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)
1139 struct bio_vec *bvec;
1141 unsigned long stripe_offset;
1142 unsigned long page_index;
1145 spin_lock_irq(&rbio->bio_list_lock);
1146 bio_list_for_each(bio, &rbio->bio_list) {
1147 start = (u64)bio->bi_iter.bi_sector << 9;
1148 stripe_offset = start - rbio->bbio->raid_map[0];
1149 page_index = stripe_offset >> PAGE_SHIFT;
1151 bio_for_each_segment_all(bvec, bio, i)
1152 rbio->bio_pages[page_index + i] = bvec->bv_page;
1154 spin_unlock_irq(&rbio->bio_list_lock);
1158 * this is called from one of two situations. We either
1159 * have a full stripe from the higher layers, or we've read all
1160 * the missing bits off disk.
1162 * This will calculate the parity and then send down any
1165 static noinline void finish_rmw(struct btrfs_raid_bio *rbio)
1167 struct btrfs_bio *bbio = rbio->bbio;
1168 void *pointers[rbio->real_stripes];
1169 int nr_data = rbio->nr_data;
1174 struct bio_list bio_list;
1178 bio_list_init(&bio_list);
1180 if (rbio->real_stripes - rbio->nr_data == 1) {
1181 p_stripe = rbio->real_stripes - 1;
1182 } else if (rbio->real_stripes - rbio->nr_data == 2) {
1183 p_stripe = rbio->real_stripes - 2;
1184 q_stripe = rbio->real_stripes - 1;
1189 /* at this point we either have a full stripe,
1190 * or we've read the full stripe from the drive.
1191 * recalculate the parity and write the new results.
1193 * We're not allowed to add any new bios to the
1194 * bio list here, anyone else that wants to
1195 * change this stripe needs to do their own rmw.
1197 spin_lock_irq(&rbio->bio_list_lock);
1198 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1199 spin_unlock_irq(&rbio->bio_list_lock);
1201 atomic_set(&rbio->error, 0);
1204 * now that we've set rmw_locked, run through the
1205 * bio list one last time and map the page pointers
1207 * We don't cache full rbios because we're assuming
1208 * the higher layers are unlikely to use this area of
1209 * the disk again soon. If they do use it again,
1210 * hopefully they will send another full bio.
1212 index_rbio_pages(rbio);
1213 if (!rbio_is_full(rbio))
1214 cache_rbio_pages(rbio);
1216 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
1218 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1220 /* first collect one page from each data stripe */
1221 for (stripe = 0; stripe < nr_data; stripe++) {
1222 p = page_in_rbio(rbio, stripe, pagenr, 0);
1223 pointers[stripe] = kmap(p);
1226 /* then add the parity stripe */
1227 p = rbio_pstripe_page(rbio, pagenr);
1229 pointers[stripe++] = kmap(p);
1231 if (q_stripe != -1) {
1234 * raid6, add the qstripe and call the
1235 * library function to fill in our p/q
1237 p = rbio_qstripe_page(rbio, pagenr);
1239 pointers[stripe++] = kmap(p);
1241 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
1245 memcpy(pointers[nr_data], pointers[0], PAGE_SIZE);
1246 run_xor(pointers + 1, nr_data - 1, PAGE_SIZE);
1250 for (stripe = 0; stripe < rbio->real_stripes; stripe++)
1251 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
1255 * time to start writing. Make bios for everything from the
1256 * higher layers (the bio_list in our rbio) and our p/q. Ignore
1259 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1260 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1262 if (stripe < rbio->nr_data) {
1263 page = page_in_rbio(rbio, stripe, pagenr, 1);
1267 page = rbio_stripe_page(rbio, stripe, pagenr);
1270 ret = rbio_add_io_page(rbio, &bio_list,
1271 page, stripe, pagenr, rbio->stripe_len);
1277 if (likely(!bbio->num_tgtdevs))
1280 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1281 if (!bbio->tgtdev_map[stripe])
1284 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1286 if (stripe < rbio->nr_data) {
1287 page = page_in_rbio(rbio, stripe, pagenr, 1);
1291 page = rbio_stripe_page(rbio, stripe, pagenr);
1294 ret = rbio_add_io_page(rbio, &bio_list, page,
1295 rbio->bbio->tgtdev_map[stripe],
1296 pagenr, rbio->stripe_len);
1303 atomic_set(&rbio->stripes_pending, bio_list_size(&bio_list));
1304 BUG_ON(atomic_read(&rbio->stripes_pending) == 0);
1307 bio = bio_list_pop(&bio_list);
1311 bio->bi_private = rbio;
1312 bio->bi_end_io = raid_write_end_io;
1313 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1320 rbio_orig_end_io(rbio, -EIO);
1324 * helper to find the stripe number for a given bio. Used to figure out which
1325 * stripe has failed. This expects the bio to correspond to a physical disk,
1326 * so it looks up based on physical sector numbers.
