]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Copyright (C) 2012 Fusion-io All rights reserved. | |
3 | * Copyright (C) 2012 Intel Corp. All rights reserved. | |
4 | * | |
5 | * This program is free software; you can redistribute it and/or | |
6 | * modify it under the terms of the GNU General Public | |
7 | * License v2 as published by the Free Software Foundation. | |
8 | * | |
9 | * This program is distributed in the hope that it will be useful, | |
10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
12 | * General Public License for more details. | |
13 | * | |
14 | * You should have received a copy of the GNU General Public | |
15 | * License along with this program; if not, write to the | |
16 | * Free Software Foundation, Inc., 59 Temple Place - Suite 330, | |
17 | * Boston, MA 021110-1307, USA. | |
18 | */ | |
19 | #include <linux/sched.h> | |
20 | #include <linux/wait.h> | |
21 | #include <linux/bio.h> | |
22 | #include <linux/slab.h> | |
23 | #include <linux/buffer_head.h> | |
24 | #include <linux/blkdev.h> | |
25 | #include <linux/random.h> | |
26 | #include <linux/iocontext.h> | |
27 | #include <linux/capability.h> | |
28 | #include <linux/ratelimit.h> | |
29 | #include <linux/kthread.h> | |
30 | #include <linux/raid/pq.h> | |
31 | #include <linux/hash.h> | |
32 | #include <linux/list_sort.h> | |
33 | #include <linux/raid/xor.h> | |
34 | #include <linux/vmalloc.h> | |
35 | #include <asm/div64.h> | |
36 | #include "ctree.h" | |
37 | #include "extent_map.h" | |
38 | #include "disk-io.h" | |
39 | #include "transaction.h" | |
40 | #include "print-tree.h" | |
41 | #include "volumes.h" | |
42 | #include "raid56.h" | |
43 | #include "async-thread.h" | |
44 | #include "check-integrity.h" | |
45 | #include "rcu-string.h" | |
46 | ||
47 | /* set when additional merges to this rbio are not allowed */ | |
48 | #define RBIO_RMW_LOCKED_BIT 1 | |
49 | ||
50 | /* | |
51 | * set when this rbio is sitting in the hash, but it is just a cache | |
52 | * of past RMW | |
53 | */ | |
54 | #define RBIO_CACHE_BIT 2 | |
55 | ||
56 | /* | |
57 | * set when it is safe to trust the stripe_pages for caching | |
58 | */ | |
59 | #define RBIO_CACHE_READY_BIT 3 | |
60 | ||
61 | #define RBIO_CACHE_SIZE 1024 | |
62 | ||
63 | enum btrfs_rbio_ops { | |
64 | BTRFS_RBIO_WRITE = 0, | |
65 | BTRFS_RBIO_READ_REBUILD = 1, | |
66 | BTRFS_RBIO_PARITY_SCRUB = 2, | |
67 | }; | |
68 | ||
69 | struct btrfs_raid_bio { | |
70 | struct btrfs_fs_info *fs_info; | |
71 | struct btrfs_bio *bbio; | |
72 | ||
73 | /* while we're doing rmw on a stripe | |
74 | * we put it into a hash table so we can | |
75 | * lock the stripe and merge more rbios | |
76 | * into it. | |
77 | */ | |
78 | struct list_head hash_list; | |
79 | ||
80 | /* | |
81 | * LRU list for the stripe cache | |
82 | */ | |
83 | struct list_head stripe_cache; | |
84 | ||
85 | /* | |
86 | * for scheduling work in the helper threads | |
87 | */ | |
88 | struct btrfs_work work; | |
89 | ||
90 | /* | |
91 | * bio list and bio_list_lock are used | |
92 | * to add more bios into the stripe | |
93 | * in hopes of avoiding the full rmw | |
94 | */ | |
95 | struct bio_list bio_list; | |
96 | spinlock_t bio_list_lock; | |
97 | ||
98 | /* also protected by the bio_list_lock, the | |
99 | * plug list is used by the plugging code | |
100 | * to collect partial bios while plugged. The | |
101 | * stripe locking code also uses it to hand off | |
102 | * the stripe lock to the next pending IO | |
103 | */ | |
104 | struct list_head plug_list; | |
105 | ||
106 | /* | |
107 | * flags that tell us if it is safe to | |
108 | * merge with this bio | |
109 | */ | |
110 | unsigned long flags; | |
111 | ||
112 | /* size of each individual stripe on disk */ | |
113 | int stripe_len; | |
114 | ||
115 | /* number of data stripes (no p/q) */ | |
116 | int nr_data; | |
117 | ||
118 | int real_stripes; | |
119 | ||
120 | int stripe_npages; | |
121 | /* | |
122 | * set if we're doing a parity rebuild | |
123 | * for a read from higher up, which is handled | |
124 | * differently from a parity rebuild as part of | |
125 | * rmw | |
126 | */ | |
127 | enum btrfs_rbio_ops operation; | |
128 | ||
129 | /* first bad stripe */ | |
130 | int faila; | |
131 | ||
132 | /* second bad stripe (for raid6 use) */ | |
133 | int failb; | |
134 | ||
135 | int scrubp; | |
136 | /* | |
137 | * number of pages needed to represent the full | |
138 | * stripe | |
139 | */ | |
140 | int nr_pages; | |
141 | ||
142 | /* | |
143 | * size of all the bios in the bio_list. This | |
144 | * helps us decide if the rbio maps to a full | |
145 | * stripe or not | |
146 | */ | |
147 | int bio_list_bytes; | |
148 | ||
149 | int generic_bio_cnt; | |
150 | ||
151 | atomic_t refs; | |
152 | ||
153 | atomic_t stripes_pending; | |
154 | ||
155 | atomic_t error; | |
156 | /* | |
157 | * these are two arrays of pointers. We allocate the | |
158 | * rbio big enough to hold them both and setup their | |
159 | * locations when the rbio is allocated | |
160 | */ | |
161 | ||
162 | /* pointers to pages that we allocated for | |
163 | * reading/writing stripes directly from the disk (including P/Q) | |
164 | */ | |
165 | struct page **stripe_pages; | |
166 | ||
167 | /* | |
168 | * pointers to the pages in the bio_list. Stored | |
169 | * here for faster lookup | |
170 | */ | |
171 | struct page **bio_pages; | |
172 | ||
173 | /* | |
174 | * bitmap to record which horizontal stripe has data | |
175 | */ | |
176 | unsigned long *dbitmap; | |
177 | }; | |
178 | ||
179 | static int __raid56_parity_recover(struct btrfs_raid_bio *rbio); | |
180 | static noinline void finish_rmw(struct btrfs_raid_bio *rbio); | |
181 | static void rmw_work(struct btrfs_work *work); | |
182 | static void read_rebuild_work(struct btrfs_work *work); | |
183 | static void async_rmw_stripe(struct btrfs_raid_bio *rbio); | |
184 | static void async_read_rebuild(struct btrfs_raid_bio *rbio); | |
185 | static int fail_bio_stripe(struct btrfs_raid_bio *rbio, struct bio *bio); | |
186 | static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed); | |
187 | static void __free_raid_bio(struct btrfs_raid_bio *rbio); | |
188 | static void index_rbio_pages(struct btrfs_raid_bio *rbio); | |
189 | static int alloc_rbio_pages(struct btrfs_raid_bio *rbio); | |
190 | ||
191 | static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio, | |
192 | int need_check); | |
193 | static void async_scrub_parity(struct btrfs_raid_bio *rbio); | |
194 | ||
195 | /* | |
196 | * the stripe hash table is used for locking, and to collect | |
197 | * bios in hopes of making a full stripe | |
198 | */ | |
199 | int btrfs_alloc_stripe_hash_table(struct btrfs_fs_info *info) | |
200 | { | |
201 | struct btrfs_stripe_hash_table *table; | |
202 | struct btrfs_stripe_hash_table *x; | |
203 | struct btrfs_stripe_hash *cur; | |
204 | struct btrfs_stripe_hash *h; | |
205 | int num_entries = 1 << BTRFS_STRIPE_HASH_TABLE_BITS; | |
206 | int i; | |
207 | int table_size; | |
208 | ||
209 | if (info->stripe_hash_table) | |
210 | return 0; | |
211 | ||
212 | /* | |
213 | * The table is large, starting with order 4 and can go as high as | |
214 | * order 7 in case lock debugging is turned on. | |
215 | * | |
216 | * Try harder to allocate and fallback to vmalloc to lower the chance | |
217 | * of a failing mount. | |
218 | */ | |
219 | table_size = sizeof(*table) + sizeof(*h) * num_entries; | |
220 | table = kzalloc(table_size, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT); | |
221 | if (!table) { | |
222 | table = vzalloc(table_size); | |
223 | if (!table) | |
224 | return -ENOMEM; | |
225 | } | |
226 | ||
227 | spin_lock_init(&table->cache_lock); | |
228 | INIT_LIST_HEAD(&table->stripe_cache); | |
229 | ||
230 | h = table->table; | |
231 | ||
232 | for (i = 0; i < num_entries; i++) { | |
233 | cur = h + i; | |
234 | INIT_LIST_HEAD(&cur->hash_list); | |
235 | spin_lock_init(&cur->lock); | |
236 | init_waitqueue_head(&cur->wait); | |
237 | } | |
238 | ||
239 | x = cmpxchg(&info->stripe_hash_table, NULL, table); | |
240 | if (x) | |
241 | kvfree(x); | |
242 | return 0; | |
243 | } | |
244 | ||
245 | /* | |
246 | * caching an rbio means to copy anything from the | |
247 | * bio_pages array into the stripe_pages array. We | |
248 | * use the page uptodate bit in the stripe cache array | |
249 | * to indicate if it has valid data | |
250 | * | |
251 | * once the caching is done, we set the cache ready | |
252 | * bit. | |
253 | */ | |
254 | static void cache_rbio_pages(struct btrfs_raid_bio *rbio) | |
255 | { | |
256 | int i; | |
257 | char *s; | |
258 | char *d; | |
259 | int ret; | |
260 | ||
261 | ret = alloc_rbio_pages(rbio); | |
262 | if (ret) | |
263 | return; | |
264 | ||
265 | for (i = 0; i < rbio->nr_pages; i++) { | |
266 | if (!rbio->bio_pages[i]) | |
267 | continue; | |
268 | ||
269 | s = kmap(rbio->bio_pages[i]); | |
270 | d = kmap(rbio->stripe_pages[i]); | |
271 | ||
272 | memcpy(d, s, PAGE_CACHE_SIZE); | |
273 | ||
274 | kunmap(rbio->bio_pages[i]); | |
275 | kunmap(rbio->stripe_pages[i]); | |
276 | SetPageUptodate(rbio->stripe_pages[i]); | |
277 | } | |
278 | set_bit(RBIO_CACHE_READY_BIT, &rbio->flags); | |
279 | } | |
280 | ||
281 | /* | |
282 | * we hash on the first logical address of the stripe | |
283 | */ | |
284 | static int rbio_bucket(struct btrfs_raid_bio *rbio) | |
285 | { | |
286 | u64 num = rbio->bbio->raid_map[0]; | |
287 | ||
288 | /* | |
289 | * we shift down quite a bit. We're using byte | |
290 | * addressing, and most of the lower bits are zeros. | |
