1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2012 Fusion-io All rights reserved.
4 * Copyright (C) 2012 Intel Corp. All rights reserved.
7 #include <linux/sched.h>
9 #include <linux/slab.h>
10 #include <linux/blkdev.h>
11 #include <linux/raid/pq.h>
12 #include <linux/hash.h>
13 #include <linux/list_sort.h>
14 #include <linux/raid/xor.h>
20 #include "async-thread.h"
22 /* set when additional merges to this rbio are not allowed */
23 #define RBIO_RMW_LOCKED_BIT 1
26 * set when this rbio is sitting in the hash, but it is just a cache
29 #define RBIO_CACHE_BIT 2
32 * set when it is safe to trust the stripe_pages for caching
34 #define RBIO_CACHE_READY_BIT 3
36 #define RBIO_CACHE_SIZE 1024
38 #define BTRFS_STRIPE_HASH_TABLE_BITS 11
40 /* Used by the raid56 code to lock stripes for read/modify/write */
41 struct btrfs_stripe_hash {
42 struct list_head hash_list;
46 /* Used by the raid56 code to lock stripes for read/modify/write */
47 struct btrfs_stripe_hash_table {
48 struct list_head stripe_cache;
49 spinlock_t cache_lock;
51 struct btrfs_stripe_hash table[];
56 BTRFS_RBIO_READ_REBUILD,
57 BTRFS_RBIO_PARITY_SCRUB,
58 BTRFS_RBIO_REBUILD_MISSING,
61 struct btrfs_raid_bio {
62 struct btrfs_fs_info *fs_info;
63 struct btrfs_bio *bbio;
65 /* while we're doing rmw on a stripe
66 * we put it into a hash table so we can
67 * lock the stripe and merge more rbios
70 struct list_head hash_list;
73 * LRU list for the stripe cache
75 struct list_head stripe_cache;
78 * for scheduling work in the helper threads
80 struct btrfs_work work;
83 * bio list and bio_list_lock are used
84 * to add more bios into the stripe
85 * in hopes of avoiding the full rmw
87 struct bio_list bio_list;
88 spinlock_t bio_list_lock;
90 /* also protected by the bio_list_lock, the
91 * plug list is used by the plugging code
92 * to collect partial bios while plugged. The
93 * stripe locking code also uses it to hand off
94 * the stripe lock to the next pending IO
96 struct list_head plug_list;
99 * flags that tell us if it is safe to
100 * merge with this bio
104 /* size of each individual stripe on disk */
107 /* number of data stripes (no p/q) */
114 * set if we're doing a parity rebuild
115 * for a read from higher up, which is handled
116 * differently from a parity rebuild as part of
119 enum btrfs_rbio_ops operation;
121 /* first bad stripe */
124 /* second bad stripe (for raid6 use) */
129 * number of pages needed to represent the full
135 * size of all the bios in the bio_list. This
136 * helps us decide if the rbio maps to a full
145 atomic_t stripes_pending;
149 * these are two arrays of pointers. We allocate the
150 * rbio big enough to hold them both and setup their
151 * locations when the rbio is allocated
154 /* pointers to pages that we allocated for
155 * reading/writing stripes directly from the disk (including P/Q)
157 struct page **stripe_pages;
160 * pointers to the pages in the bio_list. Stored
161 * here for faster lookup
163 struct page **bio_pages;
166 * bitmap to record which horizontal stripe has data
168 unsigned long *dbitmap;
170 /* allocated with real_stripes-many pointers for finish_*() calls */
171 void **finish_pointers;
173 /* allocated with stripe_npages-many bits for finish_*() calls */
174 unsigned long *finish_pbitmap;
177 static int __raid56_parity_recover(struct btrfs_raid_bio *rbio);
178 static noinline void finish_rmw(struct btrfs_raid_bio *rbio);
179 static void rmw_work(struct btrfs_work *work);
180 static void read_rebuild_work(struct btrfs_work *work);
181 static int fail_bio_stripe(struct btrfs_raid_bio *rbio, struct bio *bio);
182 static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed);
183 static void __free_raid_bio(struct btrfs_raid_bio *rbio);
184 static void index_rbio_pages(struct btrfs_raid_bio *rbio);
185 static int alloc_rbio_pages(struct btrfs_raid_bio *rbio);
187 static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
189 static void scrub_parity_work(struct btrfs_work *work);
191 static void start_async_work(struct btrfs_raid_bio *rbio, btrfs_func_t work_func)
193 btrfs_init_work(&rbio->work, work_func, NULL, NULL);
194 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
198 * the stripe hash table is used for locking, and to collect
199 * bios in hopes of making a full stripe
201 int btrfs_alloc_stripe_hash_table(struct btrfs_fs_info *info)
203 struct btrfs_stripe_hash_table *table;
204 struct btrfs_stripe_hash_table *x;
205 struct btrfs_stripe_hash *cur;
206 struct btrfs_stripe_hash *h;
207 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 = kvzalloc(struct_size(table, table, num_entries), GFP_KERNEL);
224 spin_lock_init(&table->cache_lock);
225 INIT_LIST_HEAD(&table->stripe_cache);
229 for (i = 0; i < num_entries; i++) {
231 INIT_LIST_HEAD(&cur->hash_list);
232 spin_lock_init(&cur->lock);
235 x = cmpxchg(&info->stripe_hash_table, NULL, table);
242 * caching an rbio means to copy anything from the
243 * bio_pages array into the stripe_pages array. We
244 * use the page uptodate bit in the stripe cache array
245 * to indicate if it has valid data
247 * once the caching is done, we set the cache ready
250 static void cache_rbio_pages(struct btrfs_raid_bio *rbio)
255 ret = alloc_rbio_pages(rbio);
259 for (i = 0; i < rbio->nr_pages; i++) {
260 if (!rbio->bio_pages[i])
263 copy_highpage(rbio->stripe_pages[i], rbio->bio_pages[i]);
264 SetPageUptodate(rbio->stripe_pages[i]);
266 set_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
270 * we hash on the first logical address of the stripe
272 static int rbio_bucket(struct btrfs_raid_bio *rbio)
274 u64 num = rbio->bbio->raid_map[0];
277 * we shift down quite a bit. We're using byte
278 * addressing, and most of the lower bits are zeros.
279 * This tends to upset hash_64, and it consistently
280 * returns just one or two different values.
282 * shifting off the lower bits fixes things.
284 return hash_64(num >> 16, BTRFS_STRIPE_HASH_TABLE_BITS);
288 * stealing an rbio means taking all the uptodate pages from the stripe
289 * array in the source rbio and putting them into the destination rbio
291 static void steal_rbio(struct btrfs_raid_bio *src, struct btrfs_raid_bio *dest)
297 if (!test_bit(RBIO_CACHE_READY_BIT, &src->flags))
300 for (i = 0; i < dest->nr_pages; i++) {
301 s = src->stripe_pages[i];
302 if (!s || !PageUptodate(s)) {
306 d = dest->stripe_pages[i];
310 dest->stripe_pages[i] = s;
311 src->stripe_pages[i] = NULL;
316 * merging means we take the bio_list from the victim and
317 * splice it into the destination. The victim should
318 * be discarded afterwards.
320 * must be called with dest->rbio_list_lock held
322 static void merge_rbio(struct btrfs_raid_bio *dest,
323 struct btrfs_raid_bio *victim)
325 bio_list_merge(&dest->bio_list, &victim->bio_list);
326 dest->bio_list_bytes += victim->bio_list_bytes;
327 dest->generic_bio_cnt += victim->generic_bio_cnt;
328 bio_list_init(&victim->bio_list);
332 * used to prune items that are in the cache. The caller
333 * must hold the hash table lock.
335 static void __remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
337 int bucket = rbio_bucket(rbio);
338 struct btrfs_stripe_hash_table *table;
339 struct btrfs_stripe_hash *h;
343 * check the bit again under the hash table lock.
345 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
348 table = rbio->fs_info->stripe_hash_table;
349 h = table->table + bucket;
351 /* hold the lock for the bucket because we may be
352 * removing it from the hash table
357 * hold the lock for the bio list because we need
358 * to make sure the bio list is empty
360 spin_lock(&rbio->bio_list_lock);
362 if (test_and_clear_bit(RBIO_CACHE_BIT, &rbio->flags)) {
363 list_del_init(&rbio->stripe_cache);
364 table->cache_size -= 1;
367 /* if the bio list isn't empty, this rbio is
368 * still involved in an IO. We take it out
369 * of the cache list, and drop the ref that
370 * was held for the list.
