2 * Copyright (C) 2012 Fusion-io All rights reserved.
3 * Copyright (C) 2012 Intel Corp. All rights reserved.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public
7 * License v2 as published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * General Public License for more details.
14 * You should have received a copy of the GNU General Public
15 * License along with this program; if not, write to the
16 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
17 * Boston, MA 021110-1307, USA.
19 #include <linux/sched.h>
20 #include <linux/wait.h>
21 #include <linux/bio.h>
22 #include <linux/slab.h>
23 #include <linux/buffer_head.h>
24 #include <linux/blkdev.h>
25 #include <linux/random.h>
26 #include <linux/iocontext.h>
27 #include <linux/capability.h>
28 #include <linux/ratelimit.h>
29 #include <linux/kthread.h>
30 #include <linux/raid/pq.h>
31 #include <linux/hash.h>
32 #include <linux/list_sort.h>
33 #include <linux/raid/xor.h>
34 #include <linux/vmalloc.h>
35 #include <asm/div64.h>
37 #include "extent_map.h"
39 #include "transaction.h"
40 #include "print-tree.h"
43 #include "async-thread.h"
44 #include "check-integrity.h"
45 #include "rcu-string.h"
47 /* set when additional merges to this rbio are not allowed */
48 #define RBIO_RMW_LOCKED_BIT 1
51 * set when this rbio is sitting in the hash, but it is just a cache
54 #define RBIO_CACHE_BIT 2
57 * set when it is safe to trust the stripe_pages for caching
59 #define RBIO_CACHE_READY_BIT 3
61 #define RBIO_CACHE_SIZE 1024
65 BTRFS_RBIO_READ_REBUILD,
66 BTRFS_RBIO_PARITY_SCRUB,
67 BTRFS_RBIO_REBUILD_MISSING,
70 struct btrfs_raid_bio {
71 struct btrfs_fs_info *fs_info;
72 struct btrfs_bio *bbio;
74 /* while we're doing rmw on a stripe
75 * we put it into a hash table so we can
76 * lock the stripe and merge more rbios
79 struct list_head hash_list;
82 * LRU list for the stripe cache
84 struct list_head stripe_cache;
87 * for scheduling work in the helper threads
89 struct btrfs_work work;
92 * bio list and bio_list_lock are used
93 * to add more bios into the stripe
94 * in hopes of avoiding the full rmw
96 struct bio_list bio_list;
97 spinlock_t bio_list_lock;
99 /* also protected by the bio_list_lock, the
100 * plug list is used by the plugging code
101 * to collect partial bios while plugged. The
102 * stripe locking code also uses it to hand off
103 * the stripe lock to the next pending IO
105 struct list_head plug_list;
108 * flags that tell us if it is safe to
109 * merge with this bio
113 /* size of each individual stripe on disk */
116 /* number of data stripes (no p/q) */
123 * set if we're doing a parity rebuild
124 * for a read from higher up, which is handled
125 * differently from a parity rebuild as part of
128 enum btrfs_rbio_ops operation;
130 /* first bad stripe */
133 /* second bad stripe (for raid6 use) */
138 * number of pages needed to represent the full
144 * size of all the bios in the bio_list. This
145 * helps us decide if the rbio maps to a full
154 atomic_t stripes_pending;
158 * these are two arrays of pointers. We allocate the
159 * rbio big enough to hold them both and setup their
160 * locations when the rbio is allocated
163 /* pointers to pages that we allocated for
164 * reading/writing stripes directly from the disk (including P/Q)
166 struct page **stripe_pages;
169 * pointers to the pages in the bio_list. Stored
170 * here for faster lookup
172 struct page **bio_pages;
175 * bitmap to record which horizontal stripe has data
177 unsigned long *dbitmap;
180 static int __raid56_parity_recover(struct btrfs_raid_bio *rbio);
181 static noinline void finish_rmw(struct btrfs_raid_bio *rbio);
182 static void rmw_work(struct btrfs_work *work);
183 static void read_rebuild_work(struct btrfs_work *work);
184 static void async_rmw_stripe(struct btrfs_raid_bio *rbio);
185 static void async_read_rebuild(struct btrfs_raid_bio *rbio);
186 static int fail_bio_stripe(struct btrfs_raid_bio *rbio, struct bio *bio);
187 static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed);
188 static void __free_raid_bio(struct btrfs_raid_bio *rbio);
189 static void index_rbio_pages(struct btrfs_raid_bio *rbio);
190 static int alloc_rbio_pages(struct btrfs_raid_bio *rbio);
192 static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
194 static void async_scrub_parity(struct btrfs_raid_bio *rbio);
197 * the stripe hash table is used for locking, and to collect
198 * bios in hopes of making a full stripe
200 int btrfs_alloc_stripe_hash_table(struct btrfs_fs_info *info)
202 struct btrfs_stripe_hash_table *table;
203 struct btrfs_stripe_hash_table *x;
204 struct btrfs_stripe_hash *cur;
205 struct btrfs_stripe_hash *h;
206 int num_entries = 1 << BTRFS_STRIPE_HASH_TABLE_BITS;
210 if (info->stripe_hash_table)
214 * The table is large, starting with order 4 and can go as high as
215 * order 7 in case lock debugging is turned on.
217 * Try harder to allocate and fallback to vmalloc to lower the chance
218 * of a failing mount.
220 table_size = sizeof(*table) + sizeof(*h) * num_entries;
221 table = kzalloc(table_size, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
223 table = vzalloc(table_size);
228 spin_lock_init(&table->cache_lock);
229 INIT_LIST_HEAD(&table->stripe_cache);
233 for (i = 0; i < num_entries; i++) {
235 INIT_LIST_HEAD(&cur->hash_list);
236 spin_lock_init(&cur->lock);
237 init_waitqueue_head(&cur->wait);
240 x = cmpxchg(&info->stripe_hash_table, NULL, table);
247 * caching an rbio means to copy anything from the
248 * bio_pages array into the stripe_pages array. We
249 * use the page uptodate bit in the stripe cache array
250 * to indicate if it has valid data
252 * once the caching is done, we set the cache ready
255 static void cache_rbio_pages(struct btrfs_raid_bio *rbio)
262 ret = alloc_rbio_pages(rbio);
266 for (i = 0; i < rbio->nr_pages; i++) {
267 if (!rbio->bio_pages[i])
270 s = kmap(rbio->bio_pages[i]);
271 d = kmap(rbio->stripe_pages[i]);
273 memcpy(d, s, PAGE_SIZE);
275 kunmap(rbio->bio_pages[i]);
276 kunmap(rbio->stripe_pages[i]);
277 SetPageUptodate(rbio->stripe_pages[i]);
279 set_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
283 * we hash on the first logical address of the stripe
285 static int rbio_bucket(struct btrfs_raid_bio *rbio)
287 u64 num = rbio->bbio->raid_map[0];
290 * we shift down quite a bit. We're using byte
291 * addressing, and most of the lower bits are zeros.
292 * This tends to upset hash_64, and it consistently
293 * returns just one or two different values.
295 * shifting off the lower bits fixes things.
297 return hash_64(num >> 16, BTRFS_STRIPE_HASH_TABLE_BITS);
301 * stealing an rbio means taking all the uptodate pages from the stripe
302 * array in the source rbio and putting them into the destination rbio
304 static void steal_rbio(struct btrfs_raid_bio *src, struct btrfs_raid_bio *dest)
310 if (!test_bit(RBIO_CACHE_READY_BIT, &src->flags))
313 for (i = 0; i < dest->nr_pages; i++) {
314 s = src->stripe_pages[i];
315 if (!s || !PageUptodate(s)) {
319 d = dest->stripe_pages[i];
323 dest->stripe_pages[i] = s;
324 src->stripe_pages[i] = NULL;
329 * merging means we take the bio_list from the victim and
330 * splice it into the destination. The victim should
331 * be discarded afterwards.
