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);
234 init_waitqueue_head(&cur->wait);
237 x = cmpxchg(&info->stripe_hash_table, NULL, table);
244 * caching an rbio means to copy anything from the
245 * bio_pages array into the stripe_pages array. We
246 * use the page uptodate bit in the stripe cache array
247 * to indicate if it has valid data
249 * once the caching is done, we set the cache ready
252 static void cache_rbio_pages(struct btrfs_raid_bio *rbio)
259 ret = alloc_rbio_pages(rbio);
263 for (i = 0; i < rbio->nr_pages; i++) {
264 if (!rbio->bio_pages[i])
267 s = kmap(rbio->bio_pages[i]);
268 d = kmap(rbio->stripe_pages[i]);
270 memcpy(d, s, PAGE_SIZE);
272 kunmap(rbio->bio_pages[i]);
273 kunmap(rbio->stripe_pages[i]);
274 SetPageUptodate(rbio->stripe_pages[i]);
276 set_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
280 * we hash on the first logical address of the stripe
282 static int rbio_bucket(struct btrfs_raid_bio *rbio)
284 u64 num = rbio->bbio->raid_map[0];
287 * we shift down quite a bit. We're using byte
288 * addressing, and most of the lower bits are zeros.
289 * This tends to upset hash_64, and it consistently
290 * returns just one or two different values.
292 * shifting off the lower bits fixes things.
294 return hash_64(num >> 16, BTRFS_STRIPE_HASH_TABLE_BITS);
298 * stealing an rbio means taking all the uptodate pages from the stripe
299 * array in the source rbio and putting them into the destination rbio
301 static void steal_rbio(struct btrfs_raid_bio *src, struct btrfs_raid_bio *dest)
307 if (!test_bit(RBIO_CACHE_READY_BIT, &src->flags))
310 for (i = 0; i < dest->nr_pages; i++) {
311 s = src->stripe_pages[i];
312 if (!s || !PageUptodate(s)) {
316 d = dest->stripe_pages[i];
320 dest->stripe_pages[i] = s;
321 src->stripe_pages[i] = NULL;
326 * merging means we take the bio_list from the victim and
327 * splice it into the destination. The victim should
328 * be discarded afterwards.
330 * must be called with dest->rbio_list_lock held
332 static void merge_rbio(struct btrfs_raid_bio *dest,
333 struct btrfs_raid_bio *victim)
335 bio_list_merge(&dest->bio_list, &victim->bio_list);
336 dest->bio_list_bytes += victim->bio_list_bytes;
337 dest->generic_bio_cnt += victim->generic_bio_cnt;
338 bio_list_init(&victim->bio_list);
342 * used to prune items that are in the cache. The caller
343 * must hold the hash table lock.
345 static void __remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
347 int bucket = rbio_bucket(rbio);
348 struct btrfs_stripe_hash_table *table;
349 struct btrfs_stripe_hash *h;
353 * check the bit again under the hash table lock.
355 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
358 table = rbio->fs_info->stripe_hash_table;
359 h = table->table + bucket;
361 /* hold the lock for the bucket because we may be
362 * removing it from the hash table
367 * hold the lock for the bio list because we need
368 * to make sure the bio list is empty
370 spin_lock(&rbio->bio_list_lock);
372 if (test_and_clear_bit(RBIO_CACHE_BIT, &rbio->flags)) {
373 list_del_init(&rbio->stripe_cache);
374 table->cache_size -= 1;
377 /* if the bio list isn't empty, this rbio is
378 * still involved in an IO. We take it out
379 * of the cache list, and drop the ref that
380 * was held for the list.
382 * If the bio_list was empty, we also remove
383 * the rbio from the hash_table, and drop
384 * the corresponding ref
386 if (bio_list_empty(&rbio->bio_list)) {
387 if (!list_empty(&rbio->hash_list)) {
388 list_del_init(&rbio->hash_list);
389 refcount_dec(&rbio->refs);
390 BUG_ON(!list_empty(&rbio->plug_list));
395 spin_unlock(&rbio->bio_list_lock);
396 spin_unlock(&h->lock);
399 __free_raid_bio(rbio);
403 * prune a given rbio from the cache
405 static void remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
407 struct btrfs_stripe_hash_table *table;
410 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
413 table = rbio->fs_info->stripe_hash_table;
415 spin_lock_irqsave(&table->cache_lock, flags);
416 __remove_rbio_from_cache(rbio);
417 spin_unlock_irqrestore(&table->cache_lock, flags);
421 * remove everything in the cache
423 static void btrfs_clear_rbio_cache(struct btrfs_fs_info *info)
425 struct btrfs_stripe_hash_table *table;
427 struct btrfs_raid_bio *rbio;
429 table = info->stripe_hash_table;
431 spin_lock_irqsave(&table->cache_lock, flags);
432 while (!list_empty(&table->stripe_cache)) {
433 rbio = list_entry(table->stripe_cache.next,
434 struct btrfs_raid_bio,
436 __remove_rbio_from_cache(rbio);
438 spin_unlock_irqrestore(&table->cache_lock, flags);
442 * remove all cached entries and free the hash table
445 void btrfs_free_stripe_hash_table(struct btrfs_fs_info *info)
447 if (!info->stripe_hash_table)
449 btrfs_clear_rbio_cache(info);
450 kvfree(info->stripe_hash_table);
451 info->stripe_hash_table = NULL;
455 * insert an rbio into the stripe cache. It
456 * must have already been prepared by calling
459 * If this rbio was already cached, it gets
460 * moved to the front of the lru.
462 * If the size of the rbio cache is too big, we
465 static void cache_rbio(struct btrfs_raid_bio *rbio)
467 struct btrfs_stripe_hash_table *table;
470 if (!test_bit(RBIO_CACHE_READY_BIT, &rbio->flags))
473 table = rbio->fs_info->stripe_hash_table;
475 spin_lock_irqsave(&table->cache_lock, flags);
476 spin_lock(&rbio->bio_list_lock);
478 /* bump our ref if we were not in the list before */
479 if (!test_and_set_bit(RBIO_CACHE_BIT, &rbio->flags))
480 refcount_inc(&rbio->refs);
482 if (!list_empty(&rbio->stripe_cache)){
483 list_move(&rbio->stripe_cache, &table->stripe_cache);
485 list_add(&rbio->stripe_cache, &table->stripe_cache);
486 table->cache_size += 1;
489 spin_unlock(&rbio->bio_list_lock);
491 if (table->cache_size > RBIO_CACHE_SIZE) {
492 struct btrfs_raid_bio *found;
494 found = list_entry(table->stripe_cache.prev,
495 struct btrfs_raid_bio,
499 __remove_rbio_from_cache(found);
502 spin_unlock_irqrestore(&table->cache_lock, flags);
506 * helper function to run the xor_blocks api. It is only
507 * able to do MAX_XOR_BLOCKS at a time, so we need to
510 static void run_xor(void **pages, int src_cnt, ssize_t len)
514 void *dest = pages[src_cnt];
517 xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS);
518 xor_blocks(xor_src_cnt, len, dest, pages + src_off);
520 src_cnt -= xor_src_cnt;
521 src_off += xor_src_cnt;
526 * returns true if the bio list inside this rbio
527 * covers an entire stripe (no rmw required).
528 * Must be called with the bio list lock held, or
529 * at a time when you know it is impossible to add
530 * new bios into the list
532 static int __rbio_is_full(struct btrfs_raid_bio *rbio)
534 unsigned long size = rbio->bio_list_bytes;
537 if (size != rbio->nr_data * rbio->stripe_len)
540 BUG_ON(size > rbio->nr_data * rbio->stripe_len);
544 static int rbio_is_full(struct btrfs_raid_bio *rbio)
549 spin_lock_irqsave(&rbio->bio_list_lock, flags);
550 ret = __rbio_is_full(rbio);
551 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
556 * returns 1 if it is safe to merge two rbios together.
