Loading...
1/*
2 * Copyright (C) 2012 Fusion-io All rights reserved.
3 * Copyright (C) 2012 Intel Corp. All rights reserved.
4 *
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public
7 * License v2 as published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public
15 * License along with this program; if not, write to the
16 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
17 * Boston, MA 021110-1307, USA.
18 */
19#include <linux/sched.h>
20#include <linux/wait.h>
21#include <linux/bio.h>
22#include <linux/slab.h>
23#include <linux/buffer_head.h>
24#include <linux/blkdev.h>
25#include <linux/random.h>
26#include <linux/iocontext.h>
27#include <linux/capability.h>
28#include <linux/ratelimit.h>
29#include <linux/kthread.h>
30#include <linux/raid/pq.h>
31#include <linux/hash.h>
32#include <linux/list_sort.h>
33#include <linux/raid/xor.h>
34#include <linux/vmalloc.h>
35#include <asm/div64.h>
36#include "ctree.h"
37#include "extent_map.h"
38#include "disk-io.h"
39#include "transaction.h"
40#include "print-tree.h"
41#include "volumes.h"
42#include "raid56.h"
43#include "async-thread.h"
44#include "check-integrity.h"
45#include "rcu-string.h"
46
47/* set when additional merges to this rbio are not allowed */
48#define RBIO_RMW_LOCKED_BIT 1
49
50/*
51 * set when this rbio is sitting in the hash, but it is just a cache
52 * of past RMW
53 */
54#define RBIO_CACHE_BIT 2
55
56/*
57 * set when it is safe to trust the stripe_pages for caching
58 */
59#define RBIO_CACHE_READY_BIT 3
60
61#define RBIO_CACHE_SIZE 1024
62
63enum btrfs_rbio_ops {
64 BTRFS_RBIO_WRITE,
65 BTRFS_RBIO_READ_REBUILD,
66 BTRFS_RBIO_PARITY_SCRUB,
67 BTRFS_RBIO_REBUILD_MISSING,
68};
69
70struct btrfs_raid_bio {
71 struct btrfs_fs_info *fs_info;
72 struct btrfs_bio *bbio;
73
74 /* while we're doing rmw on a stripe
75 * we put it into a hash table so we can
76 * lock the stripe and merge more rbios
77 * into it.
78 */
79 struct list_head hash_list;
80
81 /*
82 * LRU list for the stripe cache
83 */
84 struct list_head stripe_cache;
85
86 /*
87 * for scheduling work in the helper threads
88 */
89 struct btrfs_work work;
90
91 /*
92 * bio list and bio_list_lock are used
93 * to add more bios into the stripe
94 * in hopes of avoiding the full rmw
95 */
96 struct bio_list bio_list;
97 spinlock_t bio_list_lock;
98
99 /* also protected by the bio_list_lock, the
100 * plug list is used by the plugging code
101 * to collect partial bios while plugged. The
102 * stripe locking code also uses it to hand off
103 * the stripe lock to the next pending IO
104 */
105 struct list_head plug_list;
106
107 /*
108 * flags that tell us if it is safe to
109 * merge with this bio
110 */
111 unsigned long flags;
112
113 /* size of each individual stripe on disk */
114 int stripe_len;
115
116 /* number of data stripes (no p/q) */
117 int nr_data;
118
119 int real_stripes;
120
121 int stripe_npages;
122 /*
123 * set if we're doing a parity rebuild
124 * for a read from higher up, which is handled
125 * differently from a parity rebuild as part of
126 * rmw
127 */
128 enum btrfs_rbio_ops operation;
129
130 /* first bad stripe */
131 int faila;
132
133 /* second bad stripe (for raid6 use) */
134 int failb;
135
136 int scrubp;
137 /*
138 * number of pages needed to represent the full
139 * stripe
140 */
141 int nr_pages;
142
143 /*
144 * size of all the bios in the bio_list. This
145 * helps us decide if the rbio maps to a full
146 * stripe or not
147 */
148 int bio_list_bytes;
149
150 int generic_bio_cnt;
151
152 atomic_t refs;
153
154 atomic_t stripes_pending;
155
156 atomic_t error;
157 /*
158 * these are two arrays of pointers. We allocate the
159 * rbio big enough to hold them both and setup their
160 * locations when the rbio is allocated
161 */
162
163 /* pointers to pages that we allocated for
164 * reading/writing stripes directly from the disk (including P/Q)
165 */
166 struct page **stripe_pages;
167
168 /*
169 * pointers to the pages in the bio_list. Stored
170 * here for faster lookup
171 */
172 struct page **bio_pages;
173
174 /*
175 * bitmap to record which horizontal stripe has data
176 */
177 unsigned long *dbitmap;
178};
179
180static int __raid56_parity_recover(struct btrfs_raid_bio *rbio);
181static noinline void finish_rmw(struct btrfs_raid_bio *rbio);
182static void rmw_work(struct btrfs_work *work);
183static void read_rebuild_work(struct btrfs_work *work);
184static void async_rmw_stripe(struct btrfs_raid_bio *rbio);
185static void async_read_rebuild(struct btrfs_raid_bio *rbio);
186static int fail_bio_stripe(struct btrfs_raid_bio *rbio, struct bio *bio);
187static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed);
188static void __free_raid_bio(struct btrfs_raid_bio *rbio);
189static void index_rbio_pages(struct btrfs_raid_bio *rbio);
190static int alloc_rbio_pages(struct btrfs_raid_bio *rbio);
191
192static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
193 int need_check);
194static void async_scrub_parity(struct btrfs_raid_bio *rbio);
195
196/*
197 * the stripe hash table is used for locking, and to collect
198 * bios in hopes of making a full stripe
199 */
200int btrfs_alloc_stripe_hash_table(struct btrfs_fs_info *info)
201{
202 struct btrfs_stripe_hash_table *table;
203 struct btrfs_stripe_hash_table *x;
204 struct btrfs_stripe_hash *cur;
205 struct btrfs_stripe_hash *h;
206 int num_entries = 1 << BTRFS_STRIPE_HASH_TABLE_BITS;
207 int i;
208 int table_size;
209
210 if (info->stripe_hash_table)
211 return 0;
212
213 /*
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.
216 *
217 * Try harder to allocate and fallback to vmalloc to lower the chance
218 * of a failing mount.
219 */
220 table_size = sizeof(*table) + sizeof(*h) * num_entries;
221 table = kzalloc(table_size, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
222 if (!table) {
223 table = vzalloc(table_size);
224 if (!table)
225 return -ENOMEM;
226 }
227
228 spin_lock_init(&table->cache_lock);
229 INIT_LIST_HEAD(&table->stripe_cache);
230
231 h = table->table;
232
233 for (i = 0; i < num_entries; i++) {
234 cur = h + i;
235 INIT_LIST_HEAD(&cur->hash_list);
236 spin_lock_init(&cur->lock);
237 init_waitqueue_head(&cur->wait);
238 }
239
240 x = cmpxchg(&info->stripe_hash_table, NULL, table);
241 if (x)
242 kvfree(x);
243 return 0;
244}
245
246/*
247 * caching an rbio means to copy anything from the
248 * bio_pages array into the stripe_pages array. We
249 * use the page uptodate bit in the stripe cache array
250 * to indicate if it has valid data
251 *
252 * once the caching is done, we set the cache ready
253 * bit.
254 */
255static void cache_rbio_pages(struct btrfs_raid_bio *rbio)
256{
257 int i;
258 char *s;
259 char *d;
260 int ret;
261
262 ret = alloc_rbio_pages(rbio);
263 if (ret)
264 return;
265
266 for (i = 0; i < rbio->nr_pages; i++) {
267 if (!rbio->bio_pages[i])
268 continue;
269
270 s = kmap(rbio->bio_pages[i]);
271 d = kmap(rbio->stripe_pages[i]);
272
273 memcpy(d, s, PAGE_SIZE);
274
275 kunmap(rbio->bio_pages[i]);
276 kunmap(rbio->stripe_pages[i]);
277 SetPageUptodate(rbio->stripe_pages[i]);
278 }
279 set_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
280}
281
282/*
283 * we hash on the first logical address of the stripe
284 */
285static int rbio_bucket(struct btrfs_raid_bio *rbio)
286{
287 u64 num = rbio->bbio->raid_map[0];
288
289 /*
290 * we shift down quite a bit. We're using byte
291 * addressing, and most of the lower bits are zeros.
292 * This tends to upset hash_64, and it consistently
293 * returns just one or two different values.
294 *
295 * shifting off the lower bits fixes things.
296 */
297 return hash_64(num >> 16, BTRFS_STRIPE_HASH_TABLE_BITS);
298}
299
300/*
301 * stealing an rbio means taking all the uptodate pages from the stripe
302 * array in the source rbio and putting them into the destination rbio
303 */
304static void steal_rbio(struct btrfs_raid_bio *src, struct btrfs_raid_bio *dest)
305{
306 int i;
307 struct page *s;
308 struct page *d;
309
310 if (!test_bit(RBIO_CACHE_READY_BIT, &src->flags))
311 return;
312
313 for (i = 0; i < dest->nr_pages; i++) {
314 s = src->stripe_pages[i];
315 if (!s || !PageUptodate(s)) {
316 continue;
317 }
318
319 d = dest->stripe_pages[i];
320 if (d)
321 __free_page(d);
322
323 dest->stripe_pages[i] = s;
324 src->stripe_pages[i] = NULL;
325 }
326}
327
328/*
329 * merging means we take the bio_list from the victim and
330 * splice it into the destination. The victim should
331 * be discarded afterwards.
332 *
333 * must be called with dest->rbio_list_lock held
334 */
335static void merge_rbio(struct btrfs_raid_bio *dest,
336 struct btrfs_raid_bio *victim)
337{
338 bio_list_merge(&dest->bio_list, &victim->bio_list);
339 dest->bio_list_bytes += victim->bio_list_bytes;
340 dest->generic_bio_cnt += victim->generic_bio_cnt;
341 bio_list_init(&victim->bio_list);
342}
343
344/*
345 * used to prune items that are in the cache. The caller
346 * must hold the hash table lock.
347 */
348static void __remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
349{
350 int bucket = rbio_bucket(rbio);
351 struct btrfs_stripe_hash_table *table;
352 struct btrfs_stripe_hash *h;
353 int freeit = 0;
354
355 /*
356 * check the bit again under the hash table lock.
357 */
358 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
359 return;
360
361 table = rbio->fs_info->stripe_hash_table;
362 h = table->table + bucket;
363
364 /* hold the lock for the bucket because we may be
365 * removing it from the hash table
366 */
367 spin_lock(&h->lock);
368
369 /*
370 * hold the lock for the bio list because we need
371 * to make sure the bio list is empty
372 */
373 spin_lock(&rbio->bio_list_lock);
374
375 if (test_and_clear_bit(RBIO_CACHE_BIT, &rbio->flags)) {
376 list_del_init(&rbio->stripe_cache);
377 table->cache_size -= 1;
378 freeit = 1;
379
380 /* if the bio list isn't empty, this rbio is
381 * still involved in an IO. We take it out
382 * of the cache list, and drop the ref that
383 * was held for the list.
384 *
385 * If the bio_list was empty, we also remove
386 * the rbio from the hash_table, and drop
387 * the corresponding ref
388 */
389 if (bio_list_empty(&rbio->bio_list)) {
390 if (!list_empty(&rbio->hash_list)) {
391 list_del_init(&rbio->hash_list);
392 atomic_dec(&rbio->refs);
393 BUG_ON(!list_empty(&rbio->plug_list));
394 }
395 }
396 }
397
398 spin_unlock(&rbio->bio_list_lock);
399 spin_unlock(&h->lock);
400
401 if (freeit)
402 __free_raid_bio(rbio);
403}
404
405/*
406 * prune a given rbio from the cache
407 */
408static void remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
409{
410 struct btrfs_stripe_hash_table *table;
411 unsigned long flags;
412
413 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
414 return;
415
416 table = rbio->fs_info->stripe_hash_table;
417
418 spin_lock_irqsave(&table->cache_lock, flags);
419 __remove_rbio_from_cache(rbio);
420 spin_unlock_irqrestore(&table->cache_lock, flags);
421}
422
423/*
424 * remove everything in the cache
425 */
426static void btrfs_clear_rbio_cache(struct btrfs_fs_info *info)
427{
428 struct btrfs_stripe_hash_table *table;
429 unsigned long flags;
430 struct btrfs_raid_bio *rbio;
431
432 table = info->stripe_hash_table;
433
434 spin_lock_irqsave(&table->cache_lock, flags);
435 while (!list_empty(&table->stripe_cache)) {
436 rbio = list_entry(table->stripe_cache.next,
437 struct btrfs_raid_bio,
438 stripe_cache);
439 __remove_rbio_from_cache(rbio);
440 }
441 spin_unlock_irqrestore(&table->cache_lock, flags);
442}
443
444/*
445 * remove all cached entries and free the hash table
446 * used by unmount
447 */
448void btrfs_free_stripe_hash_table(struct btrfs_fs_info *info)
449{
450 if (!info->stripe_hash_table)
451 return;
452 btrfs_clear_rbio_cache(info);
453 kvfree(info->stripe_hash_table);
454 info->stripe_hash_table = NULL;
455}
456
457/*
458 * insert an rbio into the stripe cache. It
459 * must have already been prepared by calling
460 * cache_rbio_pages
461 *
462 * If this rbio was already cached, it gets
463 * moved to the front of the lru.
464 *
465 * If the size of the rbio cache is too big, we
466 * prune an item.
467 */
468static void cache_rbio(struct btrfs_raid_bio *rbio)
469{
470 struct btrfs_stripe_hash_table *table;
471 unsigned long flags;
472
473 if (!test_bit(RBIO_CACHE_READY_BIT, &rbio->flags))
474 return;
475
476 table = rbio->fs_info->stripe_hash_table;
477
478 spin_lock_irqsave(&table->cache_lock, flags);
479 spin_lock(&rbio->bio_list_lock);
480
481 /* bump our ref if we were not in the list before */
482 if (!test_and_set_bit(RBIO_CACHE_BIT, &rbio->flags))
483 atomic_inc(&rbio->refs);
484
485 if (!list_empty(&rbio->stripe_cache)){
486 list_move(&rbio->stripe_cache, &table->stripe_cache);
487 } else {
488 list_add(&rbio->stripe_cache, &table->stripe_cache);
489 table->cache_size += 1;
490 }
491
492 spin_unlock(&rbio->bio_list_lock);
493
494 if (table->cache_size > RBIO_CACHE_SIZE) {
495 struct btrfs_raid_bio *found;
496
497 found = list_entry(table->stripe_cache.prev,
498 struct btrfs_raid_bio,
499 stripe_cache);
500
501 if (found != rbio)
502 __remove_rbio_from_cache(found);
503 }
504
505 spin_unlock_irqrestore(&table->cache_lock, flags);
506}
507
508/*
509 * helper function to run the xor_blocks api. It is only
510 * able to do MAX_XOR_BLOCKS at a time, so we need to
511 * loop through.
512 */
513static void run_xor(void **pages, int src_cnt, ssize_t len)
514{
515 int src_off = 0;
516 int xor_src_cnt = 0;
517 void *dest = pages[src_cnt];
518
519 while(src_cnt > 0) {
520 xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS);
521 xor_blocks(xor_src_cnt, len, dest, pages + src_off);
522
523 src_cnt -= xor_src_cnt;
524 src_off += xor_src_cnt;
525 }
526}
527
528/*
529 * returns true if the bio list inside this rbio
530 * covers an entire stripe (no rmw required).
531 * Must be called with the bio list lock held, or
532 * at a time when you know it is impossible to add
533 * new bios into the list
534 */
535static int __rbio_is_full(struct btrfs_raid_bio *rbio)
536{
537 unsigned long size = rbio->bio_list_bytes;
538 int ret = 1;
539
540 if (size != rbio->nr_data * rbio->stripe_len)
541 ret = 0;
542
543 BUG_ON(size > rbio->nr_data * rbio->stripe_len);
544 return ret;
545}
546
547static int rbio_is_full(struct btrfs_raid_bio *rbio)
548{
549 unsigned long flags;
550 int ret;
551
552 spin_lock_irqsave(&rbio->bio_list_lock, flags);
553 ret = __rbio_is_full(rbio);
554 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
555 return ret;
556}
557
558/*
559 * returns 1 if it is safe to merge two rbios together.
560 * The merging is safe if the two rbios correspond to
561 * the same stripe and if they are both going in the same
562 * direction (read vs write), and if neither one is
563 * locked for final IO
564 *
565 * The caller is responsible for locking such that
566 * rmw_locked is safe to test
567 */
568static int rbio_can_merge(struct btrfs_raid_bio *last,
569 struct btrfs_raid_bio *cur)
570{
571 if (test_bit(RBIO_RMW_LOCKED_BIT, &last->flags) ||
572 test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags))
573 return 0;
574
575 /*
576 * we can't merge with cached rbios, since the
577 * idea is that when we merge the destination
578 * rbio is going to run our IO for us. We can
579 * steal from cached rbio's though, other functions
580 * handle that.
581 */
582 if (test_bit(RBIO_CACHE_BIT, &last->flags) ||
583 test_bit(RBIO_CACHE_BIT, &cur->flags))
584 return 0;
585
586 if (last->bbio->raid_map[0] !=
587 cur->bbio->raid_map[0])
588 return 0;
589
590 /* we can't merge with different operations */
591 if (last->operation != cur->operation)
592 return 0;
593 /*
594 * We've need read the full stripe from the drive.
595 * check and repair the parity and write the new results.
596 *
597 * We're not allowed to add any new bios to the
598 * bio list here, anyone else that wants to
599 * change this stripe needs to do their own rmw.
600 */
601 if (last->operation == BTRFS_RBIO_PARITY_SCRUB ||
602 cur->operation == BTRFS_RBIO_PARITY_SCRUB)
603 return 0;
604
605 if (last->operation == BTRFS_RBIO_REBUILD_MISSING ||
606 cur->operation == BTRFS_RBIO_REBUILD_MISSING)
607 return 0;
608
609 return 1;
610}
611
612static int rbio_stripe_page_index(struct btrfs_raid_bio *rbio, int stripe,
613 int index)
614{
615 return stripe * rbio->stripe_npages + index;
616}
617
618/*
619 * these are just the pages from the rbio array, not from anything
620 * the FS sent down to us
621 */
622static struct page *rbio_stripe_page(struct btrfs_raid_bio *rbio, int stripe,
623 int index)
624{
625 return rbio->stripe_pages[rbio_stripe_page_index(rbio, stripe, index)];
626}
627
628/*
629 * helper to index into the pstripe
630 */
631static struct page *rbio_pstripe_page(struct btrfs_raid_bio *rbio, int index)
632{
633 return rbio_stripe_page(rbio, rbio->nr_data, index);
634}
635
636/*
637 * helper to index into the qstripe, returns null
638 * if there is no qstripe
639 */
640static struct page *rbio_qstripe_page(struct btrfs_raid_bio *rbio, int index)
641{
642 if (rbio->nr_data + 1 == rbio->real_stripes)
643 return NULL;
644 return rbio_stripe_page(rbio, rbio->nr_data + 1, index);
645}
646
647/*
648 * The first stripe in the table for a logical address
649 * has the lock. rbios are added in one of three ways:
650 *
651 * 1) Nobody has the stripe locked yet. The rbio is given
652 * the lock and 0 is returned. The caller must start the IO
653 * themselves.
654 *
655 * 2) Someone has the stripe locked, but we're able to merge
656 * with the lock owner. The rbio is freed and the IO will
657 * start automatically along with the existing rbio. 1 is returned.
658 *
659 * 3) Someone has the stripe locked, but we're not able to merge.
660 * The rbio is added to the lock owner's plug list, or merged into
661 * an rbio already on the plug list. When the lock owner unlocks,
662 * the next rbio on the list is run and the IO is started automatically.
663 * 1 is returned
664 *
665 * If we return 0, the caller still owns the rbio and must continue with
666 * IO submission. If we return 1, the caller must assume the rbio has
667 * already been freed.
668 */
669static noinline int lock_stripe_add(struct btrfs_raid_bio *rbio)
670{
671 int bucket = rbio_bucket(rbio);
672 struct btrfs_stripe_hash *h = rbio->fs_info->stripe_hash_table->table + bucket;
673 struct btrfs_raid_bio *cur;
674 struct btrfs_raid_bio *pending;
675 unsigned long flags;
676 DEFINE_WAIT(wait);
677 struct btrfs_raid_bio *freeit = NULL;
678 struct btrfs_raid_bio *cache_drop = NULL;
679 int ret = 0;
680 int walk = 0;
681
682 spin_lock_irqsave(&h->lock, flags);
683 list_for_each_entry(cur, &h->hash_list, hash_list) {
684 walk++;
685 if (cur->bbio->raid_map[0] == rbio->bbio->raid_map[0]) {
686 spin_lock(&cur->bio_list_lock);
687
688 /* can we steal this cached rbio's pages? */
689 if (bio_list_empty(&cur->bio_list) &&
690 list_empty(&cur->plug_list) &&
691 test_bit(RBIO_CACHE_BIT, &cur->flags) &&
692 !test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) {
693 list_del_init(&cur->hash_list);
694 atomic_dec(&cur->refs);
695
696 steal_rbio(cur, rbio);
697 cache_drop = cur;
698 spin_unlock(&cur->bio_list_lock);
699
700 goto lockit;
701 }
702
703 /* can we merge into the lock owner? */
704 if (rbio_can_merge(cur, rbio)) {
705 merge_rbio(cur, rbio);
706 spin_unlock(&cur->bio_list_lock);
707 freeit = rbio;
708 ret = 1;
709 goto out;
710 }
711
712
713 /*
714 * we couldn't merge with the running
715 * rbio, see if we can merge with the
716 * pending ones. We don't have to
717 * check for rmw_locked because there
718 * is no way they are inside finish_rmw
719 * right now
720 */
721 list_for_each_entry(pending, &cur->plug_list,
722 plug_list) {
723 if (rbio_can_merge(pending, rbio)) {
724 merge_rbio(pending, rbio);
725 spin_unlock(&cur->bio_list_lock);
726 freeit = rbio;
727 ret = 1;
728 goto out;
729 }
730 }
731
732 /* no merging, put us on the tail of the plug list,
733 * our rbio will be started with the currently
734 * running rbio unlocks
735 */
736 list_add_tail(&rbio->plug_list, &cur->plug_list);
737 spin_unlock(&cur->bio_list_lock);
738 ret = 1;
739 goto out;
740 }
741 }
742lockit:
743 atomic_inc(&rbio->refs);
744 list_add(&rbio->hash_list, &h->hash_list);
745out:
746 spin_unlock_irqrestore(&h->lock, flags);
747 if (cache_drop)
748 remove_rbio_from_cache(cache_drop);
749 if (freeit)
750 __free_raid_bio(freeit);
751 return ret;
752}
753
754/*
755 * called as rmw or parity rebuild is completed. If the plug list has more
756 * rbios waiting for this stripe, the next one on the list will be started
757 */
758static noinline void unlock_stripe(struct btrfs_raid_bio *rbio)
759{
760 int bucket;
761 struct btrfs_stripe_hash *h;
762 unsigned long flags;
763 int keep_cache = 0;
764
765 bucket = rbio_bucket(rbio);
766 h = rbio->fs_info->stripe_hash_table->table + bucket;
767
768 if (list_empty(&rbio->plug_list))
769 cache_rbio(rbio);
770
771 spin_lock_irqsave(&h->lock, flags);
772 spin_lock(&rbio->bio_list_lock);
773
774 if (!list_empty(&rbio->hash_list)) {
775 /*
776 * if we're still cached and there is no other IO
777 * to perform, just leave this rbio here for others
778 * to steal from later
779 */
780 if (list_empty(&rbio->plug_list) &&
781 test_bit(RBIO_CACHE_BIT, &rbio->flags)) {
782 keep_cache = 1;
783 clear_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
784 BUG_ON(!bio_list_empty(&rbio->bio_list));
785 goto done;
786 }
787
788 list_del_init(&rbio->hash_list);
789 atomic_dec(&rbio->refs);
790
791 /*
792 * we use the plug list to hold all the rbios
793 * waiting for the chance to lock this stripe.
