Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2011, 2012 STRATO. All rights reserved.
4 */
5
6#include <linux/blkdev.h>
7#include <linux/ratelimit.h>
8#include <linux/sched/mm.h>
9#include <crypto/hash.h>
10#include "ctree.h"
11#include "discard.h"
12#include "volumes.h"
13#include "disk-io.h"
14#include "ordered-data.h"
15#include "transaction.h"
16#include "backref.h"
17#include "extent_io.h"
18#include "dev-replace.h"
19#include "raid56.h"
20#include "block-group.h"
21#include "zoned.h"
22#include "fs.h"
23#include "accessors.h"
24#include "file-item.h"
25#include "scrub.h"
26#include "raid-stripe-tree.h"
27
28/*
29 * This is only the first step towards a full-features scrub. It reads all
30 * extent and super block and verifies the checksums. In case a bad checksum
31 * is found or the extent cannot be read, good data will be written back if
32 * any can be found.
33 *
34 * Future enhancements:
35 * - In case an unrepairable extent is encountered, track which files are
36 * affected and report them
37 * - track and record media errors, throw out bad devices
38 * - add a mode to also read unallocated space
39 */
40
41struct scrub_ctx;
42
43/*
44 * The following value only influences the performance.
45 *
46 * This determines how many stripes would be submitted in one go,
47 * which is 512KiB (BTRFS_STRIPE_LEN * SCRUB_STRIPES_PER_GROUP).
48 */
49#define SCRUB_STRIPES_PER_GROUP 8
50
51/*
52 * How many groups we have for each sctx.
53 *
54 * This would be 8M per device, the same value as the old scrub in-flight bios
55 * size limit.
56 */
57#define SCRUB_GROUPS_PER_SCTX 16
58
59#define SCRUB_TOTAL_STRIPES (SCRUB_GROUPS_PER_SCTX * SCRUB_STRIPES_PER_GROUP)
60
61/*
62 * The following value times PAGE_SIZE needs to be large enough to match the
63 * largest node/leaf/sector size that shall be supported.
64 */
65#define SCRUB_MAX_SECTORS_PER_BLOCK (BTRFS_MAX_METADATA_BLOCKSIZE / SZ_4K)
66
67/* Represent one sector and its needed info to verify the content. */
68struct scrub_sector_verification {
69 bool is_metadata;
70
71 union {
72 /*
73 * Csum pointer for data csum verification. Should point to a
74 * sector csum inside scrub_stripe::csums.
75 *
76 * NULL if this data sector has no csum.
77 */
78 u8 *csum;
79
80 /*
81 * Extra info for metadata verification. All sectors inside a
82 * tree block share the same generation.
83 */
84 u64 generation;
85 };
86};
87
88enum scrub_stripe_flags {
89 /* Set when @mirror_num, @dev, @physical and @logical are set. */
90 SCRUB_STRIPE_FLAG_INITIALIZED,
91
92 /* Set when the read-repair is finished. */
93 SCRUB_STRIPE_FLAG_REPAIR_DONE,
94
95 /*
96 * Set for data stripes if it's triggered from P/Q stripe.
97 * During such scrub, we should not report errors in data stripes, nor
98 * update the accounting.
99 */
100 SCRUB_STRIPE_FLAG_NO_REPORT,
101};
102
103#define SCRUB_STRIPE_PAGES (BTRFS_STRIPE_LEN / PAGE_SIZE)
104
105/*
106 * Represent one contiguous range with a length of BTRFS_STRIPE_LEN.
107 */
108struct scrub_stripe {
109 struct scrub_ctx *sctx;
110 struct btrfs_block_group *bg;
111
112 struct page *pages[SCRUB_STRIPE_PAGES];
113 struct scrub_sector_verification *sectors;
114
115 struct btrfs_device *dev;
116 u64 logical;
117 u64 physical;
118
119 u16 mirror_num;
120
121 /* Should be BTRFS_STRIPE_LEN / sectorsize. */
122 u16 nr_sectors;
123
124 /*
125 * How many data/meta extents are in this stripe. Only for scrub status
126 * reporting purposes.
127 */
128 u16 nr_data_extents;
129 u16 nr_meta_extents;
130
131 atomic_t pending_io;
132 wait_queue_head_t io_wait;
133 wait_queue_head_t repair_wait;
134
135 /*
136 * Indicate the states of the stripe. Bits are defined in
137 * scrub_stripe_flags enum.
138 */
139 unsigned long state;
140
141 /* Indicate which sectors are covered by extent items. */
142 unsigned long extent_sector_bitmap;
143
144 /*
145 * The errors hit during the initial read of the stripe.
146 *
147 * Would be utilized for error reporting and repair.
148 *
149 * The remaining init_nr_* records the number of errors hit, only used
150 * by error reporting.
151 */
152 unsigned long init_error_bitmap;
153 unsigned int init_nr_io_errors;
154 unsigned int init_nr_csum_errors;
155 unsigned int init_nr_meta_errors;
156
157 /*
158 * The following error bitmaps are all for the current status.
159 * Every time we submit a new read, these bitmaps may be updated.
160 *
161 * error_bitmap = io_error_bitmap | csum_error_bitmap | meta_error_bitmap;
162 *
163 * IO and csum errors can happen for both metadata and data.
164 */
165 unsigned long error_bitmap;
166 unsigned long io_error_bitmap;
167 unsigned long csum_error_bitmap;
168 unsigned long meta_error_bitmap;
169
170 /* For writeback (repair or replace) error reporting. */
171 unsigned long write_error_bitmap;
172
173 /* Writeback can be concurrent, thus we need to protect the bitmap. */
174 spinlock_t write_error_lock;
175
176 /*
177 * Checksum for the whole stripe if this stripe is inside a data block
178 * group.
179 */
180 u8 *csums;
181
182 struct work_struct work;
183};
184
185struct scrub_ctx {
186 struct scrub_stripe stripes[SCRUB_TOTAL_STRIPES];
187 struct scrub_stripe *raid56_data_stripes;
188 struct btrfs_fs_info *fs_info;
189 struct btrfs_path extent_path;
190 struct btrfs_path csum_path;
191 int first_free;
192 int cur_stripe;
193 atomic_t cancel_req;
194 int readonly;
195
196 /* State of IO submission throttling affecting the associated device */
197 ktime_t throttle_deadline;
198 u64 throttle_sent;
199
200 int is_dev_replace;
201 u64 write_pointer;
202
203 struct mutex wr_lock;
204 struct btrfs_device *wr_tgtdev;
205
206 /*
207 * statistics
208 */
209 struct btrfs_scrub_progress stat;
210 spinlock_t stat_lock;
211
212 /*
213 * Use a ref counter to avoid use-after-free issues. Scrub workers
214 * decrement bios_in_flight and workers_pending and then do a wakeup
215 * on the list_wait wait queue. We must ensure the main scrub task
216 * doesn't free the scrub context before or while the workers are
217 * doing the wakeup() call.
218 */
219 refcount_t refs;
220};
221
222struct scrub_warning {
223 struct btrfs_path *path;
224 u64 extent_item_size;
225 const char *errstr;
226 u64 physical;
227 u64 logical;
228 struct btrfs_device *dev;
229};
230
231static void release_scrub_stripe(struct scrub_stripe *stripe)
232{
233 if (!stripe)
234 return;
235
236 for (int i = 0; i < SCRUB_STRIPE_PAGES; i++) {
237 if (stripe->pages[i])
238 __free_page(stripe->pages[i]);
239 stripe->pages[i] = NULL;
240 }
241 kfree(stripe->sectors);
242 kfree(stripe->csums);
243 stripe->sectors = NULL;
244 stripe->csums = NULL;
245 stripe->sctx = NULL;
246 stripe->state = 0;
247}
248
249static int init_scrub_stripe(struct btrfs_fs_info *fs_info,
250 struct scrub_stripe *stripe)
251{
252 int ret;
253
254 memset(stripe, 0, sizeof(*stripe));
255
256 stripe->nr_sectors = BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits;
257 stripe->state = 0;
258
259 init_waitqueue_head(&stripe->io_wait);
260 init_waitqueue_head(&stripe->repair_wait);
261 atomic_set(&stripe->pending_io, 0);
262 spin_lock_init(&stripe->write_error_lock);
263
264 ret = btrfs_alloc_page_array(SCRUB_STRIPE_PAGES, stripe->pages, 0);
265 if (ret < 0)
266 goto error;
267
268 stripe->sectors = kcalloc(stripe->nr_sectors,
269 sizeof(struct scrub_sector_verification),
270 GFP_KERNEL);
271 if (!stripe->sectors)
272 goto error;
273
274 stripe->csums = kcalloc(BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits,
275 fs_info->csum_size, GFP_KERNEL);
276 if (!stripe->csums)
277 goto error;
278 return 0;
279error:
280 release_scrub_stripe(stripe);
281 return -ENOMEM;
282}
283
284static void wait_scrub_stripe_io(struct scrub_stripe *stripe)
285{
286 wait_event(stripe->io_wait, atomic_read(&stripe->pending_io) == 0);
287}
288
289static void scrub_put_ctx(struct scrub_ctx *sctx);
290
291static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
292{
293 while (atomic_read(&fs_info->scrub_pause_req)) {
294 mutex_unlock(&fs_info->scrub_lock);
295 wait_event(fs_info->scrub_pause_wait,
296 atomic_read(&fs_info->scrub_pause_req) == 0);
297 mutex_lock(&fs_info->scrub_lock);
298 }
299}
300
301static void scrub_pause_on(struct btrfs_fs_info *fs_info)
302{
303 atomic_inc(&fs_info->scrubs_paused);
304 wake_up(&fs_info->scrub_pause_wait);
305}
306
307static void scrub_pause_off(struct btrfs_fs_info *fs_info)
308{
309 mutex_lock(&fs_info->scrub_lock);
310 __scrub_blocked_if_needed(fs_info);
311 atomic_dec(&fs_info->scrubs_paused);
312 mutex_unlock(&fs_info->scrub_lock);
313
314 wake_up(&fs_info->scrub_pause_wait);
315}
316
317static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
318{
319 scrub_pause_on(fs_info);
320 scrub_pause_off(fs_info);
321}
322
323static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
324{
325 int i;
326
327 if (!sctx)
328 return;
329
330 for (i = 0; i < SCRUB_TOTAL_STRIPES; i++)
331 release_scrub_stripe(&sctx->stripes[i]);
332
333 kvfree(sctx);
334}
335
336static void scrub_put_ctx(struct scrub_ctx *sctx)
337{
338 if (refcount_dec_and_test(&sctx->refs))
339 scrub_free_ctx(sctx);
340}
341
342static noinline_for_stack struct scrub_ctx *scrub_setup_ctx(
343 struct btrfs_fs_info *fs_info, int is_dev_replace)
344{
345 struct scrub_ctx *sctx;
346 int i;
347
348 /* Since sctx has inline 128 stripes, it can go beyond 64K easily. Use
349 * kvzalloc().
350 */
351 sctx = kvzalloc(sizeof(*sctx), GFP_KERNEL);
352 if (!sctx)
353 goto nomem;
354 refcount_set(&sctx->refs, 1);
355 sctx->is_dev_replace = is_dev_replace;
356 sctx->fs_info = fs_info;
357 sctx->extent_path.search_commit_root = 1;
358 sctx->extent_path.skip_locking = 1;
359 sctx->csum_path.search_commit_root = 1;
360 sctx->csum_path.skip_locking = 1;
361 for (i = 0; i < SCRUB_TOTAL_STRIPES; i++) {
362 int ret;
363
364 ret = init_scrub_stripe(fs_info, &sctx->stripes[i]);
365 if (ret < 0)
366 goto nomem;
367 sctx->stripes[i].sctx = sctx;
368 }
369 sctx->first_free = 0;
370 atomic_set(&sctx->cancel_req, 0);
371
372 spin_lock_init(&sctx->stat_lock);
373 sctx->throttle_deadline = 0;
374
375 mutex_init(&sctx->wr_lock);
376 if (is_dev_replace) {
377 WARN_ON(!fs_info->dev_replace.tgtdev);
378 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
379 }
380
381 return sctx;
382
383nomem:
384 scrub_free_ctx(sctx);
385 return ERR_PTR(-ENOMEM);
386}
387
388static int scrub_print_warning_inode(u64 inum, u64 offset, u64 num_bytes,
389 u64 root, void *warn_ctx)
390{
391 u32 nlink;
392 int ret;
393 int i;
394 unsigned nofs_flag;
395 struct extent_buffer *eb;
396 struct btrfs_inode_item *inode_item;
397 struct scrub_warning *swarn = warn_ctx;
398 struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
399 struct inode_fs_paths *ipath = NULL;
400 struct btrfs_root *local_root;
401 struct btrfs_key key;
402
403 local_root = btrfs_get_fs_root(fs_info, root, true);
404 if (IS_ERR(local_root)) {
405 ret = PTR_ERR(local_root);
406 goto err;
407 }
408
409 /*
410 * this makes the path point to (inum INODE_ITEM ioff)
411 */
412 key.objectid = inum;
413 key.type = BTRFS_INODE_ITEM_KEY;
414 key.offset = 0;
415
416 ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
417 if (ret) {
418 btrfs_put_root(local_root);
419 btrfs_release_path(swarn->path);
420 goto err;
421 }
422
423 eb = swarn->path->nodes[0];
424 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
425 struct btrfs_inode_item);
426 nlink = btrfs_inode_nlink(eb, inode_item);
427 btrfs_release_path(swarn->path);
428
429 /*
430 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
431 * uses GFP_NOFS in this context, so we keep it consistent but it does
432 * not seem to be strictly necessary.
433 */
434 nofs_flag = memalloc_nofs_save();
435 ipath = init_ipath(4096, local_root, swarn->path);
436 memalloc_nofs_restore(nofs_flag);
437 if (IS_ERR(ipath)) {
438 btrfs_put_root(local_root);
439 ret = PTR_ERR(ipath);
440 ipath = NULL;
441 goto err;
442 }
443 ret = paths_from_inode(inum, ipath);
444
445 if (ret < 0)
446 goto err;
447
448 /*
449 * we deliberately ignore the bit ipath might have been too small to
450 * hold all of the paths here
451 */
452 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
453 btrfs_warn_in_rcu(fs_info,
454"%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %u, links %u (path: %s)",
455 swarn->errstr, swarn->logical,
456 btrfs_dev_name(swarn->dev),
457 swarn->physical,
458 root, inum, offset,
459 fs_info->sectorsize, nlink,
460 (char *)(unsigned long)ipath->fspath->val[i]);
461
462 btrfs_put_root(local_root);
463 free_ipath(ipath);
464 return 0;
465
466err:
467 btrfs_warn_in_rcu(fs_info,
468 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
469 swarn->errstr, swarn->logical,
470 btrfs_dev_name(swarn->dev),
471 swarn->physical,
472 root, inum, offset, ret);
473
474 free_ipath(ipath);
475 return 0;
476}
477
478static void scrub_print_common_warning(const char *errstr, struct btrfs_device *dev,
479 bool is_super, u64 logical, u64 physical)
480{
481 struct btrfs_fs_info *fs_info = dev->fs_info;
482 struct btrfs_path *path;
483 struct btrfs_key found_key;
484 struct extent_buffer *eb;
485 struct btrfs_extent_item *ei;
486 struct scrub_warning swarn;
487 u64 flags = 0;
488 u32 item_size;
489 int ret;
490
491 /* Super block error, no need to search extent tree. */
492 if (is_super) {
493 btrfs_warn_in_rcu(fs_info, "%s on device %s, physical %llu",
494 errstr, btrfs_dev_name(dev), physical);
495 return;
496 }
497 path = btrfs_alloc_path();
498 if (!path)
499 return;
500
501 swarn.physical = physical;
502 swarn.logical = logical;
503 swarn.errstr = errstr;
504 swarn.dev = NULL;
505
506 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
507 &flags);
508 if (ret < 0)
509 goto out;
510
511 swarn.extent_item_size = found_key.offset;
512
513 eb = path->nodes[0];
514 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
515 item_size = btrfs_item_size(eb, path->slots[0]);
516
517 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
518 unsigned long ptr = 0;
519 u8 ref_level;
520 u64 ref_root;
521
522 while (true) {
523 ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
524 item_size, &ref_root,
525 &ref_level);
526 if (ret < 0) {
527 btrfs_warn(fs_info,
528 "failed to resolve tree backref for logical %llu: %d",
529 swarn.logical, ret);
530 break;
531 }
532 if (ret > 0)
533 break;
534 btrfs_warn_in_rcu(fs_info,
535"%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
536 errstr, swarn.logical, btrfs_dev_name(dev),
537 swarn.physical, (ref_level ? "node" : "leaf"),
538 ref_level, ref_root);
539 }
540 btrfs_release_path(path);
541 } else {
542 struct btrfs_backref_walk_ctx ctx = { 0 };
543
544 btrfs_release_path(path);
545
546 ctx.bytenr = found_key.objectid;
547 ctx.extent_item_pos = swarn.logical - found_key.objectid;
548 ctx.fs_info = fs_info;
549
550 swarn.path = path;
551 swarn.dev = dev;
552
553 iterate_extent_inodes(&ctx, true, scrub_print_warning_inode, &swarn);
554 }
555
556out:
557 btrfs_free_path(path);
558}
559
560static int fill_writer_pointer_gap(struct scrub_ctx *sctx, u64 physical)
561{
562 int ret = 0;
563 u64 length;
564
565 if (!btrfs_is_zoned(sctx->fs_info))
566 return 0;
567
568 if (!btrfs_dev_is_sequential(sctx->wr_tgtdev, physical))
569 return 0;
570
571 if (sctx->write_pointer < physical) {
572 length = physical - sctx->write_pointer;
573
574 ret = btrfs_zoned_issue_zeroout(sctx->wr_tgtdev,
575 sctx->write_pointer, length);
576 if (!ret)
577 sctx->write_pointer = physical;
578 }
579 return ret;
580}
581
582static struct page *scrub_stripe_get_page(struct scrub_stripe *stripe, int sector_nr)
583{
584 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
585 int page_index = (sector_nr << fs_info->sectorsize_bits) >> PAGE_SHIFT;
586
587 return stripe->pages[page_index];
588}
589
590static unsigned int scrub_stripe_get_page_offset(struct scrub_stripe *stripe,
591 int sector_nr)
592{
593 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
594
595 return offset_in_page(sector_nr << fs_info->sectorsize_bits);
596}
597
598static void scrub_verify_one_metadata(struct scrub_stripe *stripe, int sector_nr)
599{
600 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
601 const u32 sectors_per_tree = fs_info->nodesize >> fs_info->sectorsize_bits;
602 const u64 logical = stripe->logical + (sector_nr << fs_info->sectorsize_bits);
603 const struct page *first_page = scrub_stripe_get_page(stripe, sector_nr);
604 const unsigned int first_off = scrub_stripe_get_page_offset(stripe, sector_nr);
605 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
606 u8 on_disk_csum[BTRFS_CSUM_SIZE];
607 u8 calculated_csum[BTRFS_CSUM_SIZE];
608 struct btrfs_header *header;
609
610 /*
611 * Here we don't have a good way to attach the pages (and subpages)
612 * to a dummy extent buffer, thus we have to directly grab the members
613 * from pages.
614 */
615 header = (struct btrfs_header *)(page_address(first_page) + first_off);
616 memcpy(on_disk_csum, header->csum, fs_info->csum_size);
617
618 if (logical != btrfs_stack_header_bytenr(header)) {
619 bitmap_set(&stripe->csum_error_bitmap, sector_nr, sectors_per_tree);
620 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
621 btrfs_warn_rl(fs_info,
622 "tree block %llu mirror %u has bad bytenr, has %llu want %llu",
623 logical, stripe->mirror_num,
624 btrfs_stack_header_bytenr(header), logical);
625 return;
626 }
627 if (memcmp(header->fsid, fs_info->fs_devices->metadata_uuid,
628 BTRFS_FSID_SIZE) != 0) {
629 bitmap_set(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
630 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
631 btrfs_warn_rl(fs_info,
632 "tree block %llu mirror %u has bad fsid, has %pU want %pU",
633 logical, stripe->mirror_num,
634 header->fsid, fs_info->fs_devices->fsid);
635 return;
636 }
637 if (memcmp(header->chunk_tree_uuid, fs_info->chunk_tree_uuid,
638 BTRFS_UUID_SIZE) != 0) {
639 bitmap_set(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
640 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
641 btrfs_warn_rl(fs_info,
642 "tree block %llu mirror %u has bad chunk tree uuid, has %pU want %pU",
643 logical, stripe->mirror_num,
644 header->chunk_tree_uuid, fs_info->chunk_tree_uuid);
645 return;
646 }
647
648 /* Now check tree block csum. */
649 shash->tfm = fs_info->csum_shash;
650 crypto_shash_init(shash);
651 crypto_shash_update(shash, page_address(first_page) + first_off +
652 BTRFS_CSUM_SIZE, fs_info->sectorsize - BTRFS_CSUM_SIZE);
653
654 for (int i = sector_nr + 1; i < sector_nr + sectors_per_tree; i++) {
655 struct page *page = scrub_stripe_get_page(stripe, i);
656 unsigned int page_off = scrub_stripe_get_page_offset(stripe, i);
657
658 crypto_shash_update(shash, page_address(page) + page_off,
659 fs_info->sectorsize);
660 }
661
662 crypto_shash_final(shash, calculated_csum);
663 if (memcmp(calculated_csum, on_disk_csum, fs_info->csum_size) != 0) {
664 bitmap_set(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
665 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
666 btrfs_warn_rl(fs_info,
667 "tree block %llu mirror %u has bad csum, has " CSUM_FMT " want " CSUM_FMT,
668 logical, stripe->mirror_num,
669 CSUM_FMT_VALUE(fs_info->csum_size, on_disk_csum),
670 CSUM_FMT_VALUE(fs_info->csum_size, calculated_csum));
671 return;
672 }
673 if (stripe->sectors[sector_nr].generation !=
674 btrfs_stack_header_generation(header)) {
675 bitmap_set(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
676 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
677 btrfs_warn_rl(fs_info,
678 "tree block %llu mirror %u has bad generation, has %llu want %llu",
679 logical, stripe->mirror_num,
680 btrfs_stack_header_generation(header),
681 stripe->sectors[sector_nr].generation);
682 return;
683 }
684 bitmap_clear(&stripe->error_bitmap, sector_nr, sectors_per_tree);
685 bitmap_clear(&stripe->csum_error_bitmap, sector_nr, sectors_per_tree);
686 bitmap_clear(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
687}
688
689static void scrub_verify_one_sector(struct scrub_stripe *stripe, int sector_nr)
690{
691 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
692 struct scrub_sector_verification *sector = &stripe->sectors[sector_nr];
693 const u32 sectors_per_tree = fs_info->nodesize >> fs_info->sectorsize_bits;
694 struct page *page = scrub_stripe_get_page(stripe, sector_nr);
695 unsigned int pgoff = scrub_stripe_get_page_offset(stripe, sector_nr);
696 u8 csum_buf[BTRFS_CSUM_SIZE];
697 int ret;
698
699 ASSERT(sector_nr >= 0 && sector_nr < stripe->nr_sectors);
700
701 /* Sector not utilized, skip it. */
702 if (!test_bit(sector_nr, &stripe->extent_sector_bitmap))
703 return;
704
705 /* IO error, no need to check. */
706 if (test_bit(sector_nr, &stripe->io_error_bitmap))
707 return;
708
709 /* Metadata, verify the full tree block. */
710 if (sector->is_metadata) {
711 /*
712 * Check if the tree block crosses the stripe boundary. If
713 * crossed the boundary, we cannot verify it but only give a
714 * warning.
715 *
716 * This can only happen on a very old filesystem where chunks
717 * are not ensured to be stripe aligned.
718 */
719 if (unlikely(sector_nr + sectors_per_tree > stripe->nr_sectors)) {
720 btrfs_warn_rl(fs_info,
721 "tree block at %llu crosses stripe boundary %llu",
722 stripe->logical +
723 (sector_nr << fs_info->sectorsize_bits),
724 stripe->logical);
725 return;
726 }
727 scrub_verify_one_metadata(stripe, sector_nr);
728 return;
729 }
730
731 /*
732 * Data is easier, we just verify the data csum (if we have it). For
733 * cases without csum, we have no other choice but to trust it.
734 */
735 if (!sector->csum) {
736 clear_bit(sector_nr, &stripe->error_bitmap);
737 return;
738 }
739
740 ret = btrfs_check_sector_csum(fs_info, page, pgoff, csum_buf, sector->csum);
741 if (ret < 0) {
742 set_bit(sector_nr, &stripe->csum_error_bitmap);
743 set_bit(sector_nr, &stripe->error_bitmap);
744 } else {
745 clear_bit(sector_nr, &stripe->csum_error_bitmap);
746 clear_bit(sector_nr, &stripe->error_bitmap);
747 }
748}
749
750/* Verify specified sectors of a stripe. */
751static void scrub_verify_one_stripe(struct scrub_stripe *stripe, unsigned long bitmap)
752{
753 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
754 const u32 sectors_per_tree = fs_info->nodesize >> fs_info->sectorsize_bits;
755 int sector_nr;
756
757 for_each_set_bit(sector_nr, &bitmap, stripe->nr_sectors) {
758 scrub_verify_one_sector(stripe, sector_nr);
759 if (stripe->sectors[sector_nr].is_metadata)
760 sector_nr += sectors_per_tree - 1;
761 }
762}
763
764static int calc_sector_number(struct scrub_stripe *stripe, struct bio_vec *first_bvec)
765{
766 int i;
767
768 for (i = 0; i < stripe->nr_sectors; i++) {
769 if (scrub_stripe_get_page(stripe, i) == first_bvec->bv_page &&
770 scrub_stripe_get_page_offset(stripe, i) == first_bvec->bv_offset)
771 break;
772 }
773 ASSERT(i < stripe->nr_sectors);
774 return i;
775}
776
777/*
778 * Repair read is different to the regular read:
779 *
780 * - Only reads the failed sectors
781 * - May have extra blocksize limits
782 */
783static void scrub_repair_read_endio(struct btrfs_bio *bbio)
784{
785 struct scrub_stripe *stripe = bbio->private;
786 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
787 struct bio_vec *bvec;
788 int sector_nr = calc_sector_number(stripe, bio_first_bvec_all(&bbio->bio));
789 u32 bio_size = 0;
790 int i;
791
792 ASSERT(sector_nr < stripe->nr_sectors);
793
794 bio_for_each_bvec_all(bvec, &bbio->bio, i)
795 bio_size += bvec->bv_len;
796
797 if (bbio->bio.bi_status) {
798 bitmap_set(&stripe->io_error_bitmap, sector_nr,
799 bio_size >> fs_info->sectorsize_bits);
800 bitmap_set(&stripe->error_bitmap, sector_nr,
801 bio_size >> fs_info->sectorsize_bits);
802 } else {
803 bitmap_clear(&stripe->io_error_bitmap, sector_nr,
804 bio_size >> fs_info->sectorsize_bits);
805 }
806 bio_put(&bbio->bio);
807 if (atomic_dec_and_test(&stripe->pending_io))
808 wake_up(&stripe->io_wait);
809}
810
811static int calc_next_mirror(int mirror, int num_copies)
812{
813 ASSERT(mirror <= num_copies);
814 return (mirror + 1 > num_copies) ? 1 : mirror + 1;
815}
816
817static void scrub_stripe_submit_repair_read(struct scrub_stripe *stripe,
818 int mirror, int blocksize, bool wait)
819{
820 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
821 struct btrfs_bio *bbio = NULL;
822 const unsigned long old_error_bitmap = stripe->error_bitmap;
823 int i;
824
825 ASSERT(stripe->mirror_num >= 1);
826 ASSERT(atomic_read(&stripe->pending_io) == 0);
827
828 for_each_set_bit(i, &old_error_bitmap, stripe->nr_sectors) {
829 struct page *page;
830 int pgoff;
831 int ret;
832
833 page = scrub_stripe_get_page(stripe, i);
834 pgoff = scrub_stripe_get_page_offset(stripe, i);
835
836 /* The current sector cannot be merged, submit the bio. */
837 if (bbio && ((i > 0 && !test_bit(i - 1, &stripe->error_bitmap)) ||
838 bbio->bio.bi_iter.bi_size >= blocksize)) {
839 ASSERT(bbio->bio.bi_iter.bi_size);
840 atomic_inc(&stripe->pending_io);
841 btrfs_submit_bio(bbio, mirror);
842 if (wait)
843 wait_scrub_stripe_io(stripe);
844 bbio = NULL;
845 }
846
847 if (!bbio) {
848 bbio = btrfs_bio_alloc(stripe->nr_sectors, REQ_OP_READ,
849 fs_info, scrub_repair_read_endio, stripe);
850 bbio->bio.bi_iter.bi_sector = (stripe->logical +
851 (i << fs_info->sectorsize_bits)) >> SECTOR_SHIFT;
852 }
853
854 ret = bio_add_page(&bbio->bio, page, fs_info->sectorsize, pgoff);
855 ASSERT(ret == fs_info->sectorsize);
856 }
857 if (bbio) {
858 ASSERT(bbio->bio.bi_iter.bi_size);
859 atomic_inc(&stripe->pending_io);
860 btrfs_submit_bio(bbio, mirror);
861 if (wait)
862 wait_scrub_stripe_io(stripe);
863 }
864}
865
866static void scrub_stripe_report_errors(struct scrub_ctx *sctx,
867 struct scrub_stripe *stripe)
868{
869 static DEFINE_RATELIMIT_STATE(rs, DEFAULT_RATELIMIT_INTERVAL,
870 DEFAULT_RATELIMIT_BURST);
871 struct btrfs_fs_info *fs_info = sctx->fs_info;
872 struct btrfs_device *dev = NULL;
873 u64 physical = 0;
874 int nr_data_sectors = 0;
875 int nr_meta_sectors = 0;
876 int nr_nodatacsum_sectors = 0;
877 int nr_repaired_sectors = 0;
878 int sector_nr;
879
880 if (test_bit(SCRUB_STRIPE_FLAG_NO_REPORT, &stripe->state))
881 return;
882
883 /*
884 * Init needed infos for error reporting.
885 *
886 * Although our scrub_stripe infrastructure is mostly based on btrfs_submit_bio()
887 * thus no need for dev/physical, error reporting still needs dev and physical.
888 */
889 if (!bitmap_empty(&stripe->init_error_bitmap, stripe->nr_sectors)) {
890 u64 mapped_len = fs_info->sectorsize;
891 struct btrfs_io_context *bioc = NULL;
892 int stripe_index = stripe->mirror_num - 1;
893 int ret;
894
895 /* For scrub, our mirror_num should always start at 1. */
896 ASSERT(stripe->mirror_num >= 1);
897 ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
898 stripe->logical, &mapped_len, &bioc,
899 NULL, NULL);
900 /*
901 * If we failed, dev will be NULL, and later detailed reports
902 * will just be skipped.
