->flags &
917 BTRFS_EXTENT_FLAG_DATA);
918 have_csum = sblock_to_check->pagev[0]->have_csum;
919 dev = sblock_to_check->pagev[0]->dev;
920
921 if (sctx->is_dev_replace && !is_metadata && !have_csum) {
922 sblocks_for_recheck = NULL;
923 goto nodatasum_case;
924 }
925
926 /*
927 * read all mirrors one after the other. This includes to
928 * re-read the extent or metadata block that failed (that was
929 * the cause that this fixup code is called) another time,
930 * page by page this time in order to know which pages
931 * caused I/O errors and which ones are good (for all mirrors).
932 * It is the goal to handle the situation when more than one
933 * mirror contains I/O errors, but the errors do not
934 * overlap, i.e. the data can be repaired by selecting the
935 * pages from those mirrors without I/O error on the
936 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
937 * would be that mirror #1 has an I/O error on the first page,
938 * the second page is good, and mirror #2 has an I/O error on
939 * the second page, but the first page is good.
940 * Then the first page of the first mirror can be repaired by
941 * taking the first page of the second mirror, and the
942 * second page of the second mirror can be repaired by
943 * copying the contents of the 2nd page of the 1st mirror.
944 * One more note: if the pages of one mirror contain I/O
945 * errors, the checksum cannot be verified. In order to get
946 * the best data for repairing, the first attempt is to find
947 * a mirror without I/O errors and with a validated checksum.
948 * Only if this is not possible, the pages are picked from
949 * mirrors with I/O errors without considering the checksum.
950 * If the latter is the case, at the end, the checksum of the
951 * repaired area is verified in order to correctly maintain
952 * the statistics.
953 */
954
955 sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
956 sizeof(*sblocks_for_recheck), GFP_NOFS);
957 if (!sblocks_for_recheck) {
958 spin_lock(&sctx->stat_lock);
959 sctx->stat.malloc_errors++;
960 sctx->stat.read_errors++;
961 sctx->stat.uncorrectable_errors++;
962 spin_unlock(&sctx->stat_lock);
963 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
964 goto out;
965 }
966
967 /* setup the context, map the logical blocks and alloc the pages */
968 ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
969 if (ret) {
970 spin_lock(&sctx->stat_lock);
971 sctx->stat.read_errors++;
972 sctx->stat.uncorrectable_errors++;
973 spin_unlock(&sctx->stat_lock);
974 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
975 goto out;
976 }
977 BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
978 sblock_bad = sblocks_for_recheck + failed_mirror_index;
979
980 /* build and submit the bios for the failed mirror, check checksums */
981 scrub_recheck_block(fs_info, sblock_bad, 1);
982
983 if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
984 sblock_bad->no_io_error_seen) {
985 /*
986 * the error disappeared after reading page by page, or
987 * the area was part of a huge bio and other parts of the
988 * bio caused I/O errors, or the block layer merged several
989 * read requests into one and the error is caused by a
990 * different bio (usually one of the two latter cases is
991 * the cause)
992 */
993 spin_lock(&sctx->stat_lock);
994 sctx->stat.unverified_errors++;
995 sblock_to_check->data_corrected = 1;
996 spin_unlock(&sctx->stat_lock);
997
998 if (sctx->is_dev_replace)
999 scrub_write_block_to_dev_replace(sblock_bad);
1000 goto out;
1001 }
1002
1003 if (!sblock_bad->no_io_error_seen) {
1004 spin_lock(&sctx->stat_lock);
1005 sctx->stat.read_errors++;
1006 spin_unlock(&sctx->stat_lock);
1007 if (__ratelimit(&_rs))
1008 scrub_print_warning("i/o error", sblock_to_check);
1009 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1010 } else if (sblock_bad->checksum_error) {
1011 spin_lock(&sctx->stat_lock);
1012 sctx->stat.csum_errors++;
1013 spin_unlock(&sctx->stat_lock);
1014 if (__ratelimit(&_rs))
1015 scrub_print_warning("checksum error", sblock_to_check);
1016 btrfs_dev_stat_inc_and_print(dev,
1017 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1018 } else if (sblock_bad->header_error) {
1019 spin_lock(&sctx->stat_lock);
1020 sctx->stat.verify_errors++;
1021 spin_unlock(&sctx->stat_lock);
1022 if (__ratelimit(&_rs))
1023 scrub_print_warning("checksum/header error",
1024 sblock_to_check);
1025 if (sblock_bad->generation_error)
1026 btrfs_dev_stat_inc_and_print(dev,
1027 BTRFS_DEV_STAT_GENERATION_ERRS);
1028 else
1029 btrfs_dev_stat_inc_and_print(dev,
1030 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1031 }
1032
1033 if (sctx->readonly) {
1034 ASSERT(!sctx->is_dev_replace);
1035 goto out;
1036 }
1037
1038 if (!is_metadata && !have_csum) {
1039 struct scrub_fixup_nodatasum *fixup_nodatasum;
1040
1041 WARN_ON(sctx->is_dev_replace);
1042
1043 nodatasum_case:
1044
1045 /*
1046 * !is_metadata and !have_csum, this means that the data
1047 * might not be COWed, that it might be modified
1048 * concurrently. The general strategy to work on the
1049 * commit root does not help in the case when COW is not
1050 * used.
1051 */
1052 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
1053 if (!fixup_nodatasum)
1054 goto did_not_correct_error;
1055 fixup_nodatasum->sctx = sctx;
1056 fixup_nodatasum->dev = dev;
1057 fixup_nodatasum->logical = logical;
1058 fixup_nodatasum->root = fs_info->extent_root;
1059 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
1060 scrub_pending_trans_workers_inc(sctx);
1061 btrfs_init_work(&fixup_nodatasum->work, btrfs_scrub_helper,
1062 scrub_fixup_nodatasum, NULL, NULL);
1063 btrfs_queue_work(fs_info->scrub_workers,
1064 &fixup_nodatasum->work);
1065 goto out;
1066 }
1067
1068 /*
1069 * now build and submit the bios for the other mirrors, check
1070 * checksums.
1071 * First try to pick the mirror which is completely without I/O
1072 * errors and also does not have a checksum error.
1073 * If one is found, and if a checksum is present, the full block
1074 * that is known to contain an error is rewritten. Afterwards
1075 * the block is known to be corrected.
1076 * If a mirror is found which is completely correct, and no
1077 * checksum is present, only those pages are rewritten that had
1078 * an I/O error in the block to be repaired, since it cannot be
1079 * determined, which copy of the other pages is better (and it
1080 * could happen otherwise that a correct page would be
1081 * overwritten by a bad one).
1082 */
1083 for (mirror_index = 0;
1084 mirror_index < BTRFS_MAX_MIRRORS &&
1085 sblocks_for_recheck[mirror_index].page_count > 0;
1086 mirror_index++) {
1087 struct scrub_block *sblock_other;
1088
1089 if (mirror_index == failed_mirror_index)
1090 continue;
1091 sblock_other = sblocks_for_recheck + mirror_index;
1092
1093 /* build and submit the bios, check checksums */
1094 scrub_recheck_block(fs_info, sblock_other, 0);
1095
1096 if (!sblock_other->header_error &&
1097 !sblock_other->checksum_error &&
1098 sblock_other->no_io_error_seen) {
1099 if (sctx->is_dev_replace) {
1100 scrub_write_block_to_dev_replace(sblock_other);
1101 goto corrected_error;
1102 } else {
1103 ret = scrub_repair_block_from_good_copy(
1104 sblock_bad, sblock_other);
1105 if (!ret)
1106 goto corrected_error;
1107 }
1108 }
1109 }
1110
1111 if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1112 goto did_not_correct_error;
1113
1114 /*
1115 * In case of I/O errors in the area that is supposed to be
1116 * repaired, continue by picking good copies of those pages.
1117 * Select the good pages from mirrors to rewrite bad pages from
1118 * the area to fix. Afterwards verify the checksum of the block
1119 * that is supposed to be repaired. This verification step is
1120 * only done for the purpose of statistic counting and for the
1121 * final scrub report, whether errors remain.
1122 * A perfect algorithm could make use of the checksum and try
1123 * all possible combinations of pages from the different mirrors
1124 * until the checksum verification succeeds. For example, when
1125 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1126 * of mirror #2 is readable but the final checksum test fails,
1127 * then the 2nd page of mirror #3 could be tried, whether now
1128 * the final checksum succeeds. But this would be a rare
1129 * exception and is therefore not implemented. At least it is
1130 * avoided that the good copy is overwritten.
1131 * A more useful improvement would be to pick the sectors
1132 * without I/O error based on sector sizes (512 bytes on legacy
1133 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1134 * mirror could be repaired by taking 512 byte of a different
1135 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1136 * area are unreadable.
1137 */
1138 success = 1;
1139 for (page_num = 0; page_num < sblock_bad->page_count;
1140 page_num++) {
1141 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1142 struct scrub_block *sblock_other = NULL;
1143
1144 /* skip no-io-error page in scrub */
1145 if (!page_bad->io_error && !sctx->is_dev_replace)
1146 continue;
1147
1148 /* try to find no-io-error page in mirrors */
1149 if (page_bad->io_error) {
1150 for (mirror_index = 0;
1151 mirror_index < BTRFS_MAX_MIRRORS &&
1152 sblocks_for_recheck[mirror_index].page_count > 0;
1153 mirror_index++) {
1154 if (!sblocks_for_recheck[mirror_index].
1155 pagev[page_num]->io_error) {
1156 sblock_other = sblocks_for_recheck +
1157 mirror_index;
1158 break;
1159 }
1160 }
1161 if (!sblock_other)
1162 success = 0;
1163 }
1164
1165 if (sctx->is_dev_replace) {
1166 /*
1167 * did not find a mirror to fetch the page
1168 * from. scrub_write_page_to_dev_replace()
1169 * handles this case (page->io_error), by
1170 * filling the block with zeros before
1171 * submitting the write request
1172 */
1173 if (!sblock_other)
1174 sblock_other = sblock_bad;
1175
1176 if (scrub_write_page_to_dev_replace(sblock_other,
1177 page_num) != 0) {
1178 btrfs_dev_replace_stats_inc(
1179 &sctx->dev_root->
1180 fs_info->dev_replace.
1181 num_write_errors);
1182 success = 0;
1183 }
1184 } else if (sblock_other) {
1185 ret = scrub_repair_page_from_good_copy(sblock_bad,
1186 sblock_other,
1187 page_num, 0);
1188 if (0 == ret)
1189 page_bad->io_error = 0;
1190 else
1191 success = 0;
1192 }
1193 }
1194
1195 if (success && !sctx->is_dev_replace) {
1196 if (is_metadata || have_csum) {
1197 /*
1198 * need to verify the checksum now that all
1199 * sectors on disk are repaired (the write
1200 * request for data to be repaired is on its way).
1201 * Just be lazy and use scrub_recheck_block()
1202 * which re-reads the data before the checksum
1203 * is verified, but most likely the data comes out
1204 * of the page cache.
