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>
22 #include "async-thread.h"
23 #include "file-item.h"
24 #include "btrfs_inode.h"
26 /* set when additional merges to this rbio are not allowed */
27 #define RBIO_RMW_LOCKED_BIT 1
30 * set when this rbio is sitting in the hash, but it is just a cache
33 #define RBIO_CACHE_BIT 2
36 * set when it is safe to trust the stripe_pages for caching
38 #define RBIO_CACHE_READY_BIT 3
40 #define RBIO_CACHE_SIZE 1024
42 #define BTRFS_STRIPE_HASH_TABLE_BITS 11
44 /* Used by the raid56 code to lock stripes for read/modify/write */
45 struct btrfs_stripe_hash {
46 struct list_head hash_list;
50 /* Used by the raid56 code to lock stripes for read/modify/write */
51 struct btrfs_stripe_hash_table {
52 struct list_head stripe_cache;
53 spinlock_t cache_lock;
55 struct btrfs_stripe_hash table[];
59 * A bvec like structure to present a sector inside a page.
61 * Unlike bvec we don't need bvlen, as it's fixed to sectorsize.
65 unsigned int pgoff:24;
66 unsigned int uptodate:8;
69 static void rmw_rbio_work(struct work_struct *work);
70 static void rmw_rbio_work_locked(struct work_struct *work);
71 static void index_rbio_pages(struct btrfs_raid_bio *rbio);
72 static int alloc_rbio_pages(struct btrfs_raid_bio *rbio);
74 static int finish_parity_scrub(struct btrfs_raid_bio *rbio);
75 static void scrub_rbio_work_locked(struct work_struct *work);
77 static void free_raid_bio_pointers(struct btrfs_raid_bio *rbio)
79 bitmap_free(rbio->error_bitmap);
80 kfree(rbio->stripe_pages);
81 kfree(rbio->bio_sectors);
82 kfree(rbio->stripe_sectors);
83 kfree(rbio->finish_pointers);
86 static void free_raid_bio(struct btrfs_raid_bio *rbio)
90 if (!refcount_dec_and_test(&rbio->refs))
93 WARN_ON(!list_empty(&rbio->stripe_cache));
94 WARN_ON(!list_empty(&rbio->hash_list));
95 WARN_ON(!bio_list_empty(&rbio->bio_list));
97 for (i = 0; i < rbio->nr_pages; i++) {
98 if (rbio->stripe_pages[i]) {
99 __free_page(rbio->stripe_pages[i]);
100 rbio->stripe_pages[i] = NULL;
104 btrfs_put_bioc(rbio->bioc);
105 free_raid_bio_pointers(rbio);
109 static void start_async_work(struct btrfs_raid_bio *rbio, work_func_t work_func)
111 INIT_WORK(&rbio->work, work_func);
112 queue_work(rbio->bioc->fs_info->rmw_workers, &rbio->work);
116 * the stripe hash table is used for locking, and to collect
117 * bios in hopes of making a full stripe
119 int btrfs_alloc_stripe_hash_table(struct btrfs_fs_info *info)
121 struct btrfs_stripe_hash_table *table;
122 struct btrfs_stripe_hash_table *x;
123 struct btrfs_stripe_hash *cur;
124 struct btrfs_stripe_hash *h;
125 int num_entries = 1 << BTRFS_STRIPE_HASH_TABLE_BITS;
128 if (info->stripe_hash_table)
132 * The table is large, starting with order 4 and can go as high as
133 * order 7 in case lock debugging is turned on.
135 * Try harder to allocate and fallback to vmalloc to lower the chance
136 * of a failing mount.
138 table = kvzalloc(struct_size(table, table, num_entries), GFP_KERNEL);
142 spin_lock_init(&table->cache_lock);
143 INIT_LIST_HEAD(&table->stripe_cache);
147 for (i = 0; i < num_entries; i++) {
149 INIT_LIST_HEAD(&cur->hash_list);
150 spin_lock_init(&cur->lock);
153 x = cmpxchg(&info->stripe_hash_table, NULL, table);
159 * caching an rbio means to copy anything from the
160 * bio_sectors array into the stripe_pages array. We
161 * use the page uptodate bit in the stripe cache array
162 * to indicate if it has valid data
164 * once the caching is done, we set the cache ready
167 static void cache_rbio_pages(struct btrfs_raid_bio *rbio)
172 ret = alloc_rbio_pages(rbio);
176 for (i = 0; i < rbio->nr_sectors; i++) {
177 /* Some range not covered by bio (partial write), skip it */
178 if (!rbio->bio_sectors[i].page) {
180 * Even if the sector is not covered by bio, if it is
181 * a data sector it should still be uptodate as it is
184 if (i < rbio->nr_data * rbio->stripe_nsectors)
185 ASSERT(rbio->stripe_sectors[i].uptodate);
189 ASSERT(rbio->stripe_sectors[i].page);
190 memcpy_page(rbio->stripe_sectors[i].page,
191 rbio->stripe_sectors[i].pgoff,
192 rbio->bio_sectors[i].page,
193 rbio->bio_sectors[i].pgoff,
194 rbio->bioc->fs_info->sectorsize);
195 rbio->stripe_sectors[i].uptodate = 1;
197 set_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
201 * we hash on the first logical address of the stripe
203 static int rbio_bucket(struct btrfs_raid_bio *rbio)
205 u64 num = rbio->bioc->full_stripe_logical;
208 * we shift down quite a bit. We're using byte
209 * addressing, and most of the lower bits are zeros.
210 * This tends to upset hash_64, and it consistently
211 * returns just one or two different values.
213 * shifting off the lower bits fixes things.
215 return hash_64(num >> 16, BTRFS_STRIPE_HASH_TABLE_BITS);
218 static bool full_page_sectors_uptodate(struct btrfs_raid_bio *rbio,
219 unsigned int page_nr)
221 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
222 const u32 sectors_per_page = PAGE_SIZE / sectorsize;
225 ASSERT(page_nr < rbio->nr_pages);
227 for (i = sectors_per_page * page_nr;
228 i < sectors_per_page * page_nr + sectors_per_page;
230 if (!rbio->stripe_sectors[i].uptodate)
237 * Update the stripe_sectors[] array to use correct page and pgoff
239 * Should be called every time any page pointer in stripes_pages[] got modified.
241 static void index_stripe_sectors(struct btrfs_raid_bio *rbio)
243 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
247 for (i = 0, offset = 0; i < rbio->nr_sectors; i++, offset += sectorsize) {
248 int page_index = offset >> PAGE_SHIFT;
250 ASSERT(page_index < rbio->nr_pages);
251 rbio->stripe_sectors[i].page = rbio->stripe_pages[page_index];
252 rbio->stripe_sectors[i].pgoff = offset_in_page(offset);
256 static void steal_rbio_page(struct btrfs_raid_bio *src,
257 struct btrfs_raid_bio *dest, int page_nr)
259 const u32 sectorsize = src->bioc->fs_info->sectorsize;
260 const u32 sectors_per_page = PAGE_SIZE / sectorsize;
263 if (dest->stripe_pages[page_nr])
264 __free_page(dest->stripe_pages[page_nr]);
265 dest->stripe_pages[page_nr] = src->stripe_pages[page_nr];
266 src->stripe_pages[page_nr] = NULL;
268 /* Also update the sector->uptodate bits. */
269 for (i = sectors_per_page * page_nr;
270 i < sectors_per_page * page_nr + sectors_per_page; i++)
271 dest->stripe_sectors[i].uptodate = true;
274 static bool is_data_stripe_page(struct btrfs_raid_bio *rbio, int page_nr)
276 const int sector_nr = (page_nr << PAGE_SHIFT) >>
277 rbio->bioc->fs_info->sectorsize_bits;
280 * We have ensured PAGE_SIZE is aligned with sectorsize, thus
281 * we won't have a page which is half data half parity.
283 * Thus if the first sector of the page belongs to data stripes, then
284 * the full page belongs to data stripes.
286 return (sector_nr < rbio->nr_data * rbio->stripe_nsectors);
290 * Stealing an rbio means taking all the uptodate pages from the stripe array
291 * in the source rbio and putting them into the destination rbio.
293 * This will also update the involved stripe_sectors[] which are referring to
296 static void steal_rbio(struct btrfs_raid_bio *src, struct btrfs_raid_bio *dest)
300 if (!test_bit(RBIO_CACHE_READY_BIT, &src->flags))
303 for (i = 0; i < dest->nr_pages; i++) {
304 struct page *p = src->stripe_pages[i];
307 * We don't need to steal P/Q pages as they will always be
308 * regenerated for RMW or full write anyway.
310 if (!is_data_stripe_page(src, i))
314 * If @src already has RBIO_CACHE_READY_BIT, it should have
315 * all data stripe pages present and uptodate.
318 ASSERT(full_page_sectors_uptodate(src, i));
319 steal_rbio_page(src, dest, i);
321 index_stripe_sectors(dest);
322 index_stripe_sectors(src);
326 * merging means we take the bio_list from the victim and
327 * splice it into the destination. The victim should
328 * be discarded afterwards.
330 * must be called with dest->rbio_list_lock held
332 static void merge_rbio(struct btrfs_raid_bio *dest,
333 struct btrfs_raid_bio *victim)
335 bio_list_merge(&dest->bio_list, &victim->bio_list);
336 dest->bio_list_bytes += victim->bio_list_bytes;
337 /* Also inherit the bitmaps from @victim. */
338 bitmap_or(&dest->dbitmap, &victim->dbitmap, &dest->dbitmap,
339 dest->stripe_nsectors);
340 bio_list_init(&victim->bio_list);
344 * used to prune items that are in the cache. The caller
345 * must hold the hash table lock.
347 static void __remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
349 int bucket = rbio_bucket(rbio);
350 struct btrfs_stripe_hash_table *table;
351 struct btrfs_stripe_hash *h;
355 * check the bit again under the hash table lock.
