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, 2012 STRATO. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19#include <linux/blkdev.h>
20#include <linux/ratelimit.h>
21#include "ctree.h"
22#include "volumes.h"
23#include "disk-io.h"
24#include "ordered-data.h"
25#include "transaction.h"
26#include "backref.h"
27#include "extent_io.h"
28#include "dev-replace.h"
29#include "check-integrity.h"
30#include "rcu-string.h"
31#include "raid56.h"
32
33/*
34 * This is only the first step towards a full-features scrub. It reads all
35 * extent and super block and verifies the checksums. In case a bad checksum
36 * is found or the extent cannot be read, good data will be written back if
37 * any can be found.
38 *
39 * Future enhancements:
40 * - In case an unrepairable extent is encountered, track which files are
41 * affected and report them
42 * - track and record media errors, throw out bad devices
43 * - add a mode to also read unallocated space
44 */
45
46struct scrub_block;
47struct scrub_ctx;
48
49/*
50 * the following three values only influence the performance.
51 * The last one configures the number of parallel and outstanding I/O
52 * operations. The first two values configure an upper limit for the number
53 * of (dynamically allocated) pages that are added to a bio.
54 */
55#define SCRUB_PAGES_PER_RD_BIO 32 /* 128k per bio */
56#define SCRUB_PAGES_PER_WR_BIO 32 /* 128k per bio */
57#define SCRUB_BIOS_PER_SCTX 64 /* 8MB per device in flight */
58
59/*
60 * the following value times PAGE_SIZE needs to be large enough to match the
61 * largest node/leaf/sector size that shall be supported.
62 * Values larger than BTRFS_STRIPE_LEN are not supported.
63 */
64#define SCRUB_MAX_PAGES_PER_BLOCK 16 /* 64k per node/leaf/sector */
65
66struct scrub_page {
67 struct scrub_block *sblock;
68 struct page *page;
69 struct btrfs_device *dev;
70 u64 flags; /* extent flags */
71 u64 generation;
72 u64 logical;
73 u64 physical;
74 u64 physical_for_dev_replace;
75 atomic_t ref_count;
76 struct {
77 unsigned int mirror_num:8;
78 unsigned int have_csum:1;
79 unsigned int io_error:1;
80 };
81 u8 csum[BTRFS_CSUM_SIZE];
82};
83
84struct scrub_bio {
85 int index;
86 struct scrub_ctx *sctx;
87 struct btrfs_device *dev;
88 struct bio *bio;
89 int err;
90 u64 logical;
91 u64 physical;
92#if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
93 struct scrub_page *pagev[SCRUB_PAGES_PER_WR_BIO];
94#else
95 struct scrub_page *pagev[SCRUB_PAGES_PER_RD_BIO];
96#endif
97 int page_count;
98 int next_free;
99 struct btrfs_work work;
100};
101
102struct scrub_block {
103 struct scrub_page *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
104 int page_count;
105 atomic_t outstanding_pages;
106 atomic_t ref_count; /* free mem on transition to zero */
107 struct scrub_ctx *sctx;
108 struct {
109 unsigned int header_error:1;
110 unsigned int checksum_error:1;
111 unsigned int no_io_error_seen:1;
112 unsigned int generation_error:1; /* also sets header_error */
113 };
114};
115
116struct scrub_wr_ctx {
117 struct scrub_bio *wr_curr_bio;
118 struct btrfs_device *tgtdev;
119 int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
120 atomic_t flush_all_writes;
121 struct mutex wr_lock;
122};
123
124struct scrub_ctx {
125 struct scrub_bio *bios[SCRUB_BIOS_PER_SCTX];
126 struct btrfs_root *dev_root;
127 int first_free;
128 int curr;
129 atomic_t bios_in_flight;
130 atomic_t workers_pending;
131 spinlock_t list_lock;
132 wait_queue_head_t list_wait;
133 u16 csum_size;
134 struct list_head csum_list;
135 atomic_t cancel_req;
136 int readonly;
137 int pages_per_rd_bio;
138 u32 sectorsize;
139 u32 nodesize;
140 u32 leafsize;
141
142 int is_dev_replace;
143 struct scrub_wr_ctx wr_ctx;
144
145 /*
146 * statistics
147 */
148 struct btrfs_scrub_progress stat;
149 spinlock_t stat_lock;
150};
151
152struct scrub_fixup_nodatasum {
153 struct scrub_ctx *sctx;
154 struct btrfs_device *dev;
155 u64 logical;
156 struct btrfs_root *root;
157 struct btrfs_work work;
158 int mirror_num;
159};
160
161struct scrub_nocow_inode {
162 u64 inum;
163 u64 offset;
164 u64 root;
165 struct list_head list;
166};
167
168struct scrub_copy_nocow_ctx {
169 struct scrub_ctx *sctx;
170 u64 logical;
171 u64 len;
172 int mirror_num;
173 u64 physical_for_dev_replace;
174 struct list_head inodes;
175 struct btrfs_work work;
176};
177
178struct scrub_warning {
179 struct btrfs_path *path;
180 u64 extent_item_size;
181 char *scratch_buf;
182 char *msg_buf;
183 const char *errstr;
184 sector_t sector;
185 u64 logical;
186 struct btrfs_device *dev;
187 int msg_bufsize;
188 int scratch_bufsize;
189};
190
191
192static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
193static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
194static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
195static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
196static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
197static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
198 struct btrfs_fs_info *fs_info,
199 struct scrub_block *original_sblock,
200 u64 length, u64 logical,
201 struct scrub_block *sblocks_for_recheck);
202static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
203 struct scrub_block *sblock, int is_metadata,
204 int have_csum, u8 *csum, u64 generation,
205 u16 csum_size);
206static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
207 struct scrub_block *sblock,
208 int is_metadata, int have_csum,
209 const u8 *csum, u64 generation,
210 u16 csum_size);
211static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
212 struct scrub_block *sblock_good,
213 int force_write);
214static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
215 struct scrub_block *sblock_good,
216 int page_num, int force_write);
217static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
218static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
219 int page_num);
220static int scrub_checksum_data(struct scrub_block *sblock);
221static int scrub_checksum_tree_block(struct scrub_block *sblock);
222static int scrub_checksum_super(struct scrub_block *sblock);
223static void scrub_block_get(struct scrub_block *sblock);
224static void scrub_block_put(struct scrub_block *sblock);
225static void scrub_page_get(struct scrub_page *spage);
226static void scrub_page_put(struct scrub_page *spage);
227static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
228 struct scrub_page *spage);
229static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
230 u64 physical, struct btrfs_device *dev, u64 flags,
231 u64 gen, int mirror_num, u8 *csum, int force,
232 u64 physical_for_dev_replace);
233static void scrub_bio_end_io(struct bio *bio, int err);
234static void scrub_bio_end_io_worker(struct btrfs_work *work);
235static void scrub_block_complete(struct scrub_block *sblock);
236static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
237 u64 extent_logical, u64 extent_len,
238 u64 *extent_physical,
239 struct btrfs_device **extent_dev,
240 int *extent_mirror_num);
241static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
242 struct scrub_wr_ctx *wr_ctx,
243 struct btrfs_fs_info *fs_info,
244 struct btrfs_device *dev,
245 int is_dev_replace);
246static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx);
247static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
248 struct scrub_page *spage);
249static void scrub_wr_submit(struct scrub_ctx *sctx);
250static void scrub_wr_bio_end_io(struct bio *bio, int err);
251static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
252static int write_page_nocow(struct scrub_ctx *sctx,
253 u64 physical_for_dev_replace, struct page *page);
254static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
255 struct scrub_copy_nocow_ctx *ctx);
256static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
257 int mirror_num, u64 physical_for_dev_replace);
258static void copy_nocow_pages_worker(struct btrfs_work *work);
259static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
260static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
261
262
263static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
264{
265 atomic_inc(&sctx->bios_in_flight);
266}
267
268static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
269{
270 atomic_dec(&sctx->bios_in_flight);
271 wake_up(&sctx->list_wait);
272}
273
274static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
275{
276 while (atomic_read(&fs_info->scrub_pause_req)) {
277 mutex_unlock(&fs_info->scrub_lock);
278 wait_event(fs_info->scrub_pause_wait,
279 atomic_read(&fs_info->scrub_pause_req) == 0);
280 mutex_lock(&fs_info->scrub_lock);
281 }
282}
283
284static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
285{
286 atomic_inc(&fs_info->scrubs_paused);
287 wake_up(&fs_info->scrub_pause_wait);
288
289 mutex_lock(&fs_info->scrub_lock);
290 __scrub_blocked_if_needed(fs_info);
291 atomic_dec(&fs_info->scrubs_paused);
292 mutex_unlock(&fs_info->scrub_lock);
293
294 wake_up(&fs_info->scrub_pause_wait);
295}
296
297/*
298 * used for workers that require transaction commits (i.e., for the
299 * NOCOW case)
300 */
301static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
302{
303 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
304
305 /*
306 * increment scrubs_running to prevent cancel requests from
307 * completing as long as a worker is running. we must also
308 * increment scrubs_paused to prevent deadlocking on pause
309 * requests used for transactions commits (as the worker uses a
310 * transaction context). it is safe to regard the worker
311 * as paused for all matters practical. effectively, we only
312 * avoid cancellation requests from completing.
313 */
314 mutex_lock(&fs_info->scrub_lock);
315 atomic_inc(&fs_info->scrubs_running);
316 atomic_inc(&fs_info->scrubs_paused);
317 mutex_unlock(&fs_info->scrub_lock);
318
319 /*
320 * check if @scrubs_running=@scrubs_paused condition
321 * inside wait_event() is not an atomic operation.
322 * which means we may inc/dec @scrub_running/paused
323 * at any time. Let's wake up @scrub_pause_wait as
324 * much as we can to let commit transaction blocked less.
325 */
326 wake_up(&fs_info->scrub_pause_wait);
327
328 atomic_inc(&sctx->workers_pending);
329}
330
331/* used for workers that require transaction commits */
332static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
333{
334 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
335
336 /*
337 * see scrub_pending_trans_workers_inc() why we're pretending
338 * to be paused in the scrub counters
339 */
340 mutex_lock(&fs_info->scrub_lock);
341 atomic_dec(&fs_info->scrubs_running);
342 atomic_dec(&fs_info->scrubs_paused);
343 mutex_unlock(&fs_info->scrub_lock);
344 atomic_dec(&sctx->workers_pending);
345 wake_up(&fs_info->scrub_pause_wait);
346 wake_up(&sctx->list_wait);
347}
348
349static void scrub_free_csums(struct scrub_ctx *sctx)
350{
351 while (!list_empty(&sctx->csum_list)) {
352 struct btrfs_ordered_sum *sum;
353 sum = list_first_entry(&sctx->csum_list,
354 struct btrfs_ordered_sum, list);
355 list_del(&sum->list);
356 kfree(sum);
357 }
358}
359
360static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
361{
362 int i;
363
364 if (!sctx)
365 return;
366
367 scrub_free_wr_ctx(&sctx->wr_ctx);
368
369 /* this can happen when scrub is cancelled */
370 if (sctx->curr != -1) {
371 struct scrub_bio *sbio = sctx->bios[sctx->curr];
372
373 for (i = 0; i < sbio->page_count; i++) {
374 WARN_ON(!sbio->pagev[i]->page);
375 scrub_block_put(sbio->pagev[i]->sblock);
376 }
377 bio_put(sbio->bio);
378 }
379
380 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
381 struct scrub_bio *sbio = sctx->bios[i];
382
383 if (!sbio)
384 break;
385 kfree(sbio);
386 }
387
388 scrub_free_csums(sctx);
389 kfree(sctx);
390}
391
392static noinline_for_stack
393struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
394{
395 struct scrub_ctx *sctx;
396 int i;
397 struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
398 int pages_per_rd_bio;
399 int ret;
400
401 /*
402 * the setting of pages_per_rd_bio is correct for scrub but might
403 * be wrong for the dev_replace code where we might read from
404 * different devices in the initial huge bios. However, that
405 * code is able to correctly handle the case when adding a page
406 * to a bio fails.
