1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (C) 2012 Fusion-io All rights reserved.
4 * Copyright (C) 2012 Intel Corp. All rights reserved.
5 */
6
7 #include <linux/sched.h>
8 #include <linux/bio.h>
9 #include <linux/slab.h>
10 #include <linux/blkdev.h>
11 #include <linux/raid/pq.h>
12 #include <linux/hash.h>
13 #include <linux/list_sort.h>
14 #include <linux/raid/xor.h>
15 #include <linux/mm.h>
16 #include "messages.h"
17 #include "ctree.h"
18 #include "disk-io.h"
19 #include "volumes.h"
20 #include "raid56.h"
21 #include "async-thread.h"
22 #include "file-item.h"
23 #include "btrfs_inode.h"
24
25 /* set when additional merges to this rbio are not allowed */
26 #define RBIO_RMW_LOCKED_BIT 1
27
28 /*
29 * set when this rbio is sitting in the hash, but it is just a cache
30 * of past RMW
31 */
32 #define RBIO_CACHE_BIT 2
33
34 /*
35 * set when it is safe to trust the stripe_pages for caching
36 */
37 #define RBIO_CACHE_READY_BIT 3
38
39 #define RBIO_CACHE_SIZE 1024
40
41 #define BTRFS_STRIPE_HASH_TABLE_BITS 11
42
43 static void dump_bioc(const struct btrfs_fs_info *fs_info, const struct btrfs_io_context *bioc)
44 {
45 if (unlikely(!bioc)) {
46 btrfs_crit(fs_info, "bioc=NULL");
47 return;
48 }
49 btrfs_crit(fs_info,
50 "bioc logical=%llu full_stripe=%llu size=%llu map_type=0x%llx mirror=%u replace_nr_stripes=%u replace_stripe_src=%d num_stripes=%u",
51 bioc->logical, bioc->full_stripe_logical, bioc->size,
52 bioc->map_type, bioc->mirror_num, bioc->replace_nr_stripes,
53 bioc->replace_stripe_src, bioc->num_stripes);
54 for (int i = 0; i < bioc->num_stripes; i++) {
55 btrfs_crit(fs_info, " nr=%d devid=%llu physical=%llu",
56 i, bioc->stripes[i].dev->devid,
57 bioc->stripes[i].physical);
58 }
59 }
60
61 static void btrfs_dump_rbio(const struct btrfs_fs_info *fs_info,
62 const struct btrfs_raid_bio *rbio)
63 {
64 if (!IS_ENABLED(CONFIG_BTRFS_ASSERT))
65 return;
66
67 dump_bioc(fs_info, rbio->bioc);
68 btrfs_crit(fs_info,
69 "rbio flags=0x%lx nr_sectors=%u nr_data=%u real_stripes=%u stripe_nsectors=%u scrubp=%u dbitmap=0x%lx",
70 rbio->flags, rbio->nr_sectors, rbio->nr_data,
71 rbio->real_stripes, rbio->stripe_nsectors,
72 rbio->scrubp, rbio->dbitmap);
73 }
74
75 #define ASSERT_RBIO(expr, rbio) \
76 ({ \
77 if (IS_ENABLED(CONFIG_BTRFS_ASSERT) && unlikely(!(expr))) { \
78 const struct btrfs_fs_info *__fs_info = (rbio)->bioc ? \
79 (rbio)->bioc->fs_info : NULL; \
80 \
81 btrfs_dump_rbio(__fs_info, (rbio)); \
82 } \
83 ASSERT((expr)); \
84 })
85
86 #define ASSERT_RBIO_STRIPE(expr, rbio, stripe_nr) \
87 ({ \
88 if (IS_ENABLED(CONFIG_BTRFS_ASSERT) && unlikely(!(expr))) { \
89 const struct btrfs_fs_info *__fs_info = (rbio)->bioc ? \
90 (rbio)->bioc->fs_info : NULL; \
91 \
92 btrfs_dump_rbio(__fs_info, (rbio)); \
93 btrfs_crit(__fs_info, "stripe_nr=%d", (stripe_nr)); \
94 } \
95 ASSERT((expr)); \
96 })
97
98 #define ASSERT_RBIO_SECTOR(expr, rbio, sector_nr) \
99 ({ \
100 if (IS_ENABLED(CONFIG_BTRFS_ASSERT) && unlikely(!(expr))) { \
101 const struct btrfs_fs_info *__fs_info = (rbio)->bioc ? \
102 (rbio)->bioc->fs_info : NULL; \
103 \
104 btrfs_dump_rbio(__fs_info, (rbio)); \
105 btrfs_crit(__fs_info, "sector_nr=%d", (sector_nr)); \
106 } \
107 ASSERT((expr)); \
108 })
109
110 #define ASSERT_RBIO_LOGICAL(expr, rbio, logical) \
111 ({ \
112 if (IS_ENABLED(CONFIG_BTRFS_ASSERT) && unlikely(!(expr))) { \
113 const struct btrfs_fs_info *__fs_info = (rbio)->bioc ? \
114 (rbio)->bioc->fs_info : NULL; \
115 \
116 btrfs_dump_rbio(__fs_info, (rbio)); \
117 btrfs_crit(__fs_info, "logical=%llu", (logical)); \
118 } \
119 ASSERT((expr)); \
120 })
121
122 /* Used by the raid56 code to lock stripes for read/modify/write */
123 struct btrfs_stripe_hash {
124 struct list_head hash_list;
125 spinlock_t lock;
126 };
127
128 /* Used by the raid56 code to lock stripes for read/modify/write */
129 struct btrfs_stripe_hash_table {
130 struct list_head stripe_cache;
131 spinlock_t cache_lock;
132 int cache_size;
133 struct btrfs_stripe_hash table[];
134 };
135
136 /*
137 * A bvec like structure to present a sector inside a page.
138 *
139 * Unlike bvec we don't need bvlen, as it's fixed to sectorsize.
140 */
141 struct sector_ptr {
142 struct page *page;
143 unsigned int pgoff:24;
144 unsigned int uptodate:8;
145 };
146
147 static void rmw_rbio_work(struct work_struct *work);
148 static void rmw_rbio_work_locked(struct work_struct *work);
149 static void index_rbio_pages(struct btrfs_raid_bio *rbio);
150 static int alloc_rbio_pages(struct btrfs_raid_bio *rbio);
151
152 static int finish_parity_scrub(struct btrfs_raid_bio *rbio);
153 static void scrub_rbio_work_locked(struct work_struct *work);
154
155 static void free_raid_bio_pointers(struct btrfs_raid_bio *rbio)
156 {
157 bitmap_free(rbio->error_bitmap);
158 kfree(rbio->stripe_pages);
159 kfree(rbio->bio_sectors);
160 kfree(rbio->stripe_sectors);
161 kfree(rbio->finish_pointers);
162 }
163
164 static void free_raid_bio(struct btrfs_raid_bio *rbio)
165 {
166 int i;
167
168 if (!refcount_dec_and_test(&rbio->refs))
169 return;
170
171 WARN_ON(!list_empty(&rbio->stripe_cache));
172 WARN_ON(!list_empty(&rbio->hash_list));
173 WARN_ON(!bio_list_empty(&rbio->bio_list));
174
175 for (i = 0; i < rbio->nr_pages; i++) {
176 if (rbio->stripe_pages[i]) {
177 __free_page(rbio->stripe_pages[i]);
178 rbio->stripe_pages[i] = NULL;
179 }
180 }
181
182 btrfs_put_bioc(rbio->bioc);
183 free_raid_bio_pointers(rbio);
184 kfree(rbio);
185 }
186
187 static void start_async_work(struct btrfs_raid_bio *rbio, work_func_t work_func)
188 {
189 INIT_WORK(&rbio->work, work_func);
190 queue_work(rbio->bioc->fs_info->rmw_workers, &rbio->work);
191 }
192
193 /*
194 * the stripe hash table is used for locking, and to collect
195 * bios in hopes of making a full stripe
196 */
197 int btrfs_alloc_stripe_hash_table(struct btrfs_fs_info *info)
198 {
199 struct btrfs_stripe_hash_table *table;
200 struct btrfs_stripe_hash_table *x;
201 struct btrfs_stripe_hash *cur;
202 struct btrfs_stripe_hash *h;
203 int num_entries = 1 << BTRFS_STRIPE_HASH_TABLE_BITS;
204 int i;
205
206 if (info->stripe_hash_table)
207 return 0;
208
209 /*
210 * The table is large, starting with order 4 and can go as high as
211 * order 7 in case lock debugging is turned on.
212 *
213 * Try harder to allocate and fallback to vmalloc to lower the chance
214 * of a failing mount.
215 */
216 table = kvzalloc(struct_size(table, table, num_entries), GFP_KERNEL);
217 if (!table)
218 return -ENOMEM;
219
220 spin_lock_init(&table->cache_lock);
221 INIT_LIST_HEAD(&table->stripe_cache);
222
223 h = table->table;
224
225 for (i = 0; i < num_entries; i++) {
226 cur = h + i;
227 INIT_LIST_HEAD(&cur->hash_list);
228 spin_lock_init(&cur->lock);
229 }
230
231 x = cmpxchg(&info->stripe_hash_table, NULL, table);
232 kvfree(x);
233 return 0;
234 }
235
236 /*
237 * caching an rbio means to copy anything from the
238 * bio_sectors array into the stripe_pages array. We
239 * use the page uptodate bit in the stripe cache array
240 * to indicate if it has valid data
241 *
242 * once the caching is done, we set the cache ready
243 * bit.
244 */
245 static void cache_rbio_pages(struct btrfs_raid_bio *rbio)
246 {
247 int i;
248 int ret;
249
250 ret = alloc_rbio_pages(rbio);
251 if (ret)
252 return;
253
254 for (i = 0; i < rbio->nr_sectors; i++) {
255 /* Some range not covered by bio (partial write), skip it */
256 if (!rbio->bio_sectors[i].page) {
257 /*
258 * Even if the sector is not covered by bio, if it is
259 * a data sector it should still be uptodate as it is
260 * read from disk.
261 */
262 if (i < rbio->nr_data * rbio->stripe_nsectors)
263 ASSERT(rbio->stripe_sectors[i].uptodate);
264 continue;
265 }
266
267 ASSERT(rbio->stripe_sectors[i].page);
268 memcpy_page(rbio->stripe_sectors[i].page,
269 rbio->stripe_sectors[i].pgoff,
270 rbio->bio_sectors[i].page,
271 rbio->bio_sectors[i].pgoff,
272 rbio->bioc->fs_info->sectorsize);
273 rbio->stripe_sectors[i].uptodate = 1;
274 }
275 set_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
276 }
277
278 /*
279 * we hash on the first logical address of the stripe
280 */
281 static int rbio_bucket(struct btrfs_raid_bio *rbio)
282 {
283 u64 num = rbio->bioc->full_stripe_logical;
284
285 /*
286 * we shift down quite a bit. We're using byte
287 * addressing, and most of the lower bits are zeros.
288 * This tends to upset hash_64, and it consistently
289 * returns just one or two different values.
290 *
291 * shifting off the lower bits fixes things.
292 */
293 return hash_64(num >> 16, BTRFS_STRIPE_HASH_TABLE_BITS);
294 }
295
296 static bool full_page_sectors_uptodate(struct btrfs_raid_bio *rbio,
297 unsigned int page_nr)
298 {
299 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
300 const u32 sectors_per_page = PAGE_SIZE / sectorsize;
301 int i;
302
303 ASSERT(page_nr < rbio->nr_pages);
304
305 for (i = sectors_per_page * page_nr;
306 i < sectors_per_page * page_nr + sectors_per_page;
307 i++) {
308 if (!rbio->stripe_sectors[i].uptodate)
309 return false;
310 }
311 return true;
312 }
313
314 /*
315 * Update the stripe_sectors[] array to use correct page and pgoff
316 *
317 * Should be called every time any page pointer in stripes_pages[] got modified.
318 */
319 static void index_stripe_sectors(struct btrfs_raid_bio *rbio)
320 {
321 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
322 u32 offset;
323 int i;
324
325 for (i = 0, offset = 0; i < rbio->nr_sectors; i++, offset += sectorsize) {
326 int page_index = offset >> PAGE_SHIFT;
327
328 ASSERT(page_index < rbio->nr_pages);
329 rbio->stripe_sectors[i].page = rbio->stripe_pages[page_index];
330 rbio->stripe_sectors[i].pgoff = offset_in_page(offset);
331 }
332 }
333
334 static void steal_rbio_page(struct btrfs_raid_bio *src,
335 struct btrfs_raid_bio *dest, int page_nr)
336 {
337 const u32 sectorsize = src->bioc->fs_info->sectorsize;
338 const u32 sectors_per_page = PAGE_SIZE / sectorsize;
339 int i;
340
341 if (dest->stripe_pages[page_nr])
342 __free_page(dest->stripe_pages[page_nr]);
343 dest->stripe_pages[page_nr] = src->stripe_pages[page_nr];
344 src->stripe_pages[page_nr] = NULL;
345
346 /* Also update the sector->uptodate bits. */
347 for (i = sectors_per_page * page_nr;
348 i < sectors_per_page * page_nr + sectors_per_page; i++)
349 dest->stripe_sectors[i].uptodate = true;
350 }
351
352 static bool is_data_stripe_page(struct btrfs_raid_bio *rbio, int page_nr)
353 {
354 const int sector_nr = (page_nr << PAGE_SHIFT) >>
355 rbio->bioc->fs_info->sectorsize_bits;
356
357 /*
358 * We have ensured PAGE_SIZE is aligned with sectorsize, thus
359 * we won't have a page which is half data half parity.
