blob: c435fe0f86e485340c02ffe6c8f0b3dafb50cb2e [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
273 memcpy(d, s, PAGE_CACHE_SIZE);
274
275 kunmap(rbio->bio_pages[i]);
276 kunmap(rbio->stripe_pages[i]);
277 SetPageUptodate(rbio->stripe_pages[i]);
278 }
279 set_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
280}
281
282/*
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
579 * steal from cached rbio's though, other functions
580 * handle that.
581 */
582 if (test_bit(RBIO_CACHE_BIT, &last->flags) ||
583 test_bit(RBIO_CACHE_BIT, &cur->flags))
584 return 0;
585
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;
David Woodhouse53b381b2013-01-29 18:40:14 -0500904
905 if (err)
906 fail_bio_stripe(rbio, bio);
907
908 bio_put(bio);
909
Miao Xieb89e1b02014-10-15 11:18:44 +0800910 if (!atomic_dec_and_test(&rbio->stripes_pending))
David Woodhouse53b381b2013-01-29 18:40:14 -0500911 return;
912
913 err = 0;
914
915 /* OK, we have read all the stripes we need to. */
Miao Xieb89e1b02014-10-15 11:18:44 +0800916 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
David Woodhouse53b381b2013-01-29 18:40:14 -0500917 err = -EIO;
918
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200919 rbio_orig_end_io(rbio, err);
David Woodhouse53b381b2013-01-29 18:40:14 -0500920}
921
922/*
923 * the read/modify/write code wants to use the original bio for
924 * any pages it included, and then use the rbio for everything
925 * else. This function decides if a given index (stripe number)
926 * and page number in that stripe fall inside the original bio
927 * or the rbio.
928 *
929 * if you set bio_list_only, you'll get a NULL back for any ranges
930 * that are outside the bio_list
931 *
932 * This doesn't take any refs on anything, you get a bare page pointer
933 * and the caller must bump refs as required.
934 *
935 * You must call index_rbio_pages once before you can trust
936 * the answers from this function.
937 */
938static struct page *page_in_rbio(struct btrfs_raid_bio *rbio,
939 int index, int pagenr, int bio_list_only)
940{
941 int chunk_page;
942 struct page *p = NULL;
943
944 chunk_page = index * (rbio->stripe_len >> PAGE_SHIFT) + pagenr;
945
946 spin_lock_irq(&rbio->bio_list_lock);
947 p = rbio->bio_pages[chunk_page];
948 spin_unlock_irq(&rbio->bio_list_lock);
949
950 if (p || bio_list_only)
951 return p;
952
953 return rbio->stripe_pages[chunk_page];
954}
955
956/*
957 * number of pages we need for the entire stripe across all the
958 * drives
959 */
960static unsigned long rbio_nr_pages(unsigned long stripe_len, int nr_stripes)
961{
Zhao Leib7178a52015-03-03 20:38:46 +0800962 return DIV_ROUND_UP(stripe_len, PAGE_CACHE_SIZE) * nr_stripes;
David Woodhouse53b381b2013-01-29 18:40:14 -0500963}
964
965/*
966 * allocation and initial setup for the btrfs_raid_bio. Not
967 * this does not allocate any pages for rbio->pages.
968 */
969static struct btrfs_raid_bio *alloc_rbio(struct btrfs_root *root,
Zhao Lei8e5cfb52015-01-20 15:11:33 +0800970 struct btrfs_bio *bbio, u64 stripe_len)
David Woodhouse53b381b2013-01-29 18:40:14 -0500971{
972 struct btrfs_raid_bio *rbio;
973 int nr_data = 0;
Miao Xie2c8cdd62014-11-14 16:06:25 +0800974 int real_stripes = bbio->num_stripes - bbio->num_tgtdevs;
975 int num_pages = rbio_nr_pages(stripe_len, real_stripes);
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800976 int stripe_npages = DIV_ROUND_UP(stripe_len, PAGE_SIZE);
David Woodhouse53b381b2013-01-29 18:40:14 -0500977 void *p;
978
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800979 rbio = kzalloc(sizeof(*rbio) + num_pages * sizeof(struct page *) * 2 +
Zhao Leibfca9a62014-12-08 19:55:57 +0800980 DIV_ROUND_UP(stripe_npages, BITS_PER_LONG) *
981 sizeof(long), GFP_NOFS);
Miao Xieaf8e2d12014-10-23 14:42:50 +0800982 if (!rbio)
David Woodhouse53b381b2013-01-29 18:40:14 -0500983 return ERR_PTR(-ENOMEM);
David Woodhouse53b381b2013-01-29 18:40:14 -0500984
985 bio_list_init(&rbio->bio_list);
986 INIT_LIST_HEAD(&rbio->plug_list);
987 spin_lock_init(&rbio->bio_list_lock);
Chris Mason4ae10b32013-01-31 14:42:09 -0500988 INIT_LIST_HEAD(&rbio->stripe_cache);
David Woodhouse53b381b2013-01-29 18:40:14 -0500989 INIT_LIST_HEAD(&rbio->hash_list);
990 rbio->bbio = bbio;
David Woodhouse53b381b2013-01-29 18:40:14 -0500991 rbio->fs_info = root->fs_info;
992 rbio->stripe_len = stripe_len;
993 rbio->nr_pages = num_pages;
Miao Xie2c8cdd62014-11-14 16:06:25 +0800994 rbio->real_stripes = real_stripes;
Miao Xie5a6ac9e2014-11-06 17:20:58 +0800995 rbio->stripe_npages = stripe_npages;
David Woodhouse53b381b2013-01-29 18:40:14 -0500996 rbio->faila = -1;
997 rbio->failb = -1;
998 atomic_set(&rbio->refs, 1);
Miao Xieb89e1b02014-10-15 11:18:44 +0800999 atomic_set(&rbio->error, 0);
1000 atomic_set(&rbio->stripes_pending, 0);
David Woodhouse53b381b2013-01-29 18:40:14 -05001001
1002 /*
1003 * the stripe_pages and bio_pages array point to the extra
1004 * memory we allocated past the end of the rbio
1005 */
1006 p = rbio + 1;
1007 rbio->stripe_pages = p;
1008 rbio->bio_pages = p + sizeof(struct page *) * num_pages;
Miao Xie5a6ac9e2014-11-06 17:20:58 +08001009 rbio->dbitmap = p + sizeof(struct page *) * num_pages * 2;
David Woodhouse53b381b2013-01-29 18:40:14 -05001010
Zhao Lei10f11902015-01-20 15:11:43 +08001011 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1012 nr_data = real_stripes - 1;
1013 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
Miao Xie2c8cdd62014-11-14 16:06:25 +08001014 nr_data = real_stripes - 2;
David Woodhouse53b381b2013-01-29 18:40:14 -05001015 else
Zhao Lei10f11902015-01-20 15:11:43 +08001016 BUG();
David Woodhouse53b381b2013-01-29 18:40:14 -05001017
1018 rbio->nr_data = nr_data;
1019 return rbio;
1020}
1021
1022/* allocate pages for all the stripes in the bio, including parity */
1023static int alloc_rbio_pages(struct btrfs_raid_bio *rbio)
1024{
1025 int i;
1026 struct page *page;
1027
1028 for (i = 0; i < rbio->nr_pages; i++) {
1029 if (rbio->stripe_pages[i])
1030 continue;
1031 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1032 if (!page)
1033 return -ENOMEM;
1034 rbio->stripe_pages[i] = page;
1035 ClearPageUptodate(page);
1036 }
1037 return 0;
1038}
1039
Zhao Leib7178a52015-03-03 20:38:46 +08001040/* only allocate pages for p/q stripes */
David Woodhouse53b381b2013-01-29 18:40:14 -05001041static int alloc_rbio_parity_pages(struct btrfs_raid_bio *rbio)
1042{
1043 int i;
1044 struct page *page;
1045
Zhao Leib7178a52015-03-03 20:38:46 +08001046 i = rbio_stripe_page_index(rbio, rbio->nr_data, 0);
David Woodhouse53b381b2013-01-29 18:40:14 -05001047
1048 for (; i < rbio->nr_pages; i++) {
1049 if (rbio->stripe_pages[i])
1050 continue;
1051 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1052 if (!page)
1053 return -ENOMEM;
1054 rbio->stripe_pages[i] = page;
1055 }
1056 return 0;
1057}
1058
1059/*
1060 * add a single page from a specific stripe into our list of bios for IO
1061 * this will try to merge into existing bios if possible, and returns
1062 * zero if all went well.
