1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2012 Fusion-io All rights reserved.
4 * Copyright (C) 2012 Intel Corp. All rights reserved.
7 #include <linux/sched.h>
9 #include <linux/slab.h>
10 #include <linux/blkdev.h>
11 #include <linux/raid/pq.h>
12 #include <linux/hash.h>
13 #include <linux/list_sort.h>
14 #include <linux/raid/xor.h>
21 #include "async-thread.h"
22 #include "file-item.h"
23 #include "btrfs_inode.h"
25 /* set when additional merges to this rbio are not allowed */
26 #define RBIO_RMW_LOCKED_BIT 1
29 * set when this rbio is sitting in the hash, but it is just a cache
32 #define RBIO_CACHE_BIT 2
35 * set when it is safe to trust the stripe_pages for caching
37 #define RBIO_CACHE_READY_BIT 3
39 #define RBIO_CACHE_SIZE 1024
41 #define BTRFS_STRIPE_HASH_TABLE_BITS 11
43 /* Used by the raid56 code to lock stripes for read/modify/write */
44 struct btrfs_stripe_hash {
45 struct list_head hash_list;
49 /* Used by the raid56 code to lock stripes for read/modify/write */
50 struct btrfs_stripe_hash_table {
51 struct list_head stripe_cache;
52 spinlock_t cache_lock;
54 struct btrfs_stripe_hash table[];
58 * A bvec like structure to present a sector inside a page.
60 * Unlike bvec we don't need bvlen, as it's fixed to sectorsize.
64 unsigned int pgoff:24;
65 unsigned int uptodate:8;
68 static void rmw_rbio_work(struct work_struct *work);
69 static void rmw_rbio_work_locked(struct work_struct *work);
70 static void index_rbio_pages(struct btrfs_raid_bio *rbio);
71 static int alloc_rbio_pages(struct btrfs_raid_bio *rbio);
73 static int finish_parity_scrub(struct btrfs_raid_bio *rbio);
74 static void scrub_rbio_work_locked(struct work_struct *work);
76 static void free_raid_bio_pointers(struct btrfs_raid_bio *rbio)
78 bitmap_free(rbio->error_bitmap);
79 kfree(rbio->stripe_pages);
80 kfree(rbio->bio_sectors);
81 kfree(rbio->stripe_sectors);
82 kfree(rbio->finish_pointers);
85 static void free_raid_bio(struct btrfs_raid_bio *rbio)
89 if (!refcount_dec_and_test(&rbio->refs))
92 WARN_ON(!list_empty(&rbio->stripe_cache));
93 WARN_ON(!list_empty(&rbio->hash_list));
94 WARN_ON(!bio_list_empty(&rbio->bio_list));
96 for (i = 0; i < rbio->nr_pages; i++) {
97 if (rbio->stripe_pages[i]) {
98 __free_page(rbio->stripe_pages[i]);
99 rbio->stripe_pages[i] = NULL;
103 btrfs_put_bioc(rbio->bioc);
104 free_raid_bio_pointers(rbio);
108 static void start_async_work(struct btrfs_raid_bio *rbio, work_func_t work_func)
110 INIT_WORK(&rbio->work, work_func);
111 queue_work(rbio->bioc->fs_info->rmw_workers, &rbio->work);
115 * the stripe hash table is used for locking, and to collect
116 * bios in hopes of making a full stripe
118 int btrfs_alloc_stripe_hash_table(struct btrfs_fs_info *info)
120 struct btrfs_stripe_hash_table *table;
121 struct btrfs_stripe_hash_table *x;
122 struct btrfs_stripe_hash *cur;
123 struct btrfs_stripe_hash *h;
124 int num_entries = 1 << BTRFS_STRIPE_HASH_TABLE_BITS;
127 if (info->stripe_hash_table)
131 * The table is large, starting with order 4 and can go as high as
132 * order 7 in case lock debugging is turned on.
134 * Try harder to allocate and fallback to vmalloc to lower the chance
135 * of a failing mount.
137 table = kvzalloc(struct_size(table, table, num_entries), GFP_KERNEL);
141 spin_lock_init(&table->cache_lock);
142 INIT_LIST_HEAD(&table->stripe_cache);
146 for (i = 0; i < num_entries; i++) {
148 INIT_LIST_HEAD(&cur->hash_list);
149 spin_lock_init(&cur->lock);
152 x = cmpxchg(&info->stripe_hash_table, NULL, table);
158 * caching an rbio means to copy anything from the
159 * bio_sectors array into the stripe_pages array. We
160 * use the page uptodate bit in the stripe cache array
161 * to indicate if it has valid data
163 * once the caching is done, we set the cache ready
166 static void cache_rbio_pages(struct btrfs_raid_bio *rbio)
171 ret = alloc_rbio_pages(rbio);
175 for (i = 0; i < rbio->nr_sectors; i++) {
176 /* Some range not covered by bio (partial write), skip it */
177 if (!rbio->bio_sectors[i].page) {
179 * Even if the sector is not covered by bio, if it is
180 * a data sector it should still be uptodate as it is
183 if (i < rbio->nr_data * rbio->stripe_nsectors)
184 ASSERT(rbio->stripe_sectors[i].uptodate);
188 ASSERT(rbio->stripe_sectors[i].page);
189 memcpy_page(rbio->stripe_sectors[i].page,
190 rbio->stripe_sectors[i].pgoff,
191 rbio->bio_sectors[i].page,
192 rbio->bio_sectors[i].pgoff,
193 rbio->bioc->fs_info->sectorsize);
194 rbio->stripe_sectors[i].uptodate = 1;
196 set_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
200 * we hash on the first logical address of the stripe
202 static int rbio_bucket(struct btrfs_raid_bio *rbio)
204 u64 num = rbio->bioc->full_stripe_logical;
207 * we shift down quite a bit. We're using byte
208 * addressing, and most of the lower bits are zeros.
209 * This tends to upset hash_64, and it consistently
210 * returns just one or two different values.
212 * shifting off the lower bits fixes things.
214 return hash_64(num >> 16, BTRFS_STRIPE_HASH_TABLE_BITS);
217 static bool full_page_sectors_uptodate(struct btrfs_raid_bio *rbio,
218 unsigned int page_nr)
220 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
221 const u32 sectors_per_page = PAGE_SIZE / sectorsize;
224 ASSERT(page_nr < rbio->nr_pages);
226 for (i = sectors_per_page * page_nr;
227 i < sectors_per_page * page_nr + sectors_per_page;
229 if (!rbio->stripe_sectors[i].uptodate)
236 * Update the stripe_sectors[] array to use correct page and pgoff
238 * Should be called every time any page pointer in stripes_pages[] got modified.
240 static void index_stripe_sectors(struct btrfs_raid_bio *rbio)
242 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
246 for (i = 0, offset = 0; i < rbio->nr_sectors; i++, offset += sectorsize) {
247 int page_index = offset >> PAGE_SHIFT;
249 ASSERT(page_index < rbio->nr_pages);
250 rbio->stripe_sectors[i].page = rbio->stripe_pages[page_index];
251 rbio->stripe_sectors[i].pgoff = offset_in_page(offset);
255 static void steal_rbio_page(struct btrfs_raid_bio *src,
256 struct btrfs_raid_bio *dest, int page_nr)
258 const u32 sectorsize = src->bioc->fs_info->sectorsize;
259 const u32 sectors_per_page = PAGE_SIZE / sectorsize;
262 if (dest->stripe_pages[page_nr])
263 __free_page(dest->stripe_pages[page_nr]);
264 dest->stripe_pages[page_nr] = src->stripe_pages[page_nr];
265 src->stripe_pages[page_nr] = NULL;
267 /* Also update the sector->uptodate bits. */
268 for (i = sectors_per_page * page_nr;
269 i < sectors_per_page * page_nr + sectors_per_page; i++)
270 dest->stripe_sectors[i].uptodate = true;
273 static bool is_data_stripe_page(struct btrfs_raid_bio *rbio, int page_nr)
275 const int sector_nr = (page_nr << PAGE_SHIFT) >>
276 rbio->bioc->fs_info->sectorsize_bits;
279 * We have ensured PAGE_SIZE is aligned with sectorsize, thus
280 * we won't have a page which is half data half parity.
282 * Thus if the first sector of the page belongs to data stripes, then
283 * the full page belongs to data stripes.
285 return (sector_nr < rbio->nr_data * rbio->stripe_nsectors);
289 * Stealing an rbio means taking all the uptodate pages from the stripe array
290 * in the source rbio and putting them into the destination rbio.
292 * This will also update the involved stripe_sectors[] which are referring to
295 static void steal_rbio(struct btrfs_raid_bio *src, struct btrfs_raid_bio *dest)
299 if (!test_bit(RBIO_CACHE_READY_BIT, &src->flags))
302 for (i = 0; i < dest->nr_pages; i++) {
303 struct page *p = src->stripe_pages[i];
306 * We don't need to steal P/Q pages as they will always be
307 * regenerated for RMW or full write anyway.
309 if (!is_data_stripe_page(src, i))
313 * If @src already has RBIO_CACHE_READY_BIT, it should have
314 * all data stripe pages present and uptodate.
317 ASSERT(full_page_sectors_uptodate(src, i));
318 steal_rbio_page(src, dest, i);
320 index_stripe_sectors(dest);
321 index_stripe_sectors(src);
325 * merging means we take the bio_list from the victim and
326 * splice it into the destination. The victim should
327 * be discarded afterwards.
329 * must be called with dest->rbio_list_lock held
331 static void merge_rbio(struct btrfs_raid_bio *dest,
332 struct btrfs_raid_bio *victim)
334 bio_list_merge(&dest->bio_list, &victim->bio_list);
335 dest->bio_list_bytes += victim->bio_list_bytes;
336 /* Also inherit the bitmaps from @victim. */
337 bitmap_or(&dest->dbitmap, &victim->dbitmap, &dest->dbitmap,
338 dest->stripe_nsectors);
339 bio_list_init(&victim->bio_list);
343 * used to prune items that are in the cache. The caller
344 * must hold the hash table lock.
