]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - fs/btrfs/scrub.c
projects / mirror_ubuntu-bionic-kernel.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
btrfs: Wait for in-flight bios before freeing target device for raid56
[mirror_ubuntu-bionic-kernel.git] / fs / btrfs / scrub.c
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 refcount_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 refcount_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 refcount_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_fs_info *fs_info;
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 refcount_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_wr_ctx *wr_ctx,
286 struct btrfs_device *dev,
287 int is_dev_replace);
288 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx);
289 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
290 struct scrub_page *spage);
291 static void scrub_wr_submit(struct scrub_ctx *sctx);
292 static void scrub_wr_bio_end_io(struct bio *bio);
293 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
294 static int write_page_nocow(struct scrub_ctx *sctx,
295 u64 physical_for_dev_replace, struct page *page);
296 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
297 struct scrub_copy_nocow_ctx *ctx);
298 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
299 int mirror_num, u64 physical_for_dev_replace);
300 static void copy_nocow_pages_worker(struct btrfs_work *work);
301 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
302 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
303 static void scrub_put_ctx(struct scrub_ctx *sctx);
304
305
306 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
307 {
308 refcount_inc(&sctx->refs);
309 atomic_inc(&sctx->bios_in_flight);
310 }
311
312 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
313 {
314 atomic_dec(&sctx->bios_in_flight);
315 wake_up(&sctx->list_wait);
316 scrub_put_ctx(sctx);
317 }
318
319 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
320 {
321 while (atomic_read(&fs_info->scrub_pause_req)) {
322 mutex_unlock(&fs_info->scrub_lock);
323 wait_event(fs_info->scrub_pause_wait,
324 atomic_read(&fs_info->scrub_pause_req) == 0);
325 mutex_lock(&fs_info->scrub_lock);
326 }
327 }
328
329 static void scrub_pause_on(struct btrfs_fs_info *fs_info)
330 {
331 atomic_inc(&fs_info->scrubs_paused);
332 wake_up(&fs_info->scrub_pause_wait);
333 }
334
335 static void scrub_pause_off(struct btrfs_fs_info *fs_info)
336 {
337 mutex_lock(&fs_info->scrub_lock);
338 __scrub_blocked_if_needed(fs_info);
339 atomic_dec(&fs_info->scrubs_paused);
340 mutex_unlock(&fs_info->scrub_lock);
341
342 wake_up(&fs_info->scrub_pause_wait);
343 }
344
345 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
346 {
347 scrub_pause_on(fs_info);
348 scrub_pause_off(fs_info);
349 }
350
351 /*
352 * used for workers that require transaction commits (i.e., for the
353 * NOCOW case)
354 */
355 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
356 {
357 struct btrfs_fs_info *fs_info = sctx->fs_info;
358
359 refcount_inc(&sctx->refs);
360 /*
361 * increment scrubs_running to prevent cancel requests from
362 * completing as long as a worker is running. we must also
363 * increment scrubs_paused to prevent deadlocking on pause
364 * requests used for transactions commits (as the worker uses a
365 * transaction context). it is safe to regard the worker
366 * as paused for all matters practical. effectively, we only
367 * avoid cancellation requests from completing.
368 */
369 mutex_lock(&fs_info->scrub_lock);
370 atomic_inc(&fs_info->scrubs_running);
371 atomic_inc(&fs_info->scrubs_paused);
372 mutex_unlock(&fs_info->scrub_lock);
373
374 /*
375 * check if @scrubs_running=@scrubs_paused condition
376 * inside wait_event() is not an atomic operation.
377 * which means we may inc/dec @scrub_running/paused
378 * at any time. Let's wake up @scrub_pause_wait as
379 * much as we can to let commit transaction blocked less.
380 */
381 wake_up(&fs_info->scrub_pause_wait);
382
383 atomic_inc(&sctx->workers_pending);
384 }
385
386 /* used for workers that require transaction commits */
387 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
388 {
389 struct btrfs_fs_info *fs_info = sctx->fs_info;
390
391 /*
392 * see scrub_pending_trans_workers_inc() why we're pretending
393 * to be paused in the scrub counters
394 */
395 mutex_lock(&fs_info->scrub_lock);
396 atomic_dec(&fs_info->scrubs_running);
397 atomic_dec(&fs_info->scrubs_paused);
398 mutex_unlock(&fs_info->scrub_lock);
399 atomic_dec(&sctx->workers_pending);
400 wake_up(&fs_info->scrub_pause_wait);
401 wake_up(&sctx->list_wait);
402 scrub_put_ctx(sctx);
403 }
404
405 static void scrub_free_csums(struct scrub_ctx *sctx)
406 {
407 while (!list_empty(&sctx->csum_list)) {
408 struct btrfs_ordered_sum *sum;
409 sum = list_first_entry(&sctx->csum_list,
410 struct btrfs_ordered_sum, list);
411 list_del(&sum->list);
412 kfree(sum);
413 }
414 }
415
416 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
417 {
418 int i;
419
420 if (!sctx)
421 return;
422
423 scrub_free_wr_ctx(&sctx->wr_ctx);
424
425 /* this can happen when scrub is cancelled */
426 if (sctx->curr != -1) {
427 struct scrub_bio *sbio = sctx->bios[sctx->curr];
428
429 for (i = 0; i < sbio->page_count; i++) {
430 WARN_ON(!sbio->pagev[i]->page);
431 scrub_block_put(sbio->pagev[i]->sblock);
432 }
433 bio_put(sbio->bio);
434 }
435
436 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
437 struct scrub_bio *sbio = sctx->bios[i];
438
439 if (!sbio)
440 break;
441 kfree(sbio);
442 }
443
444 scrub_free_csums(sctx);
445 kfree(sctx);
446 }
447
448 static void scrub_put_ctx(struct scrub_ctx *sctx)
449 {
450 if (refcount_dec_and_test(&sctx->refs))
451 scrub_free_ctx(sctx);
452 }
453
454 static noinline_for_stack
455 struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
456 {
457 struct scrub_ctx *sctx;
458 int i;
459 struct btrfs_fs_info *fs_info = dev->fs_info;
460 int ret;
461
462 sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
463 if (!sctx)
464 goto nomem;
465 refcount_set(&sctx->refs, 1);
466 sctx->is_dev_replace = is_dev_replace;
467 sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
468 sctx->curr = -1;
469 sctx->fs_info = dev->fs_info;
470 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
471 struct scrub_bio *sbio;
472
473 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
474 if (!sbio)
475 goto nomem;
476 sctx->bios[i] = sbio;
477
478 sbio->index = i;
479 sbio->sctx = sctx;
480 sbio->page_count = 0;
481 btrfs_init_work(&sbio->work, btrfs_scrub_helper,
482 scrub_bio_end_io_worker, NULL, NULL);
483
484 if (i != SCRUB_BIOS_PER_SCTX - 1)
485 sctx->bios[i]->next_free = i + 1;
486 else
487 sctx->bios[i]->next_free = -1;
488 }
489 sctx->first_free = 0;
490 sctx->nodesize = fs_info->nodesize;
491 sctx->sectorsize = fs_info->sectorsize;
492 atomic_set(&sctx->bios_in_flight, 0);
493 atomic_set(&sctx->workers_pending, 0);
494 atomic_set(&sctx->cancel_req, 0);
495 sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
496 INIT_LIST_HEAD(&sctx->csum_list);
497
498 spin_lock_init(&sctx->list_lock);
499 spin_lock_init(&sctx->stat_lock);
500 init_waitqueue_head(&sctx->list_wait);
501
502 ret = scrub_setup_wr_ctx(&sctx->wr_ctx,
503 fs_info->dev_replace.tgtdev, is_dev_replace);
504 if (ret) {
505 scrub_free_ctx(sctx);
506 return ERR_PTR(ret);
507 }
508 return sctx;
509
510 nomem:
511 scrub_free_ctx(sctx);
512 return ERR_PTR(-ENOMEM);
513 }
514
515 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
516 void *warn_ctx)
517 {
518 u64 isize;
519 u32 nlink;
520 int ret;
521 int i;
522 struct extent_buffer *eb;
523 struct btrfs_inode_item *inode_item;
524 struct scrub_warning *swarn = warn_ctx;
525 struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
526 struct inode_fs_paths *ipath = NULL;
527 struct btrfs_root *local_root;
528 struct btrfs_key root_key;
529 struct btrfs_key key;
530
531 root_key.objectid = root;
532 root_key.type = BTRFS_ROOT_ITEM_KEY;
533 root_key.offset = (u64)-1;
534 local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
535 if (IS_ERR(local_root)) {
536 ret = PTR_ERR(local_root);
537 goto err;
538 }
539
540 /*
541 * this makes the path point to (inum INODE_ITEM ioff)
542 */
543 key.objectid = inum;
544 key.type = BTRFS_INODE_ITEM_KEY;
545 key.offset = 0;
546
547 ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
548 if (ret) {
549 btrfs_release_path(swarn->path);
550 goto err;
551 }
552
553 eb = swarn->path->nodes[0];
554 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
555 struct btrfs_inode_item);
556 isize = btrfs_inode_size(eb, inode_item);
557 nlink = btrfs_inode_nlink(eb, inode_item);
558 btrfs_release_path(swarn->path);
559
560 ipath = init_ipath(4096, local_root, swarn->path);
561 if (IS_ERR(ipath)) {
562 ret = PTR_ERR(ipath);
563 ipath = NULL;
564 goto err;
565 }
566 ret = paths_from_inode(inum, ipath);
567
568 if (ret < 0)
569 goto err;
570
571 /*
572 * we deliberately ignore the bit ipath might have been too small to
573 * hold all of the paths here
574 */
575 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
576 btrfs_warn_in_rcu(fs_info,
577 "%s at logical %llu on dev %s, sector %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
578 swarn->errstr, swarn->logical,
579 rcu_str_deref(swarn->dev->name),
580 (unsigned long long)swarn->sector,
581 root, inum, offset,
582 min(isize - offset, (u64)PAGE_SIZE), nlink,
583 (char *)(unsigned long)ipath->fspath->val[i]);
584
585 free_ipath(ipath);
586 return 0;
587
588 err:
589 btrfs_warn_in_rcu(fs_info,
590 "%s at logical %llu on dev %s, sector %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
591 swarn->errstr, swarn->logical,
592 rcu_str_deref(swarn->dev->name),
593 (unsigned long long)swarn->sector,
594 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->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, sector %llu: metadata %s (level %d) in tree %llu",
649 errstr, swarn.logical,
650 rcu_str_deref(dev->name),
651 (unsigned long long)swarn.sector,
652 ref_level ? "node" : "leaf",
653 ret < 0 ? -1 : ref_level,
654 ret < 0 ? -1 : ref_root);
655 } while (ret != 1);
656 btrfs_release_path(path);
657 } else {
658 btrfs_release_path(path);
659 swarn.path = path;
660 swarn.dev = dev;
661 iterate_extent_inodes(fs_info, found_key.objectid,
662 extent_item_pos, 1,
663 scrub_print_warning_inode, &swarn);
664 }
665
666 out:
667 btrfs_free_path(path);
668 }
669
670 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
671 {
672 struct page *page = NULL;
673 unsigned long index;
674 struct scrub_fixup_nodatasum *fixup = fixup_ctx;
675 int ret;
676 int corrected = 0;
677 struct btrfs_key key;
678 struct inode *inode = NULL;
679 struct btrfs_fs_info *fs_info;
680 u64 end = offset + PAGE_SIZE - 1;
681 struct btrfs_root *local_root;
682 int srcu_index;
683
684 key.objectid = root;
685 key.type = BTRFS_ROOT_ITEM_KEY;
686 key.offset = (u64)-1;
687
688 fs_info = fixup->root->fs_info;
689 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
690
691 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
692 if (IS_ERR(local_root)) {
693 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
694 return PTR_ERR(local_root);
695 }
696
697 key.type = BTRFS_INODE_ITEM_KEY;
698 key.objectid = inum;
699 key.offset = 0;
700 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
701 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
702 if (IS_ERR(inode))
703 return PTR_ERR(inode);
704
705 index = offset >> PAGE_SHIFT;
706
707 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
708 if (!page) {
709 ret = -ENOMEM;
710 goto out;
711 }
712
713 if (PageUptodate(page)) {
714 if (PageDirty(page)) {
715 /*
716 * we need to write the data to the defect sector. the
717 * data that was in that sector is not in memory,
718 * because the page was modified. we must not write the
719 * modified page to that sector.
720 *
721 * TODO: what could be done here: wait for the delalloc
722 * runner to write out that page (might involve
723 * COW) and see whether the sector is still
724 * referenced afterwards.
725 *
726 * For the meantime, we'll treat this error
727 * incorrectable, although there is a chance that a
728 * later scrub will find the bad sector again and that
729 * there's no dirty page in memory, then.
730 */
731 ret = -EIO;
732 goto out;
733 }
734 ret = repair_io_failure(BTRFS_I(inode), offset, PAGE_SIZE,
735 fixup->logical, page,
736 offset - page_offset(page),
737 fixup->mirror_num);
738 unlock_page(page);
739 corrected = !ret;
740 } else {
741 /*
742 * we need to get good data first. the general readpage path
743 * will call repair_io_failure for us, we just have to make
744 * sure we read the bad mirror.
745 */
746 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
747 EXTENT_DAMAGED);
748 if (ret) {
749 /* set_extent_bits should give proper error */
750 WARN_ON(ret > 0);
751 if (ret > 0)
752 ret = -EFAULT;
753 goto out;
754 }
755
756 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
757 btrfs_get_extent,
758 fixup->mirror_num);
759 wait_on_page_locked(page);
760
761 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
762 end, EXTENT_DAMAGED, 0, NULL);
763 if (!corrected)
764 clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
765 EXTENT_DAMAGED);
766 }
767
768 out:
769 if (page)
770 put_page(page);
771
772 iput(inode);
773
774 if (ret < 0)
775 return ret;
776
777 if (ret == 0 && corrected) {
778 /*
779 * we only need to call readpage for one of the inodes belonging
780 * to this extent. so make iterate_extent_inodes stop
781 */
782 return 1;
783 }
784
785 return -EIO;
786 }
787
788 static void scrub_fixup_nodatasum(struct btrfs_work *work)
789 {
790 struct btrfs_fs_info *fs_info;
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 fs_info = fixup->root->fs_info;
801
802 path = btrfs_alloc_path();
803 if (!path) {
804 spin_lock(&sctx->stat_lock);
805 ++sctx->stat.malloc_errors;
806 spin_unlock(&sctx->stat_lock);
807 uncorrectable = 1;
808 goto out;
809 }
810
811 trans = btrfs_join_transaction(fixup->root);
812 if (IS_ERR(trans)) {
813 uncorrectable = 1;
814 goto out;
815 }
816
817 /*
818 * the idea is to trigger a regular read through the standard path. we
819 * read a page from the (failed) logical address by specifying the
820 * corresponding copynum of the failed sector. thus, that readpage is
821 * expected to fail.
822 * that is the point where on-the-fly error correction will kick in
823 * (once it's finished) and rewrite the failed sector if a good copy
824 * can be found.
825 */
826 ret = iterate_inodes_from_logical(fixup->logical, fs_info, path,
827 scrub_fixup_readpage, 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);
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 &fs_info->dev_replace.num_uncorrectable_read_errors);
847 btrfs_err_rl_in_rcu(fs_info,
848 "unable to fixup (nodatasum) error at logical %llu on dev %s",
849 fixup->logical, rcu_str_deref(fixup->dev->name));
850 }
851
852 btrfs_free_path(path);
853 kfree(fixup);
854
855 scrub_pending_trans_workers_dec(sctx);
856 }
857
858 static inline void scrub_get_recover(struct scrub_recover *recover)
859 {
860 refcount_inc(&recover->refs);
861 }
862
863 static inline void scrub_put_recover(struct scrub_recover *recover)
864 {
865 if (refcount_dec_and_test(&recover->refs)) {
866 btrfs_put_bbio(recover->bbio);
867 kfree(recover);
868 }
869 }
870
871 /*
872 * scrub_handle_errored_block gets called when either verification of the
873 * pages failed or the bio failed to read, e.g. with EIO. In the latter
874 * case, this function handles all pages in the bio, even though only one
875 * may be bad.
876 * The goal of this function is to repair the errored block by using the
877 * contents of one of the mirrors.