333 * must be called with dest->rbio_list_lock held
335 static void merge_rbio(struct btrfs_raid_bio *dest,
336 struct btrfs_raid_bio *victim)
338 bio_list_merge(&dest->bio_list, &victim->bio_list);
339 dest->bio_list_bytes += victim->bio_list_bytes;
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 atomic_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 atomic_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
530 * covers an entire stripe (no rmw required).
531 * Must be called with the bio list lock held, or
532 * at a time when you know it is impossible to add
533 * new bios into the list
535 static int __rbio_is_full(struct btrfs_raid_bio *rbio)
537 unsigned long size = rbio->bio_list_bytes;
540 if (size != rbio->nr_data * rbio->stripe_len)
543 BUG_ON(size > rbio->nr_data * rbio->stripe_len);
547 static int rbio_is_full(struct btrfs_raid_bio *rbio)
552 spin_lock_irqsave(&rbio->bio_list_lock, flags);
553 ret = __rbio_is_full(rbio);
554 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
559 * returns 1 if it is safe to merge two rbios together.
560 * The merging is safe if the two rbios correspond to
561 * the same stripe and if they are both going in the same
562 * direction (read vs write), and if neither one is
563 * locked for final IO
565 * The caller is responsible for locking such that
566 * rmw_locked is safe to test
568 static int rbio_can_merge(struct btrfs_raid_bio *last,
569 struct btrfs_raid_bio *cur)
571 if (test_bit(RBIO_RMW_LOCKED_BIT, &last->flags) ||
572 test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags))
576 * we can't merge with cached rbios, since the
577 * idea is that when we merge the destination
578 * rbio is going to run our IO for us. We can
579 * steal from cached rbios though, other functions
582 if (test_bit(RBIO_CACHE_BIT, &last->flags) ||
583 test_bit(RBIO_CACHE_BIT, &cur->flags))
586 if (last->bbio->raid_map[0] !=
587 cur->bbio->raid_map[0])
590 /* we can't merge with different operations */
591 if (last->operation != cur->operation)
594 * We've need read the full stripe from the drive.
595 * check and repair the parity and write the new results.
597 * We're not allowed to add any new bios to the
598 * bio list here, anyone else that wants to
599 * change this stripe needs to do their own rmw.
601 if (last->operation == BTRFS_RBIO_PARITY_SCRUB ||
602 cur->operation == BTRFS_RBIO_PARITY_SCRUB)
605 if (last->operation == BTRFS_RBIO_REBUILD_MISSING ||
606 cur->operation == BTRFS_RBIO_REBUILD_MISSING)
612 static int rbio_stripe_page_index(struct btrfs_raid_bio *rbio, int stripe,
615 return stripe * rbio->stripe_npages + index;
619 * these are just the pages from the rbio array, not from anything
620 * the FS sent down to us
622 static struct page *rbio_stripe_page(struct btrfs_raid_bio *rbio, int stripe,
625 return rbio->stripe_pages[rbio_stripe_page_index(rbio, stripe, index)];
629 * helper to index into the pstripe
631 static struct page *rbio_pstripe_page(struct btrfs_raid_bio *rbio, int index)
633 return rbio_stripe_page(rbio, rbio->nr_data, index);
637 * helper to index into the qstripe, returns null
638 * if there is no qstripe
640 static struct page *rbio_qstripe_page(struct btrfs_raid_bio *rbio, int index)
642 if (rbio->nr_data + 1 == rbio->real_stripes)
644 return rbio_stripe_page(rbio, rbio->nr_data + 1, index);
648 * The first stripe in the table for a logical address
649 * has the lock. rbios are added in one of three ways:
651 * 1) Nobody has the stripe locked yet. The rbio is given
652 * the lock and 0 is returned. The caller must start the IO
655 * 2) Someone has the stripe locked, but we're able to merge
656 * with the lock owner. The rbio is freed and the IO will
657 * start automatically along with the existing rbio. 1 is returned.
659 * 3) Someone has the stripe locked, but we're not able to merge.
660 * The rbio is added to the lock owner's plug list, or merged into
661 * an rbio already on the plug list. When the lock owner unlocks,
662 * the next rbio on the list is run and the IO is started automatically.
665 * If we return 0, the caller still owns the rbio and must continue with
666 * IO submission. If we return 1, the caller must assume the rbio has
667 * already been freed.
669 static noinline int lock_stripe_add(struct btrfs_raid_bio *rbio)
671 int bucket = rbio_bucket(rbio);
672 struct btrfs_stripe_hash *h = rbio->fs_info->stripe_hash_table->table + bucket;
673 struct btrfs_raid_bio *cur;
674 struct btrfs_raid_bio *pending;
677 struct btrfs_raid_bio *freeit = NULL;
678 struct btrfs_raid_bio *cache_drop = NULL;
682 spin_lock_irqsave(&h->lock, flags);
683 list_for_each_entry(cur, &h->hash_list, hash_list) {
685 if (cur->bbio->raid_map[0] == rbio->bbio->raid_map[0]) {
686 spin_lock(&cur->bio_list_lock);
688 /* can we steal this cached rbio's pages? */
689 if (bio_list_empty(&cur->bio_list) &&
690 list_empty(&cur->plug_list) &&
691 test_bit(RBIO_CACHE_BIT, &cur->flags) &&
692 !test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) {
693 list_del_init(&cur->hash_list);
694 atomic_dec(&cur->refs);
696 steal_rbio(cur, rbio);
698 spin_unlock(&cur->bio_list_lock);
703 /* can we merge into the lock owner? */
704 if (rbio_can_merge(cur, rbio)) {
705 merge_rbio(cur, rbio);
706 spin_unlock(&cur->bio_list_lock);
714 * we couldn't merge with the running
715 * rbio, see if we can merge with the
716 * pending ones. We don't have to
717 * check for rmw_locked because there
718 * is no way they are inside finish_rmw
721 list_for_each_entry(pending, &cur->plug_list,
723 if (rbio_can_merge(pending, rbio)) {
724 merge_rbio(pending, rbio);
725 spin_unlock(&cur->bio_list_lock);
732 /* no merging, put us on the tail of the plug list,
733 * our rbio will be started with the currently
734 * running rbio unlocks
736 list_add_tail(&rbio->plug_list, &cur->plug_list);
737 spin_unlock(&cur->bio_list_lock);
743 atomic_inc(&rbio->refs);
744 list_add(&rbio->hash_list, &h->hash_list);
746 spin_unlock_irqrestore(&h->lock, flags);
748 remove_rbio_from_cache(cache_drop);
750 __free_raid_bio(freeit);
755 * called as rmw or parity rebuild is completed. If the plug list has more
756 * rbios waiting for this stripe, the next one on the list will be started
758 static noinline void unlock_stripe(struct btrfs_raid_bio *rbio)
761 struct btrfs_stripe_hash *h;
765 bucket = rbio_bucket(rbio);
766 h = rbio->fs_info->stripe_hash_table->table + bucket;
768 if (list_empty(&rbio->plug_list))
771 spin_lock_irqsave(&h->lock, flags);
772 spin_lock(&rbio->bio_list_lock);
774 if (!list_empty(&rbio->hash_list)) {
776 * if we're still cached and there is no other IO
777 * to perform, just leave this rbio here for others
778 * to steal from later
780 if (list_empty(&rbio->plug_list) &&
781 test_bit(RBIO_CACHE_BIT, &rbio->flags)) {
783 clear_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
784 BUG_ON(!bio_list_empty(&rbio->bio_list));
788 list_del_init(&rbio->hash_list);
789 atomic_dec(&rbio->refs);
792 * we use the plug list to hold all the rbios
793 * waiting for the chance to lock this stripe.
