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, int need_check);
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->raid_map[0];
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;
412 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
415 table = rbio->bioc->fs_info->stripe_hash_table;
417 spin_lock_irqsave(&table->cache_lock, flags);
418 __remove_rbio_from_cache(rbio);
419 spin_unlock_irqrestore(&table->cache_lock, flags);
423 * remove everything in the cache
425 static void btrfs_clear_rbio_cache(struct btrfs_fs_info *info)
427 struct btrfs_stripe_hash_table *table;
429 struct btrfs_raid_bio *rbio;
431 table = info->stripe_hash_table;
433 spin_lock_irqsave(&table->cache_lock, flags);
434 while (!list_empty(&table->stripe_cache)) {
435 rbio = list_entry(table->stripe_cache.next,
436 struct btrfs_raid_bio,
438 __remove_rbio_from_cache(rbio);
440 spin_unlock_irqrestore(&table->cache_lock, flags);
444 * remove all cached entries and free the hash table
447 void btrfs_free_stripe_hash_table(struct btrfs_fs_info *info)
449 if (!info->stripe_hash_table)
451 btrfs_clear_rbio_cache(info);
452 kvfree(info->stripe_hash_table);
453 info->stripe_hash_table = NULL;
457 * insert an rbio into the stripe cache. It
458 * must have already been prepared by calling
461 * If this rbio was already cached, it gets
462 * moved to the front of the lru.
464 * If the size of the rbio cache is too big, we
467 static void cache_rbio(struct btrfs_raid_bio *rbio)
469 struct btrfs_stripe_hash_table *table;
472 if (!test_bit(RBIO_CACHE_READY_BIT, &rbio->flags))
475 table = rbio->bioc->fs_info->stripe_hash_table;
477 spin_lock_irqsave(&table->cache_lock, flags);
478 spin_lock(&rbio->bio_list_lock);
480 /* bump our ref if we were not in the list before */
481 if (!test_and_set_bit(RBIO_CACHE_BIT, &rbio->flags))
482 refcount_inc(&rbio->refs);
484 if (!list_empty(&rbio->stripe_cache)){
485 list_move(&rbio->stripe_cache, &table->stripe_cache);
487 list_add(&rbio->stripe_cache, &table->stripe_cache);
488 table->cache_size += 1;
491 spin_unlock(&rbio->bio_list_lock);
493 if (table->cache_size > RBIO_CACHE_SIZE) {
494 struct btrfs_raid_bio *found;
496 found = list_entry(table->stripe_cache.prev,
497 struct btrfs_raid_bio,
501 __remove_rbio_from_cache(found);
504 spin_unlock_irqrestore(&table->cache_lock, flags);
508 * helper function to run the xor_blocks api. It is only
509 * able to do MAX_XOR_BLOCKS at a time, so we need to
512 static void run_xor(void **pages, int src_cnt, ssize_t len)
516 void *dest = pages[src_cnt];
519 xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS);
520 xor_blocks(xor_src_cnt, len, dest, pages + src_off);
522 src_cnt -= xor_src_cnt;
523 src_off += xor_src_cnt;
528 * Returns true if the bio list inside this rbio covers an entire stripe (no
531 static int rbio_is_full(struct btrfs_raid_bio *rbio)
534 unsigned long size = rbio->bio_list_bytes;
537 spin_lock_irqsave(&rbio->bio_list_lock, flags);
538 if (size != rbio->nr_data * BTRFS_STRIPE_LEN)
540 BUG_ON(size > rbio->nr_data * BTRFS_STRIPE_LEN);
541 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
547 * returns 1 if it is safe to merge two rbios together.
548 * The merging is safe if the two rbios correspond to
549 * the same stripe and if they are both going in the same
550 * direction (read vs write), and if neither one is
551 * locked for final IO
553 * The caller is responsible for locking such that
554 * rmw_locked is safe to test
556 static int rbio_can_merge(struct btrfs_raid_bio *last,
557 struct btrfs_raid_bio *cur)
559 if (test_bit(RBIO_RMW_LOCKED_BIT, &last->flags) ||
560 test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags))
564 * we can't merge with cached rbios, since the
565 * idea is that when we merge the destination
566 * rbio is going to run our IO for us. We can
567 * steal from cached rbios though, other functions
570 if (test_bit(RBIO_CACHE_BIT, &last->flags) ||
571 test_bit(RBIO_CACHE_BIT, &cur->flags))
574 if (last->bioc->raid_map[0] != cur->bioc->raid_map[0])
577 /* we can't merge with different operations */
578 if (last->operation != cur->operation)
581 * We've need read the full stripe from the drive.
582 * check and repair the parity and write the new results.
584 * We're not allowed to add any new bios to the
585 * bio list here, anyone else that wants to
586 * change this stripe needs to do their own rmw.
588 if (last->operation == BTRFS_RBIO_PARITY_SCRUB)
591 if (last->operation == BTRFS_RBIO_REBUILD_MISSING ||
592 last->operation == BTRFS_RBIO_READ_REBUILD)
598 static unsigned int rbio_stripe_sector_index(const struct btrfs_raid_bio *rbio,
599 unsigned int stripe_nr,
600 unsigned int sector_nr)
602 ASSERT(stripe_nr < rbio->real_stripes);
603 ASSERT(sector_nr < rbio->stripe_nsectors);
605 return stripe_nr * rbio->stripe_nsectors + sector_nr;
608 /* Return a sector from rbio->stripe_sectors, not from the bio list */
609 static struct sector_ptr *rbio_stripe_sector(const struct btrfs_raid_bio *rbio,
610 unsigned int stripe_nr,
611 unsigned int sector_nr)
613 return &rbio->stripe_sectors[rbio_stripe_sector_index(rbio, stripe_nr,
617 /* Grab a sector inside P stripe */
618 static struct sector_ptr *rbio_pstripe_sector(const struct btrfs_raid_bio *rbio,
619 unsigned int sector_nr)
621 return rbio_stripe_sector(rbio, rbio->nr_data, sector_nr);
624 /* Grab a sector inside Q stripe, return NULL if not RAID6 */
625 static struct sector_ptr *rbio_qstripe_sector(const struct btrfs_raid_bio *rbio,
626 unsigned int sector_nr)
628 if (rbio->nr_data + 1 == rbio->real_stripes)
630 return rbio_stripe_sector(rbio, rbio->nr_data + 1, sector_nr);
634 * The first stripe in the table for a logical address
635 * has the lock. rbios are added in one of three ways:
637 * 1) Nobody has the stripe locked yet. The rbio is given
638 * the lock and 0 is returned. The caller must start the IO
641 * 2) Someone has the stripe locked, but we're able to merge
642 * with the lock owner. The rbio is freed and the IO will
643 * start automatically along with the existing rbio. 1 is returned.
645 * 3) Someone has the stripe locked, but we're not able to merge.
646 * The rbio is added to the lock owner's plug list, or merged into
647 * an rbio already on the plug list. When the lock owner unlocks,
648 * the next rbio on the list is run and the IO is started automatically.
651 * If we return 0, the caller still owns the rbio and must continue with
652 * IO submission. If we return 1, the caller must assume the rbio has
653 * already been freed.
