blob: cd8d302a1f61588bf24e8a563a53b93ec30f9318 [file] [log] [blame]
David Woodhouse53b381b2013-01-29 18:40:14 -05001/*
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>
Geert Uytterhoevend7011f52013-03-03 04:44:41 -070034#include <linux/vmalloc.h>
David Woodhouse53b381b2013-01-29 18:40:14 -050035#include <asm/div64.h>
David Woodhouse53b381b2013-01-29 18:40:14 -050036#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
Chris Mason4ae10b32013-01-31 14:42:09 -050050/*
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
Chris Mason4ae10b32013-01-31 14:42:09 -050061#define RBIO_CACHE_SIZE 1024
62
Miao Xie1b94b552014-11-06 16:14:21 +080063enum btrfs_rbio_ops {
Omar Sandovalb4ee1782015-06-19 11:52:50 -070064 BTRFS_RBIO_WRITE,
65 BTRFS_RBIO_READ_REBUILD,
66 BTRFS_RBIO_PARITY_SCRUB,
67 BTRFS_RBIO_REBUILD_MISSING,
Miao Xie1b94b552014-11-06 16:14:21 +080068};
69
David Woodhouse53b381b2013-01-29 18:40:14 -050070struct btrfs_raid_bio {
71 struct btrfs_fs_info *fs_info;
72 struct btrfs_bio *bbio;
73
David Woodhouse53b381b2013-01-29 18:40:14 -050074 /* while we're doing rmw on a stripe
75 * we put it into a hash table so we can
76 * lock the stripe and merge more rbios
77 * into it.
78 */
79 struct list_head hash_list;
80
81 /*
Chris Mason4ae10b32013-01-31 14:42:09 -050082 * LRU list for the stripe cache
83 */
84 struct list_head stripe_cache;
85
86 /*
David Woodhouse53b381b2013-01-29 18:40:14 -050087 * for scheduling work in the helper threads
88 */
89 struct btrfs_work work;
90
91 /*
92 * bio list and bio_list_lock are used
93 * to add more bios into the stripe
94 * in hopes of avoiding the full rmw
95 */
96 struct bio_list bio_list;
97 spinlock_t bio_list_lock;
98
Chris Mason6ac0f482013-01-31 14:42:28 -050099 /* also protected by the bio_list_lock, the
100 * plug list is used by the plugging code
101 * to collect partial bios while plugged. The
102 * stripe locking code also uses it to hand off
David Woodhouse53b381b2013-01-29 18:40:14 -0500103 * the stripe lock to the next pending IO
104 */
105 struct list_head plug_list;
106
107 /*
108 * flags that tell us if it is safe to
109 * merge with this bio
110 */
111 unsigned long flags;
112
113 /* size of each individual stripe on disk */
114 int stripe_len;
115
116 /* number of data stripes (no p/q) */
117 int nr_data;
118
Miao Xie2c8cdd62014-11-14 16:06:25 +0800119 int real_stripes;
120
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800121 int stripe_npages;
David Woodhouse53b381b2013-01-29 18:40:14 -0500122 /*
123 * set if we're doing a parity rebuild
124 * for a read from higher up, which is handled
125 * differently from a parity rebuild as part of
126 * rmw
127 */
Miao Xie1b94b552014-11-06 16:14:21 +0800128 enum btrfs_rbio_ops operation;
David Woodhouse53b381b2013-01-29 18:40:14 -0500129
130 /* first bad stripe */
131 int faila;
132
133 /* second bad stripe (for raid6 use) */
134 int failb;
135
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800136 int scrubp;
David Woodhouse53b381b2013-01-29 18:40:14 -0500137 /*
138 * number of pages needed to represent the full
139 * stripe
140 */
141 int nr_pages;
142
143 /*
144 * size of all the bios in the bio_list. This
145 * helps us decide if the rbio maps to a full
146 * stripe or not
147 */
148 int bio_list_bytes;
149
Miao Xie42452152014-11-25 16:39:28 +0800150 int generic_bio_cnt;
151
David Woodhouse53b381b2013-01-29 18:40:14 -0500152 atomic_t refs;
153
Miao Xieb89e1b02014-10-15 11:18:44 +0800154 atomic_t stripes_pending;
155
156 atomic_t error;
David Woodhouse53b381b2013-01-29 18:40:14 -0500157 /*
158 * these are two arrays of pointers. We allocate the
159 * rbio big enough to hold them both and setup their
160 * locations when the rbio is allocated
161 */
162
163 /* pointers to pages that we allocated for
164 * reading/writing stripes directly from the disk (including P/Q)
165 */
166 struct page **stripe_pages;
167
168 /*
169 * pointers to the pages in the bio_list. Stored
170 * here for faster lookup
171 */
172 struct page **bio_pages;
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800173
174 /*
175 * bitmap to record which horizontal stripe has data
176 */
177 unsigned long *dbitmap;
David Woodhouse53b381b2013-01-29 18:40:14 -0500178};
179
180static int __raid56_parity_recover(struct btrfs_raid_bio *rbio);
181static noinline void finish_rmw(struct btrfs_raid_bio *rbio);
182static void rmw_work(struct btrfs_work *work);
183static void read_rebuild_work(struct btrfs_work *work);
184static void async_rmw_stripe(struct btrfs_raid_bio *rbio);
185static void async_read_rebuild(struct btrfs_raid_bio *rbio);
186static int fail_bio_stripe(struct btrfs_raid_bio *rbio, struct bio *bio);
187static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed);
188static void __free_raid_bio(struct btrfs_raid_bio *rbio);
189static void index_rbio_pages(struct btrfs_raid_bio *rbio);
190static int alloc_rbio_pages(struct btrfs_raid_bio *rbio);
191
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800192static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
193 int need_check);
194static void async_scrub_parity(struct btrfs_raid_bio *rbio);
195
David Woodhouse53b381b2013-01-29 18:40:14 -0500196/*
197 * the stripe hash table is used for locking, and to collect
198 * bios in hopes of making a full stripe
199 */
200int btrfs_alloc_stripe_hash_table(struct btrfs_fs_info *info)
201{
202 struct btrfs_stripe_hash_table *table;
203 struct btrfs_stripe_hash_table *x;
204 struct btrfs_stripe_hash *cur;
205 struct btrfs_stripe_hash *h;
206 int num_entries = 1 << BTRFS_STRIPE_HASH_TABLE_BITS;
207 int i;
David Sterba83c82662013-03-01 15:03:00 +0000208 int table_size;
David Woodhouse53b381b2013-01-29 18:40:14 -0500209
210 if (info->stripe_hash_table)
211 return 0;
212
David Sterba83c82662013-03-01 15:03:00 +0000213 /*
214 * The table is large, starting with order 4 and can go as high as
215 * order 7 in case lock debugging is turned on.
216 *
217 * Try harder to allocate and fallback to vmalloc to lower the chance
218 * of a failing mount.
219 */
220 table_size = sizeof(*table) + sizeof(*h) * num_entries;
221 table = kzalloc(table_size, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
222 if (!table) {
223 table = vzalloc(table_size);
224 if (!table)
225 return -ENOMEM;
226 }
David Woodhouse53b381b2013-01-29 18:40:14 -0500227
Chris Mason4ae10b32013-01-31 14:42:09 -0500228 spin_lock_init(&table->cache_lock);
229 INIT_LIST_HEAD(&table->stripe_cache);
230
David Woodhouse53b381b2013-01-29 18:40:14 -0500231 h = table->table;
232
233 for (i = 0; i < num_entries; i++) {
234 cur = h + i;
235 INIT_LIST_HEAD(&cur->hash_list);
236 spin_lock_init(&cur->lock);
237 init_waitqueue_head(&cur->wait);
238 }
239
240 x = cmpxchg(&info->stripe_hash_table, NULL, table);
Wang Shilongf7493032014-11-22 21:13:10 +0800241 if (x)
242 kvfree(x);
David Woodhouse53b381b2013-01-29 18:40:14 -0500243 return 0;
244}
245
246/*
Chris Mason4ae10b32013-01-31 14:42:09 -0500247 * caching an rbio means to copy anything from the
248 * bio_pages array into the stripe_pages array. We
249 * use the page uptodate bit in the stripe cache array
250 * to indicate if it has valid data
251 *
252 * once the caching is done, we set the cache ready
253 * bit.
254 */
255static void cache_rbio_pages(struct btrfs_raid_bio *rbio)
256{
257 int i;
258 char *s;
259 char *d;
260 int ret;
261
262 ret = alloc_rbio_pages(rbio);
263 if (ret)
264 return;
265
266 for (i = 0; i < rbio->nr_pages; i++) {
267 if (!rbio->bio_pages[i])
268 continue;
269
270 s = kmap(rbio->bio_pages[i]);
271 d = kmap(rbio->stripe_pages[i]);
272
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300273 memcpy(d, s, PAGE_SIZE);
Chris Mason4ae10b32013-01-31 14:42:09 -0500274
275 kunmap(rbio->bio_pages[i]);
276 kunmap(rbio->stripe_pages[i]);
277 SetPageUptodate(rbio->stripe_pages[i]);
278 }
279 set_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
280}
281
282/*
David Woodhouse53b381b2013-01-29 18:40:14 -0500283 * we hash on the first logical address of the stripe
284 */
285static int rbio_bucket(struct btrfs_raid_bio *rbio)
286{
Zhao Lei8e5cfb52015-01-20 15:11:33 +0800287 u64 num = rbio->bbio->raid_map[0];
David Woodhouse53b381b2013-01-29 18:40:14 -0500288
289 /*
290 * we shift down quite a bit. We're using byte
291 * addressing, and most of the lower bits are zeros.
292 * This tends to upset hash_64, and it consistently
293 * returns just one or two different values.
294 *
295 * shifting off the lower bits fixes things.
296 */
297 return hash_64(num >> 16, BTRFS_STRIPE_HASH_TABLE_BITS);
298}
299
300/*
Chris Mason4ae10b32013-01-31 14:42:09 -0500301 * stealing an rbio means taking all the uptodate pages from the stripe
302 * array in the source rbio and putting them into the destination rbio
303 */
304static void steal_rbio(struct btrfs_raid_bio *src, struct btrfs_raid_bio *dest)
305{
306 int i;
307 struct page *s;
308 struct page *d;
309
310 if (!test_bit(RBIO_CACHE_READY_BIT, &src->flags))
311 return;
312
313 for (i = 0; i < dest->nr_pages; i++) {
314 s = src->stripe_pages[i];
315 if (!s || !PageUptodate(s)) {
316 continue;
317 }
318
319 d = dest->stripe_pages[i];
320 if (d)
321 __free_page(d);
322
323 dest->stripe_pages[i] = s;
324 src->stripe_pages[i] = NULL;
325 }
326}
327
328/*
David Woodhouse53b381b2013-01-29 18:40:14 -0500329 * merging means we take the bio_list from the victim and
330 * splice it into the destination. The victim should
331 * be discarded afterwards.
332 *
333 * must be called with dest->rbio_list_lock held
334 */
335static void merge_rbio(struct btrfs_raid_bio *dest,
336 struct btrfs_raid_bio *victim)
337{
338 bio_list_merge(&dest->bio_list, &victim->bio_list);
339 dest->bio_list_bytes += victim->bio_list_bytes;
Miao Xie42452152014-11-25 16:39:28 +0800340 dest->generic_bio_cnt += victim->generic_bio_cnt;
David Woodhouse53b381b2013-01-29 18:40:14 -0500341 bio_list_init(&victim->bio_list);
342}
343
344/*
Chris Mason4ae10b32013-01-31 14:42:09 -0500345 * used to prune items that are in the cache. The caller
346 * must hold the hash table lock.
347 */
348static void __remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
349{
350 int bucket = rbio_bucket(rbio);
351 struct btrfs_stripe_hash_table *table;
352 struct btrfs_stripe_hash *h;
353 int freeit = 0;
354
355 /*
356 * check the bit again under the hash table lock.
357 */
358 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
359 return;
360
361 table = rbio->fs_info->stripe_hash_table;
362 h = table->table + bucket;
363
364 /* hold the lock for the bucket because we may be
365 * removing it from the hash table
366 */
367 spin_lock(&h->lock);
368
369 /*
370 * hold the lock for the bio list because we need
371 * to make sure the bio list is empty
372 */
373 spin_lock(&rbio->bio_list_lock);
374
375 if (test_and_clear_bit(RBIO_CACHE_BIT, &rbio->flags)) {
376 list_del_init(&rbio->stripe_cache);
377 table->cache_size -= 1;
378 freeit = 1;
379
380 /* if the bio list isn't empty, this rbio is
381 * still involved in an IO. We take it out
382 * of the cache list, and drop the ref that
383 * was held for the list.