557 * The merging is safe if the two rbios correspond to
558 * the same stripe and if they are both going in the same
559 * direction (read vs write), and if neither one is
560 * locked for final IO
562 * The caller is responsible for locking such that
563 * rmw_locked is safe to test
565 static int rbio_can_merge(struct btrfs_raid_bio *last,
566 struct btrfs_raid_bio *cur)
568 if (test_bit(RBIO_RMW_LOCKED_BIT, &last->flags) ||
569 test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags))
573 * we can't merge with cached rbios, since the
574 * idea is that when we merge the destination
575 * rbio is going to run our IO for us. We can
576 * steal from cached rbios though, other functions
579 if (test_bit(RBIO_CACHE_BIT, &last->flags) ||
580 test_bit(RBIO_CACHE_BIT, &cur->flags))
583 if (last->bbio->raid_map[0] !=
584 cur->bbio->raid_map[0])
587 /* we can't merge with different operations */
588 if (last->operation != cur->operation)
591 * We've need read the full stripe from the drive.
592 * check and repair the parity and write the new results.
594 * We're not allowed to add any new bios to the
595 * bio list here, anyone else that wants to
596 * change this stripe needs to do their own rmw.
598 if (last->operation == BTRFS_RBIO_PARITY_SCRUB ||
599 cur->operation == BTRFS_RBIO_PARITY_SCRUB)
602 if (last->operation == BTRFS_RBIO_REBUILD_MISSING ||
603 cur->operation == BTRFS_RBIO_REBUILD_MISSING)
609 static int rbio_stripe_page_index(struct btrfs_raid_bio *rbio, int stripe,
612 return stripe * rbio->stripe_npages + index;
616 * these are just the pages from the rbio array, not from anything
617 * the FS sent down to us
619 static struct page *rbio_stripe_page(struct btrfs_raid_bio *rbio, int stripe,
622 return rbio->stripe_pages[rbio_stripe_page_index(rbio, stripe, index)];
626 * helper to index into the pstripe
628 static struct page *rbio_pstripe_page(struct btrfs_raid_bio *rbio, int index)
630 return rbio_stripe_page(rbio, rbio->nr_data, index);
634 * helper to index into the qstripe, returns null
635 * if there is no qstripe
637 static struct page *rbio_qstripe_page(struct btrfs_raid_bio *rbio, int index)
639 if (rbio->nr_data + 1 == rbio->real_stripes)
641 return rbio_stripe_page(rbio, rbio->nr_data + 1, index);
645 * The first stripe in the table for a logical address
646 * has the lock. rbios are added in one of three ways:
648 * 1) Nobody has the stripe locked yet. The rbio is given
649 * the lock and 0 is returned. The caller must start the IO
652 * 2) Someone has the stripe locked, but we're able to merge
653 * with the lock owner. The rbio is freed and the IO will
654 * start automatically along with the existing rbio. 1 is returned.
656 * 3) Someone has the stripe locked, but we're not able to merge.
657 * The rbio is added to the lock owner's plug list, or merged into
658 * an rbio already on the plug list. When the lock owner unlocks,
659 * the next rbio on the list is run and the IO is started automatically.
662 * If we return 0, the caller still owns the rbio and must continue with
663 * IO submission. If we return 1, the caller must assume the rbio has
664 * already been freed.
666 static noinline int lock_stripe_add(struct btrfs_raid_bio *rbio)
668 int bucket = rbio_bucket(rbio);
669 struct btrfs_stripe_hash *h = rbio->fs_info->stripe_hash_table->table + bucket;
670 struct btrfs_raid_bio *cur;
671 struct btrfs_raid_bio *pending;
674 struct btrfs_raid_bio *freeit = NULL;
675 struct btrfs_raid_bio *cache_drop = NULL;
678 spin_lock_irqsave(&h->lock, flags);
679 list_for_each_entry(cur, &h->hash_list, hash_list) {
680 if (cur->bbio->raid_map[0] == rbio->bbio->raid_map[0]) {
681 spin_lock(&cur->bio_list_lock);
683 /* can we steal this cached rbio's pages? */
684 if (bio_list_empty(&cur->bio_list) &&
685 list_empty(&cur->plug_list) &&
686 test_bit(RBIO_CACHE_BIT, &cur->flags) &&
687 !test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) {
688 list_del_init(&cur->hash_list);
689 refcount_dec(&cur->refs);
691 steal_rbio(cur, rbio);
693 spin_unlock(&cur->bio_list_lock);
698 /* can we merge into the lock owner? */
699 if (rbio_can_merge(cur, rbio)) {
700 merge_rbio(cur, rbio);
701 spin_unlock(&cur->bio_list_lock);
709 * we couldn't merge with the running
710 * rbio, see if we can merge with the
711 * pending ones. We don't have to
712 * check for rmw_locked because there
713 * is no way they are inside finish_rmw
716 list_for_each_entry(pending, &cur->plug_list,
718 if (rbio_can_merge(pending, rbio)) {
719 merge_rbio(pending, rbio);
720 spin_unlock(&cur->bio_list_lock);
727 /* no merging, put us on the tail of the plug list,
728 * our rbio will be started with the currently
729 * running rbio unlocks
731 list_add_tail(&rbio->plug_list, &cur->plug_list);
732 spin_unlock(&cur->bio_list_lock);
738 refcount_inc(&rbio->refs);
739 list_add(&rbio->hash_list, &h->hash_list);
741 spin_unlock_irqrestore(&h->lock, flags);
743 remove_rbio_from_cache(cache_drop);
745 __free_raid_bio(freeit);
750 * called as rmw or parity rebuild is completed. If the plug list has more
751 * rbios waiting for this stripe, the next one on the list will be started
753 static noinline void unlock_stripe(struct btrfs_raid_bio *rbio)
756 struct btrfs_stripe_hash *h;
760 bucket = rbio_bucket(rbio);
761 h = rbio->fs_info->stripe_hash_table->table + bucket;
763 if (list_empty(&rbio->plug_list))
766 spin_lock_irqsave(&h->lock, flags);
767 spin_lock(&rbio->bio_list_lock);
769 if (!list_empty(&rbio->hash_list)) {
771 * if we're still cached and there is no other IO
772 * to perform, just leave this rbio here for others
773 * to steal from later
775 if (list_empty(&rbio->plug_list) &&
776 test_bit(RBIO_CACHE_BIT, &rbio->flags)) {
778 clear_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
779 BUG_ON(!bio_list_empty(&rbio->bio_list));
783 list_del_init(&rbio->hash_list);
784 refcount_dec(&rbio->refs);
787 * we use the plug list to hold all the rbios
788 * waiting for the chance to lock this stripe.
789 * hand the lock over to one of them.
791 if (!list_empty(&rbio->plug_list)) {
792 struct btrfs_raid_bio *next;
793 struct list_head *head = rbio->plug_list.next;
795 next = list_entry(head, struct btrfs_raid_bio,
798 list_del_init(&rbio->plug_list);
800 list_add(&next->hash_list, &h->hash_list);
801 refcount_inc(&next->refs);
802 spin_unlock(&rbio->bio_list_lock);
803 spin_unlock_irqrestore(&h->lock, flags);
805 if (next->operation == BTRFS_RBIO_READ_REBUILD)
806 async_read_rebuild(next);
807 else if (next->operation == BTRFS_RBIO_REBUILD_MISSING) {
808 steal_rbio(rbio, next);
809 async_read_rebuild(next);
810 } else if (next->operation == BTRFS_RBIO_WRITE) {
811 steal_rbio(rbio, next);
812 async_rmw_stripe(next);
813 } else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) {
814 steal_rbio(rbio, next);
815 async_scrub_parity(next);
820 * The barrier for this waitqueue_active is not needed,
821 * we're protected by h->lock and can't miss a wakeup.