357 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
360 table = rbio->bioc->fs_info->stripe_hash_table;
361 h = table->table + bucket;
363 /* hold the lock for the bucket because we may be
364 * removing it from the hash table
369 * hold the lock for the bio list because we need
370 * to make sure the bio list is empty
372 spin_lock(&rbio->bio_list_lock);
374 if (test_and_clear_bit(RBIO_CACHE_BIT, &rbio->flags)) {
375 list_del_init(&rbio->stripe_cache);
376 table->cache_size -= 1;
379 /* if the bio list isn't empty, this rbio is
380 * still involved in an IO. We take it out
381 * of the cache list, and drop the ref that
382 * was held for the list.
384 * If the bio_list was empty, we also remove
385 * the rbio from the hash_table, and drop
386 * the corresponding ref
388 if (bio_list_empty(&rbio->bio_list)) {
389 if (!list_empty(&rbio->hash_list)) {
390 list_del_init(&rbio->hash_list);
391 refcount_dec(&rbio->refs);
392 BUG_ON(!list_empty(&rbio->plug_list));
397 spin_unlock(&rbio->bio_list_lock);
398 spin_unlock(&h->lock);
405 * prune a given rbio from the cache
407 static void remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
409 struct btrfs_stripe_hash_table *table;
411 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
414 table = rbio->bioc->fs_info->stripe_hash_table;
416 spin_lock(&table->cache_lock);
417 __remove_rbio_from_cache(rbio);
418 spin_unlock(&table->cache_lock);
422 * remove everything in the cache
424 static void btrfs_clear_rbio_cache(struct btrfs_fs_info *info)
426 struct btrfs_stripe_hash_table *table;
427 struct btrfs_raid_bio *rbio;
429 table = info->stripe_hash_table;
431 spin_lock(&table->cache_lock);
432 while (!list_empty(&table->stripe_cache)) {
433 rbio = list_entry(table->stripe_cache.next,
434 struct btrfs_raid_bio,
436 __remove_rbio_from_cache(rbio);
438 spin_unlock(&table->cache_lock);
442 * remove all cached entries and free the hash table
445 void btrfs_free_stripe_hash_table(struct btrfs_fs_info *info)
447 if (!info->stripe_hash_table)
449 btrfs_clear_rbio_cache(info);
450 kvfree(info->stripe_hash_table);
451 info->stripe_hash_table = NULL;
455 * insert an rbio into the stripe cache. It
456 * must have already been prepared by calling
459 * If this rbio was already cached, it gets
460 * moved to the front of the lru.
462 * If the size of the rbio cache is too big, we
465 static void cache_rbio(struct btrfs_raid_bio *rbio)
467 struct btrfs_stripe_hash_table *table;
469 if (!test_bit(RBIO_CACHE_READY_BIT, &rbio->flags))
472 table = rbio->bioc->fs_info->stripe_hash_table;
474 spin_lock(&table->cache_lock);
475 spin_lock(&rbio->bio_list_lock);
477 /* bump our ref if we were not in the list before */
478 if (!test_and_set_bit(RBIO_CACHE_BIT, &rbio->flags))
479 refcount_inc(&rbio->refs);
481 if (!list_empty(&rbio->stripe_cache)){
482 list_move(&rbio->stripe_cache, &table->stripe_cache);
484 list_add(&rbio->stripe_cache, &table->stripe_cache);
485 table->cache_size += 1;
488 spin_unlock(&rbio->bio_list_lock);
490 if (table->cache_size > RBIO_CACHE_SIZE) {
491 struct btrfs_raid_bio *found;
493 found = list_entry(table->stripe_cache.prev,
494 struct btrfs_raid_bio,
498 __remove_rbio_from_cache(found);
501 spin_unlock(&table->cache_lock);
505 * helper function to run the xor_blocks api. It is only
506 * able to do MAX_XOR_BLOCKS at a time, so we need to
509 static void run_xor(void **pages, int src_cnt, ssize_t len)
513 void *dest = pages[src_cnt];
516 xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS);
517 xor_blocks(xor_src_cnt, len, dest, pages + src_off);
519 src_cnt -= xor_src_cnt;
520 src_off += xor_src_cnt;
525 * Returns true if the bio list inside this rbio covers an entire stripe (no
528 static int rbio_is_full(struct btrfs_raid_bio *rbio)
530 unsigned long size = rbio->bio_list_bytes;
533 spin_lock(&rbio->bio_list_lock);
534 if (size != rbio->nr_data * BTRFS_STRIPE_LEN)
536 BUG_ON(size > rbio->nr_data * BTRFS_STRIPE_LEN);
537 spin_unlock(&rbio->bio_list_lock);
543 * returns 1 if it is safe to merge two rbios together.
544 * The merging is safe if the two rbios correspond to
545 * the same stripe and if they are both going in the same
546 * direction (read vs write), and if neither one is
547 * locked for final IO
549 * The caller is responsible for locking such that
550 * rmw_locked is safe to test
552 static int rbio_can_merge(struct btrfs_raid_bio *last,
553 struct btrfs_raid_bio *cur)
555 if (test_bit(RBIO_RMW_LOCKED_BIT, &last->flags) ||
556 test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags))
560 * we can't merge with cached rbios, since the
561 * idea is that when we merge the destination
562 * rbio is going to run our IO for us. We can
563 * steal from cached rbios though, other functions
566 if (test_bit(RBIO_CACHE_BIT, &last->flags) ||
567 test_bit(RBIO_CACHE_BIT, &cur->flags))
570 if (last->bioc->full_stripe_logical != cur->bioc->full_stripe_logical)
573 /* we can't merge with different operations */
574 if (last->operation != cur->operation)
577 * We've need read the full stripe from the drive.
578 * check and repair the parity and write the new results.
580 * We're not allowed to add any new bios to the
581 * bio list here, anyone else that wants to
582 * change this stripe needs to do their own rmw.
584 if (last->operation == BTRFS_RBIO_PARITY_SCRUB)
587 if (last->operation == BTRFS_RBIO_READ_REBUILD)
593 static unsigned int rbio_stripe_sector_index(const struct btrfs_raid_bio *rbio,
594 unsigned int stripe_nr,
595 unsigned int sector_nr)
597 ASSERT(stripe_nr < rbio->real_stripes);
598 ASSERT(sector_nr < rbio->stripe_nsectors);
600 return stripe_nr * rbio->stripe_nsectors + sector_nr;
603 /* Return a sector from rbio->stripe_sectors, not from the bio list */
604 static struct sector_ptr *rbio_stripe_sector(const struct btrfs_raid_bio *rbio,
605 unsigned int stripe_nr,
606 unsigned int sector_nr)
608 return &rbio->stripe_sectors[rbio_stripe_sector_index(rbio, stripe_nr,
612 /* Grab a sector inside P stripe */
613 static struct sector_ptr *rbio_pstripe_sector(const struct btrfs_raid_bio *rbio,
614 unsigned int sector_nr)
616 return rbio_stripe_sector(rbio, rbio->nr_data, sector_nr);
619 /* Grab a sector inside Q stripe, return NULL if not RAID6 */
620 static struct sector_ptr *rbio_qstripe_sector(const struct btrfs_raid_bio *rbio,
621 unsigned int sector_nr)
623 if (rbio->nr_data + 1 == rbio->real_stripes)
625 return rbio_stripe_sector(rbio, rbio->nr_data + 1, sector_nr);
629 * The first stripe in the table for a logical address
630 * has the lock. rbios are added in one of three ways:
632 * 1) Nobody has the stripe locked yet. The rbio is given
633 * the lock and 0 is returned. The caller must start the IO
636 * 2) Someone has the stripe locked, but we're able to merge
637 * with the lock owner. The rbio is freed and the IO will
638 * start automatically along with the existing rbio. 1 is returned.
640 * 3) Someone has the stripe locked, but we're not able to merge.
641 * The rbio is added to the lock owner's plug list, or merged into
642 * an rbio already on the plug list. When the lock owner unlocks,
643 * the next rbio on the list is run and the IO is started automatically.
646 * If we return 0, the caller still owns the rbio and must continue with
647 * IO submission. If we return 1, the caller must assume the rbio has
648 * already been freed.
650 static noinline int lock_stripe_add(struct btrfs_raid_bio *rbio)
652 struct btrfs_stripe_hash *h;
653 struct btrfs_raid_bio *cur;
654 struct btrfs_raid_bio *pending;
655 struct btrfs_raid_bio *freeit = NULL;
656 struct btrfs_raid_bio *cache_drop = NULL;
659 h = rbio->bioc->fs_info->stripe_hash_table->table + rbio_bucket(rbio);
662 list_for_each_entry(cur, &h->hash_list, hash_list) {
663 if (cur->bioc->full_stripe_logical != rbio->bioc->full_stripe_logical)
666 spin_lock(&cur->bio_list_lock);
668 /* Can we steal this cached rbio's pages? */
669 if (bio_list_empty(&cur->bio_list) &&
670 list_empty(&cur->plug_list) &&
671 test_bit(RBIO_CACHE_BIT, &cur->flags) &&
672 !test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) {
673 list_del_init(&cur->hash_list);
674 refcount_dec(&cur->refs);
676 steal_rbio(cur, rbio);
678 spin_unlock(&cur->bio_list_lock);
683 /* Can we merge into the lock owner? */
684 if (rbio_can_merge(cur, rbio)) {
685 merge_rbio(cur, rbio);
686 spin_unlock(&cur->bio_list_lock);
694 * We couldn't merge with the running rbio, see if we can merge
695 * with the pending ones. We don't have to check for rmw_locked
696 * because there is no way they are inside finish_rmw right now
698 list_for_each_entry(pending, &cur->plug_list, plug_list) {
699 if (rbio_can_merge(pending, rbio)) {
700 merge_rbio(pending, rbio);
701 spin_unlock(&cur->bio_list_lock);
709 * No merging, put us on the tail of the plug list, our rbio
710 * will be started with the currently running rbio unlocks
712 list_add_tail(&rbio->plug_list, &cur->plug_list);
713 spin_unlock(&cur->bio_list_lock);
718 refcount_inc(&rbio->refs);
719 list_add(&rbio->hash_list, &h->hash_list);
721 spin_unlock(&h->lock);
723 remove_rbio_from_cache(cache_drop);
725 free_raid_bio(freeit);
729 static void recover_rbio_work_locked(struct work_struct *work);
732 * called as rmw or parity rebuild is completed. If the plug list has more
733 * rbios waiting for this stripe, the next one on the list will be started
735 static noinline void unlock_stripe(struct btrfs_raid_bio *rbio)
738 struct btrfs_stripe_hash *h;
741 bucket = rbio_bucket(rbio);
742 h = rbio->bioc->fs_info->stripe_hash_table->table + bucket;
744 if (list_empty(&rbio->plug_list))
748 spin_lock(&rbio->bio_list_lock);
750 if (!list_empty(&rbio->hash_list)) {
752 * if we're still cached and there is no other IO
753 * to perform, just leave this rbio here for others
754 * to steal from later
756 if (list_empty(&rbio->plug_list) &&
757 test_bit(RBIO_CACHE_BIT, &rbio->flags)) {
759 clear_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
760 BUG_ON(!bio_list_empty(&rbio->bio_list));
764 list_del_init(&rbio->hash_list);
765 refcount_dec(&rbio->refs);
768 * we use the plug list to hold all the rbios
769 * waiting for the chance to lock this stripe.