360 *
361 * Thus if the first sector of the page belongs to data stripes, then
362 * the full page belongs to data stripes.
363 */
364 return (sector_nr < rbio->nr_data * rbio->stripe_nsectors);
365 }
366
367 /*
368 * Stealing an rbio means taking all the uptodate pages from the stripe array
369 * in the source rbio and putting them into the destination rbio.
370 *
371 * This will also update the involved stripe_sectors[] which are referring to
372 * the old pages.
373 */
374 static void steal_rbio(struct btrfs_raid_bio *src, struct btrfs_raid_bio *dest)
375 {
376 int i;
377
378 if (!test_bit(RBIO_CACHE_READY_BIT, &src->flags))
379 return;
380
381 for (i = 0; i < dest->nr_pages; i++) {
382 struct page *p = src->stripe_pages[i];
383
384 /*
385 * We don't need to steal P/Q pages as they will always be
386 * regenerated for RMW or full write anyway.
387 */
388 if (!is_data_stripe_page(src, i))
389 continue;
390
391 /*
392 * If @src already has RBIO_CACHE_READY_BIT, it should have
393 * all data stripe pages present and uptodate.
394 */
395 ASSERT(p);
396 ASSERT(full_page_sectors_uptodate(src, i));
397 steal_rbio_page(src, dest, i);
398 }
399 index_stripe_sectors(dest);
400 index_stripe_sectors(src);
401 }
402
403 /*
404 * merging means we take the bio_list from the victim and
405 * splice it into the destination. The victim should
406 * be discarded afterwards.
407 *
408 * must be called with dest->rbio_list_lock held
409 */
410 static void merge_rbio(struct btrfs_raid_bio *dest,
411 struct btrfs_raid_bio *victim)
412 {
413 bio_list_merge_init(&dest->bio_list, &victim->bio_list);
414 dest->bio_list_bytes += victim->bio_list_bytes;
415 /* Also inherit the bitmaps from @victim. */
416 bitmap_or(&dest->dbitmap, &victim->dbitmap, &dest->dbitmap,
417 dest->stripe_nsectors);
418 }
419
420 /*
421 * used to prune items that are in the cache. The caller
422 * must hold the hash table lock.
423 */
424 static void __remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
425 {
426 int bucket = rbio_bucket(rbio);
427 struct btrfs_stripe_hash_table *table;
428 struct btrfs_stripe_hash *h;
429 int freeit = 0;
430
431 /*
432 * check the bit again under the hash table lock.
433 */
434 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
435 return;
436
437 table = rbio->bioc->fs_info->stripe_hash_table;
438 h = table->table + bucket;
439
440 /* hold the lock for the bucket because we may be
441 * removing it from the hash table
442 */
443 spin_lock(&h->lock);
444
445 /*
446 * hold the lock for the bio list because we need
447 * to make sure the bio list is empty
448 */
449 spin_lock(&rbio->bio_list_lock);
450
451 if (test_and_clear_bit(RBIO_CACHE_BIT, &rbio->flags)) {
452 list_del_init(&rbio->stripe_cache);
453 table->cache_size -= 1;
454 freeit = 1;
455
456 /* if the bio list isn't empty, this rbio is
457 * still involved in an IO. We take it out
458 * of the cache list, and drop the ref that
459 * was held for the list.
460 *
461 * If the bio_list was empty, we also remove
462 * the rbio from the hash_table, and drop
463 * the corresponding ref
464 */
465 if (bio_list_empty(&rbio->bio_list)) {
466 if (!list_empty(&rbio->hash_list)) {
467 list_del_init(&rbio->hash_list);
468 refcount_dec(&rbio->refs);
469 BUG_ON(!list_empty(&rbio->plug_list));
470 }
471 }
472 }
473
474 spin_unlock(&rbio->bio_list_lock);
475 spin_unlock(&h->lock);
476
477 if (freeit)
478 free_raid_bio(rbio);
479 }
480
481 /*
482 * prune a given rbio from the cache
483 */
484 static void remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
485 {
486 struct btrfs_stripe_hash_table *table;
487
488 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
489 return;
490
491 table = rbio->bioc->fs_info->stripe_hash_table;
492
493 spin_lock(&table->cache_lock);
494 __remove_rbio_from_cache(rbio);
495 spin_unlock(&table->cache_lock);
496 }
497
498 /*
499 * remove everything in the cache
500 */
501 static void btrfs_clear_rbio_cache(struct btrfs_fs_info *info)
502 {
503 struct btrfs_stripe_hash_table *table;
504 struct btrfs_raid_bio *rbio;
505
506 table = info->stripe_hash_table;
507
508 spin_lock(&table->cache_lock);
509 while (!list_empty(&table->stripe_cache)) {
510 rbio = list_entry(table->stripe_cache.next,
511 struct btrfs_raid_bio,
512 stripe_cache);
513 __remove_rbio_from_cache(rbio);
514 }
515 spin_unlock(&table->cache_lock);
516 }
517
518 /*
519 * remove all cached entries and free the hash table
520 * used by unmount
521 */
522 void btrfs_free_stripe_hash_table(struct btrfs_fs_info *info)
523 {
524 if (!info->stripe_hash_table)
525 return;
526 btrfs_clear_rbio_cache(info);
527 kvfree(info->stripe_hash_table);
528 info->stripe_hash_table = NULL;
529 }
530
531 /*
532 * insert an rbio into the stripe cache. It
533 * must have already been prepared by calling
534 * cache_rbio_pages
535 *
536 * If this rbio was already cached, it gets
537 * moved to the front of the lru.
538 *
539 * If the size of the rbio cache is too big, we
540 * prune an item.
541 */
542 static void cache_rbio(struct btrfs_raid_bio *rbio)
543 {
544 struct btrfs_stripe_hash_table *table;
545
546 if (!test_bit(RBIO_CACHE_READY_BIT, &rbio->flags))
547 return;
548
549 table = rbio->bioc->fs_info->stripe_hash_table;
550
551 spin_lock(&table->cache_lock);
552 spin_lock(&rbio->bio_list_lock);
553
554 /* bump our ref if we were not in the list before */
555 if (!test_and_set_bit(RBIO_CACHE_BIT, &rbio->flags))
556 refcount_inc(&rbio->refs);
557
558 if (!list_empty(&rbio->stripe_cache)){
559 list_move(&rbio->stripe_cache, &table->stripe_cache);
560 } else {
561 list_add(&rbio->stripe_cache, &table->stripe_cache);
562 table->cache_size += 1;
563 }
564
565 spin_unlock(&rbio->bio_list_lock);
566
567 if (table->cache_size > RBIO_CACHE_SIZE) {
568 struct btrfs_raid_bio *found;
569
570 found = list_entry(table->stripe_cache.prev,
571 struct btrfs_raid_bio,
572 stripe_cache);
573
574 if (found != rbio)
575 __remove_rbio_from_cache(found);
576 }
577
578 spin_unlock(&table->cache_lock);
579 }
580
581 /*
582 * helper function to run the xor_blocks api. It is only
583 * able to do MAX_XOR_BLOCKS at a time, so we need to
584 * loop through.
585 */
586 static void run_xor(void **pages, int src_cnt, ssize_t len)
587 {
588 int src_off = 0;
589 int xor_src_cnt = 0;
590 void *dest = pages[src_cnt];
591
592 while(src_cnt > 0) {
593 xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS);
594 xor_blocks(xor_src_cnt, len, dest, pages + src_off);
595
596 src_cnt -= xor_src_cnt;
597 src_off += xor_src_cnt;
598 }
599 }
600
601 /*
602 * Returns true if the bio list inside this rbio covers an entire stripe (no
603 * rmw required).
604 */
605 static int rbio_is_full(struct btrfs_raid_bio *rbio)
606 {
607 unsigned long size = rbio->bio_list_bytes;
608 int ret = 1;
609
610 spin_lock(&rbio->bio_list_lock);
611 if (size != rbio->nr_data * BTRFS_STRIPE_LEN)
612 ret = 0;
613 BUG_ON(size > rbio->nr_data * BTRFS_STRIPE_LEN);
614 spin_unlock(&rbio->bio_list_lock);
615
616 return ret;
617 }
618
619 /*
620 * returns 1 if it is safe to merge two rbios together.
621 * The merging is safe if the two rbios correspond to
622 * the same stripe and if they are both going in the same
623 * direction (read vs write), and if neither one is
624 * locked for final IO
625 *
626 * The caller is responsible for locking such that
627 * rmw_locked is safe to test
628 */
629 static int rbio_can_merge(struct btrfs_raid_bio *last,
630 struct btrfs_raid_bio *cur)
631 {
632 if (test_bit(RBIO_RMW_LOCKED_BIT, &last->flags) ||
633 test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags))
634 return 0;
635
636 /*
637 * we can't merge with cached rbios, since the
638 * idea is that when we merge the destination
639 * rbio is going to run our IO for us. We can
640 * steal from cached rbios though, other functions
641 * handle that.
642 */
643 if (test_bit(RBIO_CACHE_BIT, &last->flags) ||
644 test_bit(RBIO_CACHE_BIT, &cur->flags))
645 return 0;
646
647 if (last->bioc->full_stripe_logical != cur->bioc->full_stripe_logical)
648 return 0;
649
650 /* we can't merge with different operations */
651 if (last->operation != cur->operation)
652 return 0;
653 /*
654 * We've need read the full stripe from the drive.
655 * check and repair the parity and write the new results.
656 *
657 * We're not allowed to add any new bios to the
658 * bio list here, anyone else that wants to
659 * change this stripe needs to do their own rmw.
660 */
661 if (last->operation == BTRFS_RBIO_PARITY_SCRUB)
662 return 0;
663
664 if (last->operation == BTRFS_RBIO_READ_REBUILD)
665 return 0;
666
667 return 1;
668 }
669
670 static unsigned int rbio_stripe_sector_index(const struct btrfs_raid_bio *rbio,
671 unsigned int stripe_nr,
672 unsigned int sector_nr)
673 {
674 ASSERT_RBIO_STRIPE(stripe_nr < rbio->real_stripes, rbio, stripe_nr);
675 ASSERT_RBIO_SECTOR(sector_nr < rbio->stripe_nsectors, rbio, sector_nr);
676
677 return stripe_nr * rbio->stripe_nsectors + sector_nr;
678 }
679
680 /* Return a sector from rbio->stripe_sectors, not from the bio list */
681 static struct sector_ptr *rbio_stripe_sector(const struct btrfs_raid_bio *rbio,
682 unsigned int stripe_nr,
683 unsigned int sector_nr)
684 {
685 return &rbio->stripe_sectors[rbio_stripe_sector_index(rbio, stripe_nr,
686 sector_nr)];
687 }
688
689 /* Grab a sector inside P stripe */
690 static struct sector_ptr *rbio_pstripe_sector(const struct btrfs_raid_bio *rbio,
691 unsigned int sector_nr)
692 {
693 return rbio_stripe_sector(rbio, rbio->nr_data, sector_nr);
694 }
695
696 /* Grab a sector inside Q stripe, return NULL if not RAID6 */
697 static struct sector_ptr *rbio_qstripe_sector(const struct btrfs_raid_bio *rbio,
698 unsigned int sector_nr)
699 {
700 if (rbio->nr_data + 1 == rbio->real_stripes)
701 return NULL;
702 return rbio_stripe_sector(rbio, rbio->nr_data + 1, sector_nr);
703 }
704
705 /*
706 * The first stripe in the table for a logical address
707 * has the lock. rbios are added in one of three ways:
708 *
709 * 1) Nobody has the stripe locked yet. The rbio is given
710 * the lock and 0 is returned. The caller must start the IO
711 * themselves.
712 *
713 * 2) Someone has the stripe locked, but we're able to merge
714 * with the lock owner. The rbio is freed and the IO will
715 * start automatically along with the existing rbio. 1 is returned.
716 *
717 * 3) Someone has the stripe locked, but we're not able to merge.
718 * The rbio is added to the lock owner's plug list, or merged into
719 * an rbio already on the plug list. When the lock owner unlocks,
720 * the next rbio on the list is run and the IO is started automatically.
721 * 1 is returned
722 *
723 * If we return 0, the caller still owns the rbio and must continue with
724 * IO submission. If we return 1, the caller must assume the rbio has
725 * already been freed.