823 } else if (waitqueue_active(&h->wait)) {
824 spin_unlock(&rbio->bio_list_lock);
825 spin_unlock_irqrestore(&h->lock, flags);
831 spin_unlock(&rbio->bio_list_lock);
832 spin_unlock_irqrestore(&h->lock, flags);
836 remove_rbio_from_cache(rbio);
839 static void __free_raid_bio(struct btrfs_raid_bio *rbio)
843 if (!refcount_dec_and_test(&rbio->refs))
846 WARN_ON(!list_empty(&rbio->stripe_cache));
847 WARN_ON(!list_empty(&rbio->hash_list));
848 WARN_ON(!bio_list_empty(&rbio->bio_list));
850 for (i = 0; i < rbio->nr_pages; i++) {
851 if (rbio->stripe_pages[i]) {
852 __free_page(rbio->stripe_pages[i]);
853 rbio->stripe_pages[i] = NULL;
857 btrfs_put_bbio(rbio->bbio);
861 static void free_raid_bio(struct btrfs_raid_bio *rbio)
864 __free_raid_bio(rbio);
868 * this frees the rbio and runs through all the bios in the
869 * bio_list and calls end_io on them
871 static void rbio_orig_end_io(struct btrfs_raid_bio *rbio, blk_status_t err)
873 struct bio *cur = bio_list_get(&rbio->bio_list);
876 if (rbio->generic_bio_cnt)
877 btrfs_bio_counter_sub(rbio->fs_info, rbio->generic_bio_cnt);
884 cur->bi_status = err;
891 * end io function used by finish_rmw. When we finally
892 * get here, we've written a full stripe
894 static void raid_write_end_io(struct bio *bio)
896 struct btrfs_raid_bio *rbio = bio->bi_private;
897 blk_status_t err = bio->bi_status;
901 fail_bio_stripe(rbio, bio);
905 if (!atomic_dec_and_test(&rbio->stripes_pending))
910 /* OK, we have read all the stripes we need to. */
911 max_errors = (rbio->operation == BTRFS_RBIO_PARITY_SCRUB) ?
912 0 : rbio->bbio->max_errors;
913 if (atomic_read(&rbio->error) > max_errors)
916 rbio_orig_end_io(rbio, err);
920 * the read/modify/write code wants to use the original bio for
921 * any pages it included, and then use the rbio for everything
922 * else. This function decides if a given index (stripe number)
923 * and page number in that stripe fall inside the original bio
926 * if you set bio_list_only, you'll get a NULL back for any ranges
927 * that are outside the bio_list
929 * This doesn't take any refs on anything, you get a bare page pointer
930 * and the caller must bump refs as required.
932 * You must call index_rbio_pages once before you can trust
933 * the answers from this function.
935 static struct page *page_in_rbio(struct btrfs_raid_bio *rbio,
936 int index, int pagenr, int bio_list_only)
939 struct page *p = NULL;
941 chunk_page = index * (rbio->stripe_len >> PAGE_SHIFT) + pagenr;
943 spin_lock_irq(&rbio->bio_list_lock);
944 p = rbio->bio_pages[chunk_page];
945 spin_unlock_irq(&rbio->bio_list_lock);
947 if (p || bio_list_only)
950 return rbio->stripe_pages[chunk_page];
954 * number of pages we need for the entire stripe across all the
957 static unsigned long rbio_nr_pages(unsigned long stripe_len, int nr_stripes)
959 return DIV_ROUND_UP(stripe_len, PAGE_SIZE) * nr_stripes;
963 * allocation and initial setup for the btrfs_raid_bio. Not
964 * this does not allocate any pages for rbio->pages.
966 static struct btrfs_raid_bio *alloc_rbio(struct btrfs_fs_info *fs_info,
967 struct btrfs_bio *bbio,
970 struct btrfs_raid_bio *rbio;
972 int real_stripes = bbio->num_stripes - bbio->num_tgtdevs;
973 int num_pages = rbio_nr_pages(stripe_len, real_stripes);
974 int stripe_npages = DIV_ROUND_UP(stripe_len, PAGE_SIZE);
977 rbio = kzalloc(sizeof(*rbio) + num_pages * sizeof(struct page *) * 2 +
978 DIV_ROUND_UP(stripe_npages, BITS_PER_LONG) *
979 sizeof(long), GFP_NOFS);
981 return ERR_PTR(-ENOMEM);
983 bio_list_init(&rbio->bio_list);
984 INIT_LIST_HEAD(&rbio->plug_list);
985 spin_lock_init(&rbio->bio_list_lock);
986 INIT_LIST_HEAD(&rbio->stripe_cache);
987 INIT_LIST_HEAD(&rbio->hash_list);
989 rbio->fs_info = fs_info;
990 rbio->stripe_len = stripe_len;
991 rbio->nr_pages = num_pages;
992 rbio->real_stripes = real_stripes;
993 rbio->stripe_npages = stripe_npages;
996 refcount_set(&rbio->refs, 1);
997 atomic_set(&rbio->error, 0);
998 atomic_set(&rbio->stripes_pending, 0);
1001 * the stripe_pages and bio_pages array point to the extra
1002 * memory we allocated past the end of the rbio
1005 rbio->stripe_pages = p;
1006 rbio->bio_pages = p + sizeof(struct page *) * num_pages;
1007 rbio->dbitmap = p + sizeof(struct page *) * num_pages * 2;
1009 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1010 nr_data = real_stripes - 1;
1011 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1012 nr_data = real_stripes - 2;
1016 rbio->nr_data = nr_data;
1020 /* allocate pages for all the stripes in the bio, including parity */
1021 static int alloc_rbio_pages(struct btrfs_raid_bio *rbio)
1026 for (i = 0; i < rbio->nr_pages; i++) {
1027 if (rbio->stripe_pages[i])
1029 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1032 rbio->stripe_pages[i] = page;
1037 /* only allocate pages for p/q stripes */
1038 static int alloc_rbio_parity_pages(struct btrfs_raid_bio *rbio)
1043 i = rbio_stripe_page_index(rbio, rbio->nr_data, 0);
1045 for (; i < rbio->nr_pages; i++) {
1046 if (rbio->stripe_pages[i])
1048 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1051 rbio->stripe_pages[i] = page;
1057 * add a single page from a specific stripe into our list of bios for IO
1058 * this will try to merge into existing bios if possible, and returns
1059 * zero if all went well.
