2 * Copyright (C) 2012 Fusion-io All rights reserved.
3 * Copyright (C) 2012 Intel Corp. All rights reserved.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public
7 * License v2 as published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * General Public License for more details.
14 * You should have received a copy of the GNU General Public
15 * License along with this program; if not, write to the
16 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
17 * Boston, MA 021110-1307, USA.
19 #include <linux/sched.h>
20 #include <linux/wait.h>
21 #include <linux/bio.h>
22 #include <linux/slab.h>
23 #include <linux/buffer_head.h>
24 #include <linux/blkdev.h>
25 #include <linux/random.h>
26 #include <linux/iocontext.h>
27 #include <linux/capability.h>
28 #include <linux/ratelimit.h>
29 #include <linux/kthread.h>
30 #include <linux/raid/pq.h>
31 #include <linux/hash.h>
32 #include <linux/list_sort.h>
33 #include <linux/raid/xor.h>
34 #include <linux/vmalloc.h>
35 #include <asm/div64.h>
37 #include "extent_map.h"
39 #include "transaction.h"
40 #include "print-tree.h"
43 #include "async-thread.h"
44 #include "check-integrity.h"
45 #include "rcu-string.h"
47 /* set when additional merges to this rbio are not allowed */
48 #define RBIO_RMW_LOCKED_BIT 1
51 * set when this rbio is sitting in the hash, but it is just a cache
54 #define RBIO_CACHE_BIT 2
57 * set when it is safe to trust the stripe_pages for caching
59 #define RBIO_CACHE_READY_BIT 3
61 #define RBIO_CACHE_SIZE 1024
65 BTRFS_RBIO_READ_REBUILD
,
66 BTRFS_RBIO_PARITY_SCRUB
,
67 BTRFS_RBIO_REBUILD_MISSING
,
70 struct btrfs_raid_bio
{
71 struct btrfs_fs_info
*fs_info
;
72 struct btrfs_bio
*bbio
;
74 /* while we're doing rmw on a stripe
75 * we put it into a hash table so we can
76 * lock the stripe and merge more rbios
79 struct list_head hash_list
;
82 * LRU list for the stripe cache
84 struct list_head stripe_cache
;
87 * for scheduling work in the helper threads
89 struct btrfs_work work
;
92 * bio list and bio_list_lock are used
93 * to add more bios into the stripe
94 * in hopes of avoiding the full rmw
96 struct bio_list bio_list
;
97 spinlock_t bio_list_lock
;
99 /* also protected by the bio_list_lock, the
100 * plug list is used by the plugging code
101 * to collect partial bios while plugged. The
102 * stripe locking code also uses it to hand off
103 * the stripe lock to the next pending IO
105 struct list_head plug_list
;
108 * flags that tell us if it is safe to
109 * merge with this bio
113 /* size of each individual stripe on disk */
116 /* number of data stripes (no p/q) */
123 * set if we're doing a parity rebuild
124 * for a read from higher up, which is handled
125 * differently from a parity rebuild as part of
128 enum btrfs_rbio_ops operation
;
130 /* first bad stripe */
133 /* second bad stripe (for raid6 use) */
138 * number of pages needed to represent the full
144 * size of all the bios in the bio_list. This
145 * helps us decide if the rbio maps to a full
154 atomic_t stripes_pending
;
158 * these are two arrays of pointers. We allocate the
159 * rbio big enough to hold them both and setup their
160 * locations when the rbio is allocated
163 /* pointers to pages that we allocated for
164 * reading/writing stripes directly from the disk (including P/Q)
166 struct page
**stripe_pages
;
169 * pointers to the pages in the bio_list. Stored
170 * here for faster lookup
172 struct page
**bio_pages
;
175 * bitmap to record which horizontal stripe has data
177 unsigned long *dbitmap
;
180 static int __raid56_parity_recover(struct btrfs_raid_bio
*rbio
);
181 static noinline
void finish_rmw(struct btrfs_raid_bio
*rbio
);
182 static void rmw_work(struct btrfs_work
*work
);
183 static void read_rebuild_work(struct btrfs_work
*work
);
184 static void async_rmw_stripe(struct btrfs_raid_bio
*rbio
);
185 static void async_read_rebuild(struct btrfs_raid_bio
*rbio
);
186 static int fail_bio_stripe(struct btrfs_raid_bio
*rbio
, struct bio
*bio
);
187 static int fail_rbio_index(struct btrfs_raid_bio
*rbio
, int failed
);
188 static void __free_raid_bio(struct btrfs_raid_bio
*rbio
);
189 static void index_rbio_pages(struct btrfs_raid_bio
*rbio
);
190 static int alloc_rbio_pages(struct btrfs_raid_bio
*rbio
);
192 static noinline
void finish_parity_scrub(struct btrfs_raid_bio
*rbio
,
194 static void async_scrub_parity(struct btrfs_raid_bio
*rbio
);
197 * the stripe hash table is used for locking, and to collect
198 * bios in hopes of making a full stripe
200 int btrfs_alloc_stripe_hash_table(struct btrfs_fs_info
*info
)
202 struct btrfs_stripe_hash_table
*table
;
203 struct btrfs_stripe_hash_table
*x
;
204 struct btrfs_stripe_hash
*cur
;
205 struct btrfs_stripe_hash
*h
;
206 int num_entries
= 1 << BTRFS_STRIPE_HASH_TABLE_BITS
;
210 if (info
->stripe_hash_table
)
214 * The table is large, starting with order 4 and can go as high as
215 * order 7 in case lock debugging is turned on.
217 * Try harder to allocate and fallback to vmalloc to lower the chance
218 * of a failing mount.
220 table_size
= sizeof(*table
) + sizeof(*h
) * num_entries
;
221 table
= kzalloc(table_size
, GFP_KERNEL
| __GFP_NOWARN
| __GFP_REPEAT
);
223 table
= vzalloc(table_size
);
228 spin_lock_init(&table
->cache_lock
);
229 INIT_LIST_HEAD(&table
->stripe_cache
);
233 for (i
= 0; i
< num_entries
; i
++) {
235 INIT_LIST_HEAD(&cur
->hash_list
);
236 spin_lock_init(&cur
->lock
);
237 init_waitqueue_head(&cur
->wait
);
240 x
= cmpxchg(&info
->stripe_hash_table
, NULL
, table
);
247 * caching an rbio means to copy anything from the
248 * bio_pages array into the stripe_pages array. We
249 * use the page uptodate bit in the stripe cache array
250 * to indicate if it has valid data
252 * once the caching is done, we set the cache ready
255 static void cache_rbio_pages(struct btrfs_raid_bio
*rbio
)
262 ret
= alloc_rbio_pages(rbio
);
266 for (i
= 0; i
< rbio
->nr_pages
; i
++) {
267 if (!rbio
->bio_pages
[i
])
270 s
= kmap(rbio
->bio_pages
[i
]);
271 d
= kmap(rbio
->stripe_pages
[i
]);
273 memcpy(d
, s
, PAGE_SIZE
);
275 kunmap(rbio
->bio_pages
[i
]);
276 kunmap(rbio
->stripe_pages
[i
]);
277 SetPageUptodate(rbio
->stripe_pages
[i
]);
279 set_bit(RBIO_CACHE_READY_BIT
, &rbio
->flags
);
283 * we hash on the first logical address of the stripe
285 static int rbio_bucket(struct btrfs_raid_bio
*rbio
)
287 u64 num
= rbio
->bbio
->raid_map
[0];
290 * we shift down quite a bit. We're using byte
291 * addressing, and most of the lower bits are zeros.
292 * This tends to upset hash_64, and it consistently
293 * returns just one or two different values.
295 * shifting off the lower bits fixes things.
297 return hash_64(num
>> 16, BTRFS_STRIPE_HASH_TABLE_BITS
);
301 * stealing an rbio means taking all the uptodate pages from the stripe
302 * array in the source rbio and putting them into the destination rbio
304 static void steal_rbio(struct btrfs_raid_bio
*src
, struct btrfs_raid_bio
*dest
)
310 if (!test_bit(RBIO_CACHE_READY_BIT
, &src
->flags
))
313 for (i
= 0; i
< dest
->nr_pages
; i
++) {
314 s
= src
->stripe_pages
[i
];
315 if (!s
|| !PageUptodate(s
)) {
319 d
= dest
->stripe_pages
[i
];
323 dest
->stripe_pages
[i
] = s
;
324 src
->stripe_pages
[i
] = NULL
;
329 * merging means we take the bio_list from the victim and
330 * splice it into the destination. The victim should
331 * be discarded afterwards.
