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 "ctree.h"
10#include "volumes.h"
11#include "disk-io.h"
12#include "ordered-data.h"
13#include "transaction.h"
14#include "backref.h"
15#include "extent_io.h"
16#include "dev-replace.h"
17#include "check-integrity.h"
18#include "rcu-string.h"
19#include "raid56.h"
20
21/*
22 * This is only the first step towards a full-features scrub. It reads all
23 * extent and super block and verifies the checksums. In case a bad checksum
24 * is found or the extent cannot be read, good data will be written back if
25 * any can be found.
26 *
27 * Future enhancements:
28 * - In case an unrepairable extent is encountered, track which files are
29 * affected and report them
30 * - track and record media errors, throw out bad devices
31 * - add a mode to also read unallocated space
32 */
33
34struct scrub_block;
35struct scrub_ctx;
36
37/*
38 * the following three values only influence the performance.
39 * The last one configures the number of parallel and outstanding I/O
40 * operations. The first two values configure an upper limit for the number
41 * of (dynamically allocated) pages that are added to a bio.
42 */
43#define SCRUB_PAGES_PER_RD_BIO 32 /* 128k per bio */
44#define SCRUB_PAGES_PER_WR_BIO 32 /* 128k per bio */
45#define SCRUB_BIOS_PER_SCTX 64 /* 8MB per device in flight */
46
47/*
48 * the following value times PAGE_SIZE needs to be large enough to match the
49 * largest node/leaf/sector size that shall be supported.
50 * Values larger than BTRFS_STRIPE_LEN are not supported.
51 */
52#define SCRUB_MAX_PAGES_PER_BLOCK 16 /* 64k per node/leaf/sector */
53
54struct scrub_recover {
55 refcount_t refs;
56 struct btrfs_bio *bbio;
57 u64 map_length;
58};
59
60struct scrub_page {
61 struct scrub_block *sblock;
62 struct page *page;
63 struct btrfs_device *dev;
64 struct list_head list;
65 u64 flags; /* extent flags */
66 u64 generation;
67 u64 logical;
68 u64 physical;
69 u64 physical_for_dev_replace;
70 atomic_t refs;
71 struct {
72 unsigned int mirror_num:8;
73 unsigned int have_csum:1;
74 unsigned int io_error:1;
75 };
76 u8 csum[BTRFS_CSUM_SIZE];
77
78 struct scrub_recover *recover;
79};
80
81struct scrub_bio {
82 int index;
83 struct scrub_ctx *sctx;
84 struct btrfs_device *dev;
85 struct bio *bio;
86 blk_status_t status;
87 u64 logical;
88 u64 physical;
89#if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
90 struct scrub_page *pagev[SCRUB_PAGES_PER_WR_BIO];
91#else
92 struct scrub_page *pagev[SCRUB_PAGES_PER_RD_BIO];
93#endif
94 int page_count;
95 int next_free;
96 struct btrfs_work work;
97};
98
99struct scrub_block {
100 struct scrub_page *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
101 int page_count;
102 atomic_t outstanding_pages;
103 refcount_t refs; /* free mem on transition to zero */
104 struct scrub_ctx *sctx;
105 struct scrub_parity *sparity;
106 struct {
107 unsigned int header_error:1;
108 unsigned int checksum_error:1;
109 unsigned int no_io_error_seen:1;
110 unsigned int generation_error:1; /* also sets header_error */
111
112 /* The following is for the data used to check parity */
113 /* It is for the data with checksum */
114 unsigned int data_corrected:1;
115 };
116 struct btrfs_work work;
117};
118
119/* Used for the chunks with parity stripe such RAID5/6 */
120struct scrub_parity {
121 struct scrub_ctx *sctx;
122
123 struct btrfs_device *scrub_dev;
124
125 u64 logic_start;
126
127 u64 logic_end;
128
129 int nsectors;
130
131 u64 stripe_len;
132
133 refcount_t refs;
134
135 struct list_head spages;
136
137 /* Work of parity check and repair */
138 struct btrfs_work work;
139
140 /* Mark the parity blocks which have data */
141 unsigned long *dbitmap;
142
143 /*
144 * Mark the parity blocks which have data, but errors happen when
145 * read data or check data
146 */
147 unsigned long *ebitmap;
148
149 unsigned long bitmap[0];
150};
151
152struct scrub_ctx {
153 struct scrub_bio *bios[SCRUB_BIOS_PER_SCTX];
154 struct btrfs_fs_info *fs_info;
155 int first_free;
156 int curr;
157 atomic_t bios_in_flight;
158 atomic_t workers_pending;
159 spinlock_t list_lock;
160 wait_queue_head_t list_wait;
161 u16 csum_size;
162 struct list_head csum_list;
163 atomic_t cancel_req;
164 int readonly;
165 int pages_per_rd_bio;
166
167 int is_dev_replace;
168
169 struct scrub_bio *wr_curr_bio;
170 struct mutex wr_lock;
171 int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
172 struct btrfs_device *wr_tgtdev;
173 bool flush_all_writes;
174
175 /*
176 * statistics
177 */
178 struct btrfs_scrub_progress stat;
179 spinlock_t stat_lock;
180
181 /*
182 * Use a ref counter to avoid use-after-free issues. Scrub workers
183 * decrement bios_in_flight and workers_pending and then do a wakeup
184 * on the list_wait wait queue. We must ensure the main scrub task
185 * doesn't free the scrub context before or while the workers are
186 * doing the wakeup() call.
187 */
188 refcount_t refs;
189};
190
191struct scrub_fixup_nodatasum {
192 struct scrub_ctx *sctx;
193 struct btrfs_device *dev;
194 u64 logical;
195 struct btrfs_root *root;
196 struct btrfs_work work;
197 int mirror_num;
198};
199
200struct scrub_nocow_inode {
201 u64 inum;
202 u64 offset;
203 u64 root;
204 struct list_head list;
205};
206
207struct scrub_copy_nocow_ctx {
208 struct scrub_ctx *sctx;
209 u64 logical;
210 u64 len;
211 int mirror_num;
212 u64 physical_for_dev_replace;
213 struct list_head inodes;
214 struct btrfs_work work;
215};
216
217struct scrub_warning {
218 struct btrfs_path *path;
219 u64 extent_item_size;
220 const char *errstr;
221 u64 physical;
222 u64 logical;
223 struct btrfs_device *dev;
224};
225
226struct full_stripe_lock {
227 struct rb_node node;
228 u64 logical;
229 u64 refs;
230 struct mutex mutex;
231};
232
233static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
234static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
235static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
236static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
237static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
238static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
239 struct scrub_block *sblocks_for_recheck);
240static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
241 struct scrub_block *sblock,
242 int retry_failed_mirror);
243static void scrub_recheck_block_checksum(struct scrub_block *sblock);
244static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
245 struct scrub_block *sblock_good);
246static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
247 struct scrub_block *sblock_good,
248 int page_num, int force_write);
249static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
250static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
251 int page_num);
252static int scrub_checksum_data(struct scrub_block *sblock);
253static int scrub_checksum_tree_block(struct scrub_block *sblock);
254static int scrub_checksum_super(struct scrub_block *sblock);
255static void scrub_block_get(struct scrub_block *sblock);
256static void scrub_block_put(struct scrub_block *sblock);
257static void scrub_page_get(struct scrub_page *spage);
258static void scrub_page_put(struct scrub_page *spage);
259static void scrub_parity_get(struct scrub_parity *sparity);
260static void scrub_parity_put(struct scrub_parity *sparity);
261static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
262 struct scrub_page *spage);
263static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
264 u64 physical, struct btrfs_device *dev, u64 flags,
265 u64 gen, int mirror_num, u8 *csum, int force,
266 u64 physical_for_dev_replace);
267static void scrub_bio_end_io(struct bio *bio);
268static void scrub_bio_end_io_worker(struct btrfs_work *work);
269static void scrub_block_complete(struct scrub_block *sblock);
270static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
271 u64 extent_logical, u64 extent_len,
272 u64 *extent_physical,
273 struct btrfs_device **extent_dev,
274 int *extent_mirror_num);
275static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
276 struct scrub_page *spage);
277static void scrub_wr_submit(struct scrub_ctx *sctx);
278static void scrub_wr_bio_end_io(struct bio *bio);
279static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
280static int write_page_nocow(struct scrub_ctx *sctx,
281 u64 physical_for_dev_replace, struct page *page);
282static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
283 struct scrub_copy_nocow_ctx *ctx);
284static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
285 int mirror_num, u64 physical_for_dev_replace);
286static void copy_nocow_pages_worker(struct btrfs_work *work);
287static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
288static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
289static void scrub_put_ctx(struct scrub_ctx *sctx);
290
291static inline int scrub_is_page_on_raid56(struct scrub_page *page)
292{
293 return page->recover &&
294 (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
295}
296
297static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
298{
299 refcount_inc(&sctx->refs);
300 atomic_inc(&sctx->bios_in_flight);
301}
302
303static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
304{
305 atomic_dec(&sctx->bios_in_flight);
306 wake_up(&sctx->list_wait);
307 scrub_put_ctx(sctx);
308}
309
310static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
311{
312 while (atomic_read(&fs_info->scrub_pause_req)) {
313 mutex_unlock(&fs_info->scrub_lock);
314 wait_event(fs_info->scrub_pause_wait,
315 atomic_read(&fs_info->scrub_pause_req) == 0);
316 mutex_lock(&fs_info->scrub_lock);
317 }
318}
319
320static void scrub_pause_on(struct btrfs_fs_info *fs_info)
321{
322 atomic_inc(&fs_info->scrubs_paused);
323 wake_up(&fs_info->scrub_pause_wait);
324}
325
326static void scrub_pause_off(struct btrfs_fs_info *fs_info)
327{
328 mutex_lock(&fs_info->scrub_lock);
329 __scrub_blocked_if_needed(fs_info);
330 atomic_dec(&fs_info->scrubs_paused);
331 mutex_unlock(&fs_info->scrub_lock);
332
333 wake_up(&fs_info->scrub_pause_wait);
334}
335
336static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
337{
338 scrub_pause_on(fs_info);
339 scrub_pause_off(fs_info);
340}
341
342/*
343 * Insert new full stripe lock into full stripe locks tree
344 *
345 * Return pointer to existing or newly inserted full_stripe_lock structure if
346 * everything works well.
347 * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
348 *
349 * NOTE: caller must hold full_stripe_locks_root->lock before calling this
350 * function
351 */
352static struct full_stripe_lock *insert_full_stripe_lock(
353 struct btrfs_full_stripe_locks_tree *locks_root,
354 u64 fstripe_logical)
355{
356 struct rb_node **p;
357 struct rb_node *parent = NULL;
358 struct full_stripe_lock *entry;
359 struct full_stripe_lock *ret;
360
361 lockdep_assert_held(&locks_root->lock);
362
363 p = &locks_root->root.rb_node;
364 while (*p) {
365 parent = *p;
366 entry = rb_entry(parent, struct full_stripe_lock, node);
367 if (fstripe_logical < entry->logical) {
368 p = &(*p)->rb_left;
369 } else if (fstripe_logical > entry->logical) {
370 p = &(*p)->rb_right;
371 } else {
372 entry->refs++;
373 return entry;
374 }
375 }
376
377 /* Insert new lock */
378 ret = kmalloc(sizeof(*ret), GFP_KERNEL);
379 if (!ret)
380 return ERR_PTR(-ENOMEM);
381 ret->logical = fstripe_logical;
382 ret->refs = 1;
383 mutex_init(&ret->mutex);
384
385 rb_link_node(&ret->node, parent, p);
386 rb_insert_color(&ret->node, &locks_root->root);
387 return ret;
388}
389
390/*
391 * Search for a full stripe lock of a block group
392 *
393 * Return pointer to existing full stripe lock if found
394 * Return NULL if not found
395 */
396static struct full_stripe_lock *search_full_stripe_lock(
397 struct btrfs_full_stripe_locks_tree *locks_root,
398 u64 fstripe_logical)
399{
400 struct rb_node *node;
401 struct full_stripe_lock *entry;
402
403 lockdep_assert_held(&locks_root->lock);
404
405 node = locks_root->root.rb_node;
406 while (node) {
407 entry = rb_entry(node, struct full_stripe_lock, node);
408 if (fstripe_logical < entry->logical)
409 node = node->rb_left;
410 else if (fstripe_logical > entry->logical)
411 node = node->rb_right;
412 else
413 return entry;
414 }
415 return NULL;
416}
417
418/*
419 * Helper to get full stripe logical from a normal bytenr.
420 *
421 * Caller must ensure @cache is a RAID56 block group.
422 */
423static u64 get_full_stripe_logical(struct btrfs_block_group_cache *cache,
424 u64 bytenr)
425{
426 u64 ret;
427
428 /*
429 * Due to chunk item size limit, full stripe length should not be
430 * larger than U32_MAX. Just a sanity check here.
431 */
432 WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
433
434 /*
435 * round_down() can only handle power of 2, while RAID56 full
436 * stripe length can be 64KiB * n, so we need to manually round down.
437 */
438 ret = div64_u64(bytenr - cache->key.objectid, cache->full_stripe_len) *
439 cache->full_stripe_len + cache->key.objectid;
440 return ret;
441}
442
443/*
444 * Lock a full stripe to avoid concurrency of recovery and read
445 *
446 * It's only used for profiles with parities (RAID5/6), for other profiles it
447 * does nothing.
448 *
449 * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
450 * So caller must call unlock_full_stripe() at the same context.
451 *
452 * Return <0 if encounters error.
453 */
454static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
455 bool *locked_ret)
456{
457 struct btrfs_block_group_cache *bg_cache;
458 struct btrfs_full_stripe_locks_tree *locks_root;
459 struct full_stripe_lock *existing;
460 u64 fstripe_start;
461 int ret = 0;
462
463 *locked_ret = false;
464 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
465 if (!bg_cache) {
466 ASSERT(0);
467 return -ENOENT;
468 }
469
470 /* Profiles not based on parity don't need full stripe lock */
471 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
472 goto out;
473 locks_root = &bg_cache->full_stripe_locks_root;
474
475 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
476
477 /* Now insert the full stripe lock */
478 mutex_lock(&locks_root->lock);
479 existing = insert_full_stripe_lock(locks_root, fstripe_start);
480 mutex_unlock(&locks_root->lock);
481 if (IS_ERR(existing)) {
482 ret = PTR_ERR(existing);
483 goto out;
484 }
485 mutex_lock(&existing->mutex);
486 *locked_ret = true;
487out:
488 btrfs_put_block_group(bg_cache);
489 return ret;
490}
491
492/*
493 * Unlock a full stripe.
494 *
495 * NOTE: Caller must ensure it's the same context calling corresponding
496 * lock_full_stripe().
497 *
498 * Return 0 if we unlock full stripe without problem.
499 * Return <0 for error
500 */
501static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
502 bool locked)
503{
504 struct btrfs_block_group_cache *bg_cache;
505 struct btrfs_full_stripe_locks_tree *locks_root;
506 struct full_stripe_lock *fstripe_lock;
507 u64 fstripe_start;
508 bool freeit = false;
509 int ret = 0;
510
511 /* If we didn't acquire full stripe lock, no need to continue */
512 if (!locked)
513 return 0;
514
515 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
516 if (!bg_cache) {
517 ASSERT(0);
518 return -ENOENT;
519 }
520 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
521 goto out;
522
523 locks_root = &bg_cache->full_stripe_locks_root;
524 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
525
526 mutex_lock(&locks_root->lock);
527 fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
528 /* Unpaired unlock_full_stripe() detected */
529 if (!fstripe_lock) {
530 WARN_ON(1);
531 ret = -ENOENT;
532 mutex_unlock(&locks_root->lock);
533 goto out;
534 }
535
536 if (fstripe_lock->refs == 0) {
537 WARN_ON(1);
538 btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
539 fstripe_lock->logical);
540 } else {
541 fstripe_lock->refs--;
542 }
543
544 if (fstripe_lock->refs == 0) {
545 rb_erase(&fstripe_lock->node, &locks_root->root);
546 freeit = true;
547 }
548 mutex_unlock(&locks_root->lock);
549
550 mutex_unlock(&fstripe_lock->mutex);
551 if (freeit)
552 kfree(fstripe_lock);
553out:
554 btrfs_put_block_group(bg_cache);
555 return ret;
556}
557
558/*
559 * used for workers that require transaction commits (i.e., for the
560 * NOCOW case)
561 */
562static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
563{
564 struct btrfs_fs_info *fs_info = sctx->fs_info;
565
566 refcount_inc(&sctx->refs);
567 /*
568 * increment scrubs_running to prevent cancel requests from
569 * completing as long as a worker is running. we must also
570 * increment scrubs_paused to prevent deadlocking on pause
571 * requests used for transactions commits (as the worker uses a
572 * transaction context). it is safe to regard the worker
573 * as paused for all matters practical. effectively, we only
574 * avoid cancellation requests from completing.
575 */
576 mutex_lock(&fs_info->scrub_lock);
577 atomic_inc(&fs_info->scrubs_running);
578 atomic_inc(&fs_info->scrubs_paused);
579 mutex_unlock(&fs_info->scrub_lock);
580
581 /*
582 * check if @scrubs_running=@scrubs_paused condition
583 * inside wait_event() is not an atomic operation.
584 * which means we may inc/dec @scrub_running/paused
585 * at any time. Let's wake up @scrub_pause_wait as
586 * much as we can to let commit transaction blocked less.