407 */
408 if (dev->bdev)
409 pages_per_rd_bio = min_t(int, SCRUB_PAGES_PER_RD_BIO,
410 bio_get_nr_vecs(dev->bdev));
411 else
412 pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
413 sctx = kzalloc(sizeof(*sctx), GFP_NOFS);
414 if (!sctx)
415 goto nomem;
416 sctx->is_dev_replace = is_dev_replace;
417 sctx->pages_per_rd_bio = pages_per_rd_bio;
418 sctx->curr = -1;
419 sctx->dev_root = dev->dev_root;
420 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
421 struct scrub_bio *sbio;
422
423 sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
424 if (!sbio)
425 goto nomem;
426 sctx->bios[i] = sbio;
427
428 sbio->index = i;
429 sbio->sctx = sctx;
430 sbio->page_count = 0;
431 btrfs_init_work(&sbio->work, scrub_bio_end_io_worker,
432 NULL, NULL);
433
434 if (i != SCRUB_BIOS_PER_SCTX - 1)
435 sctx->bios[i]->next_free = i + 1;
436 else
437 sctx->bios[i]->next_free = -1;
438 }
439 sctx->first_free = 0;
440 sctx->nodesize = dev->dev_root->nodesize;
441 sctx->leafsize = dev->dev_root->leafsize;
442 sctx->sectorsize = dev->dev_root->sectorsize;
443 atomic_set(&sctx->bios_in_flight, 0);
444 atomic_set(&sctx->workers_pending, 0);
445 atomic_set(&sctx->cancel_req, 0);
446 sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
447 INIT_LIST_HEAD(&sctx->csum_list);
448
449 spin_lock_init(&sctx->list_lock);
450 spin_lock_init(&sctx->stat_lock);
451 init_waitqueue_head(&sctx->list_wait);
452
453 ret = scrub_setup_wr_ctx(sctx, &sctx->wr_ctx, fs_info,
454 fs_info->dev_replace.tgtdev, is_dev_replace);
455 if (ret) {
456 scrub_free_ctx(sctx);
457 return ERR_PTR(ret);
458 }
459 return sctx;
460
461nomem:
462 scrub_free_ctx(sctx);
463 return ERR_PTR(-ENOMEM);
464}
465
466static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
467 void *warn_ctx)
468{
469 u64 isize;
470 u32 nlink;
471 int ret;
472 int i;
473 struct extent_buffer *eb;
474 struct btrfs_inode_item *inode_item;
475 struct scrub_warning *swarn = warn_ctx;
476 struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
477 struct inode_fs_paths *ipath = NULL;
478 struct btrfs_root *local_root;
479 struct btrfs_key root_key;
480
481 root_key.objectid = root;
482 root_key.type = BTRFS_ROOT_ITEM_KEY;
483 root_key.offset = (u64)-1;
484 local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
485 if (IS_ERR(local_root)) {
486 ret = PTR_ERR(local_root);
487 goto err;
488 }
489
490 ret = inode_item_info(inum, 0, local_root, swarn->path);
491 if (ret) {
492 btrfs_release_path(swarn->path);
493 goto err;
494 }
495
496 eb = swarn->path->nodes[0];
497 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
498 struct btrfs_inode_item);
499 isize = btrfs_inode_size(eb, inode_item);
500 nlink = btrfs_inode_nlink(eb, inode_item);
501 btrfs_release_path(swarn->path);
502
503 ipath = init_ipath(4096, local_root, swarn->path);
504 if (IS_ERR(ipath)) {
505 ret = PTR_ERR(ipath);
506 ipath = NULL;
507 goto err;
508 }
509 ret = paths_from_inode(inum, ipath);
510
511 if (ret < 0)
512 goto err;
513
514 /*
515 * we deliberately ignore the bit ipath might have been too small to
516 * hold all of the paths here
517 */
518 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
519 printk_in_rcu(KERN_WARNING "BTRFS: %s at logical %llu on dev "
520 "%s, sector %llu, root %llu, inode %llu, offset %llu, "
521 "length %llu, links %u (path: %s)\n", swarn->errstr,
522 swarn->logical, rcu_str_deref(swarn->dev->name),
523 (unsigned long long)swarn->sector, root, inum, offset,
524 min(isize - offset, (u64)PAGE_SIZE), nlink,
525 (char *)(unsigned long)ipath->fspath->val[i]);
526
527 free_ipath(ipath);
528 return 0;
529
530err:
531 printk_in_rcu(KERN_WARNING "BTRFS: %s at logical %llu on dev "
532 "%s, sector %llu, root %llu, inode %llu, offset %llu: path "
533 "resolving failed with ret=%d\n", swarn->errstr,
534 swarn->logical, rcu_str_deref(swarn->dev->name),
535 (unsigned long long)swarn->sector, root, inum, offset, ret);
536
537 free_ipath(ipath);
538 return 0;
539}
540
541static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
542{
543 struct btrfs_device *dev;
544 struct btrfs_fs_info *fs_info;
545 struct btrfs_path *path;
546 struct btrfs_key found_key;
547 struct extent_buffer *eb;
548 struct btrfs_extent_item *ei;
549 struct scrub_warning swarn;
550 unsigned long ptr = 0;
551 u64 extent_item_pos;
552 u64 flags = 0;
553 u64 ref_root;
554 u32 item_size;
555 u8 ref_level;
556 const int bufsize = 4096;
557 int ret;
558
559 WARN_ON(sblock->page_count < 1);
560 dev = sblock->pagev[0]->dev;
561 fs_info = sblock->sctx->dev_root->fs_info;
562
563 path = btrfs_alloc_path();
564
565 swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
566 swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
567 swarn.sector = (sblock->pagev[0]->physical) >> 9;
568 swarn.logical = sblock->pagev[0]->logical;
569 swarn.errstr = errstr;
570 swarn.dev = NULL;
571 swarn.msg_bufsize = bufsize;
572 swarn.scratch_bufsize = bufsize;
573
574 if (!path || !swarn.scratch_buf || !swarn.msg_buf)
575 goto out;
576
577 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
578 &flags);
579 if (ret < 0)
580 goto out;
581
582 extent_item_pos = swarn.logical - found_key.objectid;
583 swarn.extent_item_size = found_key.offset;
584
585 eb = path->nodes[0];
586 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
587 item_size = btrfs_item_size_nr(eb, path->slots[0]);
588
589 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
590 do {
591 ret = tree_backref_for_extent(&ptr, eb, ei, item_size,
592 &ref_root, &ref_level);
593 printk_in_rcu(KERN_WARNING
594 "BTRFS: %s at logical %llu on dev %s, "
595 "sector %llu: metadata %s (level %d) in tree "
596 "%llu\n", errstr, swarn.logical,
597 rcu_str_deref(dev->name),
598 (unsigned long long)swarn.sector,
599 ref_level ? "node" : "leaf",
600 ret < 0 ? -1 : ref_level,
601 ret < 0 ? -1 : ref_root);
602 } while (ret != 1);
603 btrfs_release_path(path);
604 } else {
605 btrfs_release_path(path);
606 swarn.path = path;
607 swarn.dev = dev;
608 iterate_extent_inodes(fs_info, found_key.objectid,
609 extent_item_pos, 1,
610 scrub_print_warning_inode, &swarn);
611 }
612
613out:
614 btrfs_free_path(path);
615 kfree(swarn.scratch_buf);
616 kfree(swarn.msg_buf);
617}
618
619static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
620{
621 struct page *page = NULL;
622 unsigned long index;
623 struct scrub_fixup_nodatasum *fixup = fixup_ctx;
624 int ret;
625 int corrected = 0;
626 struct btrfs_key key;
627 struct inode *inode = NULL;
628 struct btrfs_fs_info *fs_info;
629 u64 end = offset + PAGE_SIZE - 1;
630 struct btrfs_root *local_root;
631 int srcu_index;
632
633 key.objectid = root;
634 key.type = BTRFS_ROOT_ITEM_KEY;
635 key.offset = (u64)-1;
636
637 fs_info = fixup->root->fs_info;
638 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
639
640 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
641 if (IS_ERR(local_root)) {
642 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
643 return PTR_ERR(local_root);
644 }
645
646 key.type = BTRFS_INODE_ITEM_KEY;
647 key.objectid = inum;
648 key.offset = 0;
649 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
650 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
651 if (IS_ERR(inode))
652 return PTR_ERR(inode);
653
654 index = offset >> PAGE_CACHE_SHIFT;
655
656 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
657 if (!page) {
658 ret = -ENOMEM;
659 goto out;
660 }
661
662 if (PageUptodate(page)) {
663 if (PageDirty(page)) {
664 /*
665 * we need to write the data to the defect sector. the
666 * data that was in that sector is not in memory,
667 * because the page was modified. we must not write the
668 * modified page to that sector.
669 *
670 * TODO: what could be done here: wait for the delalloc
671 * runner to write out that page (might involve
672 * COW) and see whether the sector is still
673 * referenced afterwards.
674 *
675 * For the meantime, we'll treat this error
676 * incorrectable, although there is a chance that a
677 * later scrub will find the bad sector again and that
678 * there's no dirty page in memory, then.
679 */
680 ret = -EIO;
681 goto out;
682 }
683 fs_info = BTRFS_I(inode)->root->fs_info;
684 ret = repair_io_failure(fs_info, offset, PAGE_SIZE,
685 fixup->logical, page,
686 fixup->mirror_num);
687 unlock_page(page);
688 corrected = !ret;
689 } else {
690 /*
691 * we need to get good data first. the general readpage path
692 * will call repair_io_failure for us, we just have to make
693 * sure we read the bad mirror.
694 */
695 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
696 EXTENT_DAMAGED, GFP_NOFS);
697 if (ret) {
698 /* set_extent_bits should give proper error */
699 WARN_ON(ret > 0);
700 if (ret > 0)
701 ret = -EFAULT;
702 goto out;
703 }
704
705 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
706 btrfs_get_extent,
707 fixup->mirror_num);
708 wait_on_page_locked(page);
709
710 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
711 end, EXTENT_DAMAGED, 0, NULL);
712 if (!corrected)
713 clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
714 EXTENT_DAMAGED, GFP_NOFS);
715 }
716
717out:
718 if (page)
719 put_page(page);
720 if (inode)
721 iput(inode);
722
723 if (ret < 0)
724 return ret;
725
726 if (ret == 0 && corrected) {
727 /*
728 * we only need to call readpage for one of the inodes belonging
729 * to this extent. so make iterate_extent_inodes stop
730 */
731 return 1;
732 }
733
734 return -EIO;
735}
736
737static void scrub_fixup_nodatasum(struct btrfs_work *work)
738{
739 int ret;
740 struct scrub_fixup_nodatasum *fixup;
741 struct scrub_ctx *sctx;
742 struct btrfs_trans_handle *trans = NULL;
743 struct btrfs_path *path;
744 int uncorrectable = 0;
745
746 fixup = container_of(work, struct scrub_fixup_nodatasum, work);
747 sctx = fixup->sctx;
748
749 path = btrfs_alloc_path();
750 if (!path) {
751 spin_lock(&sctx->stat_lock);
752 ++sctx->stat.malloc_errors;
753 spin_unlock(&sctx->stat_lock);
754 uncorrectable = 1;
755 goto out;
756 }
757
758 trans = btrfs_join_transaction(fixup->root);
759 if (IS_ERR(trans)) {
760 uncorrectable = 1;
761 goto out;
762 }
763
764 /*
765 * the idea is to trigger a regular read through the standard path. we
766 * read a page from the (failed) logical address by specifying the
767 * corresponding copynum of the failed sector. thus, that readpage is
768 * expected to fail.
769 * that is the point where on-the-fly error correction will kick in
770 * (once it's finished) and rewrite the failed sector if a good copy
771 * can be found.
772 */
773 ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
774 path, scrub_fixup_readpage,
775 fixup);
776 if (ret < 0) {
777 uncorrectable = 1;
778 goto out;
779 }
780 WARN_ON(ret != 1);
781
782 spin_lock(&sctx->stat_lock);
783 ++sctx->stat.corrected_errors;
784 spin_unlock(&sctx->stat_lock);
785
786out:
787 if (trans && !IS_ERR(trans))
788 btrfs_end_transaction(trans, fixup->root);
789 if (uncorrectable) {
790 spin_lock(&sctx->stat_lock);
791 ++sctx->stat.uncorrectable_errors;
792 spin_unlock(&sctx->stat_lock);
793 btrfs_dev_replace_stats_inc(
794 &sctx->dev_root->fs_info->dev_replace.
795 num_uncorrectable_read_errors);
796 printk_ratelimited_in_rcu(KERN_ERR "BTRFS: "
797 "unable to fixup (nodatasum) error at logical %llu on dev %s\n",
798 fixup->logical, rcu_str_deref(fixup->dev->name));
799 }
800
801 btrfs_free_path(path);
802 kfree(fixup);
803
804 scrub_pending_trans_workers_dec(sctx);
805}
806
807/*
808 * scrub_handle_errored_block gets called when either verification of the
809 * pages failed or the bio failed to read, e.g. with EIO. In the latter
810 * case, this function handles all pages in the bio, even though only one
811 * may be bad.
812 * The goal of this function is to repair the errored block by using the
813 * contents of one of the mirrors.
814 */
815static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
816{
817 struct scrub_ctx *sctx = sblock_to_check->sctx;
818 struct btrfs_device *dev;
819 struct btrfs_fs_info *fs_info;
820 u64 length;
821 u64 logical;
822 u64 generation;
823 unsigned int failed_mirror_index;
824 unsigned int is_metadata;
825 unsigned int have_csum;
826 u8 *csum;
827 struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
828 struct scrub_block *sblock_bad;
829 int ret;
830 int mirror_index;
831 int page_num;
832 int success;
833 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
834 DEFAULT_RATELIMIT_BURST);
835
836 BUG_ON(sblock_to_check->page_count < 1);
837 fs_info = sctx->dev_root->fs_info;
838 if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
839 /*
840 * if we find an error in a super block, we just report it.
841 * They will get written with the next transaction commit
842 * anyway
843 */
844 spin_lock(&sctx->stat_lock);
845 ++sctx->stat.super_errors;
846 spin_unlock(&sctx->stat_lock);
847 return 0;
848 }
849 length = sblock_to_check->page_count * PAGE_SIZE;
850 logical = sblock_to_check->pagev[0]->logical;
851 generation = sblock_to_check->pagev[0]->generation;
852 BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
853 failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
854 is_metadata = !(sblock_to_check->pagev[0]->flags &
855 BTRFS_EXTENT_FLAG_DATA);
856 have_csum = sblock_to_check->pagev[0]->have_csum;
857 csum = sblock_to_check->pagev[0]->csum;
858 dev = sblock_to_check->pagev[0]->dev;
859
860 if (sctx->is_dev_replace && !is_metadata && !have_csum) {
861 sblocks_for_recheck = NULL;
862 goto nodatasum_case;
863 }
864
865 /*
866 * read all mirrors one after the other. This includes to
867 * re-read the extent or metadata block that failed (that was
868 * the cause that this fixup code is called) another time,
869 * page by page this time in order to know which pages
870 * caused I/O errors and which ones are good (for all mirrors).
871 * It is the goal to handle the situation when more than one
872 * mirror contains I/O errors, but the errors do not
873 * overlap, i.e. the data can be repaired by selecting the
874 * pages from those mirrors without I/O error on the
875 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
876 * would be that mirror #1 has an I/O error on the first page,
877 * the second page is good, and mirror #2 has an I/O error on
878 * the second page, but the first page is good.
879 * Then the first page of the first mirror can be repaired by
880 * taking the first page of the second mirror, and the
881 * second page of the second mirror can be repaired by
882 * copying the contents of the 2nd page of the 1st mirror.
883 * One more note: if the pages of one mirror contain I/O
884 * errors, the checksum cannot be verified. In order to get
885 * the best data for repairing, the first attempt is to find
886 * a mirror without I/O errors and with a validated checksum.
887 * Only if this is not possible, the pages are picked from
888 * mirrors with I/O errors without considering the checksum.
889 * If the latter is the case, at the end, the checksum of the
890 * repaired area is verified in order to correctly maintain
891 * the statistics.