1061 static int rbio_add_io_page(struct btrfs_raid_bio *rbio,
1062 struct bio_list *bio_list,
1065 unsigned long page_index,
1066 unsigned long bio_max_len)
1068 struct bio *last = bio_list->tail;
1072 struct btrfs_bio_stripe *stripe;
1075 stripe = &rbio->bbio->stripes[stripe_nr];
1076 disk_start = stripe->physical + (page_index << PAGE_SHIFT);
1078 /* if the device is missing, just fail this stripe */
1079 if (!stripe->dev->bdev)
1080 return fail_rbio_index(rbio, stripe_nr);
1082 /* see if we can add this page onto our existing bio */
1084 last_end = (u64)last->bi_iter.bi_sector << 9;
1085 last_end += last->bi_iter.bi_size;
1088 * we can't merge these if they are from different
1089 * devices or if they are not contiguous
1091 if (last_end == disk_start && stripe->dev->bdev &&
1093 last->bi_bdev == stripe->dev->bdev) {
1094 ret = bio_add_page(last, page, PAGE_SIZE, 0);
1095 if (ret == PAGE_SIZE)
1100 /* put a new bio on the list */
1101 bio = btrfs_io_bio_alloc(bio_max_len >> PAGE_SHIFT ?: 1);
1102 bio->bi_iter.bi_size = 0;
1103 bio->bi_bdev = stripe->dev->bdev;
1104 bio->bi_iter.bi_sector = disk_start >> 9;
1106 bio_add_page(bio, page, PAGE_SIZE, 0);
1107 bio_list_add(bio_list, bio);
1112 * while we're doing the read/modify/write cycle, we could
1113 * have errors in reading pages off the disk. This checks
1114 * for errors and if we're not able to read the page it'll
1115 * trigger parity reconstruction. The rmw will be finished
1116 * after we've reconstructed the failed stripes
1118 static void validate_rbio_for_rmw(struct btrfs_raid_bio *rbio)
1120 if (rbio->faila >= 0 || rbio->failb >= 0) {
1121 BUG_ON(rbio->faila == rbio->real_stripes - 1);
1122 __raid56_parity_recover(rbio);
1129 * helper function to walk our bio list and populate the bio_pages array with
1130 * the result. This seems expensive, but it is faster than constantly
1131 * searching through the bio list as we setup the IO in finish_rmw or stripe
1134 * This must be called before you trust the answers from page_in_rbio
1136 static void index_rbio_pages(struct btrfs_raid_bio *rbio)
1140 unsigned long stripe_offset;
1141 unsigned long page_index;
1143 spin_lock_irq(&rbio->bio_list_lock);
1144 bio_list_for_each(bio, &rbio->bio_list) {
1145 struct bio_vec bvec;
1146 struct bvec_iter iter;
1149 start = (u64)bio->bi_iter.bi_sector << 9;
1150 stripe_offset = start - rbio->bbio->raid_map[0];
1151 page_index = stripe_offset >> PAGE_SHIFT;
1153 if (bio_flagged(bio, BIO_CLONED))
1154 bio->bi_iter = btrfs_io_bio(bio)->iter;
1156 bio_for_each_segment(bvec, bio, iter) {
1157 rbio->bio_pages[page_index + i] = bvec.bv_page;
1161 spin_unlock_irq(&rbio->bio_list_lock);
1165 * this is called from one of two situations. We either
1166 * have a full stripe from the higher layers, or we've read all
1167 * the missing bits off disk.
1169 * This will calculate the parity and then send down any
1172 static noinline void finish_rmw(struct btrfs_raid_bio *rbio)
1174 struct btrfs_bio *bbio = rbio->bbio;
1175 void *pointers[rbio->real_stripes];
1176 int nr_data = rbio->nr_data;
1181 struct bio_list bio_list;
1185 bio_list_init(&bio_list);
1187 if (rbio->real_stripes - rbio->nr_data == 1) {
1188 p_stripe = rbio->real_stripes - 1;
1189 } else if (rbio->real_stripes - rbio->nr_data == 2) {
1190 p_stripe = rbio->real_stripes - 2;
1191 q_stripe = rbio->real_stripes - 1;
1196 /* at this point we either have a full stripe,
1197 * or we've read the full stripe from the drive.
1198 * recalculate the parity and write the new results.
1200 * We're not allowed to add any new bios to the
1201 * bio list here, anyone else that wants to
1202 * change this stripe needs to do their own rmw.
1204 spin_lock_irq(&rbio->bio_list_lock);
1205 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1206 spin_unlock_irq(&rbio->bio_list_lock);
1208 atomic_set(&rbio->error, 0);
1211 * now that we've set rmw_locked, run through the
1212 * bio list one last time and map the page pointers
1214 * We don't cache full rbios because we're assuming
1215 * the higher layers are unlikely to use this area of
1216 * the disk again soon. If they do use it again,
1217 * hopefully they will send another full bio.
1219 index_rbio_pages(rbio);
1220 if (!rbio_is_full(rbio))
1221 cache_rbio_pages(rbio);
1223 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
1225 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1227 /* first collect one page from each data stripe */
1228 for (stripe = 0; stripe < nr_data; stripe++) {
1229 p = page_in_rbio(rbio, stripe, pagenr, 0);
1230 pointers[stripe] = kmap(p);
1233 /* then add the parity stripe */
1234 p = rbio_pstripe_page(rbio, pagenr);
1236 pointers[stripe++] = kmap(p);
1238 if (q_stripe != -1) {
1241 * raid6, add the qstripe and call the
1242 * library function to fill in our p/q
1244 p = rbio_qstripe_page(rbio, pagenr);
1246 pointers[stripe++] = kmap(p);
1248 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
1252 memcpy(pointers[nr_data], pointers[0], PAGE_SIZE);
1253 run_xor(pointers + 1, nr_data - 1, PAGE_SIZE);
1257 for (stripe = 0; stripe < rbio->real_stripes; stripe++)
1258 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
1262 * time to start writing. Make bios for everything from the
1263 * higher layers (the bio_list in our rbio) and our p/q. Ignore
1266 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1267 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1269 if (stripe < rbio->nr_data) {
1270 page = page_in_rbio(rbio, stripe, pagenr, 1);
1274 page = rbio_stripe_page(rbio, stripe, pagenr);
1277 ret = rbio_add_io_page(rbio, &bio_list,
1278 page, stripe, pagenr, rbio->stripe_len);
1284 if (likely(!bbio->num_tgtdevs))
1287 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1288 if (!bbio->tgtdev_map[stripe])
1291 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1293 if (stripe < rbio->nr_data) {
1294 page = page_in_rbio(rbio, stripe, pagenr, 1);
1298 page = rbio_stripe_page(rbio, stripe, pagenr);
1301 ret = rbio_add_io_page(rbio, &bio_list, page,
1302 rbio->bbio->tgtdev_map[stripe],
1303 pagenr, rbio->stripe_len);
1310 atomic_set(&rbio->stripes_pending, bio_list_size(&bio_list));
1311 BUG_ON(atomic_read(&rbio->stripes_pending) == 0);
1314 bio = bio_list_pop(&bio_list);
1318 bio->bi_private = rbio;
1319 bio->bi_end_io = raid_write_end_io;
1320 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1327 rbio_orig_end_io(rbio, -EIO);
1331 * helper to find the stripe number for a given bio. Used to figure out which
1332 * stripe has failed. This expects the bio to correspond to a physical disk,
1333 * so it looks up based on physical sector numbers.
1335 static int find_bio_stripe(struct btrfs_raid_bio *rbio,
1338 u64 physical = bio->bi_iter.bi_sector;
1341 struct btrfs_bio_stripe *stripe;
1345 for (i = 0; i < rbio->bbio->num_stripes; i++) {
1346 stripe = &rbio->bbio->stripes[i];
1347 stripe_start = stripe->physical;
1348 if (physical >= stripe_start &&
1349 physical < stripe_start + rbio->stripe_len &&
1350 bio->bi_bdev == stripe->dev->bdev) {
1358 * helper to find the stripe number for a given
1359 * bio (before mapping). Used to figure out which stripe has
1360 * failed. This looks up based on logical block numbers.