878 */
879 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
880 {
881 struct scrub_ctx *sctx = sblock_to_check->sctx;
882 struct btrfs_device *dev;
883 struct btrfs_fs_info *fs_info;
884 u64 length;
885 u64 logical;
886 unsigned int failed_mirror_index;
887 unsigned int is_metadata;
888 unsigned int have_csum;
889 struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
890 struct scrub_block *sblock_bad;
891 int ret;
892 int mirror_index;
893 int page_num;
894 int success;
895 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
896 DEFAULT_RATELIMIT_BURST);
897
898 BUG_ON(sblock_to_check->page_count < 1);
899 fs_info = sctx->fs_info;
900 if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
901 /*
902 * if we find an error in a super block, we just report it.
903 * They will get written with the next transaction commit
904 * anyway
905 */
906 spin_lock(&sctx->stat_lock);
907 ++sctx->stat.super_errors;
908 spin_unlock(&sctx->stat_lock);
909 return 0;
910 }
911 length = sblock_to_check->page_count * PAGE_SIZE;
912 logical = sblock_to_check->pagev[0]->logical;
913 BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
914 failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
915 is_metadata = !(sblock_to_check->pagev[0]->flags &
916 BTRFS_EXTENT_FLAG_DATA);
917 have_csum = sblock_to_check->pagev[0]->have_csum;
918 dev = sblock_to_check->pagev[0]->dev;
919
920 if (sctx->is_dev_replace && !is_metadata && !have_csum) {
921 sblocks_for_recheck = NULL;
922 goto nodatasum_case;
923 }
924
925 /*
926 * read all mirrors one after the other. This includes to
927 * re-read the extent or metadata block that failed (that was
928 * the cause that this fixup code is called) another time,
929 * page by page this time in order to know which pages
930 * caused I/O errors and which ones are good (for all mirrors).
931 * It is the goal to handle the situation when more than one
932 * mirror contains I/O errors, but the errors do not
933 * overlap, i.e. the data can be repaired by selecting the
934 * pages from those mirrors without I/O error on the
935 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
936 * would be that mirror #1 has an I/O error on the first page,
937 * the second page is good, and mirror #2 has an I/O error on
938 * the second page, but the first page is good.
939 * Then the first page of the first mirror can be repaired by
940 * taking the first page of the second mirror, and the
941 * second page of the second mirror can be repaired by
942 * copying the contents of the 2nd page of the 1st mirror.
943 * One more note: if the pages of one mirror contain I/O
944 * errors, the checksum cannot be verified. In order to get
945 * the best data for repairing, the first attempt is to find
946 * a mirror without I/O errors and with a validated checksum.
947 * Only if this is not possible, the pages are picked from
948 * mirrors with I/O errors without considering the checksum.
949 * If the latter is the case, at the end, the checksum of the
950 * repaired area is verified in order to correctly maintain
951 * the statistics.
952 */
953
954 sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
955 sizeof(*sblocks_for_recheck), GFP_NOFS);
956 if (!sblocks_for_recheck) {
957 spin_lock(&sctx->stat_lock);
958 sctx->stat.malloc_errors++;
959 sctx->stat.read_errors++;
960 sctx->stat.uncorrectable_errors++;
961 spin_unlock(&sctx->stat_lock);
962 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
963 goto out;
964 }
965
966 /* setup the context, map the logical blocks and alloc the pages */
967 ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
968 if (ret) {
969 spin_lock(&sctx->stat_lock);
970 sctx->stat.read_errors++;
971 sctx->stat.uncorrectable_errors++;
972 spin_unlock(&sctx->stat_lock);
973 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
974 goto out;
975 }
976 BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
977 sblock_bad = sblocks_for_recheck + failed_mirror_index;
978
979 /* build and submit the bios for the failed mirror, check checksums */
980 scrub_recheck_block(fs_info, sblock_bad, 1);
981
982 if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
983 sblock_bad->no_io_error_seen) {
984 /*
985 * the error disappeared after reading page by page, or
986 * the area was part of a huge bio and other parts of the
987 * bio caused I/O errors, or the block layer merged several
988 * read requests into one and the error is caused by a
989 * different bio (usually one of the two latter cases is
990 * the cause)
991 */
992 spin_lock(&sctx->stat_lock);
993 sctx->stat.unverified_errors++;
994 sblock_to_check->data_corrected = 1;
995 spin_unlock(&sctx->stat_lock);
996
997 if (sctx->is_dev_replace)
998 scrub_write_block_to_dev_replace(sblock_bad);
999 goto out;
1000 }
1001
1002 if (!sblock_bad->no_io_error_seen) {
1003 spin_lock(&sctx->stat_lock);
1004 sctx->stat.read_errors++;
1005 spin_unlock(&sctx->stat_lock);
1006 if (__ratelimit(&_rs))
1007 scrub_print_warning("i/o error", sblock_to_check);
1008 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1009 } else if (sblock_bad->checksum_error) {
1010 spin_lock(&sctx->stat_lock);
1011 sctx->stat.csum_errors++;
1012 spin_unlock(&sctx->stat_lock);
1013 if (__ratelimit(&_rs))
1014 scrub_print_warning("checksum error", sblock_to_check);
1015 btrfs_dev_stat_inc_and_print(dev,
1016 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1017 } else if (sblock_bad->header_error) {
1018 spin_lock(&sctx->stat_lock);
1019 sctx->stat.verify_errors++;
1020 spin_unlock(&sctx->stat_lock);
1021 if (__ratelimit(&_rs))
1022 scrub_print_warning("checksum/header error",
1023 sblock_to_check);
1024 if (sblock_bad->generation_error)
1025 btrfs_dev_stat_inc_and_print(dev,
1026 BTRFS_DEV_STAT_GENERATION_ERRS);
1027 else
1028 btrfs_dev_stat_inc_and_print(dev,
1029 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1030 }
1031
1032 if (sctx->readonly) {
1033 ASSERT(!sctx->is_dev_replace);
1034 goto out;
1035 }
1036
1037 if (!is_metadata && !have_csum) {
1038 struct scrub_fixup_nodatasum *fixup_nodatasum;
1039
1040 WARN_ON(sctx->is_dev_replace);
1041
1042 nodatasum_case:
1043
1044 /*
1045 * !is_metadata and !have_csum, this means that the data
1046 * might not be COWed, that it might be modified
1047 * concurrently. The general strategy to work on the
1048 * commit root does not help in the case when COW is not
1049 * used.
1050 */
1051 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
1052 if (!fixup_nodatasum)
1053 goto did_not_correct_error;
1054 fixup_nodatasum->sctx = sctx;
1055 fixup_nodatasum->dev = dev;
1056 fixup_nodatasum->logical = logical;
1057 fixup_nodatasum->root = fs_info->extent_root;
1058 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
1059 scrub_pending_trans_workers_inc(sctx);
1060 btrfs_init_work(&fixup_nodatasum->work, btrfs_scrub_helper,
1061 scrub_fixup_nodatasum, NULL, NULL);
1062 btrfs_queue_work(fs_info->scrub_workers,
1063 &fixup_nodatasum->work);
1064 goto out;
1065 }
1066
1067 /*
1068 * now build and submit the bios for the other mirrors, check
1069 * checksums.
1070 * First try to pick the mirror which is completely without I/O
1071 * errors and also does not have a checksum error.
1072 * If one is found, and if a checksum is present, the full block
1073 * that is known to contain an error is rewritten. Afterwards
1074 * the block is known to be corrected.
1075 * If a mirror is found which is completely correct, and no
1076 * checksum is present, only those pages are rewritten that had
1077 * an I/O error in the block to be repaired, since it cannot be
1078 * determined, which copy of the other pages is better (and it
1079 * could happen otherwise that a correct page would be
1080 * overwritten by a bad one).
1081 */
1082 for (mirror_index = 0;
1083 mirror_index < BTRFS_MAX_MIRRORS &&
1084 sblocks_for_recheck[mirror_index].page_count > 0;
1085 mirror_index++) {
1086 struct scrub_block *sblock_other;
1087
1088 if (mirror_index == failed_mirror_index)
1089 continue;
1090 sblock_other = sblocks_for_recheck + mirror_index;
1091
1092 /* build and submit the bios, check checksums */
1093 scrub_recheck_block(fs_info, sblock_other, 0);
1094
1095 if (!sblock_other->header_error &&
1096 !sblock_other->checksum_error &&
1097 sblock_other->no_io_error_seen) {
1098 if (sctx->is_dev_replace) {
1099 scrub_write_block_to_dev_replace(sblock_other);
1100 goto corrected_error;
1101 } else {
1102 ret = scrub_repair_block_from_good_copy(
1103 sblock_bad, sblock_other);
1104 if (!ret)
1105 goto corrected_error;
1106 }
1107 }
1108 }
1109
1110 if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1111 goto did_not_correct_error;
1112
1113 /*
1114 * In case of I/O errors in the area that is supposed to be
1115 * repaired, continue by picking good copies of those pages.
1116 * Select the good pages from mirrors to rewrite bad pages from
1117 * the area to fix. Afterwards verify the checksum of the block
1118 * that is supposed to be repaired. This verification step is
1119 * only done for the purpose of statistic counting and for the
1120 * final scrub report, whether errors remain.
1121 * A perfect algorithm could make use of the checksum and try
1122 * all possible combinations of pages from the different mirrors
1123 * until the checksum verification succeeds. For example, when
1124 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1125 * of mirror #2 is readable but the final checksum test fails,
1126 * then the 2nd page of mirror #3 could be tried, whether now
1127 * the final checksum succeeds. But this would be a rare
1128 * exception and is therefore not implemented. At least it is
1129 * avoided that the good copy is overwritten.
1130 * A more useful improvement would be to pick the sectors
1131 * without I/O error based on sector sizes (512 bytes on legacy
1132 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1133 * mirror could be repaired by taking 512 byte of a different
1134 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1135 * area are unreadable.
1136 */
1137 success = 1;
1138 for (page_num = 0; page_num < sblock_bad->page_count;
1139 page_num++) {
1140 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1141 struct scrub_block *sblock_other = NULL;
1142
1143 /* skip no-io-error page in scrub */
1144 if (!page_bad->io_error && !sctx->is_dev_replace)
1145 continue;
1146
1147 /* try to find no-io-error page in mirrors */
1148 if (page_bad->io_error) {
1149 for (mirror_index = 0;
1150 mirror_index < BTRFS_MAX_MIRRORS &&
1151 sblocks_for_recheck[mirror_index].page_count > 0;
1152 mirror_index++) {
1153 if (!sblocks_for_recheck[mirror_index].
1154 pagev[page_num]->io_error) {
1155 sblock_other = sblocks_for_recheck +
1156 mirror_index;
1157 break;
1158 }
1159 }
1160 if (!sblock_other)
1161 success = 0;
1162 }
1163
1164 if (sctx->is_dev_replace) {
1165 /*
1166 * did not find a mirror to fetch the page
1167 * from. scrub_write_page_to_dev_replace()
1168 * handles this case (page->io_error), by
1169 * filling the block with zeros before
1170 * submitting the write request
1171 */
1172 if (!sblock_other)
1173 sblock_other = sblock_bad;
1174
1175 if (scrub_write_page_to_dev_replace(sblock_other,
1176 page_num) != 0) {
1177 btrfs_dev_replace_stats_inc(
1178 &fs_info->dev_replace.num_write_errors);
1179 success = 0;
1180 }
1181 } else if (sblock_other) {
1182 ret = scrub_repair_page_from_good_copy(sblock_bad,
1183 sblock_other,
1184 page_num, 0);
1185 if (0 == ret)
1186 page_bad->io_error = 0;
1187 else
1188 success = 0;
1189 }
1190 }
1191
1192 if (success && !sctx->is_dev_replace) {
1193 if (is_metadata || have_csum) {
1194 /*
1195 * need to verify the checksum now that all
1196 * sectors on disk are repaired (the write
1197 * request for data to be repaired is on its way).
1198 * Just be lazy and use scrub_recheck_block()
1199 * which re-reads the data before the checksum
1200 * is verified, but most likely the data comes out
1201 * of the page cache.
1202 */
1203 scrub_recheck_block(fs_info, sblock_bad, 1);
1204 if (!sblock_bad->header_error &&
1205 !sblock_bad->checksum_error &&
1206 sblock_bad->no_io_error_seen)
1207 goto corrected_error;
1208 else
1209 goto did_not_correct_error;
1210 } else {
1211 corrected_error:
1212 spin_lock(&sctx->stat_lock);
1213 sctx->stat.corrected_errors++;
1214 sblock_to_check->data_corrected = 1;
1215 spin_unlock(&sctx->stat_lock);
1216 btrfs_err_rl_in_rcu(fs_info,
1217 "fixed up error at logical %llu on dev %s",
1218 logical, rcu_str_deref(dev->name));
1219 }
1220 } else {
1221 did_not_correct_error:
1222 spin_lock(&sctx->stat_lock);
1223 sctx->stat.uncorrectable_errors++;
1224 spin_unlock(&sctx->stat_lock);
1225 btrfs_err_rl_in_rcu(fs_info,
1226 "unable to fixup (regular) error at logical %llu on dev %s",
1227 logical, rcu_str_deref(dev->name));
1228 }
1229
1230 out:
1231 if (sblocks_for_recheck) {
1232 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1233 mirror_index++) {
1234 struct scrub_block *sblock = sblocks_for_recheck +
1235 mirror_index;
1236 struct scrub_recover *recover;
1237 int page_index;
1238
1239 for (page_index = 0; page_index < sblock->page_count;
1240 page_index++) {
1241 sblock->pagev[page_index]->sblock = NULL;
1242 recover = sblock->pagev[page_index]->recover;
1243 if (recover) {
1244 scrub_put_recover(recover);
1245 sblock->pagev[page_index]->recover =
1246 NULL;
1247 }
1248 scrub_page_put(sblock->pagev[page_index]);
1249 }
1250 }
1251 kfree(sblocks_for_recheck);
1252 }
1253
1254 return 0;
1255 }
1256
1257 static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1258 {
1259 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1260 return 2;
1261 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1262 return 3;
1263 else
1264 return (int)bbio->num_stripes;
1265 }
1266
1267 static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1268 u64 *raid_map,
1269 u64 mapped_length,
1270 int nstripes, int mirror,
1271 int *stripe_index,
1272 u64 *stripe_offset)
1273 {
1274 int i;
1275
1276 if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1277 /* RAID5/6 */
1278 for (i = 0; i < nstripes; i++) {
1279 if (raid_map[i] == RAID6_Q_STRIPE ||
1280 raid_map[i] == RAID5_P_STRIPE)
1281 continue;
1282
1283 if (logical >= raid_map[i] &&
1284 logical < raid_map[i] + mapped_length)
1285 break;
1286 }
1287
1288 *stripe_index = i;
1289 *stripe_offset = logical - raid_map[i];
1290 } else {
1291 /* The other RAID type */
1292 *stripe_index = mirror;
1293 *stripe_offset = 0;
1294 }
1295 }
1296
1297 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1298 struct scrub_block *sblocks_for_recheck)
1299 {
1300 struct scrub_ctx *sctx = original_sblock->sctx;
1301 struct btrfs_fs_info *fs_info = sctx->fs_info;
1302 u64 length = original_sblock->page_count * PAGE_SIZE;
1303 u64 logical = original_sblock->pagev[0]->logical;
1304 u64 generation = original_sblock->pagev[0]->generation;
1305 u64 flags = original_sblock->pagev[0]->flags;
1306 u64 have_csum = original_sblock->pagev[0]->have_csum;
1307 struct scrub_recover *recover;
1308 struct btrfs_bio *bbio;
1309 u64 sublen;
1310 u64 mapped_length;
1311 u64 stripe_offset;
1312 int stripe_index;
1313 int page_index = 0;
1314 int mirror_index;
1315 int nmirrors;
1316 int ret;
1317
1318 /*
1319 * note: the two members refs and outstanding_pages
1320 * are not used (and not set) in the blocks that are used for
1321 * the recheck procedure
1322 */
1323
1324 while (length > 0) {
1325 sublen = min_t(u64, length, PAGE_SIZE);
1326 mapped_length = sublen;
1327 bbio = NULL;
1328
1329 /*
1330 * with a length of PAGE_SIZE, each returned stripe
1331 * represents one mirror
1332 */
1333 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1334 logical, &mapped_length, &bbio);
1335 if (ret || !bbio || mapped_length < sublen) {
1336 btrfs_put_bbio(bbio);
1337 return -EIO;
1338 }
1339
1340 recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1341 if (!recover) {
1342 btrfs_put_bbio(bbio);
1343 return -ENOMEM;
1344 }
1345
1346 refcount_set(&recover->refs, 1);
1347 recover->bbio = bbio;
1348 recover->map_length = mapped_length;
1349
1350 BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1351
1352 nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1353
1354 for (mirror_index = 0; mirror_index < nmirrors;
1355 mirror_index++) {
1356 struct scrub_block *sblock;
1357 struct scrub_page *page;
1358
1359 sblock = sblocks_for_recheck + mirror_index;
1360 sblock->sctx = sctx;
1361
1362 page = kzalloc(sizeof(*page), GFP_NOFS);
1363 if (!page) {
1364 leave_nomem:
1365 spin_lock(&sctx->stat_lock);
1366 sctx->stat.malloc_errors++;
1367 spin_unlock(&sctx->stat_lock);
1368 scrub_put_recover(recover);
1369 return -ENOMEM;
1370 }
1371 scrub_page_get(page);
1372 sblock->pagev[page_index] = page;
1373 page->sblock = sblock;
1374 page->flags = flags;
1375 page->generation = generation;
1376 page->logical = logical;
1377 page->have_csum = have_csum;
1378 if (have_csum)
1379 memcpy(page->csum,
1380 original_sblock->pagev[0]->csum,
1381 sctx->csum_size);
1382
1383 scrub_stripe_index_and_offset(logical,
1384 bbio->map_type,
1385 bbio->raid_map,
1386 mapped_length,
1387 bbio->num_stripes -
1388 bbio->num_tgtdevs,
1389 mirror_index,
1390 &stripe_index,
1391 &stripe_offset);
1392 page->physical = bbio->stripes[stripe_index].physical +
1393 stripe_offset;
1394 page->dev = bbio->stripes[stripe_index].dev;
1395
1396 BUG_ON(page_index >= original_sblock->page_count);
1397 page->physical_for_dev_replace =
1398 original_sblock->pagev[page_index]->
1399 physical_for_dev_replace;
1400 /* for missing devices, dev->bdev is NULL */
1401 page->mirror_num = mirror_index + 1;
1402 sblock->page_count++;
1403 page->page = alloc_page(GFP_NOFS);
1404 if (!page->page)
1405 goto leave_nomem;
1406
1407 scrub_get_recover(recover);
1408 page->recover = recover;
1409 }
1410 scrub_put_recover(recover);
1411 length -= sublen;
1412 logical += sublen;
1413 page_index++;
1414 }
1415
1416 return 0;
1417 }
1418
1419 struct scrub_bio_ret {
1420 struct completion event;
1421 int error;
1422 };
1423
1424 static void scrub_bio_wait_endio(struct bio *bio)
1425 {
1426 struct scrub_bio_ret *ret = bio->bi_private;
1427
1428 ret->error = bio->bi_error;
1429 complete(&ret->event);
1430 }
1431
1432 static inline int scrub_is_page_on_raid56(struct scrub_page *page)
1433 {
1434 return page->recover &&
1435 (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
1436 }
1437
1438 static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1439 struct bio *bio,
1440 struct scrub_page *page)
1441 {
1442 struct scrub_bio_ret done;
1443 int ret;
1444
1445 init_completion(&done.event);
1446 done.error = 0;
1447 bio->bi_iter.bi_sector = page->logical >> 9;
1448 bio->bi_private = &done;
1449 bio->bi_end_io = scrub_bio_wait_endio;
1450
1451 ret = raid56_parity_recover(fs_info, bio, page->recover->bbio,
1452 page->recover->map_length,
1453 page->mirror_num, 0);
1454 if (ret)
1455 return ret;
1456
1457 wait_for_completion(&done.event);
1458 if (done.error)
1459 return -EIO;
1460
1461 return 0;
1462 }
1463
1464 /*
1465 * this function will check the on disk data for checksum errors, header
1466 * errors and read I/O errors. If any I/O errors happen, the exact pages
1467 * which are errored are marked as being bad. The goal is to enable scrub
1468 * to take those pages that are not errored from all the mirrors so that
1469 * the pages that are errored in the just handled mirror can be repaired.
