2 * Copyright (C) 2012 Fusion-io All rights reserved.
3 * Copyright (C) 2012 Intel Corp. All rights reserved.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public
7 * License v2 as published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * General Public License for more details.
14 * You should have received a copy of the GNU General Public
15 * License along with this program; if not, write to the
16 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
17 * Boston, MA 021110-1307, USA.
19 #include <linux/sched.h>
20 #include <linux/wait.h>
21 #include <linux/bio.h>
22 #include <linux/slab.h>
23 #include <linux/buffer_head.h>
24 #include <linux/blkdev.h>
25 #include <linux/random.h>
26 #include <linux/iocontext.h>
27 #include <linux/capability.h>
28 #include <linux/ratelimit.h>
29 #include <linux/kthread.h>
30 #include <linux/raid/pq.h>
31 #include <linux/hash.h>
32 #include <linux/list_sort.h>
33 #include <linux/raid/xor.h>
34 #include <linux/vmalloc.h>
35 #include <asm/div64.h>
37 #include "extent_map.h"
39 #include "transaction.h"
40 #include "print-tree.h"
43 #include "async-thread.h"
44 #include "check-integrity.h"
45 #include "rcu-string.h"
47 /* set when additional merges to this rbio are not allowed */
48 #define RBIO_RMW_LOCKED_BIT 1
51 * set when this rbio is sitting in the hash, but it is just a cache
54 #define RBIO_CACHE_BIT 2
57 * set when it is safe to trust the stripe_pages for caching
59 #define RBIO_CACHE_READY_BIT 3
61 #define RBIO_CACHE_SIZE 1024
65 BTRFS_RBIO_READ_REBUILD = 1,
66 BTRFS_RBIO_PARITY_SCRUB = 2,
69 struct btrfs_raid_bio {
70 struct btrfs_fs_info *fs_info;
71 struct btrfs_bio *bbio;
73 /* while we're doing rmw on a stripe
74 * we put it into a hash table so we can
75 * lock the stripe and merge more rbios
78 struct list_head hash_list;
81 * LRU list for the stripe cache
83 struct list_head stripe_cache;
86 * for scheduling work in the helper threads
88 struct btrfs_work work;
91 * bio list and bio_list_lock are used
92 * to add more bios into the stripe
93 * in hopes of avoiding the full rmw
95 struct bio_list bio_list;
96 spinlock_t bio_list_lock;
98 /* also protected by the bio_list_lock, the
99 * plug list is used by the plugging code
100 * to collect partial bios while plugged. The
101 * stripe locking code also uses it to hand off
102 * the stripe lock to the next pending IO
104 struct list_head plug_list;
107 * flags that tell us if it is safe to
108 * merge with this bio
112 /* size of each individual stripe on disk */
115 /* number of data stripes (no p/q) */
122 * set if we're doing a parity rebuild
123 * for a read from higher up, which is handled
124 * differently from a parity rebuild as part of
127 enum btrfs_rbio_ops operation;
129 /* first bad stripe */
132 /* second bad stripe (for raid6 use) */
137 * number of pages needed to represent the full
143 * size of all the bios in the bio_list. This
144 * helps us decide if the rbio maps to a full
153 atomic_t stripes_pending;
157 * these are two arrays of pointers. We allocate the
158 * rbio big enough to hold them both and setup their
159 * locations when the rbio is allocated
162 /* pointers to pages that we allocated for
163 * reading/writing stripes directly from the disk (including P/Q)
165 struct page **stripe_pages;
168 * pointers to the pages in the bio_list. Stored
169 * here for faster lookup
171 struct page **bio_pages;
174 * bitmap to record which horizontal stripe has data
176 unsigned long *dbitmap;
179 static int __raid56_parity_recover(struct btrfs_raid_bio *rbio);
180 static noinline void finish_rmw(struct btrfs_raid_bio *rbio);
181 static void rmw_work(struct btrfs_work *work);
182 static void read_rebuild_work(struct btrfs_work *work);
183 static void async_rmw_stripe(struct btrfs_raid_bio *rbio);
184 static void async_read_rebuild(struct btrfs_raid_bio *rbio);
185 static int fail_bio_stripe(struct btrfs_raid_bio *rbio, struct bio *bio);
186 static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed);
187 static void __free_raid_bio(struct btrfs_raid_bio *rbio);
188 static void index_rbio_pages(struct btrfs_raid_bio *rbio);
189 static int alloc_rbio_pages(struct btrfs_raid_bio *rbio);
191 static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
193 static void async_scrub_parity(struct btrfs_raid_bio *rbio);
196 * the stripe hash table is used for locking, and to collect
197 * bios in hopes of making a full stripe
199 int btrfs_alloc_stripe_hash_table(struct btrfs_fs_info *info)
201 struct btrfs_stripe_hash_table *table;
202 struct btrfs_stripe_hash_table *x;
203 struct btrfs_stripe_hash *cur;
204 struct btrfs_stripe_hash *h;
205 int num_entries = 1 << BTRFS_STRIPE_HASH_TABLE_BITS;
209 if (info->stripe_hash_table)
213 * The table is large, starting with order 4 and can go as high as
214 * order 7 in case lock debugging is turned on.
216 * Try harder to allocate and fallback to vmalloc to lower the chance
217 * of a failing mount.
219 table_size = sizeof(*table) + sizeof(*h) * num_entries;
220 table = kzalloc(table_size, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
222 table = vzalloc(table_size);
227 spin_lock_init(&table->cache_lock);
228 INIT_LIST_HEAD(&table->stripe_cache);
232 for (i = 0; i < num_entries; i++) {
234 INIT_LIST_HEAD(&cur->hash_list);
235 spin_lock_init(&cur->lock);
236 init_waitqueue_head(&cur->wait);
239 x = cmpxchg(&info->stripe_hash_table, NULL, table);
246 * caching an rbio means to copy anything from the
247 * bio_pages array into the stripe_pages array. We
248 * use the page uptodate bit in the stripe cache array
249 * to indicate if it has valid data
251 * once the caching is done, we set the cache ready
254 static void cache_rbio_pages(struct btrfs_raid_bio *rbio)
261 ret = alloc_rbio_pages(rbio);
265 for (i = 0; i < rbio->nr_pages; i++) {
266 if (!rbio->bio_pages[i])
269 s = kmap(rbio->bio_pages[i]);
270 d = kmap(rbio->stripe_pages[i]);
272 memcpy(d, s, PAGE_CACHE_SIZE);
274 kunmap(rbio->bio_pages[i]);
275 kunmap(rbio->stripe_pages[i]);
276 SetPageUptodate(rbio->stripe_pages[i]);
278 set_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
282 * we hash on the first logical address of the stripe
284 static int rbio_bucket(struct btrfs_raid_bio *rbio)
286 u64 num = rbio->bbio->raid_map[0];
289 * we shift down quite a bit. We're using byte
290 * addressing, and most of the lower bits are zeros.
291 * This tends to upset hash_64, and it consistently
292 * returns just one or two different values.
294 * shifting off the lower bits fixes things.
296 return hash_64(num >> 16, BTRFS_STRIPE_HASH_TABLE_BITS);
300 * stealing an rbio means taking all the uptodate pages from the stripe
301 * array in the source rbio and putting them into the destination rbio
303 static void steal_rbio(struct btrfs_raid_bio *src, struct btrfs_raid_bio *dest)
309 if (!test_bit(RBIO_CACHE_READY_BIT, &src->flags))
312 for (i = 0; i < dest->nr_pages; i++) {
313 s = src->stripe_pages[i];
314 if (!s || !PageUptodate(s)) {
318 d = dest->stripe_pages[i];
322 dest->stripe_pages[i] = s;
323 src->stripe_pages[i] = NULL;
328 * merging means we take the bio_list from the victim and
329 * splice it into the destination. The victim should
330 * be discarded afterwards.