794 * hand the lock over to one of them.
796 if (!list_empty(&rbio->plug_list)) {
797 struct btrfs_raid_bio *next;
798 struct list_head *head = rbio->plug_list.next;
800 next = list_entry(head, struct btrfs_raid_bio,
803 list_del_init(&rbio->plug_list);
805 list_add(&next->hash_list, &h->hash_list);
806 atomic_inc(&next->refs);
807 spin_unlock(&rbio->bio_list_lock);
808 spin_unlock_irqrestore(&h->lock, flags);
810 if (next->operation == BTRFS_RBIO_READ_REBUILD)
811 async_read_rebuild(next);
812 else if (next->operation == BTRFS_RBIO_REBUILD_MISSING) {
813 steal_rbio(rbio, next);
814 async_read_rebuild(next);
815 } else if (next->operation == BTRFS_RBIO_WRITE) {
816 steal_rbio(rbio, next);
817 async_rmw_stripe(next);
818 } else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) {
819 steal_rbio(rbio, next);
820 async_scrub_parity(next);
825 * The barrier for this waitqueue_active is not needed,
826 * we're protected by h->lock and can't miss a wakeup.
828 } else if (waitqueue_active(&h->wait)) {
829 spin_unlock(&rbio->bio_list_lock);
830 spin_unlock_irqrestore(&h->lock, flags);
836 spin_unlock(&rbio->bio_list_lock);
837 spin_unlock_irqrestore(&h->lock, flags);
841 remove_rbio_from_cache(rbio);
844 static void __free_raid_bio(struct btrfs_raid_bio *rbio)
848 WARN_ON(atomic_read(&rbio->refs) < 0);
849 if (!atomic_dec_and_test(&rbio->refs))
852 WARN_ON(!list_empty(&rbio->stripe_cache));
853 WARN_ON(!list_empty(&rbio->hash_list));
854 WARN_ON(!bio_list_empty(&rbio->bio_list));
856 for (i = 0; i < rbio->nr_pages; i++) {
857 if (rbio->stripe_pages[i]) {
858 __free_page(rbio->stripe_pages[i]);
859 rbio->stripe_pages[i] = NULL;
863 btrfs_put_bbio(rbio->bbio);
867 static void free_raid_bio(struct btrfs_raid_bio *rbio)
870 __free_raid_bio(rbio);
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, int 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);
897 * end io function used by finish_rmw. When we finally
898 * get here, we've written a full stripe
900 static void raid_write_end_io(struct bio *bio)
902 struct btrfs_raid_bio *rbio = bio->bi_private;
903 int err = bio->bi_error;
907 fail_bio_stripe(rbio, bio);
911 if (!atomic_dec_and_test(&rbio->stripes_pending))
916 /* OK, we have read all the stripes we need to. */
917 max_errors = (rbio->operation == BTRFS_RBIO_PARITY_SCRUB) ?
918 0 : rbio->bbio->max_errors;
919 if (atomic_read(&rbio->error) > max_errors)
922 rbio_orig_end_io(rbio, err);
926 * the read/modify/write code wants to use the original bio for
927 * any pages it included, and then use the rbio for everything
928 * else. This function decides if a given index (stripe number)
929 * and page number in that stripe fall inside the original bio
932 * if you set bio_list_only, you'll get a NULL back for any ranges
933 * that are outside the bio_list
935 * This doesn't take any refs on anything, you get a bare page pointer
936 * and the caller must bump refs as required.
938 * You must call index_rbio_pages once before you can trust
939 * the answers from this function.
941 static struct page *page_in_rbio(struct btrfs_raid_bio *rbio,
942 int index, int pagenr, int bio_list_only)
945 struct page *p = NULL;
947 chunk_page = index * (rbio->stripe_len >> PAGE_SHIFT) + pagenr;
949 spin_lock_irq(&rbio->bio_list_lock);
950 p = rbio->bio_pages[chunk_page];
951 spin_unlock_irq(&rbio->bio_list_lock);
953 if (p || bio_list_only)
956 return rbio->stripe_pages[chunk_page];
960 * number of pages we need for the entire stripe across all the
963 static unsigned long rbio_nr_pages(unsigned long stripe_len, int nr_stripes)
965 return DIV_ROUND_UP(stripe_len, PAGE_SIZE) * nr_stripes;
969 * allocation and initial setup for the btrfs_raid_bio. Not
970 * this does not allocate any pages for rbio->pages.
972 static struct btrfs_raid_bio *alloc_rbio(struct btrfs_fs_info *fs_info,
973 struct btrfs_bio *bbio,
976 struct btrfs_raid_bio *rbio;
978 int real_stripes = bbio->num_stripes - bbio->num_tgtdevs;
979 int num_pages = rbio_nr_pages(stripe_len, real_stripes);
980 int stripe_npages = DIV_ROUND_UP(stripe_len, PAGE_SIZE);
983 rbio = kzalloc(sizeof(*rbio) + num_pages * sizeof(struct page *) * 2 +
984 DIV_ROUND_UP(stripe_npages, BITS_PER_LONG) *
985 sizeof(long), GFP_NOFS);
987 return ERR_PTR(-ENOMEM);
989 bio_list_init(&rbio->bio_list);
990 INIT_LIST_HEAD(&rbio->plug_list);
991 spin_lock_init(&rbio->bio_list_lock);
992 INIT_LIST_HEAD(&rbio->stripe_cache);
993 INIT_LIST_HEAD(&rbio->hash_list);
995 rbio->fs_info = fs_info;
996 rbio->stripe_len = stripe_len;
997 rbio->nr_pages = num_pages;
998 rbio->real_stripes = real_stripes;
999 rbio->stripe_npages = stripe_npages;
1002 atomic_set(&rbio->refs, 1);
1003 atomic_set(&rbio->error, 0);
1004 atomic_set(&rbio->stripes_pending, 0);
1007 * the stripe_pages and bio_pages array point to the extra
1008 * memory we allocated past the end of the rbio
1011 rbio->stripe_pages = p;
1012 rbio->bio_pages = p + sizeof(struct page *) * num_pages;
1013 rbio->dbitmap = p + sizeof(struct page *) * num_pages * 2;
1015 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1016 nr_data = real_stripes - 1;
1017 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1018 nr_data = real_stripes - 2;
1022 rbio->nr_data = nr_data;
1026 /* allocate pages for all the stripes in the bio, including parity */
1027 static int alloc_rbio_pages(struct btrfs_raid_bio *rbio)
1032 for (i = 0; i < rbio->nr_pages; i++) {
1033 if (rbio->stripe_pages[i])
1035 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1038 rbio->stripe_pages[i] = page;
1043 /* only allocate pages for p/q stripes */
1044 static int alloc_rbio_parity_pages(struct btrfs_raid_bio *rbio)
1049 i = rbio_stripe_page_index(rbio, rbio->nr_data, 0);
1051 for (; i < rbio->nr_pages; i++) {
1052 if (rbio->stripe_pages[i])
1054 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1057 rbio->stripe_pages[i] = page;
1063 * add a single page from a specific stripe into our list of bios for IO
1064 * this will try to merge into existing bios if possible, and returns
1065 * zero if all went well.