794 * hand the lock over to one of them.
795 */
796 if (!list_empty(&rbio->plug_list)) {
797 struct btrfs_raid_bio *next;
798 struct list_head *head = rbio->plug_list.next;
799
800 next = list_entry(head, struct btrfs_raid_bio,
801 plug_list);
802
803 list_del_init(&rbio->plug_list);
804
805 list_add(&next->hash_list, &h->hash_list);
806 atomic_inc(&next->refs);
807 spin_unlock(&rbio->bio_list_lock);
808 spin_unlock_irqrestore(&h->lock, flags);
809
810 if (next->operation == BTRFS_RBIO_READ_REBUILD)
811 async_read_rebuild(next);
812 else if (next->operation == BTRFS_RBIO_REBUILD_MISSING) {
813 steal_rbio(rbio, next);
814 async_read_rebuild(next);
815 } else if (next->operation == BTRFS_RBIO_WRITE) {
816 steal_rbio(rbio, next);
817 async_rmw_stripe(next);
818 } else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) {
819 steal_rbio(rbio, next);
820 async_scrub_parity(next);
821 }
822
823 goto done_nolock;
824 /*
825 * The barrier for this waitqueue_active is not needed,
826 * we're protected by h->lock and can't miss a wakeup.
827 */
828 } else if (waitqueue_active(&h->wait)) {
829 spin_unlock(&rbio->bio_list_lock);
830 spin_unlock_irqrestore(&h->lock, flags);
831 wake_up(&h->wait);
832 goto done_nolock;
833 }
834 }
835done:
836 spin_unlock(&rbio->bio_list_lock);
837 spin_unlock_irqrestore(&h->lock, flags);
838
839done_nolock:
840 if (!keep_cache)
841 remove_rbio_from_cache(rbio);
842}
843
844static void __free_raid_bio(struct btrfs_raid_bio *rbio)
845{
846 int i;
847
848 WARN_ON(atomic_read(&rbio->refs) < 0);
849 if (!atomic_dec_and_test(&rbio->refs))
850 return;
851
852 WARN_ON(!list_empty(&rbio->stripe_cache));
853 WARN_ON(!list_empty(&rbio->hash_list));
854 WARN_ON(!bio_list_empty(&rbio->bio_list));
855
856 for (i = 0; i < rbio->nr_pages; i++) {
857 if (rbio->stripe_pages[i]) {
858 __free_page(rbio->stripe_pages[i]);
859 rbio->stripe_pages[i] = NULL;
860 }
861 }
862
863 btrfs_put_bbio(rbio->bbio);
864 kfree(rbio);
865}
866
867static void free_raid_bio(struct btrfs_raid_bio *rbio)
868{
869 unlock_stripe(rbio);
870 __free_raid_bio(rbio);
871}
872
873/*
874 * this frees the rbio and runs through all the bios in the
875 * bio_list and calls end_io on them
876 */
877static void rbio_orig_end_io(struct btrfs_raid_bio *rbio, int err)
878{
879 struct bio *cur = bio_list_get(&rbio->bio_list);
880 struct bio *next;
881
882 if (rbio->generic_bio_cnt)
883 btrfs_bio_counter_sub(rbio->fs_info, rbio->generic_bio_cnt);
884
885 free_raid_bio(rbio);
886
887 while (cur) {
888 next = cur->bi_next;
889 cur->bi_next = NULL;
890 cur->bi_error = err;
891 bio_endio(cur);
892 cur = next;
893 }
894}
895
896/*
897 * end io function used by finish_rmw. When we finally
898 * get here, we've written a full stripe
899 */
900static void raid_write_end_io(struct bio *bio)
901{
902 struct btrfs_raid_bio *rbio = bio->bi_private;
903 int err = bio->bi_error;
904 int max_errors;
905
906 if (err)
907 fail_bio_stripe(rbio, bio);
908
909 bio_put(bio);
910
911 if (!atomic_dec_and_test(&rbio->stripes_pending))
912 return;
913
914 err = 0;
915
916 /* OK, we have read all the stripes we need to. */
917 max_errors = (rbio->operation == BTRFS_RBIO_PARITY_SCRUB) ?
918 0 : rbio->bbio->max_errors;
919 if (atomic_read(&rbio->error) > max_errors)
920 err = -EIO;
921
922 rbio_orig_end_io(rbio, err);
923}
924
925/*
926 * the read/modify/write code wants to use the original bio for
927 * any pages it included, and then use the rbio for everything
928 * else. This function decides if a given index (stripe number)
929 * and page number in that stripe fall inside the original bio
930 * or the rbio.
931 *
932 * if you set bio_list_only, you'll get a NULL back for any ranges
933 * that are outside the bio_list
934 *
935 * This doesn't take any refs on anything, you get a bare page pointer
936 * and the caller must bump refs as required.
937 *
938 * You must call index_rbio_pages once before you can trust
939 * the answers from this function.
940 */
941static struct page *page_in_rbio(struct btrfs_raid_bio *rbio,
942 int index, int pagenr, int bio_list_only)
943{
944 int chunk_page;
945 struct page *p = NULL;
946
947 chunk_page = index * (rbio->stripe_len >> PAGE_SHIFT) + pagenr;
948
949 spin_lock_irq(&rbio->bio_list_lock);
950 p = rbio->bio_pages[chunk_page];
951 spin_unlock_irq(&rbio->bio_list_lock);
952
953 if (p || bio_list_only)
954 return p;
955
956 return rbio->stripe_pages[chunk_page];
957}
958
959/*
960 * number of pages we need for the entire stripe across all the
961 * drives
962 */
963static unsigned long rbio_nr_pages(unsigned long stripe_len, int nr_stripes)
964{
965 return DIV_ROUND_UP(stripe_len, PAGE_SIZE) * nr_stripes;
966}
967
968/*
969 * allocation and initial setup for the btrfs_raid_bio. Not
970 * this does not allocate any pages for rbio->pages.
971 */
972static struct btrfs_raid_bio *alloc_rbio(struct btrfs_root *root,
973 struct btrfs_bio *bbio, u64 stripe_len)
974{
975 struct btrfs_raid_bio *rbio;
976 int nr_data = 0;
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);
980 void *p;
981
982 rbio = kzalloc(sizeof(*rbio) + num_pages * sizeof(struct page *) * 2 +
983 DIV_ROUND_UP(stripe_npages, BITS_PER_LONG) *
984 sizeof(long), GFP_NOFS);
985 if (!rbio)
986 return ERR_PTR(-ENOMEM);
987
988 bio_list_init(&rbio->bio_list);
989 INIT_LIST_HEAD(&rbio->plug_list);
990 spin_lock_init(&rbio->bio_list_lock);
991 INIT_LIST_HEAD(&rbio->stripe_cache);
992 INIT_LIST_HEAD(&rbio->hash_list);
993 rbio->bbio = bbio;
994 rbio->fs_info = root->fs_info;
995 rbio->stripe_len = stripe_len;
996 rbio->nr_pages = num_pages;
997 rbio->real_stripes = real_stripes;
998 rbio->stripe_npages = stripe_npages;
999 rbio->faila = -1;
1000 rbio->failb = -1;
1001 atomic_set(&rbio->refs, 1);
1002 atomic_set(&rbio->error, 0);
1003 atomic_set(&rbio->stripes_pending, 0);
1004
1005 /*
1006 * the stripe_pages and bio_pages array point to the extra
1007 * memory we allocated past the end of the rbio
1008 */
1009 p = rbio + 1;
1010 rbio->stripe_pages = p;
1011 rbio->bio_pages = p + sizeof(struct page *) * num_pages;
1012 rbio->dbitmap = p + sizeof(struct page *) * num_pages * 2;
1013
1014 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1015 nr_data = real_stripes - 1;
1016 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1017 nr_data = real_stripes - 2;
1018 else
1019 BUG();
1020
1021 rbio->nr_data = nr_data;
1022 return rbio;
1023}
1024
1025/* allocate pages for all the stripes in the bio, including parity */
1026static int alloc_rbio_pages(struct btrfs_raid_bio *rbio)
1027{
1028 int i;
1029 struct page *page;
1030
1031 for (i = 0; i < rbio->nr_pages; i++) {
1032 if (rbio->stripe_pages[i])
1033 continue;
1034 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1035 if (!page)
1036 return -ENOMEM;
1037 rbio->stripe_pages[i] = page;
1038 }
1039 return 0;
1040}
1041
1042/* only allocate pages for p/q stripes */
1043static int alloc_rbio_parity_pages(struct btrfs_raid_bio *rbio)
1044{
1045 int i;
1046 struct page *page;
1047
1048 i = rbio_stripe_page_index(rbio, rbio->nr_data, 0);
1049
1050 for (; i < rbio->nr_pages; i++) {
1051 if (rbio->stripe_pages[i])
1052 continue;
1053 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1054 if (!page)
1055 return -ENOMEM;
1056 rbio->stripe_pages[i] = page;
1057 }
1058 return 0;
1059}
1060
1061/*
1062 * add a single page from a specific stripe into our list of bios for IO
1063 * this will try to merge into existing bios if possible, and returns
1064 * zero if all went well.
1065 */
1066static int rbio_add_io_page(struct btrfs_raid_bio *rbio,
1067 struct bio_list *bio_list,
1068 struct page *page,
1069 int stripe_nr,
1070 unsigned long page_index,
1071 unsigned long bio_max_len)
1072{
1073 struct bio *last = bio_list->tail;
1074 u64 last_end = 0;
1075 int ret;
1076 struct bio *bio;
1077 struct btrfs_bio_stripe *stripe;
1078 u64 disk_start;
1079
1080 stripe = &rbio->bbio->stripes[stripe_nr];
1081 disk_start = stripe->physical + (page_index << PAGE_SHIFT);
1082
1083 /* if the device is missing, just fail this stripe */
1084 if (!stripe->dev->bdev)
1085 return fail_rbio_index(rbio, stripe_nr);
1086
1087 /* see if we can add this page onto our existing bio */
1088 if (last) {
1089 last_end = (u64)last->bi_iter.bi_sector << 9;
1090 last_end += last->bi_iter.bi_size;
1091
1092 /*
1093 * we can't merge these if they are from different
1094 * devices or if they are not contiguous
1095 */
1096 if (last_end == disk_start && stripe->dev->bdev &&
1097 !last->bi_error &&
1098 last->bi_bdev == stripe->dev->bdev) {
1099 ret = bio_add_page(last, page, PAGE_SIZE, 0);
1100 if (ret == PAGE_SIZE)
1101 return 0;
1102 }
1103 }
1104
1105 /* put a new bio on the list */
1106 bio = btrfs_io_bio_alloc(GFP_NOFS, bio_max_len >> PAGE_SHIFT?:1);
1107 if (!bio)
1108 return -ENOMEM;
1109
1110 bio->bi_iter.bi_size = 0;
1111 bio->bi_bdev = stripe->dev->bdev;
1112 bio->bi_iter.bi_sector = disk_start >> 9;
1113
1114 bio_add_page(bio, page, PAGE_SIZE, 0);
1115 bio_list_add(bio_list, bio);
1116 return 0;
1117}
1118
1119/*
1120 * while we're doing the read/modify/write cycle, we could
1121 * have errors in reading pages off the disk. This checks
1122 * for errors and if we're not able to read the page it'll
1123 * trigger parity reconstruction. The rmw will be finished
1124 * after we've reconstructed the failed stripes
1125 */
1126static void validate_rbio_for_rmw(struct btrfs_raid_bio *rbio)
1127{
1128 if (rbio->faila >= 0 || rbio->failb >= 0) {
1129 BUG_ON(rbio->faila == rbio->real_stripes - 1);
1130 __raid56_parity_recover(rbio);
1131 } else {
1132 finish_rmw(rbio);
1133 }
1134}
1135
1136/*
1137 * helper function to walk our bio list and populate the bio_pages array with
1138 * the result. This seems expensive, but it is faster than constantly
1139 * searching through the bio list as we setup the IO in finish_rmw or stripe
1140 * reconstruction.
1141 *
1142 * This must be called before you trust the answers from page_in_rbio
1143 */
1144static void index_rbio_pages(struct btrfs_raid_bio *rbio)
1145{
1146 struct bio *bio;
1147 u64 start;
1148 unsigned long stripe_offset;
1149 unsigned long page_index;
1150 struct page *p;
1151 int i;
1152
1153 spin_lock_irq(&rbio->bio_list_lock);
1154 bio_list_for_each(bio, &rbio->bio_list) {
1155 start = (u64)bio->bi_iter.bi_sector << 9;
1156 stripe_offset = start - rbio->bbio->raid_map[0];
1157 page_index = stripe_offset >> PAGE_SHIFT;
1158
1159 for (i = 0; i < bio->bi_vcnt; i++) {
1160 p = bio->bi_io_vec[i].bv_page;
1161 rbio->bio_pages[page_index + i] = p;
1162 }
1163 }
1164 spin_unlock_irq(&rbio->bio_list_lock);
1165}
1166
1167/*
1168 * this is called from one of two situations. We either
1169 * have a full stripe from the higher layers, or we've read all
1170 * the missing bits off disk.
1171 *
1172 * This will calculate the parity and then send down any
1173 * changed blocks.
1174 */
1175static noinline void finish_rmw(struct btrfs_raid_bio *rbio)
1176{
1177 struct btrfs_bio *bbio = rbio->bbio;
1178 void *pointers[rbio->real_stripes];
1179 int nr_data = rbio->nr_data;
1180 int stripe;
1181 int pagenr;
1182 int p_stripe = -1;
1183 int q_stripe = -1;
1184 struct bio_list bio_list;
1185 struct bio *bio;
1186 int ret;
1187
1188 bio_list_init(&bio_list);
1189
1190 if (rbio->real_stripes - rbio->nr_data == 1) {
1191 p_stripe = rbio->real_stripes - 1;
1192 } else if (rbio->real_stripes - rbio->nr_data == 2) {
1193 p_stripe = rbio->real_stripes - 2;
1194 q_stripe = rbio->real_stripes - 1;
1195 } else {
1196 BUG();
1197 }
1198
1199 /* at this point we either have a full stripe,
1200 * or we've read the full stripe from the drive.
1201 * recalculate the parity and write the new results.
1202 *
1203 * We're not allowed to add any new bios to the
1204 * bio list here, anyone else that wants to
1205 * change this stripe needs to do their own rmw.
1206 */
1207 spin_lock_irq(&rbio->bio_list_lock);
1208 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1209 spin_unlock_irq(&rbio->bio_list_lock);
1210
1211 atomic_set(&rbio->error, 0);
1212
1213 /*
1214 * now that we've set rmw_locked, run through the
1215 * bio list one last time and map the page pointers
1216 *
1217 * We don't cache full rbios because we're assuming
1218 * the higher layers are unlikely to use this area of
1219 * the disk again soon. If they do use it again,
1220 * hopefully they will send another full bio.
1221 */
1222 index_rbio_pages(rbio);
1223 if (!rbio_is_full(rbio))
1224 cache_rbio_pages(rbio);
1225 else
1226 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
1227
1228 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1229 struct page *p;
1230 /* first collect one page from each data stripe */
1231 for (stripe = 0; stripe < nr_data; stripe++) {
1232 p = page_in_rbio(rbio, stripe, pagenr, 0);
1233 pointers[stripe] = kmap(p);
1234 }
1235
1236 /* then add the parity stripe */
1237 p = rbio_pstripe_page(rbio, pagenr);
1238 SetPageUptodate(p);
1239 pointers[stripe++] = kmap(p);
1240
1241 if (q_stripe != -1) {
1242
1243 /*
1244 * raid6, add the qstripe and call the
1245 * library function to fill in our p/q
1246 */
1247 p = rbio_qstripe_page(rbio, pagenr);
1248 SetPageUptodate(p);
1249 pointers[stripe++] = kmap(p);
1250
1251 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
1252 pointers);
1253 } else {
1254 /* raid5 */
1255 memcpy(pointers[nr_data], pointers[0], PAGE_SIZE);
1256 run_xor(pointers + 1, nr_data - 1, PAGE_SIZE);
1257 }
1258
1259
1260 for (stripe = 0; stripe < rbio->real_stripes; stripe++)
1261 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
1262 }
1263
1264 /*
1265 * time to start writing. Make bios for everything from the
1266 * higher layers (the bio_list in our rbio) and our p/q. Ignore
1267 * everything else.
1268 */
1269 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1270 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1271 struct page *page;
1272 if (stripe < rbio->nr_data) {
1273 page = page_in_rbio(rbio, stripe, pagenr, 1);
1274 if (!page)
1275 continue;
1276 } else {
1277 page = rbio_stripe_page(rbio, stripe, pagenr);
1278 }
1279
1280 ret = rbio_add_io_page(rbio, &bio_list,
1281 page, stripe, pagenr, rbio->stripe_len);
1282 if (ret)
1283 goto cleanup;
1284 }
1285 }
1286
1287 if (likely(!bbio->num_tgtdevs))
1288 goto write_data;
1289
1290 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1291 if (!bbio->tgtdev_map[stripe])
1292 continue;
1293
1294 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1295 struct page *page;
1296 if (stripe < rbio->nr_data) {
1297 page = page_in_rbio(rbio, stripe, pagenr, 1);
1298 if (!page)
1299 continue;
1300 } else {
1301 page = rbio_stripe_page(rbio, stripe, pagenr);
1302 }
1303
1304 ret = rbio_add_io_page(rbio, &bio_list, page,
1305 rbio->bbio->tgtdev_map[stripe],
1306 pagenr, rbio->stripe_len);
1307 if (ret)
1308 goto cleanup;
1309 }
1310 }
1311
1312write_data:
1313 atomic_set(&rbio->stripes_pending, bio_list_size(&bio_list));
1314 BUG_ON(atomic_read(&rbio->stripes_pending) == 0);
1315
1316 while (1) {
1317 bio = bio_list_pop(&bio_list);
1318 if (!bio)
1319 break;
1320
1321 bio->bi_private = rbio;
1322 bio->bi_end_io = raid_write_end_io;
1323 submit_bio(WRITE, bio);
1324 }
1325 return;
1326
1327cleanup:
1328 rbio_orig_end_io(rbio, -EIO);
1329}
1330
1331/*
1332 * helper to find the stripe number for a given bio. Used to figure out which
1333 * stripe has failed. This expects the bio to correspond to a physical disk,
1334 * so it looks up based on physical sector numbers.
1335 */
1336static int find_bio_stripe(struct btrfs_raid_bio *rbio,
1337 struct bio *bio)
1338{
1339 u64 physical = bio->bi_iter.bi_sector;
1340 u64 stripe_start;
1341 int i;
1342 struct btrfs_bio_stripe *stripe;
1343
1344 physical <<= 9;
1345
1346 for (i = 0; i < rbio->bbio->num_stripes; i++) {
1347 stripe = &rbio->bbio->stripes[i];
1348 stripe_start = stripe->physical;
1349 if (physical >= stripe_start &&
1350 physical < stripe_start + rbio->stripe_len &&
1351 bio->bi_bdev == stripe->dev->bdev) {
1352 return i;
1353 }
1354 }
1355 return -1;
1356}
1357
1358/*
1359 * helper to find the stripe number for a given
1360 * bio (before mapping). Used to figure out which stripe has
1361 * failed. This looks up based on logical block numbers.
1362 */
1363static int find_logical_bio_stripe(struct btrfs_raid_bio *rbio,
1364 struct bio *bio)
1365{
1366 u64 logical = bio->bi_iter.bi_sector;
1367 u64 stripe_start;
1368 int i;
1369
1370 logical <<= 9;
1371
1372 for (i = 0; i < rbio->nr_data; i++) {
1373 stripe_start = rbio->bbio->raid_map[i];
1374 if (logical >= stripe_start &&
1375 logical < stripe_start + rbio->stripe_len) {
1376 return i;
1377 }
1378 }
1379 return -1;
1380}
1381
1382/*
1383 * returns -EIO if we had too many failures
1384 */
1385static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed)
1386{
1387 unsigned long flags;
1388 int ret = 0;
1389
1390 spin_lock_irqsave(&rbio->bio_list_lock, flags);
1391
1392 /* we already know this stripe is bad, move on */
1393 if (rbio->faila == failed || rbio->failb == failed)
1394 goto out;
1395
1396 if (rbio->faila == -1) {
1397 /* first failure on this rbio */
1398 rbio->faila = failed;
1399 atomic_inc(&rbio->error);
1400 } else if (rbio->failb == -1) {
1401 /* second failure on this rbio */
1402 rbio->failb = failed;
1403 atomic_inc(&rbio->error);
1404 } else {
1405 ret = -EIO;
1406 }
1407out:
1408 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
1409
1410 return ret;
1411}
1412
1413/*
1414 * helper to fail a stripe based on a physical disk
1415 * bio.
1416 */
1417static int fail_bio_stripe(struct btrfs_raid_bio *rbio,
1418 struct bio *bio)
1419{
1420 int failed = find_bio_stripe(rbio, bio);
1421
1422 if (failed < 0)
1423 return -EIO;
1424
1425 return fail_rbio_index(rbio, failed);
1426}
1427
1428/*
1429 * this sets each page in the bio uptodate. It should only be used on private
1430 * rbio pages, nothing that comes in from the higher layers
1431 */
1432static void set_bio_pages_uptodate(struct bio *bio)
1433{
1434 int i;
1435 struct page *p;
1436
1437 for (i = 0; i < bio->bi_vcnt; i++) {
1438 p = bio->bi_io_vec[i].bv_page;
1439 SetPageUptodate(p);
1440 }
1441}
1442
1443/*
1444 * end io for the read phase of the rmw cycle. All the bios here are physical
1445 * stripe bios we've read from the disk so we can recalculate the parity of the
1446 * stripe.