587 */
588 wake_up(&fs_info->scrub_pause_wait);
589
590 atomic_inc(&sctx->workers_pending);
591}
592
593/* used for workers that require transaction commits */
594static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
595{
596 struct btrfs_fs_info *fs_info = sctx->fs_info;
597
598 /*
599 * see scrub_pending_trans_workers_inc() why we're pretending
600 * to be paused in the scrub counters
601 */
602 mutex_lock(&fs_info->scrub_lock);
603 atomic_dec(&fs_info->scrubs_running);
604 atomic_dec(&fs_info->scrubs_paused);
605 mutex_unlock(&fs_info->scrub_lock);
606 atomic_dec(&sctx->workers_pending);
607 wake_up(&fs_info->scrub_pause_wait);
608 wake_up(&sctx->list_wait);
609 scrub_put_ctx(sctx);
610}
611
612static void scrub_free_csums(struct scrub_ctx *sctx)
613{
614 while (!list_empty(&sctx->csum_list)) {
615 struct btrfs_ordered_sum *sum;
616 sum = list_first_entry(&sctx->csum_list,
617 struct btrfs_ordered_sum, list);
618 list_del(&sum->list);
619 kfree(sum);
620 }
621}
622
623static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
624{
625 int i;
626
627 if (!sctx)
628 return;
629
630 /* this can happen when scrub is cancelled */
631 if (sctx->curr != -1) {
632 struct scrub_bio *sbio = sctx->bios[sctx->curr];
633
634 for (i = 0; i < sbio->page_count; i++) {
635 WARN_ON(!sbio->pagev[i]->page);
636 scrub_block_put(sbio->pagev[i]->sblock);
637 }
638 bio_put(sbio->bio);
639 }
640
641 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
642 struct scrub_bio *sbio = sctx->bios[i];
643
644 if (!sbio)
645 break;
646 kfree(sbio);
647 }
648
649 kfree(sctx->wr_curr_bio);
650 scrub_free_csums(sctx);
651 kfree(sctx);
652}
653
654static void scrub_put_ctx(struct scrub_ctx *sctx)
655{
656 if (refcount_dec_and_test(&sctx->refs))
657 scrub_free_ctx(sctx);
658}
659
660static noinline_for_stack
661struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
662{
663 struct scrub_ctx *sctx;
664 int i;
665 struct btrfs_fs_info *fs_info = dev->fs_info;
666
667 sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
668 if (!sctx)
669 goto nomem;
670 refcount_set(&sctx->refs, 1);
671 sctx->is_dev_replace = is_dev_replace;
672 sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
673 sctx->curr = -1;
674 sctx->fs_info = dev->fs_info;
675 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
676 struct scrub_bio *sbio;
677
678 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
679 if (!sbio)
680 goto nomem;
681 sctx->bios[i] = sbio;
682
683 sbio->index = i;
684 sbio->sctx = sctx;
685 sbio->page_count = 0;
686 btrfs_init_work(&sbio->work, btrfs_scrub_helper,
687 scrub_bio_end_io_worker, NULL, NULL);
688
689 if (i != SCRUB_BIOS_PER_SCTX - 1)
690 sctx->bios[i]->next_free = i + 1;
691 else
692 sctx->bios[i]->next_free = -1;
693 }
694 sctx->first_free = 0;
695 atomic_set(&sctx->bios_in_flight, 0);
696 atomic_set(&sctx->workers_pending, 0);
697 atomic_set(&sctx->cancel_req, 0);
698 sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
699 INIT_LIST_HEAD(&sctx->csum_list);
700
701 spin_lock_init(&sctx->list_lock);
702 spin_lock_init(&sctx->stat_lock);
703 init_waitqueue_head(&sctx->list_wait);
704
705 WARN_ON(sctx->wr_curr_bio != NULL);
706 mutex_init(&sctx->wr_lock);
707 sctx->wr_curr_bio = NULL;
708 if (is_dev_replace) {
709 WARN_ON(!fs_info->dev_replace.tgtdev);
710 sctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
711 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
712 sctx->flush_all_writes = false;
713 }
714
715 return sctx;
716
717nomem:
718 scrub_free_ctx(sctx);
719 return ERR_PTR(-ENOMEM);
720}
721
722static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
723 void *warn_ctx)
724{
725 u64 isize;
726 u32 nlink;
727 int ret;
728 int i;
729 unsigned nofs_flag;
730 struct extent_buffer *eb;
731 struct btrfs_inode_item *inode_item;
732 struct scrub_warning *swarn = warn_ctx;
733 struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
734 struct inode_fs_paths *ipath = NULL;
735 struct btrfs_root *local_root;
736 struct btrfs_key root_key;
737 struct btrfs_key key;
738
739 root_key.objectid = root;
740 root_key.type = BTRFS_ROOT_ITEM_KEY;
741 root_key.offset = (u64)-1;
742 local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
743 if (IS_ERR(local_root)) {
744 ret = PTR_ERR(local_root);
745 goto err;
746 }
747
748 /*
749 * this makes the path point to (inum INODE_ITEM ioff)
750 */
751 key.objectid = inum;
752 key.type = BTRFS_INODE_ITEM_KEY;
753 key.offset = 0;
754
755 ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
756 if (ret) {
757 btrfs_release_path(swarn->path);
758 goto err;
759 }
760
761 eb = swarn->path->nodes[0];
762 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
763 struct btrfs_inode_item);
764 isize = btrfs_inode_size(eb, inode_item);
765 nlink = btrfs_inode_nlink(eb, inode_item);
766 btrfs_release_path(swarn->path);
767
768 /*
769 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
770 * uses GFP_NOFS in this context, so we keep it consistent but it does
771 * not seem to be strictly necessary.
772 */
773 nofs_flag = memalloc_nofs_save();
774 ipath = init_ipath(4096, local_root, swarn->path);
775 memalloc_nofs_restore(nofs_flag);
776 if (IS_ERR(ipath)) {
777 ret = PTR_ERR(ipath);
778 ipath = NULL;
779 goto err;
780 }
781 ret = paths_from_inode(inum, ipath);
782
783 if (ret < 0)
784 goto err;
785
786 /*
787 * we deliberately ignore the bit ipath might have been too small to
788 * hold all of the paths here
789 */
790 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
791 btrfs_warn_in_rcu(fs_info,
792"%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
793 swarn->errstr, swarn->logical,
794 rcu_str_deref(swarn->dev->name),
795 swarn->physical,
796 root, inum, offset,
797 min(isize - offset, (u64)PAGE_SIZE), nlink,
798 (char *)(unsigned long)ipath->fspath->val[i]);
799
800 free_ipath(ipath);
801 return 0;
802
803err:
804 btrfs_warn_in_rcu(fs_info,
805 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
806 swarn->errstr, swarn->logical,
807 rcu_str_deref(swarn->dev->name),
808 swarn->physical,
809 root, inum, offset, ret);
810
811 free_ipath(ipath);
812 return 0;
813}
814
815static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
816{
817 struct btrfs_device *dev;
818 struct btrfs_fs_info *fs_info;
819 struct btrfs_path *path;
820 struct btrfs_key found_key;
821 struct extent_buffer *eb;
822 struct btrfs_extent_item *ei;
823 struct scrub_warning swarn;
824 unsigned long ptr = 0;
825 u64 extent_item_pos;
826 u64 flags = 0;
827 u64 ref_root;
828 u32 item_size;
829 u8 ref_level = 0;
830 int ret;
831
832 WARN_ON(sblock->page_count < 1);
833 dev = sblock->pagev[0]->dev;
834 fs_info = sblock->sctx->fs_info;
835
836 path = btrfs_alloc_path();
837 if (!path)
838 return;
839
840 swarn.physical = sblock->pagev[0]->physical;
841 swarn.logical = sblock->pagev[0]->logical;
842 swarn.errstr = errstr;
843 swarn.dev = NULL;
844
845 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
846 &flags);
847 if (ret < 0)
848 goto out;
849
850 extent_item_pos = swarn.logical - found_key.objectid;
851 swarn.extent_item_size = found_key.offset;
852
853 eb = path->nodes[0];
854 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
855 item_size = btrfs_item_size_nr(eb, path->slots[0]);
856
857 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
858 do {
859 ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
860 item_size, &ref_root,
861 &ref_level);
862 btrfs_warn_in_rcu(fs_info,
863"%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
864 errstr, swarn.logical,
865 rcu_str_deref(dev->name),
866 swarn.physical,
867 ref_level ? "node" : "leaf",
868 ret < 0 ? -1 : ref_level,
869 ret < 0 ? -1 : ref_root);
870 } while (ret != 1);
871 btrfs_release_path(path);
872 } else {
873 btrfs_release_path(path);
874 swarn.path = path;
875 swarn.dev = dev;
876 iterate_extent_inodes(fs_info, found_key.objectid,
877 extent_item_pos, 1,
878 scrub_print_warning_inode, &swarn, false);
879 }
880
881out:
882 btrfs_free_path(path);
883}
884
885static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
886{
887 struct page *page = NULL;
888 unsigned long index;
889 struct scrub_fixup_nodatasum *fixup = fixup_ctx;
890 int ret;
891 int corrected = 0;
892 struct btrfs_key key;
893 struct inode *inode = NULL;
894 struct btrfs_fs_info *fs_info;
895 u64 end = offset + PAGE_SIZE - 1;
896 struct btrfs_root *local_root;
897 int srcu_index;
898
899 key.objectid = root;
900 key.type = BTRFS_ROOT_ITEM_KEY;
901 key.offset = (u64)-1;
902
903 fs_info = fixup->root->fs_info;
904 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
905
906 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
907 if (IS_ERR(local_root)) {
908 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
909 return PTR_ERR(local_root);
910 }
911
912 key.type = BTRFS_INODE_ITEM_KEY;
913 key.objectid = inum;
914 key.offset = 0;
915 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
916 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
917 if (IS_ERR(inode))
918 return PTR_ERR(inode);
919
920 index = offset >> PAGE_SHIFT;
921
922 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
923 if (!page) {
924 ret = -ENOMEM;
925 goto out;
926 }
927
928 if (PageUptodate(page)) {
929 if (PageDirty(page)) {
930 /*
931 * we need to write the data to the defect sector. the
932 * data that was in that sector is not in memory,
933 * because the page was modified. we must not write the
934 * modified page to that sector.
935 *
936 * TODO: what could be done here: wait for the delalloc
937 * runner to write out that page (might involve
938 * COW) and see whether the sector is still
939 * referenced afterwards.
940 *
941 * For the meantime, we'll treat this error
942 * incorrectable, although there is a chance that a
943 * later scrub will find the bad sector again and that
944 * there's no dirty page in memory, then.
945 */
946 ret = -EIO;
947 goto out;
948 }
949 ret = repair_io_failure(fs_info, inum, offset, PAGE_SIZE,
950 fixup->logical, page,
951 offset - page_offset(page),
952 fixup->mirror_num);
953 unlock_page(page);
954 corrected = !ret;
955 } else {
956 /*
957 * we need to get good data first. the general readpage path
958 * will call repair_io_failure for us, we just have to make
959 * sure we read the bad mirror.
960 */
961 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
962 EXTENT_DAMAGED);
963 if (ret) {
964 /* set_extent_bits should give proper error */
965 WARN_ON(ret > 0);
966 if (ret > 0)
967 ret = -EFAULT;
968 goto out;
969 }
970
971 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
972 btrfs_get_extent,
973 fixup->mirror_num);
974 wait_on_page_locked(page);
975
976 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
977 end, EXTENT_DAMAGED, 0, NULL);
978 if (!corrected)
979 clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
980 EXTENT_DAMAGED);
981 }
982
983out:
984 if (page)
985 put_page(page);
986
987 iput(inode);
988
989 if (ret < 0)
990 return ret;
991
992 if (ret == 0 && corrected) {
993 /*
994 * we only need to call readpage for one of the inodes belonging
995 * to this extent. so make iterate_extent_inodes stop
996 */
997 return 1;
998 }
999
1000 return -EIO;
1001}
1002
1003static void scrub_fixup_nodatasum(struct btrfs_work *work)
1004{
1005 struct btrfs_fs_info *fs_info;
1006 int ret;
1007 struct scrub_fixup_nodatasum *fixup;
1008 struct scrub_ctx *sctx;
1009 struct btrfs_trans_handle *trans = NULL;
1010 struct btrfs_path *path;
1011 int uncorrectable = 0;
1012
1013 fixup = container_of(work, struct scrub_fixup_nodatasum, work);
1014 sctx = fixup->sctx;
1015 fs_info = fixup->root->fs_info;
1016
1017 path = btrfs_alloc_path();
1018 if (!path) {
1019 spin_lock(&sctx->stat_lock);
1020 ++sctx->stat.malloc_errors;
1021 spin_unlock(&sctx->stat_lock);
1022 uncorrectable = 1;
1023 goto out;
1024 }
1025
1026 trans = btrfs_join_transaction(fixup->root);
1027 if (IS_ERR(trans)) {
1028 uncorrectable = 1;
1029 goto out;
1030 }
1031
1032 /*
1033 * the idea is to trigger a regular read through the standard path. we
1034 * read a page from the (failed) logical address by specifying the
1035 * corresponding copynum of the failed sector. thus, that readpage is
1036 * expected to fail.
1037 * that is the point where on-the-fly error correction will kick in
1038 * (once it's finished) and rewrite the failed sector if a good copy
1039 * can be found.
1040 */
1041 ret = iterate_inodes_from_logical(fixup->logical, fs_info, path,
1042 scrub_fixup_readpage, fixup, false);
1043 if (ret < 0) {
1044 uncorrectable = 1;
1045 goto out;
1046 }
1047 WARN_ON(ret != 1);
1048
1049 spin_lock(&sctx->stat_lock);
1050 ++sctx->stat.corrected_errors;
1051 spin_unlock(&sctx->stat_lock);
1052
1053out:
1054 if (trans && !IS_ERR(trans))
1055 btrfs_end_transaction(trans);
1056 if (uncorrectable) {
1057 spin_lock(&sctx->stat_lock);
1058 ++sctx->stat.uncorrectable_errors;
1059 spin_unlock(&sctx->stat_lock);
1060 btrfs_dev_replace_stats_inc(
1061 &fs_info->dev_replace.num_uncorrectable_read_errors);
1062 btrfs_err_rl_in_rcu(fs_info,
1063 "unable to fixup (nodatasum) error at logical %llu on dev %s",
1064 fixup->logical, rcu_str_deref(fixup->dev->name));
1065 }
1066
1067 btrfs_free_path(path);
1068 kfree(fixup);
1069
1070 scrub_pending_trans_workers_dec(sctx);
1071}
1072
1073static inline void scrub_get_recover(struct scrub_recover *recover)
1074{
1075 refcount_inc(&recover->refs);
1076}
1077
1078static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
1079 struct scrub_recover *recover)
1080{
1081 if (refcount_dec_and_test(&recover->refs)) {
1082 btrfs_bio_counter_dec(fs_info);
1083 btrfs_put_bbio(recover->bbio);
1084 kfree(recover);
1085 }
1086}
1087
1088/*
1089 * scrub_handle_errored_block gets called when either verification of the
1090 * pages failed or the bio failed to read, e.g. with EIO. In the latter
1091 * case, this function handles all pages in the bio, even though only one
1092 * may be bad.
1093 * The goal of this function is to repair the errored block by using the
1094 * contents of one of the mirrors.
1095 */
1096static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
1097{
1098 struct scrub_ctx *sctx = sblock_to_check->sctx;
1099 struct btrfs_device *dev;
1100 struct btrfs_fs_info *fs_info;
1101 u64 logical;
1102 unsigned int failed_mirror_index;
1103 unsigned int is_metadata;
1104 unsigned int have_csum;
1105 struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
1106 struct scrub_block *sblock_bad;
1107 int ret;
1108 int mirror_index;
1109 int page_num;
1110 int success;
1111 bool full_stripe_locked;
1112 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
1113 DEFAULT_RATELIMIT_BURST);
1114
1115 BUG_ON(sblock_to_check->page_count < 1);
1116 fs_info = sctx->fs_info;
1117 if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
1118 /*
1119 * if we find an error in a super block, we just report it.
1120 * They will get written with the next transaction commit
1121 * anyway
1122 */
1123 spin_lock(&sctx->stat_lock);
1124 ++sctx->stat.super_errors;
1125 spin_unlock(&sctx->stat_lock);
1126 return 0;
1127 }
1128 logical = sblock_to_check->pagev[0]->logical;
1129 BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
1130 failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
1131 is_metadata = !(sblock_to_check->pagev[0]->flags &
1132 BTRFS_EXTENT_FLAG_DATA);
1133 have_csum = sblock_to_check->pagev[0]->have_csum;
1134 dev = sblock_to_check->pagev[0]->dev;
1135
1136 /*
1137 * For RAID5/6, race can happen for a different device scrub thread.
1138 * For data corruption, Parity and Data threads will both try
1139 * to recovery the data.
1140 * Race can lead to doubly added csum error, or even unrecoverable
1141 * error.
1142 */
1143 ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
1144 if (ret < 0) {
1145 spin_lock(&sctx->stat_lock);
1146 if (ret == -ENOMEM)
1147 sctx->stat.malloc_errors++;
1148 sctx->stat.read_errors++;
1149 sctx->stat.uncorrectable_errors++;
1150 spin_unlock(&sctx->stat_lock);
1151 return ret;
1152 }
1153
1154 if (sctx->is_dev_replace && !is_metadata && !have_csum) {
1155 sblocks_for_recheck = NULL;
1156 goto nodatasum_case;
1157 }
1158
1159 /*
1160 * read all mirrors one after the other. This includes to
1161 * re-read the extent or metadata block that failed (that was
1162 * the cause that this fixup code is called) another time,
1163 * page by page this time in order to know which pages
1164 * caused I/O errors and which ones are good (for all mirrors).
1165 * It is the goal to handle the situation when more than one
1166 * mirror contains I/O errors, but the errors do not
1167 * overlap, i.e. the data can be repaired by selecting the
1168 * pages from those mirrors without I/O error on the
1169 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
1170 * would be that mirror #1 has an I/O error on the first page,
1171 * the second page is good, and mirror #2 has an I/O error on
1172 * the second page, but the first page is good.
1173 * Then the first page of the first mirror can be repaired by
1174 * taking the first page of the second mirror, and the
1175 * second page of the second mirror can be repaired by
1176 * copying the contents of the 2nd page of the 1st mirror.
1177 * One more note: if the pages of one mirror contain I/O
1178 * errors, the checksum cannot be verified. In order to get
1179 * the best data for repairing, the first attempt is to find
1180 * a mirror without I/O errors and with a validated checksum.
1181 * Only if this is not possible, the pages are picked from
1182 * mirrors with I/O errors without considering the checksum.
1183 * If the latter is the case, at the end, the checksum of the
1184 * repaired area is verified in order to correctly maintain
1185 * the statistics.