1362 static int find_logical_bio_stripe(struct btrfs_raid_bio *rbio,
1365 u64 logical = bio->bi_iter.bi_sector;
1371 for (i = 0; i < rbio->nr_data; i++) {
1372 stripe_start = rbio->bbio->raid_map[i];
1373 if (logical >= stripe_start &&
1374 logical < stripe_start + rbio->stripe_len) {
1382 * returns -EIO if we had too many failures
1384 static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed)
1386 unsigned long flags;
1389 spin_lock_irqsave(&rbio->bio_list_lock, flags);
1391 /* we already know this stripe is bad, move on */
1392 if (rbio->faila == failed || rbio->failb == failed)
1395 if (rbio->faila == -1) {
1396 /* first failure on this rbio */
1397 rbio->faila = failed;
1398 atomic_inc(&rbio->error);
1399 } else if (rbio->failb == -1) {
1400 /* second failure on this rbio */
1401 rbio->failb = failed;
1402 atomic_inc(&rbio->error);
1407 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
1413 * helper to fail a stripe based on a physical disk
1416 static int fail_bio_stripe(struct btrfs_raid_bio *rbio,
1419 int failed = find_bio_stripe(rbio, bio);
1424 return fail_rbio_index(rbio, failed);
1428 * this sets each page in the bio uptodate. It should only be used on private
1429 * rbio pages, nothing that comes in from the higher layers
1431 static void set_bio_pages_uptodate(struct bio *bio)
1433 struct bio_vec bvec;
1434 struct bvec_iter iter;
1436 if (bio_flagged(bio, BIO_CLONED))
1437 bio->bi_iter = btrfs_io_bio(bio)->iter;
1439 bio_for_each_segment(bvec, bio, iter)
1440 SetPageUptodate(bvec.bv_page);
1444 * end io for the read phase of the rmw cycle. All the bios here are physical
1445 * stripe bios we've read from the disk so we can recalculate the parity of the
1448 * This will usually kick off finish_rmw once all the bios are read in, but it
1449 * may trigger parity reconstruction if we had any errors along the way
1451 static void raid_rmw_end_io(struct bio *bio)
1453 struct btrfs_raid_bio *rbio = bio->bi_private;
1456 fail_bio_stripe(rbio, bio);
1458 set_bio_pages_uptodate(bio);
1462 if (!atomic_dec_and_test(&rbio->stripes_pending))
1465 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
1469 * this will normally call finish_rmw to start our write
1470 * but if there are any failed stripes we'll reconstruct
1473 validate_rbio_for_rmw(rbio);
1478 rbio_orig_end_io(rbio, -EIO);
1481 static void async_rmw_stripe(struct btrfs_raid_bio *rbio)
1483 btrfs_init_work(&rbio->work, btrfs_rmw_helper, rmw_work, NULL, NULL);
1484 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
1487 static void async_read_rebuild(struct btrfs_raid_bio *rbio)
1489 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
1490 read_rebuild_work, NULL, NULL);
1492 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
1496 * the stripe must be locked by the caller. It will
1497 * unlock after all the writes are done
1499 static int raid56_rmw_stripe(struct btrfs_raid_bio *rbio)
1501 int bios_to_read = 0;
1502 struct bio_list bio_list;
1508 bio_list_init(&bio_list);
1510 ret = alloc_rbio_pages(rbio);
1514 index_rbio_pages(rbio);
1516 atomic_set(&rbio->error, 0);
1518 * build a list of bios to read all the missing parts of this
1521 for (stripe = 0; stripe < rbio->nr_data; stripe++) {
1522 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1525 * we want to find all the pages missing from
1526 * the rbio and read them from the disk. If
1527 * page_in_rbio finds a page in the bio list
1528 * we don't need to read it off the stripe.
1530 page = page_in_rbio(rbio, stripe, pagenr, 1);
1534 page = rbio_stripe_page(rbio, stripe, pagenr);
1536 * the bio cache may have handed us an uptodate
1537 * page. If so, be happy and use it
1539 if (PageUptodate(page))
1542 ret = rbio_add_io_page(rbio, &bio_list, page,
1543 stripe, pagenr, rbio->stripe_len);
1549 bios_to_read = bio_list_size(&bio_list);
1550 if (!bios_to_read) {
1552 * this can happen if others have merged with
1553 * us, it means there is nothing left to read.
1554 * But if there are missing devices it may not be
1555 * safe to do the full stripe write yet.
1561 * the bbio may be freed once we submit the last bio. Make sure
1562 * not to touch it after that
1564 atomic_set(&rbio->stripes_pending, bios_to_read);
1566 bio = bio_list_pop(&bio_list);
1570 bio->bi_private = rbio;
1571 bio->bi_end_io = raid_rmw_end_io;
1572 bio_set_op_attrs(bio, REQ_OP_READ, 0);
1574 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
1578 /* the actual write will happen once the reads are done */
1582 rbio_orig_end_io(rbio, -EIO);
1586 validate_rbio_for_rmw(rbio);
1591 * if the upper layers pass in a full stripe, we thank them by only allocating
1592 * enough pages to hold the parity, and sending it all down quickly.
1594 static int full_stripe_write(struct btrfs_raid_bio *rbio)
1598 ret = alloc_rbio_parity_pages(rbio);
1600 __free_raid_bio(rbio);
1604 ret = lock_stripe_add(rbio);
1611 * partial stripe writes get handed over to async helpers.
1612 * We're really hoping to merge a few more writes into this
1613 * rbio before calculating new parity
1615 static int partial_stripe_write(struct btrfs_raid_bio *rbio)
1619 ret = lock_stripe_add(rbio);
1621 async_rmw_stripe(rbio);
1626 * sometimes while we were reading from the drive to
1627 * recalculate parity, enough new bios come into create
1628 * a full stripe. So we do a check here to see if we can
1629 * go directly to finish_rmw
1631 static int __raid56_parity_write(struct btrfs_raid_bio *rbio)
1633 /* head off into rmw land if we don't have a full stripe */
1634 if (!rbio_is_full(rbio))
1635 return partial_stripe_write(rbio);
1636 return full_stripe_write(rbio);
1640 * We use plugging call backs to collect full stripes.
1641 * Any time we get a partial stripe write while plugged
1642 * we collect it into a list. When the unplug comes down,
1643 * we sort the list by logical block number and merge
1644 * everything we can into the same rbios
1646 struct btrfs_plug_cb {
1647 struct blk_plug_cb cb;
1648 struct btrfs_fs_info *info;
1649 struct list_head rbio_list;
1650 struct btrfs_work work;
1654 * rbios on the plug list are sorted for easier merging.
1656 static int plug_cmp(void *priv, struct list_head *a, struct list_head *b)
1658 struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio,
1660 struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio,
1662 u64 a_sector = ra->bio_list.head->bi_iter.bi_sector;
1663 u64 b_sector = rb->bio_list.head->bi_iter.bi_sector;
1665 if (a_sector < b_sector)
1667 if (a_sector > b_sector)
1672 static void run_plug(struct btrfs_plug_cb *plug)
1674 struct btrfs_raid_bio *cur;
1675 struct btrfs_raid_bio *last = NULL;
1678 * sort our plug list then try to merge
1679 * everything we can in hopes of creating full
1682 list_sort(NULL, &plug->rbio_list, plug_cmp);
1683 while (!list_empty(&plug->rbio_list)) {
1684 cur = list_entry(plug->rbio_list.next,
1685 struct btrfs_raid_bio, plug_list);
1686 list_del_init(&cur->plug_list);
1688 if (rbio_is_full(cur)) {
1689 /* we have a full stripe, send it down */
1690 full_stripe_write(cur);
1694 if (rbio_can_merge(last, cur)) {
1695 merge_rbio(last, cur);
1696 __free_raid_bio(cur);
1700 __raid56_parity_write(last);
1705 __raid56_parity_write(last);
1711 * if the unplug comes from schedule, we have to push the
1712 * work off to a helper thread
1714 static void unplug_work(struct btrfs_work *work)
1716 struct btrfs_plug_cb *plug;
1717 plug = container_of(work, struct btrfs_plug_cb, work);
1721 static void btrfs_raid_unplug(struct blk_plug_cb *cb, bool from_schedule)
1723 struct btrfs_plug_cb *plug;
1724 plug = container_of(cb, struct btrfs_plug_cb, cb);
1726 if (from_schedule) {
1727 btrfs_init_work(&plug->work, btrfs_rmw_helper,
1728 unplug_work, NULL, NULL);
1729 btrfs_queue_work(plug->info->rmw_workers,
1737 * our main entry point for writes from the rest of the FS.