1447 *
1448 * This will usually kick off finish_rmw once all the bios are read in, but it
1449 * may trigger parity reconstruction if we had any errors along the way
1450 */
1451static void raid_rmw_end_io(struct bio *bio)
1452{
1453 struct btrfs_raid_bio *rbio = bio->bi_private;
1454
1455 if (bio->bi_error)
1456 fail_bio_stripe(rbio, bio);
1457 else
1458 set_bio_pages_uptodate(bio);
1459
1460 bio_put(bio);
1461
1462 if (!atomic_dec_and_test(&rbio->stripes_pending))
1463 return;
1464
1465 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
1466 goto cleanup;
1467
1468 /*
1469 * this will normally call finish_rmw to start our write
1470 * but if there are any failed stripes we'll reconstruct
1471 * from parity first
1472 */
1473 validate_rbio_for_rmw(rbio);
1474 return;
1475
1476cleanup:
1477
1478 rbio_orig_end_io(rbio, -EIO);
1479}
1480
1481static void async_rmw_stripe(struct btrfs_raid_bio *rbio)
1482{
1483 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
1484 rmw_work, NULL, NULL);
1485
1486 btrfs_queue_work(rbio->fs_info->rmw_workers,
1487 &rbio->work);
1488}
1489
1490static void async_read_rebuild(struct btrfs_raid_bio *rbio)
1491{
1492 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
1493 read_rebuild_work, NULL, NULL);
1494
1495 btrfs_queue_work(rbio->fs_info->rmw_workers,
1496 &rbio->work);
1497}
1498
1499/*
1500 * the stripe must be locked by the caller. It will
1501 * unlock after all the writes are done
1502 */
1503static int raid56_rmw_stripe(struct btrfs_raid_bio *rbio)
1504{
1505 int bios_to_read = 0;
1506 struct bio_list bio_list;
1507 int ret;
1508 int pagenr;
1509 int stripe;
1510 struct bio *bio;
1511
1512 bio_list_init(&bio_list);
1513
1514 ret = alloc_rbio_pages(rbio);
1515 if (ret)
1516 goto cleanup;
1517
1518 index_rbio_pages(rbio);
1519
1520 atomic_set(&rbio->error, 0);
1521 /*
1522 * build a list of bios to read all the missing parts of this
1523 * stripe
1524 */
1525 for (stripe = 0; stripe < rbio->nr_data; stripe++) {
1526 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1527 struct page *page;
1528 /*
1529 * we want to find all the pages missing from
1530 * the rbio and read them from the disk. If
1531 * page_in_rbio finds a page in the bio list
1532 * we don't need to read it off the stripe.
1533 */
1534 page = page_in_rbio(rbio, stripe, pagenr, 1);
1535 if (page)
1536 continue;
1537
1538 page = rbio_stripe_page(rbio, stripe, pagenr);
1539 /*
1540 * the bio cache may have handed us an uptodate
1541 * page. If so, be happy and use it
1542 */
1543 if (PageUptodate(page))
1544 continue;
1545
1546 ret = rbio_add_io_page(rbio, &bio_list, page,
1547 stripe, pagenr, rbio->stripe_len);
1548 if (ret)
1549 goto cleanup;
1550 }
1551 }
1552
1553 bios_to_read = bio_list_size(&bio_list);
1554 if (!bios_to_read) {
1555 /*
1556 * this can happen if others have merged with
1557 * us, it means there is nothing left to read.
1558 * But if there are missing devices it may not be
1559 * safe to do the full stripe write yet.
1560 */
1561 goto finish;
1562 }
1563
1564 /*
1565 * the bbio may be freed once we submit the last bio. Make sure
1566 * not to touch it after that
1567 */
1568 atomic_set(&rbio->stripes_pending, bios_to_read);
1569 while (1) {
1570 bio = bio_list_pop(&bio_list);
1571 if (!bio)
1572 break;
1573
1574 bio->bi_private = rbio;
1575 bio->bi_end_io = raid_rmw_end_io;
1576
1577 btrfs_bio_wq_end_io(rbio->fs_info, bio,
1578 BTRFS_WQ_ENDIO_RAID56);
1579
1580 submit_bio(READ, bio);
1581 }
1582 /* the actual write will happen once the reads are done */
1583 return 0;
1584
1585cleanup:
1586 rbio_orig_end_io(rbio, -EIO);
1587 return -EIO;
1588
1589finish:
1590 validate_rbio_for_rmw(rbio);
1591 return 0;
1592}
1593
1594/*
1595 * if the upper layers pass in a full stripe, we thank them by only allocating
1596 * enough pages to hold the parity, and sending it all down quickly.
1597 */
1598static int full_stripe_write(struct btrfs_raid_bio *rbio)
1599{
1600 int ret;
1601
1602 ret = alloc_rbio_parity_pages(rbio);
1603 if (ret) {
1604 __free_raid_bio(rbio);
1605 return ret;
1606 }
1607
1608 ret = lock_stripe_add(rbio);
1609 if (ret == 0)
1610 finish_rmw(rbio);
1611 return 0;
1612}
1613
1614/*
1615 * partial stripe writes get handed over to async helpers.
1616 * We're really hoping to merge a few more writes into this
1617 * rbio before calculating new parity
1618 */
1619static int partial_stripe_write(struct btrfs_raid_bio *rbio)
1620{
1621 int ret;
1622
1623 ret = lock_stripe_add(rbio);
1624 if (ret == 0)
1625 async_rmw_stripe(rbio);
1626 return 0;
1627}
1628
1629/*
1630 * sometimes while we were reading from the drive to
1631 * recalculate parity, enough new bios come into create
1632 * a full stripe. So we do a check here to see if we can
1633 * go directly to finish_rmw
1634 */
1635static int __raid56_parity_write(struct btrfs_raid_bio *rbio)
1636{
1637 /* head off into rmw land if we don't have a full stripe */
1638 if (!rbio_is_full(rbio))
1639 return partial_stripe_write(rbio);
1640 return full_stripe_write(rbio);
1641}
1642
1643/*
1644 * We use plugging call backs to collect full stripes.
1645 * Any time we get a partial stripe write while plugged
1646 * we collect it into a list. When the unplug comes down,
1647 * we sort the list by logical block number and merge
1648 * everything we can into the same rbios
1649 */
1650struct btrfs_plug_cb {
1651 struct blk_plug_cb cb;
1652 struct btrfs_fs_info *info;
1653 struct list_head rbio_list;
1654 struct btrfs_work work;
1655};
1656
1657/*
1658 * rbios on the plug list are sorted for easier merging.
1659 */
1660static int plug_cmp(void *priv, struct list_head *a, struct list_head *b)
1661{
1662 struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio,
1663 plug_list);
1664 struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio,
1665 plug_list);
1666 u64 a_sector = ra->bio_list.head->bi_iter.bi_sector;
1667 u64 b_sector = rb->bio_list.head->bi_iter.bi_sector;
1668
1669 if (a_sector < b_sector)
1670 return -1;
1671 if (a_sector > b_sector)
1672 return 1;
1673 return 0;
1674}
1675
1676static void run_plug(struct btrfs_plug_cb *plug)
1677{
1678 struct btrfs_raid_bio *cur;
1679 struct btrfs_raid_bio *last = NULL;
1680
1681 /*
1682 * sort our plug list then try to merge
1683 * everything we can in hopes of creating full
1684 * stripes.
1685 */
1686 list_sort(NULL, &plug->rbio_list, plug_cmp);
1687 while (!list_empty(&plug->rbio_list)) {
1688 cur = list_entry(plug->rbio_list.next,
1689 struct btrfs_raid_bio, plug_list);
1690 list_del_init(&cur->plug_list);
1691
1692 if (rbio_is_full(cur)) {
1693 /* we have a full stripe, send it down */
1694 full_stripe_write(cur);
1695 continue;
1696 }
1697 if (last) {
1698 if (rbio_can_merge(last, cur)) {
1699 merge_rbio(last, cur);
1700 __free_raid_bio(cur);
1701 continue;
1702
1703 }
1704 __raid56_parity_write(last);
1705 }
1706 last = cur;
1707 }
1708 if (last) {
1709 __raid56_parity_write(last);
1710 }
1711 kfree(plug);
1712}
1713
1714/*
1715 * if the unplug comes from schedule, we have to push the
1716 * work off to a helper thread
1717 */
1718static void unplug_work(struct btrfs_work *work)
1719{
1720 struct btrfs_plug_cb *plug;
1721 plug = container_of(work, struct btrfs_plug_cb, work);
1722 run_plug(plug);
1723}
1724
1725static void btrfs_raid_unplug(struct blk_plug_cb *cb, bool from_schedule)
1726{
1727 struct btrfs_plug_cb *plug;
1728 plug = container_of(cb, struct btrfs_plug_cb, cb);
1729
1730 if (from_schedule) {
1731 btrfs_init_work(&plug->work, btrfs_rmw_helper,
1732 unplug_work, NULL, NULL);
1733 btrfs_queue_work(plug->info->rmw_workers,
1734 &plug->work);
1735 return;
1736 }
1737 run_plug(plug);
1738}
1739
1740/*
1741 * our main entry point for writes from the rest of the FS.
1742 */
1743int raid56_parity_write(struct btrfs_root *root, struct bio *bio,
1744 struct btrfs_bio *bbio, u64 stripe_len)
1745{
1746 struct btrfs_raid_bio *rbio;
1747 struct btrfs_plug_cb *plug = NULL;
1748 struct blk_plug_cb *cb;
1749 int ret;
1750
1751 rbio = alloc_rbio(root, bbio, stripe_len);
1752 if (IS_ERR(rbio)) {
1753 btrfs_put_bbio(bbio);
1754 return PTR_ERR(rbio);
1755 }
1756 bio_list_add(&rbio->bio_list, bio);
1757 rbio->bio_list_bytes = bio->bi_iter.bi_size;
1758 rbio->operation = BTRFS_RBIO_WRITE;
1759
1760 btrfs_bio_counter_inc_noblocked(root->fs_info);
1761 rbio->generic_bio_cnt = 1;
1762
1763 /*
1764 * don't plug on full rbios, just get them out the door
1765 * as quickly as we can
1766 */
1767 if (rbio_is_full(rbio)) {
1768 ret = full_stripe_write(rbio);
1769 if (ret)
1770 btrfs_bio_counter_dec(root->fs_info);
1771 return ret;
1772 }
1773
1774 cb = blk_check_plugged(btrfs_raid_unplug, root->fs_info,
1775 sizeof(*plug));
1776 if (cb) {
1777 plug = container_of(cb, struct btrfs_plug_cb, cb);
1778 if (!plug->info) {
1779 plug->info = root->fs_info;
1780 INIT_LIST_HEAD(&plug->rbio_list);
1781 }
1782 list_add_tail(&rbio->plug_list, &plug->rbio_list);
1783 ret = 0;
1784 } else {
1785 ret = __raid56_parity_write(rbio);
1786 if (ret)
1787 btrfs_bio_counter_dec(root->fs_info);
1788 }
1789 return ret;
1790}
1791
1792/*
1793 * all parity reconstruction happens here. We've read in everything
1794 * we can find from the drives and this does the heavy lifting of
1795 * sorting the good from the bad.
1796 */
1797static void __raid_recover_end_io(struct btrfs_raid_bio *rbio)
1798{
1799 int pagenr, stripe;
1800 void **pointers;
1801 int faila = -1, failb = -1;
1802 struct page *page;
1803 int err;
1804 int i;
1805
1806 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
1807 if (!pointers) {
1808 err = -ENOMEM;
1809 goto cleanup_io;
1810 }
1811
1812 faila = rbio->faila;
1813 failb = rbio->failb;
1814
1815 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1816 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
1817 spin_lock_irq(&rbio->bio_list_lock);
1818 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1819 spin_unlock_irq(&rbio->bio_list_lock);
1820 }
1821
1822 index_rbio_pages(rbio);
1823
1824 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
1825 /*
1826 * Now we just use bitmap to mark the horizontal stripes in
1827 * which we have data when doing parity scrub.
1828 */
1829 if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB &&
1830 !test_bit(pagenr, rbio->dbitmap))
1831 continue;
1832
1833 /* setup our array of pointers with pages
1834 * from each stripe
1835 */
1836 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1837 /*
1838 * if we're rebuilding a read, we have to use
1839 * pages from the bio list
1840 */
1841 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1842 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
1843 (stripe == faila || stripe == failb)) {
1844 page = page_in_rbio(rbio, stripe, pagenr, 0);
1845 } else {
1846 page = rbio_stripe_page(rbio, stripe, pagenr);
1847 }
1848 pointers[stripe] = kmap(page);
1849 }
1850
1851 /* all raid6 handling here */
1852 if (rbio->bbio->map_type & BTRFS_BLOCK_GROUP_RAID6) {
1853 /*
1854 * single failure, rebuild from parity raid5
1855 * style
1856 */
1857 if (failb < 0) {
1858 if (faila == rbio->nr_data) {
1859 /*
1860 * Just the P stripe has failed, without
1861 * a bad data or Q stripe.
1862 * TODO, we should redo the xor here.
1863 */
1864 err = -EIO;
1865 goto cleanup;
1866 }
1867 /*
1868 * a single failure in raid6 is rebuilt
1869 * in the pstripe code below
1870 */
1871 goto pstripe;
1872 }
1873
1874 /* make sure our ps and qs are in order */
1875 if (faila > failb) {
1876 int tmp = failb;
1877 failb = faila;
1878 faila = tmp;
1879 }
1880
1881 /* if the q stripe is failed, do a pstripe reconstruction
1882 * from the xors.
1883 * If both the q stripe and the P stripe are failed, we're
1884 * here due to a crc mismatch and we can't give them the
1885 * data they want
1886 */
1887 if (rbio->bbio->raid_map[failb] == RAID6_Q_STRIPE) {
1888 if (rbio->bbio->raid_map[faila] ==
1889 RAID5_P_STRIPE) {
1890 err = -EIO;
1891 goto cleanup;
1892 }
1893 /*
1894 * otherwise we have one bad data stripe and
1895 * a good P stripe. raid5!
1896 */
1897 goto pstripe;
1898 }
1899
1900 if (rbio->bbio->raid_map[failb] == RAID5_P_STRIPE) {
1901 raid6_datap_recov(rbio->real_stripes,
1902 PAGE_SIZE, faila, pointers);
1903 } else {
1904 raid6_2data_recov(rbio->real_stripes,
1905 PAGE_SIZE, faila, failb,
1906 pointers);
1907 }
1908 } else {
1909 void *p;
1910
1911 /* rebuild from P stripe here (raid5 or raid6) */
1912 BUG_ON(failb != -1);
1913pstripe:
1914 /* Copy parity block into failed block to start with */
1915 memcpy(pointers[faila],
1916 pointers[rbio->nr_data],
1917 PAGE_SIZE);
1918
1919 /* rearrange the pointer array */
1920 p = pointers[faila];
1921 for (stripe = faila; stripe < rbio->nr_data - 1; stripe++)
1922 pointers[stripe] = pointers[stripe + 1];
1923 pointers[rbio->nr_data - 1] = p;
1924
1925 /* xor in the rest */
1926 run_xor(pointers, rbio->nr_data - 1, PAGE_SIZE);
1927 }
1928 /* if we're doing this rebuild as part of an rmw, go through
1929 * and set all of our private rbio pages in the
1930 * failed stripes as uptodate. This way finish_rmw will
1931 * know they can be trusted. If this was a read reconstruction,
1932 * other endio functions will fiddle the uptodate bits
1933 */
1934 if (rbio->operation == BTRFS_RBIO_WRITE) {
1935 for (i = 0; i < rbio->stripe_npages; i++) {
1936 if (faila != -1) {
1937 page = rbio_stripe_page(rbio, faila, i);
1938 SetPageUptodate(page);
1939 }
1940 if (failb != -1) {
1941 page = rbio_stripe_page(rbio, failb, i);
1942 SetPageUptodate(page);
1943 }
1944 }
1945 }
1946 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1947 /*
1948 * if we're rebuilding a read, we have to use
1949 * pages from the bio list
1950 */
1951 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1952 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
1953 (stripe == faila || stripe == failb)) {
1954 page = page_in_rbio(rbio, stripe, pagenr, 0);
1955 } else {
1956 page = rbio_stripe_page(rbio, stripe, pagenr);
1957 }
1958 kunmap(page);
1959 }
1960 }
1961
1962 err = 0;
1963cleanup:
1964 kfree(pointers);
1965
1966cleanup_io:
1967 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
1968 if (err == 0)
1969 cache_rbio_pages(rbio);
1970 else
1971 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
1972
1973 rbio_orig_end_io(rbio, err);
1974 } else if (rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
1975 rbio_orig_end_io(rbio, err);
1976 } else if (err == 0) {
1977 rbio->faila = -1;
1978 rbio->failb = -1;
1979
1980 if (rbio->operation == BTRFS_RBIO_WRITE)
1981 finish_rmw(rbio);
1982 else if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB)
1983 finish_parity_scrub(rbio, 0);
1984 else
1985 BUG();
1986 } else {
1987 rbio_orig_end_io(rbio, err);
1988 }
1989}
1990
1991/*
1992 * This is called only for stripes we've read from disk to
1993 * reconstruct the parity.
1994 */
1995static void raid_recover_end_io(struct bio *bio)
1996{
1997 struct btrfs_raid_bio *rbio = bio->bi_private;
1998
1999 /*
2000 * we only read stripe pages off the disk, set them
2001 * up to date if there were no errors
2002 */
2003 if (bio->bi_error)
2004 fail_bio_stripe(rbio, bio);
2005 else
2006 set_bio_pages_uptodate(bio);
2007 bio_put(bio);
2008
2009 if (!atomic_dec_and_test(&rbio->stripes_pending))
2010 return;
2011
2012 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
2013 rbio_orig_end_io(rbio, -EIO);
2014 else
2015 __raid_recover_end_io(rbio);
2016}
2017
2018/*
2019 * reads everything we need off the disk to reconstruct
2020 * the parity. endio handlers trigger final reconstruction
2021 * when the IO is done.
2022 *
2023 * This is used both for reads from the higher layers and for
2024 * parity construction required to finish a rmw cycle.
2025 */
2026static int __raid56_parity_recover(struct btrfs_raid_bio *rbio)
2027{
2028 int bios_to_read = 0;
2029 struct bio_list bio_list;
2030 int ret;
2031 int pagenr;
2032 int stripe;
2033 struct bio *bio;
2034
2035 bio_list_init(&bio_list);
2036
2037 ret = alloc_rbio_pages(rbio);
2038 if (ret)
2039 goto cleanup;
2040
2041 atomic_set(&rbio->error, 0);
2042
2043 /*
2044 * read everything that hasn't failed. Thanks to the
2045 * stripe cache, it is possible that some or all of these
2046 * pages are going to be uptodate.
2047 */
2048 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
2049 if (rbio->faila == stripe || rbio->failb == stripe) {
2050 atomic_inc(&rbio->error);
2051 continue;
2052 }
2053
2054 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
2055 struct page *p;
2056
2057 /*
2058 * the rmw code may have already read this
2059 * page in
2060 */
2061 p = rbio_stripe_page(rbio, stripe, pagenr);
2062 if (PageUptodate(p))
2063 continue;
2064
2065 ret = rbio_add_io_page(rbio, &bio_list,
2066 rbio_stripe_page(rbio, stripe, pagenr),
2067 stripe, pagenr, rbio->stripe_len);
2068 if (ret < 0)
2069 goto cleanup;
2070 }
2071 }
2072
2073 bios_to_read = bio_list_size(&bio_list);
2074 if (!bios_to_read) {
2075 /*
2076 * we might have no bios to read just because the pages
2077 * were up to date, or we might have no bios to read because
2078 * the devices were gone.
2079 */
2080 if (atomic_read(&rbio->error) <= rbio->bbio->max_errors) {
2081 __raid_recover_end_io(rbio);
2082 goto out;
2083 } else {
2084 goto cleanup;
2085 }
2086 }
2087
2088 /*
2089 * the bbio may be freed once we submit the last bio. Make sure
2090 * not to touch it after that
2091 */
2092 atomic_set(&rbio->stripes_pending, bios_to_read);
2093 while (1) {
2094 bio = bio_list_pop(&bio_list);
2095 if (!bio)
2096 break;
2097
2098 bio->bi_private = rbio;
2099 bio->bi_end_io = raid_recover_end_io;
2100
2101 btrfs_bio_wq_end_io(rbio->fs_info, bio,
2102 BTRFS_WQ_ENDIO_RAID56);
2103
2104 submit_bio(READ, bio);
2105 }
2106out:
2107 return 0;
2108
2109cleanup:
2110 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
2111 rbio->operation == BTRFS_RBIO_REBUILD_MISSING)
2112 rbio_orig_end_io(rbio, -EIO);
2113 return -EIO;
2114}
2115
2116/*
2117 * the main entry point for reads from the higher layers. This
2118 * is really only called when the normal read path had a failure,
2119 * so we assume the bio they send down corresponds to a failed part
2120 * of the drive.
2121 */
2122int raid56_parity_recover(struct btrfs_root *root, struct bio *bio,
2123 struct btrfs_bio *bbio, u64 stripe_len,
2124 int mirror_num, int generic_io)
2125{
2126 struct btrfs_raid_bio *rbio;
2127 int ret;
2128
2129 rbio = alloc_rbio(root, bbio, stripe_len);
2130 if (IS_ERR(rbio)) {
2131 if (generic_io)
2132 btrfs_put_bbio(bbio);
2133 return PTR_ERR(rbio);
2134 }
2135
2136 rbio->operation = BTRFS_RBIO_READ_REBUILD;
2137 bio_list_add(&rbio->bio_list, bio);
2138 rbio->bio_list_bytes = bio->bi_iter.bi_size;
2139
2140 rbio->faila = find_logical_bio_stripe(rbio, bio);
2141 if (rbio->faila == -1) {
2142 BUG();
2143 if (generic_io)
2144 btrfs_put_bbio(bbio);
2145 kfree(rbio);
2146 return -EIO;
2147 }
2148
2149 if (generic_io) {
2150 btrfs_bio_counter_inc_noblocked(root->fs_info);
2151 rbio->generic_bio_cnt = 1;
2152 } else {
2153 btrfs_get_bbio(bbio);
2154 }
2155
2156 /*
2157 * reconstruct from the q stripe if they are
2158 * asking for mirror 3
2159 */
2160 if (mirror_num == 3)
2161 rbio->failb = rbio->real_stripes - 2;
2162
2163 ret = lock_stripe_add(rbio);
2164
2165 /*
2166 * __raid56_parity_recover will end the bio with
2167 * any errors it hits. We don't want to return
2168 * its error value up the stack because our caller
2169 * will end up calling bio_endio with any nonzero
2170 * return
2171 */
2172 if (ret == 0)
2173 __raid56_parity_recover(rbio);
2174 /*
2175 * our rbio has been added to the list of
2176 * rbios that will be handled after the
2177 * currently lock owner is done
2178 */
2179 return 0;
2180
2181}
2182
2183static void rmw_work(struct btrfs_work *work)
2184{
2185 struct btrfs_raid_bio *rbio;
2186
2187 rbio = container_of(work, struct btrfs_raid_bio, work);
2188 raid56_rmw_stripe(rbio);
2189}
2190
2191static void read_rebuild_work(struct btrfs_work *work)
2192{
2193 struct btrfs_raid_bio *rbio;
2194
2195 rbio = container_of(work, struct btrfs_raid_bio, work);
2196 __raid56_parity_recover(rbio);
2197}
2198
2199/*
2200 * The following code is used to scrub/replace the parity stripe
2201 *
2202 * Note: We need make sure all the pages that add into the scrub/replace
2203 * raid bio are correct and not be changed during the scrub/replace. That
2204 * is those pages just hold metadata or file data with checksum.
