Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2011, 2012 STRATO. All rights reserved.
4 */
5
6#include <linux/blkdev.h>
7#include <linux/ratelimit.h>
8#include <linux/sched/mm.h>
9#include <crypto/hash.h>
10#include "ctree.h"
11#include "discard.h"
12#include "volumes.h"
13#include "disk-io.h"
14#include "ordered-data.h"
15#include "transaction.h"
16#include "backref.h"
17#include "extent_io.h"
18#include "dev-replace.h"
19#include "raid56.h"
20#include "block-group.h"
21#include "zoned.h"
22#include "fs.h"
23#include "accessors.h"
24#include "file-item.h"
25#include "scrub.h"
26#include "raid-stripe-tree.h"
27
28/*
29 * This is only the first step towards a full-features scrub. It reads all
30 * extent and super block and verifies the checksums. In case a bad checksum
31 * is found or the extent cannot be read, good data will be written back if
32 * any can be found.
33 *
34 * Future enhancements:
35 * - In case an unrepairable extent is encountered, track which files are
36 * affected and report them
37 * - track and record media errors, throw out bad devices
38 * - add a mode to also read unallocated space
39 */
40
41struct scrub_ctx;
42
43/*
44 * The following value only influences the performance.
45 *
46 * This determines how many stripes would be submitted in one go,
47 * which is 512KiB (BTRFS_STRIPE_LEN * SCRUB_STRIPES_PER_GROUP).
48 */
49#define SCRUB_STRIPES_PER_GROUP 8
50
51/*
52 * How many groups we have for each sctx.
53 *
54 * This would be 8M per device, the same value as the old scrub in-flight bios
55 * size limit.
56 */
57#define SCRUB_GROUPS_PER_SCTX 16
58
59#define SCRUB_TOTAL_STRIPES (SCRUB_GROUPS_PER_SCTX * SCRUB_STRIPES_PER_GROUP)
60
61/*
62 * The following value times PAGE_SIZE needs to be large enough to match the
63 * largest node/leaf/sector size that shall be supported.
64 */
65#define SCRUB_MAX_SECTORS_PER_BLOCK (BTRFS_MAX_METADATA_BLOCKSIZE / SZ_4K)
66
67/* Represent one sector and its needed info to verify the content. */
68struct scrub_sector_verification {
69 bool is_metadata;
70
71 union {
72 /*
73 * Csum pointer for data csum verification. Should point to a
74 * sector csum inside scrub_stripe::csums.
75 *
76 * NULL if this data sector has no csum.
77 */
78 u8 *csum;
79
80 /*
81 * Extra info for metadata verification. All sectors inside a
82 * tree block share the same generation.
83 */
84 u64 generation;
85 };
86};
87
88enum scrub_stripe_flags {
89 /* Set when @mirror_num, @dev, @physical and @logical are set. */
90 SCRUB_STRIPE_FLAG_INITIALIZED,
91
92 /* Set when the read-repair is finished. */
93 SCRUB_STRIPE_FLAG_REPAIR_DONE,
94
95 /*
96 * Set for data stripes if it's triggered from P/Q stripe.
97 * During such scrub, we should not report errors in data stripes, nor
98 * update the accounting.
99 */
100 SCRUB_STRIPE_FLAG_NO_REPORT,
101};
102
103#define SCRUB_STRIPE_PAGES (BTRFS_STRIPE_LEN / PAGE_SIZE)
104
105/*
106 * Represent one contiguous range with a length of BTRFS_STRIPE_LEN.
107 */
108struct scrub_stripe {
109 struct scrub_ctx *sctx;
110 struct btrfs_block_group *bg;
111
112 struct page *pages[SCRUB_STRIPE_PAGES];
113 struct scrub_sector_verification *sectors;
114
115 struct btrfs_device *dev;
116 u64 logical;
117 u64 physical;
118
119 u16 mirror_num;
120
121 /* Should be BTRFS_STRIPE_LEN / sectorsize. */
122 u16 nr_sectors;
123
124 /*
125 * How many data/meta extents are in this stripe. Only for scrub status
126 * reporting purposes.
127 */
128 u16 nr_data_extents;
129 u16 nr_meta_extents;
130
131 atomic_t pending_io;
132 wait_queue_head_t io_wait;
133 wait_queue_head_t repair_wait;
134
135 /*
136 * Indicate the states of the stripe. Bits are defined in
137 * scrub_stripe_flags enum.
138 */
139 unsigned long state;
140
141 /* Indicate which sectors are covered by extent items. */
142 unsigned long extent_sector_bitmap;
143
144 /*
145 * The errors hit during the initial read of the stripe.
146 *
147 * Would be utilized for error reporting and repair.
148 *
149 * The remaining init_nr_* records the number of errors hit, only used
150 * by error reporting.
151 */
152 unsigned long init_error_bitmap;
153 unsigned int init_nr_io_errors;
154 unsigned int init_nr_csum_errors;
155 unsigned int init_nr_meta_errors;
156
157 /*
158 * The following error bitmaps are all for the current status.
159 * Every time we submit a new read, these bitmaps may be updated.
160 *
161 * error_bitmap = io_error_bitmap | csum_error_bitmap | meta_error_bitmap;
162 *
163 * IO and csum errors can happen for both metadata and data.
164 */
165 unsigned long error_bitmap;
166 unsigned long io_error_bitmap;
167 unsigned long csum_error_bitmap;
168 unsigned long meta_error_bitmap;
169
170 /* For writeback (repair or replace) error reporting. */
171 unsigned long write_error_bitmap;
172
173 /* Writeback can be concurrent, thus we need to protect the bitmap. */
174 spinlock_t write_error_lock;
175
176 /*
177 * Checksum for the whole stripe if this stripe is inside a data block
178 * group.
179 */
180 u8 *csums;
181
182 struct work_struct work;
183};
184
185struct scrub_ctx {
186 struct scrub_stripe stripes[SCRUB_TOTAL_STRIPES];
187 struct scrub_stripe *raid56_data_stripes;
188 struct btrfs_fs_info *fs_info;
189 struct btrfs_path extent_path;
190 struct btrfs_path csum_path;
191 int first_free;
192 int cur_stripe;
193 atomic_t cancel_req;
194 int readonly;
195
196 /* State of IO submission throttling affecting the associated device */
197 ktime_t throttle_deadline;
198 u64 throttle_sent;
199
200 int is_dev_replace;
201 u64 write_pointer;
202
203 struct mutex wr_lock;
204 struct btrfs_device *wr_tgtdev;
205
206 /*
207 * statistics
208 */
209 struct btrfs_scrub_progress stat;
210 spinlock_t stat_lock;
211
212 /*
213 * Use a ref counter to avoid use-after-free issues. Scrub workers
214 * decrement bios_in_flight and workers_pending and then do a wakeup
215 * on the list_wait wait queue. We must ensure the main scrub task
216 * doesn't free the scrub context before or while the workers are
217 * doing the wakeup() call.
218 */
219 refcount_t refs;
220};
221
222struct scrub_warning {
223 struct btrfs_path *path;
224 u64 extent_item_size;
225 const char *errstr;
226 u64 physical;
227 u64 logical;
228 struct btrfs_device *dev;
229};
230
231static void release_scrub_stripe(struct scrub_stripe *stripe)
232{
233 if (!stripe)
234 return;
235
236 for (int i = 0; i < SCRUB_STRIPE_PAGES; i++) {
237 if (stripe->pages[i])
238 __free_page(stripe->pages[i]);
239 stripe->pages[i] = NULL;
240 }
241 kfree(stripe->sectors);
242 kfree(stripe->csums);
243 stripe->sectors = NULL;
244 stripe->csums = NULL;
245 stripe->sctx = NULL;
246 stripe->state = 0;
247}
248
249static int init_scrub_stripe(struct btrfs_fs_info *fs_info,
250 struct scrub_stripe *stripe)
251{
252 int ret;
253
254 memset(stripe, 0, sizeof(*stripe));
255
256 stripe->nr_sectors = BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits;
257 stripe->state = 0;
258
259 init_waitqueue_head(&stripe->io_wait);
260 init_waitqueue_head(&stripe->repair_wait);
261 atomic_set(&stripe->pending_io, 0);
262 spin_lock_init(&stripe->write_error_lock);
263
264 ret = btrfs_alloc_page_array(SCRUB_STRIPE_PAGES, stripe->pages, false);
265 if (ret < 0)
266 goto error;
267
268 stripe->sectors = kcalloc(stripe->nr_sectors,
269 sizeof(struct scrub_sector_verification),
270 GFP_KERNEL);
271 if (!stripe->sectors)
272 goto error;
273
274 stripe->csums = kcalloc(BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits,
275 fs_info->csum_size, GFP_KERNEL);
276 if (!stripe->csums)
277 goto error;
278 return 0;
279error:
280 release_scrub_stripe(stripe);
281 return -ENOMEM;
282}
283
284static void wait_scrub_stripe_io(struct scrub_stripe *stripe)
285{
286 wait_event(stripe->io_wait, atomic_read(&stripe->pending_io) == 0);
287}
288
289static void scrub_put_ctx(struct scrub_ctx *sctx);
290
291static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
292{
293 while (atomic_read(&fs_info->scrub_pause_req)) {
294 mutex_unlock(&fs_info->scrub_lock);
295 wait_event(fs_info->scrub_pause_wait,
296 atomic_read(&fs_info->scrub_pause_req) == 0);
297 mutex_lock(&fs_info->scrub_lock);
298 }
299}
300
301static void scrub_pause_on(struct btrfs_fs_info *fs_info)
302{
303 atomic_inc(&fs_info->scrubs_paused);
304 wake_up(&fs_info->scrub_pause_wait);
305}
306
307static void scrub_pause_off(struct btrfs_fs_info *fs_info)
308{
309 mutex_lock(&fs_info->scrub_lock);
310 __scrub_blocked_if_needed(fs_info);
311 atomic_dec(&fs_info->scrubs_paused);
312 mutex_unlock(&fs_info->scrub_lock);
313
314 wake_up(&fs_info->scrub_pause_wait);
315}
316
317static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
318{
319 scrub_pause_on(fs_info);
320 scrub_pause_off(fs_info);
321}
322
323static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
324{
325 int i;
326
327 if (!sctx)
328 return;
329
330 for (i = 0; i < SCRUB_TOTAL_STRIPES; i++)
331 release_scrub_stripe(&sctx->stripes[i]);
332
333 kvfree(sctx);
334}
335
336static void scrub_put_ctx(struct scrub_ctx *sctx)
337{
338 if (refcount_dec_and_test(&sctx->refs))
339 scrub_free_ctx(sctx);
340}
341
342static noinline_for_stack struct scrub_ctx *scrub_setup_ctx(
343 struct btrfs_fs_info *fs_info, int is_dev_replace)
344{
345 struct scrub_ctx *sctx;
346 int i;
347
348 /* Since sctx has inline 128 stripes, it can go beyond 64K easily. Use
349 * kvzalloc().
350 */
351 sctx = kvzalloc(sizeof(*sctx), GFP_KERNEL);
352 if (!sctx)
353 goto nomem;
354 refcount_set(&sctx->refs, 1);
355 sctx->is_dev_replace = is_dev_replace;
356 sctx->fs_info = fs_info;
357 sctx->extent_path.search_commit_root = 1;
358 sctx->extent_path.skip_locking = 1;
359 sctx->csum_path.search_commit_root = 1;
360 sctx->csum_path.skip_locking = 1;
361 for (i = 0; i < SCRUB_TOTAL_STRIPES; i++) {
362 int ret;
363
364 ret = init_scrub_stripe(fs_info, &sctx->stripes[i]);
365 if (ret < 0)
366 goto nomem;
367 sctx->stripes[i].sctx = sctx;
368 }
369 sctx->first_free = 0;
370 atomic_set(&sctx->cancel_req, 0);
371
372 spin_lock_init(&sctx->stat_lock);
373 sctx->throttle_deadline = 0;
374
375 mutex_init(&sctx->wr_lock);
376 if (is_dev_replace) {
377 WARN_ON(!fs_info->dev_replace.tgtdev);
378 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
379 }
380
381 return sctx;
382
383nomem:
384 scrub_free_ctx(sctx);
385 return ERR_PTR(-ENOMEM);
386}
387
388static int scrub_print_warning_inode(u64 inum, u64 offset, u64 num_bytes,
389 u64 root, void *warn_ctx)
390{
391 u32 nlink;
392 int ret;
393 int i;
394 unsigned nofs_flag;
395 struct extent_buffer *eb;
396 struct btrfs_inode_item *inode_item;
397 struct scrub_warning *swarn = warn_ctx;
398 struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
399 struct inode_fs_paths *ipath = NULL;
400 struct btrfs_root *local_root;
401 struct btrfs_key key;
402
403 local_root = btrfs_get_fs_root(fs_info, root, true);
404 if (IS_ERR(local_root)) {
405 ret = PTR_ERR(local_root);
406 goto err;
407 }
408
409 /*
410 * this makes the path point to (inum INODE_ITEM ioff)
411 */
412 key.objectid = inum;
413 key.type = BTRFS_INODE_ITEM_KEY;
414 key.offset = 0;
415
416 ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
417 if (ret) {
418 btrfs_put_root(local_root);
419 btrfs_release_path(swarn->path);
420 goto err;
421 }
422
423 eb = swarn->path->nodes[0];
424 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
425 struct btrfs_inode_item);
426 nlink = btrfs_inode_nlink(eb, inode_item);
427 btrfs_release_path(swarn->path);
428
429 /*
430 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
431 * uses GFP_NOFS in this context, so we keep it consistent but it does
432 * not seem to be strictly necessary.
433 */
434 nofs_flag = memalloc_nofs_save();
435 ipath = init_ipath(4096, local_root, swarn->path);
436 memalloc_nofs_restore(nofs_flag);
437 if (IS_ERR(ipath)) {
438 btrfs_put_root(local_root);
439 ret = PTR_ERR(ipath);
440 ipath = NULL;
441 goto err;
442 }
443 ret = paths_from_inode(inum, ipath);
444
445 if (ret < 0)
446 goto err;
447
448 /*
449 * we deliberately ignore the bit ipath might have been too small to
450 * hold all of the paths here
451 */
452 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
453 btrfs_warn_in_rcu(fs_info,
454"%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %u, links %u (path: %s)",
455 swarn->errstr, swarn->logical,
456 btrfs_dev_name(swarn->dev),
457 swarn->physical,
458 root, inum, offset,
459 fs_info->sectorsize, nlink,
460 (char *)(unsigned long)ipath->fspath->val[i]);
461
462 btrfs_put_root(local_root);
463 free_ipath(ipath);
464 return 0;
465
466err:
467 btrfs_warn_in_rcu(fs_info,
468 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
469 swarn->errstr, swarn->logical,
470 btrfs_dev_name(swarn->dev),
471 swarn->physical,
472 root, inum, offset, ret);
473
474 free_ipath(ipath);
475 return 0;
476}
477
478static void scrub_print_common_warning(const char *errstr, struct btrfs_device *dev,
479 bool is_super, u64 logical, u64 physical)
480{
481 struct btrfs_fs_info *fs_info = dev->fs_info;
482 struct btrfs_path *path;
483 struct btrfs_key found_key;
484 struct extent_buffer *eb;
485 struct btrfs_extent_item *ei;
486 struct scrub_warning swarn;
487 u64 flags = 0;
488 u32 item_size;
489 int ret;
490
491 /* Super block error, no need to search extent tree. */
492 if (is_super) {
493 btrfs_warn_in_rcu(fs_info, "%s on device %s, physical %llu",
494 errstr, btrfs_dev_name(dev), physical);
495 return;
496 }
497 path = btrfs_alloc_path();
498 if (!path)
499 return;
500
501 swarn.physical = physical;
502 swarn.logical = logical;
503 swarn.errstr = errstr;
504 swarn.dev = NULL;
505
506 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
507 &flags);
508 if (ret < 0)
509 goto out;
510
511 swarn.extent_item_size = found_key.offset;
512
513 eb = path->nodes[0];
514 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
515 item_size = btrfs_item_size(eb, path->slots[0]);
516
517 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
518 unsigned long ptr = 0;
519 u8 ref_level;
520 u64 ref_root;
521
522 while (true) {
523 ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
524 item_size, &ref_root,
525 &ref_level);
526 if (ret < 0) {
527 btrfs_warn(fs_info,
528 "failed to resolve tree backref for logical %llu: %d",
529 swarn.logical, ret);
530 break;
531 }
532 if (ret > 0)
533 break;
534 btrfs_warn_in_rcu(fs_info,
535"%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
536 errstr, swarn.logical, btrfs_dev_name(dev),
537 swarn.physical, (ref_level ? "node" : "leaf"),
538 ref_level, ref_root);
539 }
540 btrfs_release_path(path);
541 } else {
542 struct btrfs_backref_walk_ctx ctx = { 0 };
543
544 btrfs_release_path(path);
545
546 ctx.bytenr = found_key.objectid;
547 ctx.extent_item_pos = swarn.logical - found_key.objectid;
548 ctx.fs_info = fs_info;
549
550 swarn.path = path;
551 swarn.dev = dev;
552
553 iterate_extent_inodes(&ctx, true, scrub_print_warning_inode, &swarn);
554 }
555
556out:
557 btrfs_free_path(path);
558}
559
560static int fill_writer_pointer_gap(struct scrub_ctx *sctx, u64 physical)
561{
562 int ret = 0;
563 u64 length;
564
565 if (!btrfs_is_zoned(sctx->fs_info))
566 return 0;
567
568 if (!btrfs_dev_is_sequential(sctx->wr_tgtdev, physical))
569 return 0;
570
571 if (sctx->write_pointer < physical) {
572 length = physical - sctx->write_pointer;
573
574 ret = btrfs_zoned_issue_zeroout(sctx->wr_tgtdev,
575 sctx->write_pointer, length);
576 if (!ret)
577 sctx->write_pointer = physical;
578 }
579 return ret;
580}
581
582static struct page *scrub_stripe_get_page(struct scrub_stripe *stripe, int sector_nr)
583{
584 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
585 int page_index = (sector_nr << fs_info->sectorsize_bits) >> PAGE_SHIFT;
586
587 return stripe->pages[page_index];
588}
589
590static unsigned int scrub_stripe_get_page_offset(struct scrub_stripe *stripe,
591 int sector_nr)
592{
593 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
594
595 return offset_in_page(sector_nr << fs_info->sectorsize_bits);
596}
597
598static void scrub_verify_one_metadata(struct scrub_stripe *stripe, int sector_nr)
599{
600 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
601 const u32 sectors_per_tree = fs_info->nodesize >> fs_info->sectorsize_bits;
602 const u64 logical = stripe->logical + (sector_nr << fs_info->sectorsize_bits);
603 const struct page *first_page = scrub_stripe_get_page(stripe, sector_nr);
604 const unsigned int first_off = scrub_stripe_get_page_offset(stripe, sector_nr);
605 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
606 u8 on_disk_csum[BTRFS_CSUM_SIZE];
607 u8 calculated_csum[BTRFS_CSUM_SIZE];
608 struct btrfs_header *header;
609
610 /*
611 * Here we don't have a good way to attach the pages (and subpages)
612 * to a dummy extent buffer, thus we have to directly grab the members
613 * from pages.
614 */
615 header = (struct btrfs_header *)(page_address(first_page) + first_off);
616 memcpy(on_disk_csum, header->csum, fs_info->csum_size);
617
618 if (logical != btrfs_stack_header_bytenr(header)) {
619 bitmap_set(&stripe->csum_error_bitmap, sector_nr, sectors_per_tree);
620 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
621 btrfs_warn_rl(fs_info,
622 "tree block %llu mirror %u has bad bytenr, has %llu want %llu",
623 logical, stripe->mirror_num,
624 btrfs_stack_header_bytenr(header), logical);
625 return;
626 }
627 if (memcmp(header->fsid, fs_info->fs_devices->metadata_uuid,
628 BTRFS_FSID_SIZE) != 0) {
629 bitmap_set(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
630 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
631 btrfs_warn_rl(fs_info,
632 "tree block %llu mirror %u has bad fsid, has %pU want %pU",
633 logical, stripe->mirror_num,
634 header->fsid, fs_info->fs_devices->fsid);
635 return;
636 }
637 if (memcmp(header->chunk_tree_uuid, fs_info->chunk_tree_uuid,
638 BTRFS_UUID_SIZE) != 0) {
639 bitmap_set(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
640 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
641 btrfs_warn_rl(fs_info,
642 "tree block %llu mirror %u has bad chunk tree uuid, has %pU want %pU",
643 logical, stripe->mirror_num,
644 header->chunk_tree_uuid, fs_info->chunk_tree_uuid);
645 return;
646 }
647
648 /* Now check tree block csum. */
649 shash->tfm = fs_info->csum_shash;
650 crypto_shash_init(shash);
651 crypto_shash_update(shash, page_address(first_page) + first_off +
652 BTRFS_CSUM_SIZE, fs_info->sectorsize - BTRFS_CSUM_SIZE);
653
654 for (int i = sector_nr + 1; i < sector_nr + sectors_per_tree; i++) {
655 struct page *page = scrub_stripe_get_page(stripe, i);
656 unsigned int page_off = scrub_stripe_get_page_offset(stripe, i);
657
658 crypto_shash_update(shash, page_address(page) + page_off,
659 fs_info->sectorsize);
660 }
661
662 crypto_shash_final(shash, calculated_csum);
663 if (memcmp(calculated_csum, on_disk_csum, fs_info->csum_size) != 0) {
664 bitmap_set(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
665 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
666 btrfs_warn_rl(fs_info,
667 "tree block %llu mirror %u has bad csum, has " CSUM_FMT " want " CSUM_FMT,
668 logical, stripe->mirror_num,
669 CSUM_FMT_VALUE(fs_info->csum_size, on_disk_csum),
670 CSUM_FMT_VALUE(fs_info->csum_size, calculated_csum));
671 return;
672 }
673 if (stripe->sectors[sector_nr].generation !=
674 btrfs_stack_header_generation(header)) {
675 bitmap_set(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
676 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
677 btrfs_warn_rl(fs_info,
678 "tree block %llu mirror %u has bad generation, has %llu want %llu",
679 logical, stripe->mirror_num,
680 btrfs_stack_header_generation(header),
681 stripe->sectors[sector_nr].generation);
682 return;
683 }
684 bitmap_clear(&stripe->error_bitmap, sector_nr, sectors_per_tree);
685 bitmap_clear(&stripe->csum_error_bitmap, sector_nr, sectors_per_tree);
686 bitmap_clear(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
687}
688
689static void scrub_verify_one_sector(struct scrub_stripe *stripe, int sector_nr)
690{
691 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
692 struct scrub_sector_verification *sector = &stripe->sectors[sector_nr];
693 const u32 sectors_per_tree = fs_info->nodesize >> fs_info->sectorsize_bits;
694 struct page *page = scrub_stripe_get_page(stripe, sector_nr);
695 unsigned int pgoff = scrub_stripe_get_page_offset(stripe, sector_nr);
696 u8 csum_buf[BTRFS_CSUM_SIZE];
697 int ret;
698
699 ASSERT(sector_nr >= 0 && sector_nr < stripe->nr_sectors);
700
701 /* Sector not utilized, skip it. */
702 if (!test_bit(sector_nr, &stripe->extent_sector_bitmap))
703 return;
704
705 /* IO error, no need to check. */
706 if (test_bit(sector_nr, &stripe->io_error_bitmap))
707 return;
708
709 /* Metadata, verify the full tree block. */
710 if (sector->is_metadata) {
711 /*
712 * Check if the tree block crosses the stripe boundary. If
713 * crossed the boundary, we cannot verify it but only give a
714 * warning.
715 *
716 * This can only happen on a very old filesystem where chunks
717 * are not ensured to be stripe aligned.
718 */
719 if (unlikely(sector_nr + sectors_per_tree > stripe->nr_sectors)) {
720 btrfs_warn_rl(fs_info,
721 "tree block at %llu crosses stripe boundary %llu",
722 stripe->logical +
723 (sector_nr << fs_info->sectorsize_bits),
724 stripe->logical);
725 return;
726 }
727 scrub_verify_one_metadata(stripe, sector_nr);
728 return;
729 }
730
731 /*
732 * Data is easier, we just verify the data csum (if we have it). For
733 * cases without csum, we have no other choice but to trust it.