726 */
727 static noinline int lock_stripe_add(struct btrfs_raid_bio *rbio)
728 {
729 struct btrfs_stripe_hash *h;
730 struct btrfs_raid_bio *cur;
731 struct btrfs_raid_bio *pending;
732 struct btrfs_raid_bio *freeit = NULL;
733 struct btrfs_raid_bio *cache_drop = NULL;
734 int ret = 0;
735
736 h = rbio->bioc->fs_info->stripe_hash_table->table + rbio_bucket(rbio);
737
738 spin_lock(&h->lock);
739 list_for_each_entry(cur, &h->hash_list, hash_list) {
740 if (cur->bioc->full_stripe_logical != rbio->bioc->full_stripe_logical)
741 continue;
742
743 spin_lock(&cur->bio_list_lock);
744
745 /* Can we steal this cached rbio's pages? */
746 if (bio_list_empty(&cur->bio_list) &&
747 list_empty(&cur->plug_list) &&
748 test_bit(RBIO_CACHE_BIT, &cur->flags) &&
749 !test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) {
750 list_del_init(&cur->hash_list);
751 refcount_dec(&cur->refs);
752
753 steal_rbio(cur, rbio);
754 cache_drop = cur;
755 spin_unlock(&cur->bio_list_lock);
756
757 goto lockit;
758 }
759
760 /* Can we merge into the lock owner? */
761 if (rbio_can_merge(cur, rbio)) {
762 merge_rbio(cur, rbio);
763 spin_unlock(&cur->bio_list_lock);
764 freeit = rbio;
765 ret = 1;
766 goto out;
767 }
768
769
770 /*
771 * We couldn't merge with the running rbio, see if we can merge
772 * with the pending ones. We don't have to check for rmw_locked
773 * because there is no way they are inside finish_rmw right now
774 */
775 list_for_each_entry(pending, &cur->plug_list, plug_list) {
776 if (rbio_can_merge(pending, rbio)) {
777 merge_rbio(pending, rbio);
778 spin_unlock(&cur->bio_list_lock);
779 freeit = rbio;
780 ret = 1;
781 goto out;
782 }
783 }
784
785 /*
786 * No merging, put us on the tail of the plug list, our rbio
787 * will be started with the currently running rbio unlocks
788 */
789 list_add_tail(&rbio->plug_list, &cur->plug_list);
790 spin_unlock(&cur->bio_list_lock);
791 ret = 1;
792 goto out;
793 }
794 lockit:
795 refcount_inc(&rbio->refs);
796 list_add(&rbio->hash_list, &h->hash_list);
797 out:
798 spin_unlock(&h->lock);
799 if (cache_drop)
800 remove_rbio_from_cache(cache_drop);
801 if (freeit)
802 free_raid_bio(freeit);
803 return ret;
804 }
805
806 static void recover_rbio_work_locked(struct work_struct *work);
807
808 /*
809 * called as rmw or parity rebuild is completed. If the plug list has more
810 * rbios waiting for this stripe, the next one on the list will be started
811 */
812 static noinline void unlock_stripe(struct btrfs_raid_bio *rbio)
813 {
814 int bucket;
815 struct btrfs_stripe_hash *h;
816 int keep_cache = 0;
817
818 bucket = rbio_bucket(rbio);
819 h = rbio->bioc->fs_info->stripe_hash_table->table + bucket;
820
821 if (list_empty(&rbio->plug_list))
822 cache_rbio(rbio);
823
824 spin_lock(&h->lock);
825 spin_lock(&rbio->bio_list_lock);
826
827 if (!list_empty(&rbio->hash_list)) {
828 /*
829 * if we're still cached and there is no other IO
830 * to perform, just leave this rbio here for others
831 * to steal from later
832 */
833 if (list_empty(&rbio->plug_list) &&
834 test_bit(RBIO_CACHE_BIT, &rbio->flags)) {
835 keep_cache = 1;
836 clear_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
837 BUG_ON(!bio_list_empty(&rbio->bio_list));
838 goto done;
839 }
840
841 list_del_init(&rbio->hash_list);
842 refcount_dec(&rbio->refs);
843
844 /*
845 * we use the plug list to hold all the rbios
846 * waiting for the chance to lock this stripe.
847 * hand the lock over to one of them.
848 */
849 if (!list_empty(&rbio->plug_list)) {
850 struct btrfs_raid_bio *next;
851 struct list_head *head = rbio->plug_list.next;
852
853 next = list_entry(head, struct btrfs_raid_bio,
854 plug_list);
855
856 list_del_init(&rbio->plug_list);
857
858 list_add(&next->hash_list, &h->hash_list);
859 refcount_inc(&next->refs);
860 spin_unlock(&rbio->bio_list_lock);
861 spin_unlock(&h->lock);
862
863 if (next->operation == BTRFS_RBIO_READ_REBUILD) {
864 start_async_work(next, recover_rbio_work_locked);
865 } else if (next->operation == BTRFS_RBIO_WRITE) {
866 steal_rbio(rbio, next);
867 start_async_work(next, rmw_rbio_work_locked);
868 } else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) {
869 steal_rbio(rbio, next);
870 start_async_work(next, scrub_rbio_work_locked);
871 }
872
873 goto done_nolock;
874 }
875 }
876 done:
877 spin_unlock(&rbio->bio_list_lock);
878 spin_unlock(&h->lock);
879
880 done_nolock:
881 if (!keep_cache)
882 remove_rbio_from_cache(rbio);
883 }
884
885 static void rbio_endio_bio_list(struct bio *cur, blk_status_t err)
886 {
887 struct bio *next;
888
889 while (cur) {
890 next = cur->bi_next;
891 cur->bi_next = NULL;
892 cur->bi_status = err;
893 bio_endio(cur);
894 cur = next;
895 }
896 }
897
898 /*
899 * this frees the rbio and runs through all the bios in the
900 * bio_list and calls end_io on them
901 */
902 static void rbio_orig_end_io(struct btrfs_raid_bio *rbio, blk_status_t err)
903 {
904 struct bio *cur = bio_list_get(&rbio->bio_list);
905 struct bio *extra;
906
907 kfree(rbio->csum_buf);
908 bitmap_free(rbio->csum_bitmap);
909 rbio->csum_buf = NULL;
910 rbio->csum_bitmap = NULL;
911
912 /*
913 * Clear the data bitmap, as the rbio may be cached for later usage.
914 * do this before before unlock_stripe() so there will be no new bio
915 * for this bio.
916 */
917 bitmap_clear(&rbio->dbitmap, 0, rbio->stripe_nsectors);
918
919 /*
920 * At this moment, rbio->bio_list is empty, however since rbio does not
921 * always have RBIO_RMW_LOCKED_BIT set and rbio is still linked on the
922 * hash list, rbio may be merged with others so that rbio->bio_list
923 * becomes non-empty.
924 * Once unlock_stripe() is done, rbio->bio_list will not be updated any
925 * more and we can call bio_endio() on all queued bios.
926 */
927 unlock_stripe(rbio);
928 extra = bio_list_get(&rbio->bio_list);
929 free_raid_bio(rbio);
930
931 rbio_endio_bio_list(cur, err);
932 if (extra)
933 rbio_endio_bio_list(extra, err);
934 }
935
936 /*
937 * Get a sector pointer specified by its @stripe_nr and @sector_nr.
938 *
939 * @rbio: The raid bio
940 * @stripe_nr: Stripe number, valid range [0, real_stripe)
941 * @sector_nr: Sector number inside the stripe,
942 * valid range [0, stripe_nsectors)
943 * @bio_list_only: Whether to use sectors inside the bio list only.
944 *
945 * The read/modify/write code wants to reuse the original bio page as much
946 * as possible, and only use stripe_sectors as fallback.
947 */
948 static struct sector_ptr *sector_in_rbio(struct btrfs_raid_bio *rbio,
949 int stripe_nr, int sector_nr,
950 bool bio_list_only)
951 {
952 struct sector_ptr *sector;
953 int index;
954
955 ASSERT_RBIO_STRIPE(stripe_nr >= 0 && stripe_nr < rbio->real_stripes,
956 rbio, stripe_nr);
957 ASSERT_RBIO_SECTOR(sector_nr >= 0 && sector_nr < rbio->stripe_nsectors,
958 rbio, sector_nr);
959
960 index = stripe_nr * rbio->stripe_nsectors + sector_nr;
961 ASSERT(index >= 0 && index < rbio->nr_sectors);
962
963 spin_lock(&rbio->bio_list_lock);
964 sector = &rbio->bio_sectors[index];
965 if (sector->page || bio_list_only) {
966 /* Don't return sector without a valid page pointer */
967 if (!sector->page)
968 sector = NULL;
969 spin_unlock(&rbio->bio_list_lock);
970 return sector;
971 }
972 spin_unlock(&rbio->bio_list_lock);
973
974 return &rbio->stripe_sectors[index];
975 }
976
977 /*
978 * allocation and initial setup for the btrfs_raid_bio. Not
979 * this does not allocate any pages for rbio->pages.
980 */
981 static struct btrfs_raid_bio *alloc_rbio(struct btrfs_fs_info *fs_info,
982 struct btrfs_io_context *bioc)
983 {
984 const unsigned int real_stripes = bioc->num_stripes - bioc->replace_nr_stripes;
985 const unsigned int stripe_npages = BTRFS_STRIPE_LEN >> PAGE_SHIFT;
986 const unsigned int num_pages = stripe_npages * real_stripes;
987 const unsigned int stripe_nsectors =
988 BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits;
989 const unsigned int num_sectors = stripe_nsectors * real_stripes;
990 struct btrfs_raid_bio *rbio;
991
992 /* PAGE_SIZE must also be aligned to sectorsize for subpage support */
993 ASSERT(IS_ALIGNED(PAGE_SIZE, fs_info->sectorsize));
994 /*
995 * Our current stripe len should be fixed to 64k thus stripe_nsectors
996 * (at most 16) should be no larger than BITS_PER_LONG.
997 */
998 ASSERT(stripe_nsectors <= BITS_PER_LONG);
999
1000 /*
1001 * Real stripes must be between 2 (2 disks RAID5, aka RAID1) and 256
1002 * (limited by u8).
1003 */
1004 ASSERT(real_stripes >= 2);
1005 ASSERT(real_stripes <= U8_MAX);
1006
1007 rbio = kzalloc(sizeof(*rbio), GFP_NOFS);
1008 if (!rbio)
1009 return ERR_PTR(-ENOMEM);
1010 rbio->stripe_pages = kcalloc(num_pages, sizeof(struct page *),
1011 GFP_NOFS);
1012 rbio->bio_sectors = kcalloc(num_sectors, sizeof(struct sector_ptr),
1013 GFP_NOFS);
1014 rbio->stripe_sectors = kcalloc(num_sectors, sizeof(struct sector_ptr),
1015 GFP_NOFS);
1016 rbio->finish_pointers = kcalloc(real_stripes, sizeof(void *), GFP_NOFS);
1017 rbio->error_bitmap = bitmap_zalloc(num_sectors, GFP_NOFS);
1018
1019 if (!rbio->stripe_pages || !rbio->bio_sectors || !rbio->stripe_sectors ||
1020 !rbio->finish_pointers || !rbio->error_bitmap) {
1021 free_raid_bio_pointers(rbio);
1022 kfree(rbio);
1023 return ERR_PTR(-ENOMEM);
1024 }
1025
1026 bio_list_init(&rbio->bio_list);
1027 init_waitqueue_head(&rbio->io_wait);
1028 INIT_LIST_HEAD(&rbio->plug_list);
1029 spin_lock_init(&rbio->bio_list_lock);
1030 INIT_LIST_HEAD(&rbio->stripe_cache);
1031 INIT_LIST_HEAD(&rbio->hash_list);
1032 btrfs_get_bioc(bioc);
1033 rbio->bioc = bioc;
1034 rbio->nr_pages = num_pages;
1035 rbio->nr_sectors = num_sectors;
1036 rbio->real_stripes = real_stripes;
1037 rbio->stripe_npages = stripe_npages;
1038 rbio->stripe_nsectors = stripe_nsectors;
1039 refcount_set(&rbio->refs, 1);
1040 atomic_set(&rbio->stripes_pending, 0);
1041
1042 ASSERT(btrfs_nr_parity_stripes(bioc->map_type));
1043 rbio->nr_data = real_stripes - btrfs_nr_parity_stripes(bioc->map_type);
1044 ASSERT(rbio->nr_data > 0);
1045
1046 return rbio;
1047 }
1048
1049 /* allocate pages for all the stripes in the bio, including parity */
1050 static int alloc_rbio_pages(struct btrfs_raid_bio *rbio)
1051 {
1052 int ret;
1053
1054 ret = btrfs_alloc_page_array(rbio->nr_pages, rbio->stripe_pages, false);
1055 if (ret < 0)
1056 return ret;
1057 /* Mapping all sectors */
1058 index_stripe_sectors(rbio);
1059 return 0;
1060 }
1061
1062 /* only allocate pages for p/q stripes */
1063 static int alloc_rbio_parity_pages(struct btrfs_raid_bio *rbio)
1064 {
1065 const int data_pages = rbio->nr_data * rbio->stripe_npages;
1066 int ret;
1067
1068 ret = btrfs_alloc_page_array(rbio->nr_pages - data_pages,
1069 rbio->stripe_pages + data_pages, false);
1070 if (ret < 0)
1071 return ret;
1072
1073 index_stripe_sectors(rbio);
1074 return 0;
1075 }
1076
1077 /*
1078 * Return the total number of errors found in the vertical stripe of @sector_nr.
1079 *
1080 * @faila and @failb will also be updated to the first and second stripe
1081 * number of the errors.
