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>
20 #include "async-thread.h"
22 /* set when additional merges to this rbio are not allowed */
23 #define RBIO_RMW_LOCKED_BIT 1
26 * set when this rbio is sitting in the hash, but it is just a cache
29 #define RBIO_CACHE_BIT 2
32 * set when it is safe to trust the stripe_pages for caching
34 #define RBIO_CACHE_READY_BIT 3
36 #define RBIO_CACHE_SIZE 1024
38 #define BTRFS_STRIPE_HASH_TABLE_BITS 11
40 /* Used by the raid56 code to lock stripes for read/modify/write */
41 struct btrfs_stripe_hash {
42 struct list_head hash_list;
46 /* Used by the raid56 code to lock stripes for read/modify/write */
47 struct btrfs_stripe_hash_table {
48 struct list_head stripe_cache;
49 spinlock_t cache_lock;
51 struct btrfs_stripe_hash table[];
56 BTRFS_RBIO_READ_REBUILD,
57 BTRFS_RBIO_PARITY_SCRUB,
58 BTRFS_RBIO_REBUILD_MISSING,
61 struct btrfs_raid_bio {
62 struct btrfs_fs_info *fs_info;
63 struct btrfs_bio *bbio;
65 /* while we're doing rmw on a stripe
66 * we put it into a hash table so we can
67 * lock the stripe and merge more rbios
70 struct list_head hash_list;
73 * LRU list for the stripe cache
75 struct list_head stripe_cache;
78 * for scheduling work in the helper threads
80 struct btrfs_work work;
83 * bio list and bio_list_lock are used
84 * to add more bios into the stripe
85 * in hopes of avoiding the full rmw
87 struct bio_list bio_list;
88 spinlock_t bio_list_lock;
90 /* also protected by the bio_list_lock, the
91 * plug list is used by the plugging code
92 * to collect partial bios while plugged. The
93 * stripe locking code also uses it to hand off
94 * the stripe lock to the next pending IO
96 struct list_head plug_list;
99 * flags that tell us if it is safe to
100 * merge with this bio
104 /* size of each individual stripe on disk */
107 /* number of data stripes (no p/q) */
114 * set if we're doing a parity rebuild
115 * for a read from higher up, which is handled
116 * differently from a parity rebuild as part of
119 enum btrfs_rbio_ops operation;
121 /* first bad stripe */
124 /* second bad stripe (for raid6 use) */
129 * number of pages needed to represent the full
135 * size of all the bios in the bio_list. This
136 * helps us decide if the rbio maps to a full
145 atomic_t stripes_pending;
149 * these are two arrays of pointers. We allocate the
150 * rbio big enough to hold them both and setup their
151 * locations when the rbio is allocated
154 /* pointers to pages that we allocated for
155 * reading/writing stripes directly from the disk (including P/Q)
157 struct page **stripe_pages;
160 * pointers to the pages in the bio_list. Stored
161 * here for faster lookup
163 struct page **bio_pages;
166 * bitmap to record which horizontal stripe has data
168 unsigned long *dbitmap;
170 /* allocated with real_stripes-many pointers for finish_*() calls */
171 void **finish_pointers;
173 /* allocated with stripe_npages-many bits for finish_*() calls */
174 unsigned long *finish_pbitmap;
177 static int __raid56_parity_recover(struct btrfs_raid_bio *rbio);
178 static noinline void finish_rmw(struct btrfs_raid_bio *rbio);
179 static void rmw_work(struct btrfs_work *work);
180 static void read_rebuild_work(struct btrfs_work *work);
181 static int fail_bio_stripe(struct btrfs_raid_bio *rbio, struct bio *bio);
182 static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed);
183 static void __free_raid_bio(struct btrfs_raid_bio *rbio);
184 static void index_rbio_pages(struct btrfs_raid_bio *rbio);
185 static int alloc_rbio_pages(struct btrfs_raid_bio *rbio);
187 static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
189 static void scrub_parity_work(struct btrfs_work *work);
191 static void start_async_work(struct btrfs_raid_bio *rbio, btrfs_func_t work_func)
193 btrfs_init_work(&rbio->work, work_func, NULL, NULL);
194 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
198 * the stripe hash table is used for locking, and to collect
199 * bios in hopes of making a full stripe
201 int btrfs_alloc_stripe_hash_table(struct btrfs_fs_info *info)
203 struct btrfs_stripe_hash_table *table;
204 struct btrfs_stripe_hash_table *x;
205 struct btrfs_stripe_hash *cur;
206 struct btrfs_stripe_hash *h;
207 int num_entries = 1 << BTRFS_STRIPE_HASH_TABLE_BITS;
211 if (info->stripe_hash_table)
215 * The table is large, starting with order 4 and can go as high as
216 * order 7 in case lock debugging is turned on.
218 * Try harder to allocate and fallback to vmalloc to lower the chance
219 * of a failing mount.
221 table_size = sizeof(*table) + sizeof(*h) * num_entries;
222 table = kvzalloc(table_size, GFP_KERNEL);
226 spin_lock_init(&table->cache_lock);
227 INIT_LIST_HEAD(&table->stripe_cache);
231 for (i = 0; i < num_entries; i++) {
233 INIT_LIST_HEAD(&cur->hash_list);
234 spin_lock_init(&cur->lock);
237 x = cmpxchg(&info->stripe_hash_table, NULL, table);
244 * caching an rbio means to copy anything from the
245 * bio_pages array into the stripe_pages array. We
246 * use the page uptodate bit in the stripe cache array
247 * to indicate if it has valid data
249 * once the caching is done, we set the cache ready
252 static void cache_rbio_pages(struct btrfs_raid_bio *rbio)
259 ret = alloc_rbio_pages(rbio);
263 for (i = 0; i < rbio->nr_pages; i++) {
264 if (!rbio->bio_pages[i])
267 s = kmap(rbio->bio_pages[i]);
268 d = kmap(rbio->stripe_pages[i]);
272 kunmap(rbio->bio_pages[i]);
273 kunmap(rbio->stripe_pages[i]);
274 SetPageUptodate(rbio->stripe_pages[i]);
276 set_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
280 * we hash on the first logical address of the stripe
282 static int rbio_bucket(struct btrfs_raid_bio *rbio)
284 u64 num = rbio->bbio->raid_map[0];
287 * we shift down quite a bit. We're using byte
288 * addressing, and most of the lower bits are zeros.
289 * This tends to upset hash_64, and it consistently
290 * returns just one or two different values.
292 * shifting off the lower bits fixes things.
294 return hash_64(num >> 16, BTRFS_STRIPE_HASH_TABLE_BITS);
298 * stealing an rbio means taking all the uptodate pages from the stripe
299 * array in the source rbio and putting them into the destination rbio
301 static void steal_rbio(struct btrfs_raid_bio *src, struct btrfs_raid_bio *dest)
307 if (!test_bit(RBIO_CACHE_READY_BIT, &src->flags))
310 for (i = 0; i < dest->nr_pages; i++) {
311 s = src->stripe_pages[i];
312 if (!s || !PageUptodate(s)) {
316 d = dest->stripe_pages[i];
320 dest->stripe_pages[i] = s;
321 src->stripe_pages[i] = NULL;
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_npages);
340 dest->generic_bio_cnt += victim->generic_bio_cnt;
341 bio_list_init(&victim->bio_list);
345 * used to prune items that are in the cache. The caller
346 * must hold the hash table lock.
348 static void __remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
350 int bucket = rbio_bucket(rbio);
351 struct btrfs_stripe_hash_table *table;
352 struct btrfs_stripe_hash *h;
356 * check the bit again under the hash table lock.
358 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
361 table = rbio->fs_info->stripe_hash_table;
362 h = table->table + bucket;
364 /* hold the lock for the bucket because we may be
365 * removing it from the hash table
370 * hold the lock for the bio list because we need
371 * to make sure the bio list is empty
373 spin_lock(&rbio->bio_list_lock);
375 if (test_and_clear_bit(RBIO_CACHE_BIT, &rbio->flags)) {
376 list_del_init(&rbio->stripe_cache);
377 table->cache_size -= 1;
380 /* if the bio list isn't empty, this rbio is
381 * still involved in an IO. We take it out
382 * of the cache list, and drop the ref that
383 * was held for the list.