1328 static int find_bio_stripe(struct btrfs_raid_bio *rbio,
1331 u64 physical = bio->bi_iter.bi_sector;
1334 struct btrfs_bio_stripe *stripe;
1338 for (i = 0; i < rbio->bbio->num_stripes; i++) {
1339 stripe = &rbio->bbio->stripes[i];
1340 stripe_start = stripe->physical;
1341 if (physical >= stripe_start &&
1342 physical < stripe_start + rbio->stripe_len &&
1343 bio->bi_bdev == stripe->dev->bdev) {
1351 * helper to find the stripe number for a given
1352 * bio (before mapping). Used to figure out which stripe has
1353 * failed. This looks up based on logical block numbers.
1355 static int find_logical_bio_stripe(struct btrfs_raid_bio *rbio,
1358 u64 logical = bio->bi_iter.bi_sector;
1364 for (i = 0; i < rbio->nr_data; i++) {
1365 stripe_start = rbio->bbio->raid_map[i];
1366 if (logical >= stripe_start &&
1367 logical < stripe_start + rbio->stripe_len) {
1375 * returns -EIO if we had too many failures
1377 static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed)
1379 unsigned long flags;
1382 spin_lock_irqsave(&rbio->bio_list_lock, flags);
1384 /* we already know this stripe is bad, move on */
1385 if (rbio->faila == failed || rbio->failb == failed)
1388 if (rbio->faila == -1) {
1389 /* first failure on this rbio */
1390 rbio->faila = failed;
1391 atomic_inc(&rbio->error);
1392 } else if (rbio->failb == -1) {
1393 /* second failure on this rbio */
1394 rbio->failb = failed;
1395 atomic_inc(&rbio->error);
1400 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
1406 * helper to fail a stripe based on a physical disk
1409 static int fail_bio_stripe(struct btrfs_raid_bio *rbio,
1412 int failed = find_bio_stripe(rbio, bio);
1417 return fail_rbio_index(rbio, failed);
1421 * this sets each page in the bio uptodate. It should only be used on private
1422 * rbio pages, nothing that comes in from the higher layers
1424 static void set_bio_pages_uptodate(struct bio *bio)
1426 struct bio_vec *bvec;
1429 bio_for_each_segment_all(bvec, bio, i)
1430 SetPageUptodate(bvec->bv_page);
1434 * end io for the read phase of the rmw cycle. All the bios here are physical
1435 * stripe bios we've read from the disk so we can recalculate the parity of the
1438 * This will usually kick off finish_rmw once all the bios are read in, but it
1439 * may trigger parity reconstruction if we had any errors along the way
1441 static void raid_rmw_end_io(struct bio *bio)
1443 struct btrfs_raid_bio *rbio = bio->bi_private;
1446 fail_bio_stripe(rbio, bio);
1448 set_bio_pages_uptodate(bio);
1452 if (!atomic_dec_and_test(&rbio->stripes_pending))
1455 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
1459 * this will normally call finish_rmw to start our write
1460 * but if there are any failed stripes we'll reconstruct
1463 validate_rbio_for_rmw(rbio);
1468 rbio_orig_end_io(rbio, -EIO);
1471 static void async_rmw_stripe(struct btrfs_raid_bio *rbio)
1473 btrfs_init_work(&rbio->work, btrfs_rmw_helper, rmw_work, NULL, NULL);
1474 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
1477 static void async_read_rebuild(struct btrfs_raid_bio *rbio)
1479 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
1480 read_rebuild_work, NULL, NULL);
1482 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
1486 * the stripe must be locked by the caller. It will
1487 * unlock after all the writes are done
1489 static int raid56_rmw_stripe(struct btrfs_raid_bio *rbio)
1491 int bios_to_read = 0;
1492 struct bio_list bio_list;
1498 bio_list_init(&bio_list);
1500 ret = alloc_rbio_pages(rbio);
1504 index_rbio_pages(rbio);
1506 atomic_set(&rbio->error, 0);
1508 * build a list of bios to read all the missing parts of this
1511 for (stripe = 0; stripe < rbio->nr_data; stripe++) {
1512 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1515 * we want to find all the pages missing from
1516 * the rbio and read them from the disk. If
1517 * page_in_rbio finds a page in the bio list
1518 * we don't need to read it off the stripe.
1520 page = page_in_rbio(rbio, stripe, pagenr, 1);
1524 page = rbio_stripe_page(rbio, stripe, pagenr);
1526 * the bio cache may have handed us an uptodate
1527 * page. If so, be happy and use it
1529 if (PageUptodate(page))
1532 ret = rbio_add_io_page(rbio, &bio_list, page,
1533 stripe, pagenr, rbio->stripe_len);
1539 bios_to_read = bio_list_size(&bio_list);
1540 if (!bios_to_read) {
1542 * this can happen if others have merged with
1543 * us, it means there is nothing left to read.