1067 static int rbio_add_io_page(struct btrfs_raid_bio *rbio,
1068 struct bio_list *bio_list,
1071 unsigned long page_index,
1072 unsigned long bio_max_len)
1074 struct bio *last = bio_list->tail;
1078 struct btrfs_bio_stripe *stripe;
1081 stripe = &rbio->bbio->stripes[stripe_nr];
1082 disk_start = stripe->physical + (page_index << PAGE_SHIFT);
1084 /* if the device is missing, just fail this stripe */
1085 if (!stripe->dev->bdev)
1086 return fail_rbio_index(rbio, stripe_nr);
1088 /* see if we can add this page onto our existing bio */
1090 last_end = (u64)last->bi_iter.bi_sector << 9;
1091 last_end += last->bi_iter.bi_size;
1094 * we can't merge these if they are from different
1095 * devices or if they are not contiguous
1097 if (last_end == disk_start && stripe->dev->bdev &&
1099 last->bi_bdev == stripe->dev->bdev) {
1100 ret = bio_add_page(last, page, PAGE_SIZE, 0);
1101 if (ret == PAGE_SIZE)
1106 /* put a new bio on the list */
1107 bio = btrfs_io_bio_alloc(GFP_NOFS, bio_max_len >> PAGE_SHIFT?:1);
1111 bio->bi_iter.bi_size = 0;
1112 bio->bi_bdev = stripe->dev->bdev;
1113 bio->bi_iter.bi_sector = disk_start >> 9;
1115 bio_add_page(bio, page, PAGE_SIZE, 0);
1116 bio_list_add(bio_list, bio);
1121 * while we're doing the read/modify/write cycle, we could
1122 * have errors in reading pages off the disk. This checks
1123 * for errors and if we're not able to read the page it'll
1124 * trigger parity reconstruction. The rmw will be finished
1125 * after we've reconstructed the failed stripes
1127 static void validate_rbio_for_rmw(struct btrfs_raid_bio *rbio)
1129 if (rbio->faila >= 0 || rbio->failb >= 0) {
1130 BUG_ON(rbio->faila == rbio->real_stripes - 1);
1131 __raid56_parity_recover(rbio);
1138 * helper function to walk our bio list and populate the bio_pages array with
1139 * the result. This seems expensive, but it is faster than constantly
1140 * searching through the bio list as we setup the IO in finish_rmw or stripe
1143 * This must be called before you trust the answers from page_in_rbio
1145 static void index_rbio_pages(struct btrfs_raid_bio *rbio)
1148 struct bio_vec *bvec;
1150 unsigned long stripe_offset;
1151 unsigned long page_index;
1154 spin_lock_irq(&rbio->bio_list_lock);
1155 bio_list_for_each(bio, &rbio->bio_list) {
1156 start = (u64)bio->bi_iter.bi_sector << 9;
1157 stripe_offset = start - rbio->bbio->raid_map[0];
1158 page_index = stripe_offset >> PAGE_SHIFT;
1160 bio_for_each_segment_all(bvec, bio, i)
1161 rbio->bio_pages[page_index + i] = bvec->bv_page;
1163 spin_unlock_irq(&rbio->bio_list_lock);
1167 * this is called from one of two situations. We either
1168 * have a full stripe from the higher layers, or we've read all
1169 * the missing bits off disk.
1171 * This will calculate the parity and then send down any
1174 static noinline void finish_rmw(struct btrfs_raid_bio *rbio)
1176 struct btrfs_bio *bbio = rbio->bbio;
1177 void *pointers[rbio->real_stripes];
1178 int nr_data = rbio->nr_data;
1183 struct bio_list bio_list;
1187 bio_list_init(&bio_list);
1189 if (rbio->real_stripes - rbio->nr_data == 1) {
1190 p_stripe = rbio->real_stripes - 1;
1191 } else if (rbio->real_stripes - rbio->nr_data == 2) {
1192 p_stripe = rbio->real_stripes - 2;
1193 q_stripe = rbio->real_stripes - 1;
1198 /* at this point we either have a full stripe,
1199 * or we've read the full stripe from the drive.
1200 * recalculate the parity and write the new results.
1202 * We're not allowed to add any new bios to the
1203 * bio list here, anyone else that wants to
1204 * change this stripe needs to do their own rmw.
1206 spin_lock_irq(&rbio->bio_list_lock);
1207 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1208 spin_unlock_irq(&rbio->bio_list_lock);
1210 atomic_set(&rbio->error, 0);
1213 * now that we've set rmw_locked, run through the
1214 * bio list one last time and map the page pointers
1216 * We don't cache full rbios because we're assuming
1217 * the higher layers are unlikely to use this area of
1218 * the disk again soon. If they do use it again,
1219 * hopefully they will send another full bio.
1221 index_rbio_pages(rbio);
1222 if (!rbio_is_full(rbio))
1223 cache_rbio_pages(rbio);
1225 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
1227 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1229 /* first collect one page from each data stripe */
1230 for (stripe = 0; stripe < nr_data; stripe++) {
1231 p = page_in_rbio(rbio, stripe, pagenr, 0);
1232 pointers[stripe] = kmap(p);
1235 /* then add the parity stripe */
1236 p = rbio_pstripe_page(rbio, pagenr);
1238 pointers[stripe++] = kmap(p);
1240 if (q_stripe != -1) {
1243 * raid6, add the qstripe and call the
1244 * library function to fill in our p/q
1246 p = rbio_qstripe_page(rbio, pagenr);
1248 pointers[stripe++] = kmap(p);
1250 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
1254 memcpy(pointers[nr_data], pointers[0], PAGE_SIZE);
1255 run_xor(pointers + 1, nr_data - 1, PAGE_SIZE);
1259 for (stripe = 0; stripe < rbio->real_stripes; stripe++)
1260 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
1264 * time to start writing. Make bios for everything from the
1265 * higher layers (the bio_list in our rbio) and our p/q. Ignore
1268 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1269 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1271 if (stripe < rbio->nr_data) {
1272 page = page_in_rbio(rbio, stripe, pagenr, 1);
1276 page = rbio_stripe_page(rbio, stripe, pagenr);
1279 ret = rbio_add_io_page(rbio, &bio_list,
1280 page, stripe, pagenr, rbio->stripe_len);
1286 if (likely(!bbio->num_tgtdevs))
1289 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1290 if (!bbio->tgtdev_map[stripe])
1293 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1295 if (stripe < rbio->nr_data) {
1296 page = page_in_rbio(rbio, stripe, pagenr, 1);
1300 page = rbio_stripe_page(rbio, stripe, pagenr);
1303 ret = rbio_add_io_page(rbio, &bio_list, page,
1304 rbio->bbio->tgtdev_map[stripe],
1305 pagenr, rbio->stripe_len);
1312 atomic_set(&rbio->stripes_pending, bio_list_size(&bio_list));
1313 BUG_ON(atomic_read(&rbio->stripes_pending) == 0);
1316 bio = bio_list_pop(&bio_list);
1320 bio->bi_private = rbio;
1321 bio->bi_end_io = raid_write_end_io;
1322 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1329 rbio_orig_end_io(rbio, -EIO);
1333 * helper to find the stripe number for a given bio. Used to figure out which
1334 * stripe has failed. This expects the bio to correspond to a physical disk,
1335 * so it looks up based on physical sector numbers.
1337 static int find_bio_stripe(struct btrfs_raid_bio *rbio,
1340 u64 physical = bio->bi_iter.bi_sector;
1343 struct btrfs_bio_stripe *stripe;
1347 for (i = 0; i < rbio->bbio->num_stripes; i++) {
1348 stripe = &rbio->bbio->stripes[i];
1349 stripe_start = stripe->physical;
1350 if (physical >= stripe_start &&
1351 physical < stripe_start + rbio->stripe_len &&
1352 bio->bi_bdev == stripe->dev->bdev) {
1360 * helper to find the stripe number for a given
1361 * bio (before mapping). Used to figure out which stripe has
1362 * failed. This looks up based on logical block numbers.
