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>
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 = kvzalloc(table_size, GFP_KERNEL);
225 spin_lock_init(&table->cache_lock);
226 INIT_LIST_HEAD(&table->stripe_cache);
230 for (i = 0; i < num_entries; i++) {
232 INIT_LIST_HEAD(&cur->hash_list);
233 spin_lock_init(&cur->lock);
236 x = cmpxchg(&info->stripe_hash_table, NULL, table);
243 * caching an rbio means to copy anything from the
244 * bio_pages array into the stripe_pages array. We
245 * use the page uptodate bit in the stripe cache array
246 * to indicate if it has valid data
248 * once the caching is done, we set the cache ready
251 static void cache_rbio_pages(struct btrfs_raid_bio *rbio)
258 ret = alloc_rbio_pages(rbio);
262 for (i = 0; i < rbio->nr_pages; i++) {
263 if (!rbio->bio_pages[i])
266 s = kmap(rbio->bio_pages[i]);
267 d = kmap(rbio->stripe_pages[i]);
269 memcpy(d, s, PAGE_SIZE);
271 kunmap(rbio->bio_pages[i]);
272 kunmap(rbio->stripe_pages[i]);
273 SetPageUptodate(rbio->stripe_pages[i]);
275 set_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
279 * we hash on the first logical address of the stripe
281 static int rbio_bucket(struct btrfs_raid_bio *rbio)
283 u64 num = rbio->bbio->raid_map[0];
286 * we shift down quite a bit. We're using byte
287 * addressing, and most of the lower bits are zeros.
288 * This tends to upset hash_64, and it consistently
289 * returns just one or two different values.
291 * shifting off the lower bits fixes things.
293 return hash_64(num >> 16, BTRFS_STRIPE_HASH_TABLE_BITS);
297 * stealing an rbio means taking all the uptodate pages from the stripe
298 * array in the source rbio and putting them into the destination rbio
300 static void steal_rbio(struct btrfs_raid_bio *src, struct btrfs_raid_bio *dest)
306 if (!test_bit(RBIO_CACHE_READY_BIT, &src->flags))
309 for (i = 0; i < dest->nr_pages; i++) {
310 s = src->stripe_pages[i];
311 if (!s || !PageUptodate(s)) {
315 d = dest->stripe_pages[i];
319 dest->stripe_pages[i] = s;
320 src->stripe_pages[i] = NULL;
325 * merging means we take the bio_list from the victim and
326 * splice it into the destination. The victim should
327 * be discarded afterwards.
329 * must be called with dest->rbio_list_lock held
331 static void merge_rbio(struct btrfs_raid_bio *dest,
332 struct btrfs_raid_bio *victim)
334 bio_list_merge(&dest->bio_list, &victim->bio_list);
335 dest->bio_list_bytes += victim->bio_list_bytes;
336 dest->generic_bio_cnt += victim->generic_bio_cnt;
337 bio_list_init(&victim->bio_list);
341 * used to prune items that are in the cache. The caller
342 * must hold the hash table lock.
344 static void __remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
346 int bucket = rbio_bucket(rbio);
347 struct btrfs_stripe_hash_table *table;
348 struct btrfs_stripe_hash *h;
352 * check the bit again under the hash table lock.
354 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
357 table = rbio->fs_info->stripe_hash_table;
358 h = table->table + bucket;
360 /* hold the lock for the bucket because we may be
361 * removing it from the hash table
366 * hold the lock for the bio list because we need
367 * to make sure the bio list is empty
369 spin_lock(&rbio->bio_list_lock);
371 if (test_and_clear_bit(RBIO_CACHE_BIT, &rbio->flags)) {
372 list_del_init(&rbio->stripe_cache);
373 table->cache_size -= 1;
376 /* if the bio list isn't empty, this rbio is
377 * still involved in an IO. We take it out
378 * of the cache list, and drop the ref that
379 * was held for the list.
381 * If the bio_list was empty, we also remove
382 * the rbio from the hash_table, and drop
383 * the corresponding ref
385 if (bio_list_empty(&rbio->bio_list)) {
386 if (!list_empty(&rbio->hash_list)) {
387 list_del_init(&rbio->hash_list);
388 refcount_dec(&rbio->refs);
389 BUG_ON(!list_empty(&rbio->plug_list));
394 spin_unlock(&rbio->bio_list_lock);
395 spin_unlock(&h->lock);
398 __free_raid_bio(rbio);
402 * prune a given rbio from the cache
404 static void remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
406 struct btrfs_stripe_hash_table *table;
409 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
412 table = rbio->fs_info->stripe_hash_table;
414 spin_lock_irqsave(&table->cache_lock, flags);
415 __remove_rbio_from_cache(rbio);
416 spin_unlock_irqrestore(&table->cache_lock, flags);
420 * remove everything in the cache
422 static void btrfs_clear_rbio_cache(struct btrfs_fs_info *info)
424 struct btrfs_stripe_hash_table *table;
426 struct btrfs_raid_bio *rbio;
428 table = info->stripe_hash_table;
430 spin_lock_irqsave(&table->cache_lock, flags);
431 while (!list_empty(&table->stripe_cache)) {
432 rbio = list_entry(table->stripe_cache.next,
433 struct btrfs_raid_bio,
435 __remove_rbio_from_cache(rbio);
437 spin_unlock_irqrestore(&table->cache_lock, flags);
441 * remove all cached entries and free the hash table
444 void btrfs_free_stripe_hash_table(struct btrfs_fs_info *info)
446 if (!info->stripe_hash_table)
448 btrfs_clear_rbio_cache(info);
449 kvfree(info->stripe_hash_table);
450 info->stripe_hash_table = NULL;
454 * insert an rbio into the stripe cache. It
455 * must have already been prepared by calling
458 * If this rbio was already cached, it gets
459 * moved to the front of the lru.
461 * If the size of the rbio cache is too big, we
464 static void cache_rbio(struct btrfs_raid_bio *rbio)
466 struct btrfs_stripe_hash_table *table;
469 if (!test_bit(RBIO_CACHE_READY_BIT, &rbio->flags))
472 table = rbio->fs_info->stripe_hash_table;
474 spin_lock_irqsave(&table->cache_lock, flags);
475 spin_lock(&rbio->bio_list_lock);
477 /* bump our ref if we were not in the list before */
478 if (!test_and_set_bit(RBIO_CACHE_BIT, &rbio->flags))
479 refcount_inc(&rbio->refs);
481 if (!list_empty(&rbio->stripe_cache)){
482 list_move(&rbio->stripe_cache, &table->stripe_cache);
484 list_add(&rbio->stripe_cache, &table->stripe_cache);
485 table->cache_size += 1;
488 spin_unlock(&rbio->bio_list_lock);
490 if (table->cache_size > RBIO_CACHE_SIZE) {
491 struct btrfs_raid_bio *found;
493 found = list_entry(table->stripe_cache.prev,
494 struct btrfs_raid_bio,
498 __remove_rbio_from_cache(found);
501 spin_unlock_irqrestore(&table->cache_lock, flags);
505 * helper function to run the xor_blocks api. It is only
506 * able to do MAX_XOR_BLOCKS at a time, so we need to
509 static void run_xor(void **pages, int src_cnt, ssize_t len)
513 void *dest = pages[src_cnt];
516 xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS);
517 xor_blocks(xor_src_cnt, len, dest, pages + src_off);
519 src_cnt -= xor_src_cnt;
520 src_off += xor_src_cnt;
525 * returns true if the bio list inside this rbio
526 * covers an entire stripe (no rmw required).