291 | * This tends to upset hash_64, and it consistently | |
292 | * returns just one or two different values. | |
293 | * | |
294 | * shifting off the lower bits fixes things. | |
295 | */ | |
296 | return hash_64(num >> 16, BTRFS_STRIPE_HASH_TABLE_BITS); | |
297 | } | |
298 | ||
299 | /* | |
300 | * stealing an rbio means taking all the uptodate pages from the stripe | |
301 | * array in the source rbio and putting them into the destination rbio | |
302 | */ | |
303 | static void steal_rbio(struct btrfs_raid_bio *src, struct btrfs_raid_bio *dest) | |
304 | { | |
305 | int i; | |
306 | struct page *s; | |
307 | struct page *d; | |
308 | ||
309 | if (!test_bit(RBIO_CACHE_READY_BIT, &src->flags)) | |
310 | return; | |
311 | ||
312 | for (i = 0; i < dest->nr_pages; i++) { | |
313 | s = src->stripe_pages[i]; | |
314 | if (!s || !PageUptodate(s)) { | |
315 | continue; | |
316 | } | |
317 | ||
318 | d = dest->stripe_pages[i]; | |
319 | if (d) | |
320 | __free_page(d); | |
321 | ||
322 | dest->stripe_pages[i] = s; | |
323 | src->stripe_pages[i] = NULL; | |
324 | } | |
325 | } | |
326 | ||
327 | /* | |
328 | * merging means we take the bio_list from the victim and | |
329 | * splice it into the destination. The victim should | |
330 | * be discarded afterwards. | |
331 | * | |
332 | * must be called with dest->rbio_list_lock held | |
333 | */ | |
334 | static void merge_rbio(struct btrfs_raid_bio *dest, | |
335 | struct btrfs_raid_bio *victim) | |
336 | { | |
337 | bio_list_merge(&dest->bio_list, &victim->bio_list); | |
338 | dest->bio_list_bytes += victim->bio_list_bytes; | |
339 | dest->generic_bio_cnt += victim->generic_bio_cnt; | |
340 | bio_list_init(&victim->bio_list); | |
341 | } | |
342 | ||
343 | /* | |
344 | * used to prune items that are in the cache. The caller | |
345 | * must hold the hash table lock. | |
346 | */ | |
347 | static void __remove_rbio_from_cache(struct btrfs_raid_bio *rbio) | |
348 | { | |
349 | int bucket = rbio_bucket(rbio); | |
350 | struct btrfs_stripe_hash_table *table; | |
351 | struct btrfs_stripe_hash *h; | |
352 | int freeit = 0; | |
353 | ||
354 | /* | |
355 | * check the bit again under the hash table lock. | |
356 | */ | |
357 | if (!test_bit(RBIO_CACHE_BIT, &rbio->flags)) | |
358 | return; | |
359 | ||
360 | table = rbio->fs_info->stripe_hash_table; | |
361 | h = table->table + bucket; | |
362 | ||
363 | /* hold the lock for the bucket because we may be | |
364 | * removing it from the hash table | |
365 | */ | |
366 | spin_lock(&h->lock); | |
367 | ||
368 | /* | |
369 | * hold the lock for the bio list because we need | |
370 | * to make sure the bio list is empty | |
371 | */ | |
372 | spin_lock(&rbio->bio_list_lock); | |
373 | ||
374 | if (test_and_clear_bit(RBIO_CACHE_BIT, &rbio->flags)) { | |
375 | list_del_init(&rbio->stripe_cache); | |
376 | table->cache_size -= 1; | |
377 | freeit = 1; | |
378 | ||
379 | /* if the bio list isn't empty, this rbio is | |
380 | * still involved in an IO. We take it out | |
381 | * of the cache list, and drop the ref that | |
382 | * was held for the list. | |
383 | * | |
384 | * If the bio_list was empty, we also remove | |
385 | * the rbio from the hash_table, and drop | |
386 | * the corresponding ref | |
387 | */ | |
388 | if (bio_list_empty(&rbio->bio_list)) { | |
389 | if (!list_empty(&rbio->hash_list)) { | |
390 | list_del_init(&rbio->hash_list); | |
391 | atomic_dec(&rbio->refs); | |
392 | BUG_ON(!list_empty(&rbio->plug_list)); | |
393 | } | |
394 | } | |
395 | } | |
396 | ||
397 | spin_unlock(&rbio->bio_list_lock); | |
398 | spin_unlock(&h->lock); | |
399 | ||
400 | if (freeit) | |
401 | __free_raid_bio(rbio); | |
402 | } | |
403 | ||
404 | /* | |
405 | * prune a given rbio from the cache | |
406 | */ | |
407 | static void remove_rbio_from_cache(struct btrfs_raid_bio *rbio) | |
408 | { | |
409 | struct btrfs_stripe_hash_table *table; | |
410 | unsigned long flags; | |
411 | ||
412 | if (!test_bit(RBIO_CACHE_BIT, &rbio->flags)) | |
413 | return; | |
414 | ||
415 | table = rbio->fs_info->stripe_hash_table; | |
416 | ||
417 | spin_lock_irqsave(&table->cache_lock, flags); | |
418 | __remove_rbio_from_cache(rbio); | |
419 | spin_unlock_irqrestore(&table->cache_lock, flags); | |
420 | } | |
421 | ||
422 | /* | |
423 | * remove everything in the cache | |
424 | */ | |
425 | static void btrfs_clear_rbio_cache(struct btrfs_fs_info *info) | |
426 | { | |
427 | struct btrfs_stripe_hash_table *table; | |
428 | unsigned long flags; | |
429 | struct btrfs_raid_bio *rbio; | |
430 | ||
431 | table = info->stripe_hash_table; | |
432 | ||
433 | spin_lock_irqsave(&table->cache_lock, flags); | |
434 | while (!list_empty(&table->stripe_cache)) { | |
435 | rbio = list_entry(table->stripe_cache.next, | |
436 | struct btrfs_raid_bio, | |
437 | stripe_cache); | |
438 | __remove_rbio_from_cache(rbio); | |
439 | } | |
440 | spin_unlock_irqrestore(&table->cache_lock, flags); | |
441 | } | |
442 | ||
443 | /* | |
444 | * remove all cached entries and free the hash table | |
445 | * used by unmount | |
446 | */ | |
447 | void btrfs_free_stripe_hash_table(struct btrfs_fs_info *info) | |
448 | { | |
449 | if (!info->stripe_hash_table) | |
450 | return; | |
451 | btrfs_clear_rbio_cache(info); | |
452 | kvfree(info->stripe_hash_table); | |
453 | info->stripe_hash_table = NULL; | |
454 | } | |
455 | ||
456 | /* | |
457 | * insert an rbio into the stripe cache. It | |
458 | * must have already been prepared by calling | |
459 | * cache_rbio_pages | |
460 | * | |
461 | * If this rbio was already cached, it gets | |
462 | * moved to the front of the lru. | |
463 | * | |
464 | * If the size of the rbio cache is too big, we | |
465 | * prune an item. | |
466 | */ | |
467 | static void cache_rbio(struct btrfs_raid_bio *rbio) | |
468 | { | |
469 | struct btrfs_stripe_hash_table *table; | |
470 | unsigned long flags; | |
471 | ||
472 | if (!test_bit(RBIO_CACHE_READY_BIT, &rbio->flags)) | |
473 | return; | |
474 | ||
475 | table = rbio->fs_info->stripe_hash_table; | |
476 | ||
477 | spin_lock_irqsave(&table->cache_lock, flags); | |
478 | spin_lock(&rbio->bio_list_lock); | |
479 | ||
480 | /* bump our ref if we were not in the list before */ | |
481 | if (!test_and_set_bit(RBIO_CACHE_BIT, &rbio->flags)) | |
482 | atomic_inc(&rbio->refs); | |
483 | ||
484 | if (!list_empty(&rbio->stripe_cache)){ | |
485 | list_move(&rbio->stripe_cache, &table->stripe_cache); | |
486 | } else { | |
487 | list_add(&rbio->stripe_cache, &table->stripe_cache); | |
488 | table->cache_size += 1; | |
489 | } | |
490 | ||
491 | spin_unlock(&rbio->bio_list_lock); | |
492 | ||
493 | if (table->cache_size > RBIO_CACHE_SIZE) { | |
494 | struct btrfs_raid_bio *found; | |
495 | ||
496 | found = list_entry(table->stripe_cache.prev, | |
497 | struct btrfs_raid_bio, | |
498 | stripe_cache); | |
499 | ||
500 | if (found != rbio) | |
501 | __remove_rbio_from_cache(found); | |
502 | } | |
503 | ||
504 | spin_unlock_irqrestore(&table->cache_lock, flags); | |
505 | return; | |
506 | } | |
507 | ||
508 | /* | |
509 | * 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 | */ | |
513 | static 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 | */ | |
535 | static 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 | ||
547 | static 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 | */ | |
568 | static 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 | ||
575 | /* | |
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 | ||
586 | if (last->bbio->raid_map[0] != | |
587 | cur->bbio->raid_map[0]) | |
588 | return 0; | |
589 | ||
590 | /* we can't merge with different operations */ | |
591 | if (last->operation != cur->operation) | |
592 | return 0; | |
593 | /* | |
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; | |
604 | ||
605 | return 1; | |
606 | } | |
607 | ||
608 | /* | |
609 | * helper to index into the pstripe | |
610 | */ | |
611 | static struct page *rbio_pstripe_page(struct btrfs_raid_bio *rbio, int index) | |
612 | { | |
613 | index += (rbio->nr_data * rbio->stripe_len) >> PAGE_CACHE_SHIFT; | |
614 | return rbio->stripe_pages[index]; | |
615 | } | |
616 | ||
617 | /* | |
618 | * helper to index into the qstripe, returns null | |
619 | * if there is no qstripe | |
620 | */ | |
621 | static struct page *rbio_qstripe_page(struct btrfs_raid_bio *rbio, int index) | |
622 | { | |
623 | if (rbio->nr_data + 1 == rbio->real_stripes) | |
624 | return NULL; | |
625 | ||
626 | index += ((rbio->nr_data + 1) * rbio->stripe_len) >> | |
627 | PAGE_CACHE_SHIFT; | |
628 | return rbio->stripe_pages[index]; | |
629 | } | |
630 | ||
631 | /* | |
632 | * The first stripe in the table for a logical address | |
633 | * has the lock. rbios are added in one of three ways: | |
634 | * | |
635 | * 1) Nobody has the stripe locked yet. The rbio is given | |
636 | * the lock and 0 is returned. The caller must start the IO | |
637 | * themselves. | |
638 | * | |
639 | * 2) Someone has the stripe locked, but we're able to merge | |
640 | * with the lock owner. The rbio is freed and the IO will | |
641 | * start automatically along with the existing rbio. 1 is returned. | |
642 | * | |
643 | * 3) Someone has the stripe locked, but we're not able to merge. | |
644 | * The rbio is added to the lock owner's plug list, or merged into | |
645 | * an rbio already on the plug list. When the lock owner unlocks, | |
646 | * the next rbio on the list is run and the IO is started automatically. | |
647 | * 1 is returned | |
648 | * | |