2205 */
2206
2207struct btrfs_raid_bio *
2208raid56_parity_alloc_scrub_rbio(struct btrfs_root *root, struct bio *bio,
2209 struct btrfs_bio *bbio, u64 stripe_len,
2210 struct btrfs_device *scrub_dev,
2211 unsigned long *dbitmap, int stripe_nsectors)
2212{
2213 struct btrfs_raid_bio *rbio;
2214 int i;
2215
2216 rbio = alloc_rbio(root, bbio, stripe_len);
2217 if (IS_ERR(rbio))
2218 return NULL;
2219 bio_list_add(&rbio->bio_list, bio);
2220 /*
2221 * This is a special bio which is used to hold the completion handler
2222 * and make the scrub rbio is similar to the other types
2223 */
2224 ASSERT(!bio->bi_iter.bi_size);
2225 rbio->operation = BTRFS_RBIO_PARITY_SCRUB;
2226
2227 for (i = 0; i < rbio->real_stripes; i++) {
2228 if (bbio->stripes[i].dev == scrub_dev) {
2229 rbio->scrubp = i;
2230 break;
2231 }
2232 }
2233
2234 /* Now we just support the sectorsize equals to page size */
2235 ASSERT(root->sectorsize == PAGE_SIZE);
2236 ASSERT(rbio->stripe_npages == stripe_nsectors);
2237 bitmap_copy(rbio->dbitmap, dbitmap, stripe_nsectors);
2238
2239 return rbio;
2240}
2241
2242/* Used for both parity scrub and missing. */
2243void raid56_add_scrub_pages(struct btrfs_raid_bio *rbio, struct page *page,
2244 u64 logical)
2245{
2246 int stripe_offset;
2247 int index;
2248
2249 ASSERT(logical >= rbio->bbio->raid_map[0]);
2250 ASSERT(logical + PAGE_SIZE <= rbio->bbio->raid_map[0] +
2251 rbio->stripe_len * rbio->nr_data);
2252 stripe_offset = (int)(logical - rbio->bbio->raid_map[0]);
2253 index = stripe_offset >> PAGE_SHIFT;
2254 rbio->bio_pages[index] = page;
2255}
2256
2257/*
2258 * We just scrub the parity that we have correct data on the same horizontal,
2259 * so we needn't allocate all pages for all the stripes.
2260 */
2261static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio)
2262{
2263 int i;
2264 int bit;
2265 int index;
2266 struct page *page;
2267
2268 for_each_set_bit(bit, rbio->dbitmap, rbio->stripe_npages) {
2269 for (i = 0; i < rbio->real_stripes; i++) {
2270 index = i * rbio->stripe_npages + bit;
2271 if (rbio->stripe_pages[index])
2272 continue;
2273
2274 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2275 if (!page)
2276 return -ENOMEM;
2277 rbio->stripe_pages[index] = page;
2278 }
2279 }
2280 return 0;
2281}
2282
2283static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
2284 int need_check)
2285{
2286 struct btrfs_bio *bbio = rbio->bbio;
2287 void *pointers[rbio->real_stripes];
2288 DECLARE_BITMAP(pbitmap, rbio->stripe_npages);
2289 int nr_data = rbio->nr_data;
2290 int stripe;
2291 int pagenr;
2292 int p_stripe = -1;
2293 int q_stripe = -1;
2294 struct page *p_page = NULL;
2295 struct page *q_page = NULL;
2296 struct bio_list bio_list;
2297 struct bio *bio;
2298 int is_replace = 0;
2299 int ret;
2300
2301 bio_list_init(&bio_list);
2302
2303 if (rbio->real_stripes - rbio->nr_data == 1) {
2304 p_stripe = rbio->real_stripes - 1;
2305 } else if (rbio->real_stripes - rbio->nr_data == 2) {
2306 p_stripe = rbio->real_stripes - 2;
2307 q_stripe = rbio->real_stripes - 1;
2308 } else {
2309 BUG();
2310 }
2311
2312 if (bbio->num_tgtdevs && bbio->tgtdev_map[rbio->scrubp]) {
2313 is_replace = 1;
2314 bitmap_copy(pbitmap, rbio->dbitmap, rbio->stripe_npages);
2315 }
2316
2317 /*
2318 * Because the higher layers(scrubber) are unlikely to
2319 * use this area of the disk again soon, so don't cache
2320 * it.
2321 */
2322 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2323
2324 if (!need_check)
2325 goto writeback;
2326
2327 p_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2328 if (!p_page)
2329 goto cleanup;
2330 SetPageUptodate(p_page);
2331
2332 if (q_stripe != -1) {
2333 q_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2334 if (!q_page) {
2335 __free_page(p_page);
2336 goto cleanup;
2337 }
2338 SetPageUptodate(q_page);
2339 }
2340
2341 atomic_set(&rbio->error, 0);
2342
2343 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2344 struct page *p;
2345 void *parity;
2346 /* first collect one page from each data stripe */
2347 for (stripe = 0; stripe < nr_data; stripe++) {
2348 p = page_in_rbio(rbio, stripe, pagenr, 0);
2349 pointers[stripe] = kmap(p);
2350 }
2351
2352 /* then add the parity stripe */
2353 pointers[stripe++] = kmap(p_page);
2354
2355 if (q_stripe != -1) {
2356
2357 /*
2358 * raid6, add the qstripe and call the
2359 * library function to fill in our p/q
2360 */
2361 pointers[stripe++] = kmap(q_page);
2362
2363 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
2364 pointers);
2365 } else {
2366 /* raid5 */
2367 memcpy(pointers[nr_data], pointers[0], PAGE_SIZE);
2368 run_xor(pointers + 1, nr_data - 1, PAGE_SIZE);
2369 }
2370
2371 /* Check scrubbing pairty and repair it */
2372 p = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2373 parity = kmap(p);
2374 if (memcmp(parity, pointers[rbio->scrubp], PAGE_SIZE))
2375 memcpy(parity, pointers[rbio->scrubp], PAGE_SIZE);
2376 else
2377 /* Parity is right, needn't writeback */
2378 bitmap_clear(rbio->dbitmap, pagenr, 1);
2379 kunmap(p);
2380
2381 for (stripe = 0; stripe < rbio->real_stripes; stripe++)
2382 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
2383 }
2384
2385 __free_page(p_page);
2386 if (q_page)
2387 __free_page(q_page);
2388
2389writeback:
2390 /*
2391 * time to start writing. Make bios for everything from the
2392 * higher layers (the bio_list in our rbio) and our p/q. Ignore
2393 * everything else.
2394 */
2395 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2396 struct page *page;
2397
2398 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2399 ret = rbio_add_io_page(rbio, &bio_list,
2400 page, rbio->scrubp, pagenr, rbio->stripe_len);
2401 if (ret)
2402 goto cleanup;
2403 }
2404
2405 if (!is_replace)
2406 goto submit_write;
2407
2408 for_each_set_bit(pagenr, pbitmap, rbio->stripe_npages) {
2409 struct page *page;
2410
2411 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2412 ret = rbio_add_io_page(rbio, &bio_list, page,
2413 bbio->tgtdev_map[rbio->scrubp],
2414 pagenr, rbio->stripe_len);
2415 if (ret)
2416 goto cleanup;
2417 }
2418
2419submit_write:
2420 nr_data = bio_list_size(&bio_list);
2421 if (!nr_data) {
2422 /* Every parity is right */
2423 rbio_orig_end_io(rbio, 0);
2424 return;
2425 }
2426
2427 atomic_set(&rbio->stripes_pending, nr_data);
2428
2429 while (1) {
2430 bio = bio_list_pop(&bio_list);
2431 if (!bio)
2432 break;
2433
2434 bio->bi_private = rbio;
2435 bio->bi_end_io = raid_write_end_io;
2436 submit_bio(WRITE, bio);
2437 }
2438 return;
2439
2440cleanup:
2441 rbio_orig_end_io(rbio, -EIO);
2442}
2443
2444static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe)
2445{
2446 if (stripe >= 0 && stripe < rbio->nr_data)
2447 return 1;
2448 return 0;
2449}
2450
2451/*
2452 * While we're doing the parity check and repair, we could have errors
2453 * in reading pages off the disk. This checks for errors and if we're
2454 * not able to read the page it'll trigger parity reconstruction. The
2455 * parity scrub will be finished after we've reconstructed the failed
2456 * stripes
2457 */
2458static void validate_rbio_for_parity_scrub(struct btrfs_raid_bio *rbio)
2459{
2460 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
2461 goto cleanup;
2462
2463 if (rbio->faila >= 0 || rbio->failb >= 0) {
2464 int dfail = 0, failp = -1;
2465
2466 if (is_data_stripe(rbio, rbio->faila))
2467 dfail++;
2468 else if (is_parity_stripe(rbio->faila))
2469 failp = rbio->faila;
2470
2471 if (is_data_stripe(rbio, rbio->failb))
2472 dfail++;
2473 else if (is_parity_stripe(rbio->failb))
2474 failp = rbio->failb;
2475
2476 /*
2477 * Because we can not use a scrubbing parity to repair
2478 * the data, so the capability of the repair is declined.
2479 * (In the case of RAID5, we can not repair anything)
2480 */
2481 if (dfail > rbio->bbio->max_errors - 1)
2482 goto cleanup;
2483
2484 /*
2485 * If all data is good, only parity is correctly, just
2486 * repair the parity.
2487 */
2488 if (dfail == 0) {
2489 finish_parity_scrub(rbio, 0);
2490 return;
2491 }
2492
2493 /*
2494 * Here means we got one corrupted data stripe and one
2495 * corrupted parity on RAID6, if the corrupted parity
2496 * is scrubbing parity, luckly, use the other one to repair
2497 * the data, or we can not repair the data stripe.
2498 */
2499 if (failp != rbio->scrubp)
2500 goto cleanup;
2501
2502 __raid_recover_end_io(rbio);
2503 } else {
2504 finish_parity_scrub(rbio, 1);
2505 }
2506 return;
2507
2508cleanup:
2509 rbio_orig_end_io(rbio, -EIO);
2510}
2511
2512/*
2513 * end io for the read phase of the rmw cycle. All the bios here are physical
2514 * stripe bios we've read from the disk so we can recalculate the parity of the
2515 * stripe.
2516 *
2517 * This will usually kick off finish_rmw once all the bios are read in, but it
2518 * may trigger parity reconstruction if we had any errors along the way
2519 */
2520static void raid56_parity_scrub_end_io(struct bio *bio)
2521{
2522 struct btrfs_raid_bio *rbio = bio->bi_private;
2523
2524 if (bio->bi_error)
2525 fail_bio_stripe(rbio, bio);
2526 else
2527 set_bio_pages_uptodate(bio);
2528
2529 bio_put(bio);
2530
2531 if (!atomic_dec_and_test(&rbio->stripes_pending))
2532 return;
2533
2534 /*
2535 * this will normally call finish_rmw to start our write
2536 * but if there are any failed stripes we'll reconstruct
2537 * from parity first
2538 */
2539 validate_rbio_for_parity_scrub(rbio);
2540}
2541
2542static void raid56_parity_scrub_stripe(struct btrfs_raid_bio *rbio)
2543{
2544 int bios_to_read = 0;
2545 struct bio_list bio_list;
2546 int ret;
2547 int pagenr;
2548 int stripe;
2549 struct bio *bio;
2550
2551 ret = alloc_rbio_essential_pages(rbio);
2552 if (ret)
2553 goto cleanup;
2554
2555 bio_list_init(&bio_list);
2556
2557 atomic_set(&rbio->error, 0);
2558 /*
2559 * build a list of bios to read all the missing parts of this
2560 * stripe
2561 */
2562 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
2563 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2564 struct page *page;
2565 /*
2566 * we want to find all the pages missing from
2567 * the rbio and read them from the disk. If
2568 * page_in_rbio finds a page in the bio list
2569 * we don't need to read it off the stripe.
2570 */
2571 page = page_in_rbio(rbio, stripe, pagenr, 1);
2572 if (page)
2573 continue;
2574
2575 page = rbio_stripe_page(rbio, stripe, pagenr);
2576 /*
2577 * the bio cache may have handed us an uptodate
2578 * page. If so, be happy and use it
2579 */
2580 if (PageUptodate(page))
2581 continue;
2582
2583 ret = rbio_add_io_page(rbio, &bio_list, page,
2584 stripe, pagenr, rbio->stripe_len);
2585 if (ret)
2586 goto cleanup;
2587 }
2588 }
2589
2590 bios_to_read = bio_list_size(&bio_list);
2591 if (!bios_to_read) {
2592 /*
2593 * this can happen if others have merged with
2594 * us, it means there is nothing left to read.
2595 * But if there are missing devices it may not be
2596 * safe to do the full stripe write yet.
2597 */
2598 goto finish;
2599 }
2600
2601 /*
2602 * the bbio may be freed once we submit the last bio. Make sure
2603 * not to touch it after that
2604 */
2605 atomic_set(&rbio->stripes_pending, bios_to_read);
2606 while (1) {
2607 bio = bio_list_pop(&bio_list);
2608 if (!bio)
2609 break;
2610
2611 bio->bi_private = rbio;
2612 bio->bi_end_io = raid56_parity_scrub_end_io;
2613
2614 btrfs_bio_wq_end_io(rbio->fs_info, bio,
2615 BTRFS_WQ_ENDIO_RAID56);
2616
2617 submit_bio(READ, bio);
2618 }
2619 /* the actual write will happen once the reads are done */
2620 return;
2621
2622cleanup:
2623 rbio_orig_end_io(rbio, -EIO);
2624 return;
2625
2626finish:
2627 validate_rbio_for_parity_scrub(rbio);
2628}
2629
2630static void scrub_parity_work(struct btrfs_work *work)
2631{
2632 struct btrfs_raid_bio *rbio;
2633
2634 rbio = container_of(work, struct btrfs_raid_bio, work);
2635 raid56_parity_scrub_stripe(rbio);
2636}
2637
2638static void async_scrub_parity(struct btrfs_raid_bio *rbio)
2639{
2640 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
2641 scrub_parity_work, NULL, NULL);
2642
2643 btrfs_queue_work(rbio->fs_info->rmw_workers,
2644 &rbio->work);
2645}
2646
2647void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio)
2648{
2649 if (!lock_stripe_add(rbio))
2650 async_scrub_parity(rbio);
2651}
2652
2653/* The following code is used for dev replace of a missing RAID 5/6 device. */
2654
2655struct btrfs_raid_bio *
2656raid56_alloc_missing_rbio(struct btrfs_root *root, struct bio *bio,
2657 struct btrfs_bio *bbio, u64 length)
2658{
2659 struct btrfs_raid_bio *rbio;
2660
2661 rbio = alloc_rbio(root, bbio, length);
2662 if (IS_ERR(rbio))
2663 return NULL;
2664
2665 rbio->operation = BTRFS_RBIO_REBUILD_MISSING;
2666 bio_list_add(&rbio->bio_list, bio);
2667 /*
2668 * This is a special bio which is used to hold the completion handler
2669 * and make the scrub rbio is similar to the other types
2670 */
2671 ASSERT(!bio->bi_iter.bi_size);
2672
2673 rbio->faila = find_logical_bio_stripe(rbio, bio);
2674 if (rbio->faila == -1) {
2675 BUG();
2676 kfree(rbio);
2677 return NULL;
2678 }
2679
2680 return rbio;
2681}
2682
2683static void missing_raid56_work(struct btrfs_work *work)
2684{
2685 struct btrfs_raid_bio *rbio;
2686
2687 rbio = container_of(work, struct btrfs_raid_bio, work);
2688 __raid56_parity_recover(rbio);
2689}
2690
2691static void async_missing_raid56(struct btrfs_raid_bio *rbio)
2692{
2693 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
2694 missing_raid56_work, NULL, NULL);
2695
2696 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
2697}
2698
2699void raid56_submit_missing_rbio(struct btrfs_raid_bio *rbio)
2700{
2701 if (!lock_stripe_add(rbio))
2702 async_missing_raid56(rbio);
2703}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2012 Fusion-io All rights reserved.
4 * Copyright (C) 2012 Intel Corp. All rights reserved.
5 */
6
7#include <linux/sched.h>
8#include <linux/bio.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>
15#include <linux/mm.h>
16#include "messages.h"
17#include "ctree.h"
18#include "disk-io.h"
19#include "volumes.h"
20#include "raid56.h"
21#include "async-thread.h"
22#include "file-item.h"
23#include "btrfs_inode.h"
24
25/* set when additional merges to this rbio are not allowed */
26#define RBIO_RMW_LOCKED_BIT 1
27
28/*
29 * set when this rbio is sitting in the hash, but it is just a cache
30 * of past RMW
31 */
32#define RBIO_CACHE_BIT 2
33
34/*
35 * set when it is safe to trust the stripe_pages for caching
36 */
37#define RBIO_CACHE_READY_BIT 3
38
39#define RBIO_CACHE_SIZE 1024
40
41#define BTRFS_STRIPE_HASH_TABLE_BITS 11
42
43/* Used by the raid56 code to lock stripes for read/modify/write */
44struct btrfs_stripe_hash {
45 struct list_head hash_list;
46 spinlock_t lock;
47};
48
49/* Used by the raid56 code to lock stripes for read/modify/write */
50struct btrfs_stripe_hash_table {
51 struct list_head stripe_cache;
52 spinlock_t cache_lock;
53 int cache_size;
54 struct btrfs_stripe_hash table[];
55};
56
57/*
58 * A bvec like structure to present a sector inside a page.
59 *
60 * Unlike bvec we don't need bvlen, as it's fixed to sectorsize.
61 */
62struct sector_ptr {
63 struct page *page;
64 unsigned int pgoff:24;
65 unsigned int uptodate:8;
66};
67
68static void rmw_rbio_work(struct work_struct *work);
69static void rmw_rbio_work_locked(struct work_struct *work);
70static void index_rbio_pages(struct btrfs_raid_bio *rbio);
71static int alloc_rbio_pages(struct btrfs_raid_bio *rbio);
72
73static int finish_parity_scrub(struct btrfs_raid_bio *rbio);
74static void scrub_rbio_work_locked(struct work_struct *work);
75
76static void free_raid_bio_pointers(struct btrfs_raid_bio *rbio)
77{
78 bitmap_free(rbio->error_bitmap);
79 kfree(rbio->stripe_pages);
80 kfree(rbio->bio_sectors);
81 kfree(rbio->stripe_sectors);
82 kfree(rbio->finish_pointers);
83}
84
85static void free_raid_bio(struct btrfs_raid_bio *rbio)
86{
87 int i;
88
89 if (!refcount_dec_and_test(&rbio->refs))
90 return;
91
92 WARN_ON(!list_empty(&rbio->stripe_cache));
93 WARN_ON(!list_empty(&rbio->hash_list));
94 WARN_ON(!bio_list_empty(&rbio->bio_list));
95
96 for (i = 0; i < rbio->nr_pages; i++) {
97 if (rbio->stripe_pages[i]) {
98 __free_page(rbio->stripe_pages[i]);
99 rbio->stripe_pages[i] = NULL;
100 }
101 }
102
103 btrfs_put_bioc(rbio->bioc);
104 free_raid_bio_pointers(rbio);
105 kfree(rbio);
106}
107
108static void start_async_work(struct btrfs_raid_bio *rbio, work_func_t work_func)
109{
110 INIT_WORK(&rbio->work, work_func);
111 queue_work(rbio->bioc->fs_info->rmw_workers, &rbio->work);
112}
113
114/*
115 * the stripe hash table is used for locking, and to collect
116 * bios in hopes of making a full stripe
117 */
118int btrfs_alloc_stripe_hash_table(struct btrfs_fs_info *info)
119{
120 struct btrfs_stripe_hash_table *table;
121 struct btrfs_stripe_hash_table *x;
122 struct btrfs_stripe_hash *cur;
123 struct btrfs_stripe_hash *h;
124 int num_entries = 1 << BTRFS_STRIPE_HASH_TABLE_BITS;
125 int i;
126
127 if (info->stripe_hash_table)
128 return 0;
129
130 /*
131 * The table is large, starting with order 4 and can go as high as
132 * order 7 in case lock debugging is turned on.
133 *
134 * Try harder to allocate and fallback to vmalloc to lower the chance
135 * of a failing mount.
136 */
137 table = kvzalloc(struct_size(table, table, num_entries), GFP_KERNEL);
138 if (!table)
139 return -ENOMEM;
140
141 spin_lock_init(&table->cache_lock);
142 INIT_LIST_HEAD(&table->stripe_cache);
143
144 h = table->table;
145
146 for (i = 0; i < num_entries; i++) {
147 cur = h + i;
148 INIT_LIST_HEAD(&cur->hash_list);
149 spin_lock_init(&cur->lock);
150 }
151
152 x = cmpxchg(&info->stripe_hash_table, NULL, table);
153 kvfree(x);
154 return 0;
155}
156
157/*
158 * caching an rbio means to copy anything from the
159 * bio_sectors array into the stripe_pages array. We
160 * use the page uptodate bit in the stripe cache array
161 * to indicate if it has valid data
162 *
163 * once the caching is done, we set the cache ready
164 * bit.
165 */
166static void cache_rbio_pages(struct btrfs_raid_bio *rbio)
167{
168 int i;
169 int ret;
170
171 ret = alloc_rbio_pages(rbio);
172 if (ret)
173 return;
174
175 for (i = 0; i < rbio->nr_sectors; i++) {
176 /* Some range not covered by bio (partial write), skip it */
177 if (!rbio->bio_sectors[i].page) {
178 /*
179 * Even if the sector is not covered by bio, if it is
180 * a data sector it should still be uptodate as it is
181 * read from disk.
182 */
183 if (i < rbio->nr_data * rbio->stripe_nsectors)
184 ASSERT(rbio->stripe_sectors[i].uptodate);
185 continue;
186 }
187
188 ASSERT(rbio->stripe_sectors[i].page);
189 memcpy_page(rbio->stripe_sectors[i].page,
190 rbio->stripe_sectors[i].pgoff,
191 rbio->bio_sectors[i].page,
192 rbio->bio_sectors[i].pgoff,
193 rbio->bioc->fs_info->sectorsize);
194 rbio->stripe_sectors[i].uptodate = 1;
195 }
196 set_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
197}
198
199/*
200 * we hash on the first logical address of the stripe
201 */
202static int rbio_bucket(struct btrfs_raid_bio *rbio)
203{
204 u64 num = rbio->bioc->full_stripe_logical;
205
206 /*
207 * we shift down quite a bit. We're using byte
208 * addressing, and most of the lower bits are zeros.
209 * This tends to upset hash_64, and it consistently
210 * returns just one or two different values.