903 */
904 if (ret < 0)
905 goto skip;
906 physical = bioc->stripes[stripe_index].physical;
907 dev = bioc->stripes[stripe_index].dev;
908 btrfs_put_bioc(bioc);
909 }
910
911skip:
912 for_each_set_bit(sector_nr, &stripe->extent_sector_bitmap, stripe->nr_sectors) {
913 bool repaired = false;
914
915 if (stripe->sectors[sector_nr].is_metadata) {
916 nr_meta_sectors++;
917 } else {
918 nr_data_sectors++;
919 if (!stripe->sectors[sector_nr].csum)
920 nr_nodatacsum_sectors++;
921 }
922
923 if (test_bit(sector_nr, &stripe->init_error_bitmap) &&
924 !test_bit(sector_nr, &stripe->error_bitmap)) {
925 nr_repaired_sectors++;
926 repaired = true;
927 }
928
929 /* Good sector from the beginning, nothing need to be done. */
930 if (!test_bit(sector_nr, &stripe->init_error_bitmap))
931 continue;
932
933 /*
934 * Report error for the corrupted sectors. If repaired, just
935 * output the message of repaired message.
936 */
937 if (repaired) {
938 if (dev) {
939 btrfs_err_rl_in_rcu(fs_info,
940 "fixed up error at logical %llu on dev %s physical %llu",
941 stripe->logical, btrfs_dev_name(dev),
942 physical);
943 } else {
944 btrfs_err_rl_in_rcu(fs_info,
945 "fixed up error at logical %llu on mirror %u",
946 stripe->logical, stripe->mirror_num);
947 }
948 continue;
949 }
950
951 /* The remaining are all for unrepaired. */
952 if (dev) {
953 btrfs_err_rl_in_rcu(fs_info,
954 "unable to fixup (regular) error at logical %llu on dev %s physical %llu",
955 stripe->logical, btrfs_dev_name(dev),
956 physical);
957 } else {
958 btrfs_err_rl_in_rcu(fs_info,
959 "unable to fixup (regular) error at logical %llu on mirror %u",
960 stripe->logical, stripe->mirror_num);
961 }
962
963 if (test_bit(sector_nr, &stripe->io_error_bitmap))
964 if (__ratelimit(&rs) && dev)
965 scrub_print_common_warning("i/o error", dev, false,
966 stripe->logical, physical);
967 if (test_bit(sector_nr, &stripe->csum_error_bitmap))
968 if (__ratelimit(&rs) && dev)
969 scrub_print_common_warning("checksum error", dev, false,
970 stripe->logical, physical);
971 if (test_bit(sector_nr, &stripe->meta_error_bitmap))
972 if (__ratelimit(&rs) && dev)
973 scrub_print_common_warning("header error", dev, false,
974 stripe->logical, physical);
975 }
976
977 spin_lock(&sctx->stat_lock);
978 sctx->stat.data_extents_scrubbed += stripe->nr_data_extents;
979 sctx->stat.tree_extents_scrubbed += stripe->nr_meta_extents;
980 sctx->stat.data_bytes_scrubbed += nr_data_sectors << fs_info->sectorsize_bits;
981 sctx->stat.tree_bytes_scrubbed += nr_meta_sectors << fs_info->sectorsize_bits;
982 sctx->stat.no_csum += nr_nodatacsum_sectors;
983 sctx->stat.read_errors += stripe->init_nr_io_errors;
984 sctx->stat.csum_errors += stripe->init_nr_csum_errors;
985 sctx->stat.verify_errors += stripe->init_nr_meta_errors;
986 sctx->stat.uncorrectable_errors +=
987 bitmap_weight(&stripe->error_bitmap, stripe->nr_sectors);
988 sctx->stat.corrected_errors += nr_repaired_sectors;
989 spin_unlock(&sctx->stat_lock);
990}
991
992static void scrub_write_sectors(struct scrub_ctx *sctx, struct scrub_stripe *stripe,
993 unsigned long write_bitmap, bool dev_replace);
994
995/*
996 * The main entrance for all read related scrub work, including:
997 *
998 * - Wait for the initial read to finish
999 * - Verify and locate any bad sectors
1000 * - Go through the remaining mirrors and try to read as large blocksize as
1001 * possible
1002 * - Go through all mirrors (including the failed mirror) sector-by-sector
1003 * - Submit writeback for repaired sectors
1004 *
1005 * Writeback for dev-replace does not happen here, it needs extra
1006 * synchronization for zoned devices.
1007 */
1008static void scrub_stripe_read_repair_worker(struct work_struct *work)
1009{
1010 struct scrub_stripe *stripe = container_of(work, struct scrub_stripe, work);
1011 struct scrub_ctx *sctx = stripe->sctx;
1012 struct btrfs_fs_info *fs_info = sctx->fs_info;
1013 int num_copies = btrfs_num_copies(fs_info, stripe->bg->start,
1014 stripe->bg->length);
1015 int mirror;
1016 int i;
1017
1018 ASSERT(stripe->mirror_num > 0);
1019
1020 wait_scrub_stripe_io(stripe);
1021 scrub_verify_one_stripe(stripe, stripe->extent_sector_bitmap);
1022 /* Save the initial failed bitmap for later repair and report usage. */
1023 stripe->init_error_bitmap = stripe->error_bitmap;
1024 stripe->init_nr_io_errors = bitmap_weight(&stripe->io_error_bitmap,
1025 stripe->nr_sectors);
1026 stripe->init_nr_csum_errors = bitmap_weight(&stripe->csum_error_bitmap,
1027 stripe->nr_sectors);
1028 stripe->init_nr_meta_errors = bitmap_weight(&stripe->meta_error_bitmap,
1029 stripe->nr_sectors);
1030
1031 if (bitmap_empty(&stripe->init_error_bitmap, stripe->nr_sectors))
1032 goto out;
1033
1034 /*
1035 * Try all remaining mirrors.
1036 *
1037 * Here we still try to read as large block as possible, as this is
1038 * faster and we have extra safety nets to rely on.
1039 */
1040 for (mirror = calc_next_mirror(stripe->mirror_num, num_copies);
1041 mirror != stripe->mirror_num;
1042 mirror = calc_next_mirror(mirror, num_copies)) {
1043 const unsigned long old_error_bitmap = stripe->error_bitmap;
1044
1045 scrub_stripe_submit_repair_read(stripe, mirror,
1046 BTRFS_STRIPE_LEN, false);
1047 wait_scrub_stripe_io(stripe);
1048 scrub_verify_one_stripe(stripe, old_error_bitmap);
1049 if (bitmap_empty(&stripe->error_bitmap, stripe->nr_sectors))
1050 goto out;
1051 }
1052
1053 /*
1054 * Last safety net, try re-checking all mirrors, including the failed
1055 * one, sector-by-sector.
1056 *
1057 * As if one sector failed the drive's internal csum, the whole read
1058 * containing the offending sector would be marked as error.
1059 * Thus here we do sector-by-sector read.
1060 *
1061 * This can be slow, thus we only try it as the last resort.
1062 */
1063
1064 for (i = 0, mirror = stripe->mirror_num;
1065 i < num_copies;
1066 i++, mirror = calc_next_mirror(mirror, num_copies)) {
1067 const unsigned long old_error_bitmap = stripe->error_bitmap;
1068
1069 scrub_stripe_submit_repair_read(stripe, mirror,
1070 fs_info->sectorsize, true);
1071 wait_scrub_stripe_io(stripe);
1072 scrub_verify_one_stripe(stripe, old_error_bitmap);
1073 if (bitmap_empty(&stripe->error_bitmap, stripe->nr_sectors))
1074 goto out;
1075 }
1076out:
1077 /*
1078 * Submit the repaired sectors. For zoned case, we cannot do repair
1079 * in-place, but queue the bg to be relocated.
1080 */
1081 if (btrfs_is_zoned(fs_info)) {
1082 if (!bitmap_empty(&stripe->error_bitmap, stripe->nr_sectors))
1083 btrfs_repair_one_zone(fs_info, sctx->stripes[0].bg->start);
1084 } else if (!sctx->readonly) {
1085 unsigned long repaired;
1086
1087 bitmap_andnot(&repaired, &stripe->init_error_bitmap,
1088 &stripe->error_bitmap, stripe->nr_sectors);
1089 scrub_write_sectors(sctx, stripe, repaired, false);
1090 wait_scrub_stripe_io(stripe);
1091 }
1092
1093 scrub_stripe_report_errors(sctx, stripe);
1094 set_bit(SCRUB_STRIPE_FLAG_REPAIR_DONE, &stripe->state);
1095 wake_up(&stripe->repair_wait);
1096}
1097
1098static void scrub_read_endio(struct btrfs_bio *bbio)
1099{
1100 struct scrub_stripe *stripe = bbio->private;
1101 struct bio_vec *bvec;
1102 int sector_nr = calc_sector_number(stripe, bio_first_bvec_all(&bbio->bio));
1103 int num_sectors;
1104 u32 bio_size = 0;
1105 int i;
1106
1107 ASSERT(sector_nr < stripe->nr_sectors);
1108 bio_for_each_bvec_all(bvec, &bbio->bio, i)
1109 bio_size += bvec->bv_len;
1110 num_sectors = bio_size >> stripe->bg->fs_info->sectorsize_bits;
1111
1112 if (bbio->bio.bi_status) {
1113 bitmap_set(&stripe->io_error_bitmap, sector_nr, num_sectors);
1114 bitmap_set(&stripe->error_bitmap, sector_nr, num_sectors);
1115 } else {
1116 bitmap_clear(&stripe->io_error_bitmap, sector_nr, num_sectors);
1117 }
1118 bio_put(&bbio->bio);
1119 if (atomic_dec_and_test(&stripe->pending_io)) {
1120 wake_up(&stripe->io_wait);
1121 INIT_WORK(&stripe->work, scrub_stripe_read_repair_worker);
1122 queue_work(stripe->bg->fs_info->scrub_workers, &stripe->work);
1123 }
1124}
1125
1126static void scrub_write_endio(struct btrfs_bio *bbio)
1127{
1128 struct scrub_stripe *stripe = bbio->private;
1129 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
1130 struct bio_vec *bvec;
1131 int sector_nr = calc_sector_number(stripe, bio_first_bvec_all(&bbio->bio));
1132 u32 bio_size = 0;
1133 int i;
1134
1135 bio_for_each_bvec_all(bvec, &bbio->bio, i)
1136 bio_size += bvec->bv_len;
1137
1138 if (bbio->bio.bi_status) {
1139 unsigned long flags;
1140
1141 spin_lock_irqsave(&stripe->write_error_lock, flags);
1142 bitmap_set(&stripe->write_error_bitmap, sector_nr,
1143 bio_size >> fs_info->sectorsize_bits);
1144 spin_unlock_irqrestore(&stripe->write_error_lock, flags);
1145 }
1146 bio_put(&bbio->bio);
1147
1148 if (atomic_dec_and_test(&stripe->pending_io))
1149 wake_up(&stripe->io_wait);
1150}
1151
1152static void scrub_submit_write_bio(struct scrub_ctx *sctx,
1153 struct scrub_stripe *stripe,
1154 struct btrfs_bio *bbio, bool dev_replace)
1155{
1156 struct btrfs_fs_info *fs_info = sctx->fs_info;
1157 u32 bio_len = bbio->bio.bi_iter.bi_size;
1158 u32 bio_off = (bbio->bio.bi_iter.bi_sector << SECTOR_SHIFT) -
1159 stripe->logical;
1160
1161 fill_writer_pointer_gap(sctx, stripe->physical + bio_off);
1162 atomic_inc(&stripe->pending_io);
1163 btrfs_submit_repair_write(bbio, stripe->mirror_num, dev_replace);
1164 if (!btrfs_is_zoned(fs_info))
1165 return;
1166 /*
1167 * For zoned writeback, queue depth must be 1, thus we must wait for
1168 * the write to finish before the next write.
1169 */
1170 wait_scrub_stripe_io(stripe);
1171
1172 /*
1173 * And also need to update the write pointer if write finished
1174 * successfully.
1175 */
1176 if (!test_bit(bio_off >> fs_info->sectorsize_bits,
1177 &stripe->write_error_bitmap))
1178 sctx->write_pointer += bio_len;
1179}
1180
1181/*
1182 * Submit the write bio(s) for the sectors specified by @write_bitmap.
1183 *
1184 * Here we utilize btrfs_submit_repair_write(), which has some extra benefits:
1185 *
1186 * - Only needs logical bytenr and mirror_num
1187 * Just like the scrub read path
1188 *
1189 * - Would only result in writes to the specified mirror
1190 * Unlike the regular writeback path, which would write back to all stripes
1191 *
1192 * - Handle dev-replace and read-repair writeback differently
1193 */
1194static void scrub_write_sectors(struct scrub_ctx *sctx, struct scrub_stripe *stripe,
1195 unsigned long write_bitmap, bool dev_replace)
1196{
1197 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
1198 struct btrfs_bio *bbio = NULL;
1199 int sector_nr;
1200
1201 for_each_set_bit(sector_nr, &write_bitmap, stripe->nr_sectors) {
1202 struct page *page = scrub_stripe_get_page(stripe, sector_nr);
1203 unsigned int pgoff = scrub_stripe_get_page_offset(stripe, sector_nr);
1204 int ret;
1205
1206 /* We should only writeback sectors covered by an extent. */
1207 ASSERT(test_bit(sector_nr, &stripe->extent_sector_bitmap));
1208
1209 /* Cannot merge with previous sector, submit the current one. */
1210 if (bbio && sector_nr && !test_bit(sector_nr - 1, &write_bitmap)) {
1211 scrub_submit_write_bio(sctx, stripe, bbio, dev_replace);
1212 bbio = NULL;
1213 }
1214 if (!bbio) {
1215 bbio = btrfs_bio_alloc(stripe->nr_sectors, REQ_OP_WRITE,
1216 fs_info, scrub_write_endio, stripe);
1217 bbio->bio.bi_iter.bi_sector = (stripe->logical +
1218 (sector_nr << fs_info->sectorsize_bits)) >>
1219 SECTOR_SHIFT;
1220 }
1221 ret = bio_add_page(&bbio->bio, page, fs_info->sectorsize, pgoff);
1222 ASSERT(ret == fs_info->sectorsize);
1223 }
1224 if (bbio)
1225 scrub_submit_write_bio(sctx, stripe, bbio, dev_replace);
1226}
1227
1228/*
1229 * Throttling of IO submission, bandwidth-limit based, the timeslice is 1
1230 * second. Limit can be set via /sys/fs/UUID/devinfo/devid/scrub_speed_max.
1231 */
1232static void scrub_throttle_dev_io(struct scrub_ctx *sctx, struct btrfs_device *device,
1233 unsigned int bio_size)
1234{
1235 const int time_slice = 1000;
1236 s64 delta;
1237 ktime_t now;
1238 u32 div;
1239 u64 bwlimit;
1240
1241 bwlimit = READ_ONCE(device->scrub_speed_max);
1242 if (bwlimit == 0)
1243 return;
1244
1245 /*
1246 * Slice is divided into intervals when the IO is submitted, adjust by
1247 * bwlimit and maximum of 64 intervals.
1248 */
1249 div = max_t(u32, 1, (u32)(bwlimit / (16 * 1024 * 1024)));
1250 div = min_t(u32, 64, div);
1251
1252 /* Start new epoch, set deadline */
1253 now = ktime_get();
1254 if (sctx->throttle_deadline == 0) {
1255 sctx->throttle_deadline = ktime_add_ms(now, time_slice / div);
1256 sctx->throttle_sent = 0;
1257 }
1258
1259 /* Still in the time to send? */
1260 if (ktime_before(now, sctx->throttle_deadline)) {
1261 /* If current bio is within the limit, send it */
1262 sctx->throttle_sent += bio_size;
1263 if (sctx->throttle_sent <= div_u64(bwlimit, div))
1264 return;
1265
1266 /* We're over the limit, sleep until the rest of the slice */
1267 delta = ktime_ms_delta(sctx->throttle_deadline, now);
1268 } else {
1269 /* New request after deadline, start new epoch */
1270 delta = 0;
1271 }
1272
1273 if (delta) {
1274 long timeout;
1275
1276 timeout = div_u64(delta * HZ, 1000);
1277 schedule_timeout_interruptible(timeout);
1278 }
1279
1280 /* Next call will start the deadline period */
1281 sctx->throttle_deadline = 0;
1282}
1283
1284/*
1285 * Given a physical address, this will calculate it's
1286 * logical offset. if this is a parity stripe, it will return
1287 * the most left data stripe's logical offset.
1288 *
1289 * return 0 if it is a data stripe, 1 means parity stripe.
1290 */
1291static int get_raid56_logic_offset(u64 physical, int num,
1292 struct btrfs_chunk_map *map, u64 *offset,
1293 u64 *stripe_start)
1294{
1295 int i;
1296 int j = 0;
1297 u64 last_offset;
1298 const int data_stripes = nr_data_stripes(map);
1299
1300 last_offset = (physical - map->stripes[num].physical) * data_stripes;
1301 if (stripe_start)
1302 *stripe_start = last_offset;
1303
1304 *offset = last_offset;
1305 for (i = 0; i < data_stripes; i++) {
1306 u32 stripe_nr;
1307 u32 stripe_index;
1308 u32 rot;
1309
1310 *offset = last_offset + btrfs_stripe_nr_to_offset(i);
1311
1312 stripe_nr = (u32)(*offset >> BTRFS_STRIPE_LEN_SHIFT) / data_stripes;
1313
1314 /* Work out the disk rotation on this stripe-set */
1315 rot = stripe_nr % map->num_stripes;
1316 /* calculate which stripe this data locates */
1317 rot += i;
1318 stripe_index = rot % map->num_stripes;
1319 if (stripe_index == num)
1320 return 0;
1321 if (stripe_index < num)
1322 j++;
1323 }
1324 *offset = last_offset + btrfs_stripe_nr_to_offset(j);
1325 return 1;
1326}
1327
1328/*
1329 * Return 0 if the extent item range covers any byte of the range.
1330 * Return <0 if the extent item is before @search_start.
1331 * Return >0 if the extent item is after @start_start + @search_len.
1332 */
1333static int compare_extent_item_range(struct btrfs_path *path,
1334 u64 search_start, u64 search_len)
1335{
1336 struct btrfs_fs_info *fs_info = path->nodes[0]->fs_info;
1337 u64 len;
1338 struct btrfs_key key;
1339
1340 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1341 ASSERT(key.type == BTRFS_EXTENT_ITEM_KEY ||
1342 key.type == BTRFS_METADATA_ITEM_KEY);
1343 if (key.type == BTRFS_METADATA_ITEM_KEY)
1344 len = fs_info->nodesize;
1345 else
1346 len = key.offset;
1347
1348 if (key.objectid + len <= search_start)
1349 return -1;
1350 if (key.objectid >= search_start + search_len)
1351 return 1;
1352 return 0;
1353}
1354
1355/*
1356 * Locate one extent item which covers any byte in range
1357 * [@search_start, @search_start + @search_length)
1358 *
1359 * If the path is not initialized, we will initialize the search by doing
1360 * a btrfs_search_slot().
1361 * If the path is already initialized, we will use the path as the initial
1362 * slot, to avoid duplicated btrfs_search_slot() calls.
1363 *
1364 * NOTE: If an extent item starts before @search_start, we will still
1365 * return the extent item. This is for data extent crossing stripe boundary.
1366 *
1367 * Return 0 if we found such extent item, and @path will point to the extent item.
1368 * Return >0 if no such extent item can be found, and @path will be released.
1369 * Return <0 if hit fatal error, and @path will be released.
1370 */
1371static int find_first_extent_item(struct btrfs_root *extent_root,
1372 struct btrfs_path *path,
1373 u64 search_start, u64 search_len)
1374{
1375 struct btrfs_fs_info *fs_info = extent_root->fs_info;
1376 struct btrfs_key key;
1377 int ret;
1378
1379 /* Continue using the existing path */
1380 if (path->nodes[0])
1381 goto search_forward;
1382
1383 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1384 key.type = BTRFS_METADATA_ITEM_KEY;
1385 else
1386 key.type = BTRFS_EXTENT_ITEM_KEY;
1387 key.objectid = search_start;
1388 key.offset = (u64)-1;
1389
1390 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
1391 if (ret < 0)
1392 return ret;
1393
1394 ASSERT(ret > 0);
1395 /*
1396 * Here we intentionally pass 0 as @min_objectid, as there could be
1397 * an extent item starting before @search_start.
1398 */
1399 ret = btrfs_previous_extent_item(extent_root, path, 0);
1400 if (ret < 0)
1401 return ret;
1402 /*
1403 * No matter whether we have found an extent item, the next loop will
1404 * properly do every check on the key.
1405 */
1406search_forward:
1407 while (true) {
1408 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1409 if (key.objectid >= search_start + search_len)
1410 break;
1411 if (key.type != BTRFS_METADATA_ITEM_KEY &&
1412 key.type != BTRFS_EXTENT_ITEM_KEY)
1413 goto next;
1414
1415 ret = compare_extent_item_range(path, search_start, search_len);
1416 if (ret == 0)
1417 return ret;
1418 if (ret > 0)
1419 break;
1420next:
1421 ret = btrfs_next_item(extent_root, path);
1422 if (ret) {
1423 /* Either no more items or a fatal error. */
1424 btrfs_release_path(path);
1425 return ret;
1426 }
1427 }
1428 btrfs_release_path(path);
1429 return 1;
1430}
1431
1432static void get_extent_info(struct btrfs_path *path, u64 *extent_start_ret,
1433 u64 *size_ret, u64 *flags_ret, u64 *generation_ret)
1434{
1435 struct btrfs_key key;
1436 struct btrfs_extent_item *ei;
1437
1438 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1439 ASSERT(key.type == BTRFS_METADATA_ITEM_KEY ||
1440 key.type == BTRFS_EXTENT_ITEM_KEY);
1441 *extent_start_ret = key.objectid;
1442 if (key.type == BTRFS_METADATA_ITEM_KEY)
1443 *size_ret = path->nodes[0]->fs_info->nodesize;
1444 else
1445 *size_ret = key.offset;
1446 ei = btrfs_item_ptr(path->nodes[0], path->slots[0], struct btrfs_extent_item);
1447 *flags_ret = btrfs_extent_flags(path->nodes[0], ei);
1448 *generation_ret = btrfs_extent_generation(path->nodes[0], ei);
1449}
1450
1451static int sync_write_pointer_for_zoned(struct scrub_ctx *sctx, u64 logical,
1452 u64 physical, u64 physical_end)
1453{
1454 struct btrfs_fs_info *fs_info = sctx->fs_info;
1455 int ret = 0;
1456
1457 if (!btrfs_is_zoned(fs_info))
1458 return 0;
1459
1460 mutex_lock(&sctx->wr_lock);
1461 if (sctx->write_pointer < physical_end) {
1462 ret = btrfs_sync_zone_write_pointer(sctx->wr_tgtdev, logical,
1463 physical,
1464 sctx->write_pointer);
1465 if (ret)
1466 btrfs_err(fs_info,
1467 "zoned: failed to recover write pointer");
1468 }
1469 mutex_unlock(&sctx->wr_lock);
1470 btrfs_dev_clear_zone_empty(sctx->wr_tgtdev, physical);
1471
1472 return ret;
1473}
1474
1475static void fill_one_extent_info(struct btrfs_fs_info *fs_info,
1476 struct scrub_stripe *stripe,
1477 u64 extent_start, u64 extent_len,
1478 u64 extent_flags, u64 extent_gen)
1479{
1480 for (u64 cur_logical = max(stripe->logical, extent_start);
1481 cur_logical < min(stripe->logical + BTRFS_STRIPE_LEN,
1482 extent_start + extent_len);
1483 cur_logical += fs_info->sectorsize) {
1484 const int nr_sector = (cur_logical - stripe->logical) >>
1485 fs_info->sectorsize_bits;
1486 struct scrub_sector_verification *sector =
1487 &stripe->sectors[nr_sector];
1488
1489 set_bit(nr_sector, &stripe->extent_sector_bitmap);
1490 if (extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1491 sector->is_metadata = true;
1492 sector->generation = extent_gen;
1493 }
1494 }
1495}
1496
1497static void scrub_stripe_reset_bitmaps(struct scrub_stripe *stripe)
1498{
1499 stripe->extent_sector_bitmap = 0;
1500 stripe->init_error_bitmap = 0;
1501 stripe->init_nr_io_errors = 0;
1502 stripe->init_nr_csum_errors = 0;
1503 stripe->init_nr_meta_errors = 0;
1504 stripe->error_bitmap = 0;
1505 stripe->io_error_bitmap = 0;
1506 stripe->csum_error_bitmap = 0;
1507 stripe->meta_error_bitmap = 0;
1508}
1509
1510/*
1511 * Locate one stripe which has at least one extent in its range.
1512 *
1513 * Return 0 if found such stripe, and store its info into @stripe.
1514 * Return >0 if there is no such stripe in the specified range.
1515 * Return <0 for error.
1516 */
1517static int scrub_find_fill_first_stripe(struct btrfs_block_group *bg,
1518 struct btrfs_path *extent_path,
1519 struct btrfs_path *csum_path,
1520 struct btrfs_device *dev, u64 physical,
1521 int mirror_num, u64 logical_start,
1522 u32 logical_len,
1523 struct scrub_stripe *stripe)
1524{
1525 struct btrfs_fs_info *fs_info = bg->fs_info;
1526 struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bg->start);
1527 struct btrfs_root *csum_root = btrfs_csum_root(fs_info, bg->start);
1528 const u64 logical_end = logical_start + logical_len;
1529 u64 cur_logical = logical_start;
1530 u64 stripe_end;
1531 u64 extent_start;
1532 u64 extent_len;
1533 u64 extent_flags;
1534 u64 extent_gen;
1535 int ret;
1536
1537 memset(stripe->sectors, 0, sizeof(struct scrub_sector_verification) *
1538 stripe->nr_sectors);
1539 scrub_stripe_reset_bitmaps(stripe);
1540
1541 /* The range must be inside the bg. */
1542 ASSERT(logical_start >= bg->start && logical_end <= bg->start + bg->length);
1543
1544 ret = find_first_extent_item(extent_root, extent_path, logical_start,
1545 logical_len);
1546 /* Either error or not found. */
1547 if (ret)
1548 goto out;
1549 get_extent_info(extent_path, &extent_start, &extent_len, &extent_flags,
1550 &extent_gen);
1551 if (extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1552 stripe->nr_meta_extents++;
1553 if (extent_flags & BTRFS_EXTENT_FLAG_DATA)
1554 stripe->nr_data_extents++;
1555 cur_logical = max(extent_start, cur_logical);
1556
1557 /*
1558 * Round down to stripe boundary.
1559 *
1560 * The extra calculation against bg->start is to handle block groups
1561 * whose logical bytenr is not BTRFS_STRIPE_LEN aligned.
1562 */
1563 stripe->logical = round_down(cur_logical - bg->start, BTRFS_STRIPE_LEN) +
1564 bg->start;
1565 stripe->physical = physical + stripe->logical - logical_start;
1566 stripe->dev = dev;
1567 stripe->bg = bg;
1568 stripe->mirror_num = mirror_num;
1569 stripe_end = stripe->logical + BTRFS_STRIPE_LEN - 1;
1570
1571 /* Fill the first extent info into stripe->sectors[] array. */
1572 fill_one_extent_info(fs_info, stripe, extent_start, extent_len,
1573 extent_flags, extent_gen);
1574 cur_logical = extent_start + extent_len;
1575
1576 /* Fill the extent info for the remaining sectors. */
1577 while (cur_logical <= stripe_end) {
1578 ret = find_first_extent_item(extent_root, extent_path, cur_logical,
1579 stripe_end - cur_logical + 1);
1580 if (ret < 0)
1581 goto out;
1582 if (ret > 0) {
1583 ret = 0;
1584 break;
1585 }
1586 get_extent_info(extent_path, &extent_start, &extent_len,
1587 &extent_flags, &extent_gen);
1588 if (extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1589 stripe->nr_meta_extents++;
1590 if (extent_flags & BTRFS_EXTENT_FLAG_DATA)
1591 stripe->nr_data_extents++;
1592 fill_one_extent_info(fs_info, stripe, extent_start, extent_len,
1593 extent_flags, extent_gen);
1594 cur_logical = extent_start + extent_len;
1595 }
1596
1597 /* Now fill the data csum. */
1598 if (bg->flags & BTRFS_BLOCK_GROUP_DATA) {
1599 int sector_nr;
1600 unsigned long csum_bitmap = 0;
1601
1602 /* Csum space should have already been allocated. */
1603 ASSERT(stripe->csums);
1604
1605 /*
1606 * Our csum bitmap should be large enough, as BTRFS_STRIPE_LEN
1607 * should contain at most 16 sectors.