1544 * But if there are missing devices it may not be
1545 * safe to do the full stripe write yet.
1551 * the bbio may be freed once we submit the last bio. Make sure
1552 * not to touch it after that
1554 atomic_set(&rbio->stripes_pending, bios_to_read);
1556 bio = bio_list_pop(&bio_list);
1560 bio->bi_private = rbio;
1561 bio->bi_end_io = raid_rmw_end_io;
1562 bio_set_op_attrs(bio, REQ_OP_READ, 0);
1564 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
1568 /* the actual write will happen once the reads are done */
1572 rbio_orig_end_io(rbio, -EIO);
1576 validate_rbio_for_rmw(rbio);
1581 * if the upper layers pass in a full stripe, we thank them by only allocating
1582 * enough pages to hold the parity, and sending it all down quickly.
1584 static int full_stripe_write(struct btrfs_raid_bio *rbio)
1588 ret = alloc_rbio_parity_pages(rbio);
1590 __free_raid_bio(rbio);
1594 ret = lock_stripe_add(rbio);
1601 * partial stripe writes get handed over to async helpers.
1602 * We're really hoping to merge a few more writes into this
1603 * rbio before calculating new parity
1605 static int partial_stripe_write(struct btrfs_raid_bio *rbio)
1609 ret = lock_stripe_add(rbio);
1611 async_rmw_stripe(rbio);
1616 * sometimes while we were reading from the drive to
1617 * recalculate parity, enough new bios come into create
1618 * a full stripe. So we do a check here to see if we can
1619 * go directly to finish_rmw
1621 static int __raid56_parity_write(struct btrfs_raid_bio *rbio)
1623 /* head off into rmw land if we don't have a full stripe */
1624 if (!rbio_is_full(rbio))
1625 return partial_stripe_write(rbio);
1626 return full_stripe_write(rbio);
1630 * We use plugging call backs to collect full stripes.
1631 * Any time we get a partial stripe write while plugged
1632 * we collect it into a list. When the unplug comes down,
1633 * we sort the list by logical block number and merge
1634 * everything we can into the same rbios
1636 struct btrfs_plug_cb {
1637 struct blk_plug_cb cb;
1638 struct btrfs_fs_info *info;
1639 struct list_head rbio_list;
1640 struct btrfs_work work;
1644 * rbios on the plug list are sorted for easier merging.
1646 static int plug_cmp(void *priv, struct list_head *a, struct list_head *b)
1648 struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio,
1650 struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio,
1652 u64 a_sector = ra->bio_list.head->bi_iter.bi_sector;
1653 u64 b_sector = rb->bio_list.head->bi_iter.bi_sector;
1655 if (a_sector < b_sector)
1657 if (a_sector > b_sector)
1662 static void run_plug(struct btrfs_plug_cb *plug)
1664 struct btrfs_raid_bio *cur;
1665 struct btrfs_raid_bio *last = NULL;
1668 * sort our plug list then try to merge
1669 * everything we can in hopes of creating full
1672 list_sort(NULL, &plug->rbio_list, plug_cmp);
1673 while (!list_empty(&plug->rbio_list)) {
1674 cur = list_entry(plug->rbio_list.next,
1675 struct btrfs_raid_bio, plug_list);
1676 list_del_init(&cur->plug_list);
1678 if (rbio_is_full(cur)) {
1679 /* we have a full stripe, send it down */
1680 full_stripe_write(cur);
1684 if (rbio_can_merge(last, cur)) {
1685 merge_rbio(last, cur);
1686 __free_raid_bio(cur);
1690 __raid56_parity_write(last);
1695 __raid56_parity_write(last);
1701 * if the unplug comes from schedule, we have to push the
1702 * work off to a helper thread
1704 static void unplug_work(struct btrfs_work *work)
1706 struct btrfs_plug_cb *plug;
1707 plug = container_of(work, struct btrfs_plug_cb, work);
1711 static void btrfs_raid_unplug(struct blk_plug_cb *cb, bool from_schedule)
1713 struct btrfs_plug_cb *plug;
1714 plug = container_of(cb, struct btrfs_plug_cb, cb);
1716 if (from_schedule) {
1717 btrfs_init_work(&plug->work, btrfs_rmw_helper,
1718 unplug_work, NULL, NULL);
1719 btrfs_queue_work(plug->info->rmw_workers,
1727 * our main entry point for writes from the rest of the FS.