346 static void __remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
348 int bucket = rbio_bucket(rbio);
349 struct btrfs_stripe_hash_table *table;
350 struct btrfs_stripe_hash *h;
354 * check the bit again under the hash table lock.
356 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
359 table = rbio->bioc->fs_info->stripe_hash_table;
360 h = table->table + bucket;
362 /* hold the lock for the bucket because we may be
363 * removing it from the hash table
368 * hold the lock for the bio list because we need
369 * to make sure the bio list is empty
371 spin_lock(&rbio->bio_list_lock);
373 if (test_and_clear_bit(RBIO_CACHE_BIT, &rbio->flags)) {
374 list_del_init(&rbio->stripe_cache);
375 table->cache_size -= 1;
378 /* if the bio list isn't empty, this rbio is
379 * still involved in an IO. We take it out
380 * of the cache list, and drop the ref that
381 * was held for the list.
383 * If the bio_list was empty, we also remove
384 * the rbio from the hash_table, and drop
385 * the corresponding ref
387 if (bio_list_empty(&rbio->bio_list)) {
388 if (!list_empty(&rbio->hash_list)) {
389 list_del_init(&rbio->hash_list);
390 refcount_dec(&rbio->refs);
391 BUG_ON(!list_empty(&rbio->plug_list));
396 spin_unlock(&rbio->bio_list_lock);
397 spin_unlock(&h->lock);
404 * prune a given rbio from the cache
406 static void remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
408 struct btrfs_stripe_hash_table *table;
410 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
413 table = rbio->bioc->fs_info->stripe_hash_table;
415 spin_lock(&table->cache_lock);
416 __remove_rbio_from_cache(rbio);
417 spin_unlock(&table->cache_lock);
421 * remove everything in the cache
423 static void btrfs_clear_rbio_cache(struct btrfs_fs_info *info)
425 struct btrfs_stripe_hash_table *table;
426 struct btrfs_raid_bio *rbio;
428 table = info->stripe_hash_table;
430 spin_lock(&table->cache_lock);
431 while (!list_empty(&table->stripe_cache)) {
432 rbio = list_entry(table->stripe_cache.next,
433 struct btrfs_raid_bio,
435 __remove_rbio_from_cache(rbio);
437 spin_unlock(&table->cache_lock);
441 * remove all cached entries and free the hash table
444 void btrfs_free_stripe_hash_table(struct btrfs_fs_info *info)
446 if (!info->stripe_hash_table)
448 btrfs_clear_rbio_cache(info);
449 kvfree(info->stripe_hash_table);
450 info->stripe_hash_table = NULL;
454 * insert an rbio into the stripe cache. It
455 * must have already been prepared by calling
458 * If this rbio was already cached, it gets
459 * moved to the front of the lru.
461 * If the size of the rbio cache is too big, we
464 static void cache_rbio(struct btrfs_raid_bio *rbio)
466 struct btrfs_stripe_hash_table *table;
468 if (!test_bit(RBIO_CACHE_READY_BIT, &rbio->flags))
471 table = rbio->bioc->fs_info->stripe_hash_table;
473 spin_lock(&table->cache_lock);
474 spin_lock(&rbio->bio_list_lock);
476 /* bump our ref if we were not in the list before */
477 if (!test_and_set_bit(RBIO_CACHE_BIT, &rbio->flags))
478 refcount_inc(&rbio->refs);
480 if (!list_empty(&rbio->stripe_cache)){
481 list_move(&rbio->stripe_cache, &table->stripe_cache);
483 list_add(&rbio->stripe_cache, &table->stripe_cache);
484 table->cache_size += 1;
487 spin_unlock(&rbio->bio_list_lock);
489 if (table->cache_size > RBIO_CACHE_SIZE) {
490 struct btrfs_raid_bio *found;
492 found = list_entry(table->stripe_cache.prev,
493 struct btrfs_raid_bio,
497 __remove_rbio_from_cache(found);
500 spin_unlock(&table->cache_lock);
504 * helper function to run the xor_blocks api. It is only
505 * able to do MAX_XOR_BLOCKS at a time, so we need to
508 static void run_xor(void **pages, int src_cnt, ssize_t len)
512 void *dest = pages[src_cnt];
515 xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS);
516 xor_blocks(xor_src_cnt, len, dest, pages + src_off);
518 src_cnt -= xor_src_cnt;
519 src_off += xor_src_cnt;
524 * Returns true if the bio list inside this rbio covers an entire stripe (no
527 static int rbio_is_full(struct btrfs_raid_bio *rbio)
529 unsigned long size = rbio->bio_list_bytes;
532 spin_lock(&rbio->bio_list_lock);
533 if (size != rbio->nr_data * BTRFS_STRIPE_LEN)
535 BUG_ON(size > rbio->nr_data * BTRFS_STRIPE_LEN);
536 spin_unlock(&rbio->bio_list_lock);
542 * returns 1 if it is safe to merge two rbios together.
543 * The merging is safe if the two rbios correspond to
544 * the same stripe and if they are both going in the same
545 * direction (read vs write), and if neither one is
546 * locked for final IO
548 * The caller is responsible for locking such that
549 * rmw_locked is safe to test
551 static int rbio_can_merge(struct btrfs_raid_bio *last,
552 struct btrfs_raid_bio *cur)
554 if (test_bit(RBIO_RMW_LOCKED_BIT, &last->flags) ||
555 test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags))
559 * we can't merge with cached rbios, since the
560 * idea is that when we merge the destination
561 * rbio is going to run our IO for us. We can
562 * steal from cached rbios though, other functions
565 if (test_bit(RBIO_CACHE_BIT, &last->flags) ||
566 test_bit(RBIO_CACHE_BIT, &cur->flags))
569 if (last->bioc->full_stripe_logical != cur->bioc->full_stripe_logical)
572 /* we can't merge with different operations */
573 if (last->operation != cur->operation)
576 * We've need read the full stripe from the drive.
577 * check and repair the parity and write the new results.
579 * We're not allowed to add any new bios to the
580 * bio list here, anyone else that wants to
581 * change this stripe needs to do their own rmw.
583 if (last->operation == BTRFS_RBIO_PARITY_SCRUB)
586 if (last->operation == BTRFS_RBIO_READ_REBUILD)
592 static unsigned int rbio_stripe_sector_index(const struct btrfs_raid_bio *rbio,
593 unsigned int stripe_nr,
594 unsigned int sector_nr)
596 ASSERT(stripe_nr < rbio->real_stripes);
597 ASSERT(sector_nr < rbio->stripe_nsectors);
599 return stripe_nr * rbio->stripe_nsectors + sector_nr;
602 /* Return a sector from rbio->stripe_sectors, not from the bio list */
603 static struct sector_ptr *rbio_stripe_sector(const struct btrfs_raid_bio *rbio,
604 unsigned int stripe_nr,
605 unsigned int sector_nr)
607 return &rbio->stripe_sectors[rbio_stripe_sector_index(rbio, stripe_nr,
611 /* Grab a sector inside P stripe */
612 static struct sector_ptr *rbio_pstripe_sector(const struct btrfs_raid_bio *rbio,
613 unsigned int sector_nr)
615 return rbio_stripe_sector(rbio, rbio->nr_data, sector_nr);
618 /* Grab a sector inside Q stripe, return NULL if not RAID6 */
619 static struct sector_ptr *rbio_qstripe_sector(const struct btrfs_raid_bio *rbio,
620 unsigned int sector_nr)
622 if (rbio->nr_data + 1 == rbio->real_stripes)
624 return rbio_stripe_sector(rbio, rbio->nr_data + 1, sector_nr);
628 * The first stripe in the table for a logical address
629 * has the lock. rbios are added in one of three ways:
631 * 1) Nobody has the stripe locked yet. The rbio is given
632 * the lock and 0 is returned. The caller must start the IO
635 * 2) Someone has the stripe locked, but we're able to merge
636 * with the lock owner. The rbio is freed and the IO will
637 * start automatically along with the existing rbio. 1 is returned.
639 * 3) Someone has the stripe locked, but we're not able to merge.
640 * The rbio is added to the lock owner's plug list, or merged into
641 * an rbio already on the plug list. When the lock owner unlocks,
642 * the next rbio on the list is run and the IO is started automatically.
645 * If we return 0, the caller still owns the rbio and must continue with
646 * IO submission. If we return 1, the caller must assume the rbio has
647 * already been freed.
649 static noinline int lock_stripe_add(struct btrfs_raid_bio *rbio)
651 struct btrfs_stripe_hash *h;
652 struct btrfs_raid_bio *cur;
653 struct btrfs_raid_bio *pending;
654 struct btrfs_raid_bio *freeit = NULL;
655 struct btrfs_raid_bio *cache_drop = NULL;
658 h = rbio->bioc->fs_info->stripe_hash_table->table + rbio_bucket(rbio);
661 list_for_each_entry(cur, &h->hash_list, hash_list) {
662 if (cur->bioc->full_stripe_logical != rbio->bioc->full_stripe_logical)
665 spin_lock(&cur->bio_list_lock);
667 /* Can we steal this cached rbio's pages? */
668 if (bio_list_empty(&cur->bio_list) &&
669 list_empty(&cur->plug_list) &&
670 test_bit(RBIO_CACHE_BIT, &cur->flags) &&
671 !test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) {
672 list_del_init(&cur->hash_list);
673 refcount_dec(&cur->refs);
675 steal_rbio(cur, rbio);
677 spin_unlock(&cur->bio_list_lock);
682 /* Can we merge into the lock owner? */
683 if (rbio_can_merge(cur, rbio)) {
684 merge_rbio(cur, rbio);
685 spin_unlock(&cur->bio_list_lock);
693 * We couldn't merge with the running rbio, see if we can merge
694 * with the pending ones. We don't have to check for rmw_locked
695 * because there is no way they are inside finish_rmw right now
697 list_for_each_entry(pending, &cur->plug_list, plug_list) {
698 if (rbio_can_merge(pending, rbio)) {
699 merge_rbio(pending, rbio);
700 spin_unlock(&cur->bio_list_lock);
708 * No merging, put us on the tail of the plug list, our rbio
709 * will be started with the currently running rbio unlocks
711 list_add_tail(&rbio->plug_list, &cur->plug_list);
712 spin_unlock(&cur->bio_list_lock);
717 refcount_inc(&rbio->refs);
718 list_add(&rbio->hash_list, &h->hash_list);
720 spin_unlock(&h->lock);
722 remove_rbio_from_cache(cache_drop);
724 free_raid_bio(freeit);
728 static void recover_rbio_work_locked(struct work_struct *work);
731 * called as rmw or parity rebuild is completed. If the plug list has more
732 * rbios waiting for this stripe, the next one on the list will be started
734 static noinline void unlock_stripe(struct btrfs_raid_bio *rbio)
737 struct btrfs_stripe_hash *h;
740 bucket = rbio_bucket(rbio);
741 h = rbio->bioc->fs_info->stripe_hash_table->table + bucket;
743 if (list_empty(&rbio->plug_list))
747 spin_lock(&rbio->bio_list_lock);
749 if (!list_empty(&rbio->hash_list)) {
751 * if we're still cached and there is no other IO
752 * to perform, just leave this rbio here for others
753 * to steal from later
755 if (list_empty(&rbio->plug_list) &&
756 test_bit(RBIO_CACHE_BIT, &rbio->flags)) {
758 clear_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
759 BUG_ON(!bio_list_empty(&rbio->bio_list));
763 list_del_init(&rbio->hash_list);
764 refcount_dec(&rbio->refs);
767 * we use the plug list to hold all the rbios
768 * waiting for the chance to lock this stripe.