385 * If the bio_list was empty, we also remove
386 * the rbio from the hash_table, and drop
387 * the corresponding ref
389 if (bio_list_empty(&rbio->bio_list)) {
390 if (!list_empty(&rbio->hash_list)) {
391 list_del_init(&rbio->hash_list);
392 refcount_dec(&rbio->refs);
393 BUG_ON(!list_empty(&rbio->plug_list));
398 spin_unlock(&rbio->bio_list_lock);
399 spin_unlock(&h->lock);
402 __free_raid_bio(rbio);
406 * prune a given rbio from the cache
408 static void remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
410 struct btrfs_stripe_hash_table *table;
413 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
416 table = rbio->fs_info->stripe_hash_table;
418 spin_lock_irqsave(&table->cache_lock, flags);
419 __remove_rbio_from_cache(rbio);
420 spin_unlock_irqrestore(&table->cache_lock, flags);
424 * remove everything in the cache
426 static void btrfs_clear_rbio_cache(struct btrfs_fs_info *info)
428 struct btrfs_stripe_hash_table *table;
430 struct btrfs_raid_bio *rbio;
432 table = info->stripe_hash_table;
434 spin_lock_irqsave(&table->cache_lock, flags);
435 while (!list_empty(&table->stripe_cache)) {
436 rbio = list_entry(table->stripe_cache.next,
437 struct btrfs_raid_bio,
439 __remove_rbio_from_cache(rbio);
441 spin_unlock_irqrestore(&table->cache_lock, flags);
445 * remove all cached entries and free the hash table
448 void btrfs_free_stripe_hash_table(struct btrfs_fs_info *info)
450 if (!info->stripe_hash_table)
452 btrfs_clear_rbio_cache(info);
453 kvfree(info->stripe_hash_table);
454 info->stripe_hash_table = NULL;
458 * insert an rbio into the stripe cache. It
459 * must have already been prepared by calling
462 * If this rbio was already cached, it gets
463 * moved to the front of the lru.
465 * If the size of the rbio cache is too big, we
468 static void cache_rbio(struct btrfs_raid_bio *rbio)
470 struct btrfs_stripe_hash_table *table;
473 if (!test_bit(RBIO_CACHE_READY_BIT, &rbio->flags))
476 table = rbio->fs_info->stripe_hash_table;
478 spin_lock_irqsave(&table->cache_lock, flags);
479 spin_lock(&rbio->bio_list_lock);
481 /* bump our ref if we were not in the list before */
482 if (!test_and_set_bit(RBIO_CACHE_BIT, &rbio->flags))
483 refcount_inc(&rbio->refs);
485 if (!list_empty(&rbio->stripe_cache)){
486 list_move(&rbio->stripe_cache, &table->stripe_cache);
488 list_add(&rbio->stripe_cache, &table->stripe_cache);
489 table->cache_size += 1;
492 spin_unlock(&rbio->bio_list_lock);
494 if (table->cache_size > RBIO_CACHE_SIZE) {
495 struct btrfs_raid_bio *found;
497 found = list_entry(table->stripe_cache.prev,
498 struct btrfs_raid_bio,
502 __remove_rbio_from_cache(found);
505 spin_unlock_irqrestore(&table->cache_lock, flags);
509 * helper function to run the xor_blocks api. It is only
510 * able to do MAX_XOR_BLOCKS at a time, so we need to
513 static void run_xor(void **pages, int src_cnt, ssize_t len)
517 void *dest = pages[src_cnt];
520 xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS);
521 xor_blocks(xor_src_cnt, len, dest, pages + src_off);
523 src_cnt -= xor_src_cnt;
524 src_off += xor_src_cnt;
529 * Returns true if the bio list inside this rbio covers an entire stripe (no
532 static int rbio_is_full(struct btrfs_raid_bio *rbio)
535 unsigned long size = rbio->bio_list_bytes;
538 spin_lock_irqsave(&rbio->bio_list_lock, flags);
539 if (size != rbio->nr_data * rbio->stripe_len)
541 BUG_ON(size > rbio->nr_data * rbio->stripe_len);
542 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
548 * returns 1 if it is safe to merge two rbios together.
549 * The merging is safe if the two rbios correspond to
550 * the same stripe and if they are both going in the same
551 * direction (read vs write), and if neither one is
552 * locked for final IO
554 * The caller is responsible for locking such that
555 * rmw_locked is safe to test
557 static int rbio_can_merge(struct btrfs_raid_bio *last,
558 struct btrfs_raid_bio *cur)
560 if (test_bit(RBIO_RMW_LOCKED_BIT, &last->flags) ||
561 test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags))
565 * we can't merge with cached rbios, since the
566 * idea is that when we merge the destination
567 * rbio is going to run our IO for us. We can
568 * steal from cached rbios though, other functions
571 if (test_bit(RBIO_CACHE_BIT, &last->flags) ||
572 test_bit(RBIO_CACHE_BIT, &cur->flags))
575 if (last->bbio->raid_map[0] !=
576 cur->bbio->raid_map[0])
579 /* we can't merge with different operations */
580 if (last->operation != cur->operation)
583 * We've need read the full stripe from the drive.
584 * check and repair the parity and write the new results.
586 * We're not allowed to add any new bios to the
587 * bio list here, anyone else that wants to
588 * change this stripe needs to do their own rmw.
590 if (last->operation == BTRFS_RBIO_PARITY_SCRUB)
593 if (last->operation == BTRFS_RBIO_REBUILD_MISSING)
596 if (last->operation == BTRFS_RBIO_READ_REBUILD) {
597 int fa = last->faila;
598 int fb = last->failb;
599 int cur_fa = cur->faila;
600 int cur_fb = cur->failb;
602 if (last->faila >= last->failb) {
607 if (cur->faila >= cur->failb) {
612 if (fa != cur_fa || fb != cur_fb)
618 static int rbio_stripe_page_index(struct btrfs_raid_bio *rbio, int stripe,
621 return stripe * rbio->stripe_npages + index;
625 * these are just the pages from the rbio array, not from anything
626 * the FS sent down to us
628 static struct page *rbio_stripe_page(struct btrfs_raid_bio *rbio, int stripe,
631 return rbio->stripe_pages[rbio_stripe_page_index(rbio, stripe, index)];
635 * helper to index into the pstripe
637 static struct page *rbio_pstripe_page(struct btrfs_raid_bio *rbio, int index)
639 return rbio_stripe_page(rbio, rbio->nr_data, index);
643 * helper to index into the qstripe, returns null
644 * if there is no qstripe
646 static struct page *rbio_qstripe_page(struct btrfs_raid_bio *rbio, int index)
648 if (rbio->nr_data + 1 == rbio->real_stripes)
650 return rbio_stripe_page(rbio, rbio->nr_data + 1, index);
654 * The first stripe in the table for a logical address
655 * has the lock. rbios are added in one of three ways:
657 * 1) Nobody has the stripe locked yet. The rbio is given
658 * the lock and 0 is returned. The caller must start the IO
661 * 2) Someone has the stripe locked, but we're able to merge
662 * with the lock owner. The rbio is freed and the IO will
663 * start automatically along with the existing rbio. 1 is returned.
665 * 3) Someone has the stripe locked, but we're not able to merge.
666 * The rbio is added to the lock owner's plug list, or merged into
667 * an rbio already on the plug list. When the lock owner unlocks,
668 * the next rbio on the list is run and the IO is started automatically.
671 * If we return 0, the caller still owns the rbio and must continue with
672 * IO submission. If we return 1, the caller must assume the rbio has
673 * already been freed.
675 static noinline int lock_stripe_add(struct btrfs_raid_bio *rbio)
677 int bucket = rbio_bucket(rbio);
678 struct btrfs_stripe_hash *h = rbio->fs_info->stripe_hash_table->table + bucket;
679 struct btrfs_raid_bio *cur;
680 struct btrfs_raid_bio *pending;
682 struct btrfs_raid_bio *freeit = NULL;
683 struct btrfs_raid_bio *cache_drop = NULL;
686 spin_lock_irqsave(&h->lock, flags);
687 list_for_each_entry(cur, &h->hash_list, hash_list) {
688 if (cur->bbio->raid_map[0] == rbio->bbio->raid_map[0]) {
689 spin_lock(&cur->bio_list_lock);
691 /* can we steal this cached rbio's pages? */
692 if (bio_list_empty(&cur->bio_list) &&
693 list_empty(&cur->plug_list) &&
694 test_bit(RBIO_CACHE_BIT, &cur->flags) &&
695 !test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) {
696 list_del_init(&cur->hash_list);
697 refcount_dec(&cur->refs);
699 steal_rbio(cur, rbio);
701 spin_unlock(&cur->bio_list_lock);
706 /* can we merge into the lock owner? */
707 if (rbio_can_merge(cur, rbio)) {
708 merge_rbio(cur, rbio);
709 spin_unlock(&cur->bio_list_lock);
717 * we couldn't merge with the running
718 * rbio, see if we can merge with the
719 * pending ones. We don't have to
720 * check for rmw_locked because there
721 * is no way they are inside finish_rmw
724 list_for_each_entry(pending, &cur->plug_list,
726 if (rbio_can_merge(pending, rbio)) {
727 merge_rbio(pending, rbio);
728 spin_unlock(&cur->bio_list_lock);
735 /* no merging, put us on the tail of the plug list,
736 * our rbio will be started with the currently
737 * running rbio unlocks
739 list_add_tail(&rbio->plug_list, &cur->plug_list);
740 spin_unlock(&cur->bio_list_lock);
746 refcount_inc(&rbio->refs);
747 list_add(&rbio->hash_list, &h->hash_list);
749 spin_unlock_irqrestore(&h->lock, flags);
751 remove_rbio_from_cache(cache_drop);
753 __free_raid_bio(freeit);
758 * called as rmw or parity rebuild is completed. If the plug list has more
759 * rbios waiting for this stripe, the next one on the list will be started
761 static noinline void unlock_stripe(struct btrfs_raid_bio *rbio)
764 struct btrfs_stripe_hash *h;
768 bucket = rbio_bucket(rbio);
769 h = rbio->fs_info->stripe_hash_table->table + bucket;
771 if (list_empty(&rbio->plug_list))
774 spin_lock_irqsave(&h->lock, flags);
775 spin_lock(&rbio->bio_list_lock);
777 if (!list_empty(&rbio->hash_list)) {
779 * if we're still cached and there is no other IO
780 * to perform, just leave this rbio here for others
781 * to steal from later
783 if (list_empty(&rbio->plug_list) &&
784 test_bit(RBIO_CACHE_BIT, &rbio->flags)) {
786 clear_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
787 BUG_ON(!bio_list_empty(&rbio->bio_list));
791 list_del_init(&rbio->hash_list);
792 refcount_dec(&rbio->refs);
795 * we use the plug list to hold all the rbios
796 * waiting for the chance to lock this stripe.