211 *
212 * shifting off the lower bits fixes things.
213 */
214 return hash_64(num >> 16, BTRFS_STRIPE_HASH_TABLE_BITS);
215}
216
217static bool full_page_sectors_uptodate(struct btrfs_raid_bio *rbio,
218 unsigned int page_nr)
219{
220 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
221 const u32 sectors_per_page = PAGE_SIZE / sectorsize;
222 int i;
223
224 ASSERT(page_nr < rbio->nr_pages);
225
226 for (i = sectors_per_page * page_nr;
227 i < sectors_per_page * page_nr + sectors_per_page;
228 i++) {
229 if (!rbio->stripe_sectors[i].uptodate)
230 return false;
231 }
232 return true;
233}
234
235/*
236 * Update the stripe_sectors[] array to use correct page and pgoff
237 *
238 * Should be called every time any page pointer in stripes_pages[] got modified.
239 */
240static void index_stripe_sectors(struct btrfs_raid_bio *rbio)
241{
242 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
243 u32 offset;
244 int i;
245
246 for (i = 0, offset = 0; i < rbio->nr_sectors; i++, offset += sectorsize) {
247 int page_index = offset >> PAGE_SHIFT;
248
249 ASSERT(page_index < rbio->nr_pages);
250 rbio->stripe_sectors[i].page = rbio->stripe_pages[page_index];
251 rbio->stripe_sectors[i].pgoff = offset_in_page(offset);
252 }
253}
254
255static void steal_rbio_page(struct btrfs_raid_bio *src,
256 struct btrfs_raid_bio *dest, int page_nr)
257{
258 const u32 sectorsize = src->bioc->fs_info->sectorsize;
259 const u32 sectors_per_page = PAGE_SIZE / sectorsize;
260 int i;
261
262 if (dest->stripe_pages[page_nr])
263 __free_page(dest->stripe_pages[page_nr]);
264 dest->stripe_pages[page_nr] = src->stripe_pages[page_nr];
265 src->stripe_pages[page_nr] = NULL;
266
267 /* Also update the sector->uptodate bits. */
268 for (i = sectors_per_page * page_nr;
269 i < sectors_per_page * page_nr + sectors_per_page; i++)
270 dest->stripe_sectors[i].uptodate = true;
271}
272
273static bool is_data_stripe_page(struct btrfs_raid_bio *rbio, int page_nr)
274{
275 const int sector_nr = (page_nr << PAGE_SHIFT) >>
276 rbio->bioc->fs_info->sectorsize_bits;
277
278 /*
279 * We have ensured PAGE_SIZE is aligned with sectorsize, thus
280 * we won't have a page which is half data half parity.
281 *
282 * Thus if the first sector of the page belongs to data stripes, then
283 * the full page belongs to data stripes.
284 */
285 return (sector_nr < rbio->nr_data * rbio->stripe_nsectors);
286}
287
288/*
289 * Stealing an rbio means taking all the uptodate pages from the stripe array
290 * in the source rbio and putting them into the destination rbio.
291 *
292 * This will also update the involved stripe_sectors[] which are referring to
293 * the old pages.
294 */
295static void steal_rbio(struct btrfs_raid_bio *src, struct btrfs_raid_bio *dest)
296{
297 int i;
298
299 if (!test_bit(RBIO_CACHE_READY_BIT, &src->flags))
300 return;
301
302 for (i = 0; i < dest->nr_pages; i++) {
303 struct page *p = src->stripe_pages[i];
304
305 /*
306 * We don't need to steal P/Q pages as they will always be
307 * regenerated for RMW or full write anyway.
308 */
309 if (!is_data_stripe_page(src, i))
310 continue;
311
312 /*
313 * If @src already has RBIO_CACHE_READY_BIT, it should have
314 * all data stripe pages present and uptodate.
315 */
316 ASSERT(p);
317 ASSERT(full_page_sectors_uptodate(src, i));
318 steal_rbio_page(src, dest, i);
319 }
320 index_stripe_sectors(dest);
321 index_stripe_sectors(src);
322}
323
324/*
325 * merging means we take the bio_list from the victim and
326 * splice it into the destination. The victim should
327 * be discarded afterwards.
328 *
329 * must be called with dest->rbio_list_lock held
330 */
331static void merge_rbio(struct btrfs_raid_bio *dest,
332 struct btrfs_raid_bio *victim)
333{
334 bio_list_merge(&dest->bio_list, &victim->bio_list);
335 dest->bio_list_bytes += victim->bio_list_bytes;
336 /* Also inherit the bitmaps from @victim. */
337 bitmap_or(&dest->dbitmap, &victim->dbitmap, &dest->dbitmap,
338 dest->stripe_nsectors);
339 bio_list_init(&victim->bio_list);
340}
341
342/*
343 * used to prune items that are in the cache. The caller
344 * must hold the hash table lock.
345 */
346static void __remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
347{
348 int bucket = rbio_bucket(rbio);
349 struct btrfs_stripe_hash_table *table;
350 struct btrfs_stripe_hash *h;
351 int freeit = 0;
352
353 /*
354 * check the bit again under the hash table lock.
355 */
356 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
357 return;
358
359 table = rbio->bioc->fs_info->stripe_hash_table;
360 h = table->table + bucket;
361
362 /* hold the lock for the bucket because we may be
363 * removing it from the hash table
364 */
365 spin_lock(&h->lock);
366
367 /*
368 * hold the lock for the bio list because we need
369 * to make sure the bio list is empty
370 */
371 spin_lock(&rbio->bio_list_lock);
372
373 if (test_and_clear_bit(RBIO_CACHE_BIT, &rbio->flags)) {
374 list_del_init(&rbio->stripe_cache);
375 table->cache_size -= 1;
376 freeit = 1;
377
378 /* if the bio list isn't empty, this rbio is
379 * still involved in an IO. We take it out
380 * of the cache list, and drop the ref that
381 * was held for the list.
382 *
383 * If the bio_list was empty, we also remove
384 * the rbio from the hash_table, and drop
385 * the corresponding ref
386 */
387 if (bio_list_empty(&rbio->bio_list)) {
388 if (!list_empty(&rbio->hash_list)) {
389 list_del_init(&rbio->hash_list);
390 refcount_dec(&rbio->refs);
391 BUG_ON(!list_empty(&rbio->plug_list));
392 }
393 }
394 }
395
396 spin_unlock(&rbio->bio_list_lock);
397 spin_unlock(&h->lock);
398
399 if (freeit)
400 free_raid_bio(rbio);
401}
402
403/*
404 * prune a given rbio from the cache
405 */
406static void remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
407{
408 struct btrfs_stripe_hash_table *table;
409
410 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
411 return;
412
413 table = rbio->bioc->fs_info->stripe_hash_table;
414
415 spin_lock(&table->cache_lock);
416 __remove_rbio_from_cache(rbio);
417 spin_unlock(&table->cache_lock);
418}
419
420/*
421 * remove everything in the cache
422 */
423static void btrfs_clear_rbio_cache(struct btrfs_fs_info *info)
424{
425 struct btrfs_stripe_hash_table *table;
426 struct btrfs_raid_bio *rbio;
427
428 table = info->stripe_hash_table;
429
430 spin_lock(&table->cache_lock);
431 while (!list_empty(&table->stripe_cache)) {
432 rbio = list_entry(table->stripe_cache.next,
433 struct btrfs_raid_bio,
434 stripe_cache);
435 __remove_rbio_from_cache(rbio);
436 }
437 spin_unlock(&table->cache_lock);
438}
439
440/*
441 * remove all cached entries and free the hash table
442 * used by unmount
443 */
444void btrfs_free_stripe_hash_table(struct btrfs_fs_info *info)
445{
446 if (!info->stripe_hash_table)
447 return;
448 btrfs_clear_rbio_cache(info);
449 kvfree(info->stripe_hash_table);
450 info->stripe_hash_table = NULL;
451}
452
453/*
454 * insert an rbio into the stripe cache. It
455 * must have already been prepared by calling
456 * cache_rbio_pages
457 *
458 * If this rbio was already cached, it gets
459 * moved to the front of the lru.
460 *
461 * If the size of the rbio cache is too big, we
462 * prune an item.
463 */
464static void cache_rbio(struct btrfs_raid_bio *rbio)
465{
466 struct btrfs_stripe_hash_table *table;
467
468 if (!test_bit(RBIO_CACHE_READY_BIT, &rbio->flags))
469 return;
470
471 table = rbio->bioc->fs_info->stripe_hash_table;
472
473 spin_lock(&table->cache_lock);
474 spin_lock(&rbio->bio_list_lock);
475
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);
479
480 if (!list_empty(&rbio->stripe_cache)){
481 list_move(&rbio->stripe_cache, &table->stripe_cache);
482 } else {
483 list_add(&rbio->stripe_cache, &table->stripe_cache);
484 table->cache_size += 1;
485 }
486
487 spin_unlock(&rbio->bio_list_lock);
488
489 if (table->cache_size > RBIO_CACHE_SIZE) {
490 struct btrfs_raid_bio *found;
491
492 found = list_entry(table->stripe_cache.prev,
493 struct btrfs_raid_bio,
494 stripe_cache);
495
496 if (found != rbio)
497 __remove_rbio_from_cache(found);
498 }
499
500 spin_unlock(&table->cache_lock);
501}
502
503/*
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
506 * loop through.
507 */
508static void run_xor(void **pages, int src_cnt, ssize_t len)
509{
510 int src_off = 0;
511 int xor_src_cnt = 0;
512 void *dest = pages[src_cnt];
513
514 while(src_cnt > 0) {
515 xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS);
516 xor_blocks(xor_src_cnt, len, dest, pages + src_off);
517
518 src_cnt -= xor_src_cnt;
519 src_off += xor_src_cnt;
520 }
521}
522
523/*
524 * Returns true if the bio list inside this rbio covers an entire stripe (no
525 * rmw required).
526 */
527static int rbio_is_full(struct btrfs_raid_bio *rbio)
528{
529 unsigned long size = rbio->bio_list_bytes;
530 int ret = 1;
531
532 spin_lock(&rbio->bio_list_lock);
533 if (size != rbio->nr_data * BTRFS_STRIPE_LEN)
534 ret = 0;
535 BUG_ON(size > rbio->nr_data * BTRFS_STRIPE_LEN);
536 spin_unlock(&rbio->bio_list_lock);
537
538 return ret;
539}
540
541/*
542 * returns 1 if it is safe to merge two rbios together.
543 * The merging is safe if the two rbios correspond to
544 * the same stripe and if they are both going in the same
545 * direction (read vs write), and if neither one is
546 * locked for final IO
547 *
548 * The caller is responsible for locking such that
549 * rmw_locked is safe to test
550 */
551static int rbio_can_merge(struct btrfs_raid_bio *last,
552 struct btrfs_raid_bio *cur)
553{
554 if (test_bit(RBIO_RMW_LOCKED_BIT, &last->flags) ||
555 test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags))
556 return 0;
557
558 /*
559 * we can't merge with cached rbios, since the
560 * idea is that when we merge the destination
561 * rbio is going to run our IO for us. We can
562 * steal from cached rbios though, other functions
563 * handle that.
564 */
565 if (test_bit(RBIO_CACHE_BIT, &last->flags) ||
566 test_bit(RBIO_CACHE_BIT, &cur->flags))
567 return 0;
568
569 if (last->bioc->full_stripe_logical != cur->bioc->full_stripe_logical)
570 return 0;
571
572 /* we can't merge with different operations */
573 if (last->operation != cur->operation)
574 return 0;
575 /*
576 * We've need read the full stripe from the drive.
577 * check and repair the parity and write the new results.
578 *
579 * We're not allowed to add any new bios to the
580 * bio list here, anyone else that wants to
581 * change this stripe needs to do their own rmw.
582 */
583 if (last->operation == BTRFS_RBIO_PARITY_SCRUB)
584 return 0;
585
586 if (last->operation == BTRFS_RBIO_READ_REBUILD)
587 return 0;
588
589 return 1;
590}
591
592static unsigned int rbio_stripe_sector_index(const struct btrfs_raid_bio *rbio,
593 unsigned int stripe_nr,
594 unsigned int sector_nr)
595{
596 ASSERT(stripe_nr < rbio->real_stripes);
597 ASSERT(sector_nr < rbio->stripe_nsectors);
598
599 return stripe_nr * rbio->stripe_nsectors + sector_nr;
600}
601
602/* Return a sector from rbio->stripe_sectors, not from the bio list */
603static struct sector_ptr *rbio_stripe_sector(const struct btrfs_raid_bio *rbio,
604 unsigned int stripe_nr,
605 unsigned int sector_nr)
606{
607 return &rbio->stripe_sectors[rbio_stripe_sector_index(rbio, stripe_nr,
608 sector_nr)];
609}
610
611/* Grab a sector inside P stripe */
612static struct sector_ptr *rbio_pstripe_sector(const struct btrfs_raid_bio *rbio,
613 unsigned int sector_nr)
614{
615 return rbio_stripe_sector(rbio, rbio->nr_data, sector_nr);
616}
617
618/* Grab a sector inside Q stripe, return NULL if not RAID6 */
619static struct sector_ptr *rbio_qstripe_sector(const struct btrfs_raid_bio *rbio,
620 unsigned int sector_nr)
621{
622 if (rbio->nr_data + 1 == rbio->real_stripes)
623 return NULL;
624 return rbio_stripe_sector(rbio, rbio->nr_data + 1, sector_nr);
625}
626
627/*
628 * The first stripe in the table for a logical address
629 * has the lock. rbios are added in one of three ways:
630 *
631 * 1) Nobody has the stripe locked yet. The rbio is given
632 * the lock and 0 is returned. The caller must start the IO
633 * themselves.
634 *
635 * 2) Someone has the stripe locked, but we're able to merge
636 * with the lock owner. The rbio is freed and the IO will
637 * start automatically along with the existing rbio. 1 is returned.
638 *
639 * 3) Someone has the stripe locked, but we're not able to merge.
640 * The rbio is added to the lock owner's plug list, or merged into
641 * an rbio already on the plug list. When the lock owner unlocks,
642 * the next rbio on the list is run and the IO is started automatically.
643 * 1 is returned
644 *
645 * If we return 0, the caller still owns the rbio and must continue with
646 * IO submission. If we return 1, the caller must assume the rbio has
647 * already been freed.
648 */
649static noinline int lock_stripe_add(struct btrfs_raid_bio *rbio)
650{
651 struct btrfs_stripe_hash *h;
652 struct btrfs_raid_bio *cur;
653 struct btrfs_raid_bio *pending;
654 struct btrfs_raid_bio *freeit = NULL;
655 struct btrfs_raid_bio *cache_drop = NULL;
656 int ret = 0;
657
658 h = rbio->bioc->fs_info->stripe_hash_table->table + rbio_bucket(rbio);
659
660 spin_lock(&h->lock);
661 list_for_each_entry(cur, &h->hash_list, hash_list) {
662 if (cur->bioc->full_stripe_logical != rbio->bioc->full_stripe_logical)
663 continue;
664
665 spin_lock(&cur->bio_list_lock);
666
667 /* Can we steal this cached rbio's pages? */
668 if (bio_list_empty(&cur->bio_list) &&
669 list_empty(&cur->plug_list) &&
670 test_bit(RBIO_CACHE_BIT, &cur->flags) &&
671 !test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) {
672 list_del_init(&cur->hash_list);
673 refcount_dec(&cur->refs);
674
675 steal_rbio(cur, rbio);
676 cache_drop = cur;
677 spin_unlock(&cur->bio_list_lock);
678
679 goto lockit;
680 }
681
682 /* Can we merge into the lock owner? */
683 if (rbio_can_merge(cur, rbio)) {
684 merge_rbio(cur, rbio);
685 spin_unlock(&cur->bio_list_lock);
686 freeit = rbio;
687 ret = 1;
688 goto out;
689 }
690
691
692 /*
693 * We couldn't merge with the running rbio, see if we can merge
694 * with the pending ones. We don't have to check for rmw_locked
695 * because there is no way they are inside finish_rmw right now
696 */
697 list_for_each_entry(pending, &cur->plug_list, plug_list) {
698 if (rbio_can_merge(pending, rbio)) {
699 merge_rbio(pending, rbio);
700 spin_unlock(&cur->bio_list_lock);
701 freeit = rbio;
702 ret = 1;
703 goto out;
704 }
705 }
706
707 /*
708 * No merging, put us on the tail of the plug list, our rbio
709 * will be started with the currently running rbio unlocks
710 */
711 list_add_tail(&rbio->plug_list, &cur->plug_list);
712 spin_unlock(&cur->bio_list_lock);
713 ret = 1;
714 goto out;
715 }
716lockit:
717 refcount_inc(&rbio->refs);
718 list_add(&rbio->hash_list, &h->hash_list);
719out:
720 spin_unlock(&h->lock);
721 if (cache_drop)
722 remove_rbio_from_cache(cache_drop);
723 if (freeit)
724 free_raid_bio(freeit);
725 return ret;
726}
727
728static void recover_rbio_work_locked(struct work_struct *work);
729
730/*
731 * called as rmw or parity rebuild is completed. If the plug list has more
732 * rbios waiting for this stripe, the next one on the list will be started
733 */
734static noinline void unlock_stripe(struct btrfs_raid_bio *rbio)
735{
736 int bucket;
737 struct btrfs_stripe_hash *h;
738 int keep_cache = 0;
739
740 bucket = rbio_bucket(rbio);
741 h = rbio->bioc->fs_info->stripe_hash_table->table + bucket;
742
743 if (list_empty(&rbio->plug_list))
744 cache_rbio(rbio);
745
746 spin_lock(&h->lock);
747 spin_lock(&rbio->bio_list_lock);
748
749 if (!list_empty(&rbio->hash_list)) {
750 /*
751 * if we're still cached and there is no other IO
752 * to perform, just leave this rbio here for others
753 * to steal from later
754 */
755 if (list_empty(&rbio->plug_list) &&
756 test_bit(RBIO_CACHE_BIT, &rbio->flags)) {
757 keep_cache = 1;
758 clear_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
759 BUG_ON(!bio_list_empty(&rbio->bio_list));
760 goto done;
761 }
762
763 list_del_init(&rbio->hash_list);
764 refcount_dec(&rbio->refs);
765
766 /*
767 * we use the plug list to hold all the rbios
768 * waiting for the chance to lock this stripe.
769 * hand the lock over to one of them.
770 */
771 if (!list_empty(&rbio->plug_list)) {
772 struct btrfs_raid_bio *next;
773 struct list_head *head = rbio->plug_list.next;
774
775 next = list_entry(head, struct btrfs_raid_bio,
776 plug_list);
777
778 list_del_init(&rbio->plug_list);
779
780 list_add(&next->hash_list, &h->hash_list);
781 refcount_inc(&next->refs);
782 spin_unlock(&rbio->bio_list_lock);
783 spin_unlock(&h->lock);
784
785 if (next->operation == BTRFS_RBIO_READ_REBUILD) {
786 start_async_work(next, recover_rbio_work_locked);
787 } else if (next->operation == BTRFS_RBIO_WRITE) {
788 steal_rbio(rbio, next);
789 start_async_work(next, rmw_rbio_work_locked);
790 } else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) {
791 steal_rbio(rbio, next);
792 start_async_work(next, scrub_rbio_work_locked);
793 }
794
795 goto done_nolock;
796 }
797 }
798done:
799 spin_unlock(&rbio->bio_list_lock);
800 spin_unlock(&h->lock);
801
802done_nolock:
803 if (!keep_cache)
804 remove_rbio_from_cache(rbio);
805}
806
807static void rbio_endio_bio_list(struct bio *cur, blk_status_t err)
808{
809 struct bio *next;
810
811 while (cur) {
812 next = cur->bi_next;
813 cur->bi_next = NULL;
814 cur->bi_status = err;
815 bio_endio(cur);
816 cur = next;
817 }
818}
819
820/*
821 * this frees the rbio and runs through all the bios in the
822 * bio_list and calls end_io on them
823 */
824static void rbio_orig_end_io(struct btrfs_raid_bio *rbio, blk_status_t err)
825{
826 struct bio *cur = bio_list_get(&rbio->bio_list);
827 struct bio *extra;
828
829 kfree(rbio->csum_buf);
830 bitmap_free(rbio->csum_bitmap);
831 rbio->csum_buf = NULL;
832 rbio->csum_bitmap = NULL;
833
834 /*
835 * Clear the data bitmap, as the rbio may be cached for later usage.
836 * do this before before unlock_stripe() so there will be no new bio
837 * for this bio.
838 */
839 bitmap_clear(&rbio->dbitmap, 0, rbio->stripe_nsectors);
840
841 /*
842 * At this moment, rbio->bio_list is empty, however since rbio does not
843 * always have RBIO_RMW_LOCKED_BIT set and rbio is still linked on the
844 * hash list, rbio may be merged with others so that rbio->bio_list
845 * becomes non-empty.
846 * Once unlock_stripe() is done, rbio->bio_list will not be updated any
847 * more and we can call bio_endio() on all queued bios.
848 */
849 unlock_stripe(rbio);
850 extra = bio_list_get(&rbio->bio_list);
851 free_raid_bio(rbio);
852
853 rbio_endio_bio_list(cur, err);
854 if (extra)
855 rbio_endio_bio_list(extra, err);
856}
857
858/*
859 * Get a sector pointer specified by its @stripe_nr and @sector_nr.
860 *
861 * @rbio: The raid bio
862 * @stripe_nr: Stripe number, valid range [0, real_stripe)
863 * @sector_nr: Sector number inside the stripe,
864 * valid range [0, stripe_nsectors)
865 * @bio_list_only: Whether to use sectors inside the bio list only.
866 *
867 * The read/modify/write code wants to reuse the original bio page as much
868 * as possible, and only use stripe_sectors as fallback.
869 */
870static struct sector_ptr *sector_in_rbio(struct btrfs_raid_bio *rbio,
871 int stripe_nr, int sector_nr,
872 bool bio_list_only)
873{
874 struct sector_ptr *sector;
875 int index;
876
877 ASSERT(stripe_nr >= 0 && stripe_nr < rbio->real_stripes);
878 ASSERT(sector_nr >= 0 && sector_nr < rbio->stripe_nsectors);
879
880 index = stripe_nr * rbio->stripe_nsectors + sector_nr;
881 ASSERT(index >= 0 && index < rbio->nr_sectors);
882
883 spin_lock(&rbio->bio_list_lock);
884 sector = &rbio->bio_sectors[index];
885 if (sector->page || bio_list_only) {
886 /* Don't return sector without a valid page pointer */
887 if (!sector->page)
888 sector = NULL;
889 spin_unlock(&rbio->bio_list_lock);
890 return sector;
891 }
892 spin_unlock(&rbio->bio_list_lock);
893
894 return &rbio->stripe_sectors[index];
895}
896
897/*
898 * allocation and initial setup for the btrfs_raid_bio. Not
899 * this does not allocate any pages for rbio->pages.
