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