Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2011, 2012 STRATO. All rights reserved.
4 */
5
6#include <linux/blkdev.h>
7#include <linux/ratelimit.h>
8#include <linux/sched/mm.h>
9#include <crypto/hash.h>
10#include "ctree.h"
11#include "discard.h"
12#include "volumes.h"
13#include "disk-io.h"
14#include "ordered-data.h"
15#include "transaction.h"
16#include "backref.h"
17#include "extent_io.h"
18#include "dev-replace.h"
19#include "raid56.h"
20#include "block-group.h"
21#include "zoned.h"
22#include "fs.h"
23#include "accessors.h"
24#include "file-item.h"
25#include "scrub.h"
26#include "raid-stripe-tree.h"
27
28/*
29 * This is only the first step towards a full-features scrub. It reads all
30 * extent and super block and verifies the checksums. In case a bad checksum
31 * is found or the extent cannot be read, good data will be written back if
32 * any can be found.
33 *
34 * Future enhancements:
35 * - In case an unrepairable extent is encountered, track which files are
36 * affected and report them
37 * - track and record media errors, throw out bad devices
38 * - add a mode to also read unallocated space
39 */
40
41struct scrub_ctx;
42
43/*
44 * The following value only influences the performance.
45 *
46 * This determines how many stripes would be submitted in one go,
47 * which is 512KiB (BTRFS_STRIPE_LEN * SCRUB_STRIPES_PER_GROUP).
48 */
49#define SCRUB_STRIPES_PER_GROUP 8
50
51/*
52 * How many groups we have for each sctx.
53 *
54 * This would be 8M per device, the same value as the old scrub in-flight bios
55 * size limit.
56 */
57#define SCRUB_GROUPS_PER_SCTX 16
58
59#define SCRUB_TOTAL_STRIPES (SCRUB_GROUPS_PER_SCTX * SCRUB_STRIPES_PER_GROUP)
60
61/*
62 * The following value times PAGE_SIZE needs to be large enough to match the
63 * largest node/leaf/sector size that shall be supported.
64 */
65#define SCRUB_MAX_SECTORS_PER_BLOCK (BTRFS_MAX_METADATA_BLOCKSIZE / SZ_4K)
66
67/* Represent one sector and its needed info to verify the content. */
68struct scrub_sector_verification {
69 bool is_metadata;
70
71 union {
72 /*
73 * Csum pointer for data csum verification. Should point to a
74 * sector csum inside scrub_stripe::csums.
75 *
76 * NULL if this data sector has no csum.
77 */
78 u8 *csum;
79
80 /*
81 * Extra info for metadata verification. All sectors inside a
82 * tree block share the same generation.
83 */
84 u64 generation;
85 };
86};
87
88enum scrub_stripe_flags {
89 /* Set when @mirror_num, @dev, @physical and @logical are set. */
90 SCRUB_STRIPE_FLAG_INITIALIZED,
91
92 /* Set when the read-repair is finished. */
93 SCRUB_STRIPE_FLAG_REPAIR_DONE,
94
95 /*
96 * Set for data stripes if it's triggered from P/Q stripe.
97 * During such scrub, we should not report errors in data stripes, nor
98 * update the accounting.
99 */
100 SCRUB_STRIPE_FLAG_NO_REPORT,
101};
102
103#define SCRUB_STRIPE_PAGES (BTRFS_STRIPE_LEN / PAGE_SIZE)
104
105/*
106 * Represent one contiguous range with a length of BTRFS_STRIPE_LEN.
107 */
108struct scrub_stripe {
109 struct scrub_ctx *sctx;
110 struct btrfs_block_group *bg;
111
112 struct page *pages[SCRUB_STRIPE_PAGES];
113 struct scrub_sector_verification *sectors;
114
115 struct btrfs_device *dev;
116 u64 logical;
117 u64 physical;
118
119 u16 mirror_num;
120
121 /* Should be BTRFS_STRIPE_LEN / sectorsize. */
122 u16 nr_sectors;
123
124 /*
125 * How many data/meta extents are in this stripe. Only for scrub status
126 * reporting purposes.
127 */
128 u16 nr_data_extents;
129 u16 nr_meta_extents;
130
131 atomic_t pending_io;
132 wait_queue_head_t io_wait;
133 wait_queue_head_t repair_wait;
134
135 /*
136 * Indicate the states of the stripe. Bits are defined in
137 * scrub_stripe_flags enum.
138 */
139 unsigned long state;
140
141 /* Indicate which sectors are covered by extent items. */
142 unsigned long extent_sector_bitmap;
143
144 /*
145 * The errors hit during the initial read of the stripe.
146 *
147 * Would be utilized for error reporting and repair.
148 *
149 * The remaining init_nr_* records the number of errors hit, only used
150 * by error reporting.
151 */
152 unsigned long init_error_bitmap;
153 unsigned int init_nr_io_errors;
154 unsigned int init_nr_csum_errors;
155 unsigned int init_nr_meta_errors;
156
157 /*
158 * The following error bitmaps are all for the current status.
159 * Every time we submit a new read, these bitmaps may be updated.
160 *
161 * error_bitmap = io_error_bitmap | csum_error_bitmap | meta_error_bitmap;
162 *
163 * IO and csum errors can happen for both metadata and data.
164 */
165 unsigned long error_bitmap;
166 unsigned long io_error_bitmap;
167 unsigned long csum_error_bitmap;
168 unsigned long meta_error_bitmap;
169
170 /* For writeback (repair or replace) error reporting. */
171 unsigned long write_error_bitmap;
172
173 /* Writeback can be concurrent, thus we need to protect the bitmap. */
174 spinlock_t write_error_lock;
175
176 /*
177 * Checksum for the whole stripe if this stripe is inside a data block
178 * group.
179 */
180 u8 *csums;
181
182 struct work_struct work;
183};
184
185struct scrub_ctx {
186 struct scrub_stripe stripes[SCRUB_TOTAL_STRIPES];
187 struct scrub_stripe *raid56_data_stripes;
188 struct btrfs_fs_info *fs_info;
189 struct btrfs_path extent_path;
190 struct btrfs_path csum_path;
191 int first_free;
192 int cur_stripe;
193 atomic_t cancel_req;
194 int readonly;
195
196 /* State of IO submission throttling affecting the associated device */
197 ktime_t throttle_deadline;
198 u64 throttle_sent;
199
200 int is_dev_replace;
201 u64 write_pointer;
202
203 struct mutex wr_lock;
204 struct btrfs_device *wr_tgtdev;
205
206 /*
207 * statistics
208 */
209 struct btrfs_scrub_progress stat;
210 spinlock_t stat_lock;
211
212 /*
213 * Use a ref counter to avoid use-after-free issues. Scrub workers
214 * decrement bios_in_flight and workers_pending and then do a wakeup
215 * on the list_wait wait queue. We must ensure the main scrub task
216 * doesn't free the scrub context before or while the workers are
217 * doing the wakeup() call.
218 */
219 refcount_t refs;
220};
221
222struct scrub_warning {
223 struct btrfs_path *path;
224 u64 extent_item_size;
225 const char *errstr;
226 u64 physical;
227 u64 logical;
228 struct btrfs_device *dev;
229};
230
231static void release_scrub_stripe(struct scrub_stripe *stripe)
232{
233 if (!stripe)
234 return;
235
236 for (int i = 0; i < SCRUB_STRIPE_PAGES; i++) {
237 if (stripe->pages[i])
238 __free_page(stripe->pages[i]);
239 stripe->pages[i] = NULL;
240 }
241 kfree(stripe->sectors);
242 kfree(stripe->csums);
243 stripe->sectors = NULL;
244 stripe->csums = NULL;
245 stripe->sctx = NULL;
246 stripe->state = 0;
247}
248
249static int init_scrub_stripe(struct btrfs_fs_info *fs_info,
250 struct scrub_stripe *stripe)
251{
252 int ret;
253
254 memset(stripe, 0, sizeof(*stripe));
255
256 stripe->nr_sectors = BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits;
257 stripe->state = 0;
258
259 init_waitqueue_head(&stripe->io_wait);
260 init_waitqueue_head(&stripe->repair_wait);
261 atomic_set(&stripe->pending_io, 0);
262 spin_lock_init(&stripe->write_error_lock);
263
264 ret = btrfs_alloc_page_array(SCRUB_STRIPE_PAGES, stripe->pages, 0);
265 if (ret < 0)
266 goto error;
267
268 stripe->sectors = kcalloc(stripe->nr_sectors,
269 sizeof(struct scrub_sector_verification),
270 GFP_KERNEL);
271 if (!stripe->sectors)
272 goto error;
273
274 stripe->csums = kcalloc(BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits,
275 fs_info->csum_size, GFP_KERNEL);
276 if (!stripe->csums)
277 goto error;
278 return 0;
279error:
280 release_scrub_stripe(stripe);
281 return -ENOMEM;
282}
283
284static void wait_scrub_stripe_io(struct scrub_stripe *stripe)
285{
286 wait_event(stripe->io_wait, atomic_read(&stripe->pending_io) == 0);
287}
288
289static void scrub_put_ctx(struct scrub_ctx *sctx);
290
291static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
292{
293 while (atomic_read(&fs_info->scrub_pause_req)) {
294 mutex_unlock(&fs_info->scrub_lock);
295 wait_event(fs_info->scrub_pause_wait,
296 atomic_read(&fs_info->scrub_pause_req) == 0);
297 mutex_lock(&fs_info->scrub_lock);
298 }
299}
300
301static void scrub_pause_on(struct btrfs_fs_info *fs_info)
302{
303 atomic_inc(&fs_info->scrubs_paused);
304 wake_up(&fs_info->scrub_pause_wait);
305}
306
307static void scrub_pause_off(struct btrfs_fs_info *fs_info)
308{
309 mutex_lock(&fs_info->scrub_lock);
310 __scrub_blocked_if_needed(fs_info);
311 atomic_dec(&fs_info->scrubs_paused);
312 mutex_unlock(&fs_info->scrub_lock);
313
314 wake_up(&fs_info->scrub_pause_wait);
315}
316
317static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
318{
319 scrub_pause_on(fs_info);
320 scrub_pause_off(fs_info);
321}
322
323static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
324{
325 int i;
326
327 if (!sctx)
328 return;
329
330 for (i = 0; i < SCRUB_TOTAL_STRIPES; i++)
331 release_scrub_stripe(&sctx->stripes[i]);
332
333 kvfree(sctx);
334}
335
336static void scrub_put_ctx(struct scrub_ctx *sctx)
337{
338 if (refcount_dec_and_test(&sctx->refs))
339 scrub_free_ctx(sctx);
340}
341
342static noinline_for_stack struct scrub_ctx *scrub_setup_ctx(
343 struct btrfs_fs_info *fs_info, int is_dev_replace)
344{
345 struct scrub_ctx *sctx;
346 int i;
347
348 /* Since sctx has inline 128 stripes, it can go beyond 64K easily. Use
349 * kvzalloc().
350 */
351 sctx = kvzalloc(sizeof(*sctx), GFP_KERNEL);
352 if (!sctx)
353 goto nomem;
354 refcount_set(&sctx->refs, 1);
355 sctx->is_dev_replace = is_dev_replace;
356 sctx->fs_info = fs_info;
357 sctx->extent_path.search_commit_root = 1;
358 sctx->extent_path.skip_locking = 1;
359 sctx->csum_path.search_commit_root = 1;
360 sctx->csum_path.skip_locking = 1;
361 for (i = 0; i < SCRUB_TOTAL_STRIPES; i++) {
362 int ret;
363
364 ret = init_scrub_stripe(fs_info, &sctx->stripes[i]);
365 if (ret < 0)
366 goto nomem;
367 sctx->stripes[i].sctx = sctx;
368 }
369 sctx->first_free = 0;
370 atomic_set(&sctx->cancel_req, 0);
371
372 spin_lock_init(&sctx->stat_lock);
373 sctx->throttle_deadline = 0;
374
375 mutex_init(&sctx->wr_lock);
376 if (is_dev_replace) {
377 WARN_ON(!fs_info->dev_replace.tgtdev);
378 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
379 }
380
381 return sctx;
382
383nomem:
384 scrub_free_ctx(sctx);
385 return ERR_PTR(-ENOMEM);
386}
387
388static int scrub_print_warning_inode(u64 inum, u64 offset, u64 num_bytes,
389 u64 root, void *warn_ctx)
390{
391 u32 nlink;
392 int ret;
393 int i;
394 unsigned nofs_flag;
395 struct extent_buffer *eb;
396 struct btrfs_inode_item *inode_item;
397 struct scrub_warning *swarn = warn_ctx;
398 struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
399 struct inode_fs_paths *ipath = NULL;
400 struct btrfs_root *local_root;
401 struct btrfs_key key;
402
403 local_root = btrfs_get_fs_root(fs_info, root, true);
404 if (IS_ERR(local_root)) {
405 ret = PTR_ERR(local_root);
406 goto err;
407 }
408
409 /*
410 * this makes the path point to (inum INODE_ITEM ioff)
411 */
412 key.objectid = inum;
413 key.type = BTRFS_INODE_ITEM_KEY;
414 key.offset = 0;
415
416 ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
417 if (ret) {
418 btrfs_put_root(local_root);
419 btrfs_release_path(swarn->path);
420 goto err;
421 }
422
423 eb = swarn->path->nodes[0];
424 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
425 struct btrfs_inode_item);
426 nlink = btrfs_inode_nlink(eb, inode_item);
427 btrfs_release_path(swarn->path);
428
429 /*
430 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
431 * uses GFP_NOFS in this context, so we keep it consistent but it does
432 * not seem to be strictly necessary.
433 */
434 nofs_flag = memalloc_nofs_save();
435 ipath = init_ipath(4096, local_root, swarn->path);
436 memalloc_nofs_restore(nofs_flag);
437 if (IS_ERR(ipath)) {
438 btrfs_put_root(local_root);
439 ret = PTR_ERR(ipath);
440 ipath = NULL;
441 goto err;
442 }
443 ret = paths_from_inode(inum, ipath);
444
445 if (ret < 0)
446 goto err;
447
448 /*
449 * we deliberately ignore the bit ipath might have been too small to
450 * hold all of the paths here
451 */
452 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
453 btrfs_warn_in_rcu(fs_info,
454"%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %u, links %u (path: %s)",
455 swarn->errstr, swarn->logical,
456 btrfs_dev_name(swarn->dev),
457 swarn->physical,
458 root, inum, offset,
459 fs_info->sectorsize, nlink,
460 (char *)(unsigned long)ipath->fspath->val[i]);
461
462 btrfs_put_root(local_root);
463 free_ipath(ipath);
464 return 0;
465
466err:
467 btrfs_warn_in_rcu(fs_info,
468 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
469 swarn->errstr, swarn->logical,
470 btrfs_dev_name(swarn->dev),
471 swarn->physical,
472 root, inum, offset, ret);
473
474 free_ipath(ipath);
475 return 0;
476}
477
478static void scrub_print_common_warning(const char *errstr, struct btrfs_device *dev,
479 bool is_super, u64 logical, u64 physical)
480{
481 struct btrfs_fs_info *fs_info = dev->fs_info;
482 struct btrfs_path *path;
483 struct btrfs_key found_key;
484 struct extent_buffer *eb;
485 struct btrfs_extent_item *ei;
486 struct scrub_warning swarn;
487 u64 flags = 0;
488 u32 item_size;
489 int ret;
490
491 /* Super block error, no need to search extent tree. */
492 if (is_super) {
493 btrfs_warn_in_rcu(fs_info, "%s on device %s, physical %llu",
494 errstr, btrfs_dev_name(dev), physical);
495 return;
496 }
497 path = btrfs_alloc_path();
498 if (!path)
499 return;
500
501 swarn.physical = physical;
502 swarn.logical = logical;
503 swarn.errstr = errstr;
504 swarn.dev = NULL;
505
506 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
507 &flags);
508 if (ret < 0)
509 goto out;
510
511 swarn.extent_item_size = found_key.offset;
512
513 eb = path->nodes[0];
514 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
515 item_size = btrfs_item_size(eb, path->slots[0]);
516
517 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
518 unsigned long ptr = 0;
519 u8 ref_level;
520 u64 ref_root;
521
522 while (true) {
523 ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
524 item_size, &ref_root,
525 &ref_level);
526 if (ret < 0) {
527 btrfs_warn(fs_info,
528 "failed to resolve tree backref for logical %llu: %d",
529 swarn.logical, ret);
530 break;
531 }
532 if (ret > 0)
533 break;
534 btrfs_warn_in_rcu(fs_info,
535"%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
536 errstr, swarn.logical, btrfs_dev_name(dev),
537 swarn.physical, (ref_level ? "node" : "leaf"),
538 ref_level, ref_root);
539 }
540 btrfs_release_path(path);
541 } else {
542 struct btrfs_backref_walk_ctx ctx = { 0 };
543
544 btrfs_release_path(path);
545
546 ctx.bytenr = found_key.objectid;
547 ctx.extent_item_pos = swarn.logical - found_key.objectid;
548 ctx.fs_info = fs_info;
549
550 swarn.path = path;
551 swarn.dev = dev;
552
553 iterate_extent_inodes(&ctx, true, scrub_print_warning_inode, &swarn);
554 }
555
556out:
557 btrfs_free_path(path);
558}
559
560static int fill_writer_pointer_gap(struct scrub_ctx *sctx, u64 physical)
561{
562 int ret = 0;
563 u64 length;
564
565 if (!btrfs_is_zoned(sctx->fs_info))
566 return 0;
567
568 if (!btrfs_dev_is_sequential(sctx->wr_tgtdev, physical))
569 return 0;
570
571 if (sctx->write_pointer < physical) {
572 length = physical - sctx->write_pointer;
573
574 ret = btrfs_zoned_issue_zeroout(sctx->wr_tgtdev,
575 sctx->write_pointer, length);
576 if (!ret)
577 sctx->write_pointer = physical;
578 }
579 return ret;
580}
581
582static struct page *scrub_stripe_get_page(struct scrub_stripe *stripe, int sector_nr)
583{
584 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
585 int page_index = (sector_nr << fs_info->sectorsize_bits) >> PAGE_SHIFT;
586
587 return stripe->pages[page_index];
588}
589
590static unsigned int scrub_stripe_get_page_offset(struct scrub_stripe *stripe,
591 int sector_nr)
592{
593 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
594
595 return offset_in_page(sector_nr << fs_info->sectorsize_bits);
596}
597
598static void scrub_verify_one_metadata(struct scrub_stripe *stripe, int sector_nr)
599{
600 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
601 const u32 sectors_per_tree = fs_info->nodesize >> fs_info->sectorsize_bits;
602 const u64 logical = stripe->logical + (sector_nr << fs_info->sectorsize_bits);
603 const struct page *first_page = scrub_stripe_get_page(stripe, sector_nr);
604 const unsigned int first_off = scrub_stripe_get_page_offset(stripe, sector_nr);
605 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
606 u8 on_disk_csum[BTRFS_CSUM_SIZE];
607 u8 calculated_csum[BTRFS_CSUM_SIZE];
608 struct btrfs_header *header;
609
610 /*
611 * Here we don't have a good way to attach the pages (and subpages)
612 * to a dummy extent buffer, thus we have to directly grab the members
613 * from pages.
614 */
615 header = (struct btrfs_header *)(page_address(first_page) + first_off);
616 memcpy(on_disk_csum, header->csum, fs_info->csum_size);
617
618 if (logical != btrfs_stack_header_bytenr(header)) {
619 bitmap_set(&stripe->csum_error_bitmap, sector_nr, sectors_per_tree);
620 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
621 btrfs_warn_rl(fs_info,
622 "tree block %llu mirror %u has bad bytenr, has %llu want %llu",
623 logical, stripe->mirror_num,
624 btrfs_stack_header_bytenr(header), logical);
625 return;
626 }
627 if (memcmp(header->fsid, fs_info->fs_devices->metadata_uuid,
628 BTRFS_FSID_SIZE) != 0) {
629 bitmap_set(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
630 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
631 btrfs_warn_rl(fs_info,
632 "tree block %llu mirror %u has bad fsid, has %pU want %pU",
633 logical, stripe->mirror_num,
634 header->fsid, fs_info->fs_devices->fsid);
635 return;
636 }
637 if (memcmp(header->chunk_tree_uuid, fs_info->chunk_tree_uuid,
638 BTRFS_UUID_SIZE) != 0) {
639 bitmap_set(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
640 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
641 btrfs_warn_rl(fs_info,
642 "tree block %llu mirror %u has bad chunk tree uuid, has %pU want %pU",
643 logical, stripe->mirror_num,
644 header->chunk_tree_uuid, fs_info->chunk_tree_uuid);
645 return;
646 }
647
648 /* Now check tree block csum. */
649 shash->tfm = fs_info->csum_shash;
650 crypto_shash_init(shash);
651 crypto_shash_update(shash, page_address(first_page) + first_off +
652 BTRFS_CSUM_SIZE, fs_info->sectorsize - BTRFS_CSUM_SIZE);
653
654 for (int i = sector_nr + 1; i < sector_nr + sectors_per_tree; i++) {
655 struct page *page = scrub_stripe_get_page(stripe, i);
656 unsigned int page_off = scrub_stripe_get_page_offset(stripe, i);
657
658 crypto_shash_update(shash, page_address(page) + page_off,
659 fs_info->sectorsize);
660 }
661
662 crypto_shash_final(shash, calculated_csum);
663 if (memcmp(calculated_csum, on_disk_csum, fs_info->csum_size) != 0) {
664 bitmap_set(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
665 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
666 btrfs_warn_rl(fs_info,
667 "tree block %llu mirror %u has bad csum, has " CSUM_FMT " want " CSUM_FMT,
668 logical, stripe->mirror_num,
669 CSUM_FMT_VALUE(fs_info->csum_size, on_disk_csum),
670 CSUM_FMT_VALUE(fs_info->csum_size, calculated_csum));
671 return;
672 }
673 if (stripe->sectors[sector_nr].generation !=
674 btrfs_stack_header_generation(header)) {
675 bitmap_set(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
676 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
677 btrfs_warn_rl(fs_info,
678 "tree block %llu mirror %u has bad generation, has %llu want %llu",
679 logical, stripe->mirror_num,
680 btrfs_stack_header_generation(header),
681 stripe->sectors[sector_nr].generation);
682 return;
683 }
684 bitmap_clear(&stripe->error_bitmap, sector_nr, sectors_per_tree);
685 bitmap_clear(&stripe->csum_error_bitmap, sector_nr, sectors_per_tree);
686 bitmap_clear(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
687}
688
689static void scrub_verify_one_sector(struct scrub_stripe *stripe, int sector_nr)
690{
691 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
692 struct scrub_sector_verification *sector = &stripe->sectors[sector_nr];
693 const u32 sectors_per_tree = fs_info->nodesize >> fs_info->sectorsize_bits;
694 struct page *page = scrub_stripe_get_page(stripe, sector_nr);
695 unsigned int pgoff = scrub_stripe_get_page_offset(stripe, sector_nr);
696 u8 csum_buf[BTRFS_CSUM_SIZE];
697 int ret;
698
699 ASSERT(sector_nr >= 0 && sector_nr < stripe->nr_sectors);
700
701 /* Sector not utilized, skip it. */
702 if (!test_bit(sector_nr, &stripe->extent_sector_bitmap))
703 return;
704
705 /* IO error, no need to check. */
706 if (test_bit(sector_nr, &stripe->io_error_bitmap))
707 return;
708
709 /* Metadata, verify the full tree block. */
710 if (sector->is_metadata) {
711 /*
712 * Check if the tree block crosses the stripe boundary. If
713 * crossed the boundary, we cannot verify it but only give a
714 * warning.
715 *
716 * This can only happen on a very old filesystem where chunks
717 * are not ensured to be stripe aligned.
718 */
719 if (unlikely(sector_nr + sectors_per_tree > stripe->nr_sectors)) {
720 btrfs_warn_rl(fs_info,
721 "tree block at %llu crosses stripe boundary %llu",
722 stripe->logical +
723 (sector_nr << fs_info->sectorsize_bits),
724 stripe->logical);
725 return;
726 }
727 scrub_verify_one_metadata(stripe, sector_nr);
728 return;
729 }
730
731 /*
732 * Data is easier, we just verify the data csum (if we have it). For
733 * cases without csum, we have no other choice but to trust it.
734 */
735 if (!sector->csum) {
736 clear_bit(sector_nr, &stripe->error_bitmap);
737 return;
738 }
739
740 ret = btrfs_check_sector_csum(fs_info, page, pgoff, csum_buf, sector->csum);
741 if (ret < 0) {
742 set_bit(sector_nr, &stripe->csum_error_bitmap);
743 set_bit(sector_nr, &stripe->error_bitmap);
744 } else {
745 clear_bit(sector_nr, &stripe->csum_error_bitmap);
746 clear_bit(sector_nr, &stripe->error_bitmap);
747 }
748}
749
750/* Verify specified sectors of a stripe. */
751static void scrub_verify_one_stripe(struct scrub_stripe *stripe, unsigned long bitmap)
752{
753 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
754 const u32 sectors_per_tree = fs_info->nodesize >> fs_info->sectorsize_bits;
755 int sector_nr;
756
757 for_each_set_bit(sector_nr, &bitmap, stripe->nr_sectors) {
758 scrub_verify_one_sector(stripe, sector_nr);
759 if (stripe->sectors[sector_nr].is_metadata)
760 sector_nr += sectors_per_tree - 1;
761 }
762}
763
764static int calc_sector_number(struct scrub_stripe *stripe, struct bio_vec *first_bvec)
765{
766 int i;
767
768 for (i = 0; i < stripe->nr_sectors; i++) {
769 if (scrub_stripe_get_page(stripe, i) == first_bvec->bv_page &&
770 scrub_stripe_get_page_offset(stripe, i) == first_bvec->bv_offset)
771 break;
772 }
773 ASSERT(i < stripe->nr_sectors);
774 return i;
775}
776
777/*
778 * Repair read is different to the regular read:
779 *
780 * - Only reads the failed sectors
781 * - May have extra blocksize limits
782 */
783static void scrub_repair_read_endio(struct btrfs_bio *bbio)
784{
785 struct scrub_stripe *stripe = bbio->private;
786 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
787 struct bio_vec *bvec;
788 int sector_nr = calc_sector_number(stripe, bio_first_bvec_all(&bbio->bio));
789 u32 bio_size = 0;
790 int i;
791
792 ASSERT(sector_nr < stripe->nr_sectors);
793
794 bio_for_each_bvec_all(bvec, &bbio->bio, i)
795 bio_size += bvec->bv_len;
796
797 if (bbio->bio.bi_status) {
798 bitmap_set(&stripe->io_error_bitmap, sector_nr,
799 bio_size >> fs_info->sectorsize_bits);
800 bitmap_set(&stripe->error_bitmap, sector_nr,
801 bio_size >> fs_info->sectorsize_bits);
802 } else {
803 bitmap_clear(&stripe->io_error_bitmap, sector_nr,
804 bio_size >> fs_info->sectorsize_bits);
805 }
806 bio_put(&bbio->bio);
807 if (atomic_dec_and_test(&stripe->pending_io))
808 wake_up(&stripe->io_wait);
809}
810
811static int calc_next_mirror(int mirror, int num_copies)
812{
813 ASSERT(mirror <= num_copies);
814 return (mirror + 1 > num_copies) ? 1 : mirror + 1;
815}
816
817static void scrub_stripe_submit_repair_read(struct scrub_stripe *stripe,
818 int mirror, int blocksize, bool wait)
819{
820 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
821 struct btrfs_bio *bbio = NULL;
822 const unsigned long old_error_bitmap = stripe->error_bitmap;
823 int i;
824
825 ASSERT(stripe->mirror_num >= 1);
826 ASSERT(atomic_read(&stripe->pending_io) == 0);
827
828 for_each_set_bit(i, &old_error_bitmap, stripe->nr_sectors) {
829 struct page *page;
830 int pgoff;
831 int ret;
832
833 page = scrub_stripe_get_page(stripe, i);
834 pgoff = scrub_stripe_get_page_offset(stripe, i);
835
836 /* The current sector cannot be merged, submit the bio. */
837 if (bbio && ((i > 0 && !test_bit(i - 1, &stripe->error_bitmap)) ||
838 bbio->bio.bi_iter.bi_size >= blocksize)) {
839 ASSERT(bbio->bio.bi_iter.bi_size);
840 atomic_inc(&stripe->pending_io);
841 btrfs_submit_bio(bbio, mirror);
842 if (wait)
843 wait_scrub_stripe_io(stripe);
844 bbio = NULL;
845 }
846
847 if (!bbio) {
848 bbio = btrfs_bio_alloc(stripe->nr_sectors, REQ_OP_READ,
849 fs_info, scrub_repair_read_endio, stripe);
850 bbio->bio.bi_iter.bi_sector = (stripe->logical +
851 (i << fs_info->sectorsize_bits)) >> SECTOR_SHIFT;
852 }
853
854 ret = bio_add_page(&bbio->bio, page, fs_info->sectorsize, pgoff);
855 ASSERT(ret == fs_info->sectorsize);
856 }
857 if (bbio) {
858 ASSERT(bbio->bio.bi_iter.bi_size);
859 atomic_inc(&stripe->pending_io);
860 btrfs_submit_bio(bbio, mirror);
861 if (wait)
862 wait_scrub_stripe_io(stripe);
863 }
864}
865
866static void scrub_stripe_report_errors(struct scrub_ctx *sctx,
867 struct scrub_stripe *stripe)
868{
869 static DEFINE_RATELIMIT_STATE(rs, DEFAULT_RATELIMIT_INTERVAL,
870 DEFAULT_RATELIMIT_BURST);
871 struct btrfs_fs_info *fs_info = sctx->fs_info;
872 struct btrfs_device *dev = NULL;
873 u64 physical = 0;
874 int nr_data_sectors = 0;
875 int nr_meta_sectors = 0;
876 int nr_nodatacsum_sectors = 0;
877 int nr_repaired_sectors = 0;
878 int sector_nr;
879
880 if (test_bit(SCRUB_STRIPE_FLAG_NO_REPORT, &stripe->state))
881 return;
882
883 /*
884 * Init needed infos for error reporting.
885 *
886 * Although our scrub_stripe infrastructure is mostly based on btrfs_submit_bio()
887 * thus no need for dev/physical, error reporting still needs dev and physical.
888 */
889 if (!bitmap_empty(&stripe->init_error_bitmap, stripe->nr_sectors)) {
890 u64 mapped_len = fs_info->sectorsize;
891 struct btrfs_io_context *bioc = NULL;
892 int stripe_index = stripe->mirror_num - 1;
893 int ret;
894
895 /* For scrub, our mirror_num should always start at 1. */
896 ASSERT(stripe->mirror_num >= 1);
897 ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
898 stripe->logical, &mapped_len, &bioc,
899 NULL, NULL);
900 /*
901 * If we failed, dev will be NULL, and later detailed reports
902 * will just be skipped.
903 */
904 if (ret < 0)
905 goto skip;
906 physical = bioc->stripes[stripe_index].physical;
907 dev = bioc->stripes[stripe_index].dev;
908 btrfs_put_bioc(bioc);
909 }
910
911skip:
912 for_each_set_bit(sector_nr, &stripe->extent_sector_bitmap, stripe->nr_sectors) {
913 bool repaired = false;
914
915 if (stripe->sectors[sector_nr].is_metadata) {
916 nr_meta_sectors++;
917 } else {
918 nr_data_sectors++;
919 if (!stripe->sectors[sector_nr].csum)
920 nr_nodatacsum_sectors++;
921 }
922
923 if (test_bit(sector_nr, &stripe->init_error_bitmap) &&
924 !test_bit(sector_nr, &stripe->error_bitmap)) {
925 nr_repaired_sectors++;
926 repaired = true;
927 }
928
929 /* Good sector from the beginning, nothing need to be done. */
930 if (!test_bit(sector_nr, &stripe->init_error_bitmap))
931 continue;
932
933 /*
934 * Report error for the corrupted sectors. If repaired, just
935 * output the message of repaired message.
