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