655 static noinline int lock_stripe_add(struct btrfs_raid_bio *rbio)
657 struct btrfs_stripe_hash *h;
658 struct btrfs_raid_bio *cur;
659 struct btrfs_raid_bio *pending;
661 struct btrfs_raid_bio *freeit = NULL;
662 struct btrfs_raid_bio *cache_drop = NULL;
665 h = rbio->bioc->fs_info->stripe_hash_table->table + rbio_bucket(rbio);
667 spin_lock_irqsave(&h->lock, flags);
668 list_for_each_entry(cur, &h->hash_list, hash_list) {
669 if (cur->bioc->raid_map[0] != rbio->bioc->raid_map[0])
672 spin_lock(&cur->bio_list_lock);
674 /* Can we steal this cached rbio's pages? */
675 if (bio_list_empty(&cur->bio_list) &&
676 list_empty(&cur->plug_list) &&
677 test_bit(RBIO_CACHE_BIT, &cur->flags) &&
678 !test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) {
679 list_del_init(&cur->hash_list);
680 refcount_dec(&cur->refs);
682 steal_rbio(cur, rbio);
684 spin_unlock(&cur->bio_list_lock);
689 /* Can we merge into the lock owner? */
690 if (rbio_can_merge(cur, rbio)) {
691 merge_rbio(cur, rbio);
692 spin_unlock(&cur->bio_list_lock);
700 * We couldn't merge with the running rbio, see if we can merge
701 * with the pending ones. We don't have to check for rmw_locked
702 * because there is no way they are inside finish_rmw right now
704 list_for_each_entry(pending, &cur->plug_list, plug_list) {
705 if (rbio_can_merge(pending, rbio)) {
706 merge_rbio(pending, rbio);
707 spin_unlock(&cur->bio_list_lock);
715 * No merging, put us on the tail of the plug list, our rbio
716 * will be started with the currently running rbio unlocks
718 list_add_tail(&rbio->plug_list, &cur->plug_list);
719 spin_unlock(&cur->bio_list_lock);
724 refcount_inc(&rbio->refs);
725 list_add(&rbio->hash_list, &h->hash_list);
727 spin_unlock_irqrestore(&h->lock, flags);
729 remove_rbio_from_cache(cache_drop);
731 free_raid_bio(freeit);
735 static void recover_rbio_work_locked(struct work_struct *work);
738 * called as rmw or parity rebuild is completed. If the plug list has more
739 * rbios waiting for this stripe, the next one on the list will be started
741 static noinline void unlock_stripe(struct btrfs_raid_bio *rbio)
744 struct btrfs_stripe_hash *h;
748 bucket = rbio_bucket(rbio);
749 h = rbio->bioc->fs_info->stripe_hash_table->table + bucket;
751 if (list_empty(&rbio->plug_list))
754 spin_lock_irqsave(&h->lock, flags);
755 spin_lock(&rbio->bio_list_lock);
757 if (!list_empty(&rbio->hash_list)) {
759 * if we're still cached and there is no other IO
760 * to perform, just leave this rbio here for others
761 * to steal from later
763 if (list_empty(&rbio->plug_list) &&
764 test_bit(RBIO_CACHE_BIT, &rbio->flags)) {
766 clear_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
767 BUG_ON(!bio_list_empty(&rbio->bio_list));
771 list_del_init(&rbio->hash_list);
772 refcount_dec(&rbio->refs);
775 * we use the plug list to hold all the rbios
776 * waiting for the chance to lock this stripe.
777 * hand the lock over to one of them.
779 if (!list_empty(&rbio->plug_list)) {
780 struct btrfs_raid_bio *next;
781 struct list_head *head = rbio->plug_list.next;
783 next = list_entry(head, struct btrfs_raid_bio,
786 list_del_init(&rbio->plug_list);
788 list_add(&next->hash_list, &h->hash_list);
789 refcount_inc(&next->refs);
790 spin_unlock(&rbio->bio_list_lock);
791 spin_unlock_irqrestore(&h->lock, flags);
793 if (next->operation == BTRFS_RBIO_READ_REBUILD)
794 start_async_work(next, recover_rbio_work_locked);
795 else if (next->operation == BTRFS_RBIO_REBUILD_MISSING) {
796 steal_rbio(rbio, next);
797 start_async_work(next, recover_rbio_work_locked);
798 } else if (next->operation == BTRFS_RBIO_WRITE) {
799 steal_rbio(rbio, next);
800 start_async_work(next, rmw_rbio_work_locked);
801 } else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) {
802 steal_rbio(rbio, next);
803 start_async_work(next, scrub_rbio_work_locked);
810 spin_unlock(&rbio->bio_list_lock);
811 spin_unlock_irqrestore(&h->lock, flags);
815 remove_rbio_from_cache(rbio);
818 static void rbio_endio_bio_list(struct bio *cur, blk_status_t err)
825 cur->bi_status = err;
832 * this frees the rbio and runs through all the bios in the
833 * bio_list and calls end_io on them
835 static void rbio_orig_end_io(struct btrfs_raid_bio *rbio, blk_status_t err)
837 struct bio *cur = bio_list_get(&rbio->bio_list);
840 kfree(rbio->csum_buf);
841 bitmap_free(rbio->csum_bitmap);
842 rbio->csum_buf = NULL;
843 rbio->csum_bitmap = NULL;
846 * Clear the data bitmap, as the rbio may be cached for later usage.
847 * do this before before unlock_stripe() so there will be no new bio
850 bitmap_clear(&rbio->dbitmap, 0, rbio->stripe_nsectors);
853 * At this moment, rbio->bio_list is empty, however since rbio does not
854 * always have RBIO_RMW_LOCKED_BIT set and rbio is still linked on the
855 * hash list, rbio may be merged with others so that rbio->bio_list
857 * Once unlock_stripe() is done, rbio->bio_list will not be updated any
858 * more and we can call bio_endio() on all queued bios.
861 extra = bio_list_get(&rbio->bio_list);
864 rbio_endio_bio_list(cur, err);
866 rbio_endio_bio_list(extra, err);
870 * Get a sector pointer specified by its @stripe_nr and @sector_nr.
872 * @rbio: The raid bio
873 * @stripe_nr: Stripe number, valid range [0, real_stripe)
874 * @sector_nr: Sector number inside the stripe,
875 * valid range [0, stripe_nsectors)
876 * @bio_list_only: Whether to use sectors inside the bio list only.
878 * The read/modify/write code wants to reuse the original bio page as much
879 * as possible, and only use stripe_sectors as fallback.
881 static struct sector_ptr *sector_in_rbio(struct btrfs_raid_bio *rbio,
882 int stripe_nr, int sector_nr,
885 struct sector_ptr *sector;
888 ASSERT(stripe_nr >= 0 && stripe_nr < rbio->real_stripes);
889 ASSERT(sector_nr >= 0 && sector_nr < rbio->stripe_nsectors);
891 index = stripe_nr * rbio->stripe_nsectors + sector_nr;
892 ASSERT(index >= 0 && index < rbio->nr_sectors);
894 spin_lock_irq(&rbio->bio_list_lock);
895 sector = &rbio->bio_sectors[index];
896 if (sector->page || bio_list_only) {
897 /* Don't return sector without a valid page pointer */
900 spin_unlock_irq(&rbio->bio_list_lock);
903 spin_unlock_irq(&rbio->bio_list_lock);
905 return &rbio->stripe_sectors[index];
909 * allocation and initial setup for the btrfs_raid_bio. Not
910 * this does not allocate any pages for rbio->pages.
912 static struct btrfs_raid_bio *alloc_rbio(struct btrfs_fs_info *fs_info,
913 struct btrfs_io_context *bioc)
915 const unsigned int real_stripes = bioc->num_stripes - bioc->num_tgtdevs;
916 const unsigned int stripe_npages = BTRFS_STRIPE_LEN >> PAGE_SHIFT;
917 const unsigned int num_pages = stripe_npages * real_stripes;
918 const unsigned int stripe_nsectors =
919 BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits;
920 const unsigned int num_sectors = stripe_nsectors * real_stripes;
921 struct btrfs_raid_bio *rbio;
923 /* PAGE_SIZE must also be aligned to sectorsize for subpage support */
924 ASSERT(IS_ALIGNED(PAGE_SIZE, fs_info->sectorsize));
926 * Our current stripe len should be fixed to 64k thus stripe_nsectors
927 * (at most 16) should be no larger than BITS_PER_LONG.