1186 */
1187
1188 sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
1189 sizeof(*sblocks_for_recheck), GFP_NOFS);
1190 if (!sblocks_for_recheck) {
1191 spin_lock(&sctx->stat_lock);
1192 sctx->stat.malloc_errors++;
1193 sctx->stat.read_errors++;
1194 sctx->stat.uncorrectable_errors++;
1195 spin_unlock(&sctx->stat_lock);
1196 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1197 goto out;
1198 }
1199
1200 /* setup the context, map the logical blocks and alloc the pages */
1201 ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
1202 if (ret) {
1203 spin_lock(&sctx->stat_lock);
1204 sctx->stat.read_errors++;
1205 sctx->stat.uncorrectable_errors++;
1206 spin_unlock(&sctx->stat_lock);
1207 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1208 goto out;
1209 }
1210 BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
1211 sblock_bad = sblocks_for_recheck + failed_mirror_index;
1212
1213 /* build and submit the bios for the failed mirror, check checksums */
1214 scrub_recheck_block(fs_info, sblock_bad, 1);
1215
1216 if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
1217 sblock_bad->no_io_error_seen) {
1218 /*
1219 * the error disappeared after reading page by page, or
1220 * the area was part of a huge bio and other parts of the
1221 * bio caused I/O errors, or the block layer merged several
1222 * read requests into one and the error is caused by a
1223 * different bio (usually one of the two latter cases is
1224 * the cause)
1225 */
1226 spin_lock(&sctx->stat_lock);
1227 sctx->stat.unverified_errors++;
1228 sblock_to_check->data_corrected = 1;
1229 spin_unlock(&sctx->stat_lock);
1230
1231 if (sctx->is_dev_replace)
1232 scrub_write_block_to_dev_replace(sblock_bad);
1233 goto out;
1234 }
1235
1236 if (!sblock_bad->no_io_error_seen) {
1237 spin_lock(&sctx->stat_lock);
1238 sctx->stat.read_errors++;
1239 spin_unlock(&sctx->stat_lock);
1240 if (__ratelimit(&_rs))
1241 scrub_print_warning("i/o error", sblock_to_check);
1242 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1243 } else if (sblock_bad->checksum_error) {
1244 spin_lock(&sctx->stat_lock);
1245 sctx->stat.csum_errors++;
1246 spin_unlock(&sctx->stat_lock);
1247 if (__ratelimit(&_rs))
1248 scrub_print_warning("checksum error", sblock_to_check);
1249 btrfs_dev_stat_inc_and_print(dev,
1250 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1251 } else if (sblock_bad->header_error) {
1252 spin_lock(&sctx->stat_lock);
1253 sctx->stat.verify_errors++;
1254 spin_unlock(&sctx->stat_lock);
1255 if (__ratelimit(&_rs))
1256 scrub_print_warning("checksum/header error",
1257 sblock_to_check);
1258 if (sblock_bad->generation_error)
1259 btrfs_dev_stat_inc_and_print(dev,
1260 BTRFS_DEV_STAT_GENERATION_ERRS);
1261 else
1262 btrfs_dev_stat_inc_and_print(dev,
1263 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1264 }
1265
1266 if (sctx->readonly) {
1267 ASSERT(!sctx->is_dev_replace);
1268 goto out;
1269 }
1270
1271 if (!is_metadata && !have_csum) {
1272 struct scrub_fixup_nodatasum *fixup_nodatasum;
1273
1274 WARN_ON(sctx->is_dev_replace);
1275
1276nodatasum_case:
1277
1278 /*
1279 * !is_metadata and !have_csum, this means that the data
1280 * might not be COWed, that it might be modified
1281 * concurrently. The general strategy to work on the
1282 * commit root does not help in the case when COW is not
1283 * used.
1284 */
1285 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
1286 if (!fixup_nodatasum)
1287 goto did_not_correct_error;
1288 fixup_nodatasum->sctx = sctx;
1289 fixup_nodatasum->dev = dev;
1290 fixup_nodatasum->logical = logical;
1291 fixup_nodatasum->root = fs_info->extent_root;
1292 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
1293 scrub_pending_trans_workers_inc(sctx);
1294 btrfs_init_work(&fixup_nodatasum->work, btrfs_scrub_helper,
1295 scrub_fixup_nodatasum, NULL, NULL);
1296 btrfs_queue_work(fs_info->scrub_workers,
1297 &fixup_nodatasum->work);
1298 goto out;
1299 }
1300
1301 /*
1302 * now build and submit the bios for the other mirrors, check
1303 * checksums.
1304 * First try to pick the mirror which is completely without I/O
1305 * errors and also does not have a checksum error.
1306 * If one is found, and if a checksum is present, the full block
1307 * that is known to contain an error is rewritten. Afterwards
1308 * the block is known to be corrected.
1309 * If a mirror is found which is completely correct, and no
1310 * checksum is present, only those pages are rewritten that had
1311 * an I/O error in the block to be repaired, since it cannot be
1312 * determined, which copy of the other pages is better (and it
1313 * could happen otherwise that a correct page would be
1314 * overwritten by a bad one).
1315 */
1316 for (mirror_index = 0; ;mirror_index++) {
1317 struct scrub_block *sblock_other;
1318
1319 if (mirror_index == failed_mirror_index)
1320 continue;
1321
1322 /* raid56's mirror can be more than BTRFS_MAX_MIRRORS */
1323 if (!scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1324 if (mirror_index >= BTRFS_MAX_MIRRORS)
1325 break;
1326 if (!sblocks_for_recheck[mirror_index].page_count)
1327 break;
1328
1329 sblock_other = sblocks_for_recheck + mirror_index;
1330 } else {
1331 struct scrub_recover *r = sblock_bad->pagev[0]->recover;
1332 int max_allowed = r->bbio->num_stripes -
1333 r->bbio->num_tgtdevs;
1334
1335 if (mirror_index >= max_allowed)
1336 break;
1337 if (!sblocks_for_recheck[1].page_count)
1338 break;
1339
1340 ASSERT(failed_mirror_index == 0);
1341 sblock_other = sblocks_for_recheck + 1;
1342 sblock_other->pagev[0]->mirror_num = 1 + mirror_index;
1343 }
1344
1345 /* build and submit the bios, check checksums */
1346 scrub_recheck_block(fs_info, sblock_other, 0);
1347
1348 if (!sblock_other->header_error &&
1349 !sblock_other->checksum_error &&
1350 sblock_other->no_io_error_seen) {
1351 if (sctx->is_dev_replace) {
1352 scrub_write_block_to_dev_replace(sblock_other);
1353 goto corrected_error;
1354 } else {
1355 ret = scrub_repair_block_from_good_copy(
1356 sblock_bad, sblock_other);
1357 if (!ret)
1358 goto corrected_error;
1359 }
1360 }
1361 }
1362
1363 if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1364 goto did_not_correct_error;
1365
1366 /*
1367 * In case of I/O errors in the area that is supposed to be
1368 * repaired, continue by picking good copies of those pages.
1369 * Select the good pages from mirrors to rewrite bad pages from
1370 * the area to fix. Afterwards verify the checksum of the block
1371 * that is supposed to be repaired. This verification step is
1372 * only done for the purpose of statistic counting and for the
1373 * final scrub report, whether errors remain.
1374 * A perfect algorithm could make use of the checksum and try
1375 * all possible combinations of pages from the different mirrors
1376 * until the checksum verification succeeds. For example, when
1377 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1378 * of mirror #2 is readable but the final checksum test fails,
1379 * then the 2nd page of mirror #3 could be tried, whether now
1380 * the final checksum succeeds. But this would be a rare
1381 * exception and is therefore not implemented. At least it is
1382 * avoided that the good copy is overwritten.
1383 * A more useful improvement would be to pick the sectors
1384 * without I/O error based on sector sizes (512 bytes on legacy
1385 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1386 * mirror could be repaired by taking 512 byte of a different
1387 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1388 * area are unreadable.
1389 */
1390 success = 1;
1391 for (page_num = 0; page_num < sblock_bad->page_count;
1392 page_num++) {
1393 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1394 struct scrub_block *sblock_other = NULL;
1395
1396 /* skip no-io-error page in scrub */
1397 if (!page_bad->io_error && !sctx->is_dev_replace)
1398 continue;
1399
1400 if (scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1401 /*
1402 * In case of dev replace, if raid56 rebuild process
1403 * didn't work out correct data, then copy the content
1404 * in sblock_bad to make sure target device is identical
1405 * to source device, instead of writing garbage data in
1406 * sblock_for_recheck array to target device.
1407 */
1408 sblock_other = NULL;
1409 } else if (page_bad->io_error) {
1410 /* try to find no-io-error page in mirrors */
1411 for (mirror_index = 0;
1412 mirror_index < BTRFS_MAX_MIRRORS &&
1413 sblocks_for_recheck[mirror_index].page_count > 0;
1414 mirror_index++) {
1415 if (!sblocks_for_recheck[mirror_index].
1416 pagev[page_num]->io_error) {
1417 sblock_other = sblocks_for_recheck +
1418 mirror_index;
1419 break;
1420 }
1421 }
1422 if (!sblock_other)
1423 success = 0;
1424 }
1425
1426 if (sctx->is_dev_replace) {
1427 /*
1428 * did not find a mirror to fetch the page
1429 * from. scrub_write_page_to_dev_replace()
1430 * handles this case (page->io_error), by
1431 * filling the block with zeros before
1432 * submitting the write request
1433 */
1434 if (!sblock_other)
1435 sblock_other = sblock_bad;
1436
1437 if (scrub_write_page_to_dev_replace(sblock_other,
1438 page_num) != 0) {
1439 btrfs_dev_replace_stats_inc(
1440 &fs_info->dev_replace.num_write_errors);
1441 success = 0;
1442 }
1443 } else if (sblock_other) {
1444 ret = scrub_repair_page_from_good_copy(sblock_bad,
1445 sblock_other,
1446 page_num, 0);
1447 if (0 == ret)
1448 page_bad->io_error = 0;
1449 else
1450 success = 0;
1451 }
1452 }
1453
1454 if (success && !sctx->is_dev_replace) {
1455 if (is_metadata || have_csum) {
1456 /*
1457 * need to verify the checksum now that all
1458 * sectors on disk are repaired (the write
1459 * request for data to be repaired is on its way).
1460 * Just be lazy and use scrub_recheck_block()
1461 * which re-reads the data before the checksum
1462 * is verified, but most likely the data comes out
1463 * of the page cache.
1464 */
1465 scrub_recheck_block(fs_info, sblock_bad, 1);
1466 if (!sblock_bad->header_error &&
1467 !sblock_bad->checksum_error &&
1468 sblock_bad->no_io_error_seen)
1469 goto corrected_error;
1470 else
1471 goto did_not_correct_error;
1472 } else {
1473corrected_error:
1474 spin_lock(&sctx->stat_lock);
1475 sctx->stat.corrected_errors++;
1476 sblock_to_check->data_corrected = 1;
1477 spin_unlock(&sctx->stat_lock);
1478 btrfs_err_rl_in_rcu(fs_info,
1479 "fixed up error at logical %llu on dev %s",
1480 logical, rcu_str_deref(dev->name));
1481 }
1482 } else {
1483did_not_correct_error:
1484 spin_lock(&sctx->stat_lock);
1485 sctx->stat.uncorrectable_errors++;
1486 spin_unlock(&sctx->stat_lock);
1487 btrfs_err_rl_in_rcu(fs_info,
1488 "unable to fixup (regular) error at logical %llu on dev %s",
1489 logical, rcu_str_deref(dev->name));
1490 }
1491
1492out:
1493 if (sblocks_for_recheck) {
1494 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1495 mirror_index++) {
1496 struct scrub_block *sblock = sblocks_for_recheck +
1497 mirror_index;
1498 struct scrub_recover *recover;
1499 int page_index;
1500
1501 for (page_index = 0; page_index < sblock->page_count;
1502 page_index++) {
1503 sblock->pagev[page_index]->sblock = NULL;
1504 recover = sblock->pagev[page_index]->recover;
1505 if (recover) {
1506 scrub_put_recover(fs_info, recover);
1507 sblock->pagev[page_index]->recover =
1508 NULL;
1509 }
1510 scrub_page_put(sblock->pagev[page_index]);
1511 }
1512 }
1513 kfree(sblocks_for_recheck);
1514 }
1515
1516 ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1517 if (ret < 0)
1518 return ret;
1519 return 0;
1520}
1521
1522static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1523{
1524 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1525 return 2;
1526 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1527 return 3;
1528 else
1529 return (int)bbio->num_stripes;
1530}
1531
1532static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1533 u64 *raid_map,
1534 u64 mapped_length,
1535 int nstripes, int mirror,
1536 int *stripe_index,
1537 u64 *stripe_offset)
1538{
1539 int i;
1540
1541 if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1542 /* RAID5/6 */
1543 for (i = 0; i < nstripes; i++) {
1544 if (raid_map[i] == RAID6_Q_STRIPE ||
1545 raid_map[i] == RAID5_P_STRIPE)
1546 continue;
1547
1548 if (logical >= raid_map[i] &&
1549 logical < raid_map[i] + mapped_length)
1550 break;
1551 }
1552
1553 *stripe_index = i;
1554 *stripe_offset = logical - raid_map[i];
1555 } else {
1556 /* The other RAID type */
1557 *stripe_index = mirror;
1558 *stripe_offset = 0;
1559 }
1560}
1561
1562static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1563 struct scrub_block *sblocks_for_recheck)
1564{
1565 struct scrub_ctx *sctx = original_sblock->sctx;
1566 struct btrfs_fs_info *fs_info = sctx->fs_info;
1567 u64 length = original_sblock->page_count * PAGE_SIZE;
1568 u64 logical = original_sblock->pagev[0]->logical;
1569 u64 generation = original_sblock->pagev[0]->generation;
1570 u64 flags = original_sblock->pagev[0]->flags;
1571 u64 have_csum = original_sblock->pagev[0]->have_csum;
1572 struct scrub_recover *recover;
1573 struct btrfs_bio *bbio;
1574 u64 sublen;
1575 u64 mapped_length;
1576 u64 stripe_offset;
1577 int stripe_index;
1578 int page_index = 0;
1579 int mirror_index;
1580 int nmirrors;
1581 int ret;
1582
1583 /*
1584 * note: the two members refs and outstanding_pages
1585 * are not used (and not set) in the blocks that are used for
1586 * the recheck procedure
1587 */
1588
1589 while (length > 0) {
1590 sublen = min_t(u64, length, PAGE_SIZE);
1591 mapped_length = sublen;
1592 bbio = NULL;
1593
1594 /*
1595 * with a length of PAGE_SIZE, each returned stripe
1596 * represents one mirror
1597 */
1598 btrfs_bio_counter_inc_blocked(fs_info);
1599 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1600 logical, &mapped_length, &bbio);
1601 if (ret || !bbio || mapped_length < sublen) {
1602 btrfs_put_bbio(bbio);
1603 btrfs_bio_counter_dec(fs_info);
1604 return -EIO;
1605 }
1606
1607 recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1608 if (!recover) {
1609 btrfs_put_bbio(bbio);
1610 btrfs_bio_counter_dec(fs_info);
1611 return -ENOMEM;
1612 }
1613
1614 refcount_set(&recover->refs, 1);
1615 recover->bbio = bbio;
1616 recover->map_length = mapped_length;
1617
1618 BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1619
1620 nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1621
1622 for (mirror_index = 0; mirror_index < nmirrors;
1623 mirror_index++) {
1624 struct scrub_block *sblock;
1625 struct scrub_page *page;
1626
1627 sblock = sblocks_for_recheck + mirror_index;
1628 sblock->sctx = sctx;
1629
1630 page = kzalloc(sizeof(*page), GFP_NOFS);
1631 if (!page) {
1632leave_nomem:
1633 spin_lock(&sctx->stat_lock);
1634 sctx->stat.malloc_errors++;
1635 spin_unlock(&sctx->stat_lock);
1636 scrub_put_recover(fs_info, recover);
1637 return -ENOMEM;
1638 }
1639 scrub_page_get(page);
1640 sblock->pagev[page_index] = page;
1641 page->sblock = sblock;
1642 page->flags = flags;
1643 page->generation = generation;
1644 page->logical = logical;
1645 page->have_csum = have_csum;
1646 if (have_csum)
1647 memcpy(page->csum,
1648 original_sblock->pagev[0]->csum,
1649 sctx->csum_size);
1650
1651 scrub_stripe_index_and_offset(logical,
1652 bbio->map_type,
1653 bbio->raid_map,
1654 mapped_length,
1655 bbio->num_stripes -
1656 bbio->num_tgtdevs,
1657 mirror_index,
1658 &stripe_index,
1659 &stripe_offset);
1660 page->physical = bbio->stripes[stripe_index].physical +
1661 stripe_offset;
1662 page->dev = bbio->stripes[stripe_index].dev;
1663
1664 BUG_ON(page_index >= original_sblock->page_count);
1665 page->physical_for_dev_replace =
1666 original_sblock->pagev[page_index]->
1667 physical_for_dev_replace;
1668 /* for missing devices, dev->bdev is NULL */
1669 page->mirror_num = mirror_index + 1;
1670 sblock->page_count++;
1671 page->page = alloc_page(GFP_NOFS);
1672 if (!page->page)
1673 goto leave_nomem;
1674
1675 scrub_get_recover(recover);
1676 page->recover = recover;
1677 }
1678 scrub_put_recover(fs_info, recover);
1679 length -= sublen;
1680 logical += sublen;
1681 page_index++;
1682 }
1683
1684 return 0;
1685}
1686
1687static void scrub_bio_wait_endio(struct bio *bio)
1688{
1689 complete(bio->bi_private);
1690}
1691
1692static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1693 struct bio *bio,
1694 struct scrub_page *page)
1695{
1696 DECLARE_COMPLETION_ONSTACK(done);
1697 int ret;
1698 int mirror_num;
1699
1700 bio->bi_iter.bi_sector = page->logical >> 9;
1701 bio->bi_private = &done;
1702 bio->bi_end_io = scrub_bio_wait_endio;
1703
1704 mirror_num = page->sblock->pagev[0]->mirror_num;
1705 ret = raid56_parity_recover(fs_info, bio, page->recover->bbio,
1706 page->recover->map_length,
1707 mirror_num, 0);
1708 if (ret)
1709 return ret;
1710
1711 wait_for_completion_io(&done);
1712 return blk_status_to_errno(bio->bi_status);
1713}
1714
1715static void scrub_recheck_block_on_raid56(struct btrfs_fs_info *fs_info,
1716 struct scrub_block *sblock)
1717{
1718 struct scrub_page *first_page = sblock->pagev[0];
1719 struct bio *bio;
1720 int page_num;
1721
1722 /* All pages in sblock belong to the same stripe on the same device. */
1723 ASSERT(first_page->dev);
1724 if (!first_page->dev->bdev)
1725 goto out;
1726
1727 bio = btrfs_io_bio_alloc(BIO_MAX_PAGES);
1728 bio_set_dev(bio, first_page->dev->bdev);
1729
1730 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1731 struct scrub_page *page = sblock->pagev[page_num];
1732
1733 WARN_ON(!page->page);
1734 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1735 }
1736
1737 if (scrub_submit_raid56_bio_wait(fs_info, bio, first_page)) {
1738 bio_put(bio);
1739 goto out;
1740 }
1741
1742 bio_put(bio);
1743
1744 scrub_recheck_block_checksum(sblock);
1745
1746 return;
1747out:
1748 for (page_num = 0; page_num < sblock->page_count; page_num++)
1749 sblock->pagev[page_num]->io_error = 1;
1750
1751 sblock->no_io_error_seen = 0;
1752}
1753
1754/*
1755 * this function will check the on disk data for checksum errors, header
1756 * errors and read I/O errors. If any I/O errors happen, the exact pages
1757 * which are errored are marked as being bad. The goal is to enable scrub
1758 * to take those pages that are not errored from all the mirrors so that
1759 * the pages that are errored in the just handled mirror can be repaired.