769 * hand the lock over to one of them.
771 if (!list_empty(&rbio->plug_list)) {
772 struct btrfs_raid_bio *next;
773 struct list_head *head = rbio->plug_list.next;
775 next = list_entry(head, struct btrfs_raid_bio,
778 list_del_init(&rbio->plug_list);
780 list_add(&next->hash_list, &h->hash_list);
781 refcount_inc(&next->refs);
782 spin_unlock(&rbio->bio_list_lock);
783 spin_unlock(&h->lock);
785 if (next->operation == BTRFS_RBIO_READ_REBUILD) {
786 start_async_work(next, recover_rbio_work_locked);
787 } else if (next->operation == BTRFS_RBIO_WRITE) {
788 steal_rbio(rbio, next);
789 start_async_work(next, rmw_rbio_work_locked);
790 } else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) {
791 steal_rbio(rbio, next);
792 start_async_work(next, scrub_rbio_work_locked);
799 spin_unlock(&rbio->bio_list_lock);
800 spin_unlock(&h->lock);
804 remove_rbio_from_cache(rbio);
807 static void rbio_endio_bio_list(struct bio *cur, blk_status_t err)
814 cur->bi_status = err;
821 * this frees the rbio and runs through all the bios in the
822 * bio_list and calls end_io on them
824 static void rbio_orig_end_io(struct btrfs_raid_bio *rbio, blk_status_t err)
826 struct bio *cur = bio_list_get(&rbio->bio_list);
829 kfree(rbio->csum_buf);
830 bitmap_free(rbio->csum_bitmap);
831 rbio->csum_buf = NULL;
832 rbio->csum_bitmap = NULL;
835 * Clear the data bitmap, as the rbio may be cached for later usage.
836 * do this before before unlock_stripe() so there will be no new bio
839 bitmap_clear(&rbio->dbitmap, 0, rbio->stripe_nsectors);
842 * At this moment, rbio->bio_list is empty, however since rbio does not
843 * always have RBIO_RMW_LOCKED_BIT set and rbio is still linked on the
844 * hash list, rbio may be merged with others so that rbio->bio_list
846 * Once unlock_stripe() is done, rbio->bio_list will not be updated any
847 * more and we can call bio_endio() on all queued bios.
850 extra = bio_list_get(&rbio->bio_list);
853 rbio_endio_bio_list(cur, err);
855 rbio_endio_bio_list(extra, err);
859 * Get a sector pointer specified by its @stripe_nr and @sector_nr.
861 * @rbio: The raid bio
862 * @stripe_nr: Stripe number, valid range [0, real_stripe)
863 * @sector_nr: Sector number inside the stripe,
864 * valid range [0, stripe_nsectors)
865 * @bio_list_only: Whether to use sectors inside the bio list only.
867 * The read/modify/write code wants to reuse the original bio page as much
868 * as possible, and only use stripe_sectors as fallback.
870 static struct sector_ptr *sector_in_rbio(struct btrfs_raid_bio *rbio,
871 int stripe_nr, int sector_nr,
874 struct sector_ptr *sector;
877 ASSERT(stripe_nr >= 0 && stripe_nr < rbio->real_stripes);
878 ASSERT(sector_nr >= 0 && sector_nr < rbio->stripe_nsectors);
880 index = stripe_nr * rbio->stripe_nsectors + sector_nr;
881 ASSERT(index >= 0 && index < rbio->nr_sectors);
883 spin_lock(&rbio->bio_list_lock);
884 sector = &rbio->bio_sectors[index];
885 if (sector->page || bio_list_only) {
886 /* Don't return sector without a valid page pointer */
889 spin_unlock(&rbio->bio_list_lock);
892 spin_unlock(&rbio->bio_list_lock);
894 return &rbio->stripe_sectors[index];
898 * allocation and initial setup for the btrfs_raid_bio. Not
899 * this does not allocate any pages for rbio->pages.
901 static struct btrfs_raid_bio *alloc_rbio(struct btrfs_fs_info *fs_info,
902 struct btrfs_io_context *bioc)
904 const unsigned int real_stripes = bioc->num_stripes - bioc->replace_nr_stripes;
905 const unsigned int stripe_npages = BTRFS_STRIPE_LEN >> PAGE_SHIFT;
906 const unsigned int num_pages = stripe_npages * real_stripes;
907 const unsigned int stripe_nsectors =
908 BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits;
909 const unsigned int num_sectors = stripe_nsectors * real_stripes;
910 struct btrfs_raid_bio *rbio;
912 /* PAGE_SIZE must also be aligned to sectorsize for subpage support */
913 ASSERT(IS_ALIGNED(PAGE_SIZE, fs_info->sectorsize));
915 * Our current stripe len should be fixed to 64k thus stripe_nsectors
916 * (at most 16) should be no larger than BITS_PER_LONG.
918 ASSERT(stripe_nsectors <= BITS_PER_LONG);
921 * Real stripes must be between 2 (2 disks RAID5, aka RAID1) and 256
924 ASSERT(real_stripes >= 2);
925 ASSERT(real_stripes <= U8_MAX);
927 rbio = kzalloc(sizeof(*rbio), GFP_NOFS);
929 return ERR_PTR(-ENOMEM);
930 rbio->stripe_pages = kcalloc(num_pages, sizeof(struct page *),
932 rbio->bio_sectors = kcalloc(num_sectors, sizeof(struct sector_ptr),
934 rbio->stripe_sectors = kcalloc(num_sectors, sizeof(struct sector_ptr),
936 rbio->finish_pointers = kcalloc(real_stripes, sizeof(void *), GFP_NOFS);
937 rbio->error_bitmap = bitmap_zalloc(num_sectors, GFP_NOFS);
939 if (!rbio->stripe_pages || !rbio->bio_sectors || !rbio->stripe_sectors ||
940 !rbio->finish_pointers || !rbio->error_bitmap) {
941 free_raid_bio_pointers(rbio);
943 return ERR_PTR(-ENOMEM);
946 bio_list_init(&rbio->bio_list);
947 init_waitqueue_head(&rbio->io_wait);
948 INIT_LIST_HEAD(&rbio->plug_list);
949 spin_lock_init(&rbio->bio_list_lock);
950 INIT_LIST_HEAD(&rbio->stripe_cache);
951 INIT_LIST_HEAD(&rbio->hash_list);
952 btrfs_get_bioc(bioc);
954 rbio->nr_pages = num_pages;
955 rbio->nr_sectors = num_sectors;
956 rbio->real_stripes = real_stripes;
957 rbio->stripe_npages = stripe_npages;
958 rbio->stripe_nsectors = stripe_nsectors;
959 refcount_set(&rbio->refs, 1);
960 atomic_set(&rbio->stripes_pending, 0);
962 ASSERT(btrfs_nr_parity_stripes(bioc->map_type));
963 rbio->nr_data = real_stripes - btrfs_nr_parity_stripes(bioc->map_type);
964 ASSERT(rbio->nr_data > 0);
969 /* allocate pages for all the stripes in the bio, including parity */
970 static int alloc_rbio_pages(struct btrfs_raid_bio *rbio)
974 ret = btrfs_alloc_page_array(rbio->nr_pages, rbio->stripe_pages, 0);
977 /* Mapping all sectors */
978 index_stripe_sectors(rbio);
982 /* only allocate pages for p/q stripes */
983 static int alloc_rbio_parity_pages(struct btrfs_raid_bio *rbio)
985 const int data_pages = rbio->nr_data * rbio->stripe_npages;
988 ret = btrfs_alloc_page_array(rbio->nr_pages - data_pages,
989 rbio->stripe_pages + data_pages, 0);
993 index_stripe_sectors(rbio);
998 * Return the total number of errors found in the vertical stripe of @sector_nr.
1000 * @faila and @failb will also be updated to the first and second stripe
1001 * number of the errors.
1003 static int get_rbio_veritical_errors(struct btrfs_raid_bio *rbio, int sector_nr,
1004 int *faila, int *failb)
1007 int found_errors = 0;
1009 if (faila || failb) {
1011 * Both @faila and @failb should be valid pointers if any of
1012 * them is specified.
1014 ASSERT(faila && failb);
1019 for (stripe_nr = 0; stripe_nr < rbio->real_stripes; stripe_nr++) {
1020 int total_sector_nr = stripe_nr * rbio->stripe_nsectors + sector_nr;
1022 if (test_bit(total_sector_nr, rbio->error_bitmap)) {
1025 /* Update faila and failb. */
1028 else if (*failb < 0)
1033 return found_errors;
1037 * Add a single sector @sector into our list of bios for IO.
1039 * Return 0 if everything went well.
1040 * Return <0 for error.