936 */
937 if (repaired) {
938 if (dev) {
939 btrfs_err_rl_in_rcu(fs_info,
940 "fixed up error at logical %llu on dev %s physical %llu",
941 stripe->logical, btrfs_dev_name(dev),
942 physical);
943 } else {
944 btrfs_err_rl_in_rcu(fs_info,
945 "fixed up error at logical %llu on mirror %u",
946 stripe->logical, stripe->mirror_num);
947 }
948 continue;
949 }
950
951 /* The remaining are all for unrepaired. */
952 if (dev) {
953 btrfs_err_rl_in_rcu(fs_info,
954 "unable to fixup (regular) error at logical %llu on dev %s physical %llu",
955 stripe->logical, btrfs_dev_name(dev),
956 physical);
957 } else {
958 btrfs_err_rl_in_rcu(fs_info,
959 "unable to fixup (regular) error at logical %llu on mirror %u",
960 stripe->logical, stripe->mirror_num);
961 }
962
963 if (test_bit(sector_nr, &stripe->io_error_bitmap))
964 if (__ratelimit(&rs) && dev)
965 scrub_print_common_warning("i/o error", dev, false,
966 stripe->logical, physical);
967 if (test_bit(sector_nr, &stripe->csum_error_bitmap))
968 if (__ratelimit(&rs) && dev)
969 scrub_print_common_warning("checksum error", dev, false,
970 stripe->logical, physical);
971 if (test_bit(sector_nr, &stripe->meta_error_bitmap))
972 if (__ratelimit(&rs) && dev)
973 scrub_print_common_warning("header error", dev, false,
974 stripe->logical, physical);
975 }
976
977 spin_lock(&sctx->stat_lock);
978 sctx->stat.data_extents_scrubbed += stripe->nr_data_extents;
979 sctx->stat.tree_extents_scrubbed += stripe->nr_meta_extents;
980 sctx->stat.data_bytes_scrubbed += nr_data_sectors << fs_info->sectorsize_bits;
981 sctx->stat.tree_bytes_scrubbed += nr_meta_sectors << fs_info->sectorsize_bits;
982 sctx->stat.no_csum += nr_nodatacsum_sectors;
983 sctx->stat.read_errors += stripe->init_nr_io_errors;
984 sctx->stat.csum_errors += stripe->init_nr_csum_errors;
985 sctx->stat.verify_errors += stripe->init_nr_meta_errors;
986 sctx->stat.uncorrectable_errors +=
987 bitmap_weight(&stripe->error_bitmap, stripe->nr_sectors);
988 sctx->stat.corrected_errors += nr_repaired_sectors;
989 spin_unlock(&sctx->stat_lock);
990}
991
992static void scrub_write_sectors(struct scrub_ctx *sctx, struct scrub_stripe *stripe,
993 unsigned long write_bitmap, bool dev_replace);
994
995/*
996 * The main entrance for all read related scrub work, including:
997 *
998 * - Wait for the initial read to finish
999 * - Verify and locate any bad sectors
1000 * - Go through the remaining mirrors and try to read as large blocksize as
1001 * possible
1002 * - Go through all mirrors (including the failed mirror) sector-by-sector
1003 * - Submit writeback for repaired sectors
1004 *
1005 * Writeback for dev-replace does not happen here, it needs extra
1006 * synchronization for zoned devices.
1007 */
1008static void scrub_stripe_read_repair_worker(struct work_struct *work)
1009{
1010 struct scrub_stripe *stripe = container_of(work, struct scrub_stripe, work);
1011 struct scrub_ctx *sctx = stripe->sctx;
1012 struct btrfs_fs_info *fs_info = sctx->fs_info;
1013 int num_copies = btrfs_num_copies(fs_info, stripe->bg->start,
1014 stripe->bg->length);
1015 int mirror;
1016 int i;
1017
1018 ASSERT(stripe->mirror_num > 0);
1019
1020 wait_scrub_stripe_io(stripe);
1021 scrub_verify_one_stripe(stripe, stripe->extent_sector_bitmap);
1022 /* Save the initial failed bitmap for later repair and report usage. */
1023 stripe->init_error_bitmap = stripe->error_bitmap;
1024 stripe->init_nr_io_errors = bitmap_weight(&stripe->io_error_bitmap,
1025 stripe->nr_sectors);
1026 stripe->init_nr_csum_errors = bitmap_weight(&stripe->csum_error_bitmap,
1027 stripe->nr_sectors);
1028 stripe->init_nr_meta_errors = bitmap_weight(&stripe->meta_error_bitmap,
1029 stripe->nr_sectors);
1030
1031 if (bitmap_empty(&stripe->init_error_bitmap, stripe->nr_sectors))
1032 goto out;
1033
1034 /*
1035 * Try all remaining mirrors.
1036 *
1037 * Here we still try to read as large block as possible, as this is
1038 * faster and we have extra safety nets to rely on.
1039 */
1040 for (mirror = calc_next_mirror(stripe->mirror_num, num_copies);
1041 mirror != stripe->mirror_num;
1042 mirror = calc_next_mirror(mirror, num_copies)) {
1043 const unsigned long old_error_bitmap = stripe->error_bitmap;
1044
1045 scrub_stripe_submit_repair_read(stripe, mirror,
1046 BTRFS_STRIPE_LEN, false);
1047 wait_scrub_stripe_io(stripe);
1048 scrub_verify_one_stripe(stripe, old_error_bitmap);
1049 if (bitmap_empty(&stripe->error_bitmap, stripe->nr_sectors))
1050 goto out;
1051 }
1052
1053 /*
1054 * Last safety net, try re-checking all mirrors, including the failed
1055 * one, sector-by-sector.
1056 *
1057 * As if one sector failed the drive's internal csum, the whole read
1058 * containing the offending sector would be marked as error.
1059 * Thus here we do sector-by-sector read.
1060 *
1061 * This can be slow, thus we only try it as the last resort.
1062 */
1063
1064 for (i = 0, mirror = stripe->mirror_num;
1065 i < num_copies;
1066 i++, mirror = calc_next_mirror(mirror, num_copies)) {
1067 const unsigned long old_error_bitmap = stripe->error_bitmap;
1068
1069 scrub_stripe_submit_repair_read(stripe, mirror,
1070 fs_info->sectorsize, true);
1071 wait_scrub_stripe_io(stripe);
1072 scrub_verify_one_stripe(stripe, old_error_bitmap);
1073 if (bitmap_empty(&stripe->error_bitmap, stripe->nr_sectors))
1074 goto out;
1075 }
1076out:
1077 /*
1078 * Submit the repaired sectors. For zoned case, we cannot do repair
1079 * in-place, but queue the bg to be relocated.
1080 */
1081 if (btrfs_is_zoned(fs_info)) {
1082 if (!bitmap_empty(&stripe->error_bitmap, stripe->nr_sectors))
1083 btrfs_repair_one_zone(fs_info, sctx->stripes[0].bg->start);
1084 } else if (!sctx->readonly) {
1085 unsigned long repaired;
1086
1087 bitmap_andnot(&repaired, &stripe->init_error_bitmap,
1088 &stripe->error_bitmap, stripe->nr_sectors);
1089 scrub_write_sectors(sctx, stripe, repaired, false);
1090 wait_scrub_stripe_io(stripe);
1091 }
1092
1093 scrub_stripe_report_errors(sctx, stripe);
1094 set_bit(SCRUB_STRIPE_FLAG_REPAIR_DONE, &stripe->state);
1095 wake_up(&stripe->repair_wait);
1096}
1097
1098static void scrub_read_endio(struct btrfs_bio *bbio)
1099{
1100 struct scrub_stripe *stripe = bbio->private;
1101 struct bio_vec *bvec;
1102 int sector_nr = calc_sector_number(stripe, bio_first_bvec_all(&bbio->bio));
1103 int num_sectors;
1104 u32 bio_size = 0;
1105 int i;
1106
1107 ASSERT(sector_nr < stripe->nr_sectors);
1108 bio_for_each_bvec_all(bvec, &bbio->bio, i)
1109 bio_size += bvec->bv_len;
1110 num_sectors = bio_size >> stripe->bg->fs_info->sectorsize_bits;
1111
1112 if (bbio->bio.bi_status) {
1113 bitmap_set(&stripe->io_error_bitmap, sector_nr, num_sectors);
1114 bitmap_set(&stripe->error_bitmap, sector_nr, num_sectors);
1115 } else {
1116 bitmap_clear(&stripe->io_error_bitmap, sector_nr, num_sectors);
1117 }
1118 bio_put(&bbio->bio);
1119 if (atomic_dec_and_test(&stripe->pending_io)) {
1120 wake_up(&stripe->io_wait);
1121 INIT_WORK(&stripe->work, scrub_stripe_read_repair_worker);
1122 queue_work(stripe->bg->fs_info->scrub_workers, &stripe->work);
1123 }
1124}
1125
1126static void scrub_write_endio(struct btrfs_bio *bbio)
1127{
1128 struct scrub_stripe *stripe = bbio->private;
1129 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
1130 struct bio_vec *bvec;
1131 int sector_nr = calc_sector_number(stripe, bio_first_bvec_all(&bbio->bio));
1132 u32 bio_size = 0;
1133 int i;
1134
1135 bio_for_each_bvec_all(bvec, &bbio->bio, i)
1136 bio_size += bvec->bv_len;
1137
1138 if (bbio->bio.bi_status) {
1139 unsigned long flags;
1140
1141 spin_lock_irqsave(&stripe->write_error_lock, flags);
1142 bitmap_set(&stripe->write_error_bitmap, sector_nr,
1143 bio_size >> fs_info->sectorsize_bits);
1144 spin_unlock_irqrestore(&stripe->write_error_lock, flags);
1145 }
1146 bio_put(&bbio->bio);
1147
1148 if (atomic_dec_and_test(&stripe->pending_io))
1149 wake_up(&stripe->io_wait);
1150}
1151
1152static void scrub_submit_write_bio(struct scrub_ctx *sctx,
1153 struct scrub_stripe *stripe,
1154 struct btrfs_bio *bbio, bool dev_replace)
1155{
1156 struct btrfs_fs_info *fs_info = sctx->fs_info;
1157 u32 bio_len = bbio->bio.bi_iter.bi_size;
1158 u32 bio_off = (bbio->bio.bi_iter.bi_sector << SECTOR_SHIFT) -
1159 stripe->logical;
1160
1161 fill_writer_pointer_gap(sctx, stripe->physical + bio_off);
1162 atomic_inc(&stripe->pending_io);
1163 btrfs_submit_repair_write(bbio, stripe->mirror_num, dev_replace);
1164 if (!btrfs_is_zoned(fs_info))
1165 return;
1166 /*
1167 * For zoned writeback, queue depth must be 1, thus we must wait for
1168 * the write to finish before the next write.
1169 */
1170 wait_scrub_stripe_io(stripe);
1171
1172 /*
1173 * And also need to update the write pointer if write finished
1174 * successfully.
1175 */
1176 if (!test_bit(bio_off >> fs_info->sectorsize_bits,
1177 &stripe->write_error_bitmap))
1178 sctx->write_pointer += bio_len;
1179}
1180
1181/*
1182 * Submit the write bio(s) for the sectors specified by @write_bitmap.
1183 *
1184 * Here we utilize btrfs_submit_repair_write(), which has some extra benefits:
1185 *
1186 * - Only needs logical bytenr and mirror_num
1187 * Just like the scrub read path
1188 *
1189 * - Would only result in writes to the specified mirror
1190 * Unlike the regular writeback path, which would write back to all stripes
1191 *
1192 * - Handle dev-replace and read-repair writeback differently
1193 */
1194static void scrub_write_sectors(struct scrub_ctx *sctx, struct scrub_stripe *stripe,
1195 unsigned long write_bitmap, bool dev_replace)
1196{
1197 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
1198 struct btrfs_bio *bbio = NULL;
1199 int sector_nr;
1200
1201 for_each_set_bit(sector_nr, &write_bitmap, stripe->nr_sectors) {
1202 struct page *page = scrub_stripe_get_page(stripe, sector_nr);
1203 unsigned int pgoff = scrub_stripe_get_page_offset(stripe, sector_nr);
1204 int ret;
1205
1206 /* We should only writeback sectors covered by an extent. */
1207 ASSERT(test_bit(sector_nr, &stripe->extent_sector_bitmap));
1208
1209 /* Cannot merge with previous sector, submit the current one. */
1210 if (bbio && sector_nr && !test_bit(sector_nr - 1, &write_bitmap)) {
1211 scrub_submit_write_bio(sctx, stripe, bbio, dev_replace);
1212 bbio = NULL;
1213 }
1214 if (!bbio) {
1215 bbio = btrfs_bio_alloc(stripe->nr_sectors, REQ_OP_WRITE,
1216 fs_info, scrub_write_endio, stripe);
1217 bbio->bio.bi_iter.bi_sector = (stripe->logical +
1218 (sector_nr << fs_info->sectorsize_bits)) >>
1219 SECTOR_SHIFT;
1220 }
1221 ret = bio_add_page(&bbio->bio, page, fs_info->sectorsize, pgoff);
1222 ASSERT(ret == fs_info->sectorsize);
1223 }
1224 if (bbio)
1225 scrub_submit_write_bio(sctx, stripe, bbio, dev_replace);
1226}
1227
1228/*
1229 * Throttling of IO submission, bandwidth-limit based, the timeslice is 1
1230 * second. Limit can be set via /sys/fs/UUID/devinfo/devid/scrub_speed_max.
1231 */
1232static void scrub_throttle_dev_io(struct scrub_ctx *sctx, struct btrfs_device *device,
1233 unsigned int bio_size)
1234{
1235 const int time_slice = 1000;
1236 s64 delta;
1237 ktime_t now;
1238 u32 div;
1239 u64 bwlimit;
1240
1241 bwlimit = READ_ONCE(device->scrub_speed_max);
1242 if (bwlimit == 0)
1243 return;
1244
1245 /*
1246 * Slice is divided into intervals when the IO is submitted, adjust by
1247 * bwlimit and maximum of 64 intervals.
1248 */
1249 div = max_t(u32, 1, (u32)(bwlimit / (16 * 1024 * 1024)));
1250 div = min_t(u32, 64, div);
1251
1252 /* Start new epoch, set deadline */
1253 now = ktime_get();
1254 if (sctx->throttle_deadline == 0) {
1255 sctx->throttle_deadline = ktime_add_ms(now, time_slice / div);
1256 sctx->throttle_sent = 0;
1257 }
1258
1259 /* Still in the time to send? */
1260 if (ktime_before(now, sctx->throttle_deadline)) {
1261 /* If current bio is within the limit, send it */
1262 sctx->throttle_sent += bio_size;
1263 if (sctx->throttle_sent <= div_u64(bwlimit, div))
1264 return;
1265
1266 /* We're over the limit, sleep until the rest of the slice */
1267 delta = ktime_ms_delta(sctx->throttle_deadline, now);
1268 } else {
1269 /* New request after deadline, start new epoch */
1270 delta = 0;
1271 }
1272
1273 if (delta) {
1274 long timeout;
1275
1276 timeout = div_u64(delta * HZ, 1000);
1277 schedule_timeout_interruptible(timeout);
1278 }
1279
1280 /* Next call will start the deadline period */
1281 sctx->throttle_deadline = 0;
1282}
1283
1284/*
1285 * Given a physical address, this will calculate it's
1286 * logical offset. if this is a parity stripe, it will return
1287 * the most left data stripe's logical offset.
1288 *
1289 * return 0 if it is a data stripe, 1 means parity stripe.
1290 */
1291static int get_raid56_logic_offset(u64 physical, int num,
1292 struct btrfs_chunk_map *map, u64 *offset,
1293 u64 *stripe_start)
1294{
1295 int i;
1296 int j = 0;
1297 u64 last_offset;
1298 const int data_stripes = nr_data_stripes(map);
1299
1300 last_offset = (physical - map->stripes[num].physical) * data_stripes;
1301 if (stripe_start)
1302 *stripe_start = last_offset;
1303
1304 *offset = last_offset;
1305 for (i = 0; i < data_stripes; i++) {
1306 u32 stripe_nr;
1307 u32 stripe_index;
1308 u32 rot;
1309
1310 *offset = last_offset + btrfs_stripe_nr_to_offset(i);
1311
1312 stripe_nr = (u32)(*offset >> BTRFS_STRIPE_LEN_SHIFT) / data_stripes;
1313
1314 /* Work out the disk rotation on this stripe-set */
1315 rot = stripe_nr % map->num_stripes;
1316 /* calculate which stripe this data locates */
1317 rot += i;
1318 stripe_index = rot % map->num_stripes;
1319 if (stripe_index == num)
1320 return 0;
1321 if (stripe_index < num)
1322 j++;
1323 }
1324 *offset = last_offset + btrfs_stripe_nr_to_offset(j);
1325 return 1;
1326}
1327
1328/*
1329 * Return 0 if the extent item range covers any byte of the range.
1330 * Return <0 if the extent item is before @search_start.
1331 * Return >0 if the extent item is after @start_start + @search_len.
1332 */
1333static int compare_extent_item_range(struct btrfs_path *path,
1334 u64 search_start, u64 search_len)
1335{
1336 struct btrfs_fs_info *fs_info = path->nodes[0]->fs_info;
1337 u64 len;
1338 struct btrfs_key key;
1339
1340 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1341 ASSERT(key.type == BTRFS_EXTENT_ITEM_KEY ||
1342 key.type == BTRFS_METADATA_ITEM_KEY);
1343 if (key.type == BTRFS_METADATA_ITEM_KEY)
1344 len = fs_info->nodesize;
1345 else
1346 len = key.offset;
1347
1348 if (key.objectid + len <= search_start)
1349 return -1;
1350 if (key.objectid >= search_start + search_len)
1351 return 1;
1352 return 0;
1353}
1354
1355/*
1356 * Locate one extent item which covers any byte in range
1357 * [@search_start, @search_start + @search_length)
1358 *
1359 * If the path is not initialized, we will initialize the search by doing
1360 * a btrfs_search_slot().
1361 * If the path is already initialized, we will use the path as the initial
1362 * slot, to avoid duplicated btrfs_search_slot() calls.
1363 *
1364 * NOTE: If an extent item starts before @search_start, we will still
1365 * return the extent item. This is for data extent crossing stripe boundary.
1366 *
1367 * Return 0 if we found such extent item, and @path will point to the extent item.
1368 * Return >0 if no such extent item can be found, and @path will be released.
1369 * Return <0 if hit fatal error, and @path will be released.
1370 */
1371static int find_first_extent_item(struct btrfs_root *extent_root,
1372 struct btrfs_path *path,
1373 u64 search_start, u64 search_len)
1374{
1375 struct btrfs_fs_info *fs_info = extent_root->fs_info;
1376 struct btrfs_key key;
1377 int ret;
1378
1379 /* Continue using the existing path */
1380 if (path->nodes[0])
1381 goto search_forward;
1382
1383 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1384 key.type = BTRFS_METADATA_ITEM_KEY;
1385 else
1386 key.type = BTRFS_EXTENT_ITEM_KEY;
1387 key.objectid = search_start;
1388 key.offset = (u64)-1;
1389
1390 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
1391 if (ret < 0)
1392 return ret;
1393
1394 ASSERT(ret > 0);
1395 /*
1396 * Here we intentionally pass 0 as @min_objectid, as there could be
1397 * an extent item starting before @search_start.
1398 */
1399 ret = btrfs_previous_extent_item(extent_root, path, 0);
1400 if (ret < 0)
1401 return ret;
1402 /*
1403 * No matter whether we have found an extent item, the next loop will
1404 * properly do every check on the key.
1405 */
1406search_forward:
1407 while (true) {
1408 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1409 if (key.objectid >= search_start + search_len)
1410 break;
1411 if (key.type != BTRFS_METADATA_ITEM_KEY &&
1412 key.type != BTRFS_EXTENT_ITEM_KEY)
1413 goto next;
1414
1415 ret = compare_extent_item_range(path, search_start, search_len);
1416 if (ret == 0)
1417 return ret;
1418 if (ret > 0)
1419 break;
1420next:
1421 ret = btrfs_next_item(extent_root, path);
1422 if (ret) {
1423 /* Either no more items or a fatal error. */
1424 btrfs_release_path(path);
1425 return ret;
1426 }
1427 }
1428 btrfs_release_path(path);
1429 return 1;
1430}
1431
1432static void get_extent_info(struct btrfs_path *path, u64 *extent_start_ret,
1433 u64 *size_ret, u64 *flags_ret, u64 *generation_ret)
1434{
1435 struct btrfs_key key;
1436 struct btrfs_extent_item *ei;
1437
1438 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1439 ASSERT(key.type == BTRFS_METADATA_ITEM_KEY ||
1440 key.type == BTRFS_EXTENT_ITEM_KEY);
1441 *extent_start_ret = key.objectid;
1442 if (key.type == BTRFS_METADATA_ITEM_KEY)
1443 *size_ret = path->nodes[0]->fs_info->nodesize;
1444 else
1445 *size_ret = key.offset;
1446 ei = btrfs_item_ptr(path->nodes[0], path->slots[0], struct btrfs_extent_item);
1447 *flags_ret = btrfs_extent_flags(path->nodes[0], ei);
1448 *generation_ret = btrfs_extent_generation(path->nodes[0], ei);
1449}
1450
1451static int sync_write_pointer_for_zoned(struct scrub_ctx *sctx, u64 logical,
1452 u64 physical, u64 physical_end)
1453{
1454 struct btrfs_fs_info *fs_info = sctx->fs_info;
1455 int ret = 0;
1456
1457 if (!btrfs_is_zoned(fs_info))
1458 return 0;
1459
1460 mutex_lock(&sctx->wr_lock);
1461 if (sctx->write_pointer < physical_end) {
1462 ret = btrfs_sync_zone_write_pointer(sctx->wr_tgtdev, logical,
1463 physical,
1464 sctx->write_pointer);
1465 if (ret)
1466 btrfs_err(fs_info,
1467 "zoned: failed to recover write pointer");
1468 }
1469 mutex_unlock(&sctx->wr_lock);
1470 btrfs_dev_clear_zone_empty(sctx->wr_tgtdev, physical);
1471
1472 return ret;
1473}
1474
1475static void fill_one_extent_info(struct btrfs_fs_info *fs_info,
1476 struct scrub_stripe *stripe,
1477 u64 extent_start, u64 extent_len,
1478 u64 extent_flags, u64 extent_gen)
1479{
1480 for (u64 cur_logical = max(stripe->logical, extent_start);
1481 cur_logical < min(stripe->logical + BTRFS_STRIPE_LEN,
1482 extent_start + extent_len);
1483 cur_logical += fs_info->sectorsize) {
1484 const int nr_sector = (cur_logical - stripe->logical) >>
1485 fs_info->sectorsize_bits;
1486 struct scrub_sector_verification *sector =
1487 &stripe->sectors[nr_sector];
1488
1489 set_bit(nr_sector, &stripe->extent_sector_bitmap);
1490 if (extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1491 sector->is_metadata = true;
1492 sector->generation = extent_gen;
1493 }
1494 }
1495}
1496
1497static void scrub_stripe_reset_bitmaps(struct scrub_stripe *stripe)
1498{
1499 stripe->extent_sector_bitmap = 0;
1500 stripe->init_error_bitmap = 0;
1501 stripe->init_nr_io_errors = 0;
1502 stripe->init_nr_csum_errors = 0;
1503 stripe->init_nr_meta_errors = 0;
1504 stripe->error_bitmap = 0;
1505 stripe->io_error_bitmap = 0;
1506 stripe->csum_error_bitmap = 0;
1507 stripe->meta_error_bitmap = 0;
1508}
1509
1510/*
1511 * Locate one stripe which has at least one extent in its range.
1512 *
1513 * Return 0 if found such stripe, and store its info into @stripe.
1514 * Return >0 if there is no such stripe in the specified range.
1515 * Return <0 for error.
1516 */
1517static int scrub_find_fill_first_stripe(struct btrfs_block_group *bg,
1518 struct btrfs_path *extent_path,
1519 struct btrfs_path *csum_path,
1520 struct btrfs_device *dev, u64 physical,
1521 int mirror_num, u64 logical_start,
1522 u32 logical_len,
1523 struct scrub_stripe *stripe)
1524{
1525 struct btrfs_fs_info *fs_info = bg->fs_info;
1526 struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bg->start);
1527 struct btrfs_root *csum_root = btrfs_csum_root(fs_info, bg->start);
1528 const u64 logical_end = logical_start + logical_len;
1529 u64 cur_logical = logical_start;
1530 u64 stripe_end;
1531 u64 extent_start;
1532 u64 extent_len;
1533 u64 extent_flags;
1534 u64 extent_gen;
1535 int ret;
1536
1537 memset(stripe->sectors, 0, sizeof(struct scrub_sector_verification) *
1538 stripe->nr_sectors);
1539 scrub_stripe_reset_bitmaps(stripe);
1540
1541 /* The range must be inside the bg. */
1542 ASSERT(logical_start >= bg->start && logical_end <= bg->start + bg->length);
1543
1544 ret = find_first_extent_item(extent_root, extent_path, logical_start,
1545 logical_len);
1546 /* Either error or not found. */
1547 if (ret)
1548 goto out;
1549 get_extent_info(extent_path, &extent_start, &extent_len, &extent_flags,
1550 &extent_gen);
1551 if (extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1552 stripe->nr_meta_extents++;
1553 if (extent_flags & BTRFS_EXTENT_FLAG_DATA)
1554 stripe->nr_data_extents++;
1555 cur_logical = max(extent_start, cur_logical);
1556
1557 /*
1558 * Round down to stripe boundary.
1559 *
1560 * The extra calculation against bg->start is to handle block groups
1561 * whose logical bytenr is not BTRFS_STRIPE_LEN aligned.
1562 */
1563 stripe->logical = round_down(cur_logical - bg->start, BTRFS_STRIPE_LEN) +
1564 bg->start;
1565 stripe->physical = physical + stripe->logical - logical_start;
1566 stripe->dev = dev;
1567 stripe->bg = bg;
1568 stripe->mirror_num = mirror_num;
1569 stripe_end = stripe->logical + BTRFS_STRIPE_LEN - 1;
1570
1571 /* Fill the first extent info into stripe->sectors[] array. */
1572 fill_one_extent_info(fs_info, stripe, extent_start, extent_len,
1573 extent_flags, extent_gen);
1574 cur_logical = extent_start + extent_len;
1575
1576 /* Fill the extent info for the remaining sectors. */
1577 while (cur_logical <= stripe_end) {
1578 ret = find_first_extent_item(extent_root, extent_path, cur_logical,
1579 stripe_end - cur_logical + 1);
1580 if (ret < 0)
1581 goto out;
1582 if (ret > 0) {
1583 ret = 0;
1584 break;
1585 }
1586 get_extent_info(extent_path, &extent_start, &extent_len,
1587 &extent_flags, &extent_gen);
1588 if (extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1589 stripe->nr_meta_extents++;
1590 if (extent_flags & BTRFS_EXTENT_FLAG_DATA)
1591 stripe->nr_data_extents++;
1592 fill_one_extent_info(fs_info, stripe, extent_start, extent_len,
1593 extent_flags, extent_gen);
1594 cur_logical = extent_start + extent_len;
1595 }
1596
1597 /* Now fill the data csum. */
1598 if (bg->flags & BTRFS_BLOCK_GROUP_DATA) {
1599 int sector_nr;
1600 unsigned long csum_bitmap = 0;
1601
1602 /* Csum space should have already been allocated. */
1603 ASSERT(stripe->csums);
1604
1605 /*
1606 * Our csum bitmap should be large enough, as BTRFS_STRIPE_LEN
1607 * should contain at most 16 sectors.
1608 */
1609 ASSERT(BITS_PER_LONG >= BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits);
1610
1611 ret = btrfs_lookup_csums_bitmap(csum_root, csum_path,
1612 stripe->logical, stripe_end,
1613 stripe->csums, &csum_bitmap);
1614 if (ret < 0)
1615 goto out;
1616 if (ret > 0)
1617 ret = 0;
1618
1619 for_each_set_bit(sector_nr, &csum_bitmap, stripe->nr_sectors) {
1620 stripe->sectors[sector_nr].csum = stripe->csums +
1621 sector_nr * fs_info->csum_size;
1622 }
1623 }
1624 set_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &stripe->state);
1625out:
1626 return ret;
1627}
1628
1629static void scrub_reset_stripe(struct scrub_stripe *stripe)
1630{
1631 scrub_stripe_reset_bitmaps(stripe);
1632
1633 stripe->nr_meta_extents = 0;
1634 stripe->nr_data_extents = 0;
1635 stripe->state = 0;
1636
1637 for (int i = 0; i < stripe->nr_sectors; i++) {
1638 stripe->sectors[i].is_metadata = false;
1639 stripe->sectors[i].csum = NULL;
1640 stripe->sectors[i].generation = 0;
1641 }
1642}
1643
1644static void scrub_submit_extent_sector_read(struct scrub_ctx *sctx,
1645 struct scrub_stripe *stripe)
1646{
1647 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
1648 struct btrfs_bio *bbio = NULL;
1649 unsigned int nr_sectors = min(BTRFS_STRIPE_LEN, stripe->bg->start +
1650 stripe->bg->length - stripe->logical) >>
1651 fs_info->sectorsize_bits;
1652 u64 stripe_len = BTRFS_STRIPE_LEN;
1653 int mirror = stripe->mirror_num;
1654 int i;
1655
1656 atomic_inc(&stripe->pending_io);
1657
1658 for_each_set_bit(i, &stripe->extent_sector_bitmap, stripe->nr_sectors) {
1659 struct page *page = scrub_stripe_get_page(stripe, i);
1660 unsigned int pgoff = scrub_stripe_get_page_offset(stripe, i);
1661
1662 /* We're beyond the chunk boundary, no need to read anymore. */
1663 if (i >= nr_sectors)
1664 break;
1665
1666 /* The current sector cannot be merged, submit the bio. */
1667 if (bbio &&
1668 ((i > 0 &&
1669 !test_bit(i - 1, &stripe->extent_sector_bitmap)) ||
1670 bbio->bio.bi_iter.bi_size >= stripe_len)) {
1671 ASSERT(bbio->bio.bi_iter.bi_size);
1672 atomic_inc(&stripe->pending_io);
1673 btrfs_submit_bio(bbio, mirror);
1674 bbio = NULL;
1675 }
1676
1677 if (!bbio) {
1678 struct btrfs_io_stripe io_stripe = {};
1679 struct btrfs_io_context *bioc = NULL;
1680 const u64 logical = stripe->logical +
1681 (i << fs_info->sectorsize_bits);
1682 int err;
1683
1684 bbio = btrfs_bio_alloc(stripe->nr_sectors, REQ_OP_READ,
1685 fs_info, scrub_read_endio, stripe);
1686 bbio->bio.bi_iter.bi_sector = logical >> SECTOR_SHIFT;
1687
1688 io_stripe.is_scrub = true;
1689 err = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical,
1690 &stripe_len, &bioc, &io_stripe,
1691 &mirror);
1692 btrfs_put_bioc(bioc);
1693 if (err) {
1694 btrfs_bio_end_io(bbio,
1695 errno_to_blk_status(err));
1696 return;
1697 }
1698 }
1699
1700 __bio_add_page(&bbio->bio, page, fs_info->sectorsize, pgoff);
1701 }
1702
1703 if (bbio) {
1704 ASSERT(bbio->bio.bi_iter.bi_size);
1705 atomic_inc(&stripe->pending_io);
1706 btrfs_submit_bio(bbio, mirror);
1707 }
1708
1709 if (atomic_dec_and_test(&stripe->pending_io)) {
1710 wake_up(&stripe->io_wait);
1711 INIT_WORK(&stripe->work, scrub_stripe_read_repair_worker);
1712 queue_work(stripe->bg->fs_info->scrub_workers, &stripe->work);
1713 }
1714}
1715
1716static void scrub_submit_initial_read(struct scrub_ctx *sctx,
1717 struct scrub_stripe *stripe)
1718{
1719 struct btrfs_fs_info *fs_info = sctx->fs_info;
1720 struct btrfs_bio *bbio;
1721 unsigned int nr_sectors = min(BTRFS_STRIPE_LEN, stripe->bg->start +
1722 stripe->bg->length - stripe->logical) >>
1723 fs_info->sectorsize_bits;
1724 int mirror = stripe->mirror_num;
1725
1726 ASSERT(stripe->bg);
1727 ASSERT(stripe->mirror_num > 0);
1728 ASSERT(test_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &stripe->state));
1729
1730 if (btrfs_need_stripe_tree_update(fs_info, stripe->bg->flags)) {
1731 scrub_submit_extent_sector_read(sctx, stripe);
1732 return;
1733 }
1734
1735 bbio = btrfs_bio_alloc(SCRUB_STRIPE_PAGES, REQ_OP_READ, fs_info,
1736 scrub_read_endio, stripe);
1737
1738 bbio->bio.bi_iter.bi_sector = stripe->logical >> SECTOR_SHIFT;
1739 /* Read the whole range inside the chunk boundary. */
1740 for (unsigned int cur = 0; cur < nr_sectors; cur++) {
1741 struct page *page = scrub_stripe_get_page(stripe, cur);
1742 unsigned int pgoff = scrub_stripe_get_page_offset(stripe, cur);
1743 int ret;
1744
1745 ret = bio_add_page(&bbio->bio, page, fs_info->sectorsize, pgoff);
1746 /* We should have allocated enough bio vectors. */
1747 ASSERT(ret == fs_info->sectorsize);
1748 }
1749 atomic_inc(&stripe->pending_io);
1750
1751 /*
1752 * For dev-replace, either user asks to avoid the source dev, or
1753 * the device is missing, we try the next mirror instead.