333 * must be called with dest->rbio_list_lock held
335 static void merge_rbio(struct btrfs_raid_bio
*dest
,
336 struct btrfs_raid_bio
*victim
)
338 bio_list_merge(&dest
->bio_list
, &victim
->bio_list
);
339 dest
->bio_list_bytes
+= victim
->bio_list_bytes
;
340 dest
->generic_bio_cnt
+= victim
->generic_bio_cnt
;
341 bio_list_init(&victim
->bio_list
);
345 * used to prune items that are in the cache. The caller
346 * must hold the hash table lock.
348 static void __remove_rbio_from_cache(struct btrfs_raid_bio
*rbio
)
350 int bucket
= rbio_bucket(rbio
);
351 struct btrfs_stripe_hash_table
*table
;
352 struct btrfs_stripe_hash
*h
;
356 * check the bit again under the hash table lock.
358 if (!test_bit(RBIO_CACHE_BIT
, &rbio
->flags
))
361 table
= rbio
->fs_info
->stripe_hash_table
;
362 h
= table
->table
+ bucket
;
364 /* hold the lock for the bucket because we may be
365 * removing it from the hash table
370 * hold the lock for the bio list because we need
371 * to make sure the bio list is empty
373 spin_lock(&rbio
->bio_list_lock
);
375 if (test_and_clear_bit(RBIO_CACHE_BIT
, &rbio
->flags
)) {
376 list_del_init(&rbio
->stripe_cache
);
377 table
->cache_size
-= 1;
380 /* if the bio list isn't empty, this rbio is
381 * still involved in an IO. We take it out
382 * of the cache list, and drop the ref that
383 * was held for the list.
385 * If the bio_list was empty, we also remove
386 * the rbio from the hash_table, and drop
387 * the corresponding ref
389 if (bio_list_empty(&rbio
->bio_list
)) {
390 if (!list_empty(&rbio
->hash_list
)) {
391 list_del_init(&rbio
->hash_list
);
392 refcount_dec(&rbio
->refs
);
393 BUG_ON(!list_empty(&rbio
->plug_list
));
398 spin_unlock(&rbio
->bio_list_lock
);
399 spin_unlock(&h
->lock
);
402 __free_raid_bio(rbio
);
406 * prune a given rbio from the cache
408 static void remove_rbio_from_cache(struct btrfs_raid_bio
*rbio
)
410 struct btrfs_stripe_hash_table
*table
;
413 if (!test_bit(RBIO_CACHE_BIT
, &rbio
->flags
))
416 table
= rbio
->fs_info
->stripe_hash_table
;
418 spin_lock_irqsave(&table
->cache_lock
, flags
);
419 __remove_rbio_from_cache(rbio
);
420 spin_unlock_irqrestore(&table
->cache_lock
, flags
);
424 * remove everything in the cache
426 static void btrfs_clear_rbio_cache(struct btrfs_fs_info
*info
)
428 struct btrfs_stripe_hash_table
*table
;
430 struct btrfs_raid_bio
*rbio
;
432 table
= info
->stripe_hash_table
;
434 spin_lock_irqsave(&table
->cache_lock
, flags
);
435 while (!list_empty(&table
->stripe_cache
)) {
436 rbio
= list_entry(table
->stripe_cache
.next
,
437 struct btrfs_raid_bio
,
439 __remove_rbio_from_cache(rbio
);
441 spin_unlock_irqrestore(&table
->cache_lock
, flags
);
445 * remove all cached entries and free the hash table
448 void btrfs_free_stripe_hash_table(struct btrfs_fs_info
*info
)
450 if (!info
->stripe_hash_table
)
452 btrfs_clear_rbio_cache(info
);
453 kvfree(info
->stripe_hash_table
);
454 info
->stripe_hash_table
= NULL
;
458 * insert an rbio into the stripe cache. It
459 * must have already been prepared by calling
462 * If this rbio was already cached, it gets
463 * moved to the front of the lru.
465 * If the size of the rbio cache is too big, we
468 static void cache_rbio(struct btrfs_raid_bio
*rbio
)
470 struct btrfs_stripe_hash_table
*table
;
473 if (!test_bit(RBIO_CACHE_READY_BIT
, &rbio
->flags
))
476 table
= rbio
->fs_info
->stripe_hash_table
;
478 spin_lock_irqsave(&table
->cache_lock
, flags
);
479 spin_lock(&rbio
->bio_list_lock
);
481 /* bump our ref if we were not in the list before */
482 if (!test_and_set_bit(RBIO_CACHE_BIT
, &rbio
->flags
))
483 refcount_inc(&rbio
->refs
);
485 if (!list_empty(&rbio
->stripe_cache
)){
486 list_move(&rbio
->stripe_cache
, &table
->stripe_cache
);
488 list_add(&rbio
->stripe_cache
, &table
->stripe_cache
);
489 table
->cache_size
+= 1;
492 spin_unlock(&rbio
->bio_list_lock
);
494 if (table
->cache_size
> RBIO_CACHE_SIZE
) {
495 struct btrfs_raid_bio
*found
;
497 found
= list_entry(table
->stripe_cache
.prev
,
498 struct btrfs_raid_bio
,
502 __remove_rbio_from_cache(found
);
505 spin_unlock_irqrestore(&table
->cache_lock
, flags
);
509 * helper function to run the xor_blocks api. It is only
510 * able to do MAX_XOR_BLOCKS at a time, so we need to
513 static void run_xor(void **pages
, int src_cnt
, ssize_t len
)
517 void *dest
= pages
[src_cnt
];
520 xor_src_cnt
= min(src_cnt
, MAX_XOR_BLOCKS
);
521 xor_blocks(xor_src_cnt
, len
, dest
, pages
+ src_off
);
523 src_cnt
-= xor_src_cnt
;
524 src_off
+= xor_src_cnt
;
529 * returns true if the bio list inside this rbio
530 * covers an entire stripe (no rmw required).
531 * Must be called with the bio list lock held, or
532 * at a time when you know it is impossible to add
533 * new bios into the list
535 static int __rbio_is_full(struct btrfs_raid_bio
*rbio
)
537 unsigned long size
= rbio
->bio_list_bytes
;
540 if (size
!= rbio
->nr_data
* rbio
->stripe_len
)
543 BUG_ON(size
> rbio
->nr_data
* rbio
->stripe_len
);
547 static int rbio_is_full(struct btrfs_raid_bio
*rbio
)
552 spin_lock_irqsave(&rbio
->bio_list_lock
, flags
);
553 ret
= __rbio_is_full(rbio
);
554 spin_unlock_irqrestore(&rbio
->bio_list_lock
, flags
);
559 * returns 1 if it is safe to merge two rbios together.
560 * The merging is safe if the two rbios correspond to
561 * the same stripe and if they are both going in the same
562 * direction (read vs write), and if neither one is
563 * locked for final IO
565 * The caller is responsible for locking such that
566 * rmw_locked is safe to test
568 static int rbio_can_merge(struct btrfs_raid_bio
*last
,
569 struct btrfs_raid_bio
*cur
)
571 if (test_bit(RBIO_RMW_LOCKED_BIT
, &last
->flags
) ||
572 test_bit(RBIO_RMW_LOCKED_BIT
, &cur
->flags
))
576 * we can't merge with cached rbios, since the
577 * idea is that when we merge the destination
578 * rbio is going to run our IO for us. We can
579 * steal from cached rbios though, other functions
582 if (test_bit(RBIO_CACHE_BIT
, &last
->flags
) ||
583 test_bit(RBIO_CACHE_BIT
, &cur
->flags
))
586 if (last
->bbio
->raid_map
[0] !=
587 cur
->bbio
->raid_map
[0])
590 /* we can't merge with different operations */
591 if (last
->operation
!= cur
->operation
)
594 * We've need read the full stripe from the drive.
595 * check and repair the parity and write the new results.
597 * We're not allowed to add any new bios to the
598 * bio list here, anyone else that wants to
599 * change this stripe needs to do their own rmw.