1205 */
1206 scrub_recheck_block(fs_info, sblock_bad, 1);
1207 if (!sblock_bad->header_error &&
1208 !sblock_bad->checksum_error &&
1209 sblock_bad->no_io_error_seen)
1210 goto corrected_error;
1211 else
1212 goto did_not_correct_error;
1213 } else {
1214 corrected_error:
1215 spin_lock(&sctx->stat_lock);
1216 sctx->stat.corrected_errors++;
1217 sblock_to_check->data_corrected = 1;
1218 spin_unlock(&sctx->stat_lock);
1219 btrfs_err_rl_in_rcu(fs_info,
1220 "fixed up error at logical %llu on dev %s",
1221 logical, rcu_str_deref(dev->name));
1222 }
1223 } else {
1224 did_not_correct_error:
1225 spin_lock(&sctx->stat_lock);
1226 sctx->stat.uncorrectable_errors++;
1227 spin_unlock(&sctx->stat_lock);
1228 btrfs_err_rl_in_rcu(fs_info,
1229 "unable to fixup (regular) error at logical %llu on dev %s",
1230 logical, rcu_str_deref(dev->name));
1231 }
1232
1233 out:
1234 if (sblocks_for_recheck) {
1235 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1236 mirror_index++) {
1237 struct scrub_block *sblock = sblocks_for_recheck +
1238 mirror_index;
1239 struct scrub_recover *recover;
1240 int page_index;
1241
1242 for (page_index = 0; page_index < sblock->page_count;
1243 page_index++) {
1244 sblock->pagev[page_index]->sblock = NULL;
1245 recover = sblock->pagev[page_index]->recover;
1246 if (recover) {
1247 scrub_put_recover(recover);
1248 sblock->pagev[page_index]->recover =
1249 NULL;
1250 }
1251 scrub_page_put(sblock->pagev[page_index]);
1252 }
1253 }
1254 kfree(sblocks_for_recheck);
1255 }
1256
1257 return 0;
1258 }
1259
1260 static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1261 {
1262 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1263 return 2;
1264 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1265 return 3;
1266 else
1267 return (int)bbio->num_stripes;
1268 }
1269
1270 static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1271 u64 *raid_map,
1272 u64 mapped_length,
1273 int nstripes, int mirror,
1274 int *stripe_index,
1275 u64 *stripe_offset)
1276 {
1277 int i;
1278
1279 if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1280 /* RAID5/6 */
1281 for (i = 0; i < nstripes; i++) {
1282 if (raid_map[i] == RAID6_Q_STRIPE ||
1283 raid_map[i] == RAID5_P_STRIPE)
1284 continue;
1285
1286 if (logical >= raid_map[i] &&
1287 logical < raid_map[i] + mapped_length)
1288 break;
1289 }
1290
1291 *stripe_index = i;
1292 *stripe_offset = logical - raid_map[i];
1293 } else {
1294 /* The other RAID type */
1295 *stripe_index = mirror;
1296 *stripe_offset = 0;
1297 }
1298 }
1299
1300 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1301 struct scrub_block *sblocks_for_recheck)
1302 {
1303 struct scrub_ctx *sctx = original_sblock->sctx;
1304 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
1305 u64 length = original_sblock->page_count * PAGE_SIZE;
1306 u64 logical = original_sblock->pagev[0]->logical;
1307 u64 generation = original_sblock->pagev[0]->generation;
1308 u64 flags = original_sblock->pagev[0]->flags;
1309 u64 have_csum = original_sblock->pagev[0]->have_csum;
1310 struct scrub_recover *recover;
1311 struct btrfs_bio *bbio;
1312 u64 sublen;
1313 u64 mapped_length;
1314 u64 stripe_offset;
1315 int stripe_index;
1316 int page_index = 0;
1317 int mirror_index;
1318 int nmirrors;
1319 int ret;
1320
1321 /*
1322 * note: the two members refs and outstanding_pages
1323 * are not used (and not set) in the blocks that are used for
1324 * the recheck procedure
1325 */
1326
1327 while (length > 0) {
1328 sublen = min_t(u64, length, PAGE_SIZE);
1329 mapped_length = sublen;
1330 bbio = NULL;
1331
1332 /*
1333 * with a length of PAGE_SIZE, each returned stripe
1334 * represents one mirror
1335 */
1336 ret = btrfs_map_sblock(fs_info, REQ_GET_READ_MIRRORS, logical,
1337 &mapped_length, &bbio, 0, 1);
1338 if (ret || !bbio || mapped_length < sublen) {
1339 btrfs_put_bbio(bbio);
1340 return -EIO;
1341 }
1342
1343 recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1344 if (!recover) {
1345 btrfs_put_bbio(bbio);
1346 return -ENOMEM;
1347 }
1348
1349 atomic_set(&recover->refs, 1);
1350 recover->bbio = bbio;
1351 recover->map_length = mapped_length;
1352
1353 BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1354
1355 nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1356
1357 for (mirror_index = 0; mirror_index < nmirrors;
1358 mirror_index++) {
1359 struct scrub_block *sblock;
1360 struct scrub_page *page;
1361
1362 sblock = sblocks_for_recheck + mirror_index;
1363 sblock->sctx = sctx;
1364
1365 page = kzalloc(sizeof(*page), GFP_NOFS);
1366 if (!page) {
1367 leave_nomem:
1368 spin_lock(&sctx->stat_lock);
1369 sctx->stat.malloc_errors++;
1370 spin_unlock(&sctx->stat_lock);
1371 scrub_put_recover(recover);
1372 return -ENOMEM;
1373 }
1374 scrub_page_get(page);
1375 sblock->pagev[page_index] = page;
1376 page->sblock = sblock;
1377 page->flags = flags;
1378 page->generation = generation;
1379 page->logical = logical;
1380 page->have_csum = have_csum;
1381 if (have_csum)
1382 memcpy(page->csum,
1383 original_sblock->pagev[0]->csum,
1384 sctx->csum_size);
1385
1386 scrub_stripe_index_and_offset(logical,
1387 bbio->map_type,
1388 bbio->raid_map,
1389 mapped_length,
1390 bbio->num_stripes -
1391 bbio->num_tgtdevs,
1392 mirror_index,
1393 &stripe_index,
1394 &stripe_offset);
1395 page->physical = bbio->stripes[stripe_index].physical +
1396 stripe_offset;
1397 page->dev = bbio->stripes[stripe_index].dev;
1398
1399 BUG_ON(page_index >= original_sblock->page_count);
1400 page->physical_for_dev_replace =
1401 original_sblock->pagev[page_index]->
1402 physical_for_dev_replace;
1403 /* for missing devices, dev->bdev is NULL */
1404 page->mirror_num = mirror_index + 1;
1405 sblock->page_count++;
1406 page->page = alloc_page(GFP_NOFS);
1407 if (!page->page)
1408 goto leave_nomem;
1409
1410 scrub_get_recover(recover);
1411 page->recover = recover;
1412 }
1413 scrub_put_recover(recover);
1414 length -= sublen;
1415 logical += sublen;
1416 page_index++;
1417 }
1418
1419 return 0;
1420 }
1421
1422 struct scrub_bio_ret {
1423 struct completion event;
1424 int error;
1425 };
1426
1427 static void scrub_bio_wait_endio(struct bio *bio)
1428 {
1429 struct scrub_bio_ret *ret = bio->bi_private;
1430
1431 ret->error = bio->bi_error;
1432 complete(&ret->event);
1433 }
1434
1435 static inline int scrub_is_page_on_raid56(struct scrub_page *page)
1436 {
1437 return page->recover &&
1438 (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
1439 }
1440
1441 static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1442 struct bio *bio,
1443 struct scrub_page *page)
1444 {
1445 struct scrub_bio_ret done;
1446 int ret;
1447
1448 init_completion(&done.event);
1449 done.error = 0;
1450 bio->bi_iter.bi_sector = page->logical >> 9;
1451 bio->bi_private = &done;
1452 bio->bi_end_io = scrub_bio_wait_endio;
1453
1454 ret = raid56_parity_recover(fs_info->fs_root, bio, page->recover->bbio,
1455 page->recover->map_length,
1456 page->mirror_num, 0);
1457 if (ret)
1458 return ret;
1459
1460 wait_for_completion(&done.event);
1461 if (done.error)
1462 return -EIO;
1463
1464 return 0;
1465 }
1466
1467 /*
1468 * this function will check the on disk data for checksum errors, header
1469 * errors and read I/O errors. If any I/O errors happen, the exact pages
1470 * which are errored are marked as being bad. The goal is to enable scrub
1471 * to take those pages that are not errored from all the mirrors so that
1472 * the pages that are errored in the just handled mirror can be repaired.
1473 */
1474 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1475 struct scrub_block *sblock,
1476 int retry_failed_mirror)
1477 {
1478 int page_num;
1479
1480 sblock->no_io_error_seen = 1;
1481
1482 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1483 struct bio *bio;
1484 struct scrub_page *page = sblock->pagev[page_num];
1485
1486 if (page->dev->bdev == NULL) {
1487 page->io_error = 1;
1488 sblock->no_io_error_seen = 0;
1489 continue;
1490 }
1491
1492 WARN_ON(!page->page);
1493 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1494 if (!bio) {
1495 page->io_error = 1;
1496 sblock->no_io_error_seen = 0;
1497 continue;
1498 }
1499 bio->bi_bdev = page->dev->bdev;
1500
1501 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1502 if (!retry_failed_mirror && scrub_is_page_on_raid56(page)) {
1503 if (scrub_submit_raid56_bio_wait(fs_info, bio, page))
1504 sblock->no_io_error_seen = 0;
1505 } else {
1506 bio->bi_iter.bi_sector = page->physical >> 9;
1507
1508 if (btrfsic_submit_bio_wait(READ, bio))
1509 sblock->no_io_error_seen = 0;
1510 }
1511
1512 bio_put(bio);
1513 }
1514
1515 if (sblock->no_io_error_seen)
1516 scrub_recheck_block_checksum(sblock);
1517 }
1518
1519 static inline int scrub_check_fsid(u8 fsid[],
1520 struct scrub_page *spage)
1521 {
1522 struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1523 int ret;
1524
1525 ret = memcmp(fsid, fs_devices->fsid, BTRFS_UUID_SIZE);
1526 return !ret;
1527 }
1528
1529 static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1530 {
1531 sblock->header_error = 0;
1532 sblock->checksum_error = 0;
1533 sblock->generation_error = 0;
1534
1535 if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1536 scrub_checksum_data(sblock);
1537 else
1538 scrub_checksum_tree_block(sblock);
1539 }
1540
1541 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1542 struct scrub_block *sblock_good)
1543 {
1544 int page_num;
1545 int ret = 0;
1546
1547 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1548 int ret_sub;
1549
1550 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1551 sblock_good,
1552 page_num, 1);
1553 if (ret_sub)
1554 ret = ret_sub;
1555 }
1556
1557 return ret;
1558 }
1559
1560 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1561 struct scrub_block *sblock_good,
1562 int page_num, int force_write)
1563 {
1564 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1565 struct scrub_page *page_good = sblock_good->pagev[page_num];
1566
1567 BUG_ON(page_bad->page == NULL);
1568 BUG_ON(page_good->page == NULL);
1569 if (force_write || sblock_bad->header_error ||
1570 sblock_bad->checksum_error || page_bad->io_error) {
1571 struct bio *bio;
1572 int ret;
1573
1574 if (!page_bad->dev->bdev) {
1575 btrfs_warn_rl(sblock_bad->sctx->dev_root->fs_info,
1576 "scrub_repair_page_from_good_copy(bdev == NULL) "
1577 "is unexpected");
1578 return -EIO;
1579 }
1580
1581 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1582 if (!bio)
1583 return -EIO;
1584 bio->bi_bdev = page_bad->dev->bdev;
1585 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1586
1587 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1588 if (PAGE_SIZE != ret) {
1589 bio_put(bio);
1590 return -EIO;
1591 }
1592
1593 if (btrfsic_submit_bio_wait(WRITE, bio)) {
1594 btrfs_dev_stat_inc_and_print(page_bad->dev,
1595 BTRFS_DEV_STAT_WRITE_ERRS);
1596 btrfs_dev_replace_stats_inc(
1597 &sblock_bad->sctx->dev_root->fs_info->
1598 dev_replace.num_write_errors);
1599 bio_put(bio);
1600 return -EIO;
1601 }
1602 bio_put(bio);
1603 }
1604
1605 return 0;
1606 }
1607
1608 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1609 {
1610 int page_num;
1611
1612 /*
1613 * This block is used for the check of the parity on the source device,
1614 * so the data needn't be written into the destination device.
1615 */
1616 if (sblock->sparity)
1617 return;
1618
1619 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1620 int ret;
1621
1622 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1623 if (ret)
1624 btrfs_dev_replace_stats_inc(
1625 &sblock->sctx->dev_root->fs_info->dev_replace.
1626 num_write_errors);
1627 }
1628 }
1629
1630 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1631 int page_num)
1632 {
1633 struct scrub_page *spage = sblock->pagev[page_num];
1634
1635 BUG_ON(spage->page == NULL);
1636 if (spage->io_error) {
1637 void *mapped_buffer = kmap_atomic(spage->page);
1638
1639 memset(mapped_buffer, 0, PAGE_SIZE);
1640 flush_dcache_page(spage->page);
1641 kunmap_atomic(mapped_buffer);
1642 }
1643 return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1644 }
1645
1646 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1647 struct scrub_page *spage)
1648 {
1649 struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1650 struct scrub_bio *sbio;
1651 int ret;
1652
1653 mutex_lock(&wr_ctx->wr_lock);
1654 again:
1655 if (!wr_ctx->wr_curr_bio) {
1656 wr_ctx->wr_curr_bio = kzalloc(sizeof(*wr_ctx->wr_curr_bio),
1657 GFP_KERNEL);
1658 if (!wr_ctx->wr_curr_bio) {
1659 mutex_unlock(&wr_ctx->wr_lock);
1660 return -ENOMEM;
1661 }
1662 wr_ctx->wr_curr_bio->sctx = sctx;
1663 wr_ctx->wr_curr_bio->page_count = 0;
1664 }
1665 sbio = wr_ctx->wr_curr_bio;
1666 if (sbio->page_count == 0) {
1667 struct bio *bio;
1668
1669 sbio->physical = spage->physical_for_dev_replace;
1670 sbio->logical = spage->logical;
1671 sbio->dev = wr_ctx->tgtdev;
1672 bio = sbio->bio;
1673 if (!bio) {
1674 bio = btrfs_io_bio_alloc(GFP_KERNEL,
1675 wr_ctx->pages_per_wr_bio);
1676 if (!bio) {
1677 mutex_unlock(&wr_ctx->wr_lock);
1678 return -ENOMEM;
1679 }
1680 sbio->bio = bio;
1681 }
1682
1683 bio->bi_private = sbio;
1684 bio->bi_end_io = scrub_wr_bio_end_io;
1685 bio->bi_bdev = sbio->dev->bdev;
1686 bio->bi_iter.bi_sector = sbio->physical >> 9;
1687 sbio->err = 0;
1688 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1689 spage->physical_for_dev_replace ||
1690 sbio->logical + sbio->page_count * PAGE_SIZE !=
1691 spage->logical) {
1692 scrub_wr_submit(sctx);
1693 goto again;
1694 }
1695
1696 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1697 if (ret != PAGE_SIZE) {
1698 if (sbio->page_count < 1) {
1699 bio_put(sbio->bio);
1700 sbio->bio = NULL;
1701 mutex_unlock(&wr_ctx->wr_lock);
1702 return -EIO;
1703 }
1704 scrub_wr_submit(sctx);
1705 goto again;
1706 }
1707
1708 sbio->pagev[sbio->page_count] = spage;
1709 scrub_page_get(spage);
1710 sbio->page_count++;
1711 if (sbio->page_count == wr_ctx->pages_per_wr_bio)
1712 scrub_wr_submit(sctx);
1713 mutex_unlock(&wr_ctx->wr_lock);
1714
1715 return 0;
1716 }
1717
1718 static void scrub_wr_submit(struct scrub_ctx *sctx)
1719 {
1720 struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1721 struct scrub_bio *sbio;
1722
1723 if (!wr_ctx->wr_curr_bio)
1724 return;
1725
1726 sbio = wr_ctx->wr_curr_bio;
1727 wr_ctx->wr_curr_bio = NULL;
1728 WARN_ON(!sbio->bio->bi_bdev);
1729 scrub_pending_bio_inc(sctx);
1730 /* process all writes in a single worker thread. Then the block layer
1731 * orders the requests before sending them to the driver which
1732 * doubled the write performance on spinning disks when measured
1733 * with Linux 3.5 */
1734 btrfsic_submit_bio(WRITE, sbio->bio);
1735 }
1736
1737 static void scrub_wr_bio_end_io(struct bio *bio)
1738 {
1739 struct scrub_bio *sbio = bio->bi_private;
1740 struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
1741
1742 sbio->err = bio->bi_error;
1743 sbio->bio = bio;
1744
1745 btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
1746 scrub_wr_bio_end_io_worker, NULL, NULL);
1747 btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1748 }
1749
1750 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1751 {
1752 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1753 struct scrub_ctx *sctx = sbio->sctx;
1754 int i;
1755
1756 WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1757 if (sbio->err) {
1758 struct btrfs_dev_replace *dev_replace =
1759 &sbio->sctx->dev_root->fs_info->dev_replace;
1760
1761 for (i = 0; i < sbio->page_count; i++) {
1762 struct scrub_page *spage = sbio->pagev[i];
1763
1764 spage->io_error = 1;
1765 btrfs_dev_replace_stats_inc(&dev_replace->
1766 num_write_errors);
1767 }
1768 }
1769
1770 for (i = 0; i < sbio->page_count; i++)
1771 scrub_page_put(sbio->pagev[i]);
1772
1773 bio_put(sbio->bio);
1774 kfree(sbio);
1775 scrub_pending_bio_dec(sctx);
1776 }
1777
1778 static int scrub_checksum(struct scrub_block *sblock)
1779 {
1780 u64 flags;
1781 int ret;
1782
1783 /*
1784 * No need to initialize these stats currently,
1785 * because this function only use return value
1786 * instead of these stats value.