1754 */
1755 if (sctx->is_dev_replace &&
1756 (fs_info->dev_replace.cont_reading_from_srcdev_mode ==
1757 BTRFS_DEV_REPLACE_ITEM_CONT_READING_FROM_SRCDEV_MODE_AVOID ||
1758 !stripe->dev->bdev)) {
1759 int num_copies = btrfs_num_copies(fs_info, stripe->bg->start,
1760 stripe->bg->length);
1761
1762 mirror = calc_next_mirror(mirror, num_copies);
1763 }
1764 btrfs_submit_bio(bbio, mirror);
1765}
1766
1767static bool stripe_has_metadata_error(struct scrub_stripe *stripe)
1768{
1769 int i;
1770
1771 for_each_set_bit(i, &stripe->error_bitmap, stripe->nr_sectors) {
1772 if (stripe->sectors[i].is_metadata) {
1773 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
1774
1775 btrfs_err(fs_info,
1776 "stripe %llu has unrepaired metadata sector at %llu",
1777 stripe->logical,
1778 stripe->logical + (i << fs_info->sectorsize_bits));
1779 return true;
1780 }
1781 }
1782 return false;
1783}
1784
1785static void submit_initial_group_read(struct scrub_ctx *sctx,
1786 unsigned int first_slot,
1787 unsigned int nr_stripes)
1788{
1789 struct blk_plug plug;
1790
1791 ASSERT(first_slot < SCRUB_TOTAL_STRIPES);
1792 ASSERT(first_slot + nr_stripes <= SCRUB_TOTAL_STRIPES);
1793
1794 scrub_throttle_dev_io(sctx, sctx->stripes[0].dev,
1795 btrfs_stripe_nr_to_offset(nr_stripes));
1796 blk_start_plug(&plug);
1797 for (int i = 0; i < nr_stripes; i++) {
1798 struct scrub_stripe *stripe = &sctx->stripes[first_slot + i];
1799
1800 /* Those stripes should be initialized. */
1801 ASSERT(test_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &stripe->state));
1802 scrub_submit_initial_read(sctx, stripe);
1803 }
1804 blk_finish_plug(&plug);
1805}
1806
1807static int flush_scrub_stripes(struct scrub_ctx *sctx)
1808{
1809 struct btrfs_fs_info *fs_info = sctx->fs_info;
1810 struct scrub_stripe *stripe;
1811 const int nr_stripes = sctx->cur_stripe;
1812 int ret = 0;
1813
1814 if (!nr_stripes)
1815 return 0;
1816
1817 ASSERT(test_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &sctx->stripes[0].state));
1818
1819 /* Submit the stripes which are populated but not submitted. */
1820 if (nr_stripes % SCRUB_STRIPES_PER_GROUP) {
1821 const int first_slot = round_down(nr_stripes, SCRUB_STRIPES_PER_GROUP);
1822
1823 submit_initial_group_read(sctx, first_slot, nr_stripes - first_slot);
1824 }
1825
1826 for (int i = 0; i < nr_stripes; i++) {
1827 stripe = &sctx->stripes[i];
1828
1829 wait_event(stripe->repair_wait,
1830 test_bit(SCRUB_STRIPE_FLAG_REPAIR_DONE, &stripe->state));
1831 }
1832
1833 /* Submit for dev-replace. */
1834 if (sctx->is_dev_replace) {
1835 /*
1836 * For dev-replace, if we know there is something wrong with
1837 * metadata, we should immediately abort.
1838 */
1839 for (int i = 0; i < nr_stripes; i++) {
1840 if (stripe_has_metadata_error(&sctx->stripes[i])) {
1841 ret = -EIO;
1842 goto out;
1843 }
1844 }
1845 for (int i = 0; i < nr_stripes; i++) {
1846 unsigned long good;
1847
1848 stripe = &sctx->stripes[i];
1849
1850 ASSERT(stripe->dev == fs_info->dev_replace.srcdev);
1851
1852 bitmap_andnot(&good, &stripe->extent_sector_bitmap,
1853 &stripe->error_bitmap, stripe->nr_sectors);
1854 scrub_write_sectors(sctx, stripe, good, true);
1855 }
1856 }
1857
1858 /* Wait for the above writebacks to finish. */
1859 for (int i = 0; i < nr_stripes; i++) {
1860 stripe = &sctx->stripes[i];
1861
1862 wait_scrub_stripe_io(stripe);
1863 scrub_reset_stripe(stripe);
1864 }
1865out:
1866 sctx->cur_stripe = 0;
1867 return ret;
1868}
1869
1870static void raid56_scrub_wait_endio(struct bio *bio)
1871{
1872 complete(bio->bi_private);
1873}
1874
1875static int queue_scrub_stripe(struct scrub_ctx *sctx, struct btrfs_block_group *bg,
1876 struct btrfs_device *dev, int mirror_num,
1877 u64 logical, u32 length, u64 physical,
1878 u64 *found_logical_ret)
1879{
1880 struct scrub_stripe *stripe;
1881 int ret;
1882
1883 /*
1884 * There should always be one slot left, as caller filling the last
1885 * slot should flush them all.
1886 */
1887 ASSERT(sctx->cur_stripe < SCRUB_TOTAL_STRIPES);
1888
1889 /* @found_logical_ret must be specified. */
1890 ASSERT(found_logical_ret);
1891
1892 stripe = &sctx->stripes[sctx->cur_stripe];
1893 scrub_reset_stripe(stripe);
1894 ret = scrub_find_fill_first_stripe(bg, &sctx->extent_path,
1895 &sctx->csum_path, dev, physical,
1896 mirror_num, logical, length, stripe);
1897 /* Either >0 as no more extents or <0 for error. */
1898 if (ret)
1899 return ret;
1900 *found_logical_ret = stripe->logical;
1901 sctx->cur_stripe++;
1902
1903 /* We filled one group, submit it. */
1904 if (sctx->cur_stripe % SCRUB_STRIPES_PER_GROUP == 0) {
1905 const int first_slot = sctx->cur_stripe - SCRUB_STRIPES_PER_GROUP;
1906
1907 submit_initial_group_read(sctx, first_slot, SCRUB_STRIPES_PER_GROUP);
1908 }
1909
1910 /* Last slot used, flush them all. */
1911 if (sctx->cur_stripe == SCRUB_TOTAL_STRIPES)
1912 return flush_scrub_stripes(sctx);
1913 return 0;
1914}
1915
1916static int scrub_raid56_parity_stripe(struct scrub_ctx *sctx,
1917 struct btrfs_device *scrub_dev,
1918 struct btrfs_block_group *bg,
1919 struct btrfs_chunk_map *map,
1920 u64 full_stripe_start)
1921{
1922 DECLARE_COMPLETION_ONSTACK(io_done);
1923 struct btrfs_fs_info *fs_info = sctx->fs_info;
1924 struct btrfs_raid_bio *rbio;
1925 struct btrfs_io_context *bioc = NULL;
1926 struct btrfs_path extent_path = { 0 };
1927 struct btrfs_path csum_path = { 0 };
1928 struct bio *bio;
1929 struct scrub_stripe *stripe;
1930 bool all_empty = true;
1931 const int data_stripes = nr_data_stripes(map);
1932 unsigned long extent_bitmap = 0;
1933 u64 length = btrfs_stripe_nr_to_offset(data_stripes);
1934 int ret;
1935
1936 ASSERT(sctx->raid56_data_stripes);
1937
1938 /*
1939 * For data stripe search, we cannot re-use the same extent/csum paths,
1940 * as the data stripe bytenr may be smaller than previous extent. Thus
1941 * we have to use our own extent/csum paths.
1942 */
1943 extent_path.search_commit_root = 1;
1944 extent_path.skip_locking = 1;
1945 csum_path.search_commit_root = 1;
1946 csum_path.skip_locking = 1;
1947
1948 for (int i = 0; i < data_stripes; i++) {
1949 int stripe_index;
1950 int rot;
1951 u64 physical;
1952
1953 stripe = &sctx->raid56_data_stripes[i];
1954 rot = div_u64(full_stripe_start - bg->start,
1955 data_stripes) >> BTRFS_STRIPE_LEN_SHIFT;
1956 stripe_index = (i + rot) % map->num_stripes;
1957 physical = map->stripes[stripe_index].physical +
1958 btrfs_stripe_nr_to_offset(rot);
1959
1960 scrub_reset_stripe(stripe);
1961 set_bit(SCRUB_STRIPE_FLAG_NO_REPORT, &stripe->state);
1962 ret = scrub_find_fill_first_stripe(bg, &extent_path, &csum_path,
1963 map->stripes[stripe_index].dev, physical, 1,
1964 full_stripe_start + btrfs_stripe_nr_to_offset(i),
1965 BTRFS_STRIPE_LEN, stripe);
1966 if (ret < 0)
1967 goto out;
1968 /*
1969 * No extent in this data stripe, need to manually mark them
1970 * initialized to make later read submission happy.
1971 */
1972 if (ret > 0) {
1973 stripe->logical = full_stripe_start +
1974 btrfs_stripe_nr_to_offset(i);
1975 stripe->dev = map->stripes[stripe_index].dev;
1976 stripe->mirror_num = 1;
1977 set_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &stripe->state);
1978 }
1979 }
1980
1981 /* Check if all data stripes are empty. */
1982 for (int i = 0; i < data_stripes; i++) {
1983 stripe = &sctx->raid56_data_stripes[i];
1984 if (!bitmap_empty(&stripe->extent_sector_bitmap, stripe->nr_sectors)) {
1985 all_empty = false;
1986 break;
1987 }
1988 }
1989 if (all_empty) {
1990 ret = 0;
1991 goto out;
1992 }
1993
1994 for (int i = 0; i < data_stripes; i++) {
1995 stripe = &sctx->raid56_data_stripes[i];
1996 scrub_submit_initial_read(sctx, stripe);
1997 }
1998 for (int i = 0; i < data_stripes; i++) {
1999 stripe = &sctx->raid56_data_stripes[i];
2000
2001 wait_event(stripe->repair_wait,
2002 test_bit(SCRUB_STRIPE_FLAG_REPAIR_DONE, &stripe->state));
2003 }
2004 /* For now, no zoned support for RAID56. */
2005 ASSERT(!btrfs_is_zoned(sctx->fs_info));
2006
2007 /*
2008 * Now all data stripes are properly verified. Check if we have any
2009 * unrepaired, if so abort immediately or we could further corrupt the
2010 * P/Q stripes.
2011 *
2012 * During the loop, also populate extent_bitmap.
2013 */
2014 for (int i = 0; i < data_stripes; i++) {
2015 unsigned long error;
2016
2017 stripe = &sctx->raid56_data_stripes[i];
2018
2019 /*
2020 * We should only check the errors where there is an extent.
2021 * As we may hit an empty data stripe while it's missing.
2022 */
2023 bitmap_and(&error, &stripe->error_bitmap,
2024 &stripe->extent_sector_bitmap, stripe->nr_sectors);
2025 if (!bitmap_empty(&error, stripe->nr_sectors)) {
2026 btrfs_err(fs_info,
2027"unrepaired sectors detected, full stripe %llu data stripe %u errors %*pbl",
2028 full_stripe_start, i, stripe->nr_sectors,
2029 &error);
2030 ret = -EIO;
2031 goto out;
2032 }
2033 bitmap_or(&extent_bitmap, &extent_bitmap,
2034 &stripe->extent_sector_bitmap, stripe->nr_sectors);
2035 }
2036
2037 /* Now we can check and regenerate the P/Q stripe. */
2038 bio = bio_alloc(NULL, 1, REQ_OP_READ, GFP_NOFS);
2039 bio->bi_iter.bi_sector = full_stripe_start >> SECTOR_SHIFT;
2040 bio->bi_private = &io_done;
2041 bio->bi_end_io = raid56_scrub_wait_endio;
2042
2043 btrfs_bio_counter_inc_blocked(fs_info);
2044 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, full_stripe_start,
2045 &length, &bioc, NULL, NULL);
2046 if (ret < 0) {
2047 btrfs_put_bioc(bioc);
2048 btrfs_bio_counter_dec(fs_info);
2049 goto out;
2050 }
2051 rbio = raid56_parity_alloc_scrub_rbio(bio, bioc, scrub_dev, &extent_bitmap,
2052 BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits);
2053 btrfs_put_bioc(bioc);
2054 if (!rbio) {
2055 ret = -ENOMEM;
2056 btrfs_bio_counter_dec(fs_info);
2057 goto out;
2058 }
2059 /* Use the recovered stripes as cache to avoid read them from disk again. */
2060 for (int i = 0; i < data_stripes; i++) {
2061 stripe = &sctx->raid56_data_stripes[i];
2062
2063 raid56_parity_cache_data_pages(rbio, stripe->pages,
2064 full_stripe_start + (i << BTRFS_STRIPE_LEN_SHIFT));
2065 }
2066 raid56_parity_submit_scrub_rbio(rbio);
2067 wait_for_completion_io(&io_done);
2068 ret = blk_status_to_errno(bio->bi_status);
2069 bio_put(bio);
2070 btrfs_bio_counter_dec(fs_info);
2071
2072 btrfs_release_path(&extent_path);
2073 btrfs_release_path(&csum_path);
2074out:
2075 return ret;
2076}
2077
2078/*
2079 * Scrub one range which can only has simple mirror based profile.
2080 * (Including all range in SINGLE/DUP/RAID1/RAID1C*, and each stripe in
2081 * RAID0/RAID10).
2082 *
2083 * Since we may need to handle a subset of block group, we need @logical_start
2084 * and @logical_length parameter.
2085 */
2086static int scrub_simple_mirror(struct scrub_ctx *sctx,
2087 struct btrfs_block_group *bg,
2088 struct btrfs_chunk_map *map,
2089 u64 logical_start, u64 logical_length,
2090 struct btrfs_device *device,
2091 u64 physical, int mirror_num)
2092{
2093 struct btrfs_fs_info *fs_info = sctx->fs_info;
2094 const u64 logical_end = logical_start + logical_length;
2095 u64 cur_logical = logical_start;
2096 int ret;
2097
2098 /* The range must be inside the bg */
2099 ASSERT(logical_start >= bg->start && logical_end <= bg->start + bg->length);
2100
2101 /* Go through each extent items inside the logical range */
2102 while (cur_logical < logical_end) {
2103 u64 found_logical = U64_MAX;
2104 u64 cur_physical = physical + cur_logical - logical_start;
2105
2106 /* Canceled? */
2107 if (atomic_read(&fs_info->scrub_cancel_req) ||
2108 atomic_read(&sctx->cancel_req)) {
2109 ret = -ECANCELED;
2110 break;
2111 }
2112 /* Paused? */
2113 if (atomic_read(&fs_info->scrub_pause_req)) {
2114 /* Push queued extents */
2115 scrub_blocked_if_needed(fs_info);
2116 }
2117 /* Block group removed? */
2118 spin_lock(&bg->lock);
2119 if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &bg->runtime_flags)) {
2120 spin_unlock(&bg->lock);
2121 ret = 0;
2122 break;
2123 }
2124 spin_unlock(&bg->lock);
2125
2126 ret = queue_scrub_stripe(sctx, bg, device, mirror_num,
2127 cur_logical, logical_end - cur_logical,
2128 cur_physical, &found_logical);
2129 if (ret > 0) {
2130 /* No more extent, just update the accounting */
2131 sctx->stat.last_physical = physical + logical_length;
2132 ret = 0;
2133 break;
2134 }
2135 if (ret < 0)
2136 break;
2137
2138 /* queue_scrub_stripe() returned 0, @found_logical must be updated. */
2139 ASSERT(found_logical != U64_MAX);
2140 cur_logical = found_logical + BTRFS_STRIPE_LEN;
2141
2142 /* Don't hold CPU for too long time */
2143 cond_resched();
2144 }
2145 return ret;
2146}
2147
2148/* Calculate the full stripe length for simple stripe based profiles */
2149static u64 simple_stripe_full_stripe_len(const struct btrfs_chunk_map *map)
2150{
2151 ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2152 BTRFS_BLOCK_GROUP_RAID10));
2153
2154 return btrfs_stripe_nr_to_offset(map->num_stripes / map->sub_stripes);
2155}
2156
2157/* Get the logical bytenr for the stripe */
2158static u64 simple_stripe_get_logical(struct btrfs_chunk_map *map,
2159 struct btrfs_block_group *bg,
2160 int stripe_index)
2161{
2162 ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2163 BTRFS_BLOCK_GROUP_RAID10));
2164 ASSERT(stripe_index < map->num_stripes);
2165
2166 /*
2167 * (stripe_index / sub_stripes) gives how many data stripes we need to
2168 * skip.
2169 */
2170 return btrfs_stripe_nr_to_offset(stripe_index / map->sub_stripes) +
2171 bg->start;
2172}
2173
2174/* Get the mirror number for the stripe */
2175static int simple_stripe_mirror_num(struct btrfs_chunk_map *map, int stripe_index)
2176{
2177 ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2178 BTRFS_BLOCK_GROUP_RAID10));
2179 ASSERT(stripe_index < map->num_stripes);
2180
2181 /* For RAID0, it's fixed to 1, for RAID10 it's 0,1,0,1... */
2182 return stripe_index % map->sub_stripes + 1;
2183}
2184
2185static int scrub_simple_stripe(struct scrub_ctx *sctx,
2186 struct btrfs_block_group *bg,
2187 struct btrfs_chunk_map *map,
2188 struct btrfs_device *device,
2189 int stripe_index)
2190{
2191 const u64 logical_increment = simple_stripe_full_stripe_len(map);
2192 const u64 orig_logical = simple_stripe_get_logical(map, bg, stripe_index);
2193 const u64 orig_physical = map->stripes[stripe_index].physical;
2194 const int mirror_num = simple_stripe_mirror_num(map, stripe_index);
2195 u64 cur_logical = orig_logical;
2196 u64 cur_physical = orig_physical;
2197 int ret = 0;
2198
2199 while (cur_logical < bg->start + bg->length) {
2200 /*
2201 * Inside each stripe, RAID0 is just SINGLE, and RAID10 is
2202 * just RAID1, so we can reuse scrub_simple_mirror() to scrub
2203 * this stripe.
2204 */
2205 ret = scrub_simple_mirror(sctx, bg, map, cur_logical,
2206 BTRFS_STRIPE_LEN, device, cur_physical,
2207 mirror_num);
2208 if (ret)
2209 return ret;
2210 /* Skip to next stripe which belongs to the target device */
2211 cur_logical += logical_increment;
2212 /* For physical offset, we just go to next stripe */
2213 cur_physical += BTRFS_STRIPE_LEN;
2214 }
2215 return ret;
2216}
2217
2218static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2219 struct btrfs_block_group *bg,
2220 struct btrfs_chunk_map *map,
2221 struct btrfs_device *scrub_dev,
2222 int stripe_index)
2223{
2224 struct btrfs_fs_info *fs_info = sctx->fs_info;
2225 const u64 profile = map->type & BTRFS_BLOCK_GROUP_PROFILE_MASK;
2226 const u64 chunk_logical = bg->start;
2227 int ret;
2228 int ret2;
2229 u64 physical = map->stripes[stripe_index].physical;
2230 const u64 dev_stripe_len = btrfs_calc_stripe_length(map);
2231 const u64 physical_end = physical + dev_stripe_len;
2232 u64 logical;
2233 u64 logic_end;
2234 /* The logical increment after finishing one stripe */
2235 u64 increment;
2236 /* Offset inside the chunk */
2237 u64 offset;
2238 u64 stripe_logical;
2239 int stop_loop = 0;
2240
2241 /* Extent_path should be released by now. */
2242 ASSERT(sctx->extent_path.nodes[0] == NULL);
2243
2244 scrub_blocked_if_needed(fs_info);
2245
2246 if (sctx->is_dev_replace &&
2247 btrfs_dev_is_sequential(sctx->wr_tgtdev, physical)) {
2248 mutex_lock(&sctx->wr_lock);
2249 sctx->write_pointer = physical;
2250 mutex_unlock(&sctx->wr_lock);
2251 }
2252
2253 /* Prepare the extra data stripes used by RAID56. */
2254 if (profile & BTRFS_BLOCK_GROUP_RAID56_MASK) {
2255 ASSERT(sctx->raid56_data_stripes == NULL);
2256
2257 sctx->raid56_data_stripes = kcalloc(nr_data_stripes(map),
2258 sizeof(struct scrub_stripe),
2259 GFP_KERNEL);
2260 if (!sctx->raid56_data_stripes) {
2261 ret = -ENOMEM;
2262 goto out;
2263 }
2264 for (int i = 0; i < nr_data_stripes(map); i++) {
2265 ret = init_scrub_stripe(fs_info,
2266 &sctx->raid56_data_stripes[i]);
2267 if (ret < 0)
2268 goto out;
2269 sctx->raid56_data_stripes[i].bg = bg;
2270 sctx->raid56_data_stripes[i].sctx = sctx;
2271 }
2272 }
2273 /*
2274 * There used to be a big double loop to handle all profiles using the
2275 * same routine, which grows larger and more gross over time.
2276 *
2277 * So here we handle each profile differently, so simpler profiles
2278 * have simpler scrubbing function.
2279 */
2280 if (!(profile & (BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID10 |
2281 BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2282 /*
2283 * Above check rules out all complex profile, the remaining
2284 * profiles are SINGLE|DUP|RAID1|RAID1C*, which is simple
2285 * mirrored duplication without stripe.
2286 *
2287 * Only @physical and @mirror_num needs to calculated using
2288 * @stripe_index.
2289 */
2290 ret = scrub_simple_mirror(sctx, bg, map, bg->start, bg->length,
2291 scrub_dev, map->stripes[stripe_index].physical,
2292 stripe_index + 1);
2293 offset = 0;
2294 goto out;
2295 }
2296 if (profile & (BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID10)) {
2297 ret = scrub_simple_stripe(sctx, bg, map, scrub_dev, stripe_index);
2298 offset = btrfs_stripe_nr_to_offset(stripe_index / map->sub_stripes);
2299 goto out;
2300 }
2301
2302 /* Only RAID56 goes through the old code */
2303 ASSERT(map->type & BTRFS_BLOCK_GROUP_RAID56_MASK);
2304 ret = 0;
2305
2306 /* Calculate the logical end of the stripe */
2307 get_raid56_logic_offset(physical_end, stripe_index,
2308 map, &logic_end, NULL);
2309 logic_end += chunk_logical;
2310
2311 /* Initialize @offset in case we need to go to out: label */
2312 get_raid56_logic_offset(physical, stripe_index, map, &offset, NULL);
2313 increment = btrfs_stripe_nr_to_offset(nr_data_stripes(map));
2314
2315 /*
2316 * Due to the rotation, for RAID56 it's better to iterate each stripe
2317 * using their physical offset.
2318 */
2319 while (physical < physical_end) {
2320 ret = get_raid56_logic_offset(physical, stripe_index, map,
2321 &logical, &stripe_logical);
2322 logical += chunk_logical;
2323 if (ret) {
2324 /* it is parity strip */
2325 stripe_logical += chunk_logical;
2326 ret = scrub_raid56_parity_stripe(sctx, scrub_dev, bg,
2327 map, stripe_logical);
2328 if (ret)
2329 goto out;
2330 goto next;
2331 }
2332
2333 /*
2334 * Now we're at a data stripe, scrub each extents in the range.
2335 *
2336 * At this stage, if we ignore the repair part, inside each data
2337 * stripe it is no different than SINGLE profile.
2338 * We can reuse scrub_simple_mirror() here, as the repair part
2339 * is still based on @mirror_num.
2340 */
2341 ret = scrub_simple_mirror(sctx, bg, map, logical, BTRFS_STRIPE_LEN,
2342 scrub_dev, physical, 1);
2343 if (ret < 0)
2344 goto out;
2345next:
2346 logical += increment;
2347 physical += BTRFS_STRIPE_LEN;
2348 spin_lock(&sctx->stat_lock);
2349 if (stop_loop)
2350 sctx->stat.last_physical =
2351 map->stripes[stripe_index].physical + dev_stripe_len;
2352 else
2353 sctx->stat.last_physical = physical;
2354 spin_unlock(&sctx->stat_lock);
2355 if (stop_loop)
2356 break;
2357 }
2358out:
2359 ret2 = flush_scrub_stripes(sctx);
2360 if (!ret)
2361 ret = ret2;
2362 btrfs_release_path(&sctx->extent_path);
2363 btrfs_release_path(&sctx->csum_path);
2364
2365 if (sctx->raid56_data_stripes) {
2366 for (int i = 0; i < nr_data_stripes(map); i++)
2367 release_scrub_stripe(&sctx->raid56_data_stripes[i]);
2368 kfree(sctx->raid56_data_stripes);
2369 sctx->raid56_data_stripes = NULL;
2370 }
2371
2372 if (sctx->is_dev_replace && ret >= 0) {
2373 int ret2;
2374
2375 ret2 = sync_write_pointer_for_zoned(sctx,
2376 chunk_logical + offset,
2377 map->stripes[stripe_index].physical,
2378 physical_end);
2379 if (ret2)
2380 ret = ret2;
2381 }
2382
2383 return ret < 0 ? ret : 0;
2384}
2385
2386static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
2387 struct btrfs_block_group *bg,
2388 struct btrfs_device *scrub_dev,
2389 u64 dev_offset,
2390 u64 dev_extent_len)
2391{
2392 struct btrfs_fs_info *fs_info = sctx->fs_info;
2393 struct btrfs_chunk_map *map;
2394 int i;
2395 int ret = 0;
2396
2397 map = btrfs_find_chunk_map(fs_info, bg->start, bg->length);
2398 if (!map) {
2399 /*
2400 * Might have been an unused block group deleted by the cleaner
2401 * kthread or relocation.
2402 */
2403 spin_lock(&bg->lock);
2404 if (!test_bit(BLOCK_GROUP_FLAG_REMOVED, &bg->runtime_flags))
2405 ret = -EINVAL;
2406 spin_unlock(&bg->lock);
2407
2408 return ret;
2409 }
2410 if (map->start != bg->start)
2411 goto out;
2412 if (map->chunk_len < dev_extent_len)
2413 goto out;
2414
2415 for (i = 0; i < map->num_stripes; ++i) {
2416 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
2417 map->stripes[i].physical == dev_offset) {
2418 ret = scrub_stripe(sctx, bg, map, scrub_dev, i);
2419 if (ret)
2420 goto out;
2421 }
2422 }
2423out:
2424 btrfs_free_chunk_map(map);
2425
2426 return ret;
2427}
2428
2429static int finish_extent_writes_for_zoned(struct btrfs_root *root,
2430 struct btrfs_block_group *cache)
2431{
2432 struct btrfs_fs_info *fs_info = cache->fs_info;
2433 struct btrfs_trans_handle *trans;
2434
2435 if (!btrfs_is_zoned(fs_info))
2436 return 0;
2437
2438 btrfs_wait_block_group_reservations(cache);
2439 btrfs_wait_nocow_writers(cache);
2440 btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start, cache->length);
2441
2442 trans = btrfs_join_transaction(root);
2443 if (IS_ERR(trans))
2444 return PTR_ERR(trans);
2445 return btrfs_commit_transaction(trans);
2446}
2447
2448static noinline_for_stack
2449int scrub_enumerate_chunks(struct scrub_ctx *sctx,
2450 struct btrfs_device *scrub_dev, u64 start, u64 end)
2451{
2452 struct btrfs_dev_extent *dev_extent = NULL;
2453 struct btrfs_path *path;
2454 struct btrfs_fs_info *fs_info = sctx->fs_info;
2455 struct btrfs_root *root = fs_info->dev_root;
2456 u64 chunk_offset;
2457 int ret = 0;
2458 int ro_set;
2459 int slot;
2460 struct extent_buffer *l;
2461 struct btrfs_key key;
2462 struct btrfs_key found_key;
2463 struct btrfs_block_group *cache;
2464 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2465
2466 path = btrfs_alloc_path();
2467 if (!path)
2468 return -ENOMEM;
2469
2470 path->reada = READA_FORWARD;
2471 path->search_commit_root = 1;
2472 path->skip_locking = 1;
2473
2474 key.objectid = scrub_dev->devid;
2475 key.offset = 0ull;
2476 key.type = BTRFS_DEV_EXTENT_KEY;
2477
2478 while (1) {
2479 u64 dev_extent_len;
2480
2481 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2482 if (ret < 0)
2483 break;
2484 if (ret > 0) {
2485 if (path->slots[0] >=
2486 btrfs_header_nritems(path->nodes[0])) {
2487 ret = btrfs_next_leaf(root, path);
2488 if (ret < 0)
2489 break;
2490 if (ret > 0) {
2491 ret = 0;
2492 break;
2493 }
2494 } else {
2495 ret = 0;
2496 }
2497 }
2498
2499 l = path->nodes[0];
2500 slot = path->slots[0];
2501
2502 btrfs_item_key_to_cpu(l, &found_key, slot);
2503
2504 if (found_key.objectid != scrub_dev->devid)
2505 break;
2506
2507 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
2508 break;
2509
2510 if (found_key.offset >= end)
2511 break;
2512
2513 if (found_key.offset < key.offset)
2514 break;
2515
2516 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2517 dev_extent_len = btrfs_dev_extent_length(l, dev_extent);
2518
2519 if (found_key.offset + dev_extent_len <= start)
2520 goto skip;
2521
2522 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2523
2524 /*
2525 * get a reference on the corresponding block group to prevent
2526 * the chunk from going away while we scrub it
2527 */
2528 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2529
2530 /* some chunks are removed but not committed to disk yet,
2531 * continue scrubbing */
2532 if (!cache)
2533 goto skip;
2534
2535 ASSERT(cache->start <= chunk_offset);
2536 /*
2537 * We are using the commit root to search for device extents, so
2538 * that means we could have found a device extent item from a
2539 * block group that was deleted in the current transaction. The
2540 * logical start offset of the deleted block group, stored at
2541 * @chunk_offset, might be part of the logical address range of
2542 * a new block group (which uses different physical extents).
2543 * In this case btrfs_lookup_block_group() has returned the new
2544 * block group, and its start address is less than @chunk_offset.
2545 *
2546 * We skip such new block groups, because it's pointless to
2547 * process them, as we won't find their extents because we search
2548 * for them using the commit root of the extent tree. For a device
2549 * replace it's also fine to skip it, we won't miss copying them
2550 * to the target device because we have the write duplication
2551 * setup through the regular write path (by btrfs_map_block()),
2552 * and we have committed a transaction when we started the device
2553 * replace, right after setting up the device replace state.
2554 */
2555 if (cache->start < chunk_offset) {
2556 btrfs_put_block_group(cache);
2557 goto skip;
2558 }
2559
2560 if (sctx->is_dev_replace && btrfs_is_zoned(fs_info)) {
2561 if (!test_bit(BLOCK_GROUP_FLAG_TO_COPY, &cache->runtime_flags)) {
2562 btrfs_put_block_group(cache);
2563 goto skip;
2564 }
2565 }
2566
2567 /*
2568 * Make sure that while we are scrubbing the corresponding block
2569 * group doesn't get its logical address and its device extents
2570 * reused for another block group, which can possibly be of a
2571 * different type and different profile. We do this to prevent
2572 * false error detections and crashes due to bogus attempts to
2573 * repair extents.
2574 */
2575 spin_lock(&cache->lock);
2576 if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &cache->runtime_flags)) {
2577 spin_unlock(&cache->lock);
2578 btrfs_put_block_group(cache);
2579 goto skip;
2580 }
2581 btrfs_freeze_block_group(cache);
2582 spin_unlock(&cache->lock);
2583
2584 /*
2585 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
2586 * to avoid deadlock caused by:
2587 * btrfs_inc_block_group_ro()
2588 * -> btrfs_wait_for_commit()
2589 * -> btrfs_commit_transaction()
2590 * -> btrfs_scrub_pause()
2591 */
2592 scrub_pause_on(fs_info);
2593
2594 /*
2595 * Don't do chunk preallocation for scrub.
2596 *
2597 * This is especially important for SYSTEM bgs, or we can hit
2598 * -EFBIG from btrfs_finish_chunk_alloc() like:
2599 * 1. The only SYSTEM bg is marked RO.
2600 * Since SYSTEM bg is small, that's pretty common.
2601 * 2. New SYSTEM bg will be allocated
2602 * Due to regular version will allocate new chunk.
2603 * 3. New SYSTEM bg is empty and will get cleaned up
2604 * Before cleanup really happens, it's marked RO again.
2605 * 4. Empty SYSTEM bg get scrubbed
2606 * We go back to 2.
2607 *
2608 * This can easily boost the amount of SYSTEM chunks if cleaner
2609 * thread can't be triggered fast enough, and use up all space
2610 * of btrfs_super_block::sys_chunk_array
2611 *
2612 * While for dev replace, we need to try our best to mark block
2613 * group RO, to prevent race between:
2614 * - Write duplication
2615 * Contains latest data
2616 * - Scrub copy
2617 * Contains data from commit tree
2618 *
2619 * If target block group is not marked RO, nocow writes can
2620 * be overwritten by scrub copy, causing data corruption.
2621 * So for dev-replace, it's not allowed to continue if a block
2622 * group is not RO.
2623 */
2624 ret = btrfs_inc_block_group_ro(cache, sctx->is_dev_replace);
2625 if (!ret && sctx->is_dev_replace) {
2626 ret = finish_extent_writes_for_zoned(root, cache);
2627 if (ret) {
2628 btrfs_dec_block_group_ro(cache);
2629 scrub_pause_off(fs_info);
2630 btrfs_put_block_group(cache);
2631 break;
2632 }
2633 }
2634
2635 if (ret == 0) {
2636 ro_set = 1;
2637 } else if (ret == -ENOSPC && !sctx->is_dev_replace &&
2638 !(cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK)) {
2639 /*
2640 * btrfs_inc_block_group_ro return -ENOSPC when it
2641 * failed in creating new chunk for metadata.
2642 * It is not a problem for scrub, because
2643 * metadata are always cowed, and our scrub paused
2644 * commit_transactions.