384 *
385 * If the bio_list was empty, we also remove
386 * the rbio from the hash_table, and drop
387 * the corresponding ref
388 */
389 if (bio_list_empty(&rbio->bio_list)) {
390 if (!list_empty(&rbio->hash_list)) {
391 list_del_init(&rbio->hash_list);
392 atomic_dec(&rbio->refs);
393 BUG_ON(!list_empty(&rbio->plug_list));
394 }
395 }
396 }
397
398 spin_unlock(&rbio->bio_list_lock);
399 spin_unlock(&h->lock);
400
401 if (freeit)
402 __free_raid_bio(rbio);
403}
404
405/*
406 * prune a given rbio from the cache
407 */
408static void remove_rbio_from_cache(struct btrfs_raid_bio *rbio)
409{
410 struct btrfs_stripe_hash_table *table;
411 unsigned long flags;
412
413 if (!test_bit(RBIO_CACHE_BIT, &rbio->flags))
414 return;
415
416 table = rbio->fs_info->stripe_hash_table;
417
418 spin_lock_irqsave(&table->cache_lock, flags);
419 __remove_rbio_from_cache(rbio);
420 spin_unlock_irqrestore(&table->cache_lock, flags);
421}
422
423/*
424 * remove everything in the cache
425 */
Eric Sandeen48a3b632013-04-25 20:41:01 +0000426static void btrfs_clear_rbio_cache(struct btrfs_fs_info *info)
Chris Mason4ae10b32013-01-31 14:42:09 -0500427{
428 struct btrfs_stripe_hash_table *table;
429 unsigned long flags;
430 struct btrfs_raid_bio *rbio;
431
432 table = info->stripe_hash_table;
433
434 spin_lock_irqsave(&table->cache_lock, flags);
435 while (!list_empty(&table->stripe_cache)) {
436 rbio = list_entry(table->stripe_cache.next,
437 struct btrfs_raid_bio,
438 stripe_cache);
439 __remove_rbio_from_cache(rbio);
440 }
441 spin_unlock_irqrestore(&table->cache_lock, flags);
442}
443
444/*
445 * remove all cached entries and free the hash table
446 * used by unmount
David Woodhouse53b381b2013-01-29 18:40:14 -0500447 */
448void btrfs_free_stripe_hash_table(struct btrfs_fs_info *info)
449{
450 if (!info->stripe_hash_table)
451 return;
Chris Mason4ae10b32013-01-31 14:42:09 -0500452 btrfs_clear_rbio_cache(info);
Wang Shilongf7493032014-11-22 21:13:10 +0800453 kvfree(info->stripe_hash_table);
David Woodhouse53b381b2013-01-29 18:40:14 -0500454 info->stripe_hash_table = NULL;
455}
456
457/*
Chris Mason4ae10b32013-01-31 14:42:09 -0500458 * insert an rbio into the stripe cache. It
459 * must have already been prepared by calling
460 * cache_rbio_pages
461 *
462 * If this rbio was already cached, it gets
463 * moved to the front of the lru.
464 *
465 * If the size of the rbio cache is too big, we
466 * prune an item.
467 */
468static void cache_rbio(struct btrfs_raid_bio *rbio)
469{
470 struct btrfs_stripe_hash_table *table;
471 unsigned long flags;
472
473 if (!test_bit(RBIO_CACHE_READY_BIT, &rbio->flags))
474 return;
475
476 table = rbio->fs_info->stripe_hash_table;
477
478 spin_lock_irqsave(&table->cache_lock, flags);
479 spin_lock(&rbio->bio_list_lock);
480
481 /* bump our ref if we were not in the list before */
482 if (!test_and_set_bit(RBIO_CACHE_BIT, &rbio->flags))
483 atomic_inc(&rbio->refs);
484
485 if (!list_empty(&rbio->stripe_cache)){
486 list_move(&rbio->stripe_cache, &table->stripe_cache);
487 } else {
488 list_add(&rbio->stripe_cache, &table->stripe_cache);
489 table->cache_size += 1;
490 }
491
492 spin_unlock(&rbio->bio_list_lock);
493
494 if (table->cache_size > RBIO_CACHE_SIZE) {
495 struct btrfs_raid_bio *found;
496
497 found = list_entry(table->stripe_cache.prev,
498 struct btrfs_raid_bio,
499 stripe_cache);
500
501 if (found != rbio)
502 __remove_rbio_from_cache(found);
503 }
504
505 spin_unlock_irqrestore(&table->cache_lock, flags);
Chris Mason4ae10b32013-01-31 14:42:09 -0500506}
507
508/*
David Woodhouse53b381b2013-01-29 18:40:14 -0500509 * helper function to run the xor_blocks api. It is only
510 * able to do MAX_XOR_BLOCKS at a time, so we need to
511 * loop through.
512 */
513static void run_xor(void **pages, int src_cnt, ssize_t len)
514{
515 int src_off = 0;
516 int xor_src_cnt = 0;
517 void *dest = pages[src_cnt];
518
519 while(src_cnt > 0) {
520 xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS);
521 xor_blocks(xor_src_cnt, len, dest, pages + src_off);
522
523 src_cnt -= xor_src_cnt;
524 src_off += xor_src_cnt;
525 }
526}
527
528/*
529 * returns true if the bio list inside this rbio
530 * covers an entire stripe (no rmw required).
531 * Must be called with the bio list lock held, or
532 * at a time when you know it is impossible to add
533 * new bios into the list
534 */
535static int __rbio_is_full(struct btrfs_raid_bio *rbio)
536{
537 unsigned long size = rbio->bio_list_bytes;
538 int ret = 1;
539
540 if (size != rbio->nr_data * rbio->stripe_len)
541 ret = 0;
542
543 BUG_ON(size > rbio->nr_data * rbio->stripe_len);
544 return ret;
545}
546
547static int rbio_is_full(struct btrfs_raid_bio *rbio)
548{
549 unsigned long flags;
550 int ret;
551
552 spin_lock_irqsave(&rbio->bio_list_lock, flags);
553 ret = __rbio_is_full(rbio);
554 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
555 return ret;
556}
557
558/*
559 * returns 1 if it is safe to merge two rbios together.
560 * The merging is safe if the two rbios correspond to
561 * the same stripe and if they are both going in the same
562 * direction (read vs write), and if neither one is
563 * locked for final IO
564 *
565 * The caller is responsible for locking such that
566 * rmw_locked is safe to test
567 */
568static int rbio_can_merge(struct btrfs_raid_bio *last,
569 struct btrfs_raid_bio *cur)
570{
571 if (test_bit(RBIO_RMW_LOCKED_BIT, &last->flags) ||
572 test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags))
573 return 0;
574
Chris Mason4ae10b32013-01-31 14:42:09 -0500575 /*
576 * we can't merge with cached rbios, since the
577 * idea is that when we merge the destination
578 * rbio is going to run our IO for us. We can
Nicholas D Steeves01327612016-05-19 21:18:45 -0400579 * steal from cached rbios though, other functions
Chris Mason4ae10b32013-01-31 14:42:09 -0500580 * handle that.
581 */
582 if (test_bit(RBIO_CACHE_BIT, &last->flags) ||
583 test_bit(RBIO_CACHE_BIT, &cur->flags))
584 return 0;
585
Zhao Lei8e5cfb52015-01-20 15:11:33 +0800586 if (last->bbio->raid_map[0] !=
587 cur->bbio->raid_map[0])
David Woodhouse53b381b2013-01-29 18:40:14 -0500588 return 0;
589
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800590 /* we can't merge with different operations */
591 if (last->operation != cur->operation)
David Woodhouse53b381b2013-01-29 18:40:14 -0500592 return 0;
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800593 /*
594 * We've need read the full stripe from the drive.
595 * check and repair the parity and write the new results.
596 *
597 * We're not allowed to add any new bios to the
598 * bio list here, anyone else that wants to
599 * change this stripe needs to do their own rmw.
600 */
601 if (last->operation == BTRFS_RBIO_PARITY_SCRUB ||
602 cur->operation == BTRFS_RBIO_PARITY_SCRUB)
603 return 0;
David Woodhouse53b381b2013-01-29 18:40:14 -0500604
Omar Sandovalb4ee1782015-06-19 11:52:50 -0700605 if (last->operation == BTRFS_RBIO_REBUILD_MISSING ||
606 cur->operation == BTRFS_RBIO_REBUILD_MISSING)
607 return 0;
608
David Woodhouse53b381b2013-01-29 18:40:14 -0500609 return 1;
610}
611
Zhao Leib7178a52015-03-03 20:38:46 +0800612static int rbio_stripe_page_index(struct btrfs_raid_bio *rbio, int stripe,
613 int index)
614{
615 return stripe * rbio->stripe_npages + index;
616}
617
618/*
619 * these are just the pages from the rbio array, not from anything
620 * the FS sent down to us
621 */
622static struct page *rbio_stripe_page(struct btrfs_raid_bio *rbio, int stripe,
623 int index)
624{
625 return rbio->stripe_pages[rbio_stripe_page_index(rbio, stripe, index)];
626}
627
David Woodhouse53b381b2013-01-29 18:40:14 -0500628/*
629 * helper to index into the pstripe
630 */
631static struct page *rbio_pstripe_page(struct btrfs_raid_bio *rbio, int index)
632{
Zhao Leib7178a52015-03-03 20:38:46 +0800633 return rbio_stripe_page(rbio, rbio->nr_data, index);
David Woodhouse53b381b2013-01-29 18:40:14 -0500634}
635
636/*
637 * helper to index into the qstripe, returns null
638 * if there is no qstripe
639 */
640static struct page *rbio_qstripe_page(struct btrfs_raid_bio *rbio, int index)
641{
Miao Xie2c8cdd62014-11-14 16:06:25 +0800642 if (rbio->nr_data + 1 == rbio->real_stripes)
David Woodhouse53b381b2013-01-29 18:40:14 -0500643 return NULL;
Zhao Leib7178a52015-03-03 20:38:46 +0800644 return rbio_stripe_page(rbio, rbio->nr_data + 1, index);
David Woodhouse53b381b2013-01-29 18:40:14 -0500645}
646
647/*
648 * The first stripe in the table for a logical address
649 * has the lock. rbios are added in one of three ways:
650 *
651 * 1) Nobody has the stripe locked yet. The rbio is given
652 * the lock and 0 is returned. The caller must start the IO
653 * themselves.
654 *
655 * 2) Someone has the stripe locked, but we're able to merge
656 * with the lock owner. The rbio is freed and the IO will
657 * start automatically along with the existing rbio. 1 is returned.
658 *
659 * 3) Someone has the stripe locked, but we're not able to merge.
660 * The rbio is added to the lock owner's plug list, or merged into
661 * an rbio already on the plug list. When the lock owner unlocks,
662 * the next rbio on the list is run and the IO is started automatically.
663 * 1 is returned
664 *
665 * If we return 0, the caller still owns the rbio and must continue with
666 * IO submission. If we return 1, the caller must assume the rbio has
667 * already been freed.
668 */
669static noinline int lock_stripe_add(struct btrfs_raid_bio *rbio)
670{
671 int bucket = rbio_bucket(rbio);
672 struct btrfs_stripe_hash *h = rbio->fs_info->stripe_hash_table->table + bucket;
673 struct btrfs_raid_bio *cur;
674 struct btrfs_raid_bio *pending;
675 unsigned long flags;
676 DEFINE_WAIT(wait);
677 struct btrfs_raid_bio *freeit = NULL;
Chris Mason4ae10b32013-01-31 14:42:09 -0500678 struct btrfs_raid_bio *cache_drop = NULL;
David Woodhouse53b381b2013-01-29 18:40:14 -0500679 int ret = 0;
680 int walk = 0;
681
682 spin_lock_irqsave(&h->lock, flags);
683 list_for_each_entry(cur, &h->hash_list, hash_list) {
684 walk++;
Zhao Lei8e5cfb52015-01-20 15:11:33 +0800685 if (cur->bbio->raid_map[0] == rbio->bbio->raid_map[0]) {
David Woodhouse53b381b2013-01-29 18:40:14 -0500686 spin_lock(&cur->bio_list_lock);
687
Chris Mason4ae10b32013-01-31 14:42:09 -0500688 /* can we steal this cached rbio's pages? */
689 if (bio_list_empty(&cur->bio_list) &&
690 list_empty(&cur->plug_list) &&
691 test_bit(RBIO_CACHE_BIT, &cur->flags) &&
692 !test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) {
693 list_del_init(&cur->hash_list);
694 atomic_dec(&cur->refs);
695
696 steal_rbio(cur, rbio);
697 cache_drop = cur;
698 spin_unlock(&cur->bio_list_lock);
699
700 goto lockit;
701 }
702
David Woodhouse53b381b2013-01-29 18:40:14 -0500703 /* can we merge into the lock owner? */
704 if (rbio_can_merge(cur, rbio)) {
705 merge_rbio(cur, rbio);
706 spin_unlock(&cur->bio_list_lock);
707 freeit = rbio;
708 ret = 1;
709 goto out;
710 }
711
Chris Mason4ae10b32013-01-31 14:42:09 -0500712
David Woodhouse53b381b2013-01-29 18:40:14 -0500713 /*
714 * we couldn't merge with the running
715 * rbio, see if we can merge with the
716 * pending ones. We don't have to
717 * check for rmw_locked because there
718 * is no way they are inside finish_rmw
719 * right now
720 */
721 list_for_each_entry(pending, &cur->plug_list,
722 plug_list) {
723 if (rbio_can_merge(pending, rbio)) {
724 merge_rbio(pending, rbio);
725 spin_unlock(&cur->bio_list_lock);
726 freeit = rbio;
727 ret = 1;
728 goto out;
729 }
730 }
731
732 /* no merging, put us on the tail of the plug list,
733 * our rbio will be started with the currently
734 * running rbio unlocks
735 */
736 list_add_tail(&rbio->plug_list, &cur->plug_list);
737 spin_unlock(&cur->bio_list_lock);
738 ret = 1;
739 goto out;
740 }
741 }
Chris Mason4ae10b32013-01-31 14:42:09 -0500742lockit:
David Woodhouse53b381b2013-01-29 18:40:14 -0500743 atomic_inc(&rbio->refs);
744 list_add(&rbio->hash_list, &h->hash_list);
745out:
746 spin_unlock_irqrestore(&h->lock, flags);
Chris Mason4ae10b32013-01-31 14:42:09 -0500747 if (cache_drop)
748 remove_rbio_from_cache(cache_drop);
David Woodhouse53b381b2013-01-29 18:40:14 -0500749 if (freeit)
750 __free_raid_bio(freeit);
751 return ret;
752}
753
754/*
755 * called as rmw or parity rebuild is completed. If the plug list has more
756 * rbios waiting for this stripe, the next one on the list will be started
757 */
758static noinline void unlock_stripe(struct btrfs_raid_bio *rbio)
759{
760 int bucket;
761 struct btrfs_stripe_hash *h;
762 unsigned long flags;
Chris Mason4ae10b32013-01-31 14:42:09 -0500763 int keep_cache = 0;
David Woodhouse53b381b2013-01-29 18:40:14 -0500764
765 bucket = rbio_bucket(rbio);
766 h = rbio->fs_info->stripe_hash_table->table + bucket;
767
Chris Mason4ae10b32013-01-31 14:42:09 -0500768 if (list_empty(&rbio->plug_list))
769 cache_rbio(rbio);
770
David Woodhouse53b381b2013-01-29 18:40:14 -0500771 spin_lock_irqsave(&h->lock, flags);
772 spin_lock(&rbio->bio_list_lock);
773
774 if (!list_empty(&rbio->hash_list)) {
Chris Mason4ae10b32013-01-31 14:42:09 -0500775 /*
776 * if we're still cached and there is no other IO
777 * to perform, just leave this rbio here for others
778 * to steal from later
779 */
780 if (list_empty(&rbio->plug_list) &&
781 test_bit(RBIO_CACHE_BIT, &rbio->flags)) {
782 keep_cache = 1;
783 clear_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
784 BUG_ON(!bio_list_empty(&rbio->bio_list));
785 goto done;
786 }
David Woodhouse53b381b2013-01-29 18:40:14 -0500787
788 list_del_init(&rbio->hash_list);
789 atomic_dec(&rbio->refs);
790
791 /*
792 * we use the plug list to hold all the rbios
793 * waiting for the chance to lock this stripe.