1042 static int rbio_add_io_sector(struct btrfs_raid_bio *rbio,
1043 struct bio_list *bio_list,
1044 struct sector_ptr *sector,
1045 unsigned int stripe_nr,
1046 unsigned int sector_nr,
1049 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
1050 struct bio *last = bio_list->tail;
1053 struct btrfs_io_stripe *stripe;
1057 * Note: here stripe_nr has taken device replace into consideration,
1058 * thus it can be larger than rbio->real_stripe.
1059 * So here we check against bioc->num_stripes, not rbio->real_stripes.
1061 ASSERT(stripe_nr >= 0 && stripe_nr < rbio->bioc->num_stripes);
1062 ASSERT(sector_nr >= 0 && sector_nr < rbio->stripe_nsectors);
1063 ASSERT(sector->page);
1065 stripe = &rbio->bioc->stripes[stripe_nr];
1066 disk_start = stripe->physical + sector_nr * sectorsize;
1068 /* if the device is missing, just fail this stripe */
1069 if (!stripe->dev->bdev) {
1072 set_bit(stripe_nr * rbio->stripe_nsectors + sector_nr,
1073 rbio->error_bitmap);
1075 /* Check if we have reached tolerance early. */
1076 found_errors = get_rbio_veritical_errors(rbio, sector_nr,
1078 if (found_errors > rbio->bioc->max_errors)
1083 /* see if we can add this page onto our existing bio */
1085 u64 last_end = last->bi_iter.bi_sector << SECTOR_SHIFT;
1086 last_end += last->bi_iter.bi_size;
1089 * we can't merge these if they are from different
1090 * devices or if they are not contiguous
1092 if (last_end == disk_start && !last->bi_status &&
1093 last->bi_bdev == stripe->dev->bdev) {
1094 ret = bio_add_page(last, sector->page, sectorsize,
1096 if (ret == sectorsize)
1101 /* put a new bio on the list */
1102 bio = bio_alloc(stripe->dev->bdev,
1103 max(BTRFS_STRIPE_LEN >> PAGE_SHIFT, 1),
1105 bio->bi_iter.bi_sector = disk_start >> SECTOR_SHIFT;
1106 bio->bi_private = rbio;
1108 __bio_add_page(bio, sector->page, sectorsize, sector->pgoff);
1109 bio_list_add(bio_list, bio);
1113 static void index_one_bio(struct btrfs_raid_bio *rbio, struct bio *bio)
1115 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
1116 struct bio_vec bvec;
1117 struct bvec_iter iter;
1118 u32 offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) -
1119 rbio->bioc->full_stripe_logical;
1121 bio_for_each_segment(bvec, bio, iter) {
1124 for (bvec_offset = 0; bvec_offset < bvec.bv_len;
1125 bvec_offset += sectorsize, offset += sectorsize) {
1126 int index = offset / sectorsize;
1127 struct sector_ptr *sector = &rbio->bio_sectors[index];
1129 sector->page = bvec.bv_page;
1130 sector->pgoff = bvec.bv_offset + bvec_offset;
1131 ASSERT(sector->pgoff < PAGE_SIZE);
1137 * helper function to walk our bio list and populate the bio_pages array with
1138 * the result. This seems expensive, but it is faster than constantly
1139 * searching through the bio list as we setup the IO in finish_rmw or stripe
1142 * This must be called before you trust the answers from page_in_rbio
1144 static void index_rbio_pages(struct btrfs_raid_bio *rbio)
1148 spin_lock(&rbio->bio_list_lock);
1149 bio_list_for_each(bio, &rbio->bio_list)
1150 index_one_bio(rbio, bio);
1152 spin_unlock(&rbio->bio_list_lock);
1155 static void bio_get_trace_info(struct btrfs_raid_bio *rbio, struct bio *bio,
1156 struct raid56_bio_trace_info *trace_info)
1158 const struct btrfs_io_context *bioc = rbio->bioc;
1163 /* We rely on bio->bi_bdev to find the stripe number. */
1167 for (i = 0; i < bioc->num_stripes; i++) {
1168 if (bio->bi_bdev != bioc->stripes[i].dev->bdev)
1170 trace_info->stripe_nr = i;
1171 trace_info->devid = bioc->stripes[i].dev->devid;
1172 trace_info->offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) -
1173 bioc->stripes[i].physical;
1178 trace_info->devid = -1;
1179 trace_info->offset = -1;
1180 trace_info->stripe_nr = -1;
1183 static inline void bio_list_put(struct bio_list *bio_list)
1187 while ((bio = bio_list_pop(bio_list)))
1191 static void assert_rbio(struct btrfs_raid_bio *rbio)
1193 if (!IS_ENABLED(CONFIG_BTRFS_DEBUG) ||
1194 !IS_ENABLED(CONFIG_BTRFS_ASSERT))
1198 * At least two stripes (2 disks RAID5), and since real_stripes is U8,
1199 * we won't go beyond 256 disks anyway.
1201 ASSERT(rbio->real_stripes >= 2);
1202 ASSERT(rbio->nr_data > 0);
1205 * This is another check to make sure nr data stripes is smaller
1206 * than total stripes.
1208 ASSERT(rbio->nr_data < rbio->real_stripes);
1211 /* Generate PQ for one vertical stripe. */
1212 static void generate_pq_vertical(struct btrfs_raid_bio *rbio, int sectornr)
1214 void **pointers = rbio->finish_pointers;
1215 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
1216 struct sector_ptr *sector;
1218 const bool has_qstripe = rbio->bioc->map_type & BTRFS_BLOCK_GROUP_RAID6;
1220 /* First collect one sector from each data stripe */
1221 for (stripe = 0; stripe < rbio->nr_data; stripe++) {
1222 sector = sector_in_rbio(rbio, stripe, sectornr, 0);
1223 pointers[stripe] = kmap_local_page(sector->page) +
1227 /* Then add the parity stripe */
1228 sector = rbio_pstripe_sector(rbio, sectornr);
1229 sector->uptodate = 1;
1230 pointers[stripe++] = kmap_local_page(sector->page) + sector->pgoff;
1234 * RAID6, add the qstripe and call the library function
1235 * to fill in our p/q
1237 sector = rbio_qstripe_sector(rbio, sectornr);
1238 sector->uptodate = 1;
1239 pointers[stripe++] = kmap_local_page(sector->page) +
1243 raid6_call.gen_syndrome(rbio->real_stripes, sectorsize,
1247 memcpy(pointers[rbio->nr_data], pointers[0], sectorsize);
1248 run_xor(pointers + 1, rbio->nr_data - 1, sectorsize);
1250 for (stripe = stripe - 1; stripe >= 0; stripe--)
1251 kunmap_local(pointers[stripe]);
1254 static int rmw_assemble_write_bios(struct btrfs_raid_bio *rbio,
1255 struct bio_list *bio_list)
1257 /* The total sector number inside the full stripe. */
1258 int total_sector_nr;
1263 ASSERT(bio_list_size(bio_list) == 0);
1265 /* We should have at least one data sector. */
1266 ASSERT(bitmap_weight(&rbio->dbitmap, rbio->stripe_nsectors));
1269 * Reset errors, as we may have errors inherited from from degraded
1272 bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors);
1275 * Start assembly. Make bios for everything from the higher layers (the
1276 * bio_list in our rbio) and our P/Q. Ignore everything else.
1278 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
1279 total_sector_nr++) {
1280 struct sector_ptr *sector;
1282 stripe = total_sector_nr / rbio->stripe_nsectors;
1283 sectornr = total_sector_nr % rbio->stripe_nsectors;
1285 /* This vertical stripe has no data, skip it. */
1286 if (!test_bit(sectornr, &rbio->dbitmap))
1289 if (stripe < rbio->nr_data) {
1290 sector = sector_in_rbio(rbio, stripe, sectornr, 1);
1294 sector = rbio_stripe_sector(rbio, stripe, sectornr);
1297 ret = rbio_add_io_sector(rbio, bio_list, sector, stripe,
1298 sectornr, REQ_OP_WRITE);
1303 if (likely(!rbio->bioc->replace_nr_stripes))
1307 * Make a copy for the replace target device.
1309 * Thus the source stripe number (in replace_stripe_src) should be valid.
1311 ASSERT(rbio->bioc->replace_stripe_src >= 0);
1313 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
1314 total_sector_nr++) {
1315 struct sector_ptr *sector;
1317 stripe = total_sector_nr / rbio->stripe_nsectors;
1318 sectornr = total_sector_nr % rbio->stripe_nsectors;
1321 * For RAID56, there is only one device that can be replaced,
1322 * and replace_stripe_src[0] indicates the stripe number we
1323 * need to copy from.
1325 if (stripe != rbio->bioc->replace_stripe_src) {
1327 * We can skip the whole stripe completely, note
1328 * total_sector_nr will be increased by one anyway.
1330 ASSERT(sectornr == 0);
1331 total_sector_nr += rbio->stripe_nsectors - 1;
1335 /* This vertical stripe has no data, skip it. */
1336 if (!test_bit(sectornr, &rbio->dbitmap))
1339 if (stripe < rbio->nr_data) {
1340 sector = sector_in_rbio(rbio, stripe, sectornr, 1);
1344 sector = rbio_stripe_sector(rbio, stripe, sectornr);
1347 ret = rbio_add_io_sector(rbio, bio_list, sector,
1349 sectornr, REQ_OP_WRITE);
1356 bio_list_put(bio_list);
1360 static void set_rbio_range_error(struct btrfs_raid_bio *rbio, struct bio *bio)
1362 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1363 u32 offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) -
1364 rbio->bioc->full_stripe_logical;
1365 int total_nr_sector = offset >> fs_info->sectorsize_bits;
1367 ASSERT(total_nr_sector < rbio->nr_data * rbio->stripe_nsectors);
1369 bitmap_set(rbio->error_bitmap, total_nr_sector,
1370 bio->bi_iter.bi_size >> fs_info->sectorsize_bits);
1373 * Special handling for raid56_alloc_missing_rbio() used by
1374 * scrub/replace. Unlike call path in raid56_parity_recover(), they
1375 * pass an empty bio here. Thus we have to find out the missing device
1376 * and mark the stripe error instead.