2645 *
2646 * For RAID56 chunks, we have to mark them read-only
2647 * for scrub, as later we would use our own cache
2648 * out of RAID56 realm.
2649 * Thus we want the RAID56 bg to be marked RO to
2650 * prevent RMW from screwing up out cache.
2651 */
2652 ro_set = 0;
2653 } else if (ret == -ETXTBSY) {
2654 btrfs_warn(fs_info,
2655 "skipping scrub of block group %llu due to active swapfile",
2656 cache->start);
2657 scrub_pause_off(fs_info);
2658 ret = 0;
2659 goto skip_unfreeze;
2660 } else {
2661 btrfs_warn(fs_info,
2662 "failed setting block group ro: %d", ret);
2663 btrfs_unfreeze_block_group(cache);
2664 btrfs_put_block_group(cache);
2665 scrub_pause_off(fs_info);
2666 break;
2667 }
2668
2669 /*
2670 * Now the target block is marked RO, wait for nocow writes to
2671 * finish before dev-replace.
2672 * COW is fine, as COW never overwrites extents in commit tree.
2673 */
2674 if (sctx->is_dev_replace) {
2675 btrfs_wait_nocow_writers(cache);
2676 btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start,
2677 cache->length);
2678 }
2679
2680 scrub_pause_off(fs_info);
2681 down_write(&dev_replace->rwsem);
2682 dev_replace->cursor_right = found_key.offset + dev_extent_len;
2683 dev_replace->cursor_left = found_key.offset;
2684 dev_replace->item_needs_writeback = 1;
2685 up_write(&dev_replace->rwsem);
2686
2687 ret = scrub_chunk(sctx, cache, scrub_dev, found_key.offset,
2688 dev_extent_len);
2689 if (sctx->is_dev_replace &&
2690 !btrfs_finish_block_group_to_copy(dev_replace->srcdev,
2691 cache, found_key.offset))
2692 ro_set = 0;
2693
2694 down_write(&dev_replace->rwsem);
2695 dev_replace->cursor_left = dev_replace->cursor_right;
2696 dev_replace->item_needs_writeback = 1;
2697 up_write(&dev_replace->rwsem);
2698
2699 if (ro_set)
2700 btrfs_dec_block_group_ro(cache);
2701
2702 /*
2703 * We might have prevented the cleaner kthread from deleting
2704 * this block group if it was already unused because we raced
2705 * and set it to RO mode first. So add it back to the unused
2706 * list, otherwise it might not ever be deleted unless a manual
2707 * balance is triggered or it becomes used and unused again.
2708 */
2709 spin_lock(&cache->lock);
2710 if (!test_bit(BLOCK_GROUP_FLAG_REMOVED, &cache->runtime_flags) &&
2711 !cache->ro && cache->reserved == 0 && cache->used == 0) {
2712 spin_unlock(&cache->lock);
2713 if (btrfs_test_opt(fs_info, DISCARD_ASYNC))
2714 btrfs_discard_queue_work(&fs_info->discard_ctl,
2715 cache);
2716 else
2717 btrfs_mark_bg_unused(cache);
2718 } else {
2719 spin_unlock(&cache->lock);
2720 }
2721skip_unfreeze:
2722 btrfs_unfreeze_block_group(cache);
2723 btrfs_put_block_group(cache);
2724 if (ret)
2725 break;
2726 if (sctx->is_dev_replace &&
2727 atomic64_read(&dev_replace->num_write_errors) > 0) {
2728 ret = -EIO;
2729 break;
2730 }
2731 if (sctx->stat.malloc_errors > 0) {
2732 ret = -ENOMEM;
2733 break;
2734 }
2735skip:
2736 key.offset = found_key.offset + dev_extent_len;
2737 btrfs_release_path(path);
2738 }
2739
2740 btrfs_free_path(path);
2741
2742 return ret;
2743}
2744
2745static int scrub_one_super(struct scrub_ctx *sctx, struct btrfs_device *dev,
2746 struct page *page, u64 physical, u64 generation)
2747{
2748 struct btrfs_fs_info *fs_info = sctx->fs_info;
2749 struct bio_vec bvec;
2750 struct bio bio;
2751 struct btrfs_super_block *sb = page_address(page);
2752 int ret;
2753
2754 bio_init(&bio, dev->bdev, &bvec, 1, REQ_OP_READ);
2755 bio.bi_iter.bi_sector = physical >> SECTOR_SHIFT;
2756 __bio_add_page(&bio, page, BTRFS_SUPER_INFO_SIZE, 0);
2757 ret = submit_bio_wait(&bio);
2758 bio_uninit(&bio);
2759
2760 if (ret < 0)
2761 return ret;
2762 ret = btrfs_check_super_csum(fs_info, sb);
2763 if (ret != 0) {
2764 btrfs_err_rl(fs_info,
2765 "super block at physical %llu devid %llu has bad csum",
2766 physical, dev->devid);
2767 return -EIO;
2768 }
2769 if (btrfs_super_generation(sb) != generation) {
2770 btrfs_err_rl(fs_info,
2771"super block at physical %llu devid %llu has bad generation %llu expect %llu",
2772 physical, dev->devid,
2773 btrfs_super_generation(sb), generation);
2774 return -EUCLEAN;
2775 }
2776
2777 return btrfs_validate_super(fs_info, sb, -1);
2778}
2779
2780static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
2781 struct btrfs_device *scrub_dev)
2782{
2783 int i;
2784 u64 bytenr;
2785 u64 gen;
2786 int ret = 0;
2787 struct page *page;
2788 struct btrfs_fs_info *fs_info = sctx->fs_info;
2789
2790 if (BTRFS_FS_ERROR(fs_info))
2791 return -EROFS;
2792
2793 page = alloc_page(GFP_KERNEL);
2794 if (!page) {
2795 spin_lock(&sctx->stat_lock);
2796 sctx->stat.malloc_errors++;
2797 spin_unlock(&sctx->stat_lock);
2798 return -ENOMEM;
2799 }
2800
2801 /* Seed devices of a new filesystem has their own generation. */
2802 if (scrub_dev->fs_devices != fs_info->fs_devices)
2803 gen = scrub_dev->generation;
2804 else
2805 gen = btrfs_get_last_trans_committed(fs_info);
2806
2807 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2808 bytenr = btrfs_sb_offset(i);
2809 if (bytenr + BTRFS_SUPER_INFO_SIZE >
2810 scrub_dev->commit_total_bytes)
2811 break;
2812 if (!btrfs_check_super_location(scrub_dev, bytenr))
2813 continue;
2814
2815 ret = scrub_one_super(sctx, scrub_dev, page, bytenr, gen);
2816 if (ret) {
2817 spin_lock(&sctx->stat_lock);
2818 sctx->stat.super_errors++;
2819 spin_unlock(&sctx->stat_lock);
2820 }
2821 }
2822 __free_page(page);
2823 return 0;
2824}
2825
2826static void scrub_workers_put(struct btrfs_fs_info *fs_info)
2827{
2828 if (refcount_dec_and_mutex_lock(&fs_info->scrub_workers_refcnt,
2829 &fs_info->scrub_lock)) {
2830 struct workqueue_struct *scrub_workers = fs_info->scrub_workers;
2831
2832 fs_info->scrub_workers = NULL;
2833 mutex_unlock(&fs_info->scrub_lock);
2834
2835 if (scrub_workers)
2836 destroy_workqueue(scrub_workers);
2837 }
2838}
2839
2840/*
2841 * get a reference count on fs_info->scrub_workers. start worker if necessary
2842 */
2843static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info)
2844{
2845 struct workqueue_struct *scrub_workers = NULL;
2846 unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
2847 int max_active = fs_info->thread_pool_size;
2848 int ret = -ENOMEM;
2849
2850 if (refcount_inc_not_zero(&fs_info->scrub_workers_refcnt))
2851 return 0;
2852
2853 scrub_workers = alloc_workqueue("btrfs-scrub", flags, max_active);
2854 if (!scrub_workers)
2855 return -ENOMEM;
2856
2857 mutex_lock(&fs_info->scrub_lock);
2858 if (refcount_read(&fs_info->scrub_workers_refcnt) == 0) {
2859 ASSERT(fs_info->scrub_workers == NULL);
2860 fs_info->scrub_workers = scrub_workers;
2861 refcount_set(&fs_info->scrub_workers_refcnt, 1);
2862 mutex_unlock(&fs_info->scrub_lock);
2863 return 0;
2864 }
2865 /* Other thread raced in and created the workers for us */
2866 refcount_inc(&fs_info->scrub_workers_refcnt);
2867 mutex_unlock(&fs_info->scrub_lock);
2868
2869 ret = 0;
2870
2871 destroy_workqueue(scrub_workers);
2872 return ret;
2873}
2874
2875int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
2876 u64 end, struct btrfs_scrub_progress *progress,
2877 int readonly, int is_dev_replace)
2878{
2879 struct btrfs_dev_lookup_args args = { .devid = devid };
2880 struct scrub_ctx *sctx;
2881 int ret;
2882 struct btrfs_device *dev;
2883 unsigned int nofs_flag;
2884 bool need_commit = false;
2885
2886 if (btrfs_fs_closing(fs_info))
2887 return -EAGAIN;
2888
2889 /* At mount time we have ensured nodesize is in the range of [4K, 64K]. */
2890 ASSERT(fs_info->nodesize <= BTRFS_STRIPE_LEN);
2891
2892 /*
2893 * SCRUB_MAX_SECTORS_PER_BLOCK is calculated using the largest possible
2894 * value (max nodesize / min sectorsize), thus nodesize should always
2895 * be fine.
2896 */
2897 ASSERT(fs_info->nodesize <=
2898 SCRUB_MAX_SECTORS_PER_BLOCK << fs_info->sectorsize_bits);
2899
2900 /* Allocate outside of device_list_mutex */
2901 sctx = scrub_setup_ctx(fs_info, is_dev_replace);
2902 if (IS_ERR(sctx))
2903 return PTR_ERR(sctx);
2904
2905 ret = scrub_workers_get(fs_info);
2906 if (ret)
2907 goto out_free_ctx;
2908
2909 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2910 dev = btrfs_find_device(fs_info->fs_devices, &args);
2911 if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
2912 !is_dev_replace)) {
2913 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2914 ret = -ENODEV;
2915 goto out;
2916 }
2917
2918 if (!is_dev_replace && !readonly &&
2919 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
2920 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2921 btrfs_err_in_rcu(fs_info,
2922 "scrub on devid %llu: filesystem on %s is not writable",
2923 devid, btrfs_dev_name(dev));
2924 ret = -EROFS;
2925 goto out;
2926 }
2927
2928 mutex_lock(&fs_info->scrub_lock);
2929 if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
2930 test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
2931 mutex_unlock(&fs_info->scrub_lock);
2932 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2933 ret = -EIO;
2934 goto out;
2935 }
2936
2937 down_read(&fs_info->dev_replace.rwsem);
2938 if (dev->scrub_ctx ||
2939 (!is_dev_replace &&
2940 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
2941 up_read(&fs_info->dev_replace.rwsem);
2942 mutex_unlock(&fs_info->scrub_lock);
2943 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2944 ret = -EINPROGRESS;
2945 goto out;
2946 }
2947 up_read(&fs_info->dev_replace.rwsem);
2948
2949 sctx->readonly = readonly;
2950 dev->scrub_ctx = sctx;
2951 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2952
2953 /*
2954 * checking @scrub_pause_req here, we can avoid
2955 * race between committing transaction and scrubbing.
2956 */
2957 __scrub_blocked_if_needed(fs_info);
2958 atomic_inc(&fs_info->scrubs_running);
2959 mutex_unlock(&fs_info->scrub_lock);
2960
2961 /*
2962 * In order to avoid deadlock with reclaim when there is a transaction
2963 * trying to pause scrub, make sure we use GFP_NOFS for all the
2964 * allocations done at btrfs_scrub_sectors() and scrub_sectors_for_parity()
2965 * invoked by our callees. The pausing request is done when the
2966 * transaction commit starts, and it blocks the transaction until scrub
2967 * is paused (done at specific points at scrub_stripe() or right above
2968 * before incrementing fs_info->scrubs_running).
2969 */
2970 nofs_flag = memalloc_nofs_save();
2971 if (!is_dev_replace) {
2972 u64 old_super_errors;
2973
2974 spin_lock(&sctx->stat_lock);
2975 old_super_errors = sctx->stat.super_errors;
2976 spin_unlock(&sctx->stat_lock);
2977
2978 btrfs_info(fs_info, "scrub: started on devid %llu", devid);
2979 /*
2980 * by holding device list mutex, we can
2981 * kick off writing super in log tree sync.
2982 */
2983 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2984 ret = scrub_supers(sctx, dev);
2985 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2986
2987 spin_lock(&sctx->stat_lock);
2988 /*
2989 * Super block errors found, but we can not commit transaction
2990 * at current context, since btrfs_commit_transaction() needs
2991 * to pause the current running scrub (hold by ourselves).
2992 */
2993 if (sctx->stat.super_errors > old_super_errors && !sctx->readonly)
2994 need_commit = true;
2995 spin_unlock(&sctx->stat_lock);
2996 }
2997
2998 if (!ret)
2999 ret = scrub_enumerate_chunks(sctx, dev, start, end);
3000 memalloc_nofs_restore(nofs_flag);
3001
3002 atomic_dec(&fs_info->scrubs_running);
3003 wake_up(&fs_info->scrub_pause_wait);
3004
3005 if (progress)
3006 memcpy(progress, &sctx->stat, sizeof(*progress));
3007
3008 if (!is_dev_replace)
3009 btrfs_info(fs_info, "scrub: %s on devid %llu with status: %d",
3010 ret ? "not finished" : "finished", devid, ret);
3011
3012 mutex_lock(&fs_info->scrub_lock);
3013 dev->scrub_ctx = NULL;
3014 mutex_unlock(&fs_info->scrub_lock);
3015
3016 scrub_workers_put(fs_info);
3017 scrub_put_ctx(sctx);
3018
3019 /*
3020 * We found some super block errors before, now try to force a
3021 * transaction commit, as scrub has finished.
3022 */
3023 if (need_commit) {
3024 struct btrfs_trans_handle *trans;
3025
3026 trans = btrfs_start_transaction(fs_info->tree_root, 0);
3027 if (IS_ERR(trans)) {
3028 ret = PTR_ERR(trans);
3029 btrfs_err(fs_info,
3030 "scrub: failed to start transaction to fix super block errors: %d", ret);
3031 return ret;
3032 }
3033 ret = btrfs_commit_transaction(trans);
3034 if (ret < 0)
3035 btrfs_err(fs_info,
3036 "scrub: failed to commit transaction to fix super block errors: %d", ret);
3037 }
3038 return ret;
3039out:
3040 scrub_workers_put(fs_info);
3041out_free_ctx:
3042 scrub_free_ctx(sctx);
3043
3044 return ret;
3045}
3046
3047void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
3048{
3049 mutex_lock(&fs_info->scrub_lock);
3050 atomic_inc(&fs_info->scrub_pause_req);
3051 while (atomic_read(&fs_info->scrubs_paused) !=
3052 atomic_read(&fs_info->scrubs_running)) {
3053 mutex_unlock(&fs_info->scrub_lock);
3054 wait_event(fs_info->scrub_pause_wait,
3055 atomic_read(&fs_info->scrubs_paused) ==
3056 atomic_read(&fs_info->scrubs_running));
3057 mutex_lock(&fs_info->scrub_lock);
3058 }
3059 mutex_unlock(&fs_info->scrub_lock);
3060}
3061
3062void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
3063{
3064 atomic_dec(&fs_info->scrub_pause_req);
3065 wake_up(&fs_info->scrub_pause_wait);
3066}
3067
3068int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
3069{
3070 mutex_lock(&fs_info->scrub_lock);
3071 if (!atomic_read(&fs_info->scrubs_running)) {
3072 mutex_unlock(&fs_info->scrub_lock);
3073 return -ENOTCONN;
3074 }
3075
3076 atomic_inc(&fs_info->scrub_cancel_req);
3077 while (atomic_read(&fs_info->scrubs_running)) {
3078 mutex_unlock(&fs_info->scrub_lock);
3079 wait_event(fs_info->scrub_pause_wait,
3080 atomic_read(&fs_info->scrubs_running) == 0);
3081 mutex_lock(&fs_info->scrub_lock);
3082 }
3083 atomic_dec(&fs_info->scrub_cancel_req);
3084 mutex_unlock(&fs_info->scrub_lock);
3085
3086 return 0;
3087}
3088
3089int btrfs_scrub_cancel_dev(struct btrfs_device *dev)
3090{
3091 struct btrfs_fs_info *fs_info = dev->fs_info;
3092 struct scrub_ctx *sctx;
3093
3094 mutex_lock(&fs_info->scrub_lock);
3095 sctx = dev->scrub_ctx;
3096 if (!sctx) {
3097 mutex_unlock(&fs_info->scrub_lock);
3098 return -ENOTCONN;
3099 }
3100 atomic_inc(&sctx->cancel_req);
3101 while (dev->scrub_ctx) {
3102 mutex_unlock(&fs_info->scrub_lock);
3103 wait_event(fs_info->scrub_pause_wait,
3104 dev->scrub_ctx == NULL);
3105 mutex_lock(&fs_info->scrub_lock);
3106 }
3107 mutex_unlock(&fs_info->scrub_lock);
3108
3109 return 0;
3110}
3111
3112int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
3113 struct btrfs_scrub_progress *progress)
3114{
3115 struct btrfs_dev_lookup_args args = { .devid = devid };
3116 struct btrfs_device *dev;
3117 struct scrub_ctx *sctx = NULL;
3118
3119 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3120 dev = btrfs_find_device(fs_info->fs_devices, &args);
3121 if (dev)
3122 sctx = dev->scrub_ctx;
3123 if (sctx)
3124 memcpy(progress, &sctx->stat, sizeof(*progress));
3125 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3126
3127 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3128}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2011, 2012 STRATO. All rights reserved.
4 */
5
6#include <linux/blkdev.h>
7#include <linux/ratelimit.h>
8#include <linux/sched/mm.h>
9#include "ctree.h"
10#include "volumes.h"
11#include "disk-io.h"
12#include "ordered-data.h"
13#include "transaction.h"
14#include "backref.h"
15#include "extent_io.h"
16#include "dev-replace.h"
17#include "check-integrity.h"
18#include "rcu-string.h"
19#include "raid56.h"
20
21/*
22 * This is only the first step towards a full-features scrub. It reads all
23 * extent and super block and verifies the checksums. In case a bad checksum
24 * is found or the extent cannot be read, good data will be written back if
25 * any can be found.
26 *
27 * Future enhancements:
28 * - In case an unrepairable extent is encountered, track which files are
29 * affected and report them
30 * - track and record media errors, throw out bad devices
31 * - add a mode to also read unallocated space
32 */
33
34struct scrub_block;
35struct scrub_ctx;
36
37/*
38 * the following three values only influence the performance.
39 * The last one configures the number of parallel and outstanding I/O
40 * operations. The first two values configure an upper limit for the number
41 * of (dynamically allocated) pages that are added to a bio.
42 */
43#define SCRUB_PAGES_PER_RD_BIO 32 /* 128k per bio */
44#define SCRUB_PAGES_PER_WR_BIO 32 /* 128k per bio */
45#define SCRUB_BIOS_PER_SCTX 64 /* 8MB per device in flight */
46
47/*
48 * the following value times PAGE_SIZE needs to be large enough to match the
49 * largest node/leaf/sector size that shall be supported.
50 * Values larger than BTRFS_STRIPE_LEN are not supported.
51 */
52#define SCRUB_MAX_PAGES_PER_BLOCK 16 /* 64k per node/leaf/sector */
53
54struct scrub_recover {
55 refcount_t refs;
56 struct btrfs_bio *bbio;
57 u64 map_length;
58};
59
60struct scrub_page {
61 struct scrub_block *sblock;
62 struct page *page;
63 struct btrfs_device *dev;
64 struct list_head list;
65 u64 flags; /* extent flags */
66 u64 generation;
67 u64 logical;
68 u64 physical;
69 u64 physical_for_dev_replace;
70 atomic_t refs;
71 struct {
72 unsigned int mirror_num:8;
73 unsigned int have_csum:1;
74 unsigned int io_error:1;
75 };
76 u8 csum[BTRFS_CSUM_SIZE];
77
78 struct scrub_recover *recover;
79};
80
81struct scrub_bio {
82 int index;
83 struct scrub_ctx *sctx;
84 struct btrfs_device *dev;
85 struct bio *bio;
86 blk_status_t status;
87 u64 logical;
88 u64 physical;
89#if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
90 struct scrub_page *pagev[SCRUB_PAGES_PER_WR_BIO];
91#else
92 struct scrub_page *pagev[SCRUB_PAGES_PER_RD_BIO];
93#endif
94 int page_count;
95 int next_free;
96 struct btrfs_work work;
97};
98
99struct scrub_block {
100 struct scrub_page *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
101 int page_count;
102 atomic_t outstanding_pages;
103 refcount_t refs; /* free mem on transition to zero */
104 struct scrub_ctx *sctx;
105 struct scrub_parity *sparity;
106 struct {
107 unsigned int header_error:1;
108 unsigned int checksum_error:1;
109 unsigned int no_io_error_seen:1;
110 unsigned int generation_error:1; /* also sets header_error */
111
112 /* The following is for the data used to check parity */
113 /* It is for the data with checksum */
114 unsigned int data_corrected:1;
115 };
116 struct btrfs_work work;
117};
118
119/* Used for the chunks with parity stripe such RAID5/6 */
120struct scrub_parity {
121 struct scrub_ctx *sctx;
122
123 struct btrfs_device *scrub_dev;
124
125 u64 logic_start;
126
127 u64 logic_end;
128
129 int nsectors;
130
131 u64 stripe_len;
132
133 refcount_t refs;
134
135 struct list_head spages;
136
137 /* Work of parity check and repair */
138 struct btrfs_work work;
139
140 /* Mark the parity blocks which have data */
141 unsigned long *dbitmap;
142
143 /*
144 * Mark the parity blocks which have data, but errors happen when
145 * read data or check data
146 */
147 unsigned long *ebitmap;
148
149 unsigned long bitmap[0];
150};
151
152struct scrub_ctx {
153 struct scrub_bio *bios[SCRUB_BIOS_PER_SCTX];
154 struct btrfs_fs_info *fs_info;
155 int first_free;
156 int curr;
157 atomic_t bios_in_flight;
158 atomic_t workers_pending;
159 spinlock_t list_lock;
160 wait_queue_head_t list_wait;
161 u16 csum_size;
162 struct list_head csum_list;
163 atomic_t cancel_req;
164 int readonly;
165 int pages_per_rd_bio;
166
167 int is_dev_replace;
168
169 struct scrub_bio *wr_curr_bio;
170 struct mutex wr_lock;
171 int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
172 struct btrfs_device *wr_tgtdev;
173 bool flush_all_writes;
174
175 /*
176 * statistics
177 */
178 struct btrfs_scrub_progress stat;
179 spinlock_t stat_lock;
180
181 /*
182 * Use a ref counter to avoid use-after-free issues. Scrub workers
183 * decrement bios_in_flight and workers_pending and then do a wakeup
184 * on the list_wait wait queue. We must ensure the main scrub task
185 * doesn't free the scrub context before or while the workers are
186 * doing the wakeup() call.
187 */
188 refcount_t refs;
189};
190
191struct scrub_fixup_nodatasum {
192 struct scrub_ctx *sctx;
193 struct btrfs_device *dev;
194 u64 logical;
195 struct btrfs_root *root;
196 struct btrfs_work work;
197 int mirror_num;
198};
199
200struct scrub_nocow_inode {
201 u64 inum;
202 u64 offset;
203 u64 root;
204 struct list_head list;
205};
206
207struct scrub_copy_nocow_ctx {
208 struct scrub_ctx *sctx;
209 u64 logical;
210 u64 len;
211 int mirror_num;
212 u64 physical_for_dev_replace;
213 struct list_head inodes;
214 struct btrfs_work work;
215};
216
217struct scrub_warning {
218 struct btrfs_path *path;
219 u64 extent_item_size;
220 const char *errstr;
221 u64 physical;
222 u64 logical;
223 struct btrfs_device *dev;
224};
225
226struct full_stripe_lock {
227 struct rb_node node;
228 u64 logical;
229 u64 refs;
230 struct mutex mutex;
231};
232
233static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
234static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
235static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
236static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
237static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
238static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
239 struct scrub_block *sblocks_for_recheck);
240static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
241 struct scrub_block *sblock,
242 int retry_failed_mirror);
243static void scrub_recheck_block_checksum(struct scrub_block *sblock);
244static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
245 struct scrub_block *sblock_good);
246static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
247 struct scrub_block *sblock_good,
248 int page_num, int force_write);
249static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
250static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
251 int page_num);
252static int scrub_checksum_data(struct scrub_block *sblock);
253static int scrub_checksum_tree_block(struct scrub_block *sblock);
254static int scrub_checksum_super(struct scrub_block *sblock);
255static void scrub_block_get(struct scrub_block *sblock);
256static void scrub_block_put(struct scrub_block *sblock);
257static void scrub_page_get(struct scrub_page *spage);
258static void scrub_page_put(struct scrub_page *spage);
259static void scrub_parity_get(struct scrub_parity *sparity);
260static void scrub_parity_put(struct scrub_parity *sparity);
261static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
262 struct scrub_page *spage);
263static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
264 u64 physical, struct btrfs_device *dev, u64 flags,
265 u64 gen, int mirror_num, u8 *csum, int force,
266 u64 physical_for_dev_replace);
267static void scrub_bio_end_io(struct bio *bio);
268static void scrub_bio_end_io_worker(struct btrfs_work *work);
269static void scrub_block_complete(struct scrub_block *sblock);
270static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
271 u64 extent_logical, u64 extent_len,
272 u64 *extent_physical,
273 struct btrfs_device **extent_dev,
274 int *extent_mirror_num);
275static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
276 struct scrub_page *spage);
277static void scrub_wr_submit(struct scrub_ctx *sctx);
278static void scrub_wr_bio_end_io(struct bio *bio);
279static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
280static int write_page_nocow(struct scrub_ctx *sctx,
281 u64 physical_for_dev_replace, struct page *page);
282static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
283 struct scrub_copy_nocow_ctx *ctx);
284static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
285 int mirror_num, u64 physical_for_dev_replace);
286static void copy_nocow_pages_worker(struct btrfs_work *work);
287static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
288static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
289static void scrub_put_ctx(struct scrub_ctx *sctx);
290
291static inline int scrub_is_page_on_raid56(struct scrub_page *page)
292{
293 return page->recover &&
294 (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
295}
296
297static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
298{
299 refcount_inc(&sctx->refs);
300 atomic_inc(&sctx->bios_in_flight);
301}
302
303static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
304{
305 atomic_dec(&sctx->bios_in_flight);
306 wake_up(&sctx->list_wait);
307 scrub_put_ctx(sctx);
308}
309
310static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
311{
312 while (atomic_read(&fs_info->scrub_pause_req)) {
313 mutex_unlock(&fs_info->scrub_lock);
314 wait_event(fs_info->scrub_pause_wait,
315 atomic_read(&fs_info->scrub_pause_req) == 0);
316 mutex_lock(&fs_info->scrub_lock);
317 }
318}
319
320static void scrub_pause_on(struct btrfs_fs_info *fs_info)
321{
322 atomic_inc(&fs_info->scrubs_paused);
323 wake_up(&fs_info->scrub_pause_wait);
324}
325
326static void scrub_pause_off(struct btrfs_fs_info *fs_info)
327{
328 mutex_lock(&fs_info->scrub_lock);
329 __scrub_blocked_if_needed(fs_info);
330 atomic_dec(&fs_info->scrubs_paused);
331 mutex_unlock(&fs_info->scrub_lock);
332
333 wake_up(&fs_info->scrub_pause_wait);
334}
335
336static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
337{
338 scrub_pause_on(fs_info);
339 scrub_pause_off(fs_info);
340}
341
342/*
343 * Insert new full stripe lock into full stripe locks tree
344 *
345 * Return pointer to existing or newly inserted full_stripe_lock structure if
346 * everything works well.
347 * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
348 *
349 * NOTE: caller must hold full_stripe_locks_root->lock before calling this
350 * function
351 */
352static struct full_stripe_lock *insert_full_stripe_lock(
353 struct btrfs_full_stripe_locks_tree *locks_root,
354 u64 fstripe_logical)
355{
356 struct rb_node **p;
357 struct rb_node *parent = NULL;
358 struct full_stripe_lock *entry;
359 struct full_stripe_lock *ret;
360
361 lockdep_assert_held(&locks_root->lock);
362
363 p = &locks_root->root.rb_node;
364 while (*p) {
365 parent = *p;
366 entry = rb_entry(parent, struct full_stripe_lock, node);
367 if (fstripe_logical < entry->logical) {
368 p = &(*p)->rb_left;
369 } else if (fstripe_logical > entry->logical) {
370 p = &(*p)->rb_right;
371 } else {
372 entry->refs++;
373 return entry;
374 }
375 }
376
377 /* Insert new lock */
378 ret = kmalloc(sizeof(*ret), GFP_KERNEL);
379 if (!ret)
380 return ERR_PTR(-ENOMEM);
381 ret->logical = fstripe_logical;
382 ret->refs = 1;
383 mutex_init(&ret->mutex);
384
385 rb_link_node(&ret->node, parent, p);
386 rb_insert_color(&ret->node, &locks_root->root);
387 return ret;
388}
389
390/*
391 * Search for a full stripe lock of a block group
392 *
393 * Return pointer to existing full stripe lock if found
394 * Return NULL if not found
395 */
396static struct full_stripe_lock *search_full_stripe_lock(
397 struct btrfs_full_stripe_locks_tree *locks_root,
398 u64 fstripe_logical)
399{
400 struct rb_node *node;
401 struct full_stripe_lock *entry;
402
403 lockdep_assert_held(&locks_root->lock);
404
405 node = locks_root->root.rb_node;
406 while (node) {
407 entry = rb_entry(node, struct full_stripe_lock, node);
408 if (fstripe_logical < entry->logical)
409 node = node->rb_left;
410 else if (fstripe_logical > entry->logical)
411 node = node->rb_right;
412 else
413 return entry;
414 }
415 return NULL;
416}
417
418/*
419 * Helper to get full stripe logical from a normal bytenr.
420 *
421 * Caller must ensure @cache is a RAID56 block group.
422 */
423static u64 get_full_stripe_logical(struct btrfs_block_group_cache *cache,
424 u64 bytenr)
425{
426 u64 ret;
427
428 /*
429 * Due to chunk item size limit, full stripe length should not be
430 * larger than U32_MAX. Just a sanity check here.
431 */
432 WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
433
434 /*
435 * round_down() can only handle power of 2, while RAID56 full
436 * stripe length can be 64KiB * n, so we need to manually round down.
437 */
438 ret = div64_u64(bytenr - cache->key.objectid, cache->full_stripe_len) *
439 cache->full_stripe_len + cache->key.objectid;
440 return ret;
441}
442
443/*
444 * Lock a full stripe to avoid concurrency of recovery and read
445 *
446 * It's only used for profiles with parities (RAID5/6), for other profiles it
447 * does nothing.
448 *
449 * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
450 * So caller must call unlock_full_stripe() at the same context.
451 *
452 * Return <0 if encounters error.
453 */
454static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
455 bool *locked_ret)
456{
457 struct btrfs_block_group_cache *bg_cache;
458 struct btrfs_full_stripe_locks_tree *locks_root;
459 struct full_stripe_lock *existing;
460 u64 fstripe_start;
461 int ret = 0;
462
463 *locked_ret = false;
464 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
465 if (!bg_cache) {
466 ASSERT(0);
467 return -ENOENT;
468 }
469
470 /* Profiles not based on parity don't need full stripe lock */
471 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
472 goto out;
473 locks_root = &bg_cache->full_stripe_locks_root;
474
475 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
476
477 /* Now insert the full stripe lock */
478 mutex_lock(&locks_root->lock);
479 existing = insert_full_stripe_lock(locks_root, fstripe_start);
480 mutex_unlock(&locks_root->lock);
481 if (IS_ERR(existing)) {
482 ret = PTR_ERR(existing);
483 goto out;
484 }
485 mutex_lock(&existing->mutex);
486 *locked_ret = true;
487out:
488 btrfs_put_block_group(bg_cache);
489 return ret;
490}
491
492/*
493 * Unlock a full stripe.
494 *
495 * NOTE: Caller must ensure it's the same context calling corresponding
496 * lock_full_stripe().
497 *
498 * Return 0 if we unlock full stripe without problem.