332 * must be called with dest->rbio_list_lock held
334 static void merge_rbio(struct btrfs_raid_bio *dest,
335 struct btrfs_raid_bio *victim)
337 bio_list_merge(&dest->bio_list, &victim->bio_list);
338 dest->bio_list_bytes += victim->bio_list_bytes;
339 dest->generic_bio_cnt += victim->generic_bio_cnt;
340 bio_list_init(&victim->bio_list);
344 * used to prune items that are in the cache. The caller
345 * must hold the hash table lock.
347 static void __remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
349 int bucket = rbio_bucket(rbio);
350 struct btrfs_stripe_hash_table *table;
351 struct btrfs_stripe_hash *h;
355 * check the bit again under the hash table lock.
357 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
360 table = rbio->fs_info->stripe_hash_table;
361 h = table->table + bucket;
363 /* hold the lock for the bucket because we may be
364 * removing it from the hash table
369 * hold the lock for the bio list because we need
370 * to make sure the bio list is empty
372 spin_lock(&rbio->bio_list_lock);
374 if (test_and_clear_bit(RBIO_CACHE_BIT, &rbio->flags)) {
375 list_del_init(&rbio->stripe_cache);
376 table->cache_size -= 1;
379 /* if the bio list isn't empty, this rbio is
380 * still involved in an IO. We take it out
381 * of the cache list, and drop the ref that
382 * was held for the list.
384 * If the bio_list was empty, we also remove
385 * the rbio from the hash_table, and drop
386 * the corresponding ref
388 if (bio_list_empty(&rbio->bio_list)) {
389 if (!list_empty(&rbio->hash_list)) {
390 list_del_init(&rbio->hash_list);
391 atomic_dec(&rbio->refs);
392 BUG_ON(!list_empty(&rbio->plug_list));
397 spin_unlock(&rbio->bio_list_lock);
398 spin_unlock(&h->lock);
401 __free_raid_bio(rbio);
405 * prune a given rbio from the cache
407 static void remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
409 struct btrfs_stripe_hash_table *table;
412 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
415 table = rbio->fs_info->stripe_hash_table;
417 spin_lock_irqsave(&table->cache_lock, flags);
418 __remove_rbio_from_cache(rbio);
419 spin_unlock_irqrestore(&table->cache_lock, flags);
423 * remove everything in the cache
425 static void btrfs_clear_rbio_cache(struct btrfs_fs_info *info)
427 struct btrfs_stripe_hash_table *table;
429 struct btrfs_raid_bio *rbio;
431 table = info->stripe_hash_table;
433 spin_lock_irqsave(&table->cache_lock, flags);
434 while (!list_empty(&table->stripe_cache)) {
435 rbio = list_entry(table->stripe_cache.next,
436 struct btrfs_raid_bio,
438 __remove_rbio_from_cache(rbio);
440 spin_unlock_irqrestore(&table->cache_lock, flags);
444 * remove all cached entries and free the hash table
447 void btrfs_free_stripe_hash_table(struct btrfs_fs_info *info)
449 if (!info->stripe_hash_table)
451 btrfs_clear_rbio_cache(info);
452 kvfree(info->stripe_hash_table);
453 info->stripe_hash_table = NULL;
457 * insert an rbio into the stripe cache. It
458 * must have already been prepared by calling
461 * If this rbio was already cached, it gets
462 * moved to the front of the lru.
464 * If the size of the rbio cache is too big, we
467 static void cache_rbio(struct btrfs_raid_bio *rbio)
469 struct btrfs_stripe_hash_table *table;
472 if (!test_bit(RBIO_CACHE_READY_BIT, &rbio->flags))
475 table = rbio->fs_info->stripe_hash_table;
477 spin_lock_irqsave(&table->cache_lock, flags);
478 spin_lock(&rbio->bio_list_lock);
480 /* bump our ref if we were not in the list before */
481 if (!test_and_set_bit(RBIO_CACHE_BIT, &rbio->flags))
482 atomic_inc(&rbio->refs);
484 if (!list_empty(&rbio->stripe_cache)){
485 list_move(&rbio->stripe_cache, &table->stripe_cache);
487 list_add(&rbio->stripe_cache, &table->stripe_cache);
488 table->cache_size += 1;
491 spin_unlock(&rbio->bio_list_lock);
493 if (table->cache_size > RBIO_CACHE_SIZE) {
494 struct btrfs_raid_bio *found;
496 found = list_entry(table->stripe_cache.prev,
497 struct btrfs_raid_bio,
501 __remove_rbio_from_cache(found);
504 spin_unlock_irqrestore(&table->cache_lock, flags);
509 * helper function to run the xor_blocks api. It is only
510 * able to do MAX_XOR_BLOCKS at a time, so we need to
513 static void run_xor(void **pages, int src_cnt, ssize_t len)
517 void *dest = pages[src_cnt];
520 xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS);
521 xor_blocks(xor_src_cnt, len, dest, pages + src_off);
523 src_cnt -= xor_src_cnt;
524 src_off += xor_src_cnt;
529 * returns true if the bio list inside this rbio
530 * covers an entire stripe (no rmw required).
531 * Must be called with the bio list lock held, or
532 * at a time when you know it is impossible to add
533 * new bios into the list
535 static int __rbio_is_full(struct btrfs_raid_bio *rbio)
537 unsigned long size = rbio->bio_list_bytes;
540 if (size != rbio->nr_data * rbio->stripe_len)
543 BUG_ON(size > rbio->nr_data * rbio->stripe_len);
547 static int rbio_is_full(struct btrfs_raid_bio *rbio)
552 spin_lock_irqsave(&rbio->bio_list_lock, flags);
553 ret = __rbio_is_full(rbio);
554 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
559 * returns 1 if it is safe to merge two rbios together.
560 * The merging is safe if the two rbios correspond to
561 * the same stripe and if they are both going in the same
562 * direction (read vs write), and if neither one is
563 * locked for final IO
565 * The caller is responsible for locking such that
566 * rmw_locked is safe to test
568 static int rbio_can_merge(struct btrfs_raid_bio *last,
569 struct btrfs_raid_bio *cur)
571 if (test_bit(RBIO_RMW_LOCKED_BIT, &last->flags) ||
572 test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags))
576 * we can't merge with cached rbios, since the
577 * idea is that when we merge the destination
578 * rbio is going to run our IO for us. We can
579 * steal from cached rbio's though, other functions
582 if (test_bit(RBIO_CACHE_BIT, &last->flags) ||
583 test_bit(RBIO_CACHE_BIT, &cur->flags))
586 if (last->bbio->raid_map[0] !=
587 cur->bbio->raid_map[0])
590 /* we can't merge with different operations */
591 if (last->operation != cur->operation)
594 * We've need read the full stripe from the drive.
595 * check and repair the parity and write the new results.
597 * We're not allowed to add any new bios to the
598 * bio list here, anyone else that wants to
599 * change this stripe needs to do their own rmw.
601 if (last->operation == BTRFS_RBIO_PARITY_SCRUB ||
602 cur->operation == BTRFS_RBIO_PARITY_SCRUB)
609 * helper to index into the pstripe
611 static struct page *rbio_pstripe_page(struct btrfs_raid_bio *rbio, int index)
613 index += (rbio->nr_data * rbio->stripe_len) >> PAGE_CACHE_SHIFT;
614 return rbio->stripe_pages[index];
618 * helper to index into the qstripe, returns null
619 * if there is no qstripe
621 static struct page *rbio_qstripe_page(struct btrfs_raid_bio *rbio, int index)
623 if (rbio->nr_data + 1 == rbio->real_stripes)
626 index += ((rbio->nr_data + 1) * rbio->stripe_len) >>
628 return rbio->stripe_pages[index];
632 * The first stripe in the table for a logical address
633 * has the lock. rbios are added in one of three ways:
635 * 1) Nobody has the stripe locked yet. The rbio is given
636 * the lock and 0 is returned. The caller must start the IO
639 * 2) Someone has the stripe locked, but we're able to merge
640 * with the lock owner. The rbio is freed and the IO will
641 * start automatically along with the existing rbio. 1 is returned.