1063 */
Eric Sandeen48a3b632013-04-25 20:41:01 +00001064static int rbio_add_io_page(struct btrfs_raid_bio *rbio,
1065 struct bio_list *bio_list,
1066 struct page *page,
1067 int stripe_nr,
1068 unsigned long page_index,
1069 unsigned long bio_max_len)
David Woodhouse53b381b2013-01-29 18:40:14 -05001070{
1071 struct bio *last = bio_list->tail;
1072 u64 last_end = 0;
1073 int ret;
1074 struct bio *bio;
1075 struct btrfs_bio_stripe *stripe;
1076 u64 disk_start;
1077
1078 stripe = &rbio->bbio->stripes[stripe_nr];
1079 disk_start = stripe->physical + (page_index << PAGE_CACHE_SHIFT);
1080
1081 /* if the device is missing, just fail this stripe */
1082 if (!stripe->dev->bdev)
1083 return fail_rbio_index(rbio, stripe_nr);
1084
1085 /* see if we can add this page onto our existing bio */
1086 if (last) {
Kent Overstreet4f024f32013-10-11 15:44:27 -07001087 last_end = (u64)last->bi_iter.bi_sector << 9;
1088 last_end += last->bi_iter.bi_size;
David Woodhouse53b381b2013-01-29 18:40:14 -05001089
1090 /*
1091 * we can't merge these if they are from different
1092 * devices or if they are not contiguous
1093 */
1094 if (last_end == disk_start && stripe->dev->bdev &&
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001095 !last->bi_error &&
David Woodhouse53b381b2013-01-29 18:40:14 -05001096 last->bi_bdev == stripe->dev->bdev) {
1097 ret = bio_add_page(last, page, PAGE_CACHE_SIZE, 0);
1098 if (ret == PAGE_CACHE_SIZE)
1099 return 0;
1100 }
1101 }
1102
1103 /* put a new bio on the list */
Chris Mason9be33952013-05-17 18:30:14 -04001104 bio = btrfs_io_bio_alloc(GFP_NOFS, bio_max_len >> PAGE_SHIFT?:1);
David Woodhouse53b381b2013-01-29 18:40:14 -05001105 if (!bio)
1106 return -ENOMEM;
1107
Kent Overstreet4f024f32013-10-11 15:44:27 -07001108 bio->bi_iter.bi_size = 0;
David Woodhouse53b381b2013-01-29 18:40:14 -05001109 bio->bi_bdev = stripe->dev->bdev;
Kent Overstreet4f024f32013-10-11 15:44:27 -07001110 bio->bi_iter.bi_sector = disk_start >> 9;
David Woodhouse53b381b2013-01-29 18:40:14 -05001111
1112 bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
1113 bio_list_add(bio_list, bio);
1114 return 0;
1115}
1116
1117/*
1118 * while we're doing the read/modify/write cycle, we could
1119 * have errors in reading pages off the disk. This checks
1120 * for errors and if we're not able to read the page it'll
1121 * trigger parity reconstruction. The rmw will be finished
1122 * after we've reconstructed the failed stripes
1123 */
1124static void validate_rbio_for_rmw(struct btrfs_raid_bio *rbio)
1125{
1126 if (rbio->faila >= 0 || rbio->failb >= 0) {
Miao Xie2c8cdd62014-11-14 16:06:25 +08001127 BUG_ON(rbio->faila == rbio->real_stripes - 1);
David Woodhouse53b381b2013-01-29 18:40:14 -05001128 __raid56_parity_recover(rbio);
1129 } else {
1130 finish_rmw(rbio);
1131 }
1132}
1133
1134/*
David Woodhouse53b381b2013-01-29 18:40:14 -05001135 * helper function to walk our bio list and populate the bio_pages array with
1136 * the result. This seems expensive, but it is faster than constantly
1137 * searching through the bio list as we setup the IO in finish_rmw or stripe
1138 * reconstruction.
1139 *
1140 * This must be called before you trust the answers from page_in_rbio
1141 */
1142static void index_rbio_pages(struct btrfs_raid_bio *rbio)
1143{
1144 struct bio *bio;
1145 u64 start;
1146 unsigned long stripe_offset;
1147 unsigned long page_index;
1148 struct page *p;
1149 int i;
1150
1151 spin_lock_irq(&rbio->bio_list_lock);
1152 bio_list_for_each(bio, &rbio->bio_list) {
Kent Overstreet4f024f32013-10-11 15:44:27 -07001153 start = (u64)bio->bi_iter.bi_sector << 9;
Zhao Lei8e5cfb52015-01-20 15:11:33 +08001154 stripe_offset = start - rbio->bbio->raid_map[0];
David Woodhouse53b381b2013-01-29 18:40:14 -05001155 page_index = stripe_offset >> PAGE_CACHE_SHIFT;
1156
1157 for (i = 0; i < bio->bi_vcnt; i++) {
1158 p = bio->bi_io_vec[i].bv_page;
1159 rbio->bio_pages[page_index + i] = p;
1160 }
1161 }
1162 spin_unlock_irq(&rbio->bio_list_lock);
1163}
1164
1165/*
1166 * this is called from one of two situations. We either
1167 * have a full stripe from the higher layers, or we've read all
1168 * the missing bits off disk.
1169 *
1170 * This will calculate the parity and then send down any
1171 * changed blocks.
1172 */
1173static noinline void finish_rmw(struct btrfs_raid_bio *rbio)
1174{
1175 struct btrfs_bio *bbio = rbio->bbio;
Miao Xie2c8cdd62014-11-14 16:06:25 +08001176 void *pointers[rbio->real_stripes];
David Woodhouse53b381b2013-01-29 18:40:14 -05001177 int stripe_len = rbio->stripe_len;
1178 int nr_data = rbio->nr_data;
1179 int stripe;
1180 int pagenr;
1181 int p_stripe = -1;
1182 int q_stripe = -1;
1183 struct bio_list bio_list;
1184 struct bio *bio;
1185 int pages_per_stripe = stripe_len >> PAGE_CACHE_SHIFT;
1186 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
1228 for (pagenr = 0; pagenr < pages_per_stripe; pagenr++) {
1229 struct page *p;
1230 /* first collect one page from each data stripe */
1231 for (stripe = 0; stripe < nr_data; stripe++) {
1232 p = page_in_rbio(rbio, stripe, pagenr, 0);
1233 pointers[stripe] = kmap(p);
1234 }
1235
1236 /* then add the parity stripe */
1237 p = rbio_pstripe_page(rbio, pagenr);
1238 SetPageUptodate(p);
1239 pointers[stripe++] = kmap(p);
1240
1241 if (q_stripe != -1) {
1242
1243 /*
1244 * raid6, add the qstripe and call the
1245 * library function to fill in our p/q
1246 */
1247 p = rbio_qstripe_page(rbio, pagenr);
1248 SetPageUptodate(p);
1249 pointers[stripe++] = kmap(p);
1250
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);
1256 run_xor(pointers + 1, nr_data - 1, PAGE_CACHE_SIZE);
1257 }
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++) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001270 for (pagenr = 0; pagenr < pages_per_stripe; pagenr++) {
1271 struct page *page;
1272 if (stripe < rbio->nr_data) {
1273 page = page_in_rbio(rbio, stripe, pagenr, 1);
1274 if (!page)
1275 continue;
1276 } else {
1277 page = rbio_stripe_page(rbio, stripe, pagenr);
1278 }
1279
1280 ret = rbio_add_io_page(rbio, &bio_list,
1281 page, stripe, pagenr, rbio->stripe_len);
1282 if (ret)
1283 goto cleanup;
1284 }
1285 }
1286
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
1294 for (pagenr = 0; pagenr < pages_per_stripe; pagenr++) {
1295 struct page *page;
1296 if (stripe < rbio->nr_data) {
1297 page = page_in_rbio(rbio, stripe, pagenr, 1);
1298 if (!page)
1299 continue;
1300 } else {
1301 page = rbio_stripe_page(rbio, stripe, pagenr);
1302 }
1303
1304 ret = rbio_add_io_page(rbio, &bio_list, page,
1305 rbio->bbio->tgtdev_map[stripe],
1306 pagenr, rbio->stripe_len);
1307 if (ret)
1308 goto cleanup;
1309 }
1310 }
1311
1312write_data:
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;
David Woodhouse53b381b2013-01-29 18:40:14 -05001323 submit_bio(WRITE, bio);
1324 }
1325 return;
1326
1327cleanup:
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001328 rbio_orig_end_io(rbio, -EIO);
David Woodhouse53b381b2013-01-29 18:40:14 -05001329}
1330
1331/*
1332 * helper to find the stripe number for a given bio. Used to figure out which
1333 * stripe has failed. This expects the bio to correspond to a physical disk,
1334 * so it looks up based on physical sector numbers.