499 * Return <0 for error
500 */
501static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
502 bool locked)
503{
504 struct btrfs_block_group_cache *bg_cache;
505 struct btrfs_full_stripe_locks_tree *locks_root;
506 struct full_stripe_lock *fstripe_lock;
507 u64 fstripe_start;
508 bool freeit = false;
509 int ret = 0;
510
511 /* If we didn't acquire full stripe lock, no need to continue */
512 if (!locked)
513 return 0;
514
515 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
516 if (!bg_cache) {
517 ASSERT(0);
518 return -ENOENT;
519 }
520 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
521 goto out;
522
523 locks_root = &bg_cache->full_stripe_locks_root;
524 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
525
526 mutex_lock(&locks_root->lock);
527 fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
528 /* Unpaired unlock_full_stripe() detected */
529 if (!fstripe_lock) {
530 WARN_ON(1);
531 ret = -ENOENT;
532 mutex_unlock(&locks_root->lock);
533 goto out;
534 }
535
536 if (fstripe_lock->refs == 0) {
537 WARN_ON(1);
538 btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
539 fstripe_lock->logical);
540 } else {
541 fstripe_lock->refs--;
542 }
543
544 if (fstripe_lock->refs == 0) {
545 rb_erase(&fstripe_lock->node, &locks_root->root);
546 freeit = true;
547 }
548 mutex_unlock(&locks_root->lock);
549
550 mutex_unlock(&fstripe_lock->mutex);
551 if (freeit)
552 kfree(fstripe_lock);
553out:
554 btrfs_put_block_group(bg_cache);
555 return ret;
556}
557
558/*
559 * used for workers that require transaction commits (i.e., for the
560 * NOCOW case)
561 */
562static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
563{
564 struct btrfs_fs_info *fs_info = sctx->fs_info;
565
566 refcount_inc(&sctx->refs);
567 /*
568 * increment scrubs_running to prevent cancel requests from
569 * completing as long as a worker is running. we must also
570 * increment scrubs_paused to prevent deadlocking on pause
571 * requests used for transactions commits (as the worker uses a
572 * transaction context). it is safe to regard the worker
573 * as paused for all matters practical. effectively, we only
574 * avoid cancellation requests from completing.
575 */
576 mutex_lock(&fs_info->scrub_lock);
577 atomic_inc(&fs_info->scrubs_running);
578 atomic_inc(&fs_info->scrubs_paused);
579 mutex_unlock(&fs_info->scrub_lock);
580
581 /*
582 * check if @scrubs_running=@scrubs_paused condition
583 * inside wait_event() is not an atomic operation.
584 * which means we may inc/dec @scrub_running/paused
585 * at any time. Let's wake up @scrub_pause_wait as
586 * much as we can to let commit transaction blocked less.
587 */
588 wake_up(&fs_info->scrub_pause_wait);
589
590 atomic_inc(&sctx->workers_pending);
591}
592
593/* used for workers that require transaction commits */
594static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
595{
596 struct btrfs_fs_info *fs_info = sctx->fs_info;
597
598 /*
599 * see scrub_pending_trans_workers_inc() why we're pretending
600 * to be paused in the scrub counters
601 */
602 mutex_lock(&fs_info->scrub_lock);
603 atomic_dec(&fs_info->scrubs_running);
604 atomic_dec(&fs_info->scrubs_paused);
605 mutex_unlock(&fs_info->scrub_lock);
606 atomic_dec(&sctx->workers_pending);
607 wake_up(&fs_info->scrub_pause_wait);
608 wake_up(&sctx->list_wait);
609 scrub_put_ctx(sctx);
610}
611
612static void scrub_free_csums(struct scrub_ctx *sctx)
613{
614 while (!list_empty(&sctx->csum_list)) {
615 struct btrfs_ordered_sum *sum;
616 sum = list_first_entry(&sctx->csum_list,
617 struct btrfs_ordered_sum, list);
618 list_del(&sum->list);
619 kfree(sum);
620 }
621}
622
623static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
624{
625 int i;
626
627 if (!sctx)
628 return;
629
630 /* this can happen when scrub is cancelled */
631 if (sctx->curr != -1) {
632 struct scrub_bio *sbio = sctx->bios[sctx->curr];
633
634 for (i = 0; i < sbio->page_count; i++) {
635 WARN_ON(!sbio->pagev[i]->page);
636 scrub_block_put(sbio->pagev[i]->sblock);
637 }
638 bio_put(sbio->bio);
639 }
640
641 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
642 struct scrub_bio *sbio = sctx->bios[i];
643
644 if (!sbio)
645 break;
646 kfree(sbio);
647 }
648
649 kfree(sctx->wr_curr_bio);
650 scrub_free_csums(sctx);
651 kfree(sctx);
652}
653
654static void scrub_put_ctx(struct scrub_ctx *sctx)
655{
656 if (refcount_dec_and_test(&sctx->refs))
657 scrub_free_ctx(sctx);
658}
659
660static noinline_for_stack
661struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
662{
663 struct scrub_ctx *sctx;
664 int i;
665 struct btrfs_fs_info *fs_info = dev->fs_info;
666
667 sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
668 if (!sctx)
669 goto nomem;
670 refcount_set(&sctx->refs, 1);
671 sctx->is_dev_replace = is_dev_replace;
672 sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
673 sctx->curr = -1;
674 sctx->fs_info = dev->fs_info;
675 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
676 struct scrub_bio *sbio;
677
678 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
679 if (!sbio)
680 goto nomem;
681 sctx->bios[i] = sbio;
682
683 sbio->index = i;
684 sbio->sctx = sctx;
685 sbio->page_count = 0;
686 btrfs_init_work(&sbio->work, btrfs_scrub_helper,
687 scrub_bio_end_io_worker, NULL, NULL);
688
689 if (i != SCRUB_BIOS_PER_SCTX - 1)
690 sctx->bios[i]->next_free = i + 1;
691 else
692 sctx->bios[i]->next_free = -1;
693 }
694 sctx->first_free = 0;
695 atomic_set(&sctx->bios_in_flight, 0);
696 atomic_set(&sctx->workers_pending, 0);
697 atomic_set(&sctx->cancel_req, 0);
698 sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
699 INIT_LIST_HEAD(&sctx->csum_list);
700
701 spin_lock_init(&sctx->list_lock);
702 spin_lock_init(&sctx->stat_lock);
703 init_waitqueue_head(&sctx->list_wait);
704
705 WARN_ON(sctx->wr_curr_bio != NULL);
706 mutex_init(&sctx->wr_lock);
707 sctx->wr_curr_bio = NULL;
708 if (is_dev_replace) {
709 WARN_ON(!fs_info->dev_replace.tgtdev);
710 sctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
711 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
712 sctx->flush_all_writes = false;
713 }
714
715 return sctx;
716
717nomem:
718 scrub_free_ctx(sctx);
719 return ERR_PTR(-ENOMEM);
720}
721
722static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
723 void *warn_ctx)
724{
725 u64 isize;
726 u32 nlink;
727 int ret;
728 int i;
729 unsigned nofs_flag;
730 struct extent_buffer *eb;
731 struct btrfs_inode_item *inode_item;
732 struct scrub_warning *swarn = warn_ctx;
733 struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
734 struct inode_fs_paths *ipath = NULL;
735 struct btrfs_root *local_root;
736 struct btrfs_key root_key;
737 struct btrfs_key key;
738
739 root_key.objectid = root;
740 root_key.type = BTRFS_ROOT_ITEM_KEY;
741 root_key.offset = (u64)-1;
742 local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
743 if (IS_ERR(local_root)) {
744 ret = PTR_ERR(local_root);
745 goto err;
746 }
747
748 /*
749 * this makes the path point to (inum INODE_ITEM ioff)
750 */
751 key.objectid = inum;
752 key.type = BTRFS_INODE_ITEM_KEY;
753 key.offset = 0;
754
755 ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
756 if (ret) {
757 btrfs_release_path(swarn->path);
758 goto err;
759 }
760
761 eb = swarn->path->nodes[0];
762 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
763 struct btrfs_inode_item);
764 isize = btrfs_inode_size(eb, inode_item);
765 nlink = btrfs_inode_nlink(eb, inode_item);
766 btrfs_release_path(swarn->path);
767
768 /*
769 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
770 * uses GFP_NOFS in this context, so we keep it consistent but it does
771 * not seem to be strictly necessary.
772 */
773 nofs_flag = memalloc_nofs_save();
774 ipath = init_ipath(4096, local_root, swarn->path);
775 memalloc_nofs_restore(nofs_flag);
776 if (IS_ERR(ipath)) {
777 ret = PTR_ERR(ipath);
778 ipath = NULL;
779 goto err;
780 }
781 ret = paths_from_inode(inum, ipath);
782
783 if (ret < 0)
784 goto err;
785
786 /*
787 * we deliberately ignore the bit ipath might have been too small to
788 * hold all of the paths here
789 */
790 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
791 btrfs_warn_in_rcu(fs_info,
792"%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
793 swarn->errstr, swarn->logical,
794 rcu_str_deref(swarn->dev->name),
795 swarn->physical,
796 root, inum, offset,
797 min(isize - offset, (u64)PAGE_SIZE), nlink,
798 (char *)(unsigned long)ipath->fspath->val[i]);
799
800 free_ipath(ipath);
801 return 0;
802
803err:
804 btrfs_warn_in_rcu(fs_info,
805 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
806 swarn->errstr, swarn->logical,
807 rcu_str_deref(swarn->dev->name),
808 swarn->physical,
809 root, inum, offset, ret);
810
811 free_ipath(ipath);
812 return 0;
813}
814
815static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
816{
817 struct btrfs_device *dev;
818 struct btrfs_fs_info *fs_info;
819 struct btrfs_path *path;
820 struct btrfs_key found_key;
821 struct extent_buffer *eb;
822 struct btrfs_extent_item *ei;
823 struct scrub_warning swarn;
824 unsigned long ptr = 0;
825 u64 extent_item_pos;
826 u64 flags = 0;
827 u64 ref_root;
828 u32 item_size;
829 u8 ref_level = 0;
830 int ret;
831
832 WARN_ON(sblock->page_count < 1);
833 dev = sblock->pagev[0]->dev;
834 fs_info = sblock->sctx->fs_info;
835
836 path = btrfs_alloc_path();
837 if (!path)
838 return;
839
840 swarn.physical = sblock->pagev[0]->physical;
841 swarn.logical = sblock->pagev[0]->logical;
842 swarn.errstr = errstr;
843 swarn.dev = NULL;
844
845 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
846 &flags);
847 if (ret < 0)
848 goto out;
849
850 extent_item_pos = swarn.logical - found_key.objectid;
851 swarn.extent_item_size = found_key.offset;
852
853 eb = path->nodes[0];
854 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
855 item_size = btrfs_item_size_nr(eb, path->slots[0]);
856
857 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
858 do {
859 ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
860 item_size, &ref_root,
861 &ref_level);
862 btrfs_warn_in_rcu(fs_info,
863"%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
864 errstr, swarn.logical,
865 rcu_str_deref(dev->name),
866 swarn.physical,
867 ref_level ? "node" : "leaf",
868 ret < 0 ? -1 : ref_level,
869 ret < 0 ? -1 : ref_root);
870 } while (ret != 1);
871 btrfs_release_path(path);
872 } else {
873 btrfs_release_path(path);
874 swarn.path = path;
875 swarn.dev = dev;
876 iterate_extent_inodes(fs_info, found_key.objectid,
877 extent_item_pos, 1,
878 scrub_print_warning_inode, &swarn, false);
879 }
880
881out:
882 btrfs_free_path(path);
883}
884
885static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
886{
887 struct page *page = NULL;
888 unsigned long index;
889 struct scrub_fixup_nodatasum *fixup = fixup_ctx;
890 int ret;
891 int corrected = 0;
892 struct btrfs_key key;
893 struct inode *inode = NULL;
894 struct btrfs_fs_info *fs_info;
895 u64 end = offset + PAGE_SIZE - 1;
896 struct btrfs_root *local_root;
897 int srcu_index;
898
899 key.objectid = root;
900 key.type = BTRFS_ROOT_ITEM_KEY;
901 key.offset = (u64)-1;
902
903 fs_info = fixup->root->fs_info;
904 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
905
906 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
907 if (IS_ERR(local_root)) {
908 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
909 return PTR_ERR(local_root);
910 }
911
912 key.type = BTRFS_INODE_ITEM_KEY;
913 key.objectid = inum;
914 key.offset = 0;
915 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
916 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
917 if (IS_ERR(inode))
918 return PTR_ERR(inode);
919
920 index = offset >> PAGE_SHIFT;
921
922 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
923 if (!page) {
924 ret = -ENOMEM;
925 goto out;
926 }
927
928 if (PageUptodate(page)) {
929 if (PageDirty(page)) {
930 /*
931 * we need to write the data to the defect sector. the
932 * data that was in that sector is not in memory,
933 * because the page was modified. we must not write the
934 * modified page to that sector.
935 *
936 * TODO: what could be done here: wait for the delalloc
937 * runner to write out that page (might involve
938 * COW) and see whether the sector is still
939 * referenced afterwards.
940 *
941 * For the meantime, we'll treat this error
942 * incorrectable, although there is a chance that a
943 * later scrub will find the bad sector again and that
944 * there's no dirty page in memory, then.
945 */
946 ret = -EIO;
947 goto out;
948 }
949 ret = repair_io_failure(fs_info, inum, offset, PAGE_SIZE,
950 fixup->logical, page,
951 offset - page_offset(page),
952 fixup->mirror_num);
953 unlock_page(page);
954 corrected = !ret;
955 } else {
956 /*
957 * we need to get good data first. the general readpage path
958 * will call repair_io_failure for us, we just have to make
959 * sure we read the bad mirror.
960 */
961 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
962 EXTENT_DAMAGED);
963 if (ret) {
964 /* set_extent_bits should give proper error */
965 WARN_ON(ret > 0);
966 if (ret > 0)
967 ret = -EFAULT;
968 goto out;
969 }
970
971 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
972 btrfs_get_extent,
973 fixup->mirror_num);
974 wait_on_page_locked(page);
975
976 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
977 end, EXTENT_DAMAGED, 0, NULL);
978 if (!corrected)
979 clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
980 EXTENT_DAMAGED);
981 }
982
983out:
984 if (page)
985 put_page(page);
986
987 iput(inode);
988
989 if (ret < 0)
990 return ret;
991
992 if (ret == 0 && corrected) {
993 /*
994 * we only need to call readpage for one of the inodes belonging
995 * to this extent. so make iterate_extent_inodes stop
996 */
997 return 1;
998 }
999
1000 return -EIO;
1001}
1002
1003static void scrub_fixup_nodatasum(struct btrfs_work *work)
1004{
1005 struct btrfs_fs_info *fs_info;
1006 int ret;
1007 struct scrub_fixup_nodatasum *fixup;
1008 struct scrub_ctx *sctx;
1009 struct btrfs_trans_handle *trans = NULL;
1010 struct btrfs_path *path;
1011 int uncorrectable = 0;
1012
1013 fixup = container_of(work, struct scrub_fixup_nodatasum, work);
1014 sctx = fixup->sctx;
1015 fs_info = fixup->root->fs_info;
1016
1017 path = btrfs_alloc_path();
1018 if (!path) {
1019 spin_lock(&sctx->stat_lock);
1020 ++sctx->stat.malloc_errors;
1021 spin_unlock(&sctx->stat_lock);
1022 uncorrectable = 1;
1023 goto out;
1024 }
1025
1026 trans = btrfs_join_transaction(fixup->root);
1027 if (IS_ERR(trans)) {
1028 uncorrectable = 1;
1029 goto out;
1030 }
1031
1032 /*
1033 * the idea is to trigger a regular read through the standard path. we
1034 * read a page from the (failed) logical address by specifying the
1035 * corresponding copynum of the failed sector. thus, that readpage is
1036 * expected to fail.
1037 * that is the point where on-the-fly error correction will kick in
1038 * (once it's finished) and rewrite the failed sector if a good copy
1039 * can be found.
1040 */
1041 ret = iterate_inodes_from_logical(fixup->logical, fs_info, path,
1042 scrub_fixup_readpage, fixup, false);
1043 if (ret < 0) {
1044 uncorrectable = 1;
1045 goto out;
1046 }
1047 WARN_ON(ret != 1);
1048
1049 spin_lock(&sctx->stat_lock);
1050 ++sctx->stat.corrected_errors;
1051 spin_unlock(&sctx->stat_lock);
1052
1053out:
1054 if (trans && !IS_ERR(trans))
1055 btrfs_end_transaction(trans);
1056 if (uncorrectable) {
1057 spin_lock(&sctx->stat_lock);
1058 ++sctx->stat.uncorrectable_errors;
1059 spin_unlock(&sctx->stat_lock);
1060 btrfs_dev_replace_stats_inc(
1061 &fs_info->dev_replace.num_uncorrectable_read_errors);
1062 btrfs_err_rl_in_rcu(fs_info,
1063 "unable to fixup (nodatasum) error at logical %llu on dev %s",
1064 fixup->logical, rcu_str_deref(fixup->dev->name));
1065 }
1066
1067 btrfs_free_path(path);
1068 kfree(fixup);
1069
1070 scrub_pending_trans_workers_dec(sctx);
1071}
1072
1073static inline void scrub_get_recover(struct scrub_recover *recover)
1074{
1075 refcount_inc(&recover->refs);
1076}
1077
1078static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
1079 struct scrub_recover *recover)
1080{
1081 if (refcount_dec_and_test(&recover->refs)) {
1082 btrfs_bio_counter_dec(fs_info);
1083 btrfs_put_bbio(recover->bbio);
1084 kfree(recover);
1085 }
1086}
1087
1088/*
1089 * scrub_handle_errored_block gets called when either verification of the
1090 * pages failed or the bio failed to read, e.g. with EIO. In the latter
1091 * case, this function handles all pages in the bio, even though only one
1092 * may be bad.
1093 * The goal of this function is to repair the errored block by using the
1094 * contents of one of the mirrors.
1095 */
1096static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
1097{
1098 struct scrub_ctx *sctx = sblock_to_check->sctx;
1099 struct btrfs_device *dev;
1100 struct btrfs_fs_info *fs_info;
1101 u64 logical;
1102 unsigned int failed_mirror_index;
1103 unsigned int is_metadata;
1104 unsigned int have_csum;
1105 struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
1106 struct scrub_block *sblock_bad;
1107 int ret;
1108 int mirror_index;
1109 int page_num;
1110 int success;
1111 bool full_stripe_locked;
1112 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
1113 DEFAULT_RATELIMIT_BURST);
1114
1115 BUG_ON(sblock_to_check->page_count < 1);
1116 fs_info = sctx->fs_info;
1117 if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
1118 /*
1119 * if we find an error in a super block, we just report it.
1120 * They will get written with the next transaction commit
1121 * anyway
1122 */
1123 spin_lock(&sctx->stat_lock);
1124 ++sctx->stat.super_errors;
1125 spin_unlock(&sctx->stat_lock);
1126 return 0;
1127 }
1128 logical = sblock_to_check->pagev[0]->logical;
1129 BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
1130 failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
1131 is_metadata = !(sblock_to_check->pagev[0]->flags &
1132 BTRFS_EXTENT_FLAG_DATA);
1133 have_csum = sblock_to_check->pagev[0]->have_csum;
1134 dev = sblock_to_check->pagev[0]->dev;
1135
1136 /*
1137 * For RAID5/6, race can happen for a different device scrub thread.
1138 * For data corruption, Parity and Data threads will both try
1139 * to recovery the data.
1140 * Race can lead to doubly added csum error, or even unrecoverable
1141 * error.
1142 */
1143 ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
1144 if (ret < 0) {
1145 spin_lock(&sctx->stat_lock);
1146 if (ret == -ENOMEM)
1147 sctx->stat.malloc_errors++;
1148 sctx->stat.read_errors++;
1149 sctx->stat.uncorrectable_errors++;
1150 spin_unlock(&sctx->stat_lock);
1151 return ret;
1152 }
1153
1154 if (sctx->is_dev_replace && !is_metadata && !have_csum) {
1155 sblocks_for_recheck = NULL;
1156 goto nodatasum_case;
1157 }
1158
1159 /*
1160 * read all mirrors one after the other. This includes to
1161 * re-read the extent or metadata block that failed (that was
1162 * the cause that this fixup code is called) another time,
1163 * page by page this time in order to know which pages
1164 * caused I/O errors and which ones are good (for all mirrors).
1165 * It is the goal to handle the situation when more than one
1166 * mirror contains I/O errors, but the errors do not
1167 * overlap, i.e. the data can be repaired by selecting the
1168 * pages from those mirrors without I/O error on the
1169 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
1170 * would be that mirror #1 has an I/O error on the first page,
1171 * the second page is good, and mirror #2 has an I/O error on
1172 * the second page, but the first page is good.
1173 * Then the first page of the first mirror can be repaired by
1174 * taking the first page of the second mirror, and the
1175 * second page of the second mirror can be repaired by
1176 * copying the contents of the 2nd page of the 1st mirror.
1177 * One more note: if the pages of one mirror contain I/O
1178 * errors, the checksum cannot be verified. In order to get
1179 * the best data for repairing, the first attempt is to find
1180 * a mirror without I/O errors and with a validated checksum.
1181 * Only if this is not possible, the pages are picked from
1182 * mirrors with I/O errors without considering the checksum.
1183 * If the latter is the case, at the end, the checksum of the
1184 * repaired area is verified in order to correctly maintain
1185 * the statistics.
1186 */
1187
1188 sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
1189 sizeof(*sblocks_for_recheck), GFP_NOFS);
1190 if (!sblocks_for_recheck) {
1191 spin_lock(&sctx->stat_lock);
1192 sctx->stat.malloc_errors++;
1193 sctx->stat.read_errors++;
1194 sctx->stat.uncorrectable_errors++;
1195 spin_unlock(&sctx->stat_lock);
1196 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1197 goto out;
1198 }
1199
1200 /* setup the context, map the logical blocks and alloc the pages */
1201 ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
1202 if (ret) {
1203 spin_lock(&sctx->stat_lock);
1204 sctx->stat.read_errors++;
1205 sctx->stat.uncorrectable_errors++;
1206 spin_unlock(&sctx->stat_lock);
1207 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1208 goto out;
1209 }
1210 BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
1211 sblock_bad = sblocks_for_recheck + failed_mirror_index;
1212
1213 /* build and submit the bios for the failed mirror, check checksums */
1214 scrub_recheck_block(fs_info, sblock_bad, 1);
1215
1216 if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
1217 sblock_bad->no_io_error_seen) {
1218 /*
1219 * the error disappeared after reading page by page, or
1220 * the area was part of a huge bio and other parts of the
1221 * bio caused I/O errors, or the block layer merged several
1222 * read requests into one and the error is caused by a
1223 * different bio (usually one of the two latter cases is
1224 * the cause)
1225 */
1226 spin_lock(&sctx->stat_lock);
1227 sctx->stat.unverified_errors++;
1228 sblock_to_check->data_corrected = 1;
1229 spin_unlock(&sctx->stat_lock);
1230
1231 if (sctx->is_dev_replace)
1232 scrub_write_block_to_dev_replace(sblock_bad);
1233 goto out;
1234 }
1235
1236 if (!sblock_bad->no_io_error_seen) {
1237 spin_lock(&sctx->stat_lock);
1238 sctx->stat.read_errors++;
1239 spin_unlock(&sctx->stat_lock);
1240 if (__ratelimit(&_rs))
1241 scrub_print_warning("i/o error", sblock_to_check);
1242 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1243 } else if (sblock_bad->checksum_error) {
1244 spin_lock(&sctx->stat_lock);
1245 sctx->stat.csum_errors++;
1246 spin_unlock(&sctx->stat_lock);
1247 if (__ratelimit(&_rs))
1248 scrub_print_warning("checksum error", sblock_to_check);
1249 btrfs_dev_stat_inc_and_print(dev,
1250 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1251 } else if (sblock_bad->header_error) {
1252 spin_lock(&sctx->stat_lock);
1253 sctx->stat.verify_errors++;
1254 spin_unlock(&sctx->stat_lock);
1255 if (__ratelimit(&_rs))
1256 scrub_print_warning("checksum/header error",
1257 sblock_to_check);
1258 if (sblock_bad->generation_error)
1259 btrfs_dev_stat_inc_and_print(dev,
1260 BTRFS_DEV_STAT_GENERATION_ERRS);
1261 else
1262 btrfs_dev_stat_inc_and_print(dev,
1263 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1264 }
1265
1266 if (sctx->readonly) {
1267 ASSERT(!sctx->is_dev_replace);
1268 goto out;
1269 }
1270
1271 if (!is_metadata && !have_csum) {
1272 struct scrub_fixup_nodatasum *fixup_nodatasum;
1273
1274 WARN_ON(sctx->is_dev_replace);
1275
1276nodatasum_case:
1277
1278 /*
1279 * !is_metadata and !have_csum, this means that the data
1280 * might not be COWed, that it might be modified
1281 * concurrently. The general strategy to work on the
1282 * commit root does not help in the case when COW is not
1283 * used.
1284 */
1285 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
1286 if (!fixup_nodatasum)
1287 goto did_not_correct_error;
1288 fixup_nodatasum->sctx = sctx;
1289 fixup_nodatasum->dev = dev;
1290 fixup_nodatasum->logical = logical;
1291 fixup_nodatasum->root = fs_info->extent_root;
1292 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
1293 scrub_pending_trans_workers_inc(sctx);
1294 btrfs_init_work(&fixup_nodatasum->work, btrfs_scrub_helper,
1295 scrub_fixup_nodatasum, NULL, NULL);
1296 btrfs_queue_work(fs_info->scrub_workers,
1297 &fixup_nodatasum->work);
1298 goto out;
1299 }
1300
1301 /*
1302 * now build and submit the bios for the other mirrors, check
1303 * checksums.
1304 * First try to pick the mirror which is completely without I/O
1305 * errors and also does not have a checksum error.
1306 * If one is found, and if a checksum is present, the full block
1307 * that is known to contain an error is rewritten. Afterwards
1308 * the block is known to be corrected.
1309 * If a mirror is found which is completely correct, and no
1310 * checksum is present, only those pages are rewritten that had
1311 * an I/O error in the block to be repaired, since it cannot be
1312 * determined, which copy of the other pages is better (and it
1313 * could happen otherwise that a correct page would be
1314 * overwritten by a bad one).
1315 */
1316 for (mirror_index = 0; ;mirror_index++) {
1317 struct scrub_block *sblock_other;
1318
1319 if (mirror_index == failed_mirror_index)
1320 continue;
1321
1322 /* raid56's mirror can be more than BTRFS_MAX_MIRRORS */
1323 if (!scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1324 if (mirror_index >= BTRFS_MAX_MIRRORS)
1325 break;
1326 if (!sblocks_for_recheck[mirror_index].page_count)
1327 break;
1328
1329 sblock_other = sblocks_for_recheck + mirror_index;
1330 } else {
1331 struct scrub_recover *r = sblock_bad->pagev[0]->recover;
1332 int max_allowed = r->bbio->num_stripes -
1333 r->bbio->num_tgtdevs;
1334
1335 if (mirror_index >= max_allowed)
1336 break;
1337 if (!sblocks_for_recheck[1].page_count)
1338 break;
1339
1340 ASSERT(failed_mirror_index == 0);
1341 sblock_other = sblocks_for_recheck + 1;
1342 sblock_other->pagev[0]->mirror_num = 1 + mirror_index;
1343 }
1344
1345 /* build and submit the bios, check checksums */
1346 scrub_recheck_block(fs_info, sblock_other, 0);
1347
1348 if (!sblock_other->header_error &&
1349 !sblock_other->checksum_error &&
1350 sblock_other->no_io_error_seen) {
1351 if (sctx->is_dev_replace) {
1352 scrub_write_block_to_dev_replace(sblock_other);
1353 goto corrected_error;
1354 } else {
1355 ret = scrub_repair_block_from_good_copy(
1356 sblock_bad, sblock_other);
1357 if (!ret)
1358 goto corrected_error;
1359 }
1360 }
1361 }
1362
1363 if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1364 goto did_not_correct_error;
1365
1366 /*
1367 * In case of I/O errors in the area that is supposed to be
1368 * repaired, continue by picking good copies of those pages.
1369 * Select the good pages from mirrors to rewrite bad pages from
1370 * the area to fix. Afterwards verify the checksum of the block
1371 * that is supposed to be repaired. This verification step is
1372 * only done for the purpose of statistic counting and for the
1373 * final scrub report, whether errors remain.
1374 * A perfect algorithm could make use of the checksum and try
1375 * all possible combinations of pages from the different mirrors
1376 * until the checksum verification succeeds. For example, when
1377 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1378 * of mirror #2 is readable but the final checksum test fails,
1379 * then the 2nd page of mirror #3 could be tried, whether now
1380 * the final checksum succeeds. But this would be a rare
1381 * exception and is therefore not implemented. At least it is
1382 * avoided that the good copy is overwritten.
1383 * A more useful improvement would be to pick the sectors
1384 * without I/O error based on sector sizes (512 bytes on legacy
1385 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1386 * mirror could be repaired by taking 512 byte of a different
1387 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1388 * area are unreadable.
1389 */
1390 success = 1;
1391 for (page_num = 0; page_num < sblock_bad->page_count;
1392 page_num++) {
1393 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1394 struct scrub_block *sblock_other = NULL;
1395
1396 /* skip no-io-error page in scrub */
1397 if (!page_bad->io_error && !sctx->is_dev_replace)
1398 continue;
1399
1400 if (scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1401 /*
1402 * In case of dev replace, if raid56 rebuild process
1403 * didn't work out correct data, then copy the content
1404 * in sblock_bad to make sure target device is identical
1405 * to source device, instead of writing garbage data in
1406 * sblock_for_recheck array to target device.
1407 */
1408 sblock_other = NULL;
1409 } else if (page_bad->io_error) {
1410 /* try to find no-io-error page in mirrors */
1411 for (mirror_index = 0;
1412 mirror_index < BTRFS_MAX_MIRRORS &&
1413 sblocks_for_recheck[mirror_index].page_count > 0;
1414 mirror_index++) {
1415 if (!sblocks_for_recheck[mirror_index].
1416 pagev[page_num]->io_error) {
1417 sblock_other = sblocks_for_recheck +
1418 mirror_index;
1419 break;
1420 }
1421 }
1422 if (!sblock_other)
1423 success = 0;
1424 }
1425
1426 if (sctx->is_dev_replace) {
1427 /*
1428 * did not find a mirror to fetch the page
1429 * from. scrub_write_page_to_dev_replace()
1430 * handles this case (page->io_error), by
1431 * filling the block with zeros before
1432 * submitting the write request
1433 */
1434 if (!sblock_other)
1435 sblock_other = sblock_bad;
1436
1437 if (scrub_write_page_to_dev_replace(sblock_other,
1438 page_num) != 0) {
1439 btrfs_dev_replace_stats_inc(
1440 &fs_info->dev_replace.num_write_errors);
1441 success = 0;
1442 }
1443 } else if (sblock_other) {
1444 ret = scrub_repair_page_from_good_copy(sblock_bad,
1445 sblock_other,
1446 page_num, 0);
1447 if (0 == ret)
1448 page_bad->io_error = 0;
1449 else
1450 success = 0;
1451 }
1452 }
1453
1454 if (success && !sctx->is_dev_replace) {
1455 if (is_metadata || have_csum) {
1456 /*
1457 * need to verify the checksum now that all
1458 * sectors on disk are repaired (the write
1459 * request for data to be repaired is on its way).
1460 * Just be lazy and use scrub_recheck_block()
1461 * which re-reads the data before the checksum
1462 * is verified, but most likely the data comes out
1463 * of the page cache.