1787 *
1788 * Todo:
1789 * always use stats
1790 */
1791 sblock->header_error = 0;
1792 sblock->generation_error = 0;
1793 sblock->checksum_error = 0;
1794
1795 WARN_ON(sblock->page_count < 1);
1796 flags = sblock->pagev[0]->flags;
1797 ret = 0;
1798 if (flags & BTRFS_EXTENT_FLAG_DATA)
1799 ret = scrub_checksum_data(sblock);
1800 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1801 ret = scrub_checksum_tree_block(sblock);
1802 else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1803 (void)scrub_checksum_super(sblock);
1804 else
1805 WARN_ON(1);
1806 if (ret)
1807 scrub_handle_errored_block(sblock);
1808
1809 return ret;
1810 }
1811
1812 static int scrub_checksum_data(struct scrub_block *sblock)
1813 {
1814 struct scrub_ctx *sctx = sblock->sctx;
1815 u8 csum[BTRFS_CSUM_SIZE];
1816 u8 *on_disk_csum;
1817 struct page *page;
1818 void *buffer;
1819 u32 crc = ~(u32)0;
1820 u64 len;
1821 int index;
1822
1823 BUG_ON(sblock->page_count < 1);
1824 if (!sblock->pagev[0]->have_csum)
1825 return 0;
1826
1827 on_disk_csum = sblock->pagev[0]->csum;
1828 page = sblock->pagev[0]->page;
1829 buffer = kmap_atomic(page);
1830
1831 len = sctx->sectorsize;
1832 index = 0;
1833 for (;;) {
1834 u64 l = min_t(u64, len, PAGE_SIZE);
1835
1836 crc = btrfs_csum_data(buffer, crc, l);
1837 kunmap_atomic(buffer);
1838 len -= l;
1839 if (len == 0)
1840 break;
1841 index++;
1842 BUG_ON(index >= sblock->page_count);
1843 BUG_ON(!sblock->pagev[index]->page);
1844 page = sblock->pagev[index]->page;
1845 buffer = kmap_atomic(page);
1846 }
1847
1848 btrfs_csum_final(crc, csum);
1849 if (memcmp(csum, on_disk_csum, sctx->csum_size))
1850 sblock->checksum_error = 1;
1851
1852 return sblock->checksum_error;
1853 }
1854
1855 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1856 {
1857 struct scrub_ctx *sctx = sblock->sctx;
1858 struct btrfs_header *h;
1859 struct btrfs_root *root = sctx->dev_root;
1860 struct btrfs_fs_info *fs_info = root->fs_info;
1861 u8 calculated_csum[BTRFS_CSUM_SIZE];
1862 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1863 struct page *page;
1864 void *mapped_buffer;
1865 u64 mapped_size;
1866 void *p;
1867 u32 crc = ~(u32)0;
1868 u64 len;
1869 int index;
1870
1871 BUG_ON(sblock->page_count < 1);
1872 page = sblock->pagev[0]->page;
1873 mapped_buffer = kmap_atomic(page);
1874 h = (struct btrfs_header *)mapped_buffer;
1875 memcpy(on_disk_csum, h->csum, sctx->csum_size);
1876
1877 /*
1878 * we don't use the getter functions here, as we
1879 * a) don't have an extent buffer and
1880 * b) the page is already kmapped
1881 */
1882 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1883 sblock->header_error = 1;
1884
1885 if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
1886 sblock->header_error = 1;
1887 sblock->generation_error = 1;
1888 }
1889
1890 if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
1891 sblock->header_error = 1;
1892
1893 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1894 BTRFS_UUID_SIZE))
1895 sblock->header_error = 1;
1896
1897 len = sctx->nodesize - BTRFS_CSUM_SIZE;
1898 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1899 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1900 index = 0;
1901 for (;;) {
1902 u64 l = min_t(u64, len, mapped_size);
1903
1904 crc = btrfs_csum_data(p, crc, l);
1905 kunmap_atomic(mapped_buffer);
1906 len -= l;
1907 if (len == 0)
1908 break;
1909 index++;
1910 BUG_ON(index >= sblock->page_count);
1911 BUG_ON(!sblock->pagev[index]->page);
1912 page = sblock->pagev[index]->page;
1913 mapped_buffer = kmap_atomic(page);
1914 mapped_size = PAGE_SIZE;
1915 p = mapped_buffer;
1916 }
1917
1918 btrfs_csum_final(crc, calculated_csum);
1919 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1920 sblock->checksum_error = 1;
1921
1922 return sblock->header_error || sblock->checksum_error;
1923 }
1924
1925 static int scrub_checksum_super(struct scrub_block *sblock)
1926 {
1927 struct btrfs_super_block *s;
1928 struct scrub_ctx *sctx = sblock->sctx;
1929 u8 calculated_csum[BTRFS_CSUM_SIZE];
1930 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1931 struct page *page;
1932 void *mapped_buffer;
1933 u64 mapped_size;
1934 void *p;
1935 u32 crc = ~(u32)0;
1936 int fail_gen = 0;
1937 int fail_cor = 0;
1938 u64 len;
1939 int index;
1940
1941 BUG_ON(sblock->page_count < 1);
1942 page = sblock->pagev[0]->page;
1943 mapped_buffer = kmap_atomic(page);
1944 s = (struct btrfs_super_block *)mapped_buffer;
1945 memcpy(on_disk_csum, s->csum, sctx->csum_size);
1946
1947 if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1948 ++fail_cor;
1949
1950 if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1951 ++fail_gen;
1952
1953 if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
1954 ++fail_cor;
1955
1956 len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1957 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1958 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1959 index = 0;
1960 for (;;) {
1961 u64 l = min_t(u64, len, mapped_size);
1962
1963 crc = btrfs_csum_data(p, crc, l);
1964 kunmap_atomic(mapped_buffer);
1965 len -= l;
1966 if (len == 0)
1967 break;
1968 index++;
1969 BUG_ON(index >= sblock->page_count);
1970 BUG_ON(!sblock->pagev[index]->page);
1971 page = sblock->pagev[index]->page;
1972 mapped_buffer = kmap_atomic(page);
1973 mapped_size = PAGE_SIZE;
1974 p = mapped_buffer;
1975 }
1976
1977 btrfs_csum_final(crc, calculated_csum);
1978 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1979 ++fail_cor;
1980
1981 if (fail_cor + fail_gen) {
1982 /*
1983 * if we find an error in a super block, we just report it.
1984 * They will get written with the next transaction commit
1985 * anyway
1986 */
1987 spin_lock(&sctx->stat_lock);
1988 ++sctx->stat.super_errors;
1989 spin_unlock(&sctx->stat_lock);
1990 if (fail_cor)
1991 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1992 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1993 else
1994 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1995 BTRFS_DEV_STAT_GENERATION_ERRS);
1996 }
1997
1998 return fail_cor + fail_gen;
1999 }
2000
2001 static void scrub_block_get(struct scrub_block *sblock)
2002 {
2003 atomic_inc(&sblock->refs);
2004 }
2005
2006 static void scrub_block_put(struct scrub_block *sblock)
2007 {
2008 if (atomic_dec_and_test(&sblock->refs)) {
2009 int i;
2010
2011 if (sblock->sparity)
2012 scrub_parity_put(sblock->sparity);
2013
2014 for (i = 0; i < sblock->page_count; i++)
2015 scrub_page_put(sblock->pagev[i]);
2016 kfree(sblock);
2017 }
2018 }
2019
2020 static void scrub_page_get(struct scrub_page *spage)
2021 {
2022 atomic_inc(&spage->refs);
2023 }
2024
2025 static void scrub_page_put(struct scrub_page *spage)
2026 {
2027 if (atomic_dec_and_test(&spage->refs)) {
2028 if (spage->page)
2029 __free_page(spage->page);
2030 kfree(spage);
2031 }
2032 }
2033
2034 static void scrub_submit(struct scrub_ctx *sctx)
2035 {
2036 struct scrub_bio *sbio;
2037
2038 if (sctx->curr == -1)
2039 return;
2040
2041 sbio = sctx->bios[sctx->curr];
2042 sctx->curr = -1;
2043 scrub_pending_bio_inc(sctx);
2044 btrfsic_submit_bio(READ, sbio->bio);
2045 }
2046
2047 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2048 struct scrub_page *spage)
2049 {
2050 struct scrub_block *sblock = spage->sblock;
2051 struct scrub_bio *sbio;
2052 int ret;
2053
2054 again:
2055 /*
2056 * grab a fresh bio or wait for one to become available
2057 */
2058 while (sctx->curr == -1) {
2059 spin_lock(&sctx->list_lock);
2060 sctx->curr = sctx->first_free;
2061 if (sctx->curr != -1) {
2062 sctx->first_free = sctx->bios[sctx->curr]->next_free;
2063 sctx->bios[sctx->curr]->next_free = -1;
2064 sctx->bios[sctx->curr]->page_count = 0;
2065 spin_unlock(&sctx->list_lock);
2066 } else {
2067 spin_unlock(&sctx->list_lock);
2068 wait_event(sctx->list_wait, sctx->first_free != -1);
2069 }
2070 }
2071 sbio = sctx->bios[sctx->curr];
2072 if (sbio->page_count == 0) {
2073 struct bio *bio;
2074
2075 sbio->physical = spage->physical;
2076 sbio->logical = spage->logical;
2077 sbio->dev = spage->dev;
2078 bio = sbio->bio;
2079 if (!bio) {
2080 bio = btrfs_io_bio_alloc(GFP_KERNEL,
2081 sctx->pages_per_rd_bio);
2082 if (!bio)
2083 return -ENOMEM;
2084 sbio->bio = bio;
2085 }
2086
2087 bio->bi_private = sbio;
2088 bio->bi_end_io = scrub_bio_end_io;
2089 bio->bi_bdev = sbio->dev->bdev;
2090 bio->bi_iter.bi_sector = sbio->physical >> 9;
2091 sbio->err = 0;
2092 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2093 spage->physical ||
2094 sbio->logical + sbio->page_count * PAGE_SIZE !=
2095 spage->logical ||
2096 sbio->dev != spage->dev) {
2097 scrub_submit(sctx);
2098 goto again;
2099 }
2100
2101 sbio->pagev[sbio->page_count] = spage;
2102 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2103 if (ret != PAGE_SIZE) {
2104 if (sbio->page_count < 1) {
2105 bio_put(sbio->bio);
2106 sbio->bio = NULL;
2107 return -EIO;
2108 }
2109 scrub_submit(sctx);
2110 goto again;
2111 }
2112
2113 scrub_block_get(sblock); /* one for the page added to the bio */
2114 atomic_inc(&sblock->outstanding_pages);
2115 sbio->page_count++;
2116 if (sbio->page_count == sctx->pages_per_rd_bio)
2117 scrub_submit(sctx);
2118
2119 return 0;
2120 }
2121
2122 static void scrub_missing_raid56_end_io(struct bio *bio)
2123 {
2124 struct scrub_block *sblock = bio->bi_private;
2125 struct btrfs_fs_info *fs_info = sblock->sctx->dev_root->fs_info;
2126
2127 if (bio->bi_error)
2128 sblock->no_io_error_seen = 0;
2129
2130 bio_put(bio);
2131
2132 btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2133 }
2134
2135 static void scrub_missing_raid56_worker(struct btrfs_work *work)
2136 {
2137 struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2138 struct scrub_ctx *sctx = sblock->sctx;
2139 u64 logical;
2140 struct btrfs_device *dev;
2141
2142 logical = sblock->pagev[0]->logical;
2143 dev = sblock->pagev[0]->dev;
2144
2145 if (sblock->no_io_error_seen)
2146 scrub_recheck_block_checksum(sblock);
2147
2148 if (!sblock->no_io_error_seen) {
2149 spin_lock(&sctx->stat_lock);
2150 sctx->stat.read_errors++;
2151 spin_unlock(&sctx->stat_lock);
2152 btrfs_err_rl_in_rcu(sctx->dev_root->fs_info,
2153 "IO error rebuilding logical %llu for dev %s",
2154 logical, rcu_str_deref(dev->name));
2155 } else if (sblock->header_error || sblock->checksum_error) {
2156 spin_lock(&sctx->stat_lock);
2157 sctx->stat.uncorrectable_errors++;
2158 spin_unlock(&sctx->stat_lock);
2159 btrfs_err_rl_in_rcu(sctx->dev_root->fs_info,
2160 "failed to rebuild valid logical %llu for dev %s",
2161 logical, rcu_str_deref(dev->name));
2162 } else {
2163 scrub_write_block_to_dev_replace(sblock);
2164 }
2165
2166 scrub_block_put(sblock);
2167
2168 if (sctx->is_dev_replace &&
2169 atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2170 mutex_lock(&sctx->wr_ctx.wr_lock);
2171 scrub_wr_submit(sctx);
2172 mutex_unlock(&sctx->wr_ctx.wr_lock);
2173 }
2174
2175 scrub_pending_bio_dec(sctx);
2176 }
2177
2178 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2179 {
2180 struct scrub_ctx *sctx = sblock->sctx;
2181 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2182 u64 length = sblock->page_count * PAGE_SIZE;
2183 u64 logical = sblock->pagev[0]->logical;
2184 struct btrfs_bio *bbio = NULL;
2185 struct bio *bio;
2186 struct btrfs_raid_bio *rbio;
2187 int ret;
2188 int i;
2189
2190 ret = btrfs_map_sblock(fs_info, REQ_GET_READ_MIRRORS, logical, &length,
2191 &bbio, 0, 1);
2192 if (ret || !bbio || !bbio->raid_map)
2193 goto bbio_out;
2194
2195 if (WARN_ON(!sctx->is_dev_replace ||
2196 !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2197 /*
2198 * We shouldn't be scrubbing a missing device. Even for dev
2199 * replace, we should only get here for RAID 5/6. We either
2200 * managed to mount something with no mirrors remaining or
2201 * there's a bug in scrub_remap_extent()/btrfs_map_block().