1364 static int find_logical_bio_stripe(struct btrfs_raid_bio *rbio,
1367 u64 logical = bio->bi_iter.bi_sector;
1373 for (i = 0; i < rbio->nr_data; i++) {
1374 stripe_start = rbio->bbio->raid_map[i];
1375 if (logical >= stripe_start &&
1376 logical < stripe_start + rbio->stripe_len) {
1384 * returns -EIO if we had too many failures
1386 static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed)
1388 unsigned long flags;
1391 spin_lock_irqsave(&rbio->bio_list_lock, flags);
1393 /* we already know this stripe is bad, move on */
1394 if (rbio->faila == failed || rbio->failb == failed)
1397 if (rbio->faila == -1) {
1398 /* first failure on this rbio */
1399 rbio->faila = failed;
1400 atomic_inc(&rbio->error);
1401 } else if (rbio->failb == -1) {
1402 /* second failure on this rbio */
1403 rbio->failb = failed;
1404 atomic_inc(&rbio->error);
1409 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
1415 * helper to fail a stripe based on a physical disk
1418 static int fail_bio_stripe(struct btrfs_raid_bio *rbio,
1421 int failed = find_bio_stripe(rbio, bio);
1426 return fail_rbio_index(rbio, failed);
1430 * this sets each page in the bio uptodate. It should only be used on private
1431 * rbio pages, nothing that comes in from the higher layers
1433 static void set_bio_pages_uptodate(struct bio *bio)
1435 struct bio_vec *bvec;
1438 bio_for_each_segment_all(bvec, bio, i)
1439 SetPageUptodate(bvec->bv_page);
1443 * end io for the read phase of the rmw cycle. All the bios here are physical
1444 * stripe bios we've read from the disk so we can recalculate the parity of the
1447 * This will usually kick off finish_rmw once all the bios are read in, but it
1448 * may trigger parity reconstruction if we had any errors along the way
1450 static void raid_rmw_end_io(struct bio *bio)
1452 struct btrfs_raid_bio *rbio = bio->bi_private;
1455 fail_bio_stripe(rbio, bio);
1457 set_bio_pages_uptodate(bio);
1461 if (!atomic_dec_and_test(&rbio->stripes_pending))
1464 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
1468 * this will normally call finish_rmw to start our write
1469 * but if there are any failed stripes we'll reconstruct
1472 validate_rbio_for_rmw(rbio);
1477 rbio_orig_end_io(rbio, -EIO);
1480 static void async_rmw_stripe(struct btrfs_raid_bio *rbio)
1482 btrfs_init_work(&rbio->work, btrfs_rmw_helper, rmw_work, NULL, NULL);
1483 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
1486 static void async_read_rebuild(struct btrfs_raid_bio *rbio)
1488 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
1489 read_rebuild_work, NULL, NULL);
1491 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
1495 * the stripe must be locked by the caller. It will
1496 * unlock after all the writes are done
1498 static int raid56_rmw_stripe(struct btrfs_raid_bio *rbio)
1500 int bios_to_read = 0;
1501 struct bio_list bio_list;
1507 bio_list_init(&bio_list);
1509 ret = alloc_rbio_pages(rbio);
1513 index_rbio_pages(rbio);
1515 atomic_set(&rbio->error, 0);
1517 * build a list of bios to read all the missing parts of this
1520 for (stripe = 0; stripe < rbio->nr_data; stripe++) {
1521 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1524 * we want to find all the pages missing from
1525 * the rbio and read them from the disk. If
1526 * page_in_rbio finds a page in the bio list
1527 * we don't need to read it off the stripe.
1529 page = page_in_rbio(rbio, stripe, pagenr, 1);
1533 page = rbio_stripe_page(rbio, stripe, pagenr);
1535 * the bio cache may have handed us an uptodate
1536 * page. If so, be happy and use it
1538 if (PageUptodate(page))
1541 ret = rbio_add_io_page(rbio, &bio_list, page,
1542 stripe, pagenr, rbio->stripe_len);
1548 bios_to_read = bio_list_size(&bio_list);
1549 if (!bios_to_read) {
1551 * this can happen if others have merged with
1552 * us, it means there is nothing left to read.
1553 * But if there are missing devices it may not be
1554 * safe to do the full stripe write yet.
1560 * the bbio may be freed once we submit the last bio. Make sure
1561 * not to touch it after that
1563 atomic_set(&rbio->stripes_pending, bios_to_read);
1565 bio = bio_list_pop(&bio_list);
1569 bio->bi_private = rbio;
1570 bio->bi_end_io = raid_rmw_end_io;
1571 bio_set_op_attrs(bio, REQ_OP_READ, 0);
1573 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
1577 /* the actual write will happen once the reads are done */
1581 rbio_orig_end_io(rbio, -EIO);
1585 validate_rbio_for_rmw(rbio);
1590 * if the upper layers pass in a full stripe, we thank them by only allocating
1591 * enough pages to hold the parity, and sending it all down quickly.
1593 static int full_stripe_write(struct btrfs_raid_bio *rbio)
1597 ret = alloc_rbio_parity_pages(rbio);
1599 __free_raid_bio(rbio);
1603 ret = lock_stripe_add(rbio);
1610 * partial stripe writes get handed over to async helpers.
1611 * We're really hoping to merge a few more writes into this
1612 * rbio before calculating new parity
1614 static int partial_stripe_write(struct btrfs_raid_bio *rbio)
1618 ret = lock_stripe_add(rbio);
1620 async_rmw_stripe(rbio);
1625 * sometimes while we were reading from the drive to
1626 * recalculate parity, enough new bios come into create
1627 * a full stripe. So we do a check here to see if we can
1628 * go directly to finish_rmw
1630 static int __raid56_parity_write(struct btrfs_raid_bio *rbio)
1632 /* head off into rmw land if we don't have a full stripe */
1633 if (!rbio_is_full(rbio))
1634 return partial_stripe_write(rbio);
1635 return full_stripe_write(rbio);
1639 * We use plugging call backs to collect full stripes.
1640 * Any time we get a partial stripe write while plugged
1641 * we collect it into a list. When the unplug comes down,
1642 * we sort the list by logical block number and merge
1643 * everything we can into the same rbios
1645 struct btrfs_plug_cb {
1646 struct blk_plug_cb cb;
1647 struct btrfs_fs_info *info;
1648 struct list_head rbio_list;
1649 struct btrfs_work work;
1653 * rbios on the plug list are sorted for easier merging.
1655 static int plug_cmp(void *priv, struct list_head *a, struct list_head *b)
1657 struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio,
1659 struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio,
1661 u64 a_sector = ra->bio_list.head->bi_iter.bi_sector;
1662 u64 b_sector = rb->bio_list.head->bi_iter.bi_sector;
1664 if (a_sector < b_sector)
1666 if (a_sector > b_sector)
1671 static void run_plug(struct btrfs_plug_cb *plug)
1673 struct btrfs_raid_bio *cur;
1674 struct btrfs_raid_bio *last = NULL;
1677 * sort our plug list then try to merge
1678 * everything we can in hopes of creating full
1681 list_sort(NULL, &plug->rbio_list, plug_cmp);
1682 while (!list_empty(&plug->rbio_list)) {
1683 cur = list_entry(plug->rbio_list.next,
1684 struct btrfs_raid_bio, plug_list);
1685 list_del_init(&cur->plug_list);
1687 if (rbio_is_full(cur)) {
1688 /* we have a full stripe, send it down */
1689 full_stripe_write(cur);
1693 if (rbio_can_merge(last, cur)) {
1694 merge_rbio(last, cur);
1695 __free_raid_bio(cur);
1699 __raid56_parity_write(last);
1704 __raid56_parity_write(last);
1710 * if the unplug comes from schedule, we have to push the
1711 * work off to a helper thread
1713 static void unplug_work(struct btrfs_work *work)
1715 struct btrfs_plug_cb *plug;
1716 plug = container_of(work, struct btrfs_plug_cb, work);
1720 static void btrfs_raid_unplug(struct blk_plug_cb *cb, bool from_schedule)
1722 struct btrfs_plug_cb *plug;
1723 plug = container_of(cb, struct btrfs_plug_cb, cb);
1725 if (from_schedule) {
1726 btrfs_init_work(&plug->work, btrfs_rmw_helper,
1727 unplug_work, NULL, NULL);
1728 btrfs_queue_work(plug->info->rmw_workers,
1736 * our main entry point for writes from the rest of the FS.