797 * hand the lock over to one of them.
799 if (!list_empty(&rbio->plug_list)) {
800 struct btrfs_raid_bio *next;
801 struct list_head *head = rbio->plug_list.next;
803 next = list_entry(head, struct btrfs_raid_bio,
806 list_del_init(&rbio->plug_list);
808 list_add(&next->hash_list, &h->hash_list);
809 refcount_inc(&next->refs);
810 spin_unlock(&rbio->bio_list_lock);
811 spin_unlock_irqrestore(&h->lock, flags);
813 if (next->operation == BTRFS_RBIO_READ_REBUILD)
814 start_async_work(next, read_rebuild_work);
815 else if (next->operation == BTRFS_RBIO_REBUILD_MISSING) {
816 steal_rbio(rbio, next);
817 start_async_work(next, read_rebuild_work);
818 } else if (next->operation == BTRFS_RBIO_WRITE) {
819 steal_rbio(rbio, next);
820 start_async_work(next, rmw_work);
821 } else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) {
822 steal_rbio(rbio, next);
823 start_async_work(next, scrub_parity_work);
830 spin_unlock(&rbio->bio_list_lock);
831 spin_unlock_irqrestore(&h->lock, flags);
835 remove_rbio_from_cache(rbio);
838 static void __free_raid_bio(struct btrfs_raid_bio *rbio)
842 if (!refcount_dec_and_test(&rbio->refs))
845 WARN_ON(!list_empty(&rbio->stripe_cache));
846 WARN_ON(!list_empty(&rbio->hash_list));
847 WARN_ON(!bio_list_empty(&rbio->bio_list));
849 for (i = 0; i < rbio->nr_pages; i++) {
850 if (rbio->stripe_pages[i]) {
851 __free_page(rbio->stripe_pages[i]);
852 rbio->stripe_pages[i] = NULL;
856 btrfs_put_bbio(rbio->bbio);
860 static void rbio_endio_bio_list(struct bio *cur, blk_status_t err)
867 cur->bi_status = err;
874 * this frees the rbio and runs through all the bios in the
875 * bio_list and calls end_io on them
877 static void rbio_orig_end_io(struct btrfs_raid_bio *rbio, blk_status_t err)
879 struct bio *cur = bio_list_get(&rbio->bio_list);
882 if (rbio->generic_bio_cnt)
883 btrfs_bio_counter_sub(rbio->fs_info, rbio->generic_bio_cnt);
885 * Clear the data bitmap, as the rbio may be cached for later usage.
886 * do this before before unlock_stripe() so there will be no new bio
889 bitmap_clear(rbio->dbitmap, 0, rbio->stripe_npages);
892 * At this moment, rbio->bio_list is empty, however since rbio does not
893 * always have RBIO_RMW_LOCKED_BIT set and rbio is still linked on the
894 * hash list, rbio may be merged with others so that rbio->bio_list
896 * Once unlock_stripe() is done, rbio->bio_list will not be updated any
897 * more and we can call bio_endio() on all queued bios.
900 extra = bio_list_get(&rbio->bio_list);
901 __free_raid_bio(rbio);
903 rbio_endio_bio_list(cur, err);
905 rbio_endio_bio_list(extra, err);
909 * end io function used by finish_rmw. When we finally
910 * get here, we've written a full stripe
912 static void raid_write_end_io(struct bio *bio)
914 struct btrfs_raid_bio *rbio = bio->bi_private;
915 blk_status_t err = bio->bi_status;
919 fail_bio_stripe(rbio, bio);
923 if (!atomic_dec_and_test(&rbio->stripes_pending))
928 /* OK, we have read all the stripes we need to. */
929 max_errors = (rbio->operation == BTRFS_RBIO_PARITY_SCRUB) ?
930 0 : rbio->bbio->max_errors;
931 if (atomic_read(&rbio->error) > max_errors)
934 rbio_orig_end_io(rbio, err);
938 * the read/modify/write code wants to use the original bio for
939 * any pages it included, and then use the rbio for everything
940 * else. This function decides if a given index (stripe number)
941 * and page number in that stripe fall inside the original bio
944 * if you set bio_list_only, you'll get a NULL back for any ranges
945 * that are outside the bio_list
947 * This doesn't take any refs on anything, you get a bare page pointer
948 * and the caller must bump refs as required.
950 * You must call index_rbio_pages once before you can trust
951 * the answers from this function.
953 static struct page *page_in_rbio(struct btrfs_raid_bio *rbio,
954 int index, int pagenr, int bio_list_only)
957 struct page *p = NULL;
959 chunk_page = index * (rbio->stripe_len >> PAGE_SHIFT) + pagenr;
961 spin_lock_irq(&rbio->bio_list_lock);
962 p = rbio->bio_pages[chunk_page];
963 spin_unlock_irq(&rbio->bio_list_lock);
965 if (p || bio_list_only)
968 return rbio->stripe_pages[chunk_page];
972 * number of pages we need for the entire stripe across all the
975 static unsigned long rbio_nr_pages(unsigned long stripe_len, int nr_stripes)
977 return DIV_ROUND_UP(stripe_len, PAGE_SIZE) * nr_stripes;
981 * allocation and initial setup for the btrfs_raid_bio. Not
982 * this does not allocate any pages for rbio->pages.
984 static struct btrfs_raid_bio *alloc_rbio(struct btrfs_fs_info *fs_info,
985 struct btrfs_bio *bbio,
988 struct btrfs_raid_bio *rbio;
990 int real_stripes = bbio->num_stripes - bbio->num_tgtdevs;
991 int num_pages = rbio_nr_pages(stripe_len, real_stripes);
992 int stripe_npages = DIV_ROUND_UP(stripe_len, PAGE_SIZE);
995 rbio = kzalloc(sizeof(*rbio) +
996 sizeof(*rbio->stripe_pages) * num_pages +
997 sizeof(*rbio->bio_pages) * num_pages +
998 sizeof(*rbio->finish_pointers) * real_stripes +
999 sizeof(*rbio->dbitmap) * BITS_TO_LONGS(stripe_npages) +
1000 sizeof(*rbio->finish_pbitmap) *
1001 BITS_TO_LONGS(stripe_npages),
1004 return ERR_PTR(-ENOMEM);
1006 bio_list_init(&rbio->bio_list);
1007 INIT_LIST_HEAD(&rbio->plug_list);
1008 spin_lock_init(&rbio->bio_list_lock);
1009 INIT_LIST_HEAD(&rbio->stripe_cache);
1010 INIT_LIST_HEAD(&rbio->hash_list);
1012 rbio->fs_info = fs_info;
1013 rbio->stripe_len = stripe_len;
1014 rbio->nr_pages = num_pages;
1015 rbio->real_stripes = real_stripes;
1016 rbio->stripe_npages = stripe_npages;
1019 refcount_set(&rbio->refs, 1);
1020 atomic_set(&rbio->error, 0);
1021 atomic_set(&rbio->stripes_pending, 0);
1024 * the stripe_pages, bio_pages, etc arrays point to the extra
1025 * memory we allocated past the end of the rbio
1028 #define CONSUME_ALLOC(ptr, count) do { \
1030 p = (unsigned char *)p + sizeof(*(ptr)) * (count); \
1032 CONSUME_ALLOC(rbio->stripe_pages, num_pages);
1033 CONSUME_ALLOC(rbio->bio_pages, num_pages);
1034 CONSUME_ALLOC(rbio->finish_pointers, real_stripes);
1035 CONSUME_ALLOC(rbio->dbitmap, BITS_TO_LONGS(stripe_npages));
1036 CONSUME_ALLOC(rbio->finish_pbitmap, BITS_TO_LONGS(stripe_npages));
1037 #undef CONSUME_ALLOC
1039 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1040 nr_data = real_stripes - 1;
1041 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1042 nr_data = real_stripes - 2;
1046 rbio->nr_data = nr_data;
1050 /* allocate pages for all the stripes in the bio, including parity */
1051 static int alloc_rbio_pages(struct btrfs_raid_bio *rbio)
1056 for (i = 0; i < rbio->nr_pages; i++) {
1057 if (rbio->stripe_pages[i])
1059 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1062 rbio->stripe_pages[i] = page;
1067 /* only allocate pages for p/q stripes */
1068 static int alloc_rbio_parity_pages(struct btrfs_raid_bio *rbio)
1073 i = rbio_stripe_page_index(rbio, rbio->nr_data, 0);
1075 for (; i < rbio->nr_pages; i++) {
1076 if (rbio->stripe_pages[i])
1078 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1081 rbio->stripe_pages[i] = page;
1087 * add a single page from a specific stripe into our list of bios for IO
1088 * this will try to merge into existing bios if possible, and returns
1089 * zero if all went well.