1608 */
1609 ASSERT(BITS_PER_LONG >= BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits);
1610
1611 ret = btrfs_lookup_csums_bitmap(csum_root, csum_path,
1612 stripe->logical, stripe_end,
1613 stripe->csums, &csum_bitmap);
1614 if (ret < 0)
1615 goto out;
1616 if (ret > 0)
1617 ret = 0;
1618
1619 for_each_set_bit(sector_nr, &csum_bitmap, stripe->nr_sectors) {
1620 stripe->sectors[sector_nr].csum = stripe->csums +
1621 sector_nr * fs_info->csum_size;
1622 }
1623 }
1624 set_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &stripe->state);
1625out:
1626 return ret;
1627}
1628
1629static void scrub_reset_stripe(struct scrub_stripe *stripe)
1630{
1631 scrub_stripe_reset_bitmaps(stripe);
1632
1633 stripe->nr_meta_extents = 0;
1634 stripe->nr_data_extents = 0;
1635 stripe->state = 0;
1636
1637 for (int i = 0; i < stripe->nr_sectors; i++) {
1638 stripe->sectors[i].is_metadata = false;
1639 stripe->sectors[i].csum = NULL;
1640 stripe->sectors[i].generation = 0;
1641 }
1642}
1643
1644static void scrub_submit_extent_sector_read(struct scrub_ctx *sctx,
1645 struct scrub_stripe *stripe)
1646{
1647 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
1648 struct btrfs_bio *bbio = NULL;
1649 unsigned int nr_sectors = min(BTRFS_STRIPE_LEN, stripe->bg->start +
1650 stripe->bg->length - stripe->logical) >>
1651 fs_info->sectorsize_bits;
1652 u64 stripe_len = BTRFS_STRIPE_LEN;
1653 int mirror = stripe->mirror_num;
1654 int i;
1655
1656 atomic_inc(&stripe->pending_io);
1657
1658 for_each_set_bit(i, &stripe->extent_sector_bitmap, stripe->nr_sectors) {
1659 struct page *page = scrub_stripe_get_page(stripe, i);
1660 unsigned int pgoff = scrub_stripe_get_page_offset(stripe, i);
1661
1662 /* We're beyond the chunk boundary, no need to read anymore. */
1663 if (i >= nr_sectors)
1664 break;
1665
1666 /* The current sector cannot be merged, submit the bio. */
1667 if (bbio &&
1668 ((i > 0 &&
1669 !test_bit(i - 1, &stripe->extent_sector_bitmap)) ||
1670 bbio->bio.bi_iter.bi_size >= stripe_len)) {
1671 ASSERT(bbio->bio.bi_iter.bi_size);
1672 atomic_inc(&stripe->pending_io);
1673 btrfs_submit_bio(bbio, mirror);
1674 bbio = NULL;
1675 }
1676
1677 if (!bbio) {
1678 struct btrfs_io_stripe io_stripe = {};
1679 struct btrfs_io_context *bioc = NULL;
1680 const u64 logical = stripe->logical +
1681 (i << fs_info->sectorsize_bits);
1682 int err;
1683
1684 bbio = btrfs_bio_alloc(stripe->nr_sectors, REQ_OP_READ,
1685 fs_info, scrub_read_endio, stripe);
1686 bbio->bio.bi_iter.bi_sector = logical >> SECTOR_SHIFT;
1687
1688 io_stripe.is_scrub = true;
1689 err = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical,
1690 &stripe_len, &bioc, &io_stripe,
1691 &mirror);
1692 btrfs_put_bioc(bioc);
1693 if (err) {
1694 btrfs_bio_end_io(bbio,
1695 errno_to_blk_status(err));
1696 return;
1697 }
1698 }
1699
1700 __bio_add_page(&bbio->bio, page, fs_info->sectorsize, pgoff);
1701 }
1702
1703 if (bbio) {
1704 ASSERT(bbio->bio.bi_iter.bi_size);
1705 atomic_inc(&stripe->pending_io);
1706 btrfs_submit_bio(bbio, mirror);
1707 }
1708
1709 if (atomic_dec_and_test(&stripe->pending_io)) {
1710 wake_up(&stripe->io_wait);
1711 INIT_WORK(&stripe->work, scrub_stripe_read_repair_worker);
1712 queue_work(stripe->bg->fs_info->scrub_workers, &stripe->work);
1713 }
1714}
1715
1716static void scrub_submit_initial_read(struct scrub_ctx *sctx,
1717 struct scrub_stripe *stripe)
1718{
1719 struct btrfs_fs_info *fs_info = sctx->fs_info;
1720 struct btrfs_bio *bbio;
1721 unsigned int nr_sectors = min(BTRFS_STRIPE_LEN, stripe->bg->start +
1722 stripe->bg->length - stripe->logical) >>
1723 fs_info->sectorsize_bits;
1724 int mirror = stripe->mirror_num;
1725
1726 ASSERT(stripe->bg);
1727 ASSERT(stripe->mirror_num > 0);
1728 ASSERT(test_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &stripe->state));
1729
1730 if (btrfs_need_stripe_tree_update(fs_info, stripe->bg->flags)) {
1731 scrub_submit_extent_sector_read(sctx, stripe);
1732 return;
1733 }
1734
1735 bbio = btrfs_bio_alloc(SCRUB_STRIPE_PAGES, REQ_OP_READ, fs_info,
1736 scrub_read_endio, stripe);
1737
1738 bbio->bio.bi_iter.bi_sector = stripe->logical >> SECTOR_SHIFT;
1739 /* Read the whole range inside the chunk boundary. */
1740 for (unsigned int cur = 0; cur < nr_sectors; cur++) {
1741 struct page *page = scrub_stripe_get_page(stripe, cur);
1742 unsigned int pgoff = scrub_stripe_get_page_offset(stripe, cur);
1743 int ret;
1744
1745 ret = bio_add_page(&bbio->bio, page, fs_info->sectorsize, pgoff);
1746 /* We should have allocated enough bio vectors. */
1747 ASSERT(ret == fs_info->sectorsize);
1748 }
1749 atomic_inc(&stripe->pending_io);
1750
1751 /*
1752 * For dev-replace, either user asks to avoid the source dev, or
1753 * the device is missing, we try the next mirror instead.
1754 */
1755 if (sctx->is_dev_replace &&
1756 (fs_info->dev_replace.cont_reading_from_srcdev_mode ==
1757 BTRFS_DEV_REPLACE_ITEM_CONT_READING_FROM_SRCDEV_MODE_AVOID ||
1758 !stripe->dev->bdev)) {
1759 int num_copies = btrfs_num_copies(fs_info, stripe->bg->start,
1760 stripe->bg->length);
1761
1762 mirror = calc_next_mirror(mirror, num_copies);
1763 }
1764 btrfs_submit_bio(bbio, mirror);
1765}
1766
1767static bool stripe_has_metadata_error(struct scrub_stripe *stripe)
1768{
1769 int i;
1770
1771 for_each_set_bit(i, &stripe->error_bitmap, stripe->nr_sectors) {
1772 if (stripe->sectors[i].is_metadata) {
1773 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
1774
1775 btrfs_err(fs_info,
1776 "stripe %llu has unrepaired metadata sector at %llu",
1777 stripe->logical,
1778 stripe->logical + (i << fs_info->sectorsize_bits));
1779 return true;
1780 }
1781 }
1782 return false;
1783}
1784
1785static void submit_initial_group_read(struct scrub_ctx *sctx,
1786 unsigned int first_slot,
1787 unsigned int nr_stripes)
1788{
1789 struct blk_plug plug;
1790
1791 ASSERT(first_slot < SCRUB_TOTAL_STRIPES);
1792 ASSERT(first_slot + nr_stripes <= SCRUB_TOTAL_STRIPES);
1793
1794 scrub_throttle_dev_io(sctx, sctx->stripes[0].dev,
1795 btrfs_stripe_nr_to_offset(nr_stripes));
1796 blk_start_plug(&plug);
1797 for (int i = 0; i < nr_stripes; i++) {
1798 struct scrub_stripe *stripe = &sctx->stripes[first_slot + i];
1799
1800 /* Those stripes should be initialized. */
1801 ASSERT(test_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &stripe->state));
1802 scrub_submit_initial_read(sctx, stripe);
1803 }
1804 blk_finish_plug(&plug);
1805}
1806
1807static int flush_scrub_stripes(struct scrub_ctx *sctx)
1808{
1809 struct btrfs_fs_info *fs_info = sctx->fs_info;
1810 struct scrub_stripe *stripe;
1811 const int nr_stripes = sctx->cur_stripe;
1812 int ret = 0;
1813
1814 if (!nr_stripes)
1815 return 0;
1816
1817 ASSERT(test_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &sctx->stripes[0].state));
1818
1819 /* Submit the stripes which are populated but not submitted. */
1820 if (nr_stripes % SCRUB_STRIPES_PER_GROUP) {
1821 const int first_slot = round_down(nr_stripes, SCRUB_STRIPES_PER_GROUP);
1822
1823 submit_initial_group_read(sctx, first_slot, nr_stripes - first_slot);
1824 }
1825
1826 for (int i = 0; i < nr_stripes; i++) {
1827 stripe = &sctx->stripes[i];
1828
1829 wait_event(stripe->repair_wait,
1830 test_bit(SCRUB_STRIPE_FLAG_REPAIR_DONE, &stripe->state));
1831 }
1832
1833 /* Submit for dev-replace. */
1834 if (sctx->is_dev_replace) {
1835 /*
1836 * For dev-replace, if we know there is something wrong with
1837 * metadata, we should immediately abort.
1838 */
1839 for (int i = 0; i < nr_stripes; i++) {
1840 if (stripe_has_metadata_error(&sctx->stripes[i])) {
1841 ret = -EIO;
1842 goto out;
1843 }
1844 }
1845 for (int i = 0; i < nr_stripes; i++) {
1846 unsigned long good;
1847
1848 stripe = &sctx->stripes[i];
1849
1850 ASSERT(stripe->dev == fs_info->dev_replace.srcdev);
1851
1852 bitmap_andnot(&good, &stripe->extent_sector_bitmap,
1853 &stripe->error_bitmap, stripe->nr_sectors);
1854 scrub_write_sectors(sctx, stripe, good, true);
1855 }
1856 }
1857
1858 /* Wait for the above writebacks to finish. */
1859 for (int i = 0; i < nr_stripes; i++) {
1860 stripe = &sctx->stripes[i];
1861
1862 wait_scrub_stripe_io(stripe);
1863 scrub_reset_stripe(stripe);
1864 }
1865out:
1866 sctx->cur_stripe = 0;
1867 return ret;
1868}
1869
1870static void raid56_scrub_wait_endio(struct bio *bio)
1871{
1872 complete(bio->bi_private);
1873}
1874
1875static int queue_scrub_stripe(struct scrub_ctx *sctx, struct btrfs_block_group *bg,
1876 struct btrfs_device *dev, int mirror_num,
1877 u64 logical, u32 length, u64 physical,
1878 u64 *found_logical_ret)
1879{
1880 struct scrub_stripe *stripe;
1881 int ret;
1882
1883 /*
1884 * There should always be one slot left, as caller filling the last
1885 * slot should flush them all.
1886 */
1887 ASSERT(sctx->cur_stripe < SCRUB_TOTAL_STRIPES);
1888
1889 /* @found_logical_ret must be specified. */
1890 ASSERT(found_logical_ret);
1891
1892 stripe = &sctx->stripes[sctx->cur_stripe];
1893 scrub_reset_stripe(stripe);
1894 ret = scrub_find_fill_first_stripe(bg, &sctx->extent_path,
1895 &sctx->csum_path, dev, physical,
1896 mirror_num, logical, length, stripe);
1897 /* Either >0 as no more extents or <0 for error. */
1898 if (ret)
1899 return ret;
1900 *found_logical_ret = stripe->logical;
1901 sctx->cur_stripe++;
1902
1903 /* We filled one group, submit it. */
1904 if (sctx->cur_stripe % SCRUB_STRIPES_PER_GROUP == 0) {
1905 const int first_slot = sctx->cur_stripe - SCRUB_STRIPES_PER_GROUP;
1906
1907 submit_initial_group_read(sctx, first_slot, SCRUB_STRIPES_PER_GROUP);
1908 }
1909
1910 /* Last slot used, flush them all. */
1911 if (sctx->cur_stripe == SCRUB_TOTAL_STRIPES)
1912 return flush_scrub_stripes(sctx);
1913 return 0;
1914}
1915
1916static int scrub_raid56_parity_stripe(struct scrub_ctx *sctx,
1917 struct btrfs_device *scrub_dev,
1918 struct btrfs_block_group *bg,
1919 struct btrfs_chunk_map *map,
1920 u64 full_stripe_start)
1921{
1922 DECLARE_COMPLETION_ONSTACK(io_done);
1923 struct btrfs_fs_info *fs_info = sctx->fs_info;
1924 struct btrfs_raid_bio *rbio;
1925 struct btrfs_io_context *bioc = NULL;
1926 struct btrfs_path extent_path = { 0 };
1927 struct btrfs_path csum_path = { 0 };
1928 struct bio *bio;
1929 struct scrub_stripe *stripe;
1930 bool all_empty = true;
1931 const int data_stripes = nr_data_stripes(map);
1932 unsigned long extent_bitmap = 0;
1933 u64 length = btrfs_stripe_nr_to_offset(data_stripes);
1934 int ret;
1935
1936 ASSERT(sctx->raid56_data_stripes);
1937
1938 /*
1939 * For data stripe search, we cannot re-use the same extent/csum paths,
1940 * as the data stripe bytenr may be smaller than previous extent. Thus
1941 * we have to use our own extent/csum paths.
1942 */
1943 extent_path.search_commit_root = 1;
1944 extent_path.skip_locking = 1;
1945 csum_path.search_commit_root = 1;
1946 csum_path.skip_locking = 1;
1947
1948 for (int i = 0; i < data_stripes; i++) {
1949 int stripe_index;
1950 int rot;
1951 u64 physical;
1952
1953 stripe = &sctx->raid56_data_stripes[i];
1954 rot = div_u64(full_stripe_start - bg->start,
1955 data_stripes) >> BTRFS_STRIPE_LEN_SHIFT;
1956 stripe_index = (i + rot) % map->num_stripes;
1957 physical = map->stripes[stripe_index].physical +
1958 btrfs_stripe_nr_to_offset(rot);
1959
1960 scrub_reset_stripe(stripe);
1961 set_bit(SCRUB_STRIPE_FLAG_NO_REPORT, &stripe->state);
1962 ret = scrub_find_fill_first_stripe(bg, &extent_path, &csum_path,
1963 map->stripes[stripe_index].dev, physical, 1,
1964 full_stripe_start + btrfs_stripe_nr_to_offset(i),
1965 BTRFS_STRIPE_LEN, stripe);
1966 if (ret < 0)
1967 goto out;
1968 /*
1969 * No extent in this data stripe, need to manually mark them
1970 * initialized to make later read submission happy.
1971 */
1972 if (ret > 0) {
1973 stripe->logical = full_stripe_start +
1974 btrfs_stripe_nr_to_offset(i);
1975 stripe->dev = map->stripes[stripe_index].dev;
1976 stripe->mirror_num = 1;
1977 set_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &stripe->state);
1978 }
1979 }
1980
1981 /* Check if all data stripes are empty. */
1982 for (int i = 0; i < data_stripes; i++) {
1983 stripe = &sctx->raid56_data_stripes[i];
1984 if (!bitmap_empty(&stripe->extent_sector_bitmap, stripe->nr_sectors)) {
1985 all_empty = false;
1986 break;
1987 }
1988 }
1989 if (all_empty) {
1990 ret = 0;
1991 goto out;
1992 }
1993
1994 for (int i = 0; i < data_stripes; i++) {
1995 stripe = &sctx->raid56_data_stripes[i];
1996 scrub_submit_initial_read(sctx, stripe);
1997 }
1998 for (int i = 0; i < data_stripes; i++) {
1999 stripe = &sctx->raid56_data_stripes[i];
2000
2001 wait_event(stripe->repair_wait,
2002 test_bit(SCRUB_STRIPE_FLAG_REPAIR_DONE, &stripe->state));
2003 }
2004 /* For now, no zoned support for RAID56. */
2005 ASSERT(!btrfs_is_zoned(sctx->fs_info));
2006
2007 /*
2008 * Now all data stripes are properly verified. Check if we have any
2009 * unrepaired, if so abort immediately or we could further corrupt the
2010 * P/Q stripes.
2011 *
2012 * During the loop, also populate extent_bitmap.
2013 */
2014 for (int i = 0; i < data_stripes; i++) {
2015 unsigned long error;
2016
2017 stripe = &sctx->raid56_data_stripes[i];
2018
2019 /*
2020 * We should only check the errors where there is an extent.
2021 * As we may hit an empty data stripe while it's missing.
2022 */
2023 bitmap_and(&error, &stripe->error_bitmap,
2024 &stripe->extent_sector_bitmap, stripe->nr_sectors);
2025 if (!bitmap_empty(&error, stripe->nr_sectors)) {
2026 btrfs_err(fs_info,
2027"unrepaired sectors detected, full stripe %llu data stripe %u errors %*pbl",
2028 full_stripe_start, i, stripe->nr_sectors,
2029 &error);
2030 ret = -EIO;
2031 goto out;
2032 }
2033 bitmap_or(&extent_bitmap, &extent_bitmap,
2034 &stripe->extent_sector_bitmap, stripe->nr_sectors);
2035 }
2036
2037 /* Now we can check and regenerate the P/Q stripe. */
2038 bio = bio_alloc(NULL, 1, REQ_OP_READ, GFP_NOFS);
2039 bio->bi_iter.bi_sector = full_stripe_start >> SECTOR_SHIFT;
2040 bio->bi_private = &io_done;
2041 bio->bi_end_io = raid56_scrub_wait_endio;
2042
2043 btrfs_bio_counter_inc_blocked(fs_info);
2044 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, full_stripe_start,
2045 &length, &bioc, NULL, NULL);
2046 if (ret < 0) {
2047 btrfs_put_bioc(bioc);
2048 btrfs_bio_counter_dec(fs_info);
2049 goto out;
2050 }
2051 rbio = raid56_parity_alloc_scrub_rbio(bio, bioc, scrub_dev, &extent_bitmap,
2052 BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits);
2053 btrfs_put_bioc(bioc);
2054 if (!rbio) {
2055 ret = -ENOMEM;
2056 btrfs_bio_counter_dec(fs_info);
2057 goto out;
2058 }
2059 /* Use the recovered stripes as cache to avoid read them from disk again. */
2060 for (int i = 0; i < data_stripes; i++) {
2061 stripe = &sctx->raid56_data_stripes[i];
2062
2063 raid56_parity_cache_data_pages(rbio, stripe->pages,
2064 full_stripe_start + (i << BTRFS_STRIPE_LEN_SHIFT));
2065 }
2066 raid56_parity_submit_scrub_rbio(rbio);
2067 wait_for_completion_io(&io_done);
2068 ret = blk_status_to_errno(bio->bi_status);
2069 bio_put(bio);
2070 btrfs_bio_counter_dec(fs_info);
2071
2072 btrfs_release_path(&extent_path);
2073 btrfs_release_path(&csum_path);
2074out:
2075 return ret;
2076}
2077
2078/*
2079 * Scrub one range which can only has simple mirror based profile.
2080 * (Including all range in SINGLE/DUP/RAID1/RAID1C*, and each stripe in
2081 * RAID0/RAID10).
2082 *
2083 * Since we may need to handle a subset of block group, we need @logical_start
2084 * and @logical_length parameter.
2085 */
2086static int scrub_simple_mirror(struct scrub_ctx *sctx,
2087 struct btrfs_block_group *bg,
2088 struct btrfs_chunk_map *map,
2089 u64 logical_start, u64 logical_length,
2090 struct btrfs_device *device,
2091 u64 physical, int mirror_num)
2092{
2093 struct btrfs_fs_info *fs_info = sctx->fs_info;
2094 const u64 logical_end = logical_start + logical_length;
2095 u64 cur_logical = logical_start;
2096 int ret;
2097
2098 /* The range must be inside the bg */
2099 ASSERT(logical_start >= bg->start && logical_end <= bg->start + bg->length);
2100
2101 /* Go through each extent items inside the logical range */
2102 while (cur_logical < logical_end) {
2103 u64 found_logical = U64_MAX;
2104 u64 cur_physical = physical + cur_logical - logical_start;
2105
2106 /* Canceled? */
2107 if (atomic_read(&fs_info->scrub_cancel_req) ||
2108 atomic_read(&sctx->cancel_req)) {
2109 ret = -ECANCELED;
2110 break;
2111 }
2112 /* Paused? */
2113 if (atomic_read(&fs_info->scrub_pause_req)) {
2114 /* Push queued extents */
2115 scrub_blocked_if_needed(fs_info);
2116 }
2117 /* Block group removed? */
2118 spin_lock(&bg->lock);
2119 if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &bg->runtime_flags)) {
2120 spin_unlock(&bg->lock);
2121 ret = 0;
2122 break;
2123 }
2124 spin_unlock(&bg->lock);
2125
2126 ret = queue_scrub_stripe(sctx, bg, device, mirror_num,
2127 cur_logical, logical_end - cur_logical,
2128 cur_physical, &found_logical);
2129 if (ret > 0) {
2130 /* No more extent, just update the accounting */
2131 sctx->stat.last_physical = physical + logical_length;
2132 ret = 0;
2133 break;
2134 }
2135 if (ret < 0)
2136 break;
2137
2138 /* queue_scrub_stripe() returned 0, @found_logical must be updated. */
2139 ASSERT(found_logical != U64_MAX);
2140 cur_logical = found_logical + BTRFS_STRIPE_LEN;
2141
2142 /* Don't hold CPU for too long time */
2143 cond_resched();
2144 }
2145 return ret;
2146}
2147
2148/* Calculate the full stripe length for simple stripe based profiles */
2149static u64 simple_stripe_full_stripe_len(const struct btrfs_chunk_map *map)
2150{
2151 ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2152 BTRFS_BLOCK_GROUP_RAID10));
2153
2154 return btrfs_stripe_nr_to_offset(map->num_stripes / map->sub_stripes);
2155}
2156
2157/* Get the logical bytenr for the stripe */
2158static u64 simple_stripe_get_logical(struct btrfs_chunk_map *map,
2159 struct btrfs_block_group *bg,
2160 int stripe_index)
2161{
2162 ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2163 BTRFS_BLOCK_GROUP_RAID10));
2164 ASSERT(stripe_index < map->num_stripes);
2165
2166 /*
2167 * (stripe_index / sub_stripes) gives how many data stripes we need to
2168 * skip.
2169 */
2170 return btrfs_stripe_nr_to_offset(stripe_index / map->sub_stripes) +
2171 bg->start;
2172}
2173
2174/* Get the mirror number for the stripe */
2175static int simple_stripe_mirror_num(struct btrfs_chunk_map *map, int stripe_index)
2176{
2177 ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2178 BTRFS_BLOCK_GROUP_RAID10));
2179 ASSERT(stripe_index < map->num_stripes);
2180
2181 /* For RAID0, it's fixed to 1, for RAID10 it's 0,1,0,1... */
2182 return stripe_index % map->sub_stripes + 1;
2183}
2184
2185static int scrub_simple_stripe(struct scrub_ctx *sctx,
2186 struct btrfs_block_group *bg,
2187 struct btrfs_chunk_map *map,
2188 struct btrfs_device *device,
2189 int stripe_index)
2190{
2191 const u64 logical_increment = simple_stripe_full_stripe_len(map);
2192 const u64 orig_logical = simple_stripe_get_logical(map, bg, stripe_index);
2193 const u64 orig_physical = map->stripes[stripe_index].physical;
2194 const int mirror_num = simple_stripe_mirror_num(map, stripe_index);
2195 u64 cur_logical = orig_logical;
2196 u64 cur_physical = orig_physical;
2197 int ret = 0;
2198
2199 while (cur_logical < bg->start + bg->length) {
2200 /*
2201 * Inside each stripe, RAID0 is just SINGLE, and RAID10 is
2202 * just RAID1, so we can reuse scrub_simple_mirror() to scrub
2203 * this stripe.
2204 */
2205 ret = scrub_simple_mirror(sctx, bg, map, cur_logical,
2206 BTRFS_STRIPE_LEN, device, cur_physical,
2207 mirror_num);
2208 if (ret)
2209 return ret;
2210 /* Skip to next stripe which belongs to the target device */
2211 cur_logical += logical_increment;
2212 /* For physical offset, we just go to next stripe */
2213 cur_physical += BTRFS_STRIPE_LEN;
2214 }
2215 return ret;
2216}
2217
2218static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2219 struct btrfs_block_group *bg,
2220 struct btrfs_chunk_map *map,
2221 struct btrfs_device *scrub_dev,
2222 int stripe_index)
2223{
2224 struct btrfs_fs_info *fs_info = sctx->fs_info;
2225 const u64 profile = map->type & BTRFS_BLOCK_GROUP_PROFILE_MASK;
2226 const u64 chunk_logical = bg->start;
2227 int ret;
2228 int ret2;
2229 u64 physical = map->stripes[stripe_index].physical;
2230 const u64 dev_stripe_len = btrfs_calc_stripe_length(map);
2231 const u64 physical_end = physical + dev_stripe_len;
2232 u64 logical;
2233 u64 logic_end;
2234 /* The logical increment after finishing one stripe */
2235 u64 increment;
2236 /* Offset inside the chunk */
2237 u64 offset;
2238 u64 stripe_logical;
2239 int stop_loop = 0;
2240
2241 /* Extent_path should be released by now. */
2242 ASSERT(sctx->extent_path.nodes[0] == NULL);
2243
2244 scrub_blocked_if_needed(fs_info);
2245
2246 if (sctx->is_dev_replace &&
2247 btrfs_dev_is_sequential(sctx->wr_tgtdev, physical)) {
2248 mutex_lock(&sctx->wr_lock);
2249 sctx->write_pointer = physical;
2250 mutex_unlock(&sctx->wr_lock);
2251 }
2252
2253 /* Prepare the extra data stripes used by RAID56. */
2254 if (profile & BTRFS_BLOCK_GROUP_RAID56_MASK) {
2255 ASSERT(sctx->raid56_data_stripes == NULL);
2256
2257 sctx->raid56_data_stripes = kcalloc(nr_data_stripes(map),
2258 sizeof(struct scrub_stripe),
2259 GFP_KERNEL);
2260 if (!sctx->raid56_data_stripes) {
2261 ret = -ENOMEM;
2262 goto out;
2263 }
2264 for (int i = 0; i < nr_data_stripes(map); i++) {
2265 ret = init_scrub_stripe(fs_info,
2266 &sctx->raid56_data_stripes[i]);
2267 if (ret < 0)
2268 goto out;
2269 sctx->raid56_data_stripes[i].bg = bg;
2270 sctx->raid56_data_stripes[i].sctx = sctx;
2271 }
2272 }
2273 /*
2274 * There used to be a big double loop to handle all profiles using the
2275 * same routine, which grows larger and more gross over time.
2276 *
2277 * So here we handle each profile differently, so simpler profiles
2278 * have simpler scrubbing function.
2279 */
2280 if (!(profile & (BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID10 |
2281 BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2282 /*
2283 * Above check rules out all complex profile, the remaining
2284 * profiles are SINGLE|DUP|RAID1|RAID1C*, which is simple
2285 * mirrored duplication without stripe.
2286 *
2287 * Only @physical and @mirror_num needs to calculated using
2288 * @stripe_index.
2289 */
2290 ret = scrub_simple_mirror(sctx, bg, map, bg->start, bg->length,
2291 scrub_dev, map->stripes[stripe_index].physical,
2292 stripe_index + 1);
2293 offset = 0;
2294 goto out;
2295 }
2296 if (profile & (BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID10)) {
2297 ret = scrub_simple_stripe(sctx, bg, map, scrub_dev, stripe_index);
2298 offset = btrfs_stripe_nr_to_offset(stripe_index / map->sub_stripes);
2299 goto out;
2300 }
2301
2302 /* Only RAID56 goes through the old code */
2303 ASSERT(map->type & BTRFS_BLOCK_GROUP_RAID56_MASK);
2304 ret = 0;
2305
2306 /* Calculate the logical end of the stripe */
2307 get_raid56_logic_offset(physical_end, stripe_index,
2308 map, &logic_end, NULL);
2309 logic_end += chunk_logical;
2310
2311 /* Initialize @offset in case we need to go to out: label */
2312 get_raid56_logic_offset(physical, stripe_index, map, &offset, NULL);
2313 increment = btrfs_stripe_nr_to_offset(nr_data_stripes(map));
2314
2315 /*
2316 * Due to the rotation, for RAID56 it's better to iterate each stripe
2317 * using their physical offset.
2318 */
2319 while (physical < physical_end) {
2320 ret = get_raid56_logic_offset(physical, stripe_index, map,
2321 &logical, &stripe_logical);
2322 logical += chunk_logical;
2323 if (ret) {
2324 /* it is parity strip */
2325 stripe_logical += chunk_logical;
2326 ret = scrub_raid56_parity_stripe(sctx, scrub_dev, bg,
2327 map, stripe_logical);
2328 if (ret)
2329 goto out;
2330 goto next;
2331 }
2332
2333 /*
2334 * Now we're at a data stripe, scrub each extents in the range.
2335 *
2336 * At this stage, if we ignore the repair part, inside each data
2337 * stripe it is no different than SINGLE profile.
2338 * We can reuse scrub_simple_mirror() here, as the repair part
2339 * is still based on @mirror_num.
2340 */
2341 ret = scrub_simple_mirror(sctx, bg, map, logical, BTRFS_STRIPE_LEN,
2342 scrub_dev, physical, 1);
2343 if (ret < 0)
2344 goto out;
2345next:
2346 logical += increment;
2347 physical += BTRFS_STRIPE_LEN;
2348 spin_lock(&sctx->stat_lock);
2349 if (stop_loop)
2350 sctx->stat.last_physical =
2351 map->stripes[stripe_index].physical + dev_stripe_len;
2352 else
2353 sctx->stat.last_physical = physical;
2354 spin_unlock(&sctx->stat_lock);
2355 if (stop_loop)
2356 break;
2357 }
2358out:
2359 ret2 = flush_scrub_stripes(sctx);
2360 if (!ret)
2361 ret = ret2;
2362 btrfs_release_path(&sctx->extent_path);
2363 btrfs_release_path(&sctx->csum_path);
2364
2365 if (sctx->raid56_data_stripes) {
2366 for (int i = 0; i < nr_data_stripes(map); i++)
2367 release_scrub_stripe(&sctx->raid56_data_stripes[i]);
2368 kfree(sctx->raid56_data_stripes);
2369 sctx->raid56_data_stripes = NULL;
2370 }
2371
2372 if (sctx->is_dev_replace && ret >= 0) {
2373 int ret2;
2374
2375 ret2 = sync_write_pointer_for_zoned(sctx,
2376 chunk_logical + offset,
2377 map->stripes[stripe_index].physical,
2378 physical_end);
2379 if (ret2)
2380 ret = ret2;
2381 }
2382
2383 return ret < 0 ? ret : 0;
2384}
2385
2386static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
2387 struct btrfs_block_group *bg,
2388 struct btrfs_device *scrub_dev,
2389 u64 dev_offset,
2390 u64 dev_extent_len)
2391{
2392 struct btrfs_fs_info *fs_info = sctx->fs_info;
2393 struct btrfs_chunk_map *map;
2394 int i;
2395 int ret = 0;
2396
2397 map = btrfs_find_chunk_map(fs_info, bg->start, bg->length);
2398 if (!map) {
2399 /*
2400 * Might have been an unused block group deleted by the cleaner
2401 * kthread or relocation.