734 */
735 if (!sector->csum) {
736 clear_bit(sector_nr, &stripe->error_bitmap);
737 return;
738 }
739
740 ret = btrfs_check_sector_csum(fs_info, page, pgoff, csum_buf, sector->csum);
741 if (ret < 0) {
742 set_bit(sector_nr, &stripe->csum_error_bitmap);
743 set_bit(sector_nr, &stripe->error_bitmap);
744 } else {
745 clear_bit(sector_nr, &stripe->csum_error_bitmap);
746 clear_bit(sector_nr, &stripe->error_bitmap);
747 }
748}
749
750/* Verify specified sectors of a stripe. */
751static void scrub_verify_one_stripe(struct scrub_stripe *stripe, unsigned long bitmap)
752{
753 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
754 const u32 sectors_per_tree = fs_info->nodesize >> fs_info->sectorsize_bits;
755 int sector_nr;
756
757 for_each_set_bit(sector_nr, &bitmap, stripe->nr_sectors) {
758 scrub_verify_one_sector(stripe, sector_nr);
759 if (stripe->sectors[sector_nr].is_metadata)
760 sector_nr += sectors_per_tree - 1;
761 }
762}
763
764static int calc_sector_number(struct scrub_stripe *stripe, struct bio_vec *first_bvec)
765{
766 int i;
767
768 for (i = 0; i < stripe->nr_sectors; i++) {
769 if (scrub_stripe_get_page(stripe, i) == first_bvec->bv_page &&
770 scrub_stripe_get_page_offset(stripe, i) == first_bvec->bv_offset)
771 break;
772 }
773 ASSERT(i < stripe->nr_sectors);
774 return i;
775}
776
777/*
778 * Repair read is different to the regular read:
779 *
780 * - Only reads the failed sectors
781 * - May have extra blocksize limits
782 */
783static void scrub_repair_read_endio(struct btrfs_bio *bbio)
784{
785 struct scrub_stripe *stripe = bbio->private;
786 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
787 struct bio_vec *bvec;
788 int sector_nr = calc_sector_number(stripe, bio_first_bvec_all(&bbio->bio));
789 u32 bio_size = 0;
790 int i;
791
792 ASSERT(sector_nr < stripe->nr_sectors);
793
794 bio_for_each_bvec_all(bvec, &bbio->bio, i)
795 bio_size += bvec->bv_len;
796
797 if (bbio->bio.bi_status) {
798 bitmap_set(&stripe->io_error_bitmap, sector_nr,
799 bio_size >> fs_info->sectorsize_bits);
800 bitmap_set(&stripe->error_bitmap, sector_nr,
801 bio_size >> fs_info->sectorsize_bits);
802 } else {
803 bitmap_clear(&stripe->io_error_bitmap, sector_nr,
804 bio_size >> fs_info->sectorsize_bits);
805 }
806 bio_put(&bbio->bio);
807 if (atomic_dec_and_test(&stripe->pending_io))
808 wake_up(&stripe->io_wait);
809}
810
811static int calc_next_mirror(int mirror, int num_copies)
812{
813 ASSERT(mirror <= num_copies);
814 return (mirror + 1 > num_copies) ? 1 : mirror + 1;
815}
816
817static void scrub_stripe_submit_repair_read(struct scrub_stripe *stripe,
818 int mirror, int blocksize, bool wait)
819{
820 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
821 struct btrfs_bio *bbio = NULL;
822 const unsigned long old_error_bitmap = stripe->error_bitmap;
823 int i;
824
825 ASSERT(stripe->mirror_num >= 1);
826 ASSERT(atomic_read(&stripe->pending_io) == 0);
827
828 for_each_set_bit(i, &old_error_bitmap, stripe->nr_sectors) {
829 struct page *page;
830 int pgoff;
831 int ret;
832
833 page = scrub_stripe_get_page(stripe, i);
834 pgoff = scrub_stripe_get_page_offset(stripe, i);
835
836 /* The current sector cannot be merged, submit the bio. */
837 if (bbio && ((i > 0 && !test_bit(i - 1, &stripe->error_bitmap)) ||
838 bbio->bio.bi_iter.bi_size >= blocksize)) {
839 ASSERT(bbio->bio.bi_iter.bi_size);
840 atomic_inc(&stripe->pending_io);
841 btrfs_submit_bbio(bbio, mirror);
842 if (wait)
843 wait_scrub_stripe_io(stripe);
844 bbio = NULL;
845 }
846
847 if (!bbio) {
848 bbio = btrfs_bio_alloc(stripe->nr_sectors, REQ_OP_READ,
849 fs_info, scrub_repair_read_endio, stripe);
850 bbio->bio.bi_iter.bi_sector = (stripe->logical +
851 (i << fs_info->sectorsize_bits)) >> SECTOR_SHIFT;
852 }
853
854 ret = bio_add_page(&bbio->bio, page, fs_info->sectorsize, pgoff);
855 ASSERT(ret == fs_info->sectorsize);
856 }
857 if (bbio) {
858 ASSERT(bbio->bio.bi_iter.bi_size);
859 atomic_inc(&stripe->pending_io);
860 btrfs_submit_bbio(bbio, mirror);
861 if (wait)
862 wait_scrub_stripe_io(stripe);
863 }
864}
865
866static void scrub_stripe_report_errors(struct scrub_ctx *sctx,
867 struct scrub_stripe *stripe)
868{
869 static DEFINE_RATELIMIT_STATE(rs, DEFAULT_RATELIMIT_INTERVAL,
870 DEFAULT_RATELIMIT_BURST);
871 struct btrfs_fs_info *fs_info = sctx->fs_info;
872 struct btrfs_device *dev = NULL;
873 u64 physical = 0;
874 int nr_data_sectors = 0;
875 int nr_meta_sectors = 0;
876 int nr_nodatacsum_sectors = 0;
877 int nr_repaired_sectors = 0;
878 int sector_nr;
879
880 if (test_bit(SCRUB_STRIPE_FLAG_NO_REPORT, &stripe->state))
881 return;
882
883 /*
884 * Init needed infos for error reporting.
885 *
886 * Although our scrub_stripe infrastructure is mostly based on btrfs_submit_bio()
887 * thus no need for dev/physical, error reporting still needs dev and physical.
888 */
889 if (!bitmap_empty(&stripe->init_error_bitmap, stripe->nr_sectors)) {
890 u64 mapped_len = fs_info->sectorsize;
891 struct btrfs_io_context *bioc = NULL;
892 int stripe_index = stripe->mirror_num - 1;
893 int ret;
894
895 /* For scrub, our mirror_num should always start at 1. */
896 ASSERT(stripe->mirror_num >= 1);
897 ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
898 stripe->logical, &mapped_len, &bioc,
899 NULL, NULL);
900 /*
901 * If we failed, dev will be NULL, and later detailed reports
902 * will just be skipped.
903 */
904 if (ret < 0)
905 goto skip;
906 physical = bioc->stripes[stripe_index].physical;
907 dev = bioc->stripes[stripe_index].dev;
908 btrfs_put_bioc(bioc);
909 }
910
911skip:
912 for_each_set_bit(sector_nr, &stripe->extent_sector_bitmap, stripe->nr_sectors) {
913 bool repaired = false;
914
915 if (stripe->sectors[sector_nr].is_metadata) {
916 nr_meta_sectors++;
917 } else {
918 nr_data_sectors++;
919 if (!stripe->sectors[sector_nr].csum)
920 nr_nodatacsum_sectors++;
921 }
922
923 if (test_bit(sector_nr, &stripe->init_error_bitmap) &&
924 !test_bit(sector_nr, &stripe->error_bitmap)) {
925 nr_repaired_sectors++;
926 repaired = true;
927 }
928
929 /* Good sector from the beginning, nothing need to be done. */
930 if (!test_bit(sector_nr, &stripe->init_error_bitmap))
931 continue;
932
933 /*
934 * Report error for the corrupted sectors. If repaired, just
935 * output the message of repaired message.
936 */
937 if (repaired) {
938 if (dev) {
939 btrfs_err_rl_in_rcu(fs_info,
940 "fixed up error at logical %llu on dev %s physical %llu",
941 stripe->logical, btrfs_dev_name(dev),
942 physical);
943 } else {
944 btrfs_err_rl_in_rcu(fs_info,
945 "fixed up error at logical %llu on mirror %u",
946 stripe->logical, stripe->mirror_num);
947 }
948 continue;
949 }
950
951 /* The remaining are all for unrepaired. */
952 if (dev) {
953 btrfs_err_rl_in_rcu(fs_info,
954 "unable to fixup (regular) error at logical %llu on dev %s physical %llu",
955 stripe->logical, btrfs_dev_name(dev),
956 physical);
957 } else {
958 btrfs_err_rl_in_rcu(fs_info,
959 "unable to fixup (regular) error at logical %llu on mirror %u",
960 stripe->logical, stripe->mirror_num);
961 }
962
963 if (test_bit(sector_nr, &stripe->io_error_bitmap))
964 if (__ratelimit(&rs) && dev)
965 scrub_print_common_warning("i/o error", dev, false,
966 stripe->logical, physical);
967 if (test_bit(sector_nr, &stripe->csum_error_bitmap))
968 if (__ratelimit(&rs) && dev)
969 scrub_print_common_warning("checksum error", dev, false,
970 stripe->logical, physical);
971 if (test_bit(sector_nr, &stripe->meta_error_bitmap))
972 if (__ratelimit(&rs) && dev)
973 scrub_print_common_warning("header error", dev, false,
974 stripe->logical, physical);
975 }
976
977 spin_lock(&sctx->stat_lock);
978 sctx->stat.data_extents_scrubbed += stripe->nr_data_extents;
979 sctx->stat.tree_extents_scrubbed += stripe->nr_meta_extents;
980 sctx->stat.data_bytes_scrubbed += nr_data_sectors << fs_info->sectorsize_bits;
981 sctx->stat.tree_bytes_scrubbed += nr_meta_sectors << fs_info->sectorsize_bits;
982 sctx->stat.no_csum += nr_nodatacsum_sectors;
983 sctx->stat.read_errors += stripe->init_nr_io_errors;
984 sctx->stat.csum_errors += stripe->init_nr_csum_errors;
985 sctx->stat.verify_errors += stripe->init_nr_meta_errors;
986 sctx->stat.uncorrectable_errors +=
987 bitmap_weight(&stripe->error_bitmap, stripe->nr_sectors);
988 sctx->stat.corrected_errors += nr_repaired_sectors;
989 spin_unlock(&sctx->stat_lock);
990}
991
992static void scrub_write_sectors(struct scrub_ctx *sctx, struct scrub_stripe *stripe,
993 unsigned long write_bitmap, bool dev_replace);
994
995/*
996 * The main entrance for all read related scrub work, including:
997 *
998 * - Wait for the initial read to finish
999 * - Verify and locate any bad sectors
1000 * - Go through the remaining mirrors and try to read as large blocksize as
1001 * possible
1002 * - Go through all mirrors (including the failed mirror) sector-by-sector
1003 * - Submit writeback for repaired sectors
1004 *
1005 * Writeback for dev-replace does not happen here, it needs extra
1006 * synchronization for zoned devices.
1007 */
1008static void scrub_stripe_read_repair_worker(struct work_struct *work)
1009{
1010 struct scrub_stripe *stripe = container_of(work, struct scrub_stripe, work);
1011 struct scrub_ctx *sctx = stripe->sctx;
1012 struct btrfs_fs_info *fs_info = sctx->fs_info;
1013 int num_copies = btrfs_num_copies(fs_info, stripe->bg->start,
1014 stripe->bg->length);
1015 unsigned long repaired;
1016 int mirror;
1017 int i;
1018
1019 ASSERT(stripe->mirror_num > 0);
1020
1021 wait_scrub_stripe_io(stripe);
1022 scrub_verify_one_stripe(stripe, stripe->extent_sector_bitmap);
1023 /* Save the initial failed bitmap for later repair and report usage. */
1024 stripe->init_error_bitmap = stripe->error_bitmap;
1025 stripe->init_nr_io_errors = bitmap_weight(&stripe->io_error_bitmap,
1026 stripe->nr_sectors);
1027 stripe->init_nr_csum_errors = bitmap_weight(&stripe->csum_error_bitmap,
1028 stripe->nr_sectors);
1029 stripe->init_nr_meta_errors = bitmap_weight(&stripe->meta_error_bitmap,
1030 stripe->nr_sectors);
1031
1032 if (bitmap_empty(&stripe->init_error_bitmap, stripe->nr_sectors))
1033 goto out;
1034
1035 /*
1036 * Try all remaining mirrors.
1037 *
1038 * Here we still try to read as large block as possible, as this is
1039 * faster and we have extra safety nets to rely on.
1040 */
1041 for (mirror = calc_next_mirror(stripe->mirror_num, num_copies);
1042 mirror != stripe->mirror_num;
1043 mirror = calc_next_mirror(mirror, num_copies)) {
1044 const unsigned long old_error_bitmap = stripe->error_bitmap;
1045
1046 scrub_stripe_submit_repair_read(stripe, mirror,
1047 BTRFS_STRIPE_LEN, false);
1048 wait_scrub_stripe_io(stripe);
1049 scrub_verify_one_stripe(stripe, old_error_bitmap);
1050 if (bitmap_empty(&stripe->error_bitmap, stripe->nr_sectors))
1051 goto out;
1052 }
1053
1054 /*
1055 * Last safety net, try re-checking all mirrors, including the failed
1056 * one, sector-by-sector.
1057 *
1058 * As if one sector failed the drive's internal csum, the whole read
1059 * containing the offending sector would be marked as error.
1060 * Thus here we do sector-by-sector read.
1061 *
1062 * This can be slow, thus we only try it as the last resort.
1063 */
1064
1065 for (i = 0, mirror = stripe->mirror_num;
1066 i < num_copies;
1067 i++, mirror = calc_next_mirror(mirror, num_copies)) {
1068 const unsigned long old_error_bitmap = stripe->error_bitmap;
1069
1070 scrub_stripe_submit_repair_read(stripe, mirror,
1071 fs_info->sectorsize, true);
1072 wait_scrub_stripe_io(stripe);
1073 scrub_verify_one_stripe(stripe, old_error_bitmap);
1074 if (bitmap_empty(&stripe->error_bitmap, stripe->nr_sectors))
1075 goto out;
1076 }
1077out:
1078 /*
1079 * Submit the repaired sectors. For zoned case, we cannot do repair
1080 * in-place, but queue the bg to be relocated.
1081 */
1082 bitmap_andnot(&repaired, &stripe->init_error_bitmap, &stripe->error_bitmap,
1083 stripe->nr_sectors);
1084 if (!sctx->readonly && !bitmap_empty(&repaired, stripe->nr_sectors)) {
1085 if (btrfs_is_zoned(fs_info)) {
1086 btrfs_repair_one_zone(fs_info, sctx->stripes[0].bg->start);
1087 } else {
1088 scrub_write_sectors(sctx, stripe, repaired, false);
1089 wait_scrub_stripe_io(stripe);
1090 }
1091 }
1092
1093 scrub_stripe_report_errors(sctx, stripe);
1094 set_bit(SCRUB_STRIPE_FLAG_REPAIR_DONE, &stripe->state);
1095 wake_up(&stripe->repair_wait);
1096}
1097
1098static void scrub_read_endio(struct btrfs_bio *bbio)
1099{
1100 struct scrub_stripe *stripe = bbio->private;
1101 struct bio_vec *bvec;
1102 int sector_nr = calc_sector_number(stripe, bio_first_bvec_all(&bbio->bio));
1103 int num_sectors;
1104 u32 bio_size = 0;
1105 int i;
1106
1107 ASSERT(sector_nr < stripe->nr_sectors);
1108 bio_for_each_bvec_all(bvec, &bbio->bio, i)
1109 bio_size += bvec->bv_len;
1110 num_sectors = bio_size >> stripe->bg->fs_info->sectorsize_bits;
1111
1112 if (bbio->bio.bi_status) {
1113 bitmap_set(&stripe->io_error_bitmap, sector_nr, num_sectors);
1114 bitmap_set(&stripe->error_bitmap, sector_nr, num_sectors);
1115 } else {
1116 bitmap_clear(&stripe->io_error_bitmap, sector_nr, num_sectors);
1117 }
1118 bio_put(&bbio->bio);
1119 if (atomic_dec_and_test(&stripe->pending_io)) {
1120 wake_up(&stripe->io_wait);
1121 INIT_WORK(&stripe->work, scrub_stripe_read_repair_worker);
1122 queue_work(stripe->bg->fs_info->scrub_workers, &stripe->work);
1123 }
1124}
1125
1126static void scrub_write_endio(struct btrfs_bio *bbio)
1127{
1128 struct scrub_stripe *stripe = bbio->private;
1129 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
1130 struct bio_vec *bvec;
1131 int sector_nr = calc_sector_number(stripe, bio_first_bvec_all(&bbio->bio));
1132 u32 bio_size = 0;
1133 int i;
1134
1135 bio_for_each_bvec_all(bvec, &bbio->bio, i)
1136 bio_size += bvec->bv_len;
1137
1138 if (bbio->bio.bi_status) {
1139 unsigned long flags;
1140
1141 spin_lock_irqsave(&stripe->write_error_lock, flags);
1142 bitmap_set(&stripe->write_error_bitmap, sector_nr,
1143 bio_size >> fs_info->sectorsize_bits);
1144 spin_unlock_irqrestore(&stripe->write_error_lock, flags);
1145 }
1146 bio_put(&bbio->bio);
1147
1148 if (atomic_dec_and_test(&stripe->pending_io))
1149 wake_up(&stripe->io_wait);
1150}
1151
1152static void scrub_submit_write_bio(struct scrub_ctx *sctx,
1153 struct scrub_stripe *stripe,
1154 struct btrfs_bio *bbio, bool dev_replace)
1155{
1156 struct btrfs_fs_info *fs_info = sctx->fs_info;
1157 u32 bio_len = bbio->bio.bi_iter.bi_size;
1158 u32 bio_off = (bbio->bio.bi_iter.bi_sector << SECTOR_SHIFT) -
1159 stripe->logical;
1160
1161 fill_writer_pointer_gap(sctx, stripe->physical + bio_off);
1162 atomic_inc(&stripe->pending_io);
1163 btrfs_submit_repair_write(bbio, stripe->mirror_num, dev_replace);
1164 if (!btrfs_is_zoned(fs_info))
1165 return;
1166 /*
1167 * For zoned writeback, queue depth must be 1, thus we must wait for
1168 * the write to finish before the next write.
1169 */
1170 wait_scrub_stripe_io(stripe);
1171
1172 /*
1173 * And also need to update the write pointer if write finished
1174 * successfully.
1175 */
1176 if (!test_bit(bio_off >> fs_info->sectorsize_bits,
1177 &stripe->write_error_bitmap))
1178 sctx->write_pointer += bio_len;
1179}
1180
1181/*
1182 * Submit the write bio(s) for the sectors specified by @write_bitmap.
1183 *
1184 * Here we utilize btrfs_submit_repair_write(), which has some extra benefits:
1185 *
1186 * - Only needs logical bytenr and mirror_num
1187 * Just like the scrub read path
1188 *
1189 * - Would only result in writes to the specified mirror
1190 * Unlike the regular writeback path, which would write back to all stripes
1191 *
1192 * - Handle dev-replace and read-repair writeback differently
1193 */
1194static void scrub_write_sectors(struct scrub_ctx *sctx, struct scrub_stripe *stripe,
1195 unsigned long write_bitmap, bool dev_replace)
1196{
1197 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
1198 struct btrfs_bio *bbio = NULL;
1199 int sector_nr;
1200
1201 for_each_set_bit(sector_nr, &write_bitmap, stripe->nr_sectors) {
1202 struct page *page = scrub_stripe_get_page(stripe, sector_nr);
1203 unsigned int pgoff = scrub_stripe_get_page_offset(stripe, sector_nr);
1204 int ret;
1205
1206 /* We should only writeback sectors covered by an extent. */
1207 ASSERT(test_bit(sector_nr, &stripe->extent_sector_bitmap));
1208
1209 /* Cannot merge with previous sector, submit the current one. */
1210 if (bbio && sector_nr && !test_bit(sector_nr - 1, &write_bitmap)) {
1211 scrub_submit_write_bio(sctx, stripe, bbio, dev_replace);
1212 bbio = NULL;
1213 }
1214 if (!bbio) {
1215 bbio = btrfs_bio_alloc(stripe->nr_sectors, REQ_OP_WRITE,
1216 fs_info, scrub_write_endio, stripe);
1217 bbio->bio.bi_iter.bi_sector = (stripe->logical +
1218 (sector_nr << fs_info->sectorsize_bits)) >>
1219 SECTOR_SHIFT;
1220 }
1221 ret = bio_add_page(&bbio->bio, page, fs_info->sectorsize, pgoff);
1222 ASSERT(ret == fs_info->sectorsize);
1223 }
1224 if (bbio)
1225 scrub_submit_write_bio(sctx, stripe, bbio, dev_replace);
1226}
1227
1228/*
1229 * Throttling of IO submission, bandwidth-limit based, the timeslice is 1
1230 * second. Limit can be set via /sys/fs/UUID/devinfo/devid/scrub_speed_max.
1231 */
1232static void scrub_throttle_dev_io(struct scrub_ctx *sctx, struct btrfs_device *device,
1233 unsigned int bio_size)
1234{
1235 const int time_slice = 1000;
1236 s64 delta;
1237 ktime_t now;
1238 u32 div;
1239 u64 bwlimit;
1240
1241 bwlimit = READ_ONCE(device->scrub_speed_max);
1242 if (bwlimit == 0)
1243 return;
1244
1245 /*
1246 * Slice is divided into intervals when the IO is submitted, adjust by
1247 * bwlimit and maximum of 64 intervals.
1248 */
1249 div = max_t(u32, 1, (u32)(bwlimit / (16 * 1024 * 1024)));
1250 div = min_t(u32, 64, div);
1251
1252 /* Start new epoch, set deadline */
1253 now = ktime_get();
1254 if (sctx->throttle_deadline == 0) {
1255 sctx->throttle_deadline = ktime_add_ms(now, time_slice / div);
1256 sctx->throttle_sent = 0;
1257 }
1258
1259 /* Still in the time to send? */
1260 if (ktime_before(now, sctx->throttle_deadline)) {
1261 /* If current bio is within the limit, send it */
1262 sctx->throttle_sent += bio_size;
1263 if (sctx->throttle_sent <= div_u64(bwlimit, div))
1264 return;
1265
1266 /* We're over the limit, sleep until the rest of the slice */
1267 delta = ktime_ms_delta(sctx->throttle_deadline, now);
1268 } else {
1269 /* New request after deadline, start new epoch */
1270 delta = 0;
1271 }
1272
1273 if (delta) {
1274 long timeout;
1275
1276 timeout = div_u64(delta * HZ, 1000);
1277 schedule_timeout_interruptible(timeout);
1278 }
1279
1280 /* Next call will start the deadline period */
1281 sctx->throttle_deadline = 0;
1282}
1283
1284/*
1285 * Given a physical address, this will calculate it's
1286 * logical offset. if this is a parity stripe, it will return
1287 * the most left data stripe's logical offset.
1288 *
1289 * return 0 if it is a data stripe, 1 means parity stripe.
1290 */
1291static int get_raid56_logic_offset(u64 physical, int num,
1292 struct btrfs_chunk_map *map, u64 *offset,
1293 u64 *stripe_start)
1294{
1295 int i;
1296 int j = 0;
1297 u64 last_offset;
1298 const int data_stripes = nr_data_stripes(map);
1299
1300 last_offset = (physical - map->stripes[num].physical) * data_stripes;
1301 if (stripe_start)
1302 *stripe_start = last_offset;
1303
1304 *offset = last_offset;
1305 for (i = 0; i < data_stripes; i++) {
1306 u32 stripe_nr;
1307 u32 stripe_index;
1308 u32 rot;
1309
1310 *offset = last_offset + btrfs_stripe_nr_to_offset(i);
1311
1312 stripe_nr = (u32)(*offset >> BTRFS_STRIPE_LEN_SHIFT) / data_stripes;
1313
1314 /* Work out the disk rotation on this stripe-set */
1315 rot = stripe_nr % map->num_stripes;
1316 /* calculate which stripe this data locates */
1317 rot += i;
1318 stripe_index = rot % map->num_stripes;
1319 if (stripe_index == num)
1320 return 0;
1321 if (stripe_index < num)
1322 j++;
1323 }
1324 *offset = last_offset + btrfs_stripe_nr_to_offset(j);
1325 return 1;
1326}
1327
1328/*
1329 * Return 0 if the extent item range covers any byte of the range.
1330 * Return <0 if the extent item is before @search_start.
1331 * Return >0 if the extent item is after @start_start + @search_len.
1332 */
1333static int compare_extent_item_range(struct btrfs_path *path,
1334 u64 search_start, u64 search_len)
1335{
1336 struct btrfs_fs_info *fs_info = path->nodes[0]->fs_info;
1337 u64 len;
1338 struct btrfs_key key;
1339
1340 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1341 ASSERT(key.type == BTRFS_EXTENT_ITEM_KEY ||
1342 key.type == BTRFS_METADATA_ITEM_KEY);
1343 if (key.type == BTRFS_METADATA_ITEM_KEY)
1344 len = fs_info->nodesize;
1345 else
1346 len = key.offset;
1347
1348 if (key.objectid + len <= search_start)
1349 return -1;
1350 if (key.objectid >= search_start + search_len)
1351 return 1;
1352 return 0;
1353}
1354
1355/*
1356 * Locate one extent item which covers any byte in range
1357 * [@search_start, @search_start + @search_length)
1358 *
1359 * If the path is not initialized, we will initialize the search by doing
1360 * a btrfs_search_slot().
1361 * If the path is already initialized, we will use the path as the initial
1362 * slot, to avoid duplicated btrfs_search_slot() calls.
1363 *
1364 * NOTE: If an extent item starts before @search_start, we will still
1365 * return the extent item. This is for data extent crossing stripe boundary.
1366 *
1367 * Return 0 if we found such extent item, and @path will point to the extent item.
1368 * Return >0 if no such extent item can be found, and @path will be released.
1369 * Return <0 if hit fatal error, and @path will be released.
1370 */
1371static int find_first_extent_item(struct btrfs_root *extent_root,
1372 struct btrfs_path *path,
1373 u64 search_start, u64 search_len)
1374{
1375 struct btrfs_fs_info *fs_info = extent_root->fs_info;
1376 struct btrfs_key key;
1377 int ret;
1378
1379 /* Continue using the existing path */
1380 if (path->nodes[0])
1381 goto search_forward;
1382
1383 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1384 key.type = BTRFS_METADATA_ITEM_KEY;
1385 else
1386 key.type = BTRFS_EXTENT_ITEM_KEY;
1387 key.objectid = search_start;
1388 key.offset = (u64)-1;
1389
1390 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
1391 if (ret < 0)
1392 return ret;
1393 if (ret == 0) {
1394 /*
1395 * Key with offset -1 found, there would have to exist an extent
1396 * item with such offset, but this is out of the valid range.