527 * Must be called with the bio list lock held, or
528 * at a time when you know it is impossible to add
529 * new bios into the list
531 static int __rbio_is_full(struct btrfs_raid_bio *rbio)
533 unsigned long size = rbio->bio_list_bytes;
536 if (size != rbio->nr_data * rbio->stripe_len)
539 BUG_ON(size > rbio->nr_data * rbio->stripe_len);
543 static int rbio_is_full(struct btrfs_raid_bio *rbio)
548 spin_lock_irqsave(&rbio->bio_list_lock, flags);
549 ret = __rbio_is_full(rbio);
550 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
555 * returns 1 if it is safe to merge two rbios together.
556 * The merging is safe if the two rbios correspond to
557 * the same stripe and if they are both going in the same
558 * direction (read vs write), and if neither one is
559 * locked for final IO
561 * The caller is responsible for locking such that
562 * rmw_locked is safe to test
564 static int rbio_can_merge(struct btrfs_raid_bio *last,
565 struct btrfs_raid_bio *cur)
567 if (test_bit(RBIO_RMW_LOCKED_BIT, &last->flags) ||
568 test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags))
572 * we can't merge with cached rbios, since the
573 * idea is that when we merge the destination
574 * rbio is going to run our IO for us. We can
575 * steal from cached rbios though, other functions
578 if (test_bit(RBIO_CACHE_BIT, &last->flags) ||
579 test_bit(RBIO_CACHE_BIT, &cur->flags))
582 if (last->bbio->raid_map[0] !=
583 cur->bbio->raid_map[0])
586 /* we can't merge with different operations */
587 if (last->operation != cur->operation)
590 * We've need read the full stripe from the drive.
591 * check and repair the parity and write the new results.
593 * We're not allowed to add any new bios to the
594 * bio list here, anyone else that wants to
595 * change this stripe needs to do their own rmw.
597 if (last->operation == BTRFS_RBIO_PARITY_SCRUB)
600 if (last->operation == BTRFS_RBIO_REBUILD_MISSING)
603 if (last->operation == BTRFS_RBIO_READ_REBUILD) {
604 int fa = last->faila;
605 int fb = last->failb;
606 int cur_fa = cur->faila;
607 int cur_fb = cur->failb;
609 if (last->faila >= last->failb) {
614 if (cur->faila >= cur->failb) {
619 if (fa != cur_fa || fb != cur_fb)
625 static int rbio_stripe_page_index(struct btrfs_raid_bio *rbio, int stripe,
628 return stripe * rbio->stripe_npages + index;
632 * these are just the pages from the rbio array, not from anything
633 * the FS sent down to us
635 static struct page *rbio_stripe_page(struct btrfs_raid_bio *rbio, int stripe,
638 return rbio->stripe_pages[rbio_stripe_page_index(rbio, stripe, index)];
642 * helper to index into the pstripe
644 static struct page *rbio_pstripe_page(struct btrfs_raid_bio *rbio, int index)
646 return rbio_stripe_page(rbio, rbio->nr_data, index);
650 * helper to index into the qstripe, returns null
651 * if there is no qstripe
653 static struct page *rbio_qstripe_page(struct btrfs_raid_bio *rbio, int index)
655 if (rbio->nr_data + 1 == rbio->real_stripes)
657 return rbio_stripe_page(rbio, rbio->nr_data + 1, index);
661 * The first stripe in the table for a logical address
662 * has the lock. rbios are added in one of three ways:
664 * 1) Nobody has the stripe locked yet. The rbio is given
665 * the lock and 0 is returned. The caller must start the IO
668 * 2) Someone has the stripe locked, but we're able to merge
669 * with the lock owner. The rbio is freed and the IO will
670 * start automatically along with the existing rbio. 1 is returned.
672 * 3) Someone has the stripe locked, but we're not able to merge.
673 * The rbio is added to the lock owner's plug list, or merged into
674 * an rbio already on the plug list. When the lock owner unlocks,
675 * the next rbio on the list is run and the IO is started automatically.
678 * If we return 0, the caller still owns the rbio and must continue with
679 * IO submission. If we return 1, the caller must assume the rbio has
680 * already been freed.
682 static noinline int lock_stripe_add(struct btrfs_raid_bio *rbio)
684 int bucket = rbio_bucket(rbio);
685 struct btrfs_stripe_hash *h = rbio->fs_info->stripe_hash_table->table + bucket;
686 struct btrfs_raid_bio *cur;
687 struct btrfs_raid_bio *pending;
689 struct btrfs_raid_bio *freeit = NULL;
690 struct btrfs_raid_bio *cache_drop = NULL;
693 spin_lock_irqsave(&h->lock, flags);
694 list_for_each_entry(cur, &h->hash_list, hash_list) {
695 if (cur->bbio->raid_map[0] == rbio->bbio->raid_map[0]) {
696 spin_lock(&cur->bio_list_lock);
698 /* can we steal this cached rbio's pages? */
699 if (bio_list_empty(&cur->bio_list) &&
700 list_empty(&cur->plug_list) &&
701 test_bit(RBIO_CACHE_BIT, &cur->flags) &&
702 !test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) {
703 list_del_init(&cur->hash_list);
704 refcount_dec(&cur->refs);
706 steal_rbio(cur, rbio);
708 spin_unlock(&cur->bio_list_lock);
713 /* can we merge into the lock owner? */
714 if (rbio_can_merge(cur, rbio)) {
715 merge_rbio(cur, rbio);
716 spin_unlock(&cur->bio_list_lock);
724 * we couldn't merge with the running
725 * rbio, see if we can merge with the
726 * pending ones. We don't have to
727 * check for rmw_locked because there
728 * is no way they are inside finish_rmw
731 list_for_each_entry(pending, &cur->plug_list,
733 if (rbio_can_merge(pending, rbio)) {
734 merge_rbio(pending, rbio);
735 spin_unlock(&cur->bio_list_lock);
742 /* no merging, put us on the tail of the plug list,
743 * our rbio will be started with the currently
744 * running rbio unlocks
746 list_add_tail(&rbio->plug_list, &cur->plug_list);
747 spin_unlock(&cur->bio_list_lock);
753 refcount_inc(&rbio->refs);
754 list_add(&rbio->hash_list, &h->hash_list);
756 spin_unlock_irqrestore(&h->lock, flags);
758 remove_rbio_from_cache(cache_drop);
760 __free_raid_bio(freeit);
765 * called as rmw or parity rebuild is completed. If the plug list has more
766 * rbios waiting for this stripe, the next one on the list will be started
768 static noinline void unlock_stripe(struct btrfs_raid_bio *rbio)
771 struct btrfs_stripe_hash *h;
775 bucket = rbio_bucket(rbio);
776 h = rbio->fs_info->stripe_hash_table->table + bucket;
778 if (list_empty(&rbio->plug_list))
781 spin_lock_irqsave(&h->lock, flags);
782 spin_lock(&rbio->bio_list_lock);
784 if (!list_empty(&rbio->hash_list)) {
786 * if we're still cached and there is no other IO
787 * to perform, just leave this rbio here for others
788 * to steal from later
790 if (list_empty(&rbio->plug_list) &&
791 test_bit(RBIO_CACHE_BIT, &rbio->flags)) {
793 clear_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
794 BUG_ON(!bio_list_empty(&rbio->bio_list));
798 list_del_init(&rbio->hash_list);
799 refcount_dec(&rbio->refs);
802 * we use the plug list to hold all the rbios
803 * waiting for the chance to lock this stripe.