601 if (last
->operation
== BTRFS_RBIO_PARITY_SCRUB
||
602 cur
->operation
== BTRFS_RBIO_PARITY_SCRUB
)
605 if (last
->operation
== BTRFS_RBIO_REBUILD_MISSING
||
606 cur
->operation
== BTRFS_RBIO_REBUILD_MISSING
)
612 static int rbio_stripe_page_index(struct btrfs_raid_bio
*rbio
, int stripe
,
615 return stripe
* rbio
->stripe_npages
+ index
;
619 * these are just the pages from the rbio array, not from anything
620 * the FS sent down to us
622 static struct page
*rbio_stripe_page(struct btrfs_raid_bio
*rbio
, int stripe
,
625 return rbio
->stripe_pages
[rbio_stripe_page_index(rbio
, stripe
, index
)];
629 * helper to index into the pstripe
631 static struct page
*rbio_pstripe_page(struct btrfs_raid_bio
*rbio
, int index
)
633 return rbio_stripe_page(rbio
, rbio
->nr_data
, index
);
637 * helper to index into the qstripe, returns null
638 * if there is no qstripe
640 static struct page
*rbio_qstripe_page(struct btrfs_raid_bio
*rbio
, int index
)
642 if (rbio
->nr_data
+ 1 == rbio
->real_stripes
)
644 return rbio_stripe_page(rbio
, rbio
->nr_data
+ 1, index
);
648 * The first stripe in the table for a logical address
649 * has the lock. rbios are added in one of three ways:
651 * 1) Nobody has the stripe locked yet. The rbio is given
652 * the lock and 0 is returned. The caller must start the IO
655 * 2) Someone has the stripe locked, but we're able to merge
656 * with the lock owner. The rbio is freed and the IO will
657 * start automatically along with the existing rbio. 1 is returned.
659 * 3) Someone has the stripe locked, but we're not able to merge.
660 * The rbio is added to the lock owner's plug list, or merged into
661 * an rbio already on the plug list. When the lock owner unlocks,
662 * the next rbio on the list is run and the IO is started automatically.
665 * If we return 0, the caller still owns the rbio and must continue with
666 * IO submission. If we return 1, the caller must assume the rbio has
667 * already been freed.
669 static noinline
int lock_stripe_add(struct btrfs_raid_bio
*rbio
)
671 int bucket
= rbio_bucket(rbio
);
672 struct btrfs_stripe_hash
*h
= rbio
->fs_info
->stripe_hash_table
->table
+ bucket
;
673 struct btrfs_raid_bio
*cur
;
674 struct btrfs_raid_bio
*pending
;
677 struct btrfs_raid_bio
*freeit
= NULL
;
678 struct btrfs_raid_bio
*cache_drop
= NULL
;
681 spin_lock_irqsave(&h
->lock
, flags
);
682 list_for_each_entry(cur
, &h
->hash_list
, hash_list
) {
683 if (cur
->bbio
->raid_map
[0] == rbio
->bbio
->raid_map
[0]) {
684 spin_lock(&cur
->bio_list_lock
);
686 /* can we steal this cached rbio's pages? */
687 if (bio_list_empty(&cur
->bio_list
) &&
688 list_empty(&cur
->plug_list
) &&
689 test_bit(RBIO_CACHE_BIT
, &cur
->flags
) &&
690 !test_bit(RBIO_RMW_LOCKED_BIT
, &cur
->flags
)) {
691 list_del_init(&cur
->hash_list
);
692 refcount_dec(&cur
->refs
);
694 steal_rbio(cur
, rbio
);
696 spin_unlock(&cur
->bio_list_lock
);
701 /* can we merge into the lock owner? */
702 if (rbio_can_merge(cur
, rbio
)) {
703 merge_rbio(cur
, rbio
);
704 spin_unlock(&cur
->bio_list_lock
);
712 * we couldn't merge with the running
713 * rbio, see if we can merge with the
714 * pending ones. We don't have to
715 * check for rmw_locked because there
716 * is no way they are inside finish_rmw
719 list_for_each_entry(pending
, &cur
->plug_list
,
721 if (rbio_can_merge(pending
, rbio
)) {
722 merge_rbio(pending
, rbio
);
723 spin_unlock(&cur
->bio_list_lock
);
730 /* no merging, put us on the tail of the plug list,
731 * our rbio will be started with the currently
732 * running rbio unlocks
734 list_add_tail(&rbio
->plug_list
, &cur
->plug_list
);
735 spin_unlock(&cur
->bio_list_lock
);
741 refcount_inc(&rbio
->refs
);
742 list_add(&rbio
->hash_list
, &h
->hash_list
);
744 spin_unlock_irqrestore(&h
->lock
, flags
);
746 remove_rbio_from_cache(cache_drop
);
748 __free_raid_bio(freeit
);
753 * called as rmw or parity rebuild is completed. If the plug list has more
754 * rbios waiting for this stripe, the next one on the list will be started
756 static noinline
void unlock_stripe(struct btrfs_raid_bio
*rbio
)
759 struct btrfs_stripe_hash
*h
;
763 bucket
= rbio_bucket(rbio
);
764 h
= rbio
->fs_info
->stripe_hash_table
->table
+ bucket
;
766 if (list_empty(&rbio
->plug_list
))
769 spin_lock_irqsave(&h
->lock
, flags
);
770 spin_lock(&rbio
->bio_list_lock
);
772 if (!list_empty(&rbio
->hash_list
)) {
774 * if we're still cached and there is no other IO
775 * to perform, just leave this rbio here for others
776 * to steal from later
778 if (list_empty(&rbio
->plug_list
) &&
779 test_bit(RBIO_CACHE_BIT
, &rbio
->flags
)) {
781 clear_bit(RBIO_RMW_LOCKED_BIT
, &rbio
->flags
);
782 BUG_ON(!bio_list_empty(&rbio
->bio_list
));
786 list_del_init(&rbio
->hash_list
);
787 refcount_dec(&rbio
->refs
);
790 * we use the plug list to hold all the rbios
791 * waiting for the chance to lock this stripe.
792 * hand the lock over to one of them.
794 if (!list_empty(&rbio
->plug_list
)) {
795 struct btrfs_raid_bio
*next
;
796 struct list_head
*head
= rbio
->plug_list
.next
;
798 next
= list_entry(head
, struct btrfs_raid_bio
,
801 list_del_init(&rbio
->plug_list
);
803 list_add(&next
->hash_list
, &h
->hash_list
);
804 refcount_inc(&next
->refs
);
805 spin_unlock(&rbio
->bio_list_lock
);
806 spin_unlock_irqrestore(&h
->lock
, flags
);
808 if (next
->operation
== BTRFS_RBIO_READ_REBUILD
)
809 async_read_rebuild(next
);
810 else if (next
->operation
== BTRFS_RBIO_REBUILD_MISSING
) {
811 steal_rbio(rbio
, next
);
812 async_read_rebuild(next
);
813 } else if (next
->operation
== BTRFS_RBIO_WRITE
) {
814 steal_rbio(rbio
, next
);
815 async_rmw_stripe(next
);
816 } else if (next
->operation
== BTRFS_RBIO_PARITY_SCRUB
) {
817 steal_rbio(rbio
, next
);
818 async_scrub_parity(next
);
823 * The barrier for this waitqueue_active is not needed,
824 * we're protected by h->lock and can't miss a wakeup.
826 } else if (waitqueue_active(&h
->wait
)) {
827 spin_unlock(&rbio
->bio_list_lock
);
828 spin_unlock_irqrestore(&h
->lock
, flags
);
834 spin_unlock(&rbio
->bio_list_lock
);
835 spin_unlock_irqrestore(&h
->lock
, flags
);
839 remove_rbio_from_cache(rbio
);
842 static void __free_raid_bio(struct btrfs_raid_bio
*rbio
)
846 if (!refcount_dec_and_test(&rbio
->refs
))
849 WARN_ON(!list_empty(&rbio
->stripe_cache
));
850 WARN_ON(!list_empty(&rbio
->hash_list
));
851 WARN_ON(!bio_list_empty(&rbio
->bio_list
));
853 for (i
= 0; i
< rbio
->nr_pages
; i
++) {
854 if (rbio
->stripe_pages
[i
]) {
855 __free_page(rbio
->stripe_pages
[i
]);
856 rbio
->stripe_pages
[i
] = NULL
;
860 btrfs_put_bbio(rbio
->bbio
);
864 static void free_raid_bio(struct btrfs_raid_bio
*rbio
)
867 __free_raid_bio(rbio
);
871 * this frees the rbio and runs through all the bios in the
872 * bio_list and calls end_io on them
874 static void rbio_orig_end_io(struct btrfs_raid_bio
*rbio
, int err
)
876 struct bio
*cur
= bio_list_get(&rbio
->bio_list
);
879 if (rbio
->generic_bio_cnt
)
880 btrfs_bio_counter_sub(rbio
->fs_info
, rbio
->generic_bio_cnt
);
894 * end io function used by finish_rmw. When we finally
895 * get here, we've written a full stripe
897 static void raid_write_end_io(struct bio
*bio
)
899 struct btrfs_raid_bio
*rbio
= bio
->bi_private
;
900 int err
= bio
->bi_error
;
904 fail_bio_stripe(rbio
, bio
);
908 if (!atomic_dec_and_test(&rbio
->stripes_pending
))
913 /* OK, we have read all the stripes we need to. */
914 max_errors
= (rbio
->operation
== BTRFS_RBIO_PARITY_SCRUB
) ?