1464 */
1465 scrub_recheck_block(fs_info, sblock_bad, 1);
1466 if (!sblock_bad->header_error &&
1467 !sblock_bad->checksum_error &&
1468 sblock_bad->no_io_error_seen)
1469 goto corrected_error;
1470 else
1471 goto did_not_correct_error;
1472 } else {
1473corrected_error:
1474 spin_lock(&sctx->stat_lock);
1475 sctx->stat.corrected_errors++;
1476 sblock_to_check->data_corrected = 1;
1477 spin_unlock(&sctx->stat_lock);
1478 btrfs_err_rl_in_rcu(fs_info,
1479 "fixed up error at logical %llu on dev %s",
1480 logical, rcu_str_deref(dev->name));
1481 }
1482 } else {
1483did_not_correct_error:
1484 spin_lock(&sctx->stat_lock);
1485 sctx->stat.uncorrectable_errors++;
1486 spin_unlock(&sctx->stat_lock);
1487 btrfs_err_rl_in_rcu(fs_info,
1488 "unable to fixup (regular) error at logical %llu on dev %s",
1489 logical, rcu_str_deref(dev->name));
1490 }
1491
1492out:
1493 if (sblocks_for_recheck) {
1494 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1495 mirror_index++) {
1496 struct scrub_block *sblock = sblocks_for_recheck +
1497 mirror_index;
1498 struct scrub_recover *recover;
1499 int page_index;
1500
1501 for (page_index = 0; page_index < sblock->page_count;
1502 page_index++) {
1503 sblock->pagev[page_index]->sblock = NULL;
1504 recover = sblock->pagev[page_index]->recover;
1505 if (recover) {
1506 scrub_put_recover(fs_info, recover);
1507 sblock->pagev[page_index]->recover =
1508 NULL;
1509 }
1510 scrub_page_put(sblock->pagev[page_index]);
1511 }
1512 }
1513 kfree(sblocks_for_recheck);
1514 }
1515
1516 ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1517 if (ret < 0)
1518 return ret;
1519 return 0;
1520}
1521
1522static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1523{
1524 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1525 return 2;
1526 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1527 return 3;
1528 else
1529 return (int)bbio->num_stripes;
1530}
1531
1532static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1533 u64 *raid_map,
1534 u64 mapped_length,
1535 int nstripes, int mirror,
1536 int *stripe_index,
1537 u64 *stripe_offset)
1538{
1539 int i;
1540
1541 if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1542 /* RAID5/6 */
1543 for (i = 0; i < nstripes; i++) {
1544 if (raid_map[i] == RAID6_Q_STRIPE ||
1545 raid_map[i] == RAID5_P_STRIPE)
1546 continue;
1547
1548 if (logical >= raid_map[i] &&
1549 logical < raid_map[i] + mapped_length)
1550 break;
1551 }
1552
1553 *stripe_index = i;
1554 *stripe_offset = logical - raid_map[i];
1555 } else {
1556 /* The other RAID type */
1557 *stripe_index = mirror;
1558 *stripe_offset = 0;
1559 }
1560}
1561
1562static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1563 struct scrub_block *sblocks_for_recheck)
1564{
1565 struct scrub_ctx *sctx = original_sblock->sctx;
1566 struct btrfs_fs_info *fs_info = sctx->fs_info;
1567 u64 length = original_sblock->page_count * PAGE_SIZE;
1568 u64 logical = original_sblock->pagev[0]->logical;
1569 u64 generation = original_sblock->pagev[0]->generation;
1570 u64 flags = original_sblock->pagev[0]->flags;
1571 u64 have_csum = original_sblock->pagev[0]->have_csum;
1572 struct scrub_recover *recover;
1573 struct btrfs_bio *bbio;
1574 u64 sublen;
1575 u64 mapped_length;
1576 u64 stripe_offset;
1577 int stripe_index;
1578 int page_index = 0;
1579 int mirror_index;
1580 int nmirrors;
1581 int ret;
1582
1583 /*
1584 * note: the two members refs and outstanding_pages
1585 * are not used (and not set) in the blocks that are used for
1586 * the recheck procedure
1587 */
1588
1589 while (length > 0) {
1590 sublen = min_t(u64, length, PAGE_SIZE);
1591 mapped_length = sublen;
1592 bbio = NULL;
1593
1594 /*
1595 * with a length of PAGE_SIZE, each returned stripe
1596 * represents one mirror
1597 */
1598 btrfs_bio_counter_inc_blocked(fs_info);
1599 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1600 logical, &mapped_length, &bbio);
1601 if (ret || !bbio || mapped_length < sublen) {
1602 btrfs_put_bbio(bbio);
1603 btrfs_bio_counter_dec(fs_info);
1604 return -EIO;
1605 }
1606
1607 recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1608 if (!recover) {
1609 btrfs_put_bbio(bbio);
1610 btrfs_bio_counter_dec(fs_info);
1611 return -ENOMEM;
1612 }
1613
1614 refcount_set(&recover->refs, 1);
1615 recover->bbio = bbio;
1616 recover->map_length = mapped_length;
1617
1618 BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1619
1620 nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1621
1622 for (mirror_index = 0; mirror_index < nmirrors;
1623 mirror_index++) {
1624 struct scrub_block *sblock;
1625 struct scrub_page *page;
1626
1627 sblock = sblocks_for_recheck + mirror_index;
1628 sblock->sctx = sctx;
1629
1630 page = kzalloc(sizeof(*page), GFP_NOFS);
1631 if (!page) {
1632leave_nomem:
1633 spin_lock(&sctx->stat_lock);
1634 sctx->stat.malloc_errors++;
1635 spin_unlock(&sctx->stat_lock);
1636 scrub_put_recover(fs_info, recover);
1637 return -ENOMEM;
1638 }
1639 scrub_page_get(page);
1640 sblock->pagev[page_index] = page;
1641 page->sblock = sblock;
1642 page->flags = flags;
1643 page->generation = generation;
1644 page->logical = logical;
1645 page->have_csum = have_csum;
1646 if (have_csum)
1647 memcpy(page->csum,
1648 original_sblock->pagev[0]->csum,
1649 sctx->csum_size);
1650
1651 scrub_stripe_index_and_offset(logical,
1652 bbio->map_type,
1653 bbio->raid_map,
1654 mapped_length,
1655 bbio->num_stripes -
1656 bbio->num_tgtdevs,
1657 mirror_index,
1658 &stripe_index,
1659 &stripe_offset);
1660 page->physical = bbio->stripes[stripe_index].physical +
1661 stripe_offset;
1662 page->dev = bbio->stripes[stripe_index].dev;
1663
1664 BUG_ON(page_index >= original_sblock->page_count);
1665 page->physical_for_dev_replace =
1666 original_sblock->pagev[page_index]->
1667 physical_for_dev_replace;
1668 /* for missing devices, dev->bdev is NULL */
1669 page->mirror_num = mirror_index + 1;
1670 sblock->page_count++;
1671 page->page = alloc_page(GFP_NOFS);
1672 if (!page->page)
1673 goto leave_nomem;
1674
1675 scrub_get_recover(recover);
1676 page->recover = recover;
1677 }
1678 scrub_put_recover(fs_info, recover);
1679 length -= sublen;
1680 logical += sublen;
1681 page_index++;
1682 }
1683
1684 return 0;
1685}
1686
1687static void scrub_bio_wait_endio(struct bio *bio)
1688{
1689 complete(bio->bi_private);
1690}
1691
1692static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1693 struct bio *bio,
1694 struct scrub_page *page)
1695{
1696 DECLARE_COMPLETION_ONSTACK(done);
1697 int ret;
1698 int mirror_num;
1699
1700 bio->bi_iter.bi_sector = page->logical >> 9;
1701 bio->bi_private = &done;
1702 bio->bi_end_io = scrub_bio_wait_endio;
1703
1704 mirror_num = page->sblock->pagev[0]->mirror_num;
1705 ret = raid56_parity_recover(fs_info, bio, page->recover->bbio,
1706 page->recover->map_length,
1707 mirror_num, 0);
1708 if (ret)
1709 return ret;
1710
1711 wait_for_completion_io(&done);
1712 return blk_status_to_errno(bio->bi_status);
1713}
1714
1715static void scrub_recheck_block_on_raid56(struct btrfs_fs_info *fs_info,
1716 struct scrub_block *sblock)
1717{
1718 struct scrub_page *first_page = sblock->pagev[0];
1719 struct bio *bio;
1720 int page_num;
1721
1722 /* All pages in sblock belong to the same stripe on the same device. */
1723 ASSERT(first_page->dev);
1724 if (!first_page->dev->bdev)
1725 goto out;
1726
1727 bio = btrfs_io_bio_alloc(BIO_MAX_PAGES);
1728 bio_set_dev(bio, first_page->dev->bdev);
1729
1730 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1731 struct scrub_page *page = sblock->pagev[page_num];
1732
1733 WARN_ON(!page->page);
1734 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1735 }
1736
1737 if (scrub_submit_raid56_bio_wait(fs_info, bio, first_page)) {
1738 bio_put(bio);
1739 goto out;
1740 }
1741
1742 bio_put(bio);
1743
1744 scrub_recheck_block_checksum(sblock);
1745
1746 return;
1747out:
1748 for (page_num = 0; page_num < sblock->page_count; page_num++)
1749 sblock->pagev[page_num]->io_error = 1;
1750
1751 sblock->no_io_error_seen = 0;
1752}
1753
1754/*
1755 * this function will check the on disk data for checksum errors, header
1756 * errors and read I/O errors. If any I/O errors happen, the exact pages
1757 * which are errored are marked as being bad. The goal is to enable scrub
1758 * to take those pages that are not errored from all the mirrors so that
1759 * the pages that are errored in the just handled mirror can be repaired.
1760 */
1761static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1762 struct scrub_block *sblock,
1763 int retry_failed_mirror)
1764{
1765 int page_num;
1766
1767 sblock->no_io_error_seen = 1;
1768
1769 /* short cut for raid56 */
1770 if (!retry_failed_mirror && scrub_is_page_on_raid56(sblock->pagev[0]))
1771 return scrub_recheck_block_on_raid56(fs_info, sblock);
1772
1773 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1774 struct bio *bio;
1775 struct scrub_page *page = sblock->pagev[page_num];
1776
1777 if (page->dev->bdev == NULL) {
1778 page->io_error = 1;
1779 sblock->no_io_error_seen = 0;
1780 continue;
1781 }
1782
1783 WARN_ON(!page->page);
1784 bio = btrfs_io_bio_alloc(1);
1785 bio_set_dev(bio, page->dev->bdev);
1786
1787 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1788 bio->bi_iter.bi_sector = page->physical >> 9;
1789 bio->bi_opf = REQ_OP_READ;
1790
1791 if (btrfsic_submit_bio_wait(bio)) {
1792 page->io_error = 1;
1793 sblock->no_io_error_seen = 0;
1794 }
1795
1796 bio_put(bio);
1797 }
1798
1799 if (sblock->no_io_error_seen)
1800 scrub_recheck_block_checksum(sblock);
1801}
1802
1803static inline int scrub_check_fsid(u8 fsid[],
1804 struct scrub_page *spage)
1805{
1806 struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1807 int ret;
1808
1809 ret = memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1810 return !ret;
1811}
1812
1813static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1814{
1815 sblock->header_error = 0;
1816 sblock->checksum_error = 0;
1817 sblock->generation_error = 0;
1818
1819 if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1820 scrub_checksum_data(sblock);
1821 else
1822 scrub_checksum_tree_block(sblock);
1823}
1824
1825static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1826 struct scrub_block *sblock_good)
1827{
1828 int page_num;
1829 int ret = 0;
1830
1831 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1832 int ret_sub;
1833
1834 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1835 sblock_good,
1836 page_num, 1);
1837 if (ret_sub)
1838 ret = ret_sub;
1839 }
1840
1841 return ret;
1842}
1843
1844static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1845 struct scrub_block *sblock_good,
1846 int page_num, int force_write)
1847{
1848 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1849 struct scrub_page *page_good = sblock_good->pagev[page_num];
1850 struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1851
1852 BUG_ON(page_bad->page == NULL);
1853 BUG_ON(page_good->page == NULL);
1854 if (force_write || sblock_bad->header_error ||
1855 sblock_bad->checksum_error || page_bad->io_error) {
1856 struct bio *bio;
1857 int ret;
1858
1859 if (!page_bad->dev->bdev) {
1860 btrfs_warn_rl(fs_info,
1861 "scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1862 return -EIO;
1863 }
1864
1865 bio = btrfs_io_bio_alloc(1);
1866 bio_set_dev(bio, page_bad->dev->bdev);
1867 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1868 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1869
1870 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1871 if (PAGE_SIZE != ret) {
1872 bio_put(bio);
1873 return -EIO;
1874 }
1875
1876 if (btrfsic_submit_bio_wait(bio)) {
1877 btrfs_dev_stat_inc_and_print(page_bad->dev,
1878 BTRFS_DEV_STAT_WRITE_ERRS);
1879 btrfs_dev_replace_stats_inc(
1880 &fs_info->dev_replace.num_write_errors);
1881 bio_put(bio);
1882 return -EIO;
1883 }
1884 bio_put(bio);
1885 }
1886
1887 return 0;
1888}
1889
1890static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1891{
1892 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1893 int page_num;
1894
1895 /*
1896 * This block is used for the check of the parity on the source device,
1897 * so the data needn't be written into the destination device.
1898 */
1899 if (sblock->sparity)
1900 return;
1901
1902 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1903 int ret;
1904
1905 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1906 if (ret)
1907 btrfs_dev_replace_stats_inc(
1908 &fs_info->dev_replace.num_write_errors);
1909 }
1910}
1911
1912static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1913 int page_num)
1914{
1915 struct scrub_page *spage = sblock->pagev[page_num];
1916
1917 BUG_ON(spage->page == NULL);
1918 if (spage->io_error) {
1919 void *mapped_buffer = kmap_atomic(spage->page);
1920
1921 clear_page(mapped_buffer);
1922 flush_dcache_page(spage->page);
1923 kunmap_atomic(mapped_buffer);
1924 }
1925 return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1926}
1927
1928static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1929 struct scrub_page *spage)
1930{
1931 struct scrub_bio *sbio;
1932 int ret;
1933
1934 mutex_lock(&sctx->wr_lock);
1935again:
1936 if (!sctx->wr_curr_bio) {
1937 sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1938 GFP_KERNEL);
1939 if (!sctx->wr_curr_bio) {
1940 mutex_unlock(&sctx->wr_lock);
1941 return -ENOMEM;
1942 }
1943 sctx->wr_curr_bio->sctx = sctx;
1944 sctx->wr_curr_bio->page_count = 0;
1945 }
1946 sbio = sctx->wr_curr_bio;
1947 if (sbio->page_count == 0) {
1948 struct bio *bio;
1949
1950 sbio->physical = spage->physical_for_dev_replace;
1951 sbio->logical = spage->logical;
1952 sbio->dev = sctx->wr_tgtdev;
1953 bio = sbio->bio;
1954 if (!bio) {
1955 bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
1956 sbio->bio = bio;
1957 }
1958
1959 bio->bi_private = sbio;
1960 bio->bi_end_io = scrub_wr_bio_end_io;
1961 bio_set_dev(bio, sbio->dev->bdev);
1962 bio->bi_iter.bi_sector = sbio->physical >> 9;
1963 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1964 sbio->status = 0;
1965 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1966 spage->physical_for_dev_replace ||
1967 sbio->logical + sbio->page_count * PAGE_SIZE !=
1968 spage->logical) {
1969 scrub_wr_submit(sctx);
1970 goto again;
1971 }
1972
1973 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1974 if (ret != PAGE_SIZE) {
1975 if (sbio->page_count < 1) {
1976 bio_put(sbio->bio);
1977 sbio->bio = NULL;
1978 mutex_unlock(&sctx->wr_lock);
1979 return -EIO;
1980 }
1981 scrub_wr_submit(sctx);
1982 goto again;
1983 }
1984
1985 sbio->pagev[sbio->page_count] = spage;
1986 scrub_page_get(spage);
1987 sbio->page_count++;
1988 if (sbio->page_count == sctx->pages_per_wr_bio)
1989 scrub_wr_submit(sctx);
1990 mutex_unlock(&sctx->wr_lock);
1991
1992 return 0;
1993}
1994
1995static void scrub_wr_submit(struct scrub_ctx *sctx)
1996{
1997 struct scrub_bio *sbio;
1998
1999 if (!sctx->wr_curr_bio)
2000 return;
2001
2002 sbio = sctx->wr_curr_bio;
2003 sctx->wr_curr_bio = NULL;
2004 WARN_ON(!sbio->bio->bi_disk);
2005 scrub_pending_bio_inc(sctx);
2006 /* process all writes in a single worker thread. Then the block layer
2007 * orders the requests before sending them to the driver which
2008 * doubled the write performance on spinning disks when measured
2009 * with Linux 3.5 */
2010 btrfsic_submit_bio(sbio->bio);
2011}
2012
2013static void scrub_wr_bio_end_io(struct bio *bio)
2014{
2015 struct scrub_bio *sbio = bio->bi_private;
2016 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2017
2018 sbio->status = bio->bi_status;
2019 sbio->bio = bio;
2020
2021 btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
2022 scrub_wr_bio_end_io_worker, NULL, NULL);
2023 btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
2024}
2025
2026static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
2027{
2028 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2029 struct scrub_ctx *sctx = sbio->sctx;
2030 int i;
2031
2032 WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
2033 if (sbio->status) {
2034 struct btrfs_dev_replace *dev_replace =
2035 &sbio->sctx->fs_info->dev_replace;
2036
2037 for (i = 0; i < sbio->page_count; i++) {
2038 struct scrub_page *spage = sbio->pagev[i];
2039
2040 spage->io_error = 1;
2041 btrfs_dev_replace_stats_inc(&dev_replace->
2042 num_write_errors);
2043 }
2044 }
2045
2046 for (i = 0; i < sbio->page_count; i++)
2047 scrub_page_put(sbio->pagev[i]);
2048
2049 bio_put(sbio->bio);
2050 kfree(sbio);
2051 scrub_pending_bio_dec(sctx);
2052}
2053
2054static int scrub_checksum(struct scrub_block *sblock)
2055{
2056 u64 flags;
2057 int ret;
2058
2059 /*
2060 * No need to initialize these stats currently,
2061 * because this function only use return value
2062 * instead of these stats value.
2063 *
2064 * Todo:
2065 * always use stats
2066 */
2067 sblock->header_error = 0;
2068 sblock->generation_error = 0;
2069 sblock->checksum_error = 0;
2070
2071 WARN_ON(sblock->page_count < 1);
2072 flags = sblock->pagev[0]->flags;
2073 ret = 0;
2074 if (flags & BTRFS_EXTENT_FLAG_DATA)
2075 ret = scrub_checksum_data(sblock);
2076 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2077 ret = scrub_checksum_tree_block(sblock);
2078 else if (flags & BTRFS_EXTENT_FLAG_SUPER)
2079 (void)scrub_checksum_super(sblock);
2080 else
2081 WARN_ON(1);
2082 if (ret)
2083 scrub_handle_errored_block(sblock);
2084
2085 return ret;
2086}
2087
2088static int scrub_checksum_data(struct scrub_block *sblock)
2089{
2090 struct scrub_ctx *sctx = sblock->sctx;
2091 u8 csum[BTRFS_CSUM_SIZE];
2092 u8 *on_disk_csum;
2093 struct page *page;
2094 void *buffer;
2095 u32 crc = ~(u32)0;
2096 u64 len;
2097 int index;
2098
2099 BUG_ON(sblock->page_count < 1);
2100 if (!sblock->pagev[0]->have_csum)
2101 return 0;
2102
2103 on_disk_csum = sblock->pagev[0]->csum;
2104 page = sblock->pagev[0]->page;
2105 buffer = kmap_atomic(page);
2106
2107 len = sctx->fs_info->sectorsize;
2108 index = 0;
2109 for (;;) {
2110 u64 l = min_t(u64, len, PAGE_SIZE);
2111
2112 crc = btrfs_csum_data(buffer, crc, l);
2113 kunmap_atomic(buffer);
2114 len -= l;
2115 if (len == 0)
2116 break;
2117 index++;
2118 BUG_ON(index >= sblock->page_count);
2119 BUG_ON(!sblock->pagev[index]->page);
2120 page = sblock->pagev[index]->page;
2121 buffer = kmap_atomic(page);
2122 }
2123
2124 btrfs_csum_final(crc, csum);
2125 if (memcmp(csum, on_disk_csum, sctx->csum_size))
2126 sblock->checksum_error = 1;
2127
2128 return sblock->checksum_error;
2129}
2130
2131static int scrub_checksum_tree_block(struct scrub_block *sblock)
2132{
2133 struct scrub_ctx *sctx = sblock->sctx;
2134 struct btrfs_header *h;
2135 struct btrfs_fs_info *fs_info = sctx->fs_info;
2136 u8 calculated_csum[BTRFS_CSUM_SIZE];
2137 u8 on_disk_csum[BTRFS_CSUM_SIZE];
2138 struct page *page;
2139 void *mapped_buffer;
2140 u64 mapped_size;
2141 void *p;
2142 u32 crc = ~(u32)0;
2143 u64 len;
2144 int index;
2145
2146 BUG_ON(sblock->page_count < 1);
2147 page = sblock->pagev[0]->page;
2148 mapped_buffer = kmap_atomic(page);
2149 h = (struct btrfs_header *)mapped_buffer;
2150 memcpy(on_disk_csum, h->csum, sctx->csum_size);
2151
2152 /*
2153 * we don't use the getter functions here, as we
2154 * a) don't have an extent buffer and
2155 * b) the page is already kmapped
2156 */
2157 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
2158 sblock->header_error = 1;
2159
2160 if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
2161 sblock->header_error = 1;
2162 sblock->generation_error = 1;
2163 }
2164
2165 if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
2166 sblock->header_error = 1;
2167
2168 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
2169 BTRFS_UUID_SIZE))
2170 sblock->header_error = 1;
2171
2172 len = sctx->fs_info->nodesize - BTRFS_CSUM_SIZE;
2173 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
2174 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
2175 index = 0;
2176 for (;;) {
2177 u64 l = min_t(u64, len, mapped_size);
2178
2179 crc = btrfs_csum_data(p, crc, l);
2180 kunmap_atomic(mapped_buffer);
2181 len -= l;
2182 if (len == 0)
2183 break;
2184 index++;
2185 BUG_ON(index >= sblock->page_count);
2186 BUG_ON(!sblock->pagev[index]->page);
2187 page = sblock->pagev[index]->page;
2188 mapped_buffer = kmap_atomic(page);
2189 mapped_size = PAGE_SIZE;
2190 p = mapped_buffer;
2191 }
2192
2193 btrfs_csum_final(crc, calculated_csum);
2194 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
2195 sblock->checksum_error = 1;
2196
2197 return sblock->header_error || sblock->checksum_error;
2198}
2199
2200static int scrub_checksum_super(struct scrub_block *sblock)
2201{
2202 struct btrfs_super_block *s;
2203 struct scrub_ctx *sctx = sblock->sctx;
2204 u8 calculated_csum[BTRFS_CSUM_SIZE];
2205 u8 on_disk_csum[BTRFS_CSUM_SIZE];
2206 struct page *page;
2207 void *mapped_buffer;
2208 u64 mapped_size;
2209 void *p;
2210 u32 crc = ~(u32)0;
2211 int fail_gen = 0;
2212 int fail_cor = 0;
2213 u64 len;
2214 int index;
2215
2216 BUG_ON(sblock->page_count < 1);
2217 page = sblock->pagev[0]->page;
2218 mapped_buffer = kmap_atomic(page);
2219 s = (struct btrfs_super_block *)mapped_buffer;
2220 memcpy(on_disk_csum, s->csum, sctx->csum_size);
2221
2222 if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
2223 ++fail_cor;
2224
2225 if (sblock->pagev[0]->generation != btrfs_super_generation(s))
2226 ++fail_gen;
2227
2228 if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
2229 ++fail_cor;
2230
2231 len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
2232 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
2233 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
2234 index = 0;
2235 for (;;) {
2236 u64 l = min_t(u64, len, mapped_size);
2237
2238 crc = btrfs_csum_data(p, crc, l);
2239 kunmap_atomic(mapped_buffer);
2240 len -= l;
2241 if (len == 0)
2242 break;
2243 index++;
2244 BUG_ON(index >= sblock->page_count);
2245 BUG_ON(!sblock->pagev[index]->page);
2246 page = sblock->pagev[index]->page;
2247 mapped_buffer = kmap_atomic(page);
2248 mapped_size = PAGE_SIZE;
2249 p = mapped_buffer;
2250 }
2251
2252 btrfs_csum_final(crc, calculated_csum);
2253 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
2254 ++fail_cor;
2255
2256 if (fail_cor + fail_gen) {
2257 /*
2258 * if we find an error in a super block, we just report it.
2259 * They will get written with the next transaction commit
2260 * anyway
2261 */
2262 spin_lock(&sctx->stat_lock);
2263 ++sctx->stat.super_errors;
2264 spin_unlock(&sctx->stat_lock);
2265 if (fail_cor)
2266 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
2267 BTRFS_DEV_STAT_CORRUPTION_ERRS);
2268 else
2269 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
2270 BTRFS_DEV_STAT_GENERATION_ERRS);
2271 }
2272
2273 return fail_cor + fail_gen;
2274}
2275
2276static void scrub_block_get(struct scrub_block *sblock)
2277{
2278 refcount_inc(&sblock->refs);
2279}
2280
2281static void scrub_block_put(struct scrub_block *sblock)
2282{
2283 if (refcount_dec_and_test(&sblock->refs)) {
2284 int i;
2285
2286 if (sblock->sparity)
2287 scrub_parity_put(sblock->sparity);
2288
2289 for (i = 0; i < sblock->page_count; i++)
2290 scrub_page_put(sblock->pagev[i]);
2291 kfree(sblock);
2292 }
2293}
2294
2295static void scrub_page_get(struct scrub_page *spage)
2296{
2297 atomic_inc(&spage->refs);
2298}
2299
2300static void scrub_page_put(struct scrub_page *spage)
2301{
2302 if (atomic_dec_and_test(&spage->refs)) {
2303 if (spage->page)
2304 __free_page(spage->page);
2305 kfree(spage);
2306 }
2307}
2308
2309static void scrub_submit(struct scrub_ctx *sctx)
2310{
2311 struct scrub_bio *sbio;
2312
2313 if (sctx->curr == -1)
2314 return;
2315
2316 sbio = sctx->bios[sctx->curr];
2317 sctx->curr = -1;
2318 scrub_pending_bio_inc(sctx);
2319 btrfsic_submit_bio(sbio->bio);
2320}
2321
2322static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2323 struct scrub_page *spage)
2324{
2325 struct scrub_block *sblock = spage->sblock;
2326 struct scrub_bio *sbio;
2327 int ret;
2328
2329again:
2330 /*
2331 * grab a fresh bio or wait for one to become available
2332 */
2333 while (sctx->curr == -1) {
2334 spin_lock(&sctx->list_lock);
2335 sctx->curr = sctx->first_free;
2336 if (sctx->curr != -1) {
2337 sctx->first_free = sctx->bios[sctx->curr]->next_free;
2338 sctx->bios[sctx->curr]->next_free = -1;
2339 sctx->bios[sctx->curr]->page_count = 0;
2340 spin_unlock(&sctx->list_lock);
2341 } else {
2342 spin_unlock(&sctx->list_lock);
2343 wait_event(sctx->list_wait, sctx->first_free != -1);
2344 }
2345 }
2346 sbio = sctx->bios[sctx->curr];
2347 if (sbio->page_count == 0) {
2348 struct bio *bio;
2349
2350 sbio->physical = spage->physical;
2351 sbio->logical = spage->logical;
2352 sbio->dev = spage->dev;
2353 bio = sbio->bio;
2354 if (!bio) {
2355 bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
2356 sbio->bio = bio;
2357 }
2358
2359 bio->bi_private = sbio;
2360 bio->bi_end_io = scrub_bio_end_io;
2361 bio_set_dev(bio, sbio->dev->bdev);
2362 bio->bi_iter.bi_sector = sbio->physical >> 9;
2363 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2364 sbio->status = 0;
2365 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2366 spage->physical ||
2367 sbio->logical + sbio->page_count * PAGE_SIZE !=
2368 spage->logical ||
2369 sbio->dev != spage->dev) {
2370 scrub_submit(sctx);
2371 goto again;
2372 }
2373
2374 sbio->pagev[sbio->page_count] = spage;
2375 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2376 if (ret != PAGE_SIZE) {
2377 if (sbio->page_count < 1) {
2378 bio_put(sbio->bio);
2379 sbio->bio = NULL;
2380 return -EIO;
2381 }
2382 scrub_submit(sctx);
2383 goto again;
2384 }
2385
2386 scrub_block_get(sblock); /* one for the page added to the bio */
2387 atomic_inc(&sblock->outstanding_pages);
2388 sbio->page_count++;
2389 if (sbio->page_count == sctx->pages_per_rd_bio)
2390 scrub_submit(sctx);
2391
2392 return 0;
2393}
2394
2395static void scrub_missing_raid56_end_io(struct bio *bio)
2396{
2397 struct scrub_block *sblock = bio->bi_private;
2398 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2399
2400 if (bio->bi_status)
2401 sblock->no_io_error_seen = 0;
2402
2403 bio_put(bio);
2404
2405 btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2406}
2407
2408static void scrub_missing_raid56_worker(struct btrfs_work *work)
2409{
2410 struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2411 struct scrub_ctx *sctx = sblock->sctx;
2412 struct btrfs_fs_info *fs_info = sctx->fs_info;
2413 u64 logical;
2414 struct btrfs_device *dev;
2415
2416 logical = sblock->pagev[0]->logical;
2417 dev = sblock->pagev[0]->dev;
2418
2419 if (sblock->no_io_error_seen)
2420 scrub_recheck_block_checksum(sblock);
2421
2422 if (!sblock->no_io_error_seen) {
2423 spin_lock(&sctx->stat_lock);
2424 sctx->stat.read_errors++;
2425 spin_unlock(&sctx->stat_lock);
2426 btrfs_err_rl_in_rcu(fs_info,
2427 "IO error rebuilding logical %llu for dev %s",
2428 logical, rcu_str_deref(dev->name));
2429 } else if (sblock->header_error || sblock->checksum_error) {
2430 spin_lock(&sctx->stat_lock);
2431 sctx->stat.uncorrectable_errors++;
2432 spin_unlock(&sctx->stat_lock);
2433 btrfs_err_rl_in_rcu(fs_info,
2434 "failed to rebuild valid logical %llu for dev %s",
2435 logical, rcu_str_deref(dev->name));
2436 } else {
2437 scrub_write_block_to_dev_replace(sblock);
2438 }
2439
2440 scrub_block_put(sblock);
2441
2442 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2443 mutex_lock(&sctx->wr_lock);
2444 scrub_wr_submit(sctx);
2445 mutex_unlock(&sctx->wr_lock);
2446 }
2447
2448 scrub_pending_bio_dec(sctx);
2449}
2450
2451static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2452{
2453 struct scrub_ctx *sctx = sblock->sctx;
2454 struct btrfs_fs_info *fs_info = sctx->fs_info;
2455 u64 length = sblock->page_count * PAGE_SIZE;
2456 u64 logical = sblock->pagev[0]->logical;
2457 struct btrfs_bio *bbio = NULL;
2458 struct bio *bio;
2459 struct btrfs_raid_bio *rbio;
2460 int ret;
2461 int i;
2462
2463 btrfs_bio_counter_inc_blocked(fs_info);
2464 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2465 &length, &bbio);
2466 if (ret || !bbio || !bbio->raid_map)
2467 goto bbio_out;
2468
2469 if (WARN_ON(!sctx->is_dev_replace ||
2470 !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2471 /*
2472 * We shouldn't be scrubbing a missing device. Even for dev
2473 * replace, we should only get here for RAID 5/6. We either
2474 * managed to mount something with no mirrors remaining or
2475 * there's a bug in scrub_remap_extent()/btrfs_map_block().
2476 */
2477 goto bbio_out;
2478 }
2479
2480 bio = btrfs_io_bio_alloc(0);
2481 bio->bi_iter.bi_sector = logical >> 9;
2482 bio->bi_private = sblock;
2483 bio->bi_end_io = scrub_missing_raid56_end_io;
2484
2485 rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2486 if (!rbio)
2487 goto rbio_out;
2488
2489 for (i = 0; i < sblock->page_count; i++) {
2490 struct scrub_page *spage = sblock->pagev[i];
2491
2492 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2493 }
2494
2495 btrfs_init_work(&sblock->work, btrfs_scrub_helper,
2496 scrub_missing_raid56_worker, NULL, NULL);
2497 scrub_block_get(sblock);
2498 scrub_pending_bio_inc(sctx);
2499 raid56_submit_missing_rbio(rbio);
2500 return;
2501
2502rbio_out:
2503 bio_put(bio);
2504bbio_out:
2505 btrfs_bio_counter_dec(fs_info);
2506 btrfs_put_bbio(bbio);
2507 spin_lock(&sctx->stat_lock);
2508 sctx->stat.malloc_errors++;
2509 spin_unlock(&sctx->stat_lock);
2510}
2511
2512static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2513 u64 physical, struct btrfs_device *dev, u64 flags,
2514 u64 gen, int mirror_num, u8 *csum, int force,
2515 u64 physical_for_dev_replace)
2516{
2517 struct scrub_block *sblock;
2518 int index;
2519
2520 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2521 if (!sblock) {
2522 spin_lock(&sctx->stat_lock);
2523 sctx->stat.malloc_errors++;
2524 spin_unlock(&sctx->stat_lock);
2525 return -ENOMEM;
2526 }
2527
2528 /* one ref inside this function, plus one for each page added to
2529 * a bio later on */
2530 refcount_set(&sblock->refs, 1);
2531 sblock->sctx = sctx;
2532 sblock->no_io_error_seen = 1;
2533
2534 for (index = 0; len > 0; index++) {
2535 struct scrub_page *spage;
2536 u64 l = min_t(u64, len, PAGE_SIZE);
2537
2538 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2539 if (!spage) {
2540leave_nomem:
2541 spin_lock(&sctx->stat_lock);
2542 sctx->stat.malloc_errors++;
2543 spin_unlock(&sctx->stat_lock);
2544 scrub_block_put(sblock);
2545 return -ENOMEM;
2546 }
2547 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2548 scrub_page_get(spage);
2549 sblock->pagev[index] = spage;
2550 spage->sblock = sblock;
2551 spage->dev = dev;
2552 spage->flags = flags;
2553 spage->generation = gen;
2554 spage->logical = logical;
2555 spage->physical = physical;
2556 spage->physical_for_dev_replace = physical_for_dev_replace;
2557 spage->mirror_num = mirror_num;
2558 if (csum) {
2559 spage->have_csum = 1;
2560 memcpy(spage->csum, csum, sctx->csum_size);
2561 } else {
2562 spage->have_csum = 0;
2563 }
2564 sblock->page_count++;
2565 spage->page = alloc_page(GFP_KERNEL);
2566 if (!spage->page)
2567 goto leave_nomem;
2568 len -= l;
2569 logical += l;
2570 physical += l;
2571 physical_for_dev_replace += l;
2572 }
2573
2574 WARN_ON(sblock->page_count == 0);
2575 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2576 /*
2577 * This case should only be hit for RAID 5/6 device replace. See
2578 * the comment in scrub_missing_raid56_pages() for details.