1082 */
1083 static int get_rbio_veritical_errors(struct btrfs_raid_bio *rbio, int sector_nr,
1084 int *faila, int *failb)
1085 {
1086 int stripe_nr;
1087 int found_errors = 0;
1088
1089 if (faila || failb) {
1090 /*
1091 * Both @faila and @failb should be valid pointers if any of
1092 * them is specified.
1093 */
1094 ASSERT(faila && failb);
1095 *faila = -1;
1096 *failb = -1;
1097 }
1098
1099 for (stripe_nr = 0; stripe_nr < rbio->real_stripes; stripe_nr++) {
1100 int total_sector_nr = stripe_nr * rbio->stripe_nsectors + sector_nr;
1101
1102 if (test_bit(total_sector_nr, rbio->error_bitmap)) {
1103 found_errors++;
1104 if (faila) {
1105 /* Update faila and failb. */
1106 if (*faila < 0)
1107 *faila = stripe_nr;
1108 else if (*failb < 0)
1109 *failb = stripe_nr;
1110 }
1111 }
1112 }
1113 return found_errors;
1114 }
1115
1116 /*
1117 * Add a single sector @sector into our list of bios for IO.
1118 *
1119 * Return 0 if everything went well.
1120 * Return <0 for error.
1121 */
1122 static int rbio_add_io_sector(struct btrfs_raid_bio *rbio,
1123 struct bio_list *bio_list,
1124 struct sector_ptr *sector,
1125 unsigned int stripe_nr,
1126 unsigned int sector_nr,
1127 enum req_op op)
1128 {
1129 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
1130 struct bio *last = bio_list->tail;
1131 int ret;
1132 struct bio *bio;
1133 struct btrfs_io_stripe *stripe;
1134 u64 disk_start;
1135
1136 /*
1137 * Note: here stripe_nr has taken device replace into consideration,
1138 * thus it can be larger than rbio->real_stripe.
1139 * So here we check against bioc->num_stripes, not rbio->real_stripes.
1140 */
1141 ASSERT_RBIO_STRIPE(stripe_nr >= 0 && stripe_nr < rbio->bioc->num_stripes,
1142 rbio, stripe_nr);
1143 ASSERT_RBIO_SECTOR(sector_nr >= 0 && sector_nr < rbio->stripe_nsectors,
1144 rbio, sector_nr);
1145 ASSERT(sector->page);
1146
1147 stripe = &rbio->bioc->stripes[stripe_nr];
1148 disk_start = stripe->physical + sector_nr * sectorsize;
1149
1150 /* if the device is missing, just fail this stripe */
1151 if (!stripe->dev->bdev) {
1152 int found_errors;
1153
1154 set_bit(stripe_nr * rbio->stripe_nsectors + sector_nr,
1155 rbio->error_bitmap);
1156
1157 /* Check if we have reached tolerance early. */
1158 found_errors = get_rbio_veritical_errors(rbio, sector_nr,
1159 NULL, NULL);
1160 if (found_errors > rbio->bioc->max_errors)
1161 return -EIO;
1162 return 0;
1163 }
1164
1165 /* see if we can add this page onto our existing bio */
1166 if (last) {
1167 u64 last_end = last->bi_iter.bi_sector << SECTOR_SHIFT;
1168 last_end += last->bi_iter.bi_size;
1169
1170 /*
1171 * we can't merge these if they are from different
1172 * devices or if they are not contiguous
1173 */
1174 if (last_end == disk_start && !last->bi_status &&
1175 last->bi_bdev == stripe->dev->bdev) {
1176 ret = bio_add_page(last, sector->page, sectorsize,
1177 sector->pgoff);
1178 if (ret == sectorsize)
1179 return 0;
1180 }
1181 }
1182
1183 /* put a new bio on the list */
1184 bio = bio_alloc(stripe->dev->bdev,
1185 max(BTRFS_STRIPE_LEN >> PAGE_SHIFT, 1),
1186 op, GFP_NOFS);
1187 bio->bi_iter.bi_sector = disk_start >> SECTOR_SHIFT;
1188 bio->bi_private = rbio;
1189
1190 __bio_add_page(bio, sector->page, sectorsize, sector->pgoff);
1191 bio_list_add(bio_list, bio);
1192 return 0;
1193 }
1194
1195 static void index_one_bio(struct btrfs_raid_bio *rbio, struct bio *bio)
1196 {
1197 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
1198 struct bio_vec bvec;
1199 struct bvec_iter iter;
1200 u32 offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) -
1201 rbio->bioc->full_stripe_logical;
1202
1203 bio_for_each_segment(bvec, bio, iter) {
1204 u32 bvec_offset;
1205
1206 for (bvec_offset = 0; bvec_offset < bvec.bv_len;
1207 bvec_offset += sectorsize, offset += sectorsize) {
1208 int index = offset / sectorsize;
1209 struct sector_ptr *sector = &rbio->bio_sectors[index];
1210
1211 sector->page = bvec.bv_page;
1212 sector->pgoff = bvec.bv_offset + bvec_offset;
1213 ASSERT(sector->pgoff < PAGE_SIZE);
1214 }
1215 }
1216 }
1217
1218 /*
1219 * helper function to walk our bio list and populate the bio_pages array with
1220 * the result. This seems expensive, but it is faster than constantly
1221 * searching through the bio list as we setup the IO in finish_rmw or stripe
1222 * reconstruction.
1223 *
1224 * This must be called before you trust the answers from page_in_rbio
1225 */
1226 static void index_rbio_pages(struct btrfs_raid_bio *rbio)
1227 {
1228 struct bio *bio;
1229
1230 spin_lock(&rbio->bio_list_lock);
1231 bio_list_for_each(bio, &rbio->bio_list)
1232 index_one_bio(rbio, bio);
1233
1234 spin_unlock(&rbio->bio_list_lock);
1235 }
1236
1237 static void bio_get_trace_info(struct btrfs_raid_bio *rbio, struct bio *bio,
1238 struct raid56_bio_trace_info *trace_info)
1239 {
1240 const struct btrfs_io_context *bioc = rbio->bioc;
1241 int i;
1242
1243 ASSERT(bioc);
1244
1245 /* We rely on bio->bi_bdev to find the stripe number. */
1246 if (!bio->bi_bdev)
1247 goto not_found;
1248
1249 for (i = 0; i < bioc->num_stripes; i++) {
1250 if (bio->bi_bdev != bioc->stripes[i].dev->bdev)
1251 continue;
1252 trace_info->stripe_nr = i;
1253 trace_info->devid = bioc->stripes[i].dev->devid;
1254 trace_info->offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) -
1255 bioc->stripes[i].physical;
1256 return;
1257 }
1258
1259 not_found:
1260 trace_info->devid = -1;
1261 trace_info->offset = -1;
1262 trace_info->stripe_nr = -1;
1263 }
1264
1265 static inline void bio_list_put(struct bio_list *bio_list)
1266 {
1267 struct bio *bio;
1268
1269 while ((bio = bio_list_pop(bio_list)))
1270 bio_put(bio);
1271 }
1272
1273 static void assert_rbio(struct btrfs_raid_bio *rbio)
1274 {
1275 if (!IS_ENABLED(CONFIG_BTRFS_DEBUG) ||
1276 !IS_ENABLED(CONFIG_BTRFS_ASSERT))
1277 return;
1278
1279 /*
1280 * At least two stripes (2 disks RAID5), and since real_stripes is U8,
1281 * we won't go beyond 256 disks anyway.
1282 */
1283 ASSERT_RBIO(rbio->real_stripes >= 2, rbio);
1284 ASSERT_RBIO(rbio->nr_data > 0, rbio);
1285
1286 /*
1287 * This is another check to make sure nr data stripes is smaller
1288 * than total stripes.
1289 */
1290 ASSERT_RBIO(rbio->nr_data < rbio->real_stripes, rbio);
1291 }
1292
1293 /* Generate PQ for one vertical stripe. */
1294 static void generate_pq_vertical(struct btrfs_raid_bio *rbio, int sectornr)
1295 {
1296 void **pointers = rbio->finish_pointers;
1297 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
1298 struct sector_ptr *sector;
1299 int stripe;
1300 const bool has_qstripe = rbio->bioc->map_type & BTRFS_BLOCK_GROUP_RAID6;
1301
1302 /* First collect one sector from each data stripe */
1303 for (stripe = 0; stripe < rbio->nr_data; stripe++) {
1304 sector = sector_in_rbio(rbio, stripe, sectornr, 0);
1305 pointers[stripe] = kmap_local_page(sector->page) +
1306 sector->pgoff;
1307 }
1308
1309 /* Then add the parity stripe */
1310 sector = rbio_pstripe_sector(rbio, sectornr);
1311 sector->uptodate = 1;
1312 pointers[stripe++] = kmap_local_page(sector->page) + sector->pgoff;
1313
1314 if (has_qstripe) {
1315 /*
1316 * RAID6, add the qstripe and call the library function
1317 * to fill in our p/q
1318 */
1319 sector = rbio_qstripe_sector(rbio, sectornr);
1320 sector->uptodate = 1;
1321 pointers[stripe++] = kmap_local_page(sector->page) +
1322 sector->pgoff;
1323
1324 assert_rbio(rbio);
1325 raid6_call.gen_syndrome(rbio->real_stripes, sectorsize,
1326 pointers);
1327 } else {
1328 /* raid5 */
1329 memcpy(pointers[rbio->nr_data], pointers[0], sectorsize);
1330 run_xor(pointers + 1, rbio->nr_data - 1, sectorsize);
1331 }
1332 for (stripe = stripe - 1; stripe >= 0; stripe--)
1333 kunmap_local(pointers[stripe]);
1334 }
1335
1336 static int rmw_assemble_write_bios(struct btrfs_raid_bio *rbio,
1337 struct bio_list *bio_list)
1338 {
1339 /* The total sector number inside the full stripe. */
1340 int total_sector_nr;
1341 int sectornr;
1342 int stripe;
1343 int ret;
1344
1345 ASSERT(bio_list_size(bio_list) == 0);
1346
1347 /* We should have at least one data sector. */
1348 ASSERT(bitmap_weight(&rbio->dbitmap, rbio->stripe_nsectors));
1349
1350 /*
1351 * Reset errors, as we may have errors inherited from from degraded
1352 * write.
1353 */
1354 bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors);
1355
1356 /*
1357 * Start assembly. Make bios for everything from the higher layers (the
1358 * bio_list in our rbio) and our P/Q. Ignore everything else.
1359 */
1360 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
1361 total_sector_nr++) {
1362 struct sector_ptr *sector;
1363
1364 stripe = total_sector_nr / rbio->stripe_nsectors;
1365 sectornr = total_sector_nr % rbio->stripe_nsectors;
1366
1367 /* This vertical stripe has no data, skip it. */
1368 if (!test_bit(sectornr, &rbio->dbitmap))
1369 continue;
1370
1371 if (stripe < rbio->nr_data) {
1372 sector = sector_in_rbio(rbio, stripe, sectornr, 1);
1373 if (!sector)
1374 continue;
1375 } else {
1376 sector = rbio_stripe_sector(rbio, stripe, sectornr);
1377 }
1378
1379 ret = rbio_add_io_sector(rbio, bio_list, sector, stripe,
1380 sectornr, REQ_OP_WRITE);
1381 if (ret)
1382 goto error;
1383 }
1384
1385 if (likely(!rbio->bioc->replace_nr_stripes))
1386 return 0;
1387
1388 /*
1389 * Make a copy for the replace target device.
1390 *
1391 * Thus the source stripe number (in replace_stripe_src) should be valid.
1392 */
1393 ASSERT(rbio->bioc->replace_stripe_src >= 0);
1394
1395 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
1396 total_sector_nr++) {
1397 struct sector_ptr *sector;
1398
1399 stripe = total_sector_nr / rbio->stripe_nsectors;
1400 sectornr = total_sector_nr % rbio->stripe_nsectors;
1401
1402 /*
1403 * For RAID56, there is only one device that can be replaced,
1404 * and replace_stripe_src[0] indicates the stripe number we
1405 * need to copy from.
1406 */
1407 if (stripe != rbio->bioc->replace_stripe_src) {
1408 /*
1409 * We can skip the whole stripe completely, note
1410 * total_sector_nr will be increased by one anyway.
1411 */
1412 ASSERT(sectornr == 0);
1413 total_sector_nr += rbio->stripe_nsectors - 1;
1414 continue;
1415 }
1416
1417 /* This vertical stripe has no data, skip it. */
1418 if (!test_bit(sectornr, &rbio->dbitmap))
1419 continue;
1420
1421 if (stripe < rbio->nr_data) {
1422 sector = sector_in_rbio(rbio, stripe, sectornr, 1);
1423 if (!sector)
1424 continue;
1425 } else {
1426 sector = rbio_stripe_sector(rbio, stripe, sectornr);
1427 }
1428
1429 ret = rbio_add_io_sector(rbio, bio_list, sector,
1430 rbio->real_stripes,
1431 sectornr, REQ_OP_WRITE);
1432 if (ret)
1433 goto error;
1434 }
1435
1436 return 0;
1437 error:
1438 bio_list_put(bio_list);
1439 return -EIO;
1440 }
1441
1442 static void set_rbio_range_error(struct btrfs_raid_bio *rbio, struct bio *bio)
1443 {
1444 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1445 u32 offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) -
1446 rbio->bioc->full_stripe_logical;
1447 int total_nr_sector = offset >> fs_info->sectorsize_bits;
1448
1449 ASSERT(total_nr_sector < rbio->nr_data * rbio->stripe_nsectors);
1450
1451 bitmap_set(rbio->error_bitmap, total_nr_sector,
1452 bio->bi_iter.bi_size >> fs_info->sectorsize_bits);
1453
1454 /*
1455 * Special handling for raid56_alloc_missing_rbio() used by
1456 * scrub/replace. Unlike call path in raid56_parity_recover(), they
1457 * pass an empty bio here. Thus we have to find out the missing device
1458 * and mark the stripe error instead.