1470 */
1471 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1472 struct scrub_block *sblock,
1473 int retry_failed_mirror)
1474 {
1475 int page_num;
1476
1477 sblock->no_io_error_seen = 1;
1478
1479 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1480 struct bio *bio;
1481 struct scrub_page *page = sblock->pagev[page_num];
1482
1483 if (page->dev->bdev == NULL) {
1484 page->io_error = 1;
1485 sblock->no_io_error_seen = 0;
1486 continue;
1487 }
1488
1489 WARN_ON(!page->page);
1490 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1491 if (!bio) {
1492 page->io_error = 1;
1493 sblock->no_io_error_seen = 0;
1494 continue;
1495 }
1496 bio->bi_bdev = page->dev->bdev;
1497
1498 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1499 if (!retry_failed_mirror && scrub_is_page_on_raid56(page)) {
1500 if (scrub_submit_raid56_bio_wait(fs_info, bio, page)) {
1501 page->io_error = 1;
1502 sblock->no_io_error_seen = 0;
1503 }
1504 } else {
1505 bio->bi_iter.bi_sector = page->physical >> 9;
1506 bio_set_op_attrs(bio, REQ_OP_READ, 0);
1507
1508 if (btrfsic_submit_bio_wait(bio)) {
1509 page->io_error = 1;
1510 sblock->no_io_error_seen = 0;
1511 }
1512 }
1513
1514 bio_put(bio);
1515 }
1516
1517 if (sblock->no_io_error_seen)
1518 scrub_recheck_block_checksum(sblock);
1519 }
1520
1521 static inline int scrub_check_fsid(u8 fsid[],
1522 struct scrub_page *spage)
1523 {
1524 struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1525 int ret;
1526
1527 ret = memcmp(fsid, fs_devices->fsid, BTRFS_UUID_SIZE);
1528 return !ret;
1529 }
1530
1531 static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1532 {
1533 sblock->header_error = 0;
1534 sblock->checksum_error = 0;
1535 sblock->generation_error = 0;
1536
1537 if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1538 scrub_checksum_data(sblock);
1539 else
1540 scrub_checksum_tree_block(sblock);
1541 }
1542
1543 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1544 struct scrub_block *sblock_good)
1545 {
1546 int page_num;
1547 int ret = 0;
1548
1549 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1550 int ret_sub;
1551
1552 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1553 sblock_good,
1554 page_num, 1);
1555 if (ret_sub)
1556 ret = ret_sub;
1557 }
1558
1559 return ret;
1560 }
1561
1562 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1563 struct scrub_block *sblock_good,
1564 int page_num, int force_write)
1565 {
1566 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1567 struct scrub_page *page_good = sblock_good->pagev[page_num];
1568 struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1569
1570 BUG_ON(page_bad->page == NULL);
1571 BUG_ON(page_good->page == NULL);
1572 if (force_write || sblock_bad->header_error ||
1573 sblock_bad->checksum_error || page_bad->io_error) {
1574 struct bio *bio;
1575 int ret;
1576
1577 if (!page_bad->dev->bdev) {
1578 btrfs_warn_rl(fs_info,
1579 "scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1580 return -EIO;
1581 }
1582
1583 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1584 if (!bio)
1585 return -EIO;
1586 bio->bi_bdev = page_bad->dev->bdev;
1587 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1588 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1589
1590 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1591 if (PAGE_SIZE != ret) {
1592 bio_put(bio);
1593 return -EIO;
1594 }
1595
1596 if (btrfsic_submit_bio_wait(bio)) {
1597 btrfs_dev_stat_inc_and_print(page_bad->dev,
1598 BTRFS_DEV_STAT_WRITE_ERRS);
1599 btrfs_dev_replace_stats_inc(
1600 &fs_info->dev_replace.num_write_errors);
1601 bio_put(bio);
1602 return -EIO;
1603 }
1604 bio_put(bio);
1605 }
1606
1607 return 0;
1608 }
1609
1610 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1611 {
1612 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1613 int page_num;
1614
1615 /*
1616 * This block is used for the check of the parity on the source device,
1617 * so the data needn't be written into the destination device.
1618 */
1619 if (sblock->sparity)
1620 return;
1621
1622 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1623 int ret;
1624
1625 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1626 if (ret)
1627 btrfs_dev_replace_stats_inc(
1628 &fs_info->dev_replace.num_write_errors);
1629 }
1630 }
1631
1632 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1633 int page_num)
1634 {
1635 struct scrub_page *spage = sblock->pagev[page_num];
1636
1637 BUG_ON(spage->page == NULL);
1638 if (spage->io_error) {
1639 void *mapped_buffer = kmap_atomic(spage->page);
1640
1641 memset(mapped_buffer, 0, PAGE_SIZE);
1642 flush_dcache_page(spage->page);
1643 kunmap_atomic(mapped_buffer);
1644 }
1645 return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1646 }
1647
1648 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1649 struct scrub_page *spage)
1650 {
1651 struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1652 struct scrub_bio *sbio;
1653 int ret;
1654
1655 mutex_lock(&wr_ctx->wr_lock);
1656 again:
1657 if (!wr_ctx->wr_curr_bio) {
1658 wr_ctx->wr_curr_bio = kzalloc(sizeof(*wr_ctx->wr_curr_bio),
1659 GFP_KERNEL);
1660 if (!wr_ctx->wr_curr_bio) {
1661 mutex_unlock(&wr_ctx->wr_lock);
1662 return -ENOMEM;
1663 }
1664 wr_ctx->wr_curr_bio->sctx = sctx;
1665 wr_ctx->wr_curr_bio->page_count = 0;
1666 }
1667 sbio = wr_ctx->wr_curr_bio;
1668 if (sbio->page_count == 0) {
1669 struct bio *bio;
1670
1671 sbio->physical = spage->physical_for_dev_replace;
1672 sbio->logical = spage->logical;
1673 sbio->dev = wr_ctx->tgtdev;
1674 bio = sbio->bio;
1675 if (!bio) {
1676 bio = btrfs_io_bio_alloc(GFP_KERNEL,
1677 wr_ctx->pages_per_wr_bio);
1678 if (!bio) {
1679 mutex_unlock(&wr_ctx->wr_lock);
1680 return -ENOMEM;
1681 }
1682 sbio->bio = bio;
1683 }
1684
1685 bio->bi_private = sbio;
1686 bio->bi_end_io = scrub_wr_bio_end_io;
1687 bio->bi_bdev = sbio->dev->bdev;
1688 bio->bi_iter.bi_sector = sbio->physical >> 9;
1689 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1690 sbio->err = 0;
1691 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1692 spage->physical_for_dev_replace ||
1693 sbio->logical + sbio->page_count * PAGE_SIZE !=
1694 spage->logical) {
1695 scrub_wr_submit(sctx);
1696 goto again;
1697 }
1698
1699 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1700 if (ret != PAGE_SIZE) {
1701 if (sbio->page_count < 1) {
1702 bio_put(sbio->bio);
1703 sbio->bio = NULL;
1704 mutex_unlock(&wr_ctx->wr_lock);
1705 return -EIO;
1706 }
1707 scrub_wr_submit(sctx);
1708 goto again;
1709 }
1710
1711 sbio->pagev[sbio->page_count] = spage;
1712 scrub_page_get(spage);
1713 sbio->page_count++;
1714 if (sbio->page_count == wr_ctx->pages_per_wr_bio)
1715 scrub_wr_submit(sctx);
1716 mutex_unlock(&wr_ctx->wr_lock);
1717
1718 return 0;
1719 }
1720
1721 static void scrub_wr_submit(struct scrub_ctx *sctx)
1722 {
1723 struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1724 struct scrub_bio *sbio;
1725
1726 if (!wr_ctx->wr_curr_bio)
1727 return;
1728
1729 sbio = wr_ctx->wr_curr_bio;
1730 wr_ctx->wr_curr_bio = NULL;
1731 WARN_ON(!sbio->bio->bi_bdev);
1732 scrub_pending_bio_inc(sctx);
1733 /* process all writes in a single worker thread. Then the block layer
1734 * orders the requests before sending them to the driver which
1735 * doubled the write performance on spinning disks when measured
1736 * with Linux 3.5 */
1737 btrfsic_submit_bio(sbio->bio);
1738 }
1739
1740 static void scrub_wr_bio_end_io(struct bio *bio)
1741 {
1742 struct scrub_bio *sbio = bio->bi_private;
1743 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
1744
1745 sbio->err = bio->bi_error;
1746 sbio->bio = bio;
1747
1748 btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
1749 scrub_wr_bio_end_io_worker, NULL, NULL);
1750 btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1751 }
1752
1753 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1754 {
1755 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1756 struct scrub_ctx *sctx = sbio->sctx;
1757 int i;
1758
1759 WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1760 if (sbio->err) {
1761 struct btrfs_dev_replace *dev_replace =
1762 &sbio->sctx->fs_info->dev_replace;
1763
1764 for (i = 0; i < sbio->page_count; i++) {
1765 struct scrub_page *spage = sbio->pagev[i];
1766
1767 spage->io_error = 1;
1768 btrfs_dev_replace_stats_inc(&dev_replace->
1769 num_write_errors);
1770 }
1771 }
1772
1773 for (i = 0; i < sbio->page_count; i++)
1774 scrub_page_put(sbio->pagev[i]);
1775
1776 bio_put(sbio->bio);
1777 kfree(sbio);
1778 scrub_pending_bio_dec(sctx);
1779 }
1780
1781 static int scrub_checksum(struct scrub_block *sblock)
1782 {
1783 u64 flags;
1784 int ret;
1785
1786 /*
1787 * No need to initialize these stats currently,
1788 * because this function only use return value
1789 * instead of these stats value.