1378 if (bio->bi_iter.bi_size == 0) {
1379 bool found_missing = false;
1382 for (stripe_nr = 0; stripe_nr < rbio->real_stripes; stripe_nr++) {
1383 if (!rbio->bioc->stripes[stripe_nr].dev->bdev) {
1384 found_missing = true;
1385 bitmap_set(rbio->error_bitmap,
1386 stripe_nr * rbio->stripe_nsectors,
1387 rbio->stripe_nsectors);
1390 ASSERT(found_missing);
1395 * For subpage case, we can no longer set page Up-to-date directly for
1396 * stripe_pages[], thus we need to locate the sector.
1398 static struct sector_ptr *find_stripe_sector(struct btrfs_raid_bio *rbio,
1404 for (i = 0; i < rbio->nr_sectors; i++) {
1405 struct sector_ptr *sector = &rbio->stripe_sectors[i];
1407 if (sector->page == page && sector->pgoff == pgoff)
1414 * this sets each page in the bio uptodate. It should only be used on private
1415 * rbio pages, nothing that comes in from the higher layers
1417 static void set_bio_pages_uptodate(struct btrfs_raid_bio *rbio, struct bio *bio)
1419 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
1420 struct bio_vec *bvec;
1421 struct bvec_iter_all iter_all;
1423 ASSERT(!bio_flagged(bio, BIO_CLONED));
1425 bio_for_each_segment_all(bvec, bio, iter_all) {
1426 struct sector_ptr *sector;
1429 for (pgoff = bvec->bv_offset; pgoff - bvec->bv_offset < bvec->bv_len;
1430 pgoff += sectorsize) {
1431 sector = find_stripe_sector(rbio, bvec->bv_page, pgoff);
1434 sector->uptodate = 1;
1439 static int get_bio_sector_nr(struct btrfs_raid_bio *rbio, struct bio *bio)
1441 struct bio_vec *bv = bio_first_bvec_all(bio);
1444 for (i = 0; i < rbio->nr_sectors; i++) {
1445 struct sector_ptr *sector;
1447 sector = &rbio->stripe_sectors[i];
1448 if (sector->page == bv->bv_page && sector->pgoff == bv->bv_offset)
1450 sector = &rbio->bio_sectors[i];
1451 if (sector->page == bv->bv_page && sector->pgoff == bv->bv_offset)
1454 ASSERT(i < rbio->nr_sectors);
1458 static void rbio_update_error_bitmap(struct btrfs_raid_bio *rbio, struct bio *bio)
1460 int total_sector_nr = get_bio_sector_nr(rbio, bio);
1462 struct bio_vec *bvec;
1465 bio_for_each_bvec_all(bvec, bio, i)
1466 bio_size += bvec->bv_len;
1469 * Since we can have multiple bios touching the error_bitmap, we cannot
1470 * call bitmap_set() without protection.
1472 * Instead use set_bit() for each bit, as set_bit() itself is atomic.
1474 for (i = total_sector_nr; i < total_sector_nr +
1475 (bio_size >> rbio->bioc->fs_info->sectorsize_bits); i++)
1476 set_bit(i, rbio->error_bitmap);
1479 /* Verify the data sectors at read time. */
1480 static void verify_bio_data_sectors(struct btrfs_raid_bio *rbio,
1483 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1484 int total_sector_nr = get_bio_sector_nr(rbio, bio);
1485 struct bio_vec *bvec;
1486 struct bvec_iter_all iter_all;
1488 /* No data csum for the whole stripe, no need to verify. */
1489 if (!rbio->csum_bitmap || !rbio->csum_buf)
1492 /* P/Q stripes, they have no data csum to verify against. */
1493 if (total_sector_nr >= rbio->nr_data * rbio->stripe_nsectors)
1496 bio_for_each_segment_all(bvec, bio, iter_all) {
1499 for (bv_offset = bvec->bv_offset;
1500 bv_offset < bvec->bv_offset + bvec->bv_len;
1501 bv_offset += fs_info->sectorsize, total_sector_nr++) {
1502 u8 csum_buf[BTRFS_CSUM_SIZE];
1503 u8 *expected_csum = rbio->csum_buf +
1504 total_sector_nr * fs_info->csum_size;
1507 /* No csum for this sector, skip to the next sector. */
1508 if (!test_bit(total_sector_nr, rbio->csum_bitmap))
1511 ret = btrfs_check_sector_csum(fs_info, bvec->bv_page,
1512 bv_offset, csum_buf, expected_csum);
1514 set_bit(total_sector_nr, rbio->error_bitmap);
1519 static void raid_wait_read_end_io(struct bio *bio)
1521 struct btrfs_raid_bio *rbio = bio->bi_private;
1523 if (bio->bi_status) {
1524 rbio_update_error_bitmap(rbio, bio);
1526 set_bio_pages_uptodate(rbio, bio);
1527 verify_bio_data_sectors(rbio, bio);
1531 if (atomic_dec_and_test(&rbio->stripes_pending))
1532 wake_up(&rbio->io_wait);
1535 static void submit_read_wait_bio_list(struct btrfs_raid_bio *rbio,
1536 struct bio_list *bio_list)
1540 atomic_set(&rbio->stripes_pending, bio_list_size(bio_list));
1541 while ((bio = bio_list_pop(bio_list))) {
1542 bio->bi_end_io = raid_wait_read_end_io;
1544 if (trace_raid56_read_enabled()) {
1545 struct raid56_bio_trace_info trace_info = { 0 };
1547 bio_get_trace_info(rbio, bio, &trace_info);
1548 trace_raid56_read(rbio, bio, &trace_info);
1553 wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0);
1556 static int alloc_rbio_data_pages(struct btrfs_raid_bio *rbio)
1558 const int data_pages = rbio->nr_data * rbio->stripe_npages;
1561 ret = btrfs_alloc_page_array(data_pages, rbio->stripe_pages, 0);
1565 index_stripe_sectors(rbio);
1570 * We use plugging call backs to collect full stripes.
1571 * Any time we get a partial stripe write while plugged
1572 * we collect it into a list. When the unplug comes down,
1573 * we sort the list by logical block number and merge
1574 * everything we can into the same rbios
1576 struct btrfs_plug_cb {
1577 struct blk_plug_cb cb;
1578 struct btrfs_fs_info *info;
1579 struct list_head rbio_list;
1583 * rbios on the plug list are sorted for easier merging.
1585 static int plug_cmp(void *priv, const struct list_head *a,
1586 const struct list_head *b)
1588 const struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio,
1590 const struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio,
1592 u64 a_sector = ra->bio_list.head->bi_iter.bi_sector;
1593 u64 b_sector = rb->bio_list.head->bi_iter.bi_sector;
1595 if (a_sector < b_sector)
1597 if (a_sector > b_sector)
1602 static void raid_unplug(struct blk_plug_cb *cb, bool from_schedule)
1604 struct btrfs_plug_cb *plug = container_of(cb, struct btrfs_plug_cb, cb);
1605 struct btrfs_raid_bio *cur;
1606 struct btrfs_raid_bio *last = NULL;
1608 list_sort(NULL, &plug->rbio_list, plug_cmp);
1610 while (!list_empty(&plug->rbio_list)) {
1611 cur = list_entry(plug->rbio_list.next,
1612 struct btrfs_raid_bio, plug_list);
1613 list_del_init(&cur->plug_list);
1615 if (rbio_is_full(cur)) {
1616 /* We have a full stripe, queue it down. */
1617 start_async_work(cur, rmw_rbio_work);
1621 if (rbio_can_merge(last, cur)) {
1622 merge_rbio(last, cur);
1626 start_async_work(last, rmw_rbio_work);
1631 start_async_work(last, rmw_rbio_work);
1635 /* Add the original bio into rbio->bio_list, and update rbio::dbitmap. */
1636 static void rbio_add_bio(struct btrfs_raid_bio *rbio, struct bio *orig_bio)
1638 const struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1639 const u64 orig_logical = orig_bio->bi_iter.bi_sector << SECTOR_SHIFT;
1640 const u64 full_stripe_start = rbio->bioc->full_stripe_logical;
1641 const u32 orig_len = orig_bio->bi_iter.bi_size;
1642 const u32 sectorsize = fs_info->sectorsize;
1645 ASSERT(orig_logical >= full_stripe_start &&
1646 orig_logical + orig_len <= full_stripe_start +
1647 rbio->nr_data * BTRFS_STRIPE_LEN);
1649 bio_list_add(&rbio->bio_list, orig_bio);
1650 rbio->bio_list_bytes += orig_bio->bi_iter.bi_size;
1652 /* Update the dbitmap. */
1653 for (cur_logical = orig_logical; cur_logical < orig_logical + orig_len;
1654 cur_logical += sectorsize) {
1655 int bit = ((u32)(cur_logical - full_stripe_start) >>
1656 fs_info->sectorsize_bits) % rbio->stripe_nsectors;
1658 set_bit(bit, &rbio->dbitmap);
1663 * our main entry point for writes from the rest of the FS.
1665 void raid56_parity_write(struct bio *bio, struct btrfs_io_context *bioc)
1667 struct btrfs_fs_info *fs_info = bioc->fs_info;
1668 struct btrfs_raid_bio *rbio;
1669 struct btrfs_plug_cb *plug = NULL;
1670 struct blk_plug_cb *cb;
1672 rbio = alloc_rbio(fs_info, bioc);
1674 bio->bi_status = errno_to_blk_status(PTR_ERR(rbio));
1678 rbio->operation = BTRFS_RBIO_WRITE;
1679 rbio_add_bio(rbio, bio);
1682 * Don't plug on full rbios, just get them out the door
1683 * as quickly as we can
1685 if (!rbio_is_full(rbio)) {
1686 cb = blk_check_plugged(raid_unplug, fs_info, sizeof(*plug));
1688 plug = container_of(cb, struct btrfs_plug_cb, cb);
1690 plug->info = fs_info;
1691 INIT_LIST_HEAD(&plug->rbio_list);
1693 list_add_tail(&rbio->plug_list, &plug->rbio_list);
1699 * Either we don't have any existing plug, or we're doing a full stripe,
1700 * queue the rmw work now.