929 ASSERT(stripe_nsectors <= BITS_PER_LONG);
931 rbio = kzalloc(sizeof(*rbio), GFP_NOFS);
933 return ERR_PTR(-ENOMEM);
934 rbio->stripe_pages = kcalloc(num_pages, sizeof(struct page *),
936 rbio->bio_sectors = kcalloc(num_sectors, sizeof(struct sector_ptr),
938 rbio->stripe_sectors = kcalloc(num_sectors, sizeof(struct sector_ptr),
940 rbio->finish_pointers = kcalloc(real_stripes, sizeof(void *), GFP_NOFS);
941 rbio->error_bitmap = bitmap_zalloc(num_sectors, GFP_NOFS);
943 if (!rbio->stripe_pages || !rbio->bio_sectors || !rbio->stripe_sectors ||
944 !rbio->finish_pointers || !rbio->error_bitmap) {
945 free_raid_bio_pointers(rbio);
947 return ERR_PTR(-ENOMEM);
950 bio_list_init(&rbio->bio_list);
951 init_waitqueue_head(&rbio->io_wait);
952 INIT_LIST_HEAD(&rbio->plug_list);
953 spin_lock_init(&rbio->bio_list_lock);
954 INIT_LIST_HEAD(&rbio->stripe_cache);
955 INIT_LIST_HEAD(&rbio->hash_list);
956 btrfs_get_bioc(bioc);
958 rbio->nr_pages = num_pages;
959 rbio->nr_sectors = num_sectors;
960 rbio->real_stripes = real_stripes;
961 rbio->stripe_npages = stripe_npages;
962 rbio->stripe_nsectors = stripe_nsectors;
963 refcount_set(&rbio->refs, 1);
964 atomic_set(&rbio->stripes_pending, 0);
966 ASSERT(btrfs_nr_parity_stripes(bioc->map_type));
967 rbio->nr_data = real_stripes - btrfs_nr_parity_stripes(bioc->map_type);
972 /* allocate pages for all the stripes in the bio, including parity */
973 static int alloc_rbio_pages(struct btrfs_raid_bio *rbio)
977 ret = btrfs_alloc_page_array(rbio->nr_pages, rbio->stripe_pages);
980 /* Mapping all sectors */
981 index_stripe_sectors(rbio);
985 /* only allocate pages for p/q stripes */
986 static int alloc_rbio_parity_pages(struct btrfs_raid_bio *rbio)
988 const int data_pages = rbio->nr_data * rbio->stripe_npages;
991 ret = btrfs_alloc_page_array(rbio->nr_pages - data_pages,
992 rbio->stripe_pages + data_pages);
996 index_stripe_sectors(rbio);
1001 * Return the total number of errors found in the vertical stripe of @sector_nr.
1003 * @faila and @failb will also be updated to the first and second stripe
1004 * number of the errors.
1006 static int get_rbio_veritical_errors(struct btrfs_raid_bio *rbio, int sector_nr,
1007 int *faila, int *failb)
1010 int found_errors = 0;
1012 if (faila || failb) {
1014 * Both @faila and @failb should be valid pointers if any of
1015 * them is specified.
1017 ASSERT(faila && failb);
1022 for (stripe_nr = 0; stripe_nr < rbio->real_stripes; stripe_nr++) {
1023 int total_sector_nr = stripe_nr * rbio->stripe_nsectors + sector_nr;
1025 if (test_bit(total_sector_nr, rbio->error_bitmap)) {
1028 /* Update faila and failb. */
1031 else if (*failb < 0)
1036 return found_errors;
1040 * Add a single sector @sector into our list of bios for IO.
1042 * Return 0 if everything went well.
1043 * Return <0 for error.
1045 static int rbio_add_io_sector(struct btrfs_raid_bio *rbio,
1046 struct bio_list *bio_list,
1047 struct sector_ptr *sector,
1048 unsigned int stripe_nr,
1049 unsigned int sector_nr,
1052 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
1053 struct bio *last = bio_list->tail;
1056 struct btrfs_io_stripe *stripe;
1060 * Note: here stripe_nr has taken device replace into consideration,
1061 * thus it can be larger than rbio->real_stripe.
1062 * So here we check against bioc->num_stripes, not rbio->real_stripes.
1064 ASSERT(stripe_nr >= 0 && stripe_nr < rbio->bioc->num_stripes);
1065 ASSERT(sector_nr >= 0 && sector_nr < rbio->stripe_nsectors);
1066 ASSERT(sector->page);
1068 stripe = &rbio->bioc->stripes[stripe_nr];
1069 disk_start = stripe->physical + sector_nr * sectorsize;
1071 /* if the device is missing, just fail this stripe */
1072 if (!stripe->dev->bdev) {
1075 set_bit(stripe_nr * rbio->stripe_nsectors + sector_nr,
1076 rbio->error_bitmap);
1078 /* Check if we have reached tolerance early. */
1079 found_errors = get_rbio_veritical_errors(rbio, sector_nr,
1081 if (found_errors > rbio->bioc->max_errors)
1086 /* see if we can add this page onto our existing bio */
1088 u64 last_end = last->bi_iter.bi_sector << 9;
1089 last_end += last->bi_iter.bi_size;
1092 * we can't merge these if they are from different
1093 * devices or if they are not contiguous
1095 if (last_end == disk_start && !last->bi_status &&
1096 last->bi_bdev == stripe->dev->bdev) {
1097 ret = bio_add_page(last, sector->page, sectorsize,
1099 if (ret == sectorsize)
1104 /* put a new bio on the list */
1105 bio = bio_alloc(stripe->dev->bdev,
1106 max(BTRFS_STRIPE_LEN >> PAGE_SHIFT, 1),
1108 bio->bi_iter.bi_sector = disk_start >> 9;
1109 bio->bi_private = rbio;
1111 bio_add_page(bio, sector->page, sectorsize, sector->pgoff);
1112 bio_list_add(bio_list, bio);
1116 static void index_one_bio(struct btrfs_raid_bio *rbio, struct bio *bio)
1118 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
1119 struct bio_vec bvec;
1120 struct bvec_iter iter;
1121 u32 offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) -
1122 rbio->bioc->raid_map[0];
1124 bio_for_each_segment(bvec, bio, iter) {
1127 for (bvec_offset = 0; bvec_offset < bvec.bv_len;
1128 bvec_offset += sectorsize, offset += sectorsize) {
1129 int index = offset / sectorsize;
1130 struct sector_ptr *sector = &rbio->bio_sectors[index];
1132 sector->page = bvec.bv_page;
1133 sector->pgoff = bvec.bv_offset + bvec_offset;
1134 ASSERT(sector->pgoff < PAGE_SIZE);
1140 * helper function to walk our bio list and populate the bio_pages array with
1141 * the result. This seems expensive, but it is faster than constantly
1142 * searching through the bio list as we setup the IO in finish_rmw or stripe
1145 * This must be called before you trust the answers from page_in_rbio
1147 static void index_rbio_pages(struct btrfs_raid_bio *rbio)
1151 spin_lock_irq(&rbio->bio_list_lock);
1152 bio_list_for_each(bio, &rbio->bio_list)
1153 index_one_bio(rbio, bio);
1155 spin_unlock_irq(&rbio->bio_list_lock);
1158 static void bio_get_trace_info(struct btrfs_raid_bio *rbio, struct bio *bio,
1159 struct raid56_bio_trace_info *trace_info)
1161 const struct btrfs_io_context *bioc = rbio->bioc;
1166 /* We rely on bio->bi_bdev to find the stripe number. */
1170 for (i = 0; i < bioc->num_stripes; i++) {
1171 if (bio->bi_bdev != bioc->stripes[i].dev->bdev)
1173 trace_info->stripe_nr = i;
1174 trace_info->devid = bioc->stripes[i].dev->devid;
1175 trace_info->offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) -
1176 bioc->stripes[i].physical;
1181 trace_info->devid = -1;
1182 trace_info->offset = -1;
1183 trace_info->stripe_nr = -1;
1186 static inline void bio_list_put(struct bio_list *bio_list)
1190 while ((bio = bio_list_pop(bio_list)))
1194 /* Generate PQ for one vertical stripe. */
1195 static void generate_pq_vertical(struct btrfs_raid_bio *rbio, int sectornr)
1197 void **pointers = rbio->finish_pointers;
1198 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
1199 struct sector_ptr *sector;
1201 const bool has_qstripe = rbio->bioc->map_type & BTRFS_BLOCK_GROUP_RAID6;
1203 /* First collect one sector from each data stripe */
1204 for (stripe = 0; stripe < rbio->nr_data; stripe++) {
1205 sector = sector_in_rbio(rbio, stripe, sectornr, 0);
1206 pointers[stripe] = kmap_local_page(sector->page) +
1210 /* Then add the parity stripe */
1211 sector = rbio_pstripe_sector(rbio, sectornr);
1212 sector->uptodate = 1;
1213 pointers[stripe++] = kmap_local_page(sector->page) + sector->pgoff;
1217 * RAID6, add the qstripe and call the library function
1218 * to fill in our p/q
1220 sector = rbio_qstripe_sector(rbio, sectornr);
1221 sector->uptodate = 1;
1222 pointers[stripe++] = kmap_local_page(sector->page) +
1225 raid6_call.gen_syndrome(rbio->real_stripes, sectorsize,
1229 memcpy(pointers[rbio->nr_data], pointers[0], sectorsize);
1230 run_xor(pointers + 1, rbio->nr_data - 1, sectorsize);
1232 for (stripe = stripe - 1; stripe >= 0; stripe--)
1233 kunmap_local(pointers[stripe]);
1236 static int rmw_assemble_write_bios(struct btrfs_raid_bio *rbio,
1237 struct bio_list *bio_list)
1239 /* The total sector number inside the full stripe. */
1240 int total_sector_nr;
1245 ASSERT(bio_list_size(bio_list) == 0);
1247 /* We should have at least one data sector. */
1248 ASSERT(bitmap_weight(&rbio->dbitmap, rbio->stripe_nsectors));
1251 * Reset errors, as we may have errors inherited from from degraded
1254 bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors);
1257 * Start assembly. Make bios for everything from the higher layers (the
1258 * bio_list in our rbio) and our P/Q. Ignore everything else.