1335 */
1336static int find_bio_stripe(struct btrfs_raid_bio *rbio,
1337 struct bio *bio)
1338{
Kent Overstreet4f024f32013-10-11 15:44:27 -07001339 u64 physical = bio->bi_iter.bi_sector;
David Woodhouse53b381b2013-01-29 18:40:14 -05001340 u64 stripe_start;
1341 int i;
1342 struct btrfs_bio_stripe *stripe;
1343
1344 physical <<= 9;
1345
1346 for (i = 0; i < rbio->bbio->num_stripes; i++) {
1347 stripe = &rbio->bbio->stripes[i];
1348 stripe_start = stripe->physical;
1349 if (physical >= stripe_start &&
Miao Xie2c8cdd62014-11-14 16:06:25 +08001350 physical < stripe_start + rbio->stripe_len &&
1351 bio->bi_bdev == stripe->dev->bdev) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001352 return i;
1353 }
1354 }
1355 return -1;
1356}
1357
1358/*
1359 * helper to find the stripe number for a given
1360 * bio (before mapping). Used to figure out which stripe has
1361 * failed. This looks up based on logical block numbers.
1362 */
1363static int find_logical_bio_stripe(struct btrfs_raid_bio *rbio,
1364 struct bio *bio)
1365{
Kent Overstreet4f024f32013-10-11 15:44:27 -07001366 u64 logical = bio->bi_iter.bi_sector;
David Woodhouse53b381b2013-01-29 18:40:14 -05001367 u64 stripe_start;
1368 int i;
1369
1370 logical <<= 9;
1371
1372 for (i = 0; i < rbio->nr_data; i++) {
Zhao Lei8e5cfb52015-01-20 15:11:33 +08001373 stripe_start = rbio->bbio->raid_map[i];
David Woodhouse53b381b2013-01-29 18:40:14 -05001374 if (logical >= stripe_start &&
1375 logical < stripe_start + rbio->stripe_len) {
1376 return i;
1377 }
1378 }
1379 return -1;
1380}
1381
1382/*
1383 * returns -EIO if we had too many failures
1384 */
1385static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed)
1386{
1387 unsigned long flags;
1388 int ret = 0;
1389
1390 spin_lock_irqsave(&rbio->bio_list_lock, flags);
1391
1392 /* we already know this stripe is bad, move on */
1393 if (rbio->faila == failed || rbio->failb == failed)
1394 goto out;
1395
1396 if (rbio->faila == -1) {
1397 /* first failure on this rbio */
1398 rbio->faila = failed;
Miao Xieb89e1b02014-10-15 11:18:44 +08001399 atomic_inc(&rbio->error);
David Woodhouse53b381b2013-01-29 18:40:14 -05001400 } else if (rbio->failb == -1) {
1401 /* second failure on this rbio */
1402 rbio->failb = failed;
Miao Xieb89e1b02014-10-15 11:18:44 +08001403 atomic_inc(&rbio->error);
David Woodhouse53b381b2013-01-29 18:40:14 -05001404 } else {
1405 ret = -EIO;
1406 }
1407out:
1408 spin_unlock_irqrestore(&rbio->bio_list_lock, flags);
1409
1410 return ret;
1411}
1412
1413/*
1414 * helper to fail a stripe based on a physical disk
1415 * bio.
1416 */
1417static int fail_bio_stripe(struct btrfs_raid_bio *rbio,
1418 struct bio *bio)
1419{
1420 int failed = find_bio_stripe(rbio, bio);
1421
1422 if (failed < 0)
1423 return -EIO;
1424
1425 return fail_rbio_index(rbio, failed);
1426}
1427
1428/*
1429 * this sets each page in the bio uptodate. It should only be used on private
1430 * rbio pages, nothing that comes in from the higher layers
1431 */
1432static void set_bio_pages_uptodate(struct bio *bio)
1433{
1434 int i;
1435 struct page *p;
1436
1437 for (i = 0; i < bio->bi_vcnt; i++) {
1438 p = bio->bi_io_vec[i].bv_page;
1439 SetPageUptodate(p);
1440 }
1441}
1442
1443/*
1444 * end io for the read phase of the rmw cycle. All the bios here are physical
1445 * stripe bios we've read from the disk so we can recalculate the parity of the
1446 * stripe.
1447 *
1448 * This will usually kick off finish_rmw once all the bios are read in, but it
1449 * may trigger parity reconstruction if we had any errors along the way
1450 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001451static void raid_rmw_end_io(struct bio *bio)
David Woodhouse53b381b2013-01-29 18:40:14 -05001452{
1453 struct btrfs_raid_bio *rbio = bio->bi_private;
1454
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001455 if (bio->bi_error)
David Woodhouse53b381b2013-01-29 18:40:14 -05001456 fail_bio_stripe(rbio, bio);
1457 else
1458 set_bio_pages_uptodate(bio);
1459
1460 bio_put(bio);
1461
Miao Xieb89e1b02014-10-15 11:18:44 +08001462 if (!atomic_dec_and_test(&rbio->stripes_pending))
David Woodhouse53b381b2013-01-29 18:40:14 -05001463 return;
1464
Miao Xieb89e1b02014-10-15 11:18:44 +08001465 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
David Woodhouse53b381b2013-01-29 18:40:14 -05001466 goto cleanup;
1467
1468 /*
1469 * this will normally call finish_rmw to start our write
1470 * but if there are any failed stripes we'll reconstruct
1471 * from parity first
1472 */
1473 validate_rbio_for_rmw(rbio);
1474 return;
1475
1476cleanup:
1477
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001478 rbio_orig_end_io(rbio, -EIO);
David Woodhouse53b381b2013-01-29 18:40:14 -05001479}
1480
1481static void async_rmw_stripe(struct btrfs_raid_bio *rbio)
1482{
Liu Bo9e0af232014-08-15 23:36:53 +08001483 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
1484 rmw_work, NULL, NULL);
David Woodhouse53b381b2013-01-29 18:40:14 -05001485
Qu Wenruod05a33a2014-02-28 10:46:11 +08001486 btrfs_queue_work(rbio->fs_info->rmw_workers,
1487 &rbio->work);
David Woodhouse53b381b2013-01-29 18:40:14 -05001488}
1489
1490static void async_read_rebuild(struct btrfs_raid_bio *rbio)
1491{
Liu Bo9e0af232014-08-15 23:36:53 +08001492 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
1493 read_rebuild_work, NULL, NULL);
David Woodhouse53b381b2013-01-29 18:40:14 -05001494
Qu Wenruod05a33a2014-02-28 10:46:11 +08001495 btrfs_queue_work(rbio->fs_info->rmw_workers,
1496 &rbio->work);
David Woodhouse53b381b2013-01-29 18:40:14 -05001497}
1498
1499/*
1500 * the stripe must be locked by the caller. It will
1501 * unlock after all the writes are done
1502 */
1503static int raid56_rmw_stripe(struct btrfs_raid_bio *rbio)
1504{
1505 int bios_to_read = 0;
David Woodhouse53b381b2013-01-29 18:40:14 -05001506 struct bio_list bio_list;
1507 int ret;
David Sterbaed6078f2014-06-05 01:59:57 +02001508 int nr_pages = DIV_ROUND_UP(rbio->stripe_len, PAGE_CACHE_SIZE);
David Woodhouse53b381b2013-01-29 18:40:14 -05001509 int pagenr;
1510 int stripe;
1511 struct bio *bio;
1512
1513 bio_list_init(&bio_list);
1514
1515 ret = alloc_rbio_pages(rbio);
1516 if (ret)
1517 goto cleanup;
1518
1519 index_rbio_pages(rbio);
1520
Miao Xieb89e1b02014-10-15 11:18:44 +08001521 atomic_set(&rbio->error, 0);
David Woodhouse53b381b2013-01-29 18:40:14 -05001522 /*
1523 * build a list of bios to read all the missing parts of this
1524 * stripe
1525 */
1526 for (stripe = 0; stripe < rbio->nr_data; stripe++) {
1527 for (pagenr = 0; pagenr < nr_pages; pagenr++) {
1528 struct page *page;
1529 /*
1530 * we want to find all the pages missing from
1531 * the rbio and read them from the disk. If
1532 * page_in_rbio finds a page in the bio list
1533 * we don't need to read it off the stripe.
1534 */
1535 page = page_in_rbio(rbio, stripe, pagenr, 1);
1536 if (page)
1537 continue;
1538
1539 page = rbio_stripe_page(rbio, stripe, pagenr);
Chris Mason4ae10b32013-01-31 14:42:09 -05001540 /*
1541 * the bio cache may have handed us an uptodate
1542 * page. If so, be happy and use it
1543 */
1544 if (PageUptodate(page))
1545 continue;
1546
David Woodhouse53b381b2013-01-29 18:40:14 -05001547 ret = rbio_add_io_page(rbio, &bio_list, page,
1548 stripe, pagenr, rbio->stripe_len);
1549 if (ret)
1550 goto cleanup;
1551 }
1552 }
1553
1554 bios_to_read = bio_list_size(&bio_list);
1555 if (!bios_to_read) {
1556 /*
1557 * this can happen if others have merged with
1558 * us, it means there is nothing left to read.
1559 * But if there are missing devices it may not be
1560 * safe to do the full stripe write yet.