892 */
893
894 sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
895 sizeof(*sblocks_for_recheck),
896 GFP_NOFS);
897 if (!sblocks_for_recheck) {
898 spin_lock(&sctx->stat_lock);
899 sctx->stat.malloc_errors++;
900 sctx->stat.read_errors++;
901 sctx->stat.uncorrectable_errors++;
902 spin_unlock(&sctx->stat_lock);
903 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
904 goto out;
905 }
906
907 /* setup the context, map the logical blocks and alloc the pages */
908 ret = scrub_setup_recheck_block(sctx, fs_info, sblock_to_check, length,
909 logical, sblocks_for_recheck);
910 if (ret) {
911 spin_lock(&sctx->stat_lock);
912 sctx->stat.read_errors++;
913 sctx->stat.uncorrectable_errors++;
914 spin_unlock(&sctx->stat_lock);
915 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
916 goto out;
917 }
918 BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
919 sblock_bad = sblocks_for_recheck + failed_mirror_index;
920
921 /* build and submit the bios for the failed mirror, check checksums */
922 scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
923 csum, generation, sctx->csum_size);
924
925 if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
926 sblock_bad->no_io_error_seen) {
927 /*
928 * the error disappeared after reading page by page, or
929 * the area was part of a huge bio and other parts of the
930 * bio caused I/O errors, or the block layer merged several
931 * read requests into one and the error is caused by a
932 * different bio (usually one of the two latter cases is
933 * the cause)
934 */
935 spin_lock(&sctx->stat_lock);
936 sctx->stat.unverified_errors++;
937 spin_unlock(&sctx->stat_lock);
938
939 if (sctx->is_dev_replace)
940 scrub_write_block_to_dev_replace(sblock_bad);
941 goto out;
942 }
943
944 if (!sblock_bad->no_io_error_seen) {
945 spin_lock(&sctx->stat_lock);
946 sctx->stat.read_errors++;
947 spin_unlock(&sctx->stat_lock);
948 if (__ratelimit(&_rs))
949 scrub_print_warning("i/o error", sblock_to_check);
950 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
951 } else if (sblock_bad->checksum_error) {
952 spin_lock(&sctx->stat_lock);
953 sctx->stat.csum_errors++;
954 spin_unlock(&sctx->stat_lock);
955 if (__ratelimit(&_rs))
956 scrub_print_warning("checksum error", sblock_to_check);
957 btrfs_dev_stat_inc_and_print(dev,
958 BTRFS_DEV_STAT_CORRUPTION_ERRS);
959 } else if (sblock_bad->header_error) {
960 spin_lock(&sctx->stat_lock);
961 sctx->stat.verify_errors++;
962 spin_unlock(&sctx->stat_lock);
963 if (__ratelimit(&_rs))
964 scrub_print_warning("checksum/header error",
965 sblock_to_check);
966 if (sblock_bad->generation_error)
967 btrfs_dev_stat_inc_and_print(dev,
968 BTRFS_DEV_STAT_GENERATION_ERRS);
969 else
970 btrfs_dev_stat_inc_and_print(dev,
971 BTRFS_DEV_STAT_CORRUPTION_ERRS);
972 }
973
974 if (sctx->readonly) {
975 ASSERT(!sctx->is_dev_replace);
976 goto out;
977 }
978
979 if (!is_metadata && !have_csum) {
980 struct scrub_fixup_nodatasum *fixup_nodatasum;
981
982nodatasum_case:
983 WARN_ON(sctx->is_dev_replace);
984
985 /*
986 * !is_metadata and !have_csum, this means that the data
987 * might not be COW'ed, that it might be modified
988 * concurrently. The general strategy to work on the
989 * commit root does not help in the case when COW is not
990 * used.
991 */
992 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
993 if (!fixup_nodatasum)
994 goto did_not_correct_error;
995 fixup_nodatasum->sctx = sctx;
996 fixup_nodatasum->dev = dev;
997 fixup_nodatasum->logical = logical;
998 fixup_nodatasum->root = fs_info->extent_root;
999 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
1000 scrub_pending_trans_workers_inc(sctx);
1001 btrfs_init_work(&fixup_nodatasum->work, scrub_fixup_nodatasum,
1002 NULL, NULL);
1003 btrfs_queue_work(fs_info->scrub_workers,
1004 &fixup_nodatasum->work);
1005 goto out;
1006 }
1007
1008 /*
1009 * now build and submit the bios for the other mirrors, check
1010 * checksums.
1011 * First try to pick the mirror which is completely without I/O
1012 * errors and also does not have a checksum error.
1013 * If one is found, and if a checksum is present, the full block
1014 * that is known to contain an error is rewritten. Afterwards
1015 * the block is known to be corrected.
1016 * If a mirror is found which is completely correct, and no
1017 * checksum is present, only those pages are rewritten that had
1018 * an I/O error in the block to be repaired, since it cannot be
1019 * determined, which copy of the other pages is better (and it
1020 * could happen otherwise that a correct page would be
1021 * overwritten by a bad one).
1022 */
1023 for (mirror_index = 0;
1024 mirror_index < BTRFS_MAX_MIRRORS &&
1025 sblocks_for_recheck[mirror_index].page_count > 0;
1026 mirror_index++) {
1027 struct scrub_block *sblock_other;
1028
1029 if (mirror_index == failed_mirror_index)
1030 continue;
1031 sblock_other = sblocks_for_recheck + mirror_index;
1032
1033 /* build and submit the bios, check checksums */
1034 scrub_recheck_block(fs_info, sblock_other, is_metadata,
1035 have_csum, csum, generation,
1036 sctx->csum_size);
1037
1038 if (!sblock_other->header_error &&
1039 !sblock_other->checksum_error &&
1040 sblock_other->no_io_error_seen) {
1041 if (sctx->is_dev_replace) {
1042 scrub_write_block_to_dev_replace(sblock_other);
1043 } else {
1044 int force_write = is_metadata || have_csum;
1045
1046 ret = scrub_repair_block_from_good_copy(
1047 sblock_bad, sblock_other,
1048 force_write);
1049 }
1050 if (0 == ret)
1051 goto corrected_error;
1052 }
1053 }
1054
1055 /*
1056 * for dev_replace, pick good pages and write to the target device.
1057 */
1058 if (sctx->is_dev_replace) {
1059 success = 1;
1060 for (page_num = 0; page_num < sblock_bad->page_count;
1061 page_num++) {
1062 int sub_success;
1063
1064 sub_success = 0;
1065 for (mirror_index = 0;
1066 mirror_index < BTRFS_MAX_MIRRORS &&
1067 sblocks_for_recheck[mirror_index].page_count > 0;
1068 mirror_index++) {
1069 struct scrub_block *sblock_other =
1070 sblocks_for_recheck + mirror_index;
1071 struct scrub_page *page_other =
1072 sblock_other->pagev[page_num];
1073
1074 if (!page_other->io_error) {
1075 ret = scrub_write_page_to_dev_replace(
1076 sblock_other, page_num);
1077 if (ret == 0) {
1078 /* succeeded for this page */
1079 sub_success = 1;
1080 break;
1081 } else {
1082 btrfs_dev_replace_stats_inc(
1083 &sctx->dev_root->
1084 fs_info->dev_replace.
1085 num_write_errors);
1086 }
1087 }
1088 }
1089
1090 if (!sub_success) {
1091 /*
1092 * did not find a mirror to fetch the page
1093 * from. scrub_write_page_to_dev_replace()
1094 * handles this case (page->io_error), by
1095 * filling the block with zeros before
1096 * submitting the write request
1097 */
1098 success = 0;
1099 ret = scrub_write_page_to_dev_replace(
1100 sblock_bad, page_num);
1101 if (ret)
1102 btrfs_dev_replace_stats_inc(
1103 &sctx->dev_root->fs_info->
1104 dev_replace.num_write_errors);
1105 }
1106 }
1107
1108 goto out;
1109 }
1110
1111 /*
1112 * for regular scrub, repair those pages that are errored.
1113 * In case of I/O errors in the area that is supposed to be
1114 * repaired, continue by picking good copies of those pages.
1115 * Select the good pages from mirrors to rewrite bad pages from
1116 * the area to fix. Afterwards verify the checksum of the block
1117 * that is supposed to be repaired. This verification step is
1118 * only done for the purpose of statistic counting and for the
1119 * final scrub report, whether errors remain.
1120 * A perfect algorithm could make use of the checksum and try
1121 * all possible combinations of pages from the different mirrors
1122 * until the checksum verification succeeds. For example, when
1123 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1124 * of mirror #2 is readable but the final checksum test fails,
1125 * then the 2nd page of mirror #3 could be tried, whether now
1126 * the final checksum succeedes. But this would be a rare
1127 * exception and is therefore not implemented. At least it is
1128 * avoided that the good copy is overwritten.
1129 * A more useful improvement would be to pick the sectors
1130 * without I/O error based on sector sizes (512 bytes on legacy
1131 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1132 * mirror could be repaired by taking 512 byte of a different
1133 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1134 * area are unreadable.
1135 */
1136
1137 /* can only fix I/O errors from here on */
1138 if (sblock_bad->no_io_error_seen)
1139 goto did_not_correct_error;
1140
1141 success = 1;
1142 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1143 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1144
1145 if (!page_bad->io_error)
1146 continue;
1147
1148 for (mirror_index = 0;
1149 mirror_index < BTRFS_MAX_MIRRORS &&
1150 sblocks_for_recheck[mirror_index].page_count > 0;
1151 mirror_index++) {
1152 struct scrub_block *sblock_other = sblocks_for_recheck +
1153 mirror_index;
1154 struct scrub_page *page_other = sblock_other->pagev[
1155 page_num];
1156
1157 if (!page_other->io_error) {
1158 ret = scrub_repair_page_from_good_copy(
1159 sblock_bad, sblock_other, page_num, 0);
1160 if (0 == ret) {
1161 page_bad->io_error = 0;
1162 break; /* succeeded for this page */
1163 }
1164 }
1165 }
1166
1167 if (page_bad->io_error) {
1168 /* did not find a mirror to copy the page from */
1169 success = 0;
1170 }
1171 }
1172
1173 if (success) {
1174 if (is_metadata || have_csum) {
1175 /*
1176 * need to verify the checksum now that all
1177 * sectors on disk are repaired (the write
1178 * request for data to be repaired is on its way).
1179 * Just be lazy and use scrub_recheck_block()
1180 * which re-reads the data before the checksum
1181 * is verified, but most likely the data comes out
1182 * of the page cache.
1183 */
1184 scrub_recheck_block(fs_info, sblock_bad,
1185 is_metadata, have_csum, csum,
1186 generation, sctx->csum_size);
1187 if (!sblock_bad->header_error &&
1188 !sblock_bad->checksum_error &&
1189 sblock_bad->no_io_error_seen)
1190 goto corrected_error;
1191 else
1192 goto did_not_correct_error;
1193 } else {
1194corrected_error:
1195 spin_lock(&sctx->stat_lock);
1196 sctx->stat.corrected_errors++;
1197 spin_unlock(&sctx->stat_lock);
1198 printk_ratelimited_in_rcu(KERN_ERR
1199 "BTRFS: fixed up error at logical %llu on dev %s\n",
1200 logical, rcu_str_deref(dev->name));
1201 }
1202 } else {
1203did_not_correct_error:
1204 spin_lock(&sctx->stat_lock);
1205 sctx->stat.uncorrectable_errors++;
1206 spin_unlock(&sctx->stat_lock);
1207 printk_ratelimited_in_rcu(KERN_ERR
1208 "BTRFS: unable to fixup (regular) error at logical %llu on dev %s\n",
1209 logical, rcu_str_deref(dev->name));
1210 }
1211
1212out:
1213 if (sblocks_for_recheck) {
1214 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1215 mirror_index++) {
1216 struct scrub_block *sblock = sblocks_for_recheck +
1217 mirror_index;
1218 int page_index;
1219
1220 for (page_index = 0; page_index < sblock->page_count;
1221 page_index++) {
1222 sblock->pagev[page_index]->sblock = NULL;
1223 scrub_page_put(sblock->pagev[page_index]);
1224 }
1225 }
1226 kfree(sblocks_for_recheck);
1227 }
1228
1229 return 0;
1230}
1231
1232static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
1233 struct btrfs_fs_info *fs_info,
1234 struct scrub_block *original_sblock,
1235 u64 length, u64 logical,
1236 struct scrub_block *sblocks_for_recheck)
1237{
1238 int page_index;
1239 int mirror_index;
1240 int ret;
1241
1242 /*
1243 * note: the two members ref_count and outstanding_pages
1244 * are not used (and not set) in the blocks that are used for
1245 * the recheck procedure
1246 */
1247
1248 page_index = 0;
1249 while (length > 0) {
1250 u64 sublen = min_t(u64, length, PAGE_SIZE);
1251 u64 mapped_length = sublen;
1252 struct btrfs_bio *bbio = NULL;
1253
1254 /*
1255 * with a length of PAGE_SIZE, each returned stripe
1256 * represents one mirror
1257 */
1258 ret = btrfs_map_block(fs_info, REQ_GET_READ_MIRRORS, logical,
1259 &mapped_length, &bbio, 0);
1260 if (ret || !bbio || mapped_length < sublen) {
1261 kfree(bbio);
1262 return -EIO;
1263 }
1264
1265 BUG_ON(page_index >= SCRUB_PAGES_PER_RD_BIO);
1266 for (mirror_index = 0; mirror_index < (int)bbio->num_stripes;
1267 mirror_index++) {
1268 struct scrub_block *sblock;
1269 struct scrub_page *page;
1270
1271 if (mirror_index >= BTRFS_MAX_MIRRORS)
1272 continue;
1273
1274 sblock = sblocks_for_recheck + mirror_index;
1275 sblock->sctx = sctx;
1276 page = kzalloc(sizeof(*page), GFP_NOFS);
1277 if (!page) {
1278leave_nomem:
1279 spin_lock(&sctx->stat_lock);
1280 sctx->stat.malloc_errors++;
1281 spin_unlock(&sctx->stat_lock);
1282 kfree(bbio);
1283 return -ENOMEM;
1284 }
1285 scrub_page_get(page);
1286 sblock->pagev[page_index] = page;
1287 page->logical = logical;
1288 page->physical = bbio->stripes[mirror_index].physical;
1289 BUG_ON(page_index >= original_sblock->page_count);
1290 page->physical_for_dev_replace =
1291 original_sblock->pagev[page_index]->
1292 physical_for_dev_replace;
1293 /* for missing devices, dev->bdev is NULL */
1294 page->dev = bbio->stripes[mirror_index].dev;
1295 page->mirror_num = mirror_index + 1;
1296 sblock->page_count++;
1297 page->page = alloc_page(GFP_NOFS);
1298 if (!page->page)
1299 goto leave_nomem;
1300 }
1301 kfree(bbio);
1302 length -= sublen;
1303 logical += sublen;
1304 page_index++;
1305 }
1306
1307 return 0;
1308}
1309
1310/*
1311 * this function will check the on disk data for checksum errors, header
1312 * errors and read I/O errors. If any I/O errors happen, the exact pages
1313 * which are errored are marked as being bad. The goal is to enable scrub
1314 * to take those pages that are not errored from all the mirrors so that
1315 * the pages that are errored in the just handled mirror can be repaired.