1260 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
1261 total_sector_nr++) {
1262 struct sector_ptr *sector;
1264 stripe = total_sector_nr / rbio->stripe_nsectors;
1265 sectornr = total_sector_nr % rbio->stripe_nsectors;
1267 /* This vertical stripe has no data, skip it. */
1268 if (!test_bit(sectornr, &rbio->dbitmap))
1271 if (stripe < rbio->nr_data) {
1272 sector = sector_in_rbio(rbio, stripe, sectornr, 1);
1276 sector = rbio_stripe_sector(rbio, stripe, sectornr);
1279 ret = rbio_add_io_sector(rbio, bio_list, sector, stripe,
1280 sectornr, REQ_OP_WRITE);
1285 if (likely(!rbio->bioc->num_tgtdevs))
1288 /* Make a copy for the replace target device. */
1289 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
1290 total_sector_nr++) {
1291 struct sector_ptr *sector;
1293 stripe = total_sector_nr / rbio->stripe_nsectors;
1294 sectornr = total_sector_nr % rbio->stripe_nsectors;
1296 if (!rbio->bioc->tgtdev_map[stripe]) {
1298 * We can skip the whole stripe completely, note
1299 * total_sector_nr will be increased by one anyway.
1301 ASSERT(sectornr == 0);
1302 total_sector_nr += rbio->stripe_nsectors - 1;
1306 /* This vertical stripe has no data, skip it. */
1307 if (!test_bit(sectornr, &rbio->dbitmap))
1310 if (stripe < rbio->nr_data) {
1311 sector = sector_in_rbio(rbio, stripe, sectornr, 1);
1315 sector = rbio_stripe_sector(rbio, stripe, sectornr);
1318 ret = rbio_add_io_sector(rbio, bio_list, sector,
1319 rbio->bioc->tgtdev_map[stripe],
1320 sectornr, REQ_OP_WRITE);
1327 bio_list_put(bio_list);
1331 static void set_rbio_range_error(struct btrfs_raid_bio *rbio, struct bio *bio)
1333 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1334 u32 offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) -
1335 rbio->bioc->raid_map[0];
1336 int total_nr_sector = offset >> fs_info->sectorsize_bits;
1338 ASSERT(total_nr_sector < rbio->nr_data * rbio->stripe_nsectors);
1340 bitmap_set(rbio->error_bitmap, total_nr_sector,
1341 bio->bi_iter.bi_size >> fs_info->sectorsize_bits);
1344 * Special handling for raid56_alloc_missing_rbio() used by
1345 * scrub/replace. Unlike call path in raid56_parity_recover(), they
1346 * pass an empty bio here. Thus we have to find out the missing device
1347 * and mark the stripe error instead.
1349 if (bio->bi_iter.bi_size == 0) {
1350 bool found_missing = false;
1353 for (stripe_nr = 0; stripe_nr < rbio->real_stripes; stripe_nr++) {
1354 if (!rbio->bioc->stripes[stripe_nr].dev->bdev) {
1355 found_missing = true;
1356 bitmap_set(rbio->error_bitmap,
1357 stripe_nr * rbio->stripe_nsectors,
1358 rbio->stripe_nsectors);
1361 ASSERT(found_missing);
1366 * For subpage case, we can no longer set page Up-to-date directly for
1367 * stripe_pages[], thus we need to locate the sector.
1369 static struct sector_ptr *find_stripe_sector(struct btrfs_raid_bio *rbio,
1375 for (i = 0; i < rbio->nr_sectors; i++) {
1376 struct sector_ptr *sector = &rbio->stripe_sectors[i];
1378 if (sector->page == page && sector->pgoff == pgoff)
1385 * this sets each page in the bio uptodate. It should only be used on private
1386 * rbio pages, nothing that comes in from the higher layers
1388 static void set_bio_pages_uptodate(struct btrfs_raid_bio *rbio, struct bio *bio)
1390 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
1391 struct bio_vec *bvec;
1392 struct bvec_iter_all iter_all;
1394 ASSERT(!bio_flagged(bio, BIO_CLONED));
1396 bio_for_each_segment_all(bvec, bio, iter_all) {
1397 struct sector_ptr *sector;
1400 for (pgoff = bvec->bv_offset; pgoff - bvec->bv_offset < bvec->bv_len;
1401 pgoff += sectorsize) {
1402 sector = find_stripe_sector(rbio, bvec->bv_page, pgoff);
1405 sector->uptodate = 1;
1410 static int get_bio_sector_nr(struct btrfs_raid_bio *rbio, struct bio *bio)
1412 struct bio_vec *bv = bio_first_bvec_all(bio);
1415 for (i = 0; i < rbio->nr_sectors; i++) {
1416 struct sector_ptr *sector;
1418 sector = &rbio->stripe_sectors[i];
1419 if (sector->page == bv->bv_page && sector->pgoff == bv->bv_offset)
1421 sector = &rbio->bio_sectors[i];
1422 if (sector->page == bv->bv_page && sector->pgoff == bv->bv_offset)
1425 ASSERT(i < rbio->nr_sectors);
1429 static void rbio_update_error_bitmap(struct btrfs_raid_bio *rbio, struct bio *bio)
1431 int total_sector_nr = get_bio_sector_nr(rbio, bio);
1433 struct bio_vec *bvec;
1436 bio_for_each_bvec_all(bvec, bio, i)
1437 bio_size += bvec->bv_len;
1440 * Since we can have multiple bios touching the error_bitmap, we cannot
1441 * call bitmap_set() without protection.
1443 * Instead use set_bit() for each bit, as set_bit() itself is atomic.