900 */
901static struct btrfs_raid_bio *alloc_rbio(struct btrfs_fs_info *fs_info,
902 struct btrfs_io_context *bioc)
903{
904 const unsigned int real_stripes = bioc->num_stripes - bioc->replace_nr_stripes;
905 const unsigned int stripe_npages = BTRFS_STRIPE_LEN >> PAGE_SHIFT;
906 const unsigned int num_pages = stripe_npages * real_stripes;
907 const unsigned int stripe_nsectors =
908 BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits;
909 const unsigned int num_sectors = stripe_nsectors * real_stripes;
910 struct btrfs_raid_bio *rbio;
911
912 /* PAGE_SIZE must also be aligned to sectorsize for subpage support */
913 ASSERT(IS_ALIGNED(PAGE_SIZE, fs_info->sectorsize));
914 /*
915 * Our current stripe len should be fixed to 64k thus stripe_nsectors
916 * (at most 16) should be no larger than BITS_PER_LONG.
917 */
918 ASSERT(stripe_nsectors <= BITS_PER_LONG);
919
920 /*
921 * Real stripes must be between 2 (2 disks RAID5, aka RAID1) and 256
922 * (limited by u8).
923 */
924 ASSERT(real_stripes >= 2);
925 ASSERT(real_stripes <= U8_MAX);
926
927 rbio = kzalloc(sizeof(*rbio), GFP_NOFS);
928 if (!rbio)
929 return ERR_PTR(-ENOMEM);
930 rbio->stripe_pages = kcalloc(num_pages, sizeof(struct page *),
931 GFP_NOFS);
932 rbio->bio_sectors = kcalloc(num_sectors, sizeof(struct sector_ptr),
933 GFP_NOFS);
934 rbio->stripe_sectors = kcalloc(num_sectors, sizeof(struct sector_ptr),
935 GFP_NOFS);
936 rbio->finish_pointers = kcalloc(real_stripes, sizeof(void *), GFP_NOFS);
937 rbio->error_bitmap = bitmap_zalloc(num_sectors, GFP_NOFS);
938
939 if (!rbio->stripe_pages || !rbio->bio_sectors || !rbio->stripe_sectors ||
940 !rbio->finish_pointers || !rbio->error_bitmap) {
941 free_raid_bio_pointers(rbio);
942 kfree(rbio);
943 return ERR_PTR(-ENOMEM);
944 }
945
946 bio_list_init(&rbio->bio_list);
947 init_waitqueue_head(&rbio->io_wait);
948 INIT_LIST_HEAD(&rbio->plug_list);
949 spin_lock_init(&rbio->bio_list_lock);
950 INIT_LIST_HEAD(&rbio->stripe_cache);
951 INIT_LIST_HEAD(&rbio->hash_list);
952 btrfs_get_bioc(bioc);
953 rbio->bioc = bioc;
954 rbio->nr_pages = num_pages;
955 rbio->nr_sectors = num_sectors;
956 rbio->real_stripes = real_stripes;
957 rbio->stripe_npages = stripe_npages;
958 rbio->stripe_nsectors = stripe_nsectors;
959 refcount_set(&rbio->refs, 1);
960 atomic_set(&rbio->stripes_pending, 0);
961
962 ASSERT(btrfs_nr_parity_stripes(bioc->map_type));
963 rbio->nr_data = real_stripes - btrfs_nr_parity_stripes(bioc->map_type);
964 ASSERT(rbio->nr_data > 0);
965
966 return rbio;
967}
968
969/* allocate pages for all the stripes in the bio, including parity */
970static int alloc_rbio_pages(struct btrfs_raid_bio *rbio)
971{
972 int ret;
973
974 ret = btrfs_alloc_page_array(rbio->nr_pages, rbio->stripe_pages, 0);
975 if (ret < 0)
976 return ret;
977 /* Mapping all sectors */
978 index_stripe_sectors(rbio);
979 return 0;
980}
981
982/* only allocate pages for p/q stripes */
983static int alloc_rbio_parity_pages(struct btrfs_raid_bio *rbio)
984{
985 const int data_pages = rbio->nr_data * rbio->stripe_npages;
986 int ret;
987
988 ret = btrfs_alloc_page_array(rbio->nr_pages - data_pages,
989 rbio->stripe_pages + data_pages, 0);
990 if (ret < 0)
991 return ret;
992
993 index_stripe_sectors(rbio);
994 return 0;
995}
996
997/*
998 * Return the total number of errors found in the vertical stripe of @sector_nr.
999 *
1000 * @faila and @failb will also be updated to the first and second stripe
1001 * number of the errors.
1002 */
1003static int get_rbio_veritical_errors(struct btrfs_raid_bio *rbio, int sector_nr,
1004 int *faila, int *failb)
1005{
1006 int stripe_nr;
1007 int found_errors = 0;
1008
1009 if (faila || failb) {
1010 /*
1011 * Both @faila and @failb should be valid pointers if any of
1012 * them is specified.
1013 */
1014 ASSERT(faila && failb);
1015 *faila = -1;
1016 *failb = -1;
1017 }
1018
1019 for (stripe_nr = 0; stripe_nr < rbio->real_stripes; stripe_nr++) {
1020 int total_sector_nr = stripe_nr * rbio->stripe_nsectors + sector_nr;
1021
1022 if (test_bit(total_sector_nr, rbio->error_bitmap)) {
1023 found_errors++;
1024 if (faila) {
1025 /* Update faila and failb. */
1026 if (*faila < 0)
1027 *faila = stripe_nr;
1028 else if (*failb < 0)
1029 *failb = stripe_nr;
1030 }
1031 }
1032 }
1033 return found_errors;
1034}
1035
1036/*
1037 * Add a single sector @sector into our list of bios for IO.
1038 *
1039 * Return 0 if everything went well.
1040 * Return <0 for error.
1041 */
1042static int rbio_add_io_sector(struct btrfs_raid_bio *rbio,
1043 struct bio_list *bio_list,
1044 struct sector_ptr *sector,
1045 unsigned int stripe_nr,
1046 unsigned int sector_nr,
1047 enum req_op op)
1048{
1049 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
1050 struct bio *last = bio_list->tail;
1051 int ret;
1052 struct bio *bio;
1053 struct btrfs_io_stripe *stripe;
1054 u64 disk_start;
1055
1056 /*
1057 * Note: here stripe_nr has taken device replace into consideration,
1058 * thus it can be larger than rbio->real_stripe.
1059 * So here we check against bioc->num_stripes, not rbio->real_stripes.
1060 */
1061 ASSERT(stripe_nr >= 0 && stripe_nr < rbio->bioc->num_stripes);
1062 ASSERT(sector_nr >= 0 && sector_nr < rbio->stripe_nsectors);
1063 ASSERT(sector->page);
1064
1065 stripe = &rbio->bioc->stripes[stripe_nr];
1066 disk_start = stripe->physical + sector_nr * sectorsize;
1067
1068 /* if the device is missing, just fail this stripe */
1069 if (!stripe->dev->bdev) {
1070 int found_errors;
1071
1072 set_bit(stripe_nr * rbio->stripe_nsectors + sector_nr,
1073 rbio->error_bitmap);
1074
1075 /* Check if we have reached tolerance early. */
1076 found_errors = get_rbio_veritical_errors(rbio, sector_nr,
1077 NULL, NULL);
1078 if (found_errors > rbio->bioc->max_errors)
1079 return -EIO;
1080 return 0;
1081 }
1082
1083 /* see if we can add this page onto our existing bio */
1084 if (last) {
1085 u64 last_end = last->bi_iter.bi_sector << SECTOR_SHIFT;
1086 last_end += last->bi_iter.bi_size;
1087
1088 /*
1089 * we can't merge these if they are from different
1090 * devices or if they are not contiguous
1091 */
1092 if (last_end == disk_start && !last->bi_status &&
1093 last->bi_bdev == stripe->dev->bdev) {
1094 ret = bio_add_page(last, sector->page, sectorsize,
1095 sector->pgoff);
1096 if (ret == sectorsize)
1097 return 0;
1098 }
1099 }
1100
1101 /* put a new bio on the list */
1102 bio = bio_alloc(stripe->dev->bdev,
1103 max(BTRFS_STRIPE_LEN >> PAGE_SHIFT, 1),
1104 op, GFP_NOFS);
1105 bio->bi_iter.bi_sector = disk_start >> SECTOR_SHIFT;
1106 bio->bi_private = rbio;
1107
1108 __bio_add_page(bio, sector->page, sectorsize, sector->pgoff);
1109 bio_list_add(bio_list, bio);
1110 return 0;
1111}
1112
1113static void index_one_bio(struct btrfs_raid_bio *rbio, struct bio *bio)
1114{
1115 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
1116 struct bio_vec bvec;
1117 struct bvec_iter iter;
1118 u32 offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) -
1119 rbio->bioc->full_stripe_logical;
1120
1121 bio_for_each_segment(bvec, bio, iter) {
1122 u32 bvec_offset;
1123
1124 for (bvec_offset = 0; bvec_offset < bvec.bv_len;
1125 bvec_offset += sectorsize, offset += sectorsize) {
1126 int index = offset / sectorsize;
1127 struct sector_ptr *sector = &rbio->bio_sectors[index];
1128
1129 sector->page = bvec.bv_page;
1130 sector->pgoff = bvec.bv_offset + bvec_offset;
1131 ASSERT(sector->pgoff < PAGE_SIZE);
1132 }
1133 }
1134}
1135
1136/*
1137 * helper function to walk our bio list and populate the bio_pages array with
1138 * the result. This seems expensive, but it is faster than constantly
1139 * searching through the bio list as we setup the IO in finish_rmw or stripe
1140 * reconstruction.
1141 *
1142 * This must be called before you trust the answers from page_in_rbio
1143 */
1144static void index_rbio_pages(struct btrfs_raid_bio *rbio)
1145{
1146 struct bio *bio;
1147
1148 spin_lock(&rbio->bio_list_lock);
1149 bio_list_for_each(bio, &rbio->bio_list)
1150 index_one_bio(rbio, bio);
1151
1152 spin_unlock(&rbio->bio_list_lock);
1153}
1154
1155static void bio_get_trace_info(struct btrfs_raid_bio *rbio, struct bio *bio,
1156 struct raid56_bio_trace_info *trace_info)
1157{
1158 const struct btrfs_io_context *bioc = rbio->bioc;
1159 int i;
1160
1161 ASSERT(bioc);
1162
1163 /* We rely on bio->bi_bdev to find the stripe number. */
1164 if (!bio->bi_bdev)
1165 goto not_found;
1166
1167 for (i = 0; i < bioc->num_stripes; i++) {
1168 if (bio->bi_bdev != bioc->stripes[i].dev->bdev)
1169 continue;
1170 trace_info->stripe_nr = i;
1171 trace_info->devid = bioc->stripes[i].dev->devid;
1172 trace_info->offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) -
1173 bioc->stripes[i].physical;
1174 return;
1175 }
1176
1177not_found:
1178 trace_info->devid = -1;
1179 trace_info->offset = -1;
1180 trace_info->stripe_nr = -1;
1181}
1182
1183static inline void bio_list_put(struct bio_list *bio_list)
1184{
1185 struct bio *bio;
1186
1187 while ((bio = bio_list_pop(bio_list)))
1188 bio_put(bio);
1189}
1190
1191static void assert_rbio(struct btrfs_raid_bio *rbio)
1192{
1193 if (!IS_ENABLED(CONFIG_BTRFS_DEBUG) ||
1194 !IS_ENABLED(CONFIG_BTRFS_ASSERT))
1195 return;
1196
1197 /*
1198 * At least two stripes (2 disks RAID5), and since real_stripes is U8,
1199 * we won't go beyond 256 disks anyway.
1200 */
1201 ASSERT(rbio->real_stripes >= 2);
1202 ASSERT(rbio->nr_data > 0);
1203
1204 /*
1205 * This is another check to make sure nr data stripes is smaller
1206 * than total stripes.
1207 */
1208 ASSERT(rbio->nr_data < rbio->real_stripes);
1209}
1210
1211/* Generate PQ for one vertical stripe. */
1212static void generate_pq_vertical(struct btrfs_raid_bio *rbio, int sectornr)
1213{
1214 void **pointers = rbio->finish_pointers;
1215 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
1216 struct sector_ptr *sector;
1217 int stripe;
1218 const bool has_qstripe = rbio->bioc->map_type & BTRFS_BLOCK_GROUP_RAID6;
1219
1220 /* First collect one sector from each data stripe */
1221 for (stripe = 0; stripe < rbio->nr_data; stripe++) {
1222 sector = sector_in_rbio(rbio, stripe, sectornr, 0);
1223 pointers[stripe] = kmap_local_page(sector->page) +
1224 sector->pgoff;
1225 }
1226
1227 /* Then add the parity stripe */
1228 sector = rbio_pstripe_sector(rbio, sectornr);
1229 sector->uptodate = 1;
1230 pointers[stripe++] = kmap_local_page(sector->page) + sector->pgoff;
1231
1232 if (has_qstripe) {
1233 /*
1234 * RAID6, add the qstripe and call the library function
1235 * to fill in our p/q
1236 */
1237 sector = rbio_qstripe_sector(rbio, sectornr);
1238 sector->uptodate = 1;
1239 pointers[stripe++] = kmap_local_page(sector->page) +
1240 sector->pgoff;
1241
1242 assert_rbio(rbio);
1243 raid6_call.gen_syndrome(rbio->real_stripes, sectorsize,
1244 pointers);
1245 } else {
1246 /* raid5 */
1247 memcpy(pointers[rbio->nr_data], pointers[0], sectorsize);
1248 run_xor(pointers + 1, rbio->nr_data - 1, sectorsize);
1249 }
1250 for (stripe = stripe - 1; stripe >= 0; stripe--)
1251 kunmap_local(pointers[stripe]);
1252}
1253
1254static int rmw_assemble_write_bios(struct btrfs_raid_bio *rbio,
1255 struct bio_list *bio_list)
1256{
1257 /* The total sector number inside the full stripe. */
1258 int total_sector_nr;
1259 int sectornr;
1260 int stripe;
1261 int ret;
1262
1263 ASSERT(bio_list_size(bio_list) == 0);
1264
1265 /* We should have at least one data sector. */
1266 ASSERT(bitmap_weight(&rbio->dbitmap, rbio->stripe_nsectors));
1267
1268 /*
1269 * Reset errors, as we may have errors inherited from from degraded
1270 * write.
1271 */
1272 bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors);
1273
1274 /*
1275 * Start assembly. Make bios for everything from the higher layers (the
1276 * bio_list in our rbio) and our P/Q. Ignore everything else.
1277 */
1278 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
1279 total_sector_nr++) {
1280 struct sector_ptr *sector;
1281
1282 stripe = total_sector_nr / rbio->stripe_nsectors;
1283 sectornr = total_sector_nr % rbio->stripe_nsectors;
1284
1285 /* This vertical stripe has no data, skip it. */
1286 if (!test_bit(sectornr, &rbio->dbitmap))
1287 continue;
1288
1289 if (stripe < rbio->nr_data) {
1290 sector = sector_in_rbio(rbio, stripe, sectornr, 1);
1291 if (!sector)
1292 continue;
1293 } else {
1294 sector = rbio_stripe_sector(rbio, stripe, sectornr);
1295 }
1296
1297 ret = rbio_add_io_sector(rbio, bio_list, sector, stripe,
1298 sectornr, REQ_OP_WRITE);
1299 if (ret)
1300 goto error;
1301 }
1302
1303 if (likely(!rbio->bioc->replace_nr_stripes))
1304 return 0;
1305
1306 /*
1307 * Make a copy for the replace target device.
1308 *
1309 * Thus the source stripe number (in replace_stripe_src) should be valid.
1310 */
1311 ASSERT(rbio->bioc->replace_stripe_src >= 0);
1312
1313 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
1314 total_sector_nr++) {
1315 struct sector_ptr *sector;
1316
1317 stripe = total_sector_nr / rbio->stripe_nsectors;
1318 sectornr = total_sector_nr % rbio->stripe_nsectors;
1319
1320 /*
1321 * For RAID56, there is only one device that can be replaced,
1322 * and replace_stripe_src[0] indicates the stripe number we
1323 * need to copy from.
1324 */
1325 if (stripe != rbio->bioc->replace_stripe_src) {
1326 /*
1327 * We can skip the whole stripe completely, note
1328 * total_sector_nr will be increased by one anyway.
1329 */
1330 ASSERT(sectornr == 0);
1331 total_sector_nr += rbio->stripe_nsectors - 1;
1332 continue;
1333 }
1334
1335 /* This vertical stripe has no data, skip it. */
1336 if (!test_bit(sectornr, &rbio->dbitmap))
1337 continue;
1338
1339 if (stripe < rbio->nr_data) {
1340 sector = sector_in_rbio(rbio, stripe, sectornr, 1);
1341 if (!sector)
1342 continue;
1343 } else {
1344 sector = rbio_stripe_sector(rbio, stripe, sectornr);
1345 }
1346
1347 ret = rbio_add_io_sector(rbio, bio_list, sector,
1348 rbio->real_stripes,
1349 sectornr, REQ_OP_WRITE);
1350 if (ret)
1351 goto error;
1352 }
1353
1354 return 0;
1355error:
1356 bio_list_put(bio_list);
1357 return -EIO;
1358}
1359
1360static void set_rbio_range_error(struct btrfs_raid_bio *rbio, struct bio *bio)
1361{
1362 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1363 u32 offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) -
1364 rbio->bioc->full_stripe_logical;
1365 int total_nr_sector = offset >> fs_info->sectorsize_bits;
1366
1367 ASSERT(total_nr_sector < rbio->nr_data * rbio->stripe_nsectors);
1368
1369 bitmap_set(rbio->error_bitmap, total_nr_sector,
1370 bio->bi_iter.bi_size >> fs_info->sectorsize_bits);
1371
1372 /*
1373 * Special handling for raid56_alloc_missing_rbio() used by
1374 * scrub/replace. Unlike call path in raid56_parity_recover(), they
1375 * pass an empty bio here. Thus we have to find out the missing device
1376 * and mark the stripe error instead.
1377 */
1378 if (bio->bi_iter.bi_size == 0) {
1379 bool found_missing = false;
1380 int stripe_nr;
1381
1382 for (stripe_nr = 0; stripe_nr < rbio->real_stripes; stripe_nr++) {
1383 if (!rbio->bioc->stripes[stripe_nr].dev->bdev) {
1384 found_missing = true;
1385 bitmap_set(rbio->error_bitmap,
1386 stripe_nr * rbio->stripe_nsectors,
1387 rbio->stripe_nsectors);
1388 }
1389 }
1390 ASSERT(found_missing);
1391 }
1392}
1393
1394/*
1395 * For subpage case, we can no longer set page Up-to-date directly for
1396 * stripe_pages[], thus we need to locate the sector.
1397 */
1398static struct sector_ptr *find_stripe_sector(struct btrfs_raid_bio *rbio,
1399 struct page *page,
1400 unsigned int pgoff)
1401{
1402 int i;
1403
1404 for (i = 0; i < rbio->nr_sectors; i++) {
1405 struct sector_ptr *sector = &rbio->stripe_sectors[i];
1406
1407 if (sector->page == page && sector->pgoff == pgoff)
1408 return sector;
1409 }
1410 return NULL;
1411}
1412
1413/*
1414 * this sets each page in the bio uptodate. It should only be used on private
1415 * rbio pages, nothing that comes in from the higher layers
1416 */
1417static void set_bio_pages_uptodate(struct btrfs_raid_bio *rbio, struct bio *bio)
1418{
1419 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
1420 struct bio_vec *bvec;
1421 struct bvec_iter_all iter_all;
1422
1423 ASSERT(!bio_flagged(bio, BIO_CLONED));
1424
1425 bio_for_each_segment_all(bvec, bio, iter_all) {
1426 struct sector_ptr *sector;
1427 int pgoff;
1428
1429 for (pgoff = bvec->bv_offset; pgoff - bvec->bv_offset < bvec->bv_len;
1430 pgoff += sectorsize) {
1431 sector = find_stripe_sector(rbio, bvec->bv_page, pgoff);
1432 ASSERT(sector);
1433 if (sector)
1434 sector->uptodate = 1;
1435 }
1436 }
1437}
1438
1439static int get_bio_sector_nr(struct btrfs_raid_bio *rbio, struct bio *bio)
1440{
1441 struct bio_vec *bv = bio_first_bvec_all(bio);
1442 int i;
1443
1444 for (i = 0; i < rbio->nr_sectors; i++) {
1445 struct sector_ptr *sector;
1446
1447 sector = &rbio->stripe_sectors[i];
1448 if (sector->page == bv->bv_page && sector->pgoff == bv->bv_offset)
1449 break;
1450 sector = &rbio->bio_sectors[i];
1451 if (sector->page == bv->bv_page && sector->pgoff == bv->bv_offset)
1452 break;
1453 }
1454 ASSERT(i < rbio->nr_sectors);
1455 return i;
1456}
1457
1458static void rbio_update_error_bitmap(struct btrfs_raid_bio *rbio, struct bio *bio)
1459{
1460 int total_sector_nr = get_bio_sector_nr(rbio, bio);
1461 u32 bio_size = 0;
1462 struct bio_vec *bvec;
1463 int i;
1464
1465 bio_for_each_bvec_all(bvec, bio, i)
1466 bio_size += bvec->bv_len;
1467
1468 /*
1469 * Since we can have multiple bios touching the error_bitmap, we cannot
1470 * call bitmap_set() without protection.
1471 *
1472 * Instead use set_bit() for each bit, as set_bit() itself is atomic.