794 * hand the lock over to one of them.
795 */
796 if (!list_empty(&rbio->plug_list)) {
797 struct btrfs_raid_bio *next;
798 struct list_head *head = rbio->plug_list.next;
799
800 next = list_entry(head, struct btrfs_raid_bio,
801 plug_list);
802
803 list_del_init(&rbio->plug_list);
804
805 list_add(&next->hash_list, &h->hash_list);
806 atomic_inc(&next->refs);
807 spin_unlock(&rbio->bio_list_lock);
808 spin_unlock_irqrestore(&h->lock, flags);
809
Miao Xie1b94b552014-11-06 16:14:21 +0800810 if (next->operation == BTRFS_RBIO_READ_REBUILD)
David Woodhouse53b381b2013-01-29 18:40:14 -0500811 async_read_rebuild(next);
Omar Sandovalb4ee1782015-06-19 11:52:50 -0700812 else if (next->operation == BTRFS_RBIO_REBUILD_MISSING) {
813 steal_rbio(rbio, next);
814 async_read_rebuild(next);
815 } else if (next->operation == BTRFS_RBIO_WRITE) {
Chris Mason4ae10b32013-01-31 14:42:09 -0500816 steal_rbio(rbio, next);
David Woodhouse53b381b2013-01-29 18:40:14 -0500817 async_rmw_stripe(next);
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800818 } else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) {
819 steal_rbio(rbio, next);
820 async_scrub_parity(next);
Chris Mason4ae10b32013-01-31 14:42:09 -0500821 }
David Woodhouse53b381b2013-01-29 18:40:14 -0500822
823 goto done_nolock;
David Sterba33a9eca2015-10-10 18:35:10 +0200824 /*
825 * The barrier for this waitqueue_active is not needed,
826 * we're protected by h->lock and can't miss a wakeup.
827 */
828 } else if (waitqueue_active(&h->wait)) {
David Woodhouse53b381b2013-01-29 18:40:14 -0500829 spin_unlock(&rbio->bio_list_lock);
830 spin_unlock_irqrestore(&h->lock, flags);
831 wake_up(&h->wait);
832 goto done_nolock;
833 }
834 }
Chris Mason4ae10b32013-01-31 14:42:09 -0500835done:
David Woodhouse53b381b2013-01-29 18:40:14 -0500836 spin_unlock(&rbio->bio_list_lock);
837 spin_unlock_irqrestore(&h->lock, flags);
838
839done_nolock:
Chris Mason4ae10b32013-01-31 14:42:09 -0500840 if (!keep_cache)
841 remove_rbio_from_cache(rbio);
David Woodhouse53b381b2013-01-29 18:40:14 -0500842}
843
844static void __free_raid_bio(struct btrfs_raid_bio *rbio)
845{
846 int i;
847
848 WARN_ON(atomic_read(&rbio->refs) < 0);
849 if (!atomic_dec_and_test(&rbio->refs))
850 return;
851
Chris Mason4ae10b32013-01-31 14:42:09 -0500852 WARN_ON(!list_empty(&rbio->stripe_cache));
David Woodhouse53b381b2013-01-29 18:40:14 -0500853 WARN_ON(!list_empty(&rbio->hash_list));
854 WARN_ON(!bio_list_empty(&rbio->bio_list));
855
856 for (i = 0; i < rbio->nr_pages; i++) {
857 if (rbio->stripe_pages[i]) {
858 __free_page(rbio->stripe_pages[i]);
859 rbio->stripe_pages[i] = NULL;
860 }
861 }
Miao Xieaf8e2d12014-10-23 14:42:50 +0800862
Zhao Lei6e9606d2015-01-20 15:11:34 +0800863 btrfs_put_bbio(rbio->bbio);
David Woodhouse53b381b2013-01-29 18:40:14 -0500864 kfree(rbio);
865}
866
867static void free_raid_bio(struct btrfs_raid_bio *rbio)
868{
869 unlock_stripe(rbio);
870 __free_raid_bio(rbio);
871}
872
873/*
874 * this frees the rbio and runs through all the bios in the
875 * bio_list and calls end_io on them
876 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200877static void rbio_orig_end_io(struct btrfs_raid_bio *rbio, int err)
David Woodhouse53b381b2013-01-29 18:40:14 -0500878{
879 struct bio *cur = bio_list_get(&rbio->bio_list);
880 struct bio *next;
Miao Xie42452152014-11-25 16:39:28 +0800881
882 if (rbio->generic_bio_cnt)
883 btrfs_bio_counter_sub(rbio->fs_info, rbio->generic_bio_cnt);
884
David Woodhouse53b381b2013-01-29 18:40:14 -0500885 free_raid_bio(rbio);
886
887 while (cur) {
888 next = cur->bi_next;
889 cur->bi_next = NULL;
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200890 cur->bi_error = err;
891 bio_endio(cur);
David Woodhouse53b381b2013-01-29 18:40:14 -0500892 cur = next;
893 }
894}
895
896/*
897 * end io function used by finish_rmw. When we finally
898 * get here, we've written a full stripe
899 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200900static void raid_write_end_io(struct bio *bio)
David Woodhouse53b381b2013-01-29 18:40:14 -0500901{
902 struct btrfs_raid_bio *rbio = bio->bi_private;
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200903 int err = bio->bi_error;
Zhao Leia6111d12016-01-12 17:52:13 +0800904 int max_errors;
David Woodhouse53b381b2013-01-29 18:40:14 -0500905
906 if (err)
907 fail_bio_stripe(rbio, bio);
908
909 bio_put(bio);
910
Miao Xieb89e1b02014-10-15 11:18:44 +0800911 if (!atomic_dec_and_test(&rbio->stripes_pending))
David Woodhouse53b381b2013-01-29 18:40:14 -0500912 return;
913
914 err = 0;
915
916 /* OK, we have read all the stripes we need to. */
Zhao Leia6111d12016-01-12 17:52:13 +0800917 max_errors = (rbio->operation == BTRFS_RBIO_PARITY_SCRUB) ?
918 0 : rbio->bbio->max_errors;
919 if (atomic_read(&rbio->error) > max_errors)
David Woodhouse53b381b2013-01-29 18:40:14 -0500920 err = -EIO;
921
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200922 rbio_orig_end_io(rbio, err);
David Woodhouse53b381b2013-01-29 18:40:14 -0500923}
924
925/*
926 * the read/modify/write code wants to use the original bio for
927 * any pages it included, and then use the rbio for everything
928 * else. This function decides if a given index (stripe number)
929 * and page number in that stripe fall inside the original bio
930 * or the rbio.
931 *
932 * if you set bio_list_only, you'll get a NULL back for any ranges
933 * that are outside the bio_list
934 *
935 * This doesn't take any refs on anything, you get a bare page pointer
936 * and the caller must bump refs as required.
937 *
938 * You must call index_rbio_pages once before you can trust
939 * the answers from this function.
940 */
941static struct page *page_in_rbio(struct btrfs_raid_bio *rbio,
942 int index, int pagenr, int bio_list_only)
943{
944 int chunk_page;
945 struct page *p = NULL;
946
947 chunk_page = index * (rbio->stripe_len >> PAGE_SHIFT) + pagenr;
948
949 spin_lock_irq(&rbio->bio_list_lock);
950 p = rbio->bio_pages[chunk_page];
951 spin_unlock_irq(&rbio->bio_list_lock);
952
953 if (p || bio_list_only)
954 return p;
955
956 return rbio->stripe_pages[chunk_page];
957}
958
959/*
960 * number of pages we need for the entire stripe across all the
961 * drives
962 */
963static unsigned long rbio_nr_pages(unsigned long stripe_len, int nr_stripes)
964{
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300965 return DIV_ROUND_UP(stripe_len, PAGE_SIZE) * nr_stripes;
David Woodhouse53b381b2013-01-29 18:40:14 -0500966}
967
968/*
969 * allocation and initial setup for the btrfs_raid_bio. Not
970 * this does not allocate any pages for rbio->pages.
971 */
972static struct btrfs_raid_bio *alloc_rbio(struct btrfs_root *root,
Zhao Lei8e5cfb52015-01-20 15:11:33 +0800973 struct btrfs_bio *bbio, u64 stripe_len)
David Woodhouse53b381b2013-01-29 18:40:14 -0500974{
975 struct btrfs_raid_bio *rbio;
976 int nr_data = 0;
Miao Xie2c8cdd62014-11-14 16:06:25 +0800977 int real_stripes = bbio->num_stripes - bbio->num_tgtdevs;
978 int num_pages = rbio_nr_pages(stripe_len, real_stripes);
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800979 int stripe_npages = DIV_ROUND_UP(stripe_len, PAGE_SIZE);
David Woodhouse53b381b2013-01-29 18:40:14 -0500980 void *p;
981
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800982 rbio = kzalloc(sizeof(*rbio) + num_pages * sizeof(struct page *) * 2 +
Zhao Leibfca9a62014-12-08 19:55:57 +0800983 DIV_ROUND_UP(stripe_npages, BITS_PER_LONG) *
984 sizeof(long), GFP_NOFS);
Miao Xieaf8e2d12014-10-23 14:42:50 +0800985 if (!rbio)
David Woodhouse53b381b2013-01-29 18:40:14 -0500986 return ERR_PTR(-ENOMEM);
David Woodhouse53b381b2013-01-29 18:40:14 -0500987
988 bio_list_init(&rbio->bio_list);
989 INIT_LIST_HEAD(&rbio->plug_list);
990 spin_lock_init(&rbio->bio_list_lock);
Chris Mason4ae10b32013-01-31 14:42:09 -0500991 INIT_LIST_HEAD(&rbio->stripe_cache);
David Woodhouse53b381b2013-01-29 18:40:14 -0500992 INIT_LIST_HEAD(&rbio->hash_list);
993 rbio->bbio = bbio;
David Woodhouse53b381b2013-01-29 18:40:14 -0500994 rbio->fs_info = root->fs_info;
995 rbio->stripe_len = stripe_len;
996 rbio->nr_pages = num_pages;
Miao Xie2c8cdd62014-11-14 16:06:25 +0800997 rbio->real_stripes = real_stripes;
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800998 rbio->stripe_npages = stripe_npages;
David Woodhouse53b381b2013-01-29 18:40:14 -0500999 rbio->faila = -1;
1000 rbio->failb = -1;
1001 atomic_set(&rbio->refs, 1);
Miao Xieb89e1b02014-10-15 11:18:44 +08001002 atomic_set(&rbio->error, 0);
1003 atomic_set(&rbio->stripes_pending, 0);
David Woodhouse53b381b2013-01-29 18:40:14 -05001004
1005 /*
1006 * the stripe_pages and bio_pages array point to the extra
1007 * memory we allocated past the end of the rbio
1008 */
1009 p = rbio + 1;
1010 rbio->stripe_pages = p;
1011 rbio->bio_pages = p + sizeof(struct page *) * num_pages;
Miao Xie5a6ac9e2014-11-06 17:20:58 +08001012 rbio->dbitmap = p + sizeof(struct page *) * num_pages * 2;
David Woodhouse53b381b2013-01-29 18:40:14 -05001013
Zhao Lei10f11902015-01-20 15:11:43 +08001014 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1015 nr_data = real_stripes - 1;
1016 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
Miao Xie2c8cdd62014-11-14 16:06:25 +08001017 nr_data = real_stripes - 2;
David Woodhouse53b381b2013-01-29 18:40:14 -05001018 else
Zhao Lei10f11902015-01-20 15:11:43 +08001019 BUG();
David Woodhouse53b381b2013-01-29 18:40:14 -05001020
1021 rbio->nr_data = nr_data;
1022 return rbio;
1023}
1024
1025/* allocate pages for all the stripes in the bio, including parity */
1026static int alloc_rbio_pages(struct btrfs_raid_bio *rbio)
1027{
1028 int i;
1029 struct page *page;
1030
1031 for (i = 0; i < rbio->nr_pages; i++) {
1032 if (rbio->stripe_pages[i])
1033 continue;
1034 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1035 if (!page)
1036 return -ENOMEM;
1037 rbio->stripe_pages[i] = page;
David Woodhouse53b381b2013-01-29 18:40:14 -05001038 }
1039 return 0;
1040}
1041
Zhao Leib7178a52015-03-03 20:38:46 +08001042/* only allocate pages for p/q stripes */
David Woodhouse53b381b2013-01-29 18:40:14 -05001043static int alloc_rbio_parity_pages(struct btrfs_raid_bio *rbio)
1044{
1045 int i;
1046 struct page *page;
1047
Zhao Leib7178a52015-03-03 20:38:46 +08001048 i = rbio_stripe_page_index(rbio, rbio->nr_data, 0);
David Woodhouse53b381b2013-01-29 18:40:14 -05001049
1050 for (; i < rbio->nr_pages; i++) {
1051 if (rbio->stripe_pages[i])
1052 continue;
1053 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1054 if (!page)
1055 return -ENOMEM;
1056 rbio->stripe_pages[i] = page;
1057 }
1058 return 0;
1059}
1060
1061/*
1062 * add a single page from a specific stripe into our list of bios for IO
1063 * this will try to merge into existing bios if possible, and returns
1064 * zero if all went well.