2202 */
2203 goto bbio_out;
2204 }
2205
2206 bio = btrfs_io_bio_alloc(GFP_NOFS, 0);
2207 if (!bio)
2208 goto bbio_out;
2209
2210 bio->bi_iter.bi_sector = logical >> 9;
2211 bio->bi_private = sblock;
2212 bio->bi_end_io = scrub_missing_raid56_end_io;
2213
2214 rbio = raid56_alloc_missing_rbio(sctx->dev_root, bio, bbio, length);
2215 if (!rbio)
2216 goto rbio_out;
2217
2218 for (i = 0; i < sblock->page_count; i++) {
2219 struct scrub_page *spage = sblock->pagev[i];
2220
2221 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2222 }
2223
2224 btrfs_init_work(&sblock->work, btrfs_scrub_helper,
2225 scrub_missing_raid56_worker, NULL, NULL);
2226 scrub_block_get(sblock);
2227 scrub_pending_bio_inc(sctx);
2228 raid56_submit_missing_rbio(rbio);
2229 return;
2230
2231 rbio_out:
2232 bio_put(bio);
2233 bbio_out:
2234 btrfs_put_bbio(bbio);
2235 spin_lock(&sctx->stat_lock);
2236 sctx->stat.malloc_errors++;
2237 spin_unlock(&sctx->stat_lock);
2238 }
2239
2240 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2241 u64 physical, struct btrfs_device *dev, u64 flags,
2242 u64 gen, int mirror_num, u8 *csum, int force,
2243 u64 physical_for_dev_replace)
2244 {
2245 struct scrub_block *sblock;
2246 int index;
2247
2248 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2249 if (!sblock) {
2250 spin_lock(&sctx->stat_lock);
2251 sctx->stat.malloc_errors++;
2252 spin_unlock(&sctx->stat_lock);
2253 return -ENOMEM;
2254 }
2255
2256 /* one ref inside this function, plus one for each page added to
2257 * a bio later on */
2258 atomic_set(&sblock->refs, 1);
2259 sblock->sctx = sctx;
2260 sblock->no_io_error_seen = 1;
2261
2262 for (index = 0; len > 0; index++) {
2263 struct scrub_page *spage;
2264 u64 l = min_t(u64, len, PAGE_SIZE);
2265
2266 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2267 if (!spage) {
2268 leave_nomem:
2269 spin_lock(&sctx->stat_lock);
2270 sctx->stat.malloc_errors++;
2271 spin_unlock(&sctx->stat_lock);
2272 scrub_block_put(sblock);
2273 return -ENOMEM;
2274 }
2275 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2276 scrub_page_get(spage);
2277 sblock->pagev[index] = spage;
2278 spage->sblock = sblock;
2279 spage->dev = dev;
2280 spage->flags = flags;
2281 spage->generation = gen;
2282 spage->logical = logical;
2283 spage->physical = physical;
2284 spage->physical_for_dev_replace = physical_for_dev_replace;
2285 spage->mirror_num = mirror_num;
2286 if (csum) {
2287 spage->have_csum = 1;
2288 memcpy(spage->csum, csum, sctx->csum_size);
2289 } else {
2290 spage->have_csum = 0;
2291 }
2292 sblock->page_count++;
2293 spage->page = alloc_page(GFP_KERNEL);
2294 if (!spage->page)
2295 goto leave_nomem;
2296 len -= l;
2297 logical += l;
2298 physical += l;
2299 physical_for_dev_replace += l;
2300 }
2301
2302 WARN_ON(sblock->page_count == 0);
2303 if (dev->missing) {
2304 /*
2305 * This case should only be hit for RAID 5/6 device replace. See
2306 * the comment in scrub_missing_raid56_pages() for details.
2307 */
2308 scrub_missing_raid56_pages(sblock);
2309 } else {
2310 for (index = 0; index < sblock->page_count; index++) {
2311 struct scrub_page *spage = sblock->pagev[index];
2312 int ret;
2313
2314 ret = scrub_add_page_to_rd_bio(sctx, spage);
2315 if (ret) {
2316 scrub_block_put(sblock);
2317 return ret;
2318 }
2319 }
2320
2321 if (force)
2322 scrub_submit(sctx);
2323 }
2324
2325 /* last one frees, either here or in bio completion for last page */
2326 scrub_block_put(sblock);
2327 return 0;
2328 }
2329
2330 static void scrub_bio_end_io(struct bio *bio)
2331 {
2332 struct scrub_bio *sbio = bio->bi_private;
2333 struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
2334
2335 sbio->err = bio->bi_error;
2336 sbio->bio = bio;
2337
2338 btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2339 }
2340
2341 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2342 {
2343 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2344 struct scrub_ctx *sctx = sbio->sctx;
2345 int i;
2346
2347 BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2348 if (sbio->err) {
2349 for (i = 0; i < sbio->page_count; i++) {
2350 struct scrub_page *spage = sbio->pagev[i];
2351
2352 spage->io_error = 1;
2353 spage->sblock->no_io_error_seen = 0;
2354 }
2355 }
2356
2357 /* now complete the scrub_block items that have all pages completed */
2358 for (i = 0; i < sbio->page_count; i++) {
2359 struct scrub_page *spage = sbio->pagev[i];
2360 struct scrub_block *sblock = spage->sblock;
2361
2362 if (atomic_dec_and_test(&sblock->outstanding_pages))
2363 scrub_block_complete(sblock);
2364 scrub_block_put(sblock);
2365 }
2366
2367 bio_put(sbio->bio);
2368 sbio->bio = NULL;
2369 spin_lock(&sctx->list_lock);
2370 sbio->next_free = sctx->first_free;
2371 sctx->first_free = sbio->index;
2372 spin_unlock(&sctx->list_lock);
2373
2374 if (sctx->is_dev_replace &&
2375 atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2376 mutex_lock(&sctx->wr_ctx.wr_lock);
2377 scrub_wr_submit(sctx);
2378 mutex_unlock(&sctx->wr_ctx.wr_lock);
2379 }
2380
2381 scrub_pending_bio_dec(sctx);
2382 }
2383
2384 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2385 unsigned long *bitmap,
2386 u64 start, u64 len)
2387 {
2388 u32 offset;
2389 int nsectors;
2390 int sectorsize = sparity->sctx->dev_root->sectorsize;
2391
2392 if (len >= sparity->stripe_len) {
2393 bitmap_set(bitmap, 0, sparity->nsectors);
2394 return;
2395 }
2396
2397 start -= sparity->logic_start;
2398 start = div_u64_rem(start, sparity->stripe_len, &offset);
2399 offset /= sectorsize;
2400 nsectors = (int)len / sectorsize;
2401
2402 if (offset + nsectors <= sparity->nsectors) {
2403 bitmap_set(bitmap, offset, nsectors);
2404 return;
2405 }
2406
2407 bitmap_set(bitmap, offset, sparity->nsectors - offset);
2408 bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2409 }
2410
2411 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2412 u64 start, u64 len)
2413 {
2414 __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2415 }
2416
2417 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2418 u64 start, u64 len)
2419 {
2420 __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2421 }
2422
2423 static void scrub_block_complete(struct scrub_block *sblock)
2424 {
2425 int corrupted = 0;
2426
2427 if (!sblock->no_io_error_seen) {
2428 corrupted = 1;
2429 scrub_handle_errored_block(sblock);
2430 } else {
2431 /*
2432 * if has checksum error, write via repair mechanism in
2433 * dev replace case, otherwise write here in dev replace
2434 * case.
2435 */
2436 corrupted = scrub_checksum(sblock);
2437 if (!corrupted && sblock->sctx->is_dev_replace)
2438 scrub_write_block_to_dev_replace(sblock);
2439 }
2440
2441 if (sblock->sparity && corrupted && !sblock->data_corrected) {
2442 u64 start = sblock->pagev[0]->logical;
2443 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2444 PAGE_SIZE;
2445
2446 scrub_parity_mark_sectors_error(sblock->sparity,
2447 start, end - start);
2448 }
2449 }
2450
2451 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2452 {
2453 struct btrfs_ordered_sum *sum = NULL;
2454 unsigned long index;
2455 unsigned long num_sectors;
2456
2457 while (!list_empty(&sctx->csum_list)) {
2458 sum = list_first_entry(&sctx->csum_list,
2459 struct btrfs_ordered_sum, list);
2460 if (sum->bytenr > logical)
2461 return 0;
2462 if (sum->bytenr + sum->len > logical)
2463 break;
2464
2465 ++sctx->stat.csum_discards;
2466 list_del(&sum->list);
2467 kfree(sum);
2468 sum = NULL;
2469 }
2470 if (!sum)
2471 return 0;
2472
2473 index = ((u32)(logical - sum->bytenr)) / sctx->sectorsize;
2474 num_sectors = sum->len / sctx->sectorsize;
2475 memcpy(csum, sum->sums + index, sctx->csum_size);
2476 if (index == num_sectors - 1) {
2477 list_del(&sum->list);
2478 kfree(sum);
2479 }
2480 return 1;
2481 }
2482
2483 /* scrub extent tries to collect up to 64 kB for each bio */
2484 static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2485 u64 physical, struct btrfs_device *dev, u64 flags,
2486 u64 gen, int mirror_num, u64 physical_for_dev_replace)
2487 {
2488 int ret;
2489 u8 csum[BTRFS_CSUM_SIZE];
2490 u32 blocksize;
2491
2492 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2493 blocksize = sctx->sectorsize;
2494 spin_lock(&sctx->stat_lock);
2495 sctx->stat.data_extents_scrubbed++;
2496 sctx->stat.data_bytes_scrubbed += len;
2497 spin_unlock(&sctx->stat_lock);
2498 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2499 blocksize = sctx->nodesize;
2500 spin_lock(&sctx->stat_lock);
2501 sctx->stat.tree_extents_scrubbed++;
2502 sctx->stat.tree_bytes_scrubbed += len;
2503 spin_unlock(&sctx->stat_lock);
2504 } else {
2505 blocksize = sctx->sectorsize;
2506 WARN_ON(1);
2507 }
2508
2509 while (len) {
2510 u64 l = min_t(u64, len, blocksize);
2511 int have_csum = 0;
2512
2513 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2514 /* push csums to sbio */
2515 have_csum = scrub_find_csum(sctx, logical, csum);
2516 if (have_csum == 0)
2517 ++sctx->stat.no_csum;
2518 if (sctx->is_dev_replace && !have_csum) {
2519 ret = copy_nocow_pages(sctx, logical, l,
2520 mirror_num,
2521 physical_for_dev_replace);
2522 goto behind_scrub_pages;
2523 }
2524 }
2525 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2526 mirror_num, have_csum ? csum : NULL, 0,
2527 physical_for_dev_replace);
2528 behind_scrub_pages:
2529 if (ret)
2530 return ret;
2531 len -= l;
2532 logical += l;
2533 physical += l;
2534 physical_for_dev_replace += l;
2535 }
2536 return 0;
2537 }
2538
2539 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2540 u64 logical, u64 len,
2541 u64 physical, struct btrfs_device *dev,
2542 u64 flags, u64 gen, int mirror_num, u8 *csum)
2543 {
2544 struct scrub_ctx *sctx = sparity->sctx;
2545 struct scrub_block *sblock;
2546 int index;
2547
2548 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2549 if (!sblock) {
2550 spin_lock(&sctx->stat_lock);
2551 sctx->stat.malloc_errors++;
2552 spin_unlock(&sctx->stat_lock);
2553 return -ENOMEM;
2554 }
2555
2556 /* one ref inside this function, plus one for each page added to
2557 * a bio later on */
2558 atomic_set(&sblock->refs, 1);
2559 sblock->sctx = sctx;
2560 sblock->no_io_error_seen = 1;
2561 sblock->sparity = sparity;
2562 scrub_parity_get(sparity);
2563
2564 for (index = 0; len > 0; index++) {
2565 struct scrub_page *spage;
2566 u64 l = min_t(u64, len, PAGE_SIZE);
2567
2568 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2569 if (!spage) {
2570 leave_nomem:
2571 spin_lock(&sctx->stat_lock);
2572 sctx->stat.malloc_errors++;
2573 spin_unlock(&sctx->stat_lock);
2574 scrub_block_put(sblock);
2575 return -ENOMEM;
2576 }
2577 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2578 /* For scrub block */
2579 scrub_page_get(spage);
2580 sblock->pagev[index] = spage;
2581 /* For scrub parity */
2582 scrub_page_get(spage);
2583 list_add_tail(&spage->list, &sparity->spages);
2584 spage->sblock = sblock;
2585 spage->dev = dev;
2586 spage->flags = flags;
2587 spage->generation = gen;
2588 spage->logical = logical;
2589 spage->physical = physical;
2590 spage->mirror_num = mirror_num;
2591 if (csum) {
2592 spage->have_csum = 1;
2593 memcpy(spage->csum, csum, sctx->csum_size);
2594 } else {
2595 spage->have_csum = 0;
2596 }
2597 sblock->page_count++;
2598 spage->page = alloc_page(GFP_KERNEL);
2599 if (!spage->page)
2600 goto leave_nomem;
2601 len -= l;
2602 logical += l;
2603 physical += l;
2604 }
2605
2606 WARN_ON(sblock->page_count == 0);
2607 for (index = 0; index < sblock->page_count; index++) {
2608 struct scrub_page *spage = sblock->pagev[index];
2609 int ret;
2610
2611 ret = scrub_add_page_to_rd_bio(sctx, spage);
2612 if (ret) {
2613 scrub_block_put(sblock);
2614 return ret;
2615 }
2616 }
2617
2618 /* last one frees, either here or in bio completion for last page */
2619 scrub_block_put(sblock);
2620 return 0;
2621 }
2622
2623 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2624 u64 logical, u64 len,
2625 u64 physical, struct btrfs_device *dev,
2626 u64 flags, u64 gen, int mirror_num)
2627 {
2628 struct scrub_ctx *sctx = sparity->sctx;
2629 int ret;
2630 u8 csum[BTRFS_CSUM_SIZE];
2631 u32 blocksize;
2632
2633 if (dev->missing) {
2634 scrub_parity_mark_sectors_error(sparity, logical, len);
2635 return 0;
2636 }
2637
2638 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2639 blocksize = sctx->sectorsize;
2640 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2641 blocksize = sctx->nodesize;
2642 } else {
2643 blocksize = sctx->sectorsize;
2644 WARN_ON(1);
2645 }
2646
2647 while (len) {
2648 u64 l = min_t(u64, len, blocksize);
2649 int have_csum = 0;
2650
2651 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2652 /* push csums to sbio */
2653 have_csum = scrub_find_csum(sctx, logical, csum);
2654 if (have_csum == 0)
2655 goto skip;
2656 }
2657 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2658 flags, gen, mirror_num,
2659 have_csum ? csum : NULL);
2660 if (ret)
2661 return ret;
2662 skip:
2663 len -= l;
2664 logical += l;
2665 physical += l;
2666 }
2667 return 0;
2668 }
2669
2670 /*
2671 * Given a physical address, this will calculate it's
2672 * logical offset. if this is a parity stripe, it will return
2673 * the most left data stripe's logical offset.