770 * hand the lock over to one of them.
772 if (!list_empty(&rbio->plug_list)) {
773 struct btrfs_raid_bio *next;
774 struct list_head *head = rbio->plug_list.next;
776 next = list_entry(head, struct btrfs_raid_bio,
779 list_del_init(&rbio->plug_list);
781 list_add(&next->hash_list, &h->hash_list);
782 refcount_inc(&next->refs);
783 spin_unlock(&rbio->bio_list_lock);
784 spin_unlock(&h->lock);
786 if (next->operation == BTRFS_RBIO_READ_REBUILD) {
787 start_async_work(next, recover_rbio_work_locked);
788 } else if (next->operation == BTRFS_RBIO_WRITE) {
789 steal_rbio(rbio, next);
790 start_async_work(next, rmw_rbio_work_locked);
791 } else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) {
792 steal_rbio(rbio, next);
793 start_async_work(next, scrub_rbio_work_locked);
800 spin_unlock(&rbio->bio_list_lock);
801 spin_unlock(&h->lock);
805 remove_rbio_from_cache(rbio);
808 static void rbio_endio_bio_list(struct bio *cur, blk_status_t err)
815 cur->bi_status = err;
822 * this frees the rbio and runs through all the bios in the
823 * bio_list and calls end_io on them
825 static void rbio_orig_end_io(struct btrfs_raid_bio *rbio, blk_status_t err)
827 struct bio *cur = bio_list_get(&rbio->bio_list);
830 kfree(rbio->csum_buf);
831 bitmap_free(rbio->csum_bitmap);
832 rbio->csum_buf = NULL;
833 rbio->csum_bitmap = NULL;
836 * Clear the data bitmap, as the rbio may be cached for later usage.
837 * do this before before unlock_stripe() so there will be no new bio
840 bitmap_clear(&rbio->dbitmap, 0, rbio->stripe_nsectors);
843 * At this moment, rbio->bio_list is empty, however since rbio does not
844 * always have RBIO_RMW_LOCKED_BIT set and rbio is still linked on the
845 * hash list, rbio may be merged with others so that rbio->bio_list
847 * Once unlock_stripe() is done, rbio->bio_list will not be updated any
848 * more and we can call bio_endio() on all queued bios.
851 extra = bio_list_get(&rbio->bio_list);
854 rbio_endio_bio_list(cur, err);
856 rbio_endio_bio_list(extra, err);
860 * Get a sector pointer specified by its @stripe_nr and @sector_nr.
862 * @rbio: The raid bio
863 * @stripe_nr: Stripe number, valid range [0, real_stripe)
864 * @sector_nr: Sector number inside the stripe,
865 * valid range [0, stripe_nsectors)
866 * @bio_list_only: Whether to use sectors inside the bio list only.
868 * The read/modify/write code wants to reuse the original bio page as much
869 * as possible, and only use stripe_sectors as fallback.
871 static struct sector_ptr *sector_in_rbio(struct btrfs_raid_bio *rbio,
872 int stripe_nr, int sector_nr,
875 struct sector_ptr *sector;
878 ASSERT(stripe_nr >= 0 && stripe_nr < rbio->real_stripes);
879 ASSERT(sector_nr >= 0 && sector_nr < rbio->stripe_nsectors);
881 index = stripe_nr * rbio->stripe_nsectors + sector_nr;
882 ASSERT(index >= 0 && index < rbio->nr_sectors);
884 spin_lock(&rbio->bio_list_lock);
885 sector = &rbio->bio_sectors[index];
886 if (sector->page || bio_list_only) {
887 /* Don't return sector without a valid page pointer */
890 spin_unlock(&rbio->bio_list_lock);
893 spin_unlock(&rbio->bio_list_lock);
895 return &rbio->stripe_sectors[index];
899 * allocation and initial setup for the btrfs_raid_bio. Not
900 * this does not allocate any pages for rbio->pages.
902 static struct btrfs_raid_bio *alloc_rbio(struct btrfs_fs_info *fs_info,
903 struct btrfs_io_context *bioc)
905 const unsigned int real_stripes = bioc->num_stripes - bioc->replace_nr_stripes;
906 const unsigned int stripe_npages = BTRFS_STRIPE_LEN >> PAGE_SHIFT;
907 const unsigned int num_pages = stripe_npages * real_stripes;
908 const unsigned int stripe_nsectors =
909 BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits;
910 const unsigned int num_sectors = stripe_nsectors * real_stripes;
911 struct btrfs_raid_bio *rbio;
913 /* PAGE_SIZE must also be aligned to sectorsize for subpage support */
914 ASSERT(IS_ALIGNED(PAGE_SIZE, fs_info->sectorsize));
916 * Our current stripe len should be fixed to 64k thus stripe_nsectors
917 * (at most 16) should be no larger than BITS_PER_LONG.
919 ASSERT(stripe_nsectors <= BITS_PER_LONG);
921 rbio = kzalloc(sizeof(*rbio), GFP_NOFS);
923 return ERR_PTR(-ENOMEM);
924 rbio->stripe_pages = kcalloc(num_pages, sizeof(struct page *),
926 rbio->bio_sectors = kcalloc(num_sectors, sizeof(struct sector_ptr),
928 rbio->stripe_sectors = kcalloc(num_sectors, sizeof(struct sector_ptr),
930 rbio->finish_pointers = kcalloc(real_stripes, sizeof(void *), GFP_NOFS);
931 rbio->error_bitmap = bitmap_zalloc(num_sectors, GFP_NOFS);
933 if (!rbio->stripe_pages || !rbio->bio_sectors || !rbio->stripe_sectors ||
934 !rbio->finish_pointers || !rbio->error_bitmap) {
935 free_raid_bio_pointers(rbio);
937 return ERR_PTR(-ENOMEM);
940 bio_list_init(&rbio->bio_list);
941 init_waitqueue_head(&rbio->io_wait);
942 INIT_LIST_HEAD(&rbio->plug_list);
943 spin_lock_init(&rbio->bio_list_lock);
944 INIT_LIST_HEAD(&rbio->stripe_cache);
945 INIT_LIST_HEAD(&rbio->hash_list);
946 btrfs_get_bioc(bioc);
948 rbio->nr_pages = num_pages;
949 rbio->nr_sectors = num_sectors;
950 rbio->real_stripes = real_stripes;
951 rbio->stripe_npages = stripe_npages;
952 rbio->stripe_nsectors = stripe_nsectors;
953 refcount_set(&rbio->refs, 1);
954 atomic_set(&rbio->stripes_pending, 0);
956 ASSERT(btrfs_nr_parity_stripes(bioc->map_type));
957 rbio->nr_data = real_stripes - btrfs_nr_parity_stripes(bioc->map_type);
962 /* allocate pages for all the stripes in the bio, including parity */
963 static int alloc_rbio_pages(struct btrfs_raid_bio *rbio)
967 ret = btrfs_alloc_page_array(rbio->nr_pages, rbio->stripe_pages, 0);
970 /* Mapping all sectors */
971 index_stripe_sectors(rbio);
975 /* only allocate pages for p/q stripes */
976 static int alloc_rbio_parity_pages(struct btrfs_raid_bio *rbio)
978 const int data_pages = rbio->nr_data * rbio->stripe_npages;
981 ret = btrfs_alloc_page_array(rbio->nr_pages - data_pages,
982 rbio->stripe_pages + data_pages, 0);
986 index_stripe_sectors(rbio);
991 * Return the total number of errors found in the vertical stripe of @sector_nr.
993 * @faila and @failb will also be updated to the first and second stripe
994 * number of the errors.
996 static int get_rbio_veritical_errors(struct btrfs_raid_bio *rbio, int sector_nr,
997 int *faila, int *failb)
1000 int found_errors = 0;
1002 if (faila || failb) {
1004 * Both @faila and @failb should be valid pointers if any of
1005 * them is specified.
1007 ASSERT(faila && failb);
1012 for (stripe_nr = 0; stripe_nr < rbio->real_stripes; stripe_nr++) {
1013 int total_sector_nr = stripe_nr * rbio->stripe_nsectors + sector_nr;
1015 if (test_bit(total_sector_nr, rbio->error_bitmap)) {
1018 /* Update faila and failb. */
1021 else if (*failb < 0)
1026 return found_errors;
1030 * Add a single sector @sector into our list of bios for IO.
1032 * Return 0 if everything went well.
1033 * Return <0 for error.
1035 static int rbio_add_io_sector(struct btrfs_raid_bio *rbio,
1036 struct bio_list *bio_list,
1037 struct sector_ptr *sector,
1038 unsigned int stripe_nr,
1039 unsigned int sector_nr,
1042 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
1043 struct bio *last = bio_list->tail;
1046 struct btrfs_io_stripe *stripe;
1050 * Note: here stripe_nr has taken device replace into consideration,
1051 * thus it can be larger than rbio->real_stripe.