1065 */
Eric Sandeen48a3b632013-04-25 20:41:01 +00001066static int rbio_add_io_page(struct btrfs_raid_bio *rbio,
1067 struct bio_list *bio_list,
1068 struct page *page,
1069 int stripe_nr,
1070 unsigned long page_index,
1071 unsigned long bio_max_len)
David Woodhouse53b381b2013-01-29 18:40:14 -05001072{
1073 struct bio *last = bio_list->tail;
1074 u64 last_end = 0;
1075 int ret;
1076 struct bio *bio;
1077 struct btrfs_bio_stripe *stripe;
1078 u64 disk_start;
1079
1080 stripe = &rbio->bbio->stripes[stripe_nr];
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001081 disk_start = stripe->physical + (page_index << PAGE_SHIFT);
David Woodhouse53b381b2013-01-29 18:40:14 -05001082
1083 /* if the device is missing, just fail this stripe */
1084 if (!stripe->dev->bdev)
1085 return fail_rbio_index(rbio, stripe_nr);
1086
1087 /* see if we can add this page onto our existing bio */
1088 if (last) {
Kent Overstreet4f024f32013-10-11 15:44:27 -07001089 last_end = (u64)last->bi_iter.bi_sector << 9;
1090 last_end += last->bi_iter.bi_size;
David Woodhouse53b381b2013-01-29 18:40:14 -05001091
1092 /*
1093 * we can't merge these if they are from different
1094 * devices or if they are not contiguous
1095 */
1096 if (last_end == disk_start && stripe->dev->bdev &&
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001097 !last->bi_error &&
David Woodhouse53b381b2013-01-29 18:40:14 -05001098 last->bi_bdev == stripe->dev->bdev) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001099 ret = bio_add_page(last, page, PAGE_SIZE, 0);
1100 if (ret == PAGE_SIZE)
David Woodhouse53b381b2013-01-29 18:40:14 -05001101 return 0;
1102 }
1103 }
1104
1105 /* put a new bio on the list */
Chris Mason9be33952013-05-17 18:30:14 -04001106 bio = btrfs_io_bio_alloc(GFP_NOFS, bio_max_len >> PAGE_SHIFT?:1);
David Woodhouse53b381b2013-01-29 18:40:14 -05001107 if (!bio)
1108 return -ENOMEM;
1109
Kent Overstreet4f024f32013-10-11 15:44:27 -07001110 bio->bi_iter.bi_size = 0;
David Woodhouse53b381b2013-01-29 18:40:14 -05001111 bio->bi_bdev = stripe->dev->bdev;
Kent Overstreet4f024f32013-10-11 15:44:27 -07001112 bio->bi_iter.bi_sector = disk_start >> 9;
David Woodhouse53b381b2013-01-29 18:40:14 -05001113
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001114 bio_add_page(bio, page, PAGE_SIZE, 0);
David Woodhouse53b381b2013-01-29 18:40:14 -05001115 bio_list_add(bio_list, bio);
1116 return 0;
1117}
1118
1119/*
1120 * while we're doing the read/modify/write cycle, we could
1121 * have errors in reading pages off the disk. This checks
1122 * for errors and if we're not able to read the page it'll
1123 * trigger parity reconstruction. The rmw will be finished
1124 * after we've reconstructed the failed stripes
1125 */
1126static void validate_rbio_for_rmw(struct btrfs_raid_bio *rbio)
1127{
1128 if (rbio->faila >= 0 || rbio->failb >= 0) {
Miao Xie2c8cdd62014-11-14 16:06:25 +08001129 BUG_ON(rbio->faila == rbio->real_stripes - 1);
David Woodhouse53b381b2013-01-29 18:40:14 -05001130 __raid56_parity_recover(rbio);
1131 } else {
1132 finish_rmw(rbio);
1133 }
1134}
1135
1136/*
David Woodhouse53b381b2013-01-29 18:40:14 -05001137 * helper function to walk our bio list and populate the bio_pages array with
1138 * the result. This seems expensive, but it is faster than constantly
1139 * searching through the bio list as we setup the IO in finish_rmw or stripe
1140 * reconstruction.
1141 *
1142 * This must be called before you trust the answers from page_in_rbio
1143 */
1144static void index_rbio_pages(struct btrfs_raid_bio *rbio)
1145{
1146 struct bio *bio;
1147 u64 start;
1148 unsigned long stripe_offset;
1149 unsigned long page_index;
1150 struct page *p;
1151 int i;
1152
1153 spin_lock_irq(&rbio->bio_list_lock);
1154 bio_list_for_each(bio, &rbio->bio_list) {
Kent Overstreet4f024f32013-10-11 15:44:27 -07001155 start = (u64)bio->bi_iter.bi_sector << 9;
Zhao Lei8e5cfb52015-01-20 15:11:33 +08001156 stripe_offset = start - rbio->bbio->raid_map[0];
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001157 page_index = stripe_offset >> PAGE_SHIFT;
David Woodhouse53b381b2013-01-29 18:40:14 -05001158
1159 for (i = 0; i < bio->bi_vcnt; i++) {
1160 p = bio->bi_io_vec[i].bv_page;
1161 rbio->bio_pages[page_index + i] = p;
1162 }
1163 }
1164 spin_unlock_irq(&rbio->bio_list_lock);
1165}
1166
1167/*
1168 * this is called from one of two situations. We either
1169 * have a full stripe from the higher layers, or we've read all
1170 * the missing bits off disk.
1171 *
1172 * This will calculate the parity and then send down any
1173 * changed blocks.
1174 */
1175static noinline void finish_rmw(struct btrfs_raid_bio *rbio)
1176{
1177 struct btrfs_bio *bbio = rbio->bbio;
Miao Xie2c8cdd62014-11-14 16:06:25 +08001178 void *pointers[rbio->real_stripes];
David Woodhouse53b381b2013-01-29 18:40:14 -05001179 int nr_data = rbio->nr_data;
1180 int stripe;
1181 int pagenr;
1182 int p_stripe = -1;
1183 int q_stripe = -1;
1184 struct bio_list bio_list;
1185 struct bio *bio;
David Woodhouse53b381b2013-01-29 18:40:14 -05001186 int ret;
1187
1188 bio_list_init(&bio_list);
1189
Miao Xie2c8cdd62014-11-14 16:06:25 +08001190 if (rbio->real_stripes - rbio->nr_data == 1) {
1191 p_stripe = rbio->real_stripes - 1;
1192 } else if (rbio->real_stripes - rbio->nr_data == 2) {
1193 p_stripe = rbio->real_stripes - 2;
1194 q_stripe = rbio->real_stripes - 1;
David Woodhouse53b381b2013-01-29 18:40:14 -05001195 } else {
1196 BUG();
1197 }
1198
1199 /* at this point we either have a full stripe,
1200 * or we've read the full stripe from the drive.
1201 * recalculate the parity and write the new results.
1202 *
1203 * We're not allowed to add any new bios to the
1204 * bio list here, anyone else that wants to
1205 * change this stripe needs to do their own rmw.
1206 */
1207 spin_lock_irq(&rbio->bio_list_lock);
1208 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1209 spin_unlock_irq(&rbio->bio_list_lock);
1210
Miao Xieb89e1b02014-10-15 11:18:44 +08001211 atomic_set(&rbio->error, 0);
David Woodhouse53b381b2013-01-29 18:40:14 -05001212
1213 /*
1214 * now that we've set rmw_locked, run through the
1215 * bio list one last time and map the page pointers
Chris Mason4ae10b32013-01-31 14:42:09 -05001216 *
1217 * We don't cache full rbios because we're assuming
1218 * the higher layers are unlikely to use this area of
1219 * the disk again soon. If they do use it again,
1220 * hopefully they will send another full bio.
David Woodhouse53b381b2013-01-29 18:40:14 -05001221 */
1222 index_rbio_pages(rbio);
Chris Mason4ae10b32013-01-31 14:42:09 -05001223 if (!rbio_is_full(rbio))
1224 cache_rbio_pages(rbio);
1225 else
1226 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
David Woodhouse53b381b2013-01-29 18:40:14 -05001227
Zhao Lei915e2292015-03-03 20:42:48 +08001228 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001229 struct page *p;
1230 /* first collect one page from each data stripe */
1231 for (stripe = 0; stripe < nr_data; stripe++) {
1232 p = page_in_rbio(rbio, stripe, pagenr, 0);
1233 pointers[stripe] = kmap(p);
1234 }
1235
1236 /* then add the parity stripe */
1237 p = rbio_pstripe_page(rbio, pagenr);
1238 SetPageUptodate(p);
1239 pointers[stripe++] = kmap(p);
1240
1241 if (q_stripe != -1) {
1242
1243 /*
1244 * raid6, add the qstripe and call the
1245 * library function to fill in our p/q
1246 */
1247 p = rbio_qstripe_page(rbio, pagenr);
1248 SetPageUptodate(p);
1249 pointers[stripe++] = kmap(p);
1250
Miao Xie2c8cdd62014-11-14 16:06:25 +08001251 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
David Woodhouse53b381b2013-01-29 18:40:14 -05001252 pointers);
1253 } else {
1254 /* raid5 */
1255 memcpy(pointers[nr_data], pointers[0], PAGE_SIZE);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001256 run_xor(pointers + 1, nr_data - 1, PAGE_SIZE);
David Woodhouse53b381b2013-01-29 18:40:14 -05001257 }
1258
1259
Miao Xie2c8cdd62014-11-14 16:06:25 +08001260 for (stripe = 0; stripe < rbio->real_stripes; stripe++)
David Woodhouse53b381b2013-01-29 18:40:14 -05001261 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
1262 }
1263
1264 /*
1265 * time to start writing. Make bios for everything from the
1266 * higher layers (the bio_list in our rbio) and our p/q. Ignore
1267 * everything else.
1268 */
Miao Xie2c8cdd62014-11-14 16:06:25 +08001269 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
Zhao Lei915e2292015-03-03 20:42:48 +08001270 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001271 struct page *page;
1272 if (stripe < rbio->nr_data) {
1273 page = page_in_rbio(rbio, stripe, pagenr, 1);
1274 if (!page)
1275 continue;
1276 } else {
1277 page = rbio_stripe_page(rbio, stripe, pagenr);
1278 }
1279
1280 ret = rbio_add_io_page(rbio, &bio_list,
1281 page, stripe, pagenr, rbio->stripe_len);
1282 if (ret)
1283 goto cleanup;
1284 }
1285 }
1286
Miao Xie2c8cdd62014-11-14 16:06:25 +08001287 if (likely(!bbio->num_tgtdevs))
1288 goto write_data;
1289
1290 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
1291 if (!bbio->tgtdev_map[stripe])
1292 continue;
1293
Zhao Lei915e2292015-03-03 20:42:48 +08001294 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
Miao Xie2c8cdd62014-11-14 16:06:25 +08001295 struct page *page;
1296 if (stripe < rbio->nr_data) {
1297 page = page_in_rbio(rbio, stripe, pagenr, 1);
1298 if (!page)
1299 continue;
1300 } else {
1301 page = rbio_stripe_page(rbio, stripe, pagenr);
1302 }
1303
1304 ret = rbio_add_io_page(rbio, &bio_list, page,
1305 rbio->bbio->tgtdev_map[stripe],
1306 pagenr, rbio->stripe_len);
1307 if (ret)
1308 goto cleanup;
1309 }
1310 }
1311
1312write_data:
Miao Xieb89e1b02014-10-15 11:18:44 +08001313 atomic_set(&rbio->stripes_pending, bio_list_size(&bio_list));
1314 BUG_ON(atomic_read(&rbio->stripes_pending) == 0);
David Woodhouse53b381b2013-01-29 18:40:14 -05001315
1316 while (1) {
1317 bio = bio_list_pop(&bio_list);
1318 if (!bio)
1319 break;
1320
1321 bio->bi_private = rbio;
1322 bio->bi_end_io = raid_write_end_io;
Mike Christie37226b22016-06-05 14:31:52 -05001323 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
Mike Christie4e49ea42016-06-05 14:31:41 -05001324
1325 submit_bio(bio);
David Woodhouse53b381b2013-01-29 18:40:14 -05001326 }
1327 return;
1328
1329cleanup:
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001330 rbio_orig_end_io(rbio, -EIO);
David Woodhouse53b381b2013-01-29 18:40:14 -05001331}
1332
1333/*
1334 * helper to find the stripe number for a given bio. Used to figure out which
1335 * stripe has failed. This expects the bio to correspond to a physical disk,
1336 * so it looks up based on physical sector numbers.