804 * hand the lock over to one of them.
806 if (!list_empty(&rbio->plug_list)) {
807 struct btrfs_raid_bio *next;
808 struct list_head *head = rbio->plug_list.next;
810 next = list_entry(head, struct btrfs_raid_bio,
813 list_del_init(&rbio->plug_list);
815 list_add(&next->hash_list, &h->hash_list);
816 refcount_inc(&next->refs);
817 spin_unlock(&rbio->bio_list_lock);
818 spin_unlock_irqrestore(&h->lock, flags);
820 if (next->operation == BTRFS_RBIO_READ_REBUILD)
821 async_read_rebuild(next);
822 else if (next->operation == BTRFS_RBIO_REBUILD_MISSING) {
823 steal_rbio(rbio, next);
824 async_read_rebuild(next);
825 } else if (next->operation == BTRFS_RBIO_WRITE) {
826 steal_rbio(rbio, next);
827 async_rmw_stripe(next);
828 } else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) {
829 steal_rbio(rbio, next);
830 async_scrub_parity(next);
837 spin_unlock(&rbio->bio_list_lock);
838 spin_unlock_irqrestore(&h->lock, flags);
842 remove_rbio_from_cache(rbio);
845 static void __free_raid_bio(struct btrfs_raid_bio *rbio)
849 if (!refcount_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, 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);
890 cur->bi_status = err;
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 blk_status_t err = bio->bi_status;
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 refcount_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_disk == stripe->dev->bdev->bd_disk &&
1100 last->bi_partno == stripe->dev->bdev->bd_partno) {
1101 ret = bio_add_page(last, page, PAGE_SIZE, 0);
1102 if (ret == PAGE_SIZE)
1107 /* put a new bio on the list */
1108 bio = btrfs_io_bio_alloc(bio_max_len >> PAGE_SHIFT ?: 1);
1109 bio->bi_iter.bi_size = 0;
1110 bio_set_dev(bio, stripe->dev->bdev);
1111 bio->bi_iter.bi_sector = disk_start >> 9;
1113 bio_add_page(bio, page, PAGE_SIZE, 0);
1114 bio_list_add(bio_list, bio);
1119 * while we're doing the read/modify/write cycle, we could
1120 * have errors in reading pages off the disk. This checks
1121 * for errors and if we're not able to read the page it'll
1122 * trigger parity reconstruction. The rmw will be finished
1123 * after we've reconstructed the failed stripes
1125 static void validate_rbio_for_rmw(struct btrfs_raid_bio *rbio)
1127 if (rbio->faila >= 0 || rbio->failb >= 0) {
1128 BUG_ON(rbio->faila == rbio->real_stripes - 1);
1129 __raid56_parity_recover(rbio);
1136 * helper function to walk our bio list and populate the bio_pages array with
1137 * the result. This seems expensive, but it is faster than constantly
1138 * searching through the bio list as we setup the IO in finish_rmw or stripe
1141 * This must be called before you trust the answers from page_in_rbio
1143 static void index_rbio_pages(struct btrfs_raid_bio *rbio)
1147 unsigned long stripe_offset;
1148 unsigned long page_index;
1150 spin_lock_irq(&rbio->bio_list_lock);
1151 bio_list_for_each(bio, &rbio->bio_list) {
1152 struct bio_vec bvec;
1153 struct bvec_iter iter;
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 if (bio_flagged(bio, BIO_CLONED))
1161 bio->bi_iter = btrfs_io_bio(bio)->iter;
1163 bio_for_each_segment(bvec, bio, iter) {
1164 rbio->bio_pages[page_index + i] = bvec.bv_page;
1168 spin_unlock_irq(&rbio->bio_list_lock);
1172 * this is called from one of two situations. We either
1173 * have a full stripe from the higher layers, or we've read all
1174 * the missing bits off disk.
1176 * This will calculate the parity and then send down any
1179 static noinline void finish_rmw(struct btrfs_raid_bio *rbio)
1181 struct btrfs_bio *bbio = rbio->bbio;
1182 void *pointers[rbio->real_stripes];
1183 int nr_data = rbio->nr_data;
1188 struct bio_list bio_list;
1192 bio_list_init(&bio_list);
1194 if (rbio->real_stripes - rbio->nr_data == 1) {
1195 p_stripe = rbio->real_stripes - 1;
1196 } else if (rbio->real_stripes - rbio->nr_data == 2) {
1197 p_stripe = rbio->real_stripes - 2;
1198 q_stripe = rbio->real_stripes - 1;
1203 /* at this point we either have a full stripe,
1204 * or we've read the full stripe from the drive.
1205 * recalculate the parity and write the new results.
1207 * We're not allowed to add any new bios to the
1208 * bio list here, anyone else that wants to
1209 * change this stripe needs to do their own rmw.
1211 spin_lock_irq(&rbio->bio_list_lock);
1212 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1213 spin_unlock_irq(&rbio->bio_list_lock);
1215 atomic_set(&rbio->error, 0);
1218 * now that we've set rmw_locked, run through the
1219 * bio list one last time and map the page pointers
1221 * We don't cache full rbios because we're assuming
1222 * the higher layers are unlikely to use this area of
1223 * the disk again soon. If they do use it again,
1224 * hopefully they will send another full bio.