1445 for (i = total_sector_nr; i < total_sector_nr +
1446 (bio_size >> rbio->bioc->fs_info->sectorsize_bits); i++)
1447 set_bit(i, rbio->error_bitmap);
1450 /* Verify the data sectors at read time. */
1451 static void verify_bio_data_sectors(struct btrfs_raid_bio *rbio,
1454 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1455 int total_sector_nr = get_bio_sector_nr(rbio, bio);
1456 struct bio_vec *bvec;
1457 struct bvec_iter_all iter_all;
1459 /* No data csum for the whole stripe, no need to verify. */
1460 if (!rbio->csum_bitmap || !rbio->csum_buf)
1463 /* P/Q stripes, they have no data csum to verify against. */
1464 if (total_sector_nr >= rbio->nr_data * rbio->stripe_nsectors)
1467 bio_for_each_segment_all(bvec, bio, iter_all) {
1470 for (bv_offset = bvec->bv_offset;
1471 bv_offset < bvec->bv_offset + bvec->bv_len;
1472 bv_offset += fs_info->sectorsize, total_sector_nr++) {
1473 u8 csum_buf[BTRFS_CSUM_SIZE];
1474 u8 *expected_csum = rbio->csum_buf +
1475 total_sector_nr * fs_info->csum_size;
1478 /* No csum for this sector, skip to the next sector. */
1479 if (!test_bit(total_sector_nr, rbio->csum_bitmap))
1482 ret = btrfs_check_sector_csum(fs_info, bvec->bv_page,
1483 bv_offset, csum_buf, expected_csum);
1485 set_bit(total_sector_nr, rbio->error_bitmap);
1490 static void raid_wait_read_end_io(struct bio *bio)
1492 struct btrfs_raid_bio *rbio = bio->bi_private;
1494 if (bio->bi_status) {
1495 rbio_update_error_bitmap(rbio, bio);
1497 set_bio_pages_uptodate(rbio, bio);
1498 verify_bio_data_sectors(rbio, bio);
1502 if (atomic_dec_and_test(&rbio->stripes_pending))
1503 wake_up(&rbio->io_wait);
1506 static void submit_read_wait_bio_list(struct btrfs_raid_bio *rbio,
1507 struct bio_list *bio_list)
1511 atomic_set(&rbio->stripes_pending, bio_list_size(bio_list));
1512 while ((bio = bio_list_pop(bio_list))) {
1513 bio->bi_end_io = raid_wait_read_end_io;
1515 if (trace_raid56_scrub_read_recover_enabled()) {
1516 struct raid56_bio_trace_info trace_info = { 0 };
1518 bio_get_trace_info(rbio, bio, &trace_info);
1519 trace_raid56_scrub_read_recover(rbio, bio, &trace_info);
1524 wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0);
1527 static int alloc_rbio_data_pages(struct btrfs_raid_bio *rbio)
1529 const int data_pages = rbio->nr_data * rbio->stripe_npages;
1532 ret = btrfs_alloc_page_array(data_pages, rbio->stripe_pages);
1536 index_stripe_sectors(rbio);
1541 * We use plugging call backs to collect full stripes.
1542 * Any time we get a partial stripe write while plugged
1543 * we collect it into a list. When the unplug comes down,
1544 * we sort the list by logical block number and merge
1545 * everything we can into the same rbios
1547 struct btrfs_plug_cb {
1548 struct blk_plug_cb cb;
1549 struct btrfs_fs_info *info;
1550 struct list_head rbio_list;
1551 struct work_struct work;
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->raid_map[0];
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 rbio->operation == BTRFS_RBIO_REBUILD_MISSING)) {
1698 sector = sector_in_rbio(rbio, stripe_nr, sector_nr, 0);
1700 sector = rbio_stripe_sector(rbio, stripe_nr, sector_nr);
1703 ASSERT(sector->page);
1705 csum_expected = rbio->csum_buf +
1706 (stripe_nr * rbio->stripe_nsectors + sector_nr) *
1708 ret = btrfs_check_sector_csum(fs_info, sector->page, sector->pgoff,
1709 csum_buf, csum_expected);
1714 * Recover a vertical stripe specified by @sector_nr.
1715 * @*pointers are the pre-allocated pointers by the caller, so we don't
1716 * need to allocate/free the pointers again and again.
1718 static int recover_vertical(struct btrfs_raid_bio *rbio, int sector_nr,
1719 void **pointers, void **unmap_array)
1721 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1722 struct sector_ptr *sector;
1723 const u32 sectorsize = fs_info->sectorsize;
1731 * Now we just use bitmap to mark the horizontal stripes in
1732 * which we have data when doing parity scrub.
1734 if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB &&
1735 !test_bit(sector_nr, &rbio->dbitmap))
1738 found_errors = get_rbio_veritical_errors(rbio, sector_nr, &faila,
1741 * No errors in the vertical stripe, skip it. Can happen for recovery
1742 * which only part of a stripe failed csum check.
1747 if (found_errors > rbio->bioc->max_errors)
1751 * Setup our array of pointers with sectors from each stripe
1753 * NOTE: store a duplicate array of pointers to preserve the
1756 for (stripe_nr = 0; stripe_nr < rbio->real_stripes; stripe_nr++) {
1758 * If we're rebuilding a read, we have to use pages from the
1759 * bio list if possible.
1761 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1762 rbio->operation == BTRFS_RBIO_REBUILD_MISSING)) {
1763 sector = sector_in_rbio(rbio, stripe_nr, sector_nr, 0);
1765 sector = rbio_stripe_sector(rbio, stripe_nr, sector_nr);
1767 ASSERT(sector->page);
1768 pointers[stripe_nr] = kmap_local_page(sector->page) +
1770 unmap_array[stripe_nr] = pointers[stripe_nr];
1773 /* All raid6 handling here */
1774 if (rbio->bioc->map_type & BTRFS_BLOCK_GROUP_RAID6) {
1775 /* Single failure, rebuild from parity raid5 style */
1777 if (faila == rbio->nr_data)
1779 * Just the P stripe has failed, without
1780 * a bad data or Q stripe.
1781 * We have nothing to do, just skip the
1782 * recovery for this stripe.
1786 * a single failure in raid6 is rebuilt
1787 * in the pstripe code below
1793 * If the q stripe is failed, do a pstripe reconstruction from
1795 * If both the q stripe and the P stripe are failed, we're
1796 * here due to a crc mismatch and we can't give them the
1799 if (rbio->bioc->raid_map[failb] == RAID6_Q_STRIPE) {
1800 if (rbio->bioc->raid_map[faila] ==
1803 * Only P and Q are corrupted.
1804 * We only care about data stripes recovery,
1805 * can skip this vertical stripe.
1809 * Otherwise we have one bad data stripe and
1810 * a good P stripe. raid5!
1815 if (rbio->bioc->raid_map[failb] == RAID5_P_STRIPE) {
1816 raid6_datap_recov(rbio->real_stripes, sectorsize,
1819 raid6_2data_recov(rbio->real_stripes, sectorsize,
1820 faila, failb, pointers);
1825 /* Rebuild from P stripe here (raid5 or raid6). */
1826 ASSERT(failb == -1);
1828 /* Copy parity block into failed block to start with */
1829 memcpy(pointers[faila], pointers[rbio->nr_data], sectorsize);
1831 /* Rearrange the pointer array */
1832 p = pointers[faila];
1833 for (stripe_nr = faila; stripe_nr < rbio->nr_data - 1;
1835 pointers[stripe_nr] = pointers[stripe_nr + 1];
1836 pointers[rbio->nr_data - 1] = p;
1838 /* Xor in the rest */
1839 run_xor(pointers, rbio->nr_data - 1, sectorsize);
1844 * No matter if this is a RMW or recovery, we should have all
1845 * failed sectors repaired in the vertical stripe, thus they are now
1847 * Especially if we determine to cache the rbio, we need to
1848 * have at least all data sectors uptodate.
1850 * If possible, also check if the repaired sector matches its data
1854 ret = verify_one_sector(rbio, faila, sector_nr);
1858 sector = rbio_stripe_sector(rbio, faila, sector_nr);
1859 sector->uptodate = 1;
1862 ret = verify_one_sector(rbio, failb, sector_nr);
1866 sector = rbio_stripe_sector(rbio, failb, sector_nr);
1867 sector->uptodate = 1;
1871 for (stripe_nr = rbio->real_stripes - 1; stripe_nr >= 0; stripe_nr--)
1872 kunmap_local(unmap_array[stripe_nr]);
1876 static int recover_sectors(struct btrfs_raid_bio *rbio)
1878 void **pointers = NULL;
1879 void **unmap_array = NULL;
1884 * @pointers array stores the pointer for each sector.
1886 * @unmap_array stores copy of pointers that does not get reordered
1887 * during reconstruction so that kunmap_local works.