1337 */
1338static int find_bio_stripe(struct btrfs_raid_bio *rbio,
1339 struct bio *bio)
1340{
Kent Overstreet4f024f32013-10-11 15:44:27 -07001341 u64 physical = bio->bi_iter.bi_sector;
David Woodhouse53b381b2013-01-29 18:40:14 -05001342 u64 stripe_start;
1343 int i;
1344 struct btrfs_bio_stripe *stripe;
1345
1346 physical <<= 9;
1347
1348 for (i = 0; i < rbio->bbio->num_stripes; i++) {
1349 stripe = &rbio->bbio->stripes[i];
1350 stripe_start = stripe->physical;
1351 if (physical >= stripe_start &&
Miao Xie2c8cdd62014-11-14 16:06:25 +08001352 physical < stripe_start + rbio->stripe_len &&
1353 bio->bi_bdev == stripe->dev->bdev) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001354 return i;
1355 }
1356 }
1357 return -1;
1358}
1359
1360/*
1361 * helper to find the stripe number for a given
1362 * bio (before mapping). Used to figure out which stripe has
1363 * failed. This looks up based on logical block numbers.
1364 */
1365static int find_logical_bio_stripe(struct btrfs_raid_bio *rbio,
1366 struct bio *bio)
1367{
Kent Overstreet4f024f32013-10-11 15:44:27 -07001368 u64 logical = bio->bi_iter.bi_sector;
David Woodhouse53b381b2013-01-29 18:40:14 -05001369 u64 stripe_start;
1370 int i;
1371
1372 logical <<= 9;
1373
1374 for (i = 0; i < rbio->nr_data; i++) {
Zhao Lei8e5cfb52015-01-20 15:11:33 +08001375 stripe_start = rbio->bbio->raid_map[i];
David Woodhouse53b381b2013-01-29 18:40:14 -05001376 if (logical >= stripe_start &&
1377 logical < stripe_start + rbio->stripe_len) {
1378 return i;
1379 }
1380 }
1381 return -1;
1382}
1383
1384/*
1385 * returns -EIO if we had too many failures
1386 */
1387static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed)
1388{
1389 unsigned long flags;
1390 int ret = 0;
1391
1392 spin_lock_irqsave(&rbio->bio_list_lock, flags);
1393
1394 /* we already know this stripe is bad, move on */
1395 if (rbio->faila == failed || rbio->failb == failed)
1396 goto out;
1397
1398 if (rbio->faila == -1) {
1399 /* first failure on this rbio */
1400 rbio->faila = failed;
Miao Xieb89e1b02014-10-15 11:18:44 +08001401 atomic_inc(&rbio->error);
David Woodhouse53b381b2013-01-29 18:40:14 -05001402 } else if (rbio->failb == -1) {
1403 /* second failure on this rbio */
1404 rbio->failb = failed;
Miao Xieb89e1b02014-10-15 11:18:44 +08001405 atomic_inc(&rbio->error);
David Woodhouse53b381b2013-01-29 18:40:14 -05001406 } else {
1407 ret = -EIO;
1408 }
1409out:
1410 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
1411
1412 return ret;
1413}
1414
1415/*
1416 * helper to fail a stripe based on a physical disk
1417 * bio.
1418 */
1419static int fail_bio_stripe(struct btrfs_raid_bio *rbio,
1420 struct bio *bio)
1421{
1422 int failed = find_bio_stripe(rbio, bio);
1423
1424 if (failed < 0)
1425 return -EIO;
1426
1427 return fail_rbio_index(rbio, failed);
1428}
1429
1430/*
1431 * this sets each page in the bio uptodate. It should only be used on private
1432 * rbio pages, nothing that comes in from the higher layers
1433 */
1434static void set_bio_pages_uptodate(struct bio *bio)
1435{
1436 int i;
1437 struct page *p;
1438
1439 for (i = 0; i < bio->bi_vcnt; i++) {
1440 p = bio->bi_io_vec[i].bv_page;
1441 SetPageUptodate(p);
1442 }
1443}
1444
1445/*
1446 * end io for the read phase of the rmw cycle. All the bios here are physical
1447 * stripe bios we've read from the disk so we can recalculate the parity of the
1448 * stripe.
1449 *
1450 * This will usually kick off finish_rmw once all the bios are read in, but it
1451 * may trigger parity reconstruction if we had any errors along the way
1452 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001453static void raid_rmw_end_io(struct bio *bio)
David Woodhouse53b381b2013-01-29 18:40:14 -05001454{
1455 struct btrfs_raid_bio *rbio = bio->bi_private;
1456
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001457 if (bio->bi_error)
David Woodhouse53b381b2013-01-29 18:40:14 -05001458 fail_bio_stripe(rbio, bio);
1459 else
1460 set_bio_pages_uptodate(bio);
1461
1462 bio_put(bio);
1463
Miao Xieb89e1b02014-10-15 11:18:44 +08001464 if (!atomic_dec_and_test(&rbio->stripes_pending))
David Woodhouse53b381b2013-01-29 18:40:14 -05001465 return;
1466
Miao Xieb89e1b02014-10-15 11:18:44 +08001467 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
David Woodhouse53b381b2013-01-29 18:40:14 -05001468 goto cleanup;
1469
1470 /*
1471 * this will normally call finish_rmw to start our write
1472 * but if there are any failed stripes we'll reconstruct
1473 * from parity first
1474 */
1475 validate_rbio_for_rmw(rbio);
1476 return;
1477
1478cleanup:
1479
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001480 rbio_orig_end_io(rbio, -EIO);
David Woodhouse53b381b2013-01-29 18:40:14 -05001481}
1482
1483static void async_rmw_stripe(struct btrfs_raid_bio *rbio)
1484{
Liu Bo9e0af232014-08-15 23:36:53 +08001485 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
1486 rmw_work, NULL, NULL);
David Woodhouse53b381b2013-01-29 18:40:14 -05001487
Qu Wenruod05a33a2014-02-28 10:46:11 +08001488 btrfs_queue_work(rbio->fs_info->rmw_workers,
1489 &rbio->work);
David Woodhouse53b381b2013-01-29 18:40:14 -05001490}
1491
1492static void async_read_rebuild(struct btrfs_raid_bio *rbio)
1493{
Liu Bo9e0af232014-08-15 23:36:53 +08001494 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
1495 read_rebuild_work, NULL, NULL);
David Woodhouse53b381b2013-01-29 18:40:14 -05001496
Qu Wenruod05a33a2014-02-28 10:46:11 +08001497 btrfs_queue_work(rbio->fs_info->rmw_workers,
1498 &rbio->work);
David Woodhouse53b381b2013-01-29 18:40:14 -05001499}
1500
1501/*
1502 * the stripe must be locked by the caller. It will
1503 * unlock after all the writes are done
1504 */
1505static int raid56_rmw_stripe(struct btrfs_raid_bio *rbio)
1506{
1507 int bios_to_read = 0;
David Woodhouse53b381b2013-01-29 18:40:14 -05001508 struct bio_list bio_list;
1509 int ret;
David Woodhouse53b381b2013-01-29 18:40:14 -05001510 int pagenr;
1511 int stripe;
1512 struct bio *bio;
1513
1514 bio_list_init(&bio_list);
1515
1516 ret = alloc_rbio_pages(rbio);
1517 if (ret)
1518 goto cleanup;
1519
1520 index_rbio_pages(rbio);
1521
Miao Xieb89e1b02014-10-15 11:18:44 +08001522 atomic_set(&rbio->error, 0);
David Woodhouse53b381b2013-01-29 18:40:14 -05001523 /*
1524 * build a list of bios to read all the missing parts of this
1525 * stripe
1526 */
1527 for (stripe = 0; stripe < rbio->nr_data; stripe++) {
Zhao Lei915e2292015-03-03 20:42:48 +08001528 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001529 struct page *page;
1530 /*
1531 * we want to find all the pages missing from
1532 * the rbio and read them from the disk. If
1533 * page_in_rbio finds a page in the bio list
1534 * we don't need to read it off the stripe.
1535 */
1536 page = page_in_rbio(rbio, stripe, pagenr, 1);
1537 if (page)
1538 continue;
1539
1540 page = rbio_stripe_page(rbio, stripe, pagenr);
Chris Mason4ae10b32013-01-31 14:42:09 -05001541 /*
1542 * the bio cache may have handed us an uptodate
1543 * page. If so, be happy and use it
1544 */
1545 if (PageUptodate(page))
1546 continue;
1547
David Woodhouse53b381b2013-01-29 18:40:14 -05001548 ret = rbio_add_io_page(rbio, &bio_list, page,
1549 stripe, pagenr, rbio->stripe_len);
1550 if (ret)
1551 goto cleanup;
1552 }
1553 }
1554
1555 bios_to_read = bio_list_size(&bio_list);
1556 if (!bios_to_read) {
1557 /*
1558 * this can happen if others have merged with
1559 * us, it means there is nothing left to read.
1560 * But if there are missing devices it may not be
1561 * safe to do the full stripe write yet.
1562 */
1563 goto finish;
1564 }
1565
1566 /*
1567 * the bbio may be freed once we submit the last bio. Make sure
1568 * not to touch it after that
1569 */
Miao Xieb89e1b02014-10-15 11:18:44 +08001570 atomic_set(&rbio->stripes_pending, bios_to_read);
David Woodhouse53b381b2013-01-29 18:40:14 -05001571 while (1) {
1572 bio = bio_list_pop(&bio_list);
1573 if (!bio)
1574 break;
1575
1576 bio->bi_private = rbio;
1577 bio->bi_end_io = raid_rmw_end_io;
Mike Christie37226b22016-06-05 14:31:52 -05001578 bio_set_op_attrs(bio, REQ_OP_READ, 0);
David Woodhouse53b381b2013-01-29 18:40:14 -05001579
1580 btrfs_bio_wq_end_io(rbio->fs_info, bio,
1581 BTRFS_WQ_ENDIO_RAID56);
1582
Mike Christie4e49ea42016-06-05 14:31:41 -05001583 submit_bio(bio);
David Woodhouse53b381b2013-01-29 18:40:14 -05001584 }
1585 /* the actual write will happen once the reads are done */
1586 return 0;
1587
1588cleanup:
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001589 rbio_orig_end_io(rbio, -EIO);
David Woodhouse53b381b2013-01-29 18:40:14 -05001590 return -EIO;
1591
1592finish:
1593 validate_rbio_for_rmw(rbio);
1594 return 0;
1595}
1596
1597/*
1598 * if the upper layers pass in a full stripe, we thank them by only allocating
1599 * enough pages to hold the parity, and sending it all down quickly.
1600 */
1601static int full_stripe_write(struct btrfs_raid_bio *rbio)
1602{
1603 int ret;
1604
1605 ret = alloc_rbio_parity_pages(rbio);
Miao Xie3cd846d2013-07-22 16:36:57 +08001606 if (ret) {
1607 __free_raid_bio(rbio);
David Woodhouse53b381b2013-01-29 18:40:14 -05001608 return ret;
Miao Xie3cd846d2013-07-22 16:36:57 +08001609 }
David Woodhouse53b381b2013-01-29 18:40:14 -05001610
1611 ret = lock_stripe_add(rbio);
1612 if (ret == 0)
1613 finish_rmw(rbio);
1614 return 0;
1615}
1616
1617/*
1618 * partial stripe writes get handed over to async helpers.
1619 * We're really hoping to merge a few more writes into this
1620 * rbio before calculating new parity
1621 */
1622static int partial_stripe_write(struct btrfs_raid_bio *rbio)
1623{
1624 int ret;
1625
1626 ret = lock_stripe_add(rbio);
1627 if (ret == 0)
1628 async_rmw_stripe(rbio);
1629 return 0;
1630}
1631
1632/*
1633 * sometimes while we were reading from the drive to
1634 * recalculate parity, enough new bios come into create
1635 * a full stripe. So we do a check here to see if we can
1636 * go directly to finish_rmw
1637 */
1638static int __raid56_parity_write(struct btrfs_raid_bio *rbio)
1639{
1640 /* head off into rmw land if we don't have a full stripe */
1641 if (!rbio_is_full(rbio))
1642 return partial_stripe_write(rbio);
1643 return full_stripe_write(rbio);
1644}
1645
1646/*
Chris Mason6ac0f482013-01-31 14:42:28 -05001647 * We use plugging call backs to collect full stripes.
1648 * Any time we get a partial stripe write while plugged
1649 * we collect it into a list. When the unplug comes down,
1650 * we sort the list by logical block number and merge
1651 * everything we can into the same rbios
1652 */
1653struct btrfs_plug_cb {
1654 struct blk_plug_cb cb;
1655 struct btrfs_fs_info *info;
1656 struct list_head rbio_list;
1657 struct btrfs_work work;
1658};
1659
1660/*
1661 * rbios on the plug list are sorted for easier merging.
1662 */
1663static int plug_cmp(void *priv, struct list_head *a, struct list_head *b)
1664{
1665 struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio,
1666 plug_list);
1667 struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio,
1668 plug_list);
Kent Overstreet4f024f32013-10-11 15:44:27 -07001669 u64 a_sector = ra->bio_list.head->bi_iter.bi_sector;
1670 u64 b_sector = rb->bio_list.head->bi_iter.bi_sector;
Chris Mason6ac0f482013-01-31 14:42:28 -05001671
1672 if (a_sector < b_sector)
1673 return -1;
1674 if (a_sector > b_sector)
1675 return 1;
1676 return 0;
1677}
1678
1679static void run_plug(struct btrfs_plug_cb *plug)
1680{
1681 struct btrfs_raid_bio *cur;
1682 struct btrfs_raid_bio *last = NULL;
1683
1684 /*
1685 * sort our plug list then try to merge
1686 * everything we can in hopes of creating full
1687 * stripes.