1091 static int rbio_add_io_page(struct btrfs_raid_bio *rbio,
1092 struct bio_list *bio_list,
1095 unsigned long page_index,
1096 unsigned long bio_max_len)
1098 struct bio *last = bio_list->tail;
1102 struct btrfs_bio_stripe *stripe;
1105 stripe = &rbio->bbio->stripes[stripe_nr];
1106 disk_start = stripe->physical + (page_index << PAGE_SHIFT);
1108 /* if the device is missing, just fail this stripe */
1109 if (!stripe->dev->bdev)
1110 return fail_rbio_index(rbio, stripe_nr);
1112 /* see if we can add this page onto our existing bio */
1114 last_end = (u64)last->bi_iter.bi_sector << 9;
1115 last_end += last->bi_iter.bi_size;
1118 * we can't merge these if they are from different
1119 * devices or if they are not contiguous
1121 if (last_end == disk_start && stripe->dev->bdev &&
1123 last->bi_disk == stripe->dev->bdev->bd_disk &&
1124 last->bi_partno == stripe->dev->bdev->bd_partno) {
1125 ret = bio_add_page(last, page, PAGE_SIZE, 0);
1126 if (ret == PAGE_SIZE)
1131 /* put a new bio on the list */
1132 bio = btrfs_io_bio_alloc(bio_max_len >> PAGE_SHIFT ?: 1);
1133 bio->bi_iter.bi_size = 0;
1134 bio_set_dev(bio, stripe->dev->bdev);
1135 bio->bi_iter.bi_sector = disk_start >> 9;
1137 bio_add_page(bio, page, PAGE_SIZE, 0);
1138 bio_list_add(bio_list, bio);
1143 * while we're doing the read/modify/write cycle, we could
1144 * have errors in reading pages off the disk. This checks
1145 * for errors and if we're not able to read the page it'll
1146 * trigger parity reconstruction. The rmw will be finished
1147 * after we've reconstructed the failed stripes
1149 static void validate_rbio_for_rmw(struct btrfs_raid_bio *rbio)
1151 if (rbio->faila >= 0 || rbio->failb >= 0) {
1152 BUG_ON(rbio->faila == rbio->real_stripes - 1);
1153 __raid56_parity_recover(rbio);
1160 * helper function to walk our bio list and populate the bio_pages array with
1161 * the result. This seems expensive, but it is faster than constantly
1162 * searching through the bio list as we setup the IO in finish_rmw or stripe
1165 * This must be called before you trust the answers from page_in_rbio
1167 static void index_rbio_pages(struct btrfs_raid_bio *rbio)
1171 unsigned long stripe_offset;
1172 unsigned long page_index;
1174 spin_lock_irq(&rbio->bio_list_lock);
1175 bio_list_for_each(bio, &rbio->bio_list) {
1176 struct bio_vec bvec;
1177 struct bvec_iter iter;
1180 start = (u64)bio->bi_iter.bi_sector << 9;
1181 stripe_offset = start - rbio->bbio->raid_map[0];
1182 page_index = stripe_offset >> PAGE_SHIFT;
1184 if (bio_flagged(bio, BIO_CLONED))
1185 bio->bi_iter = btrfs_io_bio(bio)->iter;
1187 bio_for_each_segment(bvec, bio, iter) {
1188 rbio->bio_pages[page_index + i] = bvec.bv_page;
1192 spin_unlock_irq(&rbio->bio_list_lock);
1196 * this is called from one of two situations. We either
1197 * have a full stripe from the higher layers, or we've read all
1198 * the missing bits off disk.
1200 * This will calculate the parity and then send down any
1203 static noinline void finish_rmw(struct btrfs_raid_bio *rbio)
1205 struct btrfs_bio *bbio = rbio->bbio;
1206 void **pointers = rbio->finish_pointers;
1207 int nr_data = rbio->nr_data;
1211 struct bio_list bio_list;
1215 bio_list_init(&bio_list);
1217 if (rbio->real_stripes - rbio->nr_data == 1)
1218 has_qstripe = false;
1219 else if (rbio->real_stripes - rbio->nr_data == 2)
1224 /* We should have at least one data sector. */
1225 ASSERT(bitmap_weight(rbio->dbitmap, rbio->stripe_npages));
1227 /* at this point we either have a full stripe,
1228 * or we've read the full stripe from the drive.
1229 * recalculate the parity and write the new results.
1231 * We're not allowed to add any new bios to the
1232 * bio list here, anyone else that wants to
1233 * change this stripe needs to do their own rmw.
1235 spin_lock_irq(&rbio->bio_list_lock);
1236 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1237 spin_unlock_irq(&rbio->bio_list_lock);
1239 atomic_set(&rbio->error, 0);
1242 * now that we've set rmw_locked, run through the
1243 * bio list one last time and map the page pointers
1245 * We don't cache full rbios because we're assuming
1246 * the higher layers are unlikely to use this area of
1247 * the disk again soon. If they do use it again,
1248 * hopefully they will send another full bio.
1250 index_rbio_pages(rbio);
1251 if (!rbio_is_full(rbio))
1252 cache_rbio_pages(rbio);
1254 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
1256 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1258 /* first collect one page from each data stripe */
1259 for (stripe = 0; stripe < nr_data; stripe++) {
1260 p = page_in_rbio(rbio, stripe, pagenr, 0);
1261 pointers[stripe] = kmap(p);
1264 /* then add the parity stripe */
1265 p = rbio_pstripe_page(rbio, pagenr);
1267 pointers[stripe++] = kmap(p);
1272 * raid6, add the qstripe and call the
1273 * library function to fill in our p/q
1275 p = rbio_qstripe_page(rbio, pagenr);
1277 pointers[stripe++] = kmap(p);
1279 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
1283 copy_page(pointers[nr_data], pointers[0]);
1284 run_xor(pointers + 1, nr_data - 1, PAGE_SIZE);
1288 for (stripe = 0; stripe < rbio->real_stripes; stripe++)
1289 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
1293 * time to start writing. Make bios for everything from the
1294 * higher layers (the bio_list in our rbio) and our p/q. Ignore
1297 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1298 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1301 /* This vertical stripe has no data, skip it. */
1302 if (!test_bit(pagenr, rbio->dbitmap))
1305 if (stripe < rbio->nr_data) {
1306 page = page_in_rbio(rbio, stripe, pagenr, 1);
1310 page = rbio_stripe_page(rbio, stripe, pagenr);
1313 ret = rbio_add_io_page(rbio, &bio_list,
1314 page, stripe, pagenr, rbio->stripe_len);
1320 if (likely(!bbio->num_tgtdevs))
1323 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1324 if (!bbio->tgtdev_map[stripe])
1327 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1330 /* This vertical stripe has no data, skip it. */
1331 if (!test_bit(pagenr, rbio->dbitmap))
1334 if (stripe < rbio->nr_data) {
1335 page = page_in_rbio(rbio, stripe, pagenr, 1);
1339 page = rbio_stripe_page(rbio, stripe, pagenr);
1342 ret = rbio_add_io_page(rbio, &bio_list, page,
1343 rbio->bbio->tgtdev_map[stripe],
1344 pagenr, rbio->stripe_len);
1351 atomic_set(&rbio->stripes_pending, bio_list_size(&bio_list));
1352 BUG_ON(atomic_read(&rbio->stripes_pending) == 0);
1355 bio = bio_list_pop(&bio_list);
1359 bio->bi_private = rbio;
1360 bio->bi_end_io = raid_write_end_io;
1361 bio->bi_opf = REQ_OP_WRITE;
1368 rbio_orig_end_io(rbio, BLK_STS_IOERR);
1370 while ((bio = bio_list_pop(&bio_list)))
1375 * helper to find the stripe number for a given bio. Used to figure out which
1376 * stripe has failed. This expects the bio to correspond to a physical disk,
1377 * so it looks up based on physical sector numbers.