2579 */
2580 scrub_missing_raid56_pages(sblock);
2581 } else {
2582 for (index = 0; index < sblock->page_count; index++) {
2583 struct scrub_page *spage = sblock->pagev[index];
2584 int ret;
2585
2586 ret = scrub_add_page_to_rd_bio(sctx, spage);
2587 if (ret) {
2588 scrub_block_put(sblock);
2589 return ret;
2590 }
2591 }
2592
2593 if (force)
2594 scrub_submit(sctx);
2595 }
2596
2597 /* last one frees, either here or in bio completion for last page */
2598 scrub_block_put(sblock);
2599 return 0;
2600}
2601
2602static void scrub_bio_end_io(struct bio *bio)
2603{
2604 struct scrub_bio *sbio = bio->bi_private;
2605 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2606
2607 sbio->status = bio->bi_status;
2608 sbio->bio = bio;
2609
2610 btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2611}
2612
2613static void scrub_bio_end_io_worker(struct btrfs_work *work)
2614{
2615 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2616 struct scrub_ctx *sctx = sbio->sctx;
2617 int i;
2618
2619 BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2620 if (sbio->status) {
2621 for (i = 0; i < sbio->page_count; i++) {
2622 struct scrub_page *spage = sbio->pagev[i];
2623
2624 spage->io_error = 1;
2625 spage->sblock->no_io_error_seen = 0;
2626 }
2627 }
2628
2629 /* now complete the scrub_block items that have all pages completed */
2630 for (i = 0; i < sbio->page_count; i++) {
2631 struct scrub_page *spage = sbio->pagev[i];
2632 struct scrub_block *sblock = spage->sblock;
2633
2634 if (atomic_dec_and_test(&sblock->outstanding_pages))
2635 scrub_block_complete(sblock);
2636 scrub_block_put(sblock);
2637 }
2638
2639 bio_put(sbio->bio);
2640 sbio->bio = NULL;
2641 spin_lock(&sctx->list_lock);
2642 sbio->next_free = sctx->first_free;
2643 sctx->first_free = sbio->index;
2644 spin_unlock(&sctx->list_lock);
2645
2646 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2647 mutex_lock(&sctx->wr_lock);
2648 scrub_wr_submit(sctx);
2649 mutex_unlock(&sctx->wr_lock);
2650 }
2651
2652 scrub_pending_bio_dec(sctx);
2653}
2654
2655static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2656 unsigned long *bitmap,
2657 u64 start, u64 len)
2658{
2659 u64 offset;
2660 u64 nsectors64;
2661 u32 nsectors;
2662 int sectorsize = sparity->sctx->fs_info->sectorsize;
2663
2664 if (len >= sparity->stripe_len) {
2665 bitmap_set(bitmap, 0, sparity->nsectors);
2666 return;
2667 }
2668
2669 start -= sparity->logic_start;
2670 start = div64_u64_rem(start, sparity->stripe_len, &offset);
2671 offset = div_u64(offset, sectorsize);
2672 nsectors64 = div_u64(len, sectorsize);
2673
2674 ASSERT(nsectors64 < UINT_MAX);
2675 nsectors = (u32)nsectors64;
2676
2677 if (offset + nsectors <= sparity->nsectors) {
2678 bitmap_set(bitmap, offset, nsectors);
2679 return;
2680 }
2681
2682 bitmap_set(bitmap, offset, sparity->nsectors - offset);
2683 bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2684}
2685
2686static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2687 u64 start, u64 len)
2688{
2689 __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2690}
2691
2692static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2693 u64 start, u64 len)
2694{
2695 __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2696}
2697
2698static void scrub_block_complete(struct scrub_block *sblock)
2699{
2700 int corrupted = 0;
2701
2702 if (!sblock->no_io_error_seen) {
2703 corrupted = 1;
2704 scrub_handle_errored_block(sblock);
2705 } else {
2706 /*
2707 * if has checksum error, write via repair mechanism in
2708 * dev replace case, otherwise write here in dev replace
2709 * case.
2710 */
2711 corrupted = scrub_checksum(sblock);
2712 if (!corrupted && sblock->sctx->is_dev_replace)
2713 scrub_write_block_to_dev_replace(sblock);
2714 }
2715
2716 if (sblock->sparity && corrupted && !sblock->data_corrected) {
2717 u64 start = sblock->pagev[0]->logical;
2718 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2719 PAGE_SIZE;
2720
2721 scrub_parity_mark_sectors_error(sblock->sparity,
2722 start, end - start);
2723 }
2724}
2725
2726static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2727{
2728 struct btrfs_ordered_sum *sum = NULL;
2729 unsigned long index;
2730 unsigned long num_sectors;
2731
2732 while (!list_empty(&sctx->csum_list)) {
2733 sum = list_first_entry(&sctx->csum_list,
2734 struct btrfs_ordered_sum, list);
2735 if (sum->bytenr > logical)
2736 return 0;
2737 if (sum->bytenr + sum->len > logical)
2738 break;
2739
2740 ++sctx->stat.csum_discards;
2741 list_del(&sum->list);
2742 kfree(sum);
2743 sum = NULL;
2744 }
2745 if (!sum)
2746 return 0;
2747
2748 index = div_u64(logical - sum->bytenr, sctx->fs_info->sectorsize);
2749 ASSERT(index < UINT_MAX);
2750
2751 num_sectors = sum->len / sctx->fs_info->sectorsize;
2752 memcpy(csum, sum->sums + index, sctx->csum_size);
2753 if (index == num_sectors - 1) {
2754 list_del(&sum->list);
2755 kfree(sum);
2756 }
2757 return 1;
2758}
2759
2760/* scrub extent tries to collect up to 64 kB for each bio */
2761static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2762 u64 logical, u64 len,
2763 u64 physical, struct btrfs_device *dev, u64 flags,
2764 u64 gen, int mirror_num, u64 physical_for_dev_replace)
2765{
2766 int ret;
2767 u8 csum[BTRFS_CSUM_SIZE];
2768 u32 blocksize;
2769
2770 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2771 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2772 blocksize = map->stripe_len;
2773 else
2774 blocksize = sctx->fs_info->sectorsize;
2775 spin_lock(&sctx->stat_lock);
2776 sctx->stat.data_extents_scrubbed++;
2777 sctx->stat.data_bytes_scrubbed += len;
2778 spin_unlock(&sctx->stat_lock);
2779 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2780 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2781 blocksize = map->stripe_len;
2782 else
2783 blocksize = sctx->fs_info->nodesize;
2784 spin_lock(&sctx->stat_lock);
2785 sctx->stat.tree_extents_scrubbed++;
2786 sctx->stat.tree_bytes_scrubbed += len;
2787 spin_unlock(&sctx->stat_lock);
2788 } else {
2789 blocksize = sctx->fs_info->sectorsize;
2790 WARN_ON(1);
2791 }
2792
2793 while (len) {
2794 u64 l = min_t(u64, len, blocksize);
2795 int have_csum = 0;
2796
2797 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2798 /* push csums to sbio */
2799 have_csum = scrub_find_csum(sctx, logical, csum);
2800 if (have_csum == 0)
2801 ++sctx->stat.no_csum;
2802 if (sctx->is_dev_replace && !have_csum) {
2803 ret = copy_nocow_pages(sctx, logical, l,
2804 mirror_num,
2805 physical_for_dev_replace);
2806 goto behind_scrub_pages;
2807 }
2808 }
2809 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2810 mirror_num, have_csum ? csum : NULL, 0,
2811 physical_for_dev_replace);
2812behind_scrub_pages:
2813 if (ret)
2814 return ret;
2815 len -= l;
2816 logical += l;
2817 physical += l;
2818 physical_for_dev_replace += l;
2819 }
2820 return 0;
2821}
2822
2823static int scrub_pages_for_parity(struct scrub_parity *sparity,
2824 u64 logical, u64 len,
2825 u64 physical, struct btrfs_device *dev,
2826 u64 flags, u64 gen, int mirror_num, u8 *csum)
2827{
2828 struct scrub_ctx *sctx = sparity->sctx;
2829 struct scrub_block *sblock;
2830 int index;
2831
2832 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2833 if (!sblock) {
2834 spin_lock(&sctx->stat_lock);
2835 sctx->stat.malloc_errors++;
2836 spin_unlock(&sctx->stat_lock);
2837 return -ENOMEM;
2838 }
2839
2840 /* one ref inside this function, plus one for each page added to
2841 * a bio later on */
2842 refcount_set(&sblock->refs, 1);
2843 sblock->sctx = sctx;
2844 sblock->no_io_error_seen = 1;
2845 sblock->sparity = sparity;
2846 scrub_parity_get(sparity);
2847
2848 for (index = 0; len > 0; index++) {
2849 struct scrub_page *spage;
2850 u64 l = min_t(u64, len, PAGE_SIZE);
2851
2852 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2853 if (!spage) {
2854leave_nomem:
2855 spin_lock(&sctx->stat_lock);
2856 sctx->stat.malloc_errors++;
2857 spin_unlock(&sctx->stat_lock);
2858 scrub_block_put(sblock);
2859 return -ENOMEM;
2860 }
2861 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2862 /* For scrub block */
2863 scrub_page_get(spage);
2864 sblock->pagev[index] = spage;
2865 /* For scrub parity */
2866 scrub_page_get(spage);
2867 list_add_tail(&spage->list, &sparity->spages);
2868 spage->sblock = sblock;
2869 spage->dev = dev;
2870 spage->flags = flags;
2871 spage->generation = gen;
2872 spage->logical = logical;
2873 spage->physical = physical;
2874 spage->mirror_num = mirror_num;
2875 if (csum) {
2876 spage->have_csum = 1;
2877 memcpy(spage->csum, csum, sctx->csum_size);
2878 } else {
2879 spage->have_csum = 0;
2880 }
2881 sblock->page_count++;
2882 spage->page = alloc_page(GFP_KERNEL);
2883 if (!spage->page)
2884 goto leave_nomem;
2885 len -= l;
2886 logical += l;
2887 physical += l;
2888 }
2889
2890 WARN_ON(sblock->page_count == 0);
2891 for (index = 0; index < sblock->page_count; index++) {
2892 struct scrub_page *spage = sblock->pagev[index];
2893 int ret;
2894
2895 ret = scrub_add_page_to_rd_bio(sctx, spage);
2896 if (ret) {
2897 scrub_block_put(sblock);
2898 return ret;
2899 }
2900 }
2901
2902 /* last one frees, either here or in bio completion for last page */
2903 scrub_block_put(sblock);
2904 return 0;
2905}
2906
2907static int scrub_extent_for_parity(struct scrub_parity *sparity,
2908 u64 logical, u64 len,
2909 u64 physical, struct btrfs_device *dev,
2910 u64 flags, u64 gen, int mirror_num)
2911{
2912 struct scrub_ctx *sctx = sparity->sctx;
2913 int ret;
2914 u8 csum[BTRFS_CSUM_SIZE];
2915 u32 blocksize;
2916
2917 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2918 scrub_parity_mark_sectors_error(sparity, logical, len);
2919 return 0;
2920 }
2921
2922 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2923 blocksize = sparity->stripe_len;
2924 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2925 blocksize = sparity->stripe_len;
2926 } else {
2927 blocksize = sctx->fs_info->sectorsize;
2928 WARN_ON(1);
2929 }
2930
2931 while (len) {
2932 u64 l = min_t(u64, len, blocksize);
2933 int have_csum = 0;
2934
2935 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2936 /* push csums to sbio */
2937 have_csum = scrub_find_csum(sctx, logical, csum);
2938 if (have_csum == 0)
2939 goto skip;
2940 }
2941 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2942 flags, gen, mirror_num,
2943 have_csum ? csum : NULL);
2944 if (ret)
2945 return ret;
2946skip:
2947 len -= l;
2948 logical += l;
2949 physical += l;
2950 }
2951 return 0;
2952}
2953
2954/*
2955 * Given a physical address, this will calculate it's
2956 * logical offset. if this is a parity stripe, it will return
2957 * the most left data stripe's logical offset.
2958 *
2959 * return 0 if it is a data stripe, 1 means parity stripe.
2960 */
2961static int get_raid56_logic_offset(u64 physical, int num,
2962 struct map_lookup *map, u64 *offset,
2963 u64 *stripe_start)
2964{
2965 int i;
2966 int j = 0;
2967 u64 stripe_nr;
2968 u64 last_offset;
2969 u32 stripe_index;
2970 u32 rot;
2971
2972 last_offset = (physical - map->stripes[num].physical) *
2973 nr_data_stripes(map);
2974 if (stripe_start)
2975 *stripe_start = last_offset;
2976
2977 *offset = last_offset;
2978 for (i = 0; i < nr_data_stripes(map); i++) {
2979 *offset = last_offset + i * map->stripe_len;
2980
2981 stripe_nr = div64_u64(*offset, map->stripe_len);
2982 stripe_nr = div_u64(stripe_nr, nr_data_stripes(map));
2983
2984 /* Work out the disk rotation on this stripe-set */
2985 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2986 /* calculate which stripe this data locates */
2987 rot += i;
2988 stripe_index = rot % map->num_stripes;
2989 if (stripe_index == num)
2990 return 0;
2991 if (stripe_index < num)
2992 j++;
2993 }
2994 *offset = last_offset + j * map->stripe_len;
2995 return 1;
2996}
2997
2998static void scrub_free_parity(struct scrub_parity *sparity)
2999{
3000 struct scrub_ctx *sctx = sparity->sctx;
3001 struct scrub_page *curr, *next;
3002 int nbits;
3003
3004 nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
3005 if (nbits) {
3006 spin_lock(&sctx->stat_lock);
3007 sctx->stat.read_errors += nbits;
3008 sctx->stat.uncorrectable_errors += nbits;
3009 spin_unlock(&sctx->stat_lock);
3010 }
3011
3012 list_for_each_entry_safe(curr, next, &sparity->spages, list) {
3013 list_del_init(&curr->list);
3014 scrub_page_put(curr);
3015 }
3016
3017 kfree(sparity);
3018}
3019
3020static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
3021{
3022 struct scrub_parity *sparity = container_of(work, struct scrub_parity,
3023 work);
3024 struct scrub_ctx *sctx = sparity->sctx;
3025
3026 scrub_free_parity(sparity);
3027 scrub_pending_bio_dec(sctx);
3028}
3029
3030static void scrub_parity_bio_endio(struct bio *bio)
3031{
3032 struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
3033 struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
3034
3035 if (bio->bi_status)
3036 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
3037 sparity->nsectors);
3038
3039 bio_put(bio);
3040
3041 btrfs_init_work(&sparity->work, btrfs_scrubparity_helper,
3042 scrub_parity_bio_endio_worker, NULL, NULL);
3043 btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
3044}
3045
3046static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
3047{
3048 struct scrub_ctx *sctx = sparity->sctx;
3049 struct btrfs_fs_info *fs_info = sctx->fs_info;
3050 struct bio *bio;
3051 struct btrfs_raid_bio *rbio;
3052 struct btrfs_bio *bbio = NULL;
3053 u64 length;
3054 int ret;
3055
3056 if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
3057 sparity->nsectors))
3058 goto out;
3059
3060 length = sparity->logic_end - sparity->logic_start;
3061
3062 btrfs_bio_counter_inc_blocked(fs_info);
3063 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
3064 &length, &bbio);
3065 if (ret || !bbio || !bbio->raid_map)
3066 goto bbio_out;
3067
3068 bio = btrfs_io_bio_alloc(0);
3069 bio->bi_iter.bi_sector = sparity->logic_start >> 9;
3070 bio->bi_private = sparity;
3071 bio->bi_end_io = scrub_parity_bio_endio;
3072
3073 rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
3074 length, sparity->scrub_dev,
3075 sparity->dbitmap,
3076 sparity->nsectors);
3077 if (!rbio)
3078 goto rbio_out;
3079
3080 scrub_pending_bio_inc(sctx);
3081 raid56_parity_submit_scrub_rbio(rbio);
3082 return;
3083
3084rbio_out:
3085 bio_put(bio);
3086bbio_out:
3087 btrfs_bio_counter_dec(fs_info);
3088 btrfs_put_bbio(bbio);
3089 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
3090 sparity->nsectors);
3091 spin_lock(&sctx->stat_lock);
3092 sctx->stat.malloc_errors++;
3093 spin_unlock(&sctx->stat_lock);
3094out:
3095 scrub_free_parity(sparity);
3096}
3097
3098static inline int scrub_calc_parity_bitmap_len(int nsectors)
3099{
3100 return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
3101}
3102
3103static void scrub_parity_get(struct scrub_parity *sparity)
3104{
3105 refcount_inc(&sparity->refs);
3106}
3107
3108static void scrub_parity_put(struct scrub_parity *sparity)
3109{
3110 if (!refcount_dec_and_test(&sparity->refs))
3111 return;
3112
3113 scrub_parity_check_and_repair(sparity);
3114}
3115
3116static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
3117 struct map_lookup *map,
3118 struct btrfs_device *sdev,
3119 struct btrfs_path *path,
3120 u64 logic_start,
3121 u64 logic_end)
3122{
3123 struct btrfs_fs_info *fs_info = sctx->fs_info;
3124 struct btrfs_root *root = fs_info->extent_root;
3125 struct btrfs_root *csum_root = fs_info->csum_root;
3126 struct btrfs_extent_item *extent;
3127 struct btrfs_bio *bbio = NULL;
3128 u64 flags;
3129 int ret;
3130 int slot;
3131 struct extent_buffer *l;
3132 struct btrfs_key key;
3133 u64 generation;
3134 u64 extent_logical;
3135 u64 extent_physical;
3136 u64 extent_len;
3137 u64 mapped_length;
3138 struct btrfs_device *extent_dev;
3139 struct scrub_parity *sparity;
3140 int nsectors;
3141 int bitmap_len;
3142 int extent_mirror_num;
3143 int stop_loop = 0;
3144
3145 nsectors = div_u64(map->stripe_len, fs_info->sectorsize);
3146 bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
3147 sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
3148 GFP_NOFS);
3149 if (!sparity) {
3150 spin_lock(&sctx->stat_lock);
3151 sctx->stat.malloc_errors++;
3152 spin_unlock(&sctx->stat_lock);
3153 return -ENOMEM;
3154 }
3155
3156 sparity->stripe_len = map->stripe_len;
3157 sparity->nsectors = nsectors;
3158 sparity->sctx = sctx;
3159 sparity->scrub_dev = sdev;
3160 sparity->logic_start = logic_start;
3161 sparity->logic_end = logic_end;
3162 refcount_set(&sparity->refs, 1);
3163 INIT_LIST_HEAD(&sparity->spages);
3164 sparity->dbitmap = sparity->bitmap;
3165 sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
3166
3167 ret = 0;
3168 while (logic_start < logic_end) {
3169 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3170 key.type = BTRFS_METADATA_ITEM_KEY;
3171 else
3172 key.type = BTRFS_EXTENT_ITEM_KEY;
3173 key.objectid = logic_start;
3174 key.offset = (u64)-1;
3175
3176 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3177 if (ret < 0)
3178 goto out;
3179
3180 if (ret > 0) {
3181 ret = btrfs_previous_extent_item(root, path, 0);
3182 if (ret < 0)
3183 goto out;
3184 if (ret > 0) {
3185 btrfs_release_path(path);
3186 ret = btrfs_search_slot(NULL, root, &key,
3187 path, 0, 0);
3188 if (ret < 0)
3189 goto out;
3190 }
3191 }
3192
3193 stop_loop = 0;
3194 while (1) {
3195 u64 bytes;
3196
3197 l = path->nodes[0];
3198 slot = path->slots[0];
3199 if (slot >= btrfs_header_nritems(l)) {
3200 ret = btrfs_next_leaf(root, path);
3201 if (ret == 0)
3202 continue;
3203 if (ret < 0)
3204 goto out;
3205
3206 stop_loop = 1;
3207 break;
3208 }
3209 btrfs_item_key_to_cpu(l, &key, slot);
3210
3211 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3212 key.type != BTRFS_METADATA_ITEM_KEY)
3213 goto next;
3214
3215 if (key.type == BTRFS_METADATA_ITEM_KEY)
3216 bytes = fs_info->nodesize;
3217 else
3218 bytes = key.offset;
3219
3220 if (key.objectid + bytes <= logic_start)
3221 goto next;
3222
3223 if (key.objectid >= logic_end) {
3224 stop_loop = 1;
3225 break;
3226 }
3227
3228 while (key.objectid >= logic_start + map->stripe_len)
3229 logic_start += map->stripe_len;
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 < logic_start ||
3238 key.objectid + bytes >
3239 logic_start + map->stripe_len)) {
3240 btrfs_err(fs_info,
3241 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3242 key.objectid, logic_start);
3243 spin_lock(&sctx->stat_lock);
3244 sctx->stat.uncorrectable_errors++;
3245 spin_unlock(&sctx->stat_lock);
3246 goto next;
3247 }
3248again:
3249 extent_logical = key.objectid;
3250 extent_len = bytes;
3251
3252 if (extent_logical < logic_start) {
3253 extent_len -= logic_start - extent_logical;
3254 extent_logical = logic_start;
3255 }
3256
3257 if (extent_logical + extent_len >
3258 logic_start + map->stripe_len)
3259 extent_len = logic_start + map->stripe_len -
3260 extent_logical;
3261
3262 scrub_parity_mark_sectors_data(sparity, extent_logical,
3263 extent_len);
3264
3265 mapped_length = extent_len;
3266 bbio = NULL;
3267 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
3268 extent_logical, &mapped_length, &bbio,
3269 0);
3270 if (!ret) {
3271 if (!bbio || mapped_length < extent_len)
3272 ret = -EIO;
3273 }
3274 if (ret) {
3275 btrfs_put_bbio(bbio);
3276 goto out;
3277 }
3278 extent_physical = bbio->stripes[0].physical;
3279 extent_mirror_num = bbio->mirror_num;
3280 extent_dev = bbio->stripes[0].dev;
3281 btrfs_put_bbio(bbio);
3282
3283 ret = btrfs_lookup_csums_range(csum_root,
3284 extent_logical,
3285 extent_logical + extent_len - 1,
3286 &sctx->csum_list, 1);
3287 if (ret)
3288 goto out;
3289
3290 ret = scrub_extent_for_parity(sparity, extent_logical,
3291 extent_len,
3292 extent_physical,
3293 extent_dev, flags,
3294 generation,
3295 extent_mirror_num);
3296
3297 scrub_free_csums(sctx);
3298
3299 if (ret)
3300 goto out;
3301
3302 if (extent_logical + extent_len <
3303 key.objectid + bytes) {
3304 logic_start += map->stripe_len;
3305
3306 if (logic_start >= logic_end) {
3307 stop_loop = 1;
3308 break;
3309 }
3310
3311 if (logic_start < key.objectid + bytes) {
3312 cond_resched();
3313 goto again;
3314 }
3315 }
3316next:
3317 path->slots[0]++;
3318 }
3319
3320 btrfs_release_path(path);
3321
3322 if (stop_loop)
3323 break;
3324
3325 logic_start += map->stripe_len;
3326 }
3327out:
3328 if (ret < 0)
3329 scrub_parity_mark_sectors_error(sparity, logic_start,
3330 logic_end - logic_start);
3331 scrub_parity_put(sparity);
3332 scrub_submit(sctx);
3333 mutex_lock(&sctx->wr_lock);
3334 scrub_wr_submit(sctx);
3335 mutex_unlock(&sctx->wr_lock);
3336
3337 btrfs_release_path(path);
3338 return ret < 0 ? ret : 0;
3339}
3340
3341static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3342 struct map_lookup *map,
3343 struct btrfs_device *scrub_dev,
3344 int num, u64 base, u64 length,
3345 int is_dev_replace)
3346{
3347 struct btrfs_path *path, *ppath;
3348 struct btrfs_fs_info *fs_info = sctx->fs_info;
3349 struct btrfs_root *root = fs_info->extent_root;
3350 struct btrfs_root *csum_root = fs_info->csum_root;
3351 struct btrfs_extent_item *extent;
3352 struct blk_plug plug;
3353 u64 flags;
3354 int ret;
3355 int slot;
3356 u64 nstripes;
3357 struct extent_buffer *l;
3358 u64 physical;
3359 u64 logical;
3360 u64 logic_end;
3361 u64 physical_end;
3362 u64 generation;
3363 int mirror_num;
3364 struct reada_control *reada1;
3365 struct reada_control *reada2;
3366 struct btrfs_key key;
3367 struct btrfs_key key_end;
3368 u64 increment = map->stripe_len;
3369 u64 offset;
3370 u64 extent_logical;
3371 u64 extent_physical;
3372 u64 extent_len;
3373 u64 stripe_logical;
3374 u64 stripe_end;
3375 struct btrfs_device *extent_dev;
3376 int extent_mirror_num;
3377 int stop_loop = 0;
3378
3379 physical = map->stripes[num].physical;
3380 offset = 0;
3381 nstripes = div64_u64(length, map->stripe_len);
3382 if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3383 offset = map->stripe_len * num;
3384 increment = map->stripe_len * map->num_stripes;
3385 mirror_num = 1;
3386 } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3387 int factor = map->num_stripes / map->sub_stripes;
3388 offset = map->stripe_len * (num / map->sub_stripes);
3389 increment = map->stripe_len * factor;
3390 mirror_num = num % map->sub_stripes + 1;
3391 } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
3392 increment = map->stripe_len;
3393 mirror_num = num % map->num_stripes + 1;
3394 } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3395 increment = map->stripe_len;
3396 mirror_num = num % map->num_stripes + 1;
3397 } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3398 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3399 increment = map->stripe_len * nr_data_stripes(map);
3400 mirror_num = 1;
3401 } else {
3402 increment = map->stripe_len;
3403 mirror_num = 1;
3404 }
3405
3406 path = btrfs_alloc_path();
3407 if (!path)
3408 return -ENOMEM;
3409
3410 ppath = btrfs_alloc_path();
3411 if (!ppath) {
3412 btrfs_free_path(path);
3413 return -ENOMEM;
3414 }
3415
3416 /*
3417 * work on commit root. The related disk blocks are static as
3418 * long as COW is applied. This means, it is save to rewrite
3419 * them to repair disk errors without any race conditions
3420 */
3421 path->search_commit_root = 1;
3422 path->skip_locking = 1;
3423
3424 ppath->search_commit_root = 1;
3425 ppath->skip_locking = 1;
3426 /*
3427 * trigger the readahead for extent tree csum tree and wait for
3428 * completion. During readahead, the scrub is officially paused
3429 * to not hold off transaction commits
3430 */
3431 logical = base + offset;
3432 physical_end = physical + nstripes * map->stripe_len;
3433 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3434 get_raid56_logic_offset(physical_end, num,
3435 map, &logic_end, NULL);
3436 logic_end += base;
3437 } else {
3438 logic_end = logical + increment * nstripes;
3439 }
3440 wait_event(sctx->list_wait,
3441 atomic_read(&sctx->bios_in_flight) == 0);
3442 scrub_blocked_if_needed(fs_info);
3443
3444 /* FIXME it might be better to start readahead at commit root */
3445 key.objectid = logical;
3446 key.type = BTRFS_EXTENT_ITEM_KEY;
3447 key.offset = (u64)0;
3448 key_end.objectid = logic_end;
3449 key_end.type = BTRFS_METADATA_ITEM_KEY;
3450 key_end.offset = (u64)-1;
3451 reada1 = btrfs_reada_add(root, &key, &key_end);
3452
3453 key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3454 key.type = BTRFS_EXTENT_CSUM_KEY;
3455 key.offset = logical;
3456 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3457 key_end.type = BTRFS_EXTENT_CSUM_KEY;
3458 key_end.offset = logic_end;
3459 reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3460
3461 if (!IS_ERR(reada1))
3462 btrfs_reada_wait(reada1);
3463 if (!IS_ERR(reada2))
3464 btrfs_reada_wait(reada2);
3465
3466
3467 /*
3468 * collect all data csums for the stripe to avoid seeking during
3469 * the scrub. This might currently (crc32) end up to be about 1MB
3470 */
3471 blk_start_plug(&plug);
3472
3473 /*
3474 * now find all extents for each stripe and scrub them
3475 */
3476 ret = 0;
3477 while (physical < physical_end) {
3478 /*
3479 * canceled?
3480 */
3481 if (atomic_read(&fs_info->scrub_cancel_req) ||
3482 atomic_read(&sctx->cancel_req)) {
3483 ret = -ECANCELED;
3484 goto out;
3485 }
3486 /*
3487 * check to see if we have to pause
3488 */
3489 if (atomic_read(&fs_info->scrub_pause_req)) {
3490 /* push queued extents */
3491 sctx->flush_all_writes = true;
3492 scrub_submit(sctx);
3493 mutex_lock(&sctx->wr_lock);
3494 scrub_wr_submit(sctx);
3495 mutex_unlock(&sctx->wr_lock);
3496 wait_event(sctx->list_wait,
3497 atomic_read(&sctx->bios_in_flight) == 0);
3498 sctx->flush_all_writes = false;
3499 scrub_blocked_if_needed(fs_info);
3500 }
3501
3502 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3503 ret = get_raid56_logic_offset(physical, num, map,
3504 &logical,
3505 &stripe_logical);
3506 logical += base;
3507 if (ret) {
3508 /* it is parity strip */
3509 stripe_logical += base;
3510 stripe_end = stripe_logical + increment;
3511 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3512 ppath, stripe_logical,
3513 stripe_end);
3514 if (ret)
3515 goto out;
3516 goto skip;
3517 }
3518 }
3519
3520 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3521 key.type = BTRFS_METADATA_ITEM_KEY;
3522 else
3523 key.type = BTRFS_EXTENT_ITEM_KEY;
3524 key.objectid = logical;
3525 key.offset = (u64)-1;
3526
3527 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3528 if (ret < 0)
3529 goto out;
3530
3531 if (ret > 0) {
3532 ret = btrfs_previous_extent_item(root, path, 0);
3533 if (ret < 0)
3534 goto out;
3535 if (ret > 0) {
3536 /* there's no smaller item, so stick with the
3537 * larger one */
3538 btrfs_release_path(path);
3539 ret = btrfs_search_slot(NULL, root, &key,
3540 path, 0, 0);
3541 if (ret < 0)
3542 goto out;
3543 }
3544 }
3545
3546 stop_loop = 0;
3547 while (1) {
3548 u64 bytes;
3549
3550 l = path->nodes[0];
3551 slot = path->slots[0];
3552 if (slot >= btrfs_header_nritems(l)) {
3553 ret = btrfs_next_leaf(root, path);
3554 if (ret == 0)
3555 continue;
3556 if (ret < 0)
3557 goto out;
3558
3559 stop_loop = 1;
3560 break;
3561 }
3562 btrfs_item_key_to_cpu(l, &key, slot);
3563
3564 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3565 key.type != BTRFS_METADATA_ITEM_KEY)
3566 goto next;
3567
3568 if (key.type == BTRFS_METADATA_ITEM_KEY)
3569 bytes = fs_info->nodesize;
3570 else
3571 bytes = key.offset;
3572
3573 if (key.objectid + bytes <= logical)
3574 goto next;
3575
3576 if (key.objectid >= logical + map->stripe_len) {
3577 /* out of this device extent */
3578 if (key.objectid >= logic_end)
3579 stop_loop = 1;
3580 break;
3581 }
3582
3583 extent = btrfs_item_ptr(l, slot,
3584 struct btrfs_extent_item);
3585 flags = btrfs_extent_flags(l, extent);
3586 generation = btrfs_extent_generation(l, extent);
3587
3588 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3589 (key.objectid < logical ||
3590 key.objectid + bytes >
3591 logical + map->stripe_len)) {
3592 btrfs_err(fs_info,
3593 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3594 key.objectid, logical);
3595 spin_lock(&sctx->stat_lock);
3596 sctx->stat.uncorrectable_errors++;
3597 spin_unlock(&sctx->stat_lock);
3598 goto next;
3599 }
3600
3601again:
3602 extent_logical = key.objectid;
3603 extent_len = bytes;
3604
3605 /*
3606 * trim extent to this stripe
3607 */
3608 if (extent_logical < logical) {
3609 extent_len -= logical - extent_logical;
3610 extent_logical = logical;
3611 }
3612 if (extent_logical + extent_len >
3613 logical + map->stripe_len) {
3614 extent_len = logical + map->stripe_len -
3615 extent_logical;
3616 }
3617
3618 extent_physical = extent_logical - logical + physical;
3619 extent_dev = scrub_dev;
3620 extent_mirror_num = mirror_num;
3621 if (is_dev_replace)
3622 scrub_remap_extent(fs_info, extent_logical,
3623 extent_len, &extent_physical,
3624 &extent_dev,
3625 &extent_mirror_num);
3626
3627 ret = btrfs_lookup_csums_range(csum_root,
3628 extent_logical,
3629 extent_logical +
3630 extent_len - 1,
3631 &sctx->csum_list, 1);
3632 if (ret)
3633 goto out;
3634
3635 ret = scrub_extent(sctx, map, extent_logical, extent_len,
3636 extent_physical, extent_dev, flags,
3637 generation, extent_mirror_num,
3638 extent_logical - logical + physical);
3639
3640 scrub_free_csums(sctx);
3641
3642 if (ret)
3643 goto out;
3644
3645 if (extent_logical + extent_len <
3646 key.objectid + bytes) {
3647 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3648 /*
3649 * loop until we find next data stripe
3650 * or we have finished all stripes.
