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)
257 ret = alloc_rbio_pages(rbio);
261 for (i = 0; i < rbio->nr_pages; i++) {
262 if (!rbio->bio_pages[i])
265 s = kmap(rbio->bio_pages[i]);
266 d = kmap(rbio->stripe_pages[i]);
270 kunmap(rbio->bio_pages[i]);
271 kunmap(rbio->stripe_pages[i]);
272 SetPageUptodate(rbio->stripe_pages[i]);
274 set_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
278 * we hash on the first logical address of the stripe
280 static int rbio_bucket(struct btrfs_raid_bio *rbio)
282 u64 num = rbio->bbio->raid_map[0];
285 * we shift down quite a bit. We're using byte
286 * addressing, and most of the lower bits are zeros.
287 * This tends to upset hash_64, and it consistently
288 * returns just one or two different values.
290 * shifting off the lower bits fixes things.
292 return hash_64(num >> 16, BTRFS_STRIPE_HASH_TABLE_BITS);
296 * stealing an rbio means taking all the uptodate pages from the stripe
297 * array in the source rbio and putting them into the destination rbio
299 static void steal_rbio(struct btrfs_raid_bio *src, struct btrfs_raid_bio *dest)
305 if (!test_bit(RBIO_CACHE_READY_BIT, &src->flags))
308 for (i = 0; i < dest->nr_pages; i++) {
309 s = src->stripe_pages[i];
310 if (!s || !PageUptodate(s)) {
314 d = dest->stripe_pages[i];
318 dest->stripe_pages[i] = s;
319 src->stripe_pages[i] = NULL;
324 * merging means we take the bio_list from the victim and
325 * splice it into the destination. The victim should
326 * be discarded afterwards.
328 * must be called with dest->rbio_list_lock held
330 static void merge_rbio(struct btrfs_raid_bio *dest,
331 struct btrfs_raid_bio *victim)
333 bio_list_merge(&dest->bio_list, &victim->bio_list);
334 dest->bio_list_bytes += victim->bio_list_bytes;
335 dest->generic_bio_cnt += victim->generic_bio_cnt;
336 bio_list_init(&victim->bio_list);
340 * used to prune items that are in the cache. The caller
341 * must hold the hash table lock.
343 static void __remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
345 int bucket = rbio_bucket(rbio);
346 struct btrfs_stripe_hash_table *table;
347 struct btrfs_stripe_hash *h;
351 * check the bit again under the hash table lock.
353 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
356 table = rbio->fs_info->stripe_hash_table;
357 h = table->table + bucket;
359 /* hold the lock for the bucket because we may be
360 * removing it from the hash table
365 * hold the lock for the bio list because we need
366 * to make sure the bio list is empty
368 spin_lock(&rbio->bio_list_lock);
370 if (test_and_clear_bit(RBIO_CACHE_BIT, &rbio->flags)) {
371 list_del_init(&rbio->stripe_cache);
372 table->cache_size -= 1;
375 /* if the bio list isn't empty, this rbio is
376 * still involved in an IO. We take it out
377 * of the cache list, and drop the ref that
378 * was held for the list.
380 * If the bio_list was empty, we also remove
381 * the rbio from the hash_table, and drop
382 * the corresponding ref
384 if (bio_list_empty(&rbio->bio_list)) {
385 if (!list_empty(&rbio->hash_list)) {
386 list_del_init(&rbio->hash_list);
387 refcount_dec(&rbio->refs);
388 BUG_ON(!list_empty(&rbio->plug_list));
393 spin_unlock(&rbio->bio_list_lock);
394 spin_unlock(&h->lock);
397 __free_raid_bio(rbio);
401 * prune a given rbio from the cache
403 static void remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
405 struct btrfs_stripe_hash_table *table;
408 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
411 table = rbio->fs_info->stripe_hash_table;
413 spin_lock_irqsave(&table->cache_lock, flags);
414 __remove_rbio_from_cache(rbio);
415 spin_unlock_irqrestore(&table->cache_lock, flags);
419 * remove everything in the cache
421 static void btrfs_clear_rbio_cache(struct btrfs_fs_info *info)
423 struct btrfs_stripe_hash_table *table;
425 struct btrfs_raid_bio *rbio;
427 table = info->stripe_hash_table;
429 spin_lock_irqsave(&table->cache_lock, flags);
430 while (!list_empty(&table->stripe_cache)) {
431 rbio = list_entry(table->stripe_cache.next,
432 struct btrfs_raid_bio,
434 __remove_rbio_from_cache(rbio);
436 spin_unlock_irqrestore(&table->cache_lock, flags);
440 * remove all cached entries and free the hash table
443 void btrfs_free_stripe_hash_table(struct btrfs_fs_info *info)
445 if (!info->stripe_hash_table)
447 btrfs_clear_rbio_cache(info);
448 kvfree(info->stripe_hash_table);
449 info->stripe_hash_table = NULL;
453 * insert an rbio into the stripe cache. It
454 * must have already been prepared by calling
457 * If this rbio was already cached, it gets
458 * moved to the front of the lru.
460 * If the size of the rbio cache is too big, we
463 static void cache_rbio(struct btrfs_raid_bio *rbio)
465 struct btrfs_stripe_hash_table *table;
468 if (!test_bit(RBIO_CACHE_READY_BIT, &rbio->flags))
471 table = rbio->fs_info->stripe_hash_table;
473 spin_lock_irqsave(&table->cache_lock, flags);
474 spin_lock(&rbio->bio_list_lock);
476 /* bump our ref if we were not in the list before */
477 if (!test_and_set_bit(RBIO_CACHE_BIT, &rbio->flags))
478 refcount_inc(&rbio->refs);
480 if (!list_empty(&rbio->stripe_cache)){
481 list_move(&rbio->stripe_cache, &table->stripe_cache);
483 list_add(&rbio->stripe_cache, &table->stripe_cache);
484 table->cache_size += 1;
487 spin_unlock(&rbio->bio_list_lock);
489 if (table->cache_size > RBIO_CACHE_SIZE) {
490 struct btrfs_raid_bio *found;
492 found = list_entry(table->stripe_cache.prev,
493 struct btrfs_raid_bio,
497 __remove_rbio_from_cache(found);
500 spin_unlock_irqrestore(&table->cache_lock, flags);
504 * helper function to run the xor_blocks api. It is only
505 * able to do MAX_XOR_BLOCKS at a time, so we need to
508 static void run_xor(void **pages, int src_cnt, ssize_t len)
512 void *dest = pages[src_cnt];
515 xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS);
516 xor_blocks(xor_src_cnt, len, dest, pages + src_off);
518 src_cnt -= xor_src_cnt;
519 src_off += xor_src_cnt;
524 * Returns true if the bio list inside this rbio covers an entire stripe (no
527 static int rbio_is_full(struct btrfs_raid_bio *rbio)
530 unsigned long size = rbio->bio_list_bytes;
533 spin_lock_irqsave(&rbio->bio_list_lock, flags);
534 if (size != rbio->nr_data * rbio->stripe_len)
536 BUG_ON(size > rbio->nr_data * rbio->stripe_len);
537 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
543 * returns 1 if it is safe to merge two rbios together.
544 * The merging is safe if the two rbios correspond to
545 * the same stripe and if they are both going in the same
546 * direction (read vs write), and if neither one is
547 * locked for final IO
549 * The caller is responsible for locking such that
550 * rmw_locked is safe to test
552 static int rbio_can_merge(struct btrfs_raid_bio *last,
553 struct btrfs_raid_bio *cur)
555 if (test_bit(RBIO_RMW_LOCKED_BIT, &last->flags) ||
556 test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags))
560 * we can't merge with cached rbios, since the
561 * idea is that when we merge the destination
562 * rbio is going to run our IO for us. We can
563 * steal from cached rbios though, other functions
566 if (test_bit(RBIO_CACHE_BIT, &last->flags) ||
567 test_bit(RBIO_CACHE_BIT, &cur->flags))
570 if (last->bbio->raid_map[0] !=
571 cur->bbio->raid_map[0])
574 /* we can't merge with different operations */
575 if (last->operation != cur->operation)
578 * We've need read the full stripe from the drive.
