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