1226 index_rbio_pages(rbio);
1227 if (!rbio_is_full(rbio))
1228 cache_rbio_pages(rbio);
1230 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
1232 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1234 /* first collect one page from each data stripe */
1235 for (stripe = 0; stripe < nr_data; stripe++) {
1236 p = page_in_rbio(rbio, stripe, pagenr, 0);
1237 pointers[stripe] = kmap(p);
1240 /* then add the parity stripe */
1241 p = rbio_pstripe_page(rbio, pagenr);
1243 pointers[stripe++] = kmap(p);
1245 if (q_stripe != -1) {
1248 * raid6, add the qstripe and call the
1249 * library function to fill in our p/q
1251 p = rbio_qstripe_page(rbio, pagenr);
1253 pointers[stripe++] = kmap(p);
1255 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
1259 memcpy(pointers[nr_data], pointers[0], PAGE_SIZE);
1260 run_xor(pointers + 1, nr_data - 1, PAGE_SIZE);
1264 for (stripe = 0; stripe < rbio->real_stripes; stripe++)
1265 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
1269 * time to start writing. Make bios for everything from the
1270 * higher layers (the bio_list in our rbio) and our p/q. Ignore
1273 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1274 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1276 if (stripe < rbio->nr_data) {
1277 page = page_in_rbio(rbio, stripe, pagenr, 1);
1281 page = rbio_stripe_page(rbio, stripe, pagenr);
1284 ret = rbio_add_io_page(rbio, &bio_list,
1285 page, stripe, pagenr, rbio->stripe_len);
1291 if (likely(!bbio->num_tgtdevs))
1294 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1295 if (!bbio->tgtdev_map[stripe])
1298 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1300 if (stripe < rbio->nr_data) {
1301 page = page_in_rbio(rbio, stripe, pagenr, 1);
1305 page = rbio_stripe_page(rbio, stripe, pagenr);
1308 ret = rbio_add_io_page(rbio, &bio_list, page,
1309 rbio->bbio->tgtdev_map[stripe],
1310 pagenr, rbio->stripe_len);
1317 atomic_set(&rbio->stripes_pending, bio_list_size(&bio_list));
1318 BUG_ON(atomic_read(&rbio->stripes_pending) == 0);
1321 bio = bio_list_pop(&bio_list);
1325 bio->bi_private = rbio;
1326 bio->bi_end_io = raid_write_end_io;
1327 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1334 rbio_orig_end_io(rbio, BLK_STS_IOERR);
1336 while ((bio = bio_list_pop(&bio_list)))
1341 * helper to find the stripe number for a given bio. Used to figure out which
1342 * stripe has failed. This expects the bio to correspond to a physical disk,
1343 * so it looks up based on physical sector numbers.
1345 static int find_bio_stripe(struct btrfs_raid_bio *rbio,
1348 u64 physical = bio->bi_iter.bi_sector;
1351 struct btrfs_bio_stripe *stripe;
1355 for (i = 0; i < rbio->bbio->num_stripes; i++) {
1356 stripe = &rbio->bbio->stripes[i];
1357 stripe_start = stripe->physical;
1358 if (physical >= stripe_start &&
1359 physical < stripe_start + rbio->stripe_len &&
1360 bio->bi_disk == stripe->dev->bdev->bd_disk &&
1361 bio->bi_partno == stripe->dev->bdev->bd_partno) {
1369 * helper to find the stripe number for a given
1370 * bio (before mapping). Used to figure out which stripe has
1371 * failed. This looks up based on logical block numbers.
1373 static int find_logical_bio_stripe(struct btrfs_raid_bio *rbio,
1376 u64 logical = bio->bi_iter.bi_sector;
1382 for (i = 0; i < rbio->nr_data; i++) {
1383 stripe_start = rbio->bbio->raid_map[i];
1384 if (logical >= stripe_start &&
1385 logical < stripe_start + rbio->stripe_len) {
1393 * returns -EIO if we had too many failures
1395 static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed)
1397 unsigned long flags;
1400 spin_lock_irqsave(&rbio->bio_list_lock, flags);
1402 /* we already know this stripe is bad, move on */
1403 if (rbio->faila == failed || rbio->failb == failed)
1406 if (rbio->faila == -1) {
1407 /* first failure on this rbio */
1408 rbio->faila = failed;
1409 atomic_inc(&rbio->error);
1410 } else if (rbio->failb == -1) {
1411 /* second failure on this rbio */
1412 rbio->failb = failed;
1413 atomic_inc(&rbio->error);
1418 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
1424 * helper to fail a stripe based on a physical disk
1427 static int fail_bio_stripe(struct btrfs_raid_bio *rbio,
1430 int failed = find_bio_stripe(rbio, bio);
1435 return fail_rbio_index(rbio, failed);
1439 * this sets each page in the bio uptodate. It should only be used on private
1440 * rbio pages, nothing that comes in from the higher layers
1442 static void set_bio_pages_uptodate(struct bio *bio)
1444 struct bio_vec bvec;
1445 struct bvec_iter iter;
1447 if (bio_flagged(bio, BIO_CLONED))
1448 bio->bi_iter = btrfs_io_bio(bio)->iter;
1450 bio_for_each_segment(bvec, bio, iter)
1451 SetPageUptodate(bvec.bv_page);
1455 * end io for the read phase of the rmw cycle. All the bios here are physical
1456 * stripe bios we've read from the disk so we can recalculate the parity of the
1459 * This will usually kick off finish_rmw once all the bios are read in, but it
1460 * may trigger parity reconstruction if we had any errors along the way
1462 static void raid_rmw_end_io(struct bio *bio)
1464 struct btrfs_raid_bio *rbio = bio->bi_private;
1467 fail_bio_stripe(rbio, bio);
1469 set_bio_pages_uptodate(bio);
1473 if (!atomic_dec_and_test(&rbio->stripes_pending))
1476 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
1480 * this will normally call finish_rmw to start our write
1481 * but if there are any failed stripes we'll reconstruct
1484 validate_rbio_for_rmw(rbio);
1489 rbio_orig_end_io(rbio, BLK_STS_IOERR);
1492 static void async_rmw_stripe(struct btrfs_raid_bio *rbio)
1494 btrfs_init_work(&rbio->work, btrfs_rmw_helper, rmw_work, NULL, NULL);
1495 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
1498 static void async_read_rebuild(struct btrfs_raid_bio *rbio)
1500 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
1501 read_rebuild_work, NULL, NULL);
1503 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
1507 * the stripe must be locked by the caller. It will
1508 * unlock after all the writes are done
1510 static int raid56_rmw_stripe(struct btrfs_raid_bio *rbio)
1512 int bios_to_read = 0;
1513 struct bio_list bio_list;
1519 bio_list_init(&bio_list);
1521 ret = alloc_rbio_pages(rbio);
1525 index_rbio_pages(rbio);
1527 atomic_set(&rbio->error, 0);
1529 * build a list of bios to read all the missing parts of this
1532 for (stripe = 0; stripe < rbio->nr_data; stripe++) {
1533 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1536 * we want to find all the pages missing from
1537 * the rbio and read them from the disk. If
1538 * page_in_rbio finds a page in the bio list
1539 * we don't need to read it off the stripe.
1541 page = page_in_rbio(rbio, stripe, pagenr, 1);
1545 page = rbio_stripe_page(rbio, stripe, pagenr);
1547 * the bio cache may have handed us an uptodate
1548 * page. If so, be happy and use it
1550 if (PageUptodate(page))
1553 ret = rbio_add_io_page(rbio, &bio_list, page,
1554 stripe, pagenr, rbio->stripe_len);
1560 bios_to_read = bio_list_size(&bio_list);
1561 if (!bios_to_read) {
1563 * this can happen if others have merged with
1564 * us, it means there is nothing left to read.
1565 * But if there are missing devices it may not be
1566 * safe to do the full stripe write yet.
1572 * the bbio may be freed once we submit the last bio. Make sure
1573 * not to touch it after that
1575 atomic_set(&rbio->stripes_pending, bios_to_read);
1577 bio = bio_list_pop(&bio_list);
1581 bio->bi_private = rbio;
1582 bio->bi_end_io = raid_rmw_end_io;
1583 bio_set_op_attrs(bio, REQ_OP_READ, 0);
1585 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
1589 /* the actual write will happen once the reads are done */
1593 rbio_orig_end_io(rbio, BLK_STS_IOERR);
1595 while ((bio = bio_list_pop(&bio_list)))
1601 validate_rbio_for_rmw(rbio);
1606 * if the upper layers pass in a full stripe, we thank them by only allocating
1607 * enough pages to hold the parity, and sending it all down quickly.