1688 */
1689 list_sort(NULL, &plug->rbio_list, plug_cmp);
1690 while (!list_empty(&plug->rbio_list)) {
1691 cur = list_entry(plug->rbio_list.next,
1692 struct btrfs_raid_bio, plug_list);
1693 list_del_init(&cur->plug_list);
1694
1695 if (rbio_is_full(cur)) {
1696 /* we have a full stripe, send it down */
1697 full_stripe_write(cur);
1698 continue;
1699 }
1700 if (last) {
1701 if (rbio_can_merge(last, cur)) {
1702 merge_rbio(last, cur);
1703 __free_raid_bio(cur);
1704 continue;
1705
1706 }
1707 __raid56_parity_write(last);
1708 }
1709 last = cur;
1710 }
1711 if (last) {
1712 __raid56_parity_write(last);
1713 }
1714 kfree(plug);
1715}
1716
1717/*
1718 * if the unplug comes from schedule, we have to push the
1719 * work off to a helper thread
1720 */
1721static void unplug_work(struct btrfs_work *work)
1722{
1723 struct btrfs_plug_cb *plug;
1724 plug = container_of(work, struct btrfs_plug_cb, work);
1725 run_plug(plug);
1726}
1727
1728static void btrfs_raid_unplug(struct blk_plug_cb *cb, bool from_schedule)
1729{
1730 struct btrfs_plug_cb *plug;
1731 plug = container_of(cb, struct btrfs_plug_cb, cb);
1732
1733 if (from_schedule) {
Liu Bo9e0af232014-08-15 23:36:53 +08001734 btrfs_init_work(&plug->work, btrfs_rmw_helper,
1735 unplug_work, NULL, NULL);
Qu Wenruod05a33a2014-02-28 10:46:11 +08001736 btrfs_queue_work(plug->info->rmw_workers,
1737 &plug->work);
Chris Mason6ac0f482013-01-31 14:42:28 -05001738 return;
1739 }
1740 run_plug(plug);
1741}
1742
1743/*
David Woodhouse53b381b2013-01-29 18:40:14 -05001744 * our main entry point for writes from the rest of the FS.
1745 */
1746int raid56_parity_write(struct btrfs_root *root, struct bio *bio,
Zhao Lei8e5cfb52015-01-20 15:11:33 +08001747 struct btrfs_bio *bbio, u64 stripe_len)
David Woodhouse53b381b2013-01-29 18:40:14 -05001748{
1749 struct btrfs_raid_bio *rbio;
Chris Mason6ac0f482013-01-31 14:42:28 -05001750 struct btrfs_plug_cb *plug = NULL;
1751 struct blk_plug_cb *cb;
Miao Xie42452152014-11-25 16:39:28 +08001752 int ret;
David Woodhouse53b381b2013-01-29 18:40:14 -05001753
Zhao Lei8e5cfb52015-01-20 15:11:33 +08001754 rbio = alloc_rbio(root, bbio, stripe_len);
Miao Xieaf8e2d12014-10-23 14:42:50 +08001755 if (IS_ERR(rbio)) {
Zhao Lei6e9606d2015-01-20 15:11:34 +08001756 btrfs_put_bbio(bbio);
David Woodhouse53b381b2013-01-29 18:40:14 -05001757 return PTR_ERR(rbio);
Miao Xieaf8e2d12014-10-23 14:42:50 +08001758 }
David Woodhouse53b381b2013-01-29 18:40:14 -05001759 bio_list_add(&rbio->bio_list, bio);
Kent Overstreet4f024f32013-10-11 15:44:27 -07001760 rbio->bio_list_bytes = bio->bi_iter.bi_size;
Miao Xie1b94b552014-11-06 16:14:21 +08001761 rbio->operation = BTRFS_RBIO_WRITE;
Chris Mason6ac0f482013-01-31 14:42:28 -05001762
Miao Xie42452152014-11-25 16:39:28 +08001763 btrfs_bio_counter_inc_noblocked(root->fs_info);
1764 rbio->generic_bio_cnt = 1;
1765
Chris Mason6ac0f482013-01-31 14:42:28 -05001766 /*
1767 * don't plug on full rbios, just get them out the door
1768 * as quickly as we can
1769 */
Miao Xie42452152014-11-25 16:39:28 +08001770 if (rbio_is_full(rbio)) {
1771 ret = full_stripe_write(rbio);
1772 if (ret)
1773 btrfs_bio_counter_dec(root->fs_info);
1774 return ret;
1775 }
Chris Mason6ac0f482013-01-31 14:42:28 -05001776
1777 cb = blk_check_plugged(btrfs_raid_unplug, root->fs_info,
1778 sizeof(*plug));
1779 if (cb) {
1780 plug = container_of(cb, struct btrfs_plug_cb, cb);
1781 if (!plug->info) {
1782 plug->info = root->fs_info;
1783 INIT_LIST_HEAD(&plug->rbio_list);
1784 }
1785 list_add_tail(&rbio->plug_list, &plug->rbio_list);
Miao Xie42452152014-11-25 16:39:28 +08001786 ret = 0;
Chris Mason6ac0f482013-01-31 14:42:28 -05001787 } else {
Miao Xie42452152014-11-25 16:39:28 +08001788 ret = __raid56_parity_write(rbio);
1789 if (ret)
1790 btrfs_bio_counter_dec(root->fs_info);
Chris Mason6ac0f482013-01-31 14:42:28 -05001791 }
Miao Xie42452152014-11-25 16:39:28 +08001792 return ret;
David Woodhouse53b381b2013-01-29 18:40:14 -05001793}
1794
1795/*
1796 * all parity reconstruction happens here. We've read in everything
1797 * we can find from the drives and this does the heavy lifting of
1798 * sorting the good from the bad.
1799 */
1800static void __raid_recover_end_io(struct btrfs_raid_bio *rbio)
1801{
1802 int pagenr, stripe;
1803 void **pointers;
1804 int faila = -1, failb = -1;
David Woodhouse53b381b2013-01-29 18:40:14 -05001805 struct page *page;
1806 int err;
1807 int i;
1808
David Sterba31e818f2015-02-20 18:00:26 +01001809 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
David Woodhouse53b381b2013-01-29 18:40:14 -05001810 if (!pointers) {
1811 err = -ENOMEM;
1812 goto cleanup_io;
1813 }
1814
1815 faila = rbio->faila;
1816 failb = rbio->failb;
1817
Omar Sandovalb4ee1782015-06-19 11:52:50 -07001818 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1819 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001820 spin_lock_irq(&rbio->bio_list_lock);
1821 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1822 spin_unlock_irq(&rbio->bio_list_lock);
1823 }
1824
1825 index_rbio_pages(rbio);
1826
Zhao Lei915e2292015-03-03 20:42:48 +08001827 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
Miao Xie5a6ac9e2014-11-06 17:20:58 +08001828 /*
1829 * Now we just use bitmap to mark the horizontal stripes in
1830 * which we have data when doing parity scrub.
1831 */
1832 if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB &&
1833 !test_bit(pagenr, rbio->dbitmap))
1834 continue;
1835
David Woodhouse53b381b2013-01-29 18:40:14 -05001836 /* setup our array of pointers with pages
1837 * from each stripe
1838 */
Miao Xie2c8cdd62014-11-14 16:06:25 +08001839 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001840 /*
1841 * if we're rebuilding a read, we have to use
1842 * pages from the bio list
1843 */
Omar Sandovalb4ee1782015-06-19 11:52:50 -07001844 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1845 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
David Woodhouse53b381b2013-01-29 18:40:14 -05001846 (stripe == faila || stripe == failb)) {
1847 page = page_in_rbio(rbio, stripe, pagenr, 0);
1848 } else {
1849 page = rbio_stripe_page(rbio, stripe, pagenr);
1850 }
1851 pointers[stripe] = kmap(page);
1852 }
1853
1854 /* all raid6 handling here */
Zhao Lei10f11902015-01-20 15:11:43 +08001855 if (rbio->bbio->map_type & BTRFS_BLOCK_GROUP_RAID6) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001856 /*
1857 * single failure, rebuild from parity raid5
1858 * style
1859 */
1860 if (failb < 0) {
1861 if (faila == rbio->nr_data) {
1862 /*
1863 * Just the P stripe has failed, without
1864 * a bad data or Q stripe.
1865 * TODO, we should redo the xor here.
1866 */
1867 err = -EIO;
1868 goto cleanup;
1869 }
1870 /*
1871 * a single failure in raid6 is rebuilt
1872 * in the pstripe code below
1873 */
1874 goto pstripe;
1875 }
1876
1877 /* make sure our ps and qs are in order */
1878 if (faila > failb) {
1879 int tmp = failb;
1880 failb = faila;
1881 faila = tmp;
1882 }
1883
1884 /* if the q stripe is failed, do a pstripe reconstruction
1885 * from the xors.
1886 * If both the q stripe and the P stripe are failed, we're
1887 * here due to a crc mismatch and we can't give them the
1888 * data they want
1889 */
Zhao Lei8e5cfb52015-01-20 15:11:33 +08001890 if (rbio->bbio->raid_map[failb] == RAID6_Q_STRIPE) {
1891 if (rbio->bbio->raid_map[faila] ==
1892 RAID5_P_STRIPE) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001893 err = -EIO;
1894 goto cleanup;
1895 }
1896 /*
1897 * otherwise we have one bad data stripe and
1898 * a good P stripe. raid5!
1899 */
1900 goto pstripe;
1901 }
1902
Zhao Lei8e5cfb52015-01-20 15:11:33 +08001903 if (rbio->bbio->raid_map[failb] == RAID5_P_STRIPE) {
Miao Xie2c8cdd62014-11-14 16:06:25 +08001904 raid6_datap_recov(rbio->real_stripes,
David Woodhouse53b381b2013-01-29 18:40:14 -05001905 PAGE_SIZE, faila, pointers);
1906 } else {
Miao Xie2c8cdd62014-11-14 16:06:25 +08001907 raid6_2data_recov(rbio->real_stripes,
David Woodhouse53b381b2013-01-29 18:40:14 -05001908 PAGE_SIZE, faila, failb,
1909 pointers);
1910 }
1911 } else {
1912 void *p;
1913
1914 /* rebuild from P stripe here (raid5 or raid6) */
1915 BUG_ON(failb != -1);
1916pstripe:
1917 /* Copy parity block into failed block to start with */
1918 memcpy(pointers[faila],
1919 pointers[rbio->nr_data],
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001920 PAGE_SIZE);
David Woodhouse53b381b2013-01-29 18:40:14 -05001921
1922 /* rearrange the pointer array */
1923 p = pointers[faila];
1924 for (stripe = faila; stripe < rbio->nr_data - 1; stripe++)
1925 pointers[stripe] = pointers[stripe + 1];
1926 pointers[rbio->nr_data - 1] = p;
1927
1928 /* xor in the rest */
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03001929 run_xor(pointers, rbio->nr_data - 1, PAGE_SIZE);
David Woodhouse53b381b2013-01-29 18:40:14 -05001930 }
1931 /* if we're doing this rebuild as part of an rmw, go through
1932 * and set all of our private rbio pages in the
1933 * failed stripes as uptodate. This way finish_rmw will
1934 * know they can be trusted. If this was a read reconstruction,
1935 * other endio functions will fiddle the uptodate bits
1936 */
Miao Xie1b94b552014-11-06 16:14:21 +08001937 if (rbio->operation == BTRFS_RBIO_WRITE) {
Zhao Lei915e2292015-03-03 20:42:48 +08001938 for (i = 0; i < rbio->stripe_npages; i++) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001939 if (faila != -1) {
1940 page = rbio_stripe_page(rbio, faila, i);
1941 SetPageUptodate(page);
1942 }
1943 if (failb != -1) {
1944 page = rbio_stripe_page(rbio, failb, i);
1945 SetPageUptodate(page);
1946 }
1947 }
1948 }
Miao Xie2c8cdd62014-11-14 16:06:25 +08001949 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001950 /*
1951 * if we're rebuilding a read, we have to use
1952 * pages from the bio list
1953 */
Omar Sandovalb4ee1782015-06-19 11:52:50 -07001954 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1955 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
David Woodhouse53b381b2013-01-29 18:40:14 -05001956 (stripe == faila || stripe == failb)) {
1957 page = page_in_rbio(rbio, stripe, pagenr, 0);
1958 } else {
1959 page = rbio_stripe_page(rbio, stripe, pagenr);
1960 }
1961 kunmap(page);
1962 }
1963 }
1964
1965 err = 0;
1966cleanup:
1967 kfree(pointers);
1968
1969cleanup_io:
Miao Xie1b94b552014-11-06 16:14:21 +08001970 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
Zhao Lei6e9606d2015-01-20 15:11:34 +08001971 if (err == 0)
Chris Mason4ae10b32013-01-31 14:42:09 -05001972 cache_rbio_pages(rbio);
1973 else
1974 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
1975
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001976 rbio_orig_end_io(rbio, err);
Omar Sandovalb4ee1782015-06-19 11:52:50 -07001977 } else if (rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
Linus Torvalds22365972015-09-05 15:14:43 -07001978 rbio_orig_end_io(rbio, err);
David Woodhouse53b381b2013-01-29 18:40:14 -05001979 } else if (err == 0) {
1980 rbio->faila = -1;
1981 rbio->failb = -1;
Miao Xie5a6ac9e2014-11-06 17:20:58 +08001982
1983 if (rbio->operation == BTRFS_RBIO_WRITE)
1984 finish_rmw(rbio);
1985 else if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB)
1986 finish_parity_scrub(rbio, 0);
1987 else
1988 BUG();
David Woodhouse53b381b2013-01-29 18:40:14 -05001989 } else {
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001990 rbio_orig_end_io(rbio, err);
David Woodhouse53b381b2013-01-29 18:40:14 -05001991 }
1992}
1993
1994/*
1995 * This is called only for stripes we've read from disk to
1996 * reconstruct the parity.