1379 static int find_bio_stripe(struct btrfs_raid_bio *rbio,
1382 u64 physical = bio->bi_iter.bi_sector;
1385 struct btrfs_bio_stripe *stripe;
1389 for (i = 0; i < rbio->bbio->num_stripes; i++) {
1390 stripe = &rbio->bbio->stripes[i];
1391 stripe_start = stripe->physical;
1392 if (physical >= stripe_start &&
1393 physical < stripe_start + rbio->stripe_len &&
1394 stripe->dev->bdev &&
1395 bio->bi_disk == stripe->dev->bdev->bd_disk &&
1396 bio->bi_partno == stripe->dev->bdev->bd_partno) {
1404 * helper to find the stripe number for a given
1405 * bio (before mapping). Used to figure out which stripe has
1406 * failed. This looks up based on logical block numbers.
1408 static int find_logical_bio_stripe(struct btrfs_raid_bio *rbio,
1411 u64 logical = bio->bi_iter.bi_sector;
1417 for (i = 0; i < rbio->nr_data; i++) {
1418 stripe_start = rbio->bbio->raid_map[i];
1419 if (logical >= stripe_start &&
1420 logical < stripe_start + rbio->stripe_len) {
1428 * returns -EIO if we had too many failures
1430 static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed)
1432 unsigned long flags;
1435 spin_lock_irqsave(&rbio->bio_list_lock, flags);
1437 /* we already know this stripe is bad, move on */
1438 if (rbio->faila == failed || rbio->failb == failed)
1441 if (rbio->faila == -1) {
1442 /* first failure on this rbio */
1443 rbio->faila = failed;
1444 atomic_inc(&rbio->error);
1445 } else if (rbio->failb == -1) {
1446 /* second failure on this rbio */
1447 rbio->failb = failed;
1448 atomic_inc(&rbio->error);
1453 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
1459 * helper to fail a stripe based on a physical disk
1462 static int fail_bio_stripe(struct btrfs_raid_bio *rbio,
1465 int failed = find_bio_stripe(rbio, bio);
1470 return fail_rbio_index(rbio, failed);
1474 * this sets each page in the bio uptodate. It should only be used on private
1475 * rbio pages, nothing that comes in from the higher layers
1477 static void set_bio_pages_uptodate(struct bio *bio)
1479 struct bio_vec *bvec;
1480 struct bvec_iter_all iter_all;
1482 ASSERT(!bio_flagged(bio, BIO_CLONED));
1484 bio_for_each_segment_all(bvec, bio, iter_all)
1485 SetPageUptodate(bvec->bv_page);
1489 * end io for the read phase of the rmw cycle. All the bios here are physical
1490 * stripe bios we've read from the disk so we can recalculate the parity of the
1493 * This will usually kick off finish_rmw once all the bios are read in, but it
1494 * may trigger parity reconstruction if we had any errors along the way
1496 static void raid_rmw_end_io(struct bio *bio)
1498 struct btrfs_raid_bio *rbio = bio->bi_private;
1501 fail_bio_stripe(rbio, bio);
1503 set_bio_pages_uptodate(bio);
1507 if (!atomic_dec_and_test(&rbio->stripes_pending))
1510 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
1514 * this will normally call finish_rmw to start our write
1515 * but if there are any failed stripes we'll reconstruct
1518 validate_rbio_for_rmw(rbio);
1523 rbio_orig_end_io(rbio, BLK_STS_IOERR);
1527 * the stripe must be locked by the caller. It will
1528 * unlock after all the writes are done
1530 static int raid56_rmw_stripe(struct btrfs_raid_bio *rbio)
1532 int bios_to_read = 0;
1533 struct bio_list bio_list;
1539 bio_list_init(&bio_list);
1541 ret = alloc_rbio_pages(rbio);
1545 index_rbio_pages(rbio);
1547 atomic_set(&rbio->error, 0);
1549 * build a list of bios to read all the missing parts of this
1552 for (stripe = 0; stripe < rbio->nr_data; stripe++) {
1553 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1556 * we want to find all the pages missing from
1557 * the rbio and read them from the disk. If
1558 * page_in_rbio finds a page in the bio list
1559 * we don't need to read it off the stripe.
1561 page = page_in_rbio(rbio, stripe, pagenr, 1);
1565 page = rbio_stripe_page(rbio, stripe, pagenr);
1567 * the bio cache may have handed us an uptodate
1568 * page. If so, be happy and use it
1570 if (PageUptodate(page))
1573 ret = rbio_add_io_page(rbio, &bio_list, page,
1574 stripe, pagenr, rbio->stripe_len);
1580 bios_to_read = bio_list_size(&bio_list);
1581 if (!bios_to_read) {
1583 * this can happen if others have merged with
1584 * us, it means there is nothing left to read.
1585 * But if there are missing devices it may not be
1586 * safe to do the full stripe write yet.
1592 * the bbio may be freed once we submit the last bio. Make sure
1593 * not to touch it after that
1595 atomic_set(&rbio->stripes_pending, bios_to_read);
1597 bio = bio_list_pop(&bio_list);
1601 bio->bi_private = rbio;
1602 bio->bi_end_io = raid_rmw_end_io;
1603 bio->bi_opf = REQ_OP_READ;
1605 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
1609 /* the actual write will happen once the reads are done */
1613 rbio_orig_end_io(rbio, BLK_STS_IOERR);
1615 while ((bio = bio_list_pop(&bio_list)))
1621 validate_rbio_for_rmw(rbio);
1626 * if the upper layers pass in a full stripe, we thank them by only allocating
1627 * enough pages to hold the parity, and sending it all down quickly.
1629 static int full_stripe_write(struct btrfs_raid_bio *rbio)
1633 ret = alloc_rbio_parity_pages(rbio);
1635 __free_raid_bio(rbio);
1639 ret = lock_stripe_add(rbio);
1646 * partial stripe writes get handed over to async helpers.
1647 * We're really hoping to merge a few more writes into this
1648 * rbio before calculating new parity
1650 static int partial_stripe_write(struct btrfs_raid_bio *rbio)
1654 ret = lock_stripe_add(rbio);
1656 start_async_work(rbio, rmw_work);
1661 * sometimes while we were reading from the drive to
1662 * recalculate parity, enough new bios come into create
1663 * a full stripe. So we do a check here to see if we can
1664 * go directly to finish_rmw
1666 static int __raid56_parity_write(struct btrfs_raid_bio *rbio)
1668 /* head off into rmw land if we don't have a full stripe */
1669 if (!rbio_is_full(rbio))
1670 return partial_stripe_write(rbio);
1671 return full_stripe_write(rbio);
1675 * We use plugging call backs to collect full stripes.
1676 * Any time we get a partial stripe write while plugged
1677 * we collect it into a list. When the unplug comes down,
1678 * we sort the list by logical block number and merge
1679 * everything we can into the same rbios
1681 struct btrfs_plug_cb {
1682 struct blk_plug_cb cb;
1683 struct btrfs_fs_info *info;
1684 struct list_head rbio_list;
1685 struct btrfs_work work;
1689 * rbios on the plug list are sorted for easier merging.