1609 static int full_stripe_write(struct btrfs_raid_bio *rbio)
1613 ret = alloc_rbio_parity_pages(rbio);
1615 __free_raid_bio(rbio);
1619 ret = lock_stripe_add(rbio);
1626 * partial stripe writes get handed over to async helpers.
1627 * We're really hoping to merge a few more writes into this
1628 * rbio before calculating new parity
1630 static int partial_stripe_write(struct btrfs_raid_bio *rbio)
1634 ret = lock_stripe_add(rbio);
1636 async_rmw_stripe(rbio);
1641 * sometimes while we were reading from the drive to
1642 * recalculate parity, enough new bios come into create
1643 * a full stripe. So we do a check here to see if we can
1644 * go directly to finish_rmw
1646 static int __raid56_parity_write(struct btrfs_raid_bio *rbio)
1648 /* head off into rmw land if we don't have a full stripe */
1649 if (!rbio_is_full(rbio))
1650 return partial_stripe_write(rbio);
1651 return full_stripe_write(rbio);
1655 * We use plugging call backs to collect full stripes.
1656 * Any time we get a partial stripe write while plugged
1657 * we collect it into a list. When the unplug comes down,
1658 * we sort the list by logical block number and merge
1659 * everything we can into the same rbios
1661 struct btrfs_plug_cb {
1662 struct blk_plug_cb cb;
1663 struct btrfs_fs_info *info;
1664 struct list_head rbio_list;
1665 struct btrfs_work work;
1669 * rbios on the plug list are sorted for easier merging.
1671 static int plug_cmp(void *priv, struct list_head *a, struct list_head *b)
1673 struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio,
1675 struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio,
1677 u64 a_sector = ra->bio_list.head->bi_iter.bi_sector;
1678 u64 b_sector = rb->bio_list.head->bi_iter.bi_sector;
1680 if (a_sector < b_sector)
1682 if (a_sector > b_sector)
1687 static void run_plug(struct btrfs_plug_cb *plug)
1689 struct btrfs_raid_bio *cur;
1690 struct btrfs_raid_bio *last = NULL;
1693 * sort our plug list then try to merge
1694 * everything we can in hopes of creating full
1697 list_sort(NULL, &plug->rbio_list, plug_cmp);
1698 while (!list_empty(&plug->rbio_list)) {
1699 cur = list_entry(plug->rbio_list.next,
1700 struct btrfs_raid_bio, plug_list);
1701 list_del_init(&cur->plug_list);
1703 if (rbio_is_full(cur)) {
1704 /* we have a full stripe, send it down */
1705 full_stripe_write(cur);
1709 if (rbio_can_merge(last, cur)) {
1710 merge_rbio(last, cur);
1711 __free_raid_bio(cur);
1715 __raid56_parity_write(last);
1720 __raid56_parity_write(last);
1726 * if the unplug comes from schedule, we have to push the
1727 * work off to a helper thread
1729 static void unplug_work(struct btrfs_work *work)
1731 struct btrfs_plug_cb *plug;
1732 plug = container_of(work, struct btrfs_plug_cb, work);
1736 static void btrfs_raid_unplug(struct blk_plug_cb *cb, bool from_schedule)
1738 struct btrfs_plug_cb *plug;
1739 plug = container_of(cb, struct btrfs_plug_cb, cb);
1741 if (from_schedule) {
1742 btrfs_init_work(&plug->work, btrfs_rmw_helper,
1743 unplug_work, NULL, NULL);
1744 btrfs_queue_work(plug->info->rmw_workers,
1752 * our main entry point for writes from the rest of the FS.
1754 int raid56_parity_write(struct btrfs_fs_info *fs_info, struct bio *bio,
1755 struct btrfs_bio *bbio, u64 stripe_len)
1757 struct btrfs_raid_bio *rbio;
1758 struct btrfs_plug_cb *plug = NULL;
1759 struct blk_plug_cb *cb;
1762 rbio = alloc_rbio(fs_info, bbio, stripe_len);
1764 btrfs_put_bbio(bbio);
1765 return PTR_ERR(rbio);
1767 bio_list_add(&rbio->bio_list, bio);
1768 rbio->bio_list_bytes = bio->bi_iter.bi_size;
1769 rbio->operation = BTRFS_RBIO_WRITE;
1771 btrfs_bio_counter_inc_noblocked(fs_info);
1772 rbio->generic_bio_cnt = 1;
1775 * don't plug on full rbios, just get them out the door
1776 * as quickly as we can
1778 if (rbio_is_full(rbio)) {
1779 ret = full_stripe_write(rbio);
1781 btrfs_bio_counter_dec(fs_info);
1785 cb = blk_check_plugged(btrfs_raid_unplug, fs_info, sizeof(*plug));
1787 plug = container_of(cb, struct btrfs_plug_cb, cb);
1789 plug->info = fs_info;
1790 INIT_LIST_HEAD(&plug->rbio_list);
1792 list_add_tail(&rbio->plug_list, &plug->rbio_list);
1795 ret = __raid56_parity_write(rbio);
1797 btrfs_bio_counter_dec(fs_info);
1803 * all parity reconstruction happens here. We've read in everything
1804 * we can find from the drives and this does the heavy lifting of
1805 * sorting the good from the bad.
1807 static void __raid_recover_end_io(struct btrfs_raid_bio *rbio)
1811 int faila = -1, failb = -1;
1816 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
1818 err = BLK_STS_RESOURCE;
1822 faila = rbio->faila;
1823 failb = rbio->failb;
1825 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1826 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
1827 spin_lock_irq(&rbio->bio_list_lock);
1828 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1829 spin_unlock_irq(&rbio->bio_list_lock);
1832 index_rbio_pages(rbio);
1834 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1836 * Now we just use bitmap to mark the horizontal stripes in
1837 * which we have data when doing parity scrub.
1839 if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB &&
1840 !test_bit(pagenr, rbio->dbitmap))
1843 /* setup our array of pointers with pages
1846 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1848 * if we're rebuilding a read, we have to use
1849 * pages from the bio list
1851 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1852 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
1853 (stripe == faila || stripe == failb)) {
1854 page = page_in_rbio(rbio, stripe, pagenr, 0);
1856 page = rbio_stripe_page(rbio, stripe, pagenr);
1858 pointers[stripe] = kmap(page);
1861 /* all raid6 handling here */
1862 if (rbio->bbio->map_type & BTRFS_BLOCK_GROUP_RAID6) {
1864 * single failure, rebuild from parity raid5
1868 if (faila == rbio->nr_data) {
1870 * Just the P stripe has failed, without
1871 * a bad data or Q stripe.
1872 * TODO, we should redo the xor here.