1702 start_async_work(rbio, rmw_rbio_work);
1705 static int verify_one_sector(struct btrfs_raid_bio *rbio,
1706 int stripe_nr, int sector_nr)
1708 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1709 struct sector_ptr *sector;
1710 u8 csum_buf[BTRFS_CSUM_SIZE];
1714 if (!rbio->csum_bitmap || !rbio->csum_buf)
1717 /* No way to verify P/Q as they are not covered by data csum. */
1718 if (stripe_nr >= rbio->nr_data)
1721 * If we're rebuilding a read, we have to use pages from the
1722 * bio list if possible.
1724 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
1725 sector = sector_in_rbio(rbio, stripe_nr, sector_nr, 0);
1727 sector = rbio_stripe_sector(rbio, stripe_nr, sector_nr);
1730 ASSERT(sector->page);
1732 csum_expected = rbio->csum_buf +
1733 (stripe_nr * rbio->stripe_nsectors + sector_nr) *
1735 ret = btrfs_check_sector_csum(fs_info, sector->page, sector->pgoff,
1736 csum_buf, csum_expected);
1741 * Recover a vertical stripe specified by @sector_nr.
1742 * @*pointers are the pre-allocated pointers by the caller, so we don't
1743 * need to allocate/free the pointers again and again.
1745 static int recover_vertical(struct btrfs_raid_bio *rbio, int sector_nr,
1746 void **pointers, void **unmap_array)
1748 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1749 struct sector_ptr *sector;
1750 const u32 sectorsize = fs_info->sectorsize;
1758 * Now we just use bitmap to mark the horizontal stripes in
1759 * which we have data when doing parity scrub.
1761 if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB &&
1762 !test_bit(sector_nr, &rbio->dbitmap))
1765 found_errors = get_rbio_veritical_errors(rbio, sector_nr, &faila,
1768 * No errors in the vertical stripe, skip it. Can happen for recovery
1769 * which only part of a stripe failed csum check.
1774 if (found_errors > rbio->bioc->max_errors)
1778 * Setup our array of pointers with sectors from each stripe
1780 * NOTE: store a duplicate array of pointers to preserve the
1783 for (stripe_nr = 0; stripe_nr < rbio->real_stripes; stripe_nr++) {
1785 * If we're rebuilding a read, we have to use pages from the
1786 * bio list if possible.
1788 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
1789 sector = sector_in_rbio(rbio, stripe_nr, sector_nr, 0);
1791 sector = rbio_stripe_sector(rbio, stripe_nr, sector_nr);
1793 ASSERT(sector->page);
1794 pointers[stripe_nr] = kmap_local_page(sector->page) +
1796 unmap_array[stripe_nr] = pointers[stripe_nr];
1799 /* All raid6 handling here */
1800 if (rbio->bioc->map_type & BTRFS_BLOCK_GROUP_RAID6) {
1801 /* Single failure, rebuild from parity raid5 style */
1803 if (faila == rbio->nr_data)
1805 * Just the P stripe has failed, without
1806 * a bad data or Q stripe.
1807 * We have nothing to do, just skip the
1808 * recovery for this stripe.
1812 * a single failure in raid6 is rebuilt
1813 * in the pstripe code below
1819 * If the q stripe is failed, do a pstripe reconstruction from
1821 * If both the q stripe and the P stripe are failed, we're
1822 * here due to a crc mismatch and we can't give them the
1825 if (failb == rbio->real_stripes - 1) {
1826 if (faila == rbio->real_stripes - 2)
1828 * Only P and Q are corrupted.
1829 * We only care about data stripes recovery,
1830 * can skip this vertical stripe.
1834 * Otherwise we have one bad data stripe and
1835 * a good P stripe. raid5!
1840 if (failb == rbio->real_stripes - 2) {
1841 raid6_datap_recov(rbio->real_stripes, sectorsize,
1844 raid6_2data_recov(rbio->real_stripes, sectorsize,
1845 faila, failb, pointers);
1850 /* Rebuild from P stripe here (raid5 or raid6). */
1851 ASSERT(failb == -1);
1853 /* Copy parity block into failed block to start with */
1854 memcpy(pointers[faila], pointers[rbio->nr_data], sectorsize);
1856 /* Rearrange the pointer array */
1857 p = pointers[faila];
1858 for (stripe_nr = faila; stripe_nr < rbio->nr_data - 1;
1860 pointers[stripe_nr] = pointers[stripe_nr + 1];
1861 pointers[rbio->nr_data - 1] = p;
1863 /* Xor in the rest */
1864 run_xor(pointers, rbio->nr_data - 1, sectorsize);
1869 * No matter if this is a RMW or recovery, we should have all
1870 * failed sectors repaired in the vertical stripe, thus they are now
1872 * Especially if we determine to cache the rbio, we need to
1873 * have at least all data sectors uptodate.
1875 * If possible, also check if the repaired sector matches its data
1879 ret = verify_one_sector(rbio, faila, sector_nr);
1883 sector = rbio_stripe_sector(rbio, faila, sector_nr);
1884 sector->uptodate = 1;
1887 ret = verify_one_sector(rbio, failb, sector_nr);
1891 sector = rbio_stripe_sector(rbio, failb, sector_nr);
1892 sector->uptodate = 1;
1896 for (stripe_nr = rbio->real_stripes - 1; stripe_nr >= 0; stripe_nr--)
1897 kunmap_local(unmap_array[stripe_nr]);
1901 static int recover_sectors(struct btrfs_raid_bio *rbio)
1903 void **pointers = NULL;
1904 void **unmap_array = NULL;
1909 * @pointers array stores the pointer for each sector.
1911 * @unmap_array stores copy of pointers that does not get reordered
1912 * during reconstruction so that kunmap_local works.
1914 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
1915 unmap_array = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
1916 if (!pointers || !unmap_array) {
1921 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
1922 spin_lock(&rbio->bio_list_lock);
1923 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1924 spin_unlock(&rbio->bio_list_lock);
1927 index_rbio_pages(rbio);
1929 for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++) {
1930 ret = recover_vertical(rbio, sectornr, pointers, unmap_array);
1941 static void recover_rbio(struct btrfs_raid_bio *rbio)
1943 struct bio_list bio_list = BIO_EMPTY_LIST;
1944 int total_sector_nr;
1948 * Either we're doing recover for a read failure or degraded write,
1949 * caller should have set error bitmap correctly.
1951 ASSERT(bitmap_weight(rbio->error_bitmap, rbio->nr_sectors));
1953 /* For recovery, we need to read all sectors including P/Q. */
1954 ret = alloc_rbio_pages(rbio);
1958 index_rbio_pages(rbio);
1961 * Read everything that hasn't failed. However this time we will
1962 * not trust any cached sector.
1963 * As we may read out some stale data but higher layer is not reading
1966 * So here we always re-read everything in recovery path.
1968 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
1969 total_sector_nr++) {
1970 int stripe = total_sector_nr / rbio->stripe_nsectors;
1971 int sectornr = total_sector_nr % rbio->stripe_nsectors;
1972 struct sector_ptr *sector;
1975 * Skip the range which has error. It can be a range which is
1976 * marked error (for csum mismatch), or it can be a missing
1979 if (!rbio->bioc->stripes[stripe].dev->bdev ||
1980 test_bit(total_sector_nr, rbio->error_bitmap)) {
1982 * Also set the error bit for missing device, which
1983 * may not yet have its error bit set.
1985 set_bit(total_sector_nr, rbio->error_bitmap);
1989 sector = rbio_stripe_sector(rbio, stripe, sectornr);
1990 ret = rbio_add_io_sector(rbio, &bio_list, sector, stripe,
1991 sectornr, REQ_OP_READ);
1993 bio_list_put(&bio_list);
1998 submit_read_wait_bio_list(rbio, &bio_list);
1999 ret = recover_sectors(rbio);
2001 rbio_orig_end_io(rbio, errno_to_blk_status(ret));
2004 static void recover_rbio_work(struct work_struct *work)
2006 struct btrfs_raid_bio *rbio;
2008 rbio = container_of(work, struct btrfs_raid_bio, work);
2009 if (!lock_stripe_add(rbio))
2013 static void recover_rbio_work_locked(struct work_struct *work)
2015 recover_rbio(container_of(work, struct btrfs_raid_bio, work));
2018 static void set_rbio_raid6_extra_error(struct btrfs_raid_bio *rbio, int mirror_num)
2024 * This is for RAID6 extra recovery tries, thus mirror number should
2026 * Mirror 1 means read from data stripes. Mirror 2 means rebuild using
2029 ASSERT(mirror_num > 2);
2030 for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) {
2035 found_errors = get_rbio_veritical_errors(rbio, sector_nr,
2037 /* This vertical stripe doesn't have errors. */
2042 * If we found errors, there should be only one error marked
2043 * by previous set_rbio_range_error().
2045 ASSERT(found_errors == 1);
2048 /* Now select another stripe to mark as error. */
2049 failb = rbio->real_stripes - (mirror_num - 1);
2053 /* Set the extra bit in error bitmap. */
2055 set_bit(failb * rbio->stripe_nsectors + sector_nr,
2056 rbio->error_bitmap);
2059 /* We should found at least one vertical stripe with error.*/
2064 * the main entry point for reads from the higher layers. This
2065 * is really only called when the normal read path had a failure,
2066 * so we assume the bio they send down corresponds to a failed part
2069 void raid56_parity_recover(struct bio *bio, struct btrfs_io_context *bioc,
2072 struct btrfs_fs_info *fs_info = bioc->fs_info;
2073 struct btrfs_raid_bio *rbio;
2075 rbio = alloc_rbio(fs_info, bioc);
2077 bio->bi_status = errno_to_blk_status(PTR_ERR(rbio));
2082 rbio->operation = BTRFS_RBIO_READ_REBUILD;
2083 rbio_add_bio(rbio, bio);
2085 set_rbio_range_error(rbio, bio);
2089 * for 'mirror == 2', reconstruct from all other stripes.
2090 * for 'mirror_num > 2', select a stripe to fail on every retry.