643 * 3) Someone has the stripe locked, but we're not able to merge.
644 * The rbio is added to the lock owner's plug list, or merged into
645 * an rbio already on the plug list. When the lock owner unlocks,
646 * the next rbio on the list is run and the IO is started automatically.
649 * If we return 0, the caller still owns the rbio and must continue with
650 * IO submission. If we return 1, the caller must assume the rbio has
651 * already been freed.
653 static noinline int lock_stripe_add(struct btrfs_raid_bio *rbio)
655 int bucket = rbio_bucket(rbio);
656 struct btrfs_stripe_hash *h = rbio->fs_info->stripe_hash_table->table + bucket;
657 struct btrfs_raid_bio *cur;
658 struct btrfs_raid_bio *pending;
661 struct btrfs_raid_bio *freeit = NULL;
662 struct btrfs_raid_bio *cache_drop = NULL;
666 spin_lock_irqsave(&h->lock, flags);
667 list_for_each_entry(cur, &h->hash_list, hash_list) {
669 if (cur->bbio->raid_map[0] == rbio->bbio->raid_map[0]) {
670 spin_lock(&cur->bio_list_lock);
672 /* can we steal this cached rbio's pages? */
673 if (bio_list_empty(&cur->bio_list) &&
674 list_empty(&cur->plug_list) &&
675 test_bit(RBIO_CACHE_BIT, &cur->flags) &&
676 !test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) {
677 list_del_init(&cur->hash_list);
678 atomic_dec(&cur->refs);
680 steal_rbio(cur, rbio);
682 spin_unlock(&cur->bio_list_lock);
687 /* can we merge into the lock owner? */
688 if (rbio_can_merge(cur, rbio)) {
689 merge_rbio(cur, rbio);
690 spin_unlock(&cur->bio_list_lock);
698 * we couldn't merge with the running
699 * rbio, see if we can merge with the
700 * pending ones. We don't have to
701 * check for rmw_locked because there
702 * is no way they are inside finish_rmw
705 list_for_each_entry(pending, &cur->plug_list,
707 if (rbio_can_merge(pending, rbio)) {
708 merge_rbio(pending, rbio);
709 spin_unlock(&cur->bio_list_lock);
716 /* no merging, put us on the tail of the plug list,
717 * our rbio will be started with the currently
718 * running rbio unlocks
720 list_add_tail(&rbio->plug_list, &cur->plug_list);
721 spin_unlock(&cur->bio_list_lock);
727 atomic_inc(&rbio->refs);
728 list_add(&rbio->hash_list, &h->hash_list);
730 spin_unlock_irqrestore(&h->lock, flags);
732 remove_rbio_from_cache(cache_drop);
734 __free_raid_bio(freeit);
739 * called as rmw or parity rebuild is completed. If the plug list has more
740 * rbios waiting for this stripe, the next one on the list will be started
742 static noinline void unlock_stripe(struct btrfs_raid_bio *rbio)
745 struct btrfs_stripe_hash *h;
749 bucket = rbio_bucket(rbio);
750 h = rbio->fs_info->stripe_hash_table->table + bucket;
752 if (list_empty(&rbio->plug_list))
755 spin_lock_irqsave(&h->lock, flags);
756 spin_lock(&rbio->bio_list_lock);
758 if (!list_empty(&rbio->hash_list)) {
760 * if we're still cached and there is no other IO
761 * to perform, just leave this rbio here for others
762 * to steal from later
764 if (list_empty(&rbio->plug_list) &&
765 test_bit(RBIO_CACHE_BIT, &rbio->flags)) {
767 clear_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
768 BUG_ON(!bio_list_empty(&rbio->bio_list));
772 list_del_init(&rbio->hash_list);
773 atomic_dec(&rbio->refs);
776 * we use the plug list to hold all the rbios
777 * waiting for the chance to lock this stripe.
778 * hand the lock over to one of them.
780 if (!list_empty(&rbio->plug_list)) {
781 struct btrfs_raid_bio *next;
782 struct list_head *head = rbio->plug_list.next;
784 next = list_entry(head, struct btrfs_raid_bio,
787 list_del_init(&rbio->plug_list);
789 list_add(&next->hash_list, &h->hash_list);
790 atomic_inc(&next->refs);
791 spin_unlock(&rbio->bio_list_lock);
792 spin_unlock_irqrestore(&h->lock, flags);
794 if (next->operation == BTRFS_RBIO_READ_REBUILD)
795 async_read_rebuild(next);
796 else if (next->operation == BTRFS_RBIO_WRITE) {
797 steal_rbio(rbio, next);
798 async_rmw_stripe(next);
799 } else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) {
800 steal_rbio(rbio, next);
801 async_scrub_parity(next);
805 } else if (waitqueue_active(&h->wait)) {
806 spin_unlock(&rbio->bio_list_lock);
807 spin_unlock_irqrestore(&h->lock, flags);
813 spin_unlock(&rbio->bio_list_lock);
814 spin_unlock_irqrestore(&h->lock, flags);
818 remove_rbio_from_cache(rbio);
821 static void __free_raid_bio(struct btrfs_raid_bio *rbio)
825 WARN_ON(atomic_read(&rbio->refs) < 0);
826 if (!atomic_dec_and_test(&rbio->refs))
829 WARN_ON(!list_empty(&rbio->stripe_cache));
830 WARN_ON(!list_empty(&rbio->hash_list));
831 WARN_ON(!bio_list_empty(&rbio->bio_list));
833 for (i = 0; i < rbio->nr_pages; i++) {
834 if (rbio->stripe_pages[i]) {
835 __free_page(rbio->stripe_pages[i]);
836 rbio->stripe_pages[i] = NULL;
840 btrfs_put_bbio(rbio->bbio);
844 static void free_raid_bio(struct btrfs_raid_bio *rbio)
847 __free_raid_bio(rbio);
851 * this frees the rbio and runs through all the bios in the
852 * bio_list and calls end_io on them
854 static void rbio_orig_end_io(struct btrfs_raid_bio *rbio, int err)
856 struct bio *cur = bio_list_get(&rbio->bio_list);
859 if (rbio->generic_bio_cnt)
860 btrfs_bio_counter_sub(rbio->fs_info, rbio->generic_bio_cnt);
874 * end io function used by finish_rmw. When we finally
875 * get here, we've written a full stripe
877 static void raid_write_end_io(struct bio *bio)
879 struct btrfs_raid_bio *rbio = bio->bi_private;
880 int err = bio->bi_error;
883 fail_bio_stripe(rbio, bio);
887 if (!atomic_dec_and_test(&rbio->stripes_pending))
892 /* OK, we have read all the stripes we need to. */
893 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
896 rbio_orig_end_io(rbio, err);
901 * the read/modify/write code wants to use the original bio for
902 * any pages it included, and then use the rbio for everything
903 * else. This function decides if a given index (stripe number)
904 * and page number in that stripe fall inside the original bio
907 * if you set bio_list_only, you'll get a NULL back for any ranges
908 * that are outside the bio_list
910 * This doesn't take any refs on anything, you get a bare page pointer
911 * and the caller must bump refs as required.
913 * You must call index_rbio_pages once before you can trust
914 * the answers from this function.