1760 */
1761static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1762 struct scrub_block *sblock,
1763 int retry_failed_mirror)
1764{
1765 int page_num;
1766
1767 sblock->no_io_error_seen = 1;
1768
1769 /* short cut for raid56 */
1770 if (!retry_failed_mirror && scrub_is_page_on_raid56(sblock->pagev[0]))
1771 return scrub_recheck_block_on_raid56(fs_info, sblock);
1772
1773 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1774 struct bio *bio;
1775 struct scrub_page *page = sblock->pagev[page_num];
1776
1777 if (page->dev->bdev == NULL) {
1778 page->io_error = 1;
1779 sblock->no_io_error_seen = 0;
1780 continue;
1781 }
1782
1783 WARN_ON(!page->page);
1784 bio = btrfs_io_bio_alloc(1);
1785 bio_set_dev(bio, page->dev->bdev);
1786
1787 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1788 bio->bi_iter.bi_sector = page->physical >> 9;
1789 bio->bi_opf = REQ_OP_READ;
1790
1791 if (btrfsic_submit_bio_wait(bio)) {
1792 page->io_error = 1;
1793 sblock->no_io_error_seen = 0;
1794 }
1795
1796 bio_put(bio);
1797 }
1798
1799 if (sblock->no_io_error_seen)
1800 scrub_recheck_block_checksum(sblock);
1801}
1802
1803static inline int scrub_check_fsid(u8 fsid[],
1804 struct scrub_page *spage)
1805{
1806 struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1807 int ret;
1808
1809 ret = memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1810 return !ret;
1811}
1812
1813static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1814{
1815 sblock->header_error = 0;
1816 sblock->checksum_error = 0;
1817 sblock->generation_error = 0;
1818
1819 if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1820 scrub_checksum_data(sblock);
1821 else
1822 scrub_checksum_tree_block(sblock);
1823}
1824
1825static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1826 struct scrub_block *sblock_good)
1827{
1828 int page_num;
1829 int ret = 0;
1830
1831 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1832 int ret_sub;
1833
1834 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1835 sblock_good,
1836 page_num, 1);
1837 if (ret_sub)
1838 ret = ret_sub;
1839 }
1840
1841 return ret;
1842}
1843
1844static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1845 struct scrub_block *sblock_good,
1846 int page_num, int force_write)
1847{
1848 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1849 struct scrub_page *page_good = sblock_good->pagev[page_num];
1850 struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1851
1852 BUG_ON(page_bad->page == NULL);
1853 BUG_ON(page_good->page == NULL);
1854 if (force_write || sblock_bad->header_error ||
1855 sblock_bad->checksum_error || page_bad->io_error) {
1856 struct bio *bio;
1857 int ret;
1858
1859 if (!page_bad->dev->bdev) {
1860 btrfs_warn_rl(fs_info,
1861 "scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1862 return -EIO;
1863 }
1864
1865 bio = btrfs_io_bio_alloc(1);
1866 bio_set_dev(bio, page_bad->dev->bdev);
1867 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1868 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1869
1870 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1871 if (PAGE_SIZE != ret) {
1872 bio_put(bio);
1873 return -EIO;
1874 }
1875
1876 if (btrfsic_submit_bio_wait(bio)) {
1877 btrfs_dev_stat_inc_and_print(page_bad->dev,
1878 BTRFS_DEV_STAT_WRITE_ERRS);
1879 btrfs_dev_replace_stats_inc(
1880 &fs_info->dev_replace.num_write_errors);
1881 bio_put(bio);
1882 return -EIO;
1883 }
1884 bio_put(bio);
1885 }
1886
1887 return 0;
1888}
1889
1890static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1891{
1892 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1893 int page_num;
1894
1895 /*
1896 * This block is used for the check of the parity on the source device,
1897 * so the data needn't be written into the destination device.
1898 */
1899 if (sblock->sparity)
1900 return;
1901
1902 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1903 int ret;
1904
1905 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1906 if (ret)
1907 btrfs_dev_replace_stats_inc(
1908 &fs_info->dev_replace.num_write_errors);
1909 }
1910}
1911
1912static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1913 int page_num)
1914{
1915 struct scrub_page *spage = sblock->pagev[page_num];
1916
1917 BUG_ON(spage->page == NULL);
1918 if (spage->io_error) {
1919 void *mapped_buffer = kmap_atomic(spage->page);
1920
1921 clear_page(mapped_buffer);
1922 flush_dcache_page(spage->page);
1923 kunmap_atomic(mapped_buffer);
1924 }
1925 return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1926}
1927
1928static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1929 struct scrub_page *spage)
1930{
1931 struct scrub_bio *sbio;
1932 int ret;
1933
1934 mutex_lock(&sctx->wr_lock);
1935again:
1936 if (!sctx->wr_curr_bio) {
1937 sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1938 GFP_KERNEL);
1939 if (!sctx->wr_curr_bio) {
1940 mutex_unlock(&sctx->wr_lock);
1941 return -ENOMEM;
1942 }
1943 sctx->wr_curr_bio->sctx = sctx;
1944 sctx->wr_curr_bio->page_count = 0;
1945 }
1946 sbio = sctx->wr_curr_bio;
1947 if (sbio->page_count == 0) {
1948 struct bio *bio;
1949
1950 sbio->physical = spage->physical_for_dev_replace;
1951 sbio->logical = spage->logical;
1952 sbio->dev = sctx->wr_tgtdev;
1953 bio = sbio->bio;
1954 if (!bio) {
1955 bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
1956 sbio->bio = bio;
1957 }
1958
1959 bio->bi_private = sbio;
1960 bio->bi_end_io = scrub_wr_bio_end_io;
1961 bio_set_dev(bio, sbio->dev->bdev);
1962 bio->bi_iter.bi_sector = sbio->physical >> 9;
1963 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1964 sbio->status = 0;
1965 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1966 spage->physical_for_dev_replace ||
1967 sbio->logical + sbio->page_count * PAGE_SIZE !=
1968 spage->logical) {
1969 scrub_wr_submit(sctx);
1970 goto again;
1971 }
1972
1973 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1974 if (ret != PAGE_SIZE) {
1975 if (sbio->page_count < 1) {
1976 bio_put(sbio->bio);
1977 sbio->bio = NULL;
1978 mutex_unlock(&sctx->wr_lock);
1979 return -EIO;
1980 }
1981 scrub_wr_submit(sctx);
1982 goto again;
1983 }
1984
1985 sbio->pagev[sbio->page_count] = spage;
1986 scrub_page_get(spage);
1987 sbio->page_count++;
1988 if (sbio->page_count == sctx->pages_per_wr_bio)
1989 scrub_wr_submit(sctx);
1990 mutex_unlock(&sctx->wr_lock);
1991
1992 return 0;
1993}
1994
1995static void scrub_wr_submit(struct scrub_ctx *sctx)
1996{
1997 struct scrub_bio *sbio;
1998
1999 if (!sctx->wr_curr_bio)
2000 return;
2001
2002 sbio = sctx->wr_curr_bio;
2003 sctx->wr_curr_bio = NULL;
2004 WARN_ON(!sbio->bio->bi_disk);
2005 scrub_pending_bio_inc(sctx);
2006 /* process all writes in a single worker thread. Then the block layer
2007 * orders the requests before sending them to the driver which
2008 * doubled the write performance on spinning disks when measured
2009 * with Linux 3.5 */
2010 btrfsic_submit_bio(sbio->bio);
2011}
2012
2013static void scrub_wr_bio_end_io(struct bio *bio)
2014{
2015 struct scrub_bio *sbio = bio->bi_private;
2016 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2017
2018 sbio->status = bio->bi_status;
2019 sbio->bio = bio;
2020
2021 btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
2022 scrub_wr_bio_end_io_worker, NULL, NULL);
2023 btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
2024}
2025
2026static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
2027{
2028 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2029 struct scrub_ctx *sctx = sbio->sctx;
2030 int i;
2031
2032 WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
2033 if (sbio->status) {
2034 struct btrfs_dev_replace *dev_replace =
2035 &sbio->sctx->fs_info->dev_replace;
2036
2037 for (i = 0; i < sbio->page_count; i++) {
2038 struct scrub_page *spage = sbio->pagev[i];
2039
2040 spage->io_error = 1;
2041 btrfs_dev_replace_stats_inc(&dev_replace->
2042 num_write_errors);
2043 }
2044 }
2045
2046 for (i = 0; i < sbio->page_count; i++)
2047 scrub_page_put(sbio->pagev[i]);
2048
2049 bio_put(sbio->bio);
2050 kfree(sbio);
2051 scrub_pending_bio_dec(sctx);
2052}
2053
2054static int scrub_checksum(struct scrub_block *sblock)
2055{
2056 u64 flags;
2057 int ret;
2058
2059 /*
2060 * No need to initialize these stats currently,
2061 * because this function only use return value
2062 * instead of these stats value.
2063 *
2064 * Todo:
2065 * always use stats
2066 */
2067 sblock->header_error = 0;
2068 sblock->generation_error = 0;
2069 sblock->checksum_error = 0;
2070
2071 WARN_ON(sblock->page_count < 1);
2072 flags = sblock->pagev[0]->flags;
2073 ret = 0;
2074 if (flags & BTRFS_EXTENT_FLAG_DATA)
2075 ret = scrub_checksum_data(sblock);
2076 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2077 ret = scrub_checksum_tree_block(sblock);
2078 else if (flags & BTRFS_EXTENT_FLAG_SUPER)
2079 (void)scrub_checksum_super(sblock);
2080 else
2081 WARN_ON(1);
2082 if (ret)
2083 scrub_handle_errored_block(sblock);
2084
2085 return ret;
2086}
2087
2088static int scrub_checksum_data(struct scrub_block *sblock)
2089{
2090 struct scrub_ctx *sctx = sblock->sctx;
2091 u8 csum[BTRFS_CSUM_SIZE];
2092 u8 *on_disk_csum;
2093 struct page *page;
2094 void *buffer;
2095 u32 crc = ~(u32)0;
2096 u64 len;
2097 int index;
2098
2099 BUG_ON(sblock->page_count < 1);
2100 if (!sblock->pagev[0]->have_csum)
2101 return 0;
2102
2103 on_disk_csum = sblock->pagev[0]->csum;
2104 page = sblock->pagev[0]->page;
2105 buffer = kmap_atomic(page);
2106
2107 len = sctx->fs_info->sectorsize;
2108 index = 0;
2109 for (;;) {
2110 u64 l = min_t(u64, len, PAGE_SIZE);
2111
2112 crc = btrfs_csum_data(buffer, crc, l);
2113 kunmap_atomic(buffer);
2114 len -= l;
2115 if (len == 0)
2116 break;
2117 index++;
2118 BUG_ON(index >= sblock->page_count);
2119 BUG_ON(!sblock->pagev[index]->page);
2120 page = sblock->pagev[index]->page;
2121 buffer = kmap_atomic(page);
2122 }
2123
2124 btrfs_csum_final(crc, csum);
2125 if (memcmp(csum, on_disk_csum, sctx->csum_size))
2126 sblock->checksum_error = 1;
2127
2128 return sblock->checksum_error;
2129}
2130
2131static int scrub_checksum_tree_block(struct scrub_block *sblock)
2132{
2133 struct scrub_ctx *sctx = sblock->sctx;
2134 struct btrfs_header *h;
2135 struct btrfs_fs_info *fs_info = sctx->fs_info;
2136 u8 calculated_csum[BTRFS_CSUM_SIZE];
2137 u8 on_disk_csum[BTRFS_CSUM_SIZE];
2138 struct page *page;
2139 void *mapped_buffer;
2140 u64 mapped_size;
2141 void *p;
2142 u32 crc = ~(u32)0;
2143 u64 len;
2144 int index;
2145
2146 BUG_ON(sblock->page_count < 1);
2147 page = sblock->pagev[0]->page;
2148 mapped_buffer = kmap_atomic(page);
2149 h = (struct btrfs_header *)mapped_buffer;
2150 memcpy(on_disk_csum, h->csum, sctx->csum_size);
2151
2152 /*
2153 * we don't use the getter functions here, as we
2154 * a) don't have an extent buffer and
2155 * b) the page is already kmapped
2156 */
2157 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
2158 sblock->header_error = 1;
2159
2160 if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
2161 sblock->header_error = 1;
2162 sblock->generation_error = 1;
2163 }
2164
2165 if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
2166 sblock->header_error = 1;
2167
2168 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
2169 BTRFS_UUID_SIZE))
2170 sblock->header_error = 1;
2171
2172 len = sctx->fs_info->nodesize - BTRFS_CSUM_SIZE;
2173 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
2174 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
2175 index = 0;
2176 for (;;) {
2177 u64 l = min_t(u64, len, mapped_size);
2178
2179 crc = btrfs_csum_data(p, crc, l);
2180 kunmap_atomic(mapped_buffer);
2181 len -= l;
2182 if (len == 0)
2183 break;
2184 index++;
2185 BUG_ON(index >= sblock->page_count);
2186 BUG_ON(!sblock->pagev[index]->page);
2187 page = sblock->pagev[index]->page;
2188 mapped_buffer = kmap_atomic(page);
2189 mapped_size = PAGE_SIZE;
2190 p = mapped_buffer;
2191 }
2192
2193 btrfs_csum_final(crc, calculated_csum);
2194 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
2195 sblock->checksum_error = 1;
2196
2197 return sblock->header_error || sblock->checksum_error;
2198}
2199
2200static int scrub_checksum_super(struct scrub_block *sblock)
2201{
2202 struct btrfs_super_block *s;
2203 struct scrub_ctx *sctx = sblock->sctx;
2204 u8 calculated_csum[BTRFS_CSUM_SIZE];
2205 u8 on_disk_csum[BTRFS_CSUM_SIZE];
2206 struct page *page;
2207 void *mapped_buffer;
2208 u64 mapped_size;
2209 void *p;
2210 u32 crc = ~(u32)0;
2211 int fail_gen = 0;
2212 int fail_cor = 0;
2213 u64 len;
2214 int index;
2215
2216 BUG_ON(sblock->page_count < 1);
2217 page = sblock->pagev[0]->page;
2218 mapped_buffer = kmap_atomic(page);
2219 s = (struct btrfs_super_block *)mapped_buffer;
2220 memcpy(on_disk_csum, s->csum, sctx->csum_size);
2221
2222 if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
2223 ++fail_cor;
2224
2225 if (sblock->pagev[0]->generation != btrfs_super_generation(s))
2226 ++fail_gen;
2227
2228 if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
2229 ++fail_cor;
2230
2231 len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
2232 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
2233 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
2234 index = 0;
2235 for (;;) {
2236 u64 l = min_t(u64, len, mapped_size);
2237
2238 crc = btrfs_csum_data(p, crc, l);
2239 kunmap_atomic(mapped_buffer);
2240 len -= l;
2241 if (len == 0)
2242 break;
2243 index++;
2244 BUG_ON(index >= sblock->page_count);
2245 BUG_ON(!sblock->pagev[index]->page);
2246 page = sblock->pagev[index]->page;
2247 mapped_buffer = kmap_atomic(page);
2248 mapped_size = PAGE_SIZE;
2249 p = mapped_buffer;
2250 }
2251
2252 btrfs_csum_final(crc, calculated_csum);
2253 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
2254 ++fail_cor;
2255
2256 if (fail_cor + fail_gen) {
2257 /*
2258 * if we find an error in a super block, we just report it.