3651 */
3652loop:
3653 physical += map->stripe_len;
3654 ret = get_raid56_logic_offset(physical,
3655 num, map, &logical,
3656 &stripe_logical);
3657 logical += base;
3658
3659 if (ret && physical < physical_end) {
3660 stripe_logical += base;
3661 stripe_end = stripe_logical +
3662 increment;
3663 ret = scrub_raid56_parity(sctx,
3664 map, scrub_dev, ppath,
3665 stripe_logical,
3666 stripe_end);
3667 if (ret)
3668 goto out;
3669 goto loop;
3670 }
3671 } else {
3672 physical += map->stripe_len;
3673 logical += increment;
3674 }
3675 if (logical < key.objectid + bytes) {
3676 cond_resched();
3677 goto again;
3678 }
3679
3680 if (physical >= physical_end) {
3681 stop_loop = 1;
3682 break;
3683 }
3684 }
3685next:
3686 path->slots[0]++;
3687 }
3688 btrfs_release_path(path);
3689skip:
3690 logical += increment;
3691 physical += map->stripe_len;
3692 spin_lock(&sctx->stat_lock);
3693 if (stop_loop)
3694 sctx->stat.last_physical = map->stripes[num].physical +
3695 length;
3696 else
3697 sctx->stat.last_physical = physical;
3698 spin_unlock(&sctx->stat_lock);
3699 if (stop_loop)
3700 break;
3701 }
3702out:
3703 /* push queued extents */
3704 scrub_submit(sctx);
3705 mutex_lock(&sctx->wr_lock);
3706 scrub_wr_submit(sctx);
3707 mutex_unlock(&sctx->wr_lock);
3708
3709 blk_finish_plug(&plug);
3710 btrfs_free_path(path);
3711 btrfs_free_path(ppath);
3712 return ret < 0 ? ret : 0;
3713}
3714
3715static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3716 struct btrfs_device *scrub_dev,
3717 u64 chunk_offset, u64 length,
3718 u64 dev_offset,
3719 struct btrfs_block_group_cache *cache,
3720 int is_dev_replace)
3721{
3722 struct btrfs_fs_info *fs_info = sctx->fs_info;
3723 struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
3724 struct map_lookup *map;
3725 struct extent_map *em;
3726 int i;
3727 int ret = 0;
3728
3729 read_lock(&map_tree->map_tree.lock);
3730 em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
3731 read_unlock(&map_tree->map_tree.lock);
3732
3733 if (!em) {
3734 /*
3735 * Might have been an unused block group deleted by the cleaner
3736 * kthread or relocation.
3737 */
3738 spin_lock(&cache->lock);
3739 if (!cache->removed)
3740 ret = -EINVAL;
3741 spin_unlock(&cache->lock);
3742
3743 return ret;
3744 }
3745
3746 map = em->map_lookup;
3747 if (em->start != chunk_offset)
3748 goto out;
3749
3750 if (em->len < length)
3751 goto out;
3752
3753 for (i = 0; i < map->num_stripes; ++i) {
3754 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3755 map->stripes[i].physical == dev_offset) {
3756 ret = scrub_stripe(sctx, map, scrub_dev, i,
3757 chunk_offset, length,
3758 is_dev_replace);
3759 if (ret)
3760 goto out;
3761 }
3762 }
3763out:
3764 free_extent_map(em);
3765
3766 return ret;
3767}
3768
3769static noinline_for_stack
3770int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3771 struct btrfs_device *scrub_dev, u64 start, u64 end,
3772 int is_dev_replace)
3773{
3774 struct btrfs_dev_extent *dev_extent = NULL;
3775 struct btrfs_path *path;
3776 struct btrfs_fs_info *fs_info = sctx->fs_info;
3777 struct btrfs_root *root = fs_info->dev_root;
3778 u64 length;
3779 u64 chunk_offset;
3780 int ret = 0;
3781 int ro_set;
3782 int slot;
3783 struct extent_buffer *l;
3784 struct btrfs_key key;
3785 struct btrfs_key found_key;
3786 struct btrfs_block_group_cache *cache;
3787 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3788
3789 path = btrfs_alloc_path();
3790 if (!path)
3791 return -ENOMEM;
3792
3793 path->reada = READA_FORWARD;
3794 path->search_commit_root = 1;
3795 path->skip_locking = 1;
3796
3797 key.objectid = scrub_dev->devid;
3798 key.offset = 0ull;
3799 key.type = BTRFS_DEV_EXTENT_KEY;
3800
3801 while (1) {
3802 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3803 if (ret < 0)
3804 break;
3805 if (ret > 0) {
3806 if (path->slots[0] >=
3807 btrfs_header_nritems(path->nodes[0])) {
3808 ret = btrfs_next_leaf(root, path);
3809 if (ret < 0)
3810 break;
3811 if (ret > 0) {
3812 ret = 0;
3813 break;
3814 }
3815 } else {
3816 ret = 0;
3817 }
3818 }
3819
3820 l = path->nodes[0];
3821 slot = path->slots[0];
3822
3823 btrfs_item_key_to_cpu(l, &found_key, slot);
3824
3825 if (found_key.objectid != scrub_dev->devid)
3826 break;
3827
3828 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3829 break;
3830
3831 if (found_key.offset >= end)
3832 break;
3833
3834 if (found_key.offset < key.offset)
3835 break;
3836
3837 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3838 length = btrfs_dev_extent_length(l, dev_extent);
3839
3840 if (found_key.offset + length <= start)
3841 goto skip;
3842
3843 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3844
3845 /*
3846 * get a reference on the corresponding block group to prevent
3847 * the chunk from going away while we scrub it
3848 */
3849 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3850
3851 /* some chunks are removed but not committed to disk yet,
3852 * continue scrubbing */
3853 if (!cache)
3854 goto skip;
3855
3856 /*
3857 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3858 * to avoid deadlock caused by:
3859 * btrfs_inc_block_group_ro()
3860 * -> btrfs_wait_for_commit()
3861 * -> btrfs_commit_transaction()
3862 * -> btrfs_scrub_pause()
3863 */
3864 scrub_pause_on(fs_info);
3865 ret = btrfs_inc_block_group_ro(fs_info, cache);
3866 if (!ret && is_dev_replace) {
3867 /*
3868 * If we are doing a device replace wait for any tasks
3869 * that started dellaloc right before we set the block
3870 * group to RO mode, as they might have just allocated
3871 * an extent from it or decided they could do a nocow
3872 * write. And if any such tasks did that, wait for their
3873 * ordered extents to complete and then commit the
3874 * current transaction, so that we can later see the new
3875 * extent items in the extent tree - the ordered extents
3876 * create delayed data references (for cow writes) when
3877 * they complete, which will be run and insert the
3878 * corresponding extent items into the extent tree when
3879 * we commit the transaction they used when running
3880 * inode.c:btrfs_finish_ordered_io(). We later use
3881 * the commit root of the extent tree to find extents
3882 * to copy from the srcdev into the tgtdev, and we don't
3883 * want to miss any new extents.
3884 */
3885 btrfs_wait_block_group_reservations(cache);
3886 btrfs_wait_nocow_writers(cache);
3887 ret = btrfs_wait_ordered_roots(fs_info, U64_MAX,
3888 cache->key.objectid,
3889 cache->key.offset);
3890 if (ret > 0) {
3891 struct btrfs_trans_handle *trans;
3892
3893 trans = btrfs_join_transaction(root);
3894 if (IS_ERR(trans))
3895 ret = PTR_ERR(trans);
3896 else
3897 ret = btrfs_commit_transaction(trans);
3898 if (ret) {
3899 scrub_pause_off(fs_info);
3900 btrfs_put_block_group(cache);
3901 break;
3902 }
3903 }
3904 }
3905 scrub_pause_off(fs_info);
3906
3907 if (ret == 0) {
3908 ro_set = 1;
3909 } else if (ret == -ENOSPC) {
3910 /*
3911 * btrfs_inc_block_group_ro return -ENOSPC when it
3912 * failed in creating new chunk for metadata.
3913 * It is not a problem for scrub/replace, because
3914 * metadata are always cowed, and our scrub paused
3915 * commit_transactions.
3916 */
3917 ro_set = 0;
3918 } else {
3919 btrfs_warn(fs_info,
3920 "failed setting block group ro: %d", ret);
3921 btrfs_put_block_group(cache);
3922 break;
3923 }
3924
3925 btrfs_dev_replace_write_lock(&fs_info->dev_replace);
3926 dev_replace->cursor_right = found_key.offset + length;
3927 dev_replace->cursor_left = found_key.offset;
3928 dev_replace->item_needs_writeback = 1;
3929 btrfs_dev_replace_write_unlock(&fs_info->dev_replace);
3930 ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3931 found_key.offset, cache, is_dev_replace);
3932
3933 /*
3934 * flush, submit all pending read and write bios, afterwards
3935 * wait for them.
3936 * Note that in the dev replace case, a read request causes
3937 * write requests that are submitted in the read completion
3938 * worker. Therefore in the current situation, it is required
3939 * that all write requests are flushed, so that all read and
3940 * write requests are really completed when bios_in_flight
3941 * changes to 0.
3942 */
3943 sctx->flush_all_writes = true;
3944 scrub_submit(sctx);
3945 mutex_lock(&sctx->wr_lock);
3946 scrub_wr_submit(sctx);
3947 mutex_unlock(&sctx->wr_lock);
3948
3949 wait_event(sctx->list_wait,
3950 atomic_read(&sctx->bios_in_flight) == 0);
3951
3952 scrub_pause_on(fs_info);
3953
3954 /*
3955 * must be called before we decrease @scrub_paused.
3956 * make sure we don't block transaction commit while
3957 * we are waiting pending workers finished.
3958 */
3959 wait_event(sctx->list_wait,
3960 atomic_read(&sctx->workers_pending) == 0);
3961 sctx->flush_all_writes = false;
3962
3963 scrub_pause_off(fs_info);
3964
3965 btrfs_dev_replace_write_lock(&fs_info->dev_replace);
3966 dev_replace->cursor_left = dev_replace->cursor_right;
3967 dev_replace->item_needs_writeback = 1;
3968 btrfs_dev_replace_write_unlock(&fs_info->dev_replace);
3969
3970 if (ro_set)
3971 btrfs_dec_block_group_ro(cache);
3972
3973 /*
3974 * We might have prevented the cleaner kthread from deleting
3975 * this block group if it was already unused because we raced
3976 * and set it to RO mode first. So add it back to the unused
3977 * list, otherwise it might not ever be deleted unless a manual
3978 * balance is triggered or it becomes used and unused again.
3979 */
3980 spin_lock(&cache->lock);
3981 if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3982 btrfs_block_group_used(&cache->item) == 0) {
3983 spin_unlock(&cache->lock);
3984 spin_lock(&fs_info->unused_bgs_lock);
3985 if (list_empty(&cache->bg_list)) {
3986 btrfs_get_block_group(cache);
3987 list_add_tail(&cache->bg_list,
3988 &fs_info->unused_bgs);
3989 }
3990 spin_unlock(&fs_info->unused_bgs_lock);
3991 } else {
3992 spin_unlock(&cache->lock);
3993 }
3994
3995 btrfs_put_block_group(cache);
3996 if (ret)
3997 break;
3998 if (is_dev_replace &&
3999 atomic64_read(&dev_replace->num_write_errors) > 0) {
4000 ret = -EIO;
4001 break;
4002 }
4003 if (sctx->stat.malloc_errors > 0) {
4004 ret = -ENOMEM;
4005 break;
4006 }
4007skip:
4008 key.offset = found_key.offset + length;
4009 btrfs_release_path(path);
4010 }
4011
4012 btrfs_free_path(path);
4013
4014 return ret;
4015}
4016
4017static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
4018 struct btrfs_device *scrub_dev)
4019{
4020 int i;
4021 u64 bytenr;
4022 u64 gen;
4023 int ret;
4024 struct btrfs_fs_info *fs_info = sctx->fs_info;
4025
4026 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
4027 return -EIO;
4028
4029 /* Seed devices of a new filesystem has their own generation. */
4030 if (scrub_dev->fs_devices != fs_info->fs_devices)
4031 gen = scrub_dev->generation;
4032 else
4033 gen = fs_info->last_trans_committed;
4034
4035 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
4036 bytenr = btrfs_sb_offset(i);
4037 if (bytenr + BTRFS_SUPER_INFO_SIZE >
4038 scrub_dev->commit_total_bytes)
4039 break;
4040
4041 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
4042 scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
4043 NULL, 1, bytenr);
4044 if (ret)
4045 return ret;
4046 }
4047 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4048
4049 return 0;
4050}
4051
4052/*
4053 * get a reference count on fs_info->scrub_workers. start worker if necessary
4054 */
4055static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
4056 int is_dev_replace)
4057{
4058 unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
4059 int max_active = fs_info->thread_pool_size;
4060
4061 if (fs_info->scrub_workers_refcnt == 0) {
4062 fs_info->scrub_workers = btrfs_alloc_workqueue(fs_info, "scrub",
4063 flags, is_dev_replace ? 1 : max_active, 4);
4064 if (!fs_info->scrub_workers)
4065 goto fail_scrub_workers;
4066
4067 fs_info->scrub_wr_completion_workers =
4068 btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
4069 max_active, 2);
4070 if (!fs_info->scrub_wr_completion_workers)
4071 goto fail_scrub_wr_completion_workers;
4072
4073 fs_info->scrub_nocow_workers =
4074 btrfs_alloc_workqueue(fs_info, "scrubnc", flags, 1, 0);
4075 if (!fs_info->scrub_nocow_workers)
4076 goto fail_scrub_nocow_workers;
4077 fs_info->scrub_parity_workers =
4078 btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
4079 max_active, 2);
4080 if (!fs_info->scrub_parity_workers)
4081 goto fail_scrub_parity_workers;
4082 }
4083 ++fs_info->scrub_workers_refcnt;
4084 return 0;
4085
4086fail_scrub_parity_workers:
4087 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
4088fail_scrub_nocow_workers:
4089 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
4090fail_scrub_wr_completion_workers:
4091 btrfs_destroy_workqueue(fs_info->scrub_workers);
4092fail_scrub_workers:
4093 return -ENOMEM;
4094}
4095
4096static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
4097{
4098 if (--fs_info->scrub_workers_refcnt == 0) {
4099 btrfs_destroy_workqueue(fs_info->scrub_workers);
4100 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
4101 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
4102 btrfs_destroy_workqueue(fs_info->scrub_parity_workers);
4103 }
4104 WARN_ON(fs_info->scrub_workers_refcnt < 0);
4105}
4106
4107int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
4108 u64 end, struct btrfs_scrub_progress *progress,
4109 int readonly, int is_dev_replace)
4110{
4111 struct scrub_ctx *sctx;
4112 int ret;
4113 struct btrfs_device *dev;
4114 struct rcu_string *name;
4115
4116 if (btrfs_fs_closing(fs_info))
4117 return -EINVAL;
4118
4119 if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
4120 /*
4121 * in this case scrub is unable to calculate the checksum
4122 * the way scrub is implemented. Do not handle this
4123 * situation at all because it won't ever happen.
4124 */
4125 btrfs_err(fs_info,
4126 "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
4127 fs_info->nodesize,
4128 BTRFS_STRIPE_LEN);
4129 return -EINVAL;
4130 }
4131
4132 if (fs_info->sectorsize != PAGE_SIZE) {
4133 /* not supported for data w/o checksums */
4134 btrfs_err_rl(fs_info,
4135 "scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails",
4136 fs_info->sectorsize, PAGE_SIZE);
4137 return -EINVAL;
4138 }
4139
4140 if (fs_info->nodesize >
4141 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
4142 fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
4143 /*
4144 * would exhaust the array bounds of pagev member in
4145 * struct scrub_block
4146 */
4147 btrfs_err(fs_info,
4148 "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
4149 fs_info->nodesize,
4150 SCRUB_MAX_PAGES_PER_BLOCK,
4151 fs_info->sectorsize,
4152 SCRUB_MAX_PAGES_PER_BLOCK);
4153 return -EINVAL;
4154 }
4155
4156
4157 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4158 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
4159 if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
4160 !is_dev_replace)) {
4161 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4162 return -ENODEV;
4163 }
4164
4165 if (!is_dev_replace && !readonly &&
4166 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
4167 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4168 rcu_read_lock();
4169 name = rcu_dereference(dev->name);
4170 btrfs_err(fs_info, "scrub: device %s is not writable",
4171 name->str);
4172 rcu_read_unlock();
4173 return -EROFS;
4174 }
4175
4176 mutex_lock(&fs_info->scrub_lock);
4177 if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
4178 test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
4179 mutex_unlock(&fs_info->scrub_lock);
4180 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4181 return -EIO;
4182 }
4183
4184 btrfs_dev_replace_read_lock(&fs_info->dev_replace);
4185 if (dev->scrub_ctx ||
4186 (!is_dev_replace &&
4187 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
4188 btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
4189 mutex_unlock(&fs_info->scrub_lock);
4190 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4191 return -EINPROGRESS;
4192 }
4193 btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
4194
4195 ret = scrub_workers_get(fs_info, is_dev_replace);
4196 if (ret) {
4197 mutex_unlock(&fs_info->scrub_lock);
4198 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4199 return ret;
4200 }
4201
4202 sctx = scrub_setup_ctx(dev, is_dev_replace);
4203 if (IS_ERR(sctx)) {
4204 mutex_unlock(&fs_info->scrub_lock);
4205 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4206 scrub_workers_put(fs_info);
4207 return PTR_ERR(sctx);
4208 }
4209 sctx->readonly = readonly;
4210 dev->scrub_ctx = sctx;
4211 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4212
4213 /*
4214 * checking @scrub_pause_req here, we can avoid
4215 * race between committing transaction and scrubbing.
4216 */
4217 __scrub_blocked_if_needed(fs_info);
4218 atomic_inc(&fs_info->scrubs_running);
4219 mutex_unlock(&fs_info->scrub_lock);
4220
4221 if (!is_dev_replace) {
4222 /*
4223 * by holding device list mutex, we can
4224 * kick off writing super in log tree sync.
4225 */
4226 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4227 ret = scrub_supers(sctx, dev);
4228 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4229 }
4230
4231 if (!ret)
4232 ret = scrub_enumerate_chunks(sctx, dev, start, end,
4233 is_dev_replace);
4234
4235 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4236 atomic_dec(&fs_info->scrubs_running);
4237 wake_up(&fs_info->scrub_pause_wait);
4238
4239 wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
4240
4241 if (progress)
4242 memcpy(progress, &sctx->stat, sizeof(*progress));
4243
4244 mutex_lock(&fs_info->scrub_lock);
4245 dev->scrub_ctx = NULL;
4246 scrub_workers_put(fs_info);
4247 mutex_unlock(&fs_info->scrub_lock);
4248
4249 scrub_put_ctx(sctx);
4250
4251 return ret;
4252}
4253
4254void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
4255{
4256 mutex_lock(&fs_info->scrub_lock);
4257 atomic_inc(&fs_info->scrub_pause_req);
4258 while (atomic_read(&fs_info->scrubs_paused) !=
4259 atomic_read(&fs_info->scrubs_running)) {
4260 mutex_unlock(&fs_info->scrub_lock);
4261 wait_event(fs_info->scrub_pause_wait,
4262 atomic_read(&fs_info->scrubs_paused) ==
4263 atomic_read(&fs_info->scrubs_running));
4264 mutex_lock(&fs_info->scrub_lock);
4265 }
4266 mutex_unlock(&fs_info->scrub_lock);
4267}
4268
4269void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
4270{
4271 atomic_dec(&fs_info->scrub_pause_req);
4272 wake_up(&fs_info->scrub_pause_wait);
4273}
4274
4275int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
4276{
4277 mutex_lock(&fs_info->scrub_lock);
4278 if (!atomic_read(&fs_info->scrubs_running)) {
4279 mutex_unlock(&fs_info->scrub_lock);
4280 return -ENOTCONN;
4281 }
4282
4283 atomic_inc(&fs_info->scrub_cancel_req);
4284 while (atomic_read(&fs_info->scrubs_running)) {
4285 mutex_unlock(&fs_info->scrub_lock);
4286 wait_event(fs_info->scrub_pause_wait,
4287 atomic_read(&fs_info->scrubs_running) == 0);
4288 mutex_lock(&fs_info->scrub_lock);
4289 }
4290 atomic_dec(&fs_info->scrub_cancel_req);
4291 mutex_unlock(&fs_info->scrub_lock);
4292
4293 return 0;
4294}
4295
4296int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
4297 struct btrfs_device *dev)
4298{
4299 struct scrub_ctx *sctx;
4300
4301 mutex_lock(&fs_info->scrub_lock);
4302 sctx = dev->scrub_ctx;
4303 if (!sctx) {
4304 mutex_unlock(&fs_info->scrub_lock);
4305 return -ENOTCONN;
4306 }
4307 atomic_inc(&sctx->cancel_req);
4308 while (dev->scrub_ctx) {
4309 mutex_unlock(&fs_info->scrub_lock);
4310 wait_event(fs_info->scrub_pause_wait,
4311 dev->scrub_ctx == NULL);
4312 mutex_lock(&fs_info->scrub_lock);
4313 }
4314 mutex_unlock(&fs_info->scrub_lock);
4315
4316 return 0;
4317}
4318
4319int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4320 struct btrfs_scrub_progress *progress)
4321{
4322 struct btrfs_device *dev;
4323 struct scrub_ctx *sctx = NULL;
4324
4325 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4326 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
4327 if (dev)
4328 sctx = dev->scrub_ctx;
4329 if (sctx)
4330 memcpy(progress, &sctx->stat, sizeof(*progress));
4331 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4332
4333 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4334}
4335
4336static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4337 u64 extent_logical, u64 extent_len,
4338 u64 *extent_physical,
4339 struct btrfs_device **extent_dev,
4340 int *extent_mirror_num)
4341{
4342 u64 mapped_length;
4343 struct btrfs_bio *bbio = NULL;
4344 int ret;
4345
4346 mapped_length = extent_len;
4347 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4348 &mapped_length, &bbio, 0);
4349 if (ret || !bbio || mapped_length < extent_len ||
4350 !bbio->stripes[0].dev->bdev) {
4351 btrfs_put_bbio(bbio);
4352 return;
4353 }
4354
4355 *extent_physical = bbio->stripes[0].physical;
4356 *extent_mirror_num = bbio->mirror_num;
4357 *extent_dev = bbio->stripes[0].dev;
4358 btrfs_put_bbio(bbio);
4359}
4360
4361static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
4362 int mirror_num, u64 physical_for_dev_replace)
4363{
4364 struct scrub_copy_nocow_ctx *nocow_ctx;
4365 struct btrfs_fs_info *fs_info = sctx->fs_info;
4366
4367 nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
4368 if (!nocow_ctx) {
4369 spin_lock(&sctx->stat_lock);
4370 sctx->stat.malloc_errors++;
4371 spin_unlock(&sctx->stat_lock);
4372 return -ENOMEM;
4373 }
4374
4375 scrub_pending_trans_workers_inc(sctx);
4376
4377 nocow_ctx->sctx = sctx;
4378 nocow_ctx->logical = logical;
4379 nocow_ctx->len = len;
4380 nocow_ctx->mirror_num = mirror_num;
4381 nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
4382 btrfs_init_work(&nocow_ctx->work, btrfs_scrubnc_helper,
4383 copy_nocow_pages_worker, NULL, NULL);
4384 INIT_LIST_HEAD(&nocow_ctx->inodes);
4385 btrfs_queue_work(fs_info->scrub_nocow_workers,
4386 &nocow_ctx->work);
4387
4388 return 0;
4389}
4390
4391static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
4392{
4393 struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
4394 struct scrub_nocow_inode *nocow_inode;
4395
4396 nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
4397 if (!nocow_inode)
4398 return -ENOMEM;
4399 nocow_inode->inum = inum;
4400 nocow_inode->offset = offset;
4401 nocow_inode->root = root;
4402 list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
4403 return 0;
4404}
4405
4406#define COPY_COMPLETE 1
4407
4408static void copy_nocow_pages_worker(struct btrfs_work *work)
4409{
4410 struct scrub_copy_nocow_ctx *nocow_ctx =
4411 container_of(work, struct scrub_copy_nocow_ctx, work);
4412 struct scrub_ctx *sctx = nocow_ctx->sctx;
4413 struct btrfs_fs_info *fs_info = sctx->fs_info;
4414 struct btrfs_root *root = fs_info->extent_root;
4415 u64 logical = nocow_ctx->logical;
4416 u64 len = nocow_ctx->len;
4417 int mirror_num = nocow_ctx->mirror_num;
4418 u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4419 int ret;
4420 struct btrfs_trans_handle *trans = NULL;
4421 struct btrfs_path *path;
4422 int not_written = 0;
4423
4424 path = btrfs_alloc_path();
4425 if (!path) {
4426 spin_lock(&sctx->stat_lock);
4427 sctx->stat.malloc_errors++;
4428 spin_unlock(&sctx->stat_lock);
4429 not_written = 1;
4430 goto out;
4431 }
4432
4433 trans = btrfs_join_transaction(root);
4434 if (IS_ERR(trans)) {
4435 not_written = 1;
4436 goto out;
4437 }
4438
4439 ret = iterate_inodes_from_logical(logical, fs_info, path,
4440 record_inode_for_nocow, nocow_ctx, false);
4441 if (ret != 0 && ret != -ENOENT) {
4442 btrfs_warn(fs_info,
4443 "iterate_inodes_from_logical() failed: log %llu, phys %llu, len %llu, mir %u, ret %d",
4444 logical, physical_for_dev_replace, len, mirror_num,
4445 ret);
4446 not_written = 1;
4447 goto out;
4448 }
4449
4450 btrfs_end_transaction(trans);
4451 trans = NULL;
4452 while (!list_empty(&nocow_ctx->inodes)) {
4453 struct scrub_nocow_inode *entry;
4454 entry = list_first_entry(&nocow_ctx->inodes,
4455 struct scrub_nocow_inode,
4456 list);
4457 list_del_init(&entry->list);
4458 ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
4459 entry->root, nocow_ctx);
4460 kfree(entry);
4461 if (ret == COPY_COMPLETE) {
4462 ret = 0;
4463 break;
4464 } else if (ret) {
4465 break;
4466 }
4467 }
4468out:
4469 while (!list_empty(&nocow_ctx->inodes)) {
4470 struct scrub_nocow_inode *entry;
4471 entry = list_first_entry(&nocow_ctx->inodes,
4472 struct scrub_nocow_inode,
4473 list);
4474 list_del_init(&entry->list);
4475 kfree(entry);
4476 }
4477 if (trans && !IS_ERR(trans))
4478 btrfs_end_transaction(trans);
4479 if (not_written)
4480 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
4481 num_uncorrectable_read_errors);
4482
4483 btrfs_free_path(path);
4484 kfree(nocow_ctx);
4485
4486 scrub_pending_trans_workers_dec(sctx);
4487}
4488
4489static int check_extent_to_block(struct btrfs_inode *inode, u64 start, u64 len,
4490 u64 logical)
4491{
4492 struct extent_state *cached_state = NULL;
4493 struct btrfs_ordered_extent *ordered;
4494 struct extent_io_tree *io_tree;
4495 struct extent_map *em;
4496 u64 lockstart = start, lockend = start + len - 1;
4497 int ret = 0;
4498
4499 io_tree = &inode->io_tree;
4500
4501 lock_extent_bits(io_tree, lockstart, lockend, &cached_state);
4502 ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
4503 if (ordered) {
4504 btrfs_put_ordered_extent(ordered);
4505 ret = 1;
4506 goto out_unlock;
4507 }
4508
4509 em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
4510 if (IS_ERR(em)) {
4511 ret = PTR_ERR(em);
4512 goto out_unlock;
4513 }
4514
4515 /*
4516 * This extent does not actually cover the logical extent anymore,
4517 * move on to the next inode.
4518 */
4519 if (em->block_start > logical ||
4520 em->block_start + em->block_len < logical + len ||
4521 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4522 free_extent_map(em);
4523 ret = 1;
4524 goto out_unlock;
4525 }
4526 free_extent_map(em);
4527
4528out_unlock:
4529 unlock_extent_cached(io_tree, lockstart, lockend, &cached_state);
4530 return ret;
4531}
4532
4533static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
4534 struct scrub_copy_nocow_ctx *nocow_ctx)
4535{
4536 struct btrfs_fs_info *fs_info = nocow_ctx->sctx->fs_info;
4537 struct btrfs_key key;
4538 struct inode *inode;
4539 struct page *page;
4540 struct btrfs_root *local_root;
4541 struct extent_io_tree *io_tree;
4542 u64 physical_for_dev_replace;
4543 u64 nocow_ctx_logical;
4544 u64 len = nocow_ctx->len;
4545 unsigned long index;
4546 int srcu_index;
4547 int ret = 0;
4548 int err = 0;
4549
4550 key.objectid = root;
4551 key.type = BTRFS_ROOT_ITEM_KEY;
4552 key.offset = (u64)-1;
4553
4554 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
4555
4556 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
4557 if (IS_ERR(local_root)) {
4558 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4559 return PTR_ERR(local_root);
4560 }
4561
4562 key.type = BTRFS_INODE_ITEM_KEY;
4563 key.objectid = inum;
4564 key.offset = 0;
4565 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
4566 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4567 if (IS_ERR(inode))
4568 return PTR_ERR(inode);
4569
4570 /* Avoid truncate/dio/punch hole.. */
4571 inode_lock(inode);
4572 inode_dio_wait(inode);
4573
4574 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4575 io_tree = &BTRFS_I(inode)->io_tree;
4576 nocow_ctx_logical = nocow_ctx->logical;
4577
4578 ret = check_extent_to_block(BTRFS_I(inode), offset, len,
4579 nocow_ctx_logical);
4580 if (ret) {
4581 ret = ret > 0 ? 0 : ret;
4582 goto out;
4583 }
4584
4585 while (len >= PAGE_SIZE) {
4586 index = offset >> PAGE_SHIFT;
4587again:
4588 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
4589 if (!page) {
4590 btrfs_err(fs_info, "find_or_create_page() failed");
4591 ret = -ENOMEM;
4592 goto out;
4593 }
4594
4595 if (PageUptodate(page)) {
4596 if (PageDirty(page))
4597 goto next_page;
4598 } else {
4599 ClearPageError(page);
4600 err = extent_read_full_page(io_tree, page,
4601 btrfs_get_extent,
4602 nocow_ctx->mirror_num);
4603 if (err) {
4604 ret = err;
4605 goto next_page;
4606 }
4607
4608 lock_page(page);
4609 /*
4610 * If the page has been remove from the page cache,
4611 * the data on it is meaningless, because it may be
4612 * old one, the new data may be written into the new
4613 * page in the page cache.
4614 */
4615 if (page->mapping != inode->i_mapping) {
4616 unlock_page(page);
4617 put_page(page);
4618 goto again;
4619 }
4620 if (!PageUptodate(page)) {
4621 ret = -EIO;
4622 goto next_page;
4623 }
4624 }
4625
4626 ret = check_extent_to_block(BTRFS_I(inode), offset, len,
4627 nocow_ctx_logical);
4628 if (ret) {
4629 ret = ret > 0 ? 0 : ret;
4630 goto next_page;
4631 }
4632
4633 err = write_page_nocow(nocow_ctx->sctx,
4634 physical_for_dev_replace, page);
4635 if (err)
4636 ret = err;
4637next_page:
4638 unlock_page(page);
4639 put_page(page);
4640
4641 if (ret)
4642 break;
4643
4644 offset += PAGE_SIZE;
4645 physical_for_dev_replace += PAGE_SIZE;
4646 nocow_ctx_logical += PAGE_SIZE;
4647 len -= PAGE_SIZE;
4648 }
4649 ret = COPY_COMPLETE;
4650out:
4651 inode_unlock(inode);
4652 iput(inode);
4653 return ret;
4654}
4655
4656static int write_page_nocow(struct scrub_ctx *sctx,
4657 u64 physical_for_dev_replace, struct page *page)
4658{
4659 struct bio *bio;
4660 struct btrfs_device *dev;
4661
4662 dev = sctx->wr_tgtdev;
4663 if (!dev)
4664 return -EIO;
4665 if (!dev->bdev) {
4666 btrfs_warn_rl(dev->fs_info,
4667 "scrub write_page_nocow(bdev == NULL) is unexpected");
4668 return -EIO;
4669 }
4670 bio = btrfs_io_bio_alloc(1);
4671 bio->bi_iter.bi_size = 0;
4672 bio->bi_iter.bi_sector = physical_for_dev_replace >> 9;
4673 bio_set_dev(bio, dev->bdev);
4674 bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
4675 /* bio_add_page won't fail on a freshly allocated bio */
4676 bio_add_page(bio, page, PAGE_SIZE, 0);
4677
4678 if (btrfsic_submit_bio_wait(bio)) {
4679 bio_put(bio);
4680 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
4681 return -EIO;
4682 }
4683
4684 bio_put(bio);
4685 return 0;
4686}