1561 */
1562 goto finish;
1563 }
1564
1565 /*
1566 * the bbio may be freed once we submit the last bio. Make sure
1567 * not to touch it after that
1568 */
Miao Xieb89e1b02014-10-15 11:18:44 +08001569 atomic_set(&rbio->stripes_pending, bios_to_read);
David Woodhouse53b381b2013-01-29 18:40:14 -05001570 while (1) {
1571 bio = bio_list_pop(&bio_list);
1572 if (!bio)
1573 break;
1574
1575 bio->bi_private = rbio;
1576 bio->bi_end_io = raid_rmw_end_io;
1577
1578 btrfs_bio_wq_end_io(rbio->fs_info, bio,
1579 BTRFS_WQ_ENDIO_RAID56);
1580
David Woodhouse53b381b2013-01-29 18:40:14 -05001581 submit_bio(READ, bio);
1582 }
1583 /* the actual write will happen once the reads are done */
1584 return 0;
1585
1586cleanup:
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001587 rbio_orig_end_io(rbio, -EIO);
David Woodhouse53b381b2013-01-29 18:40:14 -05001588 return -EIO;
1589
1590finish:
1591 validate_rbio_for_rmw(rbio);
1592 return 0;
1593}
1594
1595/*
1596 * if the upper layers pass in a full stripe, we thank them by only allocating
1597 * enough pages to hold the parity, and sending it all down quickly.
1598 */
1599static int full_stripe_write(struct btrfs_raid_bio *rbio)
1600{
1601 int ret;
1602
1603 ret = alloc_rbio_parity_pages(rbio);
Miao Xie3cd846d2013-07-22 16:36:57 +08001604 if (ret) {
1605 __free_raid_bio(rbio);
David Woodhouse53b381b2013-01-29 18:40:14 -05001606 return ret;
Miao Xie3cd846d2013-07-22 16:36:57 +08001607 }
David Woodhouse53b381b2013-01-29 18:40:14 -05001608
1609 ret = lock_stripe_add(rbio);
1610 if (ret == 0)
1611 finish_rmw(rbio);
1612 return 0;
1613}
1614
1615/*
1616 * partial stripe writes get handed over to async helpers.
1617 * We're really hoping to merge a few more writes into this
1618 * rbio before calculating new parity
1619 */
1620static int partial_stripe_write(struct btrfs_raid_bio *rbio)
1621{
1622 int ret;
1623
1624 ret = lock_stripe_add(rbio);
1625 if (ret == 0)
1626 async_rmw_stripe(rbio);
1627 return 0;
1628}
1629
1630/*
1631 * sometimes while we were reading from the drive to
1632 * recalculate parity, enough new bios come into create
1633 * a full stripe. So we do a check here to see if we can
1634 * go directly to finish_rmw
1635 */
1636static int __raid56_parity_write(struct btrfs_raid_bio *rbio)
1637{
1638 /* head off into rmw land if we don't have a full stripe */
1639 if (!rbio_is_full(rbio))
1640 return partial_stripe_write(rbio);
1641 return full_stripe_write(rbio);
1642}
1643
1644/*
Chris Mason6ac0f482013-01-31 14:42:28 -05001645 * We use plugging call backs to collect full stripes.
1646 * Any time we get a partial stripe write while plugged
1647 * we collect it into a list. When the unplug comes down,
1648 * we sort the list by logical block number and merge
1649 * everything we can into the same rbios
1650 */
1651struct btrfs_plug_cb {
1652 struct blk_plug_cb cb;
1653 struct btrfs_fs_info *info;
1654 struct list_head rbio_list;
1655 struct btrfs_work work;
1656};
1657
1658/*
1659 * rbios on the plug list are sorted for easier merging.
1660 */
1661static int plug_cmp(void *priv, struct list_head *a, struct list_head *b)
1662{
1663 struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio,
1664 plug_list);
1665 struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio,
1666 plug_list);
Kent Overstreet4f024f32013-10-11 15:44:27 -07001667 u64 a_sector = ra->bio_list.head->bi_iter.bi_sector;
1668 u64 b_sector = rb->bio_list.head->bi_iter.bi_sector;
Chris Mason6ac0f482013-01-31 14:42:28 -05001669
1670 if (a_sector < b_sector)
1671 return -1;
1672 if (a_sector > b_sector)
1673 return 1;
1674 return 0;
1675}
1676
1677static void run_plug(struct btrfs_plug_cb *plug)
1678{
1679 struct btrfs_raid_bio *cur;
1680 struct btrfs_raid_bio *last = NULL;
1681
1682 /*
1683 * sort our plug list then try to merge
1684 * everything we can in hopes of creating full
1685 * stripes.
1686 */
1687 list_sort(NULL, &plug->rbio_list, plug_cmp);
1688 while (!list_empty(&plug->rbio_list)) {
1689 cur = list_entry(plug->rbio_list.next,
1690 struct btrfs_raid_bio, plug_list);
1691 list_del_init(&cur->plug_list);
1692
1693 if (rbio_is_full(cur)) {
1694 /* we have a full stripe, send it down */
1695 full_stripe_write(cur);
1696 continue;
1697 }
1698 if (last) {
1699 if (rbio_can_merge(last, cur)) {
1700 merge_rbio(last, cur);
1701 __free_raid_bio(cur);
1702 continue;
1703
1704 }
1705 __raid56_parity_write(last);
1706 }
1707 last = cur;
1708 }
1709 if (last) {
1710 __raid56_parity_write(last);
1711 }
1712 kfree(plug);
1713}
1714
1715/*
1716 * if the unplug comes from schedule, we have to push the
1717 * work off to a helper thread
1718 */
1719static void unplug_work(struct btrfs_work *work)
1720{
1721 struct btrfs_plug_cb *plug;
1722 plug = container_of(work, struct btrfs_plug_cb, work);
1723 run_plug(plug);
1724}
1725
1726static void btrfs_raid_unplug(struct blk_plug_cb *cb, bool from_schedule)
1727{
1728 struct btrfs_plug_cb *plug;
1729 plug = container_of(cb, struct btrfs_plug_cb, cb);
1730
1731 if (from_schedule) {
Liu Bo9e0af232014-08-15 23:36:53 +08001732 btrfs_init_work(&plug->work, btrfs_rmw_helper,
1733 unplug_work, NULL, NULL);
Qu Wenruod05a33a2014-02-28 10:46:11 +08001734 btrfs_queue_work(plug->info->rmw_workers,
1735 &plug->work);
Chris Mason6ac0f482013-01-31 14:42:28 -05001736 return;
1737 }
1738 run_plug(plug);
1739}
1740
1741/*
David Woodhouse53b381b2013-01-29 18:40:14 -05001742 * our main entry point for writes from the rest of the FS.
1743 */
1744int raid56_parity_write(struct btrfs_root *root, struct bio *bio,
Zhao Lei8e5cfb52015-01-20 15:11:33 +08001745 struct btrfs_bio *bbio, u64 stripe_len)
David Woodhouse53b381b2013-01-29 18:40:14 -05001746{
1747 struct btrfs_raid_bio *rbio;
Chris Mason6ac0f482013-01-31 14:42:28 -05001748 struct btrfs_plug_cb *plug = NULL;
1749 struct blk_plug_cb *cb;
Miao Xie42452152014-11-25 16:39:28 +08001750 int ret;
David Woodhouse53b381b2013-01-29 18:40:14 -05001751
Zhao Lei8e5cfb52015-01-20 15:11:33 +08001752 rbio = alloc_rbio(root, bbio, stripe_len);
Miao Xieaf8e2d12014-10-23 14:42:50 +08001753 if (IS_ERR(rbio)) {
Zhao Lei6e9606d2015-01-20 15:11:34 +08001754 btrfs_put_bbio(bbio);
David Woodhouse53b381b2013-01-29 18:40:14 -05001755 return PTR_ERR(rbio);
Miao Xieaf8e2d12014-10-23 14:42:50 +08001756 }
David Woodhouse53b381b2013-01-29 18:40:14 -05001757 bio_list_add(&rbio->bio_list, bio);
Kent Overstreet4f024f32013-10-11 15:44:27 -07001758 rbio->bio_list_bytes = bio->bi_iter.bi_size;
Miao Xie1b94b552014-11-06 16:14:21 +08001759 rbio->operation = BTRFS_RBIO_WRITE;
Chris Mason6ac0f482013-01-31 14:42:28 -05001760
Miao Xie42452152014-11-25 16:39:28 +08001761 btrfs_bio_counter_inc_noblocked(root->fs_info);
1762 rbio->generic_bio_cnt = 1;
1763
Chris Mason6ac0f482013-01-31 14:42:28 -05001764 /*
1765 * don't plug on full rbios, just get them out the door
1766 * as quickly as we can
1767 */
Miao Xie42452152014-11-25 16:39:28 +08001768 if (rbio_is_full(rbio)) {
1769 ret = full_stripe_write(rbio);
1770 if (ret)
1771 btrfs_bio_counter_dec(root->fs_info);
1772 return ret;
1773 }
Chris Mason6ac0f482013-01-31 14:42:28 -05001774
1775 cb = blk_check_plugged(btrfs_raid_unplug, root->fs_info,
1776 sizeof(*plug));
1777 if (cb) {
1778 plug = container_of(cb, struct btrfs_plug_cb, cb);
1779 if (!plug->info) {
1780 plug->info = root->fs_info;
1781 INIT_LIST_HEAD(&plug->rbio_list);
1782 }
1783 list_add_tail(&rbio->plug_list, &plug->rbio_list);
Miao Xie42452152014-11-25 16:39:28 +08001784 ret = 0;
Chris Mason6ac0f482013-01-31 14:42:28 -05001785 } else {
Miao Xie42452152014-11-25 16:39:28 +08001786 ret = __raid56_parity_write(rbio);
1787 if (ret)
1788 btrfs_bio_counter_dec(root->fs_info);
Chris Mason6ac0f482013-01-31 14:42:28 -05001789 }
Miao Xie42452152014-11-25 16:39:28 +08001790 return ret;
David Woodhouse53b381b2013-01-29 18:40:14 -05001791}
1792
1793/*
1794 * all parity reconstruction happens here. We've read in everything
1795 * we can find from the drives and this does the heavy lifting of
1796 * sorting the good from the bad.