2402 */
2403 spin_lock(&bg->lock);
2404 if (!test_bit(BLOCK_GROUP_FLAG_REMOVED, &bg->runtime_flags))
2405 ret = -EINVAL;
2406 spin_unlock(&bg->lock);
2407
2408 return ret;
2409 }
2410 if (map->start != bg->start)
2411 goto out;
2412 if (map->chunk_len < dev_extent_len)
2413 goto out;
2414
2415 for (i = 0; i < map->num_stripes; ++i) {
2416 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
2417 map->stripes[i].physical == dev_offset) {
2418 ret = scrub_stripe(sctx, bg, map, scrub_dev, i);
2419 if (ret)
2420 goto out;
2421 }
2422 }
2423out:
2424 btrfs_free_chunk_map(map);
2425
2426 return ret;
2427}
2428
2429static int finish_extent_writes_for_zoned(struct btrfs_root *root,
2430 struct btrfs_block_group *cache)
2431{
2432 struct btrfs_fs_info *fs_info = cache->fs_info;
2433 struct btrfs_trans_handle *trans;
2434
2435 if (!btrfs_is_zoned(fs_info))
2436 return 0;
2437
2438 btrfs_wait_block_group_reservations(cache);
2439 btrfs_wait_nocow_writers(cache);
2440 btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start, cache->length);
2441
2442 trans = btrfs_join_transaction(root);
2443 if (IS_ERR(trans))
2444 return PTR_ERR(trans);
2445 return btrfs_commit_transaction(trans);
2446}
2447
2448static noinline_for_stack
2449int scrub_enumerate_chunks(struct scrub_ctx *sctx,
2450 struct btrfs_device *scrub_dev, u64 start, u64 end)
2451{
2452 struct btrfs_dev_extent *dev_extent = NULL;
2453 struct btrfs_path *path;
2454 struct btrfs_fs_info *fs_info = sctx->fs_info;
2455 struct btrfs_root *root = fs_info->dev_root;
2456 u64 chunk_offset;
2457 int ret = 0;
2458 int ro_set;
2459 int slot;
2460 struct extent_buffer *l;
2461 struct btrfs_key key;
2462 struct btrfs_key found_key;
2463 struct btrfs_block_group *cache;
2464 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2465
2466 path = btrfs_alloc_path();
2467 if (!path)
2468 return -ENOMEM;
2469
2470 path->reada = READA_FORWARD;
2471 path->search_commit_root = 1;
2472 path->skip_locking = 1;
2473
2474 key.objectid = scrub_dev->devid;
2475 key.offset = 0ull;
2476 key.type = BTRFS_DEV_EXTENT_KEY;
2477
2478 while (1) {
2479 u64 dev_extent_len;
2480
2481 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2482 if (ret < 0)
2483 break;
2484 if (ret > 0) {
2485 if (path->slots[0] >=
2486 btrfs_header_nritems(path->nodes[0])) {
2487 ret = btrfs_next_leaf(root, path);
2488 if (ret < 0)
2489 break;
2490 if (ret > 0) {
2491 ret = 0;
2492 break;
2493 }
2494 } else {
2495 ret = 0;
2496 }
2497 }
2498
2499 l = path->nodes[0];
2500 slot = path->slots[0];
2501
2502 btrfs_item_key_to_cpu(l, &found_key, slot);
2503
2504 if (found_key.objectid != scrub_dev->devid)
2505 break;
2506
2507 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
2508 break;
2509
2510 if (found_key.offset >= end)
2511 break;
2512
2513 if (found_key.offset < key.offset)
2514 break;
2515
2516 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2517 dev_extent_len = btrfs_dev_extent_length(l, dev_extent);
2518
2519 if (found_key.offset + dev_extent_len <= start)
2520 goto skip;
2521
2522 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2523
2524 /*
2525 * get a reference on the corresponding block group to prevent
2526 * the chunk from going away while we scrub it
2527 */
2528 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2529
2530 /* some chunks are removed but not committed to disk yet,
2531 * continue scrubbing */
2532 if (!cache)
2533 goto skip;
2534
2535 ASSERT(cache->start <= chunk_offset);
2536 /*
2537 * We are using the commit root to search for device extents, so
2538 * that means we could have found a device extent item from a
2539 * block group that was deleted in the current transaction. The
2540 * logical start offset of the deleted block group, stored at
2541 * @chunk_offset, might be part of the logical address range of
2542 * a new block group (which uses different physical extents).
2543 * In this case btrfs_lookup_block_group() has returned the new
2544 * block group, and its start address is less than @chunk_offset.
2545 *
2546 * We skip such new block groups, because it's pointless to
2547 * process them, as we won't find their extents because we search
2548 * for them using the commit root of the extent tree. For a device
2549 * replace it's also fine to skip it, we won't miss copying them
2550 * to the target device because we have the write duplication
2551 * setup through the regular write path (by btrfs_map_block()),
2552 * and we have committed a transaction when we started the device
2553 * replace, right after setting up the device replace state.
2554 */
2555 if (cache->start < chunk_offset) {
2556 btrfs_put_block_group(cache);
2557 goto skip;
2558 }
2559
2560 if (sctx->is_dev_replace && btrfs_is_zoned(fs_info)) {
2561 if (!test_bit(BLOCK_GROUP_FLAG_TO_COPY, &cache->runtime_flags)) {
2562 btrfs_put_block_group(cache);
2563 goto skip;
2564 }
2565 }
2566
2567 /*
2568 * Make sure that while we are scrubbing the corresponding block
2569 * group doesn't get its logical address and its device extents
2570 * reused for another block group, which can possibly be of a
2571 * different type and different profile. We do this to prevent
2572 * false error detections and crashes due to bogus attempts to
2573 * repair extents.
2574 */
2575 spin_lock(&cache->lock);
2576 if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &cache->runtime_flags)) {
2577 spin_unlock(&cache->lock);
2578 btrfs_put_block_group(cache);
2579 goto skip;
2580 }
2581 btrfs_freeze_block_group(cache);
2582 spin_unlock(&cache->lock);
2583
2584 /*
2585 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
2586 * to avoid deadlock caused by:
2587 * btrfs_inc_block_group_ro()
2588 * -> btrfs_wait_for_commit()
2589 * -> btrfs_commit_transaction()
2590 * -> btrfs_scrub_pause()
2591 */
2592 scrub_pause_on(fs_info);
2593
2594 /*
2595 * Don't do chunk preallocation for scrub.
2596 *
2597 * This is especially important for SYSTEM bgs, or we can hit
2598 * -EFBIG from btrfs_finish_chunk_alloc() like:
2599 * 1. The only SYSTEM bg is marked RO.
2600 * Since SYSTEM bg is small, that's pretty common.
2601 * 2. New SYSTEM bg will be allocated
2602 * Due to regular version will allocate new chunk.
2603 * 3. New SYSTEM bg is empty and will get cleaned up
2604 * Before cleanup really happens, it's marked RO again.
2605 * 4. Empty SYSTEM bg get scrubbed
2606 * We go back to 2.
2607 *
2608 * This can easily boost the amount of SYSTEM chunks if cleaner
2609 * thread can't be triggered fast enough, and use up all space
2610 * of btrfs_super_block::sys_chunk_array
2611 *
2612 * While for dev replace, we need to try our best to mark block
2613 * group RO, to prevent race between:
2614 * - Write duplication
2615 * Contains latest data
2616 * - Scrub copy
2617 * Contains data from commit tree
2618 *
2619 * If target block group is not marked RO, nocow writes can
2620 * be overwritten by scrub copy, causing data corruption.
2621 * So for dev-replace, it's not allowed to continue if a block
2622 * group is not RO.
2623 */
2624 ret = btrfs_inc_block_group_ro(cache, sctx->is_dev_replace);
2625 if (!ret && sctx->is_dev_replace) {
2626 ret = finish_extent_writes_for_zoned(root, cache);
2627 if (ret) {
2628 btrfs_dec_block_group_ro(cache);
2629 scrub_pause_off(fs_info);
2630 btrfs_put_block_group(cache);
2631 break;
2632 }
2633 }
2634
2635 if (ret == 0) {
2636 ro_set = 1;
2637 } else if (ret == -ENOSPC && !sctx->is_dev_replace &&
2638 !(cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK)) {
2639 /*
2640 * btrfs_inc_block_group_ro return -ENOSPC when it
2641 * failed in creating new chunk for metadata.
2642 * It is not a problem for scrub, because
2643 * metadata are always cowed, and our scrub paused
2644 * commit_transactions.
2645 *
2646 * For RAID56 chunks, we have to mark them read-only
2647 * for scrub, as later we would use our own cache
2648 * out of RAID56 realm.
2649 * Thus we want the RAID56 bg to be marked RO to
2650 * prevent RMW from screwing up out cache.
2651 */
2652 ro_set = 0;
2653 } else if (ret == -ETXTBSY) {
2654 btrfs_warn(fs_info,
2655 "skipping scrub of block group %llu due to active swapfile",
2656 cache->start);
2657 scrub_pause_off(fs_info);
2658 ret = 0;
2659 goto skip_unfreeze;
2660 } else {
2661 btrfs_warn(fs_info,
2662 "failed setting block group ro: %d", ret);
2663 btrfs_unfreeze_block_group(cache);
2664 btrfs_put_block_group(cache);
2665 scrub_pause_off(fs_info);
2666 break;
2667 }
2668
2669 /*
2670 * Now the target block is marked RO, wait for nocow writes to
2671 * finish before dev-replace.
2672 * COW is fine, as COW never overwrites extents in commit tree.
2673 */
2674 if (sctx->is_dev_replace) {
2675 btrfs_wait_nocow_writers(cache);
2676 btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start,
2677 cache->length);
2678 }
2679
2680 scrub_pause_off(fs_info);
2681 down_write(&dev_replace->rwsem);
2682 dev_replace->cursor_right = found_key.offset + dev_extent_len;
2683 dev_replace->cursor_left = found_key.offset;
2684 dev_replace->item_needs_writeback = 1;
2685 up_write(&dev_replace->rwsem);
2686
2687 ret = scrub_chunk(sctx, cache, scrub_dev, found_key.offset,
2688 dev_extent_len);
2689 if (sctx->is_dev_replace &&
2690 !btrfs_finish_block_group_to_copy(dev_replace->srcdev,
2691 cache, found_key.offset))
2692 ro_set = 0;
2693
2694 down_write(&dev_replace->rwsem);
2695 dev_replace->cursor_left = dev_replace->cursor_right;
2696 dev_replace->item_needs_writeback = 1;
2697 up_write(&dev_replace->rwsem);
2698
2699 if (ro_set)
2700 btrfs_dec_block_group_ro(cache);
2701
2702 /*
2703 * We might have prevented the cleaner kthread from deleting
2704 * this block group if it was already unused because we raced
2705 * and set it to RO mode first. So add it back to the unused
2706 * list, otherwise it might not ever be deleted unless a manual
2707 * balance is triggered or it becomes used and unused again.
2708 */
2709 spin_lock(&cache->lock);
2710 if (!test_bit(BLOCK_GROUP_FLAG_REMOVED, &cache->runtime_flags) &&
2711 !cache->ro && cache->reserved == 0 && cache->used == 0) {
2712 spin_unlock(&cache->lock);
2713 if (btrfs_test_opt(fs_info, DISCARD_ASYNC))
2714 btrfs_discard_queue_work(&fs_info->discard_ctl,
2715 cache);
2716 else
2717 btrfs_mark_bg_unused(cache);
2718 } else {
2719 spin_unlock(&cache->lock);
2720 }
2721skip_unfreeze:
2722 btrfs_unfreeze_block_group(cache);
2723 btrfs_put_block_group(cache);
2724 if (ret)
2725 break;
2726 if (sctx->is_dev_replace &&
2727 atomic64_read(&dev_replace->num_write_errors) > 0) {
2728 ret = -EIO;
2729 break;
2730 }
2731 if (sctx->stat.malloc_errors > 0) {
2732 ret = -ENOMEM;
2733 break;
2734 }
2735skip:
2736 key.offset = found_key.offset + dev_extent_len;
2737 btrfs_release_path(path);
2738 }
2739
2740 btrfs_free_path(path);
2741
2742 return ret;
2743}
2744
2745static int scrub_one_super(struct scrub_ctx *sctx, struct btrfs_device *dev,
2746 struct page *page, u64 physical, u64 generation)
2747{
2748 struct btrfs_fs_info *fs_info = sctx->fs_info;
2749 struct bio_vec bvec;
2750 struct bio bio;
2751 struct btrfs_super_block *sb = page_address(page);
2752 int ret;
2753
2754 bio_init(&bio, dev->bdev, &bvec, 1, REQ_OP_READ);
2755 bio.bi_iter.bi_sector = physical >> SECTOR_SHIFT;
2756 __bio_add_page(&bio, page, BTRFS_SUPER_INFO_SIZE, 0);
2757 ret = submit_bio_wait(&bio);
2758 bio_uninit(&bio);
2759
2760 if (ret < 0)
2761 return ret;
2762 ret = btrfs_check_super_csum(fs_info, sb);
2763 if (ret != 0) {
2764 btrfs_err_rl(fs_info,
2765 "super block at physical %llu devid %llu has bad csum",
2766 physical, dev->devid);
2767 return -EIO;
2768 }
2769 if (btrfs_super_generation(sb) != generation) {
2770 btrfs_err_rl(fs_info,
2771"super block at physical %llu devid %llu has bad generation %llu expect %llu",
2772 physical, dev->devid,
2773 btrfs_super_generation(sb), generation);
2774 return -EUCLEAN;
2775 }
2776
2777 return btrfs_validate_super(fs_info, sb, -1);
2778}
2779
2780static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
2781 struct btrfs_device *scrub_dev)
2782{
2783 int i;
2784 u64 bytenr;
2785 u64 gen;
2786 int ret = 0;
2787 struct page *page;
2788 struct btrfs_fs_info *fs_info = sctx->fs_info;
2789
2790 if (BTRFS_FS_ERROR(fs_info))
2791 return -EROFS;
2792
2793 page = alloc_page(GFP_KERNEL);
2794 if (!page) {
2795 spin_lock(&sctx->stat_lock);
2796 sctx->stat.malloc_errors++;
2797 spin_unlock(&sctx->stat_lock);
2798 return -ENOMEM;
2799 }
2800
2801 /* Seed devices of a new filesystem has their own generation. */
2802 if (scrub_dev->fs_devices != fs_info->fs_devices)
2803 gen = scrub_dev->generation;
2804 else
2805 gen = btrfs_get_last_trans_committed(fs_info);
2806
2807 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2808 bytenr = btrfs_sb_offset(i);
2809 if (bytenr + BTRFS_SUPER_INFO_SIZE >
2810 scrub_dev->commit_total_bytes)
2811 break;
2812 if (!btrfs_check_super_location(scrub_dev, bytenr))
2813 continue;
2814
2815 ret = scrub_one_super(sctx, scrub_dev, page, bytenr, gen);
2816 if (ret) {
2817 spin_lock(&sctx->stat_lock);
2818 sctx->stat.super_errors++;
2819 spin_unlock(&sctx->stat_lock);
2820 }
2821 }
2822 __free_page(page);
2823 return 0;
2824}
2825
2826static void scrub_workers_put(struct btrfs_fs_info *fs_info)
2827{
2828 if (refcount_dec_and_mutex_lock(&fs_info->scrub_workers_refcnt,
2829 &fs_info->scrub_lock)) {
2830 struct workqueue_struct *scrub_workers = fs_info->scrub_workers;
2831
2832 fs_info->scrub_workers = NULL;
2833 mutex_unlock(&fs_info->scrub_lock);
2834
2835 if (scrub_workers)
2836 destroy_workqueue(scrub_workers);
2837 }
2838}
2839
2840/*
2841 * get a reference count on fs_info->scrub_workers. start worker if necessary
2842 */
2843static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info)
2844{
2845 struct workqueue_struct *scrub_workers = NULL;
2846 unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
2847 int max_active = fs_info->thread_pool_size;
2848 int ret = -ENOMEM;
2849
2850 if (refcount_inc_not_zero(&fs_info->scrub_workers_refcnt))
2851 return 0;
2852
2853 scrub_workers = alloc_workqueue("btrfs-scrub", flags, max_active);
2854 if (!scrub_workers)
2855 return -ENOMEM;
2856
2857 mutex_lock(&fs_info->scrub_lock);
2858 if (refcount_read(&fs_info->scrub_workers_refcnt) == 0) {
2859 ASSERT(fs_info->scrub_workers == NULL);
2860 fs_info->scrub_workers = scrub_workers;
2861 refcount_set(&fs_info->scrub_workers_refcnt, 1);
2862 mutex_unlock(&fs_info->scrub_lock);
2863 return 0;
2864 }
2865 /* Other thread raced in and created the workers for us */
2866 refcount_inc(&fs_info->scrub_workers_refcnt);
2867 mutex_unlock(&fs_info->scrub_lock);
2868
2869 ret = 0;
2870
2871 destroy_workqueue(scrub_workers);
2872 return ret;
2873}
2874
2875int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
2876 u64 end, struct btrfs_scrub_progress *progress,
2877 int readonly, int is_dev_replace)
2878{
2879 struct btrfs_dev_lookup_args args = { .devid = devid };
2880 struct scrub_ctx *sctx;
2881 int ret;
2882 struct btrfs_device *dev;
2883 unsigned int nofs_flag;
2884 bool need_commit = false;
2885
2886 if (btrfs_fs_closing(fs_info))
2887 return -EAGAIN;
2888
2889 /* At mount time we have ensured nodesize is in the range of [4K, 64K]. */
2890 ASSERT(fs_info->nodesize <= BTRFS_STRIPE_LEN);
2891
2892 /*
2893 * SCRUB_MAX_SECTORS_PER_BLOCK is calculated using the largest possible
2894 * value (max nodesize / min sectorsize), thus nodesize should always
2895 * be fine.
2896 */
2897 ASSERT(fs_info->nodesize <=
2898 SCRUB_MAX_SECTORS_PER_BLOCK << fs_info->sectorsize_bits);
2899
2900 /* Allocate outside of device_list_mutex */
2901 sctx = scrub_setup_ctx(fs_info, is_dev_replace);
2902 if (IS_ERR(sctx))
2903 return PTR_ERR(sctx);
2904
2905 ret = scrub_workers_get(fs_info);
2906 if (ret)
2907 goto out_free_ctx;
2908
2909 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2910 dev = btrfs_find_device(fs_info->fs_devices, &args);
2911 if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
2912 !is_dev_replace)) {
2913 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2914 ret = -ENODEV;
2915 goto out;
2916 }
2917
2918 if (!is_dev_replace && !readonly &&
2919 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
2920 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2921 btrfs_err_in_rcu(fs_info,
2922 "scrub on devid %llu: filesystem on %s is not writable",
2923 devid, btrfs_dev_name(dev));
2924 ret = -EROFS;
2925 goto out;
2926 }
2927
2928 mutex_lock(&fs_info->scrub_lock);
2929 if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
2930 test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
2931 mutex_unlock(&fs_info->scrub_lock);
2932 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2933 ret = -EIO;
2934 goto out;
2935 }
2936
2937 down_read(&fs_info->dev_replace.rwsem);
2938 if (dev->scrub_ctx ||
2939 (!is_dev_replace &&
2940 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
2941 up_read(&fs_info->dev_replace.rwsem);
2942 mutex_unlock(&fs_info->scrub_lock);
2943 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2944 ret = -EINPROGRESS;
2945 goto out;
2946 }
2947 up_read(&fs_info->dev_replace.rwsem);
2948
2949 sctx->readonly = readonly;
2950 dev->scrub_ctx = sctx;
2951 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2952
2953 /*
2954 * checking @scrub_pause_req here, we can avoid
2955 * race between committing transaction and scrubbing.
2956 */
2957 __scrub_blocked_if_needed(fs_info);
2958 atomic_inc(&fs_info->scrubs_running);
2959 mutex_unlock(&fs_info->scrub_lock);
2960
2961 /*
2962 * In order to avoid deadlock with reclaim when there is a transaction
2963 * trying to pause scrub, make sure we use GFP_NOFS for all the
2964 * allocations done at btrfs_scrub_sectors() and scrub_sectors_for_parity()
2965 * invoked by our callees. The pausing request is done when the
2966 * transaction commit starts, and it blocks the transaction until scrub
2967 * is paused (done at specific points at scrub_stripe() or right above
2968 * before incrementing fs_info->scrubs_running).
2969 */
2970 nofs_flag = memalloc_nofs_save();
2971 if (!is_dev_replace) {
2972 u64 old_super_errors;
2973
2974 spin_lock(&sctx->stat_lock);
2975 old_super_errors = sctx->stat.super_errors;
2976 spin_unlock(&sctx->stat_lock);
2977
2978 btrfs_info(fs_info, "scrub: started on devid %llu", devid);
2979 /*
2980 * by holding device list mutex, we can
2981 * kick off writing super in log tree sync.
2982 */
2983 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2984 ret = scrub_supers(sctx, dev);
2985 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2986
2987 spin_lock(&sctx->stat_lock);
2988 /*
2989 * Super block errors found, but we can not commit transaction
2990 * at current context, since btrfs_commit_transaction() needs
2991 * to pause the current running scrub (hold by ourselves).
2992 */
2993 if (sctx->stat.super_errors > old_super_errors && !sctx->readonly)
2994 need_commit = true;
2995 spin_unlock(&sctx->stat_lock);
2996 }
2997
2998 if (!ret)
2999 ret = scrub_enumerate_chunks(sctx, dev, start, end);
3000 memalloc_nofs_restore(nofs_flag);
3001
3002 atomic_dec(&fs_info->scrubs_running);
3003 wake_up(&fs_info->scrub_pause_wait);
3004
3005 if (progress)
3006 memcpy(progress, &sctx->stat, sizeof(*progress));
3007
3008 if (!is_dev_replace)
3009 btrfs_info(fs_info, "scrub: %s on devid %llu with status: %d",
3010 ret ? "not finished" : "finished", devid, ret);
3011
3012 mutex_lock(&fs_info->scrub_lock);
3013 dev->scrub_ctx = NULL;
3014 mutex_unlock(&fs_info->scrub_lock);
3015
3016 scrub_workers_put(fs_info);
3017 scrub_put_ctx(sctx);
3018
3019 /*
3020 * We found some super block errors before, now try to force a
3021 * transaction commit, as scrub has finished.
3022 */
3023 if (need_commit) {
3024 struct btrfs_trans_handle *trans;
3025
3026 trans = btrfs_start_transaction(fs_info->tree_root, 0);
3027 if (IS_ERR(trans)) {
3028 ret = PTR_ERR(trans);
3029 btrfs_err(fs_info,
3030 "scrub: failed to start transaction to fix super block errors: %d", ret);
3031 return ret;
3032 }
3033 ret = btrfs_commit_transaction(trans);
3034 if (ret < 0)
3035 btrfs_err(fs_info,
3036 "scrub: failed to commit transaction to fix super block errors: %d", ret);
3037 }
3038 return ret;
3039out:
3040 scrub_workers_put(fs_info);
3041out_free_ctx:
3042 scrub_free_ctx(sctx);
3043
3044 return ret;
3045}
3046
3047void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
3048{
3049 mutex_lock(&fs_info->scrub_lock);
3050 atomic_inc(&fs_info->scrub_pause_req);
3051 while (atomic_read(&fs_info->scrubs_paused) !=
3052 atomic_read(&fs_info->scrubs_running)) {
3053 mutex_unlock(&fs_info->scrub_lock);
3054 wait_event(fs_info->scrub_pause_wait,
3055 atomic_read(&fs_info->scrubs_paused) ==
3056 atomic_read(&fs_info->scrubs_running));
3057 mutex_lock(&fs_info->scrub_lock);
3058 }
3059 mutex_unlock(&fs_info->scrub_lock);
3060}
3061
3062void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
3063{
3064 atomic_dec(&fs_info->scrub_pause_req);
3065 wake_up(&fs_info->scrub_pause_wait);
3066}
3067
3068int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
3069{
3070 mutex_lock(&fs_info->scrub_lock);
3071 if (!atomic_read(&fs_info->scrubs_running)) {
3072 mutex_unlock(&fs_info->scrub_lock);
3073 return -ENOTCONN;
3074 }
3075
3076 atomic_inc(&fs_info->scrub_cancel_req);
3077 while (atomic_read(&fs_info->scrubs_running)) {
3078 mutex_unlock(&fs_info->scrub_lock);
3079 wait_event(fs_info->scrub_pause_wait,
3080 atomic_read(&fs_info->scrubs_running) == 0);
3081 mutex_lock(&fs_info->scrub_lock);
3082 }
3083 atomic_dec(&fs_info->scrub_cancel_req);
3084 mutex_unlock(&fs_info->scrub_lock);
3085
3086 return 0;
3087}
3088
3089int btrfs_scrub_cancel_dev(struct btrfs_device *dev)
3090{
3091 struct btrfs_fs_info *fs_info = dev->fs_info;
3092 struct scrub_ctx *sctx;
3093
3094 mutex_lock(&fs_info->scrub_lock);
3095 sctx = dev->scrub_ctx;
3096 if (!sctx) {
3097 mutex_unlock(&fs_info->scrub_lock);
3098 return -ENOTCONN;
3099 }
3100 atomic_inc(&sctx->cancel_req);
3101 while (dev->scrub_ctx) {
3102 mutex_unlock(&fs_info->scrub_lock);
3103 wait_event(fs_info->scrub_pause_wait,
3104 dev->scrub_ctx == NULL);
3105 mutex_lock(&fs_info->scrub_lock);
3106 }
3107 mutex_unlock(&fs_info->scrub_lock);
3108
3109 return 0;
3110}
3111
3112int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
3113 struct btrfs_scrub_progress *progress)
3114{
3115 struct btrfs_dev_lookup_args args = { .devid = devid };
3116 struct btrfs_device *dev;
3117 struct scrub_ctx *sctx = NULL;
3118
3119 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3120 dev = btrfs_find_device(fs_info->fs_devices, &args);
3121 if (dev)
3122 sctx = dev->scrub_ctx;
3123 if (sctx)
3124 memcpy(progress, &sctx->stat, sizeof(*progress));
3125 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3126
3127 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3128}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2011, 2012 STRATO. All rights reserved.
4 */
5
6#include <linux/blkdev.h>
7#include <linux/ratelimit.h>
8#include <linux/sched/mm.h>
9#include <crypto/hash.h>
10#include "ctree.h"
11#include "volumes.h"
12#include "disk-io.h"
13#include "ordered-data.h"
14#include "transaction.h"
15#include "backref.h"
16#include "extent_io.h"
17#include "dev-replace.h"
18#include "check-integrity.h"
19#include "rcu-string.h"
20#include "raid56.h"
21#include "block-group.h"
22
23/*
24 * This is only the first step towards a full-features scrub. It reads all
25 * extent and super block and verifies the checksums. In case a bad checksum
26 * is found or the extent cannot be read, good data will be written back if
27 * any can be found.
28 *
29 * Future enhancements:
30 * - In case an unrepairable extent is encountered, track which files are
31 * affected and report them
32 * - track and record media errors, throw out bad devices
33 * - add a mode to also read unallocated space
34 */
35
36struct scrub_block;
37struct scrub_ctx;
38
39/*
40 * the following three values only influence the performance.
41 * The last one configures the number of parallel and outstanding I/O
42 * operations. The first two values configure an upper limit for the number
43 * of (dynamically allocated) pages that are added to a bio.
44 */
45#define SCRUB_PAGES_PER_RD_BIO 32 /* 128k per bio */
46#define SCRUB_PAGES_PER_WR_BIO 32 /* 128k per bio */
47#define SCRUB_BIOS_PER_SCTX 64 /* 8MB per device in flight */
48
49/*
50 * the following value times PAGE_SIZE needs to be large enough to match the
51 * largest node/leaf/sector size that shall be supported.
52 * Values larger than BTRFS_STRIPE_LEN are not supported.
53 */
54#define SCRUB_MAX_PAGES_PER_BLOCK 16 /* 64k per node/leaf/sector */
55
56struct scrub_recover {
57 refcount_t refs;
58 struct btrfs_bio *bbio;
59 u64 map_length;
60};
61
62struct scrub_page {
63 struct scrub_block *sblock;
64 struct page *page;
65 struct btrfs_device *dev;
66 struct list_head list;
67 u64 flags; /* extent flags */
68 u64 generation;
69 u64 logical;
70 u64 physical;
71 u64 physical_for_dev_replace;
72 atomic_t refs;
73 struct {
74 unsigned int mirror_num:8;
75 unsigned int have_csum:1;
76 unsigned int io_error:1;
77 };
78 u8 csum[BTRFS_CSUM_SIZE];
79
80 struct scrub_recover *recover;
81};
82
83struct scrub_bio {
84 int index;
85 struct scrub_ctx *sctx;
86 struct btrfs_device *dev;
87 struct bio *bio;
88 blk_status_t status;
89 u64 logical;
90 u64 physical;
91#if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
92 struct scrub_page *pagev[SCRUB_PAGES_PER_WR_BIO];
93#else
94 struct scrub_page *pagev[SCRUB_PAGES_PER_RD_BIO];
95#endif
96 int page_count;
97 int next_free;
98 struct btrfs_work work;
99};
100
101struct scrub_block {
102 struct scrub_page *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
103 int page_count;
104 atomic_t outstanding_pages;
105 refcount_t refs; /* free mem on transition to zero */
106 struct scrub_ctx *sctx;
107 struct scrub_parity *sparity;
108 struct {
109 unsigned int header_error:1;
110 unsigned int checksum_error:1;
111 unsigned int no_io_error_seen:1;
112 unsigned int generation_error:1; /* also sets header_error */
113
114 /* The following is for the data used to check parity */
115 /* It is for the data with checksum */
116 unsigned int data_corrected:1;
117 };
118 struct btrfs_work work;
119};
120
121/* Used for the chunks with parity stripe such RAID5/6 */
122struct scrub_parity {
123 struct scrub_ctx *sctx;
124
125 struct btrfs_device *scrub_dev;
126
127 u64 logic_start;
128
129 u64 logic_end;
130
131 int nsectors;
132
133 u64 stripe_len;
134
135 refcount_t refs;
136
137 struct list_head spages;
138
139 /* Work of parity check and repair */
140 struct btrfs_work work;
141
142 /* Mark the parity blocks which have data */
143 unsigned long *dbitmap;
144
145 /*
146 * Mark the parity blocks which have data, but errors happen when
147 * read data or check data
148 */
149 unsigned long *ebitmap;
150
151 unsigned long bitmap[0];
152};
153
154struct scrub_ctx {
155 struct scrub_bio *bios[SCRUB_BIOS_PER_SCTX];
156 struct btrfs_fs_info *fs_info;
157 int first_free;
158 int curr;
159 atomic_t bios_in_flight;
160 atomic_t workers_pending;
161 spinlock_t list_lock;
162 wait_queue_head_t list_wait;
163 u16 csum_size;
164 struct list_head csum_list;
165 atomic_t cancel_req;
166 int readonly;
167 int pages_per_rd_bio;
168
169 int is_dev_replace;
170
171 struct scrub_bio *wr_curr_bio;
172 struct mutex wr_lock;
173 int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
174 struct btrfs_device *wr_tgtdev;
175 bool flush_all_writes;
176
177 /*
178 * statistics
179 */
180 struct btrfs_scrub_progress stat;
181 spinlock_t stat_lock;
182
183 /*
184 * Use a ref counter to avoid use-after-free issues. Scrub workers
185 * decrement bios_in_flight and workers_pending and then do a wakeup
186 * on the list_wait wait queue. We must ensure the main scrub task
187 * doesn't free the scrub context before or while the workers are
188 * doing the wakeup() call.