1052 * So here we check against bioc->num_stripes, not rbio->real_stripes.
1054 ASSERT(stripe_nr >= 0 && stripe_nr < rbio->bioc->num_stripes);
1055 ASSERT(sector_nr >= 0 && sector_nr < rbio->stripe_nsectors);
1056 ASSERT(sector->page);
1058 stripe = &rbio->bioc->stripes[stripe_nr];
1059 disk_start = stripe->physical + sector_nr * sectorsize;
1061 /* if the device is missing, just fail this stripe */
1062 if (!stripe->dev->bdev) {
1065 set_bit(stripe_nr * rbio->stripe_nsectors + sector_nr,
1066 rbio->error_bitmap);
1068 /* Check if we have reached tolerance early. */
1069 found_errors = get_rbio_veritical_errors(rbio, sector_nr,
1071 if (found_errors > rbio->bioc->max_errors)
1076 /* see if we can add this page onto our existing bio */
1078 u64 last_end = last->bi_iter.bi_sector << SECTOR_SHIFT;
1079 last_end += last->bi_iter.bi_size;
1082 * we can't merge these if they are from different
1083 * devices or if they are not contiguous
1085 if (last_end == disk_start && !last->bi_status &&
1086 last->bi_bdev == stripe->dev->bdev) {
1087 ret = bio_add_page(last, sector->page, sectorsize,
1089 if (ret == sectorsize)
1094 /* put a new bio on the list */
1095 bio = bio_alloc(stripe->dev->bdev,
1096 max(BTRFS_STRIPE_LEN >> PAGE_SHIFT, 1),
1098 bio->bi_iter.bi_sector = disk_start >> SECTOR_SHIFT;
1099 bio->bi_private = rbio;
1101 __bio_add_page(bio, sector->page, sectorsize, sector->pgoff);
1102 bio_list_add(bio_list, bio);
1106 static void index_one_bio(struct btrfs_raid_bio *rbio, struct bio *bio)
1108 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
1109 struct bio_vec bvec;
1110 struct bvec_iter iter;
1111 u32 offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) -
1112 rbio->bioc->full_stripe_logical;
1114 bio_for_each_segment(bvec, bio, iter) {
1117 for (bvec_offset = 0; bvec_offset < bvec.bv_len;
1118 bvec_offset += sectorsize, offset += sectorsize) {
1119 int index = offset / sectorsize;
1120 struct sector_ptr *sector = &rbio->bio_sectors[index];
1122 sector->page = bvec.bv_page;
1123 sector->pgoff = bvec.bv_offset + bvec_offset;
1124 ASSERT(sector->pgoff < PAGE_SIZE);
1130 * helper function to walk our bio list and populate the bio_pages array with
1131 * the result. This seems expensive, but it is faster than constantly
1132 * searching through the bio list as we setup the IO in finish_rmw or stripe
1135 * This must be called before you trust the answers from page_in_rbio
1137 static void index_rbio_pages(struct btrfs_raid_bio *rbio)
1141 spin_lock(&rbio->bio_list_lock);
1142 bio_list_for_each(bio, &rbio->bio_list)
1143 index_one_bio(rbio, bio);
1145 spin_unlock(&rbio->bio_list_lock);
1148 static void bio_get_trace_info(struct btrfs_raid_bio *rbio, struct bio *bio,
1149 struct raid56_bio_trace_info *trace_info)
1151 const struct btrfs_io_context *bioc = rbio->bioc;
1156 /* We rely on bio->bi_bdev to find the stripe number. */
1160 for (i = 0; i < bioc->num_stripes; i++) {
1161 if (bio->bi_bdev != bioc->stripes[i].dev->bdev)
1163 trace_info->stripe_nr = i;
1164 trace_info->devid = bioc->stripes[i].dev->devid;
1165 trace_info->offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) -
1166 bioc->stripes[i].physical;
1171 trace_info->devid = -1;
1172 trace_info->offset = -1;
1173 trace_info->stripe_nr = -1;
1176 static inline void bio_list_put(struct bio_list *bio_list)
1180 while ((bio = bio_list_pop(bio_list)))
1184 /* Generate PQ for one vertical stripe. */
1185 static void generate_pq_vertical(struct btrfs_raid_bio *rbio, int sectornr)
1187 void **pointers = rbio->finish_pointers;
1188 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
1189 struct sector_ptr *sector;
1191 const bool has_qstripe = rbio->bioc->map_type & BTRFS_BLOCK_GROUP_RAID6;
1193 /* First collect one sector from each data stripe */
1194 for (stripe = 0; stripe < rbio->nr_data; stripe++) {
1195 sector = sector_in_rbio(rbio, stripe, sectornr, 0);
1196 pointers[stripe] = kmap_local_page(sector->page) +
1200 /* Then add the parity stripe */
1201 sector = rbio_pstripe_sector(rbio, sectornr);
1202 sector->uptodate = 1;
1203 pointers[stripe++] = kmap_local_page(sector->page) + sector->pgoff;
1207 * RAID6, add the qstripe and call the library function
1208 * to fill in our p/q
1210 sector = rbio_qstripe_sector(rbio, sectornr);
1211 sector->uptodate = 1;
1212 pointers[stripe++] = kmap_local_page(sector->page) +
1215 raid6_call.gen_syndrome(rbio->real_stripes, sectorsize,
1219 memcpy(pointers[rbio->nr_data], pointers[0], sectorsize);
1220 run_xor(pointers + 1, rbio->nr_data - 1, sectorsize);
1222 for (stripe = stripe - 1; stripe >= 0; stripe--)
1223 kunmap_local(pointers[stripe]);
1226 static int rmw_assemble_write_bios(struct btrfs_raid_bio *rbio,
1227 struct bio_list *bio_list)
1229 /* The total sector number inside the full stripe. */
1230 int total_sector_nr;
1235 ASSERT(bio_list_size(bio_list) == 0);
1237 /* We should have at least one data sector. */
1238 ASSERT(bitmap_weight(&rbio->dbitmap, rbio->stripe_nsectors));
1241 * Reset errors, as we may have errors inherited from from degraded
1244 bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors);
1247 * Start assembly. Make bios for everything from the higher layers (the
1248 * bio_list in our rbio) and our P/Q. Ignore everything else.
1250 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
1251 total_sector_nr++) {
1252 struct sector_ptr *sector;
1254 stripe = total_sector_nr / rbio->stripe_nsectors;
1255 sectornr = total_sector_nr % rbio->stripe_nsectors;
1257 /* This vertical stripe has no data, skip it. */
1258 if (!test_bit(sectornr, &rbio->dbitmap))
1261 if (stripe < rbio->nr_data) {
1262 sector = sector_in_rbio(rbio, stripe, sectornr, 1);
1266 sector = rbio_stripe_sector(rbio, stripe, sectornr);
1269 ret = rbio_add_io_sector(rbio, bio_list, sector, stripe,
1270 sectornr, REQ_OP_WRITE);
1275 if (likely(!rbio->bioc->replace_nr_stripes))
1279 * Make a copy for the replace target device.
1281 * Thus the source stripe number (in replace_stripe_src) should be valid.
1283 ASSERT(rbio->bioc->replace_stripe_src >= 0);
1285 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
1286 total_sector_nr++) {
1287 struct sector_ptr *sector;
1289 stripe = total_sector_nr / rbio->stripe_nsectors;
1290 sectornr = total_sector_nr % rbio->stripe_nsectors;
1293 * For RAID56, there is only one device that can be replaced,
1294 * and replace_stripe_src[0] indicates the stripe number we
1295 * need to copy from.
1297 if (stripe != rbio->bioc->replace_stripe_src) {
1299 * We can skip the whole stripe completely, note
1300 * total_sector_nr will be increased by one anyway.
1302 ASSERT(sectornr == 0);
1303 total_sector_nr += rbio->stripe_nsectors - 1;
1307 /* This vertical stripe has no data, skip it. */
1308 if (!test_bit(sectornr, &rbio->dbitmap))
1311 if (stripe < rbio->nr_data) {
1312 sector = sector_in_rbio(rbio, stripe, sectornr, 1);
1316 sector = rbio_stripe_sector(rbio, stripe, sectornr);
1319 ret = rbio_add_io_sector(rbio, bio_list, sector,
1321 sectornr, REQ_OP_WRITE);
1328 bio_list_put(bio_list);
1332 static void set_rbio_range_error(struct btrfs_raid_bio *rbio, struct bio *bio)
1334 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1335 u32 offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) -
1336 rbio->bioc->full_stripe_logical;
1337 int total_nr_sector = offset >> fs_info->sectorsize_bits;
1339 ASSERT(total_nr_sector < rbio->nr_data * rbio->stripe_nsectors);
1341 bitmap_set(rbio->error_bitmap, total_nr_sector,
1342 bio->bi_iter.bi_size >> fs_info->sectorsize_bits);
1345 * Special handling for raid56_alloc_missing_rbio() used by
1346 * scrub/replace. Unlike call path in raid56_parity_recover(), they
1347 * pass an empty bio here. Thus we have to find out the missing device
1348 * and mark the stripe error instead.
1350 if (bio->bi_iter.bi_size == 0) {
1351 bool found_missing = false;
1354 for (stripe_nr = 0; stripe_nr < rbio->real_stripes; stripe_nr++) {
1355 if (!rbio->bioc->stripes[stripe_nr].dev->bdev) {
1356 found_missing = true;
1357 bitmap_set(rbio->error_bitmap,
1358 stripe_nr * rbio->stripe_nsectors,
1359 rbio->stripe_nsectors);
1362 ASSERT(found_missing);
1367 * For subpage case, we can no longer set page Up-to-date directly for
1368 * stripe_pages[], thus we need to locate the sector.