579 * check and repair the parity and write the new results.
581 * We're not allowed to add any new bios to the
582 * bio list here, anyone else that wants to
583 * change this stripe needs to do their own rmw.
585 if (last->operation == BTRFS_RBIO_PARITY_SCRUB)
588 if (last->operation == BTRFS_RBIO_REBUILD_MISSING)
591 if (last->operation == BTRFS_RBIO_READ_REBUILD) {
592 int fa = last->faila;
593 int fb = last->failb;
594 int cur_fa = cur->faila;
595 int cur_fb = cur->failb;
597 if (last->faila >= last->failb) {
602 if (cur->faila >= cur->failb) {
607 if (fa != cur_fa || fb != cur_fb)
613 static int rbio_stripe_page_index(struct btrfs_raid_bio *rbio, int stripe,
616 return stripe * rbio->stripe_npages + index;
620 * these are just the pages from the rbio array, not from anything
621 * the FS sent down to us
623 static struct page *rbio_stripe_page(struct btrfs_raid_bio *rbio, int stripe,
626 return rbio->stripe_pages[rbio_stripe_page_index(rbio, stripe, index)];
630 * helper to index into the pstripe
632 static struct page *rbio_pstripe_page(struct btrfs_raid_bio *rbio, int index)
634 return rbio_stripe_page(rbio, rbio->nr_data, index);
638 * helper to index into the qstripe, returns null
639 * if there is no qstripe
641 static struct page *rbio_qstripe_page(struct btrfs_raid_bio *rbio, int index)
643 if (rbio->nr_data + 1 == rbio->real_stripes)
645 return rbio_stripe_page(rbio, rbio->nr_data + 1, index);
649 * The first stripe in the table for a logical address
650 * has the lock. rbios are added in one of three ways:
652 * 1) Nobody has the stripe locked yet. The rbio is given
653 * the lock and 0 is returned. The caller must start the IO
656 * 2) Someone has the stripe locked, but we're able to merge
657 * with the lock owner. The rbio is freed and the IO will
658 * start automatically along with the existing rbio. 1 is returned.
660 * 3) Someone has the stripe locked, but we're not able to merge.
661 * The rbio is added to the lock owner's plug list, or merged into
662 * an rbio already on the plug list. When the lock owner unlocks,
663 * the next rbio on the list is run and the IO is started automatically.
666 * If we return 0, the caller still owns the rbio and must continue with
667 * IO submission. If we return 1, the caller must assume the rbio has
668 * already been freed.
670 static noinline int lock_stripe_add(struct btrfs_raid_bio *rbio)
672 struct btrfs_stripe_hash *h;
673 struct btrfs_raid_bio *cur;
674 struct btrfs_raid_bio *pending;
676 struct btrfs_raid_bio *freeit = NULL;
677 struct btrfs_raid_bio *cache_drop = NULL;
680 h = rbio->fs_info->stripe_hash_table->table + rbio_bucket(rbio);
682 spin_lock_irqsave(&h->lock, flags);
683 list_for_each_entry(cur, &h->hash_list, hash_list) {
684 if (cur->bbio->raid_map[0] != rbio->bbio->raid_map[0])
687 spin_lock(&cur->bio_list_lock);
689 /* Can we steal this cached rbio's pages? */
690 if (bio_list_empty(&cur->bio_list) &&
691 list_empty(&cur->plug_list) &&
692 test_bit(RBIO_CACHE_BIT, &cur->flags) &&
693 !test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) {
694 list_del_init(&cur->hash_list);
695 refcount_dec(&cur->refs);
697 steal_rbio(cur, rbio);
699 spin_unlock(&cur->bio_list_lock);
704 /* Can we merge into the lock owner? */
705 if (rbio_can_merge(cur, rbio)) {
706 merge_rbio(cur, rbio);
707 spin_unlock(&cur->bio_list_lock);
715 * We couldn't merge with the running rbio, see if we can merge
716 * with the pending ones. We don't have to check for rmw_locked
717 * because there is no way they are inside finish_rmw right now
719 list_for_each_entry(pending, &cur->plug_list, plug_list) {
720 if (rbio_can_merge(pending, rbio)) {
721 merge_rbio(pending, rbio);
722 spin_unlock(&cur->bio_list_lock);
730 * No merging, put us on the tail of the plug list, our rbio
731 * will be started with the currently running rbio unlocks
733 list_add_tail(&rbio->plug_list, &cur->plug_list);
734 spin_unlock(&cur->bio_list_lock);
739 refcount_inc(&rbio->refs);
740 list_add(&rbio->hash_list, &h->hash_list);
742 spin_unlock_irqrestore(&h->lock, flags);
744 remove_rbio_from_cache(cache_drop);
746 __free_raid_bio(freeit);
751 * called as rmw or parity rebuild is completed. If the plug list has more
752 * rbios waiting for this stripe, the next one on the list will be started
754 static noinline void unlock_stripe(struct btrfs_raid_bio *rbio)
757 struct btrfs_stripe_hash *h;
761 bucket = rbio_bucket(rbio);
762 h = rbio->fs_info->stripe_hash_table->table + bucket;
764 if (list_empty(&rbio->plug_list))
767 spin_lock_irqsave(&h->lock, flags);
768 spin_lock(&rbio->bio_list_lock);
770 if (!list_empty(&rbio->hash_list)) {
772 * if we're still cached and there is no other IO
773 * to perform, just leave this rbio here for others
774 * to steal from later
776 if (list_empty(&rbio->plug_list) &&
777 test_bit(RBIO_CACHE_BIT, &rbio->flags)) {
779 clear_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
780 BUG_ON(!bio_list_empty(&rbio->bio_list));
784 list_del_init(&rbio->hash_list);
785 refcount_dec(&rbio->refs);
788 * we use the plug list to hold all the rbios
789 * waiting for the chance to lock this stripe.
790 * hand the lock over to one of them.
792 if (!list_empty(&rbio->plug_list)) {
793 struct btrfs_raid_bio *next;
794 struct list_head *head = rbio->plug_list.next;
796 next = list_entry(head, struct btrfs_raid_bio,
799 list_del_init(&rbio->plug_list);
801 list_add(&next->hash_list, &h->hash_list);
802 refcount_inc(&next->refs);
803 spin_unlock(&rbio->bio_list_lock);
804 spin_unlock_irqrestore(&h->lock, flags);
806 if (next->operation == BTRFS_RBIO_READ_REBUILD)
807 start_async_work(next, read_rebuild_work);
808 else if (next->operation == BTRFS_RBIO_REBUILD_MISSING) {
809 steal_rbio(rbio, next);
810 start_async_work(next, read_rebuild_work);
811 } else if (next->operation == BTRFS_RBIO_WRITE) {
812 steal_rbio(rbio, next);
813 start_async_work(next, rmw_work);
814 } else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) {
815 steal_rbio(rbio, next);
816 start_async_work(next, scrub_parity_work);
823 spin_unlock(&rbio->bio_list_lock);
824 spin_unlock_irqrestore(&h->lock, flags);
828 remove_rbio_from_cache(rbio);
831 static void __free_raid_bio(struct btrfs_raid_bio *rbio)
835 if (!refcount_dec_and_test(&rbio->refs))
838 WARN_ON(!list_empty(&rbio->stripe_cache));
839 WARN_ON(!list_empty(&rbio->hash_list));
840 WARN_ON(!bio_list_empty(&rbio->bio_list));
842 for (i = 0; i < rbio->nr_pages; i++) {
843 if (rbio->stripe_pages[i]) {
844 __free_page(rbio->stripe_pages[i]);
845 rbio->stripe_pages[i] = NULL;
849 btrfs_put_bbio(rbio->bbio);
853 static void rbio_endio_bio_list(struct bio *cur, blk_status_t err)
860 cur->bi_status = err;
867 * this frees the rbio and runs through all the bios in the
868 * bio_list and calls end_io on them
870 static void rbio_orig_end_io(struct btrfs_raid_bio *rbio, blk_status_t err)
872 struct bio *cur = bio_list_get(&rbio->bio_list);
875 if (rbio->generic_bio_cnt)
876 btrfs_bio_counter_sub(rbio->fs_info, rbio->generic_bio_cnt);
879 * At this moment, rbio->bio_list is empty, however since rbio does not
880 * always have RBIO_RMW_LOCKED_BIT set and rbio is still linked on the
881 * hash list, rbio may be merged with others so that rbio->bio_list
883 * Once unlock_stripe() is done, rbio->bio_list will not be updated any
884 * more and we can call bio_endio() on all queued bios.