1790 *
1791 * Todo:
1792 * always use stats
1793 */
1794 sblock->header_error = 0;
1795 sblock->generation_error = 0;
1796 sblock->checksum_error = 0;
1797
1798 WARN_ON(sblock->page_count < 1);
1799 flags = sblock->pagev[0]->flags;
1800 ret = 0;
1801 if (flags & BTRFS_EXTENT_FLAG_DATA)
1802 ret = scrub_checksum_data(sblock);
1803 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1804 ret = scrub_checksum_tree_block(sblock);
1805 else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1806 (void)scrub_checksum_super(sblock);
1807 else
1808 WARN_ON(1);
1809 if (ret)
1810 scrub_handle_errored_block(sblock);
1811
1812 return ret;
1813 }
1814
1815 static int scrub_checksum_data(struct scrub_block *sblock)
1816 {
1817 struct scrub_ctx *sctx = sblock->sctx;
1818 u8 csum[BTRFS_CSUM_SIZE];
1819 u8 *on_disk_csum;
1820 struct page *page;
1821 void *buffer;
1822 u32 crc = ~(u32)0;
1823 u64 len;
1824 int index;
1825
1826 BUG_ON(sblock->page_count < 1);
1827 if (!sblock->pagev[0]->have_csum)
1828 return 0;
1829
1830 on_disk_csum = sblock->pagev[0]->csum;
1831 page = sblock->pagev[0]->page;
1832 buffer = kmap_atomic(page);
1833
1834 len = sctx->sectorsize;
1835 index = 0;
1836 for (;;) {
1837 u64 l = min_t(u64, len, PAGE_SIZE);
1838
1839 crc = btrfs_csum_data(buffer, crc, l);
1840 kunmap_atomic(buffer);
1841 len -= l;
1842 if (len == 0)
1843 break;
1844 index++;
1845 BUG_ON(index >= sblock->page_count);
1846 BUG_ON(!sblock->pagev[index]->page);
1847 page = sblock->pagev[index]->page;
1848 buffer = kmap_atomic(page);
1849 }
1850
1851 btrfs_csum_final(crc, csum);
1852 if (memcmp(csum, on_disk_csum, sctx->csum_size))
1853 sblock->checksum_error = 1;
1854
1855 return sblock->checksum_error;
1856 }
1857
1858 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1859 {
1860 struct scrub_ctx *sctx = sblock->sctx;
1861 struct btrfs_header *h;
1862 struct btrfs_fs_info *fs_info = sctx->fs_info;
1863 u8 calculated_csum[BTRFS_CSUM_SIZE];
1864 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1865 struct page *page;
1866 void *mapped_buffer;
1867 u64 mapped_size;
1868 void *p;
1869 u32 crc = ~(u32)0;
1870 u64 len;
1871 int index;
1872
1873 BUG_ON(sblock->page_count < 1);
1874 page = sblock->pagev[0]->page;
1875 mapped_buffer = kmap_atomic(page);
1876 h = (struct btrfs_header *)mapped_buffer;
1877 memcpy(on_disk_csum, h->csum, sctx->csum_size);
1878
1879 /*
1880 * we don't use the getter functions here, as we
1881 * a) don't have an extent buffer and
1882 * b) the page is already kmapped
1883 */
1884 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1885 sblock->header_error = 1;
1886
1887 if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
1888 sblock->header_error = 1;
1889 sblock->generation_error = 1;
1890 }
1891
1892 if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
1893 sblock->header_error = 1;
1894
1895 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1896 BTRFS_UUID_SIZE))
1897 sblock->header_error = 1;
1898
1899 len = sctx->nodesize - BTRFS_CSUM_SIZE;
1900 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1901 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1902 index = 0;
1903 for (;;) {
1904 u64 l = min_t(u64, len, mapped_size);
1905
1906 crc = btrfs_csum_data(p, crc, l);
1907 kunmap_atomic(mapped_buffer);
1908 len -= l;
1909 if (len == 0)
1910 break;
1911 index++;
1912 BUG_ON(index >= sblock->page_count);
1913 BUG_ON(!sblock->pagev[index]->page);
1914 page = sblock->pagev[index]->page;
1915 mapped_buffer = kmap_atomic(page);
1916 mapped_size = PAGE_SIZE;
1917 p = mapped_buffer;
1918 }
1919
1920 btrfs_csum_final(crc, calculated_csum);
1921 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1922 sblock->checksum_error = 1;
1923
1924 return sblock->header_error || sblock->checksum_error;
1925 }
1926
1927 static int scrub_checksum_super(struct scrub_block *sblock)
1928 {
1929 struct btrfs_super_block *s;
1930 struct scrub_ctx *sctx = sblock->sctx;
1931 u8 calculated_csum[BTRFS_CSUM_SIZE];
1932 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1933 struct page *page;
1934 void *mapped_buffer;
1935 u64 mapped_size;
1936 void *p;
1937 u32 crc = ~(u32)0;
1938 int fail_gen = 0;
1939 int fail_cor = 0;
1940 u64 len;
1941 int index;
1942
1943 BUG_ON(sblock->page_count < 1);
1944 page = sblock->pagev[0]->page;
1945 mapped_buffer = kmap_atomic(page);
1946 s = (struct btrfs_super_block *)mapped_buffer;
1947 memcpy(on_disk_csum, s->csum, sctx->csum_size);
1948
1949 if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1950 ++fail_cor;
1951
1952 if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1953 ++fail_gen;
1954
1955 if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
1956 ++fail_cor;
1957
1958 len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1959 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1960 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1961 index = 0;
1962 for (;;) {
1963 u64 l = min_t(u64, len, mapped_size);
1964
1965 crc = btrfs_csum_data(p, crc, l);
1966 kunmap_atomic(mapped_buffer);
1967 len -= l;
1968 if (len == 0)
1969 break;
1970 index++;
1971 BUG_ON(index >= sblock->page_count);
1972 BUG_ON(!sblock->pagev[index]->page);
1973 page = sblock->pagev[index]->page;
1974 mapped_buffer = kmap_atomic(page);
1975 mapped_size = PAGE_SIZE;
1976 p = mapped_buffer;
1977 }
1978
1979 btrfs_csum_final(crc, calculated_csum);
1980 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1981 ++fail_cor;
1982
1983 if (fail_cor + fail_gen) {
1984 /*
1985 * if we find an error in a super block, we just report it.
1986 * They will get written with the next transaction commit
1987 * anyway
1988 */
1989 spin_lock(&sctx->stat_lock);
1990 ++sctx->stat.super_errors;
1991 spin_unlock(&sctx->stat_lock);
1992 if (fail_cor)
1993 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1994 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1995 else
1996 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1997 BTRFS_DEV_STAT_GENERATION_ERRS);
1998 }
1999
2000 return fail_cor + fail_gen;
2001 }
2002
2003 static void scrub_block_get(struct scrub_block *sblock)
2004 {
2005 refcount_inc(&sblock->refs);
2006 }
2007
2008 static void scrub_block_put(struct scrub_block *sblock)
2009 {
2010 if (refcount_dec_and_test(&sblock->refs)) {
2011 int i;
2012
2013 if (sblock->sparity)
2014 scrub_parity_put(sblock->sparity);
2015
2016 for (i = 0; i < sblock->page_count; i++)
2017 scrub_page_put(sblock->pagev[i]);
2018 kfree(sblock);
2019 }
2020 }
2021
2022 static void scrub_page_get(struct scrub_page *spage)
2023 {
2024 atomic_inc(&spage->refs);
2025 }
2026
2027 static void scrub_page_put(struct scrub_page *spage)
2028 {
2029 if (atomic_dec_and_test(&spage->refs)) {
2030 if (spage->page)
2031 __free_page(spage->page);
2032 kfree(spage);
2033 }
2034 }
2035
2036 static void scrub_submit(struct scrub_ctx *sctx)
2037 {
2038 struct scrub_bio *sbio;
2039
2040 if (sctx->curr == -1)
2041 return;
2042
2043 sbio = sctx->bios[sctx->curr];
2044 sctx->curr = -1;
2045 scrub_pending_bio_inc(sctx);
2046 btrfsic_submit_bio(sbio->bio);
2047 }
2048
2049 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2050 struct scrub_page *spage)
2051 {
2052 struct scrub_block *sblock = spage->sblock;
2053 struct scrub_bio *sbio;
2054 int ret;
2055
2056 again:
2057 /*
2058 * grab a fresh bio or wait for one to become available
2059 */
2060 while (sctx->curr == -1) {
2061 spin_lock(&sctx->list_lock);
2062 sctx->curr = sctx->first_free;
2063 if (sctx->curr != -1) {
2064 sctx->first_free = sctx->bios[sctx->curr]->next_free;
2065 sctx->bios[sctx->curr]->next_free = -1;
2066 sctx->bios[sctx->curr]->page_count = 0;
2067 spin_unlock(&sctx->list_lock);
2068 } else {
2069 spin_unlock(&sctx->list_lock);
2070 wait_event(sctx->list_wait, sctx->first_free != -1);
2071 }
2072 }
2073 sbio = sctx->bios[sctx->curr];
2074 if (sbio->page_count == 0) {
2075 struct bio *bio;
2076
2077 sbio->physical = spage->physical;
2078 sbio->logical = spage->logical;
2079 sbio->dev = spage->dev;
2080 bio = sbio->bio;
2081 if (!bio) {
2082 bio = btrfs_io_bio_alloc(GFP_KERNEL,
2083 sctx->pages_per_rd_bio);
2084 if (!bio)
2085 return -ENOMEM;
2086 sbio->bio = bio;
2087 }
2088
2089 bio->bi_private = sbio;
2090 bio->bi_end_io = scrub_bio_end_io;
2091 bio->bi_bdev = sbio->dev->bdev;
2092 bio->bi_iter.bi_sector = sbio->physical >> 9;
2093 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2094 sbio->err = 0;
2095 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2096 spage->physical ||
2097 sbio->logical + sbio->page_count * PAGE_SIZE !=
2098 spage->logical ||
2099 sbio->dev != spage->dev) {
2100 scrub_submit(sctx);
2101 goto again;
2102 }
2103
2104 sbio->pagev[sbio->page_count] = spage;
2105 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2106 if (ret != PAGE_SIZE) {
2107 if (sbio->page_count < 1) {
2108 bio_put(sbio->bio);
2109 sbio->bio = NULL;
2110 return -EIO;
2111 }
2112 scrub_submit(sctx);
2113 goto again;
2114 }
2115
2116 scrub_block_get(sblock); /* one for the page added to the bio */
2117 atomic_inc(&sblock->outstanding_pages);
2118 sbio->page_count++;
2119 if (sbio->page_count == sctx->pages_per_rd_bio)
2120 scrub_submit(sctx);
2121
2122 return 0;
2123 }
2124
2125 static void scrub_missing_raid56_end_io(struct bio *bio)
2126 {
2127 struct scrub_block *sblock = bio->bi_private;
2128 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2129
2130 if (bio->bi_error)
2131 sblock->no_io_error_seen = 0;
2132
2133 bio_put(bio);
2134
2135 btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2136 }
2137
2138 static void scrub_missing_raid56_worker(struct btrfs_work *work)
2139 {
2140 struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2141 struct scrub_ctx *sctx = sblock->sctx;
2142 struct btrfs_fs_info *fs_info = sctx->fs_info;
2143 u64 logical;
2144 struct btrfs_device *dev;
2145
2146 logical = sblock->pagev[0]->logical;
2147 dev = sblock->pagev[0]->dev;
2148
2149 if (sblock->no_io_error_seen)
2150 scrub_recheck_block_checksum(sblock);
2151
2152 if (!sblock->no_io_error_seen) {
2153 spin_lock(&sctx->stat_lock);
2154 sctx->stat.read_errors++;
2155 spin_unlock(&sctx->stat_lock);
2156 btrfs_err_rl_in_rcu(fs_info,
2157 "IO error rebuilding logical %llu for dev %s",
2158 logical, rcu_str_deref(dev->name));
2159 } else if (sblock->header_error || sblock->checksum_error) {
2160 spin_lock(&sctx->stat_lock);
2161 sctx->stat.uncorrectable_errors++;
2162 spin_unlock(&sctx->stat_lock);
2163 btrfs_err_rl_in_rcu(fs_info,
2164 "failed to rebuild valid logical %llu for dev %s",
2165 logical, rcu_str_deref(dev->name));
2166 } else {
2167 scrub_write_block_to_dev_replace(sblock);
2168 }
2169
2170 scrub_block_put(sblock);
2171
2172 if (sctx->is_dev_replace &&
2173 atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2174 mutex_lock(&sctx->wr_ctx.wr_lock);
2175 scrub_wr_submit(sctx);
2176 mutex_unlock(&sctx->wr_ctx.wr_lock);
2177 }
2178
2179 scrub_pending_bio_dec(sctx);
2180 }
2181
2182 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2183 {
2184 struct scrub_ctx *sctx = sblock->sctx;
2185 struct btrfs_fs_info *fs_info = sctx->fs_info;
2186 u64 length = sblock->page_count * PAGE_SIZE;
2187 u64 logical = sblock->pagev[0]->logical;
2188 struct btrfs_bio *bbio = NULL;
2189 struct bio *bio;
2190 struct btrfs_raid_bio *rbio;
2191 int ret;
2192 int i;
2193
2194 btrfs_bio_counter_inc_blocked(fs_info);
2195 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2196 &length, &bbio);
2197 if (ret || !bbio || !bbio->raid_map)
2198 goto bbio_out;
2199
2200 if (WARN_ON(!sctx->is_dev_replace ||
2201 !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2202 /*
2203 * We shouldn't be scrubbing a missing device. Even for dev
2204 * replace, we should only get here for RAID 5/6. We either
2205 * managed to mount something with no mirrors remaining or
2206 * there's a bug in scrub_remap_extent()/btrfs_map_block().
2207 */
2208 goto bbio_out;
2209 }
2210
2211 bio = btrfs_io_bio_alloc(GFP_NOFS, 0);
2212 if (!bio)
2213 goto bbio_out;
2214
2215 bio->bi_iter.bi_sector = logical >> 9;
2216 bio->bi_private = sblock;
2217 bio->bi_end_io = scrub_missing_raid56_end_io;
2218
2219 rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2220 if (!rbio)
2221 goto rbio_out;
2222
2223 for (i = 0; i < sblock->page_count; i++) {
2224 struct scrub_page *spage = sblock->pagev[i];
2225
2226 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2227 }
2228
2229 btrfs_init_work(&sblock->work, btrfs_scrub_helper,
2230 scrub_missing_raid56_worker, NULL, NULL);
2231 scrub_block_get(sblock);
2232 scrub_pending_bio_inc(sctx);
2233 raid56_submit_missing_rbio(rbio);
2234 return;
2235
2236 rbio_out:
2237 bio_put(bio);
2238 bbio_out:
2239 btrfs_bio_counter_dec(fs_info);
2240 btrfs_put_bbio(bbio);
2241 spin_lock(&sctx->stat_lock);
2242 sctx->stat.malloc_errors++;
2243 spin_unlock(&sctx->stat_lock);
2244 }
2245
2246 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2247 u64 physical, struct btrfs_device *dev, u64 flags,
2248 u64 gen, int mirror_num, u8 *csum, int force,
2249 u64 physical_for_dev_replace)
2250 {
2251 struct scrub_block *sblock;
2252 int index;
2253
2254 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2255 if (!sblock) {
2256 spin_lock(&sctx->stat_lock);
2257 sctx->stat.malloc_errors++;
2258 spin_unlock(&sctx->stat_lock);
2259 return -ENOMEM;
2260 }
2261
2262 /* one ref inside this function, plus one for each page added to
2263 * a bio later on */
2264 refcount_set(&sblock->refs, 1);
2265 sblock->sctx = sctx;
2266 sblock->no_io_error_seen = 1;
2267
2268 for (index = 0; len > 0; index++) {
2269 struct scrub_page *spage;
2270 u64 l = min_t(u64, len, PAGE_SIZE);
2271
2272 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2273 if (!spage) {
2274 leave_nomem:
2275 spin_lock(&sctx->stat_lock);
2276 sctx->stat.malloc_errors++;
2277 spin_unlock(&sctx->stat_lock);
2278 scrub_block_put(sblock);
2279 return -ENOMEM;
2280 }
2281 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2282 scrub_page_get(spage);
2283 sblock->pagev[index] = spage;
2284 spage->sblock = sblock;
2285 spage->dev = dev;
2286 spage->flags = flags;
2287 spage->generation = gen;
2288 spage->logical = logical;
2289 spage->physical = physical;
2290 spage->physical_for_dev_replace = physical_for_dev_replace;
2291 spage->mirror_num = mirror_num;
2292 if (csum) {
2293 spage->have_csum = 1;
2294 memcpy(spage->csum, csum, sctx->csum_size);
2295 } else {
2296 spage->have_csum = 0;
2297 }
2298 sblock->page_count++;
2299 spage->page = alloc_page(GFP_KERNEL);
2300 if (!spage->page)
2301 goto leave_nomem;
2302 len -= l;
2303 logical += l;
2304 physical += l;
2305 physical_for_dev_replace += l;
2306 }
2307
2308 WARN_ON(sblock->page_count == 0);
2309 if (dev->missing) {
2310 /*
2311 * This case should only be hit for RAID 5/6 device replace. See
2312 * the comment in scrub_missing_raid56_pages() for details.
2313 */
2314 scrub_missing_raid56_pages(sblock);
2315 } else {
2316 for (index = 0; index < sblock->page_count; index++) {
2317 struct scrub_page *spage = sblock->pagev[index];
2318 int ret;
2319
2320 ret = scrub_add_page_to_rd_bio(sctx, spage);
2321 if (ret) {
2322 scrub_block_put(sblock);
2323 return ret;
2324 }
2325 }
2326
2327 if (force)
2328 scrub_submit(sctx);
2329 }
2330
2331 /* last one frees, either here or in bio completion for last page */
2332 scrub_block_put(sblock);
2333 return 0;
2334 }
2335
2336 static void scrub_bio_end_io(struct bio *bio)
2337 {
2338 struct scrub_bio *sbio = bio->bi_private;
2339 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2340
2341 sbio->err = bio->bi_error;
2342 sbio->bio = bio;
2343
2344 btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2345 }
2346
2347 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2348 {
2349 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2350 struct scrub_ctx *sctx = sbio->sctx;
2351 int i;
2352
2353 BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2354 if (sbio->err) {
2355 for (i = 0; i < sbio->page_count; i++) {
2356 struct scrub_page *spage = sbio->pagev[i];
2357
2358 spage->io_error = 1;
2359 spage->sblock->no_io_error_seen = 0;
2360 }
2361 }
2362
2363 /* now complete the scrub_block items that have all pages completed */
2364 for (i = 0; i < sbio->page_count; i++) {
2365 struct scrub_page *spage = sbio->pagev[i];
2366 struct scrub_block *sblock = spage->sblock;
2367
2368 if (atomic_dec_and_test(&sblock->outstanding_pages))
2369 scrub_block_complete(sblock);
2370 scrub_block_put(sblock);
2371 }
2372
2373 bio_put(sbio->bio);
2374 sbio->bio = NULL;
2375 spin_lock(&sctx->list_lock);
2376 sbio->next_free = sctx->first_free;
2377 sctx->first_free = sbio->index;
2378 spin_unlock(&sctx->list_lock);
2379
2380 if (sctx->is_dev_replace &&
2381 atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2382 mutex_lock(&sctx->wr_ctx.wr_lock);
2383 scrub_wr_submit(sctx);
2384 mutex_unlock(&sctx->wr_ctx.wr_lock);
2385 }
2386
2387 scrub_pending_bio_dec(sctx);
2388 }
2389
2390 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2391 unsigned long *bitmap,
2392 u64 start, u64 len)
2393 {
2394 u32 offset;
2395 int nsectors;
2396 int sectorsize = sparity->sctx->fs_info->sectorsize;
2397
2398 if (len >= sparity->stripe_len) {
2399 bitmap_set(bitmap, 0, sparity->nsectors);
2400 return;
2401 }
2402
2403 start -= sparity->logic_start;
2404 start = div_u64_rem(start, sparity->stripe_len, &offset);
2405 offset /= sectorsize;
2406 nsectors = (int)len / sectorsize;
2407
2408 if (offset + nsectors <= sparity->nsectors) {
2409 bitmap_set(bitmap, offset, nsectors);
2410 return;
2411 }
2412
2413 bitmap_set(bitmap, offset, sparity->nsectors - offset);
2414 bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2415 }
2416
2417 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2418 u64 start, u64 len)
2419 {
2420 __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2421 }
2422
2423 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2424 u64 start, u64 len)
2425 {
2426 __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2427 }
2428
2429 static void scrub_block_complete(struct scrub_block *sblock)
2430 {
2431 int corrupted = 0;
2432
2433 if (!sblock->no_io_error_seen) {
2434 corrupted = 1;
2435 scrub_handle_errored_block(sblock);
2436 } else {
2437 /*
2438 * if has checksum error, write via repair mechanism in
2439 * dev replace case, otherwise write here in dev replace
2440 * case.