1370 static struct sector_ptr *find_stripe_sector(struct btrfs_raid_bio *rbio,
1376 for (i = 0; i < rbio->nr_sectors; i++) {
1377 struct sector_ptr *sector = &rbio->stripe_sectors[i];
1379 if (sector->page == page && sector->pgoff == pgoff)
1386 * this sets each page in the bio uptodate. It should only be used on private
1387 * rbio pages, nothing that comes in from the higher layers
1389 static void set_bio_pages_uptodate(struct btrfs_raid_bio *rbio, struct bio *bio)
1391 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
1392 struct bio_vec *bvec;
1393 struct bvec_iter_all iter_all;
1395 ASSERT(!bio_flagged(bio, BIO_CLONED));
1397 bio_for_each_segment_all(bvec, bio, iter_all) {
1398 struct sector_ptr *sector;
1401 for (pgoff = bvec->bv_offset; pgoff - bvec->bv_offset < bvec->bv_len;
1402 pgoff += sectorsize) {
1403 sector = find_stripe_sector(rbio, bvec->bv_page, pgoff);
1406 sector->uptodate = 1;
1411 static int get_bio_sector_nr(struct btrfs_raid_bio *rbio, struct bio *bio)
1413 struct bio_vec *bv = bio_first_bvec_all(bio);
1416 for (i = 0; i < rbio->nr_sectors; i++) {
1417 struct sector_ptr *sector;
1419 sector = &rbio->stripe_sectors[i];
1420 if (sector->page == bv->bv_page && sector->pgoff == bv->bv_offset)
1422 sector = &rbio->bio_sectors[i];
1423 if (sector->page == bv->bv_page && sector->pgoff == bv->bv_offset)
1426 ASSERT(i < rbio->nr_sectors);
1430 static void rbio_update_error_bitmap(struct btrfs_raid_bio *rbio, struct bio *bio)
1432 int total_sector_nr = get_bio_sector_nr(rbio, bio);
1434 struct bio_vec *bvec;
1437 bio_for_each_bvec_all(bvec, bio, i)
1438 bio_size += bvec->bv_len;
1441 * Since we can have multiple bios touching the error_bitmap, we cannot
1442 * call bitmap_set() without protection.
1444 * Instead use set_bit() for each bit, as set_bit() itself is atomic.
1446 for (i = total_sector_nr; i < total_sector_nr +
1447 (bio_size >> rbio->bioc->fs_info->sectorsize_bits); i++)
1448 set_bit(i, rbio->error_bitmap);
1451 /* Verify the data sectors at read time. */
1452 static void verify_bio_data_sectors(struct btrfs_raid_bio *rbio,
1455 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1456 int total_sector_nr = get_bio_sector_nr(rbio, bio);
1457 struct bio_vec *bvec;
1458 struct bvec_iter_all iter_all;
1460 /* No data csum for the whole stripe, no need to verify. */
1461 if (!rbio->csum_bitmap || !rbio->csum_buf)
1464 /* P/Q stripes, they have no data csum to verify against. */
1465 if (total_sector_nr >= rbio->nr_data * rbio->stripe_nsectors)
1468 bio_for_each_segment_all(bvec, bio, iter_all) {
1471 for (bv_offset = bvec->bv_offset;
1472 bv_offset < bvec->bv_offset + bvec->bv_len;
1473 bv_offset += fs_info->sectorsize, total_sector_nr++) {
1474 u8 csum_buf[BTRFS_CSUM_SIZE];
1475 u8 *expected_csum = rbio->csum_buf +
1476 total_sector_nr * fs_info->csum_size;
1479 /* No csum for this sector, skip to the next sector. */
1480 if (!test_bit(total_sector_nr, rbio->csum_bitmap))
1483 ret = btrfs_check_sector_csum(fs_info, bvec->bv_page,
1484 bv_offset, csum_buf, expected_csum);
1486 set_bit(total_sector_nr, rbio->error_bitmap);
1491 static void raid_wait_read_end_io(struct bio *bio)
1493 struct btrfs_raid_bio *rbio = bio->bi_private;
1495 if (bio->bi_status) {
1496 rbio_update_error_bitmap(rbio, bio);
1498 set_bio_pages_uptodate(rbio, bio);
1499 verify_bio_data_sectors(rbio, bio);
1503 if (atomic_dec_and_test(&rbio->stripes_pending))
1504 wake_up(&rbio->io_wait);
1507 static void submit_read_wait_bio_list(struct btrfs_raid_bio *rbio,
1508 struct bio_list *bio_list)
1512 atomic_set(&rbio->stripes_pending, bio_list_size(bio_list));
1513 while ((bio = bio_list_pop(bio_list))) {
1514 bio->bi_end_io = raid_wait_read_end_io;
1516 if (trace_raid56_read_enabled()) {
1517 struct raid56_bio_trace_info trace_info = { 0 };
1519 bio_get_trace_info(rbio, bio, &trace_info);
1520 trace_raid56_read(rbio, bio, &trace_info);
1525 wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0);
1528 static int alloc_rbio_data_pages(struct btrfs_raid_bio *rbio)
1530 const int data_pages = rbio->nr_data * rbio->stripe_npages;
1533 ret = btrfs_alloc_page_array(data_pages, rbio->stripe_pages, 0);
1537 index_stripe_sectors(rbio);
1542 * We use plugging call backs to collect full stripes.
1543 * Any time we get a partial stripe write while plugged
1544 * we collect it into a list. When the unplug comes down,
1545 * we sort the list by logical block number and merge
1546 * everything we can into the same rbios
1548 struct btrfs_plug_cb {
1549 struct blk_plug_cb cb;
1550 struct btrfs_fs_info *info;
1551 struct list_head rbio_list;
1555 * rbios on the plug list are sorted for easier merging.
1557 static int plug_cmp(void *priv, const struct list_head *a,
1558 const struct list_head *b)
1560 const struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio,
1562 const struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio,
1564 u64 a_sector = ra->bio_list.head->bi_iter.bi_sector;
1565 u64 b_sector = rb->bio_list.head->bi_iter.bi_sector;
1567 if (a_sector < b_sector)
1569 if (a_sector > b_sector)
1574 static void raid_unplug(struct blk_plug_cb *cb, bool from_schedule)
1576 struct btrfs_plug_cb *plug = container_of(cb, struct btrfs_plug_cb, cb);
1577 struct btrfs_raid_bio *cur;
1578 struct btrfs_raid_bio *last = NULL;
1580 list_sort(NULL, &plug->rbio_list, plug_cmp);
1582 while (!list_empty(&plug->rbio_list)) {
1583 cur = list_entry(plug->rbio_list.next,
1584 struct btrfs_raid_bio, plug_list);
1585 list_del_init(&cur->plug_list);
1587 if (rbio_is_full(cur)) {
1588 /* We have a full stripe, queue it down. */
1589 start_async_work(cur, rmw_rbio_work);
1593 if (rbio_can_merge(last, cur)) {
1594 merge_rbio(last, cur);
1598 start_async_work(last, rmw_rbio_work);
1603 start_async_work(last, rmw_rbio_work);
1607 /* Add the original bio into rbio->bio_list, and update rbio::dbitmap. */
1608 static void rbio_add_bio(struct btrfs_raid_bio *rbio, struct bio *orig_bio)
1610 const struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1611 const u64 orig_logical = orig_bio->bi_iter.bi_sector << SECTOR_SHIFT;
1612 const u64 full_stripe_start = rbio->bioc->full_stripe_logical;
1613 const u32 orig_len = orig_bio->bi_iter.bi_size;
1614 const u32 sectorsize = fs_info->sectorsize;
1617 ASSERT(orig_logical >= full_stripe_start &&
1618 orig_logical + orig_len <= full_stripe_start +
1619 rbio->nr_data * BTRFS_STRIPE_LEN);
1621 bio_list_add(&rbio->bio_list, orig_bio);
1622 rbio->bio_list_bytes += orig_bio->bi_iter.bi_size;
1624 /* Update the dbitmap. */
1625 for (cur_logical = orig_logical; cur_logical < orig_logical + orig_len;
1626 cur_logical += sectorsize) {
1627 int bit = ((u32)(cur_logical - full_stripe_start) >>
1628 fs_info->sectorsize_bits) % rbio->stripe_nsectors;
1630 set_bit(bit, &rbio->dbitmap);
1635 * our main entry point for writes from the rest of the FS.
1637 void raid56_parity_write(struct bio *bio, struct btrfs_io_context *bioc)
1639 struct btrfs_fs_info *fs_info = bioc->fs_info;
1640 struct btrfs_raid_bio *rbio;
1641 struct btrfs_plug_cb *plug = NULL;
1642 struct blk_plug_cb *cb;
1644 rbio = alloc_rbio(fs_info, bioc);
1646 bio->bi_status = errno_to_blk_status(PTR_ERR(rbio));
1650 rbio->operation = BTRFS_RBIO_WRITE;
1651 rbio_add_bio(rbio, bio);
1654 * Don't plug on full rbios, just get them out the door
1655 * as quickly as we can
1657 if (!rbio_is_full(rbio)) {
1658 cb = blk_check_plugged(raid_unplug, fs_info, sizeof(*plug));
1660 plug = container_of(cb, struct btrfs_plug_cb, cb);
1662 plug->info = fs_info;
1663 INIT_LIST_HEAD(&plug->rbio_list);
1665 list_add_tail(&rbio->plug_list, &plug->rbio_list);
1671 * Either we don't have any existing plug, or we're doing a full stripe,
1672 * queue the rmw work now.
1674 start_async_work(rbio, rmw_rbio_work);
1677 static int verify_one_sector(struct btrfs_raid_bio *rbio,
1678 int stripe_nr, int sector_nr)
1680 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1681 struct sector_ptr *sector;
1682 u8 csum_buf[BTRFS_CSUM_SIZE];
1686 if (!rbio->csum_bitmap || !rbio->csum_buf)
1689 /* No way to verify P/Q as they are not covered by data csum. */
1690 if (stripe_nr >= rbio->nr_data)
1693 * If we're rebuilding a read, we have to use pages from the
1694 * bio list if possible.