887 extra = bio_list_get(&rbio->bio_list);
888 __free_raid_bio(rbio);
890 rbio_endio_bio_list(cur, err);
892 rbio_endio_bio_list(extra, err);
896 * end io function used by finish_rmw. When we finally
897 * get here, we've written a full stripe
899 static void raid_write_end_io(struct bio *bio)
901 struct btrfs_raid_bio *rbio = bio->bi_private;
902 blk_status_t err = bio->bi_status;
906 fail_bio_stripe(rbio, bio);
910 if (!atomic_dec_and_test(&rbio->stripes_pending))
915 /* OK, we have read all the stripes we need to. */
916 max_errors = (rbio->operation == BTRFS_RBIO_PARITY_SCRUB) ?
917 0 : rbio->bbio->max_errors;
918 if (atomic_read(&rbio->error) > max_errors)
921 rbio_orig_end_io(rbio, err);
925 * the read/modify/write code wants to use the original bio for
926 * any pages it included, and then use the rbio for everything
927 * else. This function decides if a given index (stripe number)
928 * and page number in that stripe fall inside the original bio
931 * if you set bio_list_only, you'll get a NULL back for any ranges
932 * that are outside the bio_list
934 * This doesn't take any refs on anything, you get a bare page pointer
935 * and the caller must bump refs as required.
937 * You must call index_rbio_pages once before you can trust
938 * the answers from this function.
940 static struct page *page_in_rbio(struct btrfs_raid_bio *rbio,
941 int index, int pagenr, int bio_list_only)
944 struct page *p = NULL;
946 chunk_page = index * (rbio->stripe_len >> PAGE_SHIFT) + pagenr;
948 spin_lock_irq(&rbio->bio_list_lock);
949 p = rbio->bio_pages[chunk_page];
950 spin_unlock_irq(&rbio->bio_list_lock);
952 if (p || bio_list_only)
955 return rbio->stripe_pages[chunk_page];
959 * number of pages we need for the entire stripe across all the
962 static unsigned long rbio_nr_pages(unsigned long stripe_len, int nr_stripes)
964 return DIV_ROUND_UP(stripe_len, PAGE_SIZE) * nr_stripes;
968 * allocation and initial setup for the btrfs_raid_bio. Not
969 * this does not allocate any pages for rbio->pages.
971 static struct btrfs_raid_bio *alloc_rbio(struct btrfs_fs_info *fs_info,
972 struct btrfs_bio *bbio,
975 struct btrfs_raid_bio *rbio;
977 int real_stripes = bbio->num_stripes - bbio->num_tgtdevs;
978 int num_pages = rbio_nr_pages(stripe_len, real_stripes);
979 int stripe_npages = DIV_ROUND_UP(stripe_len, PAGE_SIZE);
982 rbio = kzalloc(sizeof(*rbio) +
983 sizeof(*rbio->stripe_pages) * num_pages +
984 sizeof(*rbio->bio_pages) * num_pages +
985 sizeof(*rbio->finish_pointers) * real_stripes +
986 sizeof(*rbio->dbitmap) * BITS_TO_LONGS(stripe_npages) +
987 sizeof(*rbio->finish_pbitmap) *
988 BITS_TO_LONGS(stripe_npages),
991 return ERR_PTR(-ENOMEM);
993 bio_list_init(&rbio->bio_list);
994 INIT_LIST_HEAD(&rbio->plug_list);
995 spin_lock_init(&rbio->bio_list_lock);
996 INIT_LIST_HEAD(&rbio->stripe_cache);
997 INIT_LIST_HEAD(&rbio->hash_list);
999 rbio->fs_info = fs_info;
1000 rbio->stripe_len = stripe_len;
1001 rbio->nr_pages = num_pages;
1002 rbio->real_stripes = real_stripes;
1003 rbio->stripe_npages = stripe_npages;
1006 refcount_set(&rbio->refs, 1);
1007 atomic_set(&rbio->error, 0);
1008 atomic_set(&rbio->stripes_pending, 0);
1011 * the stripe_pages, bio_pages, etc arrays point to the extra
1012 * memory we allocated past the end of the rbio
1015 #define CONSUME_ALLOC(ptr, count) do { \
1017 p = (unsigned char *)p + sizeof(*(ptr)) * (count); \
1019 CONSUME_ALLOC(rbio->stripe_pages, num_pages);
1020 CONSUME_ALLOC(rbio->bio_pages, num_pages);
1021 CONSUME_ALLOC(rbio->finish_pointers, real_stripes);
1022 CONSUME_ALLOC(rbio->dbitmap, BITS_TO_LONGS(stripe_npages));
1023 CONSUME_ALLOC(rbio->finish_pbitmap, BITS_TO_LONGS(stripe_npages));
1024 #undef CONSUME_ALLOC
1026 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1027 nr_data = real_stripes - 1;
1028 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1029 nr_data = real_stripes - 2;
1033 rbio->nr_data = nr_data;
1037 /* allocate pages for all the stripes in the bio, including parity */
1038 static int alloc_rbio_pages(struct btrfs_raid_bio *rbio)
1043 for (i = 0; i < rbio->nr_pages; i++) {
1044 if (rbio->stripe_pages[i])
1046 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1049 rbio->stripe_pages[i] = page;
1054 /* only allocate pages for p/q stripes */
1055 static int alloc_rbio_parity_pages(struct btrfs_raid_bio *rbio)
1060 i = rbio_stripe_page_index(rbio, rbio->nr_data, 0);
1062 for (; i < rbio->nr_pages; i++) {
1063 if (rbio->stripe_pages[i])
1065 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1068 rbio->stripe_pages[i] = page;
1074 * add a single page from a specific stripe into our list of bios for IO
1075 * this will try to merge into existing bios if possible, and returns
1076 * zero if all went well.