915 0 : rbio
->bbio
->max_errors
;
916 if (atomic_read(&rbio
->error
) > max_errors
)
919 rbio_orig_end_io(rbio
, err
);
923 * the read/modify/write code wants to use the original bio for
924 * any pages it included, and then use the rbio for everything
925 * else. This function decides if a given index (stripe number)
926 * and page number in that stripe fall inside the original bio
929 * if you set bio_list_only, you'll get a NULL back for any ranges
930 * that are outside the bio_list
932 * This doesn't take any refs on anything, you get a bare page pointer
933 * and the caller must bump refs as required.
935 * You must call index_rbio_pages once before you can trust
936 * the answers from this function.
938 static struct page
*page_in_rbio(struct btrfs_raid_bio
*rbio
,
939 int index
, int pagenr
, int bio_list_only
)
942 struct page
*p
= NULL
;
944 chunk_page
= index
* (rbio
->stripe_len
>> PAGE_SHIFT
) + pagenr
;
946 spin_lock_irq(&rbio
->bio_list_lock
);
947 p
= rbio
->bio_pages
[chunk_page
];
948 spin_unlock_irq(&rbio
->bio_list_lock
);
950 if (p
|| bio_list_only
)
953 return rbio
->stripe_pages
[chunk_page
];
957 * number of pages we need for the entire stripe across all the
960 static unsigned long rbio_nr_pages(unsigned long stripe_len
, int nr_stripes
)
962 return DIV_ROUND_UP(stripe_len
, PAGE_SIZE
) * nr_stripes
;
966 * allocation and initial setup for the btrfs_raid_bio. Not
967 * this does not allocate any pages for rbio->pages.
969 static struct btrfs_raid_bio
*alloc_rbio(struct btrfs_fs_info
*fs_info
,
970 struct btrfs_bio
*bbio
,
973 struct btrfs_raid_bio
*rbio
;
975 int real_stripes
= bbio
->num_stripes
- bbio
->num_tgtdevs
;
976 int num_pages
= rbio_nr_pages(stripe_len
, real_stripes
);
977 int stripe_npages
= DIV_ROUND_UP(stripe_len
, PAGE_SIZE
);
980 rbio
= kzalloc(sizeof(*rbio
) + num_pages
* sizeof(struct page
*) * 2 +
981 DIV_ROUND_UP(stripe_npages
, BITS_PER_LONG
) *
982 sizeof(long), GFP_NOFS
);
984 return ERR_PTR(-ENOMEM
);
986 bio_list_init(&rbio
->bio_list
);
987 INIT_LIST_HEAD(&rbio
->plug_list
);
988 spin_lock_init(&rbio
->bio_list_lock
);
989 INIT_LIST_HEAD(&rbio
->stripe_cache
);
990 INIT_LIST_HEAD(&rbio
->hash_list
);
992 rbio
->fs_info
= fs_info
;
993 rbio
->stripe_len
= stripe_len
;
994 rbio
->nr_pages
= num_pages
;
995 rbio
->real_stripes
= real_stripes
;
996 rbio
->stripe_npages
= stripe_npages
;
999 refcount_set(&rbio
->refs
, 1);
1000 atomic_set(&rbio
->error
, 0);
1001 atomic_set(&rbio
->stripes_pending
, 0);
1004 * the stripe_pages and bio_pages array point to the extra
1005 * memory we allocated past the end of the rbio
1008 rbio
->stripe_pages
= p
;
1009 rbio
->bio_pages
= p
+ sizeof(struct page
*) * num_pages
;
1010 rbio
->dbitmap
= p
+ sizeof(struct page
*) * num_pages
* 2;
1012 if (bbio
->map_type
& BTRFS_BLOCK_GROUP_RAID5
)
1013 nr_data
= real_stripes
- 1;
1014 else if (bbio
->map_type
& BTRFS_BLOCK_GROUP_RAID6
)
1015 nr_data
= real_stripes
- 2;
1019 rbio
->nr_data
= nr_data
;
1023 /* allocate pages for all the stripes in the bio, including parity */
1024 static int alloc_rbio_pages(struct btrfs_raid_bio
*rbio
)
1029 for (i
= 0; i
< rbio
->nr_pages
; i
++) {
1030 if (rbio
->stripe_pages
[i
])
1032 page
= alloc_page(GFP_NOFS
| __GFP_HIGHMEM
);
1035 rbio
->stripe_pages
[i
] = page
;
1040 /* only allocate pages for p/q stripes */
1041 static int alloc_rbio_parity_pages(struct btrfs_raid_bio
*rbio
)
1046 i
= rbio_stripe_page_index(rbio
, rbio
->nr_data
, 0);
1048 for (; i
< rbio
->nr_pages
; i
++) {
1049 if (rbio
->stripe_pages
[i
])
1051 page
= alloc_page(GFP_NOFS
| __GFP_HIGHMEM
);
1054 rbio
->stripe_pages
[i
] = page
;
1060 * add a single page from a specific stripe into our list of bios for IO
1061 * this will try to merge into existing bios if possible, and returns
1062 * zero if all went well.
1064 static int rbio_add_io_page(struct btrfs_raid_bio
*rbio
,
1065 struct bio_list
*bio_list
,
1068 unsigned long page_index
,
1069 unsigned long bio_max_len
)
1071 struct bio
*last
= bio_list
->tail
;
1075 struct btrfs_bio_stripe
*stripe
;
1078 stripe
= &rbio
->bbio
->stripes
[stripe_nr
];
1079 disk_start
= stripe
->physical
+ (page_index
<< PAGE_SHIFT
);
1081 /* if the device is missing, just fail this stripe */
1082 if (!stripe
->dev
->bdev
)
1083 return fail_rbio_index(rbio
, stripe_nr
);
1085 /* see if we can add this page onto our existing bio */
1087 last_end
= (u64
)last
->bi_iter
.bi_sector
<< 9;
1088 last_end
+= last
->bi_iter
.bi_size
;
1091 * we can't merge these if they are from different
1092 * devices or if they are not contiguous
1094 if (last_end
== disk_start
&& stripe
->dev
->bdev
&&
1096 last
->bi_bdev
== stripe
->dev
->bdev
) {
1097 ret
= bio_add_page(last
, page
, PAGE_SIZE
, 0);
1098 if (ret
== PAGE_SIZE
)
1103 /* put a new bio on the list */
1104 bio
= btrfs_io_bio_alloc(GFP_NOFS
, bio_max_len
>> PAGE_SHIFT
?:1);
1108 bio
->bi_iter
.bi_size
= 0;
1109 bio
->bi_bdev
= stripe
->dev
->bdev
;
1110 bio
->bi_iter
.bi_sector
= disk_start
>> 9;
1112 bio_add_page(bio
, page
, PAGE_SIZE
, 0);
1113 bio_list_add(bio_list
, bio
);
1118 * while we're doing the read/modify/write cycle, we could
1119 * have errors in reading pages off the disk. This checks
1120 * for errors and if we're not able to read the page it'll
1121 * trigger parity reconstruction. The rmw will be finished
1122 * after we've reconstructed the failed stripes
1124 static void validate_rbio_for_rmw(struct btrfs_raid_bio
*rbio
)
1126 if (rbio
->faila
>= 0 || rbio
->failb
>= 0) {
1127 BUG_ON(rbio
->faila
== rbio
->real_stripes
- 1);
1128 __raid56_parity_recover(rbio
);
1135 * helper function to walk our bio list and populate the bio_pages array with
1136 * the result. This seems expensive, but it is faster than constantly
1137 * searching through the bio list as we setup the IO in finish_rmw or stripe
1140 * This must be called before you trust the answers from page_in_rbio
1142 static void index_rbio_pages(struct btrfs_raid_bio
*rbio
)
1145 struct bio_vec
*bvec
;
1147 unsigned long stripe_offset
;
1148 unsigned long page_index
;
1151 spin_lock_irq(&rbio
->bio_list_lock
);
1152 bio_list_for_each(bio
, &rbio
->bio_list
) {
1153 start
= (u64
)bio
->bi_iter
.bi_sector
<< 9;
1154 stripe_offset
= start
- rbio
->bbio
->raid_map
[0];
1155 page_index
= stripe_offset
>> PAGE_SHIFT
;
1157 bio_for_each_segment_all(bvec
, bio
, i
)
1158 rbio
->bio_pages
[page_index
+ i
] = bvec
->bv_page
;
1160 spin_unlock_irq(&rbio
->bio_list_lock
);
1164 * this is called from one of two situations. We either
1165 * have a full stripe from the higher layers, or we've read all
1166 * the missing bits off disk.