916 static struct page *page_in_rbio(struct btrfs_raid_bio *rbio,
917 int index, int pagenr, int bio_list_only)
920 struct page *p = NULL;
922 chunk_page = index * (rbio->stripe_len >> PAGE_SHIFT) + pagenr;
924 spin_lock_irq(&rbio->bio_list_lock);
925 p = rbio->bio_pages[chunk_page];
926 spin_unlock_irq(&rbio->bio_list_lock);
928 if (p || bio_list_only)
931 return rbio->stripe_pages[chunk_page];
935 * number of pages we need for the entire stripe across all the
938 static unsigned long rbio_nr_pages(unsigned long stripe_len, int nr_stripes)
940 unsigned long nr = stripe_len * nr_stripes;
941 return DIV_ROUND_UP(nr, PAGE_CACHE_SIZE);
945 * allocation and initial setup for the btrfs_raid_bio. Not
946 * this does not allocate any pages for rbio->pages.
948 static struct btrfs_raid_bio *alloc_rbio(struct btrfs_root *root,
949 struct btrfs_bio *bbio, u64 stripe_len)
951 struct btrfs_raid_bio *rbio;
953 int real_stripes = bbio->num_stripes - bbio->num_tgtdevs;
954 int num_pages = rbio_nr_pages(stripe_len, real_stripes);
955 int stripe_npages = DIV_ROUND_UP(stripe_len, PAGE_SIZE);
958 rbio = kzalloc(sizeof(*rbio) + num_pages * sizeof(struct page *) * 2 +
959 DIV_ROUND_UP(stripe_npages, BITS_PER_LONG / 8),
962 return ERR_PTR(-ENOMEM);
964 bio_list_init(&rbio->bio_list);
965 INIT_LIST_HEAD(&rbio->plug_list);
966 spin_lock_init(&rbio->bio_list_lock);
967 INIT_LIST_HEAD(&rbio->stripe_cache);
968 INIT_LIST_HEAD(&rbio->hash_list);
970 rbio->fs_info = root->fs_info;
971 rbio->stripe_len = stripe_len;
972 rbio->nr_pages = num_pages;
973 rbio->real_stripes = real_stripes;
974 rbio->stripe_npages = stripe_npages;
977 atomic_set(&rbio->refs, 1);
978 atomic_set(&rbio->error, 0);
979 atomic_set(&rbio->stripes_pending, 0);
982 * the stripe_pages and bio_pages array point to the extra
983 * memory we allocated past the end of the rbio
986 rbio->stripe_pages = p;
987 rbio->bio_pages = p + sizeof(struct page *) * num_pages;
988 rbio->dbitmap = p + sizeof(struct page *) * num_pages * 2;
990 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
991 nr_data = real_stripes - 1;
992 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
993 nr_data = real_stripes - 2;
997 rbio->nr_data = nr_data;
1001 /* allocate pages for all the stripes in the bio, including parity */
1002 static int alloc_rbio_pages(struct btrfs_raid_bio *rbio)
1007 for (i = 0; i < rbio->nr_pages; i++) {
1008 if (rbio->stripe_pages[i])
1010 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1013 rbio->stripe_pages[i] = page;
1014 ClearPageUptodate(page);
1019 /* allocate pages for just the p/q stripes */
1020 static int alloc_rbio_parity_pages(struct btrfs_raid_bio *rbio)
1025 i = (rbio->nr_data * rbio->stripe_len) >> PAGE_CACHE_SHIFT;
1027 for (; i < rbio->nr_pages; i++) {
1028 if (rbio->stripe_pages[i])
1030 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1033 rbio->stripe_pages[i] = page;
1039 * add a single page from a specific stripe into our list of bios for IO
1040 * this will try to merge into existing bios if possible, and returns
1041 * zero if all went well.
1043 static int rbio_add_io_page(struct btrfs_raid_bio *rbio,
1044 struct bio_list *bio_list,
1047 unsigned long page_index,
1048 unsigned long bio_max_len)
1050 struct bio *last = bio_list->tail;
1054 struct btrfs_bio_stripe *stripe;
1057 stripe = &rbio->bbio->stripes[stripe_nr];
1058 disk_start = stripe->physical + (page_index << PAGE_CACHE_SHIFT);
1060 /* if the device is missing, just fail this stripe */
1061 if (!stripe->dev->bdev)
1062 return fail_rbio_index(rbio, stripe_nr);
1064 /* see if we can add this page onto our existing bio */
1066 last_end = (u64)last->bi_iter.bi_sector << 9;
1067 last_end += last->bi_iter.bi_size;
1070 * we can't merge these if they are from different
1071 * devices or if they are not contiguous
1073 if (last_end == disk_start && stripe->dev->bdev &&
1075 last->bi_bdev == stripe->dev->bdev) {
1076 ret = bio_add_page(last, page, PAGE_CACHE_SIZE, 0);
1077 if (ret == PAGE_CACHE_SIZE)
1082 /* put a new bio on the list */
1083 bio = btrfs_io_bio_alloc(GFP_NOFS, bio_max_len >> PAGE_SHIFT?:1);
1087 bio->bi_iter.bi_size = 0;
1088 bio->bi_bdev = stripe->dev->bdev;
1089 bio->bi_iter.bi_sector = disk_start >> 9;
1091 bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
1092 bio_list_add(bio_list, bio);
1097 * while we're doing the read/modify/write cycle, we could
1098 * have errors in reading pages off the disk. This checks
1099 * for errors and if we're not able to read the page it'll
1100 * trigger parity reconstruction. The rmw will be finished
1101 * after we've reconstructed the failed stripes
1103 static void validate_rbio_for_rmw(struct btrfs_raid_bio *rbio)
1105 if (rbio->faila >= 0 || rbio->failb >= 0) {
1106 BUG_ON(rbio->faila == rbio->real_stripes - 1);
1107 __raid56_parity_recover(rbio);
1114 * these are just the pages from the rbio array, not from anything
1115 * the FS sent down to us
1117 static struct page *rbio_stripe_page(struct btrfs_raid_bio *rbio, int stripe, int page)
1120 index = stripe * (rbio->stripe_len >> PAGE_CACHE_SHIFT);
1122 return rbio->stripe_pages[index];
1126 * helper function to walk our bio list and populate the bio_pages array with
1127 * the result. This seems expensive, but it is faster than constantly
1128 * searching through the bio list as we setup the IO in finish_rmw or stripe
1131 * This must be called before you trust the answers from page_in_rbio
1133 static void index_rbio_pages(struct btrfs_raid_bio *rbio)
1137 unsigned long stripe_offset;
1138 unsigned long page_index;
1142 spin_lock_irq(&rbio->bio_list_lock);
1143 bio_list_for_each(bio, &rbio->bio_list) {
1144 start = (u64)bio->bi_iter.bi_sector << 9;
1145 stripe_offset = start - rbio->bbio->raid_map[0];
1146 page_index = stripe_offset >> PAGE_CACHE_SHIFT;
1148 for (i = 0; i < bio->bi_vcnt; i++) {
1149 p = bio->bi_io_vec[i].bv_page;
1150 rbio->bio_pages[page_index + i] = p;
1153 spin_unlock_irq(&rbio->bio_list_lock);
1157 * this is called from one of two situations. We either
1158 * have a full stripe from the higher layers, or we've read all
1159 * the missing bits off disk.
1161 * This will calculate the parity and then send down any
1164 static noinline void finish_rmw(struct btrfs_raid_bio *rbio)
1166 struct btrfs_bio *bbio = rbio->bbio;
1167 void *pointers[rbio->real_stripes];
1168 int stripe_len = rbio->stripe_len;
1169 int nr_data = rbio->nr_data;
1174 struct bio_list bio_list;
1176 int pages_per_stripe = stripe_len >> PAGE_CACHE_SHIFT;
1179 bio_list_init(&bio_list);
1181 if (rbio->real_stripes - rbio->nr_data == 1) {
1182 p_stripe = rbio->real_stripes - 1;
1183 } else if (rbio->real_stripes - rbio->nr_data == 2) {
1184 p_stripe = rbio->real_stripes - 2;
1185 q_stripe = rbio->real_stripes - 1;
1190 /* at this point we either have a full stripe,
1191 * or we've read the full stripe from the drive.