2093 set_rbio_raid6_extra_error(rbio, mirror_num);
2095 start_async_work(rbio, recover_rbio_work);
2098 static void fill_data_csums(struct btrfs_raid_bio *rbio)
2100 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
2101 struct btrfs_root *csum_root = btrfs_csum_root(fs_info,
2102 rbio->bioc->full_stripe_logical);
2103 const u64 start = rbio->bioc->full_stripe_logical;
2104 const u32 len = (rbio->nr_data * rbio->stripe_nsectors) <<
2105 fs_info->sectorsize_bits;
2108 /* The rbio should not have its csum buffer initialized. */
2109 ASSERT(!rbio->csum_buf && !rbio->csum_bitmap);
2112 * Skip the csum search if:
2114 * - The rbio doesn't belong to data block groups
2115 * Then we are doing IO for tree blocks, no need to search csums.
2117 * - The rbio belongs to mixed block groups
2118 * This is to avoid deadlock, as we're already holding the full
2119 * stripe lock, if we trigger a metadata read, and it needs to do
2120 * raid56 recovery, we will deadlock.
2122 if (!(rbio->bioc->map_type & BTRFS_BLOCK_GROUP_DATA) ||
2123 rbio->bioc->map_type & BTRFS_BLOCK_GROUP_METADATA)
2126 rbio->csum_buf = kzalloc(rbio->nr_data * rbio->stripe_nsectors *
2127 fs_info->csum_size, GFP_NOFS);
2128 rbio->csum_bitmap = bitmap_zalloc(rbio->nr_data * rbio->stripe_nsectors,
2130 if (!rbio->csum_buf || !rbio->csum_bitmap) {
2135 ret = btrfs_lookup_csums_bitmap(csum_root, NULL, start, start + len - 1,
2136 rbio->csum_buf, rbio->csum_bitmap);
2139 if (bitmap_empty(rbio->csum_bitmap, len >> fs_info->sectorsize_bits))
2145 * We failed to allocate memory or grab the csum, but it's not fatal,
2146 * we can still continue. But better to warn users that RMW is no
2147 * longer safe for this particular sub-stripe write.
2149 btrfs_warn_rl(fs_info,
2150 "sub-stripe write for full stripe %llu is not safe, failed to get csum: %d",
2151 rbio->bioc->full_stripe_logical, ret);
2153 kfree(rbio->csum_buf);
2154 bitmap_free(rbio->csum_bitmap);
2155 rbio->csum_buf = NULL;
2156 rbio->csum_bitmap = NULL;
2159 static int rmw_read_wait_recover(struct btrfs_raid_bio *rbio)
2161 struct bio_list bio_list = BIO_EMPTY_LIST;
2162 int total_sector_nr;
2166 * Fill the data csums we need for data verification. We need to fill
2167 * the csum_bitmap/csum_buf first, as our endio function will try to
2168 * verify the data sectors.
2170 fill_data_csums(rbio);
2173 * Build a list of bios to read all sectors (including data and P/Q).
2175 * This behavior is to compensate the later csum verification and recovery.
2177 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
2178 total_sector_nr++) {
2179 struct sector_ptr *sector;
2180 int stripe = total_sector_nr / rbio->stripe_nsectors;
2181 int sectornr = total_sector_nr % rbio->stripe_nsectors;
2183 sector = rbio_stripe_sector(rbio, stripe, sectornr);
2184 ret = rbio_add_io_sector(rbio, &bio_list, sector,
2185 stripe, sectornr, REQ_OP_READ);
2187 bio_list_put(&bio_list);
2193 * We may or may not have any corrupted sectors (including missing dev
2194 * and csum mismatch), just let recover_sectors() to handle them all.
2196 submit_read_wait_bio_list(rbio, &bio_list);
2197 return recover_sectors(rbio);
2200 static void raid_wait_write_end_io(struct bio *bio)
2202 struct btrfs_raid_bio *rbio = bio->bi_private;
2203 blk_status_t err = bio->bi_status;
2206 rbio_update_error_bitmap(rbio, bio);
2208 if (atomic_dec_and_test(&rbio->stripes_pending))
2209 wake_up(&rbio->io_wait);
2212 static void submit_write_bios(struct btrfs_raid_bio *rbio,
2213 struct bio_list *bio_list)
2217 atomic_set(&rbio->stripes_pending, bio_list_size(bio_list));
2218 while ((bio = bio_list_pop(bio_list))) {
2219 bio->bi_end_io = raid_wait_write_end_io;
2221 if (trace_raid56_write_enabled()) {
2222 struct raid56_bio_trace_info trace_info = { 0 };
2224 bio_get_trace_info(rbio, bio, &trace_info);
2225 trace_raid56_write(rbio, bio, &trace_info);
2232 * To determine if we need to read any sector from the disk.
2233 * Should only be utilized in RMW path, to skip cached rbio.
2235 static bool need_read_stripe_sectors(struct btrfs_raid_bio *rbio)
2239 for (i = 0; i < rbio->nr_data * rbio->stripe_nsectors; i++) {
2240 struct sector_ptr *sector = &rbio->stripe_sectors[i];
2243 * We have a sector which doesn't have page nor uptodate,
2244 * thus this rbio can not be cached one, as cached one must
2245 * have all its data sectors present and uptodate.
2247 if (!sector->page || !sector->uptodate)
2253 static void rmw_rbio(struct btrfs_raid_bio *rbio)
2255 struct bio_list bio_list;
2260 * Allocate the pages for parity first, as P/Q pages will always be
2261 * needed for both full-stripe and sub-stripe writes.
2263 ret = alloc_rbio_parity_pages(rbio);
2268 * Either full stripe write, or we have every data sector already
2269 * cached, can go to write path immediately.
2271 if (!rbio_is_full(rbio) && need_read_stripe_sectors(rbio)) {
2273 * Now we're doing sub-stripe write, also need all data stripes
2274 * to do the full RMW.
2276 ret = alloc_rbio_data_pages(rbio);
2280 index_rbio_pages(rbio);
2282 ret = rmw_read_wait_recover(rbio);
2288 * At this stage we're not allowed to add any new bios to the
2289 * bio list any more, anyone else that wants to change this stripe
2290 * needs to do their own rmw.
2292 spin_lock(&rbio->bio_list_lock);
2293 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
2294 spin_unlock(&rbio->bio_list_lock);
2296 bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors);
2298 index_rbio_pages(rbio);
2301 * We don't cache full rbios because we're assuming
2302 * the higher layers are unlikely to use this area of
2303 * the disk again soon. If they do use it again,
2304 * hopefully they will send another full bio.
2306 if (!rbio_is_full(rbio))
2307 cache_rbio_pages(rbio);
2309 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2311 for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++)
2312 generate_pq_vertical(rbio, sectornr);
2314 bio_list_init(&bio_list);
2315 ret = rmw_assemble_write_bios(rbio, &bio_list);
2319 /* We should have at least one bio assembled. */
2320 ASSERT(bio_list_size(&bio_list));
2321 submit_write_bios(rbio, &bio_list);
2322 wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0);
2324 /* We may have more errors than our tolerance during the read. */
2325 for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++) {
2328 found_errors = get_rbio_veritical_errors(rbio, sectornr, NULL, NULL);
2329 if (found_errors > rbio->bioc->max_errors) {
2335 rbio_orig_end_io(rbio, errno_to_blk_status(ret));
2338 static void rmw_rbio_work(struct work_struct *work)
2340 struct btrfs_raid_bio *rbio;
2342 rbio = container_of(work, struct btrfs_raid_bio, work);
2343 if (lock_stripe_add(rbio) == 0)
2347 static void rmw_rbio_work_locked(struct work_struct *work)
2349 rmw_rbio(container_of(work, struct btrfs_raid_bio, work));
2353 * The following code is used to scrub/replace the parity stripe
2355 * Caller must have already increased bio_counter for getting @bioc.
2357 * Note: We need make sure all the pages that add into the scrub/replace
2358 * raid bio are correct and not be changed during the scrub/replace. That
2359 * is those pages just hold metadata or file data with checksum.
2362 struct btrfs_raid_bio *raid56_parity_alloc_scrub_rbio(struct bio *bio,
2363 struct btrfs_io_context *bioc,
2364 struct btrfs_device *scrub_dev,
2365 unsigned long *dbitmap, int stripe_nsectors)
2367 struct btrfs_fs_info *fs_info = bioc->fs_info;
2368 struct btrfs_raid_bio *rbio;
2371 rbio = alloc_rbio(fs_info, bioc);
2374 bio_list_add(&rbio->bio_list, bio);
2376 * This is a special bio which is used to hold the completion handler
2377 * and make the scrub rbio is similar to the other types
2379 ASSERT(!bio->bi_iter.bi_size);
2380 rbio->operation = BTRFS_RBIO_PARITY_SCRUB;
2383 * After mapping bioc with BTRFS_MAP_WRITE, parities have been sorted
2384 * to the end position, so this search can start from the first parity
2387 for (i = rbio->nr_data; i < rbio->real_stripes; i++) {
2388 if (bioc->stripes[i].dev == scrub_dev) {
2393 ASSERT(i < rbio->real_stripes);
2395 bitmap_copy(&rbio->dbitmap, dbitmap, stripe_nsectors);
2400 * We just scrub the parity that we have correct data on the same horizontal,
2401 * so we needn't allocate all pages for all the stripes.
2403 static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio)
2405 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
2406 int total_sector_nr;
2408 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
2409 total_sector_nr++) {
2411 int sectornr = total_sector_nr % rbio->stripe_nsectors;
2412 int index = (total_sector_nr * sectorsize) >> PAGE_SHIFT;
2414 if (!test_bit(sectornr, &rbio->dbitmap))
2416 if (rbio->stripe_pages[index])
2418 page = alloc_page(GFP_NOFS);
2421 rbio->stripe_pages[index] = page;
2423 index_stripe_sectors(rbio);
2427 static int finish_parity_scrub(struct btrfs_raid_bio *rbio)
2429 struct btrfs_io_context *bioc = rbio->bioc;
2430 const u32 sectorsize = bioc->fs_info->sectorsize;
2431 void **pointers = rbio->finish_pointers;
2432 unsigned long *pbitmap = &rbio->finish_pbitmap;
2433 int nr_data = rbio->nr_data;
2437 struct sector_ptr p_sector = { 0 };
2438 struct sector_ptr q_sector = { 0 };
2439 struct bio_list bio_list;
2443 bio_list_init(&bio_list);
2445 if (rbio->real_stripes - rbio->nr_data == 1)
2446 has_qstripe = false;
2447 else if (rbio->real_stripes - rbio->nr_data == 2)
2453 * Replace is running and our P/Q stripe is being replaced, then we
2454 * need to duplicate the final write to replace target.