2674 *
2675 * return 0 if it is a data stripe, 1 means parity stripe.
2676 */
2677 static int get_raid56_logic_offset(u64 physical, int num,
2678 struct map_lookup *map, u64 *offset,
2679 u64 *stripe_start)
2680 {
2681 int i;
2682 int j = 0;
2683 u64 stripe_nr;
2684 u64 last_offset;
2685 u32 stripe_index;
2686 u32 rot;
2687
2688 last_offset = (physical - map->stripes[num].physical) *
2689 nr_data_stripes(map);
2690 if (stripe_start)
2691 *stripe_start = last_offset;
2692
2693 *offset = last_offset;
2694 for (i = 0; i < nr_data_stripes(map); i++) {
2695 *offset = last_offset + i * map->stripe_len;
2696
2697 stripe_nr = div_u64(*offset, map->stripe_len);
2698 stripe_nr = div_u64(stripe_nr, nr_data_stripes(map));
2699
2700 /* Work out the disk rotation on this stripe-set */
2701 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2702 /* calculate which stripe this data locates */
2703 rot += i;
2704 stripe_index = rot % map->num_stripes;
2705 if (stripe_index == num)
2706 return 0;
2707 if (stripe_index < num)
2708 j++;
2709 }
2710 *offset = last_offset + j * map->stripe_len;
2711 return 1;
2712 }
2713
2714 static void scrub_free_parity(struct scrub_parity *sparity)
2715 {
2716 struct scrub_ctx *sctx = sparity->sctx;
2717 struct scrub_page *curr, *next;
2718 int nbits;
2719
2720 nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2721 if (nbits) {
2722 spin_lock(&sctx->stat_lock);
2723 sctx->stat.read_errors += nbits;
2724 sctx->stat.uncorrectable_errors += nbits;
2725 spin_unlock(&sctx->stat_lock);
2726 }
2727
2728 list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2729 list_del_init(&curr->list);
2730 scrub_page_put(curr);
2731 }
2732
2733 kfree(sparity);
2734 }
2735
2736 static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2737 {
2738 struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2739 work);
2740 struct scrub_ctx *sctx = sparity->sctx;
2741
2742 scrub_free_parity(sparity);
2743 scrub_pending_bio_dec(sctx);
2744 }
2745
2746 static void scrub_parity_bio_endio(struct bio *bio)
2747 {
2748 struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
2749
2750 if (bio->bi_error)
2751 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2752 sparity->nsectors);
2753
2754 bio_put(bio);
2755
2756 btrfs_init_work(&sparity->work, btrfs_scrubparity_helper,
2757 scrub_parity_bio_endio_worker, NULL, NULL);
2758 btrfs_queue_work(sparity->sctx->dev_root->fs_info->scrub_parity_workers,
2759 &sparity->work);
2760 }
2761
2762 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2763 {
2764 struct scrub_ctx *sctx = sparity->sctx;
2765 struct bio *bio;
2766 struct btrfs_raid_bio *rbio;
2767 struct scrub_page *spage;
2768 struct btrfs_bio *bbio = NULL;
2769 u64 length;
2770 int ret;
2771
2772 if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
2773 sparity->nsectors))
2774 goto out;
2775
2776 length = sparity->logic_end - sparity->logic_start;
2777 ret = btrfs_map_sblock(sctx->dev_root->fs_info, WRITE,
2778 sparity->logic_start,
2779 &length, &bbio, 0, 1);
2780 if (ret || !bbio || !bbio->raid_map)
2781 goto bbio_out;
2782
2783 bio = btrfs_io_bio_alloc(GFP_NOFS, 0);
2784 if (!bio)
2785 goto bbio_out;
2786
2787 bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2788 bio->bi_private = sparity;
2789 bio->bi_end_io = scrub_parity_bio_endio;
2790
2791 rbio = raid56_parity_alloc_scrub_rbio(sctx->dev_root, bio, bbio,
2792 length, sparity->scrub_dev,
2793 sparity->dbitmap,
2794 sparity->nsectors);
2795 if (!rbio)
2796 goto rbio_out;
2797
2798 list_for_each_entry(spage, &sparity->spages, list)
2799 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2800
2801 scrub_pending_bio_inc(sctx);
2802 raid56_parity_submit_scrub_rbio(rbio);
2803 return;
2804
2805 rbio_out:
2806 bio_put(bio);
2807 bbio_out:
2808 btrfs_put_bbio(bbio);
2809 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2810 sparity->nsectors);
2811 spin_lock(&sctx->stat_lock);
2812 sctx->stat.malloc_errors++;
2813 spin_unlock(&sctx->stat_lock);
2814 out:
2815 scrub_free_parity(sparity);
2816 }
2817
2818 static inline int scrub_calc_parity_bitmap_len(int nsectors)
2819 {
2820 return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
2821 }
2822
2823 static void scrub_parity_get(struct scrub_parity *sparity)
2824 {
2825 atomic_inc(&sparity->refs);
2826 }
2827
2828 static void scrub_parity_put(struct scrub_parity *sparity)
2829 {
2830 if (!atomic_dec_and_test(&sparity->refs))
2831 return;
2832
2833 scrub_parity_check_and_repair(sparity);
2834 }
2835
2836 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
2837 struct map_lookup *map,
2838 struct btrfs_device *sdev,
2839 struct btrfs_path *path,
2840 u64 logic_start,
2841 u64 logic_end)
2842 {
2843 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2844 struct btrfs_root *root = fs_info->extent_root;
2845 struct btrfs_root *csum_root = fs_info->csum_root;
2846 struct btrfs_extent_item *extent;
2847 struct btrfs_bio *bbio = NULL;
2848 u64 flags;
2849 int ret;
2850 int slot;
2851 struct extent_buffer *l;
2852 struct btrfs_key key;
2853 u64 generation;
2854 u64 extent_logical;
2855 u64 extent_physical;
2856 u64 extent_len;
2857 u64 mapped_length;
2858 struct btrfs_device *extent_dev;
2859 struct scrub_parity *sparity;
2860 int nsectors;
2861 int bitmap_len;
2862 int extent_mirror_num;
2863 int stop_loop = 0;
2864
2865 nsectors = div_u64(map->stripe_len, root->sectorsize);
2866 bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
2867 sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
2868 GFP_NOFS);
2869 if (!sparity) {
2870 spin_lock(&sctx->stat_lock);
2871 sctx->stat.malloc_errors++;
2872 spin_unlock(&sctx->stat_lock);
2873 return -ENOMEM;
2874 }
2875
2876 sparity->stripe_len = map->stripe_len;
2877 sparity->nsectors = nsectors;
2878 sparity->sctx = sctx;
2879 sparity->scrub_dev = sdev;
2880 sparity->logic_start = logic_start;
2881 sparity->logic_end = logic_end;
2882 atomic_set(&sparity->refs, 1);
2883 INIT_LIST_HEAD(&sparity->spages);
2884 sparity->dbitmap = sparity->bitmap;
2885 sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
2886
2887 ret = 0;
2888 while (logic_start < logic_end) {
2889 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2890 key.type = BTRFS_METADATA_ITEM_KEY;
2891 else
2892 key.type = BTRFS_EXTENT_ITEM_KEY;
2893 key.objectid = logic_start;
2894 key.offset = (u64)-1;
2895
2896 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2897 if (ret < 0)
2898 goto out;
2899
2900 if (ret > 0) {
2901 ret = btrfs_previous_extent_item(root, path, 0);
2902 if (ret < 0)
2903 goto out;
2904 if (ret > 0) {
2905 btrfs_release_path(path);
2906 ret = btrfs_search_slot(NULL, root, &key,
2907 path, 0, 0);
2908 if (ret < 0)
2909 goto out;
2910 }
2911 }
2912
2913 stop_loop = 0;
2914 while (1) {
2915 u64 bytes;
2916
2917 l = path->nodes[0];
2918 slot = path->slots[0];
2919 if (slot >= btrfs_header_nritems(l)) {
2920 ret = btrfs_next_leaf(root, path);
2921 if (ret == 0)
2922 continue;
2923 if (ret < 0)
2924 goto out;
2925
2926 stop_loop = 1;
2927 break;
2928 }
2929 btrfs_item_key_to_cpu(l, &key, slot);
2930
2931 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2932 key.type != BTRFS_METADATA_ITEM_KEY)
2933 goto next;
2934
2935 if (key.type == BTRFS_METADATA_ITEM_KEY)
2936 bytes = root->nodesize;
2937 else
2938 bytes = key.offset;
2939
2940 if (key.objectid + bytes <= logic_start)
2941 goto next;
2942
2943 if (key.objectid >= logic_end) {
2944 stop_loop = 1;
2945 break;
2946 }
2947
2948 while (key.objectid >= logic_start + map->stripe_len)
2949 logic_start += map->stripe_len;
2950
2951 extent = btrfs_item_ptr(l, slot,
2952 struct btrfs_extent_item);
2953 flags = btrfs_extent_flags(l, extent);
2954 generation = btrfs_extent_generation(l, extent);
2955
2956 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
2957 (key.objectid < logic_start ||
2958 key.objectid + bytes >
2959 logic_start + map->stripe_len)) {
2960 btrfs_err(fs_info, "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
2961 key.objectid, logic_start);
2962 spin_lock(&sctx->stat_lock);
2963 sctx->stat.uncorrectable_errors++;
2964 spin_unlock(&sctx->stat_lock);
2965 goto next;
2966 }
2967 again:
2968 extent_logical = key.objectid;
2969 extent_len = bytes;
2970
2971 if (extent_logical < logic_start) {
2972 extent_len -= logic_start - extent_logical;
2973 extent_logical = logic_start;
2974 }
2975
2976 if (extent_logical + extent_len >
2977 logic_start + map->stripe_len)
2978 extent_len = logic_start + map->stripe_len -
2979 extent_logical;
2980
2981 scrub_parity_mark_sectors_data(sparity, extent_logical,
2982 extent_len);
2983
2984 mapped_length = extent_len;
2985 bbio = NULL;
2986 ret = btrfs_map_block(fs_info, READ, extent_logical,
2987 &mapped_length, &bbio, 0);
2988 if (!ret) {
2989 if (!bbio || mapped_length < extent_len)
2990 ret = -EIO;
2991 }
2992 if (ret) {
2993 btrfs_put_bbio(bbio);
2994 goto out;
2995 }
2996 extent_physical = bbio->stripes[0].physical;
2997 extent_mirror_num = bbio->mirror_num;
2998 extent_dev = bbio->stripes[0].dev;
2999 btrfs_put_bbio(bbio);
3000
3001 ret = btrfs_lookup_csums_range(csum_root,
3002 extent_logical,
3003 extent_logical + extent_len - 1,
3004 &sctx->csum_list, 1);
3005 if (ret)
3006 goto out;
3007
3008 ret = scrub_extent_for_parity(sparity, extent_logical,
3009 extent_len,
3010 extent_physical,
3011 extent_dev, flags,
3012 generation,
3013 extent_mirror_num);
3014
3015 scrub_free_csums(sctx);
3016
3017 if (ret)
3018 goto out;
3019
3020 if (extent_logical + extent_len <
3021 key.objectid + bytes) {
3022 logic_start += map->stripe_len;
3023
3024 if (logic_start >= logic_end) {
3025 stop_loop = 1;
3026 break;
3027 }
3028
3029 if (logic_start < key.objectid + bytes) {
3030 cond_resched();
3031 goto again;
3032 }
3033 }
3034 next:
3035 path->slots[0]++;
3036 }
3037
3038 btrfs_release_path(path);
3039
3040 if (stop_loop)
3041 break;
3042
3043 logic_start += map->stripe_len;
3044 }
3045 out:
3046 if (ret < 0)
3047 scrub_parity_mark_sectors_error(sparity, logic_start,
3048 logic_end - logic_start);
3049 scrub_parity_put(sparity);
3050 scrub_submit(sctx);
3051 mutex_lock(&sctx->wr_ctx.wr_lock);
3052 scrub_wr_submit(sctx);
3053 mutex_unlock(&sctx->wr_ctx.wr_lock);
3054
3055 btrfs_release_path(path);
3056 return ret < 0 ? ret : 0;
3057 }
3058
3059 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3060 struct map_lookup *map,
3061 struct btrfs_device *scrub_dev,
3062 int num, u64 base, u64 length,
3063 int is_dev_replace)
3064 {
3065 struct btrfs_path *path, *ppath;
3066 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
3067 struct btrfs_root *root = fs_info->extent_root;
3068 struct btrfs_root *csum_root = fs_info->csum_root;
3069 struct btrfs_extent_item *extent;
3070 struct blk_plug plug;
3071 u64 flags;
3072 int ret;
3073 int slot;
3074 u64 nstripes;
3075 struct extent_buffer *l;
3076 u64 physical;
3077 u64 logical;
3078 u64 logic_end;
3079 u64 physical_end;
3080 u64 generation;
3081 int mirror_num;
3082 struct reada_control *reada1;
3083 struct reada_control *reada2;
3084 struct btrfs_key key;
3085 struct btrfs_key key_end;
3086 u64 increment = map->stripe_len;
3087 u64 offset;
3088 u64 extent_logical;
3089 u64 extent_physical;
3090 u64 extent_len;
3091 u64 stripe_logical;
3092 u64 stripe_end;
3093 struct btrfs_device *extent_dev;
3094 int extent_mirror_num;
3095 int stop_loop = 0;
3096
3097 physical = map->stripes[num].physical;
3098 offset = 0;
3099 nstripes = div_u64(length, map->stripe_len);
3100 if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3101 offset = map->stripe_len * num;
3102 increment = map->stripe_len * map->num_stripes;
3103 mirror_num = 1;
3104 } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3105 int factor = map->num_stripes / map->sub_stripes;
3106 offset = map->stripe_len * (num / map->sub_stripes);
3107 increment = map->stripe_len * factor;
3108 mirror_num = num % map->sub_stripes + 1;
3109 } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
3110 increment = map->stripe_len;
3111 mirror_num = num % map->num_stripes + 1;
3112 } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3113 increment = map->stripe_len;
3114 mirror_num = num % map->num_stripes + 1;
3115 } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3116 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3117 increment = map->stripe_len * nr_data_stripes(map);
3118 mirror_num = 1;
3119 } else {
3120 increment = map->stripe_len;
3121 mirror_num = 1;
3122 }
3123
3124 path = btrfs_alloc_path();
3125 if (!path)
3126 return -ENOMEM;
3127
3128 ppath = btrfs_alloc_path();
3129 if (!ppath) {
3130 btrfs_free_path(path);
3131 return -ENOMEM;
3132 }
3133
3134 /*
3135 * work on commit root. The related disk blocks are static as
3136 * long as COW is applied. This means, it is save to rewrite
3137 * them to repair disk errors without any race conditions
3138 */
3139 path->search_commit_root = 1;
3140 path->skip_locking = 1;
3141
3142 ppath->search_commit_root = 1;
3143 ppath->skip_locking = 1;
3144 /*
3145 * trigger the readahead for extent tree csum tree and wait for
3146 * completion. During readahead, the scrub is officially paused
3147 * to not hold off transaction commits
3148 */
3149 logical = base + offset;
3150 physical_end = physical + nstripes * map->stripe_len;
3151 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3152 get_raid56_logic_offset(physical_end, num,
3153 map, &logic_end, NULL);
3154 logic_end += base;
3155 } else {
3156 logic_end = logical + increment * nstripes;
3157 }
3158 wait_event(sctx->list_wait,
3159 atomic_read(&sctx->bios_in_flight) == 0);
3160 scrub_blocked_if_needed(fs_info);
3161
3162 /* FIXME it might be better to start readahead at commit root */
3163 key.objectid = logical;
3164 key.type = BTRFS_EXTENT_ITEM_KEY;
3165 key.offset = (u64)0;
3166 key_end.objectid = logic_end;
3167 key_end.type = BTRFS_METADATA_ITEM_KEY;
3168 key_end.offset = (u64)-1;
3169 reada1 = btrfs_reada_add(root, &key, &key_end);
3170
3171 key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3172 key.type = BTRFS_EXTENT_CSUM_KEY;
3173 key.offset = logical;
3174 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3175 key_end.type = BTRFS_EXTENT_CSUM_KEY;
3176 key_end.offset = logic_end;
3177 reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3178
3179 if (!IS_ERR(reada1))
3180 btrfs_reada_wait(reada1);
3181 if (!IS_ERR(reada2))
3182 btrfs_reada_wait(reada2);
3183
3184
3185 /*
3186 * collect all data csums for the stripe to avoid seeking during
3187 * the scrub. This might currently (crc32) end up to be about 1MB
3188 */
3189 blk_start_plug(&plug);
3190
3191 /*
3192 * now find all extents for each stripe and scrub them
3193 */
3194 ret = 0;
3195 while (physical < physical_end) {
3196 /*
3197 * canceled?