2441 */
2442 corrupted = scrub_checksum(sblock);
2443 if (!corrupted && sblock->sctx->is_dev_replace)
2444 scrub_write_block_to_dev_replace(sblock);
2445 }
2446
2447 if (sblock->sparity && corrupted && !sblock->data_corrected) {
2448 u64 start = sblock->pagev[0]->logical;
2449 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2450 PAGE_SIZE;
2451
2452 scrub_parity_mark_sectors_error(sblock->sparity,
2453 start, end - start);
2454 }
2455 }
2456
2457 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2458 {
2459 struct btrfs_ordered_sum *sum = NULL;
2460 unsigned long index;
2461 unsigned long num_sectors;
2462
2463 while (!list_empty(&sctx->csum_list)) {
2464 sum = list_first_entry(&sctx->csum_list,
2465 struct btrfs_ordered_sum, list);
2466 if (sum->bytenr > logical)
2467 return 0;
2468 if (sum->bytenr + sum->len > logical)
2469 break;
2470
2471 ++sctx->stat.csum_discards;
2472 list_del(&sum->list);
2473 kfree(sum);
2474 sum = NULL;
2475 }
2476 if (!sum)
2477 return 0;
2478
2479 index = ((u32)(logical - sum->bytenr)) / sctx->sectorsize;
2480 num_sectors = sum->len / sctx->sectorsize;
2481 memcpy(csum, sum->sums + index, sctx->csum_size);
2482 if (index == num_sectors - 1) {
2483 list_del(&sum->list);
2484 kfree(sum);
2485 }
2486 return 1;
2487 }
2488
2489 /* scrub extent tries to collect up to 64 kB for each bio */
2490 static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2491 u64 physical, struct btrfs_device *dev, u64 flags,
2492 u64 gen, int mirror_num, u64 physical_for_dev_replace)
2493 {
2494 int ret;
2495 u8 csum[BTRFS_CSUM_SIZE];
2496 u32 blocksize;
2497
2498 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2499 blocksize = sctx->sectorsize;
2500 spin_lock(&sctx->stat_lock);
2501 sctx->stat.data_extents_scrubbed++;
2502 sctx->stat.data_bytes_scrubbed += len;
2503 spin_unlock(&sctx->stat_lock);
2504 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2505 blocksize = sctx->nodesize;
2506 spin_lock(&sctx->stat_lock);
2507 sctx->stat.tree_extents_scrubbed++;
2508 sctx->stat.tree_bytes_scrubbed += len;
2509 spin_unlock(&sctx->stat_lock);
2510 } else {
2511 blocksize = sctx->sectorsize;
2512 WARN_ON(1);
2513 }
2514
2515 while (len) {
2516 u64 l = min_t(u64, len, blocksize);
2517 int have_csum = 0;
2518
2519 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2520 /* push csums to sbio */
2521 have_csum = scrub_find_csum(sctx, logical, csum);
2522 if (have_csum == 0)
2523 ++sctx->stat.no_csum;
2524 if (sctx->is_dev_replace && !have_csum) {
2525 ret = copy_nocow_pages(sctx, logical, l,
2526 mirror_num,
2527 physical_for_dev_replace);
2528 goto behind_scrub_pages;
2529 }
2530 }
2531 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2532 mirror_num, have_csum ? csum : NULL, 0,
2533 physical_for_dev_replace);
2534 behind_scrub_pages:
2535 if (ret)
2536 return ret;
2537 len -= l;
2538 logical += l;
2539 physical += l;
2540 physical_for_dev_replace += l;
2541 }
2542 return 0;
2543 }
2544
2545 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2546 u64 logical, u64 len,
2547 u64 physical, struct btrfs_device *dev,
2548 u64 flags, u64 gen, int mirror_num, u8 *csum)
2549 {
2550 struct scrub_ctx *sctx = sparity->sctx;
2551 struct scrub_block *sblock;
2552 int index;
2553
2554 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2555 if (!sblock) {
2556 spin_lock(&sctx->stat_lock);
2557 sctx->stat.malloc_errors++;
2558 spin_unlock(&sctx->stat_lock);
2559 return -ENOMEM;
2560 }
2561
2562 /* one ref inside this function, plus one for each page added to
2563 * a bio later on */
2564 refcount_set(&sblock->refs, 1);
2565 sblock->sctx = sctx;
2566 sblock->no_io_error_seen = 1;
2567 sblock->sparity = sparity;
2568 scrub_parity_get(sparity);
2569
2570 for (index = 0; len > 0; index++) {
2571 struct scrub_page *spage;
2572 u64 l = min_t(u64, len, PAGE_SIZE);
2573
2574 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2575 if (!spage) {
2576 leave_nomem:
2577 spin_lock(&sctx->stat_lock);
2578 sctx->stat.malloc_errors++;
2579 spin_unlock(&sctx->stat_lock);
2580 scrub_block_put(sblock);
2581 return -ENOMEM;
2582 }
2583 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2584 /* For scrub block */
2585 scrub_page_get(spage);
2586 sblock->pagev[index] = spage;
2587 /* For scrub parity */
2588 scrub_page_get(spage);
2589 list_add_tail(&spage->list, &sparity->spages);
2590 spage->sblock = sblock;
2591 spage->dev = dev;
2592 spage->flags = flags;
2593 spage->generation = gen;
2594 spage->logical = logical;
2595 spage->physical = physical;
2596 spage->mirror_num = mirror_num;
2597 if (csum) {
2598 spage->have_csum = 1;
2599 memcpy(spage->csum, csum, sctx->csum_size);
2600 } else {
2601 spage->have_csum = 0;
2602 }
2603 sblock->page_count++;
2604 spage->page = alloc_page(GFP_KERNEL);
2605 if (!spage->page)
2606 goto leave_nomem;
2607 len -= l;
2608 logical += l;
2609 physical += l;
2610 }
2611
2612 WARN_ON(sblock->page_count == 0);
2613 for (index = 0; index < sblock->page_count; index++) {
2614 struct scrub_page *spage = sblock->pagev[index];
2615 int ret;
2616
2617 ret = scrub_add_page_to_rd_bio(sctx, spage);
2618 if (ret) {
2619 scrub_block_put(sblock);
2620 return ret;
2621 }
2622 }
2623
2624 /* last one frees, either here or in bio completion for last page */
2625 scrub_block_put(sblock);
2626 return 0;
2627 }
2628
2629 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2630 u64 logical, u64 len,
2631 u64 physical, struct btrfs_device *dev,
2632 u64 flags, u64 gen, int mirror_num)
2633 {
2634 struct scrub_ctx *sctx = sparity->sctx;
2635 int ret;
2636 u8 csum[BTRFS_CSUM_SIZE];
2637 u32 blocksize;
2638
2639 if (dev->missing) {
2640 scrub_parity_mark_sectors_error(sparity, logical, len);
2641 return 0;
2642 }
2643
2644 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2645 blocksize = sctx->sectorsize;
2646 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2647 blocksize = sctx->nodesize;
2648 } else {
2649 blocksize = sctx->sectorsize;
2650 WARN_ON(1);
2651 }
2652
2653 while (len) {
2654 u64 l = min_t(u64, len, blocksize);
2655 int have_csum = 0;
2656
2657 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2658 /* push csums to sbio */
2659 have_csum = scrub_find_csum(sctx, logical, csum);
2660 if (have_csum == 0)
2661 goto skip;
2662 }
2663 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2664 flags, gen, mirror_num,
2665 have_csum ? csum : NULL);
2666 if (ret)
2667 return ret;
2668 skip:
2669 len -= l;
2670 logical += l;
2671 physical += l;
2672 }
2673 return 0;
2674 }
2675
2676 /*
2677 * Given a physical address, this will calculate it's
2678 * logical offset. if this is a parity stripe, it will return
2679 * the most left data stripe's logical offset.
2680 *
2681 * return 0 if it is a data stripe, 1 means parity stripe.
2682 */
2683 static int get_raid56_logic_offset(u64 physical, int num,
2684 struct map_lookup *map, u64 *offset,
2685 u64 *stripe_start)
2686 {
2687 int i;
2688 int j = 0;
2689 u64 stripe_nr;
2690 u64 last_offset;
2691 u32 stripe_index;
2692 u32 rot;
2693
2694 last_offset = (physical - map->stripes[num].physical) *
2695 nr_data_stripes(map);
2696 if (stripe_start)
2697 *stripe_start = last_offset;
2698
2699 *offset = last_offset;
2700 for (i = 0; i < nr_data_stripes(map); i++) {
2701 *offset = last_offset + i * map->stripe_len;
2702
2703 stripe_nr = div_u64(*offset, map->stripe_len);
2704 stripe_nr = div_u64(stripe_nr, nr_data_stripes(map));
2705
2706 /* Work out the disk rotation on this stripe-set */
2707 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2708 /* calculate which stripe this data locates */
2709 rot += i;
2710 stripe_index = rot % map->num_stripes;
2711 if (stripe_index == num)
2712 return 0;
2713 if (stripe_index < num)
2714 j++;
2715 }
2716 *offset = last_offset + j * map->stripe_len;
2717 return 1;
2718 }
2719
2720 static void scrub_free_parity(struct scrub_parity *sparity)
2721 {
2722 struct scrub_ctx *sctx = sparity->sctx;
2723 struct scrub_page *curr, *next;
2724 int nbits;
2725
2726 nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2727 if (nbits) {
2728 spin_lock(&sctx->stat_lock);
2729 sctx->stat.read_errors += nbits;
2730 sctx->stat.uncorrectable_errors += nbits;
2731 spin_unlock(&sctx->stat_lock);
2732 }
2733
2734 list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2735 list_del_init(&curr->list);
2736 scrub_page_put(curr);
2737 }
2738
2739 kfree(sparity);
2740 }
2741
2742 static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2743 {
2744 struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2745 work);
2746 struct scrub_ctx *sctx = sparity->sctx;
2747
2748 scrub_free_parity(sparity);
2749 scrub_pending_bio_dec(sctx);
2750 }
2751
2752 static void scrub_parity_bio_endio(struct bio *bio)
2753 {
2754 struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
2755 struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
2756
2757 if (bio->bi_error)
2758 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2759 sparity->nsectors);
2760
2761 bio_put(bio);
2762
2763 btrfs_init_work(&sparity->work, btrfs_scrubparity_helper,
2764 scrub_parity_bio_endio_worker, NULL, NULL);
2765 btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
2766 }
2767
2768 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2769 {
2770 struct scrub_ctx *sctx = sparity->sctx;
2771 struct btrfs_fs_info *fs_info = sctx->fs_info;
2772 struct bio *bio;
2773 struct btrfs_raid_bio *rbio;
2774 struct btrfs_bio *bbio = NULL;
2775 u64 length;
2776 int ret;
2777
2778 if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
2779 sparity->nsectors))
2780 goto out;
2781
2782 length = sparity->logic_end - sparity->logic_start;
2783
2784 btrfs_bio_counter_inc_blocked(fs_info);
2785 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
2786 &length, &bbio);
2787 if (ret || !bbio || !bbio->raid_map)
2788 goto bbio_out;
2789
2790 bio = btrfs_io_bio_alloc(GFP_NOFS, 0);
2791 if (!bio)
2792 goto bbio_out;
2793
2794 bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2795 bio->bi_private = sparity;
2796 bio->bi_end_io = scrub_parity_bio_endio;
2797
2798 rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
2799 length, sparity->scrub_dev,
2800 sparity->dbitmap,
2801 sparity->nsectors);
2802 if (!rbio)
2803 goto rbio_out;
2804
2805 scrub_pending_bio_inc(sctx);
2806 raid56_parity_submit_scrub_rbio(rbio);
2807 return;
2808
2809 rbio_out:
2810 bio_put(bio);
2811 bbio_out:
2812 btrfs_bio_counter_dec(fs_info);
2813 btrfs_put_bbio(bbio);
2814 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2815 sparity->nsectors);
2816 spin_lock(&sctx->stat_lock);
2817 sctx->stat.malloc_errors++;
2818 spin_unlock(&sctx->stat_lock);
2819 out:
2820 scrub_free_parity(sparity);
2821 }
2822
2823 static inline int scrub_calc_parity_bitmap_len(int nsectors)
2824 {
2825 return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
2826 }
2827
2828 static void scrub_parity_get(struct scrub_parity *sparity)
2829 {
2830 refcount_inc(&sparity->refs);
2831 }
2832
2833 static void scrub_parity_put(struct scrub_parity *sparity)
2834 {
2835 if (!refcount_dec_and_test(&sparity->refs))
2836 return;
2837
2838 scrub_parity_check_and_repair(sparity);
2839 }
2840
2841 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
2842 struct map_lookup *map,
2843 struct btrfs_device *sdev,
2844 struct btrfs_path *path,
2845 u64 logic_start,
2846 u64 logic_end)
2847 {
2848 struct btrfs_fs_info *fs_info = sctx->fs_info;
2849 struct btrfs_root *root = fs_info->extent_root;
2850 struct btrfs_root *csum_root = fs_info->csum_root;
2851 struct btrfs_extent_item *extent;
2852 struct btrfs_bio *bbio = NULL;
2853 u64 flags;
2854 int ret;
2855 int slot;
2856 struct extent_buffer *l;
2857 struct btrfs_key key;
2858 u64 generation;
2859 u64 extent_logical;
2860 u64 extent_physical;
2861 u64 extent_len;
2862 u64 mapped_length;
2863 struct btrfs_device *extent_dev;
2864 struct scrub_parity *sparity;
2865 int nsectors;
2866 int bitmap_len;
2867 int extent_mirror_num;
2868 int stop_loop = 0;
2869
2870 nsectors = div_u64(map->stripe_len, fs_info->sectorsize);
2871 bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
2872 sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
2873 GFP_NOFS);
2874 if (!sparity) {
2875 spin_lock(&sctx->stat_lock);
2876 sctx->stat.malloc_errors++;
2877 spin_unlock(&sctx->stat_lock);
2878 return -ENOMEM;
2879 }
2880
2881 sparity->stripe_len = map->stripe_len;
2882 sparity->nsectors = nsectors;
2883 sparity->sctx = sctx;
2884 sparity->scrub_dev = sdev;
2885 sparity->logic_start = logic_start;
2886 sparity->logic_end = logic_end;
2887 refcount_set(&sparity->refs, 1);
2888 INIT_LIST_HEAD(&sparity->spages);
2889 sparity->dbitmap = sparity->bitmap;
2890 sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
2891
2892 ret = 0;
2893 while (logic_start < logic_end) {
2894 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2895 key.type = BTRFS_METADATA_ITEM_KEY;
2896 else
2897 key.type = BTRFS_EXTENT_ITEM_KEY;
2898 key.objectid = logic_start;
2899 key.offset = (u64)-1;
2900
2901 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2902 if (ret < 0)
2903 goto out;
2904
2905 if (ret > 0) {
2906 ret = btrfs_previous_extent_item(root, path, 0);
2907 if (ret < 0)
2908 goto out;
2909 if (ret > 0) {
2910 btrfs_release_path(path);
2911 ret = btrfs_search_slot(NULL, root, &key,
2912 path, 0, 0);
2913 if (ret < 0)
2914 goto out;
2915 }
2916 }
2917
2918 stop_loop = 0;
2919 while (1) {
2920 u64 bytes;
2921
2922 l = path->nodes[0];
2923 slot = path->slots[0];
2924 if (slot >= btrfs_header_nritems(l)) {
2925 ret = btrfs_next_leaf(root, path);
2926 if (ret == 0)
2927 continue;
2928 if (ret < 0)
2929 goto out;
2930
2931 stop_loop = 1;
2932 break;
2933 }
2934 btrfs_item_key_to_cpu(l, &key, slot);
2935
2936 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2937 key.type != BTRFS_METADATA_ITEM_KEY)
2938 goto next;
2939
2940 if (key.type == BTRFS_METADATA_ITEM_KEY)
2941 bytes = fs_info->nodesize;
2942 else
2943 bytes = key.offset;
2944
2945 if (key.objectid + bytes <= logic_start)