2259 * They will get written with the next transaction commit
2260 * anyway
2261 */
2262 spin_lock(&sctx->stat_lock);
2263 ++sctx->stat.super_errors;
2264 spin_unlock(&sctx->stat_lock);
2265 if (fail_cor)
2266 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
2267 BTRFS_DEV_STAT_CORRUPTION_ERRS);
2268 else
2269 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
2270 BTRFS_DEV_STAT_GENERATION_ERRS);
2271 }
2272
2273 return fail_cor + fail_gen;
2274}
2275
2276static void scrub_block_get(struct scrub_block *sblock)
2277{
2278 refcount_inc(&sblock->refs);
2279}
2280
2281static void scrub_block_put(struct scrub_block *sblock)
2282{
2283 if (refcount_dec_and_test(&sblock->refs)) {
2284 int i;
2285
2286 if (sblock->sparity)
2287 scrub_parity_put(sblock->sparity);
2288
2289 for (i = 0; i < sblock->page_count; i++)
2290 scrub_page_put(sblock->pagev[i]);
2291 kfree(sblock);
2292 }
2293}
2294
2295static void scrub_page_get(struct scrub_page *spage)
2296{
2297 atomic_inc(&spage->refs);
2298}
2299
2300static void scrub_page_put(struct scrub_page *spage)
2301{
2302 if (atomic_dec_and_test(&spage->refs)) {
2303 if (spage->page)
2304 __free_page(spage->page);
2305 kfree(spage);
2306 }
2307}
2308
2309static void scrub_submit(struct scrub_ctx *sctx)
2310{
2311 struct scrub_bio *sbio;
2312
2313 if (sctx->curr == -1)
2314 return;
2315
2316 sbio = sctx->bios[sctx->curr];
2317 sctx->curr = -1;
2318 scrub_pending_bio_inc(sctx);
2319 btrfsic_submit_bio(sbio->bio);
2320}
2321
2322static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2323 struct scrub_page *spage)
2324{
2325 struct scrub_block *sblock = spage->sblock;
2326 struct scrub_bio *sbio;
2327 int ret;
2328
2329again:
2330 /*
2331 * grab a fresh bio or wait for one to become available
2332 */
2333 while (sctx->curr == -1) {
2334 spin_lock(&sctx->list_lock);
2335 sctx->curr = sctx->first_free;
2336 if (sctx->curr != -1) {
2337 sctx->first_free = sctx->bios[sctx->curr]->next_free;
2338 sctx->bios[sctx->curr]->next_free = -1;
2339 sctx->bios[sctx->curr]->page_count = 0;
2340 spin_unlock(&sctx->list_lock);
2341 } else {
2342 spin_unlock(&sctx->list_lock);
2343 wait_event(sctx->list_wait, sctx->first_free != -1);
2344 }
2345 }
2346 sbio = sctx->bios[sctx->curr];
2347 if (sbio->page_count == 0) {
2348 struct bio *bio;
2349
2350 sbio->physical = spage->physical;
2351 sbio->logical = spage->logical;
2352 sbio->dev = spage->dev;
2353 bio = sbio->bio;
2354 if (!bio) {
2355 bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
2356 sbio->bio = bio;
2357 }
2358
2359 bio->bi_private = sbio;
2360 bio->bi_end_io = scrub_bio_end_io;
2361 bio_set_dev(bio, sbio->dev->bdev);
2362 bio->bi_iter.bi_sector = sbio->physical >> 9;
2363 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2364 sbio->status = 0;
2365 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2366 spage->physical ||
2367 sbio->logical + sbio->page_count * PAGE_SIZE !=
2368 spage->logical ||
2369 sbio->dev != spage->dev) {
2370 scrub_submit(sctx);
2371 goto again;
2372 }
2373
2374 sbio->pagev[sbio->page_count] = spage;
2375 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2376 if (ret != PAGE_SIZE) {
2377 if (sbio->page_count < 1) {
2378 bio_put(sbio->bio);
2379 sbio->bio = NULL;
2380 return -EIO;
2381 }
2382 scrub_submit(sctx);
2383 goto again;
2384 }
2385
2386 scrub_block_get(sblock); /* one for the page added to the bio */
2387 atomic_inc(&sblock->outstanding_pages);
2388 sbio->page_count++;
2389 if (sbio->page_count == sctx->pages_per_rd_bio)
2390 scrub_submit(sctx);
2391
2392 return 0;
2393}
2394
2395static void scrub_missing_raid56_end_io(struct bio *bio)
2396{
2397 struct scrub_block *sblock = bio->bi_private;
2398 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2399
2400 if (bio->bi_status)
2401 sblock->no_io_error_seen = 0;
2402
2403 bio_put(bio);
2404
2405 btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2406}
2407
2408static void scrub_missing_raid56_worker(struct btrfs_work *work)
2409{
2410 struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2411 struct scrub_ctx *sctx = sblock->sctx;
2412 struct btrfs_fs_info *fs_info = sctx->fs_info;
2413 u64 logical;
2414 struct btrfs_device *dev;
2415
2416 logical = sblock->pagev[0]->logical;
2417 dev = sblock->pagev[0]->dev;
2418
2419 if (sblock->no_io_error_seen)
2420 scrub_recheck_block_checksum(sblock);
2421
2422 if (!sblock->no_io_error_seen) {
2423 spin_lock(&sctx->stat_lock);
2424 sctx->stat.read_errors++;
2425 spin_unlock(&sctx->stat_lock);
2426 btrfs_err_rl_in_rcu(fs_info,
2427 "IO error rebuilding logical %llu for dev %s",
2428 logical, rcu_str_deref(dev->name));
2429 } else if (sblock->header_error || sblock->checksum_error) {
2430 spin_lock(&sctx->stat_lock);
2431 sctx->stat.uncorrectable_errors++;
2432 spin_unlock(&sctx->stat_lock);
2433 btrfs_err_rl_in_rcu(fs_info,
2434 "failed to rebuild valid logical %llu for dev %s",
2435 logical, rcu_str_deref(dev->name));
2436 } else {
2437 scrub_write_block_to_dev_replace(sblock);
2438 }
2439
2440 scrub_block_put(sblock);
2441
2442 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2443 mutex_lock(&sctx->wr_lock);
2444 scrub_wr_submit(sctx);
2445 mutex_unlock(&sctx->wr_lock);
2446 }
2447
2448 scrub_pending_bio_dec(sctx);
2449}
2450
2451static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2452{
2453 struct scrub_ctx *sctx = sblock->sctx;
2454 struct btrfs_fs_info *fs_info = sctx->fs_info;
2455 u64 length = sblock->page_count * PAGE_SIZE;
2456 u64 logical = sblock->pagev[0]->logical;
2457 struct btrfs_bio *bbio = NULL;
2458 struct bio *bio;
2459 struct btrfs_raid_bio *rbio;
2460 int ret;
2461 int i;
2462
2463 btrfs_bio_counter_inc_blocked(fs_info);
2464 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2465 &length, &bbio);
2466 if (ret || !bbio || !bbio->raid_map)
2467 goto bbio_out;
2468
2469 if (WARN_ON(!sctx->is_dev_replace ||
2470 !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2471 /*
2472 * We shouldn't be scrubbing a missing device. Even for dev
2473 * replace, we should only get here for RAID 5/6. We either
2474 * managed to mount something with no mirrors remaining or
2475 * there's a bug in scrub_remap_extent()/btrfs_map_block().
2476 */
2477 goto bbio_out;
2478 }
2479
2480 bio = btrfs_io_bio_alloc(0);
2481 bio->bi_iter.bi_sector = logical >> 9;
2482 bio->bi_private = sblock;
2483 bio->bi_end_io = scrub_missing_raid56_end_io;
2484
2485 rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2486 if (!rbio)
2487 goto rbio_out;
2488
2489 for (i = 0; i < sblock->page_count; i++) {
2490 struct scrub_page *spage = sblock->pagev[i];
2491
2492 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2493 }
2494
2495 btrfs_init_work(&sblock->work, btrfs_scrub_helper,
2496 scrub_missing_raid56_worker, NULL, NULL);
2497 scrub_block_get(sblock);
2498 scrub_pending_bio_inc(sctx);
2499 raid56_submit_missing_rbio(rbio);
2500 return;
2501
2502rbio_out:
2503 bio_put(bio);
2504bbio_out:
2505 btrfs_bio_counter_dec(fs_info);
2506 btrfs_put_bbio(bbio);
2507 spin_lock(&sctx->stat_lock);
2508 sctx->stat.malloc_errors++;
2509 spin_unlock(&sctx->stat_lock);
2510}
2511
2512static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2513 u64 physical, struct btrfs_device *dev, u64 flags,
2514 u64 gen, int mirror_num, u8 *csum, int force,
2515 u64 physical_for_dev_replace)
2516{
2517 struct scrub_block *sblock;
2518 int index;
2519
2520 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2521 if (!sblock) {
2522 spin_lock(&sctx->stat_lock);
2523 sctx->stat.malloc_errors++;
2524 spin_unlock(&sctx->stat_lock);
2525 return -ENOMEM;
2526 }
2527
2528 /* one ref inside this function, plus one for each page added to
2529 * a bio later on */
2530 refcount_set(&sblock->refs, 1);
2531 sblock->sctx = sctx;
2532 sblock->no_io_error_seen = 1;
2533
2534 for (index = 0; len > 0; index++) {
2535 struct scrub_page *spage;
2536 u64 l = min_t(u64, len, PAGE_SIZE);
2537
2538 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2539 if (!spage) {
2540leave_nomem:
2541 spin_lock(&sctx->stat_lock);
2542 sctx->stat.malloc_errors++;
2543 spin_unlock(&sctx->stat_lock);
2544 scrub_block_put(sblock);
2545 return -ENOMEM;
2546 }
2547 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2548 scrub_page_get(spage);
2549 sblock->pagev[index] = spage;
2550 spage->sblock = sblock;
2551 spage->dev = dev;
2552 spage->flags = flags;
2553 spage->generation = gen;
2554 spage->logical = logical;
2555 spage->physical = physical;
2556 spage->physical_for_dev_replace = physical_for_dev_replace;
2557 spage->mirror_num = mirror_num;
2558 if (csum) {
2559 spage->have_csum = 1;
2560 memcpy(spage->csum, csum, sctx->csum_size);
2561 } else {
2562 spage->have_csum = 0;
2563 }
2564 sblock->page_count++;
2565 spage->page = alloc_page(GFP_KERNEL);
2566 if (!spage->page)
2567 goto leave_nomem;
2568 len -= l;
2569 logical += l;
2570 physical += l;
2571 physical_for_dev_replace += l;
2572 }
2573
2574 WARN_ON(sblock->page_count == 0);
2575 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2576 /*
2577 * This case should only be hit for RAID 5/6 device replace. See
2578 * the comment in scrub_missing_raid56_pages() for details.
2579 */
2580 scrub_missing_raid56_pages(sblock);
2581 } else {
2582 for (index = 0; index < sblock->page_count; index++) {
2583 struct scrub_page *spage = sblock->pagev[index];
2584 int ret;
2585
2586 ret = scrub_add_page_to_rd_bio(sctx, spage);
2587 if (ret) {
2588 scrub_block_put(sblock);
2589 return ret;
2590 }
2591 }
2592
2593 if (force)
2594 scrub_submit(sctx);
2595 }
2596
2597 /* last one frees, either here or in bio completion for last page */
2598 scrub_block_put(sblock);
2599 return 0;
2600}
2601
2602static void scrub_bio_end_io(struct bio *bio)
2603{
2604 struct scrub_bio *sbio = bio->bi_private;
2605 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2606
2607 sbio->status = bio->bi_status;
2608 sbio->bio = bio;
2609
2610 btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2611}
2612
2613static void scrub_bio_end_io_worker(struct btrfs_work *work)
2614{
2615 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2616 struct scrub_ctx *sctx = sbio->sctx;
2617 int i;
2618
2619 BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2620 if (sbio->status) {
2621 for (i = 0; i < sbio->page_count; i++) {
2622 struct scrub_page *spage = sbio->pagev[i];
2623
2624 spage->io_error = 1;
2625 spage->sblock->no_io_error_seen = 0;
2626 }
2627 }
2628
2629 /* now complete the scrub_block items that have all pages completed */
2630 for (i = 0; i < sbio->page_count; i++) {
2631 struct scrub_page *spage = sbio->pagev[i];
2632 struct scrub_block *sblock = spage->sblock;
2633
2634 if (atomic_dec_and_test(&sblock->outstanding_pages))
2635 scrub_block_complete(sblock);
2636 scrub_block_put(sblock);
2637 }
2638
2639 bio_put(sbio->bio);
2640 sbio->bio = NULL;
2641 spin_lock(&sctx->list_lock);
2642 sbio->next_free = sctx->first_free;
2643 sctx->first_free = sbio->index;
2644 spin_unlock(&sctx->list_lock);
2645
2646 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2647 mutex_lock(&sctx->wr_lock);
2648 scrub_wr_submit(sctx);
2649 mutex_unlock(&sctx->wr_lock);
2650 }
2651
2652 scrub_pending_bio_dec(sctx);
2653}
2654
2655static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2656 unsigned long *bitmap,
2657 u64 start, u64 len)
2658{
2659 u64 offset;
2660 u64 nsectors64;
2661 u32 nsectors;
2662 int sectorsize = sparity->sctx->fs_info->sectorsize;
2663
2664 if (len >= sparity->stripe_len) {
2665 bitmap_set(bitmap, 0, sparity->nsectors);
2666 return;
2667 }
2668
2669 start -= sparity->logic_start;
2670 start = div64_u64_rem(start, sparity->stripe_len, &offset);
2671 offset = div_u64(offset, sectorsize);
2672 nsectors64 = div_u64(len, sectorsize);
2673
2674 ASSERT(nsectors64 < UINT_MAX);
2675 nsectors = (u32)nsectors64;
2676
2677 if (offset + nsectors <= sparity->nsectors) {
2678 bitmap_set(bitmap, offset, nsectors);
2679 return;
2680 }
2681
2682 bitmap_set(bitmap, offset, sparity->nsectors - offset);
2683 bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2684}
2685
2686static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2687 u64 start, u64 len)
2688{
2689 __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2690}
2691
2692static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2693 u64 start, u64 len)
2694{
2695 __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2696}
2697
2698static void scrub_block_complete(struct scrub_block *sblock)
2699{
2700 int corrupted = 0;
2701
2702 if (!sblock->no_io_error_seen) {
2703 corrupted = 1;
2704 scrub_handle_errored_block(sblock);
2705 } else {
2706 /*
2707 * if has checksum error, write via repair mechanism in
2708 * dev replace case, otherwise write here in dev replace
2709 * case.
2710 */
2711 corrupted = scrub_checksum(sblock);
2712 if (!corrupted && sblock->sctx->is_dev_replace)
2713 scrub_write_block_to_dev_replace(sblock);
2714 }
2715
2716 if (sblock->sparity && corrupted && !sblock->data_corrected) {
2717 u64 start = sblock->pagev[0]->logical;
2718 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2719 PAGE_SIZE;
2720
2721 scrub_parity_mark_sectors_error(sblock->sparity,
2722 start, end - start);
2723 }
2724}
2725
2726static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2727{
2728 struct btrfs_ordered_sum *sum = NULL;
2729 unsigned long index;
2730 unsigned long num_sectors;
2731
2732 while (!list_empty(&sctx->csum_list)) {
2733 sum = list_first_entry(&sctx->csum_list,
2734 struct btrfs_ordered_sum, list);
2735 if (sum->bytenr > logical)
2736 return 0;
2737 if (sum->bytenr + sum->len > logical)
2738 break;
2739
2740 ++sctx->stat.csum_discards;
2741 list_del(&sum->list);
2742 kfree(sum);
2743 sum = NULL;
2744 }
2745 if (!sum)
2746 return 0;
2747
2748 index = div_u64(logical - sum->bytenr, sctx->fs_info->sectorsize);
2749 ASSERT(index < UINT_MAX);
2750
2751 num_sectors = sum->len / sctx->fs_info->sectorsize;
2752 memcpy(csum, sum->sums + index, sctx->csum_size);
2753 if (index == num_sectors - 1) {
2754 list_del(&sum->list);
2755 kfree(sum);
2756 }
2757 return 1;
2758}
2759
2760/* scrub extent tries to collect up to 64 kB for each bio */
2761static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2762 u64 logical, u64 len,
2763 u64 physical, struct btrfs_device *dev, u64 flags,
2764 u64 gen, int mirror_num, u64 physical_for_dev_replace)
2765{
2766 int ret;
2767 u8 csum[BTRFS_CSUM_SIZE];
2768 u32 blocksize;
2769
2770 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2771 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2772 blocksize = map->stripe_len;
2773 else
2774 blocksize = sctx->fs_info->sectorsize;
2775 spin_lock(&sctx->stat_lock);
2776 sctx->stat.data_extents_scrubbed++;
2777 sctx->stat.data_bytes_scrubbed += len;
2778 spin_unlock(&sctx->stat_lock);
2779 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2780 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2781 blocksize = map->stripe_len;
2782 else
2783 blocksize = sctx->fs_info->nodesize;
2784 spin_lock(&sctx->stat_lock);
2785 sctx->stat.tree_extents_scrubbed++;
2786 sctx->stat.tree_bytes_scrubbed += len;
2787 spin_unlock(&sctx->stat_lock);
2788 } else {
2789 blocksize = sctx->fs_info->sectorsize;
2790 WARN_ON(1);
2791 }
2792
2793 while (len) {
2794 u64 l = min_t(u64, len, blocksize);
2795 int have_csum = 0;
2796
2797 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2798 /* push csums to sbio */
2799 have_csum = scrub_find_csum(sctx, logical, csum);
2800 if (have_csum == 0)
2801 ++sctx->stat.no_csum;
2802 if (sctx->is_dev_replace && !have_csum) {
2803 ret = copy_nocow_pages(sctx, logical, l,
2804 mirror_num,
2805 physical_for_dev_replace);
2806 goto behind_scrub_pages;
2807 }
2808 }
2809 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2810 mirror_num, have_csum ? csum : NULL, 0,
2811 physical_for_dev_replace);
2812behind_scrub_pages:
2813 if (ret)
2814 return ret;
2815 len -= l;
2816 logical += l;
2817 physical += l;
2818 physical_for_dev_replace += l;
2819 }
2820 return 0;
2821}
2822
2823static int scrub_pages_for_parity(struct scrub_parity *sparity,
2824 u64 logical, u64 len,
2825 u64 physical, struct btrfs_device *dev,
2826 u64 flags, u64 gen, int mirror_num, u8 *csum)
2827{
2828 struct scrub_ctx *sctx = sparity->sctx;
2829 struct scrub_block *sblock;
2830 int index;
2831
2832 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2833 if (!sblock) {
2834 spin_lock(&sctx->stat_lock);
2835 sctx->stat.malloc_errors++;
2836 spin_unlock(&sctx->stat_lock);
2837 return -ENOMEM;
2838 }
2839
2840 /* one ref inside this function, plus one for each page added to
2841 * a bio later on */
2842 refcount_set(&sblock->refs, 1);
2843 sblock->sctx = sctx;
2844 sblock->no_io_error_seen = 1;
2845 sblock->sparity = sparity;
2846 scrub_parity_get(sparity);
2847
2848 for (index = 0; len > 0; index++) {
2849 struct scrub_page *spage;
2850 u64 l = min_t(u64, len, PAGE_SIZE);
2851
2852 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2853 if (!spage) {
2854leave_nomem:
2855 spin_lock(&sctx->stat_lock);
2856 sctx->stat.malloc_errors++;
2857 spin_unlock(&sctx->stat_lock);
2858 scrub_block_put(sblock);
2859 return -ENOMEM;
2860 }
2861 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2862 /* For scrub block */
2863 scrub_page_get(spage);
2864 sblock->pagev[index] = spage;
2865 /* For scrub parity */
2866 scrub_page_get(spage);
2867 list_add_tail(&spage->list, &sparity->spages);
2868 spage->sblock = sblock;
2869 spage->dev = dev;
2870 spage->flags = flags;
2871 spage->generation = gen;
2872 spage->logical = logical;
2873 spage->physical = physical;
2874 spage->mirror_num = mirror_num;
2875 if (csum) {
2876 spage->have_csum = 1;
2877 memcpy(spage->csum, csum, sctx->csum_size);
2878 } else {
2879 spage->have_csum = 0;
2880 }
2881 sblock->page_count++;
2882 spage->page = alloc_page(GFP_KERNEL);
2883 if (!spage->page)
2884 goto leave_nomem;
2885 len -= l;
2886 logical += l;
2887 physical += l;
2888 }
2889
2890 WARN_ON(sblock->page_count == 0);
2891 for (index = 0; index < sblock->page_count; index++) {
2892 struct scrub_page *spage = sblock->pagev[index];
2893 int ret;
2894
2895 ret = scrub_add_page_to_rd_bio(sctx, spage);
2896 if (ret) {
2897 scrub_block_put(sblock);
2898 return ret;
2899 }
2900 }
2901
2902 /* last one frees, either here or in bio completion for last page */
2903 scrub_block_put(sblock);
2904 return 0;
2905}
2906
2907static int scrub_extent_for_parity(struct scrub_parity *sparity,
2908 u64 logical, u64 len,
2909 u64 physical, struct btrfs_device *dev,
2910 u64 flags, u64 gen, int mirror_num)
2911{
2912 struct scrub_ctx *sctx = sparity->sctx;
2913 int ret;
2914 u8 csum[BTRFS_CSUM_SIZE];
2915 u32 blocksize;
2916
2917 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2918 scrub_parity_mark_sectors_error(sparity, logical, len);
2919 return 0;
2920 }
2921
2922 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2923 blocksize = sparity->stripe_len;
2924 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2925 blocksize = sparity->stripe_len;
2926 } else {
2927 blocksize = sctx->fs_info->sectorsize;
2928 WARN_ON(1);
2929 }
2930
2931 while (len) {
2932 u64 l = min_t(u64, len, blocksize);
2933 int have_csum = 0;
2934
2935 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2936 /* push csums to sbio */
2937 have_csum = scrub_find_csum(sctx, logical, csum);
2938 if (have_csum == 0)
2939 goto skip;
2940 }
2941 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2942 flags, gen, mirror_num,
2943 have_csum ? csum : NULL);
2944 if (ret)
2945 return ret;
2946skip:
2947 len -= l;
2948 logical += l;
2949 physical += l;
2950 }
2951 return 0;
2952}
2953
2954/*
2955 * Given a physical address, this will calculate it's
2956 * logical offset. if this is a parity stripe, it will return
2957 * the most left data stripe's logical offset.