1739 int raid56_parity_write(struct btrfs_fs_info *fs_info, struct bio *bio,
1740 struct btrfs_bio *bbio, u64 stripe_len)
1742 struct btrfs_raid_bio *rbio;
1743 struct btrfs_plug_cb *plug = NULL;
1744 struct blk_plug_cb *cb;
1747 rbio = alloc_rbio(fs_info, bbio, stripe_len);
1749 btrfs_put_bbio(bbio);
1750 return PTR_ERR(rbio);
1752 bio_list_add(&rbio->bio_list, bio);
1753 rbio->bio_list_bytes = bio->bi_iter.bi_size;
1754 rbio->operation = BTRFS_RBIO_WRITE;
1756 btrfs_bio_counter_inc_noblocked(fs_info);
1757 rbio->generic_bio_cnt = 1;
1760 * don't plug on full rbios, just get them out the door
1761 * as quickly as we can
1763 if (rbio_is_full(rbio)) {
1764 ret = full_stripe_write(rbio);
1766 btrfs_bio_counter_dec(fs_info);
1770 cb = blk_check_plugged(btrfs_raid_unplug, fs_info, sizeof(*plug));
1772 plug = container_of(cb, struct btrfs_plug_cb, cb);
1774 plug->info = fs_info;
1775 INIT_LIST_HEAD(&plug->rbio_list);
1777 list_add_tail(&rbio->plug_list, &plug->rbio_list);
1780 ret = __raid56_parity_write(rbio);
1782 btrfs_bio_counter_dec(fs_info);
1788 * all parity reconstruction happens here. We've read in everything
1789 * we can find from the drives and this does the heavy lifting of
1790 * sorting the good from the bad.
1792 static void __raid_recover_end_io(struct btrfs_raid_bio *rbio)
1796 int faila = -1, failb = -1;
1801 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
1807 faila = rbio->faila;
1808 failb = rbio->failb;
1810 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1811 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
1812 spin_lock_irq(&rbio->bio_list_lock);
1813 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1814 spin_unlock_irq(&rbio->bio_list_lock);
1817 index_rbio_pages(rbio);
1819 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1821 * Now we just use bitmap to mark the horizontal stripes in
1822 * which we have data when doing parity scrub.
1824 if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB &&
1825 !test_bit(pagenr, rbio->dbitmap))
1828 /* setup our array of pointers with pages
1831 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1833 * if we're rebuilding a read, we have to use
1834 * pages from the bio list
1836 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1837 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
1838 (stripe == faila || stripe == failb)) {
1839 page = page_in_rbio(rbio, stripe, pagenr, 0);
1841 page = rbio_stripe_page(rbio, stripe, pagenr);
1843 pointers[stripe] = kmap(page);
1846 /* all raid6 handling here */
1847 if (rbio->bbio->map_type & BTRFS_BLOCK_GROUP_RAID6) {
1849 * single failure, rebuild from parity raid5
1853 if (faila == rbio->nr_data) {
1855 * Just the P stripe has failed, without
1856 * a bad data or Q stripe.
1857 * TODO, we should redo the xor here.
1863 * a single failure in raid6 is rebuilt
1864 * in the pstripe code below
1869 /* make sure our ps and qs are in order */
1870 if (faila > failb) {
1876 /* if the q stripe is failed, do a pstripe reconstruction
1878 * If both the q stripe and the P stripe are failed, we're
1879 * here due to a crc mismatch and we can't give them the
1882 if (rbio->bbio->raid_map[failb] == RAID6_Q_STRIPE) {
1883 if (rbio->bbio->raid_map[faila] ==
1889 * otherwise we have one bad data stripe and
1890 * a good P stripe. raid5!
1895 if (rbio->bbio->raid_map[failb] == RAID5_P_STRIPE) {
1896 raid6_datap_recov(rbio->real_stripes,
1897 PAGE_SIZE, faila, pointers);
1899 raid6_2data_recov(rbio->real_stripes,
1900 PAGE_SIZE, faila, failb,
1906 /* rebuild from P stripe here (raid5 or raid6) */
1907 BUG_ON(failb != -1);
1909 /* Copy parity block into failed block to start with */
1910 memcpy(pointers[faila],
1911 pointers[rbio->nr_data],
1914 /* rearrange the pointer array */
1915 p = pointers[faila];
1916 for (stripe = faila; stripe < rbio->nr_data - 1; stripe++)
1917 pointers[stripe] = pointers[stripe + 1];
1918 pointers[rbio->nr_data - 1] = p;
1920 /* xor in the rest */
1921 run_xor(pointers, rbio->nr_data - 1, PAGE_SIZE);
1923 /* if we're doing this rebuild as part of an rmw, go through
1924 * and set all of our private rbio pages in the
1925 * failed stripes as uptodate. This way finish_rmw will
1926 * know they can be trusted. If this was a read reconstruction,
1927 * other endio functions will fiddle the uptodate bits
1929 if (rbio->operation == BTRFS_RBIO_WRITE) {
1930 for (i = 0; i < rbio->stripe_npages; i++) {
1932 page = rbio_stripe_page(rbio, faila, i);
1933 SetPageUptodate(page);
1936 page = rbio_stripe_page(rbio, failb, i);
1937 SetPageUptodate(page);
1941 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1943 * if we're rebuilding a read, we have to use
1944 * pages from the bio list
1946 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1947 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
1948 (stripe == faila || stripe == failb)) {
1949 page = page_in_rbio(rbio, stripe, pagenr, 0);
1951 page = rbio_stripe_page(rbio, stripe, pagenr);
1962 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
1964 cache_rbio_pages(rbio);
1966 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
1968 rbio_orig_end_io(rbio, err);
1969 } else if (rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
1970 rbio_orig_end_io(rbio, err);
1971 } else if (err == 0) {
1975 if (rbio->operation == BTRFS_RBIO_WRITE)
1977 else if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB)
1978 finish_parity_scrub(rbio, 0);
1982 rbio_orig_end_io(rbio, err);
1987 * This is called only for stripes we've read from disk to
1988 * reconstruct the parity.
1990 static void raid_recover_end_io(struct bio *bio)
1992 struct btrfs_raid_bio *rbio = bio->bi_private;
1995 * we only read stripe pages off the disk, set them
1996 * up to date if there were no errors
1999 fail_bio_stripe(rbio, bio);
2001 set_bio_pages_uptodate(bio);
2004 if (!atomic_dec_and_test(&rbio->stripes_pending))
2007 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
2008 rbio_orig_end_io(rbio, -EIO);
2010 __raid_recover_end_io(rbio);
2014 * reads everything we need off the disk to reconstruct
2015 * the parity. endio handlers trigger final reconstruction
2016 * when the IO is done.
2018 * This is used both for reads from the higher layers and for
2019 * parity construction required to finish a rmw cycle.
2021 static int __raid56_parity_recover(struct btrfs_raid_bio *rbio)
2023 int bios_to_read = 0;
2024 struct bio_list bio_list;
2030 bio_list_init(&bio_list);
2032 ret = alloc_rbio_pages(rbio);
2036 atomic_set(&rbio->error, 0);
2039 * read everything that hasn't failed. Thanks to the
2040 * stripe cache, it is possible that some or all of these
2041 * pages are going to be uptodate.
2043 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
2044 if (rbio->faila == stripe || rbio->failb == stripe) {
2045 atomic_inc(&rbio->error);
2049 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
2053 * the rmw code may have already read this
2056 p = rbio_stripe_page(rbio, stripe, pagenr);
2057 if (PageUptodate(p))
2060 ret = rbio_add_io_page(rbio, &bio_list,
2061 rbio_stripe_page(rbio, stripe, pagenr),
2062 stripe, pagenr, rbio->stripe_len);
2068 bios_to_read = bio_list_size(&bio_list);
2069 if (!bios_to_read) {
2071 * we might have no bios to read just because the pages
2072 * were up to date, or we might have no bios to read because
2073 * the devices were gone.