1459 */
1460 if (bio->bi_iter.bi_size == 0) {
1461 bool found_missing = false;
1462 int stripe_nr;
1463
1464 for (stripe_nr = 0; stripe_nr < rbio->real_stripes; stripe_nr++) {
1465 if (!rbio->bioc->stripes[stripe_nr].dev->bdev) {
1466 found_missing = true;
1467 bitmap_set(rbio->error_bitmap,
1468 stripe_nr * rbio->stripe_nsectors,
1469 rbio->stripe_nsectors);
1470 }
1471 }
1472 ASSERT(found_missing);
1473 }
1474 }
1475
1476 /*
1477 * For subpage case, we can no longer set page Up-to-date directly for
1478 * stripe_pages[], thus we need to locate the sector.
1479 */
1480 static struct sector_ptr *find_stripe_sector(struct btrfs_raid_bio *rbio,
1481 struct page *page,
1482 unsigned int pgoff)
1483 {
1484 int i;
1485
1486 for (i = 0; i < rbio->nr_sectors; i++) {
1487 struct sector_ptr *sector = &rbio->stripe_sectors[i];
1488
1489 if (sector->page == page && sector->pgoff == pgoff)
1490 return sector;
1491 }
1492 return NULL;
1493 }
1494
1495 /*
1496 * this sets each page in the bio uptodate. It should only be used on private
1497 * rbio pages, nothing that comes in from the higher layers
1498 */
1499 static void set_bio_pages_uptodate(struct btrfs_raid_bio *rbio, struct bio *bio)
1500 {
1501 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
1502 struct bio_vec *bvec;
1503 struct bvec_iter_all iter_all;
1504
1505 ASSERT(!bio_flagged(bio, BIO_CLONED));
1506
1507 bio_for_each_segment_all(bvec, bio, iter_all) {
1508 struct sector_ptr *sector;
1509 int pgoff;
1510
1511 for (pgoff = bvec->bv_offset; pgoff - bvec->bv_offset < bvec->bv_len;
1512 pgoff += sectorsize) {
1513 sector = find_stripe_sector(rbio, bvec->bv_page, pgoff);
1514 ASSERT(sector);
1515 if (sector)
1516 sector->uptodate = 1;
1517 }
1518 }
1519 }
1520
1521 static int get_bio_sector_nr(struct btrfs_raid_bio *rbio, struct bio *bio)
1522 {
1523 struct bio_vec *bv = bio_first_bvec_all(bio);
1524 int i;
1525
1526 for (i = 0; i < rbio->nr_sectors; i++) {
1527 struct sector_ptr *sector;
1528
1529 sector = &rbio->stripe_sectors[i];
1530 if (sector->page == bv->bv_page && sector->pgoff == bv->bv_offset)
1531 break;
1532 sector = &rbio->bio_sectors[i];
1533 if (sector->page == bv->bv_page && sector->pgoff == bv->bv_offset)
1534 break;
1535 }
1536 ASSERT(i < rbio->nr_sectors);
1537 return i;
1538 }
1539
1540 static void rbio_update_error_bitmap(struct btrfs_raid_bio *rbio, struct bio *bio)
1541 {
1542 int total_sector_nr = get_bio_sector_nr(rbio, bio);
1543 u32 bio_size = 0;
1544 struct bio_vec *bvec;
1545 int i;
1546
1547 bio_for_each_bvec_all(bvec, bio, i)
1548 bio_size += bvec->bv_len;
1549
1550 /*
1551 * Since we can have multiple bios touching the error_bitmap, we cannot
1552 * call bitmap_set() without protection.
1553 *
1554 * Instead use set_bit() for each bit, as set_bit() itself is atomic.
1555 */
1556 for (i = total_sector_nr; i < total_sector_nr +
1557 (bio_size >> rbio->bioc->fs_info->sectorsize_bits); i++)
1558 set_bit(i, rbio->error_bitmap);
1559 }
1560
1561 /* Verify the data sectors at read time. */
1562 static void verify_bio_data_sectors(struct btrfs_raid_bio *rbio,
1563 struct bio *bio)
1564 {
1565 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1566 int total_sector_nr = get_bio_sector_nr(rbio, bio);
1567 struct bio_vec *bvec;
1568 struct bvec_iter_all iter_all;
1569
1570 /* No data csum for the whole stripe, no need to verify. */
1571 if (!rbio->csum_bitmap || !rbio->csum_buf)
1572 return;
1573
1574 /* P/Q stripes, they have no data csum to verify against. */
1575 if (total_sector_nr >= rbio->nr_data * rbio->stripe_nsectors)
1576 return;
1577
1578 bio_for_each_segment_all(bvec, bio, iter_all) {
1579 int bv_offset;
1580
1581 for (bv_offset = bvec->bv_offset;
1582 bv_offset < bvec->bv_offset + bvec->bv_len;
1583 bv_offset += fs_info->sectorsize, total_sector_nr++) {
1584 u8 csum_buf[BTRFS_CSUM_SIZE];
1585 u8 *expected_csum = rbio->csum_buf +
1586 total_sector_nr * fs_info->csum_size;
1587 int ret;
1588
1589 /* No csum for this sector, skip to the next sector. */
1590 if (!test_bit(total_sector_nr, rbio->csum_bitmap))
1591 continue;
1592
1593 ret = btrfs_check_sector_csum(fs_info, bvec->bv_page,
1594 bv_offset, csum_buf, expected_csum);
1595 if (ret < 0)
1596 set_bit(total_sector_nr, rbio->error_bitmap);
1597 }
1598 }
1599 }
1600
1601 static void raid_wait_read_end_io(struct bio *bio)
1602 {
1603 struct btrfs_raid_bio *rbio = bio->bi_private;
1604
1605 if (bio->bi_status) {
1606 rbio_update_error_bitmap(rbio, bio);
1607 } else {
1608 set_bio_pages_uptodate(rbio, bio);
1609 verify_bio_data_sectors(rbio, bio);
1610 }
1611
1612 bio_put(bio);
1613 if (atomic_dec_and_test(&rbio->stripes_pending))
1614 wake_up(&rbio->io_wait);
1615 }
1616
1617 static void submit_read_wait_bio_list(struct btrfs_raid_bio *rbio,
1618 struct bio_list *bio_list)
1619 {
1620 struct bio *bio;
1621
1622 atomic_set(&rbio->stripes_pending, bio_list_size(bio_list));
1623 while ((bio = bio_list_pop(bio_list))) {
1624 bio->bi_end_io = raid_wait_read_end_io;
1625
1626 if (trace_raid56_read_enabled()) {
1627 struct raid56_bio_trace_info trace_info = { 0 };
1628
1629 bio_get_trace_info(rbio, bio, &trace_info);
1630 trace_raid56_read(rbio, bio, &trace_info);
1631 }
1632 submit_bio(bio);
1633 }
1634
1635 wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0);
1636 }
1637
1638 static int alloc_rbio_data_pages(struct btrfs_raid_bio *rbio)
1639 {
1640 const int data_pages = rbio->nr_data * rbio->stripe_npages;
1641 int ret;
1642
1643 ret = btrfs_alloc_page_array(data_pages, rbio->stripe_pages, false);
1644 if (ret < 0)
1645 return ret;
1646
1647 index_stripe_sectors(rbio);
1648 return 0;
1649 }
1650
1651 /*
1652 * We use plugging call backs to collect full stripes.
1653 * Any time we get a partial stripe write while plugged
1654 * we collect it into a list. When the unplug comes down,
1655 * we sort the list by logical block number and merge
1656 * everything we can into the same rbios
1657 */
1658 struct btrfs_plug_cb {
1659 struct blk_plug_cb cb;
1660 struct btrfs_fs_info *info;
1661 struct list_head rbio_list;
1662 };
1663
1664 /*
1665 * rbios on the plug list are sorted for easier merging.
1666 */
1667 static int plug_cmp(void *priv, const struct list_head *a,
1668 const struct list_head *b)
1669 {
1670 const struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio,
1671 plug_list);
1672 const struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio,
1673 plug_list);
1674 u64 a_sector = ra->bio_list.head->bi_iter.bi_sector;
1675 u64 b_sector = rb->bio_list.head->bi_iter.bi_sector;
1676
1677 if (a_sector < b_sector)
1678 return -1;
1679 if (a_sector > b_sector)
1680 return 1;
1681 return 0;
1682 }
1683
1684 static void raid_unplug(struct blk_plug_cb *cb, bool from_schedule)
1685 {
1686 struct btrfs_plug_cb *plug = container_of(cb, struct btrfs_plug_cb, cb);
1687 struct btrfs_raid_bio *cur;
1688 struct btrfs_raid_bio *last = NULL;
1689
1690 list_sort(NULL, &plug->rbio_list, plug_cmp);
1691
1692 while (!list_empty(&plug->rbio_list)) {
1693 cur = list_entry(plug->rbio_list.next,
1694 struct btrfs_raid_bio, plug_list);
1695 list_del_init(&cur->plug_list);
1696
1697 if (rbio_is_full(cur)) {
1698 /* We have a full stripe, queue it down. */
1699 start_async_work(cur, rmw_rbio_work);
1700 continue;
1701 }
1702 if (last) {
1703 if (rbio_can_merge(last, cur)) {
1704 merge_rbio(last, cur);
1705 free_raid_bio(cur);
1706 continue;
1707 }
1708 start_async_work(last, rmw_rbio_work);
1709 }
1710 last = cur;
1711 }
1712 if (last)
1713 start_async_work(last, rmw_rbio_work);
1714 kfree(plug);
1715 }
1716
1717 /* Add the original bio into rbio->bio_list, and update rbio::dbitmap. */
1718 static void rbio_add_bio(struct btrfs_raid_bio *rbio, struct bio *orig_bio)
1719 {
1720 const struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1721 const u64 orig_logical = orig_bio->bi_iter.bi_sector << SECTOR_SHIFT;
1722 const u64 full_stripe_start = rbio->bioc->full_stripe_logical;
1723 const u32 orig_len = orig_bio->bi_iter.bi_size;
1724 const u32 sectorsize = fs_info->sectorsize;
1725 u64 cur_logical;
1726
1727 ASSERT_RBIO_LOGICAL(orig_logical >= full_stripe_start &&
1728 orig_logical + orig_len <= full_stripe_start +
1729 rbio->nr_data * BTRFS_STRIPE_LEN,
1730 rbio, orig_logical);
1731
1732 bio_list_add(&rbio->bio_list, orig_bio);
1733 rbio->bio_list_bytes += orig_bio->bi_iter.bi_size;
1734
1735 /* Update the dbitmap. */
1736 for (cur_logical = orig_logical; cur_logical < orig_logical + orig_len;
1737 cur_logical += sectorsize) {
1738 int bit = ((u32)(cur_logical - full_stripe_start) >>
1739 fs_info->sectorsize_bits) % rbio->stripe_nsectors;
1740
1741 set_bit(bit, &rbio->dbitmap);
1742 }
1743 }
1744
1745 /*
1746 * our main entry point for writes from the rest of the FS.
1747 */
1748 void raid56_parity_write(struct bio *bio, struct btrfs_io_context *bioc)
1749 {
1750 struct btrfs_fs_info *fs_info = bioc->fs_info;
1751 struct btrfs_raid_bio *rbio;
1752 struct btrfs_plug_cb *plug = NULL;
1753 struct blk_plug_cb *cb;
1754
1755 rbio = alloc_rbio(fs_info, bioc);
1756 if (IS_ERR(rbio)) {
1757 bio->bi_status = errno_to_blk_status(PTR_ERR(rbio));
1758 bio_endio(bio);
1759 return;
1760 }
1761 rbio->operation = BTRFS_RBIO_WRITE;
1762 rbio_add_bio(rbio, bio);
1763
1764 /*
1765 * Don't plug on full rbios, just get them out the door
1766 * as quickly as we can
1767 */
1768 if (!rbio_is_full(rbio)) {
1769 cb = blk_check_plugged(raid_unplug, fs_info, sizeof(*plug));
1770 if (cb) {
1771 plug = container_of(cb, struct btrfs_plug_cb, cb);
1772 if (!plug->info) {
1773 plug->info = fs_info;
1774 INIT_LIST_HEAD(&plug->rbio_list);
1775 }
1776 list_add_tail(&rbio->plug_list, &plug->rbio_list);
1777 return;
1778 }
1779 }
1780
1781 /*
1782 * Either we don't have any existing plug, or we're doing a full stripe,
1783 * queue the rmw work now.