1797 */
1798static void __raid_recover_end_io(struct btrfs_raid_bio *rbio)
1799{
1800 int pagenr, stripe;
1801 void **pointers;
1802 int faila = -1, failb = -1;
David Sterbaed6078f2014-06-05 01:59:57 +02001803 int nr_pages = DIV_ROUND_UP(rbio->stripe_len, PAGE_CACHE_SIZE);
David Woodhouse53b381b2013-01-29 18:40:14 -05001804 struct page *page;
1805 int err;
1806 int i;
1807
David Sterba31e818f2015-02-20 18:00:26 +01001808 pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS);
David Woodhouse53b381b2013-01-29 18:40:14 -05001809 if (!pointers) {
1810 err = -ENOMEM;
1811 goto cleanup_io;
1812 }
1813
1814 faila = rbio->faila;
1815 failb = rbio->failb;
1816
Omar Sandovalb4ee1782015-06-19 11:52:50 -07001817 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1818 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001819 spin_lock_irq(&rbio->bio_list_lock);
1820 set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags);
1821 spin_unlock_irq(&rbio->bio_list_lock);
1822 }
1823
1824 index_rbio_pages(rbio);
1825
1826 for (pagenr = 0; pagenr < nr_pages; pagenr++) {
Miao Xie5a6ac9e2014-11-06 17:20:58 +08001827 /*
1828 * Now we just use bitmap to mark the horizontal stripes in
1829 * which we have data when doing parity scrub.
1830 */
1831 if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB &&
1832 !test_bit(pagenr, rbio->dbitmap))
1833 continue;
1834
David Woodhouse53b381b2013-01-29 18:40:14 -05001835 /* setup our array of pointers with pages
1836 * from each stripe
1837 */
Miao Xie2c8cdd62014-11-14 16:06:25 +08001838 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001839 /*
1840 * if we're rebuilding a read, we have to use
1841 * pages from the bio list
1842 */
Omar Sandovalb4ee1782015-06-19 11:52:50 -07001843 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1844 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
David Woodhouse53b381b2013-01-29 18:40:14 -05001845 (stripe == faila || stripe == failb)) {
1846 page = page_in_rbio(rbio, stripe, pagenr, 0);
1847 } else {
1848 page = rbio_stripe_page(rbio, stripe, pagenr);
1849 }
1850 pointers[stripe] = kmap(page);
1851 }
1852
1853 /* all raid6 handling here */
Zhao Lei10f11902015-01-20 15:11:43 +08001854 if (rbio->bbio->map_type & BTRFS_BLOCK_GROUP_RAID6) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001855 /*
1856 * single failure, rebuild from parity raid5
1857 * style
1858 */
1859 if (failb < 0) {
1860 if (faila == rbio->nr_data) {
1861 /*
1862 * Just the P stripe has failed, without
1863 * a bad data or Q stripe.
1864 * TODO, we should redo the xor here.
1865 */
1866 err = -EIO;
1867 goto cleanup;
1868 }
1869 /*
1870 * a single failure in raid6 is rebuilt
1871 * in the pstripe code below
1872 */
1873 goto pstripe;
1874 }
1875
1876 /* make sure our ps and qs are in order */
1877 if (faila > failb) {
1878 int tmp = failb;
1879 failb = faila;
1880 faila = tmp;
1881 }
1882
1883 /* if the q stripe is failed, do a pstripe reconstruction
1884 * from the xors.
1885 * If both the q stripe and the P stripe are failed, we're
1886 * here due to a crc mismatch and we can't give them the
1887 * data they want
1888 */
Zhao Lei8e5cfb52015-01-20 15:11:33 +08001889 if (rbio->bbio->raid_map[failb] == RAID6_Q_STRIPE) {
1890 if (rbio->bbio->raid_map[faila] ==
1891 RAID5_P_STRIPE) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001892 err = -EIO;
1893 goto cleanup;
1894 }
1895 /*
1896 * otherwise we have one bad data stripe and
1897 * a good P stripe. raid5!
1898 */
1899 goto pstripe;
1900 }
1901
Zhao Lei8e5cfb52015-01-20 15:11:33 +08001902 if (rbio->bbio->raid_map[failb] == RAID5_P_STRIPE) {
Miao Xie2c8cdd62014-11-14 16:06:25 +08001903 raid6_datap_recov(rbio->real_stripes,
David Woodhouse53b381b2013-01-29 18:40:14 -05001904 PAGE_SIZE, faila, pointers);
1905 } else {
Miao Xie2c8cdd62014-11-14 16:06:25 +08001906 raid6_2data_recov(rbio->real_stripes,
David Woodhouse53b381b2013-01-29 18:40:14 -05001907 PAGE_SIZE, faila, failb,
1908 pointers);
1909 }
1910 } else {
1911 void *p;
1912
1913 /* rebuild from P stripe here (raid5 or raid6) */
1914 BUG_ON(failb != -1);
1915pstripe:
1916 /* Copy parity block into failed block to start with */
1917 memcpy(pointers[faila],
1918 pointers[rbio->nr_data],
1919 PAGE_CACHE_SIZE);
1920
1921 /* rearrange the pointer array */
1922 p = pointers[faila];
1923 for (stripe = faila; stripe < rbio->nr_data - 1; stripe++)
1924 pointers[stripe] = pointers[stripe + 1];
1925 pointers[rbio->nr_data - 1] = p;
1926
1927 /* xor in the rest */
1928 run_xor(pointers, rbio->nr_data - 1, PAGE_CACHE_SIZE);
1929 }
1930 /* if we're doing this rebuild as part of an rmw, go through
1931 * and set all of our private rbio pages in the
1932 * failed stripes as uptodate. This way finish_rmw will
1933 * know they can be trusted. If this was a read reconstruction,
1934 * other endio functions will fiddle the uptodate bits
1935 */
Miao Xie1b94b552014-11-06 16:14:21 +08001936 if (rbio->operation == BTRFS_RBIO_WRITE) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001937 for (i = 0; i < nr_pages; i++) {
1938 if (faila != -1) {
1939 page = rbio_stripe_page(rbio, faila, i);
1940 SetPageUptodate(page);
1941 }
1942 if (failb != -1) {
1943 page = rbio_stripe_page(rbio, failb, i);
1944 SetPageUptodate(page);
1945 }
1946 }
1947 }
Miao Xie2c8cdd62014-11-14 16:06:25 +08001948 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
David Woodhouse53b381b2013-01-29 18:40:14 -05001949 /*
1950 * if we're rebuilding a read, we have to use
1951 * pages from the bio list
1952 */
Omar Sandovalb4ee1782015-06-19 11:52:50 -07001953 if ((rbio->operation == BTRFS_RBIO_READ_REBUILD ||
1954 rbio->operation == BTRFS_RBIO_REBUILD_MISSING) &&
David Woodhouse53b381b2013-01-29 18:40:14 -05001955 (stripe == faila || stripe == failb)) {
1956 page = page_in_rbio(rbio, stripe, pagenr, 0);
1957 } else {
1958 page = rbio_stripe_page(rbio, stripe, pagenr);
1959 }
1960 kunmap(page);
1961 }
1962 }
1963
1964 err = 0;
1965cleanup:
1966 kfree(pointers);
1967
1968cleanup_io:
Miao Xie1b94b552014-11-06 16:14:21 +08001969 if (rbio->operation == BTRFS_RBIO_READ_REBUILD) {
Zhao Lei6e9606d2015-01-20 15:11:34 +08001970 if (err == 0)
Chris Mason4ae10b32013-01-31 14:42:09 -05001971 cache_rbio_pages(rbio);
1972 else
1973 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
1974
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001975 rbio_orig_end_io(rbio, err);
Omar Sandovalb4ee1782015-06-19 11:52:50 -07001976 } else if (rbio->operation == BTRFS_RBIO_REBUILD_MISSING) {
Linus Torvalds22365972015-09-05 15:14:43 -07001977 rbio_orig_end_io(rbio, err);
David Woodhouse53b381b2013-01-29 18:40:14 -05001978 } else if (err == 0) {
1979 rbio->faila = -1;
1980 rbio->failb = -1;
Miao Xie5a6ac9e2014-11-06 17:20:58 +08001981
1982 if (rbio->operation == BTRFS_RBIO_WRITE)
1983 finish_rmw(rbio);
1984 else if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB)
1985 finish_parity_scrub(rbio, 0);
1986 else
1987 BUG();
David Woodhouse53b381b2013-01-29 18:40:14 -05001988 } else {
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001989 rbio_orig_end_io(rbio, err);
David Woodhouse53b381b2013-01-29 18:40:14 -05001990 }
1991}
1992
1993/*
1994 * This is called only for stripes we've read from disk to
1995 * reconstruct the parity.