3198 */
3199 if (atomic_read(&fs_info->scrub_cancel_req) ||
3200 atomic_read(&sctx->cancel_req)) {
3201 ret = -ECANCELED;
3202 goto out;
3203 }
3204 /*
3205 * check to see if we have to pause
3206 */
3207 if (atomic_read(&fs_info->scrub_pause_req)) {
3208 /* push queued extents */
3209 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
3210 scrub_submit(sctx);
3211 mutex_lock(&sctx->wr_ctx.wr_lock);
3212 scrub_wr_submit(sctx);
3213 mutex_unlock(&sctx->wr_ctx.wr_lock);
3214 wait_event(sctx->list_wait,
3215 atomic_read(&sctx->bios_in_flight) == 0);
3216 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
3217 scrub_blocked_if_needed(fs_info);
3218 }
3219
3220 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3221 ret = get_raid56_logic_offset(physical, num, map,
3222 &logical,
3223 &stripe_logical);
3224 logical += base;
3225 if (ret) {
3226 /* it is parity strip */
3227 stripe_logical += base;
3228 stripe_end = stripe_logical + increment;
3229 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3230 ppath, stripe_logical,
3231 stripe_end);
3232 if (ret)
3233 goto out;
3234 goto skip;
3235 }
3236 }
3237
3238 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3239 key.type = BTRFS_METADATA_ITEM_KEY;
3240 else
3241 key.type = BTRFS_EXTENT_ITEM_KEY;
3242 key.objectid = logical;
3243 key.offset = (u64)-1;
3244
3245 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3246 if (ret < 0)
3247 goto out;
3248
3249 if (ret > 0) {
3250 ret = btrfs_previous_extent_item(root, path, 0);
3251 if (ret < 0)
3252 goto out;
3253 if (ret > 0) {
3254 /* there's no smaller item, so stick with the
3255 * larger one */
3256 btrfs_release_path(path);
3257 ret = btrfs_search_slot(NULL, root, &key,
3258 path, 0, 0);
3259 if (ret < 0)
3260 goto out;
3261 }
3262 }
3263
3264 stop_loop = 0;
3265 while (1) {
3266 u64 bytes;
3267
3268 l = path->nodes[0];
3269 slot = path->slots[0];
3270 if (slot >= btrfs_header_nritems(l)) {
3271 ret = btrfs_next_leaf(root, path);
3272 if (ret == 0)
3273 continue;
3274 if (ret < 0)
3275 goto out;
3276
3277 stop_loop = 1;
3278 break;
3279 }
3280 btrfs_item_key_to_cpu(l, &key, slot);
3281
3282 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3283 key.type != BTRFS_METADATA_ITEM_KEY)
3284 goto next;
3285
3286 if (key.type == BTRFS_METADATA_ITEM_KEY)
3287 bytes = root->nodesize;
3288 else
3289 bytes = key.offset;
3290
3291 if (key.objectid + bytes <= logical)
3292 goto next;
3293
3294 if (key.objectid >= logical + map->stripe_len) {
3295 /* out of this device extent */
3296 if (key.objectid >= logic_end)
3297 stop_loop = 1;
3298 break;
3299 }
3300
3301 extent = btrfs_item_ptr(l, slot,
3302 struct btrfs_extent_item);
3303 flags = btrfs_extent_flags(l, extent);
3304 generation = btrfs_extent_generation(l, extent);
3305
3306 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3307 (key.objectid < logical ||
3308 key.objectid + bytes >
3309 logical + map->stripe_len)) {
3310 btrfs_err(fs_info,
3311 "scrub: tree block %llu spanning "
3312 "stripes, ignored. logical=%llu",
3313 key.objectid, logical);
3314 spin_lock(&sctx->stat_lock);
3315 sctx->stat.uncorrectable_errors++;
3316 spin_unlock(&sctx->stat_lock);
3317 goto next;
3318 }
3319
3320 again:
3321 extent_logical = key.objectid;
3322 extent_len = bytes;
3323
3324 /*
3325 * trim extent to this stripe
3326 */
3327 if (extent_logical < logical) {
3328 extent_len -= logical - extent_logical;
3329 extent_logical = logical;
3330 }
3331 if (extent_logical + extent_len >
3332 logical + map->stripe_len) {
3333 extent_len = logical + map->stripe_len -
3334 extent_logical;
3335 }
3336
3337 extent_physical = extent_logical - logical + physical;
3338 extent_dev = scrub_dev;
3339 extent_mirror_num = mirror_num;
3340 if (is_dev_replace)
3341 scrub_remap_extent(fs_info, extent_logical,
3342 extent_len, &extent_physical,
3343 &extent_dev,
3344 &extent_mirror_num);
3345
3346 ret = btrfs_lookup_csums_range(csum_root,
3347 extent_logical,
3348 extent_logical +
3349 extent_len - 1,
3350 &sctx->csum_list, 1);
3351 if (ret)
3352 goto out;
3353
3354 ret = scrub_extent(sctx, extent_logical, extent_len,
3355 extent_physical, extent_dev, flags,
3356 generation, extent_mirror_num,
3357 extent_logical - logical + physical);
3358
3359 scrub_free_csums(sctx);
3360
3361 if (ret)
3362 goto out;
3363
3364 if (extent_logical + extent_len <
3365 key.objectid + bytes) {
3366 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3367 /*
3368 * loop until we find next data stripe
3369 * or we have finished all stripes.
3370 */
3371 loop:
3372 physical += map->stripe_len;
3373 ret = get_raid56_logic_offset(physical,
3374 num, map, &logical,
3375 &stripe_logical);
3376 logical += base;
3377
3378 if (ret && physical < physical_end) {
3379 stripe_logical += base;
3380 stripe_end = stripe_logical +
3381 increment;
3382 ret = scrub_raid56_parity(sctx,
3383 map, scrub_dev, ppath,
3384 stripe_logical,
3385 stripe_end);
3386 if (ret)
3387 goto out;
3388 goto loop;
3389 }
3390 } else {
3391 physical += map->stripe_len;
3392 logical += increment;
3393 }
3394 if (logical < key.objectid + bytes) {
3395 cond_resched();
3396 goto again;
3397 }
3398
3399 if (physical >= physical_end) {
3400 stop_loop = 1;
3401 break;
3402 }
3403 }
3404 next:
3405 path->slots[0]++;
3406 }
3407 btrfs_release_path(path);
3408 skip:
3409 logical += increment;
3410 physical += map->stripe_len;
3411 spin_lock(&sctx->stat_lock);
3412 if (stop_loop)
3413 sctx->stat.last_physical = map->stripes[num].physical +
3414 length;
3415 else
3416 sctx->stat.last_physical = physical;
3417 spin_unlock(&sctx->stat_lock);
3418 if (stop_loop)
3419 break;
3420 }
3421 out:
3422 /* push queued extents */
3423 scrub_submit(sctx);
3424 mutex_lock(&sctx->wr_ctx.wr_lock);
3425 scrub_wr_submit(sctx);
3426 mutex_unlock(&sctx->wr_ctx.wr_lock);
3427
3428 blk_finish_plug(&plug);
3429 btrfs_free_path(path);
3430 btrfs_free_path(ppath);
3431 return ret < 0 ? ret : 0;
3432 }
3433
3434 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3435 struct btrfs_device *scrub_dev,
3436 u64 chunk_offset, u64 length,
3437 u64 dev_offset,
3438 struct btrfs_block_group_cache *cache,
3439 int is_dev_replace)
3440 {
3441 struct btrfs_mapping_tree *map_tree =
3442 &sctx->dev_root->fs_info->mapping_tree;
3443 struct map_lookup *map;
3444 struct extent_map *em;
3445 int i;
3446 int ret = 0;
3447
3448 read_lock(&map_tree->map_tree.lock);
3449 em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
3450 read_unlock(&map_tree->map_tree.lock);
3451
3452 if (!em) {
3453 /*
3454 * Might have been an unused block group deleted by the cleaner
3455 * kthread or relocation.
3456 */
3457 spin_lock(&cache->lock);
3458 if (!cache->removed)
3459 ret = -EINVAL;
3460 spin_unlock(&cache->lock);
3461
3462 return ret;
3463 }
3464
3465 map = em->map_lookup;
3466 if (em->start != chunk_offset)
3467 goto out;
3468
3469 if (em->len < length)
3470 goto out;
3471
3472 for (i = 0; i < map->num_stripes; ++i) {
3473 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3474 map->stripes[i].physical == dev_offset) {
3475 ret = scrub_stripe(sctx, map, scrub_dev, i,
3476 chunk_offset, length,
3477 is_dev_replace);
3478 if (ret)
3479 goto out;
3480 }
3481 }
3482 out:
3483 free_extent_map(em);
3484
3485 return ret;
3486 }
3487
3488 static noinline_for_stack
3489 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3490 struct btrfs_device *scrub_dev, u64 start, u64 end,
3491 int is_dev_replace)
3492 {
3493 struct btrfs_dev_extent *dev_extent = NULL;
3494 struct btrfs_path *path;
3495 struct btrfs_root *root = sctx->dev_root;
3496 struct btrfs_fs_info *fs_info = root->fs_info;
3497 u64 length;
3498 u64 chunk_offset;
3499 int ret = 0;
3500 int ro_set;
3501 int slot;
3502 struct extent_buffer *l;
3503 struct btrfs_key key;
3504 struct btrfs_key found_key;
3505 struct btrfs_block_group_cache *cache;
3506 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3507
3508 path = btrfs_alloc_path();
3509 if (!path)
3510 return -ENOMEM;
3511
3512 path->reada = READA_FORWARD;
3513 path->search_commit_root = 1;
3514 path->skip_locking = 1;
3515
3516 key.objectid = scrub_dev->devid;
3517 key.offset = 0ull;
3518 key.type = BTRFS_DEV_EXTENT_KEY;
3519
3520 while (1) {
3521 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3522 if (ret < 0)
3523 break;
3524 if (ret > 0) {
3525 if (path->slots[0] >=
3526 btrfs_header_nritems(path->nodes[0])) {
3527 ret = btrfs_next_leaf(root, path);
3528 if (ret < 0)
3529 break;
3530 if (ret > 0) {
3531 ret = 0;
3532 break;
3533 }
3534 } else {
3535 ret = 0;
3536 }
3537 }
3538
3539 l = path->nodes[0];
3540 slot = path->slots[0];
3541
3542 btrfs_item_key_to_cpu(l, &found_key, slot);
3543
3544 if (found_key.objectid != scrub_dev->devid)
3545 break;
3546
3547 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3548 break;
3549
3550 if (found_key.offset >= end)
3551 break;
3552
3553 if (found_key.offset < key.offset)
3554 break;
3555
3556 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3557 length = btrfs_dev_extent_length(l, dev_extent);
3558
3559 if (found_key.offset + length <= start)
3560 goto skip;
3561
3562 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3563
3564 /*
3565 * get a reference on the corresponding block group to prevent
3566 * the chunk from going away while we scrub it
3567 */
3568 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3569
3570 /* some chunks are removed but not committed to disk yet,
3571 * continue scrubbing */
3572 if (!cache)
3573 goto skip;
3574
3575 /*
3576 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3577 * to avoid deadlock caused by:
3578 * btrfs_inc_block_group_ro()
3579 * -> btrfs_wait_for_commit()
3580 * -> btrfs_commit_transaction()
3581 * -> btrfs_scrub_pause()
3582 */
3583 scrub_pause_on(fs_info);
3584 ret = btrfs_inc_block_group_ro(root, cache);
3585 scrub_pause_off(fs_info);
3586
3587 if (ret == 0) {
3588 ro_set = 1;
3589 } else if (ret == -ENOSPC) {
3590 /*
3591 * btrfs_inc_block_group_ro return -ENOSPC when it
3592 * failed in creating new chunk for metadata.