1696 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
1697 sector = sector_in_rbio(rbio, stripe_nr, sector_nr, 0);
1699 sector = rbio_stripe_sector(rbio, stripe_nr, sector_nr);
1702 ASSERT(sector->page);
1704 csum_expected = rbio->csum_buf +
1705 (stripe_nr * rbio->stripe_nsectors + sector_nr) *
1707 ret = btrfs_check_sector_csum(fs_info, sector->page, sector->pgoff,
1708 csum_buf, csum_expected);
1713 * Recover a vertical stripe specified by @sector_nr.
1714 * @*pointers are the pre-allocated pointers by the caller, so we don't
1715 * need to allocate/free the pointers again and again.
1717 static int recover_vertical(struct btrfs_raid_bio *rbio, int sector_nr,
1718 void **pointers, void **unmap_array)
1720 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1721 struct sector_ptr *sector;
1722 const u32 sectorsize = fs_info->sectorsize;
1730 * Now we just use bitmap to mark the horizontal stripes in
1731 * which we have data when doing parity scrub.
1733 if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB &&
1734 !test_bit(sector_nr, &rbio->dbitmap))
1737 found_errors = get_rbio_veritical_errors(rbio, sector_nr, &faila,
1740 * No errors in the vertical stripe, skip it. Can happen for recovery
1741 * which only part of a stripe failed csum check.
1746 if (found_errors > rbio->bioc->max_errors)
1750 * Setup our array of pointers with sectors from each stripe
1752 * NOTE: store a duplicate array of pointers to preserve the
1755 for (stripe_nr = 0; stripe_nr < rbio->real_stripes; stripe_nr++) {
1757 * If we're rebuilding a read, we have to use pages from the
1758 * bio list if possible.
1760 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
1761 sector = sector_in_rbio(rbio, stripe_nr, sector_nr, 0);
1763 sector = rbio_stripe_sector(rbio, stripe_nr, sector_nr);
1765 ASSERT(sector->page);
1766 pointers[stripe_nr] = kmap_local_page(sector->page) +
1768 unmap_array[stripe_nr] = pointers[stripe_nr];
1771 /* All raid6 handling here */
1772 if (rbio->bioc->map_type & BTRFS_BLOCK_GROUP_RAID6) {
1773 /* Single failure, rebuild from parity raid5 style */
1775 if (faila == rbio->nr_data)
1777 * Just the P stripe has failed, without
1778 * a bad data or Q stripe.
1779 * We have nothing to do, just skip the
1780 * recovery for this stripe.
1784 * a single failure in raid6 is rebuilt
1785 * in the pstripe code below
1791 * If the q stripe is failed, do a pstripe reconstruction from
1793 * If both the q stripe and the P stripe are failed, we're
1794 * here due to a crc mismatch and we can't give them the
1797 if (failb == rbio->real_stripes - 1) {
1798 if (faila == rbio->real_stripes - 2)
1800 * Only P and Q are corrupted.
1801 * We only care about data stripes recovery,
1802 * can skip this vertical stripe.
1806 * Otherwise we have one bad data stripe and
1807 * a good P stripe. raid5!
1812 if (failb == rbio->real_stripes - 2) {
1813 raid6_datap_recov(rbio->real_stripes, sectorsize,
1816 raid6_2data_recov(rbio->real_stripes, sectorsize,
1817 faila, failb, pointers);
1822 /* Rebuild from P stripe here (raid5 or raid6). */
1823 ASSERT(failb == -1);
1825 /* Copy parity block into failed block to start with */
1826 memcpy(pointers[faila], pointers[rbio->nr_data], sectorsize);
1828 /* Rearrange the pointer array */
1829 p = pointers[faila];
1830 for (stripe_nr = faila; stripe_nr < rbio->nr_data - 1;
1832 pointers[stripe_nr] = pointers[stripe_nr + 1];
1833 pointers[rbio->nr_data - 1] = p;
1835 /* Xor in the rest */
1836 run_xor(pointers, rbio->nr_data - 1, sectorsize);
1841 * No matter if this is a RMW or recovery, we should have all
1842 * failed sectors repaired in the vertical stripe, thus they are now
1844 * Especially if we determine to cache the rbio, we need to
1845 * have at least all data sectors uptodate.
1847 * If possible, also check if the repaired sector matches its data
1851 ret = verify_one_sector(rbio, faila, sector_nr);
1855 sector = rbio_stripe_sector(rbio, faila, sector_nr);
1856 sector->uptodate = 1;
1859 ret = verify_one_sector(rbio, failb, sector_nr);
1863 sector = rbio_stripe_sector(rbio, failb, sector_nr);
1864 sector->uptodate = 1;
1868 for (stripe_nr = rbio->real_stripes - 1; stripe_nr >= 0; stripe_nr--)
1869 kunmap_local(unmap_array[stripe_nr]);
1873 static int recover_sectors(struct btrfs_raid_bio *rbio)
1875 void **pointers = NULL;
1876 void **unmap_array = NULL;
1881 * @pointers array stores the pointer for each sector.
1883 * @unmap_array stores copy of pointers that does not get reordered
1884 * during reconstruction so that kunmap_local works.
1886 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
1887 unmap_array = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
1888 if (!pointers || !unmap_array) {
1893 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
1894 spin_lock(&rbio->bio_list_lock);
1895 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1896 spin_unlock(&rbio->bio_list_lock);
1899 index_rbio_pages(rbio);
1901 for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++) {
1902 ret = recover_vertical(rbio, sectornr, pointers, unmap_array);
1913 static void recover_rbio(struct btrfs_raid_bio *rbio)
1915 struct bio_list bio_list = BIO_EMPTY_LIST;
1916 int total_sector_nr;
1920 * Either we're doing recover for a read failure or degraded write,
1921 * caller should have set error bitmap correctly.
1923 ASSERT(bitmap_weight(rbio->error_bitmap, rbio->nr_sectors));
1925 /* For recovery, we need to read all sectors including P/Q. */
1926 ret = alloc_rbio_pages(rbio);
1930 index_rbio_pages(rbio);
1933 * Read everything that hasn't failed. However this time we will
1934 * not trust any cached sector.
1935 * As we may read out some stale data but higher layer is not reading
1938 * So here we always re-read everything in recovery path.
1940 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
1941 total_sector_nr++) {
1942 int stripe = total_sector_nr / rbio->stripe_nsectors;
1943 int sectornr = total_sector_nr % rbio->stripe_nsectors;
1944 struct sector_ptr *sector;
1947 * Skip the range which has error. It can be a range which is
1948 * marked error (for csum mismatch), or it can be a missing
1951 if (!rbio->bioc->stripes[stripe].dev->bdev ||
1952 test_bit(total_sector_nr, rbio->error_bitmap)) {
1954 * Also set the error bit for missing device, which
1955 * may not yet have its error bit set.
1957 set_bit(total_sector_nr, rbio->error_bitmap);
1961 sector = rbio_stripe_sector(rbio, stripe, sectornr);
1962 ret = rbio_add_io_sector(rbio, &bio_list, sector, stripe,
1963 sectornr, REQ_OP_READ);
1965 bio_list_put(&bio_list);
1970 submit_read_wait_bio_list(rbio, &bio_list);
1971 ret = recover_sectors(rbio);
1973 rbio_orig_end_io(rbio, errno_to_blk_status(ret));
1976 static void recover_rbio_work(struct work_struct *work)
1978 struct btrfs_raid_bio *rbio;
1980 rbio = container_of(work, struct btrfs_raid_bio, work);
1981 if (!lock_stripe_add(rbio))
1985 static void recover_rbio_work_locked(struct work_struct *work)
1987 recover_rbio(container_of(work, struct btrfs_raid_bio, work));
1990 static void set_rbio_raid6_extra_error(struct btrfs_raid_bio *rbio, int mirror_num)
1996 * This is for RAID6 extra recovery tries, thus mirror number should
1998 * Mirror 1 means read from data stripes. Mirror 2 means rebuild using
2001 ASSERT(mirror_num > 2);
2002 for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) {
2007 found_errors = get_rbio_veritical_errors(rbio, sector_nr,
2009 /* This vertical stripe doesn't have errors. */
2014 * If we found errors, there should be only one error marked
2015 * by previous set_rbio_range_error().
2017 ASSERT(found_errors == 1);
2020 /* Now select another stripe to mark as error. */
2021 failb = rbio->real_stripes - (mirror_num - 1);
2025 /* Set the extra bit in error bitmap. */
2027 set_bit(failb * rbio->stripe_nsectors + sector_nr,
2028 rbio->error_bitmap);
2031 /* We should found at least one vertical stripe with error.*/
2036 * the main entry point for reads from the higher layers. This
2037 * is really only called when the normal read path had a failure,
2038 * so we assume the bio they send down corresponds to a failed part
2041 void raid56_parity_recover(struct bio *bio, struct btrfs_io_context *bioc,
2044 struct btrfs_fs_info *fs_info = bioc->fs_info;
2045 struct btrfs_raid_bio *rbio;
2047 rbio = alloc_rbio(fs_info, bioc);
2049 bio->bi_status = errno_to_blk_status(PTR_ERR(rbio));
2054 rbio->operation = BTRFS_RBIO_READ_REBUILD;
2055 rbio_add_bio(rbio, bio);
2057 set_rbio_range_error(rbio, bio);
2061 * for 'mirror == 2', reconstruct from all other stripes.
2062 * for 'mirror_num > 2', select a stripe to fail on every retry.
2065 set_rbio_raid6_extra_error(rbio, mirror_num);
2067 start_async_work(rbio, recover_rbio_work);
2070 static void fill_data_csums(struct btrfs_raid_bio *rbio)
2072 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
2073 struct btrfs_root *csum_root = btrfs_csum_root(fs_info,
2074 rbio->bioc->full_stripe_logical);
2075 const u64 start = rbio->bioc->full_stripe_logical;
2076 const u32 len = (rbio->nr_data * rbio->stripe_nsectors) <<
2077 fs_info->sectorsize_bits;
2080 /* The rbio should not have its csum buffer initialized. */
2081 ASSERT(!rbio->csum_buf && !rbio->csum_bitmap);
2084 * Skip the csum search if:
2086 * - The rbio doesn't belong to data block groups
2087 * Then we are doing IO for tree blocks, no need to search csums.