372 * If the bio_list was empty, we also remove
373 * the rbio from the hash_table, and drop
374 * the corresponding ref
376 if (bio_list_empty(&rbio->bio_list)) {
377 if (!list_empty(&rbio->hash_list)) {
378 list_del_init(&rbio->hash_list);
379 refcount_dec(&rbio->refs);
380 BUG_ON(!list_empty(&rbio->plug_list));
385 spin_unlock(&rbio->bio_list_lock);
386 spin_unlock(&h->lock);
389 __free_raid_bio(rbio);
393 * prune a given rbio from the cache
395 static void remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
397 struct btrfs_stripe_hash_table *table;
400 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
403 table = rbio->fs_info->stripe_hash_table;
405 spin_lock_irqsave(&table->cache_lock, flags);
406 __remove_rbio_from_cache(rbio);
407 spin_unlock_irqrestore(&table->cache_lock, flags);
411 * remove everything in the cache
413 static void btrfs_clear_rbio_cache(struct btrfs_fs_info *info)
415 struct btrfs_stripe_hash_table *table;
417 struct btrfs_raid_bio *rbio;
419 table = info->stripe_hash_table;
421 spin_lock_irqsave(&table->cache_lock, flags);
422 while (!list_empty(&table->stripe_cache)) {
423 rbio = list_entry(table->stripe_cache.next,
424 struct btrfs_raid_bio,
426 __remove_rbio_from_cache(rbio);
428 spin_unlock_irqrestore(&table->cache_lock, flags);
432 * remove all cached entries and free the hash table
435 void btrfs_free_stripe_hash_table(struct btrfs_fs_info *info)
437 if (!info->stripe_hash_table)
439 btrfs_clear_rbio_cache(info);
440 kvfree(info->stripe_hash_table);
441 info->stripe_hash_table = NULL;
445 * insert an rbio into the stripe cache. It
446 * must have already been prepared by calling
449 * If this rbio was already cached, it gets
450 * moved to the front of the lru.
452 * If the size of the rbio cache is too big, we
455 static void cache_rbio(struct btrfs_raid_bio *rbio)
457 struct btrfs_stripe_hash_table *table;
460 if (!test_bit(RBIO_CACHE_READY_BIT, &rbio->flags))
463 table = rbio->fs_info->stripe_hash_table;
465 spin_lock_irqsave(&table->cache_lock, flags);
466 spin_lock(&rbio->bio_list_lock);
468 /* bump our ref if we were not in the list before */
469 if (!test_and_set_bit(RBIO_CACHE_BIT, &rbio->flags))
470 refcount_inc(&rbio->refs);
472 if (!list_empty(&rbio->stripe_cache)){
473 list_move(&rbio->stripe_cache, &table->stripe_cache);
475 list_add(&rbio->stripe_cache, &table->stripe_cache);
476 table->cache_size += 1;
479 spin_unlock(&rbio->bio_list_lock);
481 if (table->cache_size > RBIO_CACHE_SIZE) {
482 struct btrfs_raid_bio *found;
484 found = list_entry(table->stripe_cache.prev,
485 struct btrfs_raid_bio,
489 __remove_rbio_from_cache(found);
492 spin_unlock_irqrestore(&table->cache_lock, flags);
496 * helper function to run the xor_blocks api. It is only
497 * able to do MAX_XOR_BLOCKS at a time, so we need to
500 static void run_xor(void **pages, int src_cnt, ssize_t len)
504 void *dest = pages[src_cnt];
507 xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS);
508 xor_blocks(xor_src_cnt, len, dest, pages + src_off);
510 src_cnt -= xor_src_cnt;
511 src_off += xor_src_cnt;
516 * Returns true if the bio list inside this rbio covers an entire stripe (no
519 static int rbio_is_full(struct btrfs_raid_bio *rbio)
522 unsigned long size = rbio->bio_list_bytes;
525 spin_lock_irqsave(&rbio->bio_list_lock, flags);
526 if (size != rbio->nr_data * rbio->stripe_len)
528 BUG_ON(size > rbio->nr_data * rbio->stripe_len);
529 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
535 * returns 1 if it is safe to merge two rbios together.
536 * The merging is safe if the two rbios correspond to
537 * the same stripe and if they are both going in the same
538 * direction (read vs write), and if neither one is
539 * locked for final IO
541 * The caller is responsible for locking such that
542 * rmw_locked is safe to test
544 static int rbio_can_merge(struct btrfs_raid_bio *last,
545 struct btrfs_raid_bio *cur)
547 if (test_bit(RBIO_RMW_LOCKED_BIT, &last->flags) ||
548 test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags))
552 * we can't merge with cached rbios, since the
553 * idea is that when we merge the destination
554 * rbio is going to run our IO for us. We can
555 * steal from cached rbios though, other functions
558 if (test_bit(RBIO_CACHE_BIT, &last->flags) ||
559 test_bit(RBIO_CACHE_BIT, &cur->flags))
562 if (last->bbio->raid_map[0] !=
563 cur->bbio->raid_map[0])
566 /* we can't merge with different operations */
567 if (last->operation != cur->operation)
570 * We've need read the full stripe from the drive.
571 * check and repair the parity and write the new results.
573 * We're not allowed to add any new bios to the
574 * bio list here, anyone else that wants to
575 * change this stripe needs to do their own rmw.
577 if (last->operation == BTRFS_RBIO_PARITY_SCRUB)
580 if (last->operation == BTRFS_RBIO_REBUILD_MISSING)
583 if (last->operation == BTRFS_RBIO_READ_REBUILD) {
584 int fa = last->faila;
585 int fb = last->failb;
586 int cur_fa = cur->faila;
587 int cur_fb = cur->failb;
589 if (last->faila >= last->failb) {
594 if (cur->faila >= cur->failb) {
599 if (fa != cur_fa || fb != cur_fb)
605 static int rbio_stripe_page_index(struct btrfs_raid_bio *rbio, int stripe,
608 return stripe * rbio->stripe_npages + index;
612 * these are just the pages from the rbio array, not from anything
613 * the FS sent down to us
615 static struct page *rbio_stripe_page(struct btrfs_raid_bio *rbio, int stripe,
618 return rbio->stripe_pages[rbio_stripe_page_index(rbio, stripe, index)];
622 * helper to index into the pstripe
624 static struct page *rbio_pstripe_page(struct btrfs_raid_bio *rbio, int index)
626 return rbio_stripe_page(rbio, rbio->nr_data, index);
630 * helper to index into the qstripe, returns null
631 * if there is no qstripe
633 static struct page *rbio_qstripe_page(struct btrfs_raid_bio *rbio, int index)
635 if (rbio->nr_data + 1 == rbio->real_stripes)
637 return rbio_stripe_page(rbio, rbio->nr_data + 1, index);
641 * The first stripe in the table for a logical address
642 * has the lock. rbios are added in one of three ways:
644 * 1) Nobody has the stripe locked yet. The rbio is given
645 * the lock and 0 is returned. The caller must start the IO
648 * 2) Someone has the stripe locked, but we're able to merge
649 * with the lock owner. The rbio is freed and the IO will
650 * start automatically along with the existing rbio. 1 is returned.
652 * 3) Someone has the stripe locked, but we're not able to merge.
653 * The rbio is added to the lock owner's plug list, or merged into
654 * an rbio already on the plug list. When the lock owner unlocks,
655 * the next rbio on the list is run and the IO is started automatically.
658 * If we return 0, the caller still owns the rbio and must continue with
659 * IO submission. If we return 1, the caller must assume the rbio has
660 * already been freed.