189 */
190 refcount_t refs;
191};
192
193struct scrub_warning {
194 struct btrfs_path *path;
195 u64 extent_item_size;
196 const char *errstr;
197 u64 physical;
198 u64 logical;
199 struct btrfs_device *dev;
200};
201
202struct full_stripe_lock {
203 struct rb_node node;
204 u64 logical;
205 u64 refs;
206 struct mutex mutex;
207};
208
209static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
210static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
211static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
212static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
213 struct scrub_block *sblocks_for_recheck);
214static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
215 struct scrub_block *sblock,
216 int retry_failed_mirror);
217static void scrub_recheck_block_checksum(struct scrub_block *sblock);
218static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
219 struct scrub_block *sblock_good);
220static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
221 struct scrub_block *sblock_good,
222 int page_num, int force_write);
223static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
224static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
225 int page_num);
226static int scrub_checksum_data(struct scrub_block *sblock);
227static int scrub_checksum_tree_block(struct scrub_block *sblock);
228static int scrub_checksum_super(struct scrub_block *sblock);
229static void scrub_block_get(struct scrub_block *sblock);
230static void scrub_block_put(struct scrub_block *sblock);
231static void scrub_page_get(struct scrub_page *spage);
232static void scrub_page_put(struct scrub_page *spage);
233static void scrub_parity_get(struct scrub_parity *sparity);
234static void scrub_parity_put(struct scrub_parity *sparity);
235static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
236 struct scrub_page *spage);
237static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
238 u64 physical, struct btrfs_device *dev, u64 flags,
239 u64 gen, int mirror_num, u8 *csum, int force,
240 u64 physical_for_dev_replace);
241static void scrub_bio_end_io(struct bio *bio);
242static void scrub_bio_end_io_worker(struct btrfs_work *work);
243static void scrub_block_complete(struct scrub_block *sblock);
244static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
245 u64 extent_logical, u64 extent_len,
246 u64 *extent_physical,
247 struct btrfs_device **extent_dev,
248 int *extent_mirror_num);
249static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
250 struct scrub_page *spage);
251static void scrub_wr_submit(struct scrub_ctx *sctx);
252static void scrub_wr_bio_end_io(struct bio *bio);
253static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
254static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
255static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
256static void scrub_put_ctx(struct scrub_ctx *sctx);
257
258static inline int scrub_is_page_on_raid56(struct scrub_page *page)
259{
260 return page->recover &&
261 (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
262}
263
264static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
265{
266 refcount_inc(&sctx->refs);
267 atomic_inc(&sctx->bios_in_flight);
268}
269
270static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
271{
272 atomic_dec(&sctx->bios_in_flight);
273 wake_up(&sctx->list_wait);
274 scrub_put_ctx(sctx);
275}
276
277static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
278{
279 while (atomic_read(&fs_info->scrub_pause_req)) {
280 mutex_unlock(&fs_info->scrub_lock);
281 wait_event(fs_info->scrub_pause_wait,
282 atomic_read(&fs_info->scrub_pause_req) == 0);
283 mutex_lock(&fs_info->scrub_lock);
284 }
285}
286
287static void scrub_pause_on(struct btrfs_fs_info *fs_info)
288{
289 atomic_inc(&fs_info->scrubs_paused);
290 wake_up(&fs_info->scrub_pause_wait);
291}
292
293static void scrub_pause_off(struct btrfs_fs_info *fs_info)
294{
295 mutex_lock(&fs_info->scrub_lock);
296 __scrub_blocked_if_needed(fs_info);
297 atomic_dec(&fs_info->scrubs_paused);
298 mutex_unlock(&fs_info->scrub_lock);
299
300 wake_up(&fs_info->scrub_pause_wait);
301}
302
303static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
304{
305 scrub_pause_on(fs_info);
306 scrub_pause_off(fs_info);
307}
308
309/*
310 * Insert new full stripe lock into full stripe locks tree
311 *
312 * Return pointer to existing or newly inserted full_stripe_lock structure if
313 * everything works well.
314 * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
315 *
316 * NOTE: caller must hold full_stripe_locks_root->lock before calling this
317 * function
318 */
319static struct full_stripe_lock *insert_full_stripe_lock(
320 struct btrfs_full_stripe_locks_tree *locks_root,
321 u64 fstripe_logical)
322{
323 struct rb_node **p;
324 struct rb_node *parent = NULL;
325 struct full_stripe_lock *entry;
326 struct full_stripe_lock *ret;
327
328 lockdep_assert_held(&locks_root->lock);
329
330 p = &locks_root->root.rb_node;
331 while (*p) {
332 parent = *p;
333 entry = rb_entry(parent, struct full_stripe_lock, node);
334 if (fstripe_logical < entry->logical) {
335 p = &(*p)->rb_left;
336 } else if (fstripe_logical > entry->logical) {
337 p = &(*p)->rb_right;
338 } else {
339 entry->refs++;
340 return entry;
341 }
342 }
343
344 /*
345 * Insert new lock.
346 */
347 ret = kmalloc(sizeof(*ret), GFP_KERNEL);
348 if (!ret)
349 return ERR_PTR(-ENOMEM);
350 ret->logical = fstripe_logical;
351 ret->refs = 1;
352 mutex_init(&ret->mutex);
353
354 rb_link_node(&ret->node, parent, p);
355 rb_insert_color(&ret->node, &locks_root->root);
356 return ret;
357}
358
359/*
360 * Search for a full stripe lock of a block group
361 *
362 * Return pointer to existing full stripe lock if found
363 * Return NULL if not found
364 */
365static struct full_stripe_lock *search_full_stripe_lock(
366 struct btrfs_full_stripe_locks_tree *locks_root,
367 u64 fstripe_logical)
368{
369 struct rb_node *node;
370 struct full_stripe_lock *entry;
371
372 lockdep_assert_held(&locks_root->lock);
373
374 node = locks_root->root.rb_node;
375 while (node) {
376 entry = rb_entry(node, struct full_stripe_lock, node);
377 if (fstripe_logical < entry->logical)
378 node = node->rb_left;
379 else if (fstripe_logical > entry->logical)
380 node = node->rb_right;
381 else
382 return entry;
383 }
384 return NULL;
385}
386
387/*
388 * Helper to get full stripe logical from a normal bytenr.
389 *
390 * Caller must ensure @cache is a RAID56 block group.
391 */
392static u64 get_full_stripe_logical(struct btrfs_block_group_cache *cache,
393 u64 bytenr)
394{
395 u64 ret;
396
397 /*
398 * Due to chunk item size limit, full stripe length should not be
399 * larger than U32_MAX. Just a sanity check here.
400 */
401 WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
402
403 /*
404 * round_down() can only handle power of 2, while RAID56 full
405 * stripe length can be 64KiB * n, so we need to manually round down.
406 */
407 ret = div64_u64(bytenr - cache->key.objectid, cache->full_stripe_len) *
408 cache->full_stripe_len + cache->key.objectid;
409 return ret;
410}
411
412/*
413 * Lock a full stripe to avoid concurrency of recovery and read
414 *
415 * It's only used for profiles with parities (RAID5/6), for other profiles it
416 * does nothing.
417 *
418 * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
419 * So caller must call unlock_full_stripe() at the same context.
420 *
421 * Return <0 if encounters error.
422 */
423static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
424 bool *locked_ret)
425{
426 struct btrfs_block_group_cache *bg_cache;
427 struct btrfs_full_stripe_locks_tree *locks_root;
428 struct full_stripe_lock *existing;
429 u64 fstripe_start;
430 int ret = 0;
431
432 *locked_ret = false;
433 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
434 if (!bg_cache) {
435 ASSERT(0);
436 return -ENOENT;
437 }
438
439 /* Profiles not based on parity don't need full stripe lock */
440 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
441 goto out;
442 locks_root = &bg_cache->full_stripe_locks_root;
443
444 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
445
446 /* Now insert the full stripe lock */
447 mutex_lock(&locks_root->lock);
448 existing = insert_full_stripe_lock(locks_root, fstripe_start);
449 mutex_unlock(&locks_root->lock);
450 if (IS_ERR(existing)) {
451 ret = PTR_ERR(existing);
452 goto out;
453 }
454 mutex_lock(&existing->mutex);
455 *locked_ret = true;
456out:
457 btrfs_put_block_group(bg_cache);
458 return ret;
459}
460
461/*
462 * Unlock a full stripe.
463 *
464 * NOTE: Caller must ensure it's the same context calling corresponding
465 * lock_full_stripe().
466 *
467 * Return 0 if we unlock full stripe without problem.
468 * Return <0 for error
469 */
470static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
471 bool locked)
472{
473 struct btrfs_block_group_cache *bg_cache;
474 struct btrfs_full_stripe_locks_tree *locks_root;
475 struct full_stripe_lock *fstripe_lock;
476 u64 fstripe_start;
477 bool freeit = false;
478 int ret = 0;
479
480 /* If we didn't acquire full stripe lock, no need to continue */
481 if (!locked)
482 return 0;
483
484 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
485 if (!bg_cache) {
486 ASSERT(0);
487 return -ENOENT;
488 }
489 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
490 goto out;
491
492 locks_root = &bg_cache->full_stripe_locks_root;
493 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
494
495 mutex_lock(&locks_root->lock);
496 fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
497 /* Unpaired unlock_full_stripe() detected */
498 if (!fstripe_lock) {
499 WARN_ON(1);
500 ret = -ENOENT;
501 mutex_unlock(&locks_root->lock);
502 goto out;
503 }
504
505 if (fstripe_lock->refs == 0) {
506 WARN_ON(1);
507 btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
508 fstripe_lock->logical);
509 } else {
510 fstripe_lock->refs--;
511 }
512
513 if (fstripe_lock->refs == 0) {
514 rb_erase(&fstripe_lock->node, &locks_root->root);
515 freeit = true;
516 }
517 mutex_unlock(&locks_root->lock);
518
519 mutex_unlock(&fstripe_lock->mutex);
520 if (freeit)
521 kfree(fstripe_lock);
522out:
523 btrfs_put_block_group(bg_cache);
524 return ret;
525}
526
527static void scrub_free_csums(struct scrub_ctx *sctx)
528{
529 while (!list_empty(&sctx->csum_list)) {
530 struct btrfs_ordered_sum *sum;
531 sum = list_first_entry(&sctx->csum_list,
532 struct btrfs_ordered_sum, list);
533 list_del(&sum->list);
534 kfree(sum);
535 }
536}
537
538static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
539{
540 int i;
541
542 if (!sctx)
543 return;
544
545 /* this can happen when scrub is cancelled */
546 if (sctx->curr != -1) {
547 struct scrub_bio *sbio = sctx->bios[sctx->curr];
548
549 for (i = 0; i < sbio->page_count; i++) {
550 WARN_ON(!sbio->pagev[i]->page);
551 scrub_block_put(sbio->pagev[i]->sblock);
552 }
553 bio_put(sbio->bio);
554 }
555
556 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
557 struct scrub_bio *sbio = sctx->bios[i];
558
559 if (!sbio)
560 break;
561 kfree(sbio);
562 }
563
564 kfree(sctx->wr_curr_bio);
565 scrub_free_csums(sctx);
566 kfree(sctx);
567}
568
569static void scrub_put_ctx(struct scrub_ctx *sctx)
570{
571 if (refcount_dec_and_test(&sctx->refs))
572 scrub_free_ctx(sctx);
573}
574
575static noinline_for_stack struct scrub_ctx *scrub_setup_ctx(
576 struct btrfs_fs_info *fs_info, int is_dev_replace)
577{
578 struct scrub_ctx *sctx;
579 int i;
580
581 sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
582 if (!sctx)
583 goto nomem;
584 refcount_set(&sctx->refs, 1);
585 sctx->is_dev_replace = is_dev_replace;
586 sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
587 sctx->curr = -1;
588 sctx->fs_info = fs_info;
589 INIT_LIST_HEAD(&sctx->csum_list);
590 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
591 struct scrub_bio *sbio;
592
593 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
594 if (!sbio)
595 goto nomem;
596 sctx->bios[i] = sbio;
597
598 sbio->index = i;
599 sbio->sctx = sctx;
600 sbio->page_count = 0;
601 btrfs_init_work(&sbio->work, btrfs_scrub_helper,
602 scrub_bio_end_io_worker, NULL, NULL);
603
604 if (i != SCRUB_BIOS_PER_SCTX - 1)
605 sctx->bios[i]->next_free = i + 1;
606 else
607 sctx->bios[i]->next_free = -1;
608 }
609 sctx->first_free = 0;
610 atomic_set(&sctx->bios_in_flight, 0);
611 atomic_set(&sctx->workers_pending, 0);
612 atomic_set(&sctx->cancel_req, 0);
613 sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
614
615 spin_lock_init(&sctx->list_lock);
616 spin_lock_init(&sctx->stat_lock);
617 init_waitqueue_head(&sctx->list_wait);
618
619 WARN_ON(sctx->wr_curr_bio != NULL);
620 mutex_init(&sctx->wr_lock);
621 sctx->wr_curr_bio = NULL;
622 if (is_dev_replace) {
623 WARN_ON(!fs_info->dev_replace.tgtdev);
624 sctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
625 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
626 sctx->flush_all_writes = false;
627 }
628
629 return sctx;
630
631nomem:
632 scrub_free_ctx(sctx);
633 return ERR_PTR(-ENOMEM);
634}
635
636static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
637 void *warn_ctx)
638{
639 u64 isize;
640 u32 nlink;
641 int ret;
642 int i;
643 unsigned nofs_flag;
644 struct extent_buffer *eb;
645 struct btrfs_inode_item *inode_item;
646 struct scrub_warning *swarn = warn_ctx;
647 struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
648 struct inode_fs_paths *ipath = NULL;
649 struct btrfs_root *local_root;
650 struct btrfs_key root_key;
651 struct btrfs_key key;
652
653 root_key.objectid = root;
654 root_key.type = BTRFS_ROOT_ITEM_KEY;
655 root_key.offset = (u64)-1;
656 local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
657 if (IS_ERR(local_root)) {
658 ret = PTR_ERR(local_root);
659 goto err;
660 }
661
662 /*
663 * this makes the path point to (inum INODE_ITEM ioff)
664 */
665 key.objectid = inum;
666 key.type = BTRFS_INODE_ITEM_KEY;
667 key.offset = 0;
668
669 ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
670 if (ret) {
671 btrfs_release_path(swarn->path);
672 goto err;
673 }
674
675 eb = swarn->path->nodes[0];
676 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
677 struct btrfs_inode_item);
678 isize = btrfs_inode_size(eb, inode_item);
679 nlink = btrfs_inode_nlink(eb, inode_item);
680 btrfs_release_path(swarn->path);
681
682 /*
683 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
684 * uses GFP_NOFS in this context, so we keep it consistent but it does
685 * not seem to be strictly necessary.
686 */
687 nofs_flag = memalloc_nofs_save();
688 ipath = init_ipath(4096, local_root, swarn->path);
689 memalloc_nofs_restore(nofs_flag);
690 if (IS_ERR(ipath)) {
691 ret = PTR_ERR(ipath);
692 ipath = NULL;
693 goto err;
694 }
695 ret = paths_from_inode(inum, ipath);
696
697 if (ret < 0)
698 goto err;
699
700 /*
701 * we deliberately ignore the bit ipath might have been too small to
702 * hold all of the paths here
703 */
704 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
705 btrfs_warn_in_rcu(fs_info,
706"%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
707 swarn->errstr, swarn->logical,
708 rcu_str_deref(swarn->dev->name),
709 swarn->physical,
710 root, inum, offset,
711 min(isize - offset, (u64)PAGE_SIZE), nlink,
712 (char *)(unsigned long)ipath->fspath->val[i]);
713
714 free_ipath(ipath);
715 return 0;
716
717err:
718 btrfs_warn_in_rcu(fs_info,
719 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
720 swarn->errstr, swarn->logical,
721 rcu_str_deref(swarn->dev->name),
722 swarn->physical,
723 root, inum, offset, ret);
724
725 free_ipath(ipath);
726 return 0;
727}
728
729static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
730{
731 struct btrfs_device *dev;
732 struct btrfs_fs_info *fs_info;
733 struct btrfs_path *path;
734 struct btrfs_key found_key;
735 struct extent_buffer *eb;
736 struct btrfs_extent_item *ei;
737 struct scrub_warning swarn;
738 unsigned long ptr = 0;
739 u64 extent_item_pos;
740 u64 flags = 0;
741 u64 ref_root;
742 u32 item_size;
743 u8 ref_level = 0;
744 int ret;
745
746 WARN_ON(sblock->page_count < 1);
747 dev = sblock->pagev[0]->dev;
748 fs_info = sblock->sctx->fs_info;
749
750 path = btrfs_alloc_path();
751 if (!path)
752 return;
753
754 swarn.physical = sblock->pagev[0]->physical;
755 swarn.logical = sblock->pagev[0]->logical;
756 swarn.errstr = errstr;
757 swarn.dev = NULL;
758
759 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
760 &flags);
761 if (ret < 0)
762 goto out;
763
764 extent_item_pos = swarn.logical - found_key.objectid;
765 swarn.extent_item_size = found_key.offset;
766
767 eb = path->nodes[0];
768 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
769 item_size = btrfs_item_size_nr(eb, path->slots[0]);
770
771 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
772 do {
773 ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
774 item_size, &ref_root,
775 &ref_level);
776 btrfs_warn_in_rcu(fs_info,
777"%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
778 errstr, swarn.logical,
779 rcu_str_deref(dev->name),
780 swarn.physical,
781 ref_level ? "node" : "leaf",
782 ret < 0 ? -1 : ref_level,
783 ret < 0 ? -1 : ref_root);
784 } while (ret != 1);
785 btrfs_release_path(path);
786 } else {
787 btrfs_release_path(path);
788 swarn.path = path;
789 swarn.dev = dev;
790 iterate_extent_inodes(fs_info, found_key.objectid,
791 extent_item_pos, 1,
792 scrub_print_warning_inode, &swarn, false);
793 }
794
795out:
796 btrfs_free_path(path);
797}
798
799static inline void scrub_get_recover(struct scrub_recover *recover)
800{
801 refcount_inc(&recover->refs);
802}
803
804static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
805 struct scrub_recover *recover)
806{
807 if (refcount_dec_and_test(&recover->refs)) {
808 btrfs_bio_counter_dec(fs_info);
809 btrfs_put_bbio(recover->bbio);
810 kfree(recover);
811 }
812}
813
814/*
815 * scrub_handle_errored_block gets called when either verification of the
816 * pages failed or the bio failed to read, e.g. with EIO. In the latter
817 * case, this function handles all pages in the bio, even though only one
818 * may be bad.
819 * The goal of this function is to repair the errored block by using the
820 * contents of one of the mirrors.
821 */
822static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
823{
824 struct scrub_ctx *sctx = sblock_to_check->sctx;
825 struct btrfs_device *dev;
826 struct btrfs_fs_info *fs_info;
827 u64 logical;
828 unsigned int failed_mirror_index;
829 unsigned int is_metadata;
830 unsigned int have_csum;
831 struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
832 struct scrub_block *sblock_bad;
833 int ret;
834 int mirror_index;
835 int page_num;
836 int success;
837 bool full_stripe_locked;
838 unsigned int nofs_flag;
839 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
840 DEFAULT_RATELIMIT_BURST);
841
842 BUG_ON(sblock_to_check->page_count < 1);
843 fs_info = sctx->fs_info;
844 if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
845 /*
846 * if we find an error in a super block, we just report it.
847 * They will get written with the next transaction commit
848 * anyway
849 */
850 spin_lock(&sctx->stat_lock);
851 ++sctx->stat.super_errors;
852 spin_unlock(&sctx->stat_lock);
853 return 0;
854 }
855 logical = sblock_to_check->pagev[0]->logical;
856 BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
857 failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
858 is_metadata = !(sblock_to_check->pagev[0]->flags &
859 BTRFS_EXTENT_FLAG_DATA);
860 have_csum = sblock_to_check->pagev[0]->have_csum;
861 dev = sblock_to_check->pagev[0]->dev;
862
863 /*
864 * We must use GFP_NOFS because the scrub task might be waiting for a
865 * worker task executing this function and in turn a transaction commit
866 * might be waiting the scrub task to pause (which needs to wait for all
867 * the worker tasks to complete before pausing).
868 * We do allocations in the workers through insert_full_stripe_lock()
869 * and scrub_add_page_to_wr_bio(), which happens down the call chain of
870 * this function.
871 */
872 nofs_flag = memalloc_nofs_save();
873 /*
874 * For RAID5/6, race can happen for a different device scrub thread.
875 * For data corruption, Parity and Data threads will both try
876 * to recovery the data.
877 * Race can lead to doubly added csum error, or even unrecoverable
878 * error.
879 */
880 ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
881 if (ret < 0) {
882 memalloc_nofs_restore(nofs_flag);
883 spin_lock(&sctx->stat_lock);
884 if (ret == -ENOMEM)
885 sctx->stat.malloc_errors++;
886 sctx->stat.read_errors++;
887 sctx->stat.uncorrectable_errors++;
888 spin_unlock(&sctx->stat_lock);
889 return ret;
890 }
891
892 /*
893 * read all mirrors one after the other. This includes to
894 * re-read the extent or metadata block that failed (that was
895 * the cause that this fixup code is called) another time,
896 * page by page this time in order to know which pages
897 * caused I/O errors and which ones are good (for all mirrors).
898 * It is the goal to handle the situation when more than one
899 * mirror contains I/O errors, but the errors do not
900 * overlap, i.e. the data can be repaired by selecting the
901 * pages from those mirrors without I/O error on the
902 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
903 * would be that mirror #1 has an I/O error on the first page,
904 * the second page is good, and mirror #2 has an I/O error on
905 * the second page, but the first page is good.
906 * Then the first page of the first mirror can be repaired by
907 * taking the first page of the second mirror, and the
908 * second page of the second mirror can be repaired by
909 * copying the contents of the 2nd page of the 1st mirror.
910 * One more note: if the pages of one mirror contain I/O
911 * errors, the checksum cannot be verified. In order to get
912 * the best data for repairing, the first attempt is to find
913 * a mirror without I/O errors and with a validated checksum.
914 * Only if this is not possible, the pages are picked from
915 * mirrors with I/O errors without considering the checksum.
916 * If the latter is the case, at the end, the checksum of the
917 * repaired area is verified in order to correctly maintain
918 * the statistics.
919 */
920
921 sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
922 sizeof(*sblocks_for_recheck), GFP_KERNEL);
923 if (!sblocks_for_recheck) {
924 spin_lock(&sctx->stat_lock);
925 sctx->stat.malloc_errors++;
926 sctx->stat.read_errors++;
927 sctx->stat.uncorrectable_errors++;
928 spin_unlock(&sctx->stat_lock);
929 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
930 goto out;
931 }
932
933 /* setup the context, map the logical blocks and alloc the pages */
934 ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
935 if (ret) {
936 spin_lock(&sctx->stat_lock);
937 sctx->stat.read_errors++;
938 sctx->stat.uncorrectable_errors++;
939 spin_unlock(&sctx->stat_lock);
940 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
941 goto out;
942 }
943 BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
944 sblock_bad = sblocks_for_recheck + failed_mirror_index;
945
946 /* build and submit the bios for the failed mirror, check checksums */
947 scrub_recheck_block(fs_info, sblock_bad, 1);
948
949 if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
950 sblock_bad->no_io_error_seen) {
951 /*
952 * the error disappeared after reading page by page, or
953 * the area was part of a huge bio and other parts of the
954 * bio caused I/O errors, or the block layer merged several
955 * read requests into one and the error is caused by a
956 * different bio (usually one of the two latter cases is
957 * the cause)
958 */
959 spin_lock(&sctx->stat_lock);
960 sctx->stat.unverified_errors++;
961 sblock_to_check->data_corrected = 1;
962 spin_unlock(&sctx->stat_lock);
963
964 if (sctx->is_dev_replace)
965 scrub_write_block_to_dev_replace(sblock_bad);
966 goto out;
967 }
968
969 if (!sblock_bad->no_io_error_seen) {
970 spin_lock(&sctx->stat_lock);
971 sctx->stat.read_errors++;
972 spin_unlock(&sctx->stat_lock);
973 if (__ratelimit(&_rs))
974 scrub_print_warning("i/o error", sblock_to_check);
975 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
976 } else if (sblock_bad->checksum_error) {
977 spin_lock(&sctx->stat_lock);
978 sctx->stat.csum_errors++;
979 spin_unlock(&sctx->stat_lock);
980 if (__ratelimit(&_rs))
981 scrub_print_warning("checksum error", sblock_to_check);
982 btrfs_dev_stat_inc_and_print(dev,
983 BTRFS_DEV_STAT_CORRUPTION_ERRS);
984 } else if (sblock_bad->header_error) {
985 spin_lock(&sctx->stat_lock);
986 sctx->stat.verify_errors++;
987 spin_unlock(&sctx->stat_lock);
988 if (__ratelimit(&_rs))
989 scrub_print_warning("checksum/header error",
990 sblock_to_check);
991 if (sblock_bad->generation_error)
992 btrfs_dev_stat_inc_and_print(dev,
993 BTRFS_DEV_STAT_GENERATION_ERRS);
994 else
995 btrfs_dev_stat_inc_and_print(dev,
996 BTRFS_DEV_STAT_CORRUPTION_ERRS);
997 }
998
999 if (sctx->readonly) {
1000 ASSERT(!sctx->is_dev_replace);
1001 goto out;
1002 }
1003
1004 /*
1005 * now build and submit the bios for the other mirrors, check
1006 * checksums.
1007 * First try to pick the mirror which is completely without I/O
1008 * errors and also does not have a checksum error.
1009 * If one is found, and if a checksum is present, the full block
1010 * that is known to contain an error is rewritten. Afterwards
1011 * the block is known to be corrected.
1012 * If a mirror is found which is completely correct, and no
1013 * checksum is present, only those pages are rewritten that had
1014 * an I/O error in the block to be repaired, since it cannot be
1015 * determined, which copy of the other pages is better (and it
1016 * could happen otherwise that a correct page would be
1017 * overwritten by a bad one).
1018 */
1019 for (mirror_index = 0; ;mirror_index++) {
1020 struct scrub_block *sblock_other;
1021
1022 if (mirror_index == failed_mirror_index)
1023 continue;
1024
1025 /* raid56's mirror can be more than BTRFS_MAX_MIRRORS */
1026 if (!scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1027 if (mirror_index >= BTRFS_MAX_MIRRORS)
1028 break;
1029 if (!sblocks_for_recheck[mirror_index].page_count)
1030 break;
1031
1032 sblock_other = sblocks_for_recheck + mirror_index;
1033 } else {
1034 struct scrub_recover *r = sblock_bad->pagev[0]->recover;
1035 int max_allowed = r->bbio->num_stripes -
1036 r->bbio->num_tgtdevs;
1037
1038 if (mirror_index >= max_allowed)
1039 break;
1040 if (!sblocks_for_recheck[1].page_count)
1041 break;
1042
1043 ASSERT(failed_mirror_index == 0);
1044 sblock_other = sblocks_for_recheck + 1;
1045 sblock_other->pagev[0]->mirror_num = 1 + mirror_index;
1046 }
1047
1048 /* build and submit the bios, check checksums */
1049 scrub_recheck_block(fs_info, sblock_other, 0);
1050
1051 if (!sblock_other->header_error &&
1052 !sblock_other->checksum_error &&
1053 sblock_other->no_io_error_seen) {
1054 if (sctx->is_dev_replace) {
1055 scrub_write_block_to_dev_replace(sblock_other);
1056 goto corrected_error;
1057 } else {
1058 ret = scrub_repair_block_from_good_copy(
1059 sblock_bad, sblock_other);
1060 if (!ret)
1061 goto corrected_error;
1062 }
1063 }
1064 }
1065
1066 if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1067 goto did_not_correct_error;
1068
1069 /*
1070 * In case of I/O errors in the area that is supposed to be
1071 * repaired, continue by picking good copies of those pages.
1072 * Select the good pages from mirrors to rewrite bad pages from
1073 * the area to fix. Afterwards verify the checksum of the block
1074 * that is supposed to be repaired. This verification step is
1075 * only done for the purpose of statistic counting and for the
1076 * final scrub report, whether errors remain.
1077 * A perfect algorithm could make use of the checksum and try
1078 * all possible combinations of pages from the different mirrors
1079 * until the checksum verification succeeds. For example, when
1080 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1081 * of mirror #2 is readable but the final checksum test fails,
1082 * then the 2nd page of mirror #3 could be tried, whether now
1083 * the final checksum succeeds. But this would be a rare
1084 * exception and is therefore not implemented. At least it is
1085 * avoided that the good copy is overwritten.
1086 * A more useful improvement would be to pick the sectors
1087 * without I/O error based on sector sizes (512 bytes on legacy
1088 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1089 * mirror could be repaired by taking 512 byte of a different
1090 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1091 * area are unreadable.
1092 */
1093 success = 1;
1094 for (page_num = 0; page_num < sblock_bad->page_count;
1095 page_num++) {
1096 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1097 struct scrub_block *sblock_other = NULL;
1098
1099 /* skip no-io-error page in scrub */
1100 if (!page_bad->io_error && !sctx->is_dev_replace)
1101 continue;
1102
1103 if (scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1104 /*
1105 * In case of dev replace, if raid56 rebuild process
1106 * didn't work out correct data, then copy the content
1107 * in sblock_bad to make sure target device is identical
1108 * to source device, instead of writing garbage data in
1109 * sblock_for_recheck array to target device.
1110 */
1111 sblock_other = NULL;
1112 } else if (page_bad->io_error) {
1113 /* try to find no-io-error page in mirrors */
1114 for (mirror_index = 0;
1115 mirror_index < BTRFS_MAX_MIRRORS &&
1116 sblocks_for_recheck[mirror_index].page_count > 0;
1117 mirror_index++) {
1118 if (!sblocks_for_recheck[mirror_index].
1119 pagev[page_num]->io_error) {
1120 sblock_other = sblocks_for_recheck +
1121 mirror_index;
1122 break;
1123 }
1124 }
1125 if (!sblock_other)
1126 success = 0;
1127 }
1128
1129 if (sctx->is_dev_replace) {
1130 /*
1131 * did not find a mirror to fetch the page
1132 * from. scrub_write_page_to_dev_replace()
1133 * handles this case (page->io_error), by
1134 * filling the block with zeros before
1135 * submitting the write request
1136 */
1137 if (!sblock_other)
1138 sblock_other = sblock_bad;
1139
1140 if (scrub_write_page_to_dev_replace(sblock_other,
1141 page_num) != 0) {
1142 atomic64_inc(
1143 &fs_info->dev_replace.num_write_errors);
1144 success = 0;
1145 }
1146 } else if (sblock_other) {
1147 ret = scrub_repair_page_from_good_copy(sblock_bad,
1148 sblock_other,
1149 page_num, 0);
1150 if (0 == ret)
1151 page_bad->io_error = 0;
1152 else
1153 success = 0;
1154 }
1155 }
1156
1157 if (success && !sctx->is_dev_replace) {
1158 if (is_metadata || have_csum) {
1159 /*
1160 * need to verify the checksum now that all
1161 * sectors on disk are repaired (the write
1162 * request for data to be repaired is on its way).
1163 * Just be lazy and use scrub_recheck_block()
1164 * which re-reads the data before the checksum
1165 * is verified, but most likely the data comes out
1166 * of the page cache.