1316 */
1317static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1318 struct scrub_block *sblock, int is_metadata,
1319 int have_csum, u8 *csum, u64 generation,
1320 u16 csum_size)
1321{
1322 int page_num;
1323
1324 sblock->no_io_error_seen = 1;
1325 sblock->header_error = 0;
1326 sblock->checksum_error = 0;
1327
1328 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1329 struct bio *bio;
1330 struct scrub_page *page = sblock->pagev[page_num];
1331
1332 if (page->dev->bdev == NULL) {
1333 page->io_error = 1;
1334 sblock->no_io_error_seen = 0;
1335 continue;
1336 }
1337
1338 WARN_ON(!page->page);
1339 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1340 if (!bio) {
1341 page->io_error = 1;
1342 sblock->no_io_error_seen = 0;
1343 continue;
1344 }
1345 bio->bi_bdev = page->dev->bdev;
1346 bio->bi_iter.bi_sector = page->physical >> 9;
1347
1348 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1349 if (btrfsic_submit_bio_wait(READ, bio))
1350 sblock->no_io_error_seen = 0;
1351
1352 bio_put(bio);
1353 }
1354
1355 if (sblock->no_io_error_seen)
1356 scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
1357 have_csum, csum, generation,
1358 csum_size);
1359
1360 return;
1361}
1362
1363static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1364 struct scrub_block *sblock,
1365 int is_metadata, int have_csum,
1366 const u8 *csum, u64 generation,
1367 u16 csum_size)
1368{
1369 int page_num;
1370 u8 calculated_csum[BTRFS_CSUM_SIZE];
1371 u32 crc = ~(u32)0;
1372 void *mapped_buffer;
1373
1374 WARN_ON(!sblock->pagev[0]->page);
1375 if (is_metadata) {
1376 struct btrfs_header *h;
1377
1378 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1379 h = (struct btrfs_header *)mapped_buffer;
1380
1381 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h) ||
1382 memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) ||
1383 memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1384 BTRFS_UUID_SIZE)) {
1385 sblock->header_error = 1;
1386 } else if (generation != btrfs_stack_header_generation(h)) {
1387 sblock->header_error = 1;
1388 sblock->generation_error = 1;
1389 }
1390 csum = h->csum;
1391 } else {
1392 if (!have_csum)
1393 return;
1394
1395 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1396 }
1397
1398 for (page_num = 0;;) {
1399 if (page_num == 0 && is_metadata)
1400 crc = btrfs_csum_data(
1401 ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1402 crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1403 else
1404 crc = btrfs_csum_data(mapped_buffer, crc, PAGE_SIZE);
1405
1406 kunmap_atomic(mapped_buffer);
1407 page_num++;
1408 if (page_num >= sblock->page_count)
1409 break;
1410 WARN_ON(!sblock->pagev[page_num]->page);
1411
1412 mapped_buffer = kmap_atomic(sblock->pagev[page_num]->page);
1413 }
1414
1415 btrfs_csum_final(crc, calculated_csum);
1416 if (memcmp(calculated_csum, csum, csum_size))
1417 sblock->checksum_error = 1;
1418}
1419
1420static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1421 struct scrub_block *sblock_good,
1422 int force_write)
1423{
1424 int page_num;
1425 int ret = 0;
1426
1427 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1428 int ret_sub;
1429
1430 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1431 sblock_good,
1432 page_num,
1433 force_write);
1434 if (ret_sub)
1435 ret = ret_sub;
1436 }
1437
1438 return ret;
1439}
1440
1441static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1442 struct scrub_block *sblock_good,
1443 int page_num, int force_write)
1444{
1445 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1446 struct scrub_page *page_good = sblock_good->pagev[page_num];
1447
1448 BUG_ON(page_bad->page == NULL);
1449 BUG_ON(page_good->page == NULL);
1450 if (force_write || sblock_bad->header_error ||
1451 sblock_bad->checksum_error || page_bad->io_error) {
1452 struct bio *bio;
1453 int ret;
1454
1455 if (!page_bad->dev->bdev) {
1456 printk_ratelimited(KERN_WARNING "BTRFS: "
1457 "scrub_repair_page_from_good_copy(bdev == NULL) "
1458 "is unexpected!\n");
1459 return -EIO;
1460 }
1461
1462 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1463 if (!bio)
1464 return -EIO;
1465 bio->bi_bdev = page_bad->dev->bdev;
1466 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1467
1468 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1469 if (PAGE_SIZE != ret) {
1470 bio_put(bio);
1471 return -EIO;
1472 }
1473
1474 if (btrfsic_submit_bio_wait(WRITE, bio)) {
1475 btrfs_dev_stat_inc_and_print(page_bad->dev,
1476 BTRFS_DEV_STAT_WRITE_ERRS);
1477 btrfs_dev_replace_stats_inc(
1478 &sblock_bad->sctx->dev_root->fs_info->
1479 dev_replace.num_write_errors);
1480 bio_put(bio);
1481 return -EIO;
1482 }
1483 bio_put(bio);
1484 }
1485
1486 return 0;
1487}
1488
1489static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1490{
1491 int page_num;
1492
1493 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1494 int ret;
1495
1496 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1497 if (ret)
1498 btrfs_dev_replace_stats_inc(
1499 &sblock->sctx->dev_root->fs_info->dev_replace.
1500 num_write_errors);
1501 }
1502}
1503
1504static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1505 int page_num)
1506{
1507 struct scrub_page *spage = sblock->pagev[page_num];
1508
1509 BUG_ON(spage->page == NULL);
1510 if (spage->io_error) {
1511 void *mapped_buffer = kmap_atomic(spage->page);
1512
1513 memset(mapped_buffer, 0, PAGE_CACHE_SIZE);
1514 flush_dcache_page(spage->page);
1515 kunmap_atomic(mapped_buffer);
1516 }
1517 return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1518}
1519
1520static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1521 struct scrub_page *spage)
1522{
1523 struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1524 struct scrub_bio *sbio;
1525 int ret;
1526
1527 mutex_lock(&wr_ctx->wr_lock);
1528again:
1529 if (!wr_ctx->wr_curr_bio) {
1530 wr_ctx->wr_curr_bio = kzalloc(sizeof(*wr_ctx->wr_curr_bio),
1531 GFP_NOFS);
1532 if (!wr_ctx->wr_curr_bio) {
1533 mutex_unlock(&wr_ctx->wr_lock);
1534 return -ENOMEM;
1535 }
1536 wr_ctx->wr_curr_bio->sctx = sctx;
1537 wr_ctx->wr_curr_bio->page_count = 0;
1538 }
1539 sbio = wr_ctx->wr_curr_bio;
1540 if (sbio->page_count == 0) {
1541 struct bio *bio;
1542
1543 sbio->physical = spage->physical_for_dev_replace;
1544 sbio->logical = spage->logical;
1545 sbio->dev = wr_ctx->tgtdev;
1546 bio = sbio->bio;
1547 if (!bio) {
1548 bio = btrfs_io_bio_alloc(GFP_NOFS, wr_ctx->pages_per_wr_bio);
1549 if (!bio) {
1550 mutex_unlock(&wr_ctx->wr_lock);
1551 return -ENOMEM;
1552 }
1553 sbio->bio = bio;
1554 }
1555
1556 bio->bi_private = sbio;
1557 bio->bi_end_io = scrub_wr_bio_end_io;
1558 bio->bi_bdev = sbio->dev->bdev;
1559 bio->bi_iter.bi_sector = sbio->physical >> 9;
1560 sbio->err = 0;
1561 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1562 spage->physical_for_dev_replace ||
1563 sbio->logical + sbio->page_count * PAGE_SIZE !=
1564 spage->logical) {
1565 scrub_wr_submit(sctx);
1566 goto again;
1567 }
1568
1569 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1570 if (ret != PAGE_SIZE) {
1571 if (sbio->page_count < 1) {
1572 bio_put(sbio->bio);
1573 sbio->bio = NULL;
1574 mutex_unlock(&wr_ctx->wr_lock);
1575 return -EIO;
1576 }
1577 scrub_wr_submit(sctx);
1578 goto again;
1579 }
1580
1581 sbio->pagev[sbio->page_count] = spage;
1582 scrub_page_get(spage);
1583 sbio->page_count++;
1584 if (sbio->page_count == wr_ctx->pages_per_wr_bio)
1585 scrub_wr_submit(sctx);
1586 mutex_unlock(&wr_ctx->wr_lock);
1587
1588 return 0;
1589}
1590
1591static void scrub_wr_submit(struct scrub_ctx *sctx)
1592{
1593 struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1594 struct scrub_bio *sbio;
1595
1596 if (!wr_ctx->wr_curr_bio)
1597 return;
1598
1599 sbio = wr_ctx->wr_curr_bio;
1600 wr_ctx->wr_curr_bio = NULL;
1601 WARN_ON(!sbio->bio->bi_bdev);
1602 scrub_pending_bio_inc(sctx);
1603 /* process all writes in a single worker thread. Then the block layer
1604 * orders the requests before sending them to the driver which
1605 * doubled the write performance on spinning disks when measured
1606 * with Linux 3.5 */
1607 btrfsic_submit_bio(WRITE, sbio->bio);
1608}
1609
1610static void scrub_wr_bio_end_io(struct bio *bio, int err)
1611{
1612 struct scrub_bio *sbio = bio->bi_private;
1613 struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
1614
1615 sbio->err = err;
1616 sbio->bio = bio;
1617
1618 btrfs_init_work(&sbio->work, scrub_wr_bio_end_io_worker, NULL, NULL);
1619 btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1620}
1621
1622static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1623{
1624 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1625 struct scrub_ctx *sctx = sbio->sctx;
1626 int i;
1627
1628 WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1629 if (sbio->err) {
1630 struct btrfs_dev_replace *dev_replace =
1631 &sbio->sctx->dev_root->fs_info->dev_replace;
1632
1633 for (i = 0; i < sbio->page_count; i++) {
1634 struct scrub_page *spage = sbio->pagev[i];
1635
1636 spage->io_error = 1;
1637 btrfs_dev_replace_stats_inc(&dev_replace->
1638 num_write_errors);
1639 }
1640 }
1641
1642 for (i = 0; i < sbio->page_count; i++)
1643 scrub_page_put(sbio->pagev[i]);
1644
1645 bio_put(sbio->bio);
1646 kfree(sbio);
1647 scrub_pending_bio_dec(sctx);
1648}
1649
1650static int scrub_checksum(struct scrub_block *sblock)
1651{
1652 u64 flags;
1653 int ret;
1654
1655 WARN_ON(sblock->page_count < 1);
1656 flags = sblock->pagev[0]->flags;
1657 ret = 0;
1658 if (flags & BTRFS_EXTENT_FLAG_DATA)
1659 ret = scrub_checksum_data(sblock);
1660 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1661 ret = scrub_checksum_tree_block(sblock);
1662 else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1663 (void)scrub_checksum_super(sblock);
1664 else
1665 WARN_ON(1);
1666 if (ret)
1667 scrub_handle_errored_block(sblock);
1668
1669 return ret;
1670}
1671
1672static int scrub_checksum_data(struct scrub_block *sblock)
1673{
1674 struct scrub_ctx *sctx = sblock->sctx;
1675 u8 csum[BTRFS_CSUM_SIZE];
1676 u8 *on_disk_csum;
1677 struct page *page;
1678 void *buffer;
1679 u32 crc = ~(u32)0;
1680 int fail = 0;
1681 u64 len;
1682 int index;
1683
1684 BUG_ON(sblock->page_count < 1);
1685 if (!sblock->pagev[0]->have_csum)
1686 return 0;
1687
1688 on_disk_csum = sblock->pagev[0]->csum;
1689 page = sblock->pagev[0]->page;
1690 buffer = kmap_atomic(page);
1691
1692 len = sctx->sectorsize;
1693 index = 0;
1694 for (;;) {
1695 u64 l = min_t(u64, len, PAGE_SIZE);
1696
1697 crc = btrfs_csum_data(buffer, crc, l);
1698 kunmap_atomic(buffer);
1699 len -= l;
1700 if (len == 0)
1701 break;
1702 index++;
1703 BUG_ON(index >= sblock->page_count);
1704 BUG_ON(!sblock->pagev[index]->page);
1705 page = sblock->pagev[index]->page;
1706 buffer = kmap_atomic(page);
1707 }
1708
1709 btrfs_csum_final(crc, csum);
1710 if (memcmp(csum, on_disk_csum, sctx->csum_size))
1711 fail = 1;
1712
1713 return fail;
1714}
1715
1716static int scrub_checksum_tree_block(struct scrub_block *sblock)
1717{
1718 struct scrub_ctx *sctx = sblock->sctx;
1719 struct btrfs_header *h;
1720 struct btrfs_root *root = sctx->dev_root;
1721 struct btrfs_fs_info *fs_info = root->fs_info;
1722 u8 calculated_csum[BTRFS_CSUM_SIZE];
1723 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1724 struct page *page;
1725 void *mapped_buffer;
1726 u64 mapped_size;
1727 void *p;
1728 u32 crc = ~(u32)0;
1729 int fail = 0;
1730 int crc_fail = 0;
1731 u64 len;
1732 int index;
1733
1734 BUG_ON(sblock->page_count < 1);
1735 page = sblock->pagev[0]->page;
1736 mapped_buffer = kmap_atomic(page);
1737 h = (struct btrfs_header *)mapped_buffer;
1738 memcpy(on_disk_csum, h->csum, sctx->csum_size);
1739
1740 /*
1741 * we don't use the getter functions here, as we
1742 * a) don't have an extent buffer and
1743 * b) the page is already kmapped
1744 */
1745
1746 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1747 ++fail;
1748
1749 if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h))
1750 ++fail;
1751
1752 if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1753 ++fail;
1754
1755 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1756 BTRFS_UUID_SIZE))
1757 ++fail;
1758
1759 WARN_ON(sctx->nodesize != sctx->leafsize);
1760 len = sctx->nodesize - BTRFS_CSUM_SIZE;
1761 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1762 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1763 index = 0;
1764 for (;;) {
1765 u64 l = min_t(u64, len, mapped_size);
1766
1767 crc = btrfs_csum_data(p, crc, l);
1768 kunmap_atomic(mapped_buffer);
1769 len -= l;
1770 if (len == 0)
1771 break;
1772 index++;
1773 BUG_ON(index >= sblock->page_count);
1774 BUG_ON(!sblock->pagev[index]->page);
1775 page = sblock->pagev[index]->page;
1776 mapped_buffer = kmap_atomic(page);
1777 mapped_size = PAGE_SIZE;
1778 p = mapped_buffer;
1779 }
1780
1781 btrfs_csum_final(crc, calculated_csum);
1782 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1783 ++crc_fail;
1784
1785 return fail || crc_fail;
1786}
1787
1788static int scrub_checksum_super(struct scrub_block *sblock)
1789{
1790 struct btrfs_super_block *s;
1791 struct scrub_ctx *sctx = sblock->sctx;
1792 struct btrfs_root *root = sctx->dev_root;
1793 struct btrfs_fs_info *fs_info = root->fs_info;
1794 u8 calculated_csum[BTRFS_CSUM_SIZE];
1795 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1796 struct page *page;
1797 void *mapped_buffer;
1798 u64 mapped_size;
1799 void *p;
1800 u32 crc = ~(u32)0;
1801 int fail_gen = 0;
1802 int fail_cor = 0;
1803 u64 len;
1804 int index;
1805
1806 BUG_ON(sblock->page_count < 1);
1807 page = sblock->pagev[0]->page;
1808 mapped_buffer = kmap_atomic(page);
1809 s = (struct btrfs_super_block *)mapped_buffer;
1810 memcpy(on_disk_csum, s->csum, sctx->csum_size);
1811
1812 if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1813 ++fail_cor;
1814
1815 if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1816 ++fail_gen;
1817
1818 if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1819 ++fail_cor;
1820
1821 len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1822 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1823 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1824 index = 0;
1825 for (;;) {
1826 u64 l = min_t(u64, len, mapped_size);
1827
1828 crc = btrfs_csum_data(p, crc, l);
1829 kunmap_atomic(mapped_buffer);
1830 len -= l;
1831 if (len == 0)
1832 break;
1833 index++;
1834 BUG_ON(index >= sblock->page_count);
1835 BUG_ON(!sblock->pagev[index]->page);
1836 page = sblock->pagev[index]->page;
1837 mapped_buffer = kmap_atomic(page);
1838 mapped_size = PAGE_SIZE;
1839 p = mapped_buffer;
1840 }
1841
1842 btrfs_csum_final(crc, calculated_csum);
1843 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1844 ++fail_cor;
1845
1846 if (fail_cor + fail_gen) {
1847 /*
1848 * if we find an error in a super block, we just report it.