1397 */
1398 btrfs_release_path(path);
1399 return -EUCLEAN;
1400 }
1401
1402 /*
1403 * Here we intentionally pass 0 as @min_objectid, as there could be
1404 * an extent item starting before @search_start.
1405 */
1406 ret = btrfs_previous_extent_item(extent_root, path, 0);
1407 if (ret < 0)
1408 return ret;
1409 /*
1410 * No matter whether we have found an extent item, the next loop will
1411 * properly do every check on the key.
1412 */
1413search_forward:
1414 while (true) {
1415 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1416 if (key.objectid >= search_start + search_len)
1417 break;
1418 if (key.type != BTRFS_METADATA_ITEM_KEY &&
1419 key.type != BTRFS_EXTENT_ITEM_KEY)
1420 goto next;
1421
1422 ret = compare_extent_item_range(path, search_start, search_len);
1423 if (ret == 0)
1424 return ret;
1425 if (ret > 0)
1426 break;
1427next:
1428 ret = btrfs_next_item(extent_root, path);
1429 if (ret) {
1430 /* Either no more items or a fatal error. */
1431 btrfs_release_path(path);
1432 return ret;
1433 }
1434 }
1435 btrfs_release_path(path);
1436 return 1;
1437}
1438
1439static void get_extent_info(struct btrfs_path *path, u64 *extent_start_ret,
1440 u64 *size_ret, u64 *flags_ret, u64 *generation_ret)
1441{
1442 struct btrfs_key key;
1443 struct btrfs_extent_item *ei;
1444
1445 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1446 ASSERT(key.type == BTRFS_METADATA_ITEM_KEY ||
1447 key.type == BTRFS_EXTENT_ITEM_KEY);
1448 *extent_start_ret = key.objectid;
1449 if (key.type == BTRFS_METADATA_ITEM_KEY)
1450 *size_ret = path->nodes[0]->fs_info->nodesize;
1451 else
1452 *size_ret = key.offset;
1453 ei = btrfs_item_ptr(path->nodes[0], path->slots[0], struct btrfs_extent_item);
1454 *flags_ret = btrfs_extent_flags(path->nodes[0], ei);
1455 *generation_ret = btrfs_extent_generation(path->nodes[0], ei);
1456}
1457
1458static int sync_write_pointer_for_zoned(struct scrub_ctx *sctx, u64 logical,
1459 u64 physical, u64 physical_end)
1460{
1461 struct btrfs_fs_info *fs_info = sctx->fs_info;
1462 int ret = 0;
1463
1464 if (!btrfs_is_zoned(fs_info))
1465 return 0;
1466
1467 mutex_lock(&sctx->wr_lock);
1468 if (sctx->write_pointer < physical_end) {
1469 ret = btrfs_sync_zone_write_pointer(sctx->wr_tgtdev, logical,
1470 physical,
1471 sctx->write_pointer);
1472 if (ret)
1473 btrfs_err(fs_info,
1474 "zoned: failed to recover write pointer");
1475 }
1476 mutex_unlock(&sctx->wr_lock);
1477 btrfs_dev_clear_zone_empty(sctx->wr_tgtdev, physical);
1478
1479 return ret;
1480}
1481
1482static void fill_one_extent_info(struct btrfs_fs_info *fs_info,
1483 struct scrub_stripe *stripe,
1484 u64 extent_start, u64 extent_len,
1485 u64 extent_flags, u64 extent_gen)
1486{
1487 for (u64 cur_logical = max(stripe->logical, extent_start);
1488 cur_logical < min(stripe->logical + BTRFS_STRIPE_LEN,
1489 extent_start + extent_len);
1490 cur_logical += fs_info->sectorsize) {
1491 const int nr_sector = (cur_logical - stripe->logical) >>
1492 fs_info->sectorsize_bits;
1493 struct scrub_sector_verification *sector =
1494 &stripe->sectors[nr_sector];
1495
1496 set_bit(nr_sector, &stripe->extent_sector_bitmap);
1497 if (extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1498 sector->is_metadata = true;
1499 sector->generation = extent_gen;
1500 }
1501 }
1502}
1503
1504static void scrub_stripe_reset_bitmaps(struct scrub_stripe *stripe)
1505{
1506 stripe->extent_sector_bitmap = 0;
1507 stripe->init_error_bitmap = 0;
1508 stripe->init_nr_io_errors = 0;
1509 stripe->init_nr_csum_errors = 0;
1510 stripe->init_nr_meta_errors = 0;
1511 stripe->error_bitmap = 0;
1512 stripe->io_error_bitmap = 0;
1513 stripe->csum_error_bitmap = 0;
1514 stripe->meta_error_bitmap = 0;
1515}
1516
1517/*
1518 * Locate one stripe which has at least one extent in its range.
1519 *
1520 * Return 0 if found such stripe, and store its info into @stripe.
1521 * Return >0 if there is no such stripe in the specified range.
1522 * Return <0 for error.
1523 */
1524static int scrub_find_fill_first_stripe(struct btrfs_block_group *bg,
1525 struct btrfs_path *extent_path,
1526 struct btrfs_path *csum_path,
1527 struct btrfs_device *dev, u64 physical,
1528 int mirror_num, u64 logical_start,
1529 u32 logical_len,
1530 struct scrub_stripe *stripe)
1531{
1532 struct btrfs_fs_info *fs_info = bg->fs_info;
1533 struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bg->start);
1534 struct btrfs_root *csum_root = btrfs_csum_root(fs_info, bg->start);
1535 const u64 logical_end = logical_start + logical_len;
1536 u64 cur_logical = logical_start;
1537 u64 stripe_end;
1538 u64 extent_start;
1539 u64 extent_len;
1540 u64 extent_flags;
1541 u64 extent_gen;
1542 int ret;
1543
1544 if (unlikely(!extent_root)) {
1545 btrfs_err(fs_info, "no valid extent root for scrub");
1546 return -EUCLEAN;
1547 }
1548 memset(stripe->sectors, 0, sizeof(struct scrub_sector_verification) *
1549 stripe->nr_sectors);
1550 scrub_stripe_reset_bitmaps(stripe);
1551
1552 /* The range must be inside the bg. */
1553 ASSERT(logical_start >= bg->start && logical_end <= bg->start + bg->length);
1554
1555 ret = find_first_extent_item(extent_root, extent_path, logical_start,
1556 logical_len);
1557 /* Either error or not found. */
1558 if (ret)
1559 goto out;
1560 get_extent_info(extent_path, &extent_start, &extent_len, &extent_flags,
1561 &extent_gen);
1562 if (extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1563 stripe->nr_meta_extents++;
1564 if (extent_flags & BTRFS_EXTENT_FLAG_DATA)
1565 stripe->nr_data_extents++;
1566 cur_logical = max(extent_start, cur_logical);
1567
1568 /*
1569 * Round down to stripe boundary.
1570 *
1571 * The extra calculation against bg->start is to handle block groups
1572 * whose logical bytenr is not BTRFS_STRIPE_LEN aligned.
1573 */
1574 stripe->logical = round_down(cur_logical - bg->start, BTRFS_STRIPE_LEN) +
1575 bg->start;
1576 stripe->physical = physical + stripe->logical - logical_start;
1577 stripe->dev = dev;
1578 stripe->bg = bg;
1579 stripe->mirror_num = mirror_num;
1580 stripe_end = stripe->logical + BTRFS_STRIPE_LEN - 1;
1581
1582 /* Fill the first extent info into stripe->sectors[] array. */
1583 fill_one_extent_info(fs_info, stripe, extent_start, extent_len,
1584 extent_flags, extent_gen);
1585 cur_logical = extent_start + extent_len;
1586
1587 /* Fill the extent info for the remaining sectors. */
1588 while (cur_logical <= stripe_end) {
1589 ret = find_first_extent_item(extent_root, extent_path, cur_logical,
1590 stripe_end - cur_logical + 1);
1591 if (ret < 0)
1592 goto out;
1593 if (ret > 0) {
1594 ret = 0;
1595 break;
1596 }
1597 get_extent_info(extent_path, &extent_start, &extent_len,
1598 &extent_flags, &extent_gen);
1599 if (extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1600 stripe->nr_meta_extents++;
1601 if (extent_flags & BTRFS_EXTENT_FLAG_DATA)
1602 stripe->nr_data_extents++;
1603 fill_one_extent_info(fs_info, stripe, extent_start, extent_len,
1604 extent_flags, extent_gen);
1605 cur_logical = extent_start + extent_len;
1606 }
1607
1608 /* Now fill the data csum. */
1609 if (bg->flags & BTRFS_BLOCK_GROUP_DATA) {
1610 int sector_nr;
1611 unsigned long csum_bitmap = 0;
1612
1613 /* Csum space should have already been allocated. */
1614 ASSERT(stripe->csums);
1615
1616 /*
1617 * Our csum bitmap should be large enough, as BTRFS_STRIPE_LEN
1618 * should contain at most 16 sectors.
1619 */
1620 ASSERT(BITS_PER_LONG >= BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits);
1621
1622 ret = btrfs_lookup_csums_bitmap(csum_root, csum_path,
1623 stripe->logical, stripe_end,
1624 stripe->csums, &csum_bitmap);
1625 if (ret < 0)
1626 goto out;
1627 if (ret > 0)
1628 ret = 0;
1629
1630 for_each_set_bit(sector_nr, &csum_bitmap, stripe->nr_sectors) {
1631 stripe->sectors[sector_nr].csum = stripe->csums +
1632 sector_nr * fs_info->csum_size;
1633 }
1634 }
1635 set_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &stripe->state);
1636out:
1637 return ret;
1638}
1639
1640static void scrub_reset_stripe(struct scrub_stripe *stripe)
1641{
1642 scrub_stripe_reset_bitmaps(stripe);
1643
1644 stripe->nr_meta_extents = 0;
1645 stripe->nr_data_extents = 0;
1646 stripe->state = 0;
1647
1648 for (int i = 0; i < stripe->nr_sectors; i++) {
1649 stripe->sectors[i].is_metadata = false;
1650 stripe->sectors[i].csum = NULL;
1651 stripe->sectors[i].generation = 0;
1652 }
1653}
1654
1655static u32 stripe_length(const struct scrub_stripe *stripe)
1656{
1657 ASSERT(stripe->bg);
1658
1659 return min(BTRFS_STRIPE_LEN,
1660 stripe->bg->start + stripe->bg->length - stripe->logical);
1661}
1662
1663static void scrub_submit_extent_sector_read(struct scrub_stripe *stripe)
1664{
1665 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
1666 struct btrfs_bio *bbio = NULL;
1667 unsigned int nr_sectors = stripe_length(stripe) >> fs_info->sectorsize_bits;
1668 u64 stripe_len = BTRFS_STRIPE_LEN;
1669 int mirror = stripe->mirror_num;
1670 int i;
1671
1672 atomic_inc(&stripe->pending_io);
1673
1674 for_each_set_bit(i, &stripe->extent_sector_bitmap, stripe->nr_sectors) {
1675 struct page *page = scrub_stripe_get_page(stripe, i);
1676 unsigned int pgoff = scrub_stripe_get_page_offset(stripe, i);
1677
1678 /* We're beyond the chunk boundary, no need to read anymore. */
1679 if (i >= nr_sectors)
1680 break;
1681
1682 /* The current sector cannot be merged, submit the bio. */
1683 if (bbio &&
1684 ((i > 0 &&
1685 !test_bit(i - 1, &stripe->extent_sector_bitmap)) ||
1686 bbio->bio.bi_iter.bi_size >= stripe_len)) {
1687 ASSERT(bbio->bio.bi_iter.bi_size);
1688 atomic_inc(&stripe->pending_io);
1689 btrfs_submit_bbio(bbio, mirror);
1690 bbio = NULL;
1691 }
1692
1693 if (!bbio) {
1694 struct btrfs_io_stripe io_stripe = {};
1695 struct btrfs_io_context *bioc = NULL;
1696 const u64 logical = stripe->logical +
1697 (i << fs_info->sectorsize_bits);
1698 int err;
1699
1700 io_stripe.rst_search_commit_root = true;
1701 stripe_len = (nr_sectors - i) << fs_info->sectorsize_bits;
1702 /*
1703 * For RST cases, we need to manually split the bbio to
1704 * follow the RST boundary.
1705 */
1706 err = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical,
1707 &stripe_len, &bioc, &io_stripe, &mirror);
1708 btrfs_put_bioc(bioc);
1709 if (err < 0) {
1710 if (err != -ENODATA) {
1711 /*
1712 * Earlier btrfs_get_raid_extent_offset()
1713 * returned -ENODATA, which means there's
1714 * no entry for the corresponding range
1715 * in the stripe tree. But if it's in
1716 * the extent tree, then it's a preallocated
1717 * extent and not an error.
1718 */
1719 set_bit(i, &stripe->io_error_bitmap);
1720 set_bit(i, &stripe->error_bitmap);
1721 }
1722 continue;
1723 }
1724
1725 bbio = btrfs_bio_alloc(stripe->nr_sectors, REQ_OP_READ,
1726 fs_info, scrub_read_endio, stripe);
1727 bbio->bio.bi_iter.bi_sector = logical >> SECTOR_SHIFT;
1728 }
1729
1730 __bio_add_page(&bbio->bio, page, fs_info->sectorsize, pgoff);
1731 }
1732
1733 if (bbio) {
1734 ASSERT(bbio->bio.bi_iter.bi_size);
1735 atomic_inc(&stripe->pending_io);
1736 btrfs_submit_bbio(bbio, mirror);
1737 }
1738
1739 if (atomic_dec_and_test(&stripe->pending_io)) {
1740 wake_up(&stripe->io_wait);
1741 INIT_WORK(&stripe->work, scrub_stripe_read_repair_worker);
1742 queue_work(stripe->bg->fs_info->scrub_workers, &stripe->work);
1743 }
1744}
1745
1746static void scrub_submit_initial_read(struct scrub_ctx *sctx,
1747 struct scrub_stripe *stripe)
1748{
1749 struct btrfs_fs_info *fs_info = sctx->fs_info;
1750 struct btrfs_bio *bbio;
1751 unsigned int nr_sectors = stripe_length(stripe) >> fs_info->sectorsize_bits;
1752 int mirror = stripe->mirror_num;
1753
1754 ASSERT(stripe->bg);
1755 ASSERT(stripe->mirror_num > 0);
1756 ASSERT(test_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &stripe->state));
1757
1758 if (btrfs_need_stripe_tree_update(fs_info, stripe->bg->flags)) {
1759 scrub_submit_extent_sector_read(stripe);
1760 return;
1761 }
1762
1763 bbio = btrfs_bio_alloc(SCRUB_STRIPE_PAGES, REQ_OP_READ, fs_info,
1764 scrub_read_endio, stripe);
1765
1766 bbio->bio.bi_iter.bi_sector = stripe->logical >> SECTOR_SHIFT;
1767 /* Read the whole range inside the chunk boundary. */
1768 for (unsigned int cur = 0; cur < nr_sectors; cur++) {
1769 struct page *page = scrub_stripe_get_page(stripe, cur);
1770 unsigned int pgoff = scrub_stripe_get_page_offset(stripe, cur);
1771 int ret;
1772
1773 ret = bio_add_page(&bbio->bio, page, fs_info->sectorsize, pgoff);
1774 /* We should have allocated enough bio vectors. */
1775 ASSERT(ret == fs_info->sectorsize);
1776 }
1777 atomic_inc(&stripe->pending_io);
1778
1779 /*
1780 * For dev-replace, either user asks to avoid the source dev, or
1781 * the device is missing, we try the next mirror instead.
1782 */
1783 if (sctx->is_dev_replace &&
1784 (fs_info->dev_replace.cont_reading_from_srcdev_mode ==
1785 BTRFS_DEV_REPLACE_ITEM_CONT_READING_FROM_SRCDEV_MODE_AVOID ||
1786 !stripe->dev->bdev)) {
1787 int num_copies = btrfs_num_copies(fs_info, stripe->bg->start,
1788 stripe->bg->length);
1789
1790 mirror = calc_next_mirror(mirror, num_copies);
1791 }
1792 btrfs_submit_bbio(bbio, mirror);
1793}
1794
1795static bool stripe_has_metadata_error(struct scrub_stripe *stripe)
1796{
1797 int i;
1798
1799 for_each_set_bit(i, &stripe->error_bitmap, stripe->nr_sectors) {
1800 if (stripe->sectors[i].is_metadata) {
1801 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
1802
1803 btrfs_err(fs_info,
1804 "stripe %llu has unrepaired metadata sector at %llu",
1805 stripe->logical,
1806 stripe->logical + (i << fs_info->sectorsize_bits));
1807 return true;
1808 }
1809 }
1810 return false;
1811}
1812
1813static void submit_initial_group_read(struct scrub_ctx *sctx,
1814 unsigned int first_slot,
1815 unsigned int nr_stripes)
1816{
1817 struct blk_plug plug;
1818
1819 ASSERT(first_slot < SCRUB_TOTAL_STRIPES);
1820 ASSERT(first_slot + nr_stripes <= SCRUB_TOTAL_STRIPES);
1821
1822 scrub_throttle_dev_io(sctx, sctx->stripes[0].dev,
1823 btrfs_stripe_nr_to_offset(nr_stripes));
1824 blk_start_plug(&plug);
1825 for (int i = 0; i < nr_stripes; i++) {
1826 struct scrub_stripe *stripe = &sctx->stripes[first_slot + i];
1827
1828 /* Those stripes should be initialized. */
1829 ASSERT(test_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &stripe->state));
1830 scrub_submit_initial_read(sctx, stripe);
1831 }
1832 blk_finish_plug(&plug);
1833}
1834
1835static int flush_scrub_stripes(struct scrub_ctx *sctx)
1836{
1837 struct btrfs_fs_info *fs_info = sctx->fs_info;
1838 struct scrub_stripe *stripe;
1839 const int nr_stripes = sctx->cur_stripe;
1840 int ret = 0;
1841
1842 if (!nr_stripes)
1843 return 0;
1844
1845 ASSERT(test_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &sctx->stripes[0].state));
1846
1847 /* Submit the stripes which are populated but not submitted. */
1848 if (nr_stripes % SCRUB_STRIPES_PER_GROUP) {
1849 const int first_slot = round_down(nr_stripes, SCRUB_STRIPES_PER_GROUP);
1850
1851 submit_initial_group_read(sctx, first_slot, nr_stripes - first_slot);
1852 }
1853
1854 for (int i = 0; i < nr_stripes; i++) {
1855 stripe = &sctx->stripes[i];
1856
1857 wait_event(stripe->repair_wait,
1858 test_bit(SCRUB_STRIPE_FLAG_REPAIR_DONE, &stripe->state));
1859 }
1860
1861 /* Submit for dev-replace. */
1862 if (sctx->is_dev_replace) {
1863 /*
1864 * For dev-replace, if we know there is something wrong with
1865 * metadata, we should immediately abort.
1866 */
1867 for (int i = 0; i < nr_stripes; i++) {
1868 if (stripe_has_metadata_error(&sctx->stripes[i])) {
1869 ret = -EIO;
1870 goto out;
1871 }
1872 }
1873 for (int i = 0; i < nr_stripes; i++) {
1874 unsigned long good;
1875
1876 stripe = &sctx->stripes[i];
1877
1878 ASSERT(stripe->dev == fs_info->dev_replace.srcdev);
1879
1880 bitmap_andnot(&good, &stripe->extent_sector_bitmap,
1881 &stripe->error_bitmap, stripe->nr_sectors);
1882 scrub_write_sectors(sctx, stripe, good, true);
1883 }
1884 }
1885
1886 /* Wait for the above writebacks to finish. */
1887 for (int i = 0; i < nr_stripes; i++) {
1888 stripe = &sctx->stripes[i];
1889
1890 wait_scrub_stripe_io(stripe);
1891 spin_lock(&sctx->stat_lock);
1892 sctx->stat.last_physical = stripe->physical + stripe_length(stripe);
1893 spin_unlock(&sctx->stat_lock);
1894 scrub_reset_stripe(stripe);
1895 }
1896out:
1897 sctx->cur_stripe = 0;
1898 return ret;
1899}
1900
1901static void raid56_scrub_wait_endio(struct bio *bio)
1902{
1903 complete(bio->bi_private);
1904}
1905
1906static int queue_scrub_stripe(struct scrub_ctx *sctx, struct btrfs_block_group *bg,
1907 struct btrfs_device *dev, int mirror_num,
1908 u64 logical, u32 length, u64 physical,
1909 u64 *found_logical_ret)
1910{
1911 struct scrub_stripe *stripe;
1912 int ret;
1913
1914 /*
1915 * There should always be one slot left, as caller filling the last
1916 * slot should flush them all.
1917 */
1918 ASSERT(sctx->cur_stripe < SCRUB_TOTAL_STRIPES);
1919
1920 /* @found_logical_ret must be specified. */
1921 ASSERT(found_logical_ret);
1922
1923 stripe = &sctx->stripes[sctx->cur_stripe];
1924 scrub_reset_stripe(stripe);
1925 ret = scrub_find_fill_first_stripe(bg, &sctx->extent_path,
1926 &sctx->csum_path, dev, physical,
1927 mirror_num, logical, length, stripe);
1928 /* Either >0 as no more extents or <0 for error. */
1929 if (ret)
1930 return ret;
1931 *found_logical_ret = stripe->logical;
1932 sctx->cur_stripe++;
1933
1934 /* We filled one group, submit it. */
1935 if (sctx->cur_stripe % SCRUB_STRIPES_PER_GROUP == 0) {
1936 const int first_slot = sctx->cur_stripe - SCRUB_STRIPES_PER_GROUP;
1937
1938 submit_initial_group_read(sctx, first_slot, SCRUB_STRIPES_PER_GROUP);
1939 }
1940
1941 /* Last slot used, flush them all. */
1942 if (sctx->cur_stripe == SCRUB_TOTAL_STRIPES)
1943 return flush_scrub_stripes(sctx);
1944 return 0;
1945}
1946
1947static int scrub_raid56_parity_stripe(struct scrub_ctx *sctx,
1948 struct btrfs_device *scrub_dev,
1949 struct btrfs_block_group *bg,
1950 struct btrfs_chunk_map *map,
1951 u64 full_stripe_start)
1952{
1953 DECLARE_COMPLETION_ONSTACK(io_done);
1954 struct btrfs_fs_info *fs_info = sctx->fs_info;
1955 struct btrfs_raid_bio *rbio;
1956 struct btrfs_io_context *bioc = NULL;
1957 struct btrfs_path extent_path = { 0 };
1958 struct btrfs_path csum_path = { 0 };
1959 struct bio *bio;
1960 struct scrub_stripe *stripe;
1961 bool all_empty = true;
1962 const int data_stripes = nr_data_stripes(map);
1963 unsigned long extent_bitmap = 0;
1964 u64 length = btrfs_stripe_nr_to_offset(data_stripes);
1965 int ret;
1966
1967 ASSERT(sctx->raid56_data_stripes);
1968
1969 /*
1970 * For data stripe search, we cannot reuse the same extent/csum paths,
1971 * as the data stripe bytenr may be smaller than previous extent. Thus
1972 * we have to use our own extent/csum paths.
1973 */
1974 extent_path.search_commit_root = 1;
1975 extent_path.skip_locking = 1;
1976 csum_path.search_commit_root = 1;
1977 csum_path.skip_locking = 1;
1978
1979 for (int i = 0; i < data_stripes; i++) {
1980 int stripe_index;
1981 int rot;
1982 u64 physical;
1983
1984 stripe = &sctx->raid56_data_stripes[i];
1985 rot = div_u64(full_stripe_start - bg->start,
1986 data_stripes) >> BTRFS_STRIPE_LEN_SHIFT;
1987 stripe_index = (i + rot) % map->num_stripes;
1988 physical = map->stripes[stripe_index].physical +
1989 btrfs_stripe_nr_to_offset(rot);
1990
1991 scrub_reset_stripe(stripe);
1992 set_bit(SCRUB_STRIPE_FLAG_NO_REPORT, &stripe->state);
1993 ret = scrub_find_fill_first_stripe(bg, &extent_path, &csum_path,
1994 map->stripes[stripe_index].dev, physical, 1,
1995 full_stripe_start + btrfs_stripe_nr_to_offset(i),
1996 BTRFS_STRIPE_LEN, stripe);
1997 if (ret < 0)
1998 goto out;
1999 /*
2000 * No extent in this data stripe, need to manually mark them
2001 * initialized to make later read submission happy.