1729 int raid56_parity_write(struct btrfs_fs_info *fs_info, struct bio *bio,
1730 struct btrfs_bio *bbio, u64 stripe_len)
1732 struct btrfs_raid_bio *rbio;
1733 struct btrfs_plug_cb *plug = NULL;
1734 struct blk_plug_cb *cb;
1737 rbio = alloc_rbio(fs_info, bbio, stripe_len);
1739 btrfs_put_bbio(bbio);
1740 return PTR_ERR(rbio);
1742 bio_list_add(&rbio->bio_list, bio);
1743 rbio->bio_list_bytes = bio->bi_iter.bi_size;
1744 rbio->operation = BTRFS_RBIO_WRITE;
1746 btrfs_bio_counter_inc_noblocked(fs_info);
1747 rbio->generic_bio_cnt = 1;
1750 * don't plug on full rbios, just get them out the door
1751 * as quickly as we can
1753 if (rbio_is_full(rbio)) {
1754 ret = full_stripe_write(rbio);
1756 btrfs_bio_counter_dec(fs_info);
1760 cb = blk_check_plugged(btrfs_raid_unplug, fs_info, sizeof(*plug));
1762 plug = container_of(cb, struct btrfs_plug_cb, cb);
1764 plug->info = fs_info;
1765 INIT_LIST_HEAD(&plug->rbio_list);
1767 list_add_tail(&rbio->plug_list, &plug->rbio_list);
1770 ret = __raid56_parity_write(rbio);
1772 btrfs_bio_counter_dec(fs_info);
1778 * all parity reconstruction happens here. We've read in everything
1779 * we can find from the drives and this does the heavy lifting of
1780 * sorting the good from the bad.
1782 static void __raid_recover_end_io(struct btrfs_raid_bio *rbio)
1786 int faila = -1, failb = -1;
1791 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
1797 faila = rbio->faila;
1798 failb = rbio->failb;
1800 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1801 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
1802 spin_lock_irq(&rbio->bio_list_lock);
1803 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1804 spin_unlock_irq(&rbio->bio_list_lock);
1807 index_rbio_pages(rbio);
1809 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1811 * Now we just use bitmap to mark the horizontal stripes in
1812 * which we have data when doing parity scrub.
1814 if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB &&
1815 !test_bit(pagenr, rbio->dbitmap))
1818 /* setup our array of pointers with pages
1821 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1823 * if we're rebuilding a read, we have to use
1824 * pages from the bio list
1826 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1827 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
1828 (stripe == faila || stripe == failb)) {
1829 page = page_in_rbio(rbio, stripe, pagenr, 0);
1831 page = rbio_stripe_page(rbio, stripe, pagenr);
1833 pointers[stripe] = kmap(page);
1836 /* all raid6 handling here */
1837 if (rbio->bbio->map_type & BTRFS_BLOCK_GROUP_RAID6) {
1839 * single failure, rebuild from parity raid5
1843 if (faila == rbio->nr_data) {
1845 * Just the P stripe has failed, without
1846 * a bad data or Q stripe.
1847 * TODO, we should redo the xor here.
1853 * a single failure in raid6 is rebuilt
1854 * in the pstripe code below
1859 /* make sure our ps and qs are in order */
1860 if (faila > failb) {
1866 /* if the q stripe is failed, do a pstripe reconstruction
1868 * If both the q stripe and the P stripe are failed, we're
1869 * here due to a crc mismatch and we can't give them the
1872 if (rbio->bbio->raid_map[failb] == RAID6_Q_STRIPE) {
1873 if (rbio->bbio->raid_map[faila] ==
1879 * otherwise we have one bad data stripe and
1880 * a good P stripe. raid5!