2958 *
2959 * return 0 if it is a data stripe, 1 means parity stripe.
2960 */
2961static int get_raid56_logic_offset(u64 physical, int num,
2962 struct map_lookup *map, u64 *offset,
2963 u64 *stripe_start)
2964{
2965 int i;
2966 int j = 0;
2967 u64 stripe_nr;
2968 u64 last_offset;
2969 u32 stripe_index;
2970 u32 rot;
2971
2972 last_offset = (physical - map->stripes[num].physical) *
2973 nr_data_stripes(map);
2974 if (stripe_start)
2975 *stripe_start = last_offset;
2976
2977 *offset = last_offset;
2978 for (i = 0; i < nr_data_stripes(map); i++) {
2979 *offset = last_offset + i * map->stripe_len;
2980
2981 stripe_nr = div64_u64(*offset, map->stripe_len);
2982 stripe_nr = div_u64(stripe_nr, nr_data_stripes(map));
2983
2984 /* Work out the disk rotation on this stripe-set */
2985 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2986 /* calculate which stripe this data locates */
2987 rot += i;
2988 stripe_index = rot % map->num_stripes;
2989 if (stripe_index == num)
2990 return 0;
2991 if (stripe_index < num)
2992 j++;
2993 }
2994 *offset = last_offset + j * map->stripe_len;
2995 return 1;
2996}
2997
2998static void scrub_free_parity(struct scrub_parity *sparity)
2999{
3000 struct scrub_ctx *sctx = sparity->sctx;
3001 struct scrub_page *curr, *next;
3002 int nbits;
3003
3004 nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
3005 if (nbits) {
3006 spin_lock(&sctx->stat_lock);
3007 sctx->stat.read_errors += nbits;
3008 sctx->stat.uncorrectable_errors += nbits;
3009 spin_unlock(&sctx->stat_lock);
3010 }
3011
3012 list_for_each_entry_safe(curr, next, &sparity->spages, list) {
3013 list_del_init(&curr->list);
3014 scrub_page_put(curr);
3015 }
3016
3017 kfree(sparity);
3018}
3019
3020static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
3021{
3022 struct scrub_parity *sparity = container_of(work, struct scrub_parity,
3023 work);
3024 struct scrub_ctx *sctx = sparity->sctx;
3025
3026 scrub_free_parity(sparity);
3027 scrub_pending_bio_dec(sctx);
3028}
3029
3030static void scrub_parity_bio_endio(struct bio *bio)
3031{
3032 struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
3033 struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
3034
3035 if (bio->bi_status)
3036 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
3037 sparity->nsectors);
3038
3039 bio_put(bio);
3040
3041 btrfs_init_work(&sparity->work, btrfs_scrubparity_helper,
3042 scrub_parity_bio_endio_worker, NULL, NULL);
3043 btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
3044}
3045
3046static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
3047{
3048 struct scrub_ctx *sctx = sparity->sctx;
3049 struct btrfs_fs_info *fs_info = sctx->fs_info;
3050 struct bio *bio;
3051 struct btrfs_raid_bio *rbio;
3052 struct btrfs_bio *bbio = NULL;
3053 u64 length;
3054 int ret;
3055
3056 if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
3057 sparity->nsectors))
3058 goto out;
3059
3060 length = sparity->logic_end - sparity->logic_start;
3061
3062 btrfs_bio_counter_inc_blocked(fs_info);
3063 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
3064 &length, &bbio);
3065 if (ret || !bbio || !bbio->raid_map)
3066 goto bbio_out;
3067
3068 bio = btrfs_io_bio_alloc(0);
3069 bio->bi_iter.bi_sector = sparity->logic_start >> 9;
3070 bio->bi_private = sparity;
3071 bio->bi_end_io = scrub_parity_bio_endio;
3072
3073 rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
3074 length, sparity->scrub_dev,
3075 sparity->dbitmap,
3076 sparity->nsectors);
3077 if (!rbio)
3078 goto rbio_out;
3079
3080 scrub_pending_bio_inc(sctx);
3081 raid56_parity_submit_scrub_rbio(rbio);
3082 return;
3083
3084rbio_out:
3085 bio_put(bio);
3086bbio_out:
3087 btrfs_bio_counter_dec(fs_info);
3088 btrfs_put_bbio(bbio);
3089 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
3090 sparity->nsectors);
3091 spin_lock(&sctx->stat_lock);
3092 sctx->stat.malloc_errors++;
3093 spin_unlock(&sctx->stat_lock);
3094out:
3095 scrub_free_parity(sparity);
3096}
3097
3098static inline int scrub_calc_parity_bitmap_len(int nsectors)
3099{
3100 return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
3101}
3102
3103static void scrub_parity_get(struct scrub_parity *sparity)
3104{
3105 refcount_inc(&sparity->refs);
3106}
3107
3108static void scrub_parity_put(struct scrub_parity *sparity)
3109{
3110 if (!refcount_dec_and_test(&sparity->refs))
3111 return;
3112
3113 scrub_parity_check_and_repair(sparity);
3114}
3115
3116static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
3117 struct map_lookup *map,
3118 struct btrfs_device *sdev,
3119 struct btrfs_path *path,
3120 u64 logic_start,
3121 u64 logic_end)
3122{
3123 struct btrfs_fs_info *fs_info = sctx->fs_info;
3124 struct btrfs_root *root = fs_info->extent_root;
3125 struct btrfs_root *csum_root = fs_info->csum_root;
3126 struct btrfs_extent_item *extent;
3127 struct btrfs_bio *bbio = NULL;
3128 u64 flags;
3129 int ret;
3130 int slot;
3131 struct extent_buffer *l;
3132 struct btrfs_key key;
3133 u64 generation;
3134 u64 extent_logical;
3135 u64 extent_physical;
3136 u64 extent_len;
3137 u64 mapped_length;
3138 struct btrfs_device *extent_dev;
3139 struct scrub_parity *sparity;
3140 int nsectors;
3141 int bitmap_len;
3142 int extent_mirror_num;
3143 int stop_loop = 0;
3144
3145 nsectors = div_u64(map->stripe_len, fs_info->sectorsize);
3146 bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
3147 sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
3148 GFP_NOFS);
3149 if (!sparity) {
3150 spin_lock(&sctx->stat_lock);
3151 sctx->stat.malloc_errors++;
3152 spin_unlock(&sctx->stat_lock);
3153 return -ENOMEM;
3154 }
3155
3156 sparity->stripe_len = map->stripe_len;
3157 sparity->nsectors = nsectors;
3158 sparity->sctx = sctx;
3159 sparity->scrub_dev = sdev;
3160 sparity->logic_start = logic_start;
3161 sparity->logic_end = logic_end;
3162 refcount_set(&sparity->refs, 1);
3163 INIT_LIST_HEAD(&sparity->spages);
3164 sparity->dbitmap = sparity->bitmap;
3165 sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
3166
3167 ret = 0;
3168 while (logic_start < logic_end) {
3169 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3170 key.type = BTRFS_METADATA_ITEM_KEY;
3171 else
3172 key.type = BTRFS_EXTENT_ITEM_KEY;
3173 key.objectid = logic_start;
3174 key.offset = (u64)-1;
3175
3176 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3177 if (ret < 0)
3178 goto out;
3179
3180 if (ret > 0) {
3181 ret = btrfs_previous_extent_item(root, path, 0);
3182 if (ret < 0)
3183 goto out;
3184 if (ret > 0) {
3185 btrfs_release_path(path);
3186 ret = btrfs_search_slot(NULL, root, &key,
3187 path, 0, 0);
3188 if (ret < 0)
3189 goto out;
3190 }
3191 }
3192
3193 stop_loop = 0;
3194 while (1) {
3195 u64 bytes;
3196
3197 l = path->nodes[0];
3198 slot = path->slots[0];
3199 if (slot >= btrfs_header_nritems(l)) {
3200 ret = btrfs_next_leaf(root, path);
3201 if (ret == 0)
3202 continue;
3203 if (ret < 0)
3204 goto out;
3205
3206 stop_loop = 1;
3207 break;
3208 }
3209 btrfs_item_key_to_cpu(l, &key, slot);
3210
3211 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3212 key.type != BTRFS_METADATA_ITEM_KEY)
3213 goto next;
3214
3215 if (key.type == BTRFS_METADATA_ITEM_KEY)
3216 bytes = fs_info->nodesize;
3217 else
3218 bytes = key.offset;
3219
3220 if (key.objectid + bytes <= logic_start)
3221 goto next;
3222
3223 if (key.objectid >= logic_end) {
3224 stop_loop = 1;
3225 break;
3226 }
3227
3228 while (key.objectid >= logic_start + map->stripe_len)
3229 logic_start += map->stripe_len;
3230
3231 extent = btrfs_item_ptr(l, slot,
3232 struct btrfs_extent_item);
3233 flags = btrfs_extent_flags(l, extent);
3234 generation = btrfs_extent_generation(l, extent);
3235
3236 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3237 (key.objectid < logic_start ||
3238 key.objectid + bytes >
3239 logic_start + map->stripe_len)) {
3240 btrfs_err(fs_info,
3241 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3242 key.objectid, logic_start);
3243 spin_lock(&sctx->stat_lock);
3244 sctx->stat.uncorrectable_errors++;
3245 spin_unlock(&sctx->stat_lock);
3246 goto next;
3247 }
3248again:
3249 extent_logical = key.objectid;
3250 extent_len = bytes;
3251
3252 if (extent_logical < logic_start) {
3253 extent_len -= logic_start - extent_logical;
3254 extent_logical = logic_start;
3255 }
3256
3257 if (extent_logical + extent_len >
3258 logic_start + map->stripe_len)
3259 extent_len = logic_start + map->stripe_len -
3260 extent_logical;
3261
3262 scrub_parity_mark_sectors_data(sparity, extent_logical,
3263 extent_len);
3264
3265 mapped_length = extent_len;
3266 bbio = NULL;
3267 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
3268 extent_logical, &mapped_length, &bbio,
3269 0);
3270 if (!ret) {
3271 if (!bbio || mapped_length < extent_len)
3272 ret = -EIO;
3273 }
3274 if (ret) {
3275 btrfs_put_bbio(bbio);
3276 goto out;
3277 }
3278 extent_physical = bbio->stripes[0].physical;
3279 extent_mirror_num = bbio->mirror_num;
3280 extent_dev = bbio->stripes[0].dev;
3281 btrfs_put_bbio(bbio);
3282
3283 ret = btrfs_lookup_csums_range(csum_root,
3284 extent_logical,
3285 extent_logical + extent_len - 1,
3286 &sctx->csum_list, 1);
3287 if (ret)
3288 goto out;
3289
3290 ret = scrub_extent_for_parity(sparity, extent_logical,
3291 extent_len,
3292 extent_physical,
3293 extent_dev, flags,
3294 generation,
3295 extent_mirror_num);
3296
3297 scrub_free_csums(sctx);
3298
3299 if (ret)
3300 goto out;
3301
3302 if (extent_logical + extent_len <
3303 key.objectid + bytes) {
3304 logic_start += map->stripe_len;
3305
3306 if (logic_start >= logic_end) {
3307 stop_loop = 1;
3308 break;
3309 }
3310
3311 if (logic_start < key.objectid + bytes) {
3312 cond_resched();
3313 goto again;
3314 }
3315 }
3316next:
3317 path->slots[0]++;
3318 }
3319
3320 btrfs_release_path(path);
3321
3322 if (stop_loop)
3323 break;
3324
3325 logic_start += map->stripe_len;
3326 }
3327out:
3328 if (ret < 0)
3329 scrub_parity_mark_sectors_error(sparity, logic_start,
3330 logic_end - logic_start);
3331 scrub_parity_put(sparity);
3332 scrub_submit(sctx);
3333 mutex_lock(&sctx->wr_lock);
3334 scrub_wr_submit(sctx);
3335 mutex_unlock(&sctx->wr_lock);
3336
3337 btrfs_release_path(path);
3338 return ret < 0 ? ret : 0;
3339}
3340
3341static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3342 struct map_lookup *map,
3343 struct btrfs_device *scrub_dev,
3344 int num, u64 base, u64 length,
3345 int is_dev_replace)
3346{
3347 struct btrfs_path *path, *ppath;
3348 struct btrfs_fs_info *fs_info = sctx->fs_info;
3349 struct btrfs_root *root = fs_info->extent_root;
3350 struct btrfs_root *csum_root = fs_info->csum_root;
3351 struct btrfs_extent_item *extent;
3352 struct blk_plug plug;
3353 u64 flags;
3354 int ret;
3355 int slot;
3356 u64 nstripes;
3357 struct extent_buffer *l;
3358 u64 physical;
3359 u64 logical;
3360 u64 logic_end;
3361 u64 physical_end;
3362 u64 generation;
3363 int mirror_num;
3364 struct reada_control *reada1;
3365 struct reada_control *reada2;
3366 struct btrfs_key key;
3367 struct btrfs_key key_end;
3368 u64 increment = map->stripe_len;
3369 u64 offset;
3370 u64 extent_logical;
3371 u64 extent_physical;
3372 u64 extent_len;
3373 u64 stripe_logical;
3374 u64 stripe_end;
3375 struct btrfs_device *extent_dev;
3376 int extent_mirror_num;
3377 int stop_loop = 0;
3378
3379 physical = map->stripes[num].physical;
3380 offset = 0;
3381 nstripes = div64_u64(length, map->stripe_len);
3382 if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3383 offset = map->stripe_len * num;
3384 increment = map->stripe_len * map->num_stripes;
3385 mirror_num = 1;
3386 } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3387 int factor = map->num_stripes / map->sub_stripes;
3388 offset = map->stripe_len * (num / map->sub_stripes);
3389 increment = map->stripe_len * factor;
3390 mirror_num = num % map->sub_stripes + 1;
3391 } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
3392 increment = map->stripe_len;
3393 mirror_num = num % map->num_stripes + 1;
3394 } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3395 increment = map->stripe_len;
3396 mirror_num = num % map->num_stripes + 1;
3397 } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3398 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3399 increment = map->stripe_len * nr_data_stripes(map);
3400 mirror_num = 1;
3401 } else {
3402 increment = map->stripe_len;
3403 mirror_num = 1;
3404 }
3405
3406 path = btrfs_alloc_path();
3407 if (!path)
3408 return -ENOMEM;
3409
3410 ppath = btrfs_alloc_path();
3411 if (!ppath) {
3412 btrfs_free_path(path);
3413 return -ENOMEM;
3414 }
3415
3416 /*
3417 * work on commit root. The related disk blocks are static as
3418 * long as COW is applied. This means, it is save to rewrite
3419 * them to repair disk errors without any race conditions
3420 */
3421 path->search_commit_root = 1;
3422 path->skip_locking = 1;
3423
3424 ppath->search_commit_root = 1;
3425 ppath->skip_locking = 1;
3426 /*
3427 * trigger the readahead for extent tree csum tree and wait for
3428 * completion. During readahead, the scrub is officially paused
3429 * to not hold off transaction commits
3430 */
3431 logical = base + offset;
3432 physical_end = physical + nstripes * map->stripe_len;
3433 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3434 get_raid56_logic_offset(physical_end, num,
3435 map, &logic_end, NULL);
3436 logic_end += base;
3437 } else {
3438 logic_end = logical + increment * nstripes;
3439 }
3440 wait_event(sctx->list_wait,
3441 atomic_read(&sctx->bios_in_flight) == 0);
3442 scrub_blocked_if_needed(fs_info);
3443
3444 /* FIXME it might be better to start readahead at commit root */
3445 key.objectid = logical;
3446 key.type = BTRFS_EXTENT_ITEM_KEY;
3447 key.offset = (u64)0;
3448 key_end.objectid = logic_end;
3449 key_end.type = BTRFS_METADATA_ITEM_KEY;
3450 key_end.offset = (u64)-1;
3451 reada1 = btrfs_reada_add(root, &key, &key_end);
3452
3453 key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3454 key.type = BTRFS_EXTENT_CSUM_KEY;
3455 key.offset = logical;
3456 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3457 key_end.type = BTRFS_EXTENT_CSUM_KEY;
3458 key_end.offset = logic_end;
3459 reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3460
3461 if (!IS_ERR(reada1))
3462 btrfs_reada_wait(reada1);
3463 if (!IS_ERR(reada2))
3464 btrfs_reada_wait(reada2);
3465
3466
3467 /*
3468 * collect all data csums for the stripe to avoid seeking during
3469 * the scrub. This might currently (crc32) end up to be about 1MB
3470 */
3471 blk_start_plug(&plug);
3472
3473 /*
3474 * now find all extents for each stripe and scrub them
3475 */
3476 ret = 0;
3477 while (physical < physical_end) {
3478 /*
3479 * canceled?