1738 int raid56_parity_write(struct btrfs_fs_info *fs_info, struct bio *bio,
1739 struct btrfs_bio *bbio, u64 stripe_len)
1741 struct btrfs_raid_bio *rbio;
1742 struct btrfs_plug_cb *plug = NULL;
1743 struct blk_plug_cb *cb;
1746 rbio = alloc_rbio(fs_info, bbio, stripe_len);
1748 btrfs_put_bbio(bbio);
1749 return PTR_ERR(rbio);
1751 bio_list_add(&rbio->bio_list, bio);
1752 rbio->bio_list_bytes = bio->bi_iter.bi_size;
1753 rbio->operation = BTRFS_RBIO_WRITE;
1755 btrfs_bio_counter_inc_noblocked(fs_info);
1756 rbio->generic_bio_cnt = 1;
1759 * don't plug on full rbios, just get them out the door
1760 * as quickly as we can
1762 if (rbio_is_full(rbio)) {
1763 ret = full_stripe_write(rbio);
1765 btrfs_bio_counter_dec(fs_info);
1769 cb = blk_check_plugged(btrfs_raid_unplug, fs_info, sizeof(*plug));
1771 plug = container_of(cb, struct btrfs_plug_cb, cb);
1773 plug->info = fs_info;
1774 INIT_LIST_HEAD(&plug->rbio_list);
1776 list_add_tail(&rbio->plug_list, &plug->rbio_list);
1779 ret = __raid56_parity_write(rbio);
1781 btrfs_bio_counter_dec(fs_info);
1787 * all parity reconstruction happens here. We've read in everything
1788 * we can find from the drives and this does the heavy lifting of
1789 * sorting the good from the bad.
1791 static void __raid_recover_end_io(struct btrfs_raid_bio *rbio)
1795 int faila = -1, failb = -1;
1800 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
1806 faila = rbio->faila;
1807 failb = rbio->failb;
1809 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1810 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
1811 spin_lock_irq(&rbio->bio_list_lock);
1812 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1813 spin_unlock_irq(&rbio->bio_list_lock);
1816 index_rbio_pages(rbio);
1818 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1820 * Now we just use bitmap to mark the horizontal stripes in
1821 * which we have data when doing parity scrub.
1823 if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB &&
1824 !test_bit(pagenr, rbio->dbitmap))
1827 /* setup our array of pointers with pages
1830 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1832 * if we're rebuilding a read, we have to use
1833 * pages from the bio list
1835 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1836 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
1837 (stripe == faila || stripe == failb)) {
1838 page = page_in_rbio(rbio, stripe, pagenr, 0);
1840 page = rbio_stripe_page(rbio, stripe, pagenr);
1842 pointers[stripe] = kmap(page);
1845 /* all raid6 handling here */
1846 if (rbio->bbio->map_type & BTRFS_BLOCK_GROUP_RAID6) {
1848 * single failure, rebuild from parity raid5
1852 if (faila == rbio->nr_data) {
1854 * Just the P stripe has failed, without
1855 * a bad data or Q stripe.
1856 * TODO, we should redo the xor here.
1862 * a single failure in raid6 is rebuilt
1863 * in the pstripe code below
1868 /* make sure our ps and qs are in order */
1869 if (faila > failb) {
1875 /* if the q stripe is failed, do a pstripe reconstruction
1877 * If both the q stripe and the P stripe are failed, we're
1878 * here due to a crc mismatch and we can't give them the
1881 if (rbio->bbio->raid_map[failb] == RAID6_Q_STRIPE) {
1882 if (rbio->bbio->raid_map[faila] ==
1888 * otherwise we have one bad data stripe and
1889 * a good P stripe. raid5!
1894 if (rbio->bbio->raid_map[failb] == RAID5_P_STRIPE) {
1895 raid6_datap_recov(rbio->real_stripes,
1896 PAGE_SIZE, faila, pointers);
1898 raid6_2data_recov(rbio->real_stripes,
1899 PAGE_SIZE, faila, failb,
1905 /* rebuild from P stripe here (raid5 or raid6) */
1906 BUG_ON(failb != -1);
1908 /* Copy parity block into failed block to start with */
1909 memcpy(pointers[faila],
1910 pointers[rbio->nr_data],
1913 /* rearrange the pointer array */
1914 p = pointers[faila];
1915 for (stripe = faila; stripe < rbio->nr_data - 1; stripe++)
1916 pointers[stripe] = pointers[stripe + 1];
1917 pointers[rbio->nr_data - 1] = p;
1919 /* xor in the rest */
1920 run_xor(pointers, rbio->nr_data - 1, PAGE_SIZE);
1922 /* if we're doing this rebuild as part of an rmw, go through
1923 * and set all of our private rbio pages in the
1924 * failed stripes as uptodate. This way finish_rmw will
1925 * know they can be trusted. If this was a read reconstruction,
1926 * other endio functions will fiddle the uptodate bits
1928 if (rbio->operation == BTRFS_RBIO_WRITE) {
1929 for (i = 0; i < rbio->stripe_npages; i++) {
1931 page = rbio_stripe_page(rbio, faila, i);
1932 SetPageUptodate(page);
1935 page = rbio_stripe_page(rbio, failb, i);
1936 SetPageUptodate(page);
1940 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1942 * if we're rebuilding a read, we have to use
1943 * pages from the bio list
1945 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1946 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
1947 (stripe == faila || stripe == failb)) {
1948 page = page_in_rbio(rbio, stripe, pagenr, 0);
1950 page = rbio_stripe_page(rbio, stripe, pagenr);
1961 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
1963 cache_rbio_pages(rbio);
1965 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
1967 rbio_orig_end_io(rbio, err);
1968 } else if (rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
1969 rbio_orig_end_io(rbio, err);
1970 } else if (err == 0) {
1974 if (rbio->operation == BTRFS_RBIO_WRITE)
1976 else if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB)
1977 finish_parity_scrub(rbio, 0);
1981 rbio_orig_end_io(rbio, err);
1986 * This is called only for stripes we've read from disk to
1987 * reconstruct the parity.
1989 static void raid_recover_end_io(struct bio *bio)
1991 struct btrfs_raid_bio *rbio = bio->bi_private;
1994 * we only read stripe pages off the disk, set them
1995 * up to date if there were no errors
1998 fail_bio_stripe(rbio, bio);
2000 set_bio_pages_uptodate(bio);
2003 if (!atomic_dec_and_test(&rbio->stripes_pending))
2006 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
2007 rbio_orig_end_io(rbio, -EIO);
2009 __raid_recover_end_io(rbio);
2013 * reads everything we need off the disk to reconstruct
2014 * the parity. endio handlers trigger final reconstruction
2015 * when the IO is done.
2017 * This is used both for reads from the higher layers and for
2018 * parity construction required to finish a rmw cycle.
2020 static int __raid56_parity_recover(struct btrfs_raid_bio *rbio)
2022 int bios_to_read = 0;
2023 struct bio_list bio_list;
2029 bio_list_init(&bio_list);
2031 ret = alloc_rbio_pages(rbio);
2035 atomic_set(&rbio->error, 0);
2038 * read everything that hasn't failed. Thanks to the
2039 * stripe cache, it is possible that some or all of these
2040 * pages are going to be uptodate.