1885 if (rbio->bbio->raid_map[failb] == RAID5_P_STRIPE) {
1886 raid6_datap_recov(rbio->real_stripes,
1887 PAGE_SIZE, faila, pointers);
1889 raid6_2data_recov(rbio->real_stripes,
1890 PAGE_SIZE, faila, failb,
1896 /* rebuild from P stripe here (raid5 or raid6) */
1897 BUG_ON(failb != -1);
1899 /* Copy parity block into failed block to start with */
1900 memcpy(pointers[faila],
1901 pointers[rbio->nr_data],
1904 /* rearrange the pointer array */
1905 p = pointers[faila];
1906 for (stripe = faila; stripe < rbio->nr_data - 1; stripe++)
1907 pointers[stripe] = pointers[stripe + 1];
1908 pointers[rbio->nr_data - 1] = p;
1910 /* xor in the rest */
1911 run_xor(pointers, rbio->nr_data - 1, PAGE_SIZE);
1913 /* if we're doing this rebuild as part of an rmw, go through
1914 * and set all of our private rbio pages in the
1915 * failed stripes as uptodate. This way finish_rmw will
1916 * know they can be trusted. If this was a read reconstruction,
1917 * other endio functions will fiddle the uptodate bits
1919 if (rbio->operation == BTRFS_RBIO_WRITE) {
1920 for (i = 0; i < rbio->stripe_npages; i++) {
1922 page = rbio_stripe_page(rbio, faila, i);
1923 SetPageUptodate(page);
1926 page = rbio_stripe_page(rbio, failb, i);
1927 SetPageUptodate(page);
1931 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1933 * if we're rebuilding a read, we have to use
1934 * pages from the bio list
1936 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1937 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
1938 (stripe == faila || stripe == failb)) {
1939 page = page_in_rbio(rbio, stripe, pagenr, 0);
1941 page = rbio_stripe_page(rbio, stripe, pagenr);
1952 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
1954 cache_rbio_pages(rbio);
1956 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
1958 rbio_orig_end_io(rbio, err);
1959 } else if (rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
1960 rbio_orig_end_io(rbio, err);
1961 } else if (err == 0) {
1965 if (rbio->operation == BTRFS_RBIO_WRITE)
1967 else if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB)
1968 finish_parity_scrub(rbio, 0);
1972 rbio_orig_end_io(rbio, err);
1977 * This is called only for stripes we've read from disk to
1978 * reconstruct the parity.
1980 static void raid_recover_end_io(struct bio *bio)
1982 struct btrfs_raid_bio *rbio = bio->bi_private;
1985 * we only read stripe pages off the disk, set them
1986 * up to date if there were no errors
1989 fail_bio_stripe(rbio, bio);
1991 set_bio_pages_uptodate(bio);
1994 if (!atomic_dec_and_test(&rbio->stripes_pending))
1997 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
1998 rbio_orig_end_io(rbio, -EIO);
2000 __raid_recover_end_io(rbio);
2004 * reads everything we need off the disk to reconstruct
2005 * the parity. endio handlers trigger final reconstruction
2006 * when the IO is done.
2008 * This is used both for reads from the higher layers and for
2009 * parity construction required to finish a rmw cycle.
2011 static int __raid56_parity_recover(struct btrfs_raid_bio *rbio)
2013 int bios_to_read = 0;
2014 struct bio_list bio_list;
2020 bio_list_init(&bio_list);
2022 ret = alloc_rbio_pages(rbio);
2026 atomic_set(&rbio->error, 0);
2029 * read everything that hasn't failed. Thanks to the
2030 * stripe cache, it is possible that some or all of these
2031 * pages are going to be uptodate.
2033 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
2034 if (rbio->faila == stripe || rbio->failb == stripe) {
2035 atomic_inc(&rbio->error);
2039 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
2043 * the rmw code may have already read this
2046 p = rbio_stripe_page(rbio, stripe, pagenr);
2047 if (PageUptodate(p))
2050 ret = rbio_add_io_page(rbio, &bio_list,
2051 rbio_stripe_page(rbio, stripe, pagenr),
2052 stripe, pagenr, rbio->stripe_len);
2058 bios_to_read = bio_list_size(&bio_list);
2059 if (!bios_to_read) {
2061 * we might have no bios to read just because the pages
2062 * were up to date, or we might have no bios to read because
2063 * the devices were gone.
2065 if (atomic_read(&rbio->error) <= rbio->bbio->max_errors) {
2066 __raid_recover_end_io(rbio);
2074 * the bbio may be freed once we submit the last bio. Make sure
2075 * not to touch it after that
2077 atomic_set(&rbio->stripes_pending, bios_to_read);
2079 bio = bio_list_pop(&bio_list);
2083 bio->bi_private = rbio;