649 | * If we return 0, the caller still owns the rbio and must continue with | |
650 | * IO submission. If we return 1, the caller must assume the rbio has | |
651 | * already been freed. | |
652 | */ | |
653 | static noinline int lock_stripe_add(struct btrfs_raid_bio *rbio) | |
654 | { | |
655 | int bucket = rbio_bucket(rbio); | |
656 | struct btrfs_stripe_hash *h = rbio->fs_info->stripe_hash_table->table + bucket; | |
657 | struct btrfs_raid_bio *cur; | |
658 | struct btrfs_raid_bio *pending; | |
659 | unsigned long flags; | |
660 | DEFINE_WAIT(wait); | |
661 | struct btrfs_raid_bio *freeit = NULL; | |
662 | struct btrfs_raid_bio *cache_drop = NULL; | |
663 | int ret = 0; | |
664 | int walk = 0; | |
665 | ||
666 | spin_lock_irqsave(&h->lock, flags); | |
667 | list_for_each_entry(cur, &h->hash_list, hash_list) { | |
668 | walk++; | |
669 | if (cur->bbio->raid_map[0] == rbio->bbio->raid_map[0]) { | |
670 | spin_lock(&cur->bio_list_lock); | |
671 | ||
672 | /* can we steal this cached rbio's pages? */ | |
673 | if (bio_list_empty(&cur->bio_list) && | |
674 | list_empty(&cur->plug_list) && | |
675 | test_bit(RBIO_CACHE_BIT, &cur->flags) && | |
676 | !test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) { | |
677 | list_del_init(&cur->hash_list); | |
678 | atomic_dec(&cur->refs); | |
679 | ||
680 | steal_rbio(cur, rbio); | |
681 | cache_drop = cur; | |
682 | spin_unlock(&cur->bio_list_lock); | |
683 | ||
684 | goto lockit; | |
685 | } | |
686 | ||
687 | /* can we merge into the lock owner? */ | |
688 | if (rbio_can_merge(cur, rbio)) { | |
689 | merge_rbio(cur, rbio); | |
690 | spin_unlock(&cur->bio_list_lock); | |
691 | freeit = rbio; | |
692 | ret = 1; | |
693 | goto out; | |
694 | } | |
695 | ||
696 | ||
697 | /* | |
698 | * we couldn't merge with the running | |
699 | * rbio, see if we can merge with the | |
700 | * pending ones. We don't have to | |
701 | * check for rmw_locked because there | |
702 | * is no way they are inside finish_rmw | |
703 | * right now | |
704 | */ | |
705 | list_for_each_entry(pending, &cur->plug_list, | |
706 | plug_list) { | |
707 | if (rbio_can_merge(pending, rbio)) { | |
708 | merge_rbio(pending, rbio); | |
709 | spin_unlock(&cur->bio_list_lock); | |
710 | freeit = rbio; | |
711 | ret = 1; | |
712 | goto out; | |
713 | } | |
714 | } | |
715 | ||
716 | /* no merging, put us on the tail of the plug list, | |
717 | * our rbio will be started with the currently | |
718 | * running rbio unlocks | |
719 | */ | |
720 | list_add_tail(&rbio->plug_list, &cur->plug_list); | |
721 | spin_unlock(&cur->bio_list_lock); | |
722 | ret = 1; | |
723 | goto out; | |
724 | } | |
725 | } | |
726 | lockit: | |
727 | atomic_inc(&rbio->refs); | |
728 | list_add(&rbio->hash_list, &h->hash_list); | |
729 | out: | |
730 | spin_unlock_irqrestore(&h->lock, flags); | |
731 | if (cache_drop) | |
732 | remove_rbio_from_cache(cache_drop); | |
733 | if (freeit) | |
734 | __free_raid_bio(freeit); | |
735 | return ret; | |
736 | } | |
737 | ||
738 | /* | |
739 | * called as rmw or parity rebuild is completed. If the plug list has more | |
740 | * rbios waiting for this stripe, the next one on the list will be started | |
741 | */ | |
742 | static noinline void unlock_stripe(struct btrfs_raid_bio *rbio) | |
743 | { | |
744 | int bucket; | |
745 | struct btrfs_stripe_hash *h; | |
746 | unsigned long flags; | |
747 | int keep_cache = 0; | |
748 | ||
749 | bucket = rbio_bucket(rbio); | |
750 | h = rbio->fs_info->stripe_hash_table->table + bucket; | |
751 | ||
752 | if (list_empty(&rbio->plug_list)) | |
753 | cache_rbio(rbio); | |
754 | ||
755 | spin_lock_irqsave(&h->lock, flags); | |
756 | spin_lock(&rbio->bio_list_lock); | |
757 | ||
758 | if (!list_empty(&rbio->hash_list)) { | |
759 | /* | |
760 | * if we're still cached and there is no other IO | |
761 | * to perform, just leave this rbio here for others | |
762 | * to steal from later | |
763 | */ | |
764 | if (list_empty(&rbio->plug_list) && | |
765 | test_bit(RBIO_CACHE_BIT, &rbio->flags)) { | |
766 | keep_cache = 1; | |
767 | clear_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags); | |
768 | BUG_ON(!bio_list_empty(&rbio->bio_list)); | |
769 | goto done; | |
770 | } | |
771 | ||
772 | list_del_init(&rbio->hash_list); | |
773 | atomic_dec(&rbio->refs); | |
774 | ||
775 | /* | |
776 | * we use the plug list to hold all the rbios | |
777 | * waiting for the chance to lock this stripe. | |
778 | * hand the lock over to one of them. | |
779 | */ | |
780 | if (!list_empty(&rbio->plug_list)) { | |
781 | struct btrfs_raid_bio *next; | |
782 | struct list_head *head = rbio->plug_list.next; | |
783 | ||
784 | next = list_entry(head, struct btrfs_raid_bio, | |
785 | plug_list); | |
786 | ||
787 | list_del_init(&rbio->plug_list); | |
788 | ||
789 | list_add(&next->hash_list, &h->hash_list); | |
790 | atomic_inc(&next->refs); | |
791 | spin_unlock(&rbio->bio_list_lock); | |
792 | spin_unlock_irqrestore(&h->lock, flags); | |
793 | ||
794 | if (next->operation == BTRFS_RBIO_READ_REBUILD) | |
795 | async_read_rebuild(next); | |
796 | else if (next->operation == BTRFS_RBIO_WRITE) { | |
797 | steal_rbio(rbio, next); | |
798 | async_rmw_stripe(next); | |
799 | } else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) { | |
800 | steal_rbio(rbio, next); | |
801 | async_scrub_parity(next); | |
802 | } | |
803 | ||
804 | goto done_nolock; | |
805 | } else if (waitqueue_active(&h->wait)) { | |
806 | spin_unlock(&rbio->bio_list_lock); | |
807 | spin_unlock_irqrestore(&h->lock, flags); | |
808 | wake_up(&h->wait); | |
809 | goto done_nolock; | |
810 | } | |
811 | } | |
812 | done: | |
813 | spin_unlock(&rbio->bio_list_lock); | |
814 | spin_unlock_irqrestore(&h->lock, flags); | |
815 | ||
816 | done_nolock: | |
817 | if (!keep_cache) | |
818 | remove_rbio_from_cache(rbio); | |
819 | } | |
820 | ||
821 | static void __free_raid_bio(struct btrfs_raid_bio *rbio) | |
822 | { | |
823 | int i; | |
824 | ||
825 | WARN_ON(atomic_read(&rbio->refs) < 0); | |
826 | if (!atomic_dec_and_test(&rbio->refs)) | |
827 | return; | |
828 | ||
829 | WARN_ON(!list_empty(&rbio->stripe_cache)); | |
830 | WARN_ON(!list_empty(&rbio->hash_list)); | |
831 | WARN_ON(!bio_list_empty(&rbio->bio_list)); | |
832 | ||
833 | for (i = 0; i < rbio->nr_pages; i++) { | |
834 | if (rbio->stripe_pages[i]) { | |
835 | __free_page(rbio->stripe_pages[i]); | |
836 | rbio->stripe_pages[i] = NULL; | |
837 | } | |
838 | } | |
839 | ||
840 | btrfs_put_bbio(rbio->bbio); | |
841 | kfree(rbio); | |
842 | } | |
843 | ||
844 | static void free_raid_bio(struct btrfs_raid_bio *rbio) | |
845 | { | |
846 | unlock_stripe(rbio); | |
847 | __free_raid_bio(rbio); | |
848 | } | |
849 | ||
850 | /* | |
851 | * this frees the rbio and runs through all the bios in the | |
852 | * bio_list and calls end_io on them | |
853 | */ | |
854 | static void rbio_orig_end_io(struct btrfs_raid_bio *rbio, int err) | |
855 | { | |
856 | struct bio *cur = bio_list_get(&rbio->bio_list); | |
857 | struct bio *next; | |
858 | ||
859 | if (rbio->generic_bio_cnt) | |
860 | btrfs_bio_counter_sub(rbio->fs_info, rbio->generic_bio_cnt); | |
861 | ||
862 | free_raid_bio(rbio); | |
863 | ||
864 | while (cur) { | |
865 | next = cur->bi_next; | |
866 | cur->bi_next = NULL; | |
867 | cur->bi_error = err; | |
868 | bio_endio(cur); | |
869 | cur = next; | |
870 | } | |
871 | } | |
872 | ||
873 | /* | |
874 | * end io function used by finish_rmw. When we finally | |
875 | * get here, we've written a full stripe | |
876 | */ | |
877 | static void raid_write_end_io(struct bio *bio) | |
878 | { | |
879 | struct btrfs_raid_bio *rbio = bio->bi_private; | |
880 | int err = bio->bi_error; | |
881 | ||
882 | if (err) | |
883 | fail_bio_stripe(rbio, bio); | |
884 | ||
885 | bio_put(bio); | |
886 | ||
887 | if (!atomic_dec_and_test(&rbio->stripes_pending)) | |
888 | return; | |
889 | ||
890 | err = 0; | |
891 | ||
892 | /* OK, we have read all the stripes we need to. */ | |
893 | if (atomic_read(&rbio->error) > rbio->bbio->max_errors) | |
894 | err = -EIO; | |
895 | ||
896 | rbio_orig_end_io(rbio, err); | |
897 | return; | |
898 | } | |
899 | ||
900 | /* | |
901 | * the read/modify/write code wants to use the original bio for | |
902 | * any pages it included, and then use the rbio for everything | |
903 | * else. This function decides if a given index (stripe number) | |
904 | * and page number in that stripe fall inside the original bio | |
905 | * or the rbio. | |
906 | * | |
907 | * if you set bio_list_only, you'll get a NULL back for any ranges | |
908 | * that are outside the bio_list | |
909 | * | |
910 | * This doesn't take any refs on anything, you get a bare page pointer | |
911 | * and the caller must bump refs as required. | |
912 | * | |
913 | * You must call index_rbio_pages once before you can trust | |
914 | * the answers from this function. | |
915 | */ | |
916 | static struct page *page_in_rbio(struct btrfs_raid_bio *rbio, | |
917 | int index, int pagenr, int bio_list_only) | |
918 | { | |
919 | int chunk_page; | |
920 | struct page *p = NULL; | |
921 | ||
922 | chunk_page = index * (rbio->stripe_len >> PAGE_SHIFT) + pagenr; | |
923 | ||
924 | spin_lock_irq(&rbio->bio_list_lock); | |
925 | p = rbio->bio_pages[chunk_page]; | |
926 | spin_unlock_irq(&rbio->bio_list_lock); | |
927 | ||
928 | if (p || bio_list_only) | |
929 | return p; | |
930 | ||
931 | return rbio->stripe_pages[chunk_page]; | |
932 | } | |
933 | ||
934 | /* | |
935 | * number of pages we need for the entire stripe across all the | |
936 | * drives | |