1849 * They will get written with the next transaction commit
1850 * anyway
1851 */
1852 spin_lock(&sctx->stat_lock);
1853 ++sctx->stat.super_errors;
1854 spin_unlock(&sctx->stat_lock);
1855 if (fail_cor)
1856 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1857 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1858 else
1859 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1860 BTRFS_DEV_STAT_GENERATION_ERRS);
1861 }
1862
1863 return fail_cor + fail_gen;
1864}
1865
1866static void scrub_block_get(struct scrub_block *sblock)
1867{
1868 atomic_inc(&sblock->ref_count);
1869}
1870
1871static void scrub_block_put(struct scrub_block *sblock)
1872{
1873 if (atomic_dec_and_test(&sblock->ref_count)) {
1874 int i;
1875
1876 for (i = 0; i < sblock->page_count; i++)
1877 scrub_page_put(sblock->pagev[i]);
1878 kfree(sblock);
1879 }
1880}
1881
1882static void scrub_page_get(struct scrub_page *spage)
1883{
1884 atomic_inc(&spage->ref_count);
1885}
1886
1887static void scrub_page_put(struct scrub_page *spage)
1888{
1889 if (atomic_dec_and_test(&spage->ref_count)) {
1890 if (spage->page)
1891 __free_page(spage->page);
1892 kfree(spage);
1893 }
1894}
1895
1896static void scrub_submit(struct scrub_ctx *sctx)
1897{
1898 struct scrub_bio *sbio;
1899
1900 if (sctx->curr == -1)
1901 return;
1902
1903 sbio = sctx->bios[sctx->curr];
1904 sctx->curr = -1;
1905 scrub_pending_bio_inc(sctx);
1906
1907 if (!sbio->bio->bi_bdev) {
1908 /*
1909 * this case should not happen. If btrfs_map_block() is
1910 * wrong, it could happen for dev-replace operations on
1911 * missing devices when no mirrors are available, but in
1912 * this case it should already fail the mount.
1913 * This case is handled correctly (but _very_ slowly).
1914 */
1915 printk_ratelimited(KERN_WARNING
1916 "BTRFS: scrub_submit(bio bdev == NULL) is unexpected!\n");
1917 bio_endio(sbio->bio, -EIO);
1918 } else {
1919 btrfsic_submit_bio(READ, sbio->bio);
1920 }
1921}
1922
1923static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
1924 struct scrub_page *spage)
1925{
1926 struct scrub_block *sblock = spage->sblock;
1927 struct scrub_bio *sbio;
1928 int ret;
1929
1930again:
1931 /*
1932 * grab a fresh bio or wait for one to become available
1933 */
1934 while (sctx->curr == -1) {
1935 spin_lock(&sctx->list_lock);
1936 sctx->curr = sctx->first_free;
1937 if (sctx->curr != -1) {
1938 sctx->first_free = sctx->bios[sctx->curr]->next_free;
1939 sctx->bios[sctx->curr]->next_free = -1;
1940 sctx->bios[sctx->curr]->page_count = 0;
1941 spin_unlock(&sctx->list_lock);
1942 } else {
1943 spin_unlock(&sctx->list_lock);
1944 wait_event(sctx->list_wait, sctx->first_free != -1);
1945 }
1946 }
1947 sbio = sctx->bios[sctx->curr];
1948 if (sbio->page_count == 0) {
1949 struct bio *bio;
1950
1951 sbio->physical = spage->physical;
1952 sbio->logical = spage->logical;
1953 sbio->dev = spage->dev;
1954 bio = sbio->bio;
1955 if (!bio) {
1956 bio = btrfs_io_bio_alloc(GFP_NOFS, sctx->pages_per_rd_bio);
1957 if (!bio)
1958 return -ENOMEM;
1959 sbio->bio = bio;
1960 }
1961
1962 bio->bi_private = sbio;
1963 bio->bi_end_io = scrub_bio_end_io;
1964 bio->bi_bdev = sbio->dev->bdev;
1965 bio->bi_iter.bi_sector = sbio->physical >> 9;
1966 sbio->err = 0;
1967 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1968 spage->physical ||
1969 sbio->logical + sbio->page_count * PAGE_SIZE !=
1970 spage->logical ||
1971 sbio->dev != spage->dev) {
1972 scrub_submit(sctx);
1973 goto again;
1974 }
1975
1976 sbio->pagev[sbio->page_count] = spage;
1977 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1978 if (ret != PAGE_SIZE) {
1979 if (sbio->page_count < 1) {
1980 bio_put(sbio->bio);
1981 sbio->bio = NULL;
1982 return -EIO;
1983 }
1984 scrub_submit(sctx);
1985 goto again;
1986 }
1987
1988 scrub_block_get(sblock); /* one for the page added to the bio */
1989 atomic_inc(&sblock->outstanding_pages);
1990 sbio->page_count++;
1991 if (sbio->page_count == sctx->pages_per_rd_bio)
1992 scrub_submit(sctx);
1993
1994 return 0;
1995}
1996
1997static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
1998 u64 physical, struct btrfs_device *dev, u64 flags,
1999 u64 gen, int mirror_num, u8 *csum, int force,
2000 u64 physical_for_dev_replace)
2001{
2002 struct scrub_block *sblock;
2003 int index;
2004
2005 sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
2006 if (!sblock) {
2007 spin_lock(&sctx->stat_lock);
2008 sctx->stat.malloc_errors++;
2009 spin_unlock(&sctx->stat_lock);
2010 return -ENOMEM;
2011 }
2012
2013 /* one ref inside this function, plus one for each page added to
2014 * a bio later on */
2015 atomic_set(&sblock->ref_count, 1);
2016 sblock->sctx = sctx;
2017 sblock->no_io_error_seen = 1;
2018
2019 for (index = 0; len > 0; index++) {
2020 struct scrub_page *spage;
2021 u64 l = min_t(u64, len, PAGE_SIZE);
2022
2023 spage = kzalloc(sizeof(*spage), GFP_NOFS);
2024 if (!spage) {
2025leave_nomem:
2026 spin_lock(&sctx->stat_lock);
2027 sctx->stat.malloc_errors++;
2028 spin_unlock(&sctx->stat_lock);
2029 scrub_block_put(sblock);
2030 return -ENOMEM;
2031 }
2032 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2033 scrub_page_get(spage);
2034 sblock->pagev[index] = spage;
2035 spage->sblock = sblock;
2036 spage->dev = dev;
2037 spage->flags = flags;
2038 spage->generation = gen;
2039 spage->logical = logical;
2040 spage->physical = physical;
2041 spage->physical_for_dev_replace = physical_for_dev_replace;
2042 spage->mirror_num = mirror_num;
2043 if (csum) {
2044 spage->have_csum = 1;
2045 memcpy(spage->csum, csum, sctx->csum_size);
2046 } else {
2047 spage->have_csum = 0;
2048 }
2049 sblock->page_count++;
2050 spage->page = alloc_page(GFP_NOFS);
2051 if (!spage->page)
2052 goto leave_nomem;
2053 len -= l;
2054 logical += l;
2055 physical += l;
2056 physical_for_dev_replace += l;
2057 }
2058
2059 WARN_ON(sblock->page_count == 0);
2060 for (index = 0; index < sblock->page_count; index++) {
2061 struct scrub_page *spage = sblock->pagev[index];
2062 int ret;
2063
2064 ret = scrub_add_page_to_rd_bio(sctx, spage);
2065 if (ret) {
2066 scrub_block_put(sblock);
2067 return ret;
2068 }
2069 }
2070
2071 if (force)
2072 scrub_submit(sctx);
2073
2074 /* last one frees, either here or in bio completion for last page */
2075 scrub_block_put(sblock);
2076 return 0;
2077}
2078
2079static void scrub_bio_end_io(struct bio *bio, int err)
2080{
2081 struct scrub_bio *sbio = bio->bi_private;
2082 struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
2083
2084 sbio->err = err;
2085 sbio->bio = bio;
2086
2087 btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2088}
2089
2090static void scrub_bio_end_io_worker(struct btrfs_work *work)
2091{
2092 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2093 struct scrub_ctx *sctx = sbio->sctx;
2094 int i;
2095
2096 BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2097 if (sbio->err) {
2098 for (i = 0; i < sbio->page_count; i++) {
2099 struct scrub_page *spage = sbio->pagev[i];
2100
2101 spage->io_error = 1;
2102 spage->sblock->no_io_error_seen = 0;
2103 }
2104 }
2105
2106 /* now complete the scrub_block items that have all pages completed */
2107 for (i = 0; i < sbio->page_count; i++) {
2108 struct scrub_page *spage = sbio->pagev[i];
2109 struct scrub_block *sblock = spage->sblock;
2110
2111 if (atomic_dec_and_test(&sblock->outstanding_pages))
2112 scrub_block_complete(sblock);
2113 scrub_block_put(sblock);
2114 }
2115
2116 bio_put(sbio->bio);
2117 sbio->bio = NULL;
2118 spin_lock(&sctx->list_lock);
2119 sbio->next_free = sctx->first_free;
2120 sctx->first_free = sbio->index;
2121 spin_unlock(&sctx->list_lock);
2122
2123 if (sctx->is_dev_replace &&
2124 atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2125 mutex_lock(&sctx->wr_ctx.wr_lock);
2126 scrub_wr_submit(sctx);
2127 mutex_unlock(&sctx->wr_ctx.wr_lock);
2128 }
2129
2130 scrub_pending_bio_dec(sctx);
2131}
2132
2133static void scrub_block_complete(struct scrub_block *sblock)
2134{
2135 if (!sblock->no_io_error_seen) {
2136 scrub_handle_errored_block(sblock);
2137 } else {
2138 /*
2139 * if has checksum error, write via repair mechanism in
2140 * dev replace case, otherwise write here in dev replace
2141 * case.