1784 */
1785 start_async_work(rbio, rmw_rbio_work);
1786 }
1787
1788 static int verify_one_sector(struct btrfs_raid_bio *rbio,
1789 int stripe_nr, int sector_nr)
1790 {
1791 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1792 struct sector_ptr *sector;
1793 u8 csum_buf[BTRFS_CSUM_SIZE];
1794 u8 *csum_expected;
1795 int ret;
1796
1797 if (!rbio->csum_bitmap || !rbio->csum_buf)
1798 return 0;
1799
1800 /* No way to verify P/Q as they are not covered by data csum. */
1801 if (stripe_nr >= rbio->nr_data)
1802 return 0;
1803 /*
1804 * If we're rebuilding a read, we have to use pages from the
1805 * bio list if possible.
1806 */
1807 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
1808 sector = sector_in_rbio(rbio, stripe_nr, sector_nr, 0);
1809 } else {
1810 sector = rbio_stripe_sector(rbio, stripe_nr, sector_nr);
1811 }
1812
1813 ASSERT(sector->page);
1814
1815 csum_expected = rbio->csum_buf +
1816 (stripe_nr * rbio->stripe_nsectors + sector_nr) *
1817 fs_info->csum_size;
1818 ret = btrfs_check_sector_csum(fs_info, sector->page, sector->pgoff,
1819 csum_buf, csum_expected);
1820 return ret;
1821 }
1822
1823 /*
1824 * Recover a vertical stripe specified by @sector_nr.
1825 * @*pointers are the pre-allocated pointers by the caller, so we don't
1826 * need to allocate/free the pointers again and again.
1827 */
1828 static int recover_vertical(struct btrfs_raid_bio *rbio, int sector_nr,
1829 void **pointers, void **unmap_array)
1830 {
1831 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
1832 struct sector_ptr *sector;
1833 const u32 sectorsize = fs_info->sectorsize;
1834 int found_errors;
1835 int faila;
1836 int failb;
1837 int stripe_nr;
1838 int ret = 0;
1839
1840 /*
1841 * Now we just use bitmap to mark the horizontal stripes in
1842 * which we have data when doing parity scrub.
1843 */
1844 if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB &&
1845 !test_bit(sector_nr, &rbio->dbitmap))
1846 return 0;
1847
1848 found_errors = get_rbio_veritical_errors(rbio, sector_nr, &faila,
1849 &failb);
1850 /*
1851 * No errors in the vertical stripe, skip it. Can happen for recovery
1852 * which only part of a stripe failed csum check.
1853 */
1854 if (!found_errors)
1855 return 0;
1856
1857 if (found_errors > rbio->bioc->max_errors)
1858 return -EIO;
1859
1860 /*
1861 * Setup our array of pointers with sectors from each stripe
1862 *
1863 * NOTE: store a duplicate array of pointers to preserve the
1864 * pointer order.
1865 */
1866 for (stripe_nr = 0; stripe_nr < rbio->real_stripes; stripe_nr++) {
1867 /*
1868 * If we're rebuilding a read, we have to use pages from the
1869 * bio list if possible.
1870 */
1871 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
1872 sector = sector_in_rbio(rbio, stripe_nr, sector_nr, 0);
1873 } else {
1874 sector = rbio_stripe_sector(rbio, stripe_nr, sector_nr);
1875 }
1876 ASSERT(sector->page);
1877 pointers[stripe_nr] = kmap_local_page(sector->page) +
1878 sector->pgoff;
1879 unmap_array[stripe_nr] = pointers[stripe_nr];
1880 }
1881
1882 /* All raid6 handling here */
1883 if (rbio->bioc->map_type & BTRFS_BLOCK_GROUP_RAID6) {
1884 /* Single failure, rebuild from parity raid5 style */
1885 if (failb < 0) {
1886 if (faila == rbio->nr_data)
1887 /*
1888 * Just the P stripe has failed, without
1889 * a bad data or Q stripe.
1890 * We have nothing to do, just skip the
1891 * recovery for this stripe.
1892 */
1893 goto cleanup;
1894 /*
1895 * a single failure in raid6 is rebuilt
1896 * in the pstripe code below
1897 */
1898 goto pstripe;
1899 }
1900
1901 /*
1902 * If the q stripe is failed, do a pstripe reconstruction from
1903 * the xors.
1904 * If both the q stripe and the P stripe are failed, we're
1905 * here due to a crc mismatch and we can't give them the
1906 * data they want.
1907 */
1908 if (failb == rbio->real_stripes - 1) {
1909 if (faila == rbio->real_stripes - 2)
1910 /*
1911 * Only P and Q are corrupted.
1912 * We only care about data stripes recovery,
1913 * can skip this vertical stripe.
1914 */
1915 goto cleanup;
1916 /*
1917 * Otherwise we have one bad data stripe and
1918 * a good P stripe. raid5!
1919 */
1920 goto pstripe;
1921 }
1922
1923 if (failb == rbio->real_stripes - 2) {
1924 raid6_datap_recov(rbio->real_stripes, sectorsize,
1925 faila, pointers);
1926 } else {
1927 raid6_2data_recov(rbio->real_stripes, sectorsize,
1928 faila, failb, pointers);
1929 }
1930 } else {
1931 void *p;
1932
1933 /* Rebuild from P stripe here (raid5 or raid6). */
1934 ASSERT(failb == -1);
1935 pstripe:
1936 /* Copy parity block into failed block to start with */
1937 memcpy(pointers[faila], pointers[rbio->nr_data], sectorsize);
1938
1939 /* Rearrange the pointer array */
1940 p = pointers[faila];
1941 for (stripe_nr = faila; stripe_nr < rbio->nr_data - 1;
1942 stripe_nr++)
1943 pointers[stripe_nr] = pointers[stripe_nr + 1];
1944 pointers[rbio->nr_data - 1] = p;
1945
1946 /* Xor in the rest */
1947 run_xor(pointers, rbio->nr_data - 1, sectorsize);
1948
1949 }
1950
1951 /*
1952 * No matter if this is a RMW or recovery, we should have all
1953 * failed sectors repaired in the vertical stripe, thus they are now
1954 * uptodate.
1955 * Especially if we determine to cache the rbio, we need to
1956 * have at least all data sectors uptodate.
1957 *
1958 * If possible, also check if the repaired sector matches its data
1959 * checksum.
1960 */
1961 if (faila >= 0) {
1962 ret = verify_one_sector(rbio, faila, sector_nr);
1963 if (ret < 0)
1964 goto cleanup;
1965
1966 sector = rbio_stripe_sector(rbio, faila, sector_nr);
1967 sector->uptodate = 1;
1968 }
1969 if (failb >= 0) {
1970 ret = verify_one_sector(rbio, failb, sector_nr);
1971 if (ret < 0)
1972 goto cleanup;
1973
1974 sector = rbio_stripe_sector(rbio, failb, sector_nr);
1975 sector->uptodate = 1;
1976 }
1977
1978 cleanup:
1979 for (stripe_nr = rbio->real_stripes - 1; stripe_nr >= 0; stripe_nr--)
1980 kunmap_local(unmap_array[stripe_nr]);
1981 return ret;
1982 }
1983
1984 static int recover_sectors(struct btrfs_raid_bio *rbio)
1985 {
1986 void **pointers = NULL;
1987 void **unmap_array = NULL;
1988 int sectornr;
1989 int ret = 0;
1990
1991 /*
1992 * @pointers array stores the pointer for each sector.
1993 *
1994 * @unmap_array stores copy of pointers that does not get reordered
1995 * during reconstruction so that kunmap_local works.
1996 */
1997 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
1998 unmap_array = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
1999 if (!pointers || !unmap_array) {
2000 ret = -ENOMEM;
2001 goto out;
2002 }
2003
2004 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
2005 spin_lock(&rbio->bio_list_lock);
2006 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
2007 spin_unlock(&rbio->bio_list_lock);
2008 }
2009
2010 index_rbio_pages(rbio);
2011
2012 for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++) {
2013 ret = recover_vertical(rbio, sectornr, pointers, unmap_array);
2014 if (ret < 0)
2015 break;
2016 }
2017
2018 out:
2019 kfree(pointers);
2020 kfree(unmap_array);
2021 return ret;
2022 }
2023
2024 static void recover_rbio(struct btrfs_raid_bio *rbio)
2025 {
2026 struct bio_list bio_list = BIO_EMPTY_LIST;
2027 int total_sector_nr;
2028 int ret = 0;
2029
2030 /*
2031 * Either we're doing recover for a read failure or degraded write,
2032 * caller should have set error bitmap correctly.
2033 */
2034 ASSERT(bitmap_weight(rbio->error_bitmap, rbio->nr_sectors));
2035
2036 /* For recovery, we need to read all sectors including P/Q. */
2037 ret = alloc_rbio_pages(rbio);
2038 if (ret < 0)
2039 goto out;
2040
2041 index_rbio_pages(rbio);
2042
2043 /*
2044 * Read everything that hasn't failed. However this time we will
2045 * not trust any cached sector.
2046 * As we may read out some stale data but higher layer is not reading
2047 * that stale part.
2048 *
2049 * So here we always re-read everything in recovery path.
2050 */
2051 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
2052 total_sector_nr++) {
2053 int stripe = total_sector_nr / rbio->stripe_nsectors;
2054 int sectornr = total_sector_nr % rbio->stripe_nsectors;
2055 struct sector_ptr *sector;
2056
2057 /*
2058 * Skip the range which has error. It can be a range which is
2059 * marked error (for csum mismatch), or it can be a missing
2060 * device.
2061 */
2062 if (!rbio->bioc->stripes[stripe].dev->bdev ||
2063 test_bit(total_sector_nr, rbio->error_bitmap)) {
2064 /*
2065 * Also set the error bit for missing device, which
2066 * may not yet have its error bit set.
2067 */
2068 set_bit(total_sector_nr, rbio->error_bitmap);
2069 continue;
2070 }
2071
2072 sector = rbio_stripe_sector(rbio, stripe, sectornr);
2073 ret = rbio_add_io_sector(rbio, &bio_list, sector, stripe,
2074 sectornr, REQ_OP_READ);
2075 if (ret < 0) {
2076 bio_list_put(&bio_list);
2077 goto out;
2078 }
2079 }
2080
2081 submit_read_wait_bio_list(rbio, &bio_list);
2082 ret = recover_sectors(rbio);
2083 out:
2084 rbio_orig_end_io(rbio, errno_to_blk_status(ret));
2085 }
2086
2087 static void recover_rbio_work(struct work_struct *work)
2088 {
2089 struct btrfs_raid_bio *rbio;
2090
2091 rbio = container_of(work, struct btrfs_raid_bio, work);
2092 if (!lock_stripe_add(rbio))
2093 recover_rbio(rbio);
2094 }
2095
2096 static void recover_rbio_work_locked(struct work_struct *work)
2097 {
2098 recover_rbio(container_of(work, struct btrfs_raid_bio, work));
2099 }
2100
2101 static void set_rbio_raid6_extra_error(struct btrfs_raid_bio *rbio, int mirror_num)
2102 {
2103 bool found = false;
2104 int sector_nr;
2105
2106 /*
2107 * This is for RAID6 extra recovery tries, thus mirror number should
2108 * be large than 2.
2109 * Mirror 1 means read from data stripes. Mirror 2 means rebuild using
2110 * RAID5 methods.
2111 */
2112 ASSERT(mirror_num > 2);
2113 for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) {
2114 int found_errors;
2115 int faila;
2116 int failb;
2117
2118 found_errors = get_rbio_veritical_errors(rbio, sector_nr,
2119 &faila, &failb);
2120 /* This vertical stripe doesn't have errors. */
2121 if (!found_errors)
2122 continue;
2123
2124 /*
2125 * If we found errors, there should be only one error marked
2126 * by previous set_rbio_range_error().
2127 */
2128 ASSERT(found_errors == 1);
2129 found = true;
2130
2131 /* Now select another stripe to mark as error. */
2132 failb = rbio->real_stripes - (mirror_num - 1);
2133 if (failb <= faila)
2134 failb--;
2135
2136 /* Set the extra bit in error bitmap. */
2137 if (failb >= 0)
2138 set_bit(failb * rbio->stripe_nsectors + sector_nr,
2139 rbio->error_bitmap);
2140 }
2141
2142 /* We should found at least one vertical stripe with error.*/
2143 ASSERT(found);
2144 }
2145
2146 /*
2147 * the main entry point for reads from the higher layers. This
2148 * is really only called when the normal read path had a failure,
2149 * so we assume the bio they send down corresponds to a failed part
2150 * of the drive.
2151 */
2152 void raid56_parity_recover(struct bio *bio, struct btrfs_io_context *bioc,
2153 int mirror_num)
2154 {
2155 struct btrfs_fs_info *fs_info = bioc->fs_info;
2156 struct btrfs_raid_bio *rbio;
2157
2158 rbio = alloc_rbio(fs_info, bioc);
2159 if (IS_ERR(rbio)) {
2160 bio->bi_status = errno_to_blk_status(PTR_ERR(rbio));
2161 bio_endio(bio);
2162 return;
2163 }
2164
2165 rbio->operation = BTRFS_RBIO_READ_REBUILD;
2166 rbio_add_bio(rbio, bio);
2167
2168 set_rbio_range_error(rbio, bio);
2169
2170 /*
2171 * Loop retry:
2172 * for 'mirror == 2', reconstruct from all other stripes.
2173 * for 'mirror_num > 2', select a stripe to fail on every retry.