1996 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001997static void raid_recover_end_io(struct bio *bio)
David Woodhouse53b381b2013-01-29 18:40:14 -05001998{
1999 struct btrfs_raid_bio *rbio = bio->bi_private;
2000
2001 /*
2002 * we only read stripe pages off the disk, set them
2003 * up to date if there were no errors
2004 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002005 if (bio->bi_error)
David Woodhouse53b381b2013-01-29 18:40:14 -05002006 fail_bio_stripe(rbio, bio);
2007 else
2008 set_bio_pages_uptodate(bio);
2009 bio_put(bio);
2010
Miao Xieb89e1b02014-10-15 11:18:44 +08002011 if (!atomic_dec_and_test(&rbio->stripes_pending))
David Woodhouse53b381b2013-01-29 18:40:14 -05002012 return;
2013
Miao Xieb89e1b02014-10-15 11:18:44 +08002014 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002015 rbio_orig_end_io(rbio, -EIO);
David Woodhouse53b381b2013-01-29 18:40:14 -05002016 else
2017 __raid_recover_end_io(rbio);
2018}
2019
2020/*
2021 * reads everything we need off the disk to reconstruct
2022 * the parity. endio handlers trigger final reconstruction
2023 * when the IO is done.
2024 *
2025 * This is used both for reads from the higher layers and for
2026 * parity construction required to finish a rmw cycle.
2027 */
2028static int __raid56_parity_recover(struct btrfs_raid_bio *rbio)
2029{
2030 int bios_to_read = 0;
David Woodhouse53b381b2013-01-29 18:40:14 -05002031 struct bio_list bio_list;
2032 int ret;
David Sterbaed6078f2014-06-05 01:59:57 +02002033 int nr_pages = DIV_ROUND_UP(rbio->stripe_len, PAGE_CACHE_SIZE);
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
2057 for (pagenr = 0; pagenr < nr_pages; pagenr++) {
2058 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;
2103
2104 btrfs_bio_wq_end_io(rbio->fs_info, bio,
2105 BTRFS_WQ_ENDIO_RAID56);
2106
David Woodhouse53b381b2013-01-29 18:40:14 -05002107 submit_bio(READ, bio);
2108 }
2109out:
2110 return 0;
2111
2112cleanup:
Omar Sandovalb4ee1782015-06-19 11:52:50 -07002113 if (rbio->operation == BTRFS_RBIO_READ_REBUILD ||
2114 rbio->operation == BTRFS_RBIO_REBUILD_MISSING)
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002115 rbio_orig_end_io(rbio, -EIO);
David Woodhouse53b381b2013-01-29 18:40:14 -05002116 return -EIO;
2117}
2118
2119/*
2120 * the main entry point for reads from the higher layers. This
2121 * is really only called when the normal read path had a failure,
2122 * so we assume the bio they send down corresponds to a failed part
2123 * of the drive.
2124 */
2125int raid56_parity_recover(struct btrfs_root *root, struct bio *bio,
Zhao Lei8e5cfb52015-01-20 15:11:33 +08002126 struct btrfs_bio *bbio, u64 stripe_len,
2127 int mirror_num, int generic_io)
David Woodhouse53b381b2013-01-29 18:40:14 -05002128{
2129 struct btrfs_raid_bio *rbio;
2130 int ret;
2131
Zhao Lei8e5cfb52015-01-20 15:11:33 +08002132 rbio = alloc_rbio(root, bbio, stripe_len);
Miao Xieaf8e2d12014-10-23 14:42:50 +08002133 if (IS_ERR(rbio)) {
Zhao Lei6e9606d2015-01-20 15:11:34 +08002134 if (generic_io)
2135 btrfs_put_bbio(bbio);
David Woodhouse53b381b2013-01-29 18:40:14 -05002136 return PTR_ERR(rbio);
Miao Xieaf8e2d12014-10-23 14:42:50 +08002137 }
David Woodhouse53b381b2013-01-29 18:40:14 -05002138
Miao Xie1b94b552014-11-06 16:14:21 +08002139 rbio->operation = BTRFS_RBIO_READ_REBUILD;
David Woodhouse53b381b2013-01-29 18:40:14 -05002140 bio_list_add(&rbio->bio_list, bio);
Kent Overstreet4f024f32013-10-11 15:44:27 -07002141 rbio->bio_list_bytes = bio->bi_iter.bi_size;
David Woodhouse53b381b2013-01-29 18:40:14 -05002142
2143 rbio->faila = find_logical_bio_stripe(rbio, bio);
2144 if (rbio->faila == -1) {
2145 BUG();
Zhao Lei6e9606d2015-01-20 15:11:34 +08002146 if (generic_io)
2147 btrfs_put_bbio(bbio);
David Woodhouse53b381b2013-01-29 18:40:14 -05002148 kfree(rbio);
2149 return -EIO;
2150 }
2151
Miao Xie42452152014-11-25 16:39:28 +08002152 if (generic_io) {
2153 btrfs_bio_counter_inc_noblocked(root->fs_info);
2154 rbio->generic_bio_cnt = 1;
2155 } else {
Zhao Lei6e9606d2015-01-20 15:11:34 +08002156 btrfs_get_bbio(bbio);
Miao Xie42452152014-11-25 16:39:28 +08002157 }
2158
David Woodhouse53b381b2013-01-29 18:40:14 -05002159 /*
2160 * reconstruct from the q stripe if they are
2161 * asking for mirror 3
2162 */
2163 if (mirror_num == 3)
Miao Xie2c8cdd62014-11-14 16:06:25 +08002164 rbio->failb = rbio->real_stripes - 2;
David Woodhouse53b381b2013-01-29 18:40:14 -05002165
2166 ret = lock_stripe_add(rbio);
2167
2168 /*
2169 * __raid56_parity_recover will end the bio with
2170 * any errors it hits. We don't want to return
2171 * its error value up the stack because our caller
2172 * will end up calling bio_endio with any nonzero
2173 * return
2174 */
2175 if (ret == 0)
2176 __raid56_parity_recover(rbio);
2177 /*
2178 * our rbio has been added to the list of
2179 * rbios that will be handled after the
2180 * currently lock owner is done
2181 */
2182 return 0;
2183
2184}
2185
2186static void rmw_work(struct btrfs_work *work)
2187{
2188 struct btrfs_raid_bio *rbio;
2189
2190 rbio = container_of(work, struct btrfs_raid_bio, work);
2191 raid56_rmw_stripe(rbio);
2192}
2193
2194static void read_rebuild_work(struct btrfs_work *work)
2195{
2196 struct btrfs_raid_bio *rbio;
2197
2198 rbio = container_of(work, struct btrfs_raid_bio, work);
2199 __raid56_parity_recover(rbio);
2200}
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002201
2202/*
2203 * The following code is used to scrub/replace the parity stripe
2204 *
2205 * Note: We need make sure all the pages that add into the scrub/replace
2206 * raid bio are correct and not be changed during the scrub/replace. That
2207 * is those pages just hold metadata or file data with checksum.