2946 goto next;
2947
2948 if (key.objectid >= logic_end) {
2949 stop_loop = 1;
2950 break;
2951 }
2952
2953 while (key.objectid >= logic_start + map->stripe_len)
2954 logic_start += map->stripe_len;
2955
2956 extent = btrfs_item_ptr(l, slot,
2957 struct btrfs_extent_item);
2958 flags = btrfs_extent_flags(l, extent);
2959 generation = btrfs_extent_generation(l, extent);
2960
2961 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
2962 (key.objectid < logic_start ||
2963 key.objectid + bytes >
2964 logic_start + map->stripe_len)) {
2965 btrfs_err(fs_info,
2966 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
2967 key.objectid, logic_start);
2968 spin_lock(&sctx->stat_lock);
2969 sctx->stat.uncorrectable_errors++;
2970 spin_unlock(&sctx->stat_lock);
2971 goto next;
2972 }
2973 again:
2974 extent_logical = key.objectid;
2975 extent_len = bytes;
2976
2977 if (extent_logical < logic_start) {
2978 extent_len -= logic_start - extent_logical;
2979 extent_logical = logic_start;
2980 }
2981
2982 if (extent_logical + extent_len >
2983 logic_start + map->stripe_len)
2984 extent_len = logic_start + map->stripe_len -
2985 extent_logical;
2986
2987 scrub_parity_mark_sectors_data(sparity, extent_logical,
2988 extent_len);
2989
2990 mapped_length = extent_len;
2991 bbio = NULL;
2992 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
2993 extent_logical, &mapped_length, &bbio,
2994 0);
2995 if (!ret) {
2996 if (!bbio || mapped_length < extent_len)
2997 ret = -EIO;
2998 }
2999 if (ret) {
3000 btrfs_put_bbio(bbio);
3001 goto out;
3002 }
3003 extent_physical = bbio->stripes[0].physical;
3004 extent_mirror_num = bbio->mirror_num;
3005 extent_dev = bbio->stripes[0].dev;
3006 btrfs_put_bbio(bbio);
3007
3008 ret = btrfs_lookup_csums_range(csum_root,
3009 extent_logical,
3010 extent_logical + extent_len - 1,
3011 &sctx->csum_list, 1);
3012 if (ret)
3013 goto out;
3014
3015 ret = scrub_extent_for_parity(sparity, extent_logical,
3016 extent_len,
3017 extent_physical,
3018 extent_dev, flags,
3019 generation,
3020 extent_mirror_num);
3021
3022 scrub_free_csums(sctx);
3023
3024 if (ret)
3025 goto out;
3026
3027 if (extent_logical + extent_len <
3028 key.objectid + bytes) {
3029 logic_start += map->stripe_len;
3030
3031 if (logic_start >= logic_end) {
3032 stop_loop = 1;
3033 break;
3034 }
3035
3036 if (logic_start < key.objectid + bytes) {
3037 cond_resched();
3038 goto again;
3039 }
3040 }
3041 next:
3042 path->slots[0]++;
3043 }
3044
3045 btrfs_release_path(path);
3046
3047 if (stop_loop)
3048 break;
3049
3050 logic_start += map->stripe_len;
3051 }
3052 out:
3053 if (ret < 0)
3054 scrub_parity_mark_sectors_error(sparity, logic_start,
3055 logic_end - logic_start);
3056 scrub_parity_put(sparity);
3057 scrub_submit(sctx);
3058 mutex_lock(&sctx->wr_ctx.wr_lock);
3059 scrub_wr_submit(sctx);
3060 mutex_unlock(&sctx->wr_ctx.wr_lock);
3061
3062 btrfs_release_path(path);
3063 return ret < 0 ? ret : 0;
3064 }
3065
3066 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3067 struct map_lookup *map,
3068 struct btrfs_device *scrub_dev,
3069 int num, u64 base, u64 length,
3070 int is_dev_replace)
3071 {
3072 struct btrfs_path *path, *ppath;
3073 struct btrfs_fs_info *fs_info = sctx->fs_info;
3074 struct btrfs_root *root = fs_info->extent_root;
3075 struct btrfs_root *csum_root = fs_info->csum_root;
3076 struct btrfs_extent_item *extent;
3077 struct blk_plug plug;
3078 u64 flags;
3079 int ret;
3080 int slot;
3081 u64 nstripes;
3082 struct extent_buffer *l;
3083 u64 physical;
3084 u64 logical;
3085 u64 logic_end;
3086 u64 physical_end;
3087 u64 generation;
3088 int mirror_num;
3089 struct reada_control *reada1;
3090 struct reada_control *reada2;
3091 struct btrfs_key key;
3092 struct btrfs_key key_end;
3093 u64 increment = map->stripe_len;
3094 u64 offset;
3095 u64 extent_logical;
3096 u64 extent_physical;
3097 u64 extent_len;
3098 u64 stripe_logical;
3099 u64 stripe_end;
3100 struct btrfs_device *extent_dev;
3101 int extent_mirror_num;
3102 int stop_loop = 0;
3103
3104 physical = map->stripes[num].physical;
3105 offset = 0;
3106 nstripes = div_u64(length, map->stripe_len);
3107 if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3108 offset = map->stripe_len * num;
3109 increment = map->stripe_len * map->num_stripes;
3110 mirror_num = 1;
3111 } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3112 int factor = map->num_stripes / map->sub_stripes;
3113 offset = map->stripe_len * (num / map->sub_stripes);
3114 increment = map->stripe_len * factor;
3115 mirror_num = num % map->sub_stripes + 1;
3116 } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
3117 increment = map->stripe_len;
3118 mirror_num = num % map->num_stripes + 1;
3119 } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3120 increment = map->stripe_len;
3121 mirror_num = num % map->num_stripes + 1;
3122 } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3123 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3124 increment = map->stripe_len * nr_data_stripes(map);
3125 mirror_num = 1;
3126 } else {
3127 increment = map->stripe_len;
3128 mirror_num = 1;
3129 }
3130
3131 path = btrfs_alloc_path();
3132 if (!path)
3133 return -ENOMEM;
3134
3135 ppath = btrfs_alloc_path();
3136 if (!ppath) {
3137 btrfs_free_path(path);
3138 return -ENOMEM;
3139 }
3140
3141 /*
3142 * work on commit root. The related disk blocks are static as
3143 * long as COW is applied. This means, it is save to rewrite
3144 * them to repair disk errors without any race conditions
3145 */
3146 path->search_commit_root = 1;
3147 path->skip_locking = 1;
3148
3149 ppath->search_commit_root = 1;
3150 ppath->skip_locking = 1;
3151 /*
3152 * trigger the readahead for extent tree csum tree and wait for
3153 * completion. During readahead, the scrub is officially paused
3154 * to not hold off transaction commits
3155 */
3156 logical = base + offset;
3157 physical_end = physical + nstripes * map->stripe_len;
3158 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3159 get_raid56_logic_offset(physical_end, num,
3160 map, &logic_end, NULL);
3161 logic_end += base;
3162 } else {
3163 logic_end = logical + increment * nstripes;
3164 }
3165 wait_event(sctx->list_wait,
3166 atomic_read(&sctx->bios_in_flight) == 0);
3167 scrub_blocked_if_needed(fs_info);
3168
3169 /* FIXME it might be better to start readahead at commit root */
3170 key.objectid = logical;
3171 key.type = BTRFS_EXTENT_ITEM_KEY;
3172 key.offset = (u64)0;
3173 key_end.objectid = logic_end;
3174 key_end.type = BTRFS_METADATA_ITEM_KEY;
3175 key_end.offset = (u64)-1;
3176 reada1 = btrfs_reada_add(root, &key, &key_end);
3177
3178 key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3179 key.type = BTRFS_EXTENT_CSUM_KEY;
3180 key.offset = logical;
3181 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3182 key_end.type = BTRFS_EXTENT_CSUM_KEY;
3183 key_end.offset = logic_end;
3184 reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3185
3186 if (!IS_ERR(reada1))
3187 btrfs_reada_wait(reada1);
3188 if (!IS_ERR(reada2))
3189 btrfs_reada_wait(reada2);
3190
3191
3192 /*
3193 * collect all data csums for the stripe to avoid seeking during
3194 * the scrub. This might currently (crc32) end up to be about 1MB
3195 */
3196 blk_start_plug(&plug);
3197
3198 /*
3199 * now find all extents for each stripe and scrub them
3200 */
3201 ret = 0;
3202 while (physical < physical_end) {
3203 /*
3204 * canceled?
3205 */
3206 if (atomic_read(&fs_info->scrub_cancel_req) ||
3207 atomic_read(&sctx->cancel_req)) {
3208 ret = -ECANCELED;
3209 goto out;
3210 }
3211 /*
3212 * check to see if we have to pause
3213 */
3214 if (atomic_read(&fs_info->scrub_pause_req)) {
3215 /* push queued extents */
3216 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
3217 scrub_submit(sctx);
3218 mutex_lock(&sctx->wr_ctx.wr_lock);
3219 scrub_wr_submit(sctx);
3220 mutex_unlock(&sctx->wr_ctx.wr_lock);
3221 wait_event(sctx->list_wait,
3222 atomic_read(&sctx->bios_in_flight) == 0);
3223 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
3224 scrub_blocked_if_needed(fs_info);
3225 }
3226
3227 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3228 ret = get_raid56_logic_offset(physical, num, map,
3229 &logical,
3230 &stripe_logical);
3231 logical += base;
3232 if (ret) {
3233 /* it is parity strip */
3234 stripe_logical += base;
3235 stripe_end = stripe_logical + increment;
3236 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3237 ppath, stripe_logical,
3238 stripe_end);
3239 if (ret)
3240 goto out;
3241 goto skip;
3242 }
3243 }
3244
3245 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3246 key.type = BTRFS_METADATA_ITEM_KEY;
3247 else
3248 key.type = BTRFS_EXTENT_ITEM_KEY;
3249 key.objectid = logical;
3250 key.offset = (u64)-1;
3251
3252 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3253 if (ret < 0)
3254 goto out;
3255
3256 if (ret > 0) {
3257 ret = btrfs_previous_extent_item(root, path, 0);
3258 if (ret < 0)
3259 goto out;
3260 if (ret > 0) {
3261 /* there's no smaller item, so stick with the
3262 * larger one */
3263 btrfs_release_path(path);
3264 ret = btrfs_search_slot(NULL, root, &key,
3265 path, 0, 0);
3266 if (ret < 0)
3267 goto out;
3268 }
3269 }
3270
3271 stop_loop = 0;
3272 while (1) {
3273 u64 bytes;
3274
3275 l = path->nodes[0];
3276 slot = path->slots[0];
3277 if (slot >= btrfs_header_nritems(l)) {
3278 ret = btrfs_next_leaf(root, path);
3279 if (ret == 0)
3280 continue;
3281 if (ret < 0)
3282 goto out;
3283
3284 stop_loop = 1;
3285 break;
3286 }
3287 btrfs_item_key_to_cpu(l, &key, slot);
3288
3289 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3290 key.type != BTRFS_METADATA_ITEM_KEY)
3291 goto next;
3292
3293 if (key.type == BTRFS_METADATA_ITEM_KEY)
3294 bytes = fs_info->nodesize;
3295 else
3296 bytes = key.offset;
3297
3298 if (key.objectid + bytes <= logical)
3299 goto next;
3300
3301 if (key.objectid >= logical + map->stripe_len) {
3302 /* out of this device extent */
3303 if (key.objectid >= logic_end)
3304 stop_loop = 1;
3305 break;
3306 }
3307
3308 extent = btrfs_item_ptr(l, slot,
3309 struct btrfs_extent_item);
3310 flags = btrfs_extent_flags(l, extent);
3311 generation = btrfs_extent_generation(l, extent);
3312
3313 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3314 (key.objectid < logical ||
3315 key.objectid + bytes >
3316 logical + map->stripe_len)) {
3317 btrfs_err(fs_info,
3318 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3319 key.objectid, logical);
3320 spin_lock(&sctx->stat_lock);
3321 sctx->stat.uncorrectable_errors++;
3322 spin_unlock(&sctx->stat_lock);
3323 goto next;
3324 }
3325
3326 again:
3327 extent_logical = key.objectid;
3328 extent_len = bytes;
3329
3330 /*
3331 * trim extent to this stripe
3332 */
3333 if (extent_logical < logical) {
3334 extent_len -= logical - extent_logical;
3335 extent_logical = logical;
3336 }
3337 if (extent_logical + extent_len >
3338 logical + map->stripe_len) {
3339 extent_len = logical + map->stripe_len -
3340 extent_logical;
3341 }
3342
3343 extent_physical = extent_logical - logical + physical;
3344 extent_dev = scrub_dev;
3345 extent_mirror_num = mirror_num;
3346 if (is_dev_replace)
3347 scrub_remap_extent(fs_info, extent_logical,
3348 extent_len, &extent_physical,
3349 &extent_dev,
3350 &extent_mirror_num);
3351
3352 ret = btrfs_lookup_csums_range(csum_root,
3353 extent_logical,
3354 extent_logical +
3355 extent_len - 1,
3356 &sctx->csum_list, 1);
3357 if (ret)
3358 goto out;
3359
3360 ret = scrub_extent(sctx, extent_logical, extent_len,
3361 extent_physical, extent_dev, flags,
3362 generation, extent_mirror_num,
3363 extent_logical - logical + physical);
3364
3365 scrub_free_csums(sctx);
3366
3367 if (ret)
3368 goto out;
3369
3370 if (extent_logical + extent_len <
3371 key.objectid + bytes) {
3372 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3373 /*
3374 * loop until we find next data stripe
3375 * or we have finished all stripes.
3376 */
3377 loop:
3378 physical += map->stripe_len;
3379 ret = get_raid56_logic_offset(physical,
3380 num, map, &logical,
3381 &stripe_logical);
3382 logical += base;
3383
3384 if (ret && physical < physical_end) {
3385 stripe_logical += base;
3386 stripe_end = stripe_logical +
3387 increment;
3388 ret = scrub_raid56_parity(sctx,
3389 map, scrub_dev, ppath,
3390 stripe_logical,
3391 stripe_end);
3392 if (ret)
3393 goto out;
3394 goto loop;
3395 }
3396 } else {
3397 physical += map->stripe_len;
3398 logical += increment;
3399 }
3400 if (logical < key.objectid + bytes) {
3401 cond_resched();
3402 goto again;
3403 }
3404
3405 if (physical >= physical_end) {
3406 stop_loop = 1;
3407 break;
3408 }
3409 }
3410 next:
3411 path->slots[0]++;
3412 }
3413 btrfs_release_path(path);
3414 skip:
3415 logical += increment;
3416 physical += map->stripe_len;
3417 spin_lock(&sctx->stat_lock);
3418 if (stop_loop)
3419 sctx->stat.last_physical = map->stripes[num].physical +
3420 length;
3421 else
3422 sctx->stat.last_physical = physical;
3423 spin_unlock(&sctx->stat_lock);
3424 if (stop_loop)
3425 break;
3426 }
3427 out:
3428 /* push queued extents */
3429 scrub_submit(sctx);
3430 mutex_lock(&sctx->wr_ctx.wr_lock);
3431 scrub_wr_submit(sctx);
3432 mutex_unlock(&sctx->wr_ctx.wr_lock);
3433
3434 blk_finish_plug(&plug);
3435 btrfs_free_path(path);
3436 btrfs_free_path(ppath);
3437 return ret < 0 ? ret : 0;
3438 }
3439
3440 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3441 struct btrfs_device *scrub_dev,
3442 u64 chunk_offset, u64 length,
3443 u64 dev_offset,
3444 struct btrfs_block_group_cache *cache,
3445 int is_dev_replace)
3446 {
3447 struct btrfs_fs_info *fs_info = sctx->fs_info;
3448 struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
3449 struct map_lookup *map;
3450 struct extent_map *em;
3451 int i;
3452 int ret = 0;
3453
3454 read_lock(&map_tree->map_tree.lock);
3455 em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
3456 read_unlock(&map_tree->map_tree.lock);
3457
3458 if (!em) {
3459 /*
3460 * Might have been an unused block group deleted by the cleaner
3461 * kthread or relocation.