1167 */
1168 scrub_recheck_block(fs_info, sblock_bad, 1);
1169 if (!sblock_bad->header_error &&
1170 !sblock_bad->checksum_error &&
1171 sblock_bad->no_io_error_seen)
1172 goto corrected_error;
1173 else
1174 goto did_not_correct_error;
1175 } else {
1176corrected_error:
1177 spin_lock(&sctx->stat_lock);
1178 sctx->stat.corrected_errors++;
1179 sblock_to_check->data_corrected = 1;
1180 spin_unlock(&sctx->stat_lock);
1181 btrfs_err_rl_in_rcu(fs_info,
1182 "fixed up error at logical %llu on dev %s",
1183 logical, rcu_str_deref(dev->name));
1184 }
1185 } else {
1186did_not_correct_error:
1187 spin_lock(&sctx->stat_lock);
1188 sctx->stat.uncorrectable_errors++;
1189 spin_unlock(&sctx->stat_lock);
1190 btrfs_err_rl_in_rcu(fs_info,
1191 "unable to fixup (regular) error at logical %llu on dev %s",
1192 logical, rcu_str_deref(dev->name));
1193 }
1194
1195out:
1196 if (sblocks_for_recheck) {
1197 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1198 mirror_index++) {
1199 struct scrub_block *sblock = sblocks_for_recheck +
1200 mirror_index;
1201 struct scrub_recover *recover;
1202 int page_index;
1203
1204 for (page_index = 0; page_index < sblock->page_count;
1205 page_index++) {
1206 sblock->pagev[page_index]->sblock = NULL;
1207 recover = sblock->pagev[page_index]->recover;
1208 if (recover) {
1209 scrub_put_recover(fs_info, recover);
1210 sblock->pagev[page_index]->recover =
1211 NULL;
1212 }
1213 scrub_page_put(sblock->pagev[page_index]);
1214 }
1215 }
1216 kfree(sblocks_for_recheck);
1217 }
1218
1219 ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1220 memalloc_nofs_restore(nofs_flag);
1221 if (ret < 0)
1222 return ret;
1223 return 0;
1224}
1225
1226static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1227{
1228 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1229 return 2;
1230 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1231 return 3;
1232 else
1233 return (int)bbio->num_stripes;
1234}
1235
1236static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1237 u64 *raid_map,
1238 u64 mapped_length,
1239 int nstripes, int mirror,
1240 int *stripe_index,
1241 u64 *stripe_offset)
1242{
1243 int i;
1244
1245 if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1246 /* RAID5/6 */
1247 for (i = 0; i < nstripes; i++) {
1248 if (raid_map[i] == RAID6_Q_STRIPE ||
1249 raid_map[i] == RAID5_P_STRIPE)
1250 continue;
1251
1252 if (logical >= raid_map[i] &&
1253 logical < raid_map[i] + mapped_length)
1254 break;
1255 }
1256
1257 *stripe_index = i;
1258 *stripe_offset = logical - raid_map[i];
1259 } else {
1260 /* The other RAID type */
1261 *stripe_index = mirror;
1262 *stripe_offset = 0;
1263 }
1264}
1265
1266static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1267 struct scrub_block *sblocks_for_recheck)
1268{
1269 struct scrub_ctx *sctx = original_sblock->sctx;
1270 struct btrfs_fs_info *fs_info = sctx->fs_info;
1271 u64 length = original_sblock->page_count * PAGE_SIZE;
1272 u64 logical = original_sblock->pagev[0]->logical;
1273 u64 generation = original_sblock->pagev[0]->generation;
1274 u64 flags = original_sblock->pagev[0]->flags;
1275 u64 have_csum = original_sblock->pagev[0]->have_csum;
1276 struct scrub_recover *recover;
1277 struct btrfs_bio *bbio;
1278 u64 sublen;
1279 u64 mapped_length;
1280 u64 stripe_offset;
1281 int stripe_index;
1282 int page_index = 0;
1283 int mirror_index;
1284 int nmirrors;
1285 int ret;
1286
1287 /*
1288 * note: the two members refs and outstanding_pages
1289 * are not used (and not set) in the blocks that are used for
1290 * the recheck procedure
1291 */
1292
1293 while (length > 0) {
1294 sublen = min_t(u64, length, PAGE_SIZE);
1295 mapped_length = sublen;
1296 bbio = NULL;
1297
1298 /*
1299 * with a length of PAGE_SIZE, each returned stripe
1300 * represents one mirror
1301 */
1302 btrfs_bio_counter_inc_blocked(fs_info);
1303 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1304 logical, &mapped_length, &bbio);
1305 if (ret || !bbio || mapped_length < sublen) {
1306 btrfs_put_bbio(bbio);
1307 btrfs_bio_counter_dec(fs_info);
1308 return -EIO;
1309 }
1310
1311 recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1312 if (!recover) {
1313 btrfs_put_bbio(bbio);
1314 btrfs_bio_counter_dec(fs_info);
1315 return -ENOMEM;
1316 }
1317
1318 refcount_set(&recover->refs, 1);
1319 recover->bbio = bbio;
1320 recover->map_length = mapped_length;
1321
1322 BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1323
1324 nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1325
1326 for (mirror_index = 0; mirror_index < nmirrors;
1327 mirror_index++) {
1328 struct scrub_block *sblock;
1329 struct scrub_page *page;
1330
1331 sblock = sblocks_for_recheck + mirror_index;
1332 sblock->sctx = sctx;
1333
1334 page = kzalloc(sizeof(*page), GFP_NOFS);
1335 if (!page) {
1336leave_nomem:
1337 spin_lock(&sctx->stat_lock);
1338 sctx->stat.malloc_errors++;
1339 spin_unlock(&sctx->stat_lock);
1340 scrub_put_recover(fs_info, recover);
1341 return -ENOMEM;
1342 }
1343 scrub_page_get(page);
1344 sblock->pagev[page_index] = page;
1345 page->sblock = sblock;
1346 page->flags = flags;
1347 page->generation = generation;
1348 page->logical = logical;
1349 page->have_csum = have_csum;
1350 if (have_csum)
1351 memcpy(page->csum,
1352 original_sblock->pagev[0]->csum,
1353 sctx->csum_size);
1354
1355 scrub_stripe_index_and_offset(logical,
1356 bbio->map_type,
1357 bbio->raid_map,
1358 mapped_length,
1359 bbio->num_stripes -
1360 bbio->num_tgtdevs,
1361 mirror_index,
1362 &stripe_index,
1363 &stripe_offset);
1364 page->physical = bbio->stripes[stripe_index].physical +
1365 stripe_offset;
1366 page->dev = bbio->stripes[stripe_index].dev;
1367
1368 BUG_ON(page_index >= original_sblock->page_count);
1369 page->physical_for_dev_replace =
1370 original_sblock->pagev[page_index]->
1371 physical_for_dev_replace;
1372 /* for missing devices, dev->bdev is NULL */
1373 page->mirror_num = mirror_index + 1;
1374 sblock->page_count++;
1375 page->page = alloc_page(GFP_NOFS);
1376 if (!page->page)
1377 goto leave_nomem;
1378
1379 scrub_get_recover(recover);
1380 page->recover = recover;
1381 }
1382 scrub_put_recover(fs_info, recover);
1383 length -= sublen;
1384 logical += sublen;
1385 page_index++;
1386 }
1387
1388 return 0;
1389}
1390
1391static void scrub_bio_wait_endio(struct bio *bio)
1392{
1393 complete(bio->bi_private);
1394}
1395
1396static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1397 struct bio *bio,
1398 struct scrub_page *page)
1399{
1400 DECLARE_COMPLETION_ONSTACK(done);
1401 int ret;
1402 int mirror_num;
1403
1404 bio->bi_iter.bi_sector = page->logical >> 9;
1405 bio->bi_private = &done;
1406 bio->bi_end_io = scrub_bio_wait_endio;
1407
1408 mirror_num = page->sblock->pagev[0]->mirror_num;
1409 ret = raid56_parity_recover(fs_info, bio, page->recover->bbio,
1410 page->recover->map_length,
1411 mirror_num, 0);
1412 if (ret)
1413 return ret;
1414
1415 wait_for_completion_io(&done);
1416 return blk_status_to_errno(bio->bi_status);
1417}
1418
1419static void scrub_recheck_block_on_raid56(struct btrfs_fs_info *fs_info,
1420 struct scrub_block *sblock)
1421{
1422 struct scrub_page *first_page = sblock->pagev[0];
1423 struct bio *bio;
1424 int page_num;
1425
1426 /* All pages in sblock belong to the same stripe on the same device. */
1427 ASSERT(first_page->dev);
1428 if (!first_page->dev->bdev)
1429 goto out;
1430
1431 bio = btrfs_io_bio_alloc(BIO_MAX_PAGES);
1432 bio_set_dev(bio, first_page->dev->bdev);
1433
1434 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1435 struct scrub_page *page = sblock->pagev[page_num];
1436
1437 WARN_ON(!page->page);
1438 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1439 }
1440
1441 if (scrub_submit_raid56_bio_wait(fs_info, bio, first_page)) {
1442 bio_put(bio);
1443 goto out;
1444 }
1445
1446 bio_put(bio);
1447
1448 scrub_recheck_block_checksum(sblock);
1449
1450 return;
1451out:
1452 for (page_num = 0; page_num < sblock->page_count; page_num++)
1453 sblock->pagev[page_num]->io_error = 1;
1454
1455 sblock->no_io_error_seen = 0;
1456}
1457
1458/*
1459 * this function will check the on disk data for checksum errors, header
1460 * errors and read I/O errors. If any I/O errors happen, the exact pages
1461 * which are errored are marked as being bad. The goal is to enable scrub
1462 * to take those pages that are not errored from all the mirrors so that
1463 * the pages that are errored in the just handled mirror can be repaired.
1464 */
1465static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1466 struct scrub_block *sblock,
1467 int retry_failed_mirror)
1468{
1469 int page_num;
1470
1471 sblock->no_io_error_seen = 1;
1472
1473 /* short cut for raid56 */
1474 if (!retry_failed_mirror && scrub_is_page_on_raid56(sblock->pagev[0]))
1475 return scrub_recheck_block_on_raid56(fs_info, sblock);
1476
1477 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1478 struct bio *bio;
1479 struct scrub_page *page = sblock->pagev[page_num];
1480
1481 if (page->dev->bdev == NULL) {
1482 page->io_error = 1;
1483 sblock->no_io_error_seen = 0;
1484 continue;
1485 }
1486
1487 WARN_ON(!page->page);
1488 bio = btrfs_io_bio_alloc(1);
1489 bio_set_dev(bio, page->dev->bdev);
1490
1491 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1492 bio->bi_iter.bi_sector = page->physical >> 9;
1493 bio->bi_opf = REQ_OP_READ;
1494
1495 if (btrfsic_submit_bio_wait(bio)) {
1496 page->io_error = 1;
1497 sblock->no_io_error_seen = 0;
1498 }
1499
1500 bio_put(bio);
1501 }
1502
1503 if (sblock->no_io_error_seen)
1504 scrub_recheck_block_checksum(sblock);
1505}
1506
1507static inline int scrub_check_fsid(u8 fsid[],
1508 struct scrub_page *spage)
1509{
1510 struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1511 int ret;
1512
1513 ret = memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1514 return !ret;
1515}
1516
1517static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1518{
1519 sblock->header_error = 0;
1520 sblock->checksum_error = 0;
1521 sblock->generation_error = 0;
1522
1523 if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1524 scrub_checksum_data(sblock);
1525 else
1526 scrub_checksum_tree_block(sblock);
1527}
1528
1529static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1530 struct scrub_block *sblock_good)
1531{
1532 int page_num;
1533 int ret = 0;
1534
1535 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1536 int ret_sub;
1537
1538 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1539 sblock_good,
1540 page_num, 1);
1541 if (ret_sub)
1542 ret = ret_sub;
1543 }
1544
1545 return ret;
1546}
1547
1548static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1549 struct scrub_block *sblock_good,
1550 int page_num, int force_write)
1551{
1552 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1553 struct scrub_page *page_good = sblock_good->pagev[page_num];
1554 struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1555
1556 BUG_ON(page_bad->page == NULL);
1557 BUG_ON(page_good->page == NULL);
1558 if (force_write || sblock_bad->header_error ||
1559 sblock_bad->checksum_error || page_bad->io_error) {
1560 struct bio *bio;
1561 int ret;
1562
1563 if (!page_bad->dev->bdev) {
1564 btrfs_warn_rl(fs_info,
1565 "scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1566 return -EIO;
1567 }
1568
1569 bio = btrfs_io_bio_alloc(1);
1570 bio_set_dev(bio, page_bad->dev->bdev);
1571 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1572 bio->bi_opf = REQ_OP_WRITE;
1573
1574 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1575 if (PAGE_SIZE != ret) {
1576 bio_put(bio);
1577 return -EIO;
1578 }
1579
1580 if (btrfsic_submit_bio_wait(bio)) {
1581 btrfs_dev_stat_inc_and_print(page_bad->dev,
1582 BTRFS_DEV_STAT_WRITE_ERRS);
1583 atomic64_inc(&fs_info->dev_replace.num_write_errors);
1584 bio_put(bio);
1585 return -EIO;
1586 }
1587 bio_put(bio);
1588 }
1589
1590 return 0;
1591}
1592
1593static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1594{
1595 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1596 int page_num;
1597
1598 /*
1599 * This block is used for the check of the parity on the source device,
1600 * so the data needn't be written into the destination device.
1601 */
1602 if (sblock->sparity)
1603 return;
1604
1605 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1606 int ret;
1607
1608 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1609 if (ret)
1610 atomic64_inc(&fs_info->dev_replace.num_write_errors);
1611 }
1612}
1613
1614static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1615 int page_num)
1616{
1617 struct scrub_page *spage = sblock->pagev[page_num];
1618
1619 BUG_ON(spage->page == NULL);
1620 if (spage->io_error) {
1621 void *mapped_buffer = kmap_atomic(spage->page);
1622
1623 clear_page(mapped_buffer);
1624 flush_dcache_page(spage->page);
1625 kunmap_atomic(mapped_buffer);
1626 }
1627 return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1628}
1629
1630static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1631 struct scrub_page *spage)
1632{
1633 struct scrub_bio *sbio;
1634 int ret;
1635
1636 mutex_lock(&sctx->wr_lock);
1637again:
1638 if (!sctx->wr_curr_bio) {
1639 sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1640 GFP_KERNEL);
1641 if (!sctx->wr_curr_bio) {
1642 mutex_unlock(&sctx->wr_lock);
1643 return -ENOMEM;
1644 }
1645 sctx->wr_curr_bio->sctx = sctx;
1646 sctx->wr_curr_bio->page_count = 0;
1647 }
1648 sbio = sctx->wr_curr_bio;
1649 if (sbio->page_count == 0) {
1650 struct bio *bio;
1651
1652 sbio->physical = spage->physical_for_dev_replace;
1653 sbio->logical = spage->logical;
1654 sbio->dev = sctx->wr_tgtdev;
1655 bio = sbio->bio;
1656 if (!bio) {
1657 bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
1658 sbio->bio = bio;
1659 }
1660
1661 bio->bi_private = sbio;
1662 bio->bi_end_io = scrub_wr_bio_end_io;
1663 bio_set_dev(bio, sbio->dev->bdev);
1664 bio->bi_iter.bi_sector = sbio->physical >> 9;
1665 bio->bi_opf = REQ_OP_WRITE;
1666 sbio->status = 0;
1667 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1668 spage->physical_for_dev_replace ||
1669 sbio->logical + sbio->page_count * PAGE_SIZE !=
1670 spage->logical) {
1671 scrub_wr_submit(sctx);
1672 goto again;
1673 }
1674
1675 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1676 if (ret != PAGE_SIZE) {
1677 if (sbio->page_count < 1) {
1678 bio_put(sbio->bio);
1679 sbio->bio = NULL;
1680 mutex_unlock(&sctx->wr_lock);
1681 return -EIO;
1682 }
1683 scrub_wr_submit(sctx);
1684 goto again;
1685 }
1686
1687 sbio->pagev[sbio->page_count] = spage;
1688 scrub_page_get(spage);
1689 sbio->page_count++;
1690 if (sbio->page_count == sctx->pages_per_wr_bio)
1691 scrub_wr_submit(sctx);
1692 mutex_unlock(&sctx->wr_lock);
1693
1694 return 0;
1695}
1696
1697static void scrub_wr_submit(struct scrub_ctx *sctx)
1698{
1699 struct scrub_bio *sbio;
1700
1701 if (!sctx->wr_curr_bio)
1702 return;
1703
1704 sbio = sctx->wr_curr_bio;
1705 sctx->wr_curr_bio = NULL;
1706 WARN_ON(!sbio->bio->bi_disk);
1707 scrub_pending_bio_inc(sctx);
1708 /* process all writes in a single worker thread. Then the block layer
1709 * orders the requests before sending them to the driver which
1710 * doubled the write performance on spinning disks when measured
1711 * with Linux 3.5 */
1712 btrfsic_submit_bio(sbio->bio);
1713}
1714
1715static void scrub_wr_bio_end_io(struct bio *bio)
1716{
1717 struct scrub_bio *sbio = bio->bi_private;
1718 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
1719
1720 sbio->status = bio->bi_status;
1721 sbio->bio = bio;
1722
1723 btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
1724 scrub_wr_bio_end_io_worker, NULL, NULL);
1725 btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1726}
1727
1728static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1729{
1730 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1731 struct scrub_ctx *sctx = sbio->sctx;
1732 int i;
1733
1734 WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1735 if (sbio->status) {
1736 struct btrfs_dev_replace *dev_replace =
1737 &sbio->sctx->fs_info->dev_replace;
1738
1739 for (i = 0; i < sbio->page_count; i++) {
1740 struct scrub_page *spage = sbio->pagev[i];
1741
1742 spage->io_error = 1;
1743 atomic64_inc(&dev_replace->num_write_errors);
1744 }
1745 }
1746
1747 for (i = 0; i < sbio->page_count; i++)
1748 scrub_page_put(sbio->pagev[i]);
1749
1750 bio_put(sbio->bio);
1751 kfree(sbio);
1752 scrub_pending_bio_dec(sctx);
1753}
1754
1755static int scrub_checksum(struct scrub_block *sblock)
1756{
1757 u64 flags;
1758 int ret;
1759
1760 /*
1761 * No need to initialize these stats currently,
1762 * because this function only use return value
1763 * instead of these stats value.
1764 *
1765 * Todo:
1766 * always use stats
1767 */
1768 sblock->header_error = 0;
1769 sblock->generation_error = 0;
1770 sblock->checksum_error = 0;
1771
1772 WARN_ON(sblock->page_count < 1);
1773 flags = sblock->pagev[0]->flags;
1774 ret = 0;
1775 if (flags & BTRFS_EXTENT_FLAG_DATA)
1776 ret = scrub_checksum_data(sblock);
1777 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1778 ret = scrub_checksum_tree_block(sblock);
1779 else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1780 (void)scrub_checksum_super(sblock);
1781 else
1782 WARN_ON(1);
1783 if (ret)
1784 scrub_handle_errored_block(sblock);
1785
1786 return ret;
1787}
1788
1789static int scrub_checksum_data(struct scrub_block *sblock)
1790{
1791 struct scrub_ctx *sctx = sblock->sctx;
1792 struct btrfs_fs_info *fs_info = sctx->fs_info;
1793 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1794 u8 csum[BTRFS_CSUM_SIZE];
1795 u8 *on_disk_csum;
1796 struct page *page;
1797 void *buffer;
1798 u64 len;
1799 int index;
1800
1801 BUG_ON(sblock->page_count < 1);
1802 if (!sblock->pagev[0]->have_csum)
1803 return 0;
1804
1805 shash->tfm = fs_info->csum_shash;
1806 crypto_shash_init(shash);
1807
1808 on_disk_csum = sblock->pagev[0]->csum;
1809 page = sblock->pagev[0]->page;
1810 buffer = kmap_atomic(page);
1811
1812 len = sctx->fs_info->sectorsize;
1813 index = 0;
1814 for (;;) {
1815 u64 l = min_t(u64, len, PAGE_SIZE);
1816
1817 crypto_shash_update(shash, buffer, l);
1818 kunmap_atomic(buffer);
1819 len -= l;
1820 if (len == 0)
1821 break;
1822 index++;
1823 BUG_ON(index >= sblock->page_count);
1824 BUG_ON(!sblock->pagev[index]->page);
1825 page = sblock->pagev[index]->page;
1826 buffer = kmap_atomic(page);
1827 }
1828
1829 crypto_shash_final(shash, csum);
1830 if (memcmp(csum, on_disk_csum, sctx->csum_size))
1831 sblock->checksum_error = 1;
1832
1833 return sblock->checksum_error;
1834}
1835
1836static int scrub_checksum_tree_block(struct scrub_block *sblock)
1837{
1838 struct scrub_ctx *sctx = sblock->sctx;
1839 struct btrfs_header *h;
1840 struct btrfs_fs_info *fs_info = sctx->fs_info;
1841 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1842 u8 calculated_csum[BTRFS_CSUM_SIZE];
1843 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1844 struct page *page;
1845 void *mapped_buffer;
1846 u64 mapped_size;
1847 void *p;
1848 u64 len;
1849 int index;
1850
1851 shash->tfm = fs_info->csum_shash;
1852 crypto_shash_init(shash);
1853
1854 BUG_ON(sblock->page_count < 1);
1855 page = sblock->pagev[0]->page;
1856 mapped_buffer = kmap_atomic(page);
1857 h = (struct btrfs_header *)mapped_buffer;
1858 memcpy(on_disk_csum, h->csum, sctx->csum_size);
1859
1860 /*
1861 * we don't use the getter functions here, as we
1862 * a) don't have an extent buffer and
1863 * b) the page is already kmapped
1864 */
1865 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1866 sblock->header_error = 1;
1867
1868 if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
1869 sblock->header_error = 1;
1870 sblock->generation_error = 1;
1871 }
1872
1873 if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
1874 sblock->header_error = 1;
1875
1876 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1877 BTRFS_UUID_SIZE))
1878 sblock->header_error = 1;
1879
1880 len = sctx->fs_info->nodesize - BTRFS_CSUM_SIZE;
1881 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1882 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1883 index = 0;
1884 for (;;) {
1885 u64 l = min_t(u64, len, mapped_size);
1886
1887 crypto_shash_update(shash, p, l);
1888 kunmap_atomic(mapped_buffer);
1889 len -= l;
1890 if (len == 0)
1891 break;
1892 index++;
1893 BUG_ON(index >= sblock->page_count);
1894 BUG_ON(!sblock->pagev[index]->page);
1895 page = sblock->pagev[index]->page;
1896 mapped_buffer = kmap_atomic(page);
1897 mapped_size = PAGE_SIZE;
1898 p = mapped_buffer;
1899 }
1900
1901 crypto_shash_final(shash, calculated_csum);
1902 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1903 sblock->checksum_error = 1;
1904
1905 return sblock->header_error || sblock->checksum_error;
1906}
1907
1908static int scrub_checksum_super(struct scrub_block *sblock)
1909{
1910 struct btrfs_super_block *s;
1911 struct scrub_ctx *sctx = sblock->sctx;
1912 struct btrfs_fs_info *fs_info = sctx->fs_info;
1913 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1914 u8 calculated_csum[BTRFS_CSUM_SIZE];
1915 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1916 struct page *page;
1917 void *mapped_buffer;
1918 u64 mapped_size;
1919 void *p;
1920 int fail_gen = 0;
1921 int fail_cor = 0;
1922 u64 len;
1923 int index;
1924
1925 shash->tfm = fs_info->csum_shash;
1926 crypto_shash_init(shash);
1927
1928 BUG_ON(sblock->page_count < 1);
1929 page = sblock->pagev[0]->page;
1930 mapped_buffer = kmap_atomic(page);
1931 s = (struct btrfs_super_block *)mapped_buffer;
1932 memcpy(on_disk_csum, s->csum, sctx->csum_size);
1933
1934 if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1935 ++fail_cor;
1936
1937 if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1938 ++fail_gen;
1939
1940 if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
1941 ++fail_cor;
1942
1943 len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1944 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1945 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1946 index = 0;
1947 for (;;) {
1948 u64 l = min_t(u64, len, mapped_size);
1949
1950 crypto_shash_update(shash, p, l);
1951 kunmap_atomic(mapped_buffer);
1952 len -= l;
1953 if (len == 0)
1954 break;
1955 index++;
1956 BUG_ON(index >= sblock->page_count);
1957 BUG_ON(!sblock->pagev[index]->page);
1958 page = sblock->pagev[index]->page;
1959 mapped_buffer = kmap_atomic(page);
1960 mapped_size = PAGE_SIZE;
1961 p = mapped_buffer;
1962 }
1963
1964 crypto_shash_final(shash, calculated_csum);
1965 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1966 ++fail_cor;
1967
1968 if (fail_cor + fail_gen) {
1969 /*
1970 * if we find an error in a super block, we just report it.
1971 * They will get written with the next transaction commit
1972 * anyway
1973 */
1974 spin_lock(&sctx->stat_lock);
1975 ++sctx->stat.super_errors;
1976 spin_unlock(&sctx->stat_lock);
1977 if (fail_cor)
1978 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1979 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1980 else
1981 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1982 BTRFS_DEV_STAT_GENERATION_ERRS);
1983 }
1984
1985 return fail_cor + fail_gen;
1986}
1987
1988static void scrub_block_get(struct scrub_block *sblock)
1989{
1990 refcount_inc(&sblock->refs);
1991}
1992
1993static void scrub_block_put(struct scrub_block *sblock)
1994{
1995 if (refcount_dec_and_test(&sblock->refs)) {
1996 int i;
1997
1998 if (sblock->sparity)
1999 scrub_parity_put(sblock->sparity);
2000
2001 for (i = 0; i < sblock->page_count; i++)
2002 scrub_page_put(sblock->pagev[i]);
2003 kfree(sblock);
2004 }
2005}
2006
2007static void scrub_page_get(struct scrub_page *spage)
2008{
2009 atomic_inc(&spage->refs);
2010}
2011
2012static void scrub_page_put(struct scrub_page *spage)
2013{
2014 if (atomic_dec_and_test(&spage->refs)) {
2015 if (spage->page)
2016 __free_page(spage->page);
2017 kfree(spage);
2018 }
2019}
2020
2021static void scrub_submit(struct scrub_ctx *sctx)
2022{
2023 struct scrub_bio *sbio;
2024
2025 if (sctx->curr == -1)
2026 return;
2027
2028 sbio = sctx->bios[sctx->curr];
2029 sctx->curr = -1;
2030 scrub_pending_bio_inc(sctx);
2031 btrfsic_submit_bio(sbio->bio);
2032}
2033
2034static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2035 struct scrub_page *spage)
2036{
2037 struct scrub_block *sblock = spage->sblock;
2038 struct scrub_bio *sbio;
2039 int ret;
2040
2041again:
2042 /*
2043 * grab a fresh bio or wait for one to become available
2044 */
2045 while (sctx->curr == -1) {
2046 spin_lock(&sctx->list_lock);
2047 sctx->curr = sctx->first_free;
2048 if (sctx->curr != -1) {
2049 sctx->first_free = sctx->bios[sctx->curr]->next_free;
2050 sctx->bios[sctx->curr]->next_free = -1;
2051 sctx->bios[sctx->curr]->page_count = 0;
2052 spin_unlock(&sctx->list_lock);
2053 } else {
2054 spin_unlock(&sctx->list_lock);
2055 wait_event(sctx->list_wait, sctx->first_free != -1);
2056 }
2057 }
2058 sbio = sctx->bios[sctx->curr];
2059 if (sbio->page_count == 0) {
2060 struct bio *bio;
2061
2062 sbio->physical = spage->physical;
2063 sbio->logical = spage->logical;
2064 sbio->dev = spage->dev;
2065 bio = sbio->bio;
2066 if (!bio) {
2067 bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
2068 sbio->bio = bio;
2069 }
2070
2071 bio->bi_private = sbio;
2072 bio->bi_end_io = scrub_bio_end_io;
2073 bio_set_dev(bio, sbio->dev->bdev);
2074 bio->bi_iter.bi_sector = sbio->physical >> 9;
2075 bio->bi_opf = REQ_OP_READ;
2076 sbio->status = 0;
2077 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2078 spage->physical ||
2079 sbio->logical + sbio->page_count * PAGE_SIZE !=
2080 spage->logical ||
2081 sbio->dev != spage->dev) {
2082 scrub_submit(sctx);
2083 goto again;
2084 }
2085
2086 sbio->pagev[sbio->page_count] = spage;
2087 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2088 if (ret != PAGE_SIZE) {
2089 if (sbio->page_count < 1) {
2090 bio_put(sbio->bio);
2091 sbio->bio = NULL;
2092 return -EIO;
2093 }
2094 scrub_submit(sctx);
2095 goto again;
2096 }
2097
2098 scrub_block_get(sblock); /* one for the page added to the bio */
2099 atomic_inc(&sblock->outstanding_pages);
2100 sbio->page_count++;
2101 if (sbio->page_count == sctx->pages_per_rd_bio)
2102 scrub_submit(sctx);
2103
2104 return 0;
2105}
2106
2107static void scrub_missing_raid56_end_io(struct bio *bio)
2108{
2109 struct scrub_block *sblock = bio->bi_private;
2110 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2111
2112 if (bio->bi_status)
2113 sblock->no_io_error_seen = 0;
2114
2115 bio_put(bio);
2116
2117 btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2118}
2119
2120static void scrub_missing_raid56_worker(struct btrfs_work *work)
2121{
2122 struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2123 struct scrub_ctx *sctx = sblock->sctx;
2124 struct btrfs_fs_info *fs_info = sctx->fs_info;
2125 u64 logical;
2126 struct btrfs_device *dev;
2127
2128 logical = sblock->pagev[0]->logical;
2129 dev = sblock->pagev[0]->dev;
2130
2131 if (sblock->no_io_error_seen)
2132 scrub_recheck_block_checksum(sblock);
2133
2134 if (!sblock->no_io_error_seen) {
2135 spin_lock(&sctx->stat_lock);
2136 sctx->stat.read_errors++;
2137 spin_unlock(&sctx->stat_lock);
2138 btrfs_err_rl_in_rcu(fs_info,
2139 "IO error rebuilding logical %llu for dev %s",
2140 logical, rcu_str_deref(dev->name));
2141 } else if (sblock->header_error || sblock->checksum_error) {
2142 spin_lock(&sctx->stat_lock);
2143 sctx->stat.uncorrectable_errors++;
2144 spin_unlock(&sctx->stat_lock);
2145 btrfs_err_rl_in_rcu(fs_info,
2146 "failed to rebuild valid logical %llu for dev %s",
2147 logical, rcu_str_deref(dev->name));
2148 } else {
2149 scrub_write_block_to_dev_replace(sblock);
2150 }
2151
2152 scrub_block_put(sblock);
2153
2154 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2155 mutex_lock(&sctx->wr_lock);
2156 scrub_wr_submit(sctx);
2157 mutex_unlock(&sctx->wr_lock);
2158 }
2159
2160 scrub_pending_bio_dec(sctx);
2161}
2162
2163static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2164{
2165 struct scrub_ctx *sctx = sblock->sctx;
2166 struct btrfs_fs_info *fs_info = sctx->fs_info;
2167 u64 length = sblock->page_count * PAGE_SIZE;
2168 u64 logical = sblock->pagev[0]->logical;
2169 struct btrfs_bio *bbio = NULL;
2170 struct bio *bio;
2171 struct btrfs_raid_bio *rbio;
2172 int ret;
2173 int i;
2174
2175 btrfs_bio_counter_inc_blocked(fs_info);
2176 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2177 &length, &bbio);
2178 if (ret || !bbio || !bbio->raid_map)
2179 goto bbio_out;
2180
2181 if (WARN_ON(!sctx->is_dev_replace ||
2182 !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2183 /*
2184 * We shouldn't be scrubbing a missing device. Even for dev
2185 * replace, we should only get here for RAID 5/6. We either
2186 * managed to mount something with no mirrors remaining or
2187 * there's a bug in scrub_remap_extent()/btrfs_map_block().