1691 static int plug_cmp(void *priv, struct list_head *a, struct list_head *b)
1693 struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio,
1695 struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio,
1697 u64 a_sector = ra->bio_list.head->bi_iter.bi_sector;
1698 u64 b_sector = rb->bio_list.head->bi_iter.bi_sector;
1700 if (a_sector < b_sector)
1702 if (a_sector > b_sector)
1707 static void run_plug(struct btrfs_plug_cb *plug)
1709 struct btrfs_raid_bio *cur;
1710 struct btrfs_raid_bio *last = NULL;
1713 * sort our plug list then try to merge
1714 * everything we can in hopes of creating full
1717 list_sort(NULL, &plug->rbio_list, plug_cmp);
1718 while (!list_empty(&plug->rbio_list)) {
1719 cur = list_entry(plug->rbio_list.next,
1720 struct btrfs_raid_bio, plug_list);
1721 list_del_init(&cur->plug_list);
1723 if (rbio_is_full(cur)) {
1726 /* we have a full stripe, send it down */
1727 ret = full_stripe_write(cur);
1732 if (rbio_can_merge(last, cur)) {
1733 merge_rbio(last, cur);
1734 __free_raid_bio(cur);
1738 __raid56_parity_write(last);
1743 __raid56_parity_write(last);
1749 * if the unplug comes from schedule, we have to push the
1750 * work off to a helper thread
1752 static void unplug_work(struct btrfs_work *work)
1754 struct btrfs_plug_cb *plug;
1755 plug = container_of(work, struct btrfs_plug_cb, work);
1759 static void btrfs_raid_unplug(struct blk_plug_cb *cb, bool from_schedule)
1761 struct btrfs_plug_cb *plug;
1762 plug = container_of(cb, struct btrfs_plug_cb, cb);
1764 if (from_schedule) {
1765 btrfs_init_work(&plug->work, unplug_work, NULL, NULL);
1766 btrfs_queue_work(plug->info->rmw_workers,
1773 /* Add the original bio into rbio->bio_list, and update rbio::dbitmap. */
1774 static void rbio_add_bio(struct btrfs_raid_bio *rbio, struct bio *orig_bio)
1776 const struct btrfs_fs_info *fs_info = rbio->fs_info;
1777 const u64 orig_logical = orig_bio->bi_iter.bi_sector << SECTOR_SHIFT;
1778 const u64 full_stripe_start = rbio->bbio->raid_map[0];
1779 const u32 orig_len = orig_bio->bi_iter.bi_size;
1780 const u32 sectorsize = fs_info->sectorsize;
1783 ASSERT(orig_logical >= full_stripe_start &&
1784 orig_logical + orig_len <= full_stripe_start +
1785 rbio->nr_data * rbio->stripe_len);
1787 bio_list_add(&rbio->bio_list, orig_bio);
1788 rbio->bio_list_bytes += orig_bio->bi_iter.bi_size;
1790 /* Update the dbitmap. */
1791 for (cur_logical = orig_logical; cur_logical < orig_logical + orig_len;
1792 cur_logical += sectorsize) {
1793 int bit = ((u32)(cur_logical - full_stripe_start) >>
1794 PAGE_SHIFT) % rbio->stripe_npages;
1796 set_bit(bit, rbio->dbitmap);
1801 * our main entry point for writes from the rest of the FS.
1803 int raid56_parity_write(struct btrfs_fs_info *fs_info, struct bio *bio,
1804 struct btrfs_bio *bbio, u64 stripe_len)
1806 struct btrfs_raid_bio *rbio;
1807 struct btrfs_plug_cb *plug = NULL;
1808 struct blk_plug_cb *cb;
1811 rbio = alloc_rbio(fs_info, bbio, stripe_len);
1813 btrfs_put_bbio(bbio);
1814 return PTR_ERR(rbio);
1816 rbio->operation = BTRFS_RBIO_WRITE;
1817 rbio_add_bio(rbio, bio);
1819 btrfs_bio_counter_inc_noblocked(fs_info);
1820 rbio->generic_bio_cnt = 1;
1823 * don't plug on full rbios, just get them out the door
1824 * as quickly as we can
1826 if (rbio_is_full(rbio)) {
1827 ret = full_stripe_write(rbio);
1829 btrfs_bio_counter_dec(fs_info);
1833 cb = blk_check_plugged(btrfs_raid_unplug, fs_info, sizeof(*plug));
1835 plug = container_of(cb, struct btrfs_plug_cb, cb);
1837 plug->info = fs_info;
1838 INIT_LIST_HEAD(&plug->rbio_list);
1840 list_add_tail(&rbio->plug_list, &plug->rbio_list);
1843 ret = __raid56_parity_write(rbio);
1845 btrfs_bio_counter_dec(fs_info);
1851 * all parity reconstruction happens here. We've read in everything
1852 * we can find from the drives and this does the heavy lifting of
1853 * sorting the good from the bad.
1855 static void __raid_recover_end_io(struct btrfs_raid_bio *rbio)
1859 int faila = -1, failb = -1;
1864 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
1866 err = BLK_STS_RESOURCE;
1870 faila = rbio->faila;
1871 failb = rbio->failb;
1873 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1874 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
1875 spin_lock_irq(&rbio->bio_list_lock);
1876 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1877 spin_unlock_irq(&rbio->bio_list_lock);
1880 index_rbio_pages(rbio);
1882 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1884 * Now we just use bitmap to mark the horizontal stripes in
1885 * which we have data when doing parity scrub.
1887 if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB &&
1888 !test_bit(pagenr, rbio->dbitmap))
1891 /* setup our array of pointers with pages
1894 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1896 * if we're rebuilding a read, we have to use
1897 * pages from the bio list
1899 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1900 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
1901 (stripe == faila || stripe == failb)) {
1902 page = page_in_rbio(rbio, stripe, pagenr, 0);
1904 page = rbio_stripe_page(rbio, stripe, pagenr);
1906 pointers[stripe] = kmap(page);
1909 /* all raid6 handling here */
1910 if (rbio->bbio->map_type & BTRFS_BLOCK_GROUP_RAID6) {
1912 * single failure, rebuild from parity raid5
1916 if (faila == rbio->nr_data) {
1918 * Just the P stripe has failed, without
1919 * a bad data or Q stripe.
1920 * TODO, we should redo the xor here.
1922 err = BLK_STS_IOERR;
1926 * a single failure in raid6 is rebuilt
1927 * in the pstripe code below
1932 /* make sure our ps and qs are in order */
1933 if (faila > failb) {
1939 /* if the q stripe is failed, do a pstripe reconstruction
1941 * If both the q stripe and the P stripe are failed, we're
1942 * here due to a crc mismatch and we can't give them the
1945 if (rbio->bbio->raid_map[failb] == RAID6_Q_STRIPE) {
1946 if (rbio->bbio->raid_map[faila] ==
1948 err = BLK_STS_IOERR;
1952 * otherwise we have one bad data stripe and
1953 * a good P stripe. raid5!
1958 if (rbio->bbio->raid_map[failb] == RAID5_P_STRIPE) {
1959 raid6_datap_recov(rbio->real_stripes,
1960 PAGE_SIZE, faila, pointers);
1962 raid6_2data_recov(rbio->real_stripes,
1963 PAGE_SIZE, faila, failb,
1969 /* rebuild from P stripe here (raid5 or raid6) */
1970 BUG_ON(failb != -1);
1972 /* Copy parity block into failed block to start with */
1973 copy_page(pointers[faila], pointers[rbio->nr_data]);
1975 /* rearrange the pointer array */
1976 p = pointers[faila];
1977 for (stripe = faila; stripe < rbio->nr_data - 1; stripe++)
1978 pointers[stripe] = pointers[stripe + 1];
1979 pointers[rbio->nr_data - 1] = p;
1981 /* xor in the rest */
1982 run_xor(pointers, rbio->nr_data - 1, PAGE_SIZE);
1984 /* if we're doing this rebuild as part of an rmw, go through
1985 * and set all of our private rbio pages in the
1986 * failed stripes as uptodate. This way finish_rmw will
1987 * know they can be trusted. If this was a read reconstruction,
1988 * other endio functions will fiddle the uptodate bits
1990 if (rbio->operation == BTRFS_RBIO_WRITE) {
1991 for (i = 0; i < rbio->stripe_npages; i++) {
1993 page = rbio_stripe_page(rbio, faila, i);
1994 SetPageUptodate(page);
1997 page = rbio_stripe_page(rbio, failb, i);
1998 SetPageUptodate(page);
2002 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
2004 * if we're rebuilding a read, we have to use
2005 * pages from the bio list
2007 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
2008 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
2009 (stripe == faila || stripe == failb)) {
2010 page = page_in_rbio(rbio, stripe, pagenr, 0);
2012 page = rbio_stripe_page(rbio, stripe, pagenr);
2024 * Similar to READ_REBUILD, REBUILD_MISSING at this point also has a
2025 * valid rbio which is consistent with ondisk content, thus such a
2026 * valid rbio can be cached to avoid further disk reads.
2028 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
2029 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
2031 * - In case of two failures, where rbio->failb != -1:
2033 * Do not cache this rbio since the above read reconstruction
2034 * (raid6_datap_recov() or raid6_2data_recov()) may have
2035 * changed some content of stripes which are not identical to
2036 * on-disk content any more, otherwise, a later write/recover
2037 * may steal stripe_pages from this rbio and end up with
2038 * corruptions or rebuild failures.
2040 * - In case of single failure, where rbio->failb == -1:
2042 * Cache this rbio iff the above read reconstruction is
2043 * executed without problems.
2045 if (err == BLK_STS_OK && rbio->failb < 0)
2046 cache_rbio_pages(rbio);
2048 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2050 rbio_orig_end_io(rbio, err);
2051 } else if (err == BLK_STS_OK) {
2055 if (rbio->operation == BTRFS_RBIO_WRITE)
2057 else if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB)
2058 finish_parity_scrub(rbio, 0);
2062 rbio_orig_end_io(rbio, err);
2067 * This is called only for stripes we've read from disk to
2068 * reconstruct the parity.