3462 */
3463 spin_lock(&cache->lock);
3464 if (!cache->removed)
3465 ret = -EINVAL;
3466 spin_unlock(&cache->lock);
3467
3468 return ret;
3469 }
3470
3471 map = em->map_lookup;
3472 if (em->start != chunk_offset)
3473 goto out;
3474
3475 if (em->len < length)
3476 goto out;
3477
3478 for (i = 0; i < map->num_stripes; ++i) {
3479 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3480 map->stripes[i].physical == dev_offset) {
3481 ret = scrub_stripe(sctx, map, scrub_dev, i,
3482 chunk_offset, length,
3483 is_dev_replace);
3484 if (ret)
3485 goto out;
3486 }
3487 }
3488 out:
3489 free_extent_map(em);
3490
3491 return ret;
3492 }
3493
3494 static noinline_for_stack
3495 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3496 struct btrfs_device *scrub_dev, u64 start, u64 end,
3497 int is_dev_replace)
3498 {
3499 struct btrfs_dev_extent *dev_extent = NULL;
3500 struct btrfs_path *path;
3501 struct btrfs_fs_info *fs_info = sctx->fs_info;
3502 struct btrfs_root *root = fs_info->dev_root;
3503 u64 length;
3504 u64 chunk_offset;
3505 int ret = 0;
3506 int ro_set;
3507 int slot;
3508 struct extent_buffer *l;
3509 struct btrfs_key key;
3510 struct btrfs_key found_key;
3511 struct btrfs_block_group_cache *cache;
3512 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3513
3514 path = btrfs_alloc_path();
3515 if (!path)
3516 return -ENOMEM;
3517
3518 path->reada = READA_FORWARD;
3519 path->search_commit_root = 1;
3520 path->skip_locking = 1;
3521
3522 key.objectid = scrub_dev->devid;
3523 key.offset = 0ull;
3524 key.type = BTRFS_DEV_EXTENT_KEY;
3525
3526 while (1) {
3527 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3528 if (ret < 0)
3529 break;
3530 if (ret > 0) {
3531 if (path->slots[0] >=
3532 btrfs_header_nritems(path->nodes[0])) {
3533 ret = btrfs_next_leaf(root, path);
3534 if (ret < 0)
3535 break;
3536 if (ret > 0) {
3537 ret = 0;
3538 break;
3539 }
3540 } else {
3541 ret = 0;
3542 }
3543 }
3544
3545 l = path->nodes[0];
3546 slot = path->slots[0];
3547
3548 btrfs_item_key_to_cpu(l, &found_key, slot);
3549
3550 if (found_key.objectid != scrub_dev->devid)
3551 break;
3552
3553 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3554 break;
3555
3556 if (found_key.offset >= end)
3557 break;
3558
3559 if (found_key.offset < key.offset)
3560 break;
3561
3562 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3563 length = btrfs_dev_extent_length(l, dev_extent);
3564
3565 if (found_key.offset + length <= start)
3566 goto skip;
3567
3568 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3569
3570 /*
3571 * get a reference on the corresponding block group to prevent
3572 * the chunk from going away while we scrub it
3573 */
3574 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3575
3576 /* some chunks are removed but not committed to disk yet,
3577 * continue scrubbing */
3578 if (!cache)
3579 goto skip;
3580
3581 /*
3582 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3583 * to avoid deadlock caused by:
3584 * btrfs_inc_block_group_ro()
3585 * -> btrfs_wait_for_commit()
3586 * -> btrfs_commit_transaction()
3587 * -> btrfs_scrub_pause()
3588 */
3589 scrub_pause_on(fs_info);
3590 ret = btrfs_inc_block_group_ro(fs_info, cache);
3591 if (!ret && is_dev_replace) {
3592 /*
3593 * If we are doing a device replace wait for any tasks
3594 * that started dellaloc right before we set the block
3595 * group to RO mode, as they might have just allocated
3596 * an extent from it or decided they could do a nocow
3597 * write. And if any such tasks did that, wait for their
3598 * ordered extents to complete and then commit the
3599 * current transaction, so that we can later see the new
3600 * extent items in the extent tree - the ordered extents
3601 * create delayed data references (for cow writes) when
3602 * they complete, which will be run and insert the
3603 * corresponding extent items into the extent tree when
3604 * we commit the transaction they used when running
3605 * inode.c:btrfs_finish_ordered_io(). We later use
3606 * the commit root of the extent tree to find extents
3607 * to copy from the srcdev into the tgtdev, and we don't
3608 * want to miss any new extents.
3609 */
3610 btrfs_wait_block_group_reservations(cache);
3611 btrfs_wait_nocow_writers(cache);
3612 ret = btrfs_wait_ordered_roots(fs_info, -1,
3613 cache->key.objectid,
3614 cache->key.offset);
3615 if (ret > 0) {
3616 struct btrfs_trans_handle *trans;
3617
3618 trans = btrfs_join_transaction(root);
3619 if (IS_ERR(trans))
3620 ret = PTR_ERR(trans);
3621 else
3622 ret = btrfs_commit_transaction(trans);
3623 if (ret) {
3624 scrub_pause_off(fs_info);
3625 btrfs_put_block_group(cache);
3626 break;
3627 }
3628 }
3629 }
3630 scrub_pause_off(fs_info);
3631
3632 if (ret == 0) {
3633 ro_set = 1;
3634 } else if (ret == -ENOSPC) {
3635 /*
3636 * btrfs_inc_block_group_ro return -ENOSPC when it
3637 * failed in creating new chunk for metadata.
3638 * It is not a problem for scrub/replace, because
3639 * metadata are always cowed, and our scrub paused
3640 * commit_transactions.
3641 */
3642 ro_set = 0;
3643 } else {
3644 btrfs_warn(fs_info,
3645 "failed setting block group ro, ret=%d\n",
3646 ret);
3647 btrfs_put_block_group(cache);
3648 break;
3649 }
3650
3651 btrfs_dev_replace_lock(&fs_info->dev_replace, 1);
3652 dev_replace->cursor_right = found_key.offset + length;
3653 dev_replace->cursor_left = found_key.offset;
3654 dev_replace->item_needs_writeback = 1;
3655 btrfs_dev_replace_unlock(&fs_info->dev_replace, 1);
3656 ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3657 found_key.offset, cache, is_dev_replace);
3658
3659 /*
3660 * flush, submit all pending read and write bios, afterwards
3661 * wait for them.
3662 * Note that in the dev replace case, a read request causes
3663 * write requests that are submitted in the read completion
3664 * worker. Therefore in the current situation, it is required
3665 * that all write requests are flushed, so that all read and
3666 * write requests are really completed when bios_in_flight
3667 * changes to 0.
3668 */
3669 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
3670 scrub_submit(sctx);
3671 mutex_lock(&sctx->wr_ctx.wr_lock);
3672 scrub_wr_submit(sctx);
3673 mutex_unlock(&sctx->wr_ctx.wr_lock);
3674
3675 wait_event(sctx->list_wait,
3676 atomic_read(&sctx->bios_in_flight) == 0);
3677
3678 scrub_pause_on(fs_info);
3679
3680 /*
3681 * must be called before we decrease @scrub_paused.
3682 * make sure we don't block transaction commit while
3683 * we are waiting pending workers finished.
3684 */
3685 wait_event(sctx->list_wait,
3686 atomic_read(&sctx->workers_pending) == 0);
3687 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
3688
3689 scrub_pause_off(fs_info);
3690
3691 btrfs_dev_replace_lock(&fs_info->dev_replace, 1);
3692 dev_replace->cursor_left = dev_replace->cursor_right;
3693 dev_replace->item_needs_writeback = 1;
3694 btrfs_dev_replace_unlock(&fs_info->dev_replace, 1);
3695
3696 if (ro_set)
3697 btrfs_dec_block_group_ro(cache);
3698
3699 /*
3700 * We might have prevented the cleaner kthread from deleting
3701 * this block group if it was already unused because we raced
3702 * and set it to RO mode first. So add it back to the unused
3703 * list, otherwise it might not ever be deleted unless a manual
3704 * balance is triggered or it becomes used and unused again.
3705 */
3706 spin_lock(&cache->lock);
3707 if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3708 btrfs_block_group_used(&cache->item) == 0) {
3709 spin_unlock(&cache->lock);
3710 spin_lock(&fs_info->unused_bgs_lock);
3711 if (list_empty(&cache->bg_list)) {
3712 btrfs_get_block_group(cache);
3713 list_add_tail(&cache->bg_list,
3714 &fs_info->unused_bgs);
3715 }
3716 spin_unlock(&fs_info->unused_bgs_lock);
3717 } else {
3718 spin_unlock(&cache->lock);
3719 }
3720
3721 btrfs_put_block_group(cache);
3722 if (ret)
3723 break;
3724 if (is_dev_replace &&
3725 atomic64_read(&dev_replace->num_write_errors) > 0) {
3726 ret = -EIO;
3727 break;
3728 }
3729 if (sctx->stat.malloc_errors > 0) {
3730 ret = -ENOMEM;
3731 break;
3732 }
3733 skip:
3734 key.offset = found_key.offset + length;
3735 btrfs_release_path(path);
3736 }
3737
3738 btrfs_free_path(path);
3739
3740 return ret;
3741 }
3742
3743 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3744 struct btrfs_device *scrub_dev)
3745 {
3746 int i;
3747 u64 bytenr;
3748 u64 gen;
3749 int ret;
3750 struct btrfs_fs_info *fs_info = sctx->fs_info;
3751
3752 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
3753 return -EIO;
3754
3755 /* Seed devices of a new filesystem has their own generation. */
3756 if (scrub_dev->fs_devices != fs_info->fs_devices)
3757 gen = scrub_dev->generation;
3758 else
3759 gen = fs_info->last_trans_committed;
3760
3761 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3762 bytenr = btrfs_sb_offset(i);
3763 if (bytenr + BTRFS_SUPER_INFO_SIZE >
3764 scrub_dev->commit_total_bytes)
3765 break;
3766
3767 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3768 scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3769 NULL, 1, bytenr);
3770 if (ret)
3771 return ret;
3772 }
3773 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3774
3775 return 0;
3776 }
3777
3778 /*
3779 * get a reference count on fs_info->scrub_workers. start worker if necessary
3780 */
3781 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
3782 int is_dev_replace)
3783 {
3784 unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
3785 int max_active = fs_info->thread_pool_size;
3786
3787 if (fs_info->scrub_workers_refcnt == 0) {
3788 if (is_dev_replace)
3789 fs_info->scrub_workers =
3790 btrfs_alloc_workqueue(fs_info, "scrub", flags,
3791 1, 4);
3792 else
3793 fs_info->scrub_workers =
3794 btrfs_alloc_workqueue(fs_info, "scrub", flags,
3795 max_active, 4);
3796 if (!fs_info->scrub_workers)
3797 goto fail_scrub_workers;
3798
3799 fs_info->scrub_wr_completion_workers =
3800 btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
3801 max_active, 2);
3802 if (!fs_info->scrub_wr_completion_workers)
3803 goto fail_scrub_wr_completion_workers;
3804
3805 fs_info->scrub_nocow_workers =
3806 btrfs_alloc_workqueue(fs_info, "scrubnc", flags, 1, 0);
3807 if (!fs_info->scrub_nocow_workers)
3808 goto fail_scrub_nocow_workers;
3809 fs_info->scrub_parity_workers =
3810 btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
3811 max_active, 2);
3812 if (!fs_info->scrub_parity_workers)
3813 goto fail_scrub_parity_workers;
3814 }
3815 ++fs_info->scrub_workers_refcnt;
3816 return 0;
3817
3818 fail_scrub_parity_workers:
3819 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
3820 fail_scrub_nocow_workers:
3821 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3822 fail_scrub_wr_completion_workers:
3823 btrfs_destroy_workqueue(fs_info->scrub_workers);
3824 fail_scrub_workers:
3825 return -ENOMEM;
3826 }
3827
3828 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
3829 {
3830 if (--fs_info->scrub_workers_refcnt == 0) {
3831 btrfs_destroy_workqueue(fs_info->scrub_workers);
3832 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3833 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
3834 btrfs_destroy_workqueue(fs_info->scrub_parity_workers);
3835 }
3836 WARN_ON(fs_info->scrub_workers_refcnt < 0);
3837 }
3838
3839 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
3840 u64 end, struct btrfs_scrub_progress *progress,
3841 int readonly, int is_dev_replace)
3842 {
3843 struct scrub_ctx *sctx;
3844 int ret;
3845 struct btrfs_device *dev;
3846 struct rcu_string *name;
3847
3848 if (btrfs_fs_closing(fs_info))
3849 return -EINVAL;
3850
3851 if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
3852 /*
3853 * in this case scrub is unable to calculate the checksum
3854 * the way scrub is implemented. Do not handle this
3855 * situation at all because it won't ever happen.
3856 */
3857 btrfs_err(fs_info,
3858 "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
3859 fs_info->nodesize,
3860 BTRFS_STRIPE_LEN);
3861 return -EINVAL;
3862 }
3863
3864 if (fs_info->sectorsize != PAGE_SIZE) {
3865 /* not supported for data w/o checksums */
3866 btrfs_err_rl(fs_info,
3867 "scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails",
3868 fs_info->sectorsize, PAGE_SIZE);
3869 return -EINVAL;
3870 }
3871
3872 if (fs_info->nodesize >
3873 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
3874 fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
3875 /*
3876 * would exhaust the array bounds of pagev member in
3877 * struct scrub_block
3878 */
3879 btrfs_err(fs_info,
3880 "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
3881 fs_info->nodesize,
3882 SCRUB_MAX_PAGES_PER_BLOCK,
3883 fs_info->sectorsize,
3884 SCRUB_MAX_PAGES_PER_BLOCK);
3885 return -EINVAL;
3886 }
3887
3888
3889 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3890 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
3891 if (!dev || (dev->missing && !is_dev_replace)) {
3892 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3893 return -ENODEV;
3894 }
3895
3896 if (!is_dev_replace && !readonly && !dev->writeable) {
3897 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3898 rcu_read_lock();
3899 name = rcu_dereference(dev->name);
3900 btrfs_err(fs_info, "scrub: device %s is not writable",
3901 name->str);
3902 rcu_read_unlock();
3903 return -EROFS;
3904 }
3905
3906 mutex_lock(&fs_info->scrub_lock);
3907 if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
3908 mutex_unlock(&fs_info->scrub_lock);
3909 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3910 return -EIO;
3911 }
3912
3913 btrfs_dev_replace_lock(&fs_info->dev_replace, 0);
3914 if (dev->scrub_device ||
3915 (!is_dev_replace &&
3916 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
3917 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
3918 mutex_unlock(&fs_info->scrub_lock);
3919 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3920 return -EINPROGRESS;
3921 }
3922 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
3923
3924 ret = scrub_workers_get(fs_info, is_dev_replace);
3925 if (ret) {
3926 mutex_unlock(&fs_info->scrub_lock);
3927 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3928 return ret;
3929 }
3930
3931 sctx = scrub_setup_ctx(dev, is_dev_replace);
3932 if (IS_ERR(sctx)) {
3933 mutex_unlock(&fs_info->scrub_lock);
3934 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3935 scrub_workers_put(fs_info);
3936 return PTR_ERR(sctx);
3937 }
3938 sctx->readonly = readonly;
3939 dev->scrub_device = sctx;
3940 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3941
3942 /*
3943 * checking @scrub_pause_req here, we can avoid
3944 * race between committing transaction and scrubbing.
3945 */
3946 __scrub_blocked_if_needed(fs_info);
3947 atomic_inc(&fs_info->scrubs_running);
3948 mutex_unlock(&fs_info->scrub_lock);
3949
3950 if (!is_dev_replace) {
3951 /*
3952 * by holding device list mutex, we can
3953 * kick off writing super in log tree sync.