937 | */ | |
938 | static unsigned long rbio_nr_pages(unsigned long stripe_len, int nr_stripes) | |
939 | { | |
940 | unsigned long nr = stripe_len * nr_stripes; | |
941 | return DIV_ROUND_UP(nr, PAGE_CACHE_SIZE); | |
942 | } | |
943 | ||
944 | /* | |
945 | * allocation and initial setup for the btrfs_raid_bio. Not | |
946 | * this does not allocate any pages for rbio->pages. | |
947 | */ | |
948 | static struct btrfs_raid_bio *alloc_rbio(struct btrfs_root *root, | |
949 | struct btrfs_bio *bbio, u64 stripe_len) | |
950 | { | |
951 | struct btrfs_raid_bio *rbio; | |
952 | int nr_data = 0; | |
953 | int real_stripes = bbio->num_stripes - bbio->num_tgtdevs; | |
954 | int num_pages = rbio_nr_pages(stripe_len, real_stripes); | |
955 | int stripe_npages = DIV_ROUND_UP(stripe_len, PAGE_SIZE); | |
956 | void *p; | |
957 | ||
958 | rbio = kzalloc(sizeof(*rbio) + num_pages * sizeof(struct page *) * 2 + | |
959 | DIV_ROUND_UP(stripe_npages, BITS_PER_LONG / 8), | |
960 | GFP_NOFS); | |
961 | if (!rbio) | |
962 | return ERR_PTR(-ENOMEM); | |
963 | ||
964 | bio_list_init(&rbio->bio_list); | |
965 | INIT_LIST_HEAD(&rbio->plug_list); | |
966 | spin_lock_init(&rbio->bio_list_lock); | |
967 | INIT_LIST_HEAD(&rbio->stripe_cache); | |
968 | INIT_LIST_HEAD(&rbio->hash_list); | |
969 | rbio->bbio = bbio; | |
970 | rbio->fs_info = root->fs_info; | |
971 | rbio->stripe_len = stripe_len; | |
972 | rbio->nr_pages = num_pages; | |
973 | rbio->real_stripes = real_stripes; | |
974 | rbio->stripe_npages = stripe_npages; | |
975 | rbio->faila = -1; | |
976 | rbio->failb = -1; | |
977 | atomic_set(&rbio->refs, 1); | |
978 | atomic_set(&rbio->error, 0); | |
979 | atomic_set(&rbio->stripes_pending, 0); | |
980 | ||
981 | /* | |
982 | * the stripe_pages and bio_pages array point to the extra | |
983 | * memory we allocated past the end of the rbio | |
984 | */ | |
985 | p = rbio + 1; | |
986 | rbio->stripe_pages = p; | |
987 | rbio->bio_pages = p + sizeof(struct page *) * num_pages; | |
988 | rbio->dbitmap = p + sizeof(struct page *) * num_pages * 2; | |
989 | ||
990 | if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5) | |
991 | nr_data = real_stripes - 1; | |
992 | else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6) | |
993 | nr_data = real_stripes - 2; | |
994 | else | |
995 | BUG(); | |
996 | ||
997 | rbio->nr_data = nr_data; | |
998 | return rbio; | |
999 | } | |
1000 | ||
1001 | /* allocate pages for all the stripes in the bio, including parity */ | |
1002 | static int alloc_rbio_pages(struct btrfs_raid_bio *rbio) | |
1003 | { | |
1004 | int i; | |
1005 | struct page *page; | |
1006 | ||
1007 | for (i = 0; i < rbio->nr_pages; i++) { | |
1008 | if (rbio->stripe_pages[i]) | |
1009 | continue; | |
1010 | page = alloc_page(GFP_NOFS | __GFP_HIGHMEM); | |
1011 | if (!page) | |
1012 | return -ENOMEM; | |
1013 | rbio->stripe_pages[i] = page; | |
1014 | ClearPageUptodate(page); | |
1015 | } | |
1016 | return 0; | |
1017 | } | |
1018 | ||
1019 | /* allocate pages for just the p/q stripes */ | |
1020 | static int alloc_rbio_parity_pages(struct btrfs_raid_bio *rbio) | |
1021 | { | |
1022 | int i; | |
1023 | struct page *page; | |
1024 | ||
1025 | i = (rbio->nr_data * rbio->stripe_len) >> PAGE_CACHE_SHIFT; | |
1026 | ||
1027 | for (; i < rbio->nr_pages; i++) { | |
1028 | if (rbio->stripe_pages[i]) | |
1029 | continue; | |
1030 | page = alloc_page(GFP_NOFS | __GFP_HIGHMEM); | |
1031 | if (!page) | |
1032 | return -ENOMEM; | |
1033 | rbio->stripe_pages[i] = page; | |
1034 | } | |
1035 | return 0; | |
1036 | } | |
1037 | ||
1038 | /* | |
1039 | * add a single page from a specific stripe into our list of bios for IO | |
1040 | * this will try to merge into existing bios if possible, and returns | |
1041 | * zero if all went well. | |
1042 | */ | |
1043 | static int rbio_add_io_page(struct btrfs_raid_bio *rbio, | |
1044 | struct bio_list *bio_list, | |
1045 | struct page *page, | |
1046 | int stripe_nr, | |
1047 | unsigned long page_index, | |
1048 | unsigned long bio_max_len) | |
1049 | { | |
1050 | struct bio *last = bio_list->tail; | |
1051 | u64 last_end = 0; | |
1052 | int ret; | |
1053 | struct bio *bio; | |
1054 | struct btrfs_bio_stripe *stripe; | |
1055 | u64 disk_start; | |
1056 | ||
1057 | stripe = &rbio->bbio->stripes[stripe_nr]; | |
1058 | disk_start = stripe->physical + (page_index << PAGE_CACHE_SHIFT); | |
1059 | ||
1060 | /* if the device is missing, just fail this stripe */ | |
1061 | if (!stripe->dev->bdev) | |
1062 | return fail_rbio_index(rbio, stripe_nr); | |
1063 | ||
1064 | /* see if we can add this page onto our existing bio */ | |
1065 | if (last) { | |
1066 | last_end = (u64)last->bi_iter.bi_sector << 9; | |
1067 | last_end += last->bi_iter.bi_size; | |
1068 | ||
1069 | /* | |
1070 | * we can't merge these if they are from different | |
1071 | * devices or if they are not contiguous | |
1072 | */ | |
1073 | if (last_end == disk_start && stripe->dev->bdev && | |
1074 | !last->bi_error && | |
1075 | last->bi_bdev == stripe->dev->bdev) { | |
1076 | ret = bio_add_page(last, page, PAGE_CACHE_SIZE, 0); | |
1077 | if (ret == PAGE_CACHE_SIZE) | |
1078 | return 0; | |
1079 | } | |
1080 | } | |
1081 | ||
1082 | /* put a new bio on the list */ | |
1083 | bio = btrfs_io_bio_alloc(GFP_NOFS, bio_max_len >> PAGE_SHIFT?:1); | |
1084 | if (!bio) | |
1085 | return -ENOMEM; | |
1086 | ||
1087 | bio->bi_iter.bi_size = 0; | |
1088 | bio->bi_bdev = stripe->dev->bdev; | |
1089 | bio->bi_iter.bi_sector = disk_start >> 9; | |
1090 | ||
1091 | bio_add_page(bio, page, PAGE_CACHE_SIZE, 0); | |
1092 | bio_list_add(bio_list, bio); | |
1093 | return 0; | |
1094 | } | |
1095 | ||
1096 | /* | |
1097 | * while we're doing the read/modify/write cycle, we could | |
1098 | * have errors in reading pages off the disk. This checks | |
1099 | * for errors and if we're not able to read the page it'll | |
1100 | * trigger parity reconstruction. The rmw will be finished | |
1101 | * after we've reconstructed the failed stripes | |
1102 | */ | |
1103 | static void validate_rbio_for_rmw(struct btrfs_raid_bio *rbio) | |
1104 | { | |
1105 | if (rbio->faila >= 0 || rbio->failb >= 0) { | |
1106 | BUG_ON(rbio->faila == rbio->real_stripes - 1); | |
1107 | __raid56_parity_recover(rbio); | |
1108 | } else { | |
1109 | finish_rmw(rbio); | |
1110 | } | |
1111 | } | |
1112 | ||
1113 | /* | |
1114 | * these are just the pages from the rbio array, not from anything | |
1115 | * the FS sent down to us | |
1116 | */ | |
1117 | static struct page *rbio_stripe_page(struct btrfs_raid_bio *rbio, int stripe, int page) | |
1118 | { | |
1119 | int index; | |
1120 | index = stripe * (rbio->stripe_len >> PAGE_CACHE_SHIFT); | |
1121 | index += page; | |
1122 | return rbio->stripe_pages[index]; | |
1123 | } | |
1124 | ||
1125 | /* | |
1126 | * helper function to walk our bio list and populate the bio_pages array with | |
1127 | * the result. This seems expensive, but it is faster than constantly | |
1128 | * searching through the bio list as we setup the IO in finish_rmw or stripe | |
1129 | * reconstruction. | |
1130 | * | |
1131 | * This must be called before you trust the answers from page_in_rbio | |
1132 | */ | |
1133 | static void index_rbio_pages(struct btrfs_raid_bio *rbio) | |
1134 | { | |
1135 | struct bio *bio; | |
1136 | u64 start; | |
1137 | unsigned long stripe_offset; | |
1138 | unsigned long page_index; | |
1139 | struct page *p; | |
1140 | int i; | |
1141 | ||
1142 | spin_lock_irq(&rbio->bio_list_lock); | |
1143 | bio_list_for_each(bio, &rbio->bio_list) { | |
1144 | start = (u64)bio->bi_iter.bi_sector << 9; | |
1145 | stripe_offset = start - rbio->bbio->raid_map[0]; | |
1146 | page_index = stripe_offset >> PAGE_CACHE_SHIFT; | |
1147 | ||
1148 | for (i = 0; i < bio->bi_vcnt; i++) { | |
1149 | p = bio->bi_io_vec[i].bv_page; | |
1150 | rbio->bio_pages[page_index + i] = p; | |
1151 | } | |
1152 | } | |
1153 | spin_unlock_irq(&rbio->bio_list_lock); | |
1154 | } | |
1155 | ||
1156 | /* | |
1157 | * this is called from one of two situations. We either | |
1158 | * have a full stripe from the higher layers, or we've read all | |
1159 | * the missing bits off disk. | |
1160 | * | |
1161 | * This will calculate the parity and then send down any | |
1162 | * changed blocks. | |
1163 | */ | |
1164 | static noinline void finish_rmw(struct btrfs_raid_bio *rbio) | |
1165 | { | |
1166 | struct btrfs_bio *bbio = rbio->bbio; | |
1167 | void *pointers[rbio->real_stripes]; | |
1168 | int stripe_len = rbio->stripe_len; | |
1169 | int nr_data = rbio->nr_data; | |
1170 | int stripe; | |
1171 | int pagenr; | |
1172 | int p_stripe = -1; | |
1173 | int q_stripe = -1; | |
1174 | struct bio_list bio_list; | |
1175 | struct bio *bio; | |
1176 | int pages_per_stripe = stripe_len >> PAGE_CACHE_SHIFT; | |
1177 | int ret; | |
1178 | ||
1179 | bio_list_init(&bio_list); | |
1180 | ||
1181 | if (rbio->real_stripes - rbio->nr_data == 1) { | |
1182 | p_stripe = rbio->real_stripes - 1; | |
1183 | } else if (rbio->real_stripes - rbio->nr_data == 2) { | |
1184 | p_stripe = rbio->real_stripes - 2; | |
1185 | q_stripe = rbio->real_stripes - 1; | |
1186 | } else { | |
1187 | BUG(); | |
1188 | } | |
1189 | ||
1190 | /* at this point we either have a full stripe, | |
1191 | * or we've read the full stripe from the drive. | |
1192 | * recalculate the parity and write the new results. | |
1193 | * | |
1194 | * We're not allowed to add any new bios to the | |
1195 | * bio list here, anyone else that wants to | |
1196 | * change this stripe needs to do their own rmw. | |