1078 static int rbio_add_io_page(struct btrfs_raid_bio *rbio,
1079 struct bio_list *bio_list,
1082 unsigned long page_index,
1083 unsigned long bio_max_len)
1085 struct bio *last = bio_list->tail;
1088 struct btrfs_bio_stripe *stripe;
1091 stripe = &rbio->bbio->stripes[stripe_nr];
1092 disk_start = stripe->physical + (page_index << PAGE_SHIFT);
1094 /* if the device is missing, just fail this stripe */
1095 if (!stripe->dev->bdev)
1096 return fail_rbio_index(rbio, stripe_nr);
1098 /* see if we can add this page onto our existing bio */
1100 u64 last_end = (u64)last->bi_iter.bi_sector << 9;
1101 last_end += last->bi_iter.bi_size;
1104 * we can't merge these if they are from different
1105 * devices or if they are not contiguous
1107 if (last_end == disk_start && !last->bi_status &&
1108 last->bi_disk == stripe->dev->bdev->bd_disk &&
1109 last->bi_partno == stripe->dev->bdev->bd_partno) {
1110 ret = bio_add_page(last, page, PAGE_SIZE, 0);
1111 if (ret == PAGE_SIZE)
1116 /* put a new bio on the list */
1117 bio = btrfs_io_bio_alloc(bio_max_len >> PAGE_SHIFT ?: 1);
1118 btrfs_io_bio(bio)->device = stripe->dev;
1119 bio->bi_iter.bi_size = 0;
1120 bio_set_dev(bio, stripe->dev->bdev);
1121 bio->bi_iter.bi_sector = disk_start >> 9;
1123 bio_add_page(bio, page, PAGE_SIZE, 0);
1124 bio_list_add(bio_list, bio);
1129 * while we're doing the read/modify/write cycle, we could
1130 * have errors in reading pages off the disk. This checks
1131 * for errors and if we're not able to read the page it'll
1132 * trigger parity reconstruction. The rmw will be finished
1133 * after we've reconstructed the failed stripes
1135 static void validate_rbio_for_rmw(struct btrfs_raid_bio *rbio)
1137 if (rbio->faila >= 0 || rbio->failb >= 0) {
1138 BUG_ON(rbio->faila == rbio->real_stripes - 1);
1139 __raid56_parity_recover(rbio);
1146 * helper function to walk our bio list and populate the bio_pages array with
1147 * the result. This seems expensive, but it is faster than constantly
1148 * searching through the bio list as we setup the IO in finish_rmw or stripe
1151 * This must be called before you trust the answers from page_in_rbio
1153 static void index_rbio_pages(struct btrfs_raid_bio *rbio)
1157 unsigned long stripe_offset;
1158 unsigned long page_index;
1160 spin_lock_irq(&rbio->bio_list_lock);
1161 bio_list_for_each(bio, &rbio->bio_list) {
1162 struct bio_vec bvec;
1163 struct bvec_iter iter;
1166 start = (u64)bio->bi_iter.bi_sector << 9;
1167 stripe_offset = start - rbio->bbio->raid_map[0];
1168 page_index = stripe_offset >> PAGE_SHIFT;
1170 if (bio_flagged(bio, BIO_CLONED))
1171 bio->bi_iter = btrfs_io_bio(bio)->iter;
1173 bio_for_each_segment(bvec, bio, iter) {
1174 rbio->bio_pages[page_index + i] = bvec.bv_page;
1178 spin_unlock_irq(&rbio->bio_list_lock);
1182 * this is called from one of two situations. We either
1183 * have a full stripe from the higher layers, or we've read all
1184 * the missing bits off disk.
1186 * This will calculate the parity and then send down any
1189 static noinline void finish_rmw(struct btrfs_raid_bio *rbio)
1191 struct btrfs_bio *bbio = rbio->bbio;
1192 void **pointers = rbio->finish_pointers;
1193 int nr_data = rbio->nr_data;
1197 struct bio_list bio_list;
1201 bio_list_init(&bio_list);
1203 if (rbio->real_stripes - rbio->nr_data == 1)
1204 has_qstripe = false;
1205 else if (rbio->real_stripes - rbio->nr_data == 2)
1210 /* at this point we either have a full stripe,
1211 * or we've read the full stripe from the drive.
1212 * recalculate the parity and write the new results.
1214 * We're not allowed to add any new bios to the
1215 * bio list here, anyone else that wants to
1216 * change this stripe needs to do their own rmw.
1218 spin_lock_irq(&rbio->bio_list_lock);
1219 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1220 spin_unlock_irq(&rbio->bio_list_lock);
1222 atomic_set(&rbio->error, 0);
1225 * now that we've set rmw_locked, run through the
1226 * bio list one last time and map the page pointers
1228 * We don't cache full rbios because we're assuming
1229 * the higher layers are unlikely to use this area of
1230 * the disk again soon. If they do use it again,
1231 * hopefully they will send another full bio.
1233 index_rbio_pages(rbio);
1234 if (!rbio_is_full(rbio))
1235 cache_rbio_pages(rbio);
1237 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
1239 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1241 /* first collect one page from each data stripe */
1242 for (stripe = 0; stripe < nr_data; stripe++) {
1243 p = page_in_rbio(rbio, stripe, pagenr, 0);
1244 pointers[stripe] = kmap(p);
1247 /* then add the parity stripe */
1248 p = rbio_pstripe_page(rbio, pagenr);
1250 pointers[stripe++] = kmap(p);
1255 * raid6, add the qstripe and call the
1256 * library function to fill in our p/q
1258 p = rbio_qstripe_page(rbio, pagenr);
1260 pointers[stripe++] = kmap(p);
1262 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
1266 copy_page(pointers[nr_data], pointers[0]);
1267 run_xor(pointers + 1, nr_data - 1, PAGE_SIZE);
1271 for (stripe = 0; stripe < rbio->real_stripes; stripe++)
1272 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
1276 * time to start writing. Make bios for everything from the
1277 * higher layers (the bio_list in our rbio) and our p/q. Ignore
1280 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1281 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1283 if (stripe < rbio->nr_data) {
1284 page = page_in_rbio(rbio, stripe, pagenr, 1);
1288 page = rbio_stripe_page(rbio, stripe, pagenr);
1291 ret = rbio_add_io_page(rbio, &bio_list,
1292 page, stripe, pagenr, rbio->stripe_len);
1298 if (likely(!bbio->num_tgtdevs))
1301 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1302 if (!bbio->tgtdev_map[stripe])
1305 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1307 if (stripe < rbio->nr_data) {
1308 page = page_in_rbio(rbio, stripe, pagenr, 1);
1312 page = rbio_stripe_page(rbio, stripe, pagenr);
1315 ret = rbio_add_io_page(rbio, &bio_list, page,
1316 rbio->bbio->tgtdev_map[stripe],
1317 pagenr, rbio->stripe_len);
1324 atomic_set(&rbio->stripes_pending, bio_list_size(&bio_list));
1325 BUG_ON(atomic_read(&rbio->stripes_pending) == 0);
1327 while ((bio = bio_list_pop(&bio_list))) {
1328 bio->bi_private = rbio;
1329 bio->bi_end_io = raid_write_end_io;
1330 bio->bi_opf = REQ_OP_WRITE;
1337 rbio_orig_end_io(rbio, BLK_STS_IOERR);
1339 while ((bio = bio_list_pop(&bio_list)))
1344 * helper to find the stripe number for a given bio. Used to figure out which
1345 * stripe has failed. This expects the bio to correspond to a physical disk,
1346 * so it looks up based on physical sector numbers.
1348 static int find_bio_stripe(struct btrfs_raid_bio *rbio,
1351 u64 physical = bio->bi_iter.bi_sector;
1353 struct btrfs_bio_stripe *stripe;
1357 for (i = 0; i < rbio->bbio->num_stripes; i++) {
1358 stripe = &rbio->bbio->stripes[i];
1359 if (in_range(physical, stripe->physical, rbio->stripe_len) &&
1360 stripe->dev->bdev &&
1361 bio->bi_disk == stripe->dev->bdev->bd_disk &&
1362 bio->bi_partno == stripe->dev->bdev->bd_partno) {
1370 * helper to find the stripe number for a given
1371 * bio (before mapping). Used to figure out which stripe has
1372 * failed. This looks up based on logical block numbers.