2002 */
2003 if (ret > 0) {
2004 stripe->logical = full_stripe_start +
2005 btrfs_stripe_nr_to_offset(i);
2006 stripe->dev = map->stripes[stripe_index].dev;
2007 stripe->mirror_num = 1;
2008 set_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &stripe->state);
2009 }
2010 }
2011
2012 /* Check if all data stripes are empty. */
2013 for (int i = 0; i < data_stripes; i++) {
2014 stripe = &sctx->raid56_data_stripes[i];
2015 if (!bitmap_empty(&stripe->extent_sector_bitmap, stripe->nr_sectors)) {
2016 all_empty = false;
2017 break;
2018 }
2019 }
2020 if (all_empty) {
2021 ret = 0;
2022 goto out;
2023 }
2024
2025 for (int i = 0; i < data_stripes; i++) {
2026 stripe = &sctx->raid56_data_stripes[i];
2027 scrub_submit_initial_read(sctx, stripe);
2028 }
2029 for (int i = 0; i < data_stripes; i++) {
2030 stripe = &sctx->raid56_data_stripes[i];
2031
2032 wait_event(stripe->repair_wait,
2033 test_bit(SCRUB_STRIPE_FLAG_REPAIR_DONE, &stripe->state));
2034 }
2035 /* For now, no zoned support for RAID56. */
2036 ASSERT(!btrfs_is_zoned(sctx->fs_info));
2037
2038 /*
2039 * Now all data stripes are properly verified. Check if we have any
2040 * unrepaired, if so abort immediately or we could further corrupt the
2041 * P/Q stripes.
2042 *
2043 * During the loop, also populate extent_bitmap.
2044 */
2045 for (int i = 0; i < data_stripes; i++) {
2046 unsigned long error;
2047
2048 stripe = &sctx->raid56_data_stripes[i];
2049
2050 /*
2051 * We should only check the errors where there is an extent.
2052 * As we may hit an empty data stripe while it's missing.
2053 */
2054 bitmap_and(&error, &stripe->error_bitmap,
2055 &stripe->extent_sector_bitmap, stripe->nr_sectors);
2056 if (!bitmap_empty(&error, stripe->nr_sectors)) {
2057 btrfs_err(fs_info,
2058"unrepaired sectors detected, full stripe %llu data stripe %u errors %*pbl",
2059 full_stripe_start, i, stripe->nr_sectors,
2060 &error);
2061 ret = -EIO;
2062 goto out;
2063 }
2064 bitmap_or(&extent_bitmap, &extent_bitmap,
2065 &stripe->extent_sector_bitmap, stripe->nr_sectors);
2066 }
2067
2068 /* Now we can check and regenerate the P/Q stripe. */
2069 bio = bio_alloc(NULL, 1, REQ_OP_READ, GFP_NOFS);
2070 bio->bi_iter.bi_sector = full_stripe_start >> SECTOR_SHIFT;
2071 bio->bi_private = &io_done;
2072 bio->bi_end_io = raid56_scrub_wait_endio;
2073
2074 btrfs_bio_counter_inc_blocked(fs_info);
2075 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, full_stripe_start,
2076 &length, &bioc, NULL, NULL);
2077 if (ret < 0) {
2078 btrfs_put_bioc(bioc);
2079 btrfs_bio_counter_dec(fs_info);
2080 goto out;
2081 }
2082 rbio = raid56_parity_alloc_scrub_rbio(bio, bioc, scrub_dev, &extent_bitmap,
2083 BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits);
2084 btrfs_put_bioc(bioc);
2085 if (!rbio) {
2086 ret = -ENOMEM;
2087 btrfs_bio_counter_dec(fs_info);
2088 goto out;
2089 }
2090 /* Use the recovered stripes as cache to avoid read them from disk again. */
2091 for (int i = 0; i < data_stripes; i++) {
2092 stripe = &sctx->raid56_data_stripes[i];
2093
2094 raid56_parity_cache_data_pages(rbio, stripe->pages,
2095 full_stripe_start + (i << BTRFS_STRIPE_LEN_SHIFT));
2096 }
2097 raid56_parity_submit_scrub_rbio(rbio);
2098 wait_for_completion_io(&io_done);
2099 ret = blk_status_to_errno(bio->bi_status);
2100 bio_put(bio);
2101 btrfs_bio_counter_dec(fs_info);
2102
2103 btrfs_release_path(&extent_path);
2104 btrfs_release_path(&csum_path);
2105out:
2106 return ret;
2107}
2108
2109/*
2110 * Scrub one range which can only has simple mirror based profile.
2111 * (Including all range in SINGLE/DUP/RAID1/RAID1C*, and each stripe in
2112 * RAID0/RAID10).
2113 *
2114 * Since we may need to handle a subset of block group, we need @logical_start
2115 * and @logical_length parameter.
2116 */
2117static int scrub_simple_mirror(struct scrub_ctx *sctx,
2118 struct btrfs_block_group *bg,
2119 u64 logical_start, u64 logical_length,
2120 struct btrfs_device *device,
2121 u64 physical, int mirror_num)
2122{
2123 struct btrfs_fs_info *fs_info = sctx->fs_info;
2124 const u64 logical_end = logical_start + logical_length;
2125 u64 cur_logical = logical_start;
2126 int ret = 0;
2127
2128 /* The range must be inside the bg */
2129 ASSERT(logical_start >= bg->start && logical_end <= bg->start + bg->length);
2130
2131 /* Go through each extent items inside the logical range */
2132 while (cur_logical < logical_end) {
2133 u64 found_logical = U64_MAX;
2134 u64 cur_physical = physical + cur_logical - logical_start;
2135
2136 /* Canceled? */
2137 if (atomic_read(&fs_info->scrub_cancel_req) ||
2138 atomic_read(&sctx->cancel_req)) {
2139 ret = -ECANCELED;
2140 break;
2141 }
2142 /* Paused? */
2143 if (atomic_read(&fs_info->scrub_pause_req)) {
2144 /* Push queued extents */
2145 scrub_blocked_if_needed(fs_info);
2146 }
2147 /* Block group removed? */
2148 spin_lock(&bg->lock);
2149 if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &bg->runtime_flags)) {
2150 spin_unlock(&bg->lock);
2151 ret = 0;
2152 break;
2153 }
2154 spin_unlock(&bg->lock);
2155
2156 ret = queue_scrub_stripe(sctx, bg, device, mirror_num,
2157 cur_logical, logical_end - cur_logical,
2158 cur_physical, &found_logical);
2159 if (ret > 0) {
2160 /* No more extent, just update the accounting */
2161 spin_lock(&sctx->stat_lock);
2162 sctx->stat.last_physical = physical + logical_length;
2163 spin_unlock(&sctx->stat_lock);
2164 ret = 0;
2165 break;
2166 }
2167 if (ret < 0)
2168 break;
2169
2170 /* queue_scrub_stripe() returned 0, @found_logical must be updated. */
2171 ASSERT(found_logical != U64_MAX);
2172 cur_logical = found_logical + BTRFS_STRIPE_LEN;
2173
2174 /* Don't hold CPU for too long time */
2175 cond_resched();
2176 }
2177 return ret;
2178}
2179
2180/* Calculate the full stripe length for simple stripe based profiles */
2181static u64 simple_stripe_full_stripe_len(const struct btrfs_chunk_map *map)
2182{
2183 ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2184 BTRFS_BLOCK_GROUP_RAID10));
2185
2186 return btrfs_stripe_nr_to_offset(map->num_stripes / map->sub_stripes);
2187}
2188
2189/* Get the logical bytenr for the stripe */
2190static u64 simple_stripe_get_logical(struct btrfs_chunk_map *map,
2191 struct btrfs_block_group *bg,
2192 int stripe_index)
2193{
2194 ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2195 BTRFS_BLOCK_GROUP_RAID10));
2196 ASSERT(stripe_index < map->num_stripes);
2197
2198 /*
2199 * (stripe_index / sub_stripes) gives how many data stripes we need to
2200 * skip.
2201 */
2202 return btrfs_stripe_nr_to_offset(stripe_index / map->sub_stripes) +
2203 bg->start;
2204}
2205
2206/* Get the mirror number for the stripe */
2207static int simple_stripe_mirror_num(struct btrfs_chunk_map *map, int stripe_index)
2208{
2209 ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2210 BTRFS_BLOCK_GROUP_RAID10));
2211 ASSERT(stripe_index < map->num_stripes);
2212
2213 /* For RAID0, it's fixed to 1, for RAID10 it's 0,1,0,1... */
2214 return stripe_index % map->sub_stripes + 1;
2215}
2216
2217static int scrub_simple_stripe(struct scrub_ctx *sctx,
2218 struct btrfs_block_group *bg,
2219 struct btrfs_chunk_map *map,
2220 struct btrfs_device *device,
2221 int stripe_index)
2222{
2223 const u64 logical_increment = simple_stripe_full_stripe_len(map);
2224 const u64 orig_logical = simple_stripe_get_logical(map, bg, stripe_index);
2225 const u64 orig_physical = map->stripes[stripe_index].physical;
2226 const int mirror_num = simple_stripe_mirror_num(map, stripe_index);
2227 u64 cur_logical = orig_logical;
2228 u64 cur_physical = orig_physical;
2229 int ret = 0;
2230
2231 while (cur_logical < bg->start + bg->length) {
2232 /*
2233 * Inside each stripe, RAID0 is just SINGLE, and RAID10 is
2234 * just RAID1, so we can reuse scrub_simple_mirror() to scrub
2235 * this stripe.
2236 */
2237 ret = scrub_simple_mirror(sctx, bg, cur_logical,
2238 BTRFS_STRIPE_LEN, device, cur_physical,
2239 mirror_num);
2240 if (ret)
2241 return ret;
2242 /* Skip to next stripe which belongs to the target device */
2243 cur_logical += logical_increment;
2244 /* For physical offset, we just go to next stripe */
2245 cur_physical += BTRFS_STRIPE_LEN;
2246 }
2247 return ret;
2248}
2249
2250static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2251 struct btrfs_block_group *bg,
2252 struct btrfs_chunk_map *map,
2253 struct btrfs_device *scrub_dev,
2254 int stripe_index)
2255{
2256 struct btrfs_fs_info *fs_info = sctx->fs_info;
2257 const u64 profile = map->type & BTRFS_BLOCK_GROUP_PROFILE_MASK;
2258 const u64 chunk_logical = bg->start;
2259 int ret;
2260 int ret2;
2261 u64 physical = map->stripes[stripe_index].physical;
2262 const u64 dev_stripe_len = btrfs_calc_stripe_length(map);
2263 const u64 physical_end = physical + dev_stripe_len;
2264 u64 logical;
2265 u64 logic_end;
2266 /* The logical increment after finishing one stripe */
2267 u64 increment;
2268 /* Offset inside the chunk */
2269 u64 offset;
2270 u64 stripe_logical;
2271
2272 /* Extent_path should be released by now. */
2273 ASSERT(sctx->extent_path.nodes[0] == NULL);
2274
2275 scrub_blocked_if_needed(fs_info);
2276
2277 if (sctx->is_dev_replace &&
2278 btrfs_dev_is_sequential(sctx->wr_tgtdev, physical)) {
2279 mutex_lock(&sctx->wr_lock);
2280 sctx->write_pointer = physical;
2281 mutex_unlock(&sctx->wr_lock);
2282 }
2283
2284 /* Prepare the extra data stripes used by RAID56. */
2285 if (profile & BTRFS_BLOCK_GROUP_RAID56_MASK) {
2286 ASSERT(sctx->raid56_data_stripes == NULL);
2287
2288 sctx->raid56_data_stripes = kcalloc(nr_data_stripes(map),
2289 sizeof(struct scrub_stripe),
2290 GFP_KERNEL);
2291 if (!sctx->raid56_data_stripes) {
2292 ret = -ENOMEM;
2293 goto out;
2294 }
2295 for (int i = 0; i < nr_data_stripes(map); i++) {
2296 ret = init_scrub_stripe(fs_info,
2297 &sctx->raid56_data_stripes[i]);
2298 if (ret < 0)
2299 goto out;
2300 sctx->raid56_data_stripes[i].bg = bg;
2301 sctx->raid56_data_stripes[i].sctx = sctx;
2302 }
2303 }
2304 /*
2305 * There used to be a big double loop to handle all profiles using the
2306 * same routine, which grows larger and more gross over time.
2307 *
2308 * So here we handle each profile differently, so simpler profiles
2309 * have simpler scrubbing function.
2310 */
2311 if (!(profile & (BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID10 |
2312 BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2313 /*
2314 * Above check rules out all complex profile, the remaining
2315 * profiles are SINGLE|DUP|RAID1|RAID1C*, which is simple
2316 * mirrored duplication without stripe.
2317 *
2318 * Only @physical and @mirror_num needs to calculated using
2319 * @stripe_index.
2320 */
2321 ret = scrub_simple_mirror(sctx, bg, bg->start, bg->length,
2322 scrub_dev, map->stripes[stripe_index].physical,
2323 stripe_index + 1);
2324 offset = 0;
2325 goto out;
2326 }
2327 if (profile & (BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID10)) {
2328 ret = scrub_simple_stripe(sctx, bg, map, scrub_dev, stripe_index);
2329 offset = btrfs_stripe_nr_to_offset(stripe_index / map->sub_stripes);
2330 goto out;
2331 }
2332
2333 /* Only RAID56 goes through the old code */
2334 ASSERT(map->type & BTRFS_BLOCK_GROUP_RAID56_MASK);
2335 ret = 0;
2336
2337 /* Calculate the logical end of the stripe */
2338 get_raid56_logic_offset(physical_end, stripe_index,
2339 map, &logic_end, NULL);
2340 logic_end += chunk_logical;
2341
2342 /* Initialize @offset in case we need to go to out: label */
2343 get_raid56_logic_offset(physical, stripe_index, map, &offset, NULL);
2344 increment = btrfs_stripe_nr_to_offset(nr_data_stripes(map));
2345
2346 /*
2347 * Due to the rotation, for RAID56 it's better to iterate each stripe
2348 * using their physical offset.
2349 */
2350 while (physical < physical_end) {
2351 ret = get_raid56_logic_offset(physical, stripe_index, map,
2352 &logical, &stripe_logical);
2353 logical += chunk_logical;
2354 if (ret) {
2355 /* it is parity strip */
2356 stripe_logical += chunk_logical;
2357 ret = scrub_raid56_parity_stripe(sctx, scrub_dev, bg,
2358 map, stripe_logical);
2359 spin_lock(&sctx->stat_lock);
2360 sctx->stat.last_physical = min(physical + BTRFS_STRIPE_LEN,
2361 physical_end);
2362 spin_unlock(&sctx->stat_lock);
2363 if (ret)
2364 goto out;
2365 goto next;
2366 }
2367
2368 /*
2369 * Now we're at a data stripe, scrub each extents in the range.
2370 *
2371 * At this stage, if we ignore the repair part, inside each data
2372 * stripe it is no different than SINGLE profile.
2373 * We can reuse scrub_simple_mirror() here, as the repair part
2374 * is still based on @mirror_num.
2375 */
2376 ret = scrub_simple_mirror(sctx, bg, logical, BTRFS_STRIPE_LEN,
2377 scrub_dev, physical, 1);
2378 if (ret < 0)
2379 goto out;
2380next:
2381 logical += increment;
2382 physical += BTRFS_STRIPE_LEN;
2383 spin_lock(&sctx->stat_lock);
2384 sctx->stat.last_physical = physical;
2385 spin_unlock(&sctx->stat_lock);
2386 }
2387out:
2388 ret2 = flush_scrub_stripes(sctx);
2389 if (!ret)
2390 ret = ret2;
2391 btrfs_release_path(&sctx->extent_path);
2392 btrfs_release_path(&sctx->csum_path);
2393
2394 if (sctx->raid56_data_stripes) {
2395 for (int i = 0; i < nr_data_stripes(map); i++)
2396 release_scrub_stripe(&sctx->raid56_data_stripes[i]);
2397 kfree(sctx->raid56_data_stripes);
2398 sctx->raid56_data_stripes = NULL;
2399 }
2400
2401 if (sctx->is_dev_replace && ret >= 0) {
2402 int ret2;
2403
2404 ret2 = sync_write_pointer_for_zoned(sctx,
2405 chunk_logical + offset,
2406 map->stripes[stripe_index].physical,
2407 physical_end);
2408 if (ret2)
2409 ret = ret2;
2410 }
2411
2412 return ret < 0 ? ret : 0;
2413}
2414
2415static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
2416 struct btrfs_block_group *bg,
2417 struct btrfs_device *scrub_dev,
2418 u64 dev_offset,
2419 u64 dev_extent_len)
2420{
2421 struct btrfs_fs_info *fs_info = sctx->fs_info;
2422 struct btrfs_chunk_map *map;
2423 int i;
2424 int ret = 0;
2425
2426 map = btrfs_find_chunk_map(fs_info, bg->start, bg->length);
2427 if (!map) {
2428 /*
2429 * Might have been an unused block group deleted by the cleaner
2430 * kthread or relocation.
2431 */
2432 spin_lock(&bg->lock);
2433 if (!test_bit(BLOCK_GROUP_FLAG_REMOVED, &bg->runtime_flags))
2434 ret = -EINVAL;
2435 spin_unlock(&bg->lock);
2436
2437 return ret;
2438 }
2439 if (map->start != bg->start)
2440 goto out;
2441 if (map->chunk_len < dev_extent_len)
2442 goto out;
2443
2444 for (i = 0; i < map->num_stripes; ++i) {
2445 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
2446 map->stripes[i].physical == dev_offset) {
2447 ret = scrub_stripe(sctx, bg, map, scrub_dev, i);
2448 if (ret)
2449 goto out;
2450 }
2451 }
2452out:
2453 btrfs_free_chunk_map(map);
2454
2455 return ret;
2456}
2457
2458static int finish_extent_writes_for_zoned(struct btrfs_root *root,
2459 struct btrfs_block_group *cache)
2460{
2461 struct btrfs_fs_info *fs_info = cache->fs_info;
2462
2463 if (!btrfs_is_zoned(fs_info))
2464 return 0;
2465
2466 btrfs_wait_block_group_reservations(cache);
2467 btrfs_wait_nocow_writers(cache);
2468 btrfs_wait_ordered_roots(fs_info, U64_MAX, cache);
2469
2470 return btrfs_commit_current_transaction(root);
2471}
2472
2473static noinline_for_stack
2474int scrub_enumerate_chunks(struct scrub_ctx *sctx,
2475 struct btrfs_device *scrub_dev, u64 start, u64 end)
2476{
2477 struct btrfs_dev_extent *dev_extent = NULL;
2478 struct btrfs_path *path;
2479 struct btrfs_fs_info *fs_info = sctx->fs_info;
2480 struct btrfs_root *root = fs_info->dev_root;
2481 u64 chunk_offset;
2482 int ret = 0;
2483 int ro_set;
2484 int slot;
2485 struct extent_buffer *l;
2486 struct btrfs_key key;
2487 struct btrfs_key found_key;
2488 struct btrfs_block_group *cache;
2489 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2490
2491 path = btrfs_alloc_path();
2492 if (!path)
2493 return -ENOMEM;
2494
2495 path->reada = READA_FORWARD;
2496 path->search_commit_root = 1;
2497 path->skip_locking = 1;
2498
2499 key.objectid = scrub_dev->devid;
2500 key.offset = 0ull;
2501 key.type = BTRFS_DEV_EXTENT_KEY;
2502
2503 while (1) {
2504 u64 dev_extent_len;
2505
2506 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2507 if (ret < 0)
2508 break;
2509 if (ret > 0) {
2510 if (path->slots[0] >=
2511 btrfs_header_nritems(path->nodes[0])) {
2512 ret = btrfs_next_leaf(root, path);
2513 if (ret < 0)
2514 break;
2515 if (ret > 0) {
2516 ret = 0;
2517 break;
2518 }
2519 } else {
2520 ret = 0;
2521 }
2522 }
2523
2524 l = path->nodes[0];
2525 slot = path->slots[0];
2526
2527 btrfs_item_key_to_cpu(l, &found_key, slot);
2528
2529 if (found_key.objectid != scrub_dev->devid)
2530 break;
2531
2532 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
2533 break;
2534
2535 if (found_key.offset >= end)
2536 break;
2537
2538 if (found_key.offset < key.offset)
2539 break;
2540
2541 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2542 dev_extent_len = btrfs_dev_extent_length(l, dev_extent);
2543
2544 if (found_key.offset + dev_extent_len <= start)
2545 goto skip;
2546
2547 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2548
2549 /*
2550 * get a reference on the corresponding block group to prevent
2551 * the chunk from going away while we scrub it
2552 */
2553 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2554
2555 /* some chunks are removed but not committed to disk yet,
2556 * continue scrubbing */
2557 if (!cache)
2558 goto skip;
2559
2560 ASSERT(cache->start <= chunk_offset);
2561 /*
2562 * We are using the commit root to search for device extents, so
2563 * that means we could have found a device extent item from a
2564 * block group that was deleted in the current transaction. The
2565 * logical start offset of the deleted block group, stored at
2566 * @chunk_offset, might be part of the logical address range of
2567 * a new block group (which uses different physical extents).
2568 * In this case btrfs_lookup_block_group() has returned the new
2569 * block group, and its start address is less than @chunk_offset.
2570 *
2571 * We skip such new block groups, because it's pointless to
2572 * process them, as we won't find their extents because we search
2573 * for them using the commit root of the extent tree. For a device
2574 * replace it's also fine to skip it, we won't miss copying them
2575 * to the target device because we have the write duplication
2576 * setup through the regular write path (by btrfs_map_block()),
2577 * and we have committed a transaction when we started the device
2578 * replace, right after setting up the device replace state.
2579 */
2580 if (cache->start < chunk_offset) {
2581 btrfs_put_block_group(cache);
2582 goto skip;
2583 }
2584
2585 if (sctx->is_dev_replace && btrfs_is_zoned(fs_info)) {
2586 if (!test_bit(BLOCK_GROUP_FLAG_TO_COPY, &cache->runtime_flags)) {
2587 btrfs_put_block_group(cache);
2588 goto skip;
2589 }
2590 }
2591
2592 /*
2593 * Make sure that while we are scrubbing the corresponding block
2594 * group doesn't get its logical address and its device extents
2595 * reused for another block group, which can possibly be of a
2596 * different type and different profile. We do this to prevent
2597 * false error detections and crashes due to bogus attempts to
2598 * repair extents.
2599 */
2600 spin_lock(&cache->lock);
2601 if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &cache->runtime_flags)) {
2602 spin_unlock(&cache->lock);
2603 btrfs_put_block_group(cache);
2604 goto skip;
2605 }
2606 btrfs_freeze_block_group(cache);
2607 spin_unlock(&cache->lock);
2608
2609 /*
2610 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
2611 * to avoid deadlock caused by:
2612 * btrfs_inc_block_group_ro()
2613 * -> btrfs_wait_for_commit()
2614 * -> btrfs_commit_transaction()
2615 * -> btrfs_scrub_pause()
2616 */
2617 scrub_pause_on(fs_info);
2618
2619 /*
2620 * Don't do chunk preallocation for scrub.
2621 *
2622 * This is especially important for SYSTEM bgs, or we can hit
2623 * -EFBIG from btrfs_finish_chunk_alloc() like:
2624 * 1. The only SYSTEM bg is marked RO.
2625 * Since SYSTEM bg is small, that's pretty common.
2626 * 2. New SYSTEM bg will be allocated
2627 * Due to regular version will allocate new chunk.
2628 * 3. New SYSTEM bg is empty and will get cleaned up
2629 * Before cleanup really happens, it's marked RO again.
2630 * 4. Empty SYSTEM bg get scrubbed
2631 * We go back to 2.
2632 *
2633 * This can easily boost the amount of SYSTEM chunks if cleaner
2634 * thread can't be triggered fast enough, and use up all space
2635 * of btrfs_super_block::sys_chunk_array
2636 *
2637 * While for dev replace, we need to try our best to mark block
2638 * group RO, to prevent race between:
2639 * - Write duplication
2640 * Contains latest data
2641 * - Scrub copy
2642 * Contains data from commit tree
2643 *
2644 * If target block group is not marked RO, nocow writes can
2645 * be overwritten by scrub copy, causing data corruption.
2646 * So for dev-replace, it's not allowed to continue if a block
2647 * group is not RO.
2648 */
2649 ret = btrfs_inc_block_group_ro(cache, sctx->is_dev_replace);
2650 if (!ret && sctx->is_dev_replace) {
2651 ret = finish_extent_writes_for_zoned(root, cache);
2652 if (ret) {
2653 btrfs_dec_block_group_ro(cache);
2654 scrub_pause_off(fs_info);
2655 btrfs_put_block_group(cache);
2656 break;
2657 }
2658 }
2659
2660 if (ret == 0) {
2661 ro_set = 1;
2662 } else if (ret == -ENOSPC && !sctx->is_dev_replace &&
2663 !(cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK)) {
2664 /*
2665 * btrfs_inc_block_group_ro return -ENOSPC when it
2666 * failed in creating new chunk for metadata.
2667 * It is not a problem for scrub, because
2668 * metadata are always cowed, and our scrub paused
2669 * commit_transactions.
2670 *
2671 * For RAID56 chunks, we have to mark them read-only
2672 * for scrub, as later we would use our own cache
2673 * out of RAID56 realm.
2674 * Thus we want the RAID56 bg to be marked RO to
2675 * prevent RMW from screwing up out cache.
2676 */
2677 ro_set = 0;
2678 } else if (ret == -ETXTBSY) {
2679 btrfs_warn(fs_info,
2680 "skipping scrub of block group %llu due to active swapfile",
2681 cache->start);
2682 scrub_pause_off(fs_info);
2683 ret = 0;
2684 goto skip_unfreeze;
2685 } else {
2686 btrfs_warn(fs_info,
2687 "failed setting block group ro: %d", ret);
2688 btrfs_unfreeze_block_group(cache);
2689 btrfs_put_block_group(cache);
2690 scrub_pause_off(fs_info);
2691 break;
2692 }
2693
2694 /*
2695 * Now the target block is marked RO, wait for nocow writes to
2696 * finish before dev-replace.