2070 static void raid_recover_end_io(struct bio *bio)
2072 struct btrfs_raid_bio *rbio = bio->bi_private;
2075 * we only read stripe pages off the disk, set them
2076 * up to date if there were no errors
2079 fail_bio_stripe(rbio, bio);
2081 set_bio_pages_uptodate(bio);
2084 if (!atomic_dec_and_test(&rbio->stripes_pending))
2087 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
2088 rbio_orig_end_io(rbio, BLK_STS_IOERR);
2090 __raid_recover_end_io(rbio);
2094 * reads everything we need off the disk to reconstruct
2095 * the parity. endio handlers trigger final reconstruction
2096 * when the IO is done.
2098 * This is used both for reads from the higher layers and for
2099 * parity construction required to finish a rmw cycle.
2101 static int __raid56_parity_recover(struct btrfs_raid_bio *rbio)
2103 int bios_to_read = 0;
2104 struct bio_list bio_list;
2110 bio_list_init(&bio_list);
2112 ret = alloc_rbio_pages(rbio);
2116 atomic_set(&rbio->error, 0);
2119 * Read everything that hasn't failed. However this time we will
2120 * not trust any cached sector.
2121 * As we may read out some stale data but higher layer is not reading
2124 * So here we always re-read everything in recovery path.
2126 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
2127 if (rbio->faila == stripe || rbio->failb == stripe) {
2128 atomic_inc(&rbio->error);
2132 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
2133 ret = rbio_add_io_page(rbio, &bio_list,
2134 rbio_stripe_page(rbio, stripe, pagenr),
2135 stripe, pagenr, rbio->stripe_len);
2141 bios_to_read = bio_list_size(&bio_list);
2142 if (!bios_to_read) {
2144 * we might have no bios to read just because the pages
2145 * were up to date, or we might have no bios to read because
2146 * the devices were gone.
2148 if (atomic_read(&rbio->error) <= rbio->bbio->max_errors) {
2149 __raid_recover_end_io(rbio);
2157 * the bbio may be freed once we submit the last bio. Make sure
2158 * not to touch it after that
2160 atomic_set(&rbio->stripes_pending, bios_to_read);
2162 bio = bio_list_pop(&bio_list);
2166 bio->bi_private = rbio;
2167 bio->bi_end_io = raid_recover_end_io;
2168 bio->bi_opf = REQ_OP_READ;
2170 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
2178 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
2179 rbio->operation == BTRFS_RBIO_REBUILD_MISSING)
2180 rbio_orig_end_io(rbio, BLK_STS_IOERR);
2182 while ((bio = bio_list_pop(&bio_list)))
2189 * the main entry point for reads from the higher layers. This
2190 * is really only called when the normal read path had a failure,
2191 * so we assume the bio they send down corresponds to a failed part
2194 int raid56_parity_recover(struct btrfs_fs_info *fs_info, struct bio *bio,
2195 struct btrfs_bio *bbio, u64 stripe_len,
2196 int mirror_num, int generic_io)
2198 struct btrfs_raid_bio *rbio;
2202 ASSERT(bbio->mirror_num == mirror_num);
2203 btrfs_io_bio(bio)->mirror_num = mirror_num;
2206 rbio = alloc_rbio(fs_info, bbio, stripe_len);
2209 btrfs_put_bbio(bbio);
2210 return PTR_ERR(rbio);
2213 rbio->operation = BTRFS_RBIO_READ_REBUILD;
2214 rbio_add_bio(rbio, bio);
2216 rbio->faila = find_logical_bio_stripe(rbio, bio);
2217 if (rbio->faila == -1) {
2219 "%s could not find the bad stripe in raid56 so that we cannot recover any more (bio has logical %llu len %llu, bbio has map_type %llu)",
2220 __func__, (u64)bio->bi_iter.bi_sector << 9,
2221 (u64)bio->bi_iter.bi_size, bbio->map_type);
2223 btrfs_put_bbio(bbio);
2229 btrfs_bio_counter_inc_noblocked(fs_info);
2230 rbio->generic_bio_cnt = 1;
2232 btrfs_get_bbio(bbio);
2237 * for 'mirror == 2', reconstruct from all other stripes.
2238 * for 'mirror_num > 2', select a stripe to fail on every retry.
2240 if (mirror_num > 2) {
2242 * 'mirror == 3' is to fail the p stripe and
2243 * reconstruct from the q stripe. 'mirror > 3' is to
2244 * fail a data stripe and reconstruct from p+q stripe.
2246 rbio->failb = rbio->real_stripes - (mirror_num - 1);
2247 ASSERT(rbio->failb > 0);
2248 if (rbio->failb <= rbio->faila)
2252 ret = lock_stripe_add(rbio);
2255 * __raid56_parity_recover will end the bio with
2256 * any errors it hits. We don't want to return
2257 * its error value up the stack because our caller
2258 * will end up calling bio_endio with any nonzero
2262 __raid56_parity_recover(rbio);
2264 * our rbio has been added to the list of
2265 * rbios that will be handled after the
2266 * currently lock owner is done
2272 static void rmw_work(struct btrfs_work *work)
2274 struct btrfs_raid_bio *rbio;
2276 rbio = container_of(work, struct btrfs_raid_bio, work);
2277 raid56_rmw_stripe(rbio);
2280 static void read_rebuild_work(struct btrfs_work *work)
2282 struct btrfs_raid_bio *rbio;
2284 rbio = container_of(work, struct btrfs_raid_bio, work);
2285 __raid56_parity_recover(rbio);
2289 * The following code is used to scrub/replace the parity stripe
2291 * Caller must have already increased bio_counter for getting @bbio.
2293 * Note: We need make sure all the pages that add into the scrub/replace
2294 * raid bio are correct and not be changed during the scrub/replace. That
2295 * is those pages just hold metadata or file data with checksum.
2298 struct btrfs_raid_bio *
2299 raid56_parity_alloc_scrub_rbio(struct btrfs_fs_info *fs_info, struct bio *bio,
2300 struct btrfs_bio *bbio, u64 stripe_len,
2301 struct btrfs_device *scrub_dev,
2302 unsigned long *dbitmap, int stripe_nsectors)
2304 struct btrfs_raid_bio *rbio;
2307 rbio = alloc_rbio(fs_info, bbio, stripe_len);
2310 bio_list_add(&rbio->bio_list, bio);
2312 * This is a special bio which is used to hold the completion handler
2313 * and make the scrub rbio is similar to the other types
2315 ASSERT(!bio->bi_iter.bi_size);
2316 rbio->operation = BTRFS_RBIO_PARITY_SCRUB;
2319 * After mapping bbio with BTRFS_MAP_WRITE, parities have been sorted
2320 * to the end position, so this search can start from the first parity
2323 for (i = rbio->nr_data; i < rbio->real_stripes; i++) {
2324 if (bbio->stripes[i].dev == scrub_dev) {
2329 ASSERT(i < rbio->real_stripes);
2331 /* Now we just support the sectorsize equals to page size */
2332 ASSERT(fs_info->sectorsize == PAGE_SIZE);
2333 ASSERT(rbio->stripe_npages == stripe_nsectors);
2334 bitmap_copy(rbio->dbitmap, dbitmap, stripe_nsectors);
2337 * We have already increased bio_counter when getting bbio, record it
2338 * so we can free it at rbio_orig_end_io().
2340 rbio->generic_bio_cnt = 1;
2345 /* Used for both parity scrub and missing. */
2346 void raid56_add_scrub_pages(struct btrfs_raid_bio *rbio, struct page *page,
2352 ASSERT(logical >= rbio->bbio->raid_map[0]);
2353 ASSERT(logical + PAGE_SIZE <= rbio->bbio->raid_map[0] +
2354 rbio->stripe_len * rbio->nr_data);
2355 stripe_offset = (int)(logical - rbio->bbio->raid_map[0]);
2356 index = stripe_offset >> PAGE_SHIFT;
2357 rbio->bio_pages[index] = page;
2361 * We just scrub the parity that we have correct data on the same horizontal,
2362 * so we needn't allocate all pages for all the stripes.