1192 * recalculate the parity and write the new results.
1194 * We're not allowed to add any new bios to the
1195 * bio list here, anyone else that wants to
1196 * change this stripe needs to do their own rmw.
1198 spin_lock_irq(&rbio->bio_list_lock);
1199 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1200 spin_unlock_irq(&rbio->bio_list_lock);
1202 atomic_set(&rbio->error, 0);
1205 * now that we've set rmw_locked, run through the
1206 * bio list one last time and map the page pointers
1208 * We don't cache full rbios because we're assuming
1209 * the higher layers are unlikely to use this area of
1210 * the disk again soon. If they do use it again,
1211 * hopefully they will send another full bio.
1213 index_rbio_pages(rbio);
1214 if (!rbio_is_full(rbio))
1215 cache_rbio_pages(rbio);
1217 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
1219 for (pagenr = 0; pagenr < pages_per_stripe; pagenr++) {
1221 /* first collect one page from each data stripe */
1222 for (stripe = 0; stripe < nr_data; stripe++) {
1223 p = page_in_rbio(rbio, stripe, pagenr, 0);
1224 pointers[stripe] = kmap(p);
1227 /* then add the parity stripe */
1228 p = rbio_pstripe_page(rbio, pagenr);
1230 pointers[stripe++] = kmap(p);
1232 if (q_stripe != -1) {
1235 * raid6, add the qstripe and call the
1236 * library function to fill in our p/q
1238 p = rbio_qstripe_page(rbio, pagenr);
1240 pointers[stripe++] = kmap(p);
1242 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
1246 memcpy(pointers[nr_data], pointers[0], PAGE_SIZE);
1247 run_xor(pointers + 1, nr_data - 1, PAGE_CACHE_SIZE);
1251 for (stripe = 0; stripe < rbio->real_stripes; stripe++)
1252 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
1256 * time to start writing. Make bios for everything from the
1257 * higher layers (the bio_list in our rbio) and our p/q. Ignore
1260 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1261 for (pagenr = 0; pagenr < pages_per_stripe; pagenr++) {
1263 if (stripe < rbio->nr_data) {
1264 page = page_in_rbio(rbio, stripe, pagenr, 1);
1268 page = rbio_stripe_page(rbio, stripe, pagenr);
1271 ret = rbio_add_io_page(rbio, &bio_list,
1272 page, stripe, pagenr, rbio->stripe_len);
1278 if (likely(!bbio->num_tgtdevs))
1281 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1282 if (!bbio->tgtdev_map[stripe])
1285 for (pagenr = 0; pagenr < pages_per_stripe; pagenr++) {
1287 if (stripe < rbio->nr_data) {
1288 page = page_in_rbio(rbio, stripe, pagenr, 1);
1292 page = rbio_stripe_page(rbio, stripe, pagenr);
1295 ret = rbio_add_io_page(rbio, &bio_list, page,
1296 rbio->bbio->tgtdev_map[stripe],
1297 pagenr, rbio->stripe_len);
1304 atomic_set(&rbio->stripes_pending, bio_list_size(&bio_list));
1305 BUG_ON(atomic_read(&rbio->stripes_pending) == 0);
1308 bio = bio_list_pop(&bio_list);
1312 bio->bi_private = rbio;
1313 bio->bi_end_io = raid_write_end_io;
1314 submit_bio(WRITE, bio);
1319 rbio_orig_end_io(rbio, -EIO);
1323 * helper to find the stripe number for a given bio. Used to figure out which
1324 * stripe has failed. This expects the bio to correspond to a physical disk,
1325 * so it looks up based on physical sector numbers.
1327 static int find_bio_stripe(struct btrfs_raid_bio *rbio,
1330 u64 physical = bio->bi_iter.bi_sector;
1333 struct btrfs_bio_stripe *stripe;
1337 for (i = 0; i < rbio->bbio->num_stripes; i++) {
1338 stripe = &rbio->bbio->stripes[i];
1339 stripe_start = stripe->physical;
1340 if (physical >= stripe_start &&
1341 physical < stripe_start + rbio->stripe_len &&
1342 bio->bi_bdev == stripe->dev->bdev) {
1350 * helper to find the stripe number for a given
1351 * bio (before mapping). Used to figure out which stripe has
1352 * failed. This looks up based on logical block numbers.
1354 static int find_logical_bio_stripe(struct btrfs_raid_bio *rbio,
1357 u64 logical = bio->bi_iter.bi_sector;
1363 for (i = 0; i < rbio->nr_data; i++) {
1364 stripe_start = rbio->bbio->raid_map[i];
1365 if (logical >= stripe_start &&
1366 logical < stripe_start + rbio->stripe_len) {
1374 * returns -EIO if we had too many failures
1376 static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed)
1378 unsigned long flags;
1381 spin_lock_irqsave(&rbio->bio_list_lock, flags);
1383 /* we already know this stripe is bad, move on */
1384 if (rbio->faila == failed || rbio->failb == failed)
1387 if (rbio->faila == -1) {
1388 /* first failure on this rbio */
1389 rbio->faila = failed;
1390 atomic_inc(&rbio->error);
1391 } else if (rbio->failb == -1) {
1392 /* second failure on this rbio */
1393 rbio->failb = failed;
1394 atomic_inc(&rbio->error);
1399 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
1405 * helper to fail a stripe based on a physical disk
1408 static int fail_bio_stripe(struct btrfs_raid_bio *rbio,
1411 int failed = find_bio_stripe(rbio, bio);
1416 return fail_rbio_index(rbio, failed);
1420 * this sets each page in the bio uptodate. It should only be used on private
1421 * rbio pages, nothing that comes in from the higher layers
1423 static void set_bio_pages_uptodate(struct bio *bio)
1428 for (i = 0; i < bio->bi_vcnt; i++) {
1429 p = bio->bi_io_vec[i].bv_page;
1435 * end io for the read phase of the rmw cycle. All the bios here are physical
1436 * stripe bios we've read from the disk so we can recalculate the parity of the
1439 * This will usually kick off finish_rmw once all the bios are read in, but it
1440 * may trigger parity reconstruction if we had any errors along the way
1442 static void raid_rmw_end_io(struct bio *bio)
1444 struct btrfs_raid_bio *rbio = bio->bi_private;
1447 fail_bio_stripe(rbio, bio);
1449 set_bio_pages_uptodate(bio);
1453 if (!atomic_dec_and_test(&rbio->stripes_pending))
1456 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
1460 * this will normally call finish_rmw to start our write
1461 * but if there are any failed stripes we'll reconstruct
1464 validate_rbio_for_rmw(rbio);
1469 rbio_orig_end_io(rbio, -EIO);
1472 static void async_rmw_stripe(struct btrfs_raid_bio *rbio)
1474 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
1475 rmw_work, NULL, NULL);
1477 btrfs_queue_work(rbio->fs_info->rmw_workers,
1481 static void async_read_rebuild(struct btrfs_raid_bio *rbio)
1483 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
1484 read_rebuild_work, NULL, NULL);
1486 btrfs_queue_work(rbio->fs_info->rmw_workers,
1491 * the stripe must be locked by the caller. It will
1492 * unlock after all the writes are done
1494 static int raid56_rmw_stripe(struct btrfs_raid_bio *rbio)
1496 int bios_to_read = 0;
1497 struct bio_list bio_list;
1499 int nr_pages = DIV_ROUND_UP(rbio->stripe_len, PAGE_CACHE_SIZE);
1504 bio_list_init(&bio_list);
1506 ret = alloc_rbio_pages(rbio);
1510 index_rbio_pages(rbio);
1512 atomic_set(&rbio->error, 0);
1514 * build a list of bios to read all the missing parts of this
1517 for (stripe = 0; stripe < rbio->nr_data; stripe++) {
1518 for (pagenr = 0; pagenr < nr_pages; pagenr++) {
1521 * we want to find all the pages missing from
1522 * the rbio and read them from the disk. If
1523 * page_in_rbio finds a page in the bio list
1524 * we don't need to read it off the stripe.