1197 | */ | |
1198 | spin_lock_irq(&rbio->bio_list_lock); | |
1199 | set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags); | |
1200 | spin_unlock_irq(&rbio->bio_list_lock); | |
1201 | ||
1202 | atomic_set(&rbio->error, 0); | |
1203 | ||
1204 | /* | |
1205 | * now that we've set rmw_locked, run through the | |
1206 | * bio list one last time and map the page pointers | |
1207 | * | |
1208 | * We don't cache full rbios because we're assuming | |
1209 | * the higher layers are unlikely to use this area of | |
1210 | * the disk again soon. If they do use it again, | |
1211 | * hopefully they will send another full bio. | |
1212 | */ | |
1213 | index_rbio_pages(rbio); | |
1214 | if (!rbio_is_full(rbio)) | |
1215 | cache_rbio_pages(rbio); | |
1216 | else | |
1217 | clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags); | |
1218 | ||
1219 | for (pagenr = 0; pagenr < pages_per_stripe; pagenr++) { | |
1220 | struct page *p; | |
1221 | /* first collect one page from each data stripe */ | |
1222 | for (stripe = 0; stripe < nr_data; stripe++) { | |
1223 | p = page_in_rbio(rbio, stripe, pagenr, 0); | |
1224 | pointers[stripe] = kmap(p); | |
1225 | } | |
1226 | ||
1227 | /* then add the parity stripe */ | |
1228 | p = rbio_pstripe_page(rbio, pagenr); | |
1229 | SetPageUptodate(p); | |
1230 | pointers[stripe++] = kmap(p); | |
1231 | ||
1232 | if (q_stripe != -1) { | |
1233 | ||
1234 | /* | |
1235 | * raid6, add the qstripe and call the | |
1236 | * library function to fill in our p/q | |
1237 | */ | |
1238 | p = rbio_qstripe_page(rbio, pagenr); | |
1239 | SetPageUptodate(p); | |
1240 | pointers[stripe++] = kmap(p); | |
1241 | ||
1242 | raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE, | |
1243 | pointers); | |
1244 | } else { | |
1245 | /* raid5 */ | |
1246 | memcpy(pointers[nr_data], pointers[0], PAGE_SIZE); | |
1247 | run_xor(pointers + 1, nr_data - 1, PAGE_CACHE_SIZE); | |
1248 | } | |
1249 | ||
1250 | ||
1251 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) | |
1252 | kunmap(page_in_rbio(rbio, stripe, pagenr, 0)); | |
1253 | } | |
1254 | ||
1255 | /* | |
1256 | * time to start writing. Make bios for everything from the | |
1257 | * higher layers (the bio_list in our rbio) and our p/q. Ignore | |
1258 | * everything else. | |
1259 | */ | |
1260 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) { | |
1261 | for (pagenr = 0; pagenr < pages_per_stripe; pagenr++) { | |
1262 | struct page *page; | |
1263 | if (stripe < rbio->nr_data) { | |
1264 | page = page_in_rbio(rbio, stripe, pagenr, 1); | |
1265 | if (!page) | |
1266 | continue; | |
1267 | } else { | |
1268 | page = rbio_stripe_page(rbio, stripe, pagenr); | |
1269 | } | |
1270 | ||
1271 | ret = rbio_add_io_page(rbio, &bio_list, | |
1272 | page, stripe, pagenr, rbio->stripe_len); | |
1273 | if (ret) | |
1274 | goto cleanup; | |
1275 | } | |
1276 | } | |
1277 | ||
1278 | if (likely(!bbio->num_tgtdevs)) | |
1279 | goto write_data; | |
1280 | ||
1281 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) { | |
1282 | if (!bbio->tgtdev_map[stripe]) | |
1283 | continue; | |
1284 | ||
1285 | for (pagenr = 0; pagenr < pages_per_stripe; pagenr++) { | |
1286 | struct page *page; | |
1287 | if (stripe < rbio->nr_data) { | |
1288 | page = page_in_rbio(rbio, stripe, pagenr, 1); | |
1289 | if (!page) | |
1290 | continue; | |
1291 | } else { | |
1292 | page = rbio_stripe_page(rbio, stripe, pagenr); | |
1293 | } | |
1294 | ||
1295 | ret = rbio_add_io_page(rbio, &bio_list, page, | |
1296 | rbio->bbio->tgtdev_map[stripe], | |
1297 | pagenr, rbio->stripe_len); | |
1298 | if (ret) | |
1299 | goto cleanup; | |
1300 | } | |
1301 | } | |
1302 | ||
1303 | write_data: | |
1304 | atomic_set(&rbio->stripes_pending, bio_list_size(&bio_list)); | |
1305 | BUG_ON(atomic_read(&rbio->stripes_pending) == 0); | |
1306 | ||
1307 | while (1) { | |
1308 | bio = bio_list_pop(&bio_list); | |
1309 | if (!bio) | |
1310 | break; | |
1311 | ||
1312 | bio->bi_private = rbio; | |
1313 | bio->bi_end_io = raid_write_end_io; | |
1314 | submit_bio(WRITE, bio); | |
1315 | } | |
1316 | return; | |
1317 | ||
1318 | cleanup: | |
1319 | rbio_orig_end_io(rbio, -EIO); | |
1320 | } | |
1321 | ||
1322 | /* | |
1323 | * helper to find the stripe number for a given bio. Used to figure out which | |
1324 | * stripe has failed. This expects the bio to correspond to a physical disk, | |
1325 | * so it looks up based on physical sector numbers. | |
1326 | */ | |
1327 | static int find_bio_stripe(struct btrfs_raid_bio *rbio, | |
1328 | struct bio *bio) | |
1329 | { | |
1330 | u64 physical = bio->bi_iter.bi_sector; | |
1331 | u64 stripe_start; | |
1332 | int i; | |
1333 | struct btrfs_bio_stripe *stripe; | |
1334 | ||
1335 | physical <<= 9; | |
1336 | ||
1337 | for (i = 0; i < rbio->bbio->num_stripes; i++) { | |
1338 | stripe = &rbio->bbio->stripes[i]; | |
1339 | stripe_start = stripe->physical; | |
1340 | if (physical >= stripe_start && | |
1341 | physical < stripe_start + rbio->stripe_len && | |
1342 | bio->bi_bdev == stripe->dev->bdev) { | |
1343 | return i; | |
1344 | } | |
1345 | } | |
1346 | return -1; | |
1347 | } | |
1348 | ||
1349 | /* | |
1350 | * helper to find the stripe number for a given | |
1351 | * bio (before mapping). Used to figure out which stripe has | |
1352 | * failed. This looks up based on logical block numbers. | |
1353 | */ | |
1354 | static int find_logical_bio_stripe(struct btrfs_raid_bio *rbio, | |
1355 | struct bio *bio) | |
1356 | { | |
1357 | u64 logical = bio->bi_iter.bi_sector; | |
1358 | u64 stripe_start; | |
1359 | int i; | |
1360 | ||
1361 | logical <<= 9; | |
1362 | ||
1363 | for (i = 0; i < rbio->nr_data; i++) { | |
1364 | stripe_start = rbio->bbio->raid_map[i]; | |
1365 | if (logical >= stripe_start && | |
1366 | logical < stripe_start + rbio->stripe_len) { | |
1367 | return i; | |
1368 | } | |
1369 | } | |
1370 | return -1; | |
1371 | } | |
1372 | ||
1373 | /* | |
1374 | * returns -EIO if we had too many failures | |
1375 | */ | |
1376 | static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed) | |
1377 | { | |
1378 | unsigned long flags; | |
1379 | int ret = 0; | |
1380 | ||
1381 | spin_lock_irqsave(&rbio->bio_list_lock, flags); | |
1382 | ||
1383 | /* we already know this stripe is bad, move on */ | |
1384 | if (rbio->faila == failed || rbio->failb == failed) | |
1385 | goto out; | |
1386 | ||
1387 | if (rbio->faila == -1) { | |
1388 | /* first failure on this rbio */ | |
1389 | rbio->faila = failed; | |
1390 | atomic_inc(&rbio->error); | |
1391 | } else if (rbio->failb == -1) { | |
1392 | /* second failure on this rbio */ | |
1393 | rbio->failb = failed; | |
1394 | atomic_inc(&rbio->error); | |
1395 | } else { | |
1396 | ret = -EIO; | |
1397 | } | |
1398 | out: | |
1399 | spin_unlock_irqrestore(&rbio->bio_list_lock, flags); | |
1400 | ||
1401 | return ret; | |
1402 | } | |
1403 | ||
1404 | /* | |
1405 | * helper to fail a stripe based on a physical disk | |
1406 | * bio. | |
1407 | */ | |
1408 | static int fail_bio_stripe(struct btrfs_raid_bio *rbio, | |
1409 | struct bio *bio) | |
1410 | { | |
1411 | int failed = find_bio_stripe(rbio, bio); | |
1412 | ||
1413 | if (failed < 0) | |
1414 | return -EIO; | |
1415 | ||
1416 | return fail_rbio_index(rbio, failed); | |
1417 | } | |
1418 | ||
1419 | /* | |
1420 | * this sets each page in the bio uptodate. It should only be used on private | |
1421 | * rbio pages, nothing that comes in from the higher layers | |
1422 | */ | |
1423 | static void set_bio_pages_uptodate(struct bio *bio) | |
1424 | { | |
1425 | int i; | |
1426 | struct page *p; | |
1427 | ||
1428 | for (i = 0; i < bio->bi_vcnt; i++) { | |
1429 | p = bio->bi_io_vec[i].bv_page; | |
1430 | SetPageUptodate(p); | |
1431 | } | |
1432 | } | |
1433 | ||
1434 | /* | |
1435 | * end io for the read phase of the rmw cycle. All the bios here are physical | |
1436 | * stripe bios we've read from the disk so we can recalculate the parity of the | |
1437 | * stripe. | |
1438 | * | |
1439 | * This will usually kick off finish_rmw once all the bios are read in, but it | |
1440 | * may trigger parity reconstruction if we had any errors along the way | |
1441 | */ | |
1442 | static void raid_rmw_end_io(struct bio *bio) | |
1443 | { | |
1444 | struct btrfs_raid_bio *rbio = bio->bi_private; | |
1445 | ||
1446 | if (bio->bi_error) | |
1447 | fail_bio_stripe(rbio, bio); | |
1448 | else | |
1449 | set_bio_pages_uptodate(bio); | |
1450 | ||
1451 | bio_put(bio); | |
1452 | ||
1453 | if (!atomic_dec_and_test(&rbio->stripes_pending)) | |
1454 | return; | |
1455 | ||
1456 | if (atomic_read(&rbio->error) > rbio->bbio->max_errors) | |
1457 | goto cleanup; | |
1458 | ||
1459 | /* | |
1460 | * this will normally call finish_rmw to start our write | |
1461 | * but if there are any failed stripes we'll reconstruct | |
1462 | * from parity first | |
1463 | */ | |
1464 | validate_rbio_for_rmw(rbio); | |
1465 | return; | |
1466 | ||
1467 | cleanup: | |
1468 | ||
1469 | rbio_orig_end_io(rbio, -EIO); | |
1470 | } | |
1471 | ||
1472 | static void async_rmw_stripe(struct btrfs_raid_bio *rbio) | |
1473 | { | |
1474 | btrfs_init_work(&rbio->work, btrfs_rmw_helper, | |
1475 | rmw_work, NULL, NULL); | |
1476 | ||
1477 | btrfs_queue_work(rbio->fs_info->rmw_workers, | |
1478 | &rbio->work); | |
1479 | } | |
1480 | ||
1481 | static void async_read_rebuild(struct btrfs_raid_bio *rbio) | |
1482 | { | |
1483 | btrfs_init_work(&rbio->work, btrfs_rmw_helper, | |
1484 | read_rebuild_work, NULL, NULL); | |
1485 | ||
1486 | btrfs_queue_work(rbio->fs_info->rmw_workers, | |