1997 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001998static void raid_recover_end_io(struct bio *bio)
David Woodhouse53b381b2013-01-29 18:40:14 -05001999{
2000 struct btrfs_raid_bio *rbio = bio->bi_private;
2001
2002 /*
2003 * we only read stripe pages off the disk, set them
2004 * up to date if there were no errors
2005 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002006 if (bio->bi_error)
David Woodhouse53b381b2013-01-29 18:40:14 -05002007 fail_bio_stripe(rbio, bio);
2008 else
2009 set_bio_pages_uptodate(bio);
2010 bio_put(bio);
2011
Miao Xieb89e1b02014-10-15 11:18:44 +08002012 if (!atomic_dec_and_test(&rbio->stripes_pending))
David Woodhouse53b381b2013-01-29 18:40:14 -05002013 return;
2014
Miao Xieb89e1b02014-10-15 11:18:44 +08002015 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002016 rbio_orig_end_io(rbio, -EIO);
David Woodhouse53b381b2013-01-29 18:40:14 -05002017 else
2018 __raid_recover_end_io(rbio);
2019}
2020
2021/*
2022 * reads everything we need off the disk to reconstruct
2023 * the parity. endio handlers trigger final reconstruction
2024 * when the IO is done.
2025 *
2026 * This is used both for reads from the higher layers and for
2027 * parity construction required to finish a rmw cycle.
2028 */
2029static int __raid56_parity_recover(struct btrfs_raid_bio *rbio)
2030{
2031 int bios_to_read = 0;
David Woodhouse53b381b2013-01-29 18:40:14 -05002032 struct bio_list bio_list;
2033 int ret;
David Woodhouse53b381b2013-01-29 18:40:14 -05002034 int pagenr;
2035 int stripe;
2036 struct bio *bio;
2037
2038 bio_list_init(&bio_list);
2039
2040 ret = alloc_rbio_pages(rbio);
2041 if (ret)
2042 goto cleanup;
2043
Miao Xieb89e1b02014-10-15 11:18:44 +08002044 atomic_set(&rbio->error, 0);
David Woodhouse53b381b2013-01-29 18:40:14 -05002045
2046 /*
Chris Mason4ae10b32013-01-31 14:42:09 -05002047 * read everything that hasn't failed. Thanks to the
2048 * stripe cache, it is possible that some or all of these
2049 * pages are going to be uptodate.
David Woodhouse53b381b2013-01-29 18:40:14 -05002050 */
Miao Xie2c8cdd62014-11-14 16:06:25 +08002051 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
Liu Bo55883832014-06-24 15:39:16 +08002052 if (rbio->faila == stripe || rbio->failb == stripe) {
Miao Xieb89e1b02014-10-15 11:18:44 +08002053 atomic_inc(&rbio->error);
David Woodhouse53b381b2013-01-29 18:40:14 -05002054 continue;
Liu Bo55883832014-06-24 15:39:16 +08002055 }
David Woodhouse53b381b2013-01-29 18:40:14 -05002056
Zhao Lei915e2292015-03-03 20:42:48 +08002057 for (pagenr = 0; pagenr < rbio->stripe_npages; pagenr++) {
David Woodhouse53b381b2013-01-29 18:40:14 -05002058 struct page *p;
2059
2060 /*
2061 * the rmw code may have already read this
2062 * page in
2063 */
2064 p = rbio_stripe_page(rbio, stripe, pagenr);
2065 if (PageUptodate(p))
2066 continue;
2067
2068 ret = rbio_add_io_page(rbio, &bio_list,
2069 rbio_stripe_page(rbio, stripe, pagenr),
2070 stripe, pagenr, rbio->stripe_len);
2071 if (ret < 0)
2072 goto cleanup;
2073 }
2074 }
2075
2076 bios_to_read = bio_list_size(&bio_list);
2077 if (!bios_to_read) {
2078 /*
2079 * we might have no bios to read just because the pages
2080 * were up to date, or we might have no bios to read because
2081 * the devices were gone.
2082 */
Miao Xieb89e1b02014-10-15 11:18:44 +08002083 if (atomic_read(&rbio->error) <= rbio->bbio->max_errors) {
David Woodhouse53b381b2013-01-29 18:40:14 -05002084 __raid_recover_end_io(rbio);
2085 goto out;
2086 } else {
2087 goto cleanup;
2088 }
2089 }
2090
2091 /*
2092 * the bbio may be freed once we submit the last bio. Make sure
2093 * not to touch it after that
2094 */
Miao Xieb89e1b02014-10-15 11:18:44 +08002095 atomic_set(&rbio->stripes_pending, bios_to_read);
David Woodhouse53b381b2013-01-29 18:40:14 -05002096 while (1) {
2097 bio = bio_list_pop(&bio_list);
2098 if (!bio)
2099 break;
2100
2101 bio->bi_private = rbio;
2102 bio->bi_end_io = raid_recover_end_io;
Mike Christie37226b22016-06-05 14:31:52 -05002103 bio_set_op_attrs(bio, REQ_OP_READ, 0);
David Woodhouse53b381b2013-01-29 18:40:14 -05002104
2105 btrfs_bio_wq_end_io(rbio->fs_info, bio,
2106 BTRFS_WQ_ENDIO_RAID56);
2107
Mike Christie4e49ea42016-06-05 14:31:41 -05002108 submit_bio(bio);
David Woodhouse53b381b2013-01-29 18:40:14 -05002109 }
2110out:
2111 return 0;
2112
2113cleanup:
Omar Sandovalb4ee1782015-06-19 11:52:50 -07002114 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
2115 rbio->operation == BTRFS_RBIO_REBUILD_MISSING)
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002116 rbio_orig_end_io(rbio, -EIO);
David Woodhouse53b381b2013-01-29 18:40:14 -05002117 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_root *root, struct bio *bio,
Zhao Lei8e5cfb52015-01-20 15:11:33 +08002127 struct btrfs_bio *bbio, u64 stripe_len,
2128 int mirror_num, int generic_io)
David Woodhouse53b381b2013-01-29 18:40:14 -05002129{
2130 struct btrfs_raid_bio *rbio;
2131 int ret;
2132
Zhao Lei8e5cfb52015-01-20 15:11:33 +08002133 rbio = alloc_rbio(root, bbio, stripe_len);
Miao Xieaf8e2d12014-10-23 14:42:50 +08002134 if (IS_ERR(rbio)) {
Zhao Lei6e9606d2015-01-20 15:11:34 +08002135 if (generic_io)
2136 btrfs_put_bbio(bbio);
David Woodhouse53b381b2013-01-29 18:40:14 -05002137 return PTR_ERR(rbio);
Miao Xieaf8e2d12014-10-23 14:42:50 +08002138 }
David Woodhouse53b381b2013-01-29 18:40:14 -05002139
Miao Xie1b94b552014-11-06 16:14:21 +08002140 rbio->operation = BTRFS_RBIO_READ_REBUILD;
David Woodhouse53b381b2013-01-29 18:40:14 -05002141 bio_list_add(&rbio->bio_list, bio);
Kent Overstreet4f024f32013-10-11 15:44:27 -07002142 rbio->bio_list_bytes = bio->bi_iter.bi_size;
David Woodhouse53b381b2013-01-29 18:40:14 -05002143
2144 rbio->faila = find_logical_bio_stripe(rbio, bio);
2145 if (rbio->faila == -1) {
2146 BUG();
Zhao Lei6e9606d2015-01-20 15:11:34 +08002147 if (generic_io)
2148 btrfs_put_bbio(bbio);
David Woodhouse53b381b2013-01-29 18:40:14 -05002149 kfree(rbio);
2150 return -EIO;
2151 }
2152
Miao Xie42452152014-11-25 16:39:28 +08002153 if (generic_io) {
2154 btrfs_bio_counter_inc_noblocked(root->fs_info);
2155 rbio->generic_bio_cnt = 1;
2156 } else {
Zhao Lei6e9606d2015-01-20 15:11:34 +08002157 btrfs_get_bbio(bbio);
Miao Xie42452152014-11-25 16:39:28 +08002158 }
2159
David Woodhouse53b381b2013-01-29 18:40:14 -05002160 /*
2161 * reconstruct from the q stripe if they are
2162 * asking for mirror 3
2163 */
2164 if (mirror_num == 3)
Miao Xie2c8cdd62014-11-14 16:06:25 +08002165 rbio->failb = rbio->real_stripes - 2;
David Woodhouse53b381b2013-01-29 18:40:14 -05002166
2167 ret = lock_stripe_add(rbio);
2168
2169 /*
2170 * __raid56_parity_recover will end the bio with
2171 * any errors it hits. We don't want to return
2172 * its error value up the stack because our caller
2173 * will end up calling bio_endio with any nonzero
2174 * return
2175 */
2176 if (ret == 0)
2177 __raid56_parity_recover(rbio);
2178 /*
2179 * our rbio has been added to the list of
2180 * rbios that will be handled after the
2181 * currently lock owner is done
2182 */
2183 return 0;
2184
2185}
2186
2187static void rmw_work(struct btrfs_work *work)
2188{
2189 struct btrfs_raid_bio *rbio;
2190
2191 rbio = container_of(work, struct btrfs_raid_bio, work);
2192 raid56_rmw_stripe(rbio);
2193}
2194
2195static void read_rebuild_work(struct btrfs_work *work)
2196{
2197 struct btrfs_raid_bio *rbio;
2198
2199 rbio = container_of(work, struct btrfs_raid_bio, work);
2200 __raid56_parity_recover(rbio);
2201}
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002202
2203/*
2204 * The following code is used to scrub/replace the parity stripe
2205 *
2206 * Note: We need make sure all the pages that add into the scrub/replace
2207 * raid bio are correct and not be changed during the scrub/replace. That
2208 * is those pages just hold metadata or file data with checksum.
2209 */
2210
2211struct btrfs_raid_bio *
2212raid56_parity_alloc_scrub_rbio(struct btrfs_root *root, struct bio *bio,
Zhao Lei8e5cfb52015-01-20 15:11:33 +08002213 struct btrfs_bio *bbio, u64 stripe_len,
2214 struct btrfs_device *scrub_dev,
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002215 unsigned long *dbitmap, int stripe_nsectors)
2216{
2217 struct btrfs_raid_bio *rbio;
2218 int i;
2219
Zhao Lei8e5cfb52015-01-20 15:11:33 +08002220 rbio = alloc_rbio(root, bbio, stripe_len);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002221 if (IS_ERR(rbio))
2222 return NULL;
2223 bio_list_add(&rbio->bio_list, bio);
2224 /*
2225 * This is a special bio which is used to hold the completion handler
2226 * and make the scrub rbio is similar to the other types
2227 */
2228 ASSERT(!bio->bi_iter.bi_size);
2229 rbio->operation = BTRFS_RBIO_PARITY_SCRUB;
2230
Miao Xie2c8cdd62014-11-14 16:06:25 +08002231 for (i = 0; i < rbio->real_stripes; i++) {
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002232 if (bbio->stripes[i].dev == scrub_dev) {
2233 rbio->scrubp = i;
2234 break;
2235 }
2236 }
2237
2238 /* Now we just support the sectorsize equals to page size */
2239 ASSERT(root->sectorsize == PAGE_SIZE);
2240 ASSERT(rbio->stripe_npages == stripe_nsectors);
2241 bitmap_copy(rbio->dbitmap, dbitmap, stripe_nsectors);
2242
2243 return rbio;
2244}
2245
Omar Sandovalb4ee1782015-06-19 11:52:50 -07002246/* Used for both parity scrub and missing. */
2247void raid56_add_scrub_pages(struct btrfs_raid_bio *rbio, struct page *page,
2248 u64 logical)
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002249{
2250 int stripe_offset;
2251 int index;
2252
Zhao Lei8e5cfb52015-01-20 15:11:33 +08002253 ASSERT(logical >= rbio->bbio->raid_map[0]);
2254 ASSERT(logical + PAGE_SIZE <= rbio->bbio->raid_map[0] +
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002255 rbio->stripe_len * rbio->nr_data);
Zhao Lei8e5cfb52015-01-20 15:11:33 +08002256 stripe_offset = (int)(logical - rbio->bbio->raid_map[0]);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002257 index = stripe_offset >> PAGE_SHIFT;
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002258 rbio->bio_pages[index] = page;
2259}
2260
2261/*
2262 * We just scrub the parity that we have correct data on the same horizontal,
2263 * so we needn't allocate all pages for all the stripes.
2264 */
2265static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio)
2266{
2267 int i;
2268 int bit;
2269 int index;
2270 struct page *page;
2271
2272 for_each_set_bit(bit, rbio->dbitmap, rbio->stripe_npages) {
Miao Xie2c8cdd62014-11-14 16:06:25 +08002273 for (i = 0; i < rbio->real_stripes; i++) {
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002274 index = i * rbio->stripe_npages + bit;
2275 if (rbio->stripe_pages[index])
2276 continue;
2277
2278 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2279 if (!page)
2280 return -ENOMEM;
2281 rbio->stripe_pages[index] = page;
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002282 }
2283 }
2284 return 0;
2285}
2286
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002287static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
2288 int need_check)
2289{
Miao Xie76035972014-11-14 17:45:42 +08002290 struct btrfs_bio *bbio = rbio->bbio;
Miao Xie2c8cdd62014-11-14 16:06:25 +08002291 void *pointers[rbio->real_stripes];
Miao Xie76035972014-11-14 17:45:42 +08002292 DECLARE_BITMAP(pbitmap, rbio->stripe_npages);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002293 int nr_data = rbio->nr_data;
2294 int stripe;
2295 int pagenr;
2296 int p_stripe = -1;
2297 int q_stripe = -1;
2298 struct page *p_page = NULL;
2299 struct page *q_page = NULL;
2300 struct bio_list bio_list;
2301 struct bio *bio;
Miao Xie76035972014-11-14 17:45:42 +08002302 int is_replace = 0;
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002303 int ret;
2304
2305 bio_list_init(&bio_list);
2306
Miao Xie2c8cdd62014-11-14 16:06:25 +08002307 if (rbio->real_stripes - rbio->nr_data == 1) {
2308 p_stripe = rbio->real_stripes - 1;
2309 } else if (rbio->real_stripes - rbio->nr_data == 2) {
2310 p_stripe = rbio->real_stripes - 2;
2311 q_stripe = rbio->real_stripes - 1;
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002312 } else {
2313 BUG();
2314 }
2315
Miao Xie76035972014-11-14 17:45:42 +08002316 if (bbio->num_tgtdevs && bbio->tgtdev_map[rbio->scrubp]) {
2317 is_replace = 1;
2318 bitmap_copy(pbitmap, rbio->dbitmap, rbio->stripe_npages);
2319 }
2320
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002321 /*
2322 * Because the higher layers(scrubber) are unlikely to
2323 * use this area of the disk again soon, so don't cache
2324 * it.