1168 * This will calculate the parity and then send down any
1171 static noinline
void finish_rmw(struct btrfs_raid_bio
*rbio
)
1173 struct btrfs_bio
*bbio
= rbio
->bbio
;
1174 void *pointers
[rbio
->real_stripes
];
1175 int nr_data
= rbio
->nr_data
;
1180 struct bio_list bio_list
;
1184 bio_list_init(&bio_list
);
1186 if (rbio
->real_stripes
- rbio
->nr_data
== 1) {
1187 p_stripe
= rbio
->real_stripes
- 1;
1188 } else if (rbio
->real_stripes
- rbio
->nr_data
== 2) {
1189 p_stripe
= rbio
->real_stripes
- 2;
1190 q_stripe
= rbio
->real_stripes
- 1;
1195 /* at this point we either have a full stripe,
1196 * or we've read the full stripe from the drive.
1197 * recalculate the parity and write the new results.
1199 * We're not allowed to add any new bios to the
1200 * bio list here, anyone else that wants to
1201 * change this stripe needs to do their own rmw.
1203 spin_lock_irq(&rbio
->bio_list_lock
);
1204 set_bit(RBIO_RMW_LOCKED_BIT
, &rbio
->flags
);
1205 spin_unlock_irq(&rbio
->bio_list_lock
);
1207 atomic_set(&rbio
->error
, 0);
1210 * now that we've set rmw_locked, run through the
1211 * bio list one last time and map the page pointers
1213 * We don't cache full rbios because we're assuming
1214 * the higher layers are unlikely to use this area of
1215 * the disk again soon. If they do use it again,
1216 * hopefully they will send another full bio.
1218 index_rbio_pages(rbio
);
1219 if (!rbio_is_full(rbio
))
1220 cache_rbio_pages(rbio
);
1222 clear_bit(RBIO_CACHE_READY_BIT
, &rbio
->flags
);
1224 for (pagenr
= 0; pagenr
< rbio
->stripe_npages
; pagenr
++) {
1226 /* first collect one page from each data stripe */
1227 for (stripe
= 0; stripe
< nr_data
; stripe
++) {
1228 p
= page_in_rbio(rbio
, stripe
, pagenr
, 0);
1229 pointers
[stripe
] = kmap(p
);
1232 /* then add the parity stripe */
1233 p
= rbio_pstripe_page(rbio
, pagenr
);
1235 pointers
[stripe
++] = kmap(p
);
1237 if (q_stripe
!= -1) {
1240 * raid6, add the qstripe and call the
1241 * library function to fill in our p/q
1243 p
= rbio_qstripe_page(rbio
, pagenr
);
1245 pointers
[stripe
++] = kmap(p
);
1247 raid6_call
.gen_syndrome(rbio
->real_stripes
, PAGE_SIZE
,
1251 memcpy(pointers
[nr_data
], pointers
[0], PAGE_SIZE
);
1252 run_xor(pointers
+ 1, nr_data
- 1, PAGE_SIZE
);
1256 for (stripe
= 0; stripe
< rbio
->real_stripes
; stripe
++)
1257 kunmap(page_in_rbio(rbio
, stripe
, pagenr
, 0));
1261 * time to start writing. Make bios for everything from the
1262 * higher layers (the bio_list in our rbio) and our p/q. Ignore
1265 for (stripe
= 0; stripe
< rbio
->real_stripes
; stripe
++) {
1266 for (pagenr
= 0; pagenr
< rbio
->stripe_npages
; pagenr
++) {
1268 if (stripe
< rbio
->nr_data
) {
1269 page
= page_in_rbio(rbio
, stripe
, pagenr
, 1);
1273 page
= rbio_stripe_page(rbio
, stripe
, pagenr
);
1276 ret
= rbio_add_io_page(rbio
, &bio_list
,
1277 page
, stripe
, pagenr
, rbio
->stripe_len
);
1283 if (likely(!bbio
->num_tgtdevs
))
1286 for (stripe
= 0; stripe
< rbio
->real_stripes
; stripe
++) {
1287 if (!bbio
->tgtdev_map
[stripe
])
1290 for (pagenr
= 0; pagenr
< rbio
->stripe_npages
; pagenr
++) {
1292 if (stripe
< rbio
->nr_data
) {
1293 page
= page_in_rbio(rbio
, stripe
, pagenr
, 1);
1297 page
= rbio_stripe_page(rbio
, stripe
, pagenr
);
1300 ret
= rbio_add_io_page(rbio
, &bio_list
, page
,
1301 rbio
->bbio
->tgtdev_map
[stripe
],
1302 pagenr
, rbio
->stripe_len
);
1309 atomic_set(&rbio
->stripes_pending
, bio_list_size(&bio_list
));
1310 BUG_ON(atomic_read(&rbio
->stripes_pending
) == 0);
1313 bio
= bio_list_pop(&bio_list
);
1317 bio
->bi_private
= rbio
;
1318 bio
->bi_end_io
= raid_write_end_io
;
1319 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
1326 rbio_orig_end_io(rbio
, -EIO
);
1330 * helper to find the stripe number for a given bio. Used to figure out which
1331 * stripe has failed. This expects the bio to correspond to a physical disk,
1332 * so it looks up based on physical sector numbers.
1334 static int find_bio_stripe(struct btrfs_raid_bio
*rbio
,
1337 u64 physical
= bio
->bi_iter
.bi_sector
;
1340 struct btrfs_bio_stripe
*stripe
;
1344 for (i
= 0; i
< rbio
->bbio
->num_stripes
; i
++) {
1345 stripe
= &rbio
->bbio
->stripes
[i
];
1346 stripe_start
= stripe
->physical
;
1347 if (physical
>= stripe_start
&&
1348 physical
< stripe_start
+ rbio
->stripe_len
&&
1349 bio
->bi_bdev
== stripe
->dev
->bdev
) {
1357 * helper to find the stripe number for a given
1358 * bio (before mapping). Used to figure out which stripe has
1359 * failed. This looks up based on logical block numbers.
1361 static int find_logical_bio_stripe(struct btrfs_raid_bio
*rbio
,
1364 u64 logical
= bio
->bi_iter
.bi_sector
;
1370 for (i
= 0; i
< rbio
->nr_data
; i
++) {
1371 stripe_start
= rbio
->bbio
->raid_map
[i
];
1372 if (logical
>= stripe_start
&&
1373 logical
< stripe_start
+ rbio
->stripe_len
) {
1381 * returns -EIO if we had too many failures
1383 static int fail_rbio_index(struct btrfs_raid_bio
*rbio
, int failed
)
1385 unsigned long flags
;
1388 spin_lock_irqsave(&rbio
->bio_list_lock
, flags
);
1390 /* we already know this stripe is bad, move on */
1391 if (rbio
->faila
== failed
|| rbio
->failb
== failed
)
1394 if (rbio
->faila
== -1) {
1395 /* first failure on this rbio */
1396 rbio
->faila
= failed
;
1397 atomic_inc(&rbio
->error
);
1398 } else if (rbio
->failb
== -1) {
1399 /* second failure on this rbio */
1400 rbio
->failb
= failed
;
1401 atomic_inc(&rbio
->error
);
1406 spin_unlock_irqrestore(&rbio
->bio_list_lock
, flags
);
1412 * helper to fail a stripe based on a physical disk
1415 static int fail_bio_stripe(struct btrfs_raid_bio
*rbio
,
1418 int failed
= find_bio_stripe(rbio
, bio
);
1423 return fail_rbio_index(rbio
, failed
);
1427 * this sets each page in the bio uptodate. It should only be used on private
1428 * rbio pages, nothing that comes in from the higher layers
1430 static void set_bio_pages_uptodate(struct bio
*bio
)
1432 struct bio_vec
*bvec
;
1435 bio_for_each_segment_all(bvec
, bio
, i
)
1436 SetPageUptodate(bvec
->bv_page
);
1440 * end io for the read phase of the rmw cycle. All the bios here are physical
1441 * stripe bios we've read from the disk so we can recalculate the parity of the
1444 * This will usually kick off finish_rmw once all the bios are read in, but it
1445 * may trigger parity reconstruction if we had any errors along the way
1447 static void raid_rmw_end_io(struct bio
*bio
)
1449 struct btrfs_raid_bio
*rbio
= bio
->bi_private
;
1452 fail_bio_stripe(rbio
, bio
);
1454 set_bio_pages_uptodate(bio
);
1458 if (!atomic_dec_and_test(&rbio
->stripes_pending
))
1461 if (atomic_read(&rbio
->error
) > rbio
->bbio
->max_errors
)
1465 * this will normally call finish_rmw to start our write
1466 * but if there are any failed stripes we'll reconstruct
1469 validate_rbio_for_rmw(rbio
);
1474 rbio_orig_end_io(rbio
, -EIO
);
1477 static void async_rmw_stripe(struct btrfs_raid_bio
*rbio
)
1479 btrfs_init_work(&rbio
->work
, btrfs_rmw_helper
, rmw_work
, NULL
, NULL
);
1480 btrfs_queue_work(rbio
->fs_info
->rmw_workers
, &rbio
->work
);
1483 static void async_read_rebuild(struct btrfs_raid_bio
*rbio
)
1485 btrfs_init_work(&rbio
->work
, btrfs_rmw_helper
,
1486 read_rebuild_work
, NULL
, NULL
);
1488 btrfs_queue_work(rbio
->fs_info
->rmw_workers
, &rbio
->work
);
1492 * the stripe must be locked by the caller. It will
1493 * unlock after all the writes are done
1495 static int raid56_rmw_stripe(struct btrfs_raid_bio
*rbio
)
1497 int bios_to_read
= 0;
1498 struct bio_list bio_list
;
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
< rbio
->stripe_npages
; 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
;
1568 bio_set_op_attrs(bio
, REQ_OP_READ
, 0);
1570 btrfs_bio_wq_end_io(rbio
->fs_info
, bio
, BTRFS_WQ_ENDIO_RAID56
);
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_fs_info
*fs_info
, 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(fs_info
, 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(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(fs_info
);
1766 cb
= blk_check_plugged(btrfs_raid_unplug
, fs_info
, sizeof(*plug
));
1768 plug
= container_of(cb
, struct btrfs_plug_cb
, cb
);
1770 plug
->info
= fs_info
;
1771 INIT_LIST_HEAD(&plug
->rbio_list
);
1773 list_add_tail(&rbio
->plug_list
, &plug
->rbio_list
);
1776 ret
= __raid56_parity_write(rbio
);
1778 btrfs_bio_counter_dec(fs_info
);
1784 * all parity reconstruction happens here. We've read in everything
1785 * we can find from the drives and this does the heavy lifting of
1786 * sorting the good from the bad.