1473 */
1474 for (i = total_sector_nr; i < total_sector_nr +
1475 (bio_size >> rbio->bioc->fs_info->sectorsize_bits); i++)
1476 set_bit(i, rbio->error_bitmap);
1477}
1478
1479/* Verify the data sectors at read time. */
1480static void verify_bio_data_sectors(struct btrfs_raid_bio *rbio,
1481 struct bio *bio)
1482{
1483 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1484 int total_sector_nr = get_bio_sector_nr(rbio, bio);
1485 struct bio_vec *bvec;
1486 struct bvec_iter_all iter_all;
1487
1488 /* No data csum for the whole stripe, no need to verify. */
1489 if (!rbio->csum_bitmap || !rbio->csum_buf)
1490 return;
1491
1492 /* P/Q stripes, they have no data csum to verify against. */
1493 if (total_sector_nr >= rbio->nr_data * rbio->stripe_nsectors)
1494 return;
1495
1496 bio_for_each_segment_all(bvec, bio, iter_all) {
1497 int bv_offset;
1498
1499 for (bv_offset = bvec->bv_offset;
1500 bv_offset < bvec->bv_offset + bvec->bv_len;
1501 bv_offset += fs_info->sectorsize, total_sector_nr++) {
1502 u8 csum_buf[BTRFS_CSUM_SIZE];
1503 u8 *expected_csum = rbio->csum_buf +
1504 total_sector_nr * fs_info->csum_size;
1505 int ret;
1506
1507 /* No csum for this sector, skip to the next sector. */
1508 if (!test_bit(total_sector_nr, rbio->csum_bitmap))
1509 continue;
1510
1511 ret = btrfs_check_sector_csum(fs_info, bvec->bv_page,
1512 bv_offset, csum_buf, expected_csum);
1513 if (ret < 0)
1514 set_bit(total_sector_nr, rbio->error_bitmap);
1515 }
1516 }
1517}
1518
1519static void raid_wait_read_end_io(struct bio *bio)
1520{
1521 struct btrfs_raid_bio *rbio = bio->bi_private;
1522
1523 if (bio->bi_status) {
1524 rbio_update_error_bitmap(rbio, bio);
1525 } else {
1526 set_bio_pages_uptodate(rbio, bio);
1527 verify_bio_data_sectors(rbio, bio);
1528 }
1529
1530 bio_put(bio);
1531 if (atomic_dec_and_test(&rbio->stripes_pending))
1532 wake_up(&rbio->io_wait);
1533}
1534
1535static void submit_read_wait_bio_list(struct btrfs_raid_bio *rbio,
1536 struct bio_list *bio_list)
1537{
1538 struct bio *bio;
1539
1540 atomic_set(&rbio->stripes_pending, bio_list_size(bio_list));
1541 while ((bio = bio_list_pop(bio_list))) {
1542 bio->bi_end_io = raid_wait_read_end_io;
1543
1544 if (trace_raid56_read_enabled()) {
1545 struct raid56_bio_trace_info trace_info = { 0 };
1546
1547 bio_get_trace_info(rbio, bio, &trace_info);
1548 trace_raid56_read(rbio, bio, &trace_info);
1549 }
1550 submit_bio(bio);
1551 }
1552
1553 wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0);
1554}
1555
1556static int alloc_rbio_data_pages(struct btrfs_raid_bio *rbio)
1557{
1558 const int data_pages = rbio->nr_data * rbio->stripe_npages;
1559 int ret;
1560
1561 ret = btrfs_alloc_page_array(data_pages, rbio->stripe_pages, 0);
1562 if (ret < 0)
1563 return ret;
1564
1565 index_stripe_sectors(rbio);
1566 return 0;
1567}
1568
1569/*
1570 * We use plugging call backs to collect full stripes.
1571 * Any time we get a partial stripe write while plugged
1572 * we collect it into a list. When the unplug comes down,
1573 * we sort the list by logical block number and merge
1574 * everything we can into the same rbios
1575 */
1576struct btrfs_plug_cb {
1577 struct blk_plug_cb cb;
1578 struct btrfs_fs_info *info;
1579 struct list_head rbio_list;
1580};
1581
1582/*
1583 * rbios on the plug list are sorted for easier merging.
1584 */
1585static int plug_cmp(void *priv, const struct list_head *a,
1586 const struct list_head *b)
1587{
1588 const struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio,
1589 plug_list);
1590 const struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio,
1591 plug_list);
1592 u64 a_sector = ra->bio_list.head->bi_iter.bi_sector;
1593 u64 b_sector = rb->bio_list.head->bi_iter.bi_sector;
1594
1595 if (a_sector < b_sector)
1596 return -1;
1597 if (a_sector > b_sector)
1598 return 1;
1599 return 0;
1600}
1601
1602static void raid_unplug(struct blk_plug_cb *cb, bool from_schedule)
1603{
1604 struct btrfs_plug_cb *plug = container_of(cb, struct btrfs_plug_cb, cb);
1605 struct btrfs_raid_bio *cur;
1606 struct btrfs_raid_bio *last = NULL;
1607
1608 list_sort(NULL, &plug->rbio_list, plug_cmp);
1609
1610 while (!list_empty(&plug->rbio_list)) {
1611 cur = list_entry(plug->rbio_list.next,
1612 struct btrfs_raid_bio, plug_list);
1613 list_del_init(&cur->plug_list);
1614
1615 if (rbio_is_full(cur)) {
1616 /* We have a full stripe, queue it down. */
1617 start_async_work(cur, rmw_rbio_work);
1618 continue;
1619 }
1620 if (last) {
1621 if (rbio_can_merge(last, cur)) {
1622 merge_rbio(last, cur);
1623 free_raid_bio(cur);
1624 continue;
1625 }
1626 start_async_work(last, rmw_rbio_work);
1627 }
1628 last = cur;
1629 }
1630 if (last)
1631 start_async_work(last, rmw_rbio_work);
1632 kfree(plug);
1633}
1634
1635/* Add the original bio into rbio->bio_list, and update rbio::dbitmap. */
1636static void rbio_add_bio(struct btrfs_raid_bio *rbio, struct bio *orig_bio)
1637{
1638 const struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1639 const u64 orig_logical = orig_bio->bi_iter.bi_sector << SECTOR_SHIFT;
1640 const u64 full_stripe_start = rbio->bioc->full_stripe_logical;
1641 const u32 orig_len = orig_bio->bi_iter.bi_size;
1642 const u32 sectorsize = fs_info->sectorsize;
1643 u64 cur_logical;
1644
1645 ASSERT(orig_logical >= full_stripe_start &&
1646 orig_logical + orig_len <= full_stripe_start +
1647 rbio->nr_data * BTRFS_STRIPE_LEN);
1648
1649 bio_list_add(&rbio->bio_list, orig_bio);
1650 rbio->bio_list_bytes += orig_bio->bi_iter.bi_size;
1651
1652 /* Update the dbitmap. */
1653 for (cur_logical = orig_logical; cur_logical < orig_logical + orig_len;
1654 cur_logical += sectorsize) {
1655 int bit = ((u32)(cur_logical - full_stripe_start) >>
1656 fs_info->sectorsize_bits) % rbio->stripe_nsectors;
1657
1658 set_bit(bit, &rbio->dbitmap);
1659 }
1660}
1661
1662/*
1663 * our main entry point for writes from the rest of the FS.
1664 */
1665void raid56_parity_write(struct bio *bio, struct btrfs_io_context *bioc)
1666{
1667 struct btrfs_fs_info *fs_info = bioc->fs_info;
1668 struct btrfs_raid_bio *rbio;
1669 struct btrfs_plug_cb *plug = NULL;
1670 struct blk_plug_cb *cb;
1671
1672 rbio = alloc_rbio(fs_info, bioc);
1673 if (IS_ERR(rbio)) {
1674 bio->bi_status = errno_to_blk_status(PTR_ERR(rbio));
1675 bio_endio(bio);
1676 return;
1677 }
1678 rbio->operation = BTRFS_RBIO_WRITE;
1679 rbio_add_bio(rbio, bio);
1680
1681 /*
1682 * Don't plug on full rbios, just get them out the door
1683 * as quickly as we can
1684 */
1685 if (!rbio_is_full(rbio)) {
1686 cb = blk_check_plugged(raid_unplug, fs_info, sizeof(*plug));
1687 if (cb) {
1688 plug = container_of(cb, struct btrfs_plug_cb, cb);
1689 if (!plug->info) {
1690 plug->info = fs_info;
1691 INIT_LIST_HEAD(&plug->rbio_list);
1692 }
1693 list_add_tail(&rbio->plug_list, &plug->rbio_list);
1694 return;
1695 }
1696 }
1697
1698 /*
1699 * Either we don't have any existing plug, or we're doing a full stripe,
1700 * queue the rmw work now.
1701 */
1702 start_async_work(rbio, rmw_rbio_work);
1703}
1704
1705static int verify_one_sector(struct btrfs_raid_bio *rbio,
1706 int stripe_nr, int sector_nr)
1707{
1708 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1709 struct sector_ptr *sector;
1710 u8 csum_buf[BTRFS_CSUM_SIZE];
1711 u8 *csum_expected;
1712 int ret;
1713
1714 if (!rbio->csum_bitmap || !rbio->csum_buf)
1715 return 0;
1716
1717 /* No way to verify P/Q as they are not covered by data csum. */
1718 if (stripe_nr >= rbio->nr_data)
1719 return 0;
1720 /*
1721 * If we're rebuilding a read, we have to use pages from the
1722 * bio list if possible.
1723 */
1724 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
1725 sector = sector_in_rbio(rbio, stripe_nr, sector_nr, 0);
1726 } else {
1727 sector = rbio_stripe_sector(rbio, stripe_nr, sector_nr);
1728 }
1729
1730 ASSERT(sector->page);
1731
1732 csum_expected = rbio->csum_buf +
1733 (stripe_nr * rbio->stripe_nsectors + sector_nr) *
1734 fs_info->csum_size;
1735 ret = btrfs_check_sector_csum(fs_info, sector->page, sector->pgoff,
1736 csum_buf, csum_expected);
1737 return ret;
1738}
1739
1740/*
1741 * Recover a vertical stripe specified by @sector_nr.
1742 * @*pointers are the pre-allocated pointers by the caller, so we don't
1743 * need to allocate/free the pointers again and again.
1744 */
1745static int recover_vertical(struct btrfs_raid_bio *rbio, int sector_nr,
1746 void **pointers, void **unmap_array)
1747{
1748 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1749 struct sector_ptr *sector;
1750 const u32 sectorsize = fs_info->sectorsize;
1751 int found_errors;
1752 int faila;
1753 int failb;
1754 int stripe_nr;
1755 int ret = 0;
1756
1757 /*
1758 * Now we just use bitmap to mark the horizontal stripes in
1759 * which we have data when doing parity scrub.
1760 */
1761 if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB &&
1762 !test_bit(sector_nr, &rbio->dbitmap))
1763 return 0;
1764
1765 found_errors = get_rbio_veritical_errors(rbio, sector_nr, &faila,
1766 &failb);
1767 /*
1768 * No errors in the vertical stripe, skip it. Can happen for recovery
1769 * which only part of a stripe failed csum check.
1770 */
1771 if (!found_errors)
1772 return 0;
1773
1774 if (found_errors > rbio->bioc->max_errors)
1775 return -EIO;
1776
1777 /*
1778 * Setup our array of pointers with sectors from each stripe
1779 *
1780 * NOTE: store a duplicate array of pointers to preserve the
1781 * pointer order.
1782 */
1783 for (stripe_nr = 0; stripe_nr < rbio->real_stripes; stripe_nr++) {
1784 /*
1785 * If we're rebuilding a read, we have to use pages from the
1786 * bio list if possible.
1787 */
1788 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
1789 sector = sector_in_rbio(rbio, stripe_nr, sector_nr, 0);
1790 } else {
1791 sector = rbio_stripe_sector(rbio, stripe_nr, sector_nr);
1792 }
1793 ASSERT(sector->page);
1794 pointers[stripe_nr] = kmap_local_page(sector->page) +
1795 sector->pgoff;
1796 unmap_array[stripe_nr] = pointers[stripe_nr];
1797 }
1798
1799 /* All raid6 handling here */
1800 if (rbio->bioc->map_type & BTRFS_BLOCK_GROUP_RAID6) {
1801 /* Single failure, rebuild from parity raid5 style */
1802 if (failb < 0) {
1803 if (faila == rbio->nr_data)
1804 /*
1805 * Just the P stripe has failed, without
1806 * a bad data or Q stripe.
1807 * We have nothing to do, just skip the
1808 * recovery for this stripe.
1809 */
1810 goto cleanup;
1811 /*
1812 * a single failure in raid6 is rebuilt
1813 * in the pstripe code below
1814 */
1815 goto pstripe;
1816 }
1817
1818 /*
1819 * If the q stripe is failed, do a pstripe reconstruction from
1820 * the xors.
1821 * If both the q stripe and the P stripe are failed, we're
1822 * here due to a crc mismatch and we can't give them the
1823 * data they want.
1824 */
1825 if (failb == rbio->real_stripes - 1) {
1826 if (faila == rbio->real_stripes - 2)
1827 /*
1828 * Only P and Q are corrupted.
1829 * We only care about data stripes recovery,
1830 * can skip this vertical stripe.
1831 */
1832 goto cleanup;
1833 /*
1834 * Otherwise we have one bad data stripe and
1835 * a good P stripe. raid5!
1836 */
1837 goto pstripe;
1838 }
1839
1840 if (failb == rbio->real_stripes - 2) {
1841 raid6_datap_recov(rbio->real_stripes, sectorsize,
1842 faila, pointers);
1843 } else {
1844 raid6_2data_recov(rbio->real_stripes, sectorsize,
1845 faila, failb, pointers);
1846 }
1847 } else {
1848 void *p;
1849
1850 /* Rebuild from P stripe here (raid5 or raid6). */
1851 ASSERT(failb == -1);
1852pstripe:
1853 /* Copy parity block into failed block to start with */
1854 memcpy(pointers[faila], pointers[rbio->nr_data], sectorsize);
1855
1856 /* Rearrange the pointer array */
1857 p = pointers[faila];
1858 for (stripe_nr = faila; stripe_nr < rbio->nr_data - 1;
1859 stripe_nr++)
1860 pointers[stripe_nr] = pointers[stripe_nr + 1];
1861 pointers[rbio->nr_data - 1] = p;
1862
1863 /* Xor in the rest */
1864 run_xor(pointers, rbio->nr_data - 1, sectorsize);
1865
1866 }
1867
1868 /*
1869 * No matter if this is a RMW or recovery, we should have all
1870 * failed sectors repaired in the vertical stripe, thus they are now
1871 * uptodate.
1872 * Especially if we determine to cache the rbio, we need to
1873 * have at least all data sectors uptodate.
1874 *
1875 * If possible, also check if the repaired sector matches its data
1876 * checksum.
1877 */
1878 if (faila >= 0) {
1879 ret = verify_one_sector(rbio, faila, sector_nr);
1880 if (ret < 0)
1881 goto cleanup;
1882
1883 sector = rbio_stripe_sector(rbio, faila, sector_nr);
1884 sector->uptodate = 1;
1885 }
1886 if (failb >= 0) {
1887 ret = verify_one_sector(rbio, failb, sector_nr);
1888 if (ret < 0)
1889 goto cleanup;
1890
1891 sector = rbio_stripe_sector(rbio, failb, sector_nr);
1892 sector->uptodate = 1;
1893 }
1894
1895cleanup:
1896 for (stripe_nr = rbio->real_stripes - 1; stripe_nr >= 0; stripe_nr--)
1897 kunmap_local(unmap_array[stripe_nr]);
1898 return ret;
1899}
1900
1901static int recover_sectors(struct btrfs_raid_bio *rbio)
1902{
1903 void **pointers = NULL;
1904 void **unmap_array = NULL;
1905 int sectornr;
1906 int ret = 0;
1907
1908 /*
1909 * @pointers array stores the pointer for each sector.
1910 *
1911 * @unmap_array stores copy of pointers that does not get reordered
1912 * during reconstruction so that kunmap_local works.
1913 */
1914 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
1915 unmap_array = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
1916 if (!pointers || !unmap_array) {
1917 ret = -ENOMEM;
1918 goto out;
1919 }
1920
1921 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
1922 spin_lock(&rbio->bio_list_lock);
1923 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1924 spin_unlock(&rbio->bio_list_lock);
1925 }
1926
1927 index_rbio_pages(rbio);
1928
1929 for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++) {
1930 ret = recover_vertical(rbio, sectornr, pointers, unmap_array);
1931 if (ret < 0)
1932 break;
1933 }
1934
1935out:
1936 kfree(pointers);
1937 kfree(unmap_array);
1938 return ret;
1939}
1940
1941static void recover_rbio(struct btrfs_raid_bio *rbio)
1942{
1943 struct bio_list bio_list = BIO_EMPTY_LIST;
1944 int total_sector_nr;
1945 int ret = 0;
1946
1947 /*
1948 * Either we're doing recover for a read failure or degraded write,
1949 * caller should have set error bitmap correctly.
1950 */
1951 ASSERT(bitmap_weight(rbio->error_bitmap, rbio->nr_sectors));
1952
1953 /* For recovery, we need to read all sectors including P/Q. */
1954 ret = alloc_rbio_pages(rbio);
1955 if (ret < 0)
1956 goto out;
1957
1958 index_rbio_pages(rbio);
1959
1960 /*
1961 * Read everything that hasn't failed. However this time we will
1962 * not trust any cached sector.
1963 * As we may read out some stale data but higher layer is not reading
1964 * that stale part.
1965 *
1966 * So here we always re-read everything in recovery path.
1967 */
1968 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
1969 total_sector_nr++) {
1970 int stripe = total_sector_nr / rbio->stripe_nsectors;
1971 int sectornr = total_sector_nr % rbio->stripe_nsectors;
1972 struct sector_ptr *sector;
1973
1974 /*
1975 * Skip the range which has error. It can be a range which is
1976 * marked error (for csum mismatch), or it can be a missing
1977 * device.
1978 */
1979 if (!rbio->bioc->stripes[stripe].dev->bdev ||
1980 test_bit(total_sector_nr, rbio->error_bitmap)) {
1981 /*
1982 * Also set the error bit for missing device, which
1983 * may not yet have its error bit set.
1984 */
1985 set_bit(total_sector_nr, rbio->error_bitmap);
1986 continue;
1987 }
1988
1989 sector = rbio_stripe_sector(rbio, stripe, sectornr);
1990 ret = rbio_add_io_sector(rbio, &bio_list, sector, stripe,
1991 sectornr, REQ_OP_READ);
1992 if (ret < 0) {
1993 bio_list_put(&bio_list);
1994 goto out;
1995 }
1996 }
1997
1998 submit_read_wait_bio_list(rbio, &bio_list);
1999 ret = recover_sectors(rbio);
2000out:
2001 rbio_orig_end_io(rbio, errno_to_blk_status(ret));
2002}
2003
2004static void recover_rbio_work(struct work_struct *work)
2005{
2006 struct btrfs_raid_bio *rbio;
2007
2008 rbio = container_of(work, struct btrfs_raid_bio, work);
2009 if (!lock_stripe_add(rbio))
2010 recover_rbio(rbio);
2011}
2012
2013static void recover_rbio_work_locked(struct work_struct *work)
2014{
2015 recover_rbio(container_of(work, struct btrfs_raid_bio, work));
2016}
2017
2018static void set_rbio_raid6_extra_error(struct btrfs_raid_bio *rbio, int mirror_num)
2019{
2020 bool found = false;
2021 int sector_nr;
2022
2023 /*
2024 * This is for RAID6 extra recovery tries, thus mirror number should
2025 * be large than 2.
2026 * Mirror 1 means read from data stripes. Mirror 2 means rebuild using
2027 * RAID5 methods.
2028 */
2029 ASSERT(mirror_num > 2);
2030 for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) {
2031 int found_errors;
2032 int faila;
2033 int failb;
2034
2035 found_errors = get_rbio_veritical_errors(rbio, sector_nr,
2036 &faila, &failb);
2037 /* This vertical stripe doesn't have errors. */
2038 if (!found_errors)
2039 continue;
2040
2041 /*
2042 * If we found errors, there should be only one error marked
2043 * by previous set_rbio_range_error().
2044 */
2045 ASSERT(found_errors == 1);
2046 found = true;
2047
2048 /* Now select another stripe to mark as error. */
2049 failb = rbio->real_stripes - (mirror_num - 1);
2050 if (failb <= faila)
2051 failb--;
2052
2053 /* Set the extra bit in error bitmap. */
2054 if (failb >= 0)
2055 set_bit(failb * rbio->stripe_nsectors + sector_nr,
2056 rbio->error_bitmap);
2057 }
2058
2059 /* We should found at least one vertical stripe with error.*/
2060 ASSERT(found);
2061}
2062
2063/*
2064 * the main entry point for reads from the higher layers. This
2065 * is really only called when the normal read path had a failure,
2066 * so we assume the bio they send down corresponds to a failed part
2067 * of the drive.
2068 */
2069void raid56_parity_recover(struct bio *bio, struct btrfs_io_context *bioc,
2070 int mirror_num)
2071{
2072 struct btrfs_fs_info *fs_info = bioc->fs_info;
2073 struct btrfs_raid_bio *rbio;
2074
2075 rbio = alloc_rbio(fs_info, bioc);
2076 if (IS_ERR(rbio)) {
2077 bio->bi_status = errno_to_blk_status(PTR_ERR(rbio));
2078 bio_endio(bio);
2079 return;
2080 }
2081
2082 rbio->operation = BTRFS_RBIO_READ_REBUILD;
2083 rbio_add_bio(rbio, bio);
2084
2085 set_rbio_range_error(rbio, bio);
2086
2087 /*
2088 * Loop retry:
2089 * for 'mirror == 2', reconstruct from all other stripes.
2090 * for 'mirror_num > 2', select a stripe to fail on every retry.
2091 */
2092 if (mirror_num > 2)
2093 set_rbio_raid6_extra_error(rbio, mirror_num);
2094
2095 start_async_work(rbio, recover_rbio_work);
2096}
2097
2098static void fill_data_csums(struct btrfs_raid_bio *rbio)
2099{
2100 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
2101 struct btrfs_root *csum_root = btrfs_csum_root(fs_info,
2102 rbio->bioc->full_stripe_logical);
2103 const u64 start = rbio->bioc->full_stripe_logical;
2104 const u32 len = (rbio->nr_data * rbio->stripe_nsectors) <<
2105 fs_info->sectorsize_bits;
2106 int ret;
2107
2108 /* The rbio should not have its csum buffer initialized. */
2109 ASSERT(!rbio->csum_buf && !rbio->csum_bitmap);
2110
2111 /*
2112 * Skip the csum search if:
2113 *
2114 * - The rbio doesn't belong to data block groups
2115 * Then we are doing IO for tree blocks, no need to search csums.
2116 *
2117 * - The rbio belongs to mixed block groups
2118 * This is to avoid deadlock, as we're already holding the full
2119 * stripe lock, if we trigger a metadata read, and it needs to do
2120 * raid56 recovery, we will deadlock.
2121 */
2122 if (!(rbio->bioc->map_type & BTRFS_BLOCK_GROUP_DATA) ||
2123 rbio->bioc->map_type & BTRFS_BLOCK_GROUP_METADATA)
2124 return;
2125
2126 rbio->csum_buf = kzalloc(rbio->nr_data * rbio->stripe_nsectors *
2127 fs_info->csum_size, GFP_NOFS);
2128 rbio->csum_bitmap = bitmap_zalloc(rbio->nr_data * rbio->stripe_nsectors,
2129 GFP_NOFS);
2130 if (!rbio->csum_buf || !rbio->csum_bitmap) {
2131 ret = -ENOMEM;
2132 goto error;
2133 }
2134
2135 ret = btrfs_lookup_csums_bitmap(csum_root, NULL, start, start + len - 1,
2136 rbio->csum_buf, rbio->csum_bitmap);
2137 if (ret < 0)
2138 goto error;
2139 if (bitmap_empty(rbio->csum_bitmap, len >> fs_info->sectorsize_bits))
2140 goto no_csum;
2141 return;
2142
2143error:
2144 /*
2145 * We failed to allocate memory or grab the csum, but it's not fatal,
2146 * we can still continue. But better to warn users that RMW is no
2147 * longer safe for this particular sub-stripe write.