3593 * It is not a problem for scrub/replace, because
3594 * metadata are always cowed, and our scrub paused
3595 * commit_transactions.
3596 */
3597 ro_set = 0;
3598 } else {
3599 btrfs_warn(fs_info, "failed setting block group ro, ret=%d\n",
3600 ret);
3601 btrfs_put_block_group(cache);
3602 break;
3603 }
3604
3605 dev_replace->cursor_right = found_key.offset + length;
3606 dev_replace->cursor_left = found_key.offset;
3607 dev_replace->item_needs_writeback = 1;
3608 ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3609 found_key.offset, cache, is_dev_replace);
3610
3611 /*
3612 * flush, submit all pending read and write bios, afterwards
3613 * wait for them.
3614 * Note that in the dev replace case, a read request causes
3615 * write requests that are submitted in the read completion
3616 * worker. Therefore in the current situation, it is required
3617 * that all write requests are flushed, so that all read and
3618 * write requests are really completed when bios_in_flight
3619 * changes to 0.
3620 */
3621 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
3622 scrub_submit(sctx);
3623 mutex_lock(&sctx->wr_ctx.wr_lock);
3624 scrub_wr_submit(sctx);
3625 mutex_unlock(&sctx->wr_ctx.wr_lock);
3626
3627 wait_event(sctx->list_wait,
3628 atomic_read(&sctx->bios_in_flight) == 0);
3629
3630 scrub_pause_on(fs_info);
3631
3632 /*
3633 * must be called before we decrease @scrub_paused.
3634 * make sure we don't block transaction commit while
3635 * we are waiting pending workers finished.
3636 */
3637 wait_event(sctx->list_wait,
3638 atomic_read(&sctx->workers_pending) == 0);
3639 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
3640
3641 scrub_pause_off(fs_info);
3642
3643 if (ro_set)
3644 btrfs_dec_block_group_ro(root, cache);
3645
3646 /*
3647 * We might have prevented the cleaner kthread from deleting
3648 * this block group if it was already unused because we raced
3649 * and set it to RO mode first. So add it back to the unused
3650 * list, otherwise it might not ever be deleted unless a manual
3651 * balance is triggered or it becomes used and unused again.
3652 */
3653 spin_lock(&cache->lock);
3654 if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3655 btrfs_block_group_used(&cache->item) == 0) {
3656 spin_unlock(&cache->lock);
3657 spin_lock(&fs_info->unused_bgs_lock);
3658 if (list_empty(&cache->bg_list)) {
3659 btrfs_get_block_group(cache);
3660 list_add_tail(&cache->bg_list,
3661 &fs_info->unused_bgs);
3662 }
3663 spin_unlock(&fs_info->unused_bgs_lock);
3664 } else {
3665 spin_unlock(&cache->lock);
3666 }
3667
3668 btrfs_put_block_group(cache);
3669 if (ret)
3670 break;
3671 if (is_dev_replace &&
3672 atomic64_read(&dev_replace->num_write_errors) > 0) {
3673 ret = -EIO;
3674 break;
3675 }
3676 if (sctx->stat.malloc_errors > 0) {
3677 ret = -ENOMEM;
3678 break;
3679 }
3680
3681 dev_replace->cursor_left = dev_replace->cursor_right;
3682 dev_replace->item_needs_writeback = 1;
3683 skip:
3684 key.offset = found_key.offset + length;
3685 btrfs_release_path(path);
3686 }
3687
3688 btrfs_free_path(path);
3689
3690 return ret;
3691 }
3692
3693 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3694 struct btrfs_device *scrub_dev)
3695 {
3696 int i;
3697 u64 bytenr;
3698 u64 gen;
3699 int ret;
3700 struct btrfs_root *root = sctx->dev_root;
3701
3702 if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
3703 return -EIO;
3704
3705 /* Seed devices of a new filesystem has their own generation. */
3706 if (scrub_dev->fs_devices != root->fs_info->fs_devices)
3707 gen = scrub_dev->generation;
3708 else
3709 gen = root->fs_info->last_trans_committed;
3710
3711 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3712 bytenr = btrfs_sb_offset(i);
3713 if (bytenr + BTRFS_SUPER_INFO_SIZE >
3714 scrub_dev->commit_total_bytes)
3715 break;
3716
3717 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3718 scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3719 NULL, 1, bytenr);
3720 if (ret)
3721 return ret;
3722 }
3723 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3724
3725 return 0;
3726 }
3727
3728 /*
3729 * get a reference count on fs_info->scrub_workers. start worker if necessary
3730 */
3731 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
3732 int is_dev_replace)
3733 {
3734 unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
3735 int max_active = fs_info->thread_pool_size;
3736
3737 if (fs_info->scrub_workers_refcnt == 0) {
3738 if (is_dev_replace)
3739 fs_info->scrub_workers =
3740 btrfs_alloc_workqueue("scrub", flags,
3741 1, 4);
3742 else
3743 fs_info->scrub_workers =
3744 btrfs_alloc_workqueue("scrub", flags,
3745 max_active, 4);
3746 if (!fs_info->scrub_workers)
3747 goto fail_scrub_workers;
3748
3749 fs_info->scrub_wr_completion_workers =
3750 btrfs_alloc_workqueue("scrubwrc", flags,
3751 max_active, 2);
3752 if (!fs_info->scrub_wr_completion_workers)
3753 goto fail_scrub_wr_completion_workers;
3754
3755 fs_info->scrub_nocow_workers =
3756 btrfs_alloc_workqueue("scrubnc", flags, 1, 0);
3757 if (!fs_info->scrub_nocow_workers)
3758 goto fail_scrub_nocow_workers;
3759 fs_info->scrub_parity_workers =
3760 btrfs_alloc_workqueue("scrubparity", flags,
3761 max_active, 2);
3762 if (!fs_info->scrub_parity_workers)
3763 goto fail_scrub_parity_workers;
3764 }
3765 ++fs_info->scrub_workers_refcnt;
3766 return 0;
3767
3768 fail_scrub_parity_workers:
3769 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
3770 fail_scrub_nocow_workers:
3771 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3772 fail_scrub_wr_completion_workers:
3773 btrfs_destroy_workqueue(fs_info->scrub_workers);
3774 fail_scrub_workers:
3775 return -ENOMEM;
3776 }
3777
3778 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
3779 {
3780 if (--fs_info->scrub_workers_refcnt == 0) {
3781 btrfs_destroy_workqueue(fs_info->scrub_workers);
3782 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3783 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
3784 btrfs_destroy_workqueue(fs_info->scrub_parity_workers);
3785 }
3786 WARN_ON(fs_info->scrub_workers_refcnt < 0);
3787 }
3788
3789 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
3790 u64 end, struct btrfs_scrub_progress *progress,
3791 int readonly, int is_dev_replace)
3792 {
3793 struct scrub_ctx *sctx;
3794 int ret;
3795 struct btrfs_device *dev;
3796 struct rcu_string *name;
3797
3798 if (btrfs_fs_closing(fs_info))
3799 return -EINVAL;
3800
3801 if (fs_info->chunk_root->nodesize > BTRFS_STRIPE_LEN) {
3802 /*
3803 * in this case scrub is unable to calculate the checksum
3804 * the way scrub is implemented. Do not handle this
3805 * situation at all because it won't ever happen.
3806 */
3807 btrfs_err(fs_info,
3808 "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
3809 fs_info->chunk_root->nodesize, BTRFS_STRIPE_LEN);
3810 return -EINVAL;
3811 }
3812
3813 if (fs_info->chunk_root->sectorsize != PAGE_SIZE) {
3814 /* not supported for data w/o checksums */
3815 btrfs_err(fs_info,
3816 "scrub: size assumption sectorsize != PAGE_SIZE "
3817 "(%d != %lu) fails",
3818 fs_info->chunk_root->sectorsize, PAGE_SIZE);
3819 return -EINVAL;
3820 }
3821
3822 if (fs_info->chunk_root->nodesize >
3823 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
3824 fs_info->chunk_root->sectorsize >
3825 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
3826 /*
3827 * would exhaust the array bounds of pagev member in
3828 * struct scrub_block
3829 */
3830 btrfs_err(fs_info, "scrub: size assumption nodesize and sectorsize "
3831 "<= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
3832 fs_info->chunk_root->nodesize,
3833 SCRUB_MAX_PAGES_PER_BLOCK,
3834 fs_info->chunk_root->sectorsize,
3835 SCRUB_MAX_PAGES_PER_BLOCK);
3836 return -EINVAL;
3837 }
3838
3839
3840 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3841 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
3842 if (!dev || (dev->missing && !is_dev_replace)) {
3843 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3844 return -ENODEV;
3845 }
3846
3847 if (!is_dev_replace && !readonly && !dev->writeable) {
3848 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3849 rcu_read_lock();
3850 name = rcu_dereference(dev->name);
3851 btrfs_err(fs_info, "scrub: device %s is not writable",
3852 name->str);
3853 rcu_read_unlock();
3854 return -EROFS;
3855 }
3856
3857 mutex_lock(&fs_info->scrub_lock);
3858 if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
3859 mutex_unlock(&fs_info->scrub_lock);
3860 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3861 return -EIO;
3862 }
3863
3864 btrfs_dev_replace_lock(&fs_info->dev_replace, 0);
3865 if (dev->scrub_device ||
3866 (!is_dev_replace &&
3867 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
3868 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
3869 mutex_unlock(&fs_info->scrub_lock);
3870 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3871 return -EINPROGRESS;
3872 }
3873 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
3874
3875 ret = scrub_workers_get(fs_info, is_dev_replace);
3876 if (ret) {
3877 mutex_unlock(&fs_info->scrub_lock);
3878 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3879 return ret;
3880 }
3881
3882 sctx = scrub_setup_ctx(dev, is_dev_replace);
3883 if (IS_ERR(sctx)) {
3884 mutex_unlock(&fs_info->scrub_lock);
3885 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3886 scrub_workers_put(fs_info);
3887 return PTR_ERR(sctx);
3888 }
3889 sctx->readonly = readonly;
3890 dev->scrub_device = sctx;
3891 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3892
3893 /*
3894 * checking @scrub_pause_req here, we can avoid
3895 * race between committing transaction and scrubbing.
3896 */
3897 __scrub_blocked_if_needed(fs_info);
3898 atomic_inc(&fs_info->scrubs_running);
3899 mutex_unlock(&fs_info->scrub_lock);
3900
3901 if (!is_dev_replace) {
3902 /*
3903 * by holding device list mutex, we can
3904 * kick off writing super in log tree sync.