2456 if (bioc->replace_nr_stripes && bioc->replace_stripe_src == rbio->scrubp) {
2458 bitmap_copy(pbitmap, &rbio->dbitmap, rbio->stripe_nsectors);
2462 * Because the higher layers(scrubber) are unlikely to
2463 * use this area of the disk again soon, so don't cache
2466 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2468 p_sector.page = alloc_page(GFP_NOFS);
2472 p_sector.uptodate = 1;
2475 /* RAID6, allocate and map temp space for the Q stripe */
2476 q_sector.page = alloc_page(GFP_NOFS);
2477 if (!q_sector.page) {
2478 __free_page(p_sector.page);
2479 p_sector.page = NULL;
2483 q_sector.uptodate = 1;
2484 pointers[rbio->real_stripes - 1] = kmap_local_page(q_sector.page);
2487 bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors);
2489 /* Map the parity stripe just once */
2490 pointers[nr_data] = kmap_local_page(p_sector.page);
2492 for_each_set_bit(sectornr, &rbio->dbitmap, rbio->stripe_nsectors) {
2493 struct sector_ptr *sector;
2496 /* first collect one page from each data stripe */
2497 for (stripe = 0; stripe < nr_data; stripe++) {
2498 sector = sector_in_rbio(rbio, stripe, sectornr, 0);
2499 pointers[stripe] = kmap_local_page(sector->page) +
2505 /* RAID6, call the library function to fill in our P/Q */
2506 raid6_call.gen_syndrome(rbio->real_stripes, sectorsize,
2510 memcpy(pointers[nr_data], pointers[0], sectorsize);
2511 run_xor(pointers + 1, nr_data - 1, sectorsize);
2514 /* Check scrubbing parity and repair it */
2515 sector = rbio_stripe_sector(rbio, rbio->scrubp, sectornr);
2516 parity = kmap_local_page(sector->page) + sector->pgoff;
2517 if (memcmp(parity, pointers[rbio->scrubp], sectorsize) != 0)
2518 memcpy(parity, pointers[rbio->scrubp], sectorsize);
2520 /* Parity is right, needn't writeback */
2521 bitmap_clear(&rbio->dbitmap, sectornr, 1);
2522 kunmap_local(parity);
2524 for (stripe = nr_data - 1; stripe >= 0; stripe--)
2525 kunmap_local(pointers[stripe]);
2528 kunmap_local(pointers[nr_data]);
2529 __free_page(p_sector.page);
2530 p_sector.page = NULL;
2531 if (q_sector.page) {
2532 kunmap_local(pointers[rbio->real_stripes - 1]);
2533 __free_page(q_sector.page);
2534 q_sector.page = NULL;
2538 * time to start writing. Make bios for everything from the
2539 * higher layers (the bio_list in our rbio) and our p/q. Ignore
2542 for_each_set_bit(sectornr, &rbio->dbitmap, rbio->stripe_nsectors) {
2543 struct sector_ptr *sector;
2545 sector = rbio_stripe_sector(rbio, rbio->scrubp, sectornr);
2546 ret = rbio_add_io_sector(rbio, &bio_list, sector, rbio->scrubp,
2547 sectornr, REQ_OP_WRITE);
2556 * Replace is running and our parity stripe needs to be duplicated to
2557 * the target device. Check we have a valid source stripe number.
2559 ASSERT(rbio->bioc->replace_stripe_src >= 0);
2560 for_each_set_bit(sectornr, pbitmap, rbio->stripe_nsectors) {
2561 struct sector_ptr *sector;
2563 sector = rbio_stripe_sector(rbio, rbio->scrubp, sectornr);
2564 ret = rbio_add_io_sector(rbio, &bio_list, sector,
2566 sectornr, REQ_OP_WRITE);
2572 submit_write_bios(rbio, &bio_list);
2576 bio_list_put(&bio_list);
2580 static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe)
2582 if (stripe >= 0 && stripe < rbio->nr_data)
2587 static int recover_scrub_rbio(struct btrfs_raid_bio *rbio)
2589 void **pointers = NULL;
2590 void **unmap_array = NULL;
2595 * @pointers array stores the pointer for each sector.
2597 * @unmap_array stores copy of pointers that does not get reordered
2598 * during reconstruction so that kunmap_local works.
2600 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
2601 unmap_array = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
2602 if (!pointers || !unmap_array) {
2607 for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) {
2608 int dfail = 0, failp = -1;
2613 found_errors = get_rbio_veritical_errors(rbio, sector_nr,
2615 if (found_errors > rbio->bioc->max_errors) {
2619 if (found_errors == 0)
2622 /* We should have at least one error here. */
2623 ASSERT(faila >= 0 || failb >= 0);
2625 if (is_data_stripe(rbio, faila))
2627 else if (is_parity_stripe(faila))
2630 if (is_data_stripe(rbio, failb))
2632 else if (is_parity_stripe(failb))
2635 * Because we can not use a scrubbing parity to repair the
2636 * data, so the capability of the repair is declined. (In the
2637 * case of RAID5, we can not repair anything.)
2639 if (dfail > rbio->bioc->max_errors - 1) {
2644 * If all data is good, only parity is correctly, just repair
2645 * the parity, no need to recover data stripes.
2651 * Here means we got one corrupted data stripe and one
2652 * corrupted parity on RAID6, if the corrupted parity is
2653 * scrubbing parity, luckily, use the other one to repair the
2654 * data, or we can not repair the data stripe.
2656 if (failp != rbio->scrubp) {
2661 ret = recover_vertical(rbio, sector_nr, pointers, unmap_array);
2671 static int scrub_assemble_read_bios(struct btrfs_raid_bio *rbio)
2673 struct bio_list bio_list = BIO_EMPTY_LIST;
2674 int total_sector_nr;
2677 /* Build a list of bios to read all the missing parts. */
2678 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
2679 total_sector_nr++) {
2680 int sectornr = total_sector_nr % rbio->stripe_nsectors;
2681 int stripe = total_sector_nr / rbio->stripe_nsectors;
2682 struct sector_ptr *sector;
2684 /* No data in the vertical stripe, no need to read. */
2685 if (!test_bit(sectornr, &rbio->dbitmap))
2689 * We want to find all the sectors missing from the rbio and
2690 * read them from the disk. If sector_in_rbio() finds a sector
2691 * in the bio list we don't need to read it off the stripe.
2693 sector = sector_in_rbio(rbio, stripe, sectornr, 1);
2697 sector = rbio_stripe_sector(rbio, stripe, sectornr);
2699 * The bio cache may have handed us an uptodate sector. If so,
2702 if (sector->uptodate)
2705 ret = rbio_add_io_sector(rbio, &bio_list, sector, stripe,
2706 sectornr, REQ_OP_READ);
2708 bio_list_put(&bio_list);
2713 submit_read_wait_bio_list(rbio, &bio_list);
2717 static void scrub_rbio(struct btrfs_raid_bio *rbio)
2722 ret = alloc_rbio_essential_pages(rbio);
2726 bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors);
2728 ret = scrub_assemble_read_bios(rbio);
2732 /* We may have some failures, recover the failed sectors first. */
2733 ret = recover_scrub_rbio(rbio);
2738 * We have every sector properly prepared. Can finish the scrub
2739 * and writeback the good content.
2741 ret = finish_parity_scrub(rbio);
2742 wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0);
2743 for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) {
2746 found_errors = get_rbio_veritical_errors(rbio, sector_nr, NULL, NULL);
2747 if (found_errors > rbio->bioc->max_errors) {
2753 rbio_orig_end_io(rbio, errno_to_blk_status(ret));
2756 static void scrub_rbio_work_locked(struct work_struct *work)
2758 scrub_rbio(container_of(work, struct btrfs_raid_bio, work));
2761 void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio)
2763 if (!lock_stripe_add(rbio))
2764 start_async_work(rbio, scrub_rbio_work_locked);
2768 * This is for scrub call sites where we already have correct data contents.
2769 * This allows us to avoid reading data stripes again.
2771 * Unfortunately here we have to do page copy, other than reusing the pages.
2772 * This is due to the fact rbio has its own page management for its cache.
2774 void raid56_parity_cache_data_pages(struct btrfs_raid_bio *rbio,
2775 struct page **data_pages, u64 data_logical)
2777 const u64 offset_in_full_stripe = data_logical -
2778 rbio->bioc->full_stripe_logical;
2779 const int page_index = offset_in_full_stripe >> PAGE_SHIFT;
2780 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
2781 const u32 sectors_per_page = PAGE_SIZE / sectorsize;
2785 * If we hit ENOMEM temporarily, but later at
2786 * raid56_parity_submit_scrub_rbio() time it succeeded, we just do
2787 * the extra read, not a big deal.
2789 * If we hit ENOMEM later at raid56_parity_submit_scrub_rbio() time,
2790 * the bio would got proper error number set.
2792 ret = alloc_rbio_data_pages(rbio);
2796 /* data_logical must be at stripe boundary and inside the full stripe. */
2797 ASSERT(IS_ALIGNED(offset_in_full_stripe, BTRFS_STRIPE_LEN));
2798 ASSERT(offset_in_full_stripe < (rbio->nr_data << BTRFS_STRIPE_LEN_SHIFT));
2800 for (int page_nr = 0; page_nr < (BTRFS_STRIPE_LEN >> PAGE_SHIFT); page_nr++) {
2801 struct page *dst = rbio->stripe_pages[page_nr + page_index];
2802 struct page *src = data_pages[page_nr];
2804 memcpy_page(dst, 0, src, 0, PAGE_SIZE);
2805 for (int sector_nr = sectors_per_page * page_index;
2806 sector_nr < sectors_per_page * (page_index + 1);
2808 rbio->stripe_sectors[sector_nr].uptodate = true;