2188 */
2189 goto bbio_out;
2190 }
2191
2192 bio = btrfs_io_bio_alloc(0);
2193 bio->bi_iter.bi_sector = logical >> 9;
2194 bio->bi_private = sblock;
2195 bio->bi_end_io = scrub_missing_raid56_end_io;
2196
2197 rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2198 if (!rbio)
2199 goto rbio_out;
2200
2201 for (i = 0; i < sblock->page_count; i++) {
2202 struct scrub_page *spage = sblock->pagev[i];
2203
2204 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2205 }
2206
2207 btrfs_init_work(&sblock->work, btrfs_scrub_helper,
2208 scrub_missing_raid56_worker, NULL, NULL);
2209 scrub_block_get(sblock);
2210 scrub_pending_bio_inc(sctx);
2211 raid56_submit_missing_rbio(rbio);
2212 return;
2213
2214rbio_out:
2215 bio_put(bio);
2216bbio_out:
2217 btrfs_bio_counter_dec(fs_info);
2218 btrfs_put_bbio(bbio);
2219 spin_lock(&sctx->stat_lock);
2220 sctx->stat.malloc_errors++;
2221 spin_unlock(&sctx->stat_lock);
2222}
2223
2224static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2225 u64 physical, struct btrfs_device *dev, u64 flags,
2226 u64 gen, int mirror_num, u8 *csum, int force,
2227 u64 physical_for_dev_replace)
2228{
2229 struct scrub_block *sblock;
2230 int index;
2231
2232 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2233 if (!sblock) {
2234 spin_lock(&sctx->stat_lock);
2235 sctx->stat.malloc_errors++;
2236 spin_unlock(&sctx->stat_lock);
2237 return -ENOMEM;
2238 }
2239
2240 /* one ref inside this function, plus one for each page added to
2241 * a bio later on */
2242 refcount_set(&sblock->refs, 1);
2243 sblock->sctx = sctx;
2244 sblock->no_io_error_seen = 1;
2245
2246 for (index = 0; len > 0; index++) {
2247 struct scrub_page *spage;
2248 u64 l = min_t(u64, len, PAGE_SIZE);
2249
2250 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2251 if (!spage) {
2252leave_nomem:
2253 spin_lock(&sctx->stat_lock);
2254 sctx->stat.malloc_errors++;
2255 spin_unlock(&sctx->stat_lock);
2256 scrub_block_put(sblock);
2257 return -ENOMEM;
2258 }
2259 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2260 scrub_page_get(spage);
2261 sblock->pagev[index] = spage;
2262 spage->sblock = sblock;
2263 spage->dev = dev;
2264 spage->flags = flags;
2265 spage->generation = gen;
2266 spage->logical = logical;
2267 spage->physical = physical;
2268 spage->physical_for_dev_replace = physical_for_dev_replace;
2269 spage->mirror_num = mirror_num;
2270 if (csum) {
2271 spage->have_csum = 1;
2272 memcpy(spage->csum, csum, sctx->csum_size);
2273 } else {
2274 spage->have_csum = 0;
2275 }
2276 sblock->page_count++;
2277 spage->page = alloc_page(GFP_KERNEL);
2278 if (!spage->page)
2279 goto leave_nomem;
2280 len -= l;
2281 logical += l;
2282 physical += l;
2283 physical_for_dev_replace += l;
2284 }
2285
2286 WARN_ON(sblock->page_count == 0);
2287 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2288 /*
2289 * This case should only be hit for RAID 5/6 device replace. See
2290 * the comment in scrub_missing_raid56_pages() for details.
2291 */
2292 scrub_missing_raid56_pages(sblock);
2293 } else {
2294 for (index = 0; index < sblock->page_count; index++) {
2295 struct scrub_page *spage = sblock->pagev[index];
2296 int ret;
2297
2298 ret = scrub_add_page_to_rd_bio(sctx, spage);
2299 if (ret) {
2300 scrub_block_put(sblock);
2301 return ret;
2302 }
2303 }
2304
2305 if (force)
2306 scrub_submit(sctx);
2307 }
2308
2309 /* last one frees, either here or in bio completion for last page */
2310 scrub_block_put(sblock);
2311 return 0;
2312}
2313
2314static void scrub_bio_end_io(struct bio *bio)
2315{
2316 struct scrub_bio *sbio = bio->bi_private;
2317 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2318
2319 sbio->status = bio->bi_status;
2320 sbio->bio = bio;
2321
2322 btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2323}
2324
2325static void scrub_bio_end_io_worker(struct btrfs_work *work)
2326{
2327 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2328 struct scrub_ctx *sctx = sbio->sctx;
2329 int i;
2330
2331 BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2332 if (sbio->status) {
2333 for (i = 0; i < sbio->page_count; i++) {
2334 struct scrub_page *spage = sbio->pagev[i];
2335
2336 spage->io_error = 1;
2337 spage->sblock->no_io_error_seen = 0;
2338 }
2339 }
2340
2341 /* now complete the scrub_block items that have all pages completed */
2342 for (i = 0; i < sbio->page_count; i++) {
2343 struct scrub_page *spage = sbio->pagev[i];
2344 struct scrub_block *sblock = spage->sblock;
2345
2346 if (atomic_dec_and_test(&sblock->outstanding_pages))
2347 scrub_block_complete(sblock);
2348 scrub_block_put(sblock);
2349 }
2350
2351 bio_put(sbio->bio);
2352 sbio->bio = NULL;
2353 spin_lock(&sctx->list_lock);
2354 sbio->next_free = sctx->first_free;
2355 sctx->first_free = sbio->index;
2356 spin_unlock(&sctx->list_lock);
2357
2358 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2359 mutex_lock(&sctx->wr_lock);
2360 scrub_wr_submit(sctx);
2361 mutex_unlock(&sctx->wr_lock);
2362 }
2363
2364 scrub_pending_bio_dec(sctx);
2365}
2366
2367static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2368 unsigned long *bitmap,
2369 u64 start, u64 len)
2370{
2371 u64 offset;
2372 u64 nsectors64;
2373 u32 nsectors;
2374 int sectorsize = sparity->sctx->fs_info->sectorsize;
2375
2376 if (len >= sparity->stripe_len) {
2377 bitmap_set(bitmap, 0, sparity->nsectors);
2378 return;
2379 }
2380
2381 start -= sparity->logic_start;
2382 start = div64_u64_rem(start, sparity->stripe_len, &offset);
2383 offset = div_u64(offset, sectorsize);
2384 nsectors64 = div_u64(len, sectorsize);
2385
2386 ASSERT(nsectors64 < UINT_MAX);
2387 nsectors = (u32)nsectors64;
2388
2389 if (offset + nsectors <= sparity->nsectors) {
2390 bitmap_set(bitmap, offset, nsectors);
2391 return;
2392 }
2393
2394 bitmap_set(bitmap, offset, sparity->nsectors - offset);
2395 bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2396}
2397
2398static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2399 u64 start, u64 len)
2400{
2401 __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2402}
2403
2404static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2405 u64 start, u64 len)
2406{
2407 __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2408}
2409
2410static void scrub_block_complete(struct scrub_block *sblock)
2411{
2412 int corrupted = 0;
2413
2414 if (!sblock->no_io_error_seen) {
2415 corrupted = 1;
2416 scrub_handle_errored_block(sblock);
2417 } else {
2418 /*
2419 * if has checksum error, write via repair mechanism in
2420 * dev replace case, otherwise write here in dev replace
2421 * case.
2422 */
2423 corrupted = scrub_checksum(sblock);
2424 if (!corrupted && sblock->sctx->is_dev_replace)
2425 scrub_write_block_to_dev_replace(sblock);
2426 }
2427
2428 if (sblock->sparity && corrupted && !sblock->data_corrected) {
2429 u64 start = sblock->pagev[0]->logical;
2430 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2431 PAGE_SIZE;
2432
2433 scrub_parity_mark_sectors_error(sblock->sparity,
2434 start, end - start);
2435 }
2436}
2437
2438static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2439{
2440 struct btrfs_ordered_sum *sum = NULL;
2441 unsigned long index;
2442 unsigned long num_sectors;
2443
2444 while (!list_empty(&sctx->csum_list)) {
2445 sum = list_first_entry(&sctx->csum_list,
2446 struct btrfs_ordered_sum, list);
2447 if (sum->bytenr > logical)
2448 return 0;
2449 if (sum->bytenr + sum->len > logical)
2450 break;
2451
2452 ++sctx->stat.csum_discards;
2453 list_del(&sum->list);
2454 kfree(sum);
2455 sum = NULL;
2456 }
2457 if (!sum)
2458 return 0;
2459
2460 index = div_u64(logical - sum->bytenr, sctx->fs_info->sectorsize);
2461 ASSERT(index < UINT_MAX);
2462
2463 num_sectors = sum->len / sctx->fs_info->sectorsize;
2464 memcpy(csum, sum->sums + index * sctx->csum_size, sctx->csum_size);
2465 if (index == num_sectors - 1) {
2466 list_del(&sum->list);
2467 kfree(sum);
2468 }
2469 return 1;
2470}
2471
2472/* scrub extent tries to collect up to 64 kB for each bio */
2473static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2474 u64 logical, u64 len,
2475 u64 physical, struct btrfs_device *dev, u64 flags,
2476 u64 gen, int mirror_num, u64 physical_for_dev_replace)
2477{
2478 int ret;
2479 u8 csum[BTRFS_CSUM_SIZE];
2480 u32 blocksize;
2481
2482 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2483 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2484 blocksize = map->stripe_len;
2485 else
2486 blocksize = sctx->fs_info->sectorsize;
2487 spin_lock(&sctx->stat_lock);
2488 sctx->stat.data_extents_scrubbed++;
2489 sctx->stat.data_bytes_scrubbed += len;
2490 spin_unlock(&sctx->stat_lock);
2491 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2492 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2493 blocksize = map->stripe_len;
2494 else
2495 blocksize = sctx->fs_info->nodesize;
2496 spin_lock(&sctx->stat_lock);
2497 sctx->stat.tree_extents_scrubbed++;
2498 sctx->stat.tree_bytes_scrubbed += len;
2499 spin_unlock(&sctx->stat_lock);
2500 } else {
2501 blocksize = sctx->fs_info->sectorsize;
2502 WARN_ON(1);
2503 }
2504
2505 while (len) {
2506 u64 l = min_t(u64, len, blocksize);
2507 int have_csum = 0;
2508
2509 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2510 /* push csums to sbio */
2511 have_csum = scrub_find_csum(sctx, logical, csum);
2512 if (have_csum == 0)
2513 ++sctx->stat.no_csum;
2514 }
2515 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2516 mirror_num, have_csum ? csum : NULL, 0,
2517 physical_for_dev_replace);
2518 if (ret)
2519 return ret;
2520 len -= l;
2521 logical += l;
2522 physical += l;
2523 physical_for_dev_replace += l;
2524 }
2525 return 0;
2526}
2527
2528static int scrub_pages_for_parity(struct scrub_parity *sparity,
2529 u64 logical, u64 len,
2530 u64 physical, struct btrfs_device *dev,
2531 u64 flags, u64 gen, int mirror_num, u8 *csum)
2532{
2533 struct scrub_ctx *sctx = sparity->sctx;
2534 struct scrub_block *sblock;
2535 int index;
2536
2537 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2538 if (!sblock) {
2539 spin_lock(&sctx->stat_lock);
2540 sctx->stat.malloc_errors++;
2541 spin_unlock(&sctx->stat_lock);
2542 return -ENOMEM;
2543 }
2544
2545 /* one ref inside this function, plus one for each page added to
2546 * a bio later on */
2547 refcount_set(&sblock->refs, 1);
2548 sblock->sctx = sctx;
2549 sblock->no_io_error_seen = 1;
2550 sblock->sparity = sparity;
2551 scrub_parity_get(sparity);
2552
2553 for (index = 0; len > 0; index++) {
2554 struct scrub_page *spage;
2555 u64 l = min_t(u64, len, PAGE_SIZE);
2556
2557 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2558 if (!spage) {
2559leave_nomem:
2560 spin_lock(&sctx->stat_lock);
2561 sctx->stat.malloc_errors++;
2562 spin_unlock(&sctx->stat_lock);
2563 scrub_block_put(sblock);
2564 return -ENOMEM;
2565 }
2566 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2567 /* For scrub block */
2568 scrub_page_get(spage);
2569 sblock->pagev[index] = spage;
2570 /* For scrub parity */
2571 scrub_page_get(spage);
2572 list_add_tail(&spage->list, &sparity->spages);
2573 spage->sblock = sblock;
2574 spage->dev = dev;
2575 spage->flags = flags;
2576 spage->generation = gen;
2577 spage->logical = logical;
2578 spage->physical = physical;
2579 spage->mirror_num = mirror_num;
2580 if (csum) {
2581 spage->have_csum = 1;
2582 memcpy(spage->csum, csum, sctx->csum_size);
2583 } else {
2584 spage->have_csum = 0;
2585 }
2586 sblock->page_count++;
2587 spage->page = alloc_page(GFP_KERNEL);
2588 if (!spage->page)
2589 goto leave_nomem;
2590 len -= l;
2591 logical += l;
2592 physical += l;
2593 }
2594
2595 WARN_ON(sblock->page_count == 0);
2596 for (index = 0; index < sblock->page_count; index++) {
2597 struct scrub_page *spage = sblock->pagev[index];
2598 int ret;
2599
2600 ret = scrub_add_page_to_rd_bio(sctx, spage);
2601 if (ret) {
2602 scrub_block_put(sblock);
2603 return ret;
2604 }
2605 }
2606
2607 /* last one frees, either here or in bio completion for last page */
2608 scrub_block_put(sblock);
2609 return 0;
2610}
2611
2612static int scrub_extent_for_parity(struct scrub_parity *sparity,
2613 u64 logical, u64 len,
2614 u64 physical, struct btrfs_device *dev,
2615 u64 flags, u64 gen, int mirror_num)
2616{
2617 struct scrub_ctx *sctx = sparity->sctx;
2618 int ret;
2619 u8 csum[BTRFS_CSUM_SIZE];
2620 u32 blocksize;
2621
2622 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2623 scrub_parity_mark_sectors_error(sparity, logical, len);
2624 return 0;
2625 }
2626
2627 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2628 blocksize = sparity->stripe_len;
2629 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2630 blocksize = sparity->stripe_len;
2631 } else {
2632 blocksize = sctx->fs_info->sectorsize;
2633 WARN_ON(1);
2634 }
2635
2636 while (len) {
2637 u64 l = min_t(u64, len, blocksize);
2638 int have_csum = 0;
2639
2640 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2641 /* push csums to sbio */
2642 have_csum = scrub_find_csum(sctx, logical, csum);
2643 if (have_csum == 0)
2644 goto skip;
2645 }
2646 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2647 flags, gen, mirror_num,
2648 have_csum ? csum : NULL);
2649 if (ret)
2650 return ret;
2651skip:
2652 len -= l;
2653 logical += l;
2654 physical += l;
2655 }
2656 return 0;
2657}
2658
2659/*
2660 * Given a physical address, this will calculate it's
2661 * logical offset. if this is a parity stripe, it will return
2662 * the most left data stripe's logical offset.
2663 *
2664 * return 0 if it is a data stripe, 1 means parity stripe.
2665 */
2666static int get_raid56_logic_offset(u64 physical, int num,
2667 struct map_lookup *map, u64 *offset,
2668 u64 *stripe_start)
2669{
2670 int i;
2671 int j = 0;
2672 u64 stripe_nr;
2673 u64 last_offset;
2674 u32 stripe_index;
2675 u32 rot;
2676 const int data_stripes = nr_data_stripes(map);
2677
2678 last_offset = (physical - map->stripes[num].physical) * data_stripes;
2679 if (stripe_start)
2680 *stripe_start = last_offset;
2681
2682 *offset = last_offset;
2683 for (i = 0; i < data_stripes; i++) {
2684 *offset = last_offset + i * map->stripe_len;
2685
2686 stripe_nr = div64_u64(*offset, map->stripe_len);
2687 stripe_nr = div_u64(stripe_nr, data_stripes);
2688
2689 /* Work out the disk rotation on this stripe-set */
2690 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2691 /* calculate which stripe this data locates */
2692 rot += i;
2693 stripe_index = rot % map->num_stripes;
2694 if (stripe_index == num)
2695 return 0;
2696 if (stripe_index < num)
2697 j++;
2698 }
2699 *offset = last_offset + j * map->stripe_len;
2700 return 1;
2701}
2702
2703static void scrub_free_parity(struct scrub_parity *sparity)
2704{
2705 struct scrub_ctx *sctx = sparity->sctx;
2706 struct scrub_page *curr, *next;
2707 int nbits;
2708
2709 nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2710 if (nbits) {
2711 spin_lock(&sctx->stat_lock);
2712 sctx->stat.read_errors += nbits;
2713 sctx->stat.uncorrectable_errors += nbits;
2714 spin_unlock(&sctx->stat_lock);
2715 }
2716
2717 list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2718 list_del_init(&curr->list);
2719 scrub_page_put(curr);
2720 }
2721
2722 kfree(sparity);
2723}
2724
2725static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2726{
2727 struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2728 work);
2729 struct scrub_ctx *sctx = sparity->sctx;
2730
2731 scrub_free_parity(sparity);
2732 scrub_pending_bio_dec(sctx);
2733}
2734
2735static void scrub_parity_bio_endio(struct bio *bio)
2736{
2737 struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
2738 struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
2739
2740 if (bio->bi_status)
2741 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2742 sparity->nsectors);
2743
2744 bio_put(bio);
2745
2746 btrfs_init_work(&sparity->work, btrfs_scrubparity_helper,
2747 scrub_parity_bio_endio_worker, NULL, NULL);
2748 btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
2749}
2750
2751static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2752{
2753 struct scrub_ctx *sctx = sparity->sctx;
2754 struct btrfs_fs_info *fs_info = sctx->fs_info;
2755 struct bio *bio;
2756 struct btrfs_raid_bio *rbio;
2757 struct btrfs_bio *bbio = NULL;
2758 u64 length;
2759 int ret;
2760
2761 if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
2762 sparity->nsectors))
2763 goto out;
2764
2765 length = sparity->logic_end - sparity->logic_start;
2766
2767 btrfs_bio_counter_inc_blocked(fs_info);
2768 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
2769 &length, &bbio);
2770 if (ret || !bbio || !bbio->raid_map)
2771 goto bbio_out;
2772
2773 bio = btrfs_io_bio_alloc(0);
2774 bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2775 bio->bi_private = sparity;
2776 bio->bi_end_io = scrub_parity_bio_endio;
2777
2778 rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
2779 length, sparity->scrub_dev,
2780 sparity->dbitmap,
2781 sparity->nsectors);
2782 if (!rbio)
2783 goto rbio_out;
2784
2785 scrub_pending_bio_inc(sctx);
2786 raid56_parity_submit_scrub_rbio(rbio);
2787 return;
2788
2789rbio_out:
2790 bio_put(bio);
2791bbio_out:
2792 btrfs_bio_counter_dec(fs_info);
2793 btrfs_put_bbio(bbio);
2794 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2795 sparity->nsectors);
2796 spin_lock(&sctx->stat_lock);
2797 sctx->stat.malloc_errors++;
2798 spin_unlock(&sctx->stat_lock);
2799out:
2800 scrub_free_parity(sparity);
2801}
2802
2803static inline int scrub_calc_parity_bitmap_len(int nsectors)
2804{
2805 return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
2806}
2807
2808static void scrub_parity_get(struct scrub_parity *sparity)
2809{
2810 refcount_inc(&sparity->refs);
2811}
2812
2813static void scrub_parity_put(struct scrub_parity *sparity)
2814{
2815 if (!refcount_dec_and_test(&sparity->refs))
2816 return;
2817
2818 scrub_parity_check_and_repair(sparity);
2819}
2820
2821static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
2822 struct map_lookup *map,
2823 struct btrfs_device *sdev,
2824 struct btrfs_path *path,
2825 u64 logic_start,
2826 u64 logic_end)
2827{
2828 struct btrfs_fs_info *fs_info = sctx->fs_info;
2829 struct btrfs_root *root = fs_info->extent_root;
2830 struct btrfs_root *csum_root = fs_info->csum_root;
2831 struct btrfs_extent_item *extent;
2832 struct btrfs_bio *bbio = NULL;
2833 u64 flags;
2834 int ret;
2835 int slot;
2836 struct extent_buffer *l;
2837 struct btrfs_key key;
2838 u64 generation;
2839 u64 extent_logical;
2840 u64 extent_physical;
2841 u64 extent_len;
2842 u64 mapped_length;
2843 struct btrfs_device *extent_dev;
2844 struct scrub_parity *sparity;
2845 int nsectors;
2846 int bitmap_len;
2847 int extent_mirror_num;
2848 int stop_loop = 0;
2849
2850 nsectors = div_u64(map->stripe_len, fs_info->sectorsize);
2851 bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
2852 sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
2853 GFP_NOFS);
2854 if (!sparity) {
2855 spin_lock(&sctx->stat_lock);
2856 sctx->stat.malloc_errors++;
2857 spin_unlock(&sctx->stat_lock);
2858 return -ENOMEM;
2859 }
2860
2861 sparity->stripe_len = map->stripe_len;
2862 sparity->nsectors = nsectors;
2863 sparity->sctx = sctx;
2864 sparity->scrub_dev = sdev;
2865 sparity->logic_start = logic_start;
2866 sparity->logic_end = logic_end;
2867 refcount_set(&sparity->refs, 1);
2868 INIT_LIST_HEAD(&sparity->spages);
2869 sparity->dbitmap = sparity->bitmap;
2870 sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
2871
2872 ret = 0;
2873 while (logic_start < logic_end) {
2874 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2875 key.type = BTRFS_METADATA_ITEM_KEY;
2876 else
2877 key.type = BTRFS_EXTENT_ITEM_KEY;
2878 key.objectid = logic_start;
2879 key.offset = (u64)-1;
2880
2881 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2882 if (ret < 0)
2883 goto out;
2884
2885 if (ret > 0) {
2886 ret = btrfs_previous_extent_item(root, path, 0);
2887 if (ret < 0)
2888 goto out;
2889 if (ret > 0) {
2890 btrfs_release_path(path);
2891 ret = btrfs_search_slot(NULL, root, &key,
2892 path, 0, 0);
2893 if (ret < 0)
2894 goto out;
2895 }
2896 }
2897
2898 stop_loop = 0;
2899 while (1) {
2900 u64 bytes;
2901
2902 l = path->nodes[0];
2903 slot = path->slots[0];
2904 if (slot >= btrfs_header_nritems(l)) {
2905 ret = btrfs_next_leaf(root, path);
2906 if (ret == 0)
2907 continue;
2908 if (ret < 0)
2909 goto out;
2910
2911 stop_loop = 1;
2912 break;
2913 }
2914 btrfs_item_key_to_cpu(l, &key, slot);
2915
2916 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2917 key.type != BTRFS_METADATA_ITEM_KEY)
2918 goto next;
2919
2920 if (key.type == BTRFS_METADATA_ITEM_KEY)
2921 bytes = fs_info->nodesize;
2922 else
2923 bytes = key.offset;
2924
2925 if (key.objectid + bytes <= logic_start)
2926 goto next;
2927
2928 if (key.objectid >= logic_end) {
2929 stop_loop = 1;
2930 break;
2931 }
2932
2933 while (key.objectid >= logic_start + map->stripe_len)
2934 logic_start += map->stripe_len;
2935
2936 extent = btrfs_item_ptr(l, slot,
2937 struct btrfs_extent_item);
2938 flags = btrfs_extent_flags(l, extent);
2939 generation = btrfs_extent_generation(l, extent);
2940
2941 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
2942 (key.objectid < logic_start ||
2943 key.objectid + bytes >
2944 logic_start + map->stripe_len)) {
2945 btrfs_err(fs_info,
2946 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
2947 key.objectid, logic_start);
2948 spin_lock(&sctx->stat_lock);
2949 sctx->stat.uncorrectable_errors++;
2950 spin_unlock(&sctx->stat_lock);
2951 goto next;
2952 }
2953again:
2954 extent_logical = key.objectid;
2955 extent_len = bytes;
2956
2957 if (extent_logical < logic_start) {
2958 extent_len -= logic_start - extent_logical;
2959 extent_logical = logic_start;
2960 }
2961
2962 if (extent_logical + extent_len >
2963 logic_start + map->stripe_len)
2964 extent_len = logic_start + map->stripe_len -
2965 extent_logical;
2966
2967 scrub_parity_mark_sectors_data(sparity, extent_logical,
2968 extent_len);
2969
2970 mapped_length = extent_len;
2971 bbio = NULL;
2972 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
2973 extent_logical, &mapped_length, &bbio,
2974 0);
2975 if (!ret) {
2976 if (!bbio || mapped_length < extent_len)
2977 ret = -EIO;
2978 }
2979 if (ret) {
2980 btrfs_put_bbio(bbio);
2981 goto out;
2982 }
2983 extent_physical = bbio->stripes[0].physical;
2984 extent_mirror_num = bbio->mirror_num;
2985 extent_dev = bbio->stripes[0].dev;
2986 btrfs_put_bbio(bbio);
2987
2988 ret = btrfs_lookup_csums_range(csum_root,
2989 extent_logical,
2990 extent_logical + extent_len - 1,
2991 &sctx->csum_list, 1);
2992 if (ret)
2993 goto out;
2994
2995 ret = scrub_extent_for_parity(sparity, extent_logical,
2996 extent_len,
2997 extent_physical,
2998 extent_dev, flags,
2999 generation,
3000 extent_mirror_num);
3001
3002 scrub_free_csums(sctx);
3003
3004 if (ret)
3005 goto out;
3006
3007 if (extent_logical + extent_len <
3008 key.objectid + bytes) {
3009 logic_start += map->stripe_len;
3010
3011 if (logic_start >= logic_end) {
3012 stop_loop = 1;
3013 break;
3014 }
3015
3016 if (logic_start < key.objectid + bytes) {
3017 cond_resched();
3018 goto again;
3019 }
3020 }
3021next:
3022 path->slots[0]++;
3023 }
3024
3025 btrfs_release_path(path);
3026
3027 if (stop_loop)
3028 break;
3029
3030 logic_start += map->stripe_len;
3031 }
3032out:
3033 if (ret < 0)
3034 scrub_parity_mark_sectors_error(sparity, logic_start,
3035 logic_end - logic_start);
3036 scrub_parity_put(sparity);
3037 scrub_submit(sctx);
3038 mutex_lock(&sctx->wr_lock);
3039 scrub_wr_submit(sctx);
3040 mutex_unlock(&sctx->wr_lock);
3041
3042 btrfs_release_path(path);
3043 return ret < 0 ? ret : 0;
3044}
3045
3046static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3047 struct map_lookup *map,
3048 struct btrfs_device *scrub_dev,
3049 int num, u64 base, u64 length)
3050{
3051 struct btrfs_path *path, *ppath;
3052 struct btrfs_fs_info *fs_info = sctx->fs_info;
3053 struct btrfs_root *root = fs_info->extent_root;
3054 struct btrfs_root *csum_root = fs_info->csum_root;
3055 struct btrfs_extent_item *extent;
3056 struct blk_plug plug;
3057 u64 flags;
3058 int ret;
3059 int slot;
3060 u64 nstripes;
3061 struct extent_buffer *l;
3062 u64 physical;
3063 u64 logical;
3064 u64 logic_end;
3065 u64 physical_end;
3066 u64 generation;
3067 int mirror_num;
3068 struct reada_control *reada1;
3069 struct reada_control *reada2;
3070 struct btrfs_key key;
3071 struct btrfs_key key_end;
3072 u64 increment = map->stripe_len;
3073 u64 offset;
3074 u64 extent_logical;
3075 u64 extent_physical;
3076 u64 extent_len;
3077 u64 stripe_logical;
3078 u64 stripe_end;
3079 struct btrfs_device *extent_dev;
3080 int extent_mirror_num;
3081 int stop_loop = 0;
3082
3083 physical = map->stripes[num].physical;
3084 offset = 0;
3085 nstripes = div64_u64(length, map->stripe_len);
3086 if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3087 offset = map->stripe_len * num;
3088 increment = map->stripe_len * map->num_stripes;
3089 mirror_num = 1;
3090 } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3091 int factor = map->num_stripes / map->sub_stripes;
3092 offset = map->stripe_len * (num / map->sub_stripes);
3093 increment = map->stripe_len * factor;
3094 mirror_num = num % map->sub_stripes + 1;
3095 } else if (map->type & BTRFS_BLOCK_GROUP_RAID1_MASK) {
3096 increment = map->stripe_len;
3097 mirror_num = num % map->num_stripes + 1;
3098 } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3099 increment = map->stripe_len;
3100 mirror_num = num % map->num_stripes + 1;
3101 } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3102 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3103 increment = map->stripe_len * nr_data_stripes(map);
3104 mirror_num = 1;
3105 } else {
3106 increment = map->stripe_len;
3107 mirror_num = 1;
3108 }
3109
3110 path = btrfs_alloc_path();
3111 if (!path)
3112 return -ENOMEM;
3113
3114 ppath = btrfs_alloc_path();
3115 if (!ppath) {
3116 btrfs_free_path(path);
3117 return -ENOMEM;
3118 }
3119
3120 /*
3121 * work on commit root. The related disk blocks are static as
3122 * long as COW is applied. This means, it is save to rewrite
3123 * them to repair disk errors without any race conditions
3124 */
3125 path->search_commit_root = 1;
3126 path->skip_locking = 1;
3127
3128 ppath->search_commit_root = 1;
3129 ppath->skip_locking = 1;
3130 /*
3131 * trigger the readahead for extent tree csum tree and wait for
3132 * completion. During readahead, the scrub is officially paused
3133 * to not hold off transaction commits
3134 */
3135 logical = base + offset;
3136 physical_end = physical + nstripes * map->stripe_len;
3137 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3138 get_raid56_logic_offset(physical_end, num,
3139 map, &logic_end, NULL);
3140 logic_end += base;
3141 } else {
3142 logic_end = logical + increment * nstripes;
3143 }
3144 wait_event(sctx->list_wait,
3145 atomic_read(&sctx->bios_in_flight) == 0);
3146 scrub_blocked_if_needed(fs_info);
3147
3148 /* FIXME it might be better to start readahead at commit root */
3149 key.objectid = logical;
3150 key.type = BTRFS_EXTENT_ITEM_KEY;
3151 key.offset = (u64)0;
3152 key_end.objectid = logic_end;
3153 key_end.type = BTRFS_METADATA_ITEM_KEY;
3154 key_end.offset = (u64)-1;
3155 reada1 = btrfs_reada_add(root, &key, &key_end);
3156
3157 key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3158 key.type = BTRFS_EXTENT_CSUM_KEY;
3159 key.offset = logical;
3160 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3161 key_end.type = BTRFS_EXTENT_CSUM_KEY;
3162 key_end.offset = logic_end;
3163 reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3164
3165 if (!IS_ERR(reada1))
3166 btrfs_reada_wait(reada1);
3167 if (!IS_ERR(reada2))
3168 btrfs_reada_wait(reada2);
3169
3170
3171 /*
3172 * collect all data csums for the stripe to avoid seeking during
3173 * the scrub. This might currently (crc32) end up to be about 1MB
3174 */
3175 blk_start_plug(&plug);
3176
3177 /*
3178 * now find all extents for each stripe and scrub them
3179 */
3180 ret = 0;
3181 while (physical < physical_end) {
3182 /*
3183 * canceled?