662 static noinline int lock_stripe_add(struct btrfs_raid_bio *rbio)
664 struct btrfs_stripe_hash *h;
665 struct btrfs_raid_bio *cur;
666 struct btrfs_raid_bio *pending;
668 struct btrfs_raid_bio *freeit = NULL;
669 struct btrfs_raid_bio *cache_drop = NULL;
672 h = rbio->fs_info->stripe_hash_table->table + rbio_bucket(rbio);
674 spin_lock_irqsave(&h->lock, flags);
675 list_for_each_entry(cur, &h->hash_list, hash_list) {
676 if (cur->bbio->raid_map[0] != rbio->bbio->raid_map[0])
679 spin_lock(&cur->bio_list_lock);
681 /* Can we steal this cached rbio's pages? */
682 if (bio_list_empty(&cur->bio_list) &&
683 list_empty(&cur->plug_list) &&
684 test_bit(RBIO_CACHE_BIT, &cur->flags) &&
685 !test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) {
686 list_del_init(&cur->hash_list);
687 refcount_dec(&cur->refs);
689 steal_rbio(cur, rbio);
691 spin_unlock(&cur->bio_list_lock);
696 /* Can we merge into the lock owner? */
697 if (rbio_can_merge(cur, rbio)) {
698 merge_rbio(cur, rbio);
699 spin_unlock(&cur->bio_list_lock);
707 * We couldn't merge with the running rbio, see if we can merge
708 * with the pending ones. We don't have to check for rmw_locked
709 * because there is no way they are inside finish_rmw right now
711 list_for_each_entry(pending, &cur->plug_list, plug_list) {
712 if (rbio_can_merge(pending, rbio)) {
713 merge_rbio(pending, rbio);
714 spin_unlock(&cur->bio_list_lock);
722 * No merging, put us on the tail of the plug list, our rbio
723 * will be started with the currently running rbio unlocks
725 list_add_tail(&rbio->plug_list, &cur->plug_list);
726 spin_unlock(&cur->bio_list_lock);
731 refcount_inc(&rbio->refs);
732 list_add(&rbio->hash_list, &h->hash_list);
734 spin_unlock_irqrestore(&h->lock, flags);
736 remove_rbio_from_cache(cache_drop);
738 __free_raid_bio(freeit);
743 * called as rmw or parity rebuild is completed. If the plug list has more
744 * rbios waiting for this stripe, the next one on the list will be started
746 static noinline void unlock_stripe(struct btrfs_raid_bio *rbio)
749 struct btrfs_stripe_hash *h;
753 bucket = rbio_bucket(rbio);
754 h = rbio->fs_info->stripe_hash_table->table + bucket;
756 if (list_empty(&rbio->plug_list))
759 spin_lock_irqsave(&h->lock, flags);
760 spin_lock(&rbio->bio_list_lock);
762 if (!list_empty(&rbio->hash_list)) {
764 * if we're still cached and there is no other IO
765 * to perform, just leave this rbio here for others
766 * to steal from later
768 if (list_empty(&rbio->plug_list) &&
769 test_bit(RBIO_CACHE_BIT, &rbio->flags)) {
771 clear_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
772 BUG_ON(!bio_list_empty(&rbio->bio_list));
776 list_del_init(&rbio->hash_list);
777 refcount_dec(&rbio->refs);
780 * we use the plug list to hold all the rbios
781 * waiting for the chance to lock this stripe.
782 * hand the lock over to one of them.
784 if (!list_empty(&rbio->plug_list)) {
785 struct btrfs_raid_bio *next;
786 struct list_head *head = rbio->plug_list.next;
788 next = list_entry(head, struct btrfs_raid_bio,
791 list_del_init(&rbio->plug_list);
793 list_add(&next->hash_list, &h->hash_list);
794 refcount_inc(&next->refs);
795 spin_unlock(&rbio->bio_list_lock);
796 spin_unlock_irqrestore(&h->lock, flags);
798 if (next->operation == BTRFS_RBIO_READ_REBUILD)
799 start_async_work(next, read_rebuild_work);
800 else if (next->operation == BTRFS_RBIO_REBUILD_MISSING) {
801 steal_rbio(rbio, next);
802 start_async_work(next, read_rebuild_work);
803 } else if (next->operation == BTRFS_RBIO_WRITE) {
804 steal_rbio(rbio, next);
805 start_async_work(next, rmw_work);
806 } else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) {
807 steal_rbio(rbio, next);
808 start_async_work(next, scrub_parity_work);
815 spin_unlock(&rbio->bio_list_lock);
816 spin_unlock_irqrestore(&h->lock, flags);
820 remove_rbio_from_cache(rbio);
823 static void __free_raid_bio(struct btrfs_raid_bio *rbio)
827 if (!refcount_dec_and_test(&rbio->refs))
830 WARN_ON(!list_empty(&rbio->stripe_cache));
831 WARN_ON(!list_empty(&rbio->hash_list));
832 WARN_ON(!bio_list_empty(&rbio->bio_list));
834 for (i = 0; i < rbio->nr_pages; i++) {
835 if (rbio->stripe_pages[i]) {
836 __free_page(rbio->stripe_pages[i]);
837 rbio->stripe_pages[i] = NULL;
841 btrfs_put_bbio(rbio->bbio);
845 static void rbio_endio_bio_list(struct bio *cur, blk_status_t err)
852 cur->bi_status = err;
859 * this frees the rbio and runs through all the bios in the
860 * bio_list and calls end_io on them
862 static void rbio_orig_end_io(struct btrfs_raid_bio *rbio, blk_status_t err)
864 struct bio *cur = bio_list_get(&rbio->bio_list);
867 if (rbio->generic_bio_cnt)
868 btrfs_bio_counter_sub(rbio->fs_info, rbio->generic_bio_cnt);
871 * At this moment, rbio->bio_list is empty, however since rbio does not
872 * always have RBIO_RMW_LOCKED_BIT set and rbio is still linked on the
873 * hash list, rbio may be merged with others so that rbio->bio_list
875 * Once unlock_stripe() is done, rbio->bio_list will not be updated any
876 * more and we can call bio_endio() on all queued bios.
879 extra = bio_list_get(&rbio->bio_list);
880 __free_raid_bio(rbio);
882 rbio_endio_bio_list(cur, err);
884 rbio_endio_bio_list(extra, err);
888 * end io function used by finish_rmw. When we finally
889 * get here, we've written a full stripe
891 static void raid_write_end_io(struct bio *bio)
893 struct btrfs_raid_bio *rbio = bio->bi_private;
894 blk_status_t err = bio->bi_status;
898 fail_bio_stripe(rbio, bio);
902 if (!atomic_dec_and_test(&rbio->stripes_pending))
907 /* OK, we have read all the stripes we need to. */
908 max_errors = (rbio->operation == BTRFS_RBIO_PARITY_SCRUB) ?
909 0 : rbio->bbio->max_errors;
910 if (atomic_read(&rbio->error) > max_errors)
913 rbio_orig_end_io(rbio, err);
917 * the read/modify/write code wants to use the original bio for
918 * any pages it included, and then use the rbio for everything
919 * else. This function decides if a given index (stripe number)
920 * and page number in that stripe fall inside the original bio
923 * if you set bio_list_only, you'll get a NULL back for any ranges
924 * that are outside the bio_list
926 * This doesn't take any refs on anything, you get a bare page pointer
927 * and the caller must bump refs as required.
929 * You must call index_rbio_pages once before you can trust
930 * the answers from this function.
932 static struct page *page_in_rbio(struct btrfs_raid_bio *rbio,
933 int index, int pagenr, int bio_list_only)
936 struct page *p = NULL;
938 chunk_page = index * (rbio->stripe_len >> PAGE_SHIFT) + pagenr;
940 spin_lock_irq(&rbio->bio_list_lock);
941 p = rbio->bio_pages[chunk_page];
942 spin_unlock_irq(&rbio->bio_list_lock);
944 if (p || bio_list_only)
947 return rbio->stripe_pages[chunk_page];
951 * number of pages we need for the entire stripe across all the
954 static unsigned long rbio_nr_pages(unsigned long stripe_len, int nr_stripes)
956 return DIV_ROUND_UP(stripe_len, PAGE_SIZE) * nr_stripes;
960 * allocation and initial setup for the btrfs_raid_bio. Not
961 * this does not allocate any pages for rbio->pages.