2142 */
2143 if (!scrub_checksum(sblock) && sblock->sctx->is_dev_replace)
2144 scrub_write_block_to_dev_replace(sblock);
2145 }
2146}
2147
2148static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u64 len,
2149 u8 *csum)
2150{
2151 struct btrfs_ordered_sum *sum = NULL;
2152 unsigned long index;
2153 unsigned long num_sectors;
2154
2155 while (!list_empty(&sctx->csum_list)) {
2156 sum = list_first_entry(&sctx->csum_list,
2157 struct btrfs_ordered_sum, list);
2158 if (sum->bytenr > logical)
2159 return 0;
2160 if (sum->bytenr + sum->len > logical)
2161 break;
2162
2163 ++sctx->stat.csum_discards;
2164 list_del(&sum->list);
2165 kfree(sum);
2166 sum = NULL;
2167 }
2168 if (!sum)
2169 return 0;
2170
2171 index = ((u32)(logical - sum->bytenr)) / sctx->sectorsize;
2172 num_sectors = sum->len / sctx->sectorsize;
2173 memcpy(csum, sum->sums + index, sctx->csum_size);
2174 if (index == num_sectors - 1) {
2175 list_del(&sum->list);
2176 kfree(sum);
2177 }
2178 return 1;
2179}
2180
2181/* scrub extent tries to collect up to 64 kB for each bio */
2182static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2183 u64 physical, struct btrfs_device *dev, u64 flags,
2184 u64 gen, int mirror_num, u64 physical_for_dev_replace)
2185{
2186 int ret;
2187 u8 csum[BTRFS_CSUM_SIZE];
2188 u32 blocksize;
2189
2190 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2191 blocksize = sctx->sectorsize;
2192 spin_lock(&sctx->stat_lock);
2193 sctx->stat.data_extents_scrubbed++;
2194 sctx->stat.data_bytes_scrubbed += len;
2195 spin_unlock(&sctx->stat_lock);
2196 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2197 WARN_ON(sctx->nodesize != sctx->leafsize);
2198 blocksize = sctx->nodesize;
2199 spin_lock(&sctx->stat_lock);
2200 sctx->stat.tree_extents_scrubbed++;
2201 sctx->stat.tree_bytes_scrubbed += len;
2202 spin_unlock(&sctx->stat_lock);
2203 } else {
2204 blocksize = sctx->sectorsize;
2205 WARN_ON(1);
2206 }
2207
2208 while (len) {
2209 u64 l = min_t(u64, len, blocksize);
2210 int have_csum = 0;
2211
2212 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2213 /* push csums to sbio */
2214 have_csum = scrub_find_csum(sctx, logical, l, csum);
2215 if (have_csum == 0)
2216 ++sctx->stat.no_csum;
2217 if (sctx->is_dev_replace && !have_csum) {
2218 ret = copy_nocow_pages(sctx, logical, l,
2219 mirror_num,
2220 physical_for_dev_replace);
2221 goto behind_scrub_pages;
2222 }
2223 }
2224 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2225 mirror_num, have_csum ? csum : NULL, 0,
2226 physical_for_dev_replace);
2227behind_scrub_pages:
2228 if (ret)
2229 return ret;
2230 len -= l;
2231 logical += l;
2232 physical += l;
2233 physical_for_dev_replace += l;
2234 }
2235 return 0;
2236}
2237
2238/*
2239 * Given a physical address, this will calculate it's
2240 * logical offset. if this is a parity stripe, it will return
2241 * the most left data stripe's logical offset.
2242 *
2243 * return 0 if it is a data stripe, 1 means parity stripe.
2244 */
2245static int get_raid56_logic_offset(u64 physical, int num,
2246 struct map_lookup *map, u64 *offset)
2247{
2248 int i;
2249 int j = 0;
2250 u64 stripe_nr;
2251 u64 last_offset;
2252 int stripe_index;
2253 int rot;
2254
2255 last_offset = (physical - map->stripes[num].physical) *
2256 nr_data_stripes(map);
2257 *offset = last_offset;
2258 for (i = 0; i < nr_data_stripes(map); i++) {
2259 *offset = last_offset + i * map->stripe_len;
2260
2261 stripe_nr = *offset;
2262 do_div(stripe_nr, map->stripe_len);
2263 do_div(stripe_nr, nr_data_stripes(map));
2264
2265 /* Work out the disk rotation on this stripe-set */
2266 rot = do_div(stripe_nr, map->num_stripes);
2267 /* calculate which stripe this data locates */
2268 rot += i;
2269 stripe_index = rot % map->num_stripes;
2270 if (stripe_index == num)
2271 return 0;
2272 if (stripe_index < num)
2273 j++;
2274 }
2275 *offset = last_offset + j * map->stripe_len;
2276 return 1;
2277}
2278
2279static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2280 struct map_lookup *map,
2281 struct btrfs_device *scrub_dev,
2282 int num, u64 base, u64 length,
2283 int is_dev_replace)
2284{
2285 struct btrfs_path *path;
2286 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2287 struct btrfs_root *root = fs_info->extent_root;
2288 struct btrfs_root *csum_root = fs_info->csum_root;
2289 struct btrfs_extent_item *extent;
2290 struct blk_plug plug;
2291 u64 flags;
2292 int ret;
2293 int slot;
2294 u64 nstripes;
2295 struct extent_buffer *l;
2296 struct btrfs_key key;
2297 u64 physical;
2298 u64 logical;
2299 u64 logic_end;
2300 u64 physical_end;
2301 u64 generation;
2302 int mirror_num;
2303 struct reada_control *reada1;
2304 struct reada_control *reada2;
2305 struct btrfs_key key_start;
2306 struct btrfs_key key_end;
2307 u64 increment = map->stripe_len;
2308 u64 offset;
2309 u64 extent_logical;
2310 u64 extent_physical;
2311 u64 extent_len;
2312 struct btrfs_device *extent_dev;
2313 int extent_mirror_num;
2314 int stop_loop = 0;
2315
2316 nstripes = length;
2317 physical = map->stripes[num].physical;
2318 offset = 0;
2319 do_div(nstripes, map->stripe_len);
2320 if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
2321 offset = map->stripe_len * num;
2322 increment = map->stripe_len * map->num_stripes;
2323 mirror_num = 1;
2324 } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
2325 int factor = map->num_stripes / map->sub_stripes;
2326 offset = map->stripe_len * (num / map->sub_stripes);
2327 increment = map->stripe_len * factor;
2328 mirror_num = num % map->sub_stripes + 1;
2329 } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
2330 increment = map->stripe_len;
2331 mirror_num = num % map->num_stripes + 1;
2332 } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
2333 increment = map->stripe_len;
2334 mirror_num = num % map->num_stripes + 1;
2335 } else if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2336 BTRFS_BLOCK_GROUP_RAID6)) {
2337 get_raid56_logic_offset(physical, num, map, &offset);
2338 increment = map->stripe_len * nr_data_stripes(map);
2339 mirror_num = 1;
2340 } else {
2341 increment = map->stripe_len;
2342 mirror_num = 1;
2343 }
2344
2345 path = btrfs_alloc_path();
2346 if (!path)
2347 return -ENOMEM;
2348
2349 /*
2350 * work on commit root. The related disk blocks are static as
2351 * long as COW is applied. This means, it is save to rewrite
2352 * them to repair disk errors without any race conditions
2353 */
2354 path->search_commit_root = 1;
2355 path->skip_locking = 1;
2356
2357 /*
2358 * trigger the readahead for extent tree csum tree and wait for
2359 * completion. During readahead, the scrub is officially paused
2360 * to not hold off transaction commits
2361 */
2362 logical = base + offset;
2363 physical_end = physical + nstripes * map->stripe_len;
2364 if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2365 BTRFS_BLOCK_GROUP_RAID6)) {
2366 get_raid56_logic_offset(physical_end, num,
2367 map, &logic_end);
2368 logic_end += base;
2369 } else {
2370 logic_end = logical + increment * nstripes;
2371 }
2372 wait_event(sctx->list_wait,
2373 atomic_read(&sctx->bios_in_flight) == 0);
2374 scrub_blocked_if_needed(fs_info);
2375
2376 /* FIXME it might be better to start readahead at commit root */
2377 key_start.objectid = logical;
2378 key_start.type = BTRFS_EXTENT_ITEM_KEY;
2379 key_start.offset = (u64)0;
2380 key_end.objectid = logic_end;
2381 key_end.type = BTRFS_METADATA_ITEM_KEY;
2382 key_end.offset = (u64)-1;
2383 reada1 = btrfs_reada_add(root, &key_start, &key_end);
2384
2385 key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2386 key_start.type = BTRFS_EXTENT_CSUM_KEY;
2387 key_start.offset = logical;
2388 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2389 key_end.type = BTRFS_EXTENT_CSUM_KEY;
2390 key_end.offset = logic_end;
2391 reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
2392
2393 if (!IS_ERR(reada1))
2394 btrfs_reada_wait(reada1);
2395 if (!IS_ERR(reada2))
2396 btrfs_reada_wait(reada2);
2397
2398
2399 /*
2400 * collect all data csums for the stripe to avoid seeking during
2401 * the scrub. This might currently (crc32) end up to be about 1MB
2402 */
2403 blk_start_plug(&plug);
2404
2405 /*
2406 * now find all extents for each stripe and scrub them
2407 */
2408 ret = 0;
2409 while (physical < physical_end) {
2410 /* for raid56, we skip parity stripe */
2411 if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2412 BTRFS_BLOCK_GROUP_RAID6)) {
2413 ret = get_raid56_logic_offset(physical, num,
2414 map, &logical);
2415 logical += base;
2416 if (ret)
2417 goto skip;
2418 }
2419 /*
2420 * canceled?
2421 */
2422 if (atomic_read(&fs_info->scrub_cancel_req) ||
2423 atomic_read(&sctx->cancel_req)) {
2424 ret = -ECANCELED;
2425 goto out;
2426 }
2427 /*
2428 * check to see if we have to pause
2429 */
2430 if (atomic_read(&fs_info->scrub_pause_req)) {
2431 /* push queued extents */
2432 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2433 scrub_submit(sctx);
2434 mutex_lock(&sctx->wr_ctx.wr_lock);
2435 scrub_wr_submit(sctx);
2436 mutex_unlock(&sctx->wr_ctx.wr_lock);
2437 wait_event(sctx->list_wait,
2438 atomic_read(&sctx->bios_in_flight) == 0);
2439 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2440 scrub_blocked_if_needed(fs_info);
2441 }
2442
2443 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2444 key.type = BTRFS_METADATA_ITEM_KEY;
2445 else
2446 key.type = BTRFS_EXTENT_ITEM_KEY;
2447 key.objectid = logical;
2448 key.offset = (u64)-1;
2449
2450 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2451 if (ret < 0)
2452 goto out;
2453
2454 if (ret > 0) {
2455 ret = btrfs_previous_extent_item(root, path, 0);
2456 if (ret < 0)
2457 goto out;
2458 if (ret > 0) {
2459 /* there's no smaller item, so stick with the
2460 * larger one */
2461 btrfs_release_path(path);
2462 ret = btrfs_search_slot(NULL, root, &key,
2463 path, 0, 0);
2464 if (ret < 0)
2465 goto out;
2466 }
2467 }
2468
2469 stop_loop = 0;
2470 while (1) {
2471 u64 bytes;
2472
2473 l = path->nodes[0];
2474 slot = path->slots[0];
2475 if (slot >= btrfs_header_nritems(l)) {
2476 ret = btrfs_next_leaf(root, path);
2477 if (ret == 0)
2478 continue;
2479 if (ret < 0)
2480 goto out;
2481
2482 stop_loop = 1;
2483 break;
2484 }
2485 btrfs_item_key_to_cpu(l, &key, slot);
2486
2487 if (key.type == BTRFS_METADATA_ITEM_KEY)
2488 bytes = root->leafsize;
2489 else
2490 bytes = key.offset;
2491
2492 if (key.objectid + bytes <= logical)
2493 goto next;
2494
2495 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2496 key.type != BTRFS_METADATA_ITEM_KEY)
2497 goto next;
2498
2499 if (key.objectid >= logical + map->stripe_len) {
2500 /* out of this device extent */
2501 if (key.objectid >= logic_end)
2502 stop_loop = 1;
2503 break;
2504 }
2505
2506 extent = btrfs_item_ptr(l, slot,
2507 struct btrfs_extent_item);
2508 flags = btrfs_extent_flags(l, extent);
2509 generation = btrfs_extent_generation(l, extent);
2510
2511 if (key.objectid < logical &&
2512 (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
2513 btrfs_err(fs_info,
2514 "scrub: tree block %llu spanning "
2515 "stripes, ignored. logical=%llu",
2516 key.objectid, logical);
2517 goto next;
2518 }
2519
2520again:
2521 extent_logical = key.objectid;
2522 extent_len = bytes;
2523
2524 /*
2525 * trim extent to this stripe
2526 */
2527 if (extent_logical < logical) {
2528 extent_len -= logical - extent_logical;
2529 extent_logical = logical;
2530 }
2531 if (extent_logical + extent_len >
2532 logical + map->stripe_len) {
2533 extent_len = logical + map->stripe_len -
2534 extent_logical;
2535 }
2536
2537 extent_physical = extent_logical - logical + physical;
2538 extent_dev = scrub_dev;
2539 extent_mirror_num = mirror_num;
2540 if (is_dev_replace)
2541 scrub_remap_extent(fs_info, extent_logical,
2542 extent_len, &extent_physical,
2543 &extent_dev,
2544 &extent_mirror_num);
2545
2546 ret = btrfs_lookup_csums_range(csum_root, logical,
2547 logical + map->stripe_len - 1,
2548 &sctx->csum_list, 1);
2549 if (ret)
2550 goto out;
2551
2552 ret = scrub_extent(sctx, extent_logical, extent_len,
2553 extent_physical, extent_dev, flags,
2554 generation, extent_mirror_num,
2555 extent_logical - logical + physical);
2556 if (ret)
2557 goto out;
2558
2559 scrub_free_csums(sctx);
2560 if (extent_logical + extent_len <
2561 key.objectid + bytes) {
2562 if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2563 BTRFS_BLOCK_GROUP_RAID6)) {
2564 /*
2565 * loop until we find next data stripe
2566 * or we have finished all stripes.