2089 * - The rbio belongs to mixed block groups
2090 * This is to avoid deadlock, as we're already holding the full
2091 * stripe lock, if we trigger a metadata read, and it needs to do
2092 * raid56 recovery, we will deadlock.
2094 if (!(rbio->bioc->map_type & BTRFS_BLOCK_GROUP_DATA) ||
2095 rbio->bioc->map_type & BTRFS_BLOCK_GROUP_METADATA)
2098 rbio->csum_buf = kzalloc(rbio->nr_data * rbio->stripe_nsectors *
2099 fs_info->csum_size, GFP_NOFS);
2100 rbio->csum_bitmap = bitmap_zalloc(rbio->nr_data * rbio->stripe_nsectors,
2102 if (!rbio->csum_buf || !rbio->csum_bitmap) {
2107 ret = btrfs_lookup_csums_bitmap(csum_root, NULL, start, start + len - 1,
2108 rbio->csum_buf, rbio->csum_bitmap);
2111 if (bitmap_empty(rbio->csum_bitmap, len >> fs_info->sectorsize_bits))
2117 * We failed to allocate memory or grab the csum, but it's not fatal,
2118 * we can still continue. But better to warn users that RMW is no
2119 * longer safe for this particular sub-stripe write.
2121 btrfs_warn_rl(fs_info,
2122 "sub-stripe write for full stripe %llu is not safe, failed to get csum: %d",
2123 rbio->bioc->full_stripe_logical, ret);
2125 kfree(rbio->csum_buf);
2126 bitmap_free(rbio->csum_bitmap);
2127 rbio->csum_buf = NULL;
2128 rbio->csum_bitmap = NULL;
2131 static int rmw_read_wait_recover(struct btrfs_raid_bio *rbio)
2133 struct bio_list bio_list = BIO_EMPTY_LIST;
2134 int total_sector_nr;
2138 * Fill the data csums we need for data verification. We need to fill
2139 * the csum_bitmap/csum_buf first, as our endio function will try to
2140 * verify the data sectors.
2142 fill_data_csums(rbio);
2145 * Build a list of bios to read all sectors (including data and P/Q).
2147 * This behavior is to compensate the later csum verification and recovery.
2149 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
2150 total_sector_nr++) {
2151 struct sector_ptr *sector;
2152 int stripe = total_sector_nr / rbio->stripe_nsectors;
2153 int sectornr = total_sector_nr % rbio->stripe_nsectors;
2155 sector = rbio_stripe_sector(rbio, stripe, sectornr);
2156 ret = rbio_add_io_sector(rbio, &bio_list, sector,
2157 stripe, sectornr, REQ_OP_READ);
2159 bio_list_put(&bio_list);
2165 * We may or may not have any corrupted sectors (including missing dev
2166 * and csum mismatch), just let recover_sectors() to handle them all.
2168 submit_read_wait_bio_list(rbio, &bio_list);
2169 return recover_sectors(rbio);
2172 static void raid_wait_write_end_io(struct bio *bio)
2174 struct btrfs_raid_bio *rbio = bio->bi_private;
2175 blk_status_t err = bio->bi_status;
2178 rbio_update_error_bitmap(rbio, bio);
2180 if (atomic_dec_and_test(&rbio->stripes_pending))
2181 wake_up(&rbio->io_wait);
2184 static void submit_write_bios(struct btrfs_raid_bio *rbio,
2185 struct bio_list *bio_list)
2189 atomic_set(&rbio->stripes_pending, bio_list_size(bio_list));
2190 while ((bio = bio_list_pop(bio_list))) {
2191 bio->bi_end_io = raid_wait_write_end_io;
2193 if (trace_raid56_write_enabled()) {
2194 struct raid56_bio_trace_info trace_info = { 0 };
2196 bio_get_trace_info(rbio, bio, &trace_info);
2197 trace_raid56_write(rbio, bio, &trace_info);
2204 * To determine if we need to read any sector from the disk.
2205 * Should only be utilized in RMW path, to skip cached rbio.
2207 static bool need_read_stripe_sectors(struct btrfs_raid_bio *rbio)
2211 for (i = 0; i < rbio->nr_data * rbio->stripe_nsectors; i++) {
2212 struct sector_ptr *sector = &rbio->stripe_sectors[i];
2215 * We have a sector which doesn't have page nor uptodate,
2216 * thus this rbio can not be cached one, as cached one must
2217 * have all its data sectors present and uptodate.
2219 if (!sector->page || !sector->uptodate)
2225 static void rmw_rbio(struct btrfs_raid_bio *rbio)
2227 struct bio_list bio_list;
2232 * Allocate the pages for parity first, as P/Q pages will always be
2233 * needed for both full-stripe and sub-stripe writes.
2235 ret = alloc_rbio_parity_pages(rbio);
2240 * Either full stripe write, or we have every data sector already
2241 * cached, can go to write path immediately.
2243 if (!rbio_is_full(rbio) && need_read_stripe_sectors(rbio)) {
2245 * Now we're doing sub-stripe write, also need all data stripes
2246 * to do the full RMW.
2248 ret = alloc_rbio_data_pages(rbio);
2252 index_rbio_pages(rbio);
2254 ret = rmw_read_wait_recover(rbio);
2260 * At this stage we're not allowed to add any new bios to the
2261 * bio list any more, anyone else that wants to change this stripe
2262 * needs to do their own rmw.
2264 spin_lock(&rbio->bio_list_lock);
2265 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
2266 spin_unlock(&rbio->bio_list_lock);
2268 bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors);
2270 index_rbio_pages(rbio);
2273 * We don't cache full rbios because we're assuming
2274 * the higher layers are unlikely to use this area of
2275 * the disk again soon. If they do use it again,
2276 * hopefully they will send another full bio.
2278 if (!rbio_is_full(rbio))
2279 cache_rbio_pages(rbio);
2281 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2283 for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++)
2284 generate_pq_vertical(rbio, sectornr);
2286 bio_list_init(&bio_list);
2287 ret = rmw_assemble_write_bios(rbio, &bio_list);
2291 /* We should have at least one bio assembled. */
2292 ASSERT(bio_list_size(&bio_list));
2293 submit_write_bios(rbio, &bio_list);
2294 wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0);
2296 /* We may have more errors than our tolerance during the read. */
2297 for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++) {
2300 found_errors = get_rbio_veritical_errors(rbio, sectornr, NULL, NULL);
2301 if (found_errors > rbio->bioc->max_errors) {
2307 rbio_orig_end_io(rbio, errno_to_blk_status(ret));
2310 static void rmw_rbio_work(struct work_struct *work)
2312 struct btrfs_raid_bio *rbio;
2314 rbio = container_of(work, struct btrfs_raid_bio, work);
2315 if (lock_stripe_add(rbio) == 0)
2319 static void rmw_rbio_work_locked(struct work_struct *work)
2321 rmw_rbio(container_of(work, struct btrfs_raid_bio, work));
2325 * The following code is used to scrub/replace the parity stripe
2327 * Caller must have already increased bio_counter for getting @bioc.
2329 * Note: We need make sure all the pages that add into the scrub/replace
2330 * raid bio are correct and not be changed during the scrub/replace. That
2331 * is those pages just hold metadata or file data with checksum.
2334 struct btrfs_raid_bio *raid56_parity_alloc_scrub_rbio(struct bio *bio,
2335 struct btrfs_io_context *bioc,
2336 struct btrfs_device *scrub_dev,
2337 unsigned long *dbitmap, int stripe_nsectors)
2339 struct btrfs_fs_info *fs_info = bioc->fs_info;
2340 struct btrfs_raid_bio *rbio;
2343 rbio = alloc_rbio(fs_info, bioc);
2346 bio_list_add(&rbio->bio_list, bio);
2348 * This is a special bio which is used to hold the completion handler
2349 * and make the scrub rbio is similar to the other types
2351 ASSERT(!bio->bi_iter.bi_size);
2352 rbio->operation = BTRFS_RBIO_PARITY_SCRUB;
2355 * After mapping bioc with BTRFS_MAP_WRITE, parities have been sorted
2356 * to the end position, so this search can start from the first parity
2359 for (i = rbio->nr_data; i < rbio->real_stripes; i++) {
2360 if (bioc->stripes[i].dev == scrub_dev) {
2365 ASSERT(i < rbio->real_stripes);
2367 bitmap_copy(&rbio->dbitmap, dbitmap, stripe_nsectors);
2372 * We just scrub the parity that we have correct data on the same horizontal,
2373 * so we needn't allocate all pages for all the stripes.
2375 static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio)
2377 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
2378 int total_sector_nr;
2380 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
2381 total_sector_nr++) {
2383 int sectornr = total_sector_nr % rbio->stripe_nsectors;
2384 int index = (total_sector_nr * sectorsize) >> PAGE_SHIFT;
2386 if (!test_bit(sectornr, &rbio->dbitmap))
2388 if (rbio->stripe_pages[index])
2390 page = alloc_page(GFP_NOFS);
2393 rbio->stripe_pages[index] = page;
2395 index_stripe_sectors(rbio);
2399 static int finish_parity_scrub(struct btrfs_raid_bio *rbio)
2401 struct btrfs_io_context *bioc = rbio->bioc;
2402 const u32 sectorsize = bioc->fs_info->sectorsize;
2403 void **pointers = rbio->finish_pointers;
2404 unsigned long *pbitmap = &rbio->finish_pbitmap;
2405 int nr_data = rbio->nr_data;
2409 struct sector_ptr p_sector = { 0 };
2410 struct sector_ptr q_sector = { 0 };
2411 struct bio_list bio_list;
2415 bio_list_init(&bio_list);
2417 if (rbio->real_stripes - rbio->nr_data == 1)
2418 has_qstripe = false;
2419 else if (rbio->real_stripes - rbio->nr_data == 2)
2425 * Replace is running and our P/Q stripe is being replaced, then we
2426 * need to duplicate the final write to replace target.