2208 */
2209
2210struct btrfs_raid_bio *
2211raid56_parity_alloc_scrub_rbio(struct btrfs_root *root, struct bio *bio,
Zhao Lei8e5cfb52015-01-20 15:11:33 +08002212 struct btrfs_bio *bbio, u64 stripe_len,
2213 struct btrfs_device *scrub_dev,
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002214 unsigned long *dbitmap, int stripe_nsectors)
2215{
2216 struct btrfs_raid_bio *rbio;
2217 int i;
2218
Zhao Lei8e5cfb52015-01-20 15:11:33 +08002219 rbio = alloc_rbio(root, bbio, stripe_len);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002220 if (IS_ERR(rbio))
2221 return NULL;
2222 bio_list_add(&rbio->bio_list, bio);
2223 /*
2224 * This is a special bio which is used to hold the completion handler
2225 * and make the scrub rbio is similar to the other types
2226 */
2227 ASSERT(!bio->bi_iter.bi_size);
2228 rbio->operation = BTRFS_RBIO_PARITY_SCRUB;
2229
Miao Xie2c8cdd62014-11-14 16:06:25 +08002230 for (i = 0; i < rbio->real_stripes; i++) {
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002231 if (bbio->stripes[i].dev == scrub_dev) {
2232 rbio->scrubp = i;
2233 break;
2234 }
2235 }
2236
2237 /* Now we just support the sectorsize equals to page size */
2238 ASSERT(root->sectorsize == PAGE_SIZE);
2239 ASSERT(rbio->stripe_npages == stripe_nsectors);
2240 bitmap_copy(rbio->dbitmap, dbitmap, stripe_nsectors);
2241
2242 return rbio;
2243}
2244
Omar Sandovalb4ee1782015-06-19 11:52:50 -07002245/* Used for both parity scrub and missing. */
2246void raid56_add_scrub_pages(struct btrfs_raid_bio *rbio, struct page *page,
2247 u64 logical)
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002248{
2249 int stripe_offset;
2250 int index;
2251
Zhao Lei8e5cfb52015-01-20 15:11:33 +08002252 ASSERT(logical >= rbio->bbio->raid_map[0]);
2253 ASSERT(logical + PAGE_SIZE <= rbio->bbio->raid_map[0] +
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002254 rbio->stripe_len * rbio->nr_data);
Zhao Lei8e5cfb52015-01-20 15:11:33 +08002255 stripe_offset = (int)(logical - rbio->bbio->raid_map[0]);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002256 index = stripe_offset >> PAGE_CACHE_SHIFT;
2257 rbio->bio_pages[index] = page;
2258}
2259
2260/*
2261 * We just scrub the parity that we have correct data on the same horizontal,
2262 * so we needn't allocate all pages for all the stripes.
2263 */
2264static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio)
2265{
2266 int i;
2267 int bit;
2268 int index;
2269 struct page *page;
2270
2271 for_each_set_bit(bit, rbio->dbitmap, rbio->stripe_npages) {
Miao Xie2c8cdd62014-11-14 16:06:25 +08002272 for (i = 0; i < rbio->real_stripes; i++) {
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002273 index = i * rbio->stripe_npages + bit;
2274 if (rbio->stripe_pages[index])
2275 continue;
2276
2277 page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2278 if (!page)
2279 return -ENOMEM;
2280 rbio->stripe_pages[index] = page;
2281 ClearPageUptodate(page);
2282 }
2283 }
2284 return 0;
2285}
2286
2287/*
2288 * end io function used by finish_rmw. When we finally
2289 * get here, we've written a full stripe
2290 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002291static void raid_write_parity_end_io(struct bio *bio)
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002292{
2293 struct btrfs_raid_bio *rbio = bio->bi_private;
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002294 int err = bio->bi_error;
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002295
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002296 if (bio->bi_error)
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002297 fail_bio_stripe(rbio, bio);
2298
2299 bio_put(bio);
2300
2301 if (!atomic_dec_and_test(&rbio->stripes_pending))
2302 return;
2303
2304 err = 0;
2305
2306 if (atomic_read(&rbio->error))
2307 err = -EIO;
2308
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002309 rbio_orig_end_io(rbio, err);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002310}
2311
2312static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio,
2313 int need_check)
2314{
Miao Xie76035972014-11-14 17:45:42 +08002315 struct btrfs_bio *bbio = rbio->bbio;
Miao Xie2c8cdd62014-11-14 16:06:25 +08002316 void *pointers[rbio->real_stripes];
Miao Xie76035972014-11-14 17:45:42 +08002317 DECLARE_BITMAP(pbitmap, rbio->stripe_npages);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002318 int nr_data = rbio->nr_data;
2319 int stripe;
2320 int pagenr;
2321 int p_stripe = -1;
2322 int q_stripe = -1;
2323 struct page *p_page = NULL;
2324 struct page *q_page = NULL;
2325 struct bio_list bio_list;
2326 struct bio *bio;
Miao Xie76035972014-11-14 17:45:42 +08002327 int is_replace = 0;
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002328 int ret;
2329
2330 bio_list_init(&bio_list);
2331
Miao Xie2c8cdd62014-11-14 16:06:25 +08002332 if (rbio->real_stripes - rbio->nr_data == 1) {
2333 p_stripe = rbio->real_stripes - 1;
2334 } else if (rbio->real_stripes - rbio->nr_data == 2) {
2335 p_stripe = rbio->real_stripes - 2;
2336 q_stripe = rbio->real_stripes - 1;
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002337 } else {
2338 BUG();
2339 }
2340
Miao Xie76035972014-11-14 17:45:42 +08002341 if (bbio->num_tgtdevs && bbio->tgtdev_map[rbio->scrubp]) {
2342 is_replace = 1;
2343 bitmap_copy(pbitmap, rbio->dbitmap, rbio->stripe_npages);
2344 }
2345
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002346 /*
2347 * Because the higher layers(scrubber) are unlikely to
2348 * use this area of the disk again soon, so don't cache
2349 * it.
2350 */
2351 clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags);
2352
2353 if (!need_check)
2354 goto writeback;
2355
2356 p_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2357 if (!p_page)
2358 goto cleanup;
2359 SetPageUptodate(p_page);
2360
2361 if (q_stripe != -1) {
2362 q_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
2363 if (!q_page) {
2364 __free_page(p_page);
2365 goto cleanup;
2366 }
2367 SetPageUptodate(q_page);
2368 }
2369
2370 atomic_set(&rbio->error, 0);
2371
2372 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2373 struct page *p;
2374 void *parity;
2375 /* first collect one page from each data stripe */
2376 for (stripe = 0; stripe < nr_data; stripe++) {
2377 p = page_in_rbio(rbio, stripe, pagenr, 0);
2378 pointers[stripe] = kmap(p);
2379 }
2380
2381 /* then add the parity stripe */
2382 pointers[stripe++] = kmap(p_page);
2383
2384 if (q_stripe != -1) {
2385
2386 /*
2387 * raid6, add the qstripe and call the
2388 * library function to fill in our p/q
2389 */
2390 pointers[stripe++] = kmap(q_page);
2391
Miao Xie2c8cdd62014-11-14 16:06:25 +08002392 raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE,
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002393 pointers);
2394 } else {
2395 /* raid5 */
2396 memcpy(pointers[nr_data], pointers[0], PAGE_SIZE);
2397 run_xor(pointers + 1, nr_data - 1, PAGE_CACHE_SIZE);
2398 }
2399
2400 /* Check scrubbing pairty and repair it */
2401 p = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2402 parity = kmap(p);
2403 if (memcmp(parity, pointers[rbio->scrubp], PAGE_CACHE_SIZE))
2404 memcpy(parity, pointers[rbio->scrubp], PAGE_CACHE_SIZE);
2405 else
2406 /* Parity is right, needn't writeback */
2407 bitmap_clear(rbio->dbitmap, pagenr, 1);
2408 kunmap(p);
2409
Miao Xie2c8cdd62014-11-14 16:06:25 +08002410 for (stripe = 0; stripe < rbio->real_stripes; stripe++)
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002411 kunmap(page_in_rbio(rbio, stripe, pagenr, 0));
2412 }
2413
2414 __free_page(p_page);
2415 if (q_page)
2416 __free_page(q_page);
2417
2418writeback:
2419 /*
2420 * time to start writing. Make bios for everything from the
2421 * higher layers (the bio_list in our rbio) and our p/q. Ignore
2422 * everything else.
2423 */
2424 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2425 struct page *page;
2426
2427 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2428 ret = rbio_add_io_page(rbio, &bio_list,
2429 page, rbio->scrubp, pagenr, rbio->stripe_len);
2430 if (ret)
2431 goto cleanup;
2432 }
2433
Miao Xie76035972014-11-14 17:45:42 +08002434 if (!is_replace)
2435 goto submit_write;
2436
2437 for_each_set_bit(pagenr, pbitmap, rbio->stripe_npages) {
2438 struct page *page;
2439
2440 page = rbio_stripe_page(rbio, rbio->scrubp, pagenr);
2441 ret = rbio_add_io_page(rbio, &bio_list, page,
2442 bbio->tgtdev_map[rbio->scrubp],
2443 pagenr, rbio->stripe_len);
2444 if (ret)
2445 goto cleanup;
2446 }
2447
2448submit_write:
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002449 nr_data = bio_list_size(&bio_list);
2450 if (!nr_data) {
2451 /* Every parity is right */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002452 rbio_orig_end_io(rbio, 0);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002453 return;
2454 }
2455
2456 atomic_set(&rbio->stripes_pending, nr_data);
2457
2458 while (1) {
2459 bio = bio_list_pop(&bio_list);
2460 if (!bio)
2461 break;
2462
2463 bio->bi_private = rbio;
2464 bio->bi_end_io = raid_write_parity_end_io;
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002465 submit_bio(WRITE, bio);
2466 }
2467 return;
2468
2469cleanup:
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002470 rbio_orig_end_io(rbio, -EIO);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002471}
2472
2473static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe)
2474{
2475 if (stripe >= 0 && stripe < rbio->nr_data)
2476 return 1;
2477 return 0;
2478}
2479
2480/*
2481 * While we're doing the parity check and repair, we could have errors
2482 * in reading pages off the disk. This checks for errors and if we're
2483 * not able to read the page it'll trigger parity reconstruction. The
2484 * parity scrub will be finished after we've reconstructed the failed
2485 * stripes
2486 */
2487static void validate_rbio_for_parity_scrub(struct btrfs_raid_bio *rbio)
2488{
2489 if (atomic_read(&rbio->error) > rbio->bbio->max_errors)
2490 goto cleanup;
2491
2492 if (rbio->faila >= 0 || rbio->failb >= 0) {
2493 int dfail = 0, failp = -1;
2494
2495 if (is_data_stripe(rbio, rbio->faila))
2496 dfail++;
2497 else if (is_parity_stripe(rbio->faila))
2498 failp = rbio->faila;
2499
2500 if (is_data_stripe(rbio, rbio->failb))
2501 dfail++;
2502 else if (is_parity_stripe(rbio->failb))
2503 failp = rbio->failb;
2504
2505 /*
2506 * Because we can not use a scrubbing parity to repair
2507 * the data, so the capability of the repair is declined.