1487 | &rbio->work); | |
1488 | } | |
1489 | ||
1490 | /* | |
1491 | * the stripe must be locked by the caller. It will | |
1492 | * unlock after all the writes are done | |
1493 | */ | |
1494 | static int raid56_rmw_stripe(struct btrfs_raid_bio *rbio) | |
1495 | { | |
1496 | int bios_to_read = 0; | |
1497 | struct bio_list bio_list; | |
1498 | int ret; | |
1499 | int nr_pages = DIV_ROUND_UP(rbio->stripe_len, PAGE_CACHE_SIZE); | |
1500 | int pagenr; | |
1501 | int stripe; | |
1502 | struct bio *bio; | |
1503 | ||
1504 | bio_list_init(&bio_list); | |
1505 | ||
1506 | ret = alloc_rbio_pages(rbio); | |
1507 | if (ret) | |
1508 | goto cleanup; | |
1509 | ||
1510 | index_rbio_pages(rbio); | |
1511 | ||
1512 | atomic_set(&rbio->error, 0); | |
1513 | /* | |
1514 | * build a list of bios to read all the missing parts of this | |
1515 | * stripe | |
1516 | */ | |
1517 | for (stripe = 0; stripe < rbio->nr_data; stripe++) { | |
1518 | for (pagenr = 0; pagenr < nr_pages; pagenr++) { | |
1519 | struct page *page; | |
1520 | /* | |
1521 | * we want to find all the pages missing from | |
1522 | * the rbio and read them from the disk. If | |
1523 | * page_in_rbio finds a page in the bio list | |
1524 | * we don't need to read it off the stripe. | |
1525 | */ | |
1526 | page = page_in_rbio(rbio, stripe, pagenr, 1); | |
1527 | if (page) | |
1528 | continue; | |
1529 | ||
1530 | page = rbio_stripe_page(rbio, stripe, pagenr); | |
1531 | /* | |
1532 | * the bio cache may have handed us an uptodate | |
1533 | * page. If so, be happy and use it | |
1534 | */ | |
1535 | if (PageUptodate(page)) | |
1536 | continue; | |
1537 | ||
1538 | ret = rbio_add_io_page(rbio, &bio_list, page, | |
1539 | stripe, pagenr, rbio->stripe_len); | |
1540 | if (ret) | |
1541 | goto cleanup; | |
1542 | } | |
1543 | } | |
1544 | ||
1545 | bios_to_read = bio_list_size(&bio_list); | |
1546 | if (!bios_to_read) { | |
1547 | /* | |
1548 | * this can happen if others have merged with | |
1549 | * us, it means there is nothing left to read. | |
1550 | * But if there are missing devices it may not be | |
1551 | * safe to do the full stripe write yet. | |
1552 | */ | |
1553 | goto finish; | |
1554 | } | |
1555 | ||
1556 | /* | |
1557 | * the bbio may be freed once we submit the last bio. Make sure | |
1558 | * not to touch it after that | |
1559 | */ | |
1560 | atomic_set(&rbio->stripes_pending, bios_to_read); | |
1561 | while (1) { | |
1562 | bio = bio_list_pop(&bio_list); | |
1563 | if (!bio) | |
1564 | break; | |
1565 | ||
1566 | bio->bi_private = rbio; | |
1567 | bio->bi_end_io = raid_rmw_end_io; | |
1568 | ||
1569 | btrfs_bio_wq_end_io(rbio->fs_info, bio, | |
1570 | BTRFS_WQ_ENDIO_RAID56); | |
1571 | ||
1572 | submit_bio(READ, bio); | |
1573 | } | |
1574 | /* the actual write will happen once the reads are done */ | |
1575 | return 0; | |
1576 | ||
1577 | cleanup: | |
1578 | rbio_orig_end_io(rbio, -EIO); | |
1579 | return -EIO; | |
1580 | ||
1581 | finish: | |
1582 | validate_rbio_for_rmw(rbio); | |
1583 | return 0; | |
1584 | } | |
1585 | ||
1586 | /* | |
1587 | * if the upper layers pass in a full stripe, we thank them by only allocating | |
1588 | * enough pages to hold the parity, and sending it all down quickly. | |
1589 | */ | |
1590 | static int full_stripe_write(struct btrfs_raid_bio *rbio) | |
1591 | { | |
1592 | int ret; | |
1593 | ||
1594 | ret = alloc_rbio_parity_pages(rbio); | |
1595 | if (ret) { | |
1596 | __free_raid_bio(rbio); | |
1597 | return ret; | |
1598 | } | |
1599 | ||
1600 | ret = lock_stripe_add(rbio); | |
1601 | if (ret == 0) | |
1602 | finish_rmw(rbio); | |
1603 | return 0; | |
1604 | } | |
1605 | ||
1606 | /* | |
1607 | * partial stripe writes get handed over to async helpers. | |
1608 | * We're really hoping to merge a few more writes into this | |
1609 | * rbio before calculating new parity | |
1610 | */ | |
1611 | static int partial_stripe_write(struct btrfs_raid_bio *rbio) | |
1612 | { | |
1613 | int ret; | |
1614 | ||
1615 | ret = lock_stripe_add(rbio); | |
1616 | if (ret == 0) | |
1617 | async_rmw_stripe(rbio); | |
1618 | return 0; | |
1619 | } | |
1620 | ||
1621 | /* | |
1622 | * sometimes while we were reading from the drive to | |
1623 | * recalculate parity, enough new bios come into create | |
1624 | * a full stripe. So we do a check here to see if we can | |
1625 | * go directly to finish_rmw | |
1626 | */ | |
1627 | static int __raid56_parity_write(struct btrfs_raid_bio *rbio) | |
1628 | { | |
1629 | /* head off into rmw land if we don't have a full stripe */ | |
1630 | if (!rbio_is_full(rbio)) | |
1631 | return partial_stripe_write(rbio); | |
1632 | return full_stripe_write(rbio); | |
1633 | } | |
1634 | ||
1635 | /* | |
1636 | * We use plugging call backs to collect full stripes. | |
1637 | * Any time we get a partial stripe write while plugged | |
1638 | * we collect it into a list. When the unplug comes down, | |
1639 | * we sort the list by logical block number and merge | |
1640 | * everything we can into the same rbios | |
1641 | */ | |
1642 | struct btrfs_plug_cb { | |
1643 | struct blk_plug_cb cb; | |
1644 | struct btrfs_fs_info *info; | |
1645 | struct list_head rbio_list; | |
1646 | struct btrfs_work work; | |
1647 | }; | |
1648 | ||
1649 | /* | |
1650 | * rbios on the plug list are sorted for easier merging. | |
1651 | */ | |
1652 | static int plug_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1653 | { | |
1654 | struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio, | |
1655 | plug_list); | |
1656 | struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio, | |
1657 | plug_list); | |
1658 | u64 a_sector = ra->bio_list.head->bi_iter.bi_sector; | |
1659 | u64 b_sector = rb->bio_list.head->bi_iter.bi_sector; | |
1660 | ||
1661 | if (a_sector < b_sector) | |
1662 | return -1; | |
1663 | if (a_sector > b_sector) | |
1664 | return 1; | |
1665 | return 0; | |
1666 | } | |
1667 | ||
1668 | static void run_plug(struct btrfs_plug_cb *plug) | |
1669 | { | |
1670 | struct btrfs_raid_bio *cur; | |
1671 | struct btrfs_raid_bio *last = NULL; | |
1672 | ||
1673 | /* | |
1674 | * sort our plug list then try to merge | |
1675 | * everything we can in hopes of creating full | |
1676 | * stripes. | |
1677 | */ | |
1678 | list_sort(NULL, &plug->rbio_list, plug_cmp); | |
1679 | while (!list_empty(&plug->rbio_list)) { | |
1680 | cur = list_entry(plug->rbio_list.next, | |
1681 | struct btrfs_raid_bio, plug_list); | |
1682 | list_del_init(&cur->plug_list); | |
1683 | ||
1684 | if (rbio_is_full(cur)) { | |
1685 | /* we have a full stripe, send it down */ | |
1686 | full_stripe_write(cur); | |
1687 | continue; | |
1688 | } | |
1689 | if (last) { | |
1690 | if (rbio_can_merge(last, cur)) { | |
1691 | merge_rbio(last, cur); | |
1692 | __free_raid_bio(cur); | |
1693 | continue; | |
1694 | ||
1695 | } | |
1696 | __raid56_parity_write(last); | |
1697 | } | |
1698 | last = cur; | |
1699 | } | |
1700 | if (last) { | |
1701 | __raid56_parity_write(last); | |
1702 | } | |
1703 | kfree(plug); | |
1704 | } | |
1705 | ||
1706 | /* | |
1707 | * if the unplug comes from schedule, we have to push the | |
1708 | * work off to a helper thread | |
1709 | */ | |
1710 | static void unplug_work(struct btrfs_work *work) | |
1711 | { | |
1712 | struct btrfs_plug_cb *plug; | |
1713 | plug = container_of(work, struct btrfs_plug_cb, work); | |
1714 | run_plug(plug); | |
1715 | } | |
1716 | ||
1717 | static void btrfs_raid_unplug(struct blk_plug_cb *cb, bool from_schedule) | |
1718 | { | |
1719 | struct btrfs_plug_cb *plug; | |
1720 | plug = container_of(cb, struct btrfs_plug_cb, cb); | |
1721 | ||
1722 | if (from_schedule) { | |
1723 | btrfs_init_work(&plug->work, btrfs_rmw_helper, | |
1724 | unplug_work, NULL, NULL); | |
1725 | btrfs_queue_work(plug->info->rmw_workers, | |
1726 | &plug->work); | |
1727 | return; | |
1728 | } | |
1729 | run_plug(plug); | |
1730 | } | |
1731 | ||
1732 | /* | |
1733 | * our main entry point for writes from the rest of the FS. | |
1734 | */ | |
1735 | int raid56_parity_write(struct btrfs_root *root, struct bio *bio, | |
1736 | struct btrfs_bio *bbio, u64 stripe_len) | |
1737 | { | |
1738 | struct btrfs_raid_bio *rbio; | |
1739 | struct btrfs_plug_cb *plug = NULL; | |
1740 | struct blk_plug_cb *cb; | |
1741 | int ret; | |
1742 | ||
1743 | rbio = alloc_rbio(root, bbio, stripe_len); | |
1744 | if (IS_ERR(rbio)) { | |
1745 | btrfs_put_bbio(bbio); | |
1746 | return PTR_ERR(rbio); | |
1747 | } | |
1748 | bio_list_add(&rbio->bio_list, bio); | |
1749 | rbio->bio_list_bytes = bio->bi_iter.bi_size; | |
1750 | rbio->operation = BTRFS_RBIO_WRITE; | |
1751 | ||
1752 | btrfs_bio_counter_inc_noblocked(root->fs_info); | |
1753 | rbio->generic_bio_cnt = 1; | |
1754 | ||
1755 | /* | |
1756 | * don't plug on full rbios, just get them out the door | |
1757 | * as quickly as we can | |
1758 | */ | |
1759 | if (rbio_is_full(rbio)) { | |
1760 | ret = full_stripe_write(rbio); | |
1761 | if (ret) | |
1762 | btrfs_bio_counter_dec(root->fs_info); | |
1763 | return ret; | |
1764 | } | |
1765 | ||
1766 | cb = blk_check_plugged(btrfs_raid_unplug, root->fs_info, | |
1767 | sizeof(*plug)); | |
1768 | if (cb) { | |
1769 | plug = container_of(cb, struct btrfs_plug_cb, cb); | |
1770 | if (!plug->info) { | |
1771 | plug->info = root->fs_info; | |
1772 | INIT_LIST_HEAD(&plug->rbio_list); | |
1773 | } | |
1774 | list_add_tail(&rbio->plug_list, &plug->rbio_list); | |
1775 | ret = 0; | |
1776 | } else { | |
1777 | ret = __raid56_parity_write(rbio); | |
1778 | if (ret) | |