2428 if (bioc->replace_nr_stripes && bioc->replace_stripe_src == rbio->scrubp) {
2430 bitmap_copy(pbitmap, &rbio->dbitmap, rbio->stripe_nsectors);
2434 * Because the higher layers(scrubber) are unlikely to
2435 * use this area of the disk again soon, so don't cache
2438 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2440 p_sector.page = alloc_page(GFP_NOFS);
2444 p_sector.uptodate = 1;
2447 /* RAID6, allocate and map temp space for the Q stripe */
2448 q_sector.page = alloc_page(GFP_NOFS);
2449 if (!q_sector.page) {
2450 __free_page(p_sector.page);
2451 p_sector.page = NULL;
2455 q_sector.uptodate = 1;
2456 pointers[rbio->real_stripes - 1] = kmap_local_page(q_sector.page);
2459 bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors);
2461 /* Map the parity stripe just once */
2462 pointers[nr_data] = kmap_local_page(p_sector.page);
2464 for_each_set_bit(sectornr, &rbio->dbitmap, rbio->stripe_nsectors) {
2465 struct sector_ptr *sector;
2468 /* first collect one page from each data stripe */
2469 for (stripe = 0; stripe < nr_data; stripe++) {
2470 sector = sector_in_rbio(rbio, stripe, sectornr, 0);
2471 pointers[stripe] = kmap_local_page(sector->page) +
2476 /* RAID6, call the library function to fill in our P/Q */
2477 raid6_call.gen_syndrome(rbio->real_stripes, sectorsize,
2481 memcpy(pointers[nr_data], pointers[0], sectorsize);
2482 run_xor(pointers + 1, nr_data - 1, sectorsize);
2485 /* Check scrubbing parity and repair it */
2486 sector = rbio_stripe_sector(rbio, rbio->scrubp, sectornr);
2487 parity = kmap_local_page(sector->page) + sector->pgoff;
2488 if (memcmp(parity, pointers[rbio->scrubp], sectorsize) != 0)
2489 memcpy(parity, pointers[rbio->scrubp], sectorsize);
2491 /* Parity is right, needn't writeback */
2492 bitmap_clear(&rbio->dbitmap, sectornr, 1);
2493 kunmap_local(parity);
2495 for (stripe = nr_data - 1; stripe >= 0; stripe--)
2496 kunmap_local(pointers[stripe]);
2499 kunmap_local(pointers[nr_data]);
2500 __free_page(p_sector.page);
2501 p_sector.page = NULL;
2502 if (q_sector.page) {
2503 kunmap_local(pointers[rbio->real_stripes - 1]);
2504 __free_page(q_sector.page);
2505 q_sector.page = NULL;
2509 * time to start writing. Make bios for everything from the
2510 * higher layers (the bio_list in our rbio) and our p/q. Ignore
2513 for_each_set_bit(sectornr, &rbio->dbitmap, rbio->stripe_nsectors) {
2514 struct sector_ptr *sector;
2516 sector = rbio_stripe_sector(rbio, rbio->scrubp, sectornr);
2517 ret = rbio_add_io_sector(rbio, &bio_list, sector, rbio->scrubp,
2518 sectornr, REQ_OP_WRITE);
2527 * Replace is running and our parity stripe needs to be duplicated to
2528 * the target device. Check we have a valid source stripe number.
2530 ASSERT(rbio->bioc->replace_stripe_src >= 0);
2531 for_each_set_bit(sectornr, pbitmap, rbio->stripe_nsectors) {
2532 struct sector_ptr *sector;
2534 sector = rbio_stripe_sector(rbio, rbio->scrubp, sectornr);
2535 ret = rbio_add_io_sector(rbio, &bio_list, sector,
2537 sectornr, REQ_OP_WRITE);
2543 submit_write_bios(rbio, &bio_list);
2547 bio_list_put(&bio_list);
2551 static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe)
2553 if (stripe >= 0 && stripe < rbio->nr_data)
2558 static int recover_scrub_rbio(struct btrfs_raid_bio *rbio)
2560 void **pointers = NULL;
2561 void **unmap_array = NULL;
2566 * @pointers array stores the pointer for each sector.
2568 * @unmap_array stores copy of pointers that does not get reordered
2569 * during reconstruction so that kunmap_local works.
2571 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
2572 unmap_array = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
2573 if (!pointers || !unmap_array) {
2578 for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) {
2579 int dfail = 0, failp = -1;
2584 found_errors = get_rbio_veritical_errors(rbio, sector_nr,
2586 if (found_errors > rbio->bioc->max_errors) {
2590 if (found_errors == 0)
2593 /* We should have at least one error here. */
2594 ASSERT(faila >= 0 || failb >= 0);
2596 if (is_data_stripe(rbio, faila))
2598 else if (is_parity_stripe(faila))
2601 if (is_data_stripe(rbio, failb))
2603 else if (is_parity_stripe(failb))
2606 * Because we can not use a scrubbing parity to repair the
2607 * data, so the capability of the repair is declined. (In the
2608 * case of RAID5, we can not repair anything.)
2610 if (dfail > rbio->bioc->max_errors - 1) {
2615 * If all data is good, only parity is correctly, just repair
2616 * the parity, no need to recover data stripes.
2622 * Here means we got one corrupted data stripe and one
2623 * corrupted parity on RAID6, if the corrupted parity is
2624 * scrubbing parity, luckily, use the other one to repair the
2625 * data, or we can not repair the data stripe.
2627 if (failp != rbio->scrubp) {
2632 ret = recover_vertical(rbio, sector_nr, pointers, unmap_array);
2642 static int scrub_assemble_read_bios(struct btrfs_raid_bio *rbio)
2644 struct bio_list bio_list = BIO_EMPTY_LIST;
2645 int total_sector_nr;
2648 /* Build a list of bios to read all the missing parts. */
2649 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
2650 total_sector_nr++) {
2651 int sectornr = total_sector_nr % rbio->stripe_nsectors;
2652 int stripe = total_sector_nr / rbio->stripe_nsectors;
2653 struct sector_ptr *sector;
2655 /* No data in the vertical stripe, no need to read. */
2656 if (!test_bit(sectornr, &rbio->dbitmap))
2660 * We want to find all the sectors missing from the rbio and
2661 * read them from the disk. If sector_in_rbio() finds a sector
2662 * in the bio list we don't need to read it off the stripe.
2664 sector = sector_in_rbio(rbio, stripe, sectornr, 1);
2668 sector = rbio_stripe_sector(rbio, stripe, sectornr);
2670 * The bio cache may have handed us an uptodate sector. If so,
2673 if (sector->uptodate)
2676 ret = rbio_add_io_sector(rbio, &bio_list, sector, stripe,
2677 sectornr, REQ_OP_READ);
2679 bio_list_put(&bio_list);
2684 submit_read_wait_bio_list(rbio, &bio_list);
2688 static void scrub_rbio(struct btrfs_raid_bio *rbio)
2693 ret = alloc_rbio_essential_pages(rbio);
2697 bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors);
2699 ret = scrub_assemble_read_bios(rbio);
2703 /* We may have some failures, recover the failed sectors first. */
2704 ret = recover_scrub_rbio(rbio);
2709 * We have every sector properly prepared. Can finish the scrub
2710 * and writeback the good content.
2712 ret = finish_parity_scrub(rbio);
2713 wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0);
2714 for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) {
2717 found_errors = get_rbio_veritical_errors(rbio, sector_nr, NULL, NULL);
2718 if (found_errors > rbio->bioc->max_errors) {
2724 rbio_orig_end_io(rbio, errno_to_blk_status(ret));
2727 static void scrub_rbio_work_locked(struct work_struct *work)
2729 scrub_rbio(container_of(work, struct btrfs_raid_bio, work));
2732 void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio)
2734 if (!lock_stripe_add(rbio))
2735 start_async_work(rbio, scrub_rbio_work_locked);
2739 * This is for scrub call sites where we already have correct data contents.
2740 * This allows us to avoid reading data stripes again.
2742 * Unfortunately here we have to do page copy, other than reusing the pages.
2743 * This is due to the fact rbio has its own page management for its cache.
2745 void raid56_parity_cache_data_pages(struct btrfs_raid_bio *rbio,
2746 struct page **data_pages, u64 data_logical)
2748 const u64 offset_in_full_stripe = data_logical -
2749 rbio->bioc->full_stripe_logical;
2750 const int page_index = offset_in_full_stripe >> PAGE_SHIFT;
2751 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
2752 const u32 sectors_per_page = PAGE_SIZE / sectorsize;
2756 * If we hit ENOMEM temporarily, but later at
2757 * raid56_parity_submit_scrub_rbio() time it succeeded, we just do
2758 * the extra read, not a big deal.
2760 * If we hit ENOMEM later at raid56_parity_submit_scrub_rbio() time,
2761 * the bio would got proper error number set.
2763 ret = alloc_rbio_data_pages(rbio);
2767 /* data_logical must be at stripe boundary and inside the full stripe. */
2768 ASSERT(IS_ALIGNED(offset_in_full_stripe, BTRFS_STRIPE_LEN));
2769 ASSERT(offset_in_full_stripe < (rbio->nr_data << BTRFS_STRIPE_LEN_SHIFT));
2771 for (int page_nr = 0; page_nr < (BTRFS_STRIPE_LEN >> PAGE_SHIFT); page_nr++) {
2772 struct page *dst = rbio->stripe_pages[page_nr + page_index];
2773 struct page *src = data_pages[page_nr];
2775 memcpy_page(dst, 0, src, 0, PAGE_SIZE);
2776 for (int sector_nr = sectors_per_page * page_index;
2777 sector_nr < sectors_per_page * (page_index + 1);
2779 rbio->stripe_sectors[sector_nr].uptodate = true;