1874 err = BLK_STS_IOERR;
1878 * a single failure in raid6 is rebuilt
1879 * in the pstripe code below
1884 /* make sure our ps and qs are in order */
1885 if (faila > failb) {
1891 /* if the q stripe is failed, do a pstripe reconstruction
1893 * If both the q stripe and the P stripe are failed, we're
1894 * here due to a crc mismatch and we can't give them the
1897 if (rbio->bbio->raid_map[failb] == RAID6_Q_STRIPE) {
1898 if (rbio->bbio->raid_map[faila] ==
1900 err = BLK_STS_IOERR;
1904 * otherwise we have one bad data stripe and
1905 * a good P stripe. raid5!
1910 if (rbio->bbio->raid_map[failb] == RAID5_P_STRIPE) {
1911 raid6_datap_recov(rbio->real_stripes,
1912 PAGE_SIZE, faila, pointers);
1914 raid6_2data_recov(rbio->real_stripes,
1915 PAGE_SIZE, faila, failb,
1921 /* rebuild from P stripe here (raid5 or raid6) */
1922 BUG_ON(failb != -1);
1924 /* Copy parity block into failed block to start with */
1925 memcpy(pointers[faila],
1926 pointers[rbio->nr_data],
1929 /* rearrange the pointer array */
1930 p = pointers[faila];
1931 for (stripe = faila; stripe < rbio->nr_data - 1; stripe++)
1932 pointers[stripe] = pointers[stripe + 1];
1933 pointers[rbio->nr_data - 1] = p;
1935 /* xor in the rest */
1936 run_xor(pointers, rbio->nr_data - 1, PAGE_SIZE);
1938 /* if we're doing this rebuild as part of an rmw, go through
1939 * and set all of our private rbio pages in the
1940 * failed stripes as uptodate. This way finish_rmw will
1941 * know they can be trusted. If this was a read reconstruction,
1942 * other endio functions will fiddle the uptodate bits
1944 if (rbio->operation == BTRFS_RBIO_WRITE) {
1945 for (i = 0; i < rbio->stripe_npages; i++) {
1947 page = rbio_stripe_page(rbio, faila, i);
1948 SetPageUptodate(page);
1951 page = rbio_stripe_page(rbio, failb, i);
1952 SetPageUptodate(page);
1956 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1958 * if we're rebuilding a read, we have to use
1959 * pages from the bio list
1961 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1962 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
1963 (stripe == faila || stripe == failb)) {
1964 page = page_in_rbio(rbio, stripe, pagenr, 0);
1966 page = rbio_stripe_page(rbio, stripe, pagenr);
1977 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
1978 if (err == BLK_STS_OK)
1979 cache_rbio_pages(rbio);
1981 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
1983 rbio_orig_end_io(rbio, err);
1984 } else if (rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
1985 rbio_orig_end_io(rbio, err);
1986 } else if (err == BLK_STS_OK) {
1990 if (rbio->operation == BTRFS_RBIO_WRITE)
1992 else if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB)
1993 finish_parity_scrub(rbio, 0);
1997 rbio_orig_end_io(rbio, err);
2002 * This is called only for stripes we've read from disk to
2003 * reconstruct the parity.
2005 static void raid_recover_end_io(struct bio *bio)
2007 struct btrfs_raid_bio *rbio = bio->bi_private;
2010 * we only read stripe pages off the disk, set them
2011 * up to date if there were no errors
2014 fail_bio_stripe(rbio, bio);
2016 set_bio_pages_uptodate(bio);
2019 if (!atomic_dec_and_test(&rbio->stripes_pending))
2022 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
2023 rbio_orig_end_io(rbio, BLK_STS_IOERR);
2025 __raid_recover_end_io(rbio);
2029 * reads everything we need off the disk to reconstruct
2030 * the parity. endio handlers trigger final reconstruction
2031 * when the IO is done.
2033 * This is used both for reads from the higher layers and for
2034 * parity construction required to finish a rmw cycle.
2036 static int __raid56_parity_recover(struct btrfs_raid_bio *rbio)
2038 int bios_to_read = 0;
2039 struct bio_list bio_list;
2045 bio_list_init(&bio_list);
2047 ret = alloc_rbio_pages(rbio);
2051 atomic_set(&rbio->error, 0);
2054 * read everything that hasn't failed. Thanks to the
2055 * stripe cache, it is possible that some or all of these
2056 * pages are going to be uptodate.
2058 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
2059 if (rbio->faila == stripe || rbio->failb == stripe) {
2060 atomic_inc(&rbio->error);
2064 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
2068 * the rmw code may have already read this
2071 p = rbio_stripe_page(rbio, stripe, pagenr);
2072 if (PageUptodate(p))
2075 ret = rbio_add_io_page(rbio, &bio_list,
2076 rbio_stripe_page(rbio, stripe, pagenr),
2077 stripe, pagenr, rbio->stripe_len);
2083 bios_to_read = bio_list_size(&bio_list);
2084 if (!bios_to_read) {
2086 * we might have no bios to read just because the pages
2087 * were up to date, or we might have no bios to read because
2088 * the devices were gone.
2090 if (atomic_read(&rbio->error) <= rbio->bbio->max_errors) {
2091 __raid_recover_end_io(rbio);
2099 * the bbio may be freed once we submit the last bio. Make sure
2100 * not to touch it after that
2102 atomic_set(&rbio->stripes_pending, bios_to_read);
2104 bio = bio_list_pop(&bio_list);
2108 bio->bi_private = rbio;
2109 bio->bi_end_io = raid_recover_end_io;
2110 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2112 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
2120 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
2121 rbio->operation == BTRFS_RBIO_REBUILD_MISSING)
2122 rbio_orig_end_io(rbio, BLK_STS_IOERR);
2124 while ((bio = bio_list_pop(&bio_list)))
2131 * the main entry point for reads from the higher layers. This
2132 * is really only called when the normal read path had a failure,
2133 * so we assume the bio they send down corresponds to a failed part
2136 int raid56_parity_recover(struct btrfs_fs_info *fs_info, struct bio *bio,
2137 struct btrfs_bio *bbio, u64 stripe_len,
2138 int mirror_num, int generic_io)
2140 struct btrfs_raid_bio *rbio;
2144 ASSERT(bbio->mirror_num == mirror_num);
2145 btrfs_io_bio(bio)->mirror_num = mirror_num;
2148 rbio = alloc_rbio(fs_info, bbio, stripe_len);
2151 btrfs_put_bbio(bbio);
2152 return PTR_ERR(rbio);
2155 rbio->operation = BTRFS_RBIO_READ_REBUILD;
2156 bio_list_add(&rbio->bio_list, bio);
2157 rbio->bio_list_bytes = bio->bi_iter.bi_size;
2159 rbio->faila = find_logical_bio_stripe(rbio, bio);
2160 if (rbio->faila == -1) {
2162 "%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)",
2163 __func__, (u64)bio->bi_iter.bi_sector << 9,
2164 (u64)bio->bi_iter.bi_size, bbio->map_type);
2166 btrfs_put_bbio(bbio);
2172 btrfs_bio_counter_inc_noblocked(fs_info);
2173 rbio->generic_bio_cnt = 1;
2175 btrfs_get_bbio(bbio);
2180 * for 'mirror == 2', reconstruct from all other stripes.