1374 static int find_logical_bio_stripe(struct btrfs_raid_bio *rbio,
1377 u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1380 for (i = 0; i < rbio->nr_data; i++) {
1381 u64 stripe_start = rbio->bbio->raid_map[i];
1383 if (in_range(logical, stripe_start, rbio->stripe_len))
1390 * returns -EIO if we had too many failures
1392 static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed)
1394 unsigned long flags;
1397 spin_lock_irqsave(&rbio->bio_list_lock, flags);
1399 /* we already know this stripe is bad, move on */
1400 if (rbio->faila == failed || rbio->failb == failed)
1403 if (rbio->faila == -1) {
1404 /* first failure on this rbio */
1405 rbio->faila = failed;
1406 atomic_inc(&rbio->error);
1407 } else if (rbio->failb == -1) {
1408 /* second failure on this rbio */
1409 rbio->failb = failed;
1410 atomic_inc(&rbio->error);
1415 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
1421 * helper to fail a stripe based on a physical disk
1424 static int fail_bio_stripe(struct btrfs_raid_bio *rbio,
1427 int failed = find_bio_stripe(rbio, bio);
1432 return fail_rbio_index(rbio, failed);
1436 * this sets each page in the bio uptodate. It should only be used on private
1437 * rbio pages, nothing that comes in from the higher layers
1439 static void set_bio_pages_uptodate(struct bio *bio)
1441 struct bio_vec *bvec;
1442 struct bvec_iter_all iter_all;
1444 ASSERT(!bio_flagged(bio, BIO_CLONED));
1446 bio_for_each_segment_all(bvec, bio, iter_all)
1447 SetPageUptodate(bvec->bv_page);
1451 * end io for the read phase of the rmw cycle. All the bios here are physical
1452 * stripe bios we've read from the disk so we can recalculate the parity of the
1455 * This will usually kick off finish_rmw once all the bios are read in, but it
1456 * may trigger parity reconstruction if we had any errors along the way
1458 static void raid_rmw_end_io(struct bio *bio)
1460 struct btrfs_raid_bio *rbio = bio->bi_private;
1463 fail_bio_stripe(rbio, bio);
1465 set_bio_pages_uptodate(bio);
1469 if (!atomic_dec_and_test(&rbio->stripes_pending))
1472 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
1476 * this will normally call finish_rmw to start our write
1477 * but if there are any failed stripes we'll reconstruct
1480 validate_rbio_for_rmw(rbio);
1485 rbio_orig_end_io(rbio, BLK_STS_IOERR);
1489 * the stripe must be locked by the caller. It will
1490 * unlock after all the writes are done
1492 static int raid56_rmw_stripe(struct btrfs_raid_bio *rbio)
1494 int bios_to_read = 0;
1495 struct bio_list bio_list;
1501 bio_list_init(&bio_list);
1503 ret = alloc_rbio_pages(rbio);
1507 index_rbio_pages(rbio);
1509 atomic_set(&rbio->error, 0);
1511 * build a list of bios to read all the missing parts of this
1514 for (stripe = 0; stripe < rbio->nr_data; stripe++) {
1515 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1518 * we want to find all the pages missing from
1519 * the rbio and read them from the disk. If
1520 * page_in_rbio finds a page in the bio list
1521 * we don't need to read it off the stripe.
1523 page = page_in_rbio(rbio, stripe, pagenr, 1);
1527 page = rbio_stripe_page(rbio, stripe, pagenr);
1529 * the bio cache may have handed us an uptodate
1530 * page. If so, be happy and use it
1532 if (PageUptodate(page))
1535 ret = rbio_add_io_page(rbio, &bio_list, page,
1536 stripe, pagenr, rbio->stripe_len);
1542 bios_to_read = bio_list_size(&bio_list);
1543 if (!bios_to_read) {
1545 * this can happen if others have merged with
1546 * us, it means there is nothing left to read.
1547 * But if there are missing devices it may not be
1548 * safe to do the full stripe write yet.
1554 * the bbio may be freed once we submit the last bio. Make sure
1555 * not to touch it after that
1557 atomic_set(&rbio->stripes_pending, bios_to_read);
1558 while ((bio = bio_list_pop(&bio_list))) {
1559 bio->bi_private = rbio;
1560 bio->bi_end_io = raid_rmw_end_io;
1561 bio->bi_opf = REQ_OP_READ;
1563 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
1567 /* the actual write will happen once the reads are done */
1571 rbio_orig_end_io(rbio, BLK_STS_IOERR);
1573 while ((bio = bio_list_pop(&bio_list)))
1579 validate_rbio_for_rmw(rbio);
1584 * if the upper layers pass in a full stripe, we thank them by only allocating
1585 * enough pages to hold the parity, and sending it all down quickly.
1587 static int full_stripe_write(struct btrfs_raid_bio *rbio)
1591 ret = alloc_rbio_parity_pages(rbio);
1593 __free_raid_bio(rbio);
1597 ret = lock_stripe_add(rbio);
1604 * partial stripe writes get handed over to async helpers.
1605 * We're really hoping to merge a few more writes into this
1606 * rbio before calculating new parity
1608 static int partial_stripe_write(struct btrfs_raid_bio *rbio)
1612 ret = lock_stripe_add(rbio);
1614 start_async_work(rbio, rmw_work);
1619 * sometimes while we were reading from the drive to
1620 * recalculate parity, enough new bios come into create
1621 * a full stripe. So we do a check here to see if we can
1622 * go directly to finish_rmw
1624 static int __raid56_parity_write(struct btrfs_raid_bio *rbio)
1626 /* head off into rmw land if we don't have a full stripe */
1627 if (!rbio_is_full(rbio))
1628 return partial_stripe_write(rbio);
1629 return full_stripe_write(rbio);
1633 * We use plugging call backs to collect full stripes.
1634 * Any time we get a partial stripe write while plugged
1635 * we collect it into a list. When the unplug comes down,
1636 * we sort the list by logical block number and merge
1637 * everything we can into the same rbios
1639 struct btrfs_plug_cb {
1640 struct blk_plug_cb cb;
1641 struct btrfs_fs_info *info;
1642 struct list_head rbio_list;
1643 struct btrfs_work work;
1647 * rbios on the plug list are sorted for easier merging.
1649 static int plug_cmp(void *priv, struct list_head *a, struct list_head *b)
1651 struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio,
1653 struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio,
1655 u64 a_sector = ra->bio_list.head->bi_iter.bi_sector;
1656 u64 b_sector = rb->bio_list.head->bi_iter.bi_sector;
1658 if (a_sector < b_sector)
1660 if (a_sector > b_sector)
1665 static void run_plug(struct btrfs_plug_cb *plug)
1667 struct btrfs_raid_bio *cur;
1668 struct btrfs_raid_bio *last = NULL;
1671 * sort our plug list then try to merge
1672 * everything we can in hopes of creating full
1675 list_sort(NULL, &plug->rbio_list, plug_cmp);
1676 while (!list_empty(&plug->rbio_list)) {
1677 cur = list_entry(plug->rbio_list.next,
1678 struct btrfs_raid_bio, plug_list);
1679 list_del_init(&cur->plug_list);
1681 if (rbio_is_full(cur)) {
1684 /* we have a full stripe, send it down */
1685 ret = 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, unplug_work, NULL, NULL);
1724 btrfs_queue_work(plug->info->rmw_workers,
1732 * our main entry point for writes from the rest of the FS.
1734 int raid56_parity_write(struct btrfs_fs_info *fs_info, struct bio *bio,
1735 struct btrfs_bio *bbio, u64 stripe_len)
1737 struct btrfs_raid_bio *rbio;
1738 struct btrfs_plug_cb *plug = NULL;
1739 struct blk_plug_cb *cb;
1742 rbio = alloc_rbio(fs_info, bbio, stripe_len);
1744 btrfs_put_bbio(bbio);
1745 return PTR_ERR(rbio);
1747 bio_list_add(&rbio->bio_list, bio);
1748 rbio->bio_list_bytes = bio->bi_iter.bi_size;
1749 rbio->operation = BTRFS_RBIO_WRITE;
1751 btrfs_bio_counter_inc_noblocked(fs_info);
1752 rbio->generic_bio_cnt = 1;
1755 * don't plug on full rbios, just get them out the door
1756 * as quickly as we can
1758 if (rbio_is_full(rbio)) {
1759 ret = full_stripe_write(rbio);
1761 btrfs_bio_counter_dec(fs_info);
1765 cb = blk_check_plugged(btrfs_raid_unplug, fs_info, sizeof(*plug));
1767 plug = container_of(cb, struct btrfs_plug_cb, cb);
1769 plug->info = fs_info;
1770 INIT_LIST_HEAD(&plug->rbio_list);
1772 list_add_tail(&rbio->plug_list, &plug->rbio_list);
1775 ret = __raid56_parity_write(rbio);
1777 btrfs_bio_counter_dec(fs_info);
1783 * all parity reconstruction happens here. We've read in everything
1784 * we can find from the drives and this does the heavy lifting of
1785 * sorting the good from the bad.