2174 */
2175 if (mirror_num > 2)
2176 set_rbio_raid6_extra_error(rbio, mirror_num);
2177
2178 start_async_work(rbio, recover_rbio_work);
2179 }
2180
2181 static void fill_data_csums(struct btrfs_raid_bio *rbio)
2182 {
2183 struct btrfs_fs_info *fs_info = rbio->bioc->fs_info;
2184 struct btrfs_root *csum_root = btrfs_csum_root(fs_info,
2185 rbio->bioc->full_stripe_logical);
2186 const u64 start = rbio->bioc->full_stripe_logical;
2187 const u32 len = (rbio->nr_data * rbio->stripe_nsectors) <<
2188 fs_info->sectorsize_bits;
2189 int ret;
2190
2191 /* The rbio should not have its csum buffer initialized. */
2192 ASSERT(!rbio->csum_buf && !rbio->csum_bitmap);
2193
2194 /*
2195 * Skip the csum search if:
2196 *
2197 * - The rbio doesn't belong to data block groups
2198 * Then we are doing IO for tree blocks, no need to search csums.
2199 *
2200 * - The rbio belongs to mixed block groups
2201 * This is to avoid deadlock, as we're already holding the full
2202 * stripe lock, if we trigger a metadata read, and it needs to do
2203 * raid56 recovery, we will deadlock.
2204 */
2205 if (!(rbio->bioc->map_type & BTRFS_BLOCK_GROUP_DATA) ||
2206 rbio->bioc->map_type & BTRFS_BLOCK_GROUP_METADATA)
2207 return;
2208
2209 rbio->csum_buf = kzalloc(rbio->nr_data * rbio->stripe_nsectors *
2210 fs_info->csum_size, GFP_NOFS);
2211 rbio->csum_bitmap = bitmap_zalloc(rbio->nr_data * rbio->stripe_nsectors,
2212 GFP_NOFS);
2213 if (!rbio->csum_buf || !rbio->csum_bitmap) {
2214 ret = -ENOMEM;
2215 goto error;
2216 }
2217
2218 ret = btrfs_lookup_csums_bitmap(csum_root, NULL, start, start + len - 1,
2219 rbio->csum_buf, rbio->csum_bitmap);
2220 if (ret < 0)
2221 goto error;
2222 if (bitmap_empty(rbio->csum_bitmap, len >> fs_info->sectorsize_bits))
2223 goto no_csum;
2224 return;
2225
2226 error:
2227 /*
2228 * We failed to allocate memory or grab the csum, but it's not fatal,
2229 * we can still continue. But better to warn users that RMW is no
2230 * longer safe for this particular sub-stripe write.
2231 */
2232 btrfs_warn_rl(fs_info,
2233 "sub-stripe write for full stripe %llu is not safe, failed to get csum: %d",
2234 rbio->bioc->full_stripe_logical, ret);
2235 no_csum:
2236 kfree(rbio->csum_buf);
2237 bitmap_free(rbio->csum_bitmap);
2238 rbio->csum_buf = NULL;
2239 rbio->csum_bitmap = NULL;
2240 }
2241
2242 static int rmw_read_wait_recover(struct btrfs_raid_bio *rbio)
2243 {
2244 struct bio_list bio_list = BIO_EMPTY_LIST;
2245 int total_sector_nr;
2246 int ret = 0;
2247
2248 /*
2249 * Fill the data csums we need for data verification. We need to fill
2250 * the csum_bitmap/csum_buf first, as our endio function will try to
2251 * verify the data sectors.
2252 */
2253 fill_data_csums(rbio);
2254
2255 /*
2256 * Build a list of bios to read all sectors (including data and P/Q).
2257 *
2258 * This behavior is to compensate the later csum verification and recovery.
2259 */
2260 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
2261 total_sector_nr++) {
2262 struct sector_ptr *sector;
2263 int stripe = total_sector_nr / rbio->stripe_nsectors;
2264 int sectornr = total_sector_nr % rbio->stripe_nsectors;
2265
2266 sector = rbio_stripe_sector(rbio, stripe, sectornr);
2267 ret = rbio_add_io_sector(rbio, &bio_list, sector,
2268 stripe, sectornr, REQ_OP_READ);
2269 if (ret) {
2270 bio_list_put(&bio_list);
2271 return ret;
2272 }
2273 }
2274
2275 /*
2276 * We may or may not have any corrupted sectors (including missing dev
2277 * and csum mismatch), just let recover_sectors() to handle them all.
2278 */
2279 submit_read_wait_bio_list(rbio, &bio_list);
2280 return recover_sectors(rbio);
2281 }
2282
2283 static void raid_wait_write_end_io(struct bio *bio)
2284 {
2285 struct btrfs_raid_bio *rbio = bio->bi_private;
2286 blk_status_t err = bio->bi_status;
2287
2288 if (err)
2289 rbio_update_error_bitmap(rbio, bio);
2290 bio_put(bio);
2291 if (atomic_dec_and_test(&rbio->stripes_pending))
2292 wake_up(&rbio->io_wait);
2293 }
2294
2295 static void submit_write_bios(struct btrfs_raid_bio *rbio,
2296 struct bio_list *bio_list)
2297 {
2298 struct bio *bio;
2299
2300 atomic_set(&rbio->stripes_pending, bio_list_size(bio_list));
2301 while ((bio = bio_list_pop(bio_list))) {
2302 bio->bi_end_io = raid_wait_write_end_io;
2303
2304 if (trace_raid56_write_enabled()) {
2305 struct raid56_bio_trace_info trace_info = { 0 };
2306
2307 bio_get_trace_info(rbio, bio, &trace_info);
2308 trace_raid56_write(rbio, bio, &trace_info);
2309 }
2310 submit_bio(bio);
2311 }
2312 }
2313
2314 /*
2315 * To determine if we need to read any sector from the disk.
2316 * Should only be utilized in RMW path, to skip cached rbio.
2317 */
2318 static bool need_read_stripe_sectors(struct btrfs_raid_bio *rbio)
2319 {
2320 int i;
2321
2322 for (i = 0; i < rbio->nr_data * rbio->stripe_nsectors; i++) {
2323 struct sector_ptr *sector = &rbio->stripe_sectors[i];
2324
2325 /*
2326 * We have a sector which doesn't have page nor uptodate,
2327 * thus this rbio can not be cached one, as cached one must
2328 * have all its data sectors present and uptodate.
2329 */
2330 if (!sector->page || !sector->uptodate)
2331 return true;
2332 }
2333 return false;
2334 }
2335
2336 static void rmw_rbio(struct btrfs_raid_bio *rbio)
2337 {
2338 struct bio_list bio_list;
2339 int sectornr;
2340 int ret = 0;
2341
2342 /*
2343 * Allocate the pages for parity first, as P/Q pages will always be
2344 * needed for both full-stripe and sub-stripe writes.
2345 */
2346 ret = alloc_rbio_parity_pages(rbio);
2347 if (ret < 0)
2348 goto out;
2349
2350 /*
2351 * Either full stripe write, or we have every data sector already
2352 * cached, can go to write path immediately.
2353 */
2354 if (!rbio_is_full(rbio) && need_read_stripe_sectors(rbio)) {
2355 /*
2356 * Now we're doing sub-stripe write, also need all data stripes
2357 * to do the full RMW.
2358 */
2359 ret = alloc_rbio_data_pages(rbio);
2360 if (ret < 0)
2361 goto out;
2362
2363 index_rbio_pages(rbio);
2364
2365 ret = rmw_read_wait_recover(rbio);
2366 if (ret < 0)
2367 goto out;
2368 }
2369
2370 /*
2371 * At this stage we're not allowed to add any new bios to the
2372 * bio list any more, anyone else that wants to change this stripe
2373 * needs to do their own rmw.
2374 */
2375 spin_lock(&rbio->bio_list_lock);
2376 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
2377 spin_unlock(&rbio->bio_list_lock);
2378
2379 bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors);
2380
2381 index_rbio_pages(rbio);
2382
2383 /*
2384 * We don't cache full rbios because we're assuming
2385 * the higher layers are unlikely to use this area of
2386 * the disk again soon. If they do use it again,
2387 * hopefully they will send another full bio.
2388 */
2389 if (!rbio_is_full(rbio))
2390 cache_rbio_pages(rbio);
2391 else
2392 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2393
2394 for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++)
2395 generate_pq_vertical(rbio, sectornr);
2396
2397 bio_list_init(&bio_list);
2398 ret = rmw_assemble_write_bios(rbio, &bio_list);
2399 if (ret < 0)
2400 goto out;
2401
2402 /* We should have at least one bio assembled. */
2403 ASSERT(bio_list_size(&bio_list));
2404 submit_write_bios(rbio, &bio_list);
2405 wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0);
2406
2407 /* We may have more errors than our tolerance during the read. */
2408 for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++) {
2409 int found_errors;
2410
2411 found_errors = get_rbio_veritical_errors(rbio, sectornr, NULL, NULL);
2412 if (found_errors > rbio->bioc->max_errors) {
2413 ret = -EIO;
2414 break;
2415 }
2416 }
2417 out:
2418 rbio_orig_end_io(rbio, errno_to_blk_status(ret));
2419 }
2420
2421 static void rmw_rbio_work(struct work_struct *work)
2422 {
2423 struct btrfs_raid_bio *rbio;
2424
2425 rbio = container_of(work, struct btrfs_raid_bio, work);
2426 if (lock_stripe_add(rbio) == 0)
2427 rmw_rbio(rbio);
2428 }
2429
2430 static void rmw_rbio_work_locked(struct work_struct *work)
2431 {
2432 rmw_rbio(container_of(work, struct btrfs_raid_bio, work));
2433 }
2434
2435 /*
2436 * The following code is used to scrub/replace the parity stripe
2437 *
2438 * Caller must have already increased bio_counter for getting @bioc.
2439 *
2440 * Note: We need make sure all the pages that add into the scrub/replace
2441 * raid bio are correct and not be changed during the scrub/replace. That
2442 * is those pages just hold metadata or file data with checksum.
2443 */
2444
2445 struct btrfs_raid_bio *raid56_parity_alloc_scrub_rbio(struct bio *bio,
2446 struct btrfs_io_context *bioc,
2447 struct btrfs_device *scrub_dev,
2448 unsigned long *dbitmap, int stripe_nsectors)
2449 {
2450 struct btrfs_fs_info *fs_info = bioc->fs_info;
2451 struct btrfs_raid_bio *rbio;
2452 int i;
2453
2454 rbio = alloc_rbio(fs_info, bioc);
2455 if (IS_ERR(rbio))
2456 return NULL;
2457 bio_list_add(&rbio->bio_list, bio);
2458 /*
2459 * This is a special bio which is used to hold the completion handler
2460 * and make the scrub rbio is similar to the other types
2461 */
2462 ASSERT(!bio->bi_iter.bi_size);
2463 rbio->operation = BTRFS_RBIO_PARITY_SCRUB;
2464
2465 /*
2466 * After mapping bioc with BTRFS_MAP_WRITE, parities have been sorted
2467 * to the end position, so this search can start from the first parity
2468 * stripe.
2469 */
2470 for (i = rbio->nr_data; i < rbio->real_stripes; i++) {
2471 if (bioc->stripes[i].dev == scrub_dev) {
2472 rbio->scrubp = i;
2473 break;
2474 }
2475 }
2476 ASSERT_RBIO_STRIPE(i < rbio->real_stripes, rbio, i);
2477
2478 bitmap_copy(&rbio->dbitmap, dbitmap, stripe_nsectors);
2479 return rbio;
2480 }
2481
2482 /*
2483 * We just scrub the parity that we have correct data on the same horizontal,
2484 * so we needn't allocate all pages for all the stripes.
2485 */
2486 static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio)
2487 {
2488 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
2489 int total_sector_nr;
2490
2491 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
2492 total_sector_nr++) {
2493 struct page *page;
2494 int sectornr = total_sector_nr % rbio->stripe_nsectors;
2495 int index = (total_sector_nr * sectorsize) >> PAGE_SHIFT;
2496
2497 if (!test_bit(sectornr, &rbio->dbitmap))
2498 continue;
2499 if (rbio->stripe_pages[index])
2500 continue;
2501 page = alloc_page(GFP_NOFS);
2502 if (!page)
2503 return -ENOMEM;
2504 rbio->stripe_pages[index] = page;
2505 }
2506 index_stripe_sectors(rbio);
2507 return 0;
2508 }
2509
2510 static int finish_parity_scrub(struct btrfs_raid_bio *rbio)
2511 {
2512 struct btrfs_io_context *bioc = rbio->bioc;
2513 const u32 sectorsize = bioc->fs_info->sectorsize;
2514 void **pointers = rbio->finish_pointers;
2515 unsigned long *pbitmap = &rbio->finish_pbitmap;
2516 int nr_data = rbio->nr_data;
2517 int stripe;
2518 int sectornr;
2519 bool has_qstripe;
2520 struct sector_ptr p_sector = { 0 };
2521 struct sector_ptr q_sector = { 0 };
2522 struct bio_list bio_list;
2523 int is_replace = 0;
2524 int ret;
2525
2526 bio_list_init(&bio_list);
2527
2528 if (rbio->real_stripes - rbio->nr_data == 1)
2529 has_qstripe = false;
2530 else if (rbio->real_stripes - rbio->nr_data == 2)
2531 has_qstripe = true;
2532 else
2533 BUG();
2534
2535 /*
2536 * Replace is running and our P/Q stripe is being replaced, then we
2537 * need to duplicate the final write to replace target.