2075 if (atomic_read(&rbio->error) <= rbio->bbio->max_errors) {
2076 __raid_recover_end_io(rbio);
2084 * the bbio may be freed once we submit the last bio. Make sure
2085 * not to touch it after that
2087 atomic_set(&rbio->stripes_pending, bios_to_read);
2089 bio = bio_list_pop(&bio_list);
2093 bio->bi_private = rbio;
2094 bio->bi_end_io = raid_recover_end_io;
2095 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2097 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
2105 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
2106 rbio->operation == BTRFS_RBIO_REBUILD_MISSING)
2107 rbio_orig_end_io(rbio, -EIO);
2112 * the main entry point for reads from the higher layers. This
2113 * is really only called when the normal read path had a failure,
2114 * so we assume the bio they send down corresponds to a failed part
2117 int raid56_parity_recover(struct btrfs_fs_info *fs_info, struct bio *bio,
2118 struct btrfs_bio *bbio, u64 stripe_len,
2119 int mirror_num, int generic_io)
2121 struct btrfs_raid_bio *rbio;
2125 ASSERT(bbio->mirror_num == mirror_num);
2126 btrfs_io_bio(bio)->mirror_num = mirror_num;
2129 rbio = alloc_rbio(fs_info, bbio, stripe_len);
2132 btrfs_put_bbio(bbio);
2133 return PTR_ERR(rbio);
2136 rbio->operation = BTRFS_RBIO_READ_REBUILD;
2137 bio_list_add(&rbio->bio_list, bio);
2138 rbio->bio_list_bytes = bio->bi_iter.bi_size;
2140 rbio->faila = find_logical_bio_stripe(rbio, bio);
2141 if (rbio->faila == -1) {
2143 "%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)",
2144 __func__, (u64)bio->bi_iter.bi_sector << 9,
2145 (u64)bio->bi_iter.bi_size, bbio->map_type);
2147 btrfs_put_bbio(bbio);
2153 btrfs_bio_counter_inc_noblocked(fs_info);
2154 rbio->generic_bio_cnt = 1;
2156 btrfs_get_bbio(bbio);
2160 * reconstruct from the q stripe if they are
2161 * asking for mirror 3
2163 if (mirror_num == 3)
2164 rbio->failb = rbio->real_stripes - 2;
2166 ret = lock_stripe_add(rbio);
2169 * __raid56_parity_recover will end the bio with
2170 * any errors it hits. We don't want to return
2171 * its error value up the stack because our caller
2172 * will end up calling bio_endio with any nonzero
2176 __raid56_parity_recover(rbio);
2178 * our rbio has been added to the list of
2179 * rbios that will be handled after the
2180 * currently lock owner is done
2186 static void rmw_work(struct btrfs_work *work)
2188 struct btrfs_raid_bio *rbio;
2190 rbio = container_of(work, struct btrfs_raid_bio, work);
2191 raid56_rmw_stripe(rbio);
2194 static void read_rebuild_work(struct btrfs_work *work)
2196 struct btrfs_raid_bio *rbio;
2198 rbio = container_of(work, struct btrfs_raid_bio, work);
2199 __raid56_parity_recover(rbio);
2203 * The following code is used to scrub/replace the parity stripe
2205 * Caller must have already increased bio_counter for getting @bbio.
2207 * Note: We need make sure all the pages that add into the scrub/replace
2208 * raid bio are correct and not be changed during the scrub/replace. That
2209 * is those pages just hold metadata or file data with checksum.
2212 struct btrfs_raid_bio *
2213 raid56_parity_alloc_scrub_rbio(struct btrfs_fs_info *fs_info, struct bio *bio,
2214 struct btrfs_bio *bbio, u64 stripe_len,
2215 struct btrfs_device *scrub_dev,
2216 unsigned long *dbitmap, int stripe_nsectors)
2218 struct btrfs_raid_bio *rbio;
2221 rbio = alloc_rbio(fs_info, bbio, stripe_len);
2224 bio_list_add(&rbio->bio_list, bio);
2226 * This is a special bio which is used to hold the completion handler
2227 * and make the scrub rbio is similar to the other types
2229 ASSERT(!bio->bi_iter.bi_size);
2230 rbio->operation = BTRFS_RBIO_PARITY_SCRUB;
2232 for (i = 0; i < rbio->real_stripes; i++) {
2233 if (bbio->stripes[i].dev == scrub_dev) {
2239 /* Now we just support the sectorsize equals to page size */
2240 ASSERT(fs_info->sectorsize == PAGE_SIZE);
2241 ASSERT(rbio->stripe_npages == stripe_nsectors);
2242 bitmap_copy(rbio->dbitmap, dbitmap, stripe_nsectors);
2245 * We have already increased bio_counter when getting bbio, record it
2246 * so we can free it at rbio_orig_end_io().
2248 rbio->generic_bio_cnt = 1;
2253 /* Used for both parity scrub and missing. */
2254 void raid56_add_scrub_pages(struct btrfs_raid_bio *rbio, struct page *page,
2260 ASSERT(logical >= rbio->bbio->raid_map[0]);
2261 ASSERT(logical + PAGE_SIZE <= rbio->bbio->raid_map[0] +
2262 rbio->stripe_len * rbio->nr_data);
2263 stripe_offset = (int)(logical - rbio->bbio->raid_map[0]);
2264 index = stripe_offset >> PAGE_SHIFT;
2265 rbio->bio_pages[index] = page;
2269 * We just scrub the parity that we have correct data on the same horizontal,
2270 * so we needn't allocate all pages for all the stripes.