1787 static void __raid_recover_end_io(struct btrfs_raid_bio *rbio)
1791 int faila = -1, failb = -1;
1796 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
1798 err = BLK_STS_RESOURCE;
1802 faila = rbio->faila;
1803 failb = rbio->failb;
1805 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1806 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
1807 spin_lock_irq(&rbio->bio_list_lock);
1808 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1809 spin_unlock_irq(&rbio->bio_list_lock);
1812 index_rbio_pages(rbio);
1814 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1816 * Now we just use bitmap to mark the horizontal stripes in
1817 * which we have data when doing parity scrub.
1819 if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB &&
1820 !test_bit(pagenr, rbio->dbitmap))
1823 /* setup our array of pointers with pages
1826 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1828 * if we're rebuilding a read, we have to use
1829 * pages from the bio list
1831 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1832 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
1833 (stripe == faila || stripe == failb)) {
1834 page = page_in_rbio(rbio, stripe, pagenr, 0);
1836 page = rbio_stripe_page(rbio, stripe, pagenr);
1838 pointers[stripe] = kmap(page);
1841 /* all raid6 handling here */
1842 if (rbio->bbio->map_type & BTRFS_BLOCK_GROUP_RAID6) {
1844 * single failure, rebuild from parity raid5
1848 if (faila == rbio->nr_data) {
1850 * Just the P stripe has failed, without
1851 * a bad data or Q stripe.
1852 * TODO, we should redo the xor here.
1854 err = BLK_STS_IOERR;
1858 * a single failure in raid6 is rebuilt
1859 * in the pstripe code below
1864 /* make sure our ps and qs are in order */
1868 /* if the q stripe is failed, do a pstripe reconstruction
1870 * If both the q stripe and the P stripe are failed, we're
1871 * here due to a crc mismatch and we can't give them the
1874 if (rbio->bbio->raid_map[failb] == RAID6_Q_STRIPE) {
1875 if (rbio->bbio->raid_map[faila] ==
1877 err = BLK_STS_IOERR;
1881 * otherwise we have one bad data stripe and
1882 * a good P stripe. raid5!
1887 if (rbio->bbio->raid_map[failb] == RAID5_P_STRIPE) {
1888 raid6_datap_recov(rbio->real_stripes,
1889 PAGE_SIZE, faila, pointers);
1891 raid6_2data_recov(rbio->real_stripes,
1892 PAGE_SIZE, faila, failb,
1898 /* rebuild from P stripe here (raid5 or raid6) */
1899 BUG_ON(failb != -1);
1901 /* Copy parity block into failed block to start with */
1902 copy_page(pointers[faila], pointers[rbio->nr_data]);
1904 /* rearrange the pointer array */
1905 p = pointers[faila];
1906 for (stripe = faila; stripe < rbio->nr_data - 1; stripe++)
1907 pointers[stripe] = pointers[stripe + 1];
1908 pointers[rbio->nr_data - 1] = p;
1910 /* xor in the rest */
1911 run_xor(pointers, rbio->nr_data - 1, PAGE_SIZE);
1913 /* if we're doing this rebuild as part of an rmw, go through
1914 * and set all of our private rbio pages in the
1915 * failed stripes as uptodate. This way finish_rmw will
1916 * know they can be trusted. If this was a read reconstruction,
1917 * other endio functions will fiddle the uptodate bits
1919 if (rbio->operation == BTRFS_RBIO_WRITE) {
1920 for (i = 0; i < rbio->stripe_npages; i++) {
1922 page = rbio_stripe_page(rbio, faila, i);
1923 SetPageUptodate(page);
1926 page = rbio_stripe_page(rbio, failb, i);
1927 SetPageUptodate(page);
1931 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1933 * if we're rebuilding a read, we have to use
1934 * pages from the bio list
1936 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1937 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
1938 (stripe == faila || stripe == failb)) {
1939 page = page_in_rbio(rbio, stripe, pagenr, 0);
1941 page = rbio_stripe_page(rbio, stripe, pagenr);
1953 * Similar to READ_REBUILD, REBUILD_MISSING at this point also has a
1954 * valid rbio which is consistent with ondisk content, thus such a
1955 * valid rbio can be cached to avoid further disk reads.
1957 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1958 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
1960 * - In case of two failures, where rbio->failb != -1:
1962 * Do not cache this rbio since the above read reconstruction
1963 * (raid6_datap_recov() or raid6_2data_recov()) may have
1964 * changed some content of stripes which are not identical to
1965 * on-disk content any more, otherwise, a later write/recover
1966 * may steal stripe_pages from this rbio and end up with
1967 * corruptions or rebuild failures.
1969 * - In case of single failure, where rbio->failb == -1:
1971 * Cache this rbio iff the above read reconstruction is
1972 * executed without problems.
1974 if (err == BLK_STS_OK && rbio->failb < 0)
1975 cache_rbio_pages(rbio);
1977 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
1979 rbio_orig_end_io(rbio, err);
1980 } else if (err == BLK_STS_OK) {
1984 if (rbio->operation == BTRFS_RBIO_WRITE)
1986 else if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB)
1987 finish_parity_scrub(rbio, 0);
1991 rbio_orig_end_io(rbio, err);
1996 * This is called only for stripes we've read from disk to
1997 * reconstruct the parity.
1999 static void raid_recover_end_io(struct bio *bio)
2001 struct btrfs_raid_bio *rbio = bio->bi_private;
2004 * we only read stripe pages off the disk, set them
2005 * up to date if there were no errors
2008 fail_bio_stripe(rbio, bio);
2010 set_bio_pages_uptodate(bio);
2013 if (!atomic_dec_and_test(&rbio->stripes_pending))
2016 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
2017 rbio_orig_end_io(rbio, BLK_STS_IOERR);
2019 __raid_recover_end_io(rbio);
2023 * reads everything we need off the disk to reconstruct
2024 * the parity. endio handlers trigger final reconstruction
2025 * when the IO is done.
2027 * This is used both for reads from the higher layers and for
2028 * parity construction required to finish a rmw cycle.
2030 static int __raid56_parity_recover(struct btrfs_raid_bio *rbio)
2032 int bios_to_read = 0;
2033 struct bio_list bio_list;
2039 bio_list_init(&bio_list);
2041 ret = alloc_rbio_pages(rbio);
2045 atomic_set(&rbio->error, 0);
2048 * read everything that hasn't failed. Thanks to the
2049 * stripe cache, it is possible that some or all of these
2050 * pages are going to be uptodate.
2052 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
2053 if (rbio->faila == stripe || rbio->failb == stripe) {
2054 atomic_inc(&rbio->error);
2058 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
2062 * the rmw code may have already read this
2065 p = rbio_stripe_page(rbio, stripe, pagenr);
2066 if (PageUptodate(p))
2069 ret = rbio_add_io_page(rbio, &bio_list,
2070 rbio_stripe_page(rbio, stripe, pagenr),
2071 stripe, pagenr, rbio->stripe_len);
2077 bios_to_read = bio_list_size(&bio_list);
2078 if (!bios_to_read) {
2080 * we might have no bios to read just because the pages
2081 * were up to date, or we might have no bios to read because
2082 * the devices were gone.