963 static struct btrfs_raid_bio *alloc_rbio(struct btrfs_fs_info *fs_info,
964 struct btrfs_bio *bbio,
967 struct btrfs_raid_bio *rbio;
969 int real_stripes = bbio->num_stripes - bbio->num_tgtdevs;
970 int num_pages = rbio_nr_pages(stripe_len, real_stripes);
971 int stripe_npages = DIV_ROUND_UP(stripe_len, PAGE_SIZE);
974 rbio = kzalloc(sizeof(*rbio) +
975 sizeof(*rbio->stripe_pages) * num_pages +
976 sizeof(*rbio->bio_pages) * num_pages +
977 sizeof(*rbio->finish_pointers) * real_stripes +
978 sizeof(*rbio->dbitmap) * BITS_TO_LONGS(stripe_npages) +
979 sizeof(*rbio->finish_pbitmap) *
980 BITS_TO_LONGS(stripe_npages),
983 return ERR_PTR(-ENOMEM);
985 bio_list_init(&rbio->bio_list);
986 INIT_LIST_HEAD(&rbio->plug_list);
987 spin_lock_init(&rbio->bio_list_lock);
988 INIT_LIST_HEAD(&rbio->stripe_cache);
989 INIT_LIST_HEAD(&rbio->hash_list);
991 rbio->fs_info = fs_info;
992 rbio->stripe_len = stripe_len;
993 rbio->nr_pages = num_pages;
994 rbio->real_stripes = real_stripes;
995 rbio->stripe_npages = stripe_npages;
998 refcount_set(&rbio->refs, 1);
999 atomic_set(&rbio->error, 0);
1000 atomic_set(&rbio->stripes_pending, 0);
1003 * the stripe_pages, bio_pages, etc arrays point to the extra
1004 * memory we allocated past the end of the rbio
1007 #define CONSUME_ALLOC(ptr, count) do { \
1009 p = (unsigned char *)p + sizeof(*(ptr)) * (count); \
1011 CONSUME_ALLOC(rbio->stripe_pages, num_pages);
1012 CONSUME_ALLOC(rbio->bio_pages, num_pages);
1013 CONSUME_ALLOC(rbio->finish_pointers, real_stripes);
1014 CONSUME_ALLOC(rbio->dbitmap, BITS_TO_LONGS(stripe_npages));
1015 CONSUME_ALLOC(rbio->finish_pbitmap, BITS_TO_LONGS(stripe_npages));
1016 #undef CONSUME_ALLOC
1018 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1019 nr_data = real_stripes - 1;
1020 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1021 nr_data = real_stripes - 2;
1025 rbio->nr_data = nr_data;
1029 /* allocate pages for all the stripes in the bio, including parity */
1030 static int alloc_rbio_pages(struct btrfs_raid_bio *rbio)
1035 for (i = 0; i < rbio->nr_pages; i++) {
1036 if (rbio->stripe_pages[i])
1038 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1041 rbio->stripe_pages[i] = page;
1046 /* only allocate pages for p/q stripes */
1047 static int alloc_rbio_parity_pages(struct btrfs_raid_bio *rbio)
1052 i = rbio_stripe_page_index(rbio, rbio->nr_data, 0);
1054 for (; i < rbio->nr_pages; i++) {
1055 if (rbio->stripe_pages[i])
1057 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1060 rbio->stripe_pages[i] = page;
1066 * add a single page from a specific stripe into our list of bios for IO
1067 * this will try to merge into existing bios if possible, and returns
1068 * zero if all went well.
1070 static int rbio_add_io_page(struct btrfs_raid_bio *rbio,
1071 struct bio_list *bio_list,
1074 unsigned long page_index,
1075 unsigned long bio_max_len)
1077 struct bio *last = bio_list->tail;
1080 struct btrfs_bio_stripe *stripe;
1083 stripe = &rbio->bbio->stripes[stripe_nr];
1084 disk_start = stripe->physical + (page_index << PAGE_SHIFT);
1086 /* if the device is missing, just fail this stripe */
1087 if (!stripe->dev->bdev)
1088 return fail_rbio_index(rbio, stripe_nr);
1090 /* see if we can add this page onto our existing bio */
1092 u64 last_end = last->bi_iter.bi_sector << 9;
1093 last_end += last->bi_iter.bi_size;
1096 * we can't merge these if they are from different
1097 * devices or if they are not contiguous
1099 if (last_end == disk_start && !last->bi_status &&
1100 last->bi_disk == stripe->dev->bdev->bd_disk &&
1101 last->bi_partno == stripe->dev->bdev->bd_partno) {
1102 ret = bio_add_page(last, page, PAGE_SIZE, 0);
1103 if (ret == PAGE_SIZE)
1108 /* put a new bio on the list */
1109 bio = btrfs_io_bio_alloc(bio_max_len >> PAGE_SHIFT ?: 1);
1110 btrfs_io_bio(bio)->device = stripe->dev;
1111 bio->bi_iter.bi_size = 0;
1112 bio_set_dev(bio, stripe->dev->bdev);
1113 bio->bi_iter.bi_sector = disk_start >> 9;
1115 bio_add_page(bio, page, PAGE_SIZE, 0);
1116 bio_list_add(bio_list, bio);
1121 * while we're doing the read/modify/write cycle, we could
1122 * have errors in reading pages off the disk. This checks
1123 * for errors and if we're not able to read the page it'll
1124 * trigger parity reconstruction. The rmw will be finished
1125 * after we've reconstructed the failed stripes
1127 static void validate_rbio_for_rmw(struct btrfs_raid_bio *rbio)
1129 if (rbio->faila >= 0 || rbio->failb >= 0) {
1130 BUG_ON(rbio->faila == rbio->real_stripes - 1);
1131 __raid56_parity_recover(rbio);
1138 * helper function to walk our bio list and populate the bio_pages array with
1139 * the result. This seems expensive, but it is faster than constantly
1140 * searching through the bio list as we setup the IO in finish_rmw or stripe
1143 * This must be called before you trust the answers from page_in_rbio
1145 static void index_rbio_pages(struct btrfs_raid_bio *rbio)
1149 unsigned long stripe_offset;
1150 unsigned long page_index;
1152 spin_lock_irq(&rbio->bio_list_lock);
1153 bio_list_for_each(bio, &rbio->bio_list) {
1154 struct bio_vec bvec;
1155 struct bvec_iter iter;
1158 start = bio->bi_iter.bi_sector << 9;
1159 stripe_offset = start - rbio->bbio->raid_map[0];
1160 page_index = stripe_offset >> PAGE_SHIFT;
1162 if (bio_flagged(bio, BIO_CLONED))
1163 bio->bi_iter = btrfs_io_bio(bio)->iter;
1165 bio_for_each_segment(bvec, bio, iter) {
1166 rbio->bio_pages[page_index + i] = bvec.bv_page;
1170 spin_unlock_irq(&rbio->bio_list_lock);
1174 * this is called from one of two situations. We either
1175 * have a full stripe from the higher layers, or we've read all
1176 * the missing bits off disk.
1178 * This will calculate the parity and then send down any
1181 static noinline void finish_rmw(struct btrfs_raid_bio *rbio)
1183 struct btrfs_bio *bbio = rbio->bbio;
1184 void **pointers = rbio->finish_pointers;
1185 int nr_data = rbio->nr_data;
1189 struct bio_list bio_list;
1193 bio_list_init(&bio_list);
1195 if (rbio->real_stripes - rbio->nr_data == 1)
1196 has_qstripe = false;
1197 else if (rbio->real_stripes - rbio->nr_data == 2)
1202 /* at this point we either have a full stripe,
1203 * or we've read the full stripe from the drive.
1204 * recalculate the parity and write the new results.
1206 * We're not allowed to add any new bios to the
1207 * bio list here, anyone else that wants to
1208 * change this stripe needs to do their own rmw.
1210 spin_lock_irq(&rbio->bio_list_lock);
1211 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1212 spin_unlock_irq(&rbio->bio_list_lock);
1214 atomic_set(&rbio->error, 0);
1217 * now that we've set rmw_locked, run through the
1218 * bio list one last time and map the page pointers
1220 * We don't cache full rbios because we're assuming
1221 * the higher layers are unlikely to use this area of
1222 * the disk again soon. If they do use it again,
1223 * hopefully they will send another full bio.
1225 index_rbio_pages(rbio);
1226 if (!rbio_is_full(rbio))
1227 cache_rbio_pages(rbio);
1229 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
1231 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1233 /* first collect one page from each data stripe */
1234 for (stripe = 0; stripe < nr_data; stripe++) {
1235 p = page_in_rbio(rbio, stripe, pagenr, 0);
1236 pointers[stripe] = kmap(p);
1239 /* then add the parity stripe */
1240 p = rbio_pstripe_page(rbio, pagenr);
1242 pointers[stripe++] = kmap(p);
1247 * raid6, add the qstripe and call the
1248 * library function to fill in our p/q
1250 p = rbio_qstripe_page(rbio, pagenr);
1252 pointers[stripe++] = kmap(p);
1254 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
1258 copy_page(pointers[nr_data], pointers[0]);
1259 run_xor(pointers + 1, nr_data - 1, PAGE_SIZE);
1263 for (stripe = 0; stripe < rbio->real_stripes; stripe++)
1264 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
1268 * time to start writing. Make bios for everything from the
1269 * higher layers (the bio_list in our rbio) and our p/q. Ignore
1272 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1273 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1275 if (stripe < rbio->nr_data) {
1276 page = page_in_rbio(rbio, stripe, pagenr, 1);
1280 page = rbio_stripe_page(rbio, stripe, pagenr);
1283 ret = rbio_add_io_page(rbio, &bio_list,
1284 page, stripe, pagenr, rbio->stripe_len);
1290 if (likely(!bbio->num_tgtdevs))
1293 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1294 if (!bbio->tgtdev_map[stripe])
1297 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1299 if (stripe < rbio->nr_data) {
1300 page = page_in_rbio(rbio, stripe, pagenr, 1);
1304 page = rbio_stripe_page(rbio, stripe, pagenr);
1307 ret = rbio_add_io_page(rbio, &bio_list, page,
1308 rbio->bbio->tgtdev_map[stripe],
1309 pagenr, rbio->stripe_len);
1316 atomic_set(&rbio->stripes_pending, bio_list_size(&bio_list));
1317 BUG_ON(atomic_read(&rbio->stripes_pending) == 0);
1319 while ((bio = bio_list_pop(&bio_list))) {
1320 bio->bi_private = rbio;
1321 bio->bi_end_io = raid_write_end_io;
1322 bio->bi_opf = REQ_OP_WRITE;
1329 rbio_orig_end_io(rbio, BLK_STS_IOERR);
1331 while ((bio = bio_list_pop(&bio_list)))
1336 * helper to find the stripe number for a given bio. Used to figure out which
1337 * stripe has failed. This expects the bio to correspond to a physical disk,
1338 * so it looks up based on physical sector numbers.