3905 */
3906 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3907 ret = scrub_supers(sctx, dev);
3908 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3909 }
3910
3911 if (!ret)
3912 ret = scrub_enumerate_chunks(sctx, dev, start, end,
3913 is_dev_replace);
3914
3915 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3916 atomic_dec(&fs_info->scrubs_running);
3917 wake_up(&fs_info->scrub_pause_wait);
3918
3919 wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3920
3921 if (progress)
3922 memcpy(progress, &sctx->stat, sizeof(*progress));
3923
3924 mutex_lock(&fs_info->scrub_lock);
3925 dev->scrub_device = NULL;
3926 scrub_workers_put(fs_info);
3927 mutex_unlock(&fs_info->scrub_lock);
3928
3929 scrub_put_ctx(sctx);
3930
3931 return ret;
3932 }
3933
3934 void btrfs_scrub_pause(struct btrfs_root *root)
3935 {
3936 struct btrfs_fs_info *fs_info = root->fs_info;
3937
3938 mutex_lock(&fs_info->scrub_lock);
3939 atomic_inc(&fs_info->scrub_pause_req);
3940 while (atomic_read(&fs_info->scrubs_paused) !=
3941 atomic_read(&fs_info->scrubs_running)) {
3942 mutex_unlock(&fs_info->scrub_lock);
3943 wait_event(fs_info->scrub_pause_wait,
3944 atomic_read(&fs_info->scrubs_paused) ==
3945 atomic_read(&fs_info->scrubs_running));
3946 mutex_lock(&fs_info->scrub_lock);
3947 }
3948 mutex_unlock(&fs_info->scrub_lock);
3949 }
3950
3951 void btrfs_scrub_continue(struct btrfs_root *root)
3952 {
3953 struct btrfs_fs_info *fs_info = root->fs_info;
3954
3955 atomic_dec(&fs_info->scrub_pause_req);
3956 wake_up(&fs_info->scrub_pause_wait);
3957 }
3958
3959 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
3960 {
3961 mutex_lock(&fs_info->scrub_lock);
3962 if (!atomic_read(&fs_info->scrubs_running)) {
3963 mutex_unlock(&fs_info->scrub_lock);
3964 return -ENOTCONN;
3965 }
3966
3967 atomic_inc(&fs_info->scrub_cancel_req);
3968 while (atomic_read(&fs_info->scrubs_running)) {
3969 mutex_unlock(&fs_info->scrub_lock);
3970 wait_event(fs_info->scrub_pause_wait,
3971 atomic_read(&fs_info->scrubs_running) == 0);
3972 mutex_lock(&fs_info->scrub_lock);
3973 }
3974 atomic_dec(&fs_info->scrub_cancel_req);
3975 mutex_unlock(&fs_info->scrub_lock);
3976
3977 return 0;
3978 }
3979
3980 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
3981 struct btrfs_device *dev)
3982 {
3983 struct scrub_ctx *sctx;
3984
3985 mutex_lock(&fs_info->scrub_lock);
3986 sctx = dev->scrub_device;
3987 if (!sctx) {
3988 mutex_unlock(&fs_info->scrub_lock);
3989 return -ENOTCONN;
3990 }
3991 atomic_inc(&sctx->cancel_req);
3992 while (dev->scrub_device) {
3993 mutex_unlock(&fs_info->scrub_lock);
3994 wait_event(fs_info->scrub_pause_wait,
3995 dev->scrub_device == NULL);
3996 mutex_lock(&fs_info->scrub_lock);
3997 }
3998 mutex_unlock(&fs_info->scrub_lock);
3999
4000 return 0;
4001 }
4002
4003 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
4004 struct btrfs_scrub_progress *progress)
4005 {
4006 struct btrfs_device *dev;
4007 struct scrub_ctx *sctx = NULL;
4008
4009 mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
4010 dev = btrfs_find_device(root->fs_info, devid, NULL, NULL);
4011 if (dev)
4012 sctx = dev->scrub_device;
4013 if (sctx)
4014 memcpy(progress, &sctx->stat, sizeof(*progress));
4015 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
4016
4017 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4018 }
4019
4020 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4021 u64 extent_logical, u64 extent_len,
4022 u64 *extent_physical,
4023 struct btrfs_device **extent_dev,
4024 int *extent_mirror_num)
4025 {
4026 u64 mapped_length;
4027 struct btrfs_bio *bbio = NULL;
4028 int ret;
4029
4030 mapped_length = extent_len;
4031 ret = btrfs_map_block(fs_info, READ, extent_logical,
4032 &mapped_length, &bbio, 0);
4033 if (ret || !bbio || mapped_length < extent_len ||
4034 !bbio->stripes[0].dev->bdev) {
4035 btrfs_put_bbio(bbio);
4036 return;
4037 }
4038
4039 *extent_physical = bbio->stripes[0].physical;
4040 *extent_mirror_num = bbio->mirror_num;
4041 *extent_dev = bbio->stripes[0].dev;
4042 btrfs_put_bbio(bbio);
4043 }
4044
4045 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
4046 struct scrub_wr_ctx *wr_ctx,
4047 struct btrfs_fs_info *fs_info,
4048 struct btrfs_device *dev,
4049 int is_dev_replace)
4050 {
4051 WARN_ON(wr_ctx->wr_curr_bio != NULL);
4052
4053 mutex_init(&wr_ctx->wr_lock);
4054 wr_ctx->wr_curr_bio = NULL;
4055 if (!is_dev_replace)
4056 return 0;
4057
4058 WARN_ON(!dev->bdev);
4059 wr_ctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
4060 wr_ctx->tgtdev = dev;
4061 atomic_set(&wr_ctx->flush_all_writes, 0);
4062 return 0;
4063 }
4064
4065 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
4066 {
4067 mutex_lock(&wr_ctx->wr_lock);
4068 kfree(wr_ctx->wr_curr_bio);
4069 wr_ctx->wr_curr_bio = NULL;
4070 mutex_unlock(&wr_ctx->wr_lock);
4071 }
4072
4073 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
4074 int mirror_num, u64 physical_for_dev_replace)
4075 {
4076 struct scrub_copy_nocow_ctx *nocow_ctx;
4077 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
4078
4079 nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
4080 if (!nocow_ctx) {
4081 spin_lock(&sctx->stat_lock);
4082 sctx->stat.malloc_errors++;
4083 spin_unlock(&sctx->stat_lock);
4084 return -ENOMEM;
4085 }
4086
4087 scrub_pending_trans_workers_inc(sctx);
4088
4089 nocow_ctx->sctx = sctx;
4090 nocow_ctx->logical = logical;
4091 nocow_ctx->len = len;
4092 nocow_ctx->mirror_num = mirror_num;
4093 nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
4094 btrfs_init_work(&nocow_ctx->work, btrfs_scrubnc_helper,
4095 copy_nocow_pages_worker, NULL, NULL);
4096 INIT_LIST_HEAD(&nocow_ctx->inodes);
4097 btrfs_queue_work(fs_info->scrub_nocow_workers,
4098 &nocow_ctx->work);
4099
4100 return 0;
4101 }
4102
4103 static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
4104 {
4105 struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
4106 struct scrub_nocow_inode *nocow_inode;
4107
4108 nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
4109 if (!nocow_inode)
4110 return -ENOMEM;
4111 nocow_inode->inum = inum;
4112 nocow_inode->offset = offset;
4113 nocow_inode->root = root;
4114 list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
4115 return 0;
4116 }
4117
4118 #define COPY_COMPLETE 1
4119
4120 static void copy_nocow_pages_worker(struct btrfs_work *work)
4121 {
4122 struct scrub_copy_nocow_ctx *nocow_ctx =
4123 container_of(work, struct scrub_copy_nocow_ctx, work);
4124 struct scrub_ctx *sctx = nocow_ctx->sctx;
4125 u64 logical = nocow_ctx->logical;
4126 u64 len = nocow_ctx->len;
4127 int mirror_num = nocow_ctx->mirror_num;
4128 u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4129 int ret;
4130 struct btrfs_trans_handle *trans = NULL;
4131 struct btrfs_fs_info *fs_info;
4132 struct btrfs_path *path;
4133 struct btrfs_root *root;
4134 int not_written = 0;
4135
4136 fs_info = sctx->dev_root->fs_info;
4137 root = fs_info->extent_root;
4138
4139 path = btrfs_alloc_path();
4140 if (!path) {
4141 spin_lock(&sctx->stat_lock);
4142 sctx->stat.malloc_errors++;
4143 spin_unlock(&sctx->stat_lock);
4144 not_written = 1;
4145 goto out;
4146 }
4147
4148 trans = btrfs_join_transaction(root);
4149 if (IS_ERR(trans)) {
4150 not_written = 1;
4151 goto out;
4152 }
4153
4154 ret = iterate_inodes_from_logical(logical, fs_info, path,
4155 record_inode_for_nocow, nocow_ctx);
4156 if (ret != 0 && ret != -ENOENT) {
4157 btrfs_warn(fs_info, "iterate_inodes_from_logical() failed: log %llu, "
4158 "phys %llu, len %llu, mir %u, ret %d",
4159 logical, physical_for_dev_replace, len, mirror_num,
4160 ret);
4161 not_written = 1;
4162 goto out;
4163 }
4164
4165 btrfs_end_transaction(trans, root);
4166 trans = NULL;
4167 while (!list_empty(&nocow_ctx->inodes)) {
4168 struct scrub_nocow_inode *entry;
4169 entry = list_first_entry(&nocow_ctx->inodes,
4170 struct scrub_nocow_inode,
4171 list);
4172 list_del_init(&entry->list);
4173 ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
4174 entry->root, nocow_ctx);
4175 kfree(entry);
4176 if (ret == COPY_COMPLETE) {
4177 ret = 0;
4178 break;
4179 } else if (ret) {
4180 break;
4181 }
4182 }
4183 out:
4184 while (!list_empty(&nocow_ctx->inodes)) {
4185 struct scrub_nocow_inode *entry;
4186 entry = list_first_entry(&nocow_ctx->inodes,
4187 struct scrub_nocow_inode,
4188 list);
4189 list_del_init(&entry->list);
4190 kfree(entry);
4191 }
4192 if (trans && !IS_ERR(trans))
4193 btrfs_end_transaction(trans, root);
4194 if (not_written)
4195 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
4196 num_uncorrectable_read_errors);
4197
4198 btrfs_free_path(path);
4199 kfree(nocow_ctx);
4200
4201 scrub_pending_trans_workers_dec(sctx);
4202 }
4203
4204 static int check_extent_to_block(struct inode *inode, u64 start, u64 len,
4205 u64 logical)
4206 {
4207 struct extent_state *cached_state = NULL;
4208 struct btrfs_ordered_extent *ordered;
4209 struct extent_io_tree *io_tree;
4210 struct extent_map *em;
4211 u64 lockstart = start, lockend = start + len - 1;
4212 int ret = 0;
4213
4214 io_tree = &BTRFS_I(inode)->io_tree;
4215
4216 lock_extent_bits(io_tree, lockstart, lockend, &cached_state);
4217 ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
4218 if (ordered) {
4219 btrfs_put_ordered_extent(ordered);
4220 ret = 1;
4221 goto out_unlock;
4222 }
4223
4224 em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
4225 if (IS_ERR(em)) {
4226 ret = PTR_ERR(em);
4227 goto out_unlock;
4228 }
4229
4230 /*
4231 * This extent does not actually cover the logical extent anymore,
4232 * move on to the next inode.
4233 */
4234 if (em->block_start > logical ||
4235 em->block_start + em->block_len < logical + len) {
4236 free_extent_map(em);
4237 ret = 1;
4238 goto out_unlock;
4239 }
4240 free_extent_map(em);
4241
4242 out_unlock:
4243 unlock_extent_cached(io_tree, lockstart, lockend, &cached_state,
4244 GFP_NOFS);
4245 return ret;
4246 }
4247
4248 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
4249 struct scrub_copy_nocow_ctx *nocow_ctx)
4250 {
4251 struct btrfs_fs_info *fs_info = nocow_ctx->sctx->dev_root->fs_info;
4252 struct btrfs_key key;
4253 struct inode *inode;
4254 struct page *page;
4255 struct btrfs_root *local_root;
4256 struct extent_io_tree *io_tree;
4257 u64 physical_for_dev_replace;
4258 u64 nocow_ctx_logical;
4259 u64 len = nocow_ctx->len;
4260 unsigned long index;
4261 int srcu_index;
4262 int ret = 0;
4263 int err = 0;
4264
4265 key.objectid = root;
4266 key.type = BTRFS_ROOT_ITEM_KEY;
4267 key.offset = (u64)-1;
4268
4269 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
4270
4271 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
4272 if (IS_ERR(local_root)) {
4273 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4274 return PTR_ERR(local_root);
4275 }
4276
4277 key.type = BTRFS_INODE_ITEM_KEY;
4278 key.objectid = inum;
4279 key.offset = 0;
4280 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
4281 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4282 if (IS_ERR(inode))
4283 return PTR_ERR(inode);
4284
4285 /* Avoid truncate/dio/punch hole.. */
4286 inode_lock(inode);
4287 inode_dio_wait(inode);
4288
4289 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4290 io_tree = &BTRFS_I(inode)->io_tree;
4291 nocow_ctx_logical = nocow_ctx->logical;
4292
4293 ret = check_extent_to_block(inode, offset, len, nocow_ctx_logical);
4294 if (ret) {
4295 ret = ret > 0 ? 0 : ret;
4296 goto out;
4297 }
4298
4299 while (len >= PAGE_SIZE) {
4300 index = offset >> PAGE_SHIFT;
4301 again:
4302 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
4303 if (!page) {
4304 btrfs_err(fs_info, "find_or_create_page() failed");
4305 ret = -ENOMEM;
4306 goto out;
4307 }
4308
4309 if (PageUptodate(page)) {
4310 if (PageDirty(page))
4311 goto next_page;
4312 } else {
4313 ClearPageError(page);
4314 err = extent_read_full_page(io_tree, page,
4315 btrfs_get_extent,
4316 nocow_ctx->mirror_num);
4317 if (err) {
4318 ret = err;
4319 goto next_page;
4320 }
4321
4322 lock_page(page);
4323 /*
4324 * If the page has been remove from the page cache,
4325 * the data on it is meaningless, because it may be
4326 * old one, the new data may be written into the new
4327 * page in the page cache.
4328 */
4329 if (page->mapping != inode->i_mapping) {
4330 unlock_page(page);
4331 put_page(page);
4332 goto again;
4333 }
4334 if (!PageUptodate(page)) {
4335 ret = -EIO;
4336 goto next_page;
4337 }
4338 }
4339
4340 ret = check_extent_to_block(inode, offset, len,
4341 nocow_ctx_logical);
4342 if (ret) {
4343 ret = ret > 0 ? 0 : ret;
4344 goto next_page;
4345 }
4346
4347 err = write_page_nocow(nocow_ctx->sctx,
4348 physical_for_dev_replace, page);
4349 if (err)
4350 ret = err;
4351 next_page:
4352 unlock_page(page);
4353 put_page(page);
4354
4355 if (ret)
4356 break;
4357
4358 offset += PAGE_SIZE;
4359 physical_for_dev_replace += PAGE_SIZE;
4360 nocow_ctx_logical += PAGE_SIZE;
4361 len -= PAGE_SIZE;
4362 }
4363 ret = COPY_COMPLETE;
4364 out:
4365 inode_unlock(inode);
4366 iput(inode);
4367 return ret;
4368 }
4369
4370 static int write_page_nocow(struct scrub_ctx *sctx,
4371 u64 physical_for_dev_replace, struct page *page)
4372 {
4373 struct bio *bio;
4374 struct btrfs_device *dev;
4375 int ret;
4376
4377 dev = sctx->wr_ctx.tgtdev;
4378 if (!dev)
4379 return -EIO;
4380 if (!dev->bdev) {
4381 btrfs_warn_rl(dev->dev_root->fs_info,
4382 "scrub write_page_nocow(bdev == NULL) is unexpected");
4383 return -EIO;
4384 }
4385 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
4386 if (!bio) {
4387 spin_lock(&sctx->stat_lock);
4388 sctx->stat.malloc_errors++;
4389 spin_unlock(&sctx->stat_lock);
4390 return -ENOMEM;
4391 }
4392 bio->bi_iter.bi_size = 0;
4393 bio->bi_iter.bi_sector = physical_for_dev_replace >> 9;
4394 bio->bi_bdev = dev->bdev;
4395 ret = bio_add_page(bio, page, PAGE_SIZE, 0);
4396 if (ret != PAGE_SIZE) {
4397 leave_with_eio:
4398 bio_put(bio);
4399 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
4400 return -EIO;
4401 }
4402
4403 if (btrfsic_submit_bio_wait(WRITE_SYNC, bio))
4404 goto leave_with_eio;
4405
4406 bio_put(bio);
4407 return 0;
4408 }
Linux with added FPC-III support