2697 * COW is fine, as COW never overwrites extents in commit tree.
2698 */
2699 if (sctx->is_dev_replace) {
2700 btrfs_wait_nocow_writers(cache);
2701 btrfs_wait_ordered_roots(fs_info, U64_MAX, cache);
2702 }
2703
2704 scrub_pause_off(fs_info);
2705 down_write(&dev_replace->rwsem);
2706 dev_replace->cursor_right = found_key.offset + dev_extent_len;
2707 dev_replace->cursor_left = found_key.offset;
2708 dev_replace->item_needs_writeback = 1;
2709 up_write(&dev_replace->rwsem);
2710
2711 ret = scrub_chunk(sctx, cache, scrub_dev, found_key.offset,
2712 dev_extent_len);
2713 if (sctx->is_dev_replace &&
2714 !btrfs_finish_block_group_to_copy(dev_replace->srcdev,
2715 cache, found_key.offset))
2716 ro_set = 0;
2717
2718 down_write(&dev_replace->rwsem);
2719 dev_replace->cursor_left = dev_replace->cursor_right;
2720 dev_replace->item_needs_writeback = 1;
2721 up_write(&dev_replace->rwsem);
2722
2723 if (ro_set)
2724 btrfs_dec_block_group_ro(cache);
2725
2726 /*
2727 * We might have prevented the cleaner kthread from deleting
2728 * this block group if it was already unused because we raced
2729 * and set it to RO mode first. So add it back to the unused
2730 * list, otherwise it might not ever be deleted unless a manual
2731 * balance is triggered or it becomes used and unused again.
2732 */
2733 spin_lock(&cache->lock);
2734 if (!test_bit(BLOCK_GROUP_FLAG_REMOVED, &cache->runtime_flags) &&
2735 !cache->ro && cache->reserved == 0 && cache->used == 0) {
2736 spin_unlock(&cache->lock);
2737 if (btrfs_test_opt(fs_info, DISCARD_ASYNC))
2738 btrfs_discard_queue_work(&fs_info->discard_ctl,
2739 cache);
2740 else
2741 btrfs_mark_bg_unused(cache);
2742 } else {
2743 spin_unlock(&cache->lock);
2744 }
2745skip_unfreeze:
2746 btrfs_unfreeze_block_group(cache);
2747 btrfs_put_block_group(cache);
2748 if (ret)
2749 break;
2750 if (sctx->is_dev_replace &&
2751 atomic64_read(&dev_replace->num_write_errors) > 0) {
2752 ret = -EIO;
2753 break;
2754 }
2755 if (sctx->stat.malloc_errors > 0) {
2756 ret = -ENOMEM;
2757 break;
2758 }
2759skip:
2760 key.offset = found_key.offset + dev_extent_len;
2761 btrfs_release_path(path);
2762 }
2763
2764 btrfs_free_path(path);
2765
2766 return ret;
2767}
2768
2769static int scrub_one_super(struct scrub_ctx *sctx, struct btrfs_device *dev,
2770 struct page *page, u64 physical, u64 generation)
2771{
2772 struct btrfs_fs_info *fs_info = sctx->fs_info;
2773 struct bio_vec bvec;
2774 struct bio bio;
2775 struct btrfs_super_block *sb = page_address(page);
2776 int ret;
2777
2778 bio_init(&bio, dev->bdev, &bvec, 1, REQ_OP_READ);
2779 bio.bi_iter.bi_sector = physical >> SECTOR_SHIFT;
2780 __bio_add_page(&bio, page, BTRFS_SUPER_INFO_SIZE, 0);
2781 ret = submit_bio_wait(&bio);
2782 bio_uninit(&bio);
2783
2784 if (ret < 0)
2785 return ret;
2786 ret = btrfs_check_super_csum(fs_info, sb);
2787 if (ret != 0) {
2788 btrfs_err_rl(fs_info,
2789 "super block at physical %llu devid %llu has bad csum",
2790 physical, dev->devid);
2791 return -EIO;
2792 }
2793 if (btrfs_super_generation(sb) != generation) {
2794 btrfs_err_rl(fs_info,
2795"super block at physical %llu devid %llu has bad generation %llu expect %llu",
2796 physical, dev->devid,
2797 btrfs_super_generation(sb), generation);
2798 return -EUCLEAN;
2799 }
2800
2801 return btrfs_validate_super(fs_info, sb, -1);
2802}
2803
2804static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
2805 struct btrfs_device *scrub_dev)
2806{
2807 int i;
2808 u64 bytenr;
2809 u64 gen;
2810 int ret = 0;
2811 struct page *page;
2812 struct btrfs_fs_info *fs_info = sctx->fs_info;
2813
2814 if (BTRFS_FS_ERROR(fs_info))
2815 return -EROFS;
2816
2817 page = alloc_page(GFP_KERNEL);
2818 if (!page) {
2819 spin_lock(&sctx->stat_lock);
2820 sctx->stat.malloc_errors++;
2821 spin_unlock(&sctx->stat_lock);
2822 return -ENOMEM;
2823 }
2824
2825 /* Seed devices of a new filesystem has their own generation. */
2826 if (scrub_dev->fs_devices != fs_info->fs_devices)
2827 gen = scrub_dev->generation;
2828 else
2829 gen = btrfs_get_last_trans_committed(fs_info);
2830
2831 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2832 ret = btrfs_sb_log_location(scrub_dev, i, 0, &bytenr);
2833 if (ret == -ENOENT)
2834 break;
2835
2836 if (ret) {
2837 spin_lock(&sctx->stat_lock);
2838 sctx->stat.super_errors++;
2839 spin_unlock(&sctx->stat_lock);
2840 continue;
2841 }
2842
2843 if (bytenr + BTRFS_SUPER_INFO_SIZE >
2844 scrub_dev->commit_total_bytes)
2845 break;
2846 if (!btrfs_check_super_location(scrub_dev, bytenr))
2847 continue;
2848
2849 ret = scrub_one_super(sctx, scrub_dev, page, bytenr, gen);
2850 if (ret) {
2851 spin_lock(&sctx->stat_lock);
2852 sctx->stat.super_errors++;
2853 spin_unlock(&sctx->stat_lock);
2854 }
2855 }
2856 __free_page(page);
2857 return 0;
2858}
2859
2860static void scrub_workers_put(struct btrfs_fs_info *fs_info)
2861{
2862 if (refcount_dec_and_mutex_lock(&fs_info->scrub_workers_refcnt,
2863 &fs_info->scrub_lock)) {
2864 struct workqueue_struct *scrub_workers = fs_info->scrub_workers;
2865
2866 fs_info->scrub_workers = NULL;
2867 mutex_unlock(&fs_info->scrub_lock);
2868
2869 if (scrub_workers)
2870 destroy_workqueue(scrub_workers);
2871 }
2872}
2873
2874/*
2875 * get a reference count on fs_info->scrub_workers. start worker if necessary
2876 */
2877static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info)
2878{
2879 struct workqueue_struct *scrub_workers = NULL;
2880 unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
2881 int max_active = fs_info->thread_pool_size;
2882 int ret = -ENOMEM;
2883
2884 if (refcount_inc_not_zero(&fs_info->scrub_workers_refcnt))
2885 return 0;
2886
2887 scrub_workers = alloc_workqueue("btrfs-scrub", flags, max_active);
2888 if (!scrub_workers)
2889 return -ENOMEM;
2890
2891 mutex_lock(&fs_info->scrub_lock);
2892 if (refcount_read(&fs_info->scrub_workers_refcnt) == 0) {
2893 ASSERT(fs_info->scrub_workers == NULL);
2894 fs_info->scrub_workers = scrub_workers;
2895 refcount_set(&fs_info->scrub_workers_refcnt, 1);
2896 mutex_unlock(&fs_info->scrub_lock);
2897 return 0;
2898 }
2899 /* Other thread raced in and created the workers for us */
2900 refcount_inc(&fs_info->scrub_workers_refcnt);
2901 mutex_unlock(&fs_info->scrub_lock);
2902
2903 ret = 0;
2904
2905 destroy_workqueue(scrub_workers);
2906 return ret;
2907}
2908
2909int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
2910 u64 end, struct btrfs_scrub_progress *progress,
2911 int readonly, int is_dev_replace)
2912{
2913 struct btrfs_dev_lookup_args args = { .devid = devid };
2914 struct scrub_ctx *sctx;
2915 int ret;
2916 struct btrfs_device *dev;
2917 unsigned int nofs_flag;
2918 bool need_commit = false;
2919
2920 if (btrfs_fs_closing(fs_info))
2921 return -EAGAIN;
2922
2923 /* At mount time we have ensured nodesize is in the range of [4K, 64K]. */
2924 ASSERT(fs_info->nodesize <= BTRFS_STRIPE_LEN);
2925
2926 /*
2927 * SCRUB_MAX_SECTORS_PER_BLOCK is calculated using the largest possible
2928 * value (max nodesize / min sectorsize), thus nodesize should always
2929 * be fine.
2930 */
2931 ASSERT(fs_info->nodesize <=
2932 SCRUB_MAX_SECTORS_PER_BLOCK << fs_info->sectorsize_bits);
2933
2934 /* Allocate outside of device_list_mutex */
2935 sctx = scrub_setup_ctx(fs_info, is_dev_replace);
2936 if (IS_ERR(sctx))
2937 return PTR_ERR(sctx);
2938
2939 ret = scrub_workers_get(fs_info);
2940 if (ret)
2941 goto out_free_ctx;
2942
2943 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2944 dev = btrfs_find_device(fs_info->fs_devices, &args);
2945 if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
2946 !is_dev_replace)) {
2947 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2948 ret = -ENODEV;
2949 goto out;
2950 }
2951
2952 if (!is_dev_replace && !readonly &&
2953 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
2954 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2955 btrfs_err_in_rcu(fs_info,
2956 "scrub on devid %llu: filesystem on %s is not writable",
2957 devid, btrfs_dev_name(dev));
2958 ret = -EROFS;
2959 goto out;
2960 }
2961
2962 mutex_lock(&fs_info->scrub_lock);
2963 if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
2964 test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
2965 mutex_unlock(&fs_info->scrub_lock);
2966 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2967 ret = -EIO;
2968 goto out;
2969 }
2970
2971 down_read(&fs_info->dev_replace.rwsem);
2972 if (dev->scrub_ctx ||
2973 (!is_dev_replace &&
2974 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
2975 up_read(&fs_info->dev_replace.rwsem);
2976 mutex_unlock(&fs_info->scrub_lock);
2977 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2978 ret = -EINPROGRESS;
2979 goto out;
2980 }
2981 up_read(&fs_info->dev_replace.rwsem);
2982
2983 sctx->readonly = readonly;
2984 dev->scrub_ctx = sctx;
2985 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2986
2987 /*
2988 * checking @scrub_pause_req here, we can avoid
2989 * race between committing transaction and scrubbing.
2990 */
2991 __scrub_blocked_if_needed(fs_info);
2992 atomic_inc(&fs_info->scrubs_running);
2993 mutex_unlock(&fs_info->scrub_lock);
2994
2995 /*
2996 * In order to avoid deadlock with reclaim when there is a transaction
2997 * trying to pause scrub, make sure we use GFP_NOFS for all the
2998 * allocations done at btrfs_scrub_sectors() and scrub_sectors_for_parity()
2999 * invoked by our callees. The pausing request is done when the
3000 * transaction commit starts, and it blocks the transaction until scrub
3001 * is paused (done at specific points at scrub_stripe() or right above
3002 * before incrementing fs_info->scrubs_running).
3003 */
3004 nofs_flag = memalloc_nofs_save();
3005 if (!is_dev_replace) {
3006 u64 old_super_errors;
3007
3008 spin_lock(&sctx->stat_lock);
3009 old_super_errors = sctx->stat.super_errors;
3010 spin_unlock(&sctx->stat_lock);
3011
3012 btrfs_info(fs_info, "scrub: started on devid %llu", devid);
3013 /*
3014 * by holding device list mutex, we can
3015 * kick off writing super in log tree sync.
3016 */
3017 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3018 ret = scrub_supers(sctx, dev);
3019 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3020
3021 spin_lock(&sctx->stat_lock);
3022 /*
3023 * Super block errors found, but we can not commit transaction
3024 * at current context, since btrfs_commit_transaction() needs
3025 * to pause the current running scrub (hold by ourselves).
3026 */
3027 if (sctx->stat.super_errors > old_super_errors && !sctx->readonly)
3028 need_commit = true;
3029 spin_unlock(&sctx->stat_lock);
3030 }
3031
3032 if (!ret)
3033 ret = scrub_enumerate_chunks(sctx, dev, start, end);
3034 memalloc_nofs_restore(nofs_flag);
3035
3036 atomic_dec(&fs_info->scrubs_running);
3037 wake_up(&fs_info->scrub_pause_wait);
3038
3039 if (progress)
3040 memcpy(progress, &sctx->stat, sizeof(*progress));
3041
3042 if (!is_dev_replace)
3043 btrfs_info(fs_info, "scrub: %s on devid %llu with status: %d",
3044 ret ? "not finished" : "finished", devid, ret);
3045
3046 mutex_lock(&fs_info->scrub_lock);
3047 dev->scrub_ctx = NULL;
3048 mutex_unlock(&fs_info->scrub_lock);
3049
3050 scrub_workers_put(fs_info);
3051 scrub_put_ctx(sctx);
3052
3053 /*
3054 * We found some super block errors before, now try to force a
3055 * transaction commit, as scrub has finished.
3056 */
3057 if (need_commit) {
3058 struct btrfs_trans_handle *trans;
3059
3060 trans = btrfs_start_transaction(fs_info->tree_root, 0);
3061 if (IS_ERR(trans)) {
3062 ret = PTR_ERR(trans);
3063 btrfs_err(fs_info,
3064 "scrub: failed to start transaction to fix super block errors: %d", ret);
3065 return ret;
3066 }
3067 ret = btrfs_commit_transaction(trans);
3068 if (ret < 0)
3069 btrfs_err(fs_info,
3070 "scrub: failed to commit transaction to fix super block errors: %d", ret);
3071 }
3072 return ret;
3073out:
3074 scrub_workers_put(fs_info);
3075out_free_ctx:
3076 scrub_free_ctx(sctx);
3077
3078 return ret;
3079}
3080
3081void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
3082{
3083 mutex_lock(&fs_info->scrub_lock);
3084 atomic_inc(&fs_info->scrub_pause_req);
3085 while (atomic_read(&fs_info->scrubs_paused) !=
3086 atomic_read(&fs_info->scrubs_running)) {
3087 mutex_unlock(&fs_info->scrub_lock);
3088 wait_event(fs_info->scrub_pause_wait,
3089 atomic_read(&fs_info->scrubs_paused) ==
3090 atomic_read(&fs_info->scrubs_running));
3091 mutex_lock(&fs_info->scrub_lock);
3092 }
3093 mutex_unlock(&fs_info->scrub_lock);
3094}
3095
3096void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
3097{
3098 atomic_dec(&fs_info->scrub_pause_req);
3099 wake_up(&fs_info->scrub_pause_wait);
3100}
3101
3102int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
3103{
3104 mutex_lock(&fs_info->scrub_lock);
3105 if (!atomic_read(&fs_info->scrubs_running)) {
3106 mutex_unlock(&fs_info->scrub_lock);
3107 return -ENOTCONN;
3108 }
3109
3110 atomic_inc(&fs_info->scrub_cancel_req);
3111 while (atomic_read(&fs_info->scrubs_running)) {
3112 mutex_unlock(&fs_info->scrub_lock);
3113 wait_event(fs_info->scrub_pause_wait,
3114 atomic_read(&fs_info->scrubs_running) == 0);
3115 mutex_lock(&fs_info->scrub_lock);
3116 }
3117 atomic_dec(&fs_info->scrub_cancel_req);
3118 mutex_unlock(&fs_info->scrub_lock);
3119
3120 return 0;
3121}
3122
3123int btrfs_scrub_cancel_dev(struct btrfs_device *dev)
3124{
3125 struct btrfs_fs_info *fs_info = dev->fs_info;
3126 struct scrub_ctx *sctx;
3127
3128 mutex_lock(&fs_info->scrub_lock);
3129 sctx = dev->scrub_ctx;
3130 if (!sctx) {
3131 mutex_unlock(&fs_info->scrub_lock);
3132 return -ENOTCONN;
3133 }
3134 atomic_inc(&sctx->cancel_req);
3135 while (dev->scrub_ctx) {
3136 mutex_unlock(&fs_info->scrub_lock);
3137 wait_event(fs_info->scrub_pause_wait,
3138 dev->scrub_ctx == NULL);
3139 mutex_lock(&fs_info->scrub_lock);
3140 }
3141 mutex_unlock(&fs_info->scrub_lock);
3142
3143 return 0;
3144}
3145
3146int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
3147 struct btrfs_scrub_progress *progress)
3148{
3149 struct btrfs_dev_lookup_args args = { .devid = devid };
3150 struct btrfs_device *dev;
3151 struct scrub_ctx *sctx = NULL;
3152
3153 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3154 dev = btrfs_find_device(fs_info->fs_devices, &args);
3155 if (dev)
3156 sctx = dev->scrub_ctx;
3157 if (sctx)
3158 memcpy(progress, &sctx->stat, sizeof(*progress));
3159 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3160
3161 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3162}
1/*
2 * Copyright (C) 2011 STRATO. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19#include <linux/blkdev.h>
20#include <linux/ratelimit.h>
21#include "ctree.h"
22#include "volumes.h"
23#include "disk-io.h"
24#include "ordered-data.h"
25#include "transaction.h"
26#include "backref.h"
27#include "extent_io.h"
28#include "check-integrity.h"
29#include "rcu-string.h"
30
31/*
32 * This is only the first step towards a full-features scrub. It reads all
33 * extent and super block and verifies the checksums. In case a bad checksum
34 * is found or the extent cannot be read, good data will be written back if
35 * any can be found.
36 *
37 * Future enhancements:
38 * - In case an unrepairable extent is encountered, track which files are
39 * affected and report them
40 * - track and record media errors, throw out bad devices
41 * - add a mode to also read unallocated space
42 */
43
44struct scrub_block;
45struct scrub_dev;
46
47#define SCRUB_PAGES_PER_BIO 16 /* 64k per bio */
48#define SCRUB_BIOS_PER_DEV 16 /* 1 MB per device in flight */
49#define SCRUB_MAX_PAGES_PER_BLOCK 16 /* 64k per node/leaf/sector */
50
51struct scrub_page {
52 struct scrub_block *sblock;
53 struct page *page;
54 struct btrfs_device *dev;
55 u64 flags; /* extent flags */
56 u64 generation;
57 u64 logical;
58 u64 physical;
59 struct {
60 unsigned int mirror_num:8;
61 unsigned int have_csum:1;
62 unsigned int io_error:1;
63 };
64 u8 csum[BTRFS_CSUM_SIZE];
65};
66
67struct scrub_bio {
68 int index;
69 struct scrub_dev *sdev;
70 struct bio *bio;
71 int err;
72 u64 logical;
73 u64 physical;
74 struct scrub_page *pagev[SCRUB_PAGES_PER_BIO];
75 int page_count;
76 int next_free;
77 struct btrfs_work work;
78};
79
80struct scrub_block {
81 struct scrub_page pagev[SCRUB_MAX_PAGES_PER_BLOCK];
82 int page_count;
83 atomic_t outstanding_pages;
84 atomic_t ref_count; /* free mem on transition to zero */
85 struct scrub_dev *sdev;
86 struct {
87 unsigned int header_error:1;
88 unsigned int checksum_error:1;
89 unsigned int no_io_error_seen:1;
90 unsigned int generation_error:1; /* also sets header_error */
91 };
92};
93
94struct scrub_dev {
95 struct scrub_bio *bios[SCRUB_BIOS_PER_DEV];
96 struct btrfs_device *dev;
97 int first_free;
98 int curr;
99 atomic_t in_flight;
100 atomic_t fixup_cnt;
101 spinlock_t list_lock;
102 wait_queue_head_t list_wait;
103 u16 csum_size;
104 struct list_head csum_list;
105 atomic_t cancel_req;
106 int readonly;
107 int pages_per_bio; /* <= SCRUB_PAGES_PER_BIO */
108 u32 sectorsize;
109 u32 nodesize;
110 u32 leafsize;
111 /*
112 * statistics
113 */
114 struct btrfs_scrub_progress stat;
115 spinlock_t stat_lock;
116};
117
118struct scrub_fixup_nodatasum {
119 struct scrub_dev *sdev;
120 u64 logical;
121 struct btrfs_root *root;
122 struct btrfs_work work;
123 int mirror_num;
124};
125
126struct scrub_warning {
127 struct btrfs_path *path;
128 u64 extent_item_size;
129 char *scratch_buf;
130 char *msg_buf;
131 const char *errstr;
132 sector_t sector;
133 u64 logical;
134 struct btrfs_device *dev;
135 int msg_bufsize;
136 int scratch_bufsize;
137};
138
139
140static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
141static int scrub_setup_recheck_block(struct scrub_dev *sdev,
142 struct btrfs_mapping_tree *map_tree,
143 u64 length, u64 logical,
144 struct scrub_block *sblock);
145static int scrub_recheck_block(struct btrfs_fs_info *fs_info,
146 struct scrub_block *sblock, int is_metadata,
147 int have_csum, u8 *csum, u64 generation,
148 u16 csum_size);
149static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
150 struct scrub_block *sblock,
151 int is_metadata, int have_csum,
152 const u8 *csum, u64 generation,
153 u16 csum_size);
154static void scrub_complete_bio_end_io(struct bio *bio, int err);
155static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
156 struct scrub_block *sblock_good,
157 int force_write);
158static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
159 struct scrub_block *sblock_good,
160 int page_num, int force_write);
161static int scrub_checksum_data(struct scrub_block *sblock);
162static int scrub_checksum_tree_block(struct scrub_block *sblock);
163static int scrub_checksum_super(struct scrub_block *sblock);
164static void scrub_block_get(struct scrub_block *sblock);
165static void scrub_block_put(struct scrub_block *sblock);
166static int scrub_add_page_to_bio(struct scrub_dev *sdev,
167 struct scrub_page *spage);
168static int scrub_pages(struct scrub_dev *sdev, u64 logical, u64 len,
169 u64 physical, u64 flags, u64 gen, int mirror_num,
170 u8 *csum, int force);
171static void scrub_bio_end_io(struct bio *bio, int err);
172static void scrub_bio_end_io_worker(struct btrfs_work *work);
173static void scrub_block_complete(struct scrub_block *sblock);
174
175
176static void scrub_free_csums(struct scrub_dev *sdev)
177{
178 while (!list_empty(&sdev->csum_list)) {
179 struct btrfs_ordered_sum *sum;
180 sum = list_first_entry(&sdev->csum_list,
181 struct btrfs_ordered_sum, list);
182 list_del(&sum->list);
183 kfree(sum);
184 }
185}
186
187static noinline_for_stack void scrub_free_dev(struct scrub_dev *sdev)
188{
189 int i;
190
191 if (!sdev)
192 return;
193
194 /* this can happen when scrub is cancelled */
195 if (sdev->curr != -1) {
196 struct scrub_bio *sbio = sdev->bios[sdev->curr];
197
198 for (i = 0; i < sbio->page_count; i++) {
199 BUG_ON(!sbio->pagev[i]);
200 BUG_ON(!sbio->pagev[i]->page);
201 scrub_block_put(sbio->pagev[i]->sblock);
202 }
203 bio_put(sbio->bio);
204 }
205
206 for (i = 0; i < SCRUB_BIOS_PER_DEV; ++i) {
207 struct scrub_bio *sbio = sdev->bios[i];
208
209 if (!sbio)
210 break;
211 kfree(sbio);
212 }
213
214 scrub_free_csums(sdev);
215 kfree(sdev);
216}
217
218static noinline_for_stack
219struct scrub_dev *scrub_setup_dev(struct btrfs_device *dev)
220{
221 struct scrub_dev *sdev;
222 int i;
223 struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
224 int pages_per_bio;
225
226 pages_per_bio = min_t(int, SCRUB_PAGES_PER_BIO,
227 bio_get_nr_vecs(dev->bdev));
228 sdev = kzalloc(sizeof(*sdev), GFP_NOFS);
229 if (!sdev)
230 goto nomem;
231 sdev->dev = dev;
232 sdev->pages_per_bio = pages_per_bio;
233 sdev->curr = -1;
234 for (i = 0; i < SCRUB_BIOS_PER_DEV; ++i) {
235 struct scrub_bio *sbio;
236
237 sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
238 if (!sbio)
239 goto nomem;
240 sdev->bios[i] = sbio;
241
242 sbio->index = i;
243 sbio->sdev = sdev;
244 sbio->page_count = 0;
245 sbio->work.func = scrub_bio_end_io_worker;
246
247 if (i != SCRUB_BIOS_PER_DEV-1)
248 sdev->bios[i]->next_free = i + 1;
249 else
250 sdev->bios[i]->next_free = -1;
251 }
252 sdev->first_free = 0;
253 sdev->nodesize = dev->dev_root->nodesize;
254 sdev->leafsize = dev->dev_root->leafsize;
255 sdev->sectorsize = dev->dev_root->sectorsize;
256 atomic_set(&sdev->in_flight, 0);
257 atomic_set(&sdev->fixup_cnt, 0);
258 atomic_set(&sdev->cancel_req, 0);
259 sdev->csum_size = btrfs_super_csum_size(fs_info->super_copy);
260 INIT_LIST_HEAD(&sdev->csum_list);
261
262 spin_lock_init(&sdev->list_lock);
263 spin_lock_init(&sdev->stat_lock);
264 init_waitqueue_head(&sdev->list_wait);
265 return sdev;
266
267nomem:
268 scrub_free_dev(sdev);
269 return ERR_PTR(-ENOMEM);
270}
271
272static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root, void *ctx)
273{
274 u64 isize;
275 u32 nlink;
276 int ret;
277 int i;
278 struct extent_buffer *eb;
279 struct btrfs_inode_item *inode_item;
280 struct scrub_warning *swarn = ctx;
281 struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
282 struct inode_fs_paths *ipath = NULL;
283 struct btrfs_root *local_root;
284 struct btrfs_key root_key;
285
286 root_key.objectid = root;
287 root_key.type = BTRFS_ROOT_ITEM_KEY;
288 root_key.offset = (u64)-1;
289 local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
290 if (IS_ERR(local_root)) {
291 ret = PTR_ERR(local_root);
292 goto err;
293 }
294
295 ret = inode_item_info(inum, 0, local_root, swarn->path);
296 if (ret) {
297 btrfs_release_path(swarn->path);
298 goto err;
299 }
300
301 eb = swarn->path->nodes[0];
302 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
303 struct btrfs_inode_item);
304 isize = btrfs_inode_size(eb, inode_item);
305 nlink = btrfs_inode_nlink(eb, inode_item);
306 btrfs_release_path(swarn->path);
307
308 ipath = init_ipath(4096, local_root, swarn->path);
309 if (IS_ERR(ipath)) {
310 ret = PTR_ERR(ipath);
311 ipath = NULL;
312 goto err;
313 }
314 ret = paths_from_inode(inum, ipath);
315
316 if (ret < 0)
317 goto err;
318
319 /*
320 * we deliberately ignore the bit ipath might have been too small to
321 * hold all of the paths here
322 */
323 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
324 printk_in_rcu(KERN_WARNING "btrfs: %s at logical %llu on dev "
325 "%s, sector %llu, root %llu, inode %llu, offset %llu, "
326 "length %llu, links %u (path: %s)\n", swarn->errstr,
327 swarn->logical, rcu_str_deref(swarn->dev->name),
328 (unsigned long long)swarn->sector, root, inum, offset,
329 min(isize - offset, (u64)PAGE_SIZE), nlink,
330 (char *)(unsigned long)ipath->fspath->val[i]);
331
332 free_ipath(ipath);
333 return 0;
334
335err:
336 printk_in_rcu(KERN_WARNING "btrfs: %s at logical %llu on dev "
337 "%s, sector %llu, root %llu, inode %llu, offset %llu: path "
338 "resolving failed with ret=%d\n", swarn->errstr,
339 swarn->logical, rcu_str_deref(swarn->dev->name),
340 (unsigned long long)swarn->sector, root, inum, offset, ret);
341
342 free_ipath(ipath);
343 return 0;
344}
345
346static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
347{
348 struct btrfs_device *dev = sblock->sdev->dev;
349 struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
350 struct btrfs_path *path;
351 struct btrfs_key found_key;
352 struct extent_buffer *eb;
353 struct btrfs_extent_item *ei;
354 struct scrub_warning swarn;
355 u32 item_size;
356 int ret;
357 u64 ref_root;
358 u8 ref_level;
359 unsigned long ptr = 0;
360 const int bufsize = 4096;
361 u64 extent_item_pos;
362
363 path = btrfs_alloc_path();
364
365 swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
366 swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
367 BUG_ON(sblock->page_count < 1);
368 swarn.sector = (sblock->pagev[0].physical) >> 9;
369 swarn.logical = sblock->pagev[0].logical;
370 swarn.errstr = errstr;
371 swarn.dev = dev;
372 swarn.msg_bufsize = bufsize;
373 swarn.scratch_bufsize = bufsize;
374
375 if (!path || !swarn.scratch_buf || !swarn.msg_buf)
376 goto out;
377
378 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key);
379 if (ret < 0)
380 goto out;
381
382 extent_item_pos = swarn.logical - found_key.objectid;
383 swarn.extent_item_size = found_key.offset;
384
385 eb = path->nodes[0];
386 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
387 item_size = btrfs_item_size_nr(eb, path->slots[0]);
388 btrfs_release_path(path);
389
390 if (ret & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
391 do {
392 ret = tree_backref_for_extent(&ptr, eb, ei, item_size,
393 &ref_root, &ref_level);
394 printk_in_rcu(KERN_WARNING
395 "btrfs: %s at logical %llu on dev %s, "
396 "sector %llu: metadata %s (level %d) in tree "
397 "%llu\n", errstr, swarn.logical,
398 rcu_str_deref(dev->name),
399 (unsigned long long)swarn.sector,
400 ref_level ? "node" : "leaf",
401 ret < 0 ? -1 : ref_level,
402 ret < 0 ? -1 : ref_root);
403 } while (ret != 1);
404 } else {
405 swarn.path = path;
406 iterate_extent_inodes(fs_info, found_key.objectid,
407 extent_item_pos, 1,
408 scrub_print_warning_inode, &swarn);
409 }
410
411out:
412 btrfs_free_path(path);
413 kfree(swarn.scratch_buf);
414 kfree(swarn.msg_buf);
415}
416
417static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *ctx)
418{
419 struct page *page = NULL;
420 unsigned long index;
421 struct scrub_fixup_nodatasum *fixup = ctx;
422 int ret;
423 int corrected = 0;
424 struct btrfs_key key;
425 struct inode *inode = NULL;
426 u64 end = offset + PAGE_SIZE - 1;
427 struct btrfs_root *local_root;
428
429 key.objectid = root;
430 key.type = BTRFS_ROOT_ITEM_KEY;
431 key.offset = (u64)-1;
432 local_root = btrfs_read_fs_root_no_name(fixup->root->fs_info, &key);
433 if (IS_ERR(local_root))
434 return PTR_ERR(local_root);
435
436 key.type = BTRFS_INODE_ITEM_KEY;
437 key.objectid = inum;
438 key.offset = 0;
439 inode = btrfs_iget(fixup->root->fs_info->sb, &key, local_root, NULL);
440 if (IS_ERR(inode))
441 return PTR_ERR(inode);
442
443 index = offset >> PAGE_CACHE_SHIFT;
444
445 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
446 if (!page) {
447 ret = -ENOMEM;
448 goto out;
449 }
450
451 if (PageUptodate(page)) {
452 struct btrfs_mapping_tree *map_tree;
453 if (PageDirty(page)) {
454 /*
455 * we need to write the data to the defect sector. the
456 * data that was in that sector is not in memory,
457 * because the page was modified. we must not write the
458 * modified page to that sector.