2567 */
2568 do {
2569 physical += map->stripe_len;
2570 ret = get_raid56_logic_offset(
2571 physical, num,
2572 map, &logical);
2573 logical += base;
2574 } while (physical < physical_end && ret);
2575 } else {
2576 physical += map->stripe_len;
2577 logical += increment;
2578 }
2579 if (logical < key.objectid + bytes) {
2580 cond_resched();
2581 goto again;
2582 }
2583
2584 if (physical >= physical_end) {
2585 stop_loop = 1;
2586 break;
2587 }
2588 }
2589next:
2590 path->slots[0]++;
2591 }
2592 btrfs_release_path(path);
2593skip:
2594 logical += increment;
2595 physical += map->stripe_len;
2596 spin_lock(&sctx->stat_lock);
2597 if (stop_loop)
2598 sctx->stat.last_physical = map->stripes[num].physical +
2599 length;
2600 else
2601 sctx->stat.last_physical = physical;
2602 spin_unlock(&sctx->stat_lock);
2603 if (stop_loop)
2604 break;
2605 }
2606out:
2607 /* push queued extents */
2608 scrub_submit(sctx);
2609 mutex_lock(&sctx->wr_ctx.wr_lock);
2610 scrub_wr_submit(sctx);
2611 mutex_unlock(&sctx->wr_ctx.wr_lock);
2612
2613 blk_finish_plug(&plug);
2614 btrfs_free_path(path);
2615 return ret < 0 ? ret : 0;
2616}
2617
2618static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
2619 struct btrfs_device *scrub_dev,
2620 u64 chunk_tree, u64 chunk_objectid,
2621 u64 chunk_offset, u64 length,
2622 u64 dev_offset, int is_dev_replace)
2623{
2624 struct btrfs_mapping_tree *map_tree =
2625 &sctx->dev_root->fs_info->mapping_tree;
2626 struct map_lookup *map;
2627 struct extent_map *em;
2628 int i;
2629 int ret = 0;
2630
2631 read_lock(&map_tree->map_tree.lock);
2632 em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
2633 read_unlock(&map_tree->map_tree.lock);
2634
2635 if (!em)
2636 return -EINVAL;
2637
2638 map = (struct map_lookup *)em->bdev;
2639 if (em->start != chunk_offset)
2640 goto out;
2641
2642 if (em->len < length)
2643 goto out;
2644
2645 for (i = 0; i < map->num_stripes; ++i) {
2646 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
2647 map->stripes[i].physical == dev_offset) {
2648 ret = scrub_stripe(sctx, map, scrub_dev, i,
2649 chunk_offset, length,
2650 is_dev_replace);
2651 if (ret)
2652 goto out;
2653 }
2654 }
2655out:
2656 free_extent_map(em);
2657
2658 return ret;
2659}
2660
2661static noinline_for_stack
2662int scrub_enumerate_chunks(struct scrub_ctx *sctx,
2663 struct btrfs_device *scrub_dev, u64 start, u64 end,
2664 int is_dev_replace)
2665{
2666 struct btrfs_dev_extent *dev_extent = NULL;
2667 struct btrfs_path *path;
2668 struct btrfs_root *root = sctx->dev_root;
2669 struct btrfs_fs_info *fs_info = root->fs_info;
2670 u64 length;
2671 u64 chunk_tree;
2672 u64 chunk_objectid;
2673 u64 chunk_offset;
2674 int ret;
2675 int slot;
2676 struct extent_buffer *l;
2677 struct btrfs_key key;
2678 struct btrfs_key found_key;
2679 struct btrfs_block_group_cache *cache;
2680 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2681
2682 path = btrfs_alloc_path();
2683 if (!path)
2684 return -ENOMEM;
2685
2686 path->reada = 2;
2687 path->search_commit_root = 1;
2688 path->skip_locking = 1;
2689
2690 key.objectid = scrub_dev->devid;
2691 key.offset = 0ull;
2692 key.type = BTRFS_DEV_EXTENT_KEY;
2693
2694 while (1) {
2695 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2696 if (ret < 0)
2697 break;
2698 if (ret > 0) {
2699 if (path->slots[0] >=
2700 btrfs_header_nritems(path->nodes[0])) {
2701 ret = btrfs_next_leaf(root, path);
2702 if (ret)
2703 break;
2704 }
2705 }
2706
2707 l = path->nodes[0];
2708 slot = path->slots[0];
2709
2710 btrfs_item_key_to_cpu(l, &found_key, slot);
2711
2712 if (found_key.objectid != scrub_dev->devid)
2713 break;
2714
2715 if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
2716 break;
2717
2718 if (found_key.offset >= end)
2719 break;
2720
2721 if (found_key.offset < key.offset)
2722 break;
2723
2724 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2725 length = btrfs_dev_extent_length(l, dev_extent);
2726
2727 if (found_key.offset + length <= start) {
2728 key.offset = found_key.offset + length;
2729 btrfs_release_path(path);
2730 continue;
2731 }
2732
2733 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
2734 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
2735 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2736
2737 /*
2738 * get a reference on the corresponding block group to prevent
2739 * the chunk from going away while we scrub it
2740 */
2741 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2742 if (!cache) {
2743 ret = -ENOENT;
2744 break;
2745 }
2746 dev_replace->cursor_right = found_key.offset + length;
2747 dev_replace->cursor_left = found_key.offset;
2748 dev_replace->item_needs_writeback = 1;
2749 ret = scrub_chunk(sctx, scrub_dev, chunk_tree, chunk_objectid,
2750 chunk_offset, length, found_key.offset,
2751 is_dev_replace);
2752
2753 /*
2754 * flush, submit all pending read and write bios, afterwards
2755 * wait for them.
2756 * Note that in the dev replace case, a read request causes
2757 * write requests that are submitted in the read completion
2758 * worker. Therefore in the current situation, it is required
2759 * that all write requests are flushed, so that all read and
2760 * write requests are really completed when bios_in_flight
2761 * changes to 0.
2762 */
2763 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2764 scrub_submit(sctx);
2765 mutex_lock(&sctx->wr_ctx.wr_lock);
2766 scrub_wr_submit(sctx);
2767 mutex_unlock(&sctx->wr_ctx.wr_lock);
2768
2769 wait_event(sctx->list_wait,
2770 atomic_read(&sctx->bios_in_flight) == 0);
2771 atomic_inc(&fs_info->scrubs_paused);
2772 wake_up(&fs_info->scrub_pause_wait);
2773
2774 /*
2775 * must be called before we decrease @scrub_paused.
2776 * make sure we don't block transaction commit while
2777 * we are waiting pending workers finished.
2778 */
2779 wait_event(sctx->list_wait,
2780 atomic_read(&sctx->workers_pending) == 0);
2781 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2782
2783 mutex_lock(&fs_info->scrub_lock);
2784 __scrub_blocked_if_needed(fs_info);
2785 atomic_dec(&fs_info->scrubs_paused);
2786 mutex_unlock(&fs_info->scrub_lock);
2787 wake_up(&fs_info->scrub_pause_wait);
2788
2789 btrfs_put_block_group(cache);
2790 if (ret)
2791 break;
2792 if (is_dev_replace &&
2793 atomic64_read(&dev_replace->num_write_errors) > 0) {
2794 ret = -EIO;
2795 break;
2796 }
2797 if (sctx->stat.malloc_errors > 0) {
2798 ret = -ENOMEM;
2799 break;
2800 }
2801
2802 dev_replace->cursor_left = dev_replace->cursor_right;
2803 dev_replace->item_needs_writeback = 1;
2804
2805 key.offset = found_key.offset + length;
2806 btrfs_release_path(path);
2807 }
2808
2809 btrfs_free_path(path);
2810
2811 /*
2812 * ret can still be 1 from search_slot or next_leaf,
2813 * that's not an error
2814 */
2815 return ret < 0 ? ret : 0;
2816}
2817
2818static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
2819 struct btrfs_device *scrub_dev)
2820{
2821 int i;
2822 u64 bytenr;
2823 u64 gen;
2824 int ret;
2825 struct btrfs_root *root = sctx->dev_root;
2826
2827 if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
2828 return -EIO;
2829
2830 gen = root->fs_info->last_trans_committed;
2831
2832 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2833 bytenr = btrfs_sb_offset(i);
2834 if (bytenr + BTRFS_SUPER_INFO_SIZE > scrub_dev->total_bytes)
2835 break;
2836
2837 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
2838 scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
2839 NULL, 1, bytenr);
2840 if (ret)
2841 return ret;
2842 }
2843 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2844
2845 return 0;
2846}
2847
2848/*
2849 * get a reference count on fs_info->scrub_workers. start worker if necessary
2850 */
2851static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
2852 int is_dev_replace)
2853{
2854 int ret = 0;
2855 int flags = WQ_FREEZABLE | WQ_UNBOUND;
2856 int max_active = fs_info->thread_pool_size;
2857
2858 if (fs_info->scrub_workers_refcnt == 0) {
2859 if (is_dev_replace)
2860 fs_info->scrub_workers =
2861 btrfs_alloc_workqueue("btrfs-scrub", flags,
2862 1, 4);
2863 else
2864 fs_info->scrub_workers =
2865 btrfs_alloc_workqueue("btrfs-scrub", flags,
2866 max_active, 4);
2867 if (!fs_info->scrub_workers) {
2868 ret = -ENOMEM;
2869 goto out;
2870 }
2871 fs_info->scrub_wr_completion_workers =
2872 btrfs_alloc_workqueue("btrfs-scrubwrc", flags,
2873 max_active, 2);
2874 if (!fs_info->scrub_wr_completion_workers) {
2875 ret = -ENOMEM;
2876 goto out;
2877 }
2878 fs_info->scrub_nocow_workers =
2879 btrfs_alloc_workqueue("btrfs-scrubnc", flags, 1, 0);
2880 if (!fs_info->scrub_nocow_workers) {
2881 ret = -ENOMEM;
2882 goto out;
2883 }
2884 }
2885 ++fs_info->scrub_workers_refcnt;
2886out:
2887 return ret;
2888}
2889
2890static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
2891{
2892 if (--fs_info->scrub_workers_refcnt == 0) {
2893 btrfs_destroy_workqueue(fs_info->scrub_workers);
2894 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
2895 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
2896 }
2897 WARN_ON(fs_info->scrub_workers_refcnt < 0);
2898}
2899
2900int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
2901 u64 end, struct btrfs_scrub_progress *progress,
2902 int readonly, int is_dev_replace)
2903{
2904 struct scrub_ctx *sctx;
2905 int ret;
2906 struct btrfs_device *dev;
2907
2908 if (btrfs_fs_closing(fs_info))
2909 return -EINVAL;
2910
2911 /*
2912 * check some assumptions
2913 */
2914 if (fs_info->chunk_root->nodesize != fs_info->chunk_root->leafsize) {
2915 btrfs_err(fs_info,
2916 "scrub: size assumption nodesize == leafsize (%d == %d) fails",
2917 fs_info->chunk_root->nodesize,
2918 fs_info->chunk_root->leafsize);
2919 return -EINVAL;
2920 }
2921
2922 if (fs_info->chunk_root->nodesize > BTRFS_STRIPE_LEN) {
2923 /*
2924 * in this case scrub is unable to calculate the checksum
2925 * the way scrub is implemented. Do not handle this
2926 * situation at all because it won't ever happen.
2927 */
2928 btrfs_err(fs_info,
2929 "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
2930 fs_info->chunk_root->nodesize, BTRFS_STRIPE_LEN);
2931 return -EINVAL;
2932 }
2933
2934 if (fs_info->chunk_root->sectorsize != PAGE_SIZE) {
2935 /* not supported for data w/o checksums */
2936 btrfs_err(fs_info,
2937 "scrub: size assumption sectorsize != PAGE_SIZE "
2938 "(%d != %lu) fails",
2939 fs_info->chunk_root->sectorsize, PAGE_SIZE);
2940 return -EINVAL;
2941 }
2942
2943 if (fs_info->chunk_root->nodesize >
2944 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
2945 fs_info->chunk_root->sectorsize >
2946 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
2947 /*
2948 * would exhaust the array bounds of pagev member in
2949 * struct scrub_block
2950 */
2951 btrfs_err(fs_info, "scrub: size assumption nodesize and sectorsize "
2952 "<= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
2953 fs_info->chunk_root->nodesize,
2954 SCRUB_MAX_PAGES_PER_BLOCK,
2955 fs_info->chunk_root->sectorsize,
2956 SCRUB_MAX_PAGES_PER_BLOCK);
2957 return -EINVAL;
2958 }
2959
2960
2961 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2962 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
2963 if (!dev || (dev->missing && !is_dev_replace)) {
2964 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2965 return -ENODEV;
2966 }
2967
2968 mutex_lock(&fs_info->scrub_lock);
2969 if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
2970 mutex_unlock(&fs_info->scrub_lock);
2971 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2972 return -EIO;
2973 }
2974
2975 btrfs_dev_replace_lock(&fs_info->dev_replace);
2976 if (dev->scrub_device ||
2977 (!is_dev_replace &&
2978 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
2979 btrfs_dev_replace_unlock(&fs_info->dev_replace);
2980 mutex_unlock(&fs_info->scrub_lock);
2981 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2982 return -EINPROGRESS;
2983 }
2984 btrfs_dev_replace_unlock(&fs_info->dev_replace);
2985
2986 ret = scrub_workers_get(fs_info, is_dev_replace);
2987 if (ret) {
2988 mutex_unlock(&fs_info->scrub_lock);
2989 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2990 return ret;
2991 }
2992
2993 sctx = scrub_setup_ctx(dev, is_dev_replace);
2994 if (IS_ERR(sctx)) {
2995 mutex_unlock(&fs_info->scrub_lock);
2996 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2997 scrub_workers_put(fs_info);
2998 return PTR_ERR(sctx);
2999 }
3000 sctx->readonly = readonly;
3001 dev->scrub_device = sctx;
3002 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3003
3004 /*
3005 * checking @scrub_pause_req here, we can avoid
3006 * race between committing transaction and scrubbing.