2042 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
2043 if (rbio->faila == stripe || rbio->failb == stripe) {
2044 atomic_inc(&rbio->error);
2048 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
2052 * the rmw code may have already read this
2055 p = rbio_stripe_page(rbio, stripe, pagenr);
2056 if (PageUptodate(p))
2059 ret = rbio_add_io_page(rbio, &bio_list,
2060 rbio_stripe_page(rbio, stripe, pagenr),
2061 stripe, pagenr, rbio->stripe_len);
2067 bios_to_read = bio_list_size(&bio_list);
2068 if (!bios_to_read) {
2070 * we might have no bios to read just because the pages
2071 * were up to date, or we might have no bios to read because
2072 * the devices were gone.
2074 if (atomic_read(&rbio->error) <= rbio->bbio->max_errors) {
2075 __raid_recover_end_io(rbio);
2083 * the bbio may be freed once we submit the last bio. Make sure
2084 * not to touch it after that
2086 atomic_set(&rbio->stripes_pending, bios_to_read);
2088 bio = bio_list_pop(&bio_list);
2092 bio->bi_private = rbio;
2093 bio->bi_end_io = raid_recover_end_io;
2094 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2096 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
2104 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
2105 rbio->operation == BTRFS_RBIO_REBUILD_MISSING)
2106 rbio_orig_end_io(rbio, -EIO);
2111 * the main entry point for reads from the higher layers. This
2112 * is really only called when the normal read path had a failure,
2113 * so we assume the bio they send down corresponds to a failed part
2116 int raid56_parity_recover(struct btrfs_fs_info *fs_info, struct bio *bio,
2117 struct btrfs_bio *bbio, u64 stripe_len,
2118 int mirror_num, int generic_io)
2120 struct btrfs_raid_bio *rbio;
2123 rbio = alloc_rbio(fs_info, bbio, stripe_len);
2126 btrfs_put_bbio(bbio);
2127 return PTR_ERR(rbio);
2130 rbio->operation = BTRFS_RBIO_READ_REBUILD;
2131 bio_list_add(&rbio->bio_list, bio);
2132 rbio->bio_list_bytes = bio->bi_iter.bi_size;
2134 rbio->faila = find_logical_bio_stripe(rbio, bio);
2135 if (rbio->faila == -1) {
2137 "%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)",
2138 __func__, (u64)bio->bi_iter.bi_sector << 9,
2139 (u64)bio->bi_iter.bi_size, bbio->map_type);
2141 btrfs_put_bbio(bbio);
2147 btrfs_bio_counter_inc_noblocked(fs_info);
2148 rbio->generic_bio_cnt = 1;
2150 btrfs_get_bbio(bbio);
2154 * reconstruct from the q stripe if they are
2155 * asking for mirror 3
2157 if (mirror_num == 3)
2158 rbio->failb = rbio->real_stripes - 2;
2160 ret = lock_stripe_add(rbio);
2163 * __raid56_parity_recover will end the bio with
2164 * any errors it hits. We don't want to return
2165 * its error value up the stack because our caller
2166 * will end up calling bio_endio with any nonzero
2170 __raid56_parity_recover(rbio);
2172 * our rbio has been added to the list of
2173 * rbios that will be handled after the
2174 * currently lock owner is done
2180 static void rmw_work(struct btrfs_work *work)
2182 struct btrfs_raid_bio *rbio;
2184 rbio = container_of(work, struct btrfs_raid_bio, work);
2185 raid56_rmw_stripe(rbio);
2188 static void read_rebuild_work(struct btrfs_work *work)
2190 struct btrfs_raid_bio *rbio;
2192 rbio = container_of(work, struct btrfs_raid_bio, work);
2193 __raid56_parity_recover(rbio);
2197 * The following code is used to scrub/replace the parity stripe
2199 * Note: We need make sure all the pages that add into the scrub/replace
2200 * raid bio are correct and not be changed during the scrub/replace. That
2201 * is those pages just hold metadata or file data with checksum.
2204 struct btrfs_raid_bio *
2205 raid56_parity_alloc_scrub_rbio(struct btrfs_fs_info *fs_info, struct bio *bio,
2206 struct btrfs_bio *bbio, u64 stripe_len,
2207 struct btrfs_device *scrub_dev,
2208 unsigned long *dbitmap, int stripe_nsectors)
2210 struct btrfs_raid_bio *rbio;
2213 rbio = alloc_rbio(fs_info, bbio, stripe_len);
2216 bio_list_add(&rbio->bio_list, bio);
2218 * This is a special bio which is used to hold the completion handler
2219 * and make the scrub rbio is similar to the other types
2221 ASSERT(!bio->bi_iter.bi_size);
2222 rbio->operation = BTRFS_RBIO_PARITY_SCRUB;
2224 for (i = 0; i < rbio->real_stripes; i++) {
2225 if (bbio->stripes[i].dev == scrub_dev) {
2231 /* Now we just support the sectorsize equals to page size */
2232 ASSERT(fs_info->sectorsize == PAGE_SIZE);
2233 ASSERT(rbio->stripe_npages == stripe_nsectors);
2234 bitmap_copy(rbio->dbitmap, dbitmap, stripe_nsectors);
2239 /* Used for both parity scrub and missing. */
2240 void raid56_add_scrub_pages(struct btrfs_raid_bio *rbio, struct page *page,
2246 ASSERT(logical >= rbio->bbio->raid_map[0]);
2247 ASSERT(logical + PAGE_SIZE <= rbio->bbio->raid_map[0] +
2248 rbio->stripe_len * rbio->nr_data);
2249 stripe_offset = (int)(logical - rbio->bbio->raid_map[0]);
2250 index = stripe_offset >> PAGE_SHIFT;
2251 rbio->bio_pages[index] = page;
2255 * We just scrub the parity that we have correct data on the same horizontal,
2256 * so we needn't allocate all pages for all the stripes.