1889 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
1890 unmap_array = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
1891 if (!pointers || !unmap_array) {
1896 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1897 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
1898 spin_lock_irq(&rbio->bio_list_lock);
1899 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1900 spin_unlock_irq(&rbio->bio_list_lock);
1903 index_rbio_pages(rbio);
1905 for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++) {
1906 ret = recover_vertical(rbio, sectornr, pointers, unmap_array);
1917 static void recover_rbio(struct btrfs_raid_bio *rbio)
1919 struct bio_list bio_list = BIO_EMPTY_LIST;
1920 int total_sector_nr;
1924 * Either we're doing recover for a read failure or degraded write,
1925 * caller should have set error bitmap correctly.
1927 ASSERT(bitmap_weight(rbio->error_bitmap, rbio->nr_sectors));
1929 /* For recovery, we need to read all sectors including P/Q. */
1930 ret = alloc_rbio_pages(rbio);
1934 index_rbio_pages(rbio);
1937 * Read everything that hasn't failed. However this time we will
1938 * not trust any cached sector.
1939 * As we may read out some stale data but higher layer is not reading
1942 * So here we always re-read everything in recovery path.
1944 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
1945 total_sector_nr++) {
1946 int stripe = total_sector_nr / rbio->stripe_nsectors;
1947 int sectornr = total_sector_nr % rbio->stripe_nsectors;
1948 struct sector_ptr *sector;
1951 * Skip the range which has error. It can be a range which is
1952 * marked error (for csum mismatch), or it can be a missing
1955 if (!rbio->bioc->stripes[stripe].dev->bdev ||
1956 test_bit(total_sector_nr, rbio->error_bitmap)) {
1958 * Also set the error bit for missing device, which
1959 * may not yet have its error bit set.
1961 set_bit(total_sector_nr, rbio->error_bitmap);
1965 sector = rbio_stripe_sector(rbio, stripe, sectornr);
1966 ret = rbio_add_io_sector(rbio, &bio_list, sector, stripe,
1967 sectornr, REQ_OP_READ);
1969 bio_list_put(&bio_list);
1974 submit_read_wait_bio_list(rbio, &bio_list);
1975 ret = recover_sectors(rbio);
1977 rbio_orig_end_io(rbio, errno_to_blk_status(ret));
1980 static void recover_rbio_work(struct work_struct *work)
1982 struct btrfs_raid_bio *rbio;
1984 rbio = container_of(work, struct btrfs_raid_bio, work);
1985 if (!lock_stripe_add(rbio))
1989 static void recover_rbio_work_locked(struct work_struct *work)
1991 recover_rbio(container_of(work, struct btrfs_raid_bio, work));
1994 static void set_rbio_raid6_extra_error(struct btrfs_raid_bio *rbio, int mirror_num)
2000 * This is for RAID6 extra recovery tries, thus mirror number should
2002 * Mirror 1 means read from data stripes. Mirror 2 means rebuild using
2005 ASSERT(mirror_num > 2);
2006 for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) {
2011 found_errors = get_rbio_veritical_errors(rbio, sector_nr,
2013 /* This vertical stripe doesn't have errors. */
2018 * If we found errors, there should be only one error marked
2019 * by previous set_rbio_range_error().
2021 ASSERT(found_errors == 1);
2024 /* Now select another stripe to mark as error. */
2025 failb = rbio->real_stripes - (mirror_num - 1);
2029 /* Set the extra bit in error bitmap. */
2031 set_bit(failb * rbio->stripe_nsectors + sector_nr,
2032 rbio->error_bitmap);
2035 /* We should found at least one vertical stripe with error.*/
2040 * the main entry point for reads from the higher layers. This
2041 * is really only called when the normal read path had a failure,
2042 * so we assume the bio they send down corresponds to a failed part
2045 void raid56_parity_recover(struct bio *bio, struct btrfs_io_context *bioc,
2048 struct btrfs_fs_info *fs_info = bioc->fs_info;
2049 struct btrfs_raid_bio *rbio;
2051 rbio = alloc_rbio(fs_info, bioc);
2053 bio->bi_status = errno_to_blk_status(PTR_ERR(rbio));
2058 rbio->operation = BTRFS_RBIO_READ_REBUILD;
2059 rbio_add_bio(rbio, bio);
2061 set_rbio_range_error(rbio, bio);
2065 * for 'mirror == 2', reconstruct from all other stripes.
2066 * for 'mirror_num > 2', select a stripe to fail on every retry.
2069 set_rbio_raid6_extra_error(rbio, mirror_num);
2071 start_async_work(rbio, recover_rbio_work);
2074 static void fill_data_csums(struct btrfs_raid_bio *rbio)
2076 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
2077 struct btrfs_root *csum_root = btrfs_csum_root(fs_info,
2078 rbio->bioc->raid_map[0]);
2079 const u64 start = rbio->bioc->raid_map[0];
2080 const u32 len = (rbio->nr_data * rbio->stripe_nsectors) <<
2081 fs_info->sectorsize_bits;
2084 /* The rbio should not have its csum buffer initialized. */
2085 ASSERT(!rbio->csum_buf && !rbio->csum_bitmap);
2088 * Skip the csum search if:
2090 * - The rbio doesn't belong to data block groups
2091 * Then we are doing IO for tree blocks, no need to search csums.
2093 * - The rbio belongs to mixed block groups
2094 * This is to avoid deadlock, as we're already holding the full
2095 * stripe lock, if we trigger a metadata read, and it needs to do
2096 * raid56 recovery, we will deadlock.
2098 if (!(rbio->bioc->map_type & BTRFS_BLOCK_GROUP_DATA) ||
2099 rbio->bioc->map_type & BTRFS_BLOCK_GROUP_METADATA)
2102 rbio->csum_buf = kzalloc(rbio->nr_data * rbio->stripe_nsectors *
2103 fs_info->csum_size, GFP_NOFS);
2104 rbio->csum_bitmap = bitmap_zalloc(rbio->nr_data * rbio->stripe_nsectors,
2106 if (!rbio->csum_buf || !rbio->csum_bitmap) {
2111 ret = btrfs_lookup_csums_bitmap(csum_root, start, start + len - 1,
2112 rbio->csum_buf, rbio->csum_bitmap);
2115 if (bitmap_empty(rbio->csum_bitmap, len >> fs_info->sectorsize_bits))
2121 * We failed to allocate memory or grab the csum, but it's not fatal,
2122 * we can still continue. But better to warn users that RMW is no
2123 * longer safe for this particular sub-stripe write.
2125 btrfs_warn_rl(fs_info,
2126 "sub-stripe write for full stripe %llu is not safe, failed to get csum: %d",
2127 rbio->bioc->raid_map[0], ret);
2129 kfree(rbio->csum_buf);
2130 bitmap_free(rbio->csum_bitmap);
2131 rbio->csum_buf = NULL;
2132 rbio->csum_bitmap = NULL;
2135 static int rmw_read_wait_recover(struct btrfs_raid_bio *rbio)
2137 struct bio_list bio_list = BIO_EMPTY_LIST;
2138 int total_sector_nr;
2142 * Fill the data csums we need for data verification. We need to fill
2143 * the csum_bitmap/csum_buf first, as our endio function will try to
2144 * verify the data sectors.
2146 fill_data_csums(rbio);
2149 * Build a list of bios to read all sectors (including data and P/Q).
2151 * This behavior is to compensate the later csum verification and recovery.
2153 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
2154 total_sector_nr++) {
2155 struct sector_ptr *sector;
2156 int stripe = total_sector_nr / rbio->stripe_nsectors;
2157 int sectornr = total_sector_nr % rbio->stripe_nsectors;
2159 sector = rbio_stripe_sector(rbio, stripe, sectornr);
2160 ret = rbio_add_io_sector(rbio, &bio_list, sector,
2161 stripe, sectornr, REQ_OP_READ);
2163 bio_list_put(&bio_list);
2169 * We may or may not have any corrupted sectors (including missing dev
2170 * and csum mismatch), just let recover_sectors() to handle them all.