1340 static int find_bio_stripe(struct btrfs_raid_bio *rbio,
1343 u64 physical = bio->bi_iter.bi_sector;
1345 struct btrfs_bio_stripe *stripe;
1349 for (i = 0; i < rbio->bbio->num_stripes; i++) {
1350 stripe = &rbio->bbio->stripes[i];
1351 if (in_range(physical, stripe->physical, rbio->stripe_len) &&
1352 stripe->dev->bdev &&
1353 bio->bi_disk == stripe->dev->bdev->bd_disk &&
1354 bio->bi_partno == stripe->dev->bdev->bd_partno) {
1362 * helper to find the stripe number for a given
1363 * bio (before mapping). Used to figure out which stripe has
1364 * failed. This looks up based on logical block numbers.
1366 static int find_logical_bio_stripe(struct btrfs_raid_bio *rbio,
1369 u64 logical = bio->bi_iter.bi_sector << 9;
1372 for (i = 0; i < rbio->nr_data; i++) {
1373 u64 stripe_start = rbio->bbio->raid_map[i];
1375 if (in_range(logical, stripe_start, rbio->stripe_len))
1382 * returns -EIO if we had too many failures
1384 static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed)
1386 unsigned long flags;
1389 spin_lock_irqsave(&rbio->bio_list_lock, flags);
1391 /* we already know this stripe is bad, move on */
1392 if (rbio->faila == failed || rbio->failb == failed)
1395 if (rbio->faila == -1) {
1396 /* first failure on this rbio */
1397 rbio->faila = failed;
1398 atomic_inc(&rbio->error);
1399 } else if (rbio->failb == -1) {
1400 /* second failure on this rbio */
1401 rbio->failb = failed;
1402 atomic_inc(&rbio->error);
1407 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
1413 * helper to fail a stripe based on a physical disk
1416 static int fail_bio_stripe(struct btrfs_raid_bio *rbio,
1419 int failed = find_bio_stripe(rbio, bio);
1424 return fail_rbio_index(rbio, failed);
1428 * this sets each page in the bio uptodate. It should only be used on private
1429 * rbio pages, nothing that comes in from the higher layers
1431 static void set_bio_pages_uptodate(struct bio *bio)
1433 struct bio_vec *bvec;
1434 struct bvec_iter_all iter_all;
1436 ASSERT(!bio_flagged(bio, BIO_CLONED));
1438 bio_for_each_segment_all(bvec, bio, iter_all)
1439 SetPageUptodate(bvec->bv_page);
1443 * end io for the read phase of the rmw cycle. All the bios here are physical
1444 * stripe bios we've read from the disk so we can recalculate the parity of the
1447 * This will usually kick off finish_rmw once all the bios are read in, but it
1448 * may trigger parity reconstruction if we had any errors along the way
1450 static void raid_rmw_end_io(struct bio *bio)
1452 struct btrfs_raid_bio *rbio = bio->bi_private;
1455 fail_bio_stripe(rbio, bio);
1457 set_bio_pages_uptodate(bio);
1461 if (!atomic_dec_and_test(&rbio->stripes_pending))
1464 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
1468 * this will normally call finish_rmw to start our write
1469 * but if there are any failed stripes we'll reconstruct
1472 validate_rbio_for_rmw(rbio);
1477 rbio_orig_end_io(rbio, BLK_STS_IOERR);
1481 * the stripe must be locked by the caller. It will
1482 * unlock after all the writes are done
1484 static int raid56_rmw_stripe(struct btrfs_raid_bio *rbio)
1486 int bios_to_read = 0;
1487 struct bio_list bio_list;
1493 bio_list_init(&bio_list);
1495 ret = alloc_rbio_pages(rbio);
1499 index_rbio_pages(rbio);
1501 atomic_set(&rbio->error, 0);
1503 * build a list of bios to read all the missing parts of this
1506 for (stripe = 0; stripe < rbio->nr_data; stripe++) {
1507 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1510 * we want to find all the pages missing from
1511 * the rbio and read them from the disk. If
1512 * page_in_rbio finds a page in the bio list
1513 * we don't need to read it off the stripe.
1515 page = page_in_rbio(rbio, stripe, pagenr, 1);
1519 page = rbio_stripe_page(rbio, stripe, pagenr);
1521 * the bio cache may have handed us an uptodate
1522 * page. If so, be happy and use it
1524 if (PageUptodate(page))
1527 ret = rbio_add_io_page(rbio, &bio_list, page,
1528 stripe, pagenr, rbio->stripe_len);
1534 bios_to_read = bio_list_size(&bio_list);
1535 if (!bios_to_read) {
1537 * this can happen if others have merged with
1538 * us, it means there is nothing left to read.
1539 * But if there are missing devices it may not be
1540 * safe to do the full stripe write yet.
1546 * the bbio may be freed once we submit the last bio. Make sure
1547 * not to touch it after that
1549 atomic_set(&rbio->stripes_pending, bios_to_read);
1550 while ((bio = bio_list_pop(&bio_list))) {
1551 bio->bi_private = rbio;
1552 bio->bi_end_io = raid_rmw_end_io;
1553 bio->bi_opf = REQ_OP_READ;
1555 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
1559 /* the actual write will happen once the reads are done */
1563 rbio_orig_end_io(rbio, BLK_STS_IOERR);
1565 while ((bio = bio_list_pop(&bio_list)))
1571 validate_rbio_for_rmw(rbio);
1576 * if the upper layers pass in a full stripe, we thank them by only allocating
1577 * enough pages to hold the parity, and sending it all down quickly.
1579 static int full_stripe_write(struct btrfs_raid_bio *rbio)
1583 ret = alloc_rbio_parity_pages(rbio);
1585 __free_raid_bio(rbio);
1589 ret = lock_stripe_add(rbio);
1596 * partial stripe writes get handed over to async helpers.
1597 * We're really hoping to merge a few more writes into this
1598 * rbio before calculating new parity
1600 static int partial_stripe_write(struct btrfs_raid_bio *rbio)
1604 ret = lock_stripe_add(rbio);
1606 start_async_work(rbio, rmw_work);
1611 * sometimes while we were reading from the drive to
1612 * recalculate parity, enough new bios come into create
1613 * a full stripe. So we do a check here to see if we can
1614 * go directly to finish_rmw
1616 static int __raid56_parity_write(struct btrfs_raid_bio *rbio)
1618 /* head off into rmw land if we don't have a full stripe */
1619 if (!rbio_is_full(rbio))
1620 return partial_stripe_write(rbio);
1621 return full_stripe_write(rbio);
1625 * We use plugging call backs to collect full stripes.
1626 * Any time we get a partial stripe write while plugged
1627 * we collect it into a list. When the unplug comes down,
1628 * we sort the list by logical block number and merge
1629 * everything we can into the same rbios
1631 struct btrfs_plug_cb {
1632 struct blk_plug_cb cb;
1633 struct btrfs_fs_info *info;
1634 struct list_head rbio_list;
1635 struct btrfs_work work;
1639 * rbios on the plug list are sorted for easier merging.