459 *
460 * TODO: what could be done here: wait for the delalloc
461 * runner to write out that page (might involve
462 * COW) and see whether the sector is still
463 * referenced afterwards.
464 *
465 * For the meantime, we'll treat this error
466 * incorrectable, although there is a chance that a
467 * later scrub will find the bad sector again and that
468 * there's no dirty page in memory, then.
469 */
470 ret = -EIO;
471 goto out;
472 }
473 map_tree = &BTRFS_I(inode)->root->fs_info->mapping_tree;
474 ret = repair_io_failure(map_tree, offset, PAGE_SIZE,
475 fixup->logical, page,
476 fixup->mirror_num);
477 unlock_page(page);
478 corrected = !ret;
479 } else {
480 /*
481 * we need to get good data first. the general readpage path
482 * will call repair_io_failure for us, we just have to make
483 * sure we read the bad mirror.
484 */
485 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
486 EXTENT_DAMAGED, GFP_NOFS);
487 if (ret) {
488 /* set_extent_bits should give proper error */
489 WARN_ON(ret > 0);
490 if (ret > 0)
491 ret = -EFAULT;
492 goto out;
493 }
494
495 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
496 btrfs_get_extent,
497 fixup->mirror_num);
498 wait_on_page_locked(page);
499
500 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
501 end, EXTENT_DAMAGED, 0, NULL);
502 if (!corrected)
503 clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
504 EXTENT_DAMAGED, GFP_NOFS);
505 }
506
507out:
508 if (page)
509 put_page(page);
510 if (inode)
511 iput(inode);
512
513 if (ret < 0)
514 return ret;
515
516 if (ret == 0 && corrected) {
517 /*
518 * we only need to call readpage for one of the inodes belonging
519 * to this extent. so make iterate_extent_inodes stop
520 */
521 return 1;
522 }
523
524 return -EIO;
525}
526
527static void scrub_fixup_nodatasum(struct btrfs_work *work)
528{
529 int ret;
530 struct scrub_fixup_nodatasum *fixup;
531 struct scrub_dev *sdev;
532 struct btrfs_trans_handle *trans = NULL;
533 struct btrfs_fs_info *fs_info;
534 struct btrfs_path *path;
535 int uncorrectable = 0;
536
537 fixup = container_of(work, struct scrub_fixup_nodatasum, work);
538 sdev = fixup->sdev;
539 fs_info = fixup->root->fs_info;
540
541 path = btrfs_alloc_path();
542 if (!path) {
543 spin_lock(&sdev->stat_lock);
544 ++sdev->stat.malloc_errors;
545 spin_unlock(&sdev->stat_lock);
546 uncorrectable = 1;
547 goto out;
548 }
549
550 trans = btrfs_join_transaction(fixup->root);
551 if (IS_ERR(trans)) {
552 uncorrectable = 1;
553 goto out;
554 }
555
556 /*
557 * the idea is to trigger a regular read through the standard path. we
558 * read a page from the (failed) logical address by specifying the
559 * corresponding copynum of the failed sector. thus, that readpage is
560 * expected to fail.
561 * that is the point where on-the-fly error correction will kick in
562 * (once it's finished) and rewrite the failed sector if a good copy
563 * can be found.
564 */
565 ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
566 path, scrub_fixup_readpage,
567 fixup);
568 if (ret < 0) {
569 uncorrectable = 1;
570 goto out;
571 }
572 WARN_ON(ret != 1);
573
574 spin_lock(&sdev->stat_lock);
575 ++sdev->stat.corrected_errors;
576 spin_unlock(&sdev->stat_lock);
577
578out:
579 if (trans && !IS_ERR(trans))
580 btrfs_end_transaction(trans, fixup->root);
581 if (uncorrectable) {
582 spin_lock(&sdev->stat_lock);
583 ++sdev->stat.uncorrectable_errors;
584 spin_unlock(&sdev->stat_lock);
585
586 printk_ratelimited_in_rcu(KERN_ERR
587 "btrfs: unable to fixup (nodatasum) error at logical %llu on dev %s\n",
588 (unsigned long long)fixup->logical,
589 rcu_str_deref(sdev->dev->name));
590 }
591
592 btrfs_free_path(path);
593 kfree(fixup);
594
595 /* see caller why we're pretending to be paused in the scrub counters */
596 mutex_lock(&fs_info->scrub_lock);
597 atomic_dec(&fs_info->scrubs_running);
598 atomic_dec(&fs_info->scrubs_paused);
599 mutex_unlock(&fs_info->scrub_lock);
600 atomic_dec(&sdev->fixup_cnt);
601 wake_up(&fs_info->scrub_pause_wait);
602 wake_up(&sdev->list_wait);
603}
604
605/*
606 * scrub_handle_errored_block gets called when either verification of the
607 * pages failed or the bio failed to read, e.g. with EIO. In the latter
608 * case, this function handles all pages in the bio, even though only one
609 * may be bad.
610 * The goal of this function is to repair the errored block by using the
611 * contents of one of the mirrors.
612 */
613static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
614{
615 struct scrub_dev *sdev = sblock_to_check->sdev;
616 struct btrfs_fs_info *fs_info;
617 u64 length;
618 u64 logical;
619 u64 generation;
620 unsigned int failed_mirror_index;
621 unsigned int is_metadata;
622 unsigned int have_csum;
623 u8 *csum;
624 struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
625 struct scrub_block *sblock_bad;
626 int ret;
627 int mirror_index;
628 int page_num;
629 int success;
630 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
631 DEFAULT_RATELIMIT_BURST);
632
633 BUG_ON(sblock_to_check->page_count < 1);
634 fs_info = sdev->dev->dev_root->fs_info;
635 length = sblock_to_check->page_count * PAGE_SIZE;
636 logical = sblock_to_check->pagev[0].logical;
637 generation = sblock_to_check->pagev[0].generation;
638 BUG_ON(sblock_to_check->pagev[0].mirror_num < 1);
639 failed_mirror_index = sblock_to_check->pagev[0].mirror_num - 1;
640 is_metadata = !(sblock_to_check->pagev[0].flags &
641 BTRFS_EXTENT_FLAG_DATA);
642 have_csum = sblock_to_check->pagev[0].have_csum;
643 csum = sblock_to_check->pagev[0].csum;
644
645 /*
646 * read all mirrors one after the other. This includes to
647 * re-read the extent or metadata block that failed (that was
648 * the cause that this fixup code is called) another time,
649 * page by page this time in order to know which pages
650 * caused I/O errors and which ones are good (for all mirrors).
651 * It is the goal to handle the situation when more than one
652 * mirror contains I/O errors, but the errors do not
653 * overlap, i.e. the data can be repaired by selecting the
654 * pages from those mirrors without I/O error on the
655 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
656 * would be that mirror #1 has an I/O error on the first page,
657 * the second page is good, and mirror #2 has an I/O error on
658 * the second page, but the first page is good.
659 * Then the first page of the first mirror can be repaired by
660 * taking the first page of the second mirror, and the
661 * second page of the second mirror can be repaired by
662 * copying the contents of the 2nd page of the 1st mirror.
663 * One more note: if the pages of one mirror contain I/O
664 * errors, the checksum cannot be verified. In order to get
665 * the best data for repairing, the first attempt is to find
666 * a mirror without I/O errors and with a validated checksum.
667 * Only if this is not possible, the pages are picked from
668 * mirrors with I/O errors without considering the checksum.
669 * If the latter is the case, at the end, the checksum of the
670 * repaired area is verified in order to correctly maintain
671 * the statistics.
672 */
673
674 sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
675 sizeof(*sblocks_for_recheck),
676 GFP_NOFS);
677 if (!sblocks_for_recheck) {
678 spin_lock(&sdev->stat_lock);
679 sdev->stat.malloc_errors++;
680 sdev->stat.read_errors++;
681 sdev->stat.uncorrectable_errors++;
682 spin_unlock(&sdev->stat_lock);
683 btrfs_dev_stat_inc_and_print(sdev->dev,
684 BTRFS_DEV_STAT_READ_ERRS);
685 goto out;
686 }
687
688 /* setup the context, map the logical blocks and alloc the pages */
689 ret = scrub_setup_recheck_block(sdev, &fs_info->mapping_tree, length,
690 logical, sblocks_for_recheck);
691 if (ret) {
692 spin_lock(&sdev->stat_lock);
693 sdev->stat.read_errors++;
694 sdev->stat.uncorrectable_errors++;
695 spin_unlock(&sdev->stat_lock);
696 btrfs_dev_stat_inc_and_print(sdev->dev,
697 BTRFS_DEV_STAT_READ_ERRS);
698 goto out;
699 }
700 BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
701 sblock_bad = sblocks_for_recheck + failed_mirror_index;
702
703 /* build and submit the bios for the failed mirror, check checksums */
704 ret = scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
705 csum, generation, sdev->csum_size);
706 if (ret) {
707 spin_lock(&sdev->stat_lock);
708 sdev->stat.read_errors++;
709 sdev->stat.uncorrectable_errors++;
710 spin_unlock(&sdev->stat_lock);
711 btrfs_dev_stat_inc_and_print(sdev->dev,
712 BTRFS_DEV_STAT_READ_ERRS);
713 goto out;
714 }
715
716 if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
717 sblock_bad->no_io_error_seen) {
718 /*
719 * the error disappeared after reading page by page, or
720 * the area was part of a huge bio and other parts of the
721 * bio caused I/O errors, or the block layer merged several
722 * read requests into one and the error is caused by a
723 * different bio (usually one of the two latter cases is
724 * the cause)
725 */
726 spin_lock(&sdev->stat_lock);
727 sdev->stat.unverified_errors++;
728 spin_unlock(&sdev->stat_lock);
729
730 goto out;
731 }
732
733 if (!sblock_bad->no_io_error_seen) {
734 spin_lock(&sdev->stat_lock);
735 sdev->stat.read_errors++;
736 spin_unlock(&sdev->stat_lock);
737 if (__ratelimit(&_rs))
738 scrub_print_warning("i/o error", sblock_to_check);
739 btrfs_dev_stat_inc_and_print(sdev->dev,
740 BTRFS_DEV_STAT_READ_ERRS);
741 } else if (sblock_bad->checksum_error) {
742 spin_lock(&sdev->stat_lock);
743 sdev->stat.csum_errors++;
744 spin_unlock(&sdev->stat_lock);
745 if (__ratelimit(&_rs))
746 scrub_print_warning("checksum error", sblock_to_check);
747 btrfs_dev_stat_inc_and_print(sdev->dev,
748 BTRFS_DEV_STAT_CORRUPTION_ERRS);
749 } else if (sblock_bad->header_error) {
750 spin_lock(&sdev->stat_lock);
751 sdev->stat.verify_errors++;
752 spin_unlock(&sdev->stat_lock);
753 if (__ratelimit(&_rs))
754 scrub_print_warning("checksum/header error",
755 sblock_to_check);
756 if (sblock_bad->generation_error)
757 btrfs_dev_stat_inc_and_print(sdev->dev,
758 BTRFS_DEV_STAT_GENERATION_ERRS);
759 else
760 btrfs_dev_stat_inc_and_print(sdev->dev,
761 BTRFS_DEV_STAT_CORRUPTION_ERRS);
762 }
763
764 if (sdev->readonly)
765 goto did_not_correct_error;
766
767 if (!is_metadata && !have_csum) {
768 struct scrub_fixup_nodatasum *fixup_nodatasum;
769
770 /*
771 * !is_metadata and !have_csum, this means that the data
772 * might not be COW'ed, that it might be modified
773 * concurrently. The general strategy to work on the
774 * commit root does not help in the case when COW is not
775 * used.
776 */
777 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
778 if (!fixup_nodatasum)
779 goto did_not_correct_error;
780 fixup_nodatasum->sdev = sdev;
781 fixup_nodatasum->logical = logical;
782 fixup_nodatasum->root = fs_info->extent_root;
783 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
784 /*
785 * increment scrubs_running to prevent cancel requests from
786 * completing as long as a fixup worker is running. we must also
787 * increment scrubs_paused to prevent deadlocking on pause
788 * requests used for transactions commits (as the worker uses a
789 * transaction context). it is safe to regard the fixup worker
790 * as paused for all matters practical. effectively, we only
791 * avoid cancellation requests from completing.
792 */
793 mutex_lock(&fs_info->scrub_lock);
794 atomic_inc(&fs_info->scrubs_running);
795 atomic_inc(&fs_info->scrubs_paused);
796 mutex_unlock(&fs_info->scrub_lock);
797 atomic_inc(&sdev->fixup_cnt);
798 fixup_nodatasum->work.func = scrub_fixup_nodatasum;
799 btrfs_queue_worker(&fs_info->scrub_workers,
800 &fixup_nodatasum->work);
801 goto out;
802 }
803
804 /*
805 * now build and submit the bios for the other mirrors, check
806 * checksums
807 */
808 for (mirror_index = 0;
809 mirror_index < BTRFS_MAX_MIRRORS &&
810 sblocks_for_recheck[mirror_index].page_count > 0;
811 mirror_index++) {
812 if (mirror_index == failed_mirror_index)
813 continue;
814
815 /* build and submit the bios, check checksums */
816 ret = scrub_recheck_block(fs_info,
817 sblocks_for_recheck + mirror_index,
818 is_metadata, have_csum, csum,
819 generation, sdev->csum_size);
820 if (ret)
821 goto did_not_correct_error;
822 }
823
824 /*
825 * first try to pick the mirror which is completely without I/O
826 * errors and also does not have a checksum error.
827 * If one is found, and if a checksum is present, the full block
828 * that is known to contain an error is rewritten. Afterwards
829 * the block is known to be corrected.
830 * If a mirror is found which is completely correct, and no
831 * checksum is present, only those pages are rewritten that had
832 * an I/O error in the block to be repaired, since it cannot be
833 * determined, which copy of the other pages is better (and it
834 * could happen otherwise that a correct page would be
835 * overwritten by a bad one).
836 */
837 for (mirror_index = 0;
838 mirror_index < BTRFS_MAX_MIRRORS &&
839 sblocks_for_recheck[mirror_index].page_count > 0;
840 mirror_index++) {
841 struct scrub_block *sblock_other = sblocks_for_recheck +
842 mirror_index;
843
844 if (!sblock_other->header_error &&
845 !sblock_other->checksum_error &&
846 sblock_other->no_io_error_seen) {
847 int force_write = is_metadata || have_csum;
848
849 ret = scrub_repair_block_from_good_copy(sblock_bad,
850 sblock_other,
851 force_write);
852 if (0 == ret)
853 goto corrected_error;
854 }
855 }
856
857 /*
858 * in case of I/O errors in the area that is supposed to be
859 * repaired, continue by picking good copies of those pages.
860 * Select the good pages from mirrors to rewrite bad pages from
861 * the area to fix. Afterwards verify the checksum of the block
862 * that is supposed to be repaired. This verification step is
863 * only done for the purpose of statistic counting and for the
864 * final scrub report, whether errors remain.
865 * A perfect algorithm could make use of the checksum and try
866 * all possible combinations of pages from the different mirrors
867 * until the checksum verification succeeds. For example, when
868 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
869 * of mirror #2 is readable but the final checksum test fails,
870 * then the 2nd page of mirror #3 could be tried, whether now
871 * the final checksum succeedes. But this would be a rare
872 * exception and is therefore not implemented. At least it is
873 * avoided that the good copy is overwritten.
874 * A more useful improvement would be to pick the sectors
875 * without I/O error based on sector sizes (512 bytes on legacy
876 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
877 * mirror could be repaired by taking 512 byte of a different
878 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
879 * area are unreadable.
880 */
881
882 /* can only fix I/O errors from here on */
883 if (sblock_bad->no_io_error_seen)
884 goto did_not_correct_error;
885
886 success = 1;
887 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
888 struct scrub_page *page_bad = sblock_bad->pagev + page_num;
889
890 if (!page_bad->io_error)
891 continue;
892
893 for (mirror_index = 0;
894 mirror_index < BTRFS_MAX_MIRRORS &&
895 sblocks_for_recheck[mirror_index].page_count > 0;
896 mirror_index++) {
897 struct scrub_block *sblock_other = sblocks_for_recheck +
898 mirror_index;
899 struct scrub_page *page_other = sblock_other->pagev +
900 page_num;
901
902 if (!page_other->io_error) {
903 ret = scrub_repair_page_from_good_copy(
904 sblock_bad, sblock_other, page_num, 0);
905 if (0 == ret) {
906 page_bad->io_error = 0;
907 break; /* succeeded for this page */
908 }
909 }
910 }
911
912 if (page_bad->io_error) {
913 /* did not find a mirror to copy the page from */
914 success = 0;
915 }
916 }
917
918 if (success) {
919 if (is_metadata || have_csum) {
920 /*
921 * need to verify the checksum now that all
922 * sectors on disk are repaired (the write
923 * request for data to be repaired is on its way).
924 * Just be lazy and use scrub_recheck_block()
925 * which re-reads the data before the checksum
926 * is verified, but most likely the data comes out
927 * of the page cache.