2172 submit_read_wait_bio_list(rbio, &bio_list);
2173 return recover_sectors(rbio);
2176 static void raid_wait_write_end_io(struct bio *bio)
2178 struct btrfs_raid_bio *rbio = bio->bi_private;
2179 blk_status_t err = bio->bi_status;
2182 rbio_update_error_bitmap(rbio, bio);
2184 if (atomic_dec_and_test(&rbio->stripes_pending))
2185 wake_up(&rbio->io_wait);
2188 static void submit_write_bios(struct btrfs_raid_bio *rbio,
2189 struct bio_list *bio_list)
2193 atomic_set(&rbio->stripes_pending, bio_list_size(bio_list));
2194 while ((bio = bio_list_pop(bio_list))) {
2195 bio->bi_end_io = raid_wait_write_end_io;
2197 if (trace_raid56_write_stripe_enabled()) {
2198 struct raid56_bio_trace_info trace_info = { 0 };
2200 bio_get_trace_info(rbio, bio, &trace_info);
2201 trace_raid56_write_stripe(rbio, bio, &trace_info);
2208 * To determine if we need to read any sector from the disk.
2209 * Should only be utilized in RMW path, to skip cached rbio.
2211 static bool need_read_stripe_sectors(struct btrfs_raid_bio *rbio)
2215 for (i = 0; i < rbio->nr_data * rbio->stripe_nsectors; i++) {
2216 struct sector_ptr *sector = &rbio->stripe_sectors[i];
2219 * We have a sector which doesn't have page nor uptodate,
2220 * thus this rbio can not be cached one, as cached one must
2221 * have all its data sectors present and uptodate.
2223 if (!sector->page || !sector->uptodate)
2229 static void rmw_rbio(struct btrfs_raid_bio *rbio)
2231 struct bio_list bio_list;
2236 * Allocate the pages for parity first, as P/Q pages will always be
2237 * needed for both full-stripe and sub-stripe writes.
2239 ret = alloc_rbio_parity_pages(rbio);
2244 * Either full stripe write, or we have every data sector already
2245 * cached, can go to write path immediately.
2247 if (!rbio_is_full(rbio) && need_read_stripe_sectors(rbio)) {
2249 * Now we're doing sub-stripe write, also need all data stripes
2250 * to do the full RMW.
2252 ret = alloc_rbio_data_pages(rbio);
2256 index_rbio_pages(rbio);
2258 ret = rmw_read_wait_recover(rbio);
2264 * At this stage we're not allowed to add any new bios to the
2265 * bio list any more, anyone else that wants to change this stripe
2266 * needs to do their own rmw.
2268 spin_lock_irq(&rbio->bio_list_lock);
2269 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
2270 spin_unlock_irq(&rbio->bio_list_lock);
2272 bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors);
2274 index_rbio_pages(rbio);
2277 * We don't cache full rbios because we're assuming
2278 * the higher layers are unlikely to use this area of
2279 * the disk again soon. If they do use it again,
2280 * hopefully they will send another full bio.
2282 if (!rbio_is_full(rbio))
2283 cache_rbio_pages(rbio);
2285 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2287 for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++)
2288 generate_pq_vertical(rbio, sectornr);
2290 bio_list_init(&bio_list);
2291 ret = rmw_assemble_write_bios(rbio, &bio_list);
2295 /* We should have at least one bio assembled. */
2296 ASSERT(bio_list_size(&bio_list));
2297 submit_write_bios(rbio, &bio_list);
2298 wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0);
2300 /* We may have more errors than our tolerance during the read. */
2301 for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++) {
2304 found_errors = get_rbio_veritical_errors(rbio, sectornr, NULL, NULL);
2305 if (found_errors > rbio->bioc->max_errors) {
2311 rbio_orig_end_io(rbio, errno_to_blk_status(ret));
2314 static void rmw_rbio_work(struct work_struct *work)
2316 struct btrfs_raid_bio *rbio;
2318 rbio = container_of(work, struct btrfs_raid_bio, work);
2319 if (lock_stripe_add(rbio) == 0)
2323 static void rmw_rbio_work_locked(struct work_struct *work)
2325 rmw_rbio(container_of(work, struct btrfs_raid_bio, work));
2329 * The following code is used to scrub/replace the parity stripe
2331 * Caller must have already increased bio_counter for getting @bioc.
2333 * Note: We need make sure all the pages that add into the scrub/replace
2334 * raid bio are correct and not be changed during the scrub/replace. That
2335 * is those pages just hold metadata or file data with checksum.
2338 struct btrfs_raid_bio *raid56_parity_alloc_scrub_rbio(struct bio *bio,
2339 struct btrfs_io_context *bioc,
2340 struct btrfs_device *scrub_dev,
2341 unsigned long *dbitmap, int stripe_nsectors)
2343 struct btrfs_fs_info *fs_info = bioc->fs_info;
2344 struct btrfs_raid_bio *rbio;
2347 rbio = alloc_rbio(fs_info, bioc);
2350 bio_list_add(&rbio->bio_list, bio);
2352 * This is a special bio which is used to hold the completion handler
2353 * and make the scrub rbio is similar to the other types
2355 ASSERT(!bio->bi_iter.bi_size);
2356 rbio->operation = BTRFS_RBIO_PARITY_SCRUB;
2359 * After mapping bioc with BTRFS_MAP_WRITE, parities have been sorted
2360 * to the end position, so this search can start from the first parity
2363 for (i = rbio->nr_data; i < rbio->real_stripes; i++) {
2364 if (bioc->stripes[i].dev == scrub_dev) {
2369 ASSERT(i < rbio->real_stripes);
2371 bitmap_copy(&rbio->dbitmap, dbitmap, stripe_nsectors);
2375 /* Used for both parity scrub and missing. */
2376 void raid56_add_scrub_pages(struct btrfs_raid_bio *rbio, struct page *page,
2377 unsigned int pgoff, u64 logical)
2379 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
2383 ASSERT(logical >= rbio->bioc->raid_map[0]);
2384 ASSERT(logical + sectorsize <= rbio->bioc->raid_map[0] +
2385 BTRFS_STRIPE_LEN * rbio->nr_data);
2386 stripe_offset = (int)(logical - rbio->bioc->raid_map[0]);
2387 index = stripe_offset / sectorsize;
2388 rbio->bio_sectors[index].page = page;
2389 rbio->bio_sectors[index].pgoff = pgoff;
2393 * We just scrub the parity that we have correct data on the same horizontal,
2394 * so we needn't allocate all pages for all the stripes.