3007 */
3008 __scrub_blocked_if_needed(fs_info);
3009 atomic_inc(&fs_info->scrubs_running);
3010 mutex_unlock(&fs_info->scrub_lock);
3011
3012 if (!is_dev_replace) {
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
3022 if (!ret)
3023 ret = scrub_enumerate_chunks(sctx, dev, start, end,
3024 is_dev_replace);
3025
3026 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3027 atomic_dec(&fs_info->scrubs_running);
3028 wake_up(&fs_info->scrub_pause_wait);
3029
3030 wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3031
3032 if (progress)
3033 memcpy(progress, &sctx->stat, sizeof(*progress));
3034
3035 mutex_lock(&fs_info->scrub_lock);
3036 dev->scrub_device = NULL;
3037 scrub_workers_put(fs_info);
3038 mutex_unlock(&fs_info->scrub_lock);
3039
3040 scrub_free_ctx(sctx);
3041
3042 return ret;
3043}
3044
3045void btrfs_scrub_pause(struct btrfs_root *root)
3046{
3047 struct btrfs_fs_info *fs_info = root->fs_info;
3048
3049 mutex_lock(&fs_info->scrub_lock);
3050 atomic_inc(&fs_info->scrub_pause_req);
3051 while (atomic_read(&fs_info->scrubs_paused) !=
3052 atomic_read(&fs_info->scrubs_running)) {
3053 mutex_unlock(&fs_info->scrub_lock);
3054 wait_event(fs_info->scrub_pause_wait,
3055 atomic_read(&fs_info->scrubs_paused) ==
3056 atomic_read(&fs_info->scrubs_running));
3057 mutex_lock(&fs_info->scrub_lock);
3058 }
3059 mutex_unlock(&fs_info->scrub_lock);
3060}
3061
3062void btrfs_scrub_continue(struct btrfs_root *root)
3063{
3064 struct btrfs_fs_info *fs_info = root->fs_info;
3065
3066 atomic_dec(&fs_info->scrub_pause_req);
3067 wake_up(&fs_info->scrub_pause_wait);
3068}
3069
3070int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
3071{
3072 mutex_lock(&fs_info->scrub_lock);
3073 if (!atomic_read(&fs_info->scrubs_running)) {
3074 mutex_unlock(&fs_info->scrub_lock);
3075 return -ENOTCONN;
3076 }
3077
3078 atomic_inc(&fs_info->scrub_cancel_req);
3079 while (atomic_read(&fs_info->scrubs_running)) {
3080 mutex_unlock(&fs_info->scrub_lock);
3081 wait_event(fs_info->scrub_pause_wait,
3082 atomic_read(&fs_info->scrubs_running) == 0);
3083 mutex_lock(&fs_info->scrub_lock);
3084 }
3085 atomic_dec(&fs_info->scrub_cancel_req);
3086 mutex_unlock(&fs_info->scrub_lock);
3087
3088 return 0;
3089}
3090
3091int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
3092 struct btrfs_device *dev)
3093{
3094 struct scrub_ctx *sctx;
3095
3096 mutex_lock(&fs_info->scrub_lock);
3097 sctx = dev->scrub_device;
3098 if (!sctx) {
3099 mutex_unlock(&fs_info->scrub_lock);
3100 return -ENOTCONN;
3101 }
3102 atomic_inc(&sctx->cancel_req);
3103 while (dev->scrub_device) {
3104 mutex_unlock(&fs_info->scrub_lock);
3105 wait_event(fs_info->scrub_pause_wait,
3106 dev->scrub_device == NULL);
3107 mutex_lock(&fs_info->scrub_lock);
3108 }
3109 mutex_unlock(&fs_info->scrub_lock);
3110
3111 return 0;
3112}
3113
3114int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
3115 struct btrfs_scrub_progress *progress)
3116{
3117 struct btrfs_device *dev;
3118 struct scrub_ctx *sctx = NULL;
3119
3120 mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
3121 dev = btrfs_find_device(root->fs_info, devid, NULL, NULL);
3122 if (dev)
3123 sctx = dev->scrub_device;
3124 if (sctx)
3125 memcpy(progress, &sctx->stat, sizeof(*progress));
3126 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
3127
3128 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3129}
3130
3131static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
3132 u64 extent_logical, u64 extent_len,
3133 u64 *extent_physical,
3134 struct btrfs_device **extent_dev,
3135 int *extent_mirror_num)
3136{
3137 u64 mapped_length;
3138 struct btrfs_bio *bbio = NULL;
3139 int ret;
3140
3141 mapped_length = extent_len;
3142 ret = btrfs_map_block(fs_info, READ, extent_logical,
3143 &mapped_length, &bbio, 0);
3144 if (ret || !bbio || mapped_length < extent_len ||
3145 !bbio->stripes[0].dev->bdev) {
3146 kfree(bbio);
3147 return;
3148 }
3149
3150 *extent_physical = bbio->stripes[0].physical;
3151 *extent_mirror_num = bbio->mirror_num;
3152 *extent_dev = bbio->stripes[0].dev;
3153 kfree(bbio);
3154}
3155
3156static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
3157 struct scrub_wr_ctx *wr_ctx,
3158 struct btrfs_fs_info *fs_info,
3159 struct btrfs_device *dev,
3160 int is_dev_replace)
3161{
3162 WARN_ON(wr_ctx->wr_curr_bio != NULL);
3163
3164 mutex_init(&wr_ctx->wr_lock);
3165 wr_ctx->wr_curr_bio = NULL;
3166 if (!is_dev_replace)
3167 return 0;
3168
3169 WARN_ON(!dev->bdev);
3170 wr_ctx->pages_per_wr_bio = min_t(int, SCRUB_PAGES_PER_WR_BIO,
3171 bio_get_nr_vecs(dev->bdev));
3172 wr_ctx->tgtdev = dev;
3173 atomic_set(&wr_ctx->flush_all_writes, 0);
3174 return 0;
3175}
3176
3177static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
3178{
3179 mutex_lock(&wr_ctx->wr_lock);
3180 kfree(wr_ctx->wr_curr_bio);
3181 wr_ctx->wr_curr_bio = NULL;
3182 mutex_unlock(&wr_ctx->wr_lock);
3183}
3184
3185static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
3186 int mirror_num, u64 physical_for_dev_replace)
3187{
3188 struct scrub_copy_nocow_ctx *nocow_ctx;
3189 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
3190
3191 nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
3192 if (!nocow_ctx) {
3193 spin_lock(&sctx->stat_lock);
3194 sctx->stat.malloc_errors++;
3195 spin_unlock(&sctx->stat_lock);
3196 return -ENOMEM;
3197 }
3198
3199 scrub_pending_trans_workers_inc(sctx);
3200
3201 nocow_ctx->sctx = sctx;
3202 nocow_ctx->logical = logical;
3203 nocow_ctx->len = len;
3204 nocow_ctx->mirror_num = mirror_num;
3205 nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
3206 btrfs_init_work(&nocow_ctx->work, copy_nocow_pages_worker, NULL, NULL);
3207 INIT_LIST_HEAD(&nocow_ctx->inodes);
3208 btrfs_queue_work(fs_info->scrub_nocow_workers,
3209 &nocow_ctx->work);
3210
3211 return 0;
3212}
3213
3214static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
3215{
3216 struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
3217 struct scrub_nocow_inode *nocow_inode;
3218
3219 nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
3220 if (!nocow_inode)
3221 return -ENOMEM;
3222 nocow_inode->inum = inum;
3223 nocow_inode->offset = offset;
3224 nocow_inode->root = root;
3225 list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
3226 return 0;
3227}
3228
3229#define COPY_COMPLETE 1
3230
3231static void copy_nocow_pages_worker(struct btrfs_work *work)
3232{
3233 struct scrub_copy_nocow_ctx *nocow_ctx =
3234 container_of(work, struct scrub_copy_nocow_ctx, work);
3235 struct scrub_ctx *sctx = nocow_ctx->sctx;
3236 u64 logical = nocow_ctx->logical;
3237 u64 len = nocow_ctx->len;
3238 int mirror_num = nocow_ctx->mirror_num;
3239 u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3240 int ret;
3241 struct btrfs_trans_handle *trans = NULL;
3242 struct btrfs_fs_info *fs_info;
3243 struct btrfs_path *path;
3244 struct btrfs_root *root;
3245 int not_written = 0;
3246
3247 fs_info = sctx->dev_root->fs_info;
3248 root = fs_info->extent_root;
3249
3250 path = btrfs_alloc_path();
3251 if (!path) {
3252 spin_lock(&sctx->stat_lock);
3253 sctx->stat.malloc_errors++;
3254 spin_unlock(&sctx->stat_lock);
3255 not_written = 1;
3256 goto out;
3257 }
3258
3259 trans = btrfs_join_transaction(root);
3260 if (IS_ERR(trans)) {
3261 not_written = 1;
3262 goto out;
3263 }
3264
3265 ret = iterate_inodes_from_logical(logical, fs_info, path,
3266 record_inode_for_nocow, nocow_ctx);
3267 if (ret != 0 && ret != -ENOENT) {
3268 btrfs_warn(fs_info, "iterate_inodes_from_logical() failed: log %llu, "
3269 "phys %llu, len %llu, mir %u, ret %d",
3270 logical, physical_for_dev_replace, len, mirror_num,
3271 ret);
3272 not_written = 1;
3273 goto out;
3274 }
3275
3276 btrfs_end_transaction(trans, root);
3277 trans = NULL;
3278 while (!list_empty(&nocow_ctx->inodes)) {
3279 struct scrub_nocow_inode *entry;
3280 entry = list_first_entry(&nocow_ctx->inodes,
3281 struct scrub_nocow_inode,
3282 list);
3283 list_del_init(&entry->list);
3284 ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
3285 entry->root, nocow_ctx);
3286 kfree(entry);
3287 if (ret == COPY_COMPLETE) {
3288 ret = 0;
3289 break;
3290 } else if (ret) {
3291 break;
3292 }
3293 }
3294out:
3295 while (!list_empty(&nocow_ctx->inodes)) {
3296 struct scrub_nocow_inode *entry;
3297 entry = list_first_entry(&nocow_ctx->inodes,
3298 struct scrub_nocow_inode,
3299 list);
3300 list_del_init(&entry->list);
3301 kfree(entry);
3302 }
3303 if (trans && !IS_ERR(trans))
3304 btrfs_end_transaction(trans, root);
3305 if (not_written)
3306 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
3307 num_uncorrectable_read_errors);
3308
3309 btrfs_free_path(path);
3310 kfree(nocow_ctx);
3311
3312 scrub_pending_trans_workers_dec(sctx);
3313}
3314
3315static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
3316 struct scrub_copy_nocow_ctx *nocow_ctx)
3317{
3318 struct btrfs_fs_info *fs_info = nocow_ctx->sctx->dev_root->fs_info;
3319 struct btrfs_key key;
3320 struct inode *inode;
3321 struct page *page;
3322 struct btrfs_root *local_root;
3323 struct btrfs_ordered_extent *ordered;
3324 struct extent_map *em;
3325 struct extent_state *cached_state = NULL;
3326 struct extent_io_tree *io_tree;
3327 u64 physical_for_dev_replace;
3328 u64 len = nocow_ctx->len;
3329 u64 lockstart = offset, lockend = offset + len - 1;
3330 unsigned long index;
3331 int srcu_index;
3332 int ret = 0;
3333 int err = 0;
3334
3335 key.objectid = root;
3336 key.type = BTRFS_ROOT_ITEM_KEY;
3337 key.offset = (u64)-1;
3338
3339 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
3340
3341 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
3342 if (IS_ERR(local_root)) {
3343 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3344 return PTR_ERR(local_root);
3345 }
3346
3347 key.type = BTRFS_INODE_ITEM_KEY;
3348 key.objectid = inum;
3349 key.offset = 0;
3350 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
3351 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3352 if (IS_ERR(inode))
3353 return PTR_ERR(inode);
3354
3355 /* Avoid truncate/dio/punch hole.. */
3356 mutex_lock(&inode->i_mutex);
3357 inode_dio_wait(inode);
3358
3359 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3360 io_tree = &BTRFS_I(inode)->io_tree;
3361
3362 lock_extent_bits(io_tree, lockstart, lockend, 0, &cached_state);
3363 ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
3364 if (ordered) {
3365 btrfs_put_ordered_extent(ordered);
3366 goto out_unlock;
3367 }
3368
3369 em = btrfs_get_extent(inode, NULL, 0, lockstart, len, 0);
3370 if (IS_ERR(em)) {
3371 ret = PTR_ERR(em);
3372 goto out_unlock;
3373 }
3374
3375 /*
3376 * This extent does not actually cover the logical extent anymore,
3377 * move on to the next inode.
3378 */
3379 if (em->block_start > nocow_ctx->logical ||
3380 em->block_start + em->block_len < nocow_ctx->logical + len) {
3381 free_extent_map(em);
3382 goto out_unlock;
3383 }
3384 free_extent_map(em);
3385
3386 while (len >= PAGE_CACHE_SIZE) {
3387 index = offset >> PAGE_CACHE_SHIFT;
3388again:
3389 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
3390 if (!page) {
3391 btrfs_err(fs_info, "find_or_create_page() failed");
3392 ret = -ENOMEM;
3393 goto out;
3394 }
3395
3396 if (PageUptodate(page)) {
3397 if (PageDirty(page))
3398 goto next_page;
3399 } else {
3400 ClearPageError(page);
3401 err = extent_read_full_page_nolock(io_tree, page,
3402 btrfs_get_extent,
3403 nocow_ctx->mirror_num);
3404 if (err) {
3405 ret = err;
3406 goto next_page;
3407 }
3408
3409 lock_page(page);
3410 /*
3411 * If the page has been remove from the page cache,
3412 * the data on it is meaningless, because it may be
3413 * old one, the new data may be written into the new
3414 * page in the page cache.
3415 */
3416 if (page->mapping != inode->i_mapping) {
3417 unlock_page(page);
3418 page_cache_release(page);
3419 goto again;
3420 }
3421 if (!PageUptodate(page)) {
3422 ret = -EIO;
3423 goto next_page;
3424 }
3425 }
3426 err = write_page_nocow(nocow_ctx->sctx,
3427 physical_for_dev_replace, page);
3428 if (err)
3429 ret = err;
3430next_page:
3431 unlock_page(page);
3432 page_cache_release(page);
3433
3434 if (ret)
3435 break;
3436
3437 offset += PAGE_CACHE_SIZE;
3438 physical_for_dev_replace += PAGE_CACHE_SIZE;
3439 len -= PAGE_CACHE_SIZE;
3440 }
3441 ret = COPY_COMPLETE;
3442out_unlock:
3443 unlock_extent_cached(io_tree, lockstart, lockend, &cached_state,
3444 GFP_NOFS);
3445out:
3446 mutex_unlock(&inode->i_mutex);
3447 iput(inode);
3448 return ret;
3449}
3450
3451static int write_page_nocow(struct scrub_ctx *sctx,
3452 u64 physical_for_dev_replace, struct page *page)
3453{
3454 struct bio *bio;
3455 struct btrfs_device *dev;
3456 int ret;
3457
3458 dev = sctx->wr_ctx.tgtdev;
3459 if (!dev)
3460 return -EIO;
3461 if (!dev->bdev) {
3462 printk_ratelimited(KERN_WARNING
3463 "BTRFS: scrub write_page_nocow(bdev == NULL) is unexpected!\n");
3464 return -EIO;
3465 }
3466 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
3467 if (!bio) {
3468 spin_lock(&sctx->stat_lock);
3469 sctx->stat.malloc_errors++;
3470 spin_unlock(&sctx->stat_lock);
3471 return -ENOMEM;
3472 }
3473 bio->bi_iter.bi_size = 0;
3474 bio->bi_iter.bi_sector = physical_for_dev_replace >> 9;
3475 bio->bi_bdev = dev->bdev;
3476 ret = bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
3477 if (ret != PAGE_CACHE_SIZE) {
3478leave_with_eio:
3479 bio_put(bio);
3480 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
3481 return -EIO;
3482 }
3483
3484 if (btrfsic_submit_bio_wait(WRITE_SYNC, bio))
3485 goto leave_with_eio;
3486
3487 bio_put(bio);
3488 return 0;
3489}