928 */
929 ret = scrub_recheck_block(fs_info, sblock_bad,
930 is_metadata, have_csum, csum,
931 generation, sdev->csum_size);
932 if (!ret && !sblock_bad->header_error &&
933 !sblock_bad->checksum_error &&
934 sblock_bad->no_io_error_seen)
935 goto corrected_error;
936 else
937 goto did_not_correct_error;
938 } else {
939corrected_error:
940 spin_lock(&sdev->stat_lock);
941 sdev->stat.corrected_errors++;
942 spin_unlock(&sdev->stat_lock);
943 printk_ratelimited_in_rcu(KERN_ERR
944 "btrfs: fixed up error at logical %llu on dev %s\n",
945 (unsigned long long)logical,
946 rcu_str_deref(sdev->dev->name));
947 }
948 } else {
949did_not_correct_error:
950 spin_lock(&sdev->stat_lock);
951 sdev->stat.uncorrectable_errors++;
952 spin_unlock(&sdev->stat_lock);
953 printk_ratelimited_in_rcu(KERN_ERR
954 "btrfs: unable to fixup (regular) error at logical %llu on dev %s\n",
955 (unsigned long long)logical,
956 rcu_str_deref(sdev->dev->name));
957 }
958
959out:
960 if (sblocks_for_recheck) {
961 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
962 mirror_index++) {
963 struct scrub_block *sblock = sblocks_for_recheck +
964 mirror_index;
965 int page_index;
966
967 for (page_index = 0; page_index < SCRUB_PAGES_PER_BIO;
968 page_index++)
969 if (sblock->pagev[page_index].page)
970 __free_page(
971 sblock->pagev[page_index].page);
972 }
973 kfree(sblocks_for_recheck);
974 }
975
976 return 0;
977}
978
979static int scrub_setup_recheck_block(struct scrub_dev *sdev,
980 struct btrfs_mapping_tree *map_tree,
981 u64 length, u64 logical,
982 struct scrub_block *sblocks_for_recheck)
983{
984 int page_index;
985 int mirror_index;
986 int ret;
987
988 /*
989 * note: the three members sdev, ref_count and outstanding_pages
990 * are not used (and not set) in the blocks that are used for
991 * the recheck procedure
992 */
993
994 page_index = 0;
995 while (length > 0) {
996 u64 sublen = min_t(u64, length, PAGE_SIZE);
997 u64 mapped_length = sublen;
998 struct btrfs_bio *bbio = NULL;
999
1000 /*
1001 * with a length of PAGE_SIZE, each returned stripe
1002 * represents one mirror
1003 */
1004 ret = btrfs_map_block(map_tree, WRITE, logical, &mapped_length,
1005 &bbio, 0);
1006 if (ret || !bbio || mapped_length < sublen) {
1007 kfree(bbio);
1008 return -EIO;
1009 }
1010
1011 BUG_ON(page_index >= SCRUB_PAGES_PER_BIO);
1012 for (mirror_index = 0; mirror_index < (int)bbio->num_stripes;
1013 mirror_index++) {
1014 struct scrub_block *sblock;
1015 struct scrub_page *page;
1016
1017 if (mirror_index >= BTRFS_MAX_MIRRORS)
1018 continue;
1019
1020 sblock = sblocks_for_recheck + mirror_index;
1021 page = sblock->pagev + page_index;
1022 page->logical = logical;
1023 page->physical = bbio->stripes[mirror_index].physical;
1024 /* for missing devices, dev->bdev is NULL */
1025 page->dev = bbio->stripes[mirror_index].dev;
1026 page->mirror_num = mirror_index + 1;
1027 page->page = alloc_page(GFP_NOFS);
1028 if (!page->page) {
1029 spin_lock(&sdev->stat_lock);
1030 sdev->stat.malloc_errors++;
1031 spin_unlock(&sdev->stat_lock);
1032 return -ENOMEM;
1033 }
1034 sblock->page_count++;
1035 }
1036 kfree(bbio);
1037 length -= sublen;
1038 logical += sublen;
1039 page_index++;
1040 }
1041
1042 return 0;
1043}
1044
1045/*
1046 * this function will check the on disk data for checksum errors, header
1047 * errors and read I/O errors. If any I/O errors happen, the exact pages
1048 * which are errored are marked as being bad. The goal is to enable scrub
1049 * to take those pages that are not errored from all the mirrors so that
1050 * the pages that are errored in the just handled mirror can be repaired.
1051 */
1052static int scrub_recheck_block(struct btrfs_fs_info *fs_info,
1053 struct scrub_block *sblock, int is_metadata,
1054 int have_csum, u8 *csum, u64 generation,
1055 u16 csum_size)
1056{
1057 int page_num;
1058
1059 sblock->no_io_error_seen = 1;
1060 sblock->header_error = 0;
1061 sblock->checksum_error = 0;
1062
1063 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1064 struct bio *bio;
1065 int ret;
1066 struct scrub_page *page = sblock->pagev + page_num;
1067 DECLARE_COMPLETION_ONSTACK(complete);
1068
1069 if (page->dev->bdev == NULL) {
1070 page->io_error = 1;
1071 sblock->no_io_error_seen = 0;
1072 continue;
1073 }
1074
1075 BUG_ON(!page->page);
1076 bio = bio_alloc(GFP_NOFS, 1);
1077 if (!bio)
1078 return -EIO;
1079 bio->bi_bdev = page->dev->bdev;
1080 bio->bi_sector = page->physical >> 9;
1081 bio->bi_end_io = scrub_complete_bio_end_io;
1082 bio->bi_private = &complete;
1083
1084 ret = bio_add_page(bio, page->page, PAGE_SIZE, 0);
1085 if (PAGE_SIZE != ret) {
1086 bio_put(bio);
1087 return -EIO;
1088 }
1089 btrfsic_submit_bio(READ, bio);
1090
1091 /* this will also unplug the queue */
1092 wait_for_completion(&complete);
1093
1094 page->io_error = !test_bit(BIO_UPTODATE, &bio->bi_flags);
1095 if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1096 sblock->no_io_error_seen = 0;
1097 bio_put(bio);
1098 }
1099
1100 if (sblock->no_io_error_seen)
1101 scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
1102 have_csum, csum, generation,
1103 csum_size);
1104
1105 return 0;
1106}
1107
1108static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1109 struct scrub_block *sblock,
1110 int is_metadata, int have_csum,
1111 const u8 *csum, u64 generation,
1112 u16 csum_size)
1113{
1114 int page_num;
1115 u8 calculated_csum[BTRFS_CSUM_SIZE];
1116 u32 crc = ~(u32)0;
1117 struct btrfs_root *root = fs_info->extent_root;
1118 void *mapped_buffer;
1119
1120 BUG_ON(!sblock->pagev[0].page);
1121 if (is_metadata) {
1122 struct btrfs_header *h;
1123
1124 mapped_buffer = kmap_atomic(sblock->pagev[0].page);
1125 h = (struct btrfs_header *)mapped_buffer;
1126
1127 if (sblock->pagev[0].logical != le64_to_cpu(h->bytenr) ||
1128 memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) ||
1129 memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1130 BTRFS_UUID_SIZE)) {
1131 sblock->header_error = 1;
1132 } else if (generation != le64_to_cpu(h->generation)) {
1133 sblock->header_error = 1;
1134 sblock->generation_error = 1;
1135 }
1136 csum = h->csum;
1137 } else {
1138 if (!have_csum)
1139 return;
1140
1141 mapped_buffer = kmap_atomic(sblock->pagev[0].page);
1142 }
1143
1144 for (page_num = 0;;) {
1145 if (page_num == 0 && is_metadata)
1146 crc = btrfs_csum_data(root,
1147 ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1148 crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1149 else
1150 crc = btrfs_csum_data(root, mapped_buffer, crc,
1151 PAGE_SIZE);
1152
1153 kunmap_atomic(mapped_buffer);
1154 page_num++;
1155 if (page_num >= sblock->page_count)
1156 break;
1157 BUG_ON(!sblock->pagev[page_num].page);
1158
1159 mapped_buffer = kmap_atomic(sblock->pagev[page_num].page);
1160 }
1161
1162 btrfs_csum_final(crc, calculated_csum);
1163 if (memcmp(calculated_csum, csum, csum_size))
1164 sblock->checksum_error = 1;
1165}
1166
1167static void scrub_complete_bio_end_io(struct bio *bio, int err)
1168{
1169 complete((struct completion *)bio->bi_private);
1170}
1171
1172static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1173 struct scrub_block *sblock_good,
1174 int force_write)
1175{
1176 int page_num;
1177 int ret = 0;
1178
1179 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1180 int ret_sub;
1181
1182 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1183 sblock_good,
1184 page_num,
1185 force_write);
1186 if (ret_sub)
1187 ret = ret_sub;
1188 }
1189
1190 return ret;
1191}
1192
1193static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1194 struct scrub_block *sblock_good,
1195 int page_num, int force_write)
1196{
1197 struct scrub_page *page_bad = sblock_bad->pagev + page_num;
1198 struct scrub_page *page_good = sblock_good->pagev + page_num;
1199
1200 BUG_ON(sblock_bad->pagev[page_num].page == NULL);
1201 BUG_ON(sblock_good->pagev[page_num].page == NULL);
1202 if (force_write || sblock_bad->header_error ||
1203 sblock_bad->checksum_error || page_bad->io_error) {
1204 struct bio *bio;
1205 int ret;
1206 DECLARE_COMPLETION_ONSTACK(complete);
1207
1208 bio = bio_alloc(GFP_NOFS, 1);
1209 if (!bio)
1210 return -EIO;
1211 bio->bi_bdev = page_bad->dev->bdev;
1212 bio->bi_sector = page_bad->physical >> 9;
1213 bio->bi_end_io = scrub_complete_bio_end_io;
1214 bio->bi_private = &complete;
1215
1216 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1217 if (PAGE_SIZE != ret) {
1218 bio_put(bio);
1219 return -EIO;
1220 }
1221 btrfsic_submit_bio(WRITE, bio);
1222
1223 /* this will also unplug the queue */
1224 wait_for_completion(&complete);
1225 if (!bio_flagged(bio, BIO_UPTODATE)) {
1226 btrfs_dev_stat_inc_and_print(page_bad->dev,
1227 BTRFS_DEV_STAT_WRITE_ERRS);
1228 bio_put(bio);
1229 return -EIO;
1230 }
1231 bio_put(bio);
1232 }
1233
1234 return 0;
1235}
1236
1237static void scrub_checksum(struct scrub_block *sblock)
1238{
1239 u64 flags;
1240 int ret;
1241
1242 BUG_ON(sblock->page_count < 1);
1243 flags = sblock->pagev[0].flags;
1244 ret = 0;
1245 if (flags & BTRFS_EXTENT_FLAG_DATA)
1246 ret = scrub_checksum_data(sblock);
1247 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1248 ret = scrub_checksum_tree_block(sblock);
1249 else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1250 (void)scrub_checksum_super(sblock);
1251 else
1252 WARN_ON(1);
1253 if (ret)
1254 scrub_handle_errored_block(sblock);
1255}
1256
1257static int scrub_checksum_data(struct scrub_block *sblock)
1258{
1259 struct scrub_dev *sdev = sblock->sdev;
1260 u8 csum[BTRFS_CSUM_SIZE];
1261 u8 *on_disk_csum;
1262 struct page *page;
1263 void *buffer;
1264 u32 crc = ~(u32)0;
1265 int fail = 0;
1266 struct btrfs_root *root = sdev->dev->dev_root;
1267 u64 len;
1268 int index;
1269
1270 BUG_ON(sblock->page_count < 1);
1271 if (!sblock->pagev[0].have_csum)
1272 return 0;
1273
1274 on_disk_csum = sblock->pagev[0].csum;
1275 page = sblock->pagev[0].page;
1276 buffer = kmap_atomic(page);
1277
1278 len = sdev->sectorsize;
1279 index = 0;
1280 for (;;) {
1281 u64 l = min_t(u64, len, PAGE_SIZE);
1282
1283 crc = btrfs_csum_data(root, buffer, crc, l);
1284 kunmap_atomic(buffer);
1285 len -= l;
1286 if (len == 0)
1287 break;
1288 index++;
1289 BUG_ON(index >= sblock->page_count);
1290 BUG_ON(!sblock->pagev[index].page);
1291 page = sblock->pagev[index].page;
1292 buffer = kmap_atomic(page);
1293 }
1294
1295 btrfs_csum_final(crc, csum);
1296 if (memcmp(csum, on_disk_csum, sdev->csum_size))
1297 fail = 1;
1298
1299 return fail;
1300}
1301
1302static int scrub_checksum_tree_block(struct scrub_block *sblock)
1303{
1304 struct scrub_dev *sdev = sblock->sdev;
1305 struct btrfs_header *h;
1306 struct btrfs_root *root = sdev->dev->dev_root;
1307 struct btrfs_fs_info *fs_info = root->fs_info;
1308 u8 calculated_csum[BTRFS_CSUM_SIZE];
1309 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1310 struct page *page;
1311 void *mapped_buffer;
1312 u64 mapped_size;
1313 void *p;
1314 u32 crc = ~(u32)0;
1315 int fail = 0;
1316 int crc_fail = 0;
1317 u64 len;
1318 int index;
1319
1320 BUG_ON(sblock->page_count < 1);
1321 page = sblock->pagev[0].page;
1322 mapped_buffer = kmap_atomic(page);
1323 h = (struct btrfs_header *)mapped_buffer;
1324 memcpy(on_disk_csum, h->csum, sdev->csum_size);
1325
1326 /*
1327 * we don't use the getter functions here, as we
1328 * a) don't have an extent buffer and
1329 * b) the page is already kmapped
1330 */
1331
1332 if (sblock->pagev[0].logical != le64_to_cpu(h->bytenr))
1333 ++fail;
1334
1335 if (sblock->pagev[0].generation != le64_to_cpu(h->generation))
1336 ++fail;
1337
1338 if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1339 ++fail;
1340
1341 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1342 BTRFS_UUID_SIZE))
1343 ++fail;
1344
1345 BUG_ON(sdev->nodesize != sdev->leafsize);
1346 len = sdev->nodesize - BTRFS_CSUM_SIZE;
1347 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1348 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1349 index = 0;
1350 for (;;) {
1351 u64 l = min_t(u64, len, mapped_size);
1352
1353 crc = btrfs_csum_data(root, p, crc, l);
1354 kunmap_atomic(mapped_buffer);
1355 len -= l;
1356 if (len == 0)
1357 break;
1358 index++;
1359 BUG_ON(index >= sblock->page_count);
1360 BUG_ON(!sblock->pagev[index].page);
1361 page = sblock->pagev[index].page;
1362 mapped_buffer = kmap_atomic(page);
1363 mapped_size = PAGE_SIZE;
1364 p = mapped_buffer;
1365 }
1366
1367 btrfs_csum_final(crc, calculated_csum);
1368 if (memcmp(calculated_csum, on_disk_csum, sdev->csum_size))
1369 ++crc_fail;
1370
1371 return fail || crc_fail;
1372}
1373
1374static int scrub_checksum_super(struct scrub_block *sblock)
1375{
1376 struct btrfs_super_block *s;
1377 struct scrub_dev *sdev = sblock->sdev;
1378 struct btrfs_root *root = sdev->dev->dev_root;
1379 struct btrfs_fs_info *fs_info = root->fs_info;
1380 u8 calculated_csum[BTRFS_CSUM_SIZE];
1381 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1382 struct page *page;
1383 void *mapped_buffer;
1384 u64 mapped_size;
1385 void *p;
1386 u32 crc = ~(u32)0;
1387 int fail_gen = 0;
1388 int fail_cor = 0;
1389 u64 len;
1390 int index;
1391
1392 BUG_ON(sblock->page_count < 1);
1393 page = sblock->pagev[0].page;
1394 mapped_buffer = kmap_atomic(page);
1395 s = (struct btrfs_super_block *)mapped_buffer;
1396 memcpy(on_disk_csum, s->csum, sdev->csum_size);
1397
1398 if (sblock->pagev[0].logical != le64_to_cpu(s->bytenr))
1399 ++fail_cor;
1400
1401 if (sblock->pagev[0].generation != le64_to_cpu(s->generation))
1402 ++fail_gen;
1403
1404 if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1405 ++fail_cor;
1406
1407 len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1408 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1409 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1410 index = 0;
1411 for (;;) {
1412 u64 l = min_t(u64, len, mapped_size);
1413
1414 crc = btrfs_csum_data(root, p, crc, l);
1415 kunmap_atomic(mapped_buffer);
1416 len -= l;
1417 if (len == 0)
1418 break;
1419 index++;
1420 BUG_ON(index >= sblock->page_count);
1421 BUG_ON(!sblock->pagev[index].page);
1422 page = sblock->pagev[index].page;
1423 mapped_buffer = kmap_atomic(page);
1424 mapped_size = PAGE_SIZE;
1425 p = mapped_buffer;
1426 }
1427
1428 btrfs_csum_final(crc, calculated_csum);
1429 if (memcmp(calculated_csum, on_disk_csum, sdev->csum_size))
1430 ++fail_cor;
1431
1432 if (fail_cor + fail_gen) {
1433 /*
1434 * if we find an error in a super block, we just report it.
1435 * They will get written with the next transaction commit
1436 * anyway
1437 */
1438 spin_lock(&sdev->stat_lock);
1439 ++sdev->stat.super_errors;
1440 spin_unlock(&sdev->stat_lock);
1441 if (fail_cor)
1442 btrfs_dev_stat_inc_and_print(sdev->dev,
1443 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1444 else
1445 btrfs_dev_stat_inc_and_print(sdev->dev,
1446 BTRFS_DEV_STAT_GENERATION_ERRS);
1447 }
1448
1449 return fail_cor + fail_gen;
1450}
1451
1452static void scrub_block_get(struct scrub_block *sblock)
1453{
1454 atomic_inc(&sblock->ref_count);
1455}
1456
1457static void scrub_block_put(struct scrub_block *sblock)
1458{
1459 if (atomic_dec_and_test(&sblock->ref_count)) {
1460 int i;
1461
1462 for (i = 0; i < sblock->page_count; i++)
1463 if (sblock->pagev[i].page)
1464 __free_page(sblock->pagev[i].page);
1465 kfree(sblock);
1466 }
1467}
1468
1469static void scrub_submit(struct scrub_dev *sdev)
1470{
1471 struct scrub_bio *sbio;
1472
1473 if (sdev->curr == -1)
1474 return;
1475
1476 sbio = sdev->bios[sdev->curr];
1477 sdev->curr = -1;
1478 atomic_inc(&sdev->in_flight);
1479
1480 btrfsic_submit_bio(READ, sbio->bio);
1481}
1482
1483static int scrub_add_page_to_bio(struct scrub_dev *sdev,
1484 struct scrub_page *spage)
1485{
1486 struct scrub_block *sblock = spage->sblock;
1487 struct scrub_bio *sbio;
1488 int ret;
1489
1490again:
1491 /*
1492 * grab a fresh bio or wait for one to become available
1493 */
1494 while (sdev->curr == -1) {
1495 spin_lock(&sdev->list_lock);
1496 sdev->curr = sdev->first_free;
1497 if (sdev->curr != -1) {
1498 sdev->first_free = sdev->bios[sdev->curr]->next_free;
1499 sdev->bios[sdev->curr]->next_free = -1;
1500 sdev->bios[sdev->curr]->page_count = 0;
1501 spin_unlock(&sdev->list_lock);
1502 } else {
1503 spin_unlock(&sdev->list_lock);
1504 wait_event(sdev->list_wait, sdev->first_free != -1);
1505 }
1506 }
1507 sbio = sdev->bios[sdev->curr];
1508 if (sbio->page_count == 0) {
1509 struct bio *bio;
1510
1511 sbio->physical = spage->physical;
1512 sbio->logical = spage->logical;
1513 bio = sbio->bio;
1514 if (!bio) {
1515 bio = bio_alloc(GFP_NOFS, sdev->pages_per_bio);
1516 if (!bio)
1517 return -ENOMEM;
1518 sbio->bio = bio;
1519 }
1520
1521 bio->bi_private = sbio;
1522 bio->bi_end_io = scrub_bio_end_io;
1523 bio->bi_bdev = sdev->dev->bdev;
1524 bio->bi_sector = spage->physical >> 9;
1525 sbio->err = 0;
1526 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1527 spage->physical ||
1528 sbio->logical + sbio->page_count * PAGE_SIZE !=
1529 spage->logical) {
1530 scrub_submit(sdev);
1531 goto again;
1532 }
1533
1534 sbio->pagev[sbio->page_count] = spage;
1535 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1536 if (ret != PAGE_SIZE) {
1537 if (sbio->page_count < 1) {
1538 bio_put(sbio->bio);
1539 sbio->bio = NULL;
1540 return -EIO;
1541 }
1542 scrub_submit(sdev);
1543 goto again;
1544 }
1545
1546 scrub_block_get(sblock); /* one for the added page */
1547 atomic_inc(&sblock->outstanding_pages);
1548 sbio->page_count++;
1549 if (sbio->page_count == sdev->pages_per_bio)
1550 scrub_submit(sdev);
1551
1552 return 0;
1553}
1554
1555static int scrub_pages(struct scrub_dev *sdev, u64 logical, u64 len,
1556 u64 physical, u64 flags, u64 gen, int mirror_num,
1557 u8 *csum, int force)
1558{
1559 struct scrub_block *sblock;
1560 int index;
1561
1562 sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
1563 if (!sblock) {
1564 spin_lock(&sdev->stat_lock);
1565 sdev->stat.malloc_errors++;
1566 spin_unlock(&sdev->stat_lock);
1567 return -ENOMEM;
1568 }
1569
1570 /* one ref inside this function, plus one for each page later on */
1571 atomic_set(&sblock->ref_count, 1);
1572 sblock->sdev = sdev;
1573 sblock->no_io_error_seen = 1;
1574
1575 for (index = 0; len > 0; index++) {
1576 struct scrub_page *spage = sblock->pagev + index;
1577 u64 l = min_t(u64, len, PAGE_SIZE);
1578
1579 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
1580 spage->page = alloc_page(GFP_NOFS);
1581 if (!spage->page) {
1582 spin_lock(&sdev->stat_lock);
1583 sdev->stat.malloc_errors++;
1584 spin_unlock(&sdev->stat_lock);
1585 while (index > 0) {
1586 index--;
1587 __free_page(sblock->pagev[index].page);
1588 }
1589 kfree(sblock);
1590 return -ENOMEM;
1591 }
1592 spage->sblock = sblock;
1593 spage->dev = sdev->dev;
1594 spage->flags = flags;
1595 spage->generation = gen;
1596 spage->logical = logical;
1597 spage->physical = physical;
1598 spage->mirror_num = mirror_num;
1599 if (csum) {
1600 spage->have_csum = 1;
1601 memcpy(spage->csum, csum, sdev->csum_size);
1602 } else {
1603 spage->have_csum = 0;
1604 }
1605 sblock->page_count++;
1606 len -= l;
1607 logical += l;
1608 physical += l;
1609 }
1610
1611 BUG_ON(sblock->page_count == 0);
1612 for (index = 0; index < sblock->page_count; index++) {
1613 struct scrub_page *spage = sblock->pagev + index;
1614 int ret;
1615
1616 ret = scrub_add_page_to_bio(sdev, spage);
1617 if (ret) {
1618 scrub_block_put(sblock);
1619 return ret;
1620 }
1621 }
1622
1623 if (force)
1624 scrub_submit(sdev);
1625
1626 /* last one frees, either here or in bio completion for last page */
1627 scrub_block_put(sblock);
1628 return 0;
1629}
1630
1631static void scrub_bio_end_io(struct bio *bio, int err)
1632{
1633 struct scrub_bio *sbio = bio->bi_private;
1634 struct scrub_dev *sdev = sbio->sdev;
1635 struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
1636
1637 sbio->err = err;
1638 sbio->bio = bio;
1639
1640 btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work);
1641}
1642
1643static void scrub_bio_end_io_worker(struct btrfs_work *work)
1644{
1645 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1646 struct scrub_dev *sdev = sbio->sdev;
1647 int i;
1648
1649 BUG_ON(sbio->page_count > SCRUB_PAGES_PER_BIO);
1650 if (sbio->err) {
1651 for (i = 0; i < sbio->page_count; i++) {
1652 struct scrub_page *spage = sbio->pagev[i];
1653
1654 spage->io_error = 1;
1655 spage->sblock->no_io_error_seen = 0;
1656 }
1657 }
1658
1659 /* now complete the scrub_block items that have all pages completed */
1660 for (i = 0; i < sbio->page_count; i++) {
1661 struct scrub_page *spage = sbio->pagev[i];
1662 struct scrub_block *sblock = spage->sblock;
1663
1664 if (atomic_dec_and_test(&sblock->outstanding_pages))
1665 scrub_block_complete(sblock);
1666 scrub_block_put(sblock);
1667 }
1668
1669 if (sbio->err) {
1670 /* what is this good for??? */
1671 sbio->bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1672 sbio->bio->bi_flags |= 1 << BIO_UPTODATE;
1673 sbio->bio->bi_phys_segments = 0;
1674 sbio->bio->bi_idx = 0;
1675
1676 for (i = 0; i < sbio->page_count; i++) {
1677 struct bio_vec *bi;
1678 bi = &sbio->bio->bi_io_vec[i];
1679 bi->bv_offset = 0;
1680 bi->bv_len = PAGE_SIZE;
1681 }
1682 }
1683
1684 bio_put(sbio->bio);
1685 sbio->bio = NULL;
1686 spin_lock(&sdev->list_lock);
1687 sbio->next_free = sdev->first_free;
1688 sdev->first_free = sbio->index;
1689 spin_unlock(&sdev->list_lock);
1690 atomic_dec(&sdev->in_flight);
1691 wake_up(&sdev->list_wait);
1692}
1693
1694static void scrub_block_complete(struct scrub_block *sblock)
1695{
1696 if (!sblock->no_io_error_seen)
1697 scrub_handle_errored_block(sblock);
1698 else
1699 scrub_checksum(sblock);
1700}
1701
1702static int scrub_find_csum(struct scrub_dev *sdev, u64 logical, u64 len,
1703 u8 *csum)
1704{
1705 struct btrfs_ordered_sum *sum = NULL;
1706 int ret = 0;
1707 unsigned long i;
1708 unsigned long num_sectors;
1709
1710 while (!list_empty(&sdev->csum_list)) {
1711 sum = list_first_entry(&sdev->csum_list,
1712 struct btrfs_ordered_sum, list);
1713 if (sum->bytenr > logical)
1714 return 0;
1715 if (sum->bytenr + sum->len > logical)
1716 break;
1717
1718 ++sdev->stat.csum_discards;
1719 list_del(&sum->list);
1720 kfree(sum);
1721 sum = NULL;
1722 }
1723 if (!sum)
1724 return 0;
1725
1726 num_sectors = sum->len / sdev->sectorsize;
1727 for (i = 0; i < num_sectors; ++i) {
1728 if (sum->sums[i].bytenr == logical) {
1729 memcpy(csum, &sum->sums[i].sum, sdev->csum_size);
1730 ret = 1;
1731 break;
1732 }
1733 }
1734 if (ret && i == num_sectors - 1) {
1735 list_del(&sum->list);
1736 kfree(sum);
1737 }
1738 return ret;
1739}
1740
1741/* scrub extent tries to collect up to 64 kB for each bio */
1742static int scrub_extent(struct scrub_dev *sdev, u64 logical, u64 len,
1743 u64 physical, u64 flags, u64 gen, int mirror_num)
1744{
1745 int ret;
1746 u8 csum[BTRFS_CSUM_SIZE];
1747 u32 blocksize;
1748
1749 if (flags & BTRFS_EXTENT_FLAG_DATA) {
1750 blocksize = sdev->sectorsize;
1751 spin_lock(&sdev->stat_lock);
1752 sdev->stat.data_extents_scrubbed++;
1753 sdev->stat.data_bytes_scrubbed += len;
1754 spin_unlock(&sdev->stat_lock);
1755 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1756 BUG_ON(sdev->nodesize != sdev->leafsize);
1757 blocksize = sdev->nodesize;
1758 spin_lock(&sdev->stat_lock);
1759 sdev->stat.tree_extents_scrubbed++;
1760 sdev->stat.tree_bytes_scrubbed += len;
1761 spin_unlock(&sdev->stat_lock);
1762 } else {
1763 blocksize = sdev->sectorsize;
1764 BUG_ON(1);
1765 }
1766
1767 while (len) {
1768 u64 l = min_t(u64, len, blocksize);
1769 int have_csum = 0;
1770
1771 if (flags & BTRFS_EXTENT_FLAG_DATA) {
1772 /* push csums to sbio */
1773 have_csum = scrub_find_csum(sdev, logical, l, csum);
1774 if (have_csum == 0)
1775 ++sdev->stat.no_csum;
1776 }
1777 ret = scrub_pages(sdev, logical, l, physical, flags, gen,
1778 mirror_num, have_csum ? csum : NULL, 0);
1779 if (ret)
1780 return ret;
1781 len -= l;
1782 logical += l;
1783 physical += l;
1784 }
1785 return 0;
1786}
1787
1788static noinline_for_stack int scrub_stripe(struct scrub_dev *sdev,
1789 struct map_lookup *map, int num, u64 base, u64 length)
1790{
1791 struct btrfs_path *path;
1792 struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
1793 struct btrfs_root *root = fs_info->extent_root;
1794 struct btrfs_root *csum_root = fs_info->csum_root;
1795 struct btrfs_extent_item *extent;
1796 struct blk_plug plug;
1797 u64 flags;
1798 int ret;
1799 int slot;
1800 int i;
1801 u64 nstripes;
1802 struct extent_buffer *l;
1803 struct btrfs_key key;
1804 u64 physical;
1805 u64 logical;
1806 u64 generation;
1807 int mirror_num;
1808 struct reada_control *reada1;
1809 struct reada_control *reada2;
1810 struct btrfs_key key_start;
1811 struct btrfs_key key_end;
1812
1813 u64 increment = map->stripe_len;
1814 u64 offset;
1815
1816 nstripes = length;
1817 offset = 0;
1818 do_div(nstripes, map->stripe_len);
1819 if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
1820 offset = map->stripe_len * num;
1821 increment = map->stripe_len * map->num_stripes;
1822 mirror_num = 1;
1823 } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
1824 int factor = map->num_stripes / map->sub_stripes;
1825 offset = map->stripe_len * (num / map->sub_stripes);
1826 increment = map->stripe_len * factor;
1827 mirror_num = num % map->sub_stripes + 1;
1828 } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
1829 increment = map->stripe_len;
1830 mirror_num = num % map->num_stripes + 1;
1831 } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
1832 increment = map->stripe_len;
1833 mirror_num = num % map->num_stripes + 1;
1834 } else {
1835 increment = map->stripe_len;
1836 mirror_num = 1;
1837 }
1838
1839 path = btrfs_alloc_path();
1840 if (!path)
1841 return -ENOMEM;
1842
1843 /*
1844 * work on commit root. The related disk blocks are static as
1845 * long as COW is applied. This means, it is save to rewrite
1846 * them to repair disk errors without any race conditions
1847 */
1848 path->search_commit_root = 1;
1849 path->skip_locking = 1;
1850
1851 /*
1852 * trigger the readahead for extent tree csum tree and wait for
1853 * completion. During readahead, the scrub is officially paused
1854 * to not hold off transaction commits
1855 */
1856 logical = base + offset;
1857
1858 wait_event(sdev->list_wait,
1859 atomic_read(&sdev->in_flight) == 0);
1860 atomic_inc(&fs_info->scrubs_paused);
1861 wake_up(&fs_info->scrub_pause_wait);
1862
1863 /* FIXME it might be better to start readahead at commit root */
1864 key_start.objectid = logical;
1865 key_start.type = BTRFS_EXTENT_ITEM_KEY;
1866 key_start.offset = (u64)0;
1867 key_end.objectid = base + offset + nstripes * increment;
1868 key_end.type = BTRFS_EXTENT_ITEM_KEY;
1869 key_end.offset = (u64)0;
1870 reada1 = btrfs_reada_add(root, &key_start, &key_end);
1871
1872 key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
1873 key_start.type = BTRFS_EXTENT_CSUM_KEY;
1874 key_start.offset = logical;
1875 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
1876 key_end.type = BTRFS_EXTENT_CSUM_KEY;
1877 key_end.offset = base + offset + nstripes * increment;
1878 reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
1879
1880 if (!IS_ERR(reada1))
1881 btrfs_reada_wait(reada1);
1882 if (!IS_ERR(reada2))
1883 btrfs_reada_wait(reada2);
1884
1885 mutex_lock(&fs_info->scrub_lock);
1886 while (atomic_read(&fs_info->scrub_pause_req)) {
1887 mutex_unlock(&fs_info->scrub_lock);
1888 wait_event(fs_info->scrub_pause_wait,
1889 atomic_read(&fs_info->scrub_pause_req) == 0);
1890 mutex_lock(&fs_info->scrub_lock);
1891 }
1892 atomic_dec(&fs_info->scrubs_paused);
1893 mutex_unlock(&fs_info->scrub_lock);
1894 wake_up(&fs_info->scrub_pause_wait);
1895
1896 /*
1897 * collect all data csums for the stripe to avoid seeking during
1898 * the scrub. This might currently (crc32) end up to be about 1MB
1899 */
1900 blk_start_plug(&plug);
1901
1902 /*
1903 * now find all extents for each stripe and scrub them
1904 */
1905 logical = base + offset;
1906 physical = map->stripes[num].physical;
1907 ret = 0;
1908 for (i = 0; i < nstripes; ++i) {
1909 /*
1910 * canceled?