2325 */
2326 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2327
2328 if (!need_check)
2329 goto writeback;
2330
2331 p_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2332 if (!p_page)
2333 goto cleanup;
2334 SetPageUptodate(p_page);
2335
2336 if (q_stripe != -1) {
2337 q_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2338 if (!q_page) {
2339 __free_page(p_page);
2340 goto cleanup;
2341 }
2342 SetPageUptodate(q_page);
2343 }
2344
2345 atomic_set(&rbio->error, 0);
2346
2347 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2348 struct page *p;
2349 void *parity;
2350 /* first collect one page from each data stripe */
2351 for (stripe = 0; stripe < nr_data; stripe++) {
2352 p = page_in_rbio(rbio, stripe, pagenr, 0);
2353 pointers[stripe] = kmap(p);
2354 }
2355
2356 /* then add the parity stripe */
2357 pointers[stripe++] = kmap(p_page);
2358
2359 if (q_stripe != -1) {
2360
2361 /*
2362 * raid6, add the qstripe and call the
2363 * library function to fill in our p/q
2364 */
2365 pointers[stripe++] = kmap(q_page);
2366
Miao Xie2c8cdd62014-11-14 16:06:25 +08002367 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002368 pointers);
2369 } else {
2370 /* raid5 */
2371 memcpy(pointers[nr_data], pointers[0], PAGE_SIZE);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002372 run_xor(pointers + 1, nr_data - 1, PAGE_SIZE);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002373 }
2374
Nicholas D Steeves01327612016-05-19 21:18:45 -04002375 /* Check scrubbing parity and repair it */
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002376 p = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2377 parity = kmap(p);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002378 if (memcmp(parity, pointers[rbio->scrubp], PAGE_SIZE))
2379 memcpy(parity, pointers[rbio->scrubp], PAGE_SIZE);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002380 else
2381 /* Parity is right, needn't writeback */
2382 bitmap_clear(rbio->dbitmap, pagenr, 1);
2383 kunmap(p);
2384
Miao Xie2c8cdd62014-11-14 16:06:25 +08002385 for (stripe = 0; stripe < rbio->real_stripes; stripe++)
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002386 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
2387 }
2388
2389 __free_page(p_page);
2390 if (q_page)
2391 __free_page(q_page);
2392
2393writeback:
2394 /*
2395 * time to start writing. Make bios for everything from the
2396 * higher layers (the bio_list in our rbio) and our p/q. Ignore
2397 * everything else.
2398 */
2399 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2400 struct page *page;
2401
2402 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2403 ret = rbio_add_io_page(rbio, &bio_list,
2404 page, rbio->scrubp, pagenr, rbio->stripe_len);
2405 if (ret)
2406 goto cleanup;
2407 }
2408
Miao Xie76035972014-11-14 17:45:42 +08002409 if (!is_replace)
2410 goto submit_write;
2411
2412 for_each_set_bit(pagenr, pbitmap, rbio->stripe_npages) {
2413 struct page *page;
2414
2415 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2416 ret = rbio_add_io_page(rbio, &bio_list, page,
2417 bbio->tgtdev_map[rbio->scrubp],
2418 pagenr, rbio->stripe_len);
2419 if (ret)
2420 goto cleanup;
2421 }
2422
2423submit_write:
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002424 nr_data = bio_list_size(&bio_list);
2425 if (!nr_data) {
2426 /* Every parity is right */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002427 rbio_orig_end_io(rbio, 0);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002428 return;
2429 }
2430
2431 atomic_set(&rbio->stripes_pending, nr_data);
2432
2433 while (1) {
2434 bio = bio_list_pop(&bio_list);
2435 if (!bio)
2436 break;
2437
2438 bio->bi_private = rbio;
Zhao Leia6111d12016-01-12 17:52:13 +08002439 bio->bi_end_io = raid_write_end_io;
Mike Christie37226b22016-06-05 14:31:52 -05002440 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
Mike Christie4e49ea42016-06-05 14:31:41 -05002441
2442 submit_bio(bio);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002443 }
2444 return;
2445
2446cleanup:
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002447 rbio_orig_end_io(rbio, -EIO);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002448}
2449
2450static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe)
2451{
2452 if (stripe >= 0 && stripe < rbio->nr_data)
2453 return 1;
2454 return 0;
2455}
2456
2457/*
2458 * While we're doing the parity check and repair, we could have errors
2459 * in reading pages off the disk. This checks for errors and if we're
2460 * not able to read the page it'll trigger parity reconstruction. The
2461 * parity scrub will be finished after we've reconstructed the failed
2462 * stripes
2463 */
2464static void validate_rbio_for_parity_scrub(struct btrfs_raid_bio *rbio)
2465{
2466 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
2467 goto cleanup;
2468
2469 if (rbio->faila >= 0 || rbio->failb >= 0) {
2470 int dfail = 0, failp = -1;
2471
2472 if (is_data_stripe(rbio, rbio->faila))
2473 dfail++;
2474 else if (is_parity_stripe(rbio->faila))
2475 failp = rbio->faila;
2476
2477 if (is_data_stripe(rbio, rbio->failb))
2478 dfail++;
2479 else if (is_parity_stripe(rbio->failb))
2480 failp = rbio->failb;
2481
2482 /*
2483 * Because we can not use a scrubbing parity to repair
2484 * the data, so the capability of the repair is declined.
2485 * (In the case of RAID5, we can not repair anything)
2486 */
2487 if (dfail > rbio->bbio->max_errors - 1)
2488 goto cleanup;
2489
2490 /*
2491 * If all data is good, only parity is correctly, just
2492 * repair the parity.
2493 */
2494 if (dfail == 0) {
2495 finish_parity_scrub(rbio, 0);
2496 return;
2497 }
2498
2499 /*
2500 * Here means we got one corrupted data stripe and one
2501 * corrupted parity on RAID6, if the corrupted parity
Nicholas D Steeves01327612016-05-19 21:18:45 -04002502 * is scrubbing parity, luckily, use the other one to repair
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002503 * the data, or we can not repair the data stripe.
2504 */
2505 if (failp != rbio->scrubp)
2506 goto cleanup;
2507
2508 __raid_recover_end_io(rbio);
2509 } else {
2510 finish_parity_scrub(rbio, 1);
2511 }
2512 return;
2513
2514cleanup:
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002515 rbio_orig_end_io(rbio, -EIO);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002516}
2517
2518/*
2519 * end io for the read phase of the rmw cycle. All the bios here are physical
2520 * stripe bios we've read from the disk so we can recalculate the parity of the
2521 * stripe.
2522 *
2523 * This will usually kick off finish_rmw once all the bios are read in, but it
2524 * may trigger parity reconstruction if we had any errors along the way
2525 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002526static void raid56_parity_scrub_end_io(struct bio *bio)
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002527{
2528 struct btrfs_raid_bio *rbio = bio->bi_private;
2529
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002530 if (bio->bi_error)
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002531 fail_bio_stripe(rbio, bio);
2532 else
2533 set_bio_pages_uptodate(bio);
2534
2535 bio_put(bio);
2536
2537 if (!atomic_dec_and_test(&rbio->stripes_pending))
2538 return;
2539
2540 /*
2541 * this will normally call finish_rmw to start our write
2542 * but if there are any failed stripes we'll reconstruct
2543 * from parity first
2544 */
2545 validate_rbio_for_parity_scrub(rbio);
2546}
2547
2548static void raid56_parity_scrub_stripe(struct btrfs_raid_bio *rbio)
2549{
2550 int bios_to_read = 0;
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002551 struct bio_list bio_list;
2552 int ret;
2553 int pagenr;
2554 int stripe;
2555 struct bio *bio;
2556
2557 ret = alloc_rbio_essential_pages(rbio);
2558 if (ret)
2559 goto cleanup;
2560
2561 bio_list_init(&bio_list);
2562
2563 atomic_set(&rbio->error, 0);
2564 /*
2565 * build a list of bios to read all the missing parts of this
2566 * stripe
2567 */
Miao Xie2c8cdd62014-11-14 16:06:25 +08002568 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002569 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2570 struct page *page;
2571 /*
2572 * we want to find all the pages missing from
2573 * the rbio and read them from the disk. If
2574 * page_in_rbio finds a page in the bio list
2575 * we don't need to read it off the stripe.
2576 */
2577 page = page_in_rbio(rbio, stripe, pagenr, 1);
2578 if (page)
2579 continue;
2580
2581 page = rbio_stripe_page(rbio, stripe, pagenr);
2582 /*
2583 * the bio cache may have handed us an uptodate
2584 * page. If so, be happy and use it
2585 */
2586 if (PageUptodate(page))
2587 continue;
2588
2589 ret = rbio_add_io_page(rbio, &bio_list, page,
2590 stripe, pagenr, rbio->stripe_len);
2591 if (ret)
2592 goto cleanup;
2593 }
2594 }
2595
2596 bios_to_read = bio_list_size(&bio_list);
2597 if (!bios_to_read) {
2598 /*
2599 * this can happen if others have merged with
2600 * us, it means there is nothing left to read.
2601 * But if there are missing devices it may not be
2602 * safe to do the full stripe write yet.
2603 */
2604 goto finish;
2605 }
2606
2607 /*
2608 * the bbio may be freed once we submit the last bio. Make sure
2609 * not to touch it after that
2610 */
2611 atomic_set(&rbio->stripes_pending, bios_to_read);
2612 while (1) {
2613 bio = bio_list_pop(&bio_list);
2614 if (!bio)
2615 break;
2616
2617 bio->bi_private = rbio;
2618 bio->bi_end_io = raid56_parity_scrub_end_io;
Mike Christie37226b22016-06-05 14:31:52 -05002619 bio_set_op_attrs(bio, REQ_OP_READ, 0);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002620
2621 btrfs_bio_wq_end_io(rbio->fs_info, bio,
2622 BTRFS_WQ_ENDIO_RAID56);
2623
Mike Christie4e49ea42016-06-05 14:31:41 -05002624 submit_bio(bio);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002625 }
2626 /* the actual write will happen once the reads are done */
2627 return;
2628
2629cleanup:
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002630 rbio_orig_end_io(rbio, -EIO);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002631 return;
2632
2633finish:
2634 validate_rbio_for_parity_scrub(rbio);
2635}
2636
2637static void scrub_parity_work(struct btrfs_work *work)
2638{
2639 struct btrfs_raid_bio *rbio;
2640
2641 rbio = container_of(work, struct btrfs_raid_bio, work);
2642 raid56_parity_scrub_stripe(rbio);
2643}
2644
2645static void async_scrub_parity(struct btrfs_raid_bio *rbio)
2646{
2647 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
2648 scrub_parity_work, NULL, NULL);
2649
2650 btrfs_queue_work(rbio->fs_info->rmw_workers,
2651 &rbio->work);
2652}
2653
2654void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio)
2655{
2656 if (!lock_stripe_add(rbio))
2657 async_scrub_parity(rbio);
2658}
Omar Sandovalb4ee1782015-06-19 11:52:50 -07002659
2660/* The following code is used for dev replace of a missing RAID 5/6 device. */
2661
2662struct btrfs_raid_bio *
2663raid56_alloc_missing_rbio(struct btrfs_root *root, struct bio *bio,
2664 struct btrfs_bio *bbio, u64 length)
2665{
2666 struct btrfs_raid_bio *rbio;
2667
2668 rbio = alloc_rbio(root, bbio, length);
2669 if (IS_ERR(rbio))
2670 return NULL;
2671
2672 rbio->operation = BTRFS_RBIO_REBUILD_MISSING;
2673 bio_list_add(&rbio->bio_list, bio);
2674 /*
2675 * This is a special bio which is used to hold the completion handler
2676 * and make the scrub rbio is similar to the other types
2677 */
2678 ASSERT(!bio->bi_iter.bi_size);
2679
2680 rbio->faila = find_logical_bio_stripe(rbio, bio);
2681 if (rbio->faila == -1) {
2682 BUG();
2683 kfree(rbio);
2684 return NULL;
2685 }
2686
2687 return rbio;
2688}
2689
2690static void missing_raid56_work(struct btrfs_work *work)
2691{
2692 struct btrfs_raid_bio *rbio;
2693
2694 rbio = container_of(work, struct btrfs_raid_bio, work);
2695 __raid56_parity_recover(rbio);
2696}
2697
2698static void async_missing_raid56(struct btrfs_raid_bio *rbio)
2699{
2700 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
2701 missing_raid56_work, NULL, NULL);
2702
2703 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
2704}
2705
2706void raid56_submit_missing_rbio(struct btrfs_raid_bio *rbio)
2707{
2708 if (!lock_stripe_add(rbio))
2709 async_missing_raid56(rbio);
2710}