2538 */
2539 if (bioc->replace_nr_stripes && bioc->replace_stripe_src == rbio->scrubp) {
2540 is_replace = 1;
2541 bitmap_copy(pbitmap, &rbio->dbitmap, rbio->stripe_nsectors);
2542 }
2543
2544 /*
2545 * Because the higher layers(scrubber) are unlikely to
2546 * use this area of the disk again soon, so don't cache
2547 * it.
2548 */
2549 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2550
2551 p_sector.page = alloc_page(GFP_NOFS);
2552 if (!p_sector.page)
2553 return -ENOMEM;
2554 p_sector.pgoff = 0;
2555 p_sector.uptodate = 1;
2556
2557 if (has_qstripe) {
2558 /* RAID6, allocate and map temp space for the Q stripe */
2559 q_sector.page = alloc_page(GFP_NOFS);
2560 if (!q_sector.page) {
2561 __free_page(p_sector.page);
2562 p_sector.page = NULL;
2563 return -ENOMEM;
2564 }
2565 q_sector.pgoff = 0;
2566 q_sector.uptodate = 1;
2567 pointers[rbio->real_stripes - 1] = kmap_local_page(q_sector.page);
2568 }
2569
2570 bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors);
2571
2572 /* Map the parity stripe just once */
2573 pointers[nr_data] = kmap_local_page(p_sector.page);
2574
2575 for_each_set_bit(sectornr, &rbio->dbitmap, rbio->stripe_nsectors) {
2576 struct sector_ptr *sector;
2577 void *parity;
2578
2579 /* first collect one page from each data stripe */
2580 for (stripe = 0; stripe < nr_data; stripe++) {
2581 sector = sector_in_rbio(rbio, stripe, sectornr, 0);
2582 pointers[stripe] = kmap_local_page(sector->page) +
2583 sector->pgoff;
2584 }
2585
2586 if (has_qstripe) {
2587 assert_rbio(rbio);
2588 /* RAID6, call the library function to fill in our P/Q */
2589 raid6_call.gen_syndrome(rbio->real_stripes, sectorsize,
2590 pointers);
2591 } else {
2592 /* raid5 */
2593 memcpy(pointers[nr_data], pointers[0], sectorsize);
2594 run_xor(pointers + 1, nr_data - 1, sectorsize);
2595 }
2596
2597 /* Check scrubbing parity and repair it */
2598 sector = rbio_stripe_sector(rbio, rbio->scrubp, sectornr);
2599 parity = kmap_local_page(sector->page) + sector->pgoff;
2600 if (memcmp(parity, pointers[rbio->scrubp], sectorsize) != 0)
2601 memcpy(parity, pointers[rbio->scrubp], sectorsize);
2602 else
2603 /* Parity is right, needn't writeback */
2604 bitmap_clear(&rbio->dbitmap, sectornr, 1);
2605 kunmap_local(parity);
2606
2607 for (stripe = nr_data - 1; stripe >= 0; stripe--)
2608 kunmap_local(pointers[stripe]);
2609 }
2610
2611 kunmap_local(pointers[nr_data]);
2612 __free_page(p_sector.page);
2613 p_sector.page = NULL;
2614 if (q_sector.page) {
2615 kunmap_local(pointers[rbio->real_stripes - 1]);
2616 __free_page(q_sector.page);
2617 q_sector.page = NULL;
2618 }
2619
2620 /*
2621 * time to start writing. Make bios for everything from the
2622 * higher layers (the bio_list in our rbio) and our p/q. Ignore
2623 * everything else.
2624 */
2625 for_each_set_bit(sectornr, &rbio->dbitmap, rbio->stripe_nsectors) {
2626 struct sector_ptr *sector;
2627
2628 sector = rbio_stripe_sector(rbio, rbio->scrubp, sectornr);
2629 ret = rbio_add_io_sector(rbio, &bio_list, sector, rbio->scrubp,
2630 sectornr, REQ_OP_WRITE);
2631 if (ret)
2632 goto cleanup;
2633 }
2634
2635 if (!is_replace)
2636 goto submit_write;
2637
2638 /*
2639 * Replace is running and our parity stripe needs to be duplicated to
2640 * the target device. Check we have a valid source stripe number.
2641 */
2642 ASSERT_RBIO(rbio->bioc->replace_stripe_src >= 0, rbio);
2643 for_each_set_bit(sectornr, pbitmap, rbio->stripe_nsectors) {
2644 struct sector_ptr *sector;
2645
2646 sector = rbio_stripe_sector(rbio, rbio->scrubp, sectornr);
2647 ret = rbio_add_io_sector(rbio, &bio_list, sector,
2648 rbio->real_stripes,
2649 sectornr, REQ_OP_WRITE);
2650 if (ret)
2651 goto cleanup;
2652 }
2653
2654 submit_write:
2655 submit_write_bios(rbio, &bio_list);
2656 return 0;
2657
2658 cleanup:
2659 bio_list_put(&bio_list);
2660 return ret;
2661 }
2662
2663 static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe)
2664 {
2665 if (stripe >= 0 && stripe < rbio->nr_data)
2666 return 1;
2667 return 0;
2668 }
2669
2670 static int recover_scrub_rbio(struct btrfs_raid_bio *rbio)
2671 {
2672 void **pointers = NULL;
2673 void **unmap_array = NULL;
2674 int sector_nr;
2675 int ret = 0;
2676
2677 /*
2678 * @pointers array stores the pointer for each sector.
2679 *
2680 * @unmap_array stores copy of pointers that does not get reordered
2681 * during reconstruction so that kunmap_local works.
2682 */
2683 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
2684 unmap_array = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
2685 if (!pointers || !unmap_array) {
2686 ret = -ENOMEM;
2687 goto out;
2688 }
2689
2690 for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) {
2691 int dfail = 0, failp = -1;
2692 int faila;
2693 int failb;
2694 int found_errors;
2695
2696 found_errors = get_rbio_veritical_errors(rbio, sector_nr,
2697 &faila, &failb);
2698 if (found_errors > rbio->bioc->max_errors) {
2699 ret = -EIO;
2700 goto out;
2701 }
2702 if (found_errors == 0)
2703 continue;
2704
2705 /* We should have at least one error here. */
2706 ASSERT(faila >= 0 || failb >= 0);
2707
2708 if (is_data_stripe(rbio, faila))
2709 dfail++;
2710 else if (is_parity_stripe(faila))
2711 failp = faila;
2712
2713 if (is_data_stripe(rbio, failb))
2714 dfail++;
2715 else if (is_parity_stripe(failb))
2716 failp = failb;
2717 /*
2718 * Because we can not use a scrubbing parity to repair the
2719 * data, so the capability of the repair is declined. (In the
2720 * case of RAID5, we can not repair anything.)
2721 */
2722 if (dfail > rbio->bioc->max_errors - 1) {
2723 ret = -EIO;
2724 goto out;
2725 }
2726 /*
2727 * If all data is good, only parity is correctly, just repair
2728 * the parity, no need to recover data stripes.
2729 */
2730 if (dfail == 0)
2731 continue;
2732
2733 /*
2734 * Here means we got one corrupted data stripe and one
2735 * corrupted parity on RAID6, if the corrupted parity is
2736 * scrubbing parity, luckily, use the other one to repair the
2737 * data, or we can not repair the data stripe.
2738 */
2739 if (failp != rbio->scrubp) {
2740 ret = -EIO;
2741 goto out;
2742 }
2743
2744 ret = recover_vertical(rbio, sector_nr, pointers, unmap_array);
2745 if (ret < 0)
2746 goto out;
2747 }
2748 out:
2749 kfree(pointers);
2750 kfree(unmap_array);
2751 return ret;
2752 }
2753
2754 static int scrub_assemble_read_bios(struct btrfs_raid_bio *rbio)
2755 {
2756 struct bio_list bio_list = BIO_EMPTY_LIST;
2757 int total_sector_nr;
2758 int ret = 0;
2759
2760 /* Build a list of bios to read all the missing parts. */
2761 for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors;
2762 total_sector_nr++) {
2763 int sectornr = total_sector_nr % rbio->stripe_nsectors;
2764 int stripe = total_sector_nr / rbio->stripe_nsectors;
2765 struct sector_ptr *sector;
2766
2767 /* No data in the vertical stripe, no need to read. */
2768 if (!test_bit(sectornr, &rbio->dbitmap))
2769 continue;
2770
2771 /*
2772 * We want to find all the sectors missing from the rbio and
2773 * read them from the disk. If sector_in_rbio() finds a sector
2774 * in the bio list we don't need to read it off the stripe.
2775 */
2776 sector = sector_in_rbio(rbio, stripe, sectornr, 1);
2777 if (sector)
2778 continue;
2779
2780 sector = rbio_stripe_sector(rbio, stripe, sectornr);
2781 /*
2782 * The bio cache may have handed us an uptodate sector. If so,
2783 * use it.
2784 */
2785 if (sector->uptodate)
2786 continue;
2787
2788 ret = rbio_add_io_sector(rbio, &bio_list, sector, stripe,
2789 sectornr, REQ_OP_READ);
2790 if (ret) {
2791 bio_list_put(&bio_list);
2792 return ret;
2793 }
2794 }
2795
2796 submit_read_wait_bio_list(rbio, &bio_list);
2797 return 0;
2798 }
2799
2800 static void scrub_rbio(struct btrfs_raid_bio *rbio)
2801 {
2802 int sector_nr;
2803 int ret;
2804
2805 ret = alloc_rbio_essential_pages(rbio);
2806 if (ret)
2807 goto out;
2808
2809 bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors);
2810
2811 ret = scrub_assemble_read_bios(rbio);
2812 if (ret < 0)
2813 goto out;
2814
2815 /* We may have some failures, recover the failed sectors first. */
2816 ret = recover_scrub_rbio(rbio);
2817 if (ret < 0)
2818 goto out;
2819
2820 /*
2821 * We have every sector properly prepared. Can finish the scrub
2822 * and writeback the good content.
2823 */
2824 ret = finish_parity_scrub(rbio);
2825 wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0);
2826 for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) {
2827 int found_errors;
2828
2829 found_errors = get_rbio_veritical_errors(rbio, sector_nr, NULL, NULL);
2830 if (found_errors > rbio->bioc->max_errors) {
2831 ret = -EIO;
2832 break;
2833 }
2834 }
2835 out:
2836 rbio_orig_end_io(rbio, errno_to_blk_status(ret));
2837 }
2838
2839 static void scrub_rbio_work_locked(struct work_struct *work)
2840 {
2841 scrub_rbio(container_of(work, struct btrfs_raid_bio, work));
2842 }
2843
2844 void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio)
2845 {
2846 if (!lock_stripe_add(rbio))
2847 start_async_work(rbio, scrub_rbio_work_locked);
2848 }
2849
2850 /*
2851 * This is for scrub call sites where we already have correct data contents.
2852 * This allows us to avoid reading data stripes again.
2853 *
2854 * Unfortunately here we have to do page copy, other than reusing the pages.
2855 * This is due to the fact rbio has its own page management for its cache.
2856 */
2857 void raid56_parity_cache_data_pages(struct btrfs_raid_bio *rbio,
2858 struct page **data_pages, u64 data_logical)
2859 {
2860 const u64 offset_in_full_stripe = data_logical -
2861 rbio->bioc->full_stripe_logical;
2862 const int page_index = offset_in_full_stripe >> PAGE_SHIFT;
2863 const u32 sectorsize = rbio->bioc->fs_info->sectorsize;
2864 const u32 sectors_per_page = PAGE_SIZE / sectorsize;
2865 int ret;
2866
2867 /*
2868 * If we hit ENOMEM temporarily, but later at
2869 * raid56_parity_submit_scrub_rbio() time it succeeded, we just do
2870 * the extra read, not a big deal.
2871 *
2872 * If we hit ENOMEM later at raid56_parity_submit_scrub_rbio() time,
2873 * the bio would got proper error number set.
2874 */
2875 ret = alloc_rbio_data_pages(rbio);
2876 if (ret < 0)
2877 return;
2878
2879 /* data_logical must be at stripe boundary and inside the full stripe. */
2880 ASSERT(IS_ALIGNED(offset_in_full_stripe, BTRFS_STRIPE_LEN));
2881 ASSERT(offset_in_full_stripe < (rbio->nr_data << BTRFS_STRIPE_LEN_SHIFT));
2882
2883 for (int page_nr = 0; page_nr < (BTRFS_STRIPE_LEN >> PAGE_SHIFT); page_nr++) {
2884 struct page *dst = rbio->stripe_pages[page_nr + page_index];
2885 struct page *src = data_pages[page_nr];
2886
2887 memcpy_page(dst, 0, src, 0, PAGE_SIZE);
2888 for (int sector_nr = sectors_per_page * page_index;
2889 sector_nr < sectors_per_page * (page_index + 1);
2890 sector_nr++)
2891 rbio->stripe_sectors[sector_nr].uptodate = true;
2892 }
2893 }
2894
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.