2272 static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio)
2279 for_each_set_bit(bit, rbio->dbitmap, rbio->stripe_npages) {
2280 for (i = 0; i < rbio->real_stripes; i++) {
2281 index = i * rbio->stripe_npages + bit;
2282 if (rbio->stripe_pages[index])
2285 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2288 rbio->stripe_pages[index] = page;
2294 static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
2297 struct btrfs_bio *bbio = rbio->bbio;
2298 void *pointers[rbio->real_stripes];
2299 DECLARE_BITMAP(pbitmap, rbio->stripe_npages);
2300 int nr_data = rbio->nr_data;
2305 struct page *p_page = NULL;
2306 struct page *q_page = NULL;
2307 struct bio_list bio_list;
2312 bio_list_init(&bio_list);
2314 if (rbio->real_stripes - rbio->nr_data == 1) {
2315 p_stripe = rbio->real_stripes - 1;
2316 } else if (rbio->real_stripes - rbio->nr_data == 2) {
2317 p_stripe = rbio->real_stripes - 2;
2318 q_stripe = rbio->real_stripes - 1;
2323 if (bbio->num_tgtdevs && bbio->tgtdev_map[rbio->scrubp]) {
2325 bitmap_copy(pbitmap, rbio->dbitmap, rbio->stripe_npages);
2329 * Because the higher layers(scrubber) are unlikely to
2330 * use this area of the disk again soon, so don't cache
2333 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2338 p_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2341 SetPageUptodate(p_page);
2343 if (q_stripe != -1) {
2344 q_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2346 __free_page(p_page);
2349 SetPageUptodate(q_page);
2352 atomic_set(&rbio->error, 0);
2354 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2357 /* first collect one page from each data stripe */
2358 for (stripe = 0; stripe < nr_data; stripe++) {
2359 p = page_in_rbio(rbio, stripe, pagenr, 0);
2360 pointers[stripe] = kmap(p);
2363 /* then add the parity stripe */
2364 pointers[stripe++] = kmap(p_page);
2366 if (q_stripe != -1) {
2369 * raid6, add the qstripe and call the
2370 * library function to fill in our p/q
2372 pointers[stripe++] = kmap(q_page);
2374 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
2378 memcpy(pointers[nr_data], pointers[0], PAGE_SIZE);
2379 run_xor(pointers + 1, nr_data - 1, PAGE_SIZE);
2382 /* Check scrubbing parity and repair it */
2383 p = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2385 if (memcmp(parity, pointers[rbio->scrubp], PAGE_SIZE))
2386 memcpy(parity, pointers[rbio->scrubp], PAGE_SIZE);
2388 /* Parity is right, needn't writeback */
2389 bitmap_clear(rbio->dbitmap, pagenr, 1);
2392 for (stripe = 0; stripe < rbio->real_stripes; stripe++)
2393 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
2396 __free_page(p_page);
2398 __free_page(q_page);
2402 * time to start writing. Make bios for everything from the
2403 * higher layers (the bio_list in our rbio) and our p/q. Ignore
2406 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2409 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2410 ret = rbio_add_io_page(rbio, &bio_list,
2411 page, rbio->scrubp, pagenr, rbio->stripe_len);
2419 for_each_set_bit(pagenr, pbitmap, rbio->stripe_npages) {
2422 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2423 ret = rbio_add_io_page(rbio, &bio_list, page,
2424 bbio->tgtdev_map[rbio->scrubp],
2425 pagenr, rbio->stripe_len);
2431 nr_data = bio_list_size(&bio_list);
2433 /* Every parity is right */
2434 rbio_orig_end_io(rbio, 0);
2438 atomic_set(&rbio->stripes_pending, nr_data);
2441 bio = bio_list_pop(&bio_list);
2445 bio->bi_private = rbio;
2446 bio->bi_end_io = raid_write_end_io;
2447 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2454 rbio_orig_end_io(rbio, -EIO);
2457 static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe)
2459 if (stripe >= 0 && stripe < rbio->nr_data)
2465 * While we're doing the parity check and repair, we could have errors
2466 * in reading pages off the disk. This checks for errors and if we're
2467 * not able to read the page it'll trigger parity reconstruction. The
2468 * parity scrub will be finished after we've reconstructed the failed
2471 static void validate_rbio_for_parity_scrub(struct btrfs_raid_bio *rbio)
2473 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
2476 if (rbio->faila >= 0 || rbio->failb >= 0) {
2477 int dfail = 0, failp = -1;
2479 if (is_data_stripe(rbio, rbio->faila))
2481 else if (is_parity_stripe(rbio->faila))
2482 failp = rbio->faila;
2484 if (is_data_stripe(rbio, rbio->failb))
2486 else if (is_parity_stripe(rbio->failb))
2487 failp = rbio->failb;
2490 * Because we can not use a scrubbing parity to repair
2491 * the data, so the capability of the repair is declined.
2492 * (In the case of RAID5, we can not repair anything)
2494 if (dfail > rbio->bbio->max_errors - 1)
2498 * If all data is good, only parity is correctly, just
2499 * repair the parity.
2502 finish_parity_scrub(rbio, 0);
2507 * Here means we got one corrupted data stripe and one
2508 * corrupted parity on RAID6, if the corrupted parity
2509 * is scrubbing parity, luckily, use the other one to repair
2510 * the data, or we can not repair the data stripe.
2512 if (failp != rbio->scrubp)
2515 __raid_recover_end_io(rbio);
2517 finish_parity_scrub(rbio, 1);
2522 rbio_orig_end_io(rbio, -EIO);
2526 * end io for the read phase of the rmw cycle. All the bios here are physical
2527 * stripe bios we've read from the disk so we can recalculate the parity of the
2530 * This will usually kick off finish_rmw once all the bios are read in, but it
2531 * may trigger parity reconstruction if we had any errors along the way
2533 static void raid56_parity_scrub_end_io(struct bio *bio)
2535 struct btrfs_raid_bio *rbio = bio->bi_private;
2538 fail_bio_stripe(rbio, bio);
2540 set_bio_pages_uptodate(bio);
2544 if (!atomic_dec_and_test(&rbio->stripes_pending))
2548 * this will normally call finish_rmw to start our write
2549 * but if there are any failed stripes we'll reconstruct
2552 validate_rbio_for_parity_scrub(rbio);
2555 static void raid56_parity_scrub_stripe(struct btrfs_raid_bio *rbio)
2557 int bios_to_read = 0;
2558 struct bio_list bio_list;
2564 ret = alloc_rbio_essential_pages(rbio);
2568 bio_list_init(&bio_list);
2570 atomic_set(&rbio->error, 0);
2572 * build a list of bios to read all the missing parts of this
2575 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
2576 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2579 * we want to find all the pages missing from
2580 * the rbio and read them from the disk. If
2581 * page_in_rbio finds a page in the bio list
2582 * we don't need to read it off the stripe.
2584 page = page_in_rbio(rbio, stripe, pagenr, 1);
2588 page = rbio_stripe_page(rbio, stripe, pagenr);
2590 * the bio cache may have handed us an uptodate
2591 * page. If so, be happy and use it
2593 if (PageUptodate(page))
2596 ret = rbio_add_io_page(rbio, &bio_list, page,
2597 stripe, pagenr, rbio->stripe_len);
2603 bios_to_read = bio_list_size(&bio_list);
2604 if (!bios_to_read) {
2606 * this can happen if others have merged with
2607 * us, it means there is nothing left to read.
2608 * But if there are missing devices it may not be
2609 * safe to do the full stripe write yet.
2615 * the bbio may be freed once we submit the last bio. Make sure
2616 * not to touch it after that
2618 atomic_set(&rbio->stripes_pending, bios_to_read);
2620 bio = bio_list_pop(&bio_list);
2624 bio->bi_private = rbio;
2625 bio->bi_end_io = raid56_parity_scrub_end_io;
2626 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2628 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
2632 /* the actual write will happen once the reads are done */
2636 rbio_orig_end_io(rbio, -EIO);
2640 validate_rbio_for_parity_scrub(rbio);
2643 static void scrub_parity_work(struct btrfs_work *work)
2645 struct btrfs_raid_bio *rbio;
2647 rbio = container_of(work, struct btrfs_raid_bio, work);
2648 raid56_parity_scrub_stripe(rbio);
2651 static void async_scrub_parity(struct btrfs_raid_bio *rbio)
2653 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
2654 scrub_parity_work, NULL, NULL);
2656 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
2659 void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio)
2661 if (!lock_stripe_add(rbio))
2662 async_scrub_parity(rbio);
2665 /* The following code is used for dev replace of a missing RAID 5/6 device. */
2667 struct btrfs_raid_bio *
2668 raid56_alloc_missing_rbio(struct btrfs_fs_info *fs_info, struct bio *bio,
2669 struct btrfs_bio *bbio, u64 length)
2671 struct btrfs_raid_bio *rbio;
2673 rbio = alloc_rbio(fs_info, bbio, length);
2677 rbio->operation = BTRFS_RBIO_REBUILD_MISSING;
2678 bio_list_add(&rbio->bio_list, bio);
2680 * This is a special bio which is used to hold the completion handler
2681 * and make the scrub rbio is similar to the other types
2683 ASSERT(!bio->bi_iter.bi_size);
2685 rbio->faila = find_logical_bio_stripe(rbio, bio);
2686 if (rbio->faila == -1) {
2693 * When we get bbio, we have already increased bio_counter, record it
2694 * so we can free it at rbio_orig_end_io()
2696 rbio->generic_bio_cnt = 1;
2701 static void missing_raid56_work(struct btrfs_work *work)
2703 struct btrfs_raid_bio *rbio;
2705 rbio = container_of(work, struct btrfs_raid_bio, work);
2706 __raid56_parity_recover(rbio);
2709 static void async_missing_raid56(struct btrfs_raid_bio *rbio)
2711 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
2712 missing_raid56_work, NULL, NULL);
2714 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
2717 void raid56_submit_missing_rbio(struct btrfs_raid_bio *rbio)
2719 if (!lock_stripe_add(rbio))
2720 async_missing_raid56(rbio);