1788 static void __raid_recover_end_io(struct btrfs_raid_bio
*rbio
)
1792 int faila
= -1, failb
= -1;
1797 pointers
= kcalloc(rbio
->real_stripes
, sizeof(void *), GFP_NOFS
);
1803 faila
= rbio
->faila
;
1804 failb
= rbio
->failb
;
1806 if (rbio
->operation
== BTRFS_RBIO_READ_REBUILD
||
1807 rbio
->operation
== BTRFS_RBIO_REBUILD_MISSING
) {
1808 spin_lock_irq(&rbio
->bio_list_lock
);
1809 set_bit(RBIO_RMW_LOCKED_BIT
, &rbio
->flags
);
1810 spin_unlock_irq(&rbio
->bio_list_lock
);
1813 index_rbio_pages(rbio
);
1815 for (pagenr
= 0; pagenr
< rbio
->stripe_npages
; pagenr
++) {
1817 * Now we just use bitmap to mark the horizontal stripes in
1818 * which we have data when doing parity scrub.
1820 if (rbio
->operation
== BTRFS_RBIO_PARITY_SCRUB
&&
1821 !test_bit(pagenr
, rbio
->dbitmap
))
1824 /* setup our array of pointers with pages
1827 for (stripe
= 0; stripe
< rbio
->real_stripes
; stripe
++) {
1829 * if we're rebuilding a read, we have to use
1830 * pages from the bio list
1832 if ((rbio
->operation
== BTRFS_RBIO_READ_REBUILD
||
1833 rbio
->operation
== BTRFS_RBIO_REBUILD_MISSING
) &&
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_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
< rbio
->stripe_npages
; 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 rbio
->operation
== BTRFS_RBIO_REBUILD_MISSING
) &&
1944 (stripe
== faila
|| stripe
== failb
)) {
1945 page
= page_in_rbio(rbio
, stripe
, pagenr
, 0);
1947 page
= rbio_stripe_page(rbio
, stripe
, pagenr
);
1958 if (rbio
->operation
== BTRFS_RBIO_READ_REBUILD
) {
1960 cache_rbio_pages(rbio
);
1962 clear_bit(RBIO_CACHE_READY_BIT
, &rbio
->flags
);
1964 rbio_orig_end_io(rbio
, err
);
1965 } else if (rbio
->operation
== BTRFS_RBIO_REBUILD_MISSING
) {
1966 rbio_orig_end_io(rbio
, err
);
1967 } else if (err
== 0) {
1971 if (rbio
->operation
== BTRFS_RBIO_WRITE
)
1973 else if (rbio
->operation
== BTRFS_RBIO_PARITY_SCRUB
)
1974 finish_parity_scrub(rbio
, 0);
1978 rbio_orig_end_io(rbio
, err
);
1983 * This is called only for stripes we've read from disk to
1984 * reconstruct the parity.
1986 static void raid_recover_end_io(struct bio
*bio
)
1988 struct btrfs_raid_bio
*rbio
= bio
->bi_private
;
1991 * we only read stripe pages off the disk, set them
1992 * up to date if there were no errors
1995 fail_bio_stripe(rbio
, bio
);
1997 set_bio_pages_uptodate(bio
);
2000 if (!atomic_dec_and_test(&rbio
->stripes_pending
))
2003 if (atomic_read(&rbio
->error
) > rbio
->bbio
->max_errors
)
2004 rbio_orig_end_io(rbio
, -EIO
);
2006 __raid_recover_end_io(rbio
);
2010 * reads everything we need off the disk to reconstruct
2011 * the parity. endio handlers trigger final reconstruction
2012 * when the IO is done.
2014 * This is used both for reads from the higher layers and for
2015 * parity construction required to finish a rmw cycle.
2017 static int __raid56_parity_recover(struct btrfs_raid_bio
*rbio
)
2019 int bios_to_read
= 0;
2020 struct bio_list bio_list
;
2026 bio_list_init(&bio_list
);
2028 ret
= alloc_rbio_pages(rbio
);
2032 atomic_set(&rbio
->error
, 0);
2035 * read everything that hasn't failed. Thanks to the
2036 * stripe cache, it is possible that some or all of these
2037 * pages are going to be uptodate.
2039 for (stripe
= 0; stripe
< rbio
->real_stripes
; stripe
++) {
2040 if (rbio
->faila
== stripe
|| rbio
->failb
== stripe
) {
2041 atomic_inc(&rbio
->error
);
2045 for (pagenr
= 0; pagenr
< rbio
->stripe_npages
; pagenr
++) {
2049 * the rmw code may have already read this
2052 p
= rbio_stripe_page(rbio
, stripe
, pagenr
);
2053 if (PageUptodate(p
))
2056 ret
= rbio_add_io_page(rbio
, &bio_list
,
2057 rbio_stripe_page(rbio
, stripe
, pagenr
),
2058 stripe
, pagenr
, rbio
->stripe_len
);
2064 bios_to_read
= bio_list_size(&bio_list
);
2065 if (!bios_to_read
) {
2067 * we might have no bios to read just because the pages
2068 * were up to date, or we might have no bios to read because
2069 * the devices were gone.