3184 */
3185 if (atomic_read(&fs_info->scrub_cancel_req) ||
3186 atomic_read(&sctx->cancel_req)) {
3187 ret = -ECANCELED;
3188 goto out;
3189 }
3190 /*
3191 * check to see if we have to pause
3192 */
3193 if (atomic_read(&fs_info->scrub_pause_req)) {
3194 /* push queued extents */
3195 sctx->flush_all_writes = true;
3196 scrub_submit(sctx);
3197 mutex_lock(&sctx->wr_lock);
3198 scrub_wr_submit(sctx);
3199 mutex_unlock(&sctx->wr_lock);
3200 wait_event(sctx->list_wait,
3201 atomic_read(&sctx->bios_in_flight) == 0);
3202 sctx->flush_all_writes = false;
3203 scrub_blocked_if_needed(fs_info);
3204 }
3205
3206 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3207 ret = get_raid56_logic_offset(physical, num, map,
3208 &logical,
3209 &stripe_logical);
3210 logical += base;
3211 if (ret) {
3212 /* it is parity strip */
3213 stripe_logical += base;
3214 stripe_end = stripe_logical + increment;
3215 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3216 ppath, stripe_logical,
3217 stripe_end);
3218 if (ret)
3219 goto out;
3220 goto skip;
3221 }
3222 }
3223
3224 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3225 key.type = BTRFS_METADATA_ITEM_KEY;
3226 else
3227 key.type = BTRFS_EXTENT_ITEM_KEY;
3228 key.objectid = logical;
3229 key.offset = (u64)-1;
3230
3231 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3232 if (ret < 0)
3233 goto out;
3234
3235 if (ret > 0) {
3236 ret = btrfs_previous_extent_item(root, path, 0);
3237 if (ret < 0)
3238 goto out;
3239 if (ret > 0) {
3240 /* there's no smaller item, so stick with the
3241 * larger one */
3242 btrfs_release_path(path);
3243 ret = btrfs_search_slot(NULL, root, &key,
3244 path, 0, 0);
3245 if (ret < 0)
3246 goto out;
3247 }
3248 }
3249
3250 stop_loop = 0;
3251 while (1) {
3252 u64 bytes;
3253
3254 l = path->nodes[0];
3255 slot = path->slots[0];
3256 if (slot >= btrfs_header_nritems(l)) {
3257 ret = btrfs_next_leaf(root, path);
3258 if (ret == 0)
3259 continue;
3260 if (ret < 0)
3261 goto out;
3262
3263 stop_loop = 1;
3264 break;
3265 }
3266 btrfs_item_key_to_cpu(l, &key, slot);
3267
3268 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3269 key.type != BTRFS_METADATA_ITEM_KEY)
3270 goto next;
3271
3272 if (key.type == BTRFS_METADATA_ITEM_KEY)
3273 bytes = fs_info->nodesize;
3274 else
3275 bytes = key.offset;
3276
3277 if (key.objectid + bytes <= logical)
3278 goto next;
3279
3280 if (key.objectid >= logical + map->stripe_len) {
3281 /* out of this device extent */
3282 if (key.objectid >= logic_end)
3283 stop_loop = 1;
3284 break;
3285 }
3286
3287 extent = btrfs_item_ptr(l, slot,
3288 struct btrfs_extent_item);
3289 flags = btrfs_extent_flags(l, extent);
3290 generation = btrfs_extent_generation(l, extent);
3291
3292 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3293 (key.objectid < logical ||
3294 key.objectid + bytes >
3295 logical + map->stripe_len)) {
3296 btrfs_err(fs_info,
3297 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3298 key.objectid, logical);
3299 spin_lock(&sctx->stat_lock);
3300 sctx->stat.uncorrectable_errors++;
3301 spin_unlock(&sctx->stat_lock);
3302 goto next;
3303 }
3304
3305again:
3306 extent_logical = key.objectid;
3307 extent_len = bytes;
3308
3309 /*
3310 * trim extent to this stripe
3311 */
3312 if (extent_logical < logical) {
3313 extent_len -= logical - extent_logical;
3314 extent_logical = logical;
3315 }
3316 if (extent_logical + extent_len >
3317 logical + map->stripe_len) {
3318 extent_len = logical + map->stripe_len -
3319 extent_logical;
3320 }
3321
3322 extent_physical = extent_logical - logical + physical;
3323 extent_dev = scrub_dev;
3324 extent_mirror_num = mirror_num;
3325 if (sctx->is_dev_replace)
3326 scrub_remap_extent(fs_info, extent_logical,
3327 extent_len, &extent_physical,
3328 &extent_dev,
3329 &extent_mirror_num);
3330
3331 ret = btrfs_lookup_csums_range(csum_root,
3332 extent_logical,
3333 extent_logical +
3334 extent_len - 1,
3335 &sctx->csum_list, 1);
3336 if (ret)
3337 goto out;
3338
3339 ret = scrub_extent(sctx, map, extent_logical, extent_len,
3340 extent_physical, extent_dev, flags,
3341 generation, extent_mirror_num,
3342 extent_logical - logical + physical);
3343
3344 scrub_free_csums(sctx);
3345
3346 if (ret)
3347 goto out;
3348
3349 if (extent_logical + extent_len <
3350 key.objectid + bytes) {
3351 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3352 /*
3353 * loop until we find next data stripe
3354 * or we have finished all stripes.
3355 */
3356loop:
3357 physical += map->stripe_len;
3358 ret = get_raid56_logic_offset(physical,
3359 num, map, &logical,
3360 &stripe_logical);
3361 logical += base;
3362
3363 if (ret && physical < physical_end) {
3364 stripe_logical += base;
3365 stripe_end = stripe_logical +
3366 increment;
3367 ret = scrub_raid56_parity(sctx,
3368 map, scrub_dev, ppath,
3369 stripe_logical,
3370 stripe_end);
3371 if (ret)
3372 goto out;
3373 goto loop;
3374 }
3375 } else {
3376 physical += map->stripe_len;
3377 logical += increment;
3378 }
3379 if (logical < key.objectid + bytes) {
3380 cond_resched();
3381 goto again;
3382 }
3383
3384 if (physical >= physical_end) {
3385 stop_loop = 1;
3386 break;
3387 }
3388 }
3389next:
3390 path->slots[0]++;
3391 }
3392 btrfs_release_path(path);
3393skip:
3394 logical += increment;
3395 physical += map->stripe_len;
3396 spin_lock(&sctx->stat_lock);
3397 if (stop_loop)
3398 sctx->stat.last_physical = map->stripes[num].physical +
3399 length;
3400 else
3401 sctx->stat.last_physical = physical;
3402 spin_unlock(&sctx->stat_lock);
3403 if (stop_loop)
3404 break;
3405 }
3406out:
3407 /* push queued extents */
3408 scrub_submit(sctx);
3409 mutex_lock(&sctx->wr_lock);
3410 scrub_wr_submit(sctx);
3411 mutex_unlock(&sctx->wr_lock);
3412
3413 blk_finish_plug(&plug);
3414 btrfs_free_path(path);
3415 btrfs_free_path(ppath);
3416 return ret < 0 ? ret : 0;
3417}
3418
3419static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3420 struct btrfs_device *scrub_dev,
3421 u64 chunk_offset, u64 length,
3422 u64 dev_offset,
3423 struct btrfs_block_group_cache *cache)
3424{
3425 struct btrfs_fs_info *fs_info = sctx->fs_info;
3426 struct extent_map_tree *map_tree = &fs_info->mapping_tree;
3427 struct map_lookup *map;
3428 struct extent_map *em;
3429 int i;
3430 int ret = 0;
3431
3432 read_lock(&map_tree->lock);
3433 em = lookup_extent_mapping(map_tree, chunk_offset, 1);
3434 read_unlock(&map_tree->lock);
3435
3436 if (!em) {
3437 /*
3438 * Might have been an unused block group deleted by the cleaner
3439 * kthread or relocation.
3440 */
3441 spin_lock(&cache->lock);
3442 if (!cache->removed)
3443 ret = -EINVAL;
3444 spin_unlock(&cache->lock);
3445
3446 return ret;
3447 }
3448
3449 map = em->map_lookup;
3450 if (em->start != chunk_offset)
3451 goto out;
3452
3453 if (em->len < length)
3454 goto out;
3455
3456 for (i = 0; i < map->num_stripes; ++i) {
3457 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3458 map->stripes[i].physical == dev_offset) {
3459 ret = scrub_stripe(sctx, map, scrub_dev, i,
3460 chunk_offset, length);
3461 if (ret)
3462 goto out;
3463 }
3464 }
3465out:
3466 free_extent_map(em);
3467
3468 return ret;
3469}
3470
3471static noinline_for_stack
3472int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3473 struct btrfs_device *scrub_dev, u64 start, u64 end)
3474{
3475 struct btrfs_dev_extent *dev_extent = NULL;
3476 struct btrfs_path *path;
3477 struct btrfs_fs_info *fs_info = sctx->fs_info;
3478 struct btrfs_root *root = fs_info->dev_root;
3479 u64 length;
3480 u64 chunk_offset;
3481 int ret = 0;
3482 int ro_set;
3483 int slot;
3484 struct extent_buffer *l;
3485 struct btrfs_key key;
3486 struct btrfs_key found_key;
3487 struct btrfs_block_group_cache *cache;
3488 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3489
3490 path = btrfs_alloc_path();
3491 if (!path)
3492 return -ENOMEM;
3493
3494 path->reada = READA_FORWARD;
3495 path->search_commit_root = 1;
3496 path->skip_locking = 1;
3497
3498 key.objectid = scrub_dev->devid;
3499 key.offset = 0ull;
3500 key.type = BTRFS_DEV_EXTENT_KEY;
3501
3502 while (1) {
3503 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3504 if (ret < 0)
3505 break;
3506 if (ret > 0) {
3507 if (path->slots[0] >=
3508 btrfs_header_nritems(path->nodes[0])) {
3509 ret = btrfs_next_leaf(root, path);
3510 if (ret < 0)
3511 break;
3512 if (ret > 0) {
3513 ret = 0;
3514 break;
3515 }
3516 } else {
3517 ret = 0;
3518 }
3519 }
3520
3521 l = path->nodes[0];
3522 slot = path->slots[0];
3523
3524 btrfs_item_key_to_cpu(l, &found_key, slot);
3525
3526 if (found_key.objectid != scrub_dev->devid)
3527 break;
3528
3529 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3530 break;
3531
3532 if (found_key.offset >= end)
3533 break;
3534
3535 if (found_key.offset < key.offset)
3536 break;
3537
3538 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3539 length = btrfs_dev_extent_length(l, dev_extent);
3540
3541 if (found_key.offset + length <= start)
3542 goto skip;
3543
3544 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3545
3546 /*
3547 * get a reference on the corresponding block group to prevent
3548 * the chunk from going away while we scrub it
3549 */
3550 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3551
3552 /* some chunks are removed but not committed to disk yet,
3553 * continue scrubbing */
3554 if (!cache)
3555 goto skip;
3556
3557 /*
3558 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3559 * to avoid deadlock caused by:
3560 * btrfs_inc_block_group_ro()
3561 * -> btrfs_wait_for_commit()
3562 * -> btrfs_commit_transaction()
3563 * -> btrfs_scrub_pause()
3564 */
3565 scrub_pause_on(fs_info);
3566 ret = btrfs_inc_block_group_ro(cache);
3567 if (!ret && sctx->is_dev_replace) {
3568 /*
3569 * If we are doing a device replace wait for any tasks
3570 * that started delalloc right before we set the block
3571 * group to RO mode, as they might have just allocated
3572 * an extent from it or decided they could do a nocow
3573 * write. And if any such tasks did that, wait for their
3574 * ordered extents to complete and then commit the
3575 * current transaction, so that we can later see the new
3576 * extent items in the extent tree - the ordered extents
3577 * create delayed data references (for cow writes) when
3578 * they complete, which will be run and insert the
3579 * corresponding extent items into the extent tree when
3580 * we commit the transaction they used when running
3581 * inode.c:btrfs_finish_ordered_io(). We later use
3582 * the commit root of the extent tree to find extents
3583 * to copy from the srcdev into the tgtdev, and we don't
3584 * want to miss any new extents.
3585 */
3586 btrfs_wait_block_group_reservations(cache);
3587 btrfs_wait_nocow_writers(cache);
3588 ret = btrfs_wait_ordered_roots(fs_info, U64_MAX,
3589 cache->key.objectid,
3590 cache->key.offset);
3591 if (ret > 0) {
3592 struct btrfs_trans_handle *trans;
3593
3594 trans = btrfs_join_transaction(root);
3595 if (IS_ERR(trans))
3596 ret = PTR_ERR(trans);
3597 else
3598 ret = btrfs_commit_transaction(trans);
3599 if (ret) {
3600 scrub_pause_off(fs_info);
3601 btrfs_put_block_group(cache);
3602 break;
3603 }
3604 }
3605 }
3606 scrub_pause_off(fs_info);
3607
3608 if (ret == 0) {
3609 ro_set = 1;
3610 } else if (ret == -ENOSPC) {
3611 /*
3612 * btrfs_inc_block_group_ro return -ENOSPC when it
3613 * failed in creating new chunk for metadata.
3614 * It is not a problem for scrub/replace, because
3615 * metadata are always cowed, and our scrub paused
3616 * commit_transactions.
3617 */
3618 ro_set = 0;
3619 } else {
3620 btrfs_warn(fs_info,
3621 "failed setting block group ro: %d", ret);
3622 btrfs_put_block_group(cache);
3623 break;
3624 }
3625
3626 down_write(&fs_info->dev_replace.rwsem);
3627 dev_replace->cursor_right = found_key.offset + length;
3628 dev_replace->cursor_left = found_key.offset;
3629 dev_replace->item_needs_writeback = 1;
3630 up_write(&dev_replace->rwsem);
3631
3632 ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3633 found_key.offset, cache);
3634
3635 /*
3636 * flush, submit all pending read and write bios, afterwards
3637 * wait for them.
3638 * Note that in the dev replace case, a read request causes
3639 * write requests that are submitted in the read completion
3640 * worker. Therefore in the current situation, it is required
3641 * that all write requests are flushed, so that all read and
3642 * write requests are really completed when bios_in_flight
3643 * changes to 0.
3644 */
3645 sctx->flush_all_writes = true;
3646 scrub_submit(sctx);
3647 mutex_lock(&sctx->wr_lock);
3648 scrub_wr_submit(sctx);
3649 mutex_unlock(&sctx->wr_lock);
3650
3651 wait_event(sctx->list_wait,
3652 atomic_read(&sctx->bios_in_flight) == 0);
3653
3654 scrub_pause_on(fs_info);
3655
3656 /*
3657 * must be called before we decrease @scrub_paused.
3658 * make sure we don't block transaction commit while
3659 * we are waiting pending workers finished.
3660 */
3661 wait_event(sctx->list_wait,
3662 atomic_read(&sctx->workers_pending) == 0);
3663 sctx->flush_all_writes = false;
3664
3665 scrub_pause_off(fs_info);
3666
3667 down_write(&fs_info->dev_replace.rwsem);
3668 dev_replace->cursor_left = dev_replace->cursor_right;
3669 dev_replace->item_needs_writeback = 1;
3670 up_write(&fs_info->dev_replace.rwsem);
3671
3672 if (ro_set)
3673 btrfs_dec_block_group_ro(cache);
3674
3675 /*
3676 * We might have prevented the cleaner kthread from deleting
3677 * this block group if it was already unused because we raced
3678 * and set it to RO mode first. So add it back to the unused
3679 * list, otherwise it might not ever be deleted unless a manual
3680 * balance is triggered or it becomes used and unused again.
3681 */
3682 spin_lock(&cache->lock);
3683 if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3684 btrfs_block_group_used(&cache->item) == 0) {
3685 spin_unlock(&cache->lock);
3686 btrfs_mark_bg_unused(cache);
3687 } else {
3688 spin_unlock(&cache->lock);
3689 }
3690
3691 btrfs_put_block_group(cache);
3692 if (ret)
3693 break;
3694 if (sctx->is_dev_replace &&
3695 atomic64_read(&dev_replace->num_write_errors) > 0) {
3696 ret = -EIO;
3697 break;
3698 }
3699 if (sctx->stat.malloc_errors > 0) {
3700 ret = -ENOMEM;
3701 break;
3702 }
3703skip:
3704 key.offset = found_key.offset + length;
3705 btrfs_release_path(path);
3706 }
3707
3708 btrfs_free_path(path);
3709
3710 return ret;
3711}
3712
3713static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3714 struct btrfs_device *scrub_dev)
3715{
3716 int i;
3717 u64 bytenr;
3718 u64 gen;
3719 int ret;
3720 struct btrfs_fs_info *fs_info = sctx->fs_info;
3721
3722 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
3723 return -EIO;
3724
3725 /* Seed devices of a new filesystem has their own generation. */
3726 if (scrub_dev->fs_devices != fs_info->fs_devices)
3727 gen = scrub_dev->generation;
3728 else
3729 gen = fs_info->last_trans_committed;
3730
3731 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3732 bytenr = btrfs_sb_offset(i);
3733 if (bytenr + BTRFS_SUPER_INFO_SIZE >
3734 scrub_dev->commit_total_bytes)
3735 break;
3736
3737 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3738 scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3739 NULL, 1, bytenr);
3740 if (ret)
3741 return ret;
3742 }
3743 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3744
3745 return 0;
3746}
3747
3748/*
3749 * get a reference count on fs_info->scrub_workers. start worker if necessary
3750 */
3751static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
3752 int is_dev_replace)
3753{
3754 unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
3755 int max_active = fs_info->thread_pool_size;
3756
3757 lockdep_assert_held(&fs_info->scrub_lock);
3758
3759 if (refcount_read(&fs_info->scrub_workers_refcnt) == 0) {
3760 ASSERT(fs_info->scrub_workers == NULL);
3761 fs_info->scrub_workers = btrfs_alloc_workqueue(fs_info, "scrub",
3762 flags, is_dev_replace ? 1 : max_active, 4);
3763 if (!fs_info->scrub_workers)
3764 goto fail_scrub_workers;
3765
3766 ASSERT(fs_info->scrub_wr_completion_workers == NULL);
3767 fs_info->scrub_wr_completion_workers =
3768 btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
3769 max_active, 2);
3770 if (!fs_info->scrub_wr_completion_workers)
3771 goto fail_scrub_wr_completion_workers;
3772
3773 ASSERT(fs_info->scrub_parity_workers == NULL);
3774 fs_info->scrub_parity_workers =
3775 btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
3776 max_active, 2);
3777 if (!fs_info->scrub_parity_workers)
3778 goto fail_scrub_parity_workers;
3779
3780 refcount_set(&fs_info->scrub_workers_refcnt, 1);
3781 } else {
3782 refcount_inc(&fs_info->scrub_workers_refcnt);
3783 }
3784 return 0;
3785
3786fail_scrub_parity_workers:
3787 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3788fail_scrub_wr_completion_workers:
3789 btrfs_destroy_workqueue(fs_info->scrub_workers);
3790fail_scrub_workers:
3791 return -ENOMEM;
3792}
3793
3794int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
3795 u64 end, struct btrfs_scrub_progress *progress,
3796 int readonly, int is_dev_replace)
3797{
3798 struct scrub_ctx *sctx;
3799 int ret;
3800 struct btrfs_device *dev;
3801 unsigned int nofs_flag;
3802 struct btrfs_workqueue *scrub_workers = NULL;
3803 struct btrfs_workqueue *scrub_wr_comp = NULL;
3804 struct btrfs_workqueue *scrub_parity = NULL;
3805
3806 if (btrfs_fs_closing(fs_info))
3807 return -EAGAIN;
3808
3809 if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
3810 /*
3811 * in this case scrub is unable to calculate the checksum
3812 * the way scrub is implemented. Do not handle this
3813 * situation at all because it won't ever happen.
3814 */
3815 btrfs_err(fs_info,
3816 "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
3817 fs_info->nodesize,
3818 BTRFS_STRIPE_LEN);
3819 return -EINVAL;
3820 }
3821
3822 if (fs_info->sectorsize != PAGE_SIZE) {
3823 /* not supported for data w/o checksums */
3824 btrfs_err_rl(fs_info,
3825 "scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails",
3826 fs_info->sectorsize, PAGE_SIZE);
3827 return -EINVAL;
3828 }
3829
3830 if (fs_info->nodesize >
3831 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
3832 fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
3833 /*
3834 * would exhaust the array bounds of pagev member in
3835 * struct scrub_block
3836 */
3837 btrfs_err(fs_info,
3838 "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
3839 fs_info->nodesize,
3840 SCRUB_MAX_PAGES_PER_BLOCK,
3841 fs_info->sectorsize,
3842 SCRUB_MAX_PAGES_PER_BLOCK);
3843 return -EINVAL;
3844 }
3845
3846 /* Allocate outside of device_list_mutex */
3847 sctx = scrub_setup_ctx(fs_info, is_dev_replace);
3848 if (IS_ERR(sctx))
3849 return PTR_ERR(sctx);
3850
3851 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3852 dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL, true);
3853 if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
3854 !is_dev_replace)) {
3855 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3856 ret = -ENODEV;
3857 goto out_free_ctx;
3858 }
3859
3860 if (!is_dev_replace && !readonly &&
3861 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
3862 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3863 btrfs_err_in_rcu(fs_info, "scrub: device %s is not writable",
3864 rcu_str_deref(dev->name));
3865 ret = -EROFS;
3866 goto out_free_ctx;
3867 }
3868
3869 mutex_lock(&fs_info->scrub_lock);
3870 if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
3871 test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
3872 mutex_unlock(&fs_info->scrub_lock);
3873 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3874 ret = -EIO;
3875 goto out_free_ctx;
3876 }
3877
3878 down_read(&fs_info->dev_replace.rwsem);
3879 if (dev->scrub_ctx ||
3880 (!is_dev_replace &&
3881 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
3882 up_read(&fs_info->dev_replace.rwsem);
3883 mutex_unlock(&fs_info->scrub_lock);
3884 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3885 ret = -EINPROGRESS;
3886 goto out_free_ctx;
3887 }
3888 up_read(&fs_info->dev_replace.rwsem);
3889
3890 ret = scrub_workers_get(fs_info, is_dev_replace);
3891 if (ret) {
3892 mutex_unlock(&fs_info->scrub_lock);
3893 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3894 goto out_free_ctx;
3895 }
3896
3897 sctx->readonly = readonly;
3898 dev->scrub_ctx = sctx;
3899 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3900
3901 /*
3902 * checking @scrub_pause_req here, we can avoid
3903 * race between committing transaction and scrubbing.
3904 */
3905 __scrub_blocked_if_needed(fs_info);
3906 atomic_inc(&fs_info->scrubs_running);
3907 mutex_unlock(&fs_info->scrub_lock);
3908
3909 /*
3910 * In order to avoid deadlock with reclaim when there is a transaction
3911 * trying to pause scrub, make sure we use GFP_NOFS for all the
3912 * allocations done at btrfs_scrub_pages() and scrub_pages_for_parity()
3913 * invoked by our callees. The pausing request is done when the
3914 * transaction commit starts, and it blocks the transaction until scrub
3915 * is paused (done at specific points at scrub_stripe() or right above
3916 * before incrementing fs_info->scrubs_running).
3917 */
3918 nofs_flag = memalloc_nofs_save();
3919 if (!is_dev_replace) {
3920 btrfs_info(fs_info, "scrub: started on devid %llu", devid);
3921 /*
3922 * by holding device list mutex, we can
3923 * kick off writing super in log tree sync.
3924 */
3925 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3926 ret = scrub_supers(sctx, dev);
3927 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3928 }
3929
3930 if (!ret)
3931 ret = scrub_enumerate_chunks(sctx, dev, start, end);
3932 memalloc_nofs_restore(nofs_flag);
3933
3934 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3935 atomic_dec(&fs_info->scrubs_running);
3936 wake_up(&fs_info->scrub_pause_wait);
3937
3938 wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3939
3940 if (progress)
3941 memcpy(progress, &sctx->stat, sizeof(*progress));
3942
3943 if (!is_dev_replace)
3944 btrfs_info(fs_info, "scrub: %s on devid %llu with status: %d",
3945 ret ? "not finished" : "finished", devid, ret);
3946
3947 mutex_lock(&fs_info->scrub_lock);
3948 dev->scrub_ctx = NULL;
3949 if (refcount_dec_and_test(&fs_info->scrub_workers_refcnt)) {
3950 scrub_workers = fs_info->scrub_workers;
3951 scrub_wr_comp = fs_info->scrub_wr_completion_workers;
3952 scrub_parity = fs_info->scrub_parity_workers;
3953
3954 fs_info->scrub_workers = NULL;
3955 fs_info->scrub_wr_completion_workers = NULL;
3956 fs_info->scrub_parity_workers = NULL;
3957 }
3958 mutex_unlock(&fs_info->scrub_lock);
3959
3960 btrfs_destroy_workqueue(scrub_workers);
3961 btrfs_destroy_workqueue(scrub_wr_comp);
3962 btrfs_destroy_workqueue(scrub_parity);
3963 scrub_put_ctx(sctx);
3964
3965 return ret;
3966
3967out_free_ctx:
3968 scrub_free_ctx(sctx);
3969
3970 return ret;
3971}
3972
3973void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
3974{
3975 mutex_lock(&fs_info->scrub_lock);
3976 atomic_inc(&fs_info->scrub_pause_req);
3977 while (atomic_read(&fs_info->scrubs_paused) !=
3978 atomic_read(&fs_info->scrubs_running)) {
3979 mutex_unlock(&fs_info->scrub_lock);
3980 wait_event(fs_info->scrub_pause_wait,
3981 atomic_read(&fs_info->scrubs_paused) ==
3982 atomic_read(&fs_info->scrubs_running));
3983 mutex_lock(&fs_info->scrub_lock);
3984 }
3985 mutex_unlock(&fs_info->scrub_lock);
3986}
3987
3988void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
3989{
3990 atomic_dec(&fs_info->scrub_pause_req);
3991 wake_up(&fs_info->scrub_pause_wait);
3992}
3993
3994int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
3995{
3996 mutex_lock(&fs_info->scrub_lock);
3997 if (!atomic_read(&fs_info->scrubs_running)) {
3998 mutex_unlock(&fs_info->scrub_lock);
3999 return -ENOTCONN;
4000 }
4001
4002 atomic_inc(&fs_info->scrub_cancel_req);
4003 while (atomic_read(&fs_info->scrubs_running)) {
4004 mutex_unlock(&fs_info->scrub_lock);
4005 wait_event(fs_info->scrub_pause_wait,
4006 atomic_read(&fs_info->scrubs_running) == 0);
4007 mutex_lock(&fs_info->scrub_lock);
4008 }
4009 atomic_dec(&fs_info->scrub_cancel_req);
4010 mutex_unlock(&fs_info->scrub_lock);
4011
4012 return 0;
4013}
4014
4015int btrfs_scrub_cancel_dev(struct btrfs_device *dev)
4016{
4017 struct btrfs_fs_info *fs_info = dev->fs_info;
4018 struct scrub_ctx *sctx;
4019
4020 mutex_lock(&fs_info->scrub_lock);
4021 sctx = dev->scrub_ctx;
4022 if (!sctx) {
4023 mutex_unlock(&fs_info->scrub_lock);
4024 return -ENOTCONN;
4025 }
4026 atomic_inc(&sctx->cancel_req);
4027 while (dev->scrub_ctx) {
4028 mutex_unlock(&fs_info->scrub_lock);
4029 wait_event(fs_info->scrub_pause_wait,
4030 dev->scrub_ctx == NULL);
4031 mutex_lock(&fs_info->scrub_lock);
4032 }
4033 mutex_unlock(&fs_info->scrub_lock);
4034
4035 return 0;
4036}
4037
4038int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4039 struct btrfs_scrub_progress *progress)
4040{
4041 struct btrfs_device *dev;
4042 struct scrub_ctx *sctx = NULL;
4043
4044 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4045 dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL, true);
4046 if (dev)
4047 sctx = dev->scrub_ctx;
4048 if (sctx)
4049 memcpy(progress, &sctx->stat, sizeof(*progress));
4050 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4051
4052 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4053}
4054
4055static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4056 u64 extent_logical, u64 extent_len,
4057 u64 *extent_physical,
4058 struct btrfs_device **extent_dev,
4059 int *extent_mirror_num)
4060{
4061 u64 mapped_length;
4062 struct btrfs_bio *bbio = NULL;
4063 int ret;
4064
4065 mapped_length = extent_len;
4066 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4067 &mapped_length, &bbio, 0);
4068 if (ret || !bbio || mapped_length < extent_len ||
4069 !bbio->stripes[0].dev->bdev) {
4070 btrfs_put_bbio(bbio);
4071 return;
4072 }
4073
4074 *extent_physical = bbio->stripes[0].physical;
4075 *extent_mirror_num = bbio->mirror_num;
4076 *extent_dev = bbio->stripes[0].dev;
4077 btrfs_put_bbio(bbio);
4078}