1779 | btrfs_bio_counter_dec(root->fs_info); | |
1780 | } | |
1781 | return ret; | |
1782 | } | |
1783 | ||
1784 | /* | |
1785 | * all parity reconstruction happens here. We've read in everything | |
1786 | * we can find from the drives and this does the heavy lifting of | |
1787 | * sorting the good from the bad. | |
1788 | */ | |
1789 | static void __raid_recover_end_io(struct btrfs_raid_bio *rbio) | |
1790 | { | |
1791 | int pagenr, stripe; | |
1792 | void **pointers; | |
1793 | int faila = -1, failb = -1; | |
1794 | int nr_pages = DIV_ROUND_UP(rbio->stripe_len, PAGE_CACHE_SIZE); | |
1795 | struct page *page; | |
1796 | int err; | |
1797 | int i; | |
1798 | ||
1799 | pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS); | |
1800 | if (!pointers) { | |
1801 | err = -ENOMEM; | |
1802 | goto cleanup_io; | |
1803 | } | |
1804 | ||
1805 | faila = rbio->faila; | |
1806 | failb = rbio->failb; | |
1807 | ||
1808 | if (rbio->operation == BTRFS_RBIO_READ_REBUILD) { | |
1809 | spin_lock_irq(&rbio->bio_list_lock); | |
1810 | set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags); | |
1811 | spin_unlock_irq(&rbio->bio_list_lock); | |
1812 | } | |
1813 | ||
1814 | index_rbio_pages(rbio); | |
1815 | ||
1816 | for (pagenr = 0; pagenr < nr_pages; pagenr++) { | |
1817 | /* | |
1818 | * Now we just use bitmap to mark the horizontal stripes in | |
1819 | * which we have data when doing parity scrub. | |
1820 | */ | |
1821 | if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB && | |
1822 | !test_bit(pagenr, rbio->dbitmap)) | |
1823 | continue; | |
1824 | ||
1825 | /* setup our array of pointers with pages | |
1826 | * from each stripe | |
1827 | */ | |
1828 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) { | |
1829 | /* | |
1830 | * if we're rebuilding a read, we have to use | |
1831 | * pages from the bio list | |
1832 | */ | |
1833 | if (rbio->operation == BTRFS_RBIO_READ_REBUILD && | |
1834 | (stripe == faila || stripe == failb)) { | |
1835 | page = page_in_rbio(rbio, stripe, pagenr, 0); | |
1836 | } else { | |
1837 | page = rbio_stripe_page(rbio, stripe, pagenr); | |
1838 | } | |
1839 | pointers[stripe] = kmap(page); | |
1840 | } | |
1841 | ||
1842 | /* all raid6 handling here */ | |
1843 | if (rbio->bbio->map_type & BTRFS_BLOCK_GROUP_RAID6) { | |
1844 | /* | |
1845 | * single failure, rebuild from parity raid5 | |
1846 | * style | |
1847 | */ | |
1848 | if (failb < 0) { | |
1849 | if (faila == rbio->nr_data) { | |
1850 | /* | |
1851 | * Just the P stripe has failed, without | |
1852 | * a bad data or Q stripe. | |
1853 | * TODO, we should redo the xor here. | |
1854 | */ | |
1855 | err = -EIO; | |
1856 | goto cleanup; | |
1857 | } | |
1858 | /* | |
1859 | * a single failure in raid6 is rebuilt | |
1860 | * in the pstripe code below | |
1861 | */ | |
1862 | goto pstripe; | |
1863 | } | |
1864 | ||
1865 | /* make sure our ps and qs are in order */ | |
1866 | if (faila > failb) { | |
1867 | int tmp = failb; | |
1868 | failb = faila; | |
1869 | faila = tmp; | |
1870 | } | |
1871 | ||
1872 | /* if the q stripe is failed, do a pstripe reconstruction | |
1873 | * from the xors. | |
1874 | * If both the q stripe and the P stripe are failed, we're | |
1875 | * here due to a crc mismatch and we can't give them the | |
1876 | * data they want | |
1877 | */ | |
1878 | if (rbio->bbio->raid_map[failb] == RAID6_Q_STRIPE) { | |
1879 | if (rbio->bbio->raid_map[faila] == | |
1880 | RAID5_P_STRIPE) { | |
1881 | err = -EIO; | |
1882 | goto cleanup; | |
1883 | } | |
1884 | /* | |
1885 | * otherwise we have one bad data stripe and | |
1886 | * a good P stripe. raid5! | |
1887 | */ | |
1888 | goto pstripe; | |
1889 | } | |
1890 | ||
1891 | if (rbio->bbio->raid_map[failb] == RAID5_P_STRIPE) { | |
1892 | raid6_datap_recov(rbio->real_stripes, | |
1893 | PAGE_SIZE, faila, pointers); | |
1894 | } else { | |
1895 | raid6_2data_recov(rbio->real_stripes, | |
1896 | PAGE_SIZE, faila, failb, | |
1897 | pointers); | |
1898 | } | |
1899 | } else { | |
1900 | void *p; | |
1901 | ||
1902 | /* rebuild from P stripe here (raid5 or raid6) */ | |
1903 | BUG_ON(failb != -1); | |
1904 | pstripe: | |
1905 | /* Copy parity block into failed block to start with */ | |
1906 | memcpy(pointers[faila], | |
1907 | pointers[rbio->nr_data], | |
1908 | PAGE_CACHE_SIZE); | |
1909 | ||
1910 | /* rearrange the pointer array */ | |
1911 | p = pointers[faila]; | |
1912 | for (stripe = faila; stripe < rbio->nr_data - 1; stripe++) | |
1913 | pointers[stripe] = pointers[stripe + 1]; | |
1914 | pointers[rbio->nr_data - 1] = p; | |
1915 | ||
1916 | /* xor in the rest */ | |
1917 | run_xor(pointers, rbio->nr_data - 1, PAGE_CACHE_SIZE); | |
1918 | } | |
1919 | /* if we're doing this rebuild as part of an rmw, go through | |
1920 | * and set all of our private rbio pages in the | |
1921 | * failed stripes as uptodate. This way finish_rmw will | |
1922 | * know they can be trusted. If this was a read reconstruction, | |
1923 | * other endio functions will fiddle the uptodate bits | |
1924 | */ | |
1925 | if (rbio->operation == BTRFS_RBIO_WRITE) { | |
1926 | for (i = 0; i < nr_pages; i++) { | |
1927 | if (faila != -1) { | |
1928 | page = rbio_stripe_page(rbio, faila, i); | |
1929 | SetPageUptodate(page); | |
1930 | } | |
1931 | if (failb != -1) { | |
1932 | page = rbio_stripe_page(rbio, failb, i); | |
1933 | SetPageUptodate(page); | |
1934 | } | |
1935 | } | |
1936 | } | |
1937 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) { | |
1938 | /* | |
1939 | * if we're rebuilding a read, we have to use | |
1940 | * pages from the bio list | |
1941 | */ | |
1942 | if (rbio->operation == BTRFS_RBIO_READ_REBUILD && | |
1943 | (stripe == faila || stripe == failb)) { | |
1944 | page = page_in_rbio(rbio, stripe, pagenr, 0); | |
1945 | } else { | |
1946 | page = rbio_stripe_page(rbio, stripe, pagenr); | |
1947 | } | |
1948 | kunmap(page); | |
1949 | } | |
1950 | } | |
1951 | ||
1952 | err = 0; | |
1953 | cleanup: | |
1954 | kfree(pointers); | |
1955 | ||
1956 | cleanup_io: | |
1957 | if (rbio->operation == BTRFS_RBIO_READ_REBUILD) { | |
1958 | if (err == 0) | |
1959 | cache_rbio_pages(rbio); | |
1960 | else | |
1961 | clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags); | |
1962 | ||
1963 | rbio_orig_end_io(rbio, err); | |
1964 | } else if (err == 0) { | |
1965 | rbio->faila = -1; | |
1966 | rbio->failb = -1; | |
1967 | ||
1968 | if (rbio->operation == BTRFS_RBIO_WRITE) | |
1969 | finish_rmw(rbio); | |
1970 | else if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB) | |
1971 | finish_parity_scrub(rbio, 0); | |
1972 | else | |
1973 | BUG(); | |
1974 | } else { | |
1975 | rbio_orig_end_io(rbio, err); | |
1976 | } | |
1977 | } | |
1978 | ||
1979 | /* | |
1980 | * This is called only for stripes we've read from disk to | |
1981 | * reconstruct the parity. | |
1982 | */ | |
1983 | static void raid_recover_end_io(struct bio *bio) | |
1984 | { | |
1985 | struct btrfs_raid_bio *rbio = bio->bi_private; | |
1986 | ||
1987 | /* | |
1988 | * we only read stripe pages off the disk, set them | |
1989 | * up to date if there were no errors | |
1990 | */ | |
1991 | if (bio->bi_error) | |
1992 | fail_bio_stripe(rbio, bio); | |
1993 | else | |
1994 | set_bio_pages_uptodate(bio); | |
1995 | bio_put(bio); | |
1996 | ||
1997 | if (!atomic_dec_and_test(&rbio->stripes_pending)) | |
1998 | return; | |
1999 | ||
2000 | if (atomic_read(&rbio->error) > rbio->bbio->max_errors) | |
2001 | rbio_orig_end_io(rbio, -EIO); | |
2002 | else | |
2003 | __raid_recover_end_io(rbio); | |
2004 | } | |
2005 | ||
2006 | /* | |
2007 | * reads everything we need off the disk to reconstruct | |
2008 | * the parity. endio handlers trigger final reconstruction | |
2009 | * when the IO is done. | |
2010 | * | |
2011 | * This is used both for reads from the higher layers and for | |
2012 | * parity construction required to finish a rmw cycle. | |
2013 | */ | |
2014 | static int __raid56_parity_recover(struct btrfs_raid_bio *rbio) | |
2015 | { | |
2016 | int bios_to_read = 0; | |
2017 | struct bio_list bio_list; | |
2018 | int ret; | |
2019 | int nr_pages = DIV_ROUND_UP(rbio->stripe_len, PAGE_CACHE_SIZE); | |
2020 | int pagenr; | |
2021 | int stripe; | |
2022 | struct bio *bio; | |
2023 | ||
2024 | bio_list_init(&bio_list); | |
2025 | ||
2026 | ret = alloc_rbio_pages(rbio); | |
2027 | if (ret) | |
2028 | goto cleanup; | |
2029 | ||
2030 | atomic_set(&rbio->error, 0); | |
2031 | ||
2032 | /* | |
2033 | * read everything that hasn't failed. Thanks to the | |
2034 | * stripe cache, it is possible that some or all of these | |
2035 | * pages are going to be uptodate. | |
2036 | */ | |
2037 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) { | |
2038 | if (rbio->faila == stripe || rbio->failb == stripe) { | |
2039 | atomic_inc(&rbio->error); | |
2040 | continue; | |
2041 | } | |
2042 | ||
2043 | for (pagenr = 0; pagenr < nr_pages; pagenr++) { | |
2044 | struct page *p; | |
2045 | ||
2046 | /* | |
2047 | * the rmw code may have already read this | |
2048 | * page in | |
2049 | */ | |
2050 | p = rbio_stripe_page(rbio, stripe, pagenr); | |
2051 | if (PageUptodate(p)) | |
2052 | continue; | |
2053 | ||
2054 | ret = rbio_add_io_page(rbio, &bio_list, | |
2055 | rbio_stripe_page(rbio, stripe, pagenr), | |
2056 | stripe, pagenr, rbio->stripe_len); | |
2057 | if (ret < 0) | |
2058 | goto cleanup; | |
2059 | } | |
2060 | } | |
2061 | ||
2062 | bios_to_read = bio_list_size(&bio_list); | |
2063 | if (!bios_to_read) { | |
2064 | /* | |