2258 static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio)
2265 for_each_set_bit(bit, rbio->dbitmap, rbio->stripe_npages) {
2266 for (i = 0; i < rbio->real_stripes; i++) {
2267 index = i * rbio->stripe_npages + bit;
2268 if (rbio->stripe_pages[index])
2271 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2274 rbio->stripe_pages[index] = page;
2280 static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
2283 struct btrfs_bio *bbio = rbio->bbio;
2284 void *pointers[rbio->real_stripes];
2285 DECLARE_BITMAP(pbitmap, rbio->stripe_npages);
2286 int nr_data = rbio->nr_data;
2291 struct page *p_page = NULL;
2292 struct page *q_page = NULL;
2293 struct bio_list bio_list;
2298 bio_list_init(&bio_list);
2300 if (rbio->real_stripes - rbio->nr_data == 1) {
2301 p_stripe = rbio->real_stripes - 1;
2302 } else if (rbio->real_stripes - rbio->nr_data == 2) {
2303 p_stripe = rbio->real_stripes - 2;
2304 q_stripe = rbio->real_stripes - 1;
2309 if (bbio->num_tgtdevs && bbio->tgtdev_map[rbio->scrubp]) {
2311 bitmap_copy(pbitmap, rbio->dbitmap, rbio->stripe_npages);
2315 * Because the higher layers(scrubber) are unlikely to
2316 * use this area of the disk again soon, so don't cache
2319 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2324 p_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2327 SetPageUptodate(p_page);
2329 if (q_stripe != -1) {
2330 q_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2332 __free_page(p_page);
2335 SetPageUptodate(q_page);
2338 atomic_set(&rbio->error, 0);
2340 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2343 /* first collect one page from each data stripe */
2344 for (stripe = 0; stripe < nr_data; stripe++) {
2345 p = page_in_rbio(rbio, stripe, pagenr, 0);
2346 pointers[stripe] = kmap(p);
2349 /* then add the parity stripe */
2350 pointers[stripe++] = kmap(p_page);
2352 if (q_stripe != -1) {
2355 * raid6, add the qstripe and call the
2356 * library function to fill in our p/q
2358 pointers[stripe++] = kmap(q_page);
2360 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
2364 memcpy(pointers[nr_data], pointers[0], PAGE_SIZE);
2365 run_xor(pointers + 1, nr_data - 1, PAGE_SIZE);
2368 /* Check scrubbing parity and repair it */
2369 p = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2371 if (memcmp(parity, pointers[rbio->scrubp], PAGE_SIZE))
2372 memcpy(parity, pointers[rbio->scrubp], PAGE_SIZE);
2374 /* Parity is right, needn't writeback */
2375 bitmap_clear(rbio->dbitmap, pagenr, 1);
2378 for (stripe = 0; stripe < rbio->real_stripes; stripe++)
2379 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
2382 __free_page(p_page);
2384 __free_page(q_page);
2388 * time to start writing. Make bios for everything from the
2389 * higher layers (the bio_list in our rbio) and our p/q. Ignore
2392 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2395 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2396 ret = rbio_add_io_page(rbio, &bio_list,
2397 page, rbio->scrubp, pagenr, rbio->stripe_len);
2405 for_each_set_bit(pagenr, pbitmap, rbio->stripe_npages) {
2408 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2409 ret = rbio_add_io_page(rbio, &bio_list, page,
2410 bbio->tgtdev_map[rbio->scrubp],
2411 pagenr, rbio->stripe_len);
2417 nr_data = bio_list_size(&bio_list);
2419 /* Every parity is right */
2420 rbio_orig_end_io(rbio, 0);
2424 atomic_set(&rbio->stripes_pending, nr_data);
2427 bio = bio_list_pop(&bio_list);
2431 bio->bi_private = rbio;
2432 bio->bi_end_io = raid_write_end_io;
2433 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2440 rbio_orig_end_io(rbio, -EIO);
2443 static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe)
2445 if (stripe >= 0 && stripe < rbio->nr_data)
2451 * While we're doing the parity check and repair, we could have errors
2452 * in reading pages off the disk. This checks for errors and if we're
2453 * not able to read the page it'll trigger parity reconstruction. The
2454 * parity scrub will be finished after we've reconstructed the failed
2457 static void validate_rbio_for_parity_scrub(struct btrfs_raid_bio *rbio)
2459 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
2462 if (rbio->faila >= 0 || rbio->failb >= 0) {
2463 int dfail = 0, failp = -1;
2465 if (is_data_stripe(rbio, rbio->faila))
2467 else if (is_parity_stripe(rbio->faila))
2468 failp = rbio->faila;
2470 if (is_data_stripe(rbio, rbio->failb))
2472 else if (is_parity_stripe(rbio->failb))
2473 failp = rbio->failb;
2476 * Because we can not use a scrubbing parity to repair
2477 * the data, so the capability of the repair is declined.
2478 * (In the case of RAID5, we can not repair anything)
2480 if (dfail > rbio->bbio->max_errors - 1)
2484 * If all data is good, only parity is correctly, just
2485 * repair the parity.
2488 finish_parity_scrub(rbio, 0);
2493 * Here means we got one corrupted data stripe and one
2494 * corrupted parity on RAID6, if the corrupted parity
2495 * is scrubbing parity, luckily, use the other one to repair
2496 * the data, or we can not repair the data stripe.
2498 if (failp != rbio->scrubp)
2501 __raid_recover_end_io(rbio);
2503 finish_parity_scrub(rbio, 1);
2508 rbio_orig_end_io(rbio, -EIO);
2512 * end io for the read phase of the rmw cycle. All the bios here are physical
2513 * stripe bios we've read from the disk so we can recalculate the parity of the
2516 * This will usually kick off finish_rmw once all the bios are read in, but it
2517 * may trigger parity reconstruction if we had any errors along the way
2519 static void raid56_parity_scrub_end_io(struct bio *bio)
2521 struct btrfs_raid_bio *rbio = bio->bi_private;
2524 fail_bio_stripe(rbio, bio);
2526 set_bio_pages_uptodate(bio);
2530 if (!atomic_dec_and_test(&rbio->stripes_pending))
2534 * this will normally call finish_rmw to start our write
2535 * but if there are any failed stripes we'll reconstruct
2538 validate_rbio_for_parity_scrub(rbio);
2541 static void raid56_parity_scrub_stripe(struct btrfs_raid_bio *rbio)
2543 int bios_to_read = 0;
2544 struct bio_list bio_list;
2550 ret = alloc_rbio_essential_pages(rbio);
2554 bio_list_init(&bio_list);
2556 atomic_set(&rbio->error, 0);
2558 * build a list of bios to read all the missing parts of this
2561 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
2562 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2565 * we want to find all the pages missing from
2566 * the rbio and read them from the disk. If
2567 * page_in_rbio finds a page in the bio list
2568 * we don't need to read it off the stripe.
2570 page = page_in_rbio(rbio, stripe, pagenr, 1);
2574 page = rbio_stripe_page(rbio, stripe, pagenr);
2576 * the bio cache may have handed us an uptodate
2577 * page. If so, be happy and use it
2579 if (PageUptodate(page))
2582 ret = rbio_add_io_page(rbio, &bio_list, page,
2583 stripe, pagenr, rbio->stripe_len);
2589 bios_to_read = bio_list_size(&bio_list);
2590 if (!bios_to_read) {
2592 * this can happen if others have merged with
2593 * us, it means there is nothing left to read.
2594 * But if there are missing devices it may not be
2595 * safe to do the full stripe write yet.
2601 * the bbio may be freed once we submit the last bio. Make sure
2602 * not to touch it after that
2604 atomic_set(&rbio->stripes_pending, bios_to_read);
2606 bio = bio_list_pop(&bio_list);
2610 bio->bi_private = rbio;
2611 bio->bi_end_io = raid56_parity_scrub_end_io;
2612 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2614 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
2618 /* the actual write will happen once the reads are done */
2622 rbio_orig_end_io(rbio, -EIO);
2626 validate_rbio_for_parity_scrub(rbio);
2629 static void scrub_parity_work(struct btrfs_work *work)
2631 struct btrfs_raid_bio *rbio;
2633 rbio = container_of(work, struct btrfs_raid_bio, work);
2634 raid56_parity_scrub_stripe(rbio);
2637 static void async_scrub_parity(struct btrfs_raid_bio *rbio)
2639 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
2640 scrub_parity_work, NULL, NULL);
2642 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
2645 void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio)
2647 if (!lock_stripe_add(rbio))
2648 async_scrub_parity(rbio);
2651 /* The following code is used for dev replace of a missing RAID 5/6 device. */
2653 struct btrfs_raid_bio *
2654 raid56_alloc_missing_rbio(struct btrfs_fs_info *fs_info, struct bio *bio,
2655 struct btrfs_bio *bbio, u64 length)
2657 struct btrfs_raid_bio *rbio;
2659 rbio = alloc_rbio(fs_info, bbio, length);
2663 rbio->operation = BTRFS_RBIO_REBUILD_MISSING;
2664 bio_list_add(&rbio->bio_list, bio);
2666 * This is a special bio which is used to hold the completion handler
2667 * and make the scrub rbio is similar to the other types
2669 ASSERT(!bio->bi_iter.bi_size);
2671 rbio->faila = find_logical_bio_stripe(rbio, bio);
2672 if (rbio->faila == -1) {
2681 static void missing_raid56_work(struct btrfs_work *work)
2683 struct btrfs_raid_bio *rbio;
2685 rbio = container_of(work, struct btrfs_raid_bio, work);
2686 __raid56_parity_recover(rbio);
2689 static void async_missing_raid56(struct btrfs_raid_bio *rbio)
2691 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
2692 missing_raid56_work, NULL, NULL);
2694 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
2697 void raid56_submit_missing_rbio(struct btrfs_raid_bio *rbio)
2699 if (!lock_stripe_add(rbio))
2700 async_missing_raid56(rbio);