2181 * for 'mirror_num > 2', select a stripe to fail on every retry.
2183 if (mirror_num > 2) {
2185 * 'mirror == 3' is to fail the p stripe and
2186 * reconstruct from the q stripe. 'mirror > 3' is to
2187 * fail a data stripe and reconstruct from p+q stripe.
2189 rbio->failb = rbio->real_stripes - (mirror_num - 1);
2190 ASSERT(rbio->failb > 0);
2191 if (rbio->failb <= rbio->faila)
2195 ret = lock_stripe_add(rbio);
2198 * __raid56_parity_recover will end the bio with
2199 * any errors it hits. We don't want to return
2200 * its error value up the stack because our caller
2201 * will end up calling bio_endio with any nonzero
2205 __raid56_parity_recover(rbio);
2207 * our rbio has been added to the list of
2208 * rbios that will be handled after the
2209 * currently lock owner is done
2215 static void rmw_work(struct btrfs_work *work)
2217 struct btrfs_raid_bio *rbio;
2219 rbio = container_of(work, struct btrfs_raid_bio, work);
2220 raid56_rmw_stripe(rbio);
2223 static void read_rebuild_work(struct btrfs_work *work)
2225 struct btrfs_raid_bio *rbio;
2227 rbio = container_of(work, struct btrfs_raid_bio, work);
2228 __raid56_parity_recover(rbio);
2232 * The following code is used to scrub/replace the parity stripe
2234 * Caller must have already increased bio_counter for getting @bbio.
2236 * Note: We need make sure all the pages that add into the scrub/replace
2237 * raid bio are correct and not be changed during the scrub/replace. That
2238 * is those pages just hold metadata or file data with checksum.
2241 struct btrfs_raid_bio *
2242 raid56_parity_alloc_scrub_rbio(struct btrfs_fs_info *fs_info, struct bio *bio,
2243 struct btrfs_bio *bbio, u64 stripe_len,
2244 struct btrfs_device *scrub_dev,
2245 unsigned long *dbitmap, int stripe_nsectors)
2247 struct btrfs_raid_bio *rbio;
2250 rbio = alloc_rbio(fs_info, bbio, stripe_len);
2253 bio_list_add(&rbio->bio_list, bio);
2255 * This is a special bio which is used to hold the completion handler
2256 * and make the scrub rbio is similar to the other types
2258 ASSERT(!bio->bi_iter.bi_size);
2259 rbio->operation = BTRFS_RBIO_PARITY_SCRUB;
2262 * After mapping bbio with BTRFS_MAP_WRITE, parities have been sorted
2263 * to the end position, so this search can start from the first parity
2266 for (i = rbio->nr_data; i < rbio->real_stripes; i++) {
2267 if (bbio->stripes[i].dev == scrub_dev) {
2272 ASSERT(i < rbio->real_stripes);
2274 /* Now we just support the sectorsize equals to page size */
2275 ASSERT(fs_info->sectorsize == PAGE_SIZE);
2276 ASSERT(rbio->stripe_npages == stripe_nsectors);
2277 bitmap_copy(rbio->dbitmap, dbitmap, stripe_nsectors);
2280 * We have already increased bio_counter when getting bbio, record it
2281 * so we can free it at rbio_orig_end_io().
2283 rbio->generic_bio_cnt = 1;
2288 /* Used for both parity scrub and missing. */
2289 void raid56_add_scrub_pages(struct btrfs_raid_bio *rbio, struct page *page,
2295 ASSERT(logical >= rbio->bbio->raid_map[0]);
2296 ASSERT(logical + PAGE_SIZE <= rbio->bbio->raid_map[0] +
2297 rbio->stripe_len * rbio->nr_data);
2298 stripe_offset = (int)(logical - rbio->bbio->raid_map[0]);
2299 index = stripe_offset >> PAGE_SHIFT;
2300 rbio->bio_pages[index] = page;
2304 * We just scrub the parity that we have correct data on the same horizontal,
2305 * so we needn't allocate all pages for all the stripes.
2307 static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio)
2314 for_each_set_bit(bit, rbio->dbitmap, rbio->stripe_npages) {
2315 for (i = 0; i < rbio->real_stripes; i++) {
2316 index = i * rbio->stripe_npages + bit;
2317 if (rbio->stripe_pages[index])
2320 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2323 rbio->stripe_pages[index] = page;
2329 static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
2332 struct btrfs_bio *bbio = rbio->bbio;
2333 void *pointers[rbio->real_stripes];
2334 DECLARE_BITMAP(pbitmap, rbio->stripe_npages);
2335 int nr_data = rbio->nr_data;
2340 struct page *p_page = NULL;
2341 struct page *q_page = NULL;
2342 struct bio_list bio_list;
2347 bio_list_init(&bio_list);
2349 if (rbio->real_stripes - rbio->nr_data == 1) {
2350 p_stripe = rbio->real_stripes - 1;
2351 } else if (rbio->real_stripes - rbio->nr_data == 2) {
2352 p_stripe = rbio->real_stripes - 2;
2353 q_stripe = rbio->real_stripes - 1;
2358 if (bbio->num_tgtdevs && bbio->tgtdev_map[rbio->scrubp]) {
2360 bitmap_copy(pbitmap, rbio->dbitmap, rbio->stripe_npages);
2364 * Because the higher layers(scrubber) are unlikely to
2365 * use this area of the disk again soon, so don't cache
2368 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2373 p_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2376 SetPageUptodate(p_page);
2378 if (q_stripe != -1) {
2379 q_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2381 __free_page(p_page);
2384 SetPageUptodate(q_page);
2387 atomic_set(&rbio->error, 0);
2389 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2392 /* first collect one page from each data stripe */
2393 for (stripe = 0; stripe < nr_data; stripe++) {
2394 p = page_in_rbio(rbio, stripe, pagenr, 0);
2395 pointers[stripe] = kmap(p);
2398 /* then add the parity stripe */
2399 pointers[stripe++] = kmap(p_page);
2401 if (q_stripe != -1) {
2404 * raid6, add the qstripe and call the
2405 * library function to fill in our p/q
2407 pointers[stripe++] = kmap(q_page);
2409 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
2413 memcpy(pointers[nr_data], pointers[0], PAGE_SIZE);
2414 run_xor(pointers + 1, nr_data - 1, PAGE_SIZE);
2417 /* Check scrubbing parity and repair it */
2418 p = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2420 if (memcmp(parity, pointers[rbio->scrubp], PAGE_SIZE))
2421 memcpy(parity, pointers[rbio->scrubp], PAGE_SIZE);
2423 /* Parity is right, needn't writeback */
2424 bitmap_clear(rbio->dbitmap, pagenr, 1);
2427 for (stripe = 0; stripe < rbio->real_stripes; stripe++)
2428 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
2431 __free_page(p_page);
2433 __free_page(q_page);
2437 * time to start writing. Make bios for everything from the
2438 * higher layers (the bio_list in our rbio) and our p/q. Ignore
2441 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2444 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2445 ret = rbio_add_io_page(rbio, &bio_list,
2446 page, rbio->scrubp, pagenr, rbio->stripe_len);
2454 for_each_set_bit(pagenr, pbitmap, rbio->stripe_npages) {
2457 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2458 ret = rbio_add_io_page(rbio, &bio_list, page,
2459 bbio->tgtdev_map[rbio->scrubp],
2460 pagenr, rbio->stripe_len);
2466 nr_data = bio_list_size(&bio_list);
2468 /* Every parity is right */
2469 rbio_orig_end_io(rbio, BLK_STS_OK);
2473 atomic_set(&rbio->stripes_pending, nr_data);
2476 bio = bio_list_pop(&bio_list);
2480 bio->bi_private = rbio;
2481 bio->bi_end_io = raid_write_end_io;
2482 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2489 rbio_orig_end_io(rbio, BLK_STS_IOERR);
2491 while ((bio = bio_list_pop(&bio_list)))
2495 static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe)
2497 if (stripe >= 0 && stripe < rbio->nr_data)
2503 * While we're doing the parity check and repair, we could have errors
2504 * in reading pages off the disk. This checks for errors and if we're
2505 * not able to read the page it'll trigger parity reconstruction. The
2506 * parity scrub will be finished after we've reconstructed the failed
2509 static void validate_rbio_for_parity_scrub(struct btrfs_raid_bio *rbio)
2511 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
2514 if (rbio->faila >= 0 || rbio->failb >= 0) {
2515 int dfail = 0, failp = -1;
2517 if (is_data_stripe(rbio, rbio->faila))
2519 else if (is_parity_stripe(rbio->faila))
2520 failp = rbio->faila;
2522 if (is_data_stripe(rbio, rbio->failb))
2524 else if (is_parity_stripe(rbio->failb))
2525 failp = rbio->failb;
2528 * Because we can not use a scrubbing parity to repair
2529 * the data, so the capability of the repair is declined.