1641 static int plug_cmp(void *priv, struct list_head *a, struct list_head *b)
1643 struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio,
1645 struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio,
1647 u64 a_sector = ra->bio_list.head->bi_iter.bi_sector;
1648 u64 b_sector = rb->bio_list.head->bi_iter.bi_sector;
1650 if (a_sector < b_sector)
1652 if (a_sector > b_sector)
1657 static void run_plug(struct btrfs_plug_cb *plug)
1659 struct btrfs_raid_bio *cur;
1660 struct btrfs_raid_bio *last = NULL;
1663 * sort our plug list then try to merge
1664 * everything we can in hopes of creating full
1667 list_sort(NULL, &plug->rbio_list, plug_cmp);
1668 while (!list_empty(&plug->rbio_list)) {
1669 cur = list_entry(plug->rbio_list.next,
1670 struct btrfs_raid_bio, plug_list);
1671 list_del_init(&cur->plug_list);
1673 if (rbio_is_full(cur)) {
1676 /* we have a full stripe, send it down */
1677 ret = full_stripe_write(cur);
1682 if (rbio_can_merge(last, cur)) {
1683 merge_rbio(last, cur);
1684 __free_raid_bio(cur);
1688 __raid56_parity_write(last);
1693 __raid56_parity_write(last);
1699 * if the unplug comes from schedule, we have to push the
1700 * work off to a helper thread
1702 static void unplug_work(struct btrfs_work *work)
1704 struct btrfs_plug_cb *plug;
1705 plug = container_of(work, struct btrfs_plug_cb, work);
1709 static void btrfs_raid_unplug(struct blk_plug_cb *cb, bool from_schedule)
1711 struct btrfs_plug_cb *plug;
1712 plug = container_of(cb, struct btrfs_plug_cb, cb);
1714 if (from_schedule) {
1715 btrfs_init_work(&plug->work, unplug_work, NULL, NULL);
1716 btrfs_queue_work(plug->info->rmw_workers,
1724 * our main entry point for writes from the rest of the FS.
1726 int raid56_parity_write(struct btrfs_fs_info *fs_info, struct bio *bio,
1727 struct btrfs_bio *bbio, u64 stripe_len)
1729 struct btrfs_raid_bio *rbio;
1730 struct btrfs_plug_cb *plug = NULL;
1731 struct blk_plug_cb *cb;
1734 rbio = alloc_rbio(fs_info, bbio, stripe_len);
1736 btrfs_put_bbio(bbio);
1737 return PTR_ERR(rbio);
1739 bio_list_add(&rbio->bio_list, bio);
1740 rbio->bio_list_bytes = bio->bi_iter.bi_size;
1741 rbio->operation = BTRFS_RBIO_WRITE;
1743 btrfs_bio_counter_inc_noblocked(fs_info);
1744 rbio->generic_bio_cnt = 1;
1747 * don't plug on full rbios, just get them out the door
1748 * as quickly as we can
1750 if (rbio_is_full(rbio)) {
1751 ret = full_stripe_write(rbio);
1753 btrfs_bio_counter_dec(fs_info);
1757 cb = blk_check_plugged(btrfs_raid_unplug, fs_info, sizeof(*plug));
1759 plug = container_of(cb, struct btrfs_plug_cb, cb);
1761 plug->info = fs_info;
1762 INIT_LIST_HEAD(&plug->rbio_list);
1764 list_add_tail(&rbio->plug_list, &plug->rbio_list);
1767 ret = __raid56_parity_write(rbio);
1769 btrfs_bio_counter_dec(fs_info);
1775 * all parity reconstruction happens here. We've read in everything
1776 * we can find from the drives and this does the heavy lifting of
1777 * sorting the good from the bad.
1779 static void __raid_recover_end_io(struct btrfs_raid_bio *rbio)
1783 int faila = -1, failb = -1;
1788 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
1790 err = BLK_STS_RESOURCE;
1794 faila = rbio->faila;
1795 failb = rbio->failb;
1797 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1798 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
1799 spin_lock_irq(&rbio->bio_list_lock);
1800 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1801 spin_unlock_irq(&rbio->bio_list_lock);
1804 index_rbio_pages(rbio);
1806 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1808 * Now we just use bitmap to mark the horizontal stripes in
1809 * which we have data when doing parity scrub.
1811 if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB &&
1812 !test_bit(pagenr, rbio->dbitmap))
1815 /* setup our array of pointers with pages
1818 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1820 * if we're rebuilding a read, we have to use
1821 * pages from the bio list
1823 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1824 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
1825 (stripe == faila || stripe == failb)) {
1826 page = page_in_rbio(rbio, stripe, pagenr, 0);
1828 page = rbio_stripe_page(rbio, stripe, pagenr);
1830 pointers[stripe] = kmap(page);
1833 /* all raid6 handling here */
1834 if (rbio->bbio->map_type & BTRFS_BLOCK_GROUP_RAID6) {
1836 * single failure, rebuild from parity raid5
1840 if (faila == rbio->nr_data) {
1842 * Just the P stripe has failed, without
1843 * a bad data or Q stripe.
1844 * TODO, we should redo the xor here.
1846 err = BLK_STS_IOERR;
1850 * a single failure in raid6 is rebuilt
1851 * in the pstripe code below
1856 /* make sure our ps and qs are in order */
1860 /* if the q stripe is failed, do a pstripe reconstruction
1862 * If both the q stripe and the P stripe are failed, we're
1863 * here due to a crc mismatch and we can't give them the
1866 if (rbio->bbio->raid_map[failb] == RAID6_Q_STRIPE) {
1867 if (rbio->bbio->raid_map[faila] ==
1869 err = BLK_STS_IOERR;
1873 * otherwise we have one bad data stripe and
1874 * a good P stripe. raid5!
1879 if (rbio->bbio->raid_map[failb] == RAID5_P_STRIPE) {
1880 raid6_datap_recov(rbio->real_stripes,
1881 PAGE_SIZE, faila, pointers);
1883 raid6_2data_recov(rbio->real_stripes,
1884 PAGE_SIZE, faila, failb,
1890 /* rebuild from P stripe here (raid5 or raid6) */
1891 BUG_ON(failb != -1);
1893 /* Copy parity block into failed block to start with */
1894 copy_page(pointers[faila], pointers[rbio->nr_data]);
1896 /* rearrange the pointer array */
1897 p = pointers[faila];
1898 for (stripe = faila; stripe < rbio->nr_data - 1; stripe++)
1899 pointers[stripe] = pointers[stripe + 1];
1900 pointers[rbio->nr_data - 1] = p;
1902 /* xor in the rest */
1903 run_xor(pointers, rbio->nr_data - 1, PAGE_SIZE);
1905 /* if we're doing this rebuild as part of an rmw, go through
1906 * and set all of our private rbio pages in the
1907 * failed stripes as uptodate. This way finish_rmw will
1908 * know they can be trusted. If this was a read reconstruction,
1909 * other endio functions will fiddle the uptodate bits
1911 if (rbio->operation == BTRFS_RBIO_WRITE) {
1912 for (i = 0; i < rbio->stripe_npages; i++) {
1914 page = rbio_stripe_page(rbio, faila, i);
1915 SetPageUptodate(page);
1918 page = rbio_stripe_page(rbio, failb, i);
1919 SetPageUptodate(page);
1923 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1925 * if we're rebuilding a read, we have to use
1926 * pages from the bio list
1928 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1929 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
1930 (stripe == faila || stripe == failb)) {
1931 page = page_in_rbio(rbio, stripe, pagenr, 0);
1933 page = rbio_stripe_page(rbio, stripe, pagenr);
1945 * Similar to READ_REBUILD, REBUILD_MISSING at this point also has a
1946 * valid rbio which is consistent with ondisk content, thus such a
1947 * valid rbio can be cached to avoid further disk reads.
1949 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1950 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
1952 * - In case of two failures, where rbio->failb != -1:
1954 * Do not cache this rbio since the above read reconstruction
1955 * (raid6_datap_recov() or raid6_2data_recov()) may have
1956 * changed some content of stripes which are not identical to
1957 * on-disk content any more, otherwise, a later write/recover
1958 * may steal stripe_pages from this rbio and end up with
1959 * corruptions or rebuild failures.
1961 * - In case of single failure, where rbio->failb == -1:
1963 * Cache this rbio iff the above read reconstruction is
1964 * executed without problems.
1966 if (err == BLK_STS_OK && rbio->failb < 0)
1967 cache_rbio_pages(rbio);
1969 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
1971 rbio_orig_end_io(rbio, err);
1972 } else if (err == BLK_STS_OK) {
1976 if (rbio->operation == BTRFS_RBIO_WRITE)
1978 else if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB)
1979 finish_parity_scrub(rbio, 0);
1983 rbio_orig_end_io(rbio, err);
1988 * This is called only for stripes we've read from disk to
1989 * reconstruct the parity.
1991 static void raid_recover_end_io(struct bio *bio)
1993 struct btrfs_raid_bio *rbio = bio->bi_private;
1996 * we only read stripe pages off the disk, set them
1997 * up to date if there were no errors
2000 fail_bio_stripe(rbio, bio);
2002 set_bio_pages_uptodate(bio);
2005 if (!atomic_dec_and_test(&rbio->stripes_pending))
2008 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
2009 rbio_orig_end_io(rbio, BLK_STS_IOERR);
2011 __raid_recover_end_io(rbio);
2015 * reads everything we need off the disk to reconstruct
2016 * the parity. endio handlers trigger final reconstruction
2017 * when the IO is done.