2364 static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio)
2371 for_each_set_bit(bit, rbio->dbitmap, rbio->stripe_npages) {
2372 for (i = 0; i < rbio->real_stripes; i++) {
2373 index = i * rbio->stripe_npages + bit;
2374 if (rbio->stripe_pages[index])
2377 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2380 rbio->stripe_pages[index] = page;
2386 static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
2389 struct btrfs_bio *bbio = rbio->bbio;
2390 void **pointers = rbio->finish_pointers;
2391 unsigned long *pbitmap = rbio->finish_pbitmap;
2392 int nr_data = rbio->nr_data;
2396 struct page *p_page = NULL;
2397 struct page *q_page = NULL;
2398 struct bio_list bio_list;
2403 bio_list_init(&bio_list);
2405 if (rbio->real_stripes - rbio->nr_data == 1)
2406 has_qstripe = false;
2407 else if (rbio->real_stripes - rbio->nr_data == 2)
2412 if (bbio->num_tgtdevs && bbio->tgtdev_map[rbio->scrubp]) {
2414 bitmap_copy(pbitmap, rbio->dbitmap, rbio->stripe_npages);
2418 * Because the higher layers(scrubber) are unlikely to
2419 * use this area of the disk again soon, so don't cache
2422 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2427 p_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2430 SetPageUptodate(p_page);
2433 /* RAID6, allocate and map temp space for the Q stripe */
2434 q_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2436 __free_page(p_page);
2439 SetPageUptodate(q_page);
2440 pointers[rbio->real_stripes - 1] = kmap(q_page);
2443 atomic_set(&rbio->error, 0);
2445 /* Map the parity stripe just once */
2446 pointers[nr_data] = kmap(p_page);
2448 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2451 /* first collect one page from each data stripe */
2452 for (stripe = 0; stripe < nr_data; stripe++) {
2453 p = page_in_rbio(rbio, stripe, pagenr, 0);
2454 pointers[stripe] = kmap(p);
2458 /* RAID6, call the library function to fill in our P/Q */
2459 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
2463 copy_page(pointers[nr_data], pointers[0]);
2464 run_xor(pointers + 1, nr_data - 1, PAGE_SIZE);
2467 /* Check scrubbing parity and repair it */
2468 p = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2470 if (memcmp(parity, pointers[rbio->scrubp], PAGE_SIZE))
2471 copy_page(parity, pointers[rbio->scrubp]);
2473 /* Parity is right, needn't writeback */
2474 bitmap_clear(rbio->dbitmap, pagenr, 1);
2477 for (stripe = 0; stripe < nr_data; stripe++)
2478 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
2482 __free_page(p_page);
2485 __free_page(q_page);
2490 * time to start writing. Make bios for everything from the
2491 * higher layers (the bio_list in our rbio) and our p/q. Ignore
2494 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2497 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2498 ret = rbio_add_io_page(rbio, &bio_list,
2499 page, rbio->scrubp, pagenr, rbio->stripe_len);
2507 for_each_set_bit(pagenr, pbitmap, rbio->stripe_npages) {
2510 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2511 ret = rbio_add_io_page(rbio, &bio_list, page,
2512 bbio->tgtdev_map[rbio->scrubp],
2513 pagenr, rbio->stripe_len);
2519 nr_data = bio_list_size(&bio_list);
2521 /* Every parity is right */
2522 rbio_orig_end_io(rbio, BLK_STS_OK);
2526 atomic_set(&rbio->stripes_pending, nr_data);
2529 bio = bio_list_pop(&bio_list);
2533 bio->bi_private = rbio;
2534 bio->bi_end_io = raid_write_end_io;
2535 bio->bi_opf = REQ_OP_WRITE;
2542 rbio_orig_end_io(rbio, BLK_STS_IOERR);
2544 while ((bio = bio_list_pop(&bio_list)))
2548 static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe)
2550 if (stripe >= 0 && stripe < rbio->nr_data)
2556 * While we're doing the parity check and repair, we could have errors
2557 * in reading pages off the disk. This checks for errors and if we're
2558 * not able to read the page it'll trigger parity reconstruction. The
2559 * parity scrub will be finished after we've reconstructed the failed
2562 static void validate_rbio_for_parity_scrub(struct btrfs_raid_bio *rbio)
2564 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
2567 if (rbio->faila >= 0 || rbio->failb >= 0) {
2568 int dfail = 0, failp = -1;
2570 if (is_data_stripe(rbio, rbio->faila))
2572 else if (is_parity_stripe(rbio->faila))
2573 failp = rbio->faila;
2575 if (is_data_stripe(rbio, rbio->failb))
2577 else if (is_parity_stripe(rbio->failb))
2578 failp = rbio->failb;
2581 * Because we can not use a scrubbing parity to repair
2582 * the data, so the capability of the repair is declined.
2583 * (In the case of RAID5, we can not repair anything)
2585 if (dfail > rbio->bbio->max_errors - 1)
2589 * If all data is good, only parity is correctly, just
2590 * repair the parity.
2593 finish_parity_scrub(rbio, 0);
2598 * Here means we got one corrupted data stripe and one
2599 * corrupted parity on RAID6, if the corrupted parity
2600 * is scrubbing parity, luckily, use the other one to repair
2601 * the data, or we can not repair the data stripe.
2603 if (failp != rbio->scrubp)
2606 __raid_recover_end_io(rbio);
2608 finish_parity_scrub(rbio, 1);
2613 rbio_orig_end_io(rbio, BLK_STS_IOERR);
2617 * end io for the read phase of the rmw cycle. All the bios here are physical
2618 * stripe bios we've read from the disk so we can recalculate the parity of the
2621 * This will usually kick off finish_rmw once all the bios are read in, but it
2622 * may trigger parity reconstruction if we had any errors along the way
2624 static void raid56_parity_scrub_end_io(struct bio *bio)
2626 struct btrfs_raid_bio *rbio = bio->bi_private;
2629 fail_bio_stripe(rbio, bio);
2631 set_bio_pages_uptodate(bio);
2635 if (!atomic_dec_and_test(&rbio->stripes_pending))
2639 * this will normally call finish_rmw to start our write
2640 * but if there are any failed stripes we'll reconstruct
2643 validate_rbio_for_parity_scrub(rbio);
2646 static void raid56_parity_scrub_stripe(struct btrfs_raid_bio *rbio)
2648 int bios_to_read = 0;
2649 struct bio_list bio_list;
2655 bio_list_init(&bio_list);
2657 ret = alloc_rbio_essential_pages(rbio);
2661 atomic_set(&rbio->error, 0);
2663 * build a list of bios to read all the missing parts of this
2666 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
2667 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2670 * we want to find all the pages missing from
2671 * the rbio and read them from the disk. If
2672 * page_in_rbio finds a page in the bio list
2673 * we don't need to read it off the stripe.
2675 page = page_in_rbio(rbio, stripe, pagenr, 1);
2679 page = rbio_stripe_page(rbio, stripe, pagenr);
2681 * the bio cache may have handed us an uptodate
2682 * page. If so, be happy and use it
2684 if (PageUptodate(page))
2687 ret = rbio_add_io_page(rbio, &bio_list, page,
2688 stripe, pagenr, rbio->stripe_len);
2694 bios_to_read = bio_list_size(&bio_list);
2695 if (!bios_to_read) {
2697 * this can happen if others have merged with
2698 * us, it means there is nothing left to read.
2699 * But if there are missing devices it may not be
2700 * safe to do the full stripe write yet.
2706 * the bbio may be freed once we submit the last bio. Make sure
2707 * not to touch it after that
2709 atomic_set(&rbio->stripes_pending, bios_to_read);
2711 bio = bio_list_pop(&bio_list);
2715 bio->bi_private = rbio;
2716 bio->bi_end_io = raid56_parity_scrub_end_io;
2717 bio->bi_opf = REQ_OP_READ;
2719 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
2723 /* the actual write will happen once the reads are done */
2727 rbio_orig_end_io(rbio, BLK_STS_IOERR);
2729 while ((bio = bio_list_pop(&bio_list)))
2735 validate_rbio_for_parity_scrub(rbio);
2738 static void scrub_parity_work(struct btrfs_work *work)
2740 struct btrfs_raid_bio *rbio;
2742 rbio = container_of(work, struct btrfs_raid_bio, work);
2743 raid56_parity_scrub_stripe(rbio);
2746 void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio)
2748 if (!lock_stripe_add(rbio))
2749 start_async_work(rbio, scrub_parity_work);
2752 /* The following code is used for dev replace of a missing RAID 5/6 device. */
2754 struct btrfs_raid_bio *
2755 raid56_alloc_missing_rbio(struct btrfs_fs_info *fs_info, struct bio *bio,
2756 struct btrfs_bio *bbio, u64 length)
2758 struct btrfs_raid_bio *rbio;
2760 rbio = alloc_rbio(fs_info, bbio, length);
2764 rbio->operation = BTRFS_RBIO_REBUILD_MISSING;
2765 bio_list_add(&rbio->bio_list, bio);
2767 * This is a special bio which is used to hold the completion handler
2768 * and make the scrub rbio is similar to the other types
2770 ASSERT(!bio->bi_iter.bi_size);
2772 rbio->faila = find_logical_bio_stripe(rbio, bio);
2773 if (rbio->faila == -1) {
2780 * When we get bbio, we have already increased bio_counter, record it
2781 * so we can free it at rbio_orig_end_io()
2783 rbio->generic_bio_cnt = 1;
2788 void raid56_submit_missing_rbio(struct btrfs_raid_bio *rbio)
2790 if (!lock_stripe_add(rbio))
2791 start_async_work(rbio, read_rebuild_work);