2530 * (In the case of RAID5, we can not repair anything)
2532 if (dfail > rbio->bbio->max_errors - 1)
2536 * If all data is good, only parity is correctly, just
2537 * repair the parity.
2540 finish_parity_scrub(rbio, 0);
2545 * Here means we got one corrupted data stripe and one
2546 * corrupted parity on RAID6, if the corrupted parity
2547 * is scrubbing parity, luckily, use the other one to repair
2548 * the data, or we can not repair the data stripe.
2550 if (failp != rbio->scrubp)
2553 __raid_recover_end_io(rbio);
2555 finish_parity_scrub(rbio, 1);
2560 rbio_orig_end_io(rbio, BLK_STS_IOERR);
2564 * end io for the read phase of the rmw cycle. All the bios here are physical
2565 * stripe bios we've read from the disk so we can recalculate the parity of the
2568 * This will usually kick off finish_rmw once all the bios are read in, but it
2569 * may trigger parity reconstruction if we had any errors along the way
2571 static void raid56_parity_scrub_end_io(struct bio *bio)
2573 struct btrfs_raid_bio *rbio = bio->bi_private;
2576 fail_bio_stripe(rbio, bio);
2578 set_bio_pages_uptodate(bio);
2582 if (!atomic_dec_and_test(&rbio->stripes_pending))
2586 * this will normally call finish_rmw to start our write
2587 * but if there are any failed stripes we'll reconstruct
2590 validate_rbio_for_parity_scrub(rbio);
2593 static void raid56_parity_scrub_stripe(struct btrfs_raid_bio *rbio)
2595 int bios_to_read = 0;
2596 struct bio_list bio_list;
2602 bio_list_init(&bio_list);
2604 ret = alloc_rbio_essential_pages(rbio);
2608 atomic_set(&rbio->error, 0);
2610 * build a list of bios to read all the missing parts of this
2613 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
2614 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2617 * we want to find all the pages missing from
2618 * the rbio and read them from the disk. If
2619 * page_in_rbio finds a page in the bio list
2620 * we don't need to read it off the stripe.
2622 page = page_in_rbio(rbio, stripe, pagenr, 1);
2626 page = rbio_stripe_page(rbio, stripe, pagenr);
2628 * the bio cache may have handed us an uptodate
2629 * page. If so, be happy and use it
2631 if (PageUptodate(page))
2634 ret = rbio_add_io_page(rbio, &bio_list, page,
2635 stripe, pagenr, rbio->stripe_len);
2641 bios_to_read = bio_list_size(&bio_list);
2642 if (!bios_to_read) {
2644 * this can happen if others have merged with
2645 * us, it means there is nothing left to read.
2646 * But if there are missing devices it may not be
2647 * safe to do the full stripe write yet.
2653 * the bbio may be freed once we submit the last bio. Make sure
2654 * not to touch it after that
2656 atomic_set(&rbio->stripes_pending, bios_to_read);
2658 bio = bio_list_pop(&bio_list);
2662 bio->bi_private = rbio;
2663 bio->bi_end_io = raid56_parity_scrub_end_io;
2664 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2666 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
2670 /* the actual write will happen once the reads are done */
2674 rbio_orig_end_io(rbio, BLK_STS_IOERR);
2676 while ((bio = bio_list_pop(&bio_list)))
2682 validate_rbio_for_parity_scrub(rbio);
2685 static void scrub_parity_work(struct btrfs_work *work)
2687 struct btrfs_raid_bio *rbio;
2689 rbio = container_of(work, struct btrfs_raid_bio, work);
2690 raid56_parity_scrub_stripe(rbio);
2693 static void async_scrub_parity(struct btrfs_raid_bio *rbio)
2695 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
2696 scrub_parity_work, NULL, NULL);
2698 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
2701 void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio)
2703 if (!lock_stripe_add(rbio))
2704 async_scrub_parity(rbio);
2707 /* The following code is used for dev replace of a missing RAID 5/6 device. */
2709 struct btrfs_raid_bio *
2710 raid56_alloc_missing_rbio(struct btrfs_fs_info *fs_info, struct bio *bio,
2711 struct btrfs_bio *bbio, u64 length)
2713 struct btrfs_raid_bio *rbio;
2715 rbio = alloc_rbio(fs_info, bbio, length);
2719 rbio->operation = BTRFS_RBIO_REBUILD_MISSING;
2720 bio_list_add(&rbio->bio_list, bio);
2722 * This is a special bio which is used to hold the completion handler
2723 * and make the scrub rbio is similar to the other types
2725 ASSERT(!bio->bi_iter.bi_size);
2727 rbio->faila = find_logical_bio_stripe(rbio, bio);
2728 if (rbio->faila == -1) {
2735 * When we get bbio, we have already increased bio_counter, record it
2736 * so we can free it at rbio_orig_end_io()
2738 rbio->generic_bio_cnt = 1;
2743 static void missing_raid56_work(struct btrfs_work *work)
2745 struct btrfs_raid_bio *rbio;
2747 rbio = container_of(work, struct btrfs_raid_bio, work);
2748 __raid56_parity_recover(rbio);
2751 static void async_missing_raid56(struct btrfs_raid_bio *rbio)
2753 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
2754 missing_raid56_work, NULL, NULL);
2756 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
2759 void raid56_submit_missing_rbio(struct btrfs_raid_bio *rbio)
2761 if (!lock_stripe_add(rbio))
2762 async_missing_raid56(rbio);