2019 * This is used both for reads from the higher layers and for
2020 * parity construction required to finish a rmw cycle.
2022 static int __raid56_parity_recover(struct btrfs_raid_bio *rbio)
2024 int bios_to_read = 0;
2025 struct bio_list bio_list;
2031 bio_list_init(&bio_list);
2033 ret = alloc_rbio_pages(rbio);
2037 atomic_set(&rbio->error, 0);
2040 * read everything that hasn't failed. Thanks to the
2041 * stripe cache, it is possible that some or all of these
2042 * pages are going to be uptodate.
2044 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
2045 if (rbio->faila == stripe || rbio->failb == stripe) {
2046 atomic_inc(&rbio->error);
2050 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
2054 * the rmw code may have already read this
2057 p = rbio_stripe_page(rbio, stripe, pagenr);
2058 if (PageUptodate(p))
2061 ret = rbio_add_io_page(rbio, &bio_list,
2062 rbio_stripe_page(rbio, stripe, pagenr),
2063 stripe, pagenr, rbio->stripe_len);
2069 bios_to_read = bio_list_size(&bio_list);
2070 if (!bios_to_read) {
2072 * we might have no bios to read just because the pages
2073 * were up to date, or we might have no bios to read because
2074 * the devices were gone.
2076 if (atomic_read(&rbio->error) <= rbio->bbio->max_errors) {
2077 __raid_recover_end_io(rbio);
2085 * the bbio may be freed once we submit the last bio. Make sure
2086 * not to touch it after that
2088 atomic_set(&rbio->stripes_pending, bios_to_read);
2089 while ((bio = bio_list_pop(&bio_list))) {
2090 bio->bi_private = rbio;
2091 bio->bi_end_io = raid_recover_end_io;
2092 bio->bi_opf = REQ_OP_READ;
2094 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
2102 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
2103 rbio->operation == BTRFS_RBIO_REBUILD_MISSING)
2104 rbio_orig_end_io(rbio, BLK_STS_IOERR);
2106 while ((bio = bio_list_pop(&bio_list)))
2113 * the main entry point for reads from the higher layers. This
2114 * is really only called when the normal read path had a failure,
2115 * so we assume the bio they send down corresponds to a failed part
2118 int raid56_parity_recover(struct btrfs_fs_info *fs_info, struct bio *bio,
2119 struct btrfs_bio *bbio, u64 stripe_len,
2120 int mirror_num, int generic_io)
2122 struct btrfs_raid_bio *rbio;
2126 ASSERT(bbio->mirror_num == mirror_num);
2127 btrfs_io_bio(bio)->mirror_num = mirror_num;
2130 rbio = alloc_rbio(fs_info, bbio, stripe_len);
2133 btrfs_put_bbio(bbio);
2134 return PTR_ERR(rbio);
2137 rbio->operation = BTRFS_RBIO_READ_REBUILD;
2138 bio_list_add(&rbio->bio_list, bio);
2139 rbio->bio_list_bytes = bio->bi_iter.bi_size;
2141 rbio->faila = find_logical_bio_stripe(rbio, bio);
2142 if (rbio->faila == -1) {
2144 "%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)",
2145 __func__, bio->bi_iter.bi_sector << 9,
2146 (u64)bio->bi_iter.bi_size, bbio->map_type);
2148 btrfs_put_bbio(bbio);
2154 btrfs_bio_counter_inc_noblocked(fs_info);
2155 rbio->generic_bio_cnt = 1;
2157 btrfs_get_bbio(bbio);
2162 * for 'mirror == 2', reconstruct from all other stripes.
2163 * for 'mirror_num > 2', select a stripe to fail on every retry.
2165 if (mirror_num > 2) {
2167 * 'mirror == 3' is to fail the p stripe and
2168 * reconstruct from the q stripe. 'mirror > 3' is to
2169 * fail a data stripe and reconstruct from p+q stripe.
2171 rbio->failb = rbio->real_stripes - (mirror_num - 1);
2172 ASSERT(rbio->failb > 0);
2173 if (rbio->failb <= rbio->faila)
2177 ret = lock_stripe_add(rbio);
2180 * __raid56_parity_recover will end the bio with
2181 * any errors it hits. We don't want to return
2182 * its error value up the stack because our caller
2183 * will end up calling bio_endio with any nonzero
2187 __raid56_parity_recover(rbio);
2189 * our rbio has been added to the list of
2190 * rbios that will be handled after the
2191 * currently lock owner is done
2197 static void rmw_work(struct btrfs_work *work)
2199 struct btrfs_raid_bio *rbio;
2201 rbio = container_of(work, struct btrfs_raid_bio, work);
2202 raid56_rmw_stripe(rbio);
2205 static void read_rebuild_work(struct btrfs_work *work)
2207 struct btrfs_raid_bio *rbio;
2209 rbio = container_of(work, struct btrfs_raid_bio, work);
2210 __raid56_parity_recover(rbio);
2214 * The following code is used to scrub/replace the parity stripe
2216 * Caller must have already increased bio_counter for getting @bbio.
2218 * Note: We need make sure all the pages that add into the scrub/replace
2219 * raid bio are correct and not be changed during the scrub/replace. That
2220 * is those pages just hold metadata or file data with checksum.
2223 struct btrfs_raid_bio *
2224 raid56_parity_alloc_scrub_rbio(struct btrfs_fs_info *fs_info, struct bio *bio,
2225 struct btrfs_bio *bbio, u64 stripe_len,
2226 struct btrfs_device *scrub_dev,
2227 unsigned long *dbitmap, int stripe_nsectors)
2229 struct btrfs_raid_bio *rbio;
2232 rbio = alloc_rbio(fs_info, bbio, stripe_len);
2235 bio_list_add(&rbio->bio_list, bio);
2237 * This is a special bio which is used to hold the completion handler
2238 * and make the scrub rbio is similar to the other types
2240 ASSERT(!bio->bi_iter.bi_size);
2241 rbio->operation = BTRFS_RBIO_PARITY_SCRUB;
2244 * After mapping bbio with BTRFS_MAP_WRITE, parities have been sorted
2245 * to the end position, so this search can start from the first parity
2248 for (i = rbio->nr_data; i < rbio->real_stripes; i++) {
2249 if (bbio->stripes[i].dev == scrub_dev) {
2254 ASSERT(i < rbio->real_stripes);
2256 /* Now we just support the sectorsize equals to page size */
2257 ASSERT(fs_info->sectorsize == PAGE_SIZE);
2258 ASSERT(rbio->stripe_npages == stripe_nsectors);
2259 bitmap_copy(rbio->dbitmap, dbitmap, stripe_nsectors);
2262 * We have already increased bio_counter when getting bbio, record it
2263 * so we can free it at rbio_orig_end_io().
2265 rbio->generic_bio_cnt = 1;
2270 /* Used for both parity scrub and missing. */
2271 void raid56_add_scrub_pages(struct btrfs_raid_bio *rbio, struct page *page,
2277 ASSERT(logical >= rbio->bbio->raid_map[0]);
2278 ASSERT(logical + PAGE_SIZE <= rbio->bbio->raid_map[0] +
2279 rbio->stripe_len * rbio->nr_data);
2280 stripe_offset = (int)(logical - rbio->bbio->raid_map[0]);
2281 index = stripe_offset >> PAGE_SHIFT;
2282 rbio->bio_pages[index] = page;
2286 * We just scrub the parity that we have correct data on the same horizontal,
2287 * so we needn't allocate all pages for all the stripes.