2084 bio->bi_end_io = raid_recover_end_io;
2085 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2087 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
2095 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
2096 rbio->operation == BTRFS_RBIO_REBUILD_MISSING)
2097 rbio_orig_end_io(rbio, -EIO);
2102 * the main entry point for reads from the higher layers. This
2103 * is really only called when the normal read path had a failure,
2104 * so we assume the bio they send down corresponds to a failed part
2107 int raid56_parity_recover(struct btrfs_fs_info *fs_info, struct bio *bio,
2108 struct btrfs_bio *bbio, u64 stripe_len,
2109 int mirror_num, int generic_io)
2111 struct btrfs_raid_bio *rbio;
2115 ASSERT(bbio->mirror_num == mirror_num);
2116 btrfs_io_bio(bio)->mirror_num = mirror_num;
2119 rbio = alloc_rbio(fs_info, bbio, stripe_len);
2122 btrfs_put_bbio(bbio);
2123 return PTR_ERR(rbio);
2126 rbio->operation = BTRFS_RBIO_READ_REBUILD;
2127 bio_list_add(&rbio->bio_list, bio);
2128 rbio->bio_list_bytes = bio->bi_iter.bi_size;
2130 rbio->faila = find_logical_bio_stripe(rbio, bio);
2131 if (rbio->faila == -1) {
2133 "%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)",
2134 __func__, (u64)bio->bi_iter.bi_sector << 9,
2135 (u64)bio->bi_iter.bi_size, bbio->map_type);
2137 btrfs_put_bbio(bbio);
2143 btrfs_bio_counter_inc_noblocked(fs_info);
2144 rbio->generic_bio_cnt = 1;
2146 btrfs_get_bbio(bbio);
2150 * reconstruct from the q stripe if they are
2151 * asking for mirror 3
2153 if (mirror_num == 3)
2154 rbio->failb = rbio->real_stripes - 2;
2156 ret = lock_stripe_add(rbio);
2159 * __raid56_parity_recover will end the bio with
2160 * any errors it hits. We don't want to return
2161 * its error value up the stack because our caller
2162 * will end up calling bio_endio with any nonzero
2166 __raid56_parity_recover(rbio);
2168 * our rbio has been added to the list of
2169 * rbios that will be handled after the
2170 * currently lock owner is done
2176 static void rmw_work(struct btrfs_work *work)
2178 struct btrfs_raid_bio *rbio;
2180 rbio = container_of(work, struct btrfs_raid_bio, work);
2181 raid56_rmw_stripe(rbio);
2184 static void read_rebuild_work(struct btrfs_work *work)
2186 struct btrfs_raid_bio *rbio;
2188 rbio = container_of(work, struct btrfs_raid_bio, work);
2189 __raid56_parity_recover(rbio);
2193 * The following code is used to scrub/replace the parity stripe
2195 * Caller must have already increased bio_counter for getting @bbio.
2197 * Note: We need make sure all the pages that add into the scrub/replace
2198 * raid bio are correct and not be changed during the scrub/replace. That
2199 * is those pages just hold metadata or file data with checksum.
2202 struct btrfs_raid_bio *
2203 raid56_parity_alloc_scrub_rbio(struct btrfs_fs_info *fs_info, struct bio *bio,
2204 struct btrfs_bio *bbio, u64 stripe_len,
2205 struct btrfs_device *scrub_dev,
2206 unsigned long *dbitmap, int stripe_nsectors)
2208 struct btrfs_raid_bio *rbio;
2211 rbio = alloc_rbio(fs_info, bbio, stripe_len);
2214 bio_list_add(&rbio->bio_list, bio);
2216 * This is a special bio which is used to hold the completion handler
2217 * and make the scrub rbio is similar to the other types
2219 ASSERT(!bio->bi_iter.bi_size);
2220 rbio->operation = BTRFS_RBIO_PARITY_SCRUB;
2222 for (i = 0; i < rbio->real_stripes; i++) {
2223 if (bbio->stripes[i].dev == scrub_dev) {
2229 /* Now we just support the sectorsize equals to page size */
2230 ASSERT(fs_info->sectorsize == PAGE_SIZE);
2231 ASSERT(rbio->stripe_npages == stripe_nsectors);
2232 bitmap_copy(rbio->dbitmap, dbitmap, stripe_nsectors);
2235 * We have already increased bio_counter when getting bbio, record it
2236 * so we can free it at rbio_orig_end_io().
2238 rbio->generic_bio_cnt = 1;
2243 /* Used for both parity scrub and missing. */
2244 void raid56_add_scrub_pages(struct btrfs_raid_bio *rbio, struct page *page,
2250 ASSERT(logical >= rbio->bbio->raid_map[0]);
2251 ASSERT(logical + PAGE_SIZE <= rbio->bbio->raid_map[0] +
2252 rbio->stripe_len * rbio->nr_data);
2253 stripe_offset = (int)(logical - rbio->bbio->raid_map[0]);
2254 index = stripe_offset >> PAGE_SHIFT;
2255 rbio->bio_pages[index] = page;
2259 * We just scrub the parity that we have correct data on the same horizontal,
2260 * so we needn't allocate all pages for all the stripes.