2396 static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio)
2398 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
2399 int total_sector_nr;
2401 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
2402 total_sector_nr++) {
2404 int sectornr = total_sector_nr % rbio->stripe_nsectors;
2405 int index = (total_sector_nr * sectorsize) >> PAGE_SHIFT;
2407 if (!test_bit(sectornr, &rbio->dbitmap))
2409 if (rbio->stripe_pages[index])
2411 page = alloc_page(GFP_NOFS);
2414 rbio->stripe_pages[index] = page;
2416 index_stripe_sectors(rbio);
2420 static int finish_parity_scrub(struct btrfs_raid_bio *rbio, int need_check)
2422 struct btrfs_io_context *bioc = rbio->bioc;
2423 const u32 sectorsize = bioc->fs_info->sectorsize;
2424 void **pointers = rbio->finish_pointers;
2425 unsigned long *pbitmap = &rbio->finish_pbitmap;
2426 int nr_data = rbio->nr_data;
2430 struct sector_ptr p_sector = { 0 };
2431 struct sector_ptr q_sector = { 0 };
2432 struct bio_list bio_list;
2436 bio_list_init(&bio_list);
2438 if (rbio->real_stripes - rbio->nr_data == 1)
2439 has_qstripe = false;
2440 else if (rbio->real_stripes - rbio->nr_data == 2)
2445 if (bioc->num_tgtdevs && bioc->tgtdev_map[rbio->scrubp]) {
2447 bitmap_copy(pbitmap, &rbio->dbitmap, rbio->stripe_nsectors);
2451 * Because the higher layers(scrubber) are unlikely to
2452 * use this area of the disk again soon, so don't cache
2455 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2460 p_sector.page = alloc_page(GFP_NOFS);
2464 p_sector.uptodate = 1;
2467 /* RAID6, allocate and map temp space for the Q stripe */
2468 q_sector.page = alloc_page(GFP_NOFS);
2469 if (!q_sector.page) {
2470 __free_page(p_sector.page);
2471 p_sector.page = NULL;
2475 q_sector.uptodate = 1;
2476 pointers[rbio->real_stripes - 1] = kmap_local_page(q_sector.page);
2479 bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors);
2481 /* Map the parity stripe just once */
2482 pointers[nr_data] = kmap_local_page(p_sector.page);
2484 for_each_set_bit(sectornr, &rbio->dbitmap, rbio->stripe_nsectors) {
2485 struct sector_ptr *sector;
2488 /* first collect one page from each data stripe */
2489 for (stripe = 0; stripe < nr_data; stripe++) {
2490 sector = sector_in_rbio(rbio, stripe, sectornr, 0);
2491 pointers[stripe] = kmap_local_page(sector->page) +
2496 /* RAID6, call the library function to fill in our P/Q */
2497 raid6_call.gen_syndrome(rbio->real_stripes, sectorsize,
2501 memcpy(pointers[nr_data], pointers[0], sectorsize);
2502 run_xor(pointers + 1, nr_data - 1, sectorsize);
2505 /* Check scrubbing parity and repair it */
2506 sector = rbio_stripe_sector(rbio, rbio->scrubp, sectornr);
2507 parity = kmap_local_page(sector->page) + sector->pgoff;
2508 if (memcmp(parity, pointers[rbio->scrubp], sectorsize) != 0)
2509 memcpy(parity, pointers[rbio->scrubp], sectorsize);
2511 /* Parity is right, needn't writeback */
2512 bitmap_clear(&rbio->dbitmap, sectornr, 1);
2513 kunmap_local(parity);
2515 for (stripe = nr_data - 1; stripe >= 0; stripe--)
2516 kunmap_local(pointers[stripe]);
2519 kunmap_local(pointers[nr_data]);
2520 __free_page(p_sector.page);
2521 p_sector.page = NULL;
2522 if (q_sector.page) {
2523 kunmap_local(pointers[rbio->real_stripes - 1]);
2524 __free_page(q_sector.page);
2525 q_sector.page = NULL;
2530 * time to start writing. Make bios for everything from the
2531 * higher layers (the bio_list in our rbio) and our p/q. Ignore
2534 for_each_set_bit(sectornr, &rbio->dbitmap, rbio->stripe_nsectors) {
2535 struct sector_ptr *sector;
2537 sector = rbio_stripe_sector(rbio, rbio->scrubp, sectornr);
2538 ret = rbio_add_io_sector(rbio, &bio_list, sector, rbio->scrubp,
2539 sectornr, REQ_OP_WRITE);
2547 for_each_set_bit(sectornr, pbitmap, rbio->stripe_nsectors) {
2548 struct sector_ptr *sector;
2550 sector = rbio_stripe_sector(rbio, rbio->scrubp, sectornr);
2551 ret = rbio_add_io_sector(rbio, &bio_list, sector,
2552 bioc->tgtdev_map[rbio->scrubp],
2553 sectornr, REQ_OP_WRITE);
2559 submit_write_bios(rbio, &bio_list);
2563 bio_list_put(&bio_list);
2567 static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe)
2569 if (stripe >= 0 && stripe < rbio->nr_data)
2574 static int recover_scrub_rbio(struct btrfs_raid_bio *rbio)
2576 void **pointers = NULL;
2577 void **unmap_array = NULL;
2582 * @pointers array stores the pointer for each sector.
2584 * @unmap_array stores copy of pointers that does not get reordered
2585 * during reconstruction so that kunmap_local works.
2587 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
2588 unmap_array = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
2589 if (!pointers || !unmap_array) {
2594 for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) {
2595 int dfail = 0, failp = -1;
2600 found_errors = get_rbio_veritical_errors(rbio, sector_nr,
2602 if (found_errors > rbio->bioc->max_errors) {
2606 if (found_errors == 0)
2609 /* We should have at least one error here. */
2610 ASSERT(faila >= 0 || failb >= 0);
2612 if (is_data_stripe(rbio, faila))
2614 else if (is_parity_stripe(faila))
2617 if (is_data_stripe(rbio, failb))
2619 else if (is_parity_stripe(failb))
2622 * Because we can not use a scrubbing parity to repair the
2623 * data, so the capability of the repair is declined. (In the
2624 * case of RAID5, we can not repair anything.)
2626 if (dfail > rbio->bioc->max_errors - 1) {
2631 * If all data is good, only parity is correctly, just repair
2632 * the parity, no need to recover data stripes.
2638 * Here means we got one corrupted data stripe and one
2639 * corrupted parity on RAID6, if the corrupted parity is
2640 * scrubbing parity, luckily, use the other one to repair the
2641 * data, or we can not repair the data stripe.
2643 if (failp != rbio->scrubp) {
2648 ret = recover_vertical(rbio, sector_nr, pointers, unmap_array);
2658 static int scrub_assemble_read_bios(struct btrfs_raid_bio *rbio)
2660 struct bio_list bio_list = BIO_EMPTY_LIST;
2661 int total_sector_nr;
2664 /* Build a list of bios to read all the missing parts. */
2665 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
2666 total_sector_nr++) {
2667 int sectornr = total_sector_nr % rbio->stripe_nsectors;
2668 int stripe = total_sector_nr / rbio->stripe_nsectors;
2669 struct sector_ptr *sector;
2671 /* No data in the vertical stripe, no need to read. */
2672 if (!test_bit(sectornr, &rbio->dbitmap))
2676 * We want to find all the sectors missing from the rbio and
2677 * read them from the disk. If sector_in_rbio() finds a sector
2678 * in the bio list we don't need to read it off the stripe.
2680 sector = sector_in_rbio(rbio, stripe, sectornr, 1);
2684 sector = rbio_stripe_sector(rbio, stripe, sectornr);
2686 * The bio cache may have handed us an uptodate sector. If so,
2689 if (sector->uptodate)
2692 ret = rbio_add_io_sector(rbio, &bio_list, sector, stripe,
2693 sectornr, REQ_OP_READ);
2695 bio_list_put(&bio_list);
2700 submit_read_wait_bio_list(rbio, &bio_list);
2704 static void scrub_rbio(struct btrfs_raid_bio *rbio)
2706 bool need_check = false;
2710 ret = alloc_rbio_essential_pages(rbio);
2714 bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors);
2716 ret = scrub_assemble_read_bios(rbio);
2720 /* We may have some failures, recover the failed sectors first. */
2721 ret = recover_scrub_rbio(rbio);
2726 * We have every sector properly prepared. Can finish the scrub
2727 * and writeback the good content.
2729 ret = finish_parity_scrub(rbio, need_check);
2730 wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0);
2731 for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) {
2734 found_errors = get_rbio_veritical_errors(rbio, sector_nr, NULL, NULL);
2735 if (found_errors > rbio->bioc->max_errors) {
2741 rbio_orig_end_io(rbio, errno_to_blk_status(ret));
2744 static void scrub_rbio_work_locked(struct work_struct *work)
2746 scrub_rbio(container_of(work, struct btrfs_raid_bio, work));
2749 void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio)
2751 if (!lock_stripe_add(rbio))
2752 start_async_work(rbio, scrub_rbio_work_locked);
2755 /* The following code is used for dev replace of a missing RAID 5/6 device. */
2757 struct btrfs_raid_bio *
2758 raid56_alloc_missing_rbio(struct bio *bio, struct btrfs_io_context *bioc)
2760 struct btrfs_fs_info *fs_info = bioc->fs_info;
2761 struct btrfs_raid_bio *rbio;
2763 rbio = alloc_rbio(fs_info, bioc);
2767 rbio->operation = BTRFS_RBIO_REBUILD_MISSING;
2768 bio_list_add(&rbio->bio_list, bio);
2770 * This is a special bio which is used to hold the completion handler
2771 * and make the scrub rbio is similar to the other types
2773 ASSERT(!bio->bi_iter.bi_size);
2775 set_rbio_range_error(rbio, bio);
2780 void raid56_submit_missing_rbio(struct btrfs_raid_bio *rbio)
2782 start_async_work(rbio, recover_rbio_work);