1526 page = page_in_rbio(rbio, stripe, pagenr, 1);
1530 page = rbio_stripe_page(rbio, stripe, pagenr);
1532 * the bio cache may have handed us an uptodate
1533 * page. If so, be happy and use it
1535 if (PageUptodate(page))
1538 ret = rbio_add_io_page(rbio, &bio_list, page,
1539 stripe, pagenr, rbio->stripe_len);
1545 bios_to_read = bio_list_size(&bio_list);
1546 if (!bios_to_read) {
1548 * this can happen if others have merged with
1549 * us, it means there is nothing left to read.
1550 * But if there are missing devices it may not be
1551 * safe to do the full stripe write yet.
1557 * the bbio may be freed once we submit the last bio. Make sure
1558 * not to touch it after that
1560 atomic_set(&rbio->stripes_pending, bios_to_read);
1562 bio = bio_list_pop(&bio_list);
1566 bio->bi_private = rbio;
1567 bio->bi_end_io = raid_rmw_end_io;
1569 btrfs_bio_wq_end_io(rbio->fs_info, bio,
1570 BTRFS_WQ_ENDIO_RAID56);
1572 submit_bio(READ, bio);
1574 /* the actual write will happen once the reads are done */
1578 rbio_orig_end_io(rbio, -EIO);
1582 validate_rbio_for_rmw(rbio);
1587 * if the upper layers pass in a full stripe, we thank them by only allocating
1588 * enough pages to hold the parity, and sending it all down quickly.
1590 static int full_stripe_write(struct btrfs_raid_bio *rbio)
1594 ret = alloc_rbio_parity_pages(rbio);
1596 __free_raid_bio(rbio);
1600 ret = lock_stripe_add(rbio);
1607 * partial stripe writes get handed over to async helpers.
1608 * We're really hoping to merge a few more writes into this
1609 * rbio before calculating new parity
1611 static int partial_stripe_write(struct btrfs_raid_bio *rbio)
1615 ret = lock_stripe_add(rbio);
1617 async_rmw_stripe(rbio);
1622 * sometimes while we were reading from the drive to
1623 * recalculate parity, enough new bios come into create
1624 * a full stripe. So we do a check here to see if we can
1625 * go directly to finish_rmw
1627 static int __raid56_parity_write(struct btrfs_raid_bio *rbio)
1629 /* head off into rmw land if we don't have a full stripe */
1630 if (!rbio_is_full(rbio))
1631 return partial_stripe_write(rbio);
1632 return full_stripe_write(rbio);
1636 * We use plugging call backs to collect full stripes.
1637 * Any time we get a partial stripe write while plugged
1638 * we collect it into a list. When the unplug comes down,
1639 * we sort the list by logical block number and merge
1640 * everything we can into the same rbios
1642 struct btrfs_plug_cb {
1643 struct blk_plug_cb cb;
1644 struct btrfs_fs_info *info;
1645 struct list_head rbio_list;
1646 struct btrfs_work work;
1650 * rbios on the plug list are sorted for easier merging.
1652 static int plug_cmp(void *priv, struct list_head *a, struct list_head *b)
1654 struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio,
1656 struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio,
1658 u64 a_sector = ra->bio_list.head->bi_iter.bi_sector;
1659 u64 b_sector = rb->bio_list.head->bi_iter.bi_sector;
1661 if (a_sector < b_sector)
1663 if (a_sector > b_sector)
1668 static void run_plug(struct btrfs_plug_cb *plug)
1670 struct btrfs_raid_bio *cur;
1671 struct btrfs_raid_bio *last = NULL;
1674 * sort our plug list then try to merge
1675 * everything we can in hopes of creating full
1678 list_sort(NULL, &plug->rbio_list, plug_cmp);
1679 while (!list_empty(&plug->rbio_list)) {
1680 cur = list_entry(plug->rbio_list.next,
1681 struct btrfs_raid_bio, plug_list);
1682 list_del_init(&cur->plug_list);
1684 if (rbio_is_full(cur)) {
1685 /* we have a full stripe, send it down */
1686 full_stripe_write(cur);
1690 if (rbio_can_merge(last, cur)) {
1691 merge_rbio(last, cur);
1692 __free_raid_bio(cur);
1696 __raid56_parity_write(last);
1701 __raid56_parity_write(last);
1707 * if the unplug comes from schedule, we have to push the
1708 * work off to a helper thread
1710 static void unplug_work(struct btrfs_work *work)
1712 struct btrfs_plug_cb *plug;
1713 plug = container_of(work, struct btrfs_plug_cb, work);
1717 static void btrfs_raid_unplug(struct blk_plug_cb *cb, bool from_schedule)
1719 struct btrfs_plug_cb *plug;
1720 plug = container_of(cb, struct btrfs_plug_cb, cb);
1722 if (from_schedule) {
1723 btrfs_init_work(&plug->work, btrfs_rmw_helper,
1724 unplug_work, NULL, NULL);
1725 btrfs_queue_work(plug->info->rmw_workers,
1733 * our main entry point for writes from the rest of the FS.
1735 int raid56_parity_write(struct btrfs_root *root, struct bio *bio,
1736 struct btrfs_bio *bbio, u64 stripe_len)
1738 struct btrfs_raid_bio *rbio;
1739 struct btrfs_plug_cb *plug = NULL;
1740 struct blk_plug_cb *cb;
1743 rbio = alloc_rbio(root, bbio, stripe_len);
1745 btrfs_put_bbio(bbio);
1746 return PTR_ERR(rbio);
1748 bio_list_add(&rbio->bio_list, bio);
1749 rbio->bio_list_bytes = bio->bi_iter.bi_size;
1750 rbio->operation = BTRFS_RBIO_WRITE;
1752 btrfs_bio_counter_inc_noblocked(root->fs_info);
1753 rbio->generic_bio_cnt = 1;
1756 * don't plug on full rbios, just get them out the door
1757 * as quickly as we can
1759 if (rbio_is_full(rbio)) {
1760 ret = full_stripe_write(rbio);
1762 btrfs_bio_counter_dec(root->fs_info);
1766 cb = blk_check_plugged(btrfs_raid_unplug, root->fs_info,
1769 plug = container_of(cb, struct btrfs_plug_cb, cb);
1771 plug->info = root->fs_info;
1772 INIT_LIST_HEAD(&plug->rbio_list);
1774 list_add_tail(&rbio->plug_list, &plug->rbio_list);
1777 ret = __raid56_parity_write(rbio);
1779 btrfs_bio_counter_dec(root->fs_info);
1785 * all parity reconstruction happens here. We've read in everything
1786 * we can find from the drives and this does the heavy lifting of
1787 * sorting the good from the bad.
1789 static void __raid_recover_end_io(struct btrfs_raid_bio *rbio)
1793 int faila = -1, failb = -1;
1794 int nr_pages = DIV_ROUND_UP(rbio->stripe_len, PAGE_CACHE_SIZE);
1799 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
1805 faila = rbio->faila;
1806 failb = rbio->failb;
1808 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
1809 spin_lock_irq(&rbio->bio_list_lock);
1810 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1811 spin_unlock_irq(&rbio->bio_list_lock);
1814 index_rbio_pages(rbio);
1816 for (pagenr = 0; pagenr < nr_pages; pagenr++) {
1818 * Now we just use bitmap to mark the horizontal stripes in
1819 * which we have data when doing parity scrub.