2262 static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio)
2269 for_each_set_bit(bit, rbio->dbitmap, rbio->stripe_npages) {
2270 for (i = 0; i < rbio->real_stripes; i++) {
2271 index = i * rbio->stripe_npages + bit;
2272 if (rbio->stripe_pages[index])
2275 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2278 rbio->stripe_pages[index] = page;
2284 static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
2287 struct btrfs_bio *bbio = rbio->bbio;
2288 void *pointers[rbio->real_stripes];
2289 DECLARE_BITMAP(pbitmap, rbio->stripe_npages);
2290 int nr_data = rbio->nr_data;
2295 struct page *p_page = NULL;
2296 struct page *q_page = NULL;
2297 struct bio_list bio_list;
2302 bio_list_init(&bio_list);
2304 if (rbio->real_stripes - rbio->nr_data == 1) {
2305 p_stripe = rbio->real_stripes - 1;
2306 } else if (rbio->real_stripes - rbio->nr_data == 2) {
2307 p_stripe = rbio->real_stripes - 2;
2308 q_stripe = rbio->real_stripes - 1;
2313 if (bbio->num_tgtdevs && bbio->tgtdev_map[rbio->scrubp]) {
2315 bitmap_copy(pbitmap, rbio->dbitmap, rbio->stripe_npages);
2319 * Because the higher layers(scrubber) are unlikely to
2320 * use this area of the disk again soon, so don't cache
2323 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2328 p_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2331 SetPageUptodate(p_page);
2333 if (q_stripe != -1) {
2334 q_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2336 __free_page(p_page);
2339 SetPageUptodate(q_page);
2342 atomic_set(&rbio->error, 0);
2344 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2347 /* first collect one page from each data stripe */
2348 for (stripe = 0; stripe < nr_data; stripe++) {
2349 p = page_in_rbio(rbio, stripe, pagenr, 0);
2350 pointers[stripe] = kmap(p);
2353 /* then add the parity stripe */
2354 pointers[stripe++] = kmap(p_page);
2356 if (q_stripe != -1) {
2359 * raid6, add the qstripe and call the
2360 * library function to fill in our p/q
2362 pointers[stripe++] = kmap(q_page);
2364 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
2368 memcpy(pointers[nr_data], pointers[0], PAGE_SIZE);
2369 run_xor(pointers + 1, nr_data - 1, PAGE_SIZE);
2372 /* Check scrubbing parity and repair it */
2373 p = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2375 if (memcmp(parity, pointers[rbio->scrubp], PAGE_SIZE))
2376 memcpy(parity, pointers[rbio->scrubp], PAGE_SIZE);
2378 /* Parity is right, needn't writeback */
2379 bitmap_clear(rbio->dbitmap, pagenr, 1);
2382 for (stripe = 0; stripe < rbio->real_stripes; stripe++)
2383 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
2386 __free_page(p_page);
2388 __free_page(q_page);
2392 * time to start writing. Make bios for everything from the
2393 * higher layers (the bio_list in our rbio) and our p/q. Ignore
2396 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2399 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2400 ret = rbio_add_io_page(rbio, &bio_list,
2401 page, rbio->scrubp, pagenr, rbio->stripe_len);
2409 for_each_set_bit(pagenr, pbitmap, rbio->stripe_npages) {
2412 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2413 ret = rbio_add_io_page(rbio, &bio_list, page,
2414 bbio->tgtdev_map[rbio->scrubp],
2415 pagenr, rbio->stripe_len);
2421 nr_data = bio_list_size(&bio_list);
2423 /* Every parity is right */
2424 rbio_orig_end_io(rbio, 0);
2428 atomic_set(&rbio->stripes_pending, nr_data);
2431 bio = bio_list_pop(&bio_list);
2435 bio->bi_private = rbio;
2436 bio->bi_end_io = raid_write_end_io;
2437 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2444 rbio_orig_end_io(rbio, -EIO);
2447 static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe)
2449 if (stripe >= 0 && stripe < rbio->nr_data)
2455 * While we're doing the parity check and repair, we could have errors
2456 * in reading pages off the disk. This checks for errors and if we're
2457 * not able to read the page it'll trigger parity reconstruction. The
2458 * parity scrub will be finished after we've reconstructed the failed
2461 static void validate_rbio_for_parity_scrub(struct btrfs_raid_bio *rbio)
2463 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
2466 if (rbio->faila >= 0 || rbio->failb >= 0) {
2467 int dfail = 0, failp = -1;
2469 if (is_data_stripe(rbio, rbio->faila))
2471 else if (is_parity_stripe(rbio->faila))
2472 failp = rbio->faila;
2474 if (is_data_stripe(rbio, rbio->failb))
2476 else if (is_parity_stripe(rbio->failb))
2477 failp = rbio->failb;
2480 * Because we can not use a scrubbing parity to repair
2481 * the data, so the capability of the repair is declined.
2482 * (In the case of RAID5, we can not repair anything)
2484 if (dfail > rbio->bbio->max_errors - 1)
2488 * If all data is good, only parity is correctly, just
2489 * repair the parity.
2492 finish_parity_scrub(rbio, 0);
2497 * Here means we got one corrupted data stripe and one
2498 * corrupted parity on RAID6, if the corrupted parity
2499 * is scrubbing parity, luckily, use the other one to repair
2500 * the data, or we can not repair the data stripe.