1911 */
1912 if (atomic_read(&fs_info->scrub_cancel_req) ||
1913 atomic_read(&sdev->cancel_req)) {
1914 ret = -ECANCELED;
1915 goto out;
1916 }
1917 /*
1918 * check to see if we have to pause
1919 */
1920 if (atomic_read(&fs_info->scrub_pause_req)) {
1921 /* push queued extents */
1922 scrub_submit(sdev);
1923 wait_event(sdev->list_wait,
1924 atomic_read(&sdev->in_flight) == 0);
1925 atomic_inc(&fs_info->scrubs_paused);
1926 wake_up(&fs_info->scrub_pause_wait);
1927 mutex_lock(&fs_info->scrub_lock);
1928 while (atomic_read(&fs_info->scrub_pause_req)) {
1929 mutex_unlock(&fs_info->scrub_lock);
1930 wait_event(fs_info->scrub_pause_wait,
1931 atomic_read(&fs_info->scrub_pause_req) == 0);
1932 mutex_lock(&fs_info->scrub_lock);
1933 }
1934 atomic_dec(&fs_info->scrubs_paused);
1935 mutex_unlock(&fs_info->scrub_lock);
1936 wake_up(&fs_info->scrub_pause_wait);
1937 }
1938
1939 ret = btrfs_lookup_csums_range(csum_root, logical,
1940 logical + map->stripe_len - 1,
1941 &sdev->csum_list, 1);
1942 if (ret)
1943 goto out;
1944
1945 key.objectid = logical;
1946 key.type = BTRFS_EXTENT_ITEM_KEY;
1947 key.offset = (u64)0;
1948
1949 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1950 if (ret < 0)
1951 goto out;
1952 if (ret > 0) {
1953 ret = btrfs_previous_item(root, path, 0,
1954 BTRFS_EXTENT_ITEM_KEY);
1955 if (ret < 0)
1956 goto out;
1957 if (ret > 0) {
1958 /* there's no smaller item, so stick with the
1959 * larger one */
1960 btrfs_release_path(path);
1961 ret = btrfs_search_slot(NULL, root, &key,
1962 path, 0, 0);
1963 if (ret < 0)
1964 goto out;
1965 }
1966 }
1967
1968 while (1) {
1969 l = path->nodes[0];
1970 slot = path->slots[0];
1971 if (slot >= btrfs_header_nritems(l)) {
1972 ret = btrfs_next_leaf(root, path);
1973 if (ret == 0)
1974 continue;
1975 if (ret < 0)
1976 goto out;
1977
1978 break;
1979 }
1980 btrfs_item_key_to_cpu(l, &key, slot);
1981
1982 if (key.objectid + key.offset <= logical)
1983 goto next;
1984
1985 if (key.objectid >= logical + map->stripe_len)
1986 break;
1987
1988 if (btrfs_key_type(&key) != BTRFS_EXTENT_ITEM_KEY)
1989 goto next;
1990
1991 extent = btrfs_item_ptr(l, slot,
1992 struct btrfs_extent_item);
1993 flags = btrfs_extent_flags(l, extent);
1994 generation = btrfs_extent_generation(l, extent);
1995
1996 if (key.objectid < logical &&
1997 (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
1998 printk(KERN_ERR
1999 "btrfs scrub: tree block %llu spanning "
2000 "stripes, ignored. logical=%llu\n",
2001 (unsigned long long)key.objectid,
2002 (unsigned long long)logical);
2003 goto next;
2004 }
2005
2006 /*
2007 * trim extent to this stripe
2008 */
2009 if (key.objectid < logical) {
2010 key.offset -= logical - key.objectid;
2011 key.objectid = logical;
2012 }
2013 if (key.objectid + key.offset >
2014 logical + map->stripe_len) {
2015 key.offset = logical + map->stripe_len -
2016 key.objectid;
2017 }
2018
2019 ret = scrub_extent(sdev, key.objectid, key.offset,
2020 key.objectid - logical + physical,
2021 flags, generation, mirror_num);
2022 if (ret)
2023 goto out;
2024
2025next:
2026 path->slots[0]++;
2027 }
2028 btrfs_release_path(path);
2029 logical += increment;
2030 physical += map->stripe_len;
2031 spin_lock(&sdev->stat_lock);
2032 sdev->stat.last_physical = physical;
2033 spin_unlock(&sdev->stat_lock);
2034 }
2035 /* push queued extents */
2036 scrub_submit(sdev);
2037
2038out:
2039 blk_finish_plug(&plug);
2040 btrfs_free_path(path);
2041 return ret < 0 ? ret : 0;
2042}
2043
2044static noinline_for_stack int scrub_chunk(struct scrub_dev *sdev,
2045 u64 chunk_tree, u64 chunk_objectid, u64 chunk_offset, u64 length,
2046 u64 dev_offset)
2047{
2048 struct btrfs_mapping_tree *map_tree =
2049 &sdev->dev->dev_root->fs_info->mapping_tree;
2050 struct map_lookup *map;
2051 struct extent_map *em;
2052 int i;
2053 int ret = -EINVAL;
2054
2055 read_lock(&map_tree->map_tree.lock);
2056 em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
2057 read_unlock(&map_tree->map_tree.lock);
2058
2059 if (!em)
2060 return -EINVAL;
2061
2062 map = (struct map_lookup *)em->bdev;
2063 if (em->start != chunk_offset)
2064 goto out;
2065
2066 if (em->len < length)
2067 goto out;
2068
2069 for (i = 0; i < map->num_stripes; ++i) {
2070 if (map->stripes[i].dev == sdev->dev &&
2071 map->stripes[i].physical == dev_offset) {
2072 ret = scrub_stripe(sdev, map, i, chunk_offset, length);
2073 if (ret)
2074 goto out;
2075 }
2076 }
2077out:
2078 free_extent_map(em);
2079
2080 return ret;
2081}
2082
2083static noinline_for_stack
2084int scrub_enumerate_chunks(struct scrub_dev *sdev, u64 start, u64 end)
2085{
2086 struct btrfs_dev_extent *dev_extent = NULL;
2087 struct btrfs_path *path;
2088 struct btrfs_root *root = sdev->dev->dev_root;
2089 struct btrfs_fs_info *fs_info = root->fs_info;
2090 u64 length;
2091 u64 chunk_tree;
2092 u64 chunk_objectid;
2093 u64 chunk_offset;
2094 int ret;
2095 int slot;
2096 struct extent_buffer *l;
2097 struct btrfs_key key;
2098 struct btrfs_key found_key;
2099 struct btrfs_block_group_cache *cache;
2100
2101 path = btrfs_alloc_path();
2102 if (!path)
2103 return -ENOMEM;
2104
2105 path->reada = 2;
2106 path->search_commit_root = 1;
2107 path->skip_locking = 1;
2108
2109 key.objectid = sdev->dev->devid;
2110 key.offset = 0ull;
2111 key.type = BTRFS_DEV_EXTENT_KEY;
2112
2113
2114 while (1) {
2115 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2116 if (ret < 0)
2117 break;
2118 if (ret > 0) {
2119 if (path->slots[0] >=
2120 btrfs_header_nritems(path->nodes[0])) {
2121 ret = btrfs_next_leaf(root, path);
2122 if (ret)
2123 break;
2124 }
2125 }
2126
2127 l = path->nodes[0];
2128 slot = path->slots[0];
2129
2130 btrfs_item_key_to_cpu(l, &found_key, slot);
2131
2132 if (found_key.objectid != sdev->dev->devid)
2133 break;
2134
2135 if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
2136 break;
2137
2138 if (found_key.offset >= end)
2139 break;
2140
2141 if (found_key.offset < key.offset)
2142 break;
2143
2144 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2145 length = btrfs_dev_extent_length(l, dev_extent);
2146
2147 if (found_key.offset + length <= start) {
2148 key.offset = found_key.offset + length;
2149 btrfs_release_path(path);
2150 continue;
2151 }
2152
2153 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
2154 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
2155 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2156
2157 /*
2158 * get a reference on the corresponding block group to prevent
2159 * the chunk from going away while we scrub it
2160 */
2161 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2162 if (!cache) {
2163 ret = -ENOENT;
2164 break;
2165 }
2166 ret = scrub_chunk(sdev, chunk_tree, chunk_objectid,
2167 chunk_offset, length, found_key.offset);
2168 btrfs_put_block_group(cache);
2169 if (ret)
2170 break;
2171
2172 key.offset = found_key.offset + length;
2173 btrfs_release_path(path);
2174 }
2175
2176 btrfs_free_path(path);
2177
2178 /*
2179 * ret can still be 1 from search_slot or next_leaf,
2180 * that's not an error
2181 */
2182 return ret < 0 ? ret : 0;
2183}
2184
2185static noinline_for_stack int scrub_supers(struct scrub_dev *sdev)
2186{
2187 int i;
2188 u64 bytenr;
2189 u64 gen;
2190 int ret;
2191 struct btrfs_device *device = sdev->dev;
2192 struct btrfs_root *root = device->dev_root;
2193
2194 if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
2195 return -EIO;
2196
2197 gen = root->fs_info->last_trans_committed;
2198
2199 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2200 bytenr = btrfs_sb_offset(i);
2201 if (bytenr + BTRFS_SUPER_INFO_SIZE > device->total_bytes)
2202 break;
2203
2204 ret = scrub_pages(sdev, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
2205 BTRFS_EXTENT_FLAG_SUPER, gen, i, NULL, 1);
2206 if (ret)
2207 return ret;
2208 }
2209 wait_event(sdev->list_wait, atomic_read(&sdev->in_flight) == 0);
2210
2211 return 0;
2212}
2213
2214/*
2215 * get a reference count on fs_info->scrub_workers. start worker if necessary
2216 */
2217static noinline_for_stack int scrub_workers_get(struct btrfs_root *root)
2218{
2219 struct btrfs_fs_info *fs_info = root->fs_info;
2220 int ret = 0;
2221
2222 mutex_lock(&fs_info->scrub_lock);
2223 if (fs_info->scrub_workers_refcnt == 0) {
2224 btrfs_init_workers(&fs_info->scrub_workers, "scrub",
2225 fs_info->thread_pool_size, &fs_info->generic_worker);
2226 fs_info->scrub_workers.idle_thresh = 4;
2227 ret = btrfs_start_workers(&fs_info->scrub_workers);
2228 if (ret)
2229 goto out;
2230 }
2231 ++fs_info->scrub_workers_refcnt;
2232out:
2233 mutex_unlock(&fs_info->scrub_lock);
2234
2235 return ret;
2236}
2237
2238static noinline_for_stack void scrub_workers_put(struct btrfs_root *root)
2239{
2240 struct btrfs_fs_info *fs_info = root->fs_info;
2241
2242 mutex_lock(&fs_info->scrub_lock);
2243 if (--fs_info->scrub_workers_refcnt == 0)
2244 btrfs_stop_workers(&fs_info->scrub_workers);
2245 WARN_ON(fs_info->scrub_workers_refcnt < 0);
2246 mutex_unlock(&fs_info->scrub_lock);
2247}
2248
2249
2250int btrfs_scrub_dev(struct btrfs_root *root, u64 devid, u64 start, u64 end,
2251 struct btrfs_scrub_progress *progress, int readonly)
2252{
2253 struct scrub_dev *sdev;
2254 struct btrfs_fs_info *fs_info = root->fs_info;
2255 int ret;
2256 struct btrfs_device *dev;
2257
2258 if (btrfs_fs_closing(root->fs_info))
2259 return -EINVAL;
2260
2261 /*
2262 * check some assumptions
2263 */
2264 if (root->nodesize != root->leafsize) {
2265 printk(KERN_ERR
2266 "btrfs_scrub: size assumption nodesize == leafsize (%d == %d) fails\n",
2267 root->nodesize, root->leafsize);
2268 return -EINVAL;
2269 }
2270
2271 if (root->nodesize > BTRFS_STRIPE_LEN) {
2272 /*
2273 * in this case scrub is unable to calculate the checksum
2274 * the way scrub is implemented. Do not handle this
2275 * situation at all because it won't ever happen.
2276 */
2277 printk(KERN_ERR
2278 "btrfs_scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails\n",
2279 root->nodesize, BTRFS_STRIPE_LEN);
2280 return -EINVAL;
2281 }
2282
2283 if (root->sectorsize != PAGE_SIZE) {
2284 /* not supported for data w/o checksums */
2285 printk(KERN_ERR
2286 "btrfs_scrub: size assumption sectorsize != PAGE_SIZE (%d != %lld) fails\n",
2287 root->sectorsize, (unsigned long long)PAGE_SIZE);
2288 return -EINVAL;
2289 }
2290
2291 ret = scrub_workers_get(root);
2292 if (ret)
2293 return ret;
2294
2295 mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
2296 dev = btrfs_find_device(root, devid, NULL, NULL);
2297 if (!dev || dev->missing) {
2298 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2299 scrub_workers_put(root);
2300 return -ENODEV;
2301 }
2302 mutex_lock(&fs_info->scrub_lock);
2303
2304 if (!dev->in_fs_metadata) {
2305 mutex_unlock(&fs_info->scrub_lock);
2306 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2307 scrub_workers_put(root);
2308 return -ENODEV;
2309 }
2310
2311 if (dev->scrub_device) {
2312 mutex_unlock(&fs_info->scrub_lock);
2313 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2314 scrub_workers_put(root);
2315 return -EINPROGRESS;
2316 }
2317 sdev = scrub_setup_dev(dev);
2318 if (IS_ERR(sdev)) {
2319 mutex_unlock(&fs_info->scrub_lock);
2320 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2321 scrub_workers_put(root);
2322 return PTR_ERR(sdev);
2323 }
2324 sdev->readonly = readonly;
2325 dev->scrub_device = sdev;
2326
2327 atomic_inc(&fs_info->scrubs_running);
2328 mutex_unlock(&fs_info->scrub_lock);
2329 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2330
2331 down_read(&fs_info->scrub_super_lock);
2332 ret = scrub_supers(sdev);
2333 up_read(&fs_info->scrub_super_lock);
2334
2335 if (!ret)
2336 ret = scrub_enumerate_chunks(sdev, start, end);
2337
2338 wait_event(sdev->list_wait, atomic_read(&sdev->in_flight) == 0);
2339 atomic_dec(&fs_info->scrubs_running);
2340 wake_up(&fs_info->scrub_pause_wait);
2341
2342 wait_event(sdev->list_wait, atomic_read(&sdev->fixup_cnt) == 0);
2343
2344 if (progress)
2345 memcpy(progress, &sdev->stat, sizeof(*progress));
2346
2347 mutex_lock(&fs_info->scrub_lock);
2348 dev->scrub_device = NULL;
2349 mutex_unlock(&fs_info->scrub_lock);
2350
2351 scrub_free_dev(sdev);
2352 scrub_workers_put(root);
2353
2354 return ret;
2355}
2356
2357void btrfs_scrub_pause(struct btrfs_root *root)
2358{
2359 struct btrfs_fs_info *fs_info = root->fs_info;
2360
2361 mutex_lock(&fs_info->scrub_lock);
2362 atomic_inc(&fs_info->scrub_pause_req);
2363 while (atomic_read(&fs_info->scrubs_paused) !=
2364 atomic_read(&fs_info->scrubs_running)) {
2365 mutex_unlock(&fs_info->scrub_lock);
2366 wait_event(fs_info->scrub_pause_wait,
2367 atomic_read(&fs_info->scrubs_paused) ==
2368 atomic_read(&fs_info->scrubs_running));
2369 mutex_lock(&fs_info->scrub_lock);
2370 }
2371 mutex_unlock(&fs_info->scrub_lock);
2372}
2373
2374void btrfs_scrub_continue(struct btrfs_root *root)
2375{
2376 struct btrfs_fs_info *fs_info = root->fs_info;
2377
2378 atomic_dec(&fs_info->scrub_pause_req);
2379 wake_up(&fs_info->scrub_pause_wait);
2380}
2381
2382void btrfs_scrub_pause_super(struct btrfs_root *root)
2383{
2384 down_write(&root->fs_info->scrub_super_lock);
2385}
2386
2387void btrfs_scrub_continue_super(struct btrfs_root *root)
2388{
2389 up_write(&root->fs_info->scrub_super_lock);
2390}
2391
2392int __btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
2393{
2394
2395 mutex_lock(&fs_info->scrub_lock);
2396 if (!atomic_read(&fs_info->scrubs_running)) {
2397 mutex_unlock(&fs_info->scrub_lock);
2398 return -ENOTCONN;
2399 }
2400
2401 atomic_inc(&fs_info->scrub_cancel_req);
2402 while (atomic_read(&fs_info->scrubs_running)) {
2403 mutex_unlock(&fs_info->scrub_lock);
2404 wait_event(fs_info->scrub_pause_wait,
2405 atomic_read(&fs_info->scrubs_running) == 0);
2406 mutex_lock(&fs_info->scrub_lock);
2407 }
2408 atomic_dec(&fs_info->scrub_cancel_req);
2409 mutex_unlock(&fs_info->scrub_lock);
2410
2411 return 0;
2412}
2413
2414int btrfs_scrub_cancel(struct btrfs_root *root)
2415{
2416 return __btrfs_scrub_cancel(root->fs_info);
2417}
2418
2419int btrfs_scrub_cancel_dev(struct btrfs_root *root, struct btrfs_device *dev)
2420{
2421 struct btrfs_fs_info *fs_info = root->fs_info;
2422 struct scrub_dev *sdev;
2423
2424 mutex_lock(&fs_info->scrub_lock);
2425 sdev = dev->scrub_device;
2426 if (!sdev) {
2427 mutex_unlock(&fs_info->scrub_lock);
2428 return -ENOTCONN;
2429 }
2430 atomic_inc(&sdev->cancel_req);
2431 while (dev->scrub_device) {
2432 mutex_unlock(&fs_info->scrub_lock);
2433 wait_event(fs_info->scrub_pause_wait,
2434 dev->scrub_device == NULL);
2435 mutex_lock(&fs_info->scrub_lock);
2436 }
2437 mutex_unlock(&fs_info->scrub_lock);
2438
2439 return 0;
2440}
2441
2442int btrfs_scrub_cancel_devid(struct btrfs_root *root, u64 devid)
2443{
2444 struct btrfs_fs_info *fs_info = root->fs_info;
2445 struct btrfs_device *dev;
2446 int ret;
2447
2448 /*
2449 * we have to hold the device_list_mutex here so the device
2450 * does not go away in cancel_dev. FIXME: find a better solution
2451 */
2452 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2453 dev = btrfs_find_device(root, devid, NULL, NULL);
2454 if (!dev) {
2455 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2456 return -ENODEV;
2457 }
2458 ret = btrfs_scrub_cancel_dev(root, dev);
2459 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2460
2461 return ret;
2462}
2463
2464int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
2465 struct btrfs_scrub_progress *progress)
2466{
2467 struct btrfs_device *dev;
2468 struct scrub_dev *sdev = NULL;
2469
2470 mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
2471 dev = btrfs_find_device(root, devid, NULL, NULL);
2472 if (dev)
2473 sdev = dev->scrub_device;
2474 if (sdev)
2475 memcpy(progress, &sdev->stat, sizeof(*progress));
2476 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2477
2478 return dev ? (sdev ? 0 : -ENOTCONN) : -ENODEV;
2479}