1821 if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB &&
1822 !test_bit(pagenr, rbio->dbitmap))
1825 /* setup our array of pointers with pages
1828 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1830 * if we're rebuilding a read, we have to use
1831 * pages from the bio list
1833 if (rbio->operation == BTRFS_RBIO_READ_REBUILD &&
1834 (stripe == faila || stripe == failb)) {
1835 page = page_in_rbio(rbio, stripe, pagenr, 0);
1837 page = rbio_stripe_page(rbio, stripe, pagenr);
1839 pointers[stripe] = kmap(page);
1842 /* all raid6 handling here */
1843 if (rbio->bbio->map_type & BTRFS_BLOCK_GROUP_RAID6) {
1845 * single failure, rebuild from parity raid5
1849 if (faila == rbio->nr_data) {
1851 * Just the P stripe has failed, without
1852 * a bad data or Q stripe.
1853 * TODO, we should redo the xor here.
1859 * a single failure in raid6 is rebuilt
1860 * in the pstripe code below
1865 /* make sure our ps and qs are in order */
1866 if (faila > failb) {
1872 /* if the q stripe is failed, do a pstripe reconstruction
1874 * If both the q stripe and the P stripe are failed, we're
1875 * here due to a crc mismatch and we can't give them the
1878 if (rbio->bbio->raid_map[failb] == RAID6_Q_STRIPE) {
1879 if (rbio->bbio->raid_map[faila] ==
1885 * otherwise we have one bad data stripe and
1886 * a good P stripe. raid5!
1891 if (rbio->bbio->raid_map[failb] == RAID5_P_STRIPE) {
1892 raid6_datap_recov(rbio->real_stripes,
1893 PAGE_SIZE, faila, pointers);
1895 raid6_2data_recov(rbio->real_stripes,
1896 PAGE_SIZE, faila, failb,
1902 /* rebuild from P stripe here (raid5 or raid6) */
1903 BUG_ON(failb != -1);
1905 /* Copy parity block into failed block to start with */
1906 memcpy(pointers[faila],
1907 pointers[rbio->nr_data],
1910 /* rearrange the pointer array */
1911 p = pointers[faila];
1912 for (stripe = faila; stripe < rbio->nr_data - 1; stripe++)
1913 pointers[stripe] = pointers[stripe + 1];
1914 pointers[rbio->nr_data - 1] = p;
1916 /* xor in the rest */
1917 run_xor(pointers, rbio->nr_data - 1, PAGE_CACHE_SIZE);
1919 /* if we're doing this rebuild as part of an rmw, go through
1920 * and set all of our private rbio pages in the
1921 * failed stripes as uptodate. This way finish_rmw will
1922 * know they can be trusted. If this was a read reconstruction,
1923 * other endio functions will fiddle the uptodate bits
1925 if (rbio->operation == BTRFS_RBIO_WRITE) {
1926 for (i = 0; i < nr_pages; i++) {
1928 page = rbio_stripe_page(rbio, faila, i);
1929 SetPageUptodate(page);
1932 page = rbio_stripe_page(rbio, failb, i);
1933 SetPageUptodate(page);
1937 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1939 * if we're rebuilding a read, we have to use
1940 * pages from the bio list
1942 if (rbio->operation == BTRFS_RBIO_READ_REBUILD &&
1943 (stripe == faila || stripe == failb)) {
1944 page = page_in_rbio(rbio, stripe, pagenr, 0);
1946 page = rbio_stripe_page(rbio, stripe, pagenr);
1957 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
1959 cache_rbio_pages(rbio);
1961 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
1963 rbio_orig_end_io(rbio, err);
1964 } else if (err == 0) {
1968 if (rbio->operation == BTRFS_RBIO_WRITE)
1970 else if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB)
1971 finish_parity_scrub(rbio, 0);
1975 rbio_orig_end_io(rbio, err);
1980 * This is called only for stripes we've read from disk to
1981 * reconstruct the parity.
1983 static void raid_recover_end_io(struct bio *bio)
1985 struct btrfs_raid_bio *rbio = bio->bi_private;
1988 * we only read stripe pages off the disk, set them
1989 * up to date if there were no errors
1992 fail_bio_stripe(rbio, bio);
1994 set_bio_pages_uptodate(bio);
1997 if (!atomic_dec_and_test(&rbio->stripes_pending))
2000 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
2001 rbio_orig_end_io(rbio, -EIO);
2003 __raid_recover_end_io(rbio);
2007 * reads everything we need off the disk to reconstruct
2008 * the parity. endio handlers trigger final reconstruction
2009 * when the IO is done.
2011 * This is used both for reads from the higher layers and for
2012 * parity construction required to finish a rmw cycle.
2014 static int __raid56_parity_recover(struct btrfs_raid_bio *rbio)
2016 int bios_to_read = 0;
2017 struct bio_list bio_list;
2019 int nr_pages = DIV_ROUND_UP(rbio->stripe_len, PAGE_CACHE_SIZE);
2024 bio_list_init(&bio_list);
2026 ret = alloc_rbio_pages(rbio);
2030 atomic_set(&rbio->error, 0);
2033 * read everything that hasn't failed. Thanks to the
2034 * stripe cache, it is possible that some or all of these
2035 * pages are going to be uptodate.
2037 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
2038 if (rbio->faila == stripe || rbio->failb == stripe) {
2039 atomic_inc(&rbio->error);
2043 for (pagenr = 0; pagenr < nr_pages; pagenr++) {
2047 * the rmw code may have already read this
2050 p = rbio_stripe_page(rbio, stripe, pagenr);
2051 if (PageUptodate(p))
2054 ret = rbio_add_io_page(rbio, &bio_list,
2055 rbio_stripe_page(rbio, stripe, pagenr),
2056 stripe, pagenr, rbio->stripe_len);
2062 bios_to_read = bio_list_size(&bio_list);
2063 if (!bios_to_read) {
2065 * we might have no bios to read just because the pages
2066 * were up to date, or we might have no bios to read because
2067 * the devices were gone.
2069 if (atomic_read(&rbio->error) <= rbio->bbio->max_errors) {
2070 __raid_recover_end_io(rbio);
2078 * the bbio may be freed once we submit the last bio. Make sure
2079 * not to touch it after that
2081 atomic_set(&rbio->stripes_pending, bios_to_read);
2083 bio = bio_list_pop(&bio_list);
2087 bio->bi_private = rbio;
2088 bio->bi_end_io = raid_recover_end_io;
2090 btrfs_bio_wq_end_io(rbio->fs_info, bio,
2091 BTRFS_WQ_ENDIO_RAID56);
2093 submit_bio(READ, bio);
2099 if (rbio->operation == BTRFS_RBIO_READ_REBUILD)
2100 rbio_orig_end_io(rbio, -EIO);
2105 * the main entry point for reads from the higher layers. This
2106 * is really only called when the normal read path had a failure,
2107 * so we assume the bio they send down corresponds to a failed part
2110 int raid56_parity_recover(struct btrfs_root *root, struct bio *bio,
2111 struct btrfs_bio *bbio, u64 stripe_len,
2112 int mirror_num, int generic_io)
2114 struct btrfs_raid_bio *rbio;
2117 rbio = alloc_rbio(root, bbio, stripe_len);
2120 btrfs_put_bbio(bbio);
2121 return PTR_ERR(rbio);
2124 rbio->operation = BTRFS_RBIO_READ_REBUILD;
2125 bio_list_add(&rbio->bio_list, bio);
2126 rbio->bio_list_bytes = bio->bi_iter.bi_size;
2128 rbio->faila = find_logical_bio_stripe(rbio, bio);
2129 if (rbio->faila == -1) {
2132 btrfs_put_bbio(bbio);
2138 btrfs_bio_counter_inc_noblocked(root->fs_info);
2139 rbio->generic_bio_cnt = 1;
2141 btrfs_get_bbio(bbio);
2145 * reconstruct from the q stripe if they are
2146 * asking for mirror 3
2148 if (mirror_num == 3)
2149 rbio->failb = rbio->real_stripes - 2;
2151 ret = lock_stripe_add(rbio);
2154 * __raid56_parity_recover will end the bio with
2155 * any errors it hits. We don't want to return
2156 * its error value up the stack because our caller
2157 * will end up calling bio_endio with any nonzero
2161 __raid56_parity_recover(rbio);
2163 * our rbio has been added to the list of
2164 * rbios that will be handled after the
2165 * currently lock owner is done
2171 static void rmw_work(struct btrfs_work *work)
2173 struct btrfs_raid_bio *rbio;
2175 rbio = container_of(work, struct btrfs_raid_bio, work);
2176 raid56_rmw_stripe(rbio);
2179 static void read_rebuild_work(struct btrfs_work *work)
2181 struct btrfs_raid_bio *rbio;
2183 rbio = container_of(work, struct btrfs_raid_bio, work);
2184 __raid56_parity_recover(rbio);
2188 * The following code is used to scrub/replace the parity stripe
2190 * Note: We need make sure all the pages that add into the scrub/replace
2191 * raid bio are correct and not be changed during the scrub/replace. That
2192 * is those pages just hold metadata or file data with checksum.