2502 if (failp != rbio->scrubp)
2505 __raid_recover_end_io(rbio);
2507 finish_parity_scrub(rbio, 1);
2512 rbio_orig_end_io(rbio, -EIO);
2516 * end io for the read phase of the rmw cycle. All the bios here are physical
2517 * stripe bios we've read from the disk so we can recalculate the parity of the
2520 * This will usually kick off finish_rmw once all the bios are read in, but it
2521 * may trigger parity reconstruction if we had any errors along the way
2523 static void raid56_parity_scrub_end_io(struct bio *bio)
2525 struct btrfs_raid_bio *rbio = bio->bi_private;
2528 fail_bio_stripe(rbio, bio);
2530 set_bio_pages_uptodate(bio);
2534 if (!atomic_dec_and_test(&rbio->stripes_pending))
2538 * this will normally call finish_rmw to start our write
2539 * but if there are any failed stripes we'll reconstruct
2542 validate_rbio_for_parity_scrub(rbio);
2545 static void raid56_parity_scrub_stripe(struct btrfs_raid_bio *rbio)
2547 int bios_to_read = 0;
2548 struct bio_list bio_list;
2554 ret = alloc_rbio_essential_pages(rbio);
2558 bio_list_init(&bio_list);
2560 atomic_set(&rbio->error, 0);
2562 * build a list of bios to read all the missing parts of this
2565 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
2566 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2569 * we want to find all the pages missing from
2570 * the rbio and read them from the disk. If
2571 * page_in_rbio finds a page in the bio list
2572 * we don't need to read it off the stripe.
2574 page = page_in_rbio(rbio, stripe, pagenr, 1);
2578 page = rbio_stripe_page(rbio, stripe, pagenr);
2580 * the bio cache may have handed us an uptodate
2581 * page. If so, be happy and use it
2583 if (PageUptodate(page))
2586 ret = rbio_add_io_page(rbio, &bio_list, page,
2587 stripe, pagenr, rbio->stripe_len);
2593 bios_to_read = bio_list_size(&bio_list);
2594 if (!bios_to_read) {
2596 * this can happen if others have merged with
2597 * us, it means there is nothing left to read.
2598 * But if there are missing devices it may not be
2599 * safe to do the full stripe write yet.
2605 * the bbio may be freed once we submit the last bio. Make sure
2606 * not to touch it after that
2608 atomic_set(&rbio->stripes_pending, bios_to_read);
2610 bio = bio_list_pop(&bio_list);
2614 bio->bi_private = rbio;
2615 bio->bi_end_io = raid56_parity_scrub_end_io;
2616 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2618 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
2622 /* the actual write will happen once the reads are done */
2626 rbio_orig_end_io(rbio, -EIO);
2630 validate_rbio_for_parity_scrub(rbio);
2633 static void scrub_parity_work(struct btrfs_work *work)
2635 struct btrfs_raid_bio *rbio;
2637 rbio = container_of(work, struct btrfs_raid_bio, work);
2638 raid56_parity_scrub_stripe(rbio);
2641 static void async_scrub_parity(struct btrfs_raid_bio *rbio)
2643 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
2644 scrub_parity_work, NULL, NULL);
2646 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
2649 void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio)
2651 if (!lock_stripe_add(rbio))
2652 async_scrub_parity(rbio);
2655 /* The following code is used for dev replace of a missing RAID 5/6 device. */
2657 struct btrfs_raid_bio *
2658 raid56_alloc_missing_rbio(struct btrfs_fs_info *fs_info, struct bio *bio,
2659 struct btrfs_bio *bbio, u64 length)
2661 struct btrfs_raid_bio *rbio;
2663 rbio = alloc_rbio(fs_info, bbio, length);
2667 rbio->operation = BTRFS_RBIO_REBUILD_MISSING;
2668 bio_list_add(&rbio->bio_list, bio);
2670 * This is a special bio which is used to hold the completion handler
2671 * and make the scrub rbio is similar to the other types
2673 ASSERT(!bio->bi_iter.bi_size);
2675 rbio->faila = find_logical_bio_stripe(rbio, bio);
2676 if (rbio->faila == -1) {
2683 * When we get bbio, we have already increased bio_counter, record it
2684 * so we can free it at rbio_orig_end_io()
2686 rbio->generic_bio_cnt = 1;
2691 static void missing_raid56_work(struct btrfs_work *work)
2693 struct btrfs_raid_bio *rbio;
2695 rbio = container_of(work, struct btrfs_raid_bio, work);
2696 __raid56_parity_recover(rbio);
2699 static void async_missing_raid56(struct btrfs_raid_bio *rbio)
2701 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
2702 missing_raid56_work, NULL, NULL);
2704 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
2707 void raid56_submit_missing_rbio(struct btrfs_raid_bio *rbio)
2709 if (!lock_stripe_add(rbio))
2710 async_missing_raid56(rbio);