2071 if (atomic_read(&rbio
->error
) <= rbio
->bbio
->max_errors
) {
2072 __raid_recover_end_io(rbio
);
2080 * the bbio may be freed once we submit the last bio. Make sure
2081 * not to touch it after that
2083 atomic_set(&rbio
->stripes_pending
, bios_to_read
);
2085 bio
= bio_list_pop(&bio_list
);
2089 bio
->bi_private
= rbio
;
2090 bio
->bi_end_io
= raid_recover_end_io
;
2091 bio_set_op_attrs(bio
, REQ_OP_READ
, 0);
2093 btrfs_bio_wq_end_io(rbio
->fs_info
, bio
, BTRFS_WQ_ENDIO_RAID56
);
2101 if (rbio
->operation
== BTRFS_RBIO_READ_REBUILD
||
2102 rbio
->operation
== BTRFS_RBIO_REBUILD_MISSING
)
2103 rbio_orig_end_io(rbio
, -EIO
);
2108 * the main entry point for reads from the higher layers. This
2109 * is really only called when the normal read path had a failure,
2110 * so we assume the bio they send down corresponds to a failed part
2113 int raid56_parity_recover(struct btrfs_fs_info
*fs_info
, struct bio
*bio
,
2114 struct btrfs_bio
*bbio
, u64 stripe_len
,
2115 int mirror_num
, int generic_io
)
2117 struct btrfs_raid_bio
*rbio
;
2121 ASSERT(bbio
->mirror_num
== mirror_num
);
2122 btrfs_io_bio(bio
)->mirror_num
= mirror_num
;
2125 rbio
= alloc_rbio(fs_info
, bbio
, stripe_len
);
2128 btrfs_put_bbio(bbio
);
2129 return PTR_ERR(rbio
);
2132 rbio
->operation
= BTRFS_RBIO_READ_REBUILD
;
2133 bio_list_add(&rbio
->bio_list
, bio
);
2134 rbio
->bio_list_bytes
= bio
->bi_iter
.bi_size
;
2136 rbio
->faila
= find_logical_bio_stripe(rbio
, bio
);
2137 if (rbio
->faila
== -1) {
2139 "%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)",
2140 __func__
, (u64
)bio
->bi_iter
.bi_sector
<< 9,
2141 (u64
)bio
->bi_iter
.bi_size
, bbio
->map_type
);
2143 btrfs_put_bbio(bbio
);
2149 btrfs_bio_counter_inc_noblocked(fs_info
);
2150 rbio
->generic_bio_cnt
= 1;
2152 btrfs_get_bbio(bbio
);
2156 * reconstruct from the q stripe if they are
2157 * asking for mirror 3
2159 if (mirror_num
== 3)
2160 rbio
->failb
= rbio
->real_stripes
- 2;
2162 ret
= lock_stripe_add(rbio
);
2165 * __raid56_parity_recover will end the bio with
2166 * any errors it hits. We don't want to return
2167 * its error value up the stack because our caller
2168 * will end up calling bio_endio with any nonzero
2172 __raid56_parity_recover(rbio
);
2174 * our rbio has been added to the list of
2175 * rbios that will be handled after the
2176 * currently lock owner is done
2182 static void rmw_work(struct btrfs_work
*work
)
2184 struct btrfs_raid_bio
*rbio
;
2186 rbio
= container_of(work
, struct btrfs_raid_bio
, work
);
2187 raid56_rmw_stripe(rbio
);
2190 static void read_rebuild_work(struct btrfs_work
*work
)
2192 struct btrfs_raid_bio
*rbio
;
2194 rbio
= container_of(work
, struct btrfs_raid_bio
, work
);
2195 __raid56_parity_recover(rbio
);
2199 * The following code is used to scrub/replace the parity stripe
2201 * Caller must have already increased bio_counter for getting @bbio.
2203 * Note: We need make sure all the pages that add into the scrub/replace
2204 * raid bio are correct and not be changed during the scrub/replace. That
2205 * is those pages just hold metadata or file data with checksum.
2208 struct btrfs_raid_bio
*
2209 raid56_parity_alloc_scrub_rbio(struct btrfs_fs_info
*fs_info
, struct bio
*bio
,
2210 struct btrfs_bio
*bbio
, u64 stripe_len
,
2211 struct btrfs_device
*scrub_dev
,
2212 unsigned long *dbitmap
, int stripe_nsectors
)
2214 struct btrfs_raid_bio
*rbio
;
2217 rbio
= alloc_rbio(fs_info
, bbio
, stripe_len
);
2220 bio_list_add(&rbio
->bio_list
, bio
);
2222 * This is a special bio which is used to hold the completion handler
2223 * and make the scrub rbio is similar to the other types
2225 ASSERT(!bio
->bi_iter
.bi_size
);
2226 rbio
->operation
= BTRFS_RBIO_PARITY_SCRUB
;
2228 for (i
= 0; i
< rbio
->real_stripes
; i
++) {
2229 if (bbio
->stripes
[i
].dev
== scrub_dev
) {
2235 /* Now we just support the sectorsize equals to page size */
2236 ASSERT(fs_info
->sectorsize
== PAGE_SIZE
);
2237 ASSERT(rbio
->stripe_npages
== stripe_nsectors
);
2238 bitmap_copy(rbio
->dbitmap
, dbitmap
, stripe_nsectors
);
2241 * We have already increased bio_counter when getting bbio, record it
2242 * so we can free it at rbio_orig_end_io().
2244 rbio
->generic_bio_cnt
= 1;
2249 /* Used for both parity scrub and missing. */
2250 void raid56_add_scrub_pages(struct btrfs_raid_bio
*rbio
, struct page
*page
,
2256 ASSERT(logical
>= rbio
->bbio
->raid_map
[0]);
2257 ASSERT(logical
+ PAGE_SIZE
<= rbio
->bbio
->raid_map
[0] +
2258 rbio
->stripe_len
* rbio
->nr_data
);
2259 stripe_offset
= (int)(logical
- rbio
->bbio
->raid_map
[0]);
2260 index
= stripe_offset
>> PAGE_SHIFT
;
2261 rbio
->bio_pages
[index
] = page
;
2265 * We just scrub the parity that we have correct data on the same horizontal,
2266 * so we needn't allocate all pages for all the stripes.
2268 static int alloc_rbio_essential_pages(struct btrfs_raid_bio
*rbio
)
2275 for_each_set_bit(bit
, rbio
->dbitmap
, rbio
->stripe_npages
) {
2276 for (i
= 0; i
< rbio
->real_stripes
; i
++) {
2277 index
= i
* rbio
->stripe_npages
+ bit
;
2278 if (rbio
->stripe_pages
[index
])
2281 page
= alloc_page(GFP_NOFS
| __GFP_HIGHMEM
);
2284 rbio
->stripe_pages
[index
] = page
;
2290 static noinline
void finish_parity_scrub(struct btrfs_raid_bio
*rbio
,
2293 struct btrfs_bio
*bbio
= rbio
->bbio
;
2294 void *pointers
[rbio
->real_stripes
];
2295 DECLARE_BITMAP(pbitmap
, rbio
->stripe_npages
);
2296 int nr_data
= rbio
->nr_data
;
2301 struct page
*p_page
= NULL
;
2302 struct page
*q_page
= NULL
;
2303 struct bio_list bio_list
;
2308 bio_list_init(&bio_list
);
2310 if (rbio
->real_stripes
- rbio
->nr_data
== 1) {
2311 p_stripe
= rbio
->real_stripes
- 1;
2312 } else if (rbio
->real_stripes
- rbio
->nr_data
== 2) {
2313 p_stripe
= rbio
->real_stripes
- 2;
2314 q_stripe
= rbio
->real_stripes
- 1;
2319 if (bbio
->num_tgtdevs
&& bbio
->tgtdev_map
[rbio
->scrubp
]) {
2321 bitmap_copy(pbitmap
, rbio
->dbitmap
, rbio
->stripe_npages
);
2325 * Because the higher layers(scrubber) are unlikely to
2326 * use this area of the disk again soon, so don't cache
2329 clear_bit(RBIO_CACHE_READY_BIT
, &rbio
->flags
);
2334 p_page
= alloc_page(GFP_NOFS
| __GFP_HIGHMEM
);
2337 SetPageUptodate(p_page
);
2339 if (q_stripe
!= -1) {
2340 q_page
= alloc_page(GFP_NOFS
| __GFP_HIGHMEM
);
2342 __free_page(p_page
);
2345 SetPageUptodate(q_page
);
2348 atomic_set(&rbio
->error
, 0);
2350 for_each_set_bit(pagenr
, rbio
->dbitmap
, rbio
->stripe_npages
) {
2353 /* first collect one page from each data stripe */
2354 for (stripe
= 0; stripe
< nr_data
; stripe
++) {
2355 p
= page_in_rbio(rbio
, stripe
, pagenr
, 0);
2356 pointers
[stripe
] = kmap(p
);
2359 /* then add the parity stripe */
2360 pointers
[stripe
++] = kmap(p_page
);
2362 if (q_stripe
!= -1) {
2365 * raid6, add the qstripe and call the
2366 * library function to fill in our p/q
2368 pointers
[stripe
++] = kmap(q_page
);
2370 raid6_call
.gen_syndrome(rbio
->real_stripes
, PAGE_SIZE
,
2374 memcpy(pointers
[nr_data
], pointers
[0], PAGE_SIZE
);
2375 run_xor(pointers
+ 1, nr_data
- 1, PAGE_SIZE
);
2378 /* Check scrubbing parity and repair it */
2379 p
= rbio_stripe_page(rbio
, rbio
->scrubp
, pagenr
);
2381 if (memcmp(parity
, pointers
[rbio
->scrubp
], PAGE_SIZE
))
2382 memcpy(parity
, pointers
[rbio
->scrubp
], PAGE_SIZE
);
2384 /* Parity is right, needn't writeback */
2385 bitmap_clear(rbio
->dbitmap
, pagenr
, 1);
2388 for (stripe
= 0; stripe
< rbio
->real_stripes
; stripe
++)
2389 kunmap(page_in_rbio(rbio
, stripe
, pagenr
, 0));
2392 __free_page(p_page
);
2394 __free_page(q_page
);
2398 * time to start writing. Make bios for everything from the
2399 * higher layers (the bio_list in our rbio) and our p/q. Ignore
2402 for_each_set_bit(pagenr
, rbio
->dbitmap
, rbio
->stripe_npages
) {
2405 page
= rbio_stripe_page(rbio
, rbio
->scrubp
, pagenr
);
2406 ret
= rbio_add_io_page(rbio
, &bio_list
,
2407 page
, rbio
->scrubp
, pagenr
, rbio
->stripe_len
);
2415 for_each_set_bit(pagenr
, pbitmap
, rbio
->stripe_npages
) {
2418 page
= rbio_stripe_page(rbio
, rbio
->scrubp
, pagenr
);
2419 ret
= rbio_add_io_page(rbio
, &bio_list
, page
,
2420 bbio
->tgtdev_map
[rbio
->scrubp
],
2421 pagenr
, rbio
->stripe_len
);
2427 nr_data
= bio_list_size(&bio_list
);
2429 /* Every parity is right */
2430 rbio_orig_end_io(rbio
, 0);
2434 atomic_set(&rbio
->stripes_pending
, nr_data
);
2437 bio
= bio_list_pop(&bio_list
);
2441 bio
->bi_private
= rbio
;
2442 bio
->bi_end_io
= raid_write_end_io
;
2443 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
2450 rbio_orig_end_io(rbio
, -EIO
);
2453 static inline int is_data_stripe(struct btrfs_raid_bio
*rbio
, int stripe
)
2455 if (stripe
>= 0 && stripe
< rbio
->nr_data
)
2461 * While we're doing the parity check and repair, we could have errors
2462 * in reading pages off the disk. This checks for errors and if we're
2463 * not able to read the page it'll trigger parity reconstruction. The
2464 * parity scrub will be finished after we've reconstructed the failed
2467 static void validate_rbio_for_parity_scrub(struct btrfs_raid_bio
*rbio
)
2469 if (atomic_read(&rbio
->error
) > rbio
->bbio
->max_errors
)
2472 if (rbio
->faila
>= 0 || rbio
->failb
>= 0) {
2473 int dfail
= 0, failp
= -1;
2475 if (is_data_stripe(rbio
, rbio
->faila
))
2477 else if (is_parity_stripe(rbio
->faila
))
2478 failp
= rbio
->faila
;
2480 if (is_data_stripe(rbio
, rbio
->failb
))
2482 else if (is_parity_stripe(rbio
->failb
))
2483 failp
= rbio
->failb
;
2486 * Because we can not use a scrubbing parity to repair
2487 * the data, so the capability of the repair is declined.