2508 * (In the case of RAID5, we can not repair anything)
2509 */
2510 if (dfail > rbio->bbio->max_errors - 1)
2511 goto cleanup;
2512
2513 /*
2514 * If all data is good, only parity is correctly, just
2515 * repair the parity.
2516 */
2517 if (dfail == 0) {
2518 finish_parity_scrub(rbio, 0);
2519 return;
2520 }
2521
2522 /*
2523 * Here means we got one corrupted data stripe and one
2524 * corrupted parity on RAID6, if the corrupted parity
2525 * is scrubbing parity, luckly, use the other one to repair
2526 * the data, or we can not repair the data stripe.
2527 */
2528 if (failp != rbio->scrubp)
2529 goto cleanup;
2530
2531 __raid_recover_end_io(rbio);
2532 } else {
2533 finish_parity_scrub(rbio, 1);
2534 }
2535 return;
2536
2537cleanup:
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002538 rbio_orig_end_io(rbio, -EIO);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002539}
2540
2541/*
2542 * end io for the read phase of the rmw cycle. All the bios here are physical
2543 * stripe bios we've read from the disk so we can recalculate the parity of the
2544 * stripe.
2545 *
2546 * This will usually kick off finish_rmw once all the bios are read in, but it
2547 * may trigger parity reconstruction if we had any errors along the way
2548 */
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002549static void raid56_parity_scrub_end_io(struct bio *bio)
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002550{
2551 struct btrfs_raid_bio *rbio = bio->bi_private;
2552
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002553 if (bio->bi_error)
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002554 fail_bio_stripe(rbio, bio);
2555 else
2556 set_bio_pages_uptodate(bio);
2557
2558 bio_put(bio);
2559
2560 if (!atomic_dec_and_test(&rbio->stripes_pending))
2561 return;
2562
2563 /*
2564 * this will normally call finish_rmw to start our write
2565 * but if there are any failed stripes we'll reconstruct
2566 * from parity first
2567 */
2568 validate_rbio_for_parity_scrub(rbio);
2569}
2570
2571static void raid56_parity_scrub_stripe(struct btrfs_raid_bio *rbio)
2572{
2573 int bios_to_read = 0;
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002574 struct bio_list bio_list;
2575 int ret;
2576 int pagenr;
2577 int stripe;
2578 struct bio *bio;
2579
2580 ret = alloc_rbio_essential_pages(rbio);
2581 if (ret)
2582 goto cleanup;
2583
2584 bio_list_init(&bio_list);
2585
2586 atomic_set(&rbio->error, 0);
2587 /*
2588 * build a list of bios to read all the missing parts of this
2589 * stripe
2590 */
Miao Xie2c8cdd62014-11-14 16:06:25 +08002591 for (stripe = 0; stripe < rbio->real_stripes; stripe++) {
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002592 for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) {
2593 struct page *page;
2594 /*
2595 * we want to find all the pages missing from
2596 * the rbio and read them from the disk. If
2597 * page_in_rbio finds a page in the bio list
2598 * we don't need to read it off the stripe.
2599 */
2600 page = page_in_rbio(rbio, stripe, pagenr, 1);
2601 if (page)
2602 continue;
2603
2604 page = rbio_stripe_page(rbio, stripe, pagenr);
2605 /*
2606 * the bio cache may have handed us an uptodate
2607 * page. If so, be happy and use it
2608 */
2609 if (PageUptodate(page))
2610 continue;
2611
2612 ret = rbio_add_io_page(rbio, &bio_list, page,
2613 stripe, pagenr, rbio->stripe_len);
2614 if (ret)
2615 goto cleanup;
2616 }
2617 }
2618
2619 bios_to_read = bio_list_size(&bio_list);
2620 if (!bios_to_read) {
2621 /*
2622 * this can happen if others have merged with
2623 * us, it means there is nothing left to read.
2624 * But if there are missing devices it may not be
2625 * safe to do the full stripe write yet.
2626 */
2627 goto finish;
2628 }
2629
2630 /*
2631 * the bbio may be freed once we submit the last bio. Make sure
2632 * not to touch it after that
2633 */
2634 atomic_set(&rbio->stripes_pending, bios_to_read);
2635 while (1) {
2636 bio = bio_list_pop(&bio_list);
2637 if (!bio)
2638 break;
2639
2640 bio->bi_private = rbio;
2641 bio->bi_end_io = raid56_parity_scrub_end_io;
2642
2643 btrfs_bio_wq_end_io(rbio->fs_info, bio,
2644 BTRFS_WQ_ENDIO_RAID56);
2645
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002646 submit_bio(READ, bio);
2647 }
2648 /* the actual write will happen once the reads are done */
2649 return;
2650
2651cleanup:
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02002652 rbio_orig_end_io(rbio, -EIO);
Miao Xie5a6ac9e2014-11-06 17:20:58 +08002653 return;
2654
2655finish:
2656 validate_rbio_for_parity_scrub(rbio);
2657}
2658
2659static void scrub_parity_work(struct btrfs_work *work)
2660{
2661 struct btrfs_raid_bio *rbio;
2662
2663 rbio = container_of(work, struct btrfs_raid_bio, work);
2664 raid56_parity_scrub_stripe(rbio);
2665}
2666
2667static void async_scrub_parity(struct btrfs_raid_bio *rbio)
2668{
2669 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
2670 scrub_parity_work, NULL, NULL);
2671
2672 btrfs_queue_work(rbio->fs_info->rmw_workers,
2673 &rbio->work);
2674}
2675
2676void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio)
2677{
2678 if (!lock_stripe_add(rbio))
2679 async_scrub_parity(rbio);
2680}
Omar Sandovalb4ee1782015-06-19 11:52:50 -07002681
2682/* The following code is used for dev replace of a missing RAID 5/6 device. */
2683
2684struct btrfs_raid_bio *
2685raid56_alloc_missing_rbio(struct btrfs_root *root, struct bio *bio,
2686 struct btrfs_bio *bbio, u64 length)
2687{
2688 struct btrfs_raid_bio *rbio;
2689
2690 rbio = alloc_rbio(root, bbio, length);
2691 if (IS_ERR(rbio))
2692 return NULL;
2693
2694 rbio->operation = BTRFS_RBIO_REBUILD_MISSING;
2695 bio_list_add(&rbio->bio_list, bio);
2696 /*
2697 * This is a special bio which is used to hold the completion handler
2698 * and make the scrub rbio is similar to the other types
2699 */
2700 ASSERT(!bio->bi_iter.bi_size);
2701
2702 rbio->faila = find_logical_bio_stripe(rbio, bio);
2703 if (rbio->faila == -1) {
2704 BUG();
2705 kfree(rbio);
2706 return NULL;
2707 }
2708
2709 return rbio;
2710}
2711
2712static void missing_raid56_work(struct btrfs_work *work)
2713{
2714 struct btrfs_raid_bio *rbio;
2715
2716 rbio = container_of(work, struct btrfs_raid_bio, work);
2717 __raid56_parity_recover(rbio);
2718}
2719
2720static void async_missing_raid56(struct btrfs_raid_bio *rbio)
2721{
2722 btrfs_init_work(&rbio->work, btrfs_rmw_helper,
2723 missing_raid56_work, NULL, NULL);
2724
2725 btrfs_queue_work(rbio->fs_info->rmw_workers, &rbio->work);
2726}
2727
2728void raid56_submit_missing_rbio(struct btrfs_raid_bio *rbio)
2729{
2730 if (!lock_stripe_add(rbio))
2731 async_missing_raid56(rbio);
2732}