2195 struct btrfs_raid_bio *
2196 raid56_parity_alloc_scrub_rbio(struct btrfs_root *root, struct bio *bio,
2197 struct btrfs_bio *bbio, u64 stripe_len,
2198 struct btrfs_device *scrub_dev,
2199 unsigned long *dbitmap, int stripe_nsectors)
2201 struct btrfs_raid_bio *rbio;
2204 rbio = alloc_rbio(root, bbio, stripe_len);
2207 bio_list_add(&rbio->bio_list, bio);
2209 * This is a special bio which is used to hold the completion handler
2210 * and make the scrub rbio is similar to the other types
2212 ASSERT(!bio->bi_iter.bi_size);
2213 rbio->operation = BTRFS_RBIO_PARITY_SCRUB;
2215 for (i = 0; i < rbio->real_stripes; i++) {
2216 if (bbio->stripes[i].dev == scrub_dev) {
2222 /* Now we just support the sectorsize equals to page size */
2223 ASSERT(root->sectorsize == PAGE_SIZE);
2224 ASSERT(rbio->stripe_npages == stripe_nsectors);
2225 bitmap_copy(rbio->dbitmap, dbitmap, stripe_nsectors);
2230 void raid56_parity_add_scrub_pages(struct btrfs_raid_bio *rbio,
2231 struct page *page, u64 logical)
2236 ASSERT(logical >= rbio->bbio->raid_map[0]);
2237 ASSERT(logical + PAGE_SIZE <= rbio->bbio->raid_map[0] +
2238 rbio->stripe_len * rbio->nr_data);
2239 stripe_offset = (int)(logical - rbio->bbio->raid_map[0]);
2240 index = stripe_offset >> PAGE_CACHE_SHIFT;
2241 rbio->bio_pages[index] = page;
2245 * We just scrub the parity that we have correct data on the same horizontal,
2246 * so we needn't allocate all pages for all the stripes.
2248 static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio)
2255 for_each_set_bit(bit, rbio->dbitmap, rbio->stripe_npages) {
2256 for (i = 0; i < rbio->real_stripes; i++) {
2257 index = i * rbio->stripe_npages + bit;
2258 if (rbio->stripe_pages[index])
2261 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2264 rbio->stripe_pages[index] = page;
2265 ClearPageUptodate(page);
2272 * end io function used by finish_rmw. When we finally
2273 * get here, we've written a full stripe
2275 static void raid_write_parity_end_io(struct bio *bio)
2277 struct btrfs_raid_bio *rbio = bio->bi_private;
2278 int err = bio->bi_error;
2281 fail_bio_stripe(rbio, bio);
2285 if (!atomic_dec_and_test(&rbio->stripes_pending))
2290 if (atomic_read(&rbio->error))
2293 rbio_orig_end_io(rbio, err);
2296 static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
2299 struct btrfs_bio *bbio = rbio->bbio;
2300 void *pointers[rbio->real_stripes];
2301 DECLARE_BITMAP(pbitmap, rbio->stripe_npages);
2302 int nr_data = rbio->nr_data;
2307 struct page *p_page = NULL;
2308 struct page *q_page = NULL;
2309 struct bio_list bio_list;
2314 bio_list_init(&bio_list);
2316 if (rbio->real_stripes - rbio->nr_data == 1) {
2317 p_stripe = rbio->real_stripes - 1;
2318 } else if (rbio->real_stripes - rbio->nr_data == 2) {
2319 p_stripe = rbio->real_stripes - 2;
2320 q_stripe = rbio->real_stripes - 1;
2325 if (bbio->num_tgtdevs && bbio->tgtdev_map[rbio->scrubp]) {
2327 bitmap_copy(pbitmap, rbio->dbitmap, rbio->stripe_npages);
2331 * Because the higher layers(scrubber) are unlikely to
2332 * use this area of the disk again soon, so don't cache
2335 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2340 p_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2343 SetPageUptodate(p_page);
2345 if (q_stripe != -1) {
2346 q_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2348 __free_page(p_page);
2351 SetPageUptodate(q_page);
2354 atomic_set(&rbio->error, 0);
2356 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2359 /* first collect one page from each data stripe */
2360 for (stripe = 0; stripe < nr_data; stripe++) {
2361 p = page_in_rbio(rbio, stripe, pagenr, 0);
2362 pointers[stripe] = kmap(p);
2365 /* then add the parity stripe */
2366 pointers[stripe++] = kmap(p_page);
2368 if (q_stripe != -1) {
2371 * raid6, add the qstripe and call the
2372 * library function to fill in our p/q
2374 pointers[stripe++] = kmap(q_page);
2376 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
2380 memcpy(pointers[nr_data], pointers[0], PAGE_SIZE);
2381 run_xor(pointers + 1, nr_data - 1, PAGE_CACHE_SIZE);
2384 /* Check scrubbing pairty and repair it */
2385 p = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2387 if (memcmp(parity, pointers[rbio->scrubp], PAGE_CACHE_SIZE))
2388 memcpy(parity, pointers[rbio->scrubp], PAGE_CACHE_SIZE);
2390 /* Parity is right, needn't writeback */
2391 bitmap_clear(rbio->dbitmap, pagenr, 1);
2394 for (stripe = 0; stripe < rbio->real_stripes; stripe++)
2395 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
2398 __free_page(p_page);
2400 __free_page(q_page);
2404 * time to start writing. Make bios for everything from the
2405 * higher layers (the bio_list in our rbio) and our p/q. Ignore
2408 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2411 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2412 ret = rbio_add_io_page(rbio, &bio_list,
2413 page, rbio->scrubp, pagenr, rbio->stripe_len);
2421 for_each_set_bit(pagenr, pbitmap, rbio->stripe_npages) {
2424 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2425 ret = rbio_add_io_page(rbio, &bio_list, page,
2426 bbio->tgtdev_map[rbio->scrubp],
2427 pagenr, rbio->stripe_len);
2433 nr_data = bio_list_size(&bio_list);
2435 /* Every parity is right */
2436 rbio_orig_end_io(rbio, 0);
2440 atomic_set(&rbio->stripes_pending, nr_data);
2443 bio = bio_list_pop(&bio_list);
2447 bio->bi_private = rbio;
2448 bio->bi_end_io = raid_write_parity_end_io;
2449 submit_bio(WRITE, bio);
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, luckly, 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;
2627 btrfs_bio_wq_end_io(rbio->fs_info, bio,
2628 BTRFS_WQ_ENDIO_RAID56);
2630 submit_bio(READ, bio);
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,
2660 void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio)
2662 if (!lock_stripe_add(rbio))
2663 async_scrub_parity(rbio);