2488 * (In the case of RAID5, we can not repair anything)
2490 if (dfail
> rbio
->bbio
->max_errors
- 1)
2494 * If all data is good, only parity is correctly, just
2495 * repair the parity.
2498 finish_parity_scrub(rbio
, 0);
2503 * Here means we got one corrupted data stripe and one
2504 * corrupted parity on RAID6, if the corrupted parity
2505 * is scrubbing parity, luckily, use the other one to repair
2506 * the data, or we can not repair the data stripe.
2508 if (failp
!= rbio
->scrubp
)
2511 __raid_recover_end_io(rbio
);
2513 finish_parity_scrub(rbio
, 1);
2518 rbio_orig_end_io(rbio
, -EIO
);
2522 * end io for the read phase of the rmw cycle. All the bios here are physical
2523 * stripe bios we've read from the disk so we can recalculate the parity of the
2526 * This will usually kick off finish_rmw once all the bios are read in, but it
2527 * may trigger parity reconstruction if we had any errors along the way
2529 static void raid56_parity_scrub_end_io(struct bio
*bio
)
2531 struct btrfs_raid_bio
*rbio
= bio
->bi_private
;
2534 fail_bio_stripe(rbio
, bio
);
2536 set_bio_pages_uptodate(bio
);
2540 if (!atomic_dec_and_test(&rbio
->stripes_pending
))
2544 * this will normally call finish_rmw to start our write
2545 * but if there are any failed stripes we'll reconstruct
2548 validate_rbio_for_parity_scrub(rbio
);
2551 static void raid56_parity_scrub_stripe(struct btrfs_raid_bio
*rbio
)
2553 int bios_to_read
= 0;
2554 struct bio_list bio_list
;
2560 ret
= alloc_rbio_essential_pages(rbio
);
2564 bio_list_init(&bio_list
);
2566 atomic_set(&rbio
->error
, 0);
2568 * build a list of bios to read all the missing parts of this
2571 for (stripe
= 0; stripe
< rbio
->real_stripes
; stripe
++) {
2572 for_each_set_bit(pagenr
, rbio
->dbitmap
, rbio
->stripe_npages
) {
2575 * we want to find all the pages missing from
2576 * the rbio and read them from the disk. If
2577 * page_in_rbio finds a page in the bio list
2578 * we don't need to read it off the stripe.
2580 page
= page_in_rbio(rbio
, stripe
, pagenr
, 1);
2584 page
= rbio_stripe_page(rbio
, stripe
, pagenr
);
2586 * the bio cache may have handed us an uptodate
2587 * page. If so, be happy and use it
2589 if (PageUptodate(page
))
2592 ret
= rbio_add_io_page(rbio
, &bio_list
, page
,
2593 stripe
, pagenr
, rbio
->stripe_len
);
2599 bios_to_read
= bio_list_size(&bio_list
);
2600 if (!bios_to_read
) {
2602 * this can happen if others have merged with
2603 * us, it means there is nothing left to read.
2604 * But if there are missing devices it may not be
2605 * safe to do the full stripe write yet.
2611 * the bbio may be freed once we submit the last bio. Make sure
2612 * not to touch it after that
2614 atomic_set(&rbio
->stripes_pending
, bios_to_read
);
2616 bio
= bio_list_pop(&bio_list
);
2620 bio
->bi_private
= rbio
;
2621 bio
->bi_end_io
= raid56_parity_scrub_end_io
;
2622 bio_set_op_attrs(bio
, REQ_OP_READ
, 0);
2624 btrfs_bio_wq_end_io(rbio
->fs_info
, bio
, BTRFS_WQ_ENDIO_RAID56
);
2628 /* the actual write will happen once the reads are done */
2632 rbio_orig_end_io(rbio
, -EIO
);
2636 validate_rbio_for_parity_scrub(rbio
);
2639 static void scrub_parity_work(struct btrfs_work
*work
)
2641 struct btrfs_raid_bio
*rbio
;
2643 rbio
= container_of(work
, struct btrfs_raid_bio
, work
);
2644 raid56_parity_scrub_stripe(rbio
);
2647 static void async_scrub_parity(struct btrfs_raid_bio
*rbio
)
2649 btrfs_init_work(&rbio
->work
, btrfs_rmw_helper
,
2650 scrub_parity_work
, NULL
, NULL
);
2652 btrfs_queue_work(rbio
->fs_info
->rmw_workers
, &rbio
->work
);
2655 void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio
*rbio
)
2657 if (!lock_stripe_add(rbio
))
2658 async_scrub_parity(rbio
);
2661 /* The following code is used for dev replace of a missing RAID 5/6 device. */
2663 struct btrfs_raid_bio
*
2664 raid56_alloc_missing_rbio(struct btrfs_fs_info
*fs_info
, struct bio
*bio
,
2665 struct btrfs_bio
*bbio
, u64 length
)
2667 struct btrfs_raid_bio
*rbio
;
2669 rbio
= alloc_rbio(fs_info
, bbio
, length
);
2673 rbio
->operation
= BTRFS_RBIO_REBUILD_MISSING
;
2674 bio_list_add(&rbio
->bio_list
, bio
);
2676 * This is a special bio which is used to hold the completion handler
2677 * and make the scrub rbio is similar to the other types
2679 ASSERT(!bio
->bi_iter
.bi_size
);
2681 rbio
->faila
= find_logical_bio_stripe(rbio
, bio
);
2682 if (rbio
->faila
== -1) {
2689 * When we get bbio, we have already increased bio_counter, record it
2690 * so we can free it at rbio_orig_end_io()
2692 rbio
->generic_bio_cnt
= 1;
2697 static void missing_raid56_work(struct btrfs_work
*work
)
2699 struct btrfs_raid_bio
*rbio
;
2701 rbio
= container_of(work
, struct btrfs_raid_bio
, work
);
2702 __raid56_parity_recover(rbio
);
2705 static void async_missing_raid56(struct btrfs_raid_bio
*rbio
)
2707 btrfs_init_work(&rbio
->work
, btrfs_rmw_helper
,
2708 missing_raid56_work
, NULL
, NULL
);
2710 btrfs_queue_work(rbio
->fs_info
->rmw_workers
, &rbio
->work
);
2713 void raid56_submit_missing_rbio(struct btrfs_raid_bio
*rbio
)
2715 if (!lock_stripe_add(rbio
))
2716 async_missing_raid56(rbio
);