3480 */
3481 if (atomic_read(&fs_info->scrub_cancel_req) ||
3482 atomic_read(&sctx->cancel_req)) {
3483 ret = -ECANCELED;
3484 goto out;
3485 }
3486 /*
3487 * check to see if we have to pause
3488 */
3489 if (atomic_read(&fs_info->scrub_pause_req)) {
3490 /* push queued extents */
3491 sctx->flush_all_writes = true;
3492 scrub_submit(sctx);
3493 mutex_lock(&sctx->wr_lock);
3494 scrub_wr_submit(sctx);
3495 mutex_unlock(&sctx->wr_lock);
3496 wait_event(sctx->list_wait,
3497 atomic_read(&sctx->bios_in_flight) == 0);
3498 sctx->flush_all_writes = false;
3499 scrub_blocked_if_needed(fs_info);
3500 }
3501
3502 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3503 ret = get_raid56_logic_offset(physical, num, map,
3504 &logical,
3505 &stripe_logical);
3506 logical += base;
3507 if (ret) {
3508 /* it is parity strip */
3509 stripe_logical += base;
3510 stripe_end = stripe_logical + increment;
3511 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3512 ppath, stripe_logical,
3513 stripe_end);
3514 if (ret)
3515 goto out;
3516 goto skip;
3517 }
3518 }
3519
3520 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3521 key.type = BTRFS_METADATA_ITEM_KEY;
3522 else
3523 key.type = BTRFS_EXTENT_ITEM_KEY;
3524 key.objectid = logical;
3525 key.offset = (u64)-1;
3526
3527 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3528 if (ret < 0)
3529 goto out;
3530
3531 if (ret > 0) {
3532 ret = btrfs_previous_extent_item(root, path, 0);
3533 if (ret < 0)
3534 goto out;
3535 if (ret > 0) {
3536 /* there's no smaller item, so stick with the
3537 * larger one */
3538 btrfs_release_path(path);
3539 ret = btrfs_search_slot(NULL, root, &key,
3540 path, 0, 0);
3541 if (ret < 0)
3542 goto out;
3543 }
3544 }
3545
3546 stop_loop = 0;
3547 while (1) {
3548 u64 bytes;
3549
3550 l = path->nodes[0];
3551 slot = path->slots[0];
3552 if (slot >= btrfs_header_nritems(l)) {
3553 ret = btrfs_next_leaf(root, path);
3554 if (ret == 0)
3555 continue;
3556 if (ret < 0)
3557 goto out;
3558
3559 stop_loop = 1;
3560 break;
3561 }
3562 btrfs_item_key_to_cpu(l, &key, slot);
3563
3564 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3565 key.type != BTRFS_METADATA_ITEM_KEY)
3566 goto next;
3567
3568 if (key.type == BTRFS_METADATA_ITEM_KEY)
3569 bytes = fs_info->nodesize;
3570 else
3571 bytes = key.offset;
3572
3573 if (key.objectid + bytes <= logical)
3574 goto next;
3575
3576 if (key.objectid >= logical + map->stripe_len) {
3577 /* out of this device extent */
3578 if (key.objectid >= logic_end)
3579 stop_loop = 1;
3580 break;
3581 }
3582
3583 extent = btrfs_item_ptr(l, slot,
3584 struct btrfs_extent_item);
3585 flags = btrfs_extent_flags(l, extent);
3586 generation = btrfs_extent_generation(l, extent);
3587
3588 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3589 (key.objectid < logical ||
3590 key.objectid + bytes >
3591 logical + map->stripe_len)) {
3592 btrfs_err(fs_info,
3593 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3594 key.objectid, logical);
3595 spin_lock(&sctx->stat_lock);
3596 sctx->stat.uncorrectable_errors++;
3597 spin_unlock(&sctx->stat_lock);
3598 goto next;
3599 }
3600
3601again:
3602 extent_logical = key.objectid;
3603 extent_len = bytes;
3604
3605 /*
3606 * trim extent to this stripe
3607 */
3608 if (extent_logical < logical) {
3609 extent_len -= logical - extent_logical;
3610 extent_logical = logical;
3611 }
3612 if (extent_logical + extent_len >
3613 logical + map->stripe_len) {
3614 extent_len = logical + map->stripe_len -
3615 extent_logical;
3616 }
3617
3618 extent_physical = extent_logical - logical + physical;
3619 extent_dev = scrub_dev;
3620 extent_mirror_num = mirror_num;
3621 if (is_dev_replace)
3622 scrub_remap_extent(fs_info, extent_logical,
3623 extent_len, &extent_physical,
3624 &extent_dev,
3625 &extent_mirror_num);
3626
3627 ret = btrfs_lookup_csums_range(csum_root,
3628 extent_logical,
3629 extent_logical +
3630 extent_len - 1,
3631 &sctx->csum_list, 1);
3632 if (ret)
3633 goto out;
3634
3635 ret = scrub_extent(sctx, map, extent_logical, extent_len,
3636 extent_physical, extent_dev, flags,
3637 generation, extent_mirror_num,
3638 extent_logical - logical + physical);
3639
3640 scrub_free_csums(sctx);
3641
3642 if (ret)
3643 goto out;
3644
3645 if (extent_logical + extent_len <
3646 key.objectid + bytes) {
3647 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3648 /*
3649 * loop until we find next data stripe
3650 * or we have finished all stripes.
3651 */
3652loop:
3653 physical += map->stripe_len;
3654 ret = get_raid56_logic_offset(physical,
3655 num, map, &logical,
3656 &stripe_logical);
3657 logical += base;
3658
3659 if (ret && physical < physical_end) {
3660 stripe_logical += base;
3661 stripe_end = stripe_logical +
3662 increment;
3663 ret = scrub_raid56_parity(sctx,
3664 map, scrub_dev, ppath,
3665 stripe_logical,
3666 stripe_end);
3667 if (ret)
3668 goto out;
3669 goto loop;
3670 }
3671 } else {
3672 physical += map->stripe_len;
3673 logical += increment;
3674 }
3675 if (logical < key.objectid + bytes) {
3676 cond_resched();
3677 goto again;
3678 }
3679
3680 if (physical >= physical_end) {
3681 stop_loop = 1;
3682 break;
3683 }
3684 }
3685next:
3686 path->slots[0]++;
3687 }
3688 btrfs_release_path(path);
3689skip:
3690 logical += increment;
3691 physical += map->stripe_len;
3692 spin_lock(&sctx->stat_lock);
3693 if (stop_loop)
3694 sctx->stat.last_physical = map->stripes[num].physical +
3695 length;
3696 else
3697 sctx->stat.last_physical = physical;
3698 spin_unlock(&sctx->stat_lock);
3699 if (stop_loop)
3700 break;
3701 }
3702out:
3703 /* push queued extents */
3704 scrub_submit(sctx);
3705 mutex_lock(&sctx->wr_lock);
3706 scrub_wr_submit(sctx);
3707 mutex_unlock(&sctx->wr_lock);
3708
3709 blk_finish_plug(&plug);
3710 btrfs_free_path(path);
3711 btrfs_free_path(ppath);
3712 return ret < 0 ? ret : 0;
3713}
3714
3715static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3716 struct btrfs_device *scrub_dev,
3717 u64 chunk_offset, u64 length,
3718 u64 dev_offset,
3719 struct btrfs_block_group_cache *cache,
3720 int is_dev_replace)
3721{
3722 struct btrfs_fs_info *fs_info = sctx->fs_info;
3723 struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
3724 struct map_lookup *map;
3725 struct extent_map *em;
3726 int i;
3727 int ret = 0;
3728
3729 read_lock(&map_tree->map_tree.lock);
3730 em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
3731 read_unlock(&map_tree->map_tree.lock);
3732
3733 if (!em) {
3734 /*
3735 * Might have been an unused block group deleted by the cleaner
3736 * kthread or relocation.
3737 */
3738 spin_lock(&cache->lock);
3739 if (!cache->removed)
3740 ret = -EINVAL;
3741 spin_unlock(&cache->lock);
3742
3743 return ret;
3744 }
3745
3746 map = em->map_lookup;
3747 if (em->start != chunk_offset)
3748 goto out;
3749
3750 if (em->len < length)
3751 goto out;
3752
3753 for (i = 0; i < map->num_stripes; ++i) {
3754 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3755 map->stripes[i].physical == dev_offset) {
3756 ret = scrub_stripe(sctx, map, scrub_dev, i,
3757 chunk_offset, length,
3758 is_dev_replace);
3759 if (ret)
3760 goto out;
3761 }
3762 }
3763out:
3764 free_extent_map(em);
3765
3766 return ret;
3767}
3768
3769static noinline_for_stack
3770int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3771 struct btrfs_device *scrub_dev, u64 start, u64 end,
3772 int is_dev_replace)
3773{
3774 struct btrfs_dev_extent *dev_extent = NULL;
3775 struct btrfs_path *path;
3776 struct btrfs_fs_info *fs_info = sctx->fs_info;
3777 struct btrfs_root *root = fs_info->dev_root;
3778 u64 length;
3779 u64 chunk_offset;
3780 int ret = 0;
3781 int ro_set;
3782 int slot;
3783 struct extent_buffer *l;
3784 struct btrfs_key key;
3785 struct btrfs_key found_key;
3786 struct btrfs_block_group_cache *cache;
3787 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3788
3789 path = btrfs_alloc_path();
3790 if (!path)
3791 return -ENOMEM;
3792
3793 path->reada = READA_FORWARD;
3794 path->search_commit_root = 1;
3795 path->skip_locking = 1;
3796
3797 key.objectid = scrub_dev->devid;
3798 key.offset = 0ull;
3799 key.type = BTRFS_DEV_EXTENT_KEY;
3800
3801 while (1) {
3802 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3803 if (ret < 0)
3804 break;
3805 if (ret > 0) {
3806 if (path->slots[0] >=
3807 btrfs_header_nritems(path->nodes[0])) {
3808 ret = btrfs_next_leaf(root, path);
3809 if (ret < 0)
3810 break;
3811 if (ret > 0) {
3812 ret = 0;
3813 break;
3814 }
3815 } else {
3816 ret = 0;
3817 }
3818 }
3819
3820 l = path->nodes[0];
3821 slot = path->slots[0];
3822
3823 btrfs_item_key_to_cpu(l, &found_key, slot);
3824
3825 if (found_key.objectid != scrub_dev->devid)
3826 break;
3827
3828 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3829 break;
3830
3831 if (found_key.offset >= end)
3832 break;
3833
3834 if (found_key.offset < key.offset)
3835 break;
3836
3837 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3838 length = btrfs_dev_extent_length(l, dev_extent);
3839
3840 if (found_key.offset + length <= start)
3841 goto skip;
3842
3843 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3844
3845 /*
3846 * get a reference on the corresponding block group to prevent
3847 * the chunk from going away while we scrub it
3848 */
3849 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3850
3851 /* some chunks are removed but not committed to disk yet,
3852 * continue scrubbing */
3853 if (!cache)
3854 goto skip;
3855
3856 /*
3857 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3858 * to avoid deadlock caused by:
3859 * btrfs_inc_block_group_ro()
3860 * -> btrfs_wait_for_commit()
3861 * -> btrfs_commit_transaction()
3862 * -> btrfs_scrub_pause()
3863 */
3864 scrub_pause_on(fs_info);
3865 ret = btrfs_inc_block_group_ro(fs_info, cache);
3866 if (!ret && is_dev_replace) {
3867 /*
3868 * If we are doing a device replace wait for any tasks
3869 * that started dellaloc right before we set the block
3870 * group to RO mode, as they might have just allocated
3871 * an extent from it or decided they could do a nocow
3872 * write. And if any such tasks did that, wait for their
3873 * ordered extents to complete and then commit the
3874 * current transaction, so that we can later see the new
3875 * extent items in the extent tree - the ordered extents
3876 * create delayed data references (for cow writes) when
3877 * they complete, which will be run and insert the
3878 * corresponding extent items into the extent tree when
3879 * we commit the transaction they used when running
3880 * inode.c:btrfs_finish_ordered_io(). We later use
3881 * the commit root of the extent tree to find extents
3882 * to copy from the srcdev into the tgtdev, and we don't
3883 * want to miss any new extents.
3884 */
3885 btrfs_wait_block_group_reservations(cache);
3886 btrfs_wait_nocow_writers(cache);
3887 ret = btrfs_wait_ordered_roots(fs_info, U64_MAX,
3888 cache->key.objectid,
3889 cache->key.offset);
3890 if (ret > 0) {
3891 struct btrfs_trans_handle *trans;
3892
3893 trans = btrfs_join_transaction(root);
3894 if (IS_ERR(trans))
3895 ret = PTR_ERR(trans);
3896 else
3897 ret = btrfs_commit_transaction(trans);
3898 if (ret) {
3899 scrub_pause_off(fs_info);
3900 btrfs_put_block_group(cache);
3901 break;
3902 }
3903 }
3904 }
3905 scrub_pause_off(fs_info);
3906
3907 if (ret == 0) {
3908 ro_set = 1;
3909 } else if (ret == -ENOSPC) {
3910 /*
3911 * btrfs_inc_block_group_ro return -ENOSPC when it
3912 * failed in creating new chunk for metadata.
3913 * It is not a problem for scrub/replace, because
3914 * metadata are always cowed, and our scrub paused
3915 * commit_transactions.
3916 */
3917 ro_set = 0;
3918 } else {
3919 btrfs_warn(fs_info,
3920 "failed setting block group ro: %d", ret);
3921 btrfs_put_block_group(cache);
3922 break;
3923 }
3924
3925 btrfs_dev_replace_write_lock(&fs_info->dev_replace);
3926 dev_replace->cursor_right = found_key.offset + length;
3927 dev_replace->cursor_left = found_key.offset;
3928 dev_replace->item_needs_writeback = 1;
3929 btrfs_dev_replace_write_unlock(&fs_info->dev_replace);
3930 ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3931 found_key.offset, cache, is_dev_replace);
3932
3933 /*
3934 * flush, submit all pending read and write bios, afterwards
3935 * wait for them.
3936 * Note that in the dev replace case, a read request causes
3937 * write requests that are submitted in the read completion
3938 * worker. Therefore in the current situation, it is required
3939 * that all write requests are flushed, so that all read and
3940 * write requests are really completed when bios_in_flight
3941 * changes to 0.
3942 */
3943 sctx->flush_all_writes = true;
3944 scrub_submit(sctx);
3945 mutex_lock(&sctx->wr_lock);
3946 scrub_wr_submit(sctx);
3947 mutex_unlock(&sctx->wr_lock);
3948
3949 wait_event(sctx->list_wait,
3950 atomic_read(&sctx->bios_in_flight) == 0);
3951
3952 scrub_pause_on(fs_info);
3953
3954 /*
3955 * must be called before we decrease @scrub_paused.
3956 * make sure we don't block transaction commit while
3957 * we are waiting pending workers finished.
3958 */
3959 wait_event(sctx->list_wait,
3960 atomic_read(&sctx->workers_pending) == 0);
3961 sctx->flush_all_writes = false;
3962
3963 scrub_pause_off(fs_info);
3964
3965 btrfs_dev_replace_write_lock(&fs_info->dev_replace);
3966 dev_replace->cursor_left = dev_replace->cursor_right;
3967 dev_replace->item_needs_writeback = 1;
3968 btrfs_dev_replace_write_unlock(&fs_info->dev_replace);
3969
3970 if (ro_set)
3971 btrfs_dec_block_group_ro(cache);
3972
3973 /*
3974 * We might have prevented the cleaner kthread from deleting
3975 * this block group if it was already unused because we raced
3976 * and set it to RO mode first. So add it back to the unused
3977 * list, otherwise it might not ever be deleted unless a manual
3978 * balance is triggered or it becomes used and unused again.
3979 */
3980 spin_lock(&cache->lock);
3981 if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3982 btrfs_block_group_used(&cache->item) == 0) {
3983 spin_unlock(&cache->lock);
3984 spin_lock(&fs_info->unused_bgs_lock);
3985 if (list_empty(&cache->bg_list)) {
3986 btrfs_get_block_group(cache);
3987 list_add_tail(&cache->bg_list,
3988 &fs_info->unused_bgs);
3989 }
3990 spin_unlock(&fs_info->unused_bgs_lock);
3991 } else {
3992 spin_unlock(&cache->lock);
3993 }
3994
3995 btrfs_put_block_group(cache);
3996 if (ret)
3997 break;
3998 if (is_dev_replace &&
3999 atomic64_read(&dev_replace->num_write_errors) > 0) {
4000 ret = -EIO;
4001 break;
4002 }
4003 if (sctx->stat.malloc_errors > 0) {
4004 ret = -ENOMEM;
4005 break;
4006 }
4007skip:
4008 key.offset = found_key.offset + length;
4009 btrfs_release_path(path);
4010 }
4011
4012 btrfs_free_path(path);
4013
4014 return ret;
4015}
4016
4017static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
4018 struct btrfs_device *scrub_dev)
4019{
4020 int i;
4021 u64 bytenr;
4022 u64 gen;
4023 int ret;
4024 struct btrfs_fs_info *fs_info = sctx->fs_info;
4025
4026 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
4027 return -EIO;
4028
4029 /* Seed devices of a new filesystem has their own generation. */
4030 if (scrub_dev->fs_devices != fs_info->fs_devices)
4031 gen = scrub_dev->generation;
4032 else
4033 gen = fs_info->last_trans_committed;
4034
4035 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
4036 bytenr = btrfs_sb_offset(i);
4037 if (bytenr + BTRFS_SUPER_INFO_SIZE >
4038 scrub_dev->commit_total_bytes)
4039 break;
4040
4041 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
4042 scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
4043 NULL, 1, bytenr);
4044 if (ret)
4045 return ret;
4046 }
4047 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4048
4049 return 0;
4050}
4051
4052/*
4053 * get a reference count on fs_info->scrub_workers. start worker if necessary
4054 */
4055static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
4056 int is_dev_replace)
4057{
4058 unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
4059 int max_active = fs_info->thread_pool_size;
4060
4061 if (fs_info->scrub_workers_refcnt == 0) {
4062 fs_info->scrub_workers = btrfs_alloc_workqueue(fs_info, "scrub",
4063 flags, is_dev_replace ? 1 : max_active, 4);
4064 if (!fs_info->scrub_workers)
4065 goto fail_scrub_workers;
4066
4067 fs_info->scrub_wr_completion_workers =
4068 btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
4069 max_active, 2);
4070 if (!fs_info->scrub_wr_completion_workers)
4071 goto fail_scrub_wr_completion_workers;
4072
4073 fs_info->scrub_nocow_workers =
4074 btrfs_alloc_workqueue(fs_info, "scrubnc", flags, 1, 0);
4075 if (!fs_info->scrub_nocow_workers)
4076 goto fail_scrub_nocow_workers;
4077 fs_info->scrub_parity_workers =
4078 btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
4079 max_active, 2);
4080 if (!fs_info->scrub_parity_workers)
4081 goto fail_scrub_parity_workers;
4082 }
4083 ++fs_info->scrub_workers_refcnt;
4084 return 0;
4085
4086fail_scrub_parity_workers:
4087 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
4088fail_scrub_nocow_workers:
4089 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
4090fail_scrub_wr_completion_workers:
4091 btrfs_destroy_workqueue(fs_info->scrub_workers);
4092fail_scrub_workers:
4093 return -ENOMEM;
4094}
4095
4096static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
4097{
4098 if (--fs_info->scrub_workers_refcnt == 0) {
4099 btrfs_destroy_workqueue(fs_info->scrub_workers);
4100 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
4101 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
4102 btrfs_destroy_workqueue(fs_info->scrub_parity_workers);
4103 }
4104 WARN_ON(fs_info->scrub_workers_refcnt < 0);
4105}
4106
4107int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
4108 u64 end, struct btrfs_scrub_progress *progress,
4109 int readonly, int is_dev_replace)
4110{
4111 struct scrub_ctx *sctx;
4112 int ret;
4113 struct btrfs_device *dev;
4114 struct rcu_string *name;
4115
4116 if (btrfs_fs_closing(fs_info))
4117 return -EINVAL;
4118
4119 if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
4120 /*
4121 * in this case scrub is unable to calculate the checksum
4122 * the way scrub is implemented. Do not handle this
4123 * situation at all because it won't ever happen.