2289 static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio)
2296 for_each_set_bit(bit, rbio->dbitmap, rbio->stripe_npages) {
2297 for (i = 0; i < rbio->real_stripes; i++) {
2298 index = i * rbio->stripe_npages + bit;
2299 if (rbio->stripe_pages[index])
2302 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2305 rbio->stripe_pages[index] = page;
2311 static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
2314 struct btrfs_bio *bbio = rbio->bbio;
2315 void **pointers = rbio->finish_pointers;
2316 unsigned long *pbitmap = rbio->finish_pbitmap;
2317 int nr_data = rbio->nr_data;
2321 struct page *p_page = NULL;
2322 struct page *q_page = NULL;
2323 struct bio_list bio_list;
2328 bio_list_init(&bio_list);
2330 if (rbio->real_stripes - rbio->nr_data == 1)
2331 has_qstripe = false;
2332 else if (rbio->real_stripes - rbio->nr_data == 2)
2337 if (bbio->num_tgtdevs && bbio->tgtdev_map[rbio->scrubp]) {
2339 bitmap_copy(pbitmap, rbio->dbitmap, rbio->stripe_npages);
2343 * Because the higher layers(scrubber) are unlikely to
2344 * use this area of the disk again soon, so don't cache
2347 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2352 p_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2355 SetPageUptodate(p_page);
2358 q_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2360 __free_page(p_page);
2363 SetPageUptodate(q_page);
2366 atomic_set(&rbio->error, 0);
2368 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2371 /* first collect one page from each data stripe */
2372 for (stripe = 0; stripe < nr_data; stripe++) {
2373 p = page_in_rbio(rbio, stripe, pagenr, 0);
2374 pointers[stripe] = kmap(p);
2377 /* then add the parity stripe */
2378 pointers[stripe++] = kmap(p_page);
2382 * raid6, add the qstripe and call the
2383 * library function to fill in our p/q
2385 pointers[stripe++] = kmap(q_page);
2387 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
2391 copy_page(pointers[nr_data], pointers[0]);
2392 run_xor(pointers + 1, nr_data - 1, PAGE_SIZE);
2395 /* Check scrubbing parity and repair it */
2396 p = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2398 if (memcmp(parity, pointers[rbio->scrubp], PAGE_SIZE))
2399 copy_page(parity, pointers[rbio->scrubp]);
2401 /* Parity is right, needn't writeback */
2402 bitmap_clear(rbio->dbitmap, pagenr, 1);
2405 for (stripe = 0; stripe < nr_data; stripe++)
2406 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
2410 __free_page(p_page);
2412 __free_page(q_page);
2416 * time to start writing. Make bios for everything from the
2417 * higher layers (the bio_list in our rbio) and our p/q. Ignore
2420 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2423 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2424 ret = rbio_add_io_page(rbio, &bio_list,
2425 page, rbio->scrubp, pagenr, rbio->stripe_len);
2433 for_each_set_bit(pagenr, pbitmap, rbio->stripe_npages) {
2436 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2437 ret = rbio_add_io_page(rbio, &bio_list, page,
2438 bbio->tgtdev_map[rbio->scrubp],
2439 pagenr, rbio->stripe_len);
2445 nr_data = bio_list_size(&bio_list);
2447 /* Every parity is right */
2448 rbio_orig_end_io(rbio, BLK_STS_OK);
2452 atomic_set(&rbio->stripes_pending, nr_data);
2454 while ((bio = bio_list_pop(&bio_list))) {
2455 bio->bi_private = rbio;
2456 bio->bi_end_io = raid_write_end_io;
2457 bio->bi_opf = REQ_OP_WRITE;
2464 rbio_orig_end_io(rbio, BLK_STS_IOERR);
2466 while ((bio = bio_list_pop(&bio_list)))
2470 static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe)
2472 if (stripe >= 0 && stripe < rbio->nr_data)
2478 * While we're doing the parity check and repair, we could have errors
2479 * in reading pages off the disk. This checks for errors and if we're
2480 * not able to read the page it'll trigger parity reconstruction. The
2481 * parity scrub will be finished after we've reconstructed the failed
2484 static void validate_rbio_for_parity_scrub(struct btrfs_raid_bio *rbio)
2486 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
2489 if (rbio->faila >= 0 || rbio->failb >= 0) {
2490 int dfail = 0, failp = -1;
2492 if (is_data_stripe(rbio, rbio->faila))
2494 else if (is_parity_stripe(rbio->faila))
2495 failp = rbio->faila;
2497 if (is_data_stripe(rbio, rbio->failb))
2499 else if (is_parity_stripe(rbio->failb))
2500 failp = rbio->failb;
2503 * Because we can not use a scrubbing parity to repair
2504 * the data, so the capability of the repair is declined.
2505 * (In the case of RAID5, we can not repair anything)
2507 if (dfail > rbio->bbio->max_errors - 1)
2511 * If all data is good, only parity is correctly, just
2512 * repair the parity.
2515 finish_parity_scrub(rbio, 0);
2520 * Here means we got one corrupted data stripe and one
2521 * corrupted parity on RAID6, if the corrupted parity
2522 * is scrubbing parity, luckily, use the other one to repair
2523 * the data, or we can not repair the data stripe.
2525 if (failp != rbio->scrubp)
2528 __raid_recover_end_io(rbio);
2530 finish_parity_scrub(rbio, 1);
2535 rbio_orig_end_io(rbio, BLK_STS_IOERR);
2539 * end io for the read phase of the rmw cycle. All the bios here are physical
2540 * stripe bios we've read from the disk so we can recalculate the parity of the
2543 * This will usually kick off finish_rmw once all the bios are read in, but it
2544 * may trigger parity reconstruction if we had any errors along the way
2546 static void raid56_parity_scrub_end_io(struct bio *bio)
2548 struct btrfs_raid_bio *rbio = bio->bi_private;
2551 fail_bio_stripe(rbio, bio);
2553 set_bio_pages_uptodate(bio);
2557 if (!atomic_dec_and_test(&rbio->stripes_pending))
2561 * this will normally call finish_rmw to start our write
2562 * but if there are any failed stripes we'll reconstruct
2565 validate_rbio_for_parity_scrub(rbio);
2568 static void raid56_parity_scrub_stripe(struct btrfs_raid_bio *rbio)
2570 int bios_to_read = 0;
2571 struct bio_list bio_list;
2577 bio_list_init(&bio_list);
2579 ret = alloc_rbio_essential_pages(rbio);
2583 atomic_set(&rbio->error, 0);
2585 * build a list of bios to read all the missing parts of this
2588 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
2589 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2592 * we want to find all the pages missing from
2593 * the rbio and read them from the disk. If
2594 * page_in_rbio finds a page in the bio list
2595 * we don't need to read it off the stripe.
2597 page = page_in_rbio(rbio, stripe, pagenr, 1);
2601 page = rbio_stripe_page(rbio, stripe, pagenr);
2603 * the bio cache may have handed us an uptodate
2604 * page. If so, be happy and use it
2606 if (PageUptodate(page))
2609 ret = rbio_add_io_page(rbio, &bio_list, page,
2610 stripe, pagenr, rbio->stripe_len);
2616 bios_to_read = bio_list_size(&bio_list);
2617 if (!bios_to_read) {
2619 * this can happen if others have merged with
2620 * us, it means there is nothing left to read.
2621 * But if there are missing devices it may not be
2622 * safe to do the full stripe write yet.
2628 * the bbio may be freed once we submit the last bio. Make sure
2629 * not to touch it after that
2631 atomic_set(&rbio->stripes_pending, bios_to_read);
2632 while ((bio = bio_list_pop(&bio_list))) {
2633 bio->bi_private = rbio;
2634 bio->bi_end_io = raid56_parity_scrub_end_io;
2635 bio->bi_opf = REQ_OP_READ;
2637 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
2641 /* the actual write will happen once the reads are done */
2645 rbio_orig_end_io(rbio, BLK_STS_IOERR);
2647 while ((bio = bio_list_pop(&bio_list)))
2653 validate_rbio_for_parity_scrub(rbio);
2656 static void scrub_parity_work(struct btrfs_work *work)
2658 struct btrfs_raid_bio *rbio;
2660 rbio = container_of(work, struct btrfs_raid_bio, work);
2661 raid56_parity_scrub_stripe(rbio);
2664 void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio)
2666 if (!lock_stripe_add(rbio))
2667 start_async_work(rbio, scrub_parity_work);
2670 /* The following code is used for dev replace of a missing RAID 5/6 device. */
2672 struct btrfs_raid_bio *
2673 raid56_alloc_missing_rbio(struct btrfs_fs_info *fs_info, struct bio *bio,
2674 struct btrfs_bio *bbio, u64 length)
2676 struct btrfs_raid_bio *rbio;
2678 rbio = alloc_rbio(fs_info, bbio, length);
2682 rbio->operation = BTRFS_RBIO_REBUILD_MISSING;
2683 bio_list_add(&rbio->bio_list, bio);
2685 * This is a special bio which is used to hold the completion handler
2686 * and make the scrub rbio is similar to the other types
2688 ASSERT(!bio->bi_iter.bi_size);
2690 rbio->faila = find_logical_bio_stripe(rbio, bio);
2691 if (rbio->faila == -1) {
2698 * When we get bbio, we have already increased bio_counter, record it
2699 * so we can free it at rbio_orig_end_io()
2701 rbio->generic_bio_cnt = 1;
2706 void raid56_submit_missing_rbio(struct btrfs_raid_bio *rbio)
2708 if (!lock_stripe_add(rbio))
2709 start_async_work(rbio, read_rebuild_work);