2084 if (atomic_read(&rbio->error) <= rbio->bbio->max_errors) {
2085 __raid_recover_end_io(rbio);
2093 * the bbio may be freed once we submit the last bio. Make sure
2094 * not to touch it after that
2096 atomic_set(&rbio->stripes_pending, bios_to_read);
2097 while ((bio = bio_list_pop(&bio_list))) {
2098 bio->bi_private = rbio;
2099 bio->bi_end_io = raid_recover_end_io;
2100 bio->bi_opf = REQ_OP_READ;
2102 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
2110 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
2111 rbio->operation == BTRFS_RBIO_REBUILD_MISSING)
2112 rbio_orig_end_io(rbio, BLK_STS_IOERR);
2114 while ((bio = bio_list_pop(&bio_list)))
2121 * the main entry point for reads from the higher layers. This
2122 * is really only called when the normal read path had a failure,
2123 * so we assume the bio they send down corresponds to a failed part
2126 int raid56_parity_recover(struct btrfs_fs_info *fs_info, struct bio *bio,
2127 struct btrfs_bio *bbio, u64 stripe_len,
2128 int mirror_num, int generic_io)
2130 struct btrfs_raid_bio *rbio;
2134 ASSERT(bbio->mirror_num == mirror_num);
2135 btrfs_io_bio(bio)->mirror_num = mirror_num;
2138 rbio = alloc_rbio(fs_info, bbio, stripe_len);
2141 btrfs_put_bbio(bbio);
2142 return PTR_ERR(rbio);
2145 rbio->operation = BTRFS_RBIO_READ_REBUILD;
2146 bio_list_add(&rbio->bio_list, bio);
2147 rbio->bio_list_bytes = bio->bi_iter.bi_size;
2149 rbio->faila = find_logical_bio_stripe(rbio, bio);
2150 if (rbio->faila == -1) {
2152 "%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)",
2153 __func__, (u64)bio->bi_iter.bi_sector << 9,
2154 (u64)bio->bi_iter.bi_size, bbio->map_type);
2156 btrfs_put_bbio(bbio);
2162 btrfs_bio_counter_inc_noblocked(fs_info);
2163 rbio->generic_bio_cnt = 1;
2165 btrfs_get_bbio(bbio);
2170 * for 'mirror == 2', reconstruct from all other stripes.
2171 * for 'mirror_num > 2', select a stripe to fail on every retry.
2173 if (mirror_num > 2) {
2175 * 'mirror == 3' is to fail the p stripe and
2176 * reconstruct from the q stripe. 'mirror > 3' is to
2177 * fail a data stripe and reconstruct from p+q stripe.
2179 rbio->failb = rbio->real_stripes - (mirror_num - 1);
2180 ASSERT(rbio->failb > 0);
2181 if (rbio->failb <= rbio->faila)
2185 ret = lock_stripe_add(rbio);
2188 * __raid56_parity_recover will end the bio with
2189 * any errors it hits. We don't want to return
2190 * its error value up the stack because our caller
2191 * will end up calling bio_endio with any nonzero
2195 __raid56_parity_recover(rbio);
2197 * our rbio has been added to the list of
2198 * rbios that will be handled after the
2199 * currently lock owner is done
2205 static void rmw_work(struct btrfs_work *work)
2207 struct btrfs_raid_bio *rbio;
2209 rbio = container_of(work, struct btrfs_raid_bio, work);
2210 raid56_rmw_stripe(rbio);
2213 static void read_rebuild_work(struct btrfs_work *work)
2215 struct btrfs_raid_bio *rbio;
2217 rbio = container_of(work, struct btrfs_raid_bio, work);
2218 __raid56_parity_recover(rbio);
2222 * The following code is used to scrub/replace the parity stripe
2224 * Caller must have already increased bio_counter for getting @bbio.
2226 * Note: We need make sure all the pages that add into the scrub/replace
2227 * raid bio are correct and not be changed during the scrub/replace. That
2228 * is those pages just hold metadata or file data with checksum.
2231 struct btrfs_raid_bio *
2232 raid56_parity_alloc_scrub_rbio(struct btrfs_fs_info *fs_info, struct bio *bio,
2233 struct btrfs_bio *bbio, u64 stripe_len,
2234 struct btrfs_device *scrub_dev,
2235 unsigned long *dbitmap, int stripe_nsectors)
2237 struct btrfs_raid_bio *rbio;
2240 rbio = alloc_rbio(fs_info, bbio, stripe_len);
2243 bio_list_add(&rbio->bio_list, bio);
2245 * This is a special bio which is used to hold the completion handler
2246 * and make the scrub rbio is similar to the other types
2248 ASSERT(!bio->bi_iter.bi_size);
2249 rbio->operation = BTRFS_RBIO_PARITY_SCRUB;
2252 * After mapping bbio with BTRFS_MAP_WRITE, parities have been sorted
2253 * to the end position, so this search can start from the first parity
2256 for (i = rbio->nr_data; i < rbio->real_stripes; i++) {
2257 if (bbio->stripes[i].dev == scrub_dev) {
2262 ASSERT(i < rbio->real_stripes);
2264 /* Now we just support the sectorsize equals to page size */
2265 ASSERT(fs_info->sectorsize == PAGE_SIZE);
2266 ASSERT(rbio->stripe_npages == stripe_nsectors);
2267 bitmap_copy(rbio->dbitmap, dbitmap, stripe_nsectors);
2270 * We have already increased bio_counter when getting bbio, record it
2271 * so we can free it at rbio_orig_end_io().
2273 rbio->generic_bio_cnt = 1;
2278 /* Used for both parity scrub and missing. */
2279 void raid56_add_scrub_pages(struct btrfs_raid_bio *rbio, struct page *page,
2285 ASSERT(logical >= rbio->bbio->raid_map[0]);
2286 ASSERT(logical + PAGE_SIZE <= rbio->bbio->raid_map[0] +
2287 rbio->stripe_len * rbio->nr_data);
2288 stripe_offset = (int)(logical - rbio->bbio->raid_map[0]);
2289 index = stripe_offset >> PAGE_SHIFT;
2290 rbio->bio_pages[index] = page;
2294 * We just scrub the parity that we have correct data on the same horizontal,
2295 * so we needn't allocate all pages for all the stripes.