4124 */
4125 btrfs_err(fs_info,
4126 "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
4127 fs_info->nodesize,
4128 BTRFS_STRIPE_LEN);
4129 return -EINVAL;
4130 }
4131
4132 if (fs_info->sectorsize != PAGE_SIZE) {
4133 /* not supported for data w/o checksums */
4134 btrfs_err_rl(fs_info,
4135 "scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails",
4136 fs_info->sectorsize, PAGE_SIZE);
4137 return -EINVAL;
4138 }
4139
4140 if (fs_info->nodesize >
4141 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
4142 fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
4143 /*
4144 * would exhaust the array bounds of pagev member in
4145 * struct scrub_block
4146 */
4147 btrfs_err(fs_info,
4148 "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
4149 fs_info->nodesize,
4150 SCRUB_MAX_PAGES_PER_BLOCK,
4151 fs_info->sectorsize,
4152 SCRUB_MAX_PAGES_PER_BLOCK);
4153 return -EINVAL;
4154 }
4155
4156
4157 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4158 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
4159 if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
4160 !is_dev_replace)) {
4161 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4162 return -ENODEV;
4163 }
4164
4165 if (!is_dev_replace && !readonly &&
4166 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
4167 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4168 rcu_read_lock();
4169 name = rcu_dereference(dev->name);
4170 btrfs_err(fs_info, "scrub: device %s is not writable",
4171 name->str);
4172 rcu_read_unlock();
4173 return -EROFS;
4174 }
4175
4176 mutex_lock(&fs_info->scrub_lock);
4177 if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
4178 test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
4179 mutex_unlock(&fs_info->scrub_lock);
4180 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4181 return -EIO;
4182 }
4183
4184 btrfs_dev_replace_read_lock(&fs_info->dev_replace);
4185 if (dev->scrub_ctx ||
4186 (!is_dev_replace &&
4187 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
4188 btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
4189 mutex_unlock(&fs_info->scrub_lock);
4190 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4191 return -EINPROGRESS;
4192 }
4193 btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
4194
4195 ret = scrub_workers_get(fs_info, is_dev_replace);
4196 if (ret) {
4197 mutex_unlock(&fs_info->scrub_lock);
4198 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4199 return ret;
4200 }
4201
4202 sctx = scrub_setup_ctx(dev, is_dev_replace);
4203 if (IS_ERR(sctx)) {
4204 mutex_unlock(&fs_info->scrub_lock);
4205 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4206 scrub_workers_put(fs_info);
4207 return PTR_ERR(sctx);
4208 }
4209 sctx->readonly = readonly;
4210 dev->scrub_ctx = sctx;
4211 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4212
4213 /*
4214 * checking @scrub_pause_req here, we can avoid
4215 * race between committing transaction and scrubbing.
4216 */
4217 __scrub_blocked_if_needed(fs_info);
4218 atomic_inc(&fs_info->scrubs_running);
4219 mutex_unlock(&fs_info->scrub_lock);
4220
4221 if (!is_dev_replace) {
4222 /*
4223 * by holding device list mutex, we can
4224 * kick off writing super in log tree sync.
4225 */
4226 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4227 ret = scrub_supers(sctx, dev);
4228 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4229 }
4230
4231 if (!ret)
4232 ret = scrub_enumerate_chunks(sctx, dev, start, end,
4233 is_dev_replace);
4234
4235 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4236 atomic_dec(&fs_info->scrubs_running);
4237 wake_up(&fs_info->scrub_pause_wait);
4238
4239 wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
4240
4241 if (progress)
4242 memcpy(progress, &sctx->stat, sizeof(*progress));
4243
4244 mutex_lock(&fs_info->scrub_lock);
4245 dev->scrub_ctx = NULL;
4246 scrub_workers_put(fs_info);
4247 mutex_unlock(&fs_info->scrub_lock);
4248
4249 scrub_put_ctx(sctx);
4250
4251 return ret;
4252}
4253
4254void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
4255{
4256 mutex_lock(&fs_info->scrub_lock);
4257 atomic_inc(&fs_info->scrub_pause_req);
4258 while (atomic_read(&fs_info->scrubs_paused) !=
4259 atomic_read(&fs_info->scrubs_running)) {
4260 mutex_unlock(&fs_info->scrub_lock);
4261 wait_event(fs_info->scrub_pause_wait,
4262 atomic_read(&fs_info->scrubs_paused) ==
4263 atomic_read(&fs_info->scrubs_running));
4264 mutex_lock(&fs_info->scrub_lock);
4265 }
4266 mutex_unlock(&fs_info->scrub_lock);
4267}
4268
4269void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
4270{
4271 atomic_dec(&fs_info->scrub_pause_req);
4272 wake_up(&fs_info->scrub_pause_wait);
4273}
4274
4275int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
4276{
4277 mutex_lock(&fs_info->scrub_lock);
4278 if (!atomic_read(&fs_info->scrubs_running)) {
4279 mutex_unlock(&fs_info->scrub_lock);
4280 return -ENOTCONN;
4281 }
4282
4283 atomic_inc(&fs_info->scrub_cancel_req);
4284 while (atomic_read(&fs_info->scrubs_running)) {
4285 mutex_unlock(&fs_info->scrub_lock);
4286 wait_event(fs_info->scrub_pause_wait,
4287 atomic_read(&fs_info->scrubs_running) == 0);
4288 mutex_lock(&fs_info->scrub_lock);
4289 }
4290 atomic_dec(&fs_info->scrub_cancel_req);
4291 mutex_unlock(&fs_info->scrub_lock);
4292
4293 return 0;
4294}
4295
4296int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
4297 struct btrfs_device *dev)
4298{
4299 struct scrub_ctx *sctx;
4300
4301 mutex_lock(&fs_info->scrub_lock);
4302 sctx = dev->scrub_ctx;
4303 if (!sctx) {
4304 mutex_unlock(&fs_info->scrub_lock);
4305 return -ENOTCONN;
4306 }
4307 atomic_inc(&sctx->cancel_req);
4308 while (dev->scrub_ctx) {
4309 mutex_unlock(&fs_info->scrub_lock);
4310 wait_event(fs_info->scrub_pause_wait,
4311 dev->scrub_ctx == NULL);
4312 mutex_lock(&fs_info->scrub_lock);
4313 }
4314 mutex_unlock(&fs_info->scrub_lock);
4315
4316 return 0;
4317}
4318
4319int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4320 struct btrfs_scrub_progress *progress)
4321{
4322 struct btrfs_device *dev;
4323 struct scrub_ctx *sctx = NULL;
4324
4325 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4326 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
4327 if (dev)
4328 sctx = dev->scrub_ctx;
4329 if (sctx)
4330 memcpy(progress, &sctx->stat, sizeof(*progress));
4331 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4332
4333 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4334}
4335
4336static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4337 u64 extent_logical, u64 extent_len,
4338 u64 *extent_physical,
4339 struct btrfs_device **extent_dev,
4340 int *extent_mirror_num)
4341{
4342 u64 mapped_length;
4343 struct btrfs_bio *bbio = NULL;
4344 int ret;
4345
4346 mapped_length = extent_len;
4347 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4348 &mapped_length, &bbio, 0);
4349 if (ret || !bbio || mapped_length < extent_len ||
4350 !bbio->stripes[0].dev->bdev) {
4351 btrfs_put_bbio(bbio);
4352 return;
4353 }
4354
4355 *extent_physical = bbio->stripes[0].physical;
4356 *extent_mirror_num = bbio->mirror_num;
4357 *extent_dev = bbio->stripes[0].dev;
4358 btrfs_put_bbio(bbio);
4359}
4360
4361static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
4362 int mirror_num, u64 physical_for_dev_replace)
4363{
4364 struct scrub_copy_nocow_ctx *nocow_ctx;
4365 struct btrfs_fs_info *fs_info = sctx->fs_info;
4366
4367 nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
4368 if (!nocow_ctx) {
4369 spin_lock(&sctx->stat_lock);
4370 sctx->stat.malloc_errors++;
4371 spin_unlock(&sctx->stat_lock);
4372 return -ENOMEM;
4373 }
4374
4375 scrub_pending_trans_workers_inc(sctx);
4376
4377 nocow_ctx->sctx = sctx;
4378 nocow_ctx->logical = logical;
4379 nocow_ctx->len = len;
4380 nocow_ctx->mirror_num = mirror_num;
4381 nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
4382 btrfs_init_work(&nocow_ctx->work, btrfs_scrubnc_helper,
4383 copy_nocow_pages_worker, NULL, NULL);
4384 INIT_LIST_HEAD(&nocow_ctx->inodes);
4385 btrfs_queue_work(fs_info->scrub_nocow_workers,
4386 &nocow_ctx->work);
4387
4388 return 0;
4389}
4390
4391static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
4392{
4393 struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
4394 struct scrub_nocow_inode *nocow_inode;
4395
4396 nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
4397 if (!nocow_inode)
4398 return -ENOMEM;
4399 nocow_inode->inum = inum;
4400 nocow_inode->offset = offset;
4401 nocow_inode->root = root;
4402 list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
4403 return 0;
4404}
4405
4406#define COPY_COMPLETE 1
4407
4408static void copy_nocow_pages_worker(struct btrfs_work *work)
4409{
4410 struct scrub_copy_nocow_ctx *nocow_ctx =
4411 container_of(work, struct scrub_copy_nocow_ctx, work);
4412 struct scrub_ctx *sctx = nocow_ctx->sctx;
4413 struct btrfs_fs_info *fs_info = sctx->fs_info;
4414 struct btrfs_root *root = fs_info->extent_root;
4415 u64 logical = nocow_ctx->logical;
4416 u64 len = nocow_ctx->len;
4417 int mirror_num = nocow_ctx->mirror_num;
4418 u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4419 int ret;
4420 struct btrfs_trans_handle *trans = NULL;
4421 struct btrfs_path *path;
4422 int not_written = 0;
4423
4424 path = btrfs_alloc_path();
4425 if (!path) {
4426 spin_lock(&sctx->stat_lock);
4427 sctx->stat.malloc_errors++;
4428 spin_unlock(&sctx->stat_lock);
4429 not_written = 1;
4430 goto out;
4431 }
4432
4433 trans = btrfs_join_transaction(root);
4434 if (IS_ERR(trans)) {
4435 not_written = 1;
4436 goto out;
4437 }
4438
4439 ret = iterate_inodes_from_logical(logical, fs_info, path,
4440 record_inode_for_nocow, nocow_ctx, false);
4441 if (ret != 0 && ret != -ENOENT) {
4442 btrfs_warn(fs_info,
4443 "iterate_inodes_from_logical() failed: log %llu, phys %llu, len %llu, mir %u, ret %d",
4444 logical, physical_for_dev_replace, len, mirror_num,
4445 ret);
4446 not_written = 1;
4447 goto out;
4448 }
4449
4450 btrfs_end_transaction(trans);
4451 trans = NULL;
4452 while (!list_empty(&nocow_ctx->inodes)) {
4453 struct scrub_nocow_inode *entry;
4454 entry = list_first_entry(&nocow_ctx->inodes,
4455 struct scrub_nocow_inode,
4456 list);
4457 list_del_init(&entry->list);
4458 ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
4459 entry->root, nocow_ctx);
4460 kfree(entry);
4461 if (ret == COPY_COMPLETE) {
4462 ret = 0;
4463 break;
4464 } else if (ret) {
4465 break;
4466 }
4467 }
4468out:
4469 while (!list_empty(&nocow_ctx->inodes)) {
4470 struct scrub_nocow_inode *entry;
4471 entry = list_first_entry(&nocow_ctx->inodes,
4472 struct scrub_nocow_inode,
4473 list);
4474 list_del_init(&entry->list);
4475 kfree(entry);
4476 }
4477 if (trans && !IS_ERR(trans))
4478 btrfs_end_transaction(trans);
4479 if (not_written)
4480 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
4481 num_uncorrectable_read_errors);
4482
4483 btrfs_free_path(path);
4484 kfree(nocow_ctx);
4485
4486 scrub_pending_trans_workers_dec(sctx);
4487}
4488
4489static int check_extent_to_block(struct btrfs_inode *inode, u64 start, u64 len,
4490 u64 logical)
4491{
4492 struct extent_state *cached_state = NULL;
4493 struct btrfs_ordered_extent *ordered;
4494 struct extent_io_tree *io_tree;
4495 struct extent_map *em;
4496 u64 lockstart = start, lockend = start + len - 1;
4497 int ret = 0;
4498
4499 io_tree = &inode->io_tree;
4500
4501 lock_extent_bits(io_tree, lockstart, lockend, &cached_state);
4502 ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
4503 if (ordered) {
4504 btrfs_put_ordered_extent(ordered);
4505 ret = 1;
4506 goto out_unlock;
4507 }
4508
4509 em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
4510 if (IS_ERR(em)) {
4511 ret = PTR_ERR(em);
4512 goto out_unlock;
4513 }
4514
4515 /*
4516 * This extent does not actually cover the logical extent anymore,
4517 * move on to the next inode.
4518 */
4519 if (em->block_start > logical ||
4520 em->block_start + em->block_len < logical + len ||
4521 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4522 free_extent_map(em);
4523 ret = 1;
4524 goto out_unlock;
4525 }
4526 free_extent_map(em);
4527
4528out_unlock:
4529 unlock_extent_cached(io_tree, lockstart, lockend, &cached_state);
4530 return ret;
4531}
4532
4533static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
4534 struct scrub_copy_nocow_ctx *nocow_ctx)
4535{
4536 struct btrfs_fs_info *fs_info = nocow_ctx->sctx->fs_info;
4537 struct btrfs_key key;
4538 struct inode *inode;
4539 struct page *page;
4540 struct btrfs_root *local_root;
4541 struct extent_io_tree *io_tree;
4542 u64 physical_for_dev_replace;
4543 u64 nocow_ctx_logical;
4544 u64 len = nocow_ctx->len;
4545 unsigned long index;
4546 int srcu_index;
4547 int ret = 0;
4548 int err = 0;
4549
4550 key.objectid = root;
4551 key.type = BTRFS_ROOT_ITEM_KEY;
4552 key.offset = (u64)-1;
4553
4554 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
4555
4556 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
4557 if (IS_ERR(local_root)) {
4558 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4559 return PTR_ERR(local_root);
4560 }
4561
4562 key.type = BTRFS_INODE_ITEM_KEY;
4563 key.objectid = inum;
4564 key.offset = 0;
4565 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
4566 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4567 if (IS_ERR(inode))
4568 return PTR_ERR(inode);
4569
4570 /* Avoid truncate/dio/punch hole.. */
4571 inode_lock(inode);
4572 inode_dio_wait(inode);
4573
4574 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4575 io_tree = &BTRFS_I(inode)->io_tree;
4576 nocow_ctx_logical = nocow_ctx->logical;
4577
4578 ret = check_extent_to_block(BTRFS_I(inode), offset, len,
4579 nocow_ctx_logical);
4580 if (ret) {
4581 ret = ret > 0 ? 0 : ret;
4582 goto out;
4583 }
4584
4585 while (len >= PAGE_SIZE) {
4586 index = offset >> PAGE_SHIFT;
4587again:
4588 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
4589 if (!page) {
4590 btrfs_err(fs_info, "find_or_create_page() failed");
4591 ret = -ENOMEM;
4592 goto out;
4593 }
4594
4595 if (PageUptodate(page)) {
4596 if (PageDirty(page))
4597 goto next_page;
4598 } else {
4599 ClearPageError(page);
4600 err = extent_read_full_page(io_tree, page,
4601 btrfs_get_extent,
4602 nocow_ctx->mirror_num);
4603 if (err) {
4604 ret = err;
4605 goto next_page;
4606 }
4607
4608 lock_page(page);
4609 /*
4610 * If the page has been remove from the page cache,
4611 * the data on it is meaningless, because it may be
4612 * old one, the new data may be written into the new
4613 * page in the page cache.
4614 */
4615 if (page->mapping != inode->i_mapping) {
4616 unlock_page(page);
4617 put_page(page);
4618 goto again;
4619 }
4620 if (!PageUptodate(page)) {
4621 ret = -EIO;
4622 goto next_page;
4623 }
4624 }
4625
4626 ret = check_extent_to_block(BTRFS_I(inode), offset, len,
4627 nocow_ctx_logical);
4628 if (ret) {
4629 ret = ret > 0 ? 0 : ret;
4630 goto next_page;
4631 }
4632
4633 err = write_page_nocow(nocow_ctx->sctx,
4634 physical_for_dev_replace, page);
4635 if (err)
4636 ret = err;
4637next_page:
4638 unlock_page(page);
4639 put_page(page);
4640
4641 if (ret)
4642 break;
4643
4644 offset += PAGE_SIZE;
4645 physical_for_dev_replace += PAGE_SIZE;
4646 nocow_ctx_logical += PAGE_SIZE;
4647 len -= PAGE_SIZE;
4648 }
4649 ret = COPY_COMPLETE;
4650out:
4651 inode_unlock(inode);
4652 iput(inode);
4653 return ret;
4654}
4655
4656static int write_page_nocow(struct scrub_ctx *sctx,
4657 u64 physical_for_dev_replace, struct page *page)
4658{
4659 struct bio *bio;
4660 struct btrfs_device *dev;
4661
4662 dev = sctx->wr_tgtdev;
4663 if (!dev)
4664 return -EIO;
4665 if (!dev->bdev) {
4666 btrfs_warn_rl(dev->fs_info,
4667 "scrub write_page_nocow(bdev == NULL) is unexpected");
4668 return -EIO;
4669 }
4670 bio = btrfs_io_bio_alloc(1);
4671 bio->bi_iter.bi_size = 0;
4672 bio->bi_iter.bi_sector = physical_for_dev_replace >> 9;
4673 bio_set_dev(bio, dev->bdev);
4674 bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
4675 /* bio_add_page won't fail on a freshly allocated bio */
4676 bio_add_page(bio, page, PAGE_SIZE, 0);
4677
4678 if (btrfsic_submit_bio_wait(bio)) {
4679 bio_put(bio);
4680 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
4681 return -EIO;
4682 }
4683
4684 bio_put(bio);
4685 return 0;
4686}
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}