3954 */
3955 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3956 ret = scrub_supers(sctx, dev);
3957 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3958 }
3959
3960 if (!ret)
3961 ret = scrub_enumerate_chunks(sctx, dev, start, end,
3962 is_dev_replace);
3963
3964 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3965 atomic_dec(&fs_info->scrubs_running);
3966 wake_up(&fs_info->scrub_pause_wait);
3967
3968 wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3969
3970 if (progress)
3971 memcpy(progress, &sctx->stat, sizeof(*progress));
3972
3973 mutex_lock(&fs_info->scrub_lock);
3974 dev->scrub_device = NULL;
3975 scrub_workers_put(fs_info);
3976 mutex_unlock(&fs_info->scrub_lock);
3977
3978 scrub_put_ctx(sctx);
3979
3980 return ret;
3981 }
3982
3983 void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
3984 {
3985 mutex_lock(&fs_info->scrub_lock);
3986 atomic_inc(&fs_info->scrub_pause_req);
3987 while (atomic_read(&fs_info->scrubs_paused) !=
3988 atomic_read(&fs_info->scrubs_running)) {
3989 mutex_unlock(&fs_info->scrub_lock);
3990 wait_event(fs_info->scrub_pause_wait,
3991 atomic_read(&fs_info->scrubs_paused) ==
3992 atomic_read(&fs_info->scrubs_running));
3993 mutex_lock(&fs_info->scrub_lock);
3994 }
3995 mutex_unlock(&fs_info->scrub_lock);
3996 }
3997
3998 void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
3999 {
4000 atomic_dec(&fs_info->scrub_pause_req);
4001 wake_up(&fs_info->scrub_pause_wait);
4002 }
4003
4004 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
4005 {
4006 mutex_lock(&fs_info->scrub_lock);
4007 if (!atomic_read(&fs_info->scrubs_running)) {
4008 mutex_unlock(&fs_info->scrub_lock);
4009 return -ENOTCONN;
4010 }
4011
4012 atomic_inc(&fs_info->scrub_cancel_req);
4013 while (atomic_read(&fs_info->scrubs_running)) {
4014 mutex_unlock(&fs_info->scrub_lock);
4015 wait_event(fs_info->scrub_pause_wait,
4016 atomic_read(&fs_info->scrubs_running) == 0);
4017 mutex_lock(&fs_info->scrub_lock);
4018 }
4019 atomic_dec(&fs_info->scrub_cancel_req);
4020 mutex_unlock(&fs_info->scrub_lock);
4021
4022 return 0;
4023 }
4024
4025 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
4026 struct btrfs_device *dev)
4027 {
4028 struct scrub_ctx *sctx;
4029
4030 mutex_lock(&fs_info->scrub_lock);
4031 sctx = dev->scrub_device;
4032 if (!sctx) {
4033 mutex_unlock(&fs_info->scrub_lock);
4034 return -ENOTCONN;
4035 }
4036 atomic_inc(&sctx->cancel_req);
4037 while (dev->scrub_device) {
4038 mutex_unlock(&fs_info->scrub_lock);
4039 wait_event(fs_info->scrub_pause_wait,
4040 dev->scrub_device == NULL);
4041 mutex_lock(&fs_info->scrub_lock);
4042 }
4043 mutex_unlock(&fs_info->scrub_lock);
4044
4045 return 0;
4046 }
4047
4048 int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4049 struct btrfs_scrub_progress *progress)
4050 {
4051 struct btrfs_device *dev;
4052 struct scrub_ctx *sctx = NULL;
4053
4054 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4055 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
4056 if (dev)
4057 sctx = dev->scrub_device;
4058 if (sctx)
4059 memcpy(progress, &sctx->stat, sizeof(*progress));
4060 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4061
4062 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4063 }
4064
4065 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4066 u64 extent_logical, u64 extent_len,
4067 u64 *extent_physical,
4068 struct btrfs_device **extent_dev,
4069 int *extent_mirror_num)
4070 {
4071 u64 mapped_length;
4072 struct btrfs_bio *bbio = NULL;
4073 int ret;
4074
4075 mapped_length = extent_len;
4076 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4077 &mapped_length, &bbio, 0);
4078 if (ret || !bbio || mapped_length < extent_len ||
4079 !bbio->stripes[0].dev->bdev) {
4080 btrfs_put_bbio(bbio);
4081 return;
4082 }
4083
4084 *extent_physical = bbio->stripes[0].physical;
4085 *extent_mirror_num = bbio->mirror_num;
4086 *extent_dev = bbio->stripes[0].dev;
4087 btrfs_put_bbio(bbio);
4088 }
4089
4090 static int scrub_setup_wr_ctx(struct scrub_wr_ctx *wr_ctx,
4091 struct btrfs_device *dev,
4092 int is_dev_replace)
4093 {
4094 WARN_ON(wr_ctx->wr_curr_bio != NULL);
4095
4096 mutex_init(&wr_ctx->wr_lock);
4097 wr_ctx->wr_curr_bio = NULL;
4098 if (!is_dev_replace)
4099 return 0;
4100
4101 WARN_ON(!dev->bdev);
4102 wr_ctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
4103 wr_ctx->tgtdev = dev;
4104 atomic_set(&wr_ctx->flush_all_writes, 0);
4105 return 0;
4106 }
4107
4108 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
4109 {
4110 mutex_lock(&wr_ctx->wr_lock);
4111 kfree(wr_ctx->wr_curr_bio);
4112 wr_ctx->wr_curr_bio = NULL;
4113 mutex_unlock(&wr_ctx->wr_lock);
4114 }
4115
4116 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
4117 int mirror_num, u64 physical_for_dev_replace)
4118 {
4119 struct scrub_copy_nocow_ctx *nocow_ctx;
4120 struct btrfs_fs_info *fs_info = sctx->fs_info;
4121
4122 nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
4123 if (!nocow_ctx) {
4124 spin_lock(&sctx->stat_lock);
4125 sctx->stat.malloc_errors++;
4126 spin_unlock(&sctx->stat_lock);
4127 return -ENOMEM;
4128 }
4129
4130 scrub_pending_trans_workers_inc(sctx);
4131
4132 nocow_ctx->sctx = sctx;
4133 nocow_ctx->logical = logical;
4134 nocow_ctx->len = len;
4135 nocow_ctx->mirror_num = mirror_num;
4136 nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
4137 btrfs_init_work(&nocow_ctx->work, btrfs_scrubnc_helper,
4138 copy_nocow_pages_worker, NULL, NULL);
4139 INIT_LIST_HEAD(&nocow_ctx->inodes);
4140 btrfs_queue_work(fs_info->scrub_nocow_workers,
4141 &nocow_ctx->work);
4142
4143 return 0;
4144 }
4145
4146 static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
4147 {
4148 struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
4149 struct scrub_nocow_inode *nocow_inode;
4150
4151 nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
4152 if (!nocow_inode)
4153 return -ENOMEM;
4154 nocow_inode->inum = inum;
4155 nocow_inode->offset = offset;
4156 nocow_inode->root = root;
4157 list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
4158 return 0;
4159 }
4160
4161 #define COPY_COMPLETE 1
4162
4163 static void copy_nocow_pages_worker(struct btrfs_work *work)
4164 {
4165 struct scrub_copy_nocow_ctx *nocow_ctx =
4166 container_of(work, struct scrub_copy_nocow_ctx, work);
4167 struct scrub_ctx *sctx = nocow_ctx->sctx;
4168 struct btrfs_fs_info *fs_info = sctx->fs_info;
4169 struct btrfs_root *root = fs_info->extent_root;
4170 u64 logical = nocow_ctx->logical;
4171 u64 len = nocow_ctx->len;
4172 int mirror_num = nocow_ctx->mirror_num;
4173 u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4174 int ret;
4175 struct btrfs_trans_handle *trans = NULL;
4176 struct btrfs_path *path;
4177 int not_written = 0;
4178
4179 path = btrfs_alloc_path();
4180 if (!path) {
4181 spin_lock(&sctx->stat_lock);
4182 sctx->stat.malloc_errors++;
4183 spin_unlock(&sctx->stat_lock);
4184 not_written = 1;
4185 goto out;
4186 }
4187
4188 trans = btrfs_join_transaction(root);
4189 if (IS_ERR(trans)) {
4190 not_written = 1;
4191 goto out;
4192 }
4193
4194 ret = iterate_inodes_from_logical(logical, fs_info, path,
4195 record_inode_for_nocow, nocow_ctx);
4196 if (ret != 0 && ret != -ENOENT) {
4197 btrfs_warn(fs_info,
4198 "iterate_inodes_from_logical() failed: log %llu, phys %llu, len %llu, mir %u, ret %d",
4199 logical, physical_for_dev_replace, len, mirror_num,
4200 ret);
4201 not_written = 1;
4202 goto out;
4203 }
4204
4205 btrfs_end_transaction(trans);
4206 trans = NULL;
4207 while (!list_empty(&nocow_ctx->inodes)) {
4208 struct scrub_nocow_inode *entry;
4209 entry = list_first_entry(&nocow_ctx->inodes,
4210 struct scrub_nocow_inode,
4211 list);
4212 list_del_init(&entry->list);
4213 ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
4214 entry->root, nocow_ctx);
4215 kfree(entry);
4216 if (ret == COPY_COMPLETE) {
4217 ret = 0;
4218 break;
4219 } else if (ret) {
4220 break;
4221 }
4222 }
4223 out:
4224 while (!list_empty(&nocow_ctx->inodes)) {
4225 struct scrub_nocow_inode *entry;
4226 entry = list_first_entry(&nocow_ctx->inodes,
4227 struct scrub_nocow_inode,
4228 list);
4229 list_del_init(&entry->list);
4230 kfree(entry);
4231 }
4232 if (trans && !IS_ERR(trans))
4233 btrfs_end_transaction(trans);
4234 if (not_written)
4235 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
4236 num_uncorrectable_read_errors);
4237
4238 btrfs_free_path(path);
4239 kfree(nocow_ctx);
4240
4241 scrub_pending_trans_workers_dec(sctx);
4242 }
4243
4244 static int check_extent_to_block(struct btrfs_inode *inode, u64 start, u64 len,
4245 u64 logical)
4246 {
4247 struct extent_state *cached_state = NULL;
4248 struct btrfs_ordered_extent *ordered;
4249 struct extent_io_tree *io_tree;
4250 struct extent_map *em;
4251 u64 lockstart = start, lockend = start + len - 1;
4252 int ret = 0;
4253
4254 io_tree = &inode->io_tree;
4255
4256 lock_extent_bits(io_tree, lockstart, lockend, &cached_state);
4257 ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
4258 if (ordered) {
4259 btrfs_put_ordered_extent(ordered);
4260 ret = 1;
4261 goto out_unlock;
4262 }
4263
4264 em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
4265 if (IS_ERR(em)) {
4266 ret = PTR_ERR(em);
4267 goto out_unlock;
4268 }
4269
4270 /*
4271 * This extent does not actually cover the logical extent anymore,
4272 * move on to the next inode.
4273 */
4274 if (em->block_start > logical ||
4275 em->block_start + em->block_len < logical + len) {
4276 free_extent_map(em);
4277 ret = 1;
4278 goto out_unlock;
4279 }
4280 free_extent_map(em);
4281
4282 out_unlock:
4283 unlock_extent_cached(io_tree, lockstart, lockend, &cached_state,
4284 GFP_NOFS);
4285 return ret;
4286 }
4287
4288 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
4289 struct scrub_copy_nocow_ctx *nocow_ctx)
4290 {
4291 struct btrfs_fs_info *fs_info = nocow_ctx->sctx->fs_info;
4292 struct btrfs_key key;
4293 struct inode *inode;
4294 struct page *page;
4295 struct btrfs_root *local_root;
4296 struct extent_io_tree *io_tree;
4297 u64 physical_for_dev_replace;
4298 u64 nocow_ctx_logical;
4299 u64 len = nocow_ctx->len;
4300 unsigned long index;
4301 int srcu_index;
4302 int ret = 0;
4303 int err = 0;
4304
4305 key.objectid = root;
4306 key.type = BTRFS_ROOT_ITEM_KEY;
4307 key.offset = (u64)-1;
4308
4309 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
4310
4311 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
4312 if (IS_ERR(local_root)) {
4313 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4314 return PTR_ERR(local_root);
4315 }
4316
4317 key.type = BTRFS_INODE_ITEM_KEY;
4318 key.objectid = inum;
4319 key.offset = 0;
4320 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
4321 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4322 if (IS_ERR(inode))
4323 return PTR_ERR(inode);
4324
4325 /* Avoid truncate/dio/punch hole.. */
4326 inode_lock(inode);
4327 inode_dio_wait(inode);
4328
4329 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4330 io_tree = &BTRFS_I(inode)->io_tree;
4331 nocow_ctx_logical = nocow_ctx->logical;
4332
4333 ret = check_extent_to_block(BTRFS_I(inode), offset, len,
4334 nocow_ctx_logical);
4335 if (ret) {
4336 ret = ret > 0 ? 0 : ret;
4337 goto out;
4338 }
4339
4340 while (len >= PAGE_SIZE) {
4341 index = offset >> PAGE_SHIFT;
4342 again:
4343 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
4344 if (!page) {
4345 btrfs_err(fs_info, "find_or_create_page() failed");
4346 ret = -ENOMEM;
4347 goto out;
4348 }
4349
4350 if (PageUptodate(page)) {
4351 if (PageDirty(page))
4352 goto next_page;
4353 } else {
4354 ClearPageError(page);
4355 err = extent_read_full_page(io_tree, page,
4356 btrfs_get_extent,
4357 nocow_ctx->mirror_num);
4358 if (err) {
4359 ret = err;
4360 goto next_page;
4361 }
4362
4363 lock_page(page);
4364 /*
4365 * If the page has been remove from the page cache,
4366 * the data on it is meaningless, because it may be
4367 * old one, the new data may be written into the new
4368 * page in the page cache.
4369 */
4370 if (page->mapping != inode->i_mapping) {
4371 unlock_page(page);
4372 put_page(page);
4373 goto again;
4374 }
4375 if (!PageUptodate(page)) {
4376 ret = -EIO;
4377 goto next_page;
4378 }
4379 }
4380
4381 ret = check_extent_to_block(BTRFS_I(inode), offset, len,
4382 nocow_ctx_logical);
4383 if (ret) {
4384 ret = ret > 0 ? 0 : ret;
4385 goto next_page;
4386 }
4387
4388 err = write_page_nocow(nocow_ctx->sctx,
4389 physical_for_dev_replace, page);
4390 if (err)
4391 ret = err;
4392 next_page:
4393 unlock_page(page);
4394 put_page(page);
4395
4396 if (ret)
4397 break;
4398
4399 offset += PAGE_SIZE;
4400 physical_for_dev_replace += PAGE_SIZE;
4401 nocow_ctx_logical += PAGE_SIZE;
4402 len -= PAGE_SIZE;
4403 }
4404 ret = COPY_COMPLETE;
4405 out:
4406 inode_unlock(inode);
4407 iput(inode);
4408 return ret;
4409 }
4410
4411 static int write_page_nocow(struct scrub_ctx *sctx,
4412 u64 physical_for_dev_replace, struct page *page)
4413 {
4414 struct bio *bio;
4415 struct btrfs_device *dev;
4416 int ret;
4417
4418 dev = sctx->wr_ctx.tgtdev;
4419 if (!dev)
4420 return -EIO;
4421 if (!dev->bdev) {
4422 btrfs_warn_rl(dev->fs_info,
4423 "scrub write_page_nocow(bdev == NULL) is unexpected");
4424 return -EIO;
4425 }
4426 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
4427 if (!bio) {
4428 spin_lock(&sctx->stat_lock);
4429 sctx->stat.malloc_errors++;
4430 spin_unlock(&sctx->stat_lock);
4431 return -ENOMEM;
4432 }
4433 bio->bi_iter.bi_size = 0;
4434 bio->bi_iter.bi_sector = physical_for_dev_replace >> 9;
4435 bio->bi_bdev = dev->bdev;
4436 bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
4437 ret = bio_add_page(bio, page, PAGE_SIZE, 0);
4438 if (ret != PAGE_SIZE) {
4439 leave_with_eio:
4440 bio_put(bio);
4441 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
4442 return -EIO;
4443 }
4444
4445 if (btrfsic_submit_bio_wait(bio))
4446 goto leave_with_eio;
4447
4448 bio_put(bio);
4449 return 0;
4450 }
ubuntu-bionic-kernel mirror