2148 */
2149 btrfs_warn_rl(fs_info,
2150"sub-stripe write for full stripe %llu is not safe, failed to get csum: %d",
2151 rbio->bioc->full_stripe_logical, ret);
2152no_csum:
2153 kfree(rbio->csum_buf);
2154 bitmap_free(rbio->csum_bitmap);
2155 rbio->csum_buf = NULL;
2156 rbio->csum_bitmap = NULL;
2157}
2158
2159static int rmw_read_wait_recover(struct btrfs_raid_bio *rbio)
2160{
2161 struct bio_list bio_list = BIO_EMPTY_LIST;
2162 int total_sector_nr;
2163 int ret = 0;
2164
2165 /*
2166 * Fill the data csums we need for data verification. We need to fill
2167 * the csum_bitmap/csum_buf first, as our endio function will try to
2168 * verify the data sectors.
2169 */
2170 fill_data_csums(rbio);
2171
2172 /*
2173 * Build a list of bios to read all sectors (including data and P/Q).
2174 *
2175 * This behavior is to compensate the later csum verification and recovery.
2176 */
2177 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
2178 total_sector_nr++) {
2179 struct sector_ptr *sector;
2180 int stripe = total_sector_nr / rbio->stripe_nsectors;
2181 int sectornr = total_sector_nr % rbio->stripe_nsectors;
2182
2183 sector = rbio_stripe_sector(rbio, stripe, sectornr);
2184 ret = rbio_add_io_sector(rbio, &bio_list, sector,
2185 stripe, sectornr, REQ_OP_READ);
2186 if (ret) {
2187 bio_list_put(&bio_list);
2188 return ret;
2189 }
2190 }
2191
2192 /*
2193 * We may or may not have any corrupted sectors (including missing dev
2194 * and csum mismatch), just let recover_sectors() to handle them all.
2195 */
2196 submit_read_wait_bio_list(rbio, &bio_list);
2197 return recover_sectors(rbio);
2198}
2199
2200static void raid_wait_write_end_io(struct bio *bio)
2201{
2202 struct btrfs_raid_bio *rbio = bio->bi_private;
2203 blk_status_t err = bio->bi_status;
2204
2205 if (err)
2206 rbio_update_error_bitmap(rbio, bio);
2207 bio_put(bio);
2208 if (atomic_dec_and_test(&rbio->stripes_pending))
2209 wake_up(&rbio->io_wait);
2210}
2211
2212static void submit_write_bios(struct btrfs_raid_bio *rbio,
2213 struct bio_list *bio_list)
2214{
2215 struct bio *bio;
2216
2217 atomic_set(&rbio->stripes_pending, bio_list_size(bio_list));
2218 while ((bio = bio_list_pop(bio_list))) {
2219 bio->bi_end_io = raid_wait_write_end_io;
2220
2221 if (trace_raid56_write_enabled()) {
2222 struct raid56_bio_trace_info trace_info = { 0 };
2223
2224 bio_get_trace_info(rbio, bio, &trace_info);
2225 trace_raid56_write(rbio, bio, &trace_info);
2226 }
2227 submit_bio(bio);
2228 }
2229}
2230
2231/*
2232 * To determine if we need to read any sector from the disk.
2233 * Should only be utilized in RMW path, to skip cached rbio.
2234 */
2235static bool need_read_stripe_sectors(struct btrfs_raid_bio *rbio)
2236{
2237 int i;
2238
2239 for (i = 0; i < rbio->nr_data * rbio->stripe_nsectors; i++) {
2240 struct sector_ptr *sector = &rbio->stripe_sectors[i];
2241
2242 /*
2243 * We have a sector which doesn't have page nor uptodate,
2244 * thus this rbio can not be cached one, as cached one must
2245 * have all its data sectors present and uptodate.
2246 */
2247 if (!sector->page || !sector->uptodate)
2248 return true;
2249 }
2250 return false;
2251}
2252
2253static void rmw_rbio(struct btrfs_raid_bio *rbio)
2254{
2255 struct bio_list bio_list;
2256 int sectornr;
2257 int ret = 0;
2258
2259 /*
2260 * Allocate the pages for parity first, as P/Q pages will always be
2261 * needed for both full-stripe and sub-stripe writes.
2262 */
2263 ret = alloc_rbio_parity_pages(rbio);
2264 if (ret < 0)
2265 goto out;
2266
2267 /*
2268 * Either full stripe write, or we have every data sector already
2269 * cached, can go to write path immediately.
2270 */
2271 if (!rbio_is_full(rbio) && need_read_stripe_sectors(rbio)) {
2272 /*
2273 * Now we're doing sub-stripe write, also need all data stripes
2274 * to do the full RMW.
2275 */
2276 ret = alloc_rbio_data_pages(rbio);
2277 if (ret < 0)
2278 goto out;
2279
2280 index_rbio_pages(rbio);
2281
2282 ret = rmw_read_wait_recover(rbio);
2283 if (ret < 0)
2284 goto out;
2285 }
2286
2287 /*
2288 * At this stage we're not allowed to add any new bios to the
2289 * bio list any more, anyone else that wants to change this stripe
2290 * needs to do their own rmw.
2291 */
2292 spin_lock(&rbio->bio_list_lock);
2293 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
2294 spin_unlock(&rbio->bio_list_lock);
2295
2296 bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors);
2297
2298 index_rbio_pages(rbio);
2299
2300 /*
2301 * We don't cache full rbios because we're assuming
2302 * the higher layers are unlikely to use this area of
2303 * the disk again soon. If they do use it again,
2304 * hopefully they will send another full bio.
2305 */
2306 if (!rbio_is_full(rbio))
2307 cache_rbio_pages(rbio);
2308 else
2309 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2310
2311 for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++)
2312 generate_pq_vertical(rbio, sectornr);
2313
2314 bio_list_init(&bio_list);
2315 ret = rmw_assemble_write_bios(rbio, &bio_list);
2316 if (ret < 0)
2317 goto out;
2318
2319 /* We should have at least one bio assembled. */
2320 ASSERT(bio_list_size(&bio_list));
2321 submit_write_bios(rbio, &bio_list);
2322 wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0);
2323
2324 /* We may have more errors than our tolerance during the read. */
2325 for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++) {
2326 int found_errors;
2327
2328 found_errors = get_rbio_veritical_errors(rbio, sectornr, NULL, NULL);
2329 if (found_errors > rbio->bioc->max_errors) {
2330 ret = -EIO;
2331 break;
2332 }
2333 }
2334out:
2335 rbio_orig_end_io(rbio, errno_to_blk_status(ret));
2336}
2337
2338static void rmw_rbio_work(struct work_struct *work)
2339{
2340 struct btrfs_raid_bio *rbio;
2341
2342 rbio = container_of(work, struct btrfs_raid_bio, work);
2343 if (lock_stripe_add(rbio) == 0)
2344 rmw_rbio(rbio);
2345}
2346
2347static void rmw_rbio_work_locked(struct work_struct *work)
2348{
2349 rmw_rbio(container_of(work, struct btrfs_raid_bio, work));
2350}
2351
2352/*
2353 * The following code is used to scrub/replace the parity stripe
2354 *
2355 * Caller must have already increased bio_counter for getting @bioc.
2356 *
2357 * Note: We need make sure all the pages that add into the scrub/replace
2358 * raid bio are correct and not be changed during the scrub/replace. That
2359 * is those pages just hold metadata or file data with checksum.
2360 */
2361
2362struct btrfs_raid_bio *raid56_parity_alloc_scrub_rbio(struct bio *bio,
2363 struct btrfs_io_context *bioc,
2364 struct btrfs_device *scrub_dev,
2365 unsigned long *dbitmap, int stripe_nsectors)
2366{
2367 struct btrfs_fs_info *fs_info = bioc->fs_info;
2368 struct btrfs_raid_bio *rbio;
2369 int i;
2370
2371 rbio = alloc_rbio(fs_info, bioc);
2372 if (IS_ERR(rbio))
2373 return NULL;
2374 bio_list_add(&rbio->bio_list, bio);
2375 /*
2376 * This is a special bio which is used to hold the completion handler
2377 * and make the scrub rbio is similar to the other types
2378 */
2379 ASSERT(!bio->bi_iter.bi_size);
2380 rbio->operation = BTRFS_RBIO_PARITY_SCRUB;
2381
2382 /*
2383 * After mapping bioc with BTRFS_MAP_WRITE, parities have been sorted
2384 * to the end position, so this search can start from the first parity
2385 * stripe.
2386 */
2387 for (i = rbio->nr_data; i < rbio->real_stripes; i++) {
2388 if (bioc->stripes[i].dev == scrub_dev) {
2389 rbio->scrubp = i;
2390 break;
2391 }
2392 }
2393 ASSERT(i < rbio->real_stripes);
2394
2395 bitmap_copy(&rbio->dbitmap, dbitmap, stripe_nsectors);
2396 return rbio;
2397}
2398
2399/*
2400 * We just scrub the parity that we have correct data on the same horizontal,
2401 * so we needn't allocate all pages for all the stripes.
2402 */
2403static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio)
2404{
2405 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
2406 int total_sector_nr;
2407
2408 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
2409 total_sector_nr++) {
2410 struct page *page;
2411 int sectornr = total_sector_nr % rbio->stripe_nsectors;
2412 int index = (total_sector_nr * sectorsize) >> PAGE_SHIFT;
2413
2414 if (!test_bit(sectornr, &rbio->dbitmap))
2415 continue;
2416 if (rbio->stripe_pages[index])
2417 continue;
2418 page = alloc_page(GFP_NOFS);
2419 if (!page)
2420 return -ENOMEM;
2421 rbio->stripe_pages[index] = page;
2422 }
2423 index_stripe_sectors(rbio);
2424 return 0;
2425}
2426
2427static int finish_parity_scrub(struct btrfs_raid_bio *rbio)
2428{
2429 struct btrfs_io_context *bioc = rbio->bioc;
2430 const u32 sectorsize = bioc->fs_info->sectorsize;
2431 void **pointers = rbio->finish_pointers;
2432 unsigned long *pbitmap = &rbio->finish_pbitmap;
2433 int nr_data = rbio->nr_data;
2434 int stripe;
2435 int sectornr;
2436 bool has_qstripe;
2437 struct sector_ptr p_sector = { 0 };
2438 struct sector_ptr q_sector = { 0 };
2439 struct bio_list bio_list;
2440 int is_replace = 0;
2441 int ret;
2442
2443 bio_list_init(&bio_list);
2444
2445 if (rbio->real_stripes - rbio->nr_data == 1)
2446 has_qstripe = false;
2447 else if (rbio->real_stripes - rbio->nr_data == 2)
2448 has_qstripe = true;
2449 else
2450 BUG();
2451
2452 /*
2453 * Replace is running and our P/Q stripe is being replaced, then we
2454 * need to duplicate the final write to replace target.
2455 */
2456 if (bioc->replace_nr_stripes && bioc->replace_stripe_src == rbio->scrubp) {
2457 is_replace = 1;
2458 bitmap_copy(pbitmap, &rbio->dbitmap, rbio->stripe_nsectors);
2459 }
2460
2461 /*
2462 * Because the higher layers(scrubber) are unlikely to
2463 * use this area of the disk again soon, so don't cache
2464 * it.
2465 */
2466 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2467
2468 p_sector.page = alloc_page(GFP_NOFS);
2469 if (!p_sector.page)
2470 return -ENOMEM;
2471 p_sector.pgoff = 0;
2472 p_sector.uptodate = 1;
2473
2474 if (has_qstripe) {
2475 /* RAID6, allocate and map temp space for the Q stripe */
2476 q_sector.page = alloc_page(GFP_NOFS);
2477 if (!q_sector.page) {
2478 __free_page(p_sector.page);
2479 p_sector.page = NULL;
2480 return -ENOMEM;
2481 }
2482 q_sector.pgoff = 0;
2483 q_sector.uptodate = 1;
2484 pointers[rbio->real_stripes - 1] = kmap_local_page(q_sector.page);
2485 }
2486
2487 bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors);
2488
2489 /* Map the parity stripe just once */
2490 pointers[nr_data] = kmap_local_page(p_sector.page);
2491
2492 for_each_set_bit(sectornr, &rbio->dbitmap, rbio->stripe_nsectors) {
2493 struct sector_ptr *sector;
2494 void *parity;
2495
2496 /* first collect one page from each data stripe */
2497 for (stripe = 0; stripe < nr_data; stripe++) {
2498 sector = sector_in_rbio(rbio, stripe, sectornr, 0);
2499 pointers[stripe] = kmap_local_page(sector->page) +
2500 sector->pgoff;
2501 }
2502
2503 if (has_qstripe) {
2504 assert_rbio(rbio);
2505 /* RAID6, call the library function to fill in our P/Q */
2506 raid6_call.gen_syndrome(rbio->real_stripes, sectorsize,
2507 pointers);
2508 } else {
2509 /* raid5 */
2510 memcpy(pointers[nr_data], pointers[0], sectorsize);
2511 run_xor(pointers + 1, nr_data - 1, sectorsize);
2512 }
2513
2514 /* Check scrubbing parity and repair it */
2515 sector = rbio_stripe_sector(rbio, rbio->scrubp, sectornr);
2516 parity = kmap_local_page(sector->page) + sector->pgoff;
2517 if (memcmp(parity, pointers[rbio->scrubp], sectorsize) != 0)
2518 memcpy(parity, pointers[rbio->scrubp], sectorsize);
2519 else
2520 /* Parity is right, needn't writeback */
2521 bitmap_clear(&rbio->dbitmap, sectornr, 1);
2522 kunmap_local(parity);
2523
2524 for (stripe = nr_data - 1; stripe >= 0; stripe--)
2525 kunmap_local(pointers[stripe]);
2526 }
2527
2528 kunmap_local(pointers[nr_data]);
2529 __free_page(p_sector.page);
2530 p_sector.page = NULL;
2531 if (q_sector.page) {
2532 kunmap_local(pointers[rbio->real_stripes - 1]);
2533 __free_page(q_sector.page);
2534 q_sector.page = NULL;
2535 }
2536
2537 /*
2538 * time to start writing. Make bios for everything from the
2539 * higher layers (the bio_list in our rbio) and our p/q. Ignore
2540 * everything else.
2541 */
2542 for_each_set_bit(sectornr, &rbio->dbitmap, rbio->stripe_nsectors) {
2543 struct sector_ptr *sector;
2544
2545 sector = rbio_stripe_sector(rbio, rbio->scrubp, sectornr);
2546 ret = rbio_add_io_sector(rbio, &bio_list, sector, rbio->scrubp,
2547 sectornr, REQ_OP_WRITE);
2548 if (ret)
2549 goto cleanup;
2550 }
2551
2552 if (!is_replace)
2553 goto submit_write;
2554
2555 /*
2556 * Replace is running and our parity stripe needs to be duplicated to
2557 * the target device. Check we have a valid source stripe number.
2558 */
2559 ASSERT(rbio->bioc->replace_stripe_src >= 0);
2560 for_each_set_bit(sectornr, pbitmap, rbio->stripe_nsectors) {
2561 struct sector_ptr *sector;
2562
2563 sector = rbio_stripe_sector(rbio, rbio->scrubp, sectornr);
2564 ret = rbio_add_io_sector(rbio, &bio_list, sector,
2565 rbio->real_stripes,
2566 sectornr, REQ_OP_WRITE);
2567 if (ret)
2568 goto cleanup;
2569 }
2570
2571submit_write:
2572 submit_write_bios(rbio, &bio_list);
2573 return 0;
2574
2575cleanup:
2576 bio_list_put(&bio_list);
2577 return ret;
2578}
2579
2580static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe)
2581{
2582 if (stripe >= 0 && stripe < rbio->nr_data)
2583 return 1;
2584 return 0;
2585}
2586
2587static int recover_scrub_rbio(struct btrfs_raid_bio *rbio)
2588{
2589 void **pointers = NULL;
2590 void **unmap_array = NULL;
2591 int sector_nr;
2592 int ret = 0;
2593
2594 /*
2595 * @pointers array stores the pointer for each sector.
2596 *
2597 * @unmap_array stores copy of pointers that does not get reordered
2598 * during reconstruction so that kunmap_local works.
2599 */
2600 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
2601 unmap_array = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
2602 if (!pointers || !unmap_array) {
2603 ret = -ENOMEM;
2604 goto out;
2605 }
2606
2607 for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) {
2608 int dfail = 0, failp = -1;
2609 int faila;
2610 int failb;
2611 int found_errors;
2612
2613 found_errors = get_rbio_veritical_errors(rbio, sector_nr,
2614 &faila, &failb);
2615 if (found_errors > rbio->bioc->max_errors) {
2616 ret = -EIO;
2617 goto out;
2618 }
2619 if (found_errors == 0)
2620 continue;
2621
2622 /* We should have at least one error here. */
2623 ASSERT(faila >= 0 || failb >= 0);
2624
2625 if (is_data_stripe(rbio, faila))
2626 dfail++;
2627 else if (is_parity_stripe(faila))
2628 failp = faila;
2629
2630 if (is_data_stripe(rbio, failb))
2631 dfail++;
2632 else if (is_parity_stripe(failb))
2633 failp = failb;
2634 /*
2635 * Because we can not use a scrubbing parity to repair the
2636 * data, so the capability of the repair is declined. (In the
2637 * case of RAID5, we can not repair anything.)
2638 */
2639 if (dfail > rbio->bioc->max_errors - 1) {
2640 ret = -EIO;
2641 goto out;
2642 }
2643 /*
2644 * If all data is good, only parity is correctly, just repair
2645 * the parity, no need to recover data stripes.
2646 */
2647 if (dfail == 0)
2648 continue;
2649
2650 /*
2651 * Here means we got one corrupted data stripe and one
2652 * corrupted parity on RAID6, if the corrupted parity is
2653 * scrubbing parity, luckily, use the other one to repair the
2654 * data, or we can not repair the data stripe.
2655 */
2656 if (failp != rbio->scrubp) {
2657 ret = -EIO;
2658 goto out;
2659 }
2660
2661 ret = recover_vertical(rbio, sector_nr, pointers, unmap_array);
2662 if (ret < 0)
2663 goto out;
2664 }
2665out:
2666 kfree(pointers);
2667 kfree(unmap_array);
2668 return ret;
2669}
2670
2671static int scrub_assemble_read_bios(struct btrfs_raid_bio *rbio)
2672{
2673 struct bio_list bio_list = BIO_EMPTY_LIST;
2674 int total_sector_nr;
2675 int ret = 0;
2676
2677 /* Build a list of bios to read all the missing parts. */
2678 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
2679 total_sector_nr++) {
2680 int sectornr = total_sector_nr % rbio->stripe_nsectors;
2681 int stripe = total_sector_nr / rbio->stripe_nsectors;
2682 struct sector_ptr *sector;
2683
2684 /* No data in the vertical stripe, no need to read. */
2685 if (!test_bit(sectornr, &rbio->dbitmap))
2686 continue;
2687
2688 /*
2689 * We want to find all the sectors missing from the rbio and
2690 * read them from the disk. If sector_in_rbio() finds a sector
2691 * in the bio list we don't need to read it off the stripe.
2692 */
2693 sector = sector_in_rbio(rbio, stripe, sectornr, 1);
2694 if (sector)
2695 continue;
2696
2697 sector = rbio_stripe_sector(rbio, stripe, sectornr);
2698 /*
2699 * The bio cache may have handed us an uptodate sector. If so,
2700 * use it.
2701 */
2702 if (sector->uptodate)
2703 continue;
2704
2705 ret = rbio_add_io_sector(rbio, &bio_list, sector, stripe,
2706 sectornr, REQ_OP_READ);
2707 if (ret) {
2708 bio_list_put(&bio_list);
2709 return ret;
2710 }
2711 }
2712
2713 submit_read_wait_bio_list(rbio, &bio_list);
2714 return 0;
2715}
2716
2717static void scrub_rbio(struct btrfs_raid_bio *rbio)
2718{
2719 int sector_nr;
2720 int ret;
2721
2722 ret = alloc_rbio_essential_pages(rbio);
2723 if (ret)
2724 goto out;
2725
2726 bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors);
2727
2728 ret = scrub_assemble_read_bios(rbio);
2729 if (ret < 0)
2730 goto out;
2731
2732 /* We may have some failures, recover the failed sectors first. */
2733 ret = recover_scrub_rbio(rbio);
2734 if (ret < 0)
2735 goto out;
2736
2737 /*
2738 * We have every sector properly prepared. Can finish the scrub
2739 * and writeback the good content.
2740 */
2741 ret = finish_parity_scrub(rbio);
2742 wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0);
2743 for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) {
2744 int found_errors;
2745
2746 found_errors = get_rbio_veritical_errors(rbio, sector_nr, NULL, NULL);
2747 if (found_errors > rbio->bioc->max_errors) {
2748 ret = -EIO;
2749 break;
2750 }
2751 }
2752out:
2753 rbio_orig_end_io(rbio, errno_to_blk_status(ret));
2754}
2755
2756static void scrub_rbio_work_locked(struct work_struct *work)
2757{
2758 scrub_rbio(container_of(work, struct btrfs_raid_bio, work));
2759}
2760
2761void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio)
2762{
2763 if (!lock_stripe_add(rbio))
2764 start_async_work(rbio, scrub_rbio_work_locked);
2765}
2766
2767/*
2768 * This is for scrub call sites where we already have correct data contents.
2769 * This allows us to avoid reading data stripes again.
2770 *
2771 * Unfortunately here we have to do page copy, other than reusing the pages.
2772 * This is due to the fact rbio has its own page management for its cache.
2773 */
2774void raid56_parity_cache_data_pages(struct btrfs_raid_bio *rbio,
2775 struct page **data_pages, u64 data_logical)
2776{
2777 const u64 offset_in_full_stripe = data_logical -
2778 rbio->bioc->full_stripe_logical;
2779 const int page_index = offset_in_full_stripe >> PAGE_SHIFT;
2780 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
2781 const u32 sectors_per_page = PAGE_SIZE / sectorsize;
2782 int ret;
2783
2784 /*
2785 * If we hit ENOMEM temporarily, but later at
2786 * raid56_parity_submit_scrub_rbio() time it succeeded, we just do
2787 * the extra read, not a big deal.
2788 *
2789 * If we hit ENOMEM later at raid56_parity_submit_scrub_rbio() time,
2790 * the bio would got proper error number set.
2791 */
2792 ret = alloc_rbio_data_pages(rbio);
2793 if (ret < 0)
2794 return;
2795
2796 /* data_logical must be at stripe boundary and inside the full stripe. */
2797 ASSERT(IS_ALIGNED(offset_in_full_stripe, BTRFS_STRIPE_LEN));
2798 ASSERT(offset_in_full_stripe < (rbio->nr_data << BTRFS_STRIPE_LEN_SHIFT));
2799
2800 for (int page_nr = 0; page_nr < (BTRFS_STRIPE_LEN >> PAGE_SHIFT); page_nr++) {
2801 struct page *dst = rbio->stripe_pages[page_nr + page_index];
2802 struct page *src = data_pages[page_nr];
2803
2804 memcpy_page(dst, 0, src, 0, PAGE_SIZE);
2805 for (int sector_nr = sectors_per_page * page_index;
2806 sector_nr < sectors_per_page * (page_index + 1);
2807 sector_nr++)
2808 rbio->stripe_sectors[sector_nr].uptodate = true;
2809 }
2810}