2297 static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio)
2304 for_each_set_bit(bit, rbio->dbitmap, rbio->stripe_npages) {
2305 for (i = 0; i < rbio->real_stripes; i++) {
2306 index = i * rbio->stripe_npages + bit;
2307 if (rbio->stripe_pages[index])
2310 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2313 rbio->stripe_pages[index] = page;
2319 static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
2322 struct btrfs_bio *bbio = rbio->bbio;
2323 void **pointers = rbio->finish_pointers;
2324 unsigned long *pbitmap = rbio->finish_pbitmap;
2325 int nr_data = rbio->nr_data;
2329 struct page *p_page = NULL;
2330 struct page *q_page = NULL;
2331 struct bio_list bio_list;
2336 bio_list_init(&bio_list);
2338 if (rbio->real_stripes - rbio->nr_data == 1)
2339 has_qstripe = false;
2340 else if (rbio->real_stripes - rbio->nr_data == 2)
2345 if (bbio->num_tgtdevs && bbio->tgtdev_map[rbio->scrubp]) {
2347 bitmap_copy(pbitmap, rbio->dbitmap, rbio->stripe_npages);
2351 * Because the higher layers(scrubber) are unlikely to
2352 * use this area of the disk again soon, so don't cache
2355 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2360 p_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2363 SetPageUptodate(p_page);
2366 q_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2368 __free_page(p_page);
2371 SetPageUptodate(q_page);
2374 atomic_set(&rbio->error, 0);
2376 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2379 /* first collect one page from each data stripe */
2380 for (stripe = 0; stripe < nr_data; stripe++) {
2381 p = page_in_rbio(rbio, stripe, pagenr, 0);
2382 pointers[stripe] = kmap(p);
2385 /* then add the parity stripe */
2386 pointers[stripe++] = kmap(p_page);
2390 * raid6, add the qstripe and call the
2391 * library function to fill in our p/q
2393 pointers[stripe++] = kmap(q_page);
2395 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
2399 copy_page(pointers[nr_data], pointers[0]);
2400 run_xor(pointers + 1, nr_data - 1, PAGE_SIZE);
2403 /* Check scrubbing parity and repair it */
2404 p = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2406 if (memcmp(parity, pointers[rbio->scrubp], PAGE_SIZE))
2407 copy_page(parity, pointers[rbio->scrubp]);
2409 /* Parity is right, needn't writeback */
2410 bitmap_clear(rbio->dbitmap, pagenr, 1);
2413 for (stripe = 0; stripe < nr_data; stripe++)
2414 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
2418 __free_page(p_page);
2420 __free_page(q_page);
2424 * time to start writing. Make bios for everything from the
2425 * higher layers (the bio_list in our rbio) and our p/q. Ignore
2428 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2431 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2432 ret = rbio_add_io_page(rbio, &bio_list,
2433 page, rbio->scrubp, pagenr, rbio->stripe_len);
2441 for_each_set_bit(pagenr, pbitmap, rbio->stripe_npages) {
2444 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2445 ret = rbio_add_io_page(rbio, &bio_list, page,
2446 bbio->tgtdev_map[rbio->scrubp],
2447 pagenr, rbio->stripe_len);
2453 nr_data = bio_list_size(&bio_list);
2455 /* Every parity is right */
2456 rbio_orig_end_io(rbio, BLK_STS_OK);
2460 atomic_set(&rbio->stripes_pending, nr_data);
2462 while ((bio = bio_list_pop(&bio_list))) {
2463 bio->bi_private = rbio;
2464 bio->bi_end_io = raid_write_end_io;
2465 bio->bi_opf = REQ_OP_WRITE;
2472 rbio_orig_end_io(rbio, BLK_STS_IOERR);
2474 while ((bio = bio_list_pop(&bio_list)))
2478 static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe)
2480 if (stripe >= 0 && stripe < rbio->nr_data)
2486 * While we're doing the parity check and repair, we could have errors
2487 * in reading pages off the disk. This checks for errors and if we're
2488 * not able to read the page it'll trigger parity reconstruction. The
2489 * parity scrub will be finished after we've reconstructed the failed
2492 static void validate_rbio_for_parity_scrub(struct btrfs_raid_bio *rbio)
2494 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
2497 if (rbio->faila >= 0 || rbio->failb >= 0) {
2498 int dfail = 0, failp = -1;
2500 if (is_data_stripe(rbio, rbio->faila))
2502 else if (is_parity_stripe(rbio->faila))
2503 failp = rbio->faila;
2505 if (is_data_stripe(rbio, rbio->failb))
2507 else if (is_parity_stripe(rbio->failb))
2508 failp = rbio->failb;
2511 * Because we can not use a scrubbing parity to repair
2512 * the data, so the capability of the repair is declined.
2513 * (In the case of RAID5, we can not repair anything)
2515 if (dfail > rbio->bbio->max_errors - 1)
2519 * If all data is good, only parity is correctly, just
2520 * repair the parity.
2523 finish_parity_scrub(rbio, 0);
2528 * Here means we got one corrupted data stripe and one
2529 * corrupted parity on RAID6, if the corrupted parity
2530 * is scrubbing parity, luckily, use the other one to repair
2531 * the data, or we can not repair the data stripe.
2533 if (failp != rbio->scrubp)
2536 __raid_recover_end_io(rbio);
2538 finish_parity_scrub(rbio, 1);
2543 rbio_orig_end_io(rbio, BLK_STS_IOERR);
2547 * end io for the read phase of the rmw cycle. All the bios here are physical
2548 * stripe bios we've read from the disk so we can recalculate the parity of the
2551 * This will usually kick off finish_rmw once all the bios are read in, but it
2552 * may trigger parity reconstruction if we had any errors along the way
2554 static void raid56_parity_scrub_end_io(struct bio *bio)
2556 struct btrfs_raid_bio *rbio = bio->bi_private;
2559 fail_bio_stripe(rbio, bio);
2561 set_bio_pages_uptodate(bio);
2565 if (!atomic_dec_and_test(&rbio->stripes_pending))
2569 * this will normally call finish_rmw to start our write
2570 * but if there are any failed stripes we'll reconstruct
2573 validate_rbio_for_parity_scrub(rbio);
2576 static void raid56_parity_scrub_stripe(struct btrfs_raid_bio *rbio)
2578 int bios_to_read = 0;
2579 struct bio_list bio_list;
2585 bio_list_init(&bio_list);
2587 ret = alloc_rbio_essential_pages(rbio);
2591 atomic_set(&rbio->error, 0);
2593 * build a list of bios to read all the missing parts of this
2596 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
2597 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2600 * we want to find all the pages missing from
2601 * the rbio and read them from the disk. If
2602 * page_in_rbio finds a page in the bio list
2603 * we don't need to read it off the stripe.
2605 page = page_in_rbio(rbio, stripe, pagenr, 1);
2609 page = rbio_stripe_page(rbio, stripe, pagenr);
2611 * the bio cache may have handed us an uptodate
2612 * page. If so, be happy and use it
2614 if (PageUptodate(page))
2617 ret = rbio_add_io_page(rbio, &bio_list, page,
2618 stripe, pagenr, rbio->stripe_len);
2624 bios_to_read = bio_list_size(&bio_list);
2625 if (!bios_to_read) {
2627 * this can happen if others have merged with
2628 * us, it means there is nothing left to read.
2629 * But if there are missing devices it may not be
2630 * safe to do the full stripe write yet.
2636 * the bbio may be freed once we submit the last bio. Make sure
2637 * not to touch it after that
2639 atomic_set(&rbio->stripes_pending, bios_to_read);
2640 while ((bio = bio_list_pop(&bio_list))) {
2641 bio->bi_private = rbio;
2642 bio->bi_end_io = raid56_parity_scrub_end_io;
2643 bio->bi_opf = REQ_OP_READ;
2645 btrfs_bio_wq_end_io(rbio->fs_info, bio, BTRFS_WQ_ENDIO_RAID56);
2649 /* the actual write will happen once the reads are done */
2653 rbio_orig_end_io(rbio, BLK_STS_IOERR);
2655 while ((bio = bio_list_pop(&bio_list)))
2661 validate_rbio_for_parity_scrub(rbio);
2664 static void scrub_parity_work(struct btrfs_work *work)
2666 struct btrfs_raid_bio *rbio;
2668 rbio = container_of(work, struct btrfs_raid_bio, work);
2669 raid56_parity_scrub_stripe(rbio);
2672 void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio)
2674 if (!lock_stripe_add(rbio))
2675 start_async_work(rbio, scrub_parity_work);
2678 /* The following code is used for dev replace of a missing RAID 5/6 device. */
2680 struct btrfs_raid_bio *
2681 raid56_alloc_missing_rbio(struct btrfs_fs_info *fs_info, struct bio *bio,
2682 struct btrfs_bio *bbio, u64 length)
2684 struct btrfs_raid_bio *rbio;
2686 rbio = alloc_rbio(fs_info, bbio, length);
2690 rbio->operation = BTRFS_RBIO_REBUILD_MISSING;
2691 bio_list_add(&rbio->bio_list, bio);
2693 * This is a special bio which is used to hold the completion handler
2694 * and make the scrub rbio is similar to the other types
2696 ASSERT(!bio->bi_iter.bi_size);
2698 rbio->faila = find_logical_bio_stripe(rbio, bio);
2699 if (rbio->faila == -1) {
2706 * When we get bbio, we have already increased bio_counter, record it
2707 * so we can free it at rbio_orig_end_io()
2709 rbio->generic_bio_cnt = 1;
2714 void raid56_submit_missing_rbio(struct btrfs_raid_bio *rbio)
2716 if (!lock_stripe_add(rbio))
2717 start_async_work(rbio, read_rebuild_work);