2065 | * we might have no bios to read just because the pages | |
2066 | * were up to date, or we might have no bios to read because | |
2067 | * the devices were gone. | |
2068 | */ | |
2069 | if (atomic_read(&rbio->error) <= rbio->bbio->max_errors) { | |
2070 | __raid_recover_end_io(rbio); | |
2071 | goto out; | |
2072 | } else { | |
2073 | goto cleanup; | |
2074 | } | |
2075 | } | |
2076 | ||
2077 | /* | |
2078 | * the bbio may be freed once we submit the last bio. Make sure | |
2079 | * not to touch it after that | |
2080 | */ | |
2081 | atomic_set(&rbio->stripes_pending, bios_to_read); | |
2082 | while (1) { | |
2083 | bio = bio_list_pop(&bio_list); | |
2084 | if (!bio) | |
2085 | break; | |
2086 | ||
2087 | bio->bi_private = rbio; | |
2088 | bio->bi_end_io = raid_recover_end_io; | |
2089 | ||
2090 | btrfs_bio_wq_end_io(rbio->fs_info, bio, | |
2091 | BTRFS_WQ_ENDIO_RAID56); | |
2092 | ||
2093 | submit_bio(READ, bio); | |
2094 | } | |
2095 | out: | |
2096 | return 0; | |
2097 | ||
2098 | cleanup: | |
2099 | if (rbio->operation == BTRFS_RBIO_READ_REBUILD) | |
2100 | rbio_orig_end_io(rbio, -EIO); | |
2101 | return -EIO; | |
2102 | } | |
2103 | ||
2104 | /* | |
2105 | * the main entry point for reads from the higher layers. This | |
2106 | * is really only called when the normal read path had a failure, | |
2107 | * so we assume the bio they send down corresponds to a failed part | |
2108 | * of the drive. | |
2109 | */ | |
2110 | int raid56_parity_recover(struct btrfs_root *root, struct bio *bio, | |
2111 | struct btrfs_bio *bbio, u64 stripe_len, | |
2112 | int mirror_num, int generic_io) | |
2113 | { | |
2114 | struct btrfs_raid_bio *rbio; | |
2115 | int ret; | |
2116 | ||
2117 | rbio = alloc_rbio(root, bbio, stripe_len); | |
2118 | if (IS_ERR(rbio)) { | |
2119 | if (generic_io) | |
2120 | btrfs_put_bbio(bbio); | |
2121 | return PTR_ERR(rbio); | |
2122 | } | |
2123 | ||
2124 | rbio->operation = BTRFS_RBIO_READ_REBUILD; | |
2125 | bio_list_add(&rbio->bio_list, bio); | |
2126 | rbio->bio_list_bytes = bio->bi_iter.bi_size; | |
2127 | ||
2128 | rbio->faila = find_logical_bio_stripe(rbio, bio); | |
2129 | if (rbio->faila == -1) { | |
2130 | BUG(); | |
2131 | if (generic_io) | |
2132 | btrfs_put_bbio(bbio); | |
2133 | kfree(rbio); | |
2134 | return -EIO; | |
2135 | } | |
2136 | ||
2137 | if (generic_io) { | |
2138 | btrfs_bio_counter_inc_noblocked(root->fs_info); | |
2139 | rbio->generic_bio_cnt = 1; | |
2140 | } else { | |
2141 | btrfs_get_bbio(bbio); | |
2142 | } | |
2143 | ||
2144 | /* | |
2145 | * reconstruct from the q stripe if they are | |
2146 | * asking for mirror 3 | |
2147 | */ | |
2148 | if (mirror_num == 3) | |
2149 | rbio->failb = rbio->real_stripes - 2; | |
2150 | ||
2151 | ret = lock_stripe_add(rbio); | |
2152 | ||
2153 | /* | |
2154 | * __raid56_parity_recover will end the bio with | |
2155 | * any errors it hits. We don't want to return | |
2156 | * its error value up the stack because our caller | |
2157 | * will end up calling bio_endio with any nonzero | |
2158 | * return | |
2159 | */ | |
2160 | if (ret == 0) | |
2161 | __raid56_parity_recover(rbio); | |
2162 | /* | |
2163 | * our rbio has been added to the list of | |
2164 | * rbios that will be handled after the | |
2165 | * currently lock owner is done | |
2166 | */ | |
2167 | return 0; | |
2168 | ||
2169 | } | |
2170 | ||
2171 | static void rmw_work(struct btrfs_work *work) | |
2172 | { | |
2173 | struct btrfs_raid_bio *rbio; | |
2174 | ||
2175 | rbio = container_of(work, struct btrfs_raid_bio, work); | |
2176 | raid56_rmw_stripe(rbio); | |
2177 | } | |
2178 | ||
2179 | static void read_rebuild_work(struct btrfs_work *work) | |
2180 | { | |
2181 | struct btrfs_raid_bio *rbio; | |
2182 | ||
2183 | rbio = container_of(work, struct btrfs_raid_bio, work); | |
2184 | __raid56_parity_recover(rbio); | |
2185 | } | |
2186 | ||
2187 | /* | |
2188 | * The following code is used to scrub/replace the parity stripe | |
2189 | * | |
2190 | * Note: We need make sure all the pages that add into the scrub/replace | |
2191 | * raid bio are correct and not be changed during the scrub/replace. That | |
2192 | * is those pages just hold metadata or file data with checksum. | |
2193 | */ | |
2194 | ||
2195 | struct btrfs_raid_bio * | |
2196 | raid56_parity_alloc_scrub_rbio(struct btrfs_root *root, struct bio *bio, | |
2197 | struct btrfs_bio *bbio, u64 stripe_len, | |
2198 | struct btrfs_device *scrub_dev, | |
2199 | unsigned long *dbitmap, int stripe_nsectors) | |
2200 | { | |
2201 | struct btrfs_raid_bio *rbio; | |
2202 | int i; | |
2203 | ||
2204 | rbio = alloc_rbio(root, bbio, stripe_len); | |
2205 | if (IS_ERR(rbio)) | |
2206 | return NULL; | |
2207 | bio_list_add(&rbio->bio_list, bio); | |
2208 | /* | |
2209 | * This is a special bio which is used to hold the completion handler | |
2210 | * and make the scrub rbio is similar to the other types | |
2211 | */ | |
2212 | ASSERT(!bio->bi_iter.bi_size); | |
2213 | rbio->operation = BTRFS_RBIO_PARITY_SCRUB; | |
2214 | ||
2215 | for (i = 0; i < rbio->real_stripes; i++) { | |
2216 | if (bbio->stripes[i].dev == scrub_dev) { | |
2217 | rbio->scrubp = i; | |
2218 | break; | |
2219 | } | |
2220 | } | |
2221 | ||
2222 | /* Now we just support the sectorsize equals to page size */ | |
2223 | ASSERT(root->sectorsize == PAGE_SIZE); | |
2224 | ASSERT(rbio->stripe_npages == stripe_nsectors); | |
2225 | bitmap_copy(rbio->dbitmap, dbitmap, stripe_nsectors); | |
2226 | ||
2227 | return rbio; | |
2228 | } | |
2229 | ||
2230 | void raid56_parity_add_scrub_pages(struct btrfs_raid_bio *rbio, | |
2231 | struct page *page, u64 logical) | |
2232 | { | |
2233 | int stripe_offset; | |
2234 | int index; | |
2235 | ||
2236 | ASSERT(logical >= rbio->bbio->raid_map[0]); | |
2237 | ASSERT(logical + PAGE_SIZE <= rbio->bbio->raid_map[0] + | |
2238 | rbio->stripe_len * rbio->nr_data); | |
2239 | stripe_offset = (int)(logical - rbio->bbio->raid_map[0]); | |
2240 | index = stripe_offset >> PAGE_CACHE_SHIFT; | |
2241 | rbio->bio_pages[index] = page; | |
2242 | } | |
2243 | ||
2244 | /* | |
2245 | * We just scrub the parity that we have correct data on the same horizontal, | |
2246 | * so we needn't allocate all pages for all the stripes. | |
2247 | */ | |
2248 | static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio) | |
2249 | { | |
2250 | int i; | |
2251 | int bit; | |
2252 | int index; | |
2253 | struct page *page; | |
2254 | ||
2255 | for_each_set_bit(bit, rbio->dbitmap, rbio->stripe_npages) { | |
2256 | for (i = 0; i < rbio->real_stripes; i++) { | |
2257 | index = i * rbio->stripe_npages + bit; | |
2258 | if (rbio->stripe_pages[index]) | |
2259 | continue; | |
2260 | ||
2261 | page = alloc_page(GFP_NOFS | __GFP_HIGHMEM); | |
2262 | if (!page) | |
2263 | return -ENOMEM; | |
2264 | rbio->stripe_pages[index] = page; | |
2265 | ClearPageUptodate(page); | |
2266 | } | |
2267 | } | |
2268 | return 0; | |
2269 | } | |
2270 | ||
2271 | /* | |
2272 | * end io function used by finish_rmw. When we finally | |
2273 | * get here, we've written a full stripe | |
2274 | */ | |
2275 | static void raid_write_parity_end_io(struct bio *bio) | |
2276 | { | |
2277 | struct btrfs_raid_bio *rbio = bio->bi_private; | |
2278 | int err = bio->bi_error; | |
2279 | ||
2280 | if (bio->bi_error) | |
2281 | fail_bio_stripe(rbio, bio); | |
2282 | ||
2283 | bio_put(bio); | |
2284 | ||
2285 | if (!atomic_dec_and_test(&rbio->stripes_pending)) | |
2286 | return; | |
2287 | ||
2288 | err = 0; | |
2289 | ||
2290 | if (atomic_read(&rbio->error)) | |
2291 | err = -EIO; | |
2292 | ||
2293 | rbio_orig_end_io(rbio, err); | |
2294 | } | |
2295 | ||
2296 | static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio, | |
2297 | int need_check) | |
2298 | { | |
2299 | struct btrfs_bio *bbio = rbio->bbio; | |
2300 | void *pointers[rbio->real_stripes]; | |
2301 | DECLARE_BITMAP(pbitmap, rbio->stripe_npages); | |
2302 | int nr_data = rbio->nr_data; | |
2303 | int stripe; | |
2304 | int pagenr; | |
2305 | int p_stripe = -1; | |
2306 | int q_stripe = -1; | |
2307 | struct page *p_page = NULL; | |
2308 | struct page *q_page = NULL; | |
2309 | struct bio_list bio_list; | |
2310 | struct bio *bio; | |
2311 | int is_replace = 0; | |
2312 | int ret; | |
2313 | ||
2314 | bio_list_init(&bio_list); | |
2315 | ||
2316 | if (rbio->real_stripes - rbio->nr_data == 1) { | |
2317 | p_stripe = rbio->real_stripes - 1; | |
2318 | } else if (rbio->real_stripes - rbio->nr_data == 2) { | |
2319 | p_stripe = rbio->real_stripes - 2; | |
2320 | q_stripe = rbio->real_stripes - 1; | |
2321 | } else { | |
2322 | BUG(); | |
2323 | } | |
2324 | ||
2325 | if (bbio->num_tgtdevs && bbio->tgtdev_map[rbio->scrubp]) { | |
2326 | is_replace = 1; | |
2327 | bitmap_copy(pbitmap, rbio->dbitmap, rbio->stripe_npages); | |
2328 | } | |
2329 | ||
2330 | /* | |
2331 | * Because the higher layers(scrubber) are unlikely to | |
2332 | * use this area of the disk again soon, so don't cache | |
2333 | * it. | |
2334 | */ | |
2335 | clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags); | |
2336 | ||
2337 | if (!need_check) | |
2338 | goto writeback; | |
2339 | ||
2340 | p_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM); | |
2341 | if (!p_page) | |
2342 | goto cleanup; | |
2343 | SetPageUptodate(p_page); | |
2344 | ||
2345 | if (q_stripe != -1) { | |
2346 | q_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM); | |
2347 | if (!q_page) { | |
2348 | __free_page(p_page); | |
2349 | goto cleanup; | |
2350 | } | |
2351 | SetPageUptodate(q_page); | |
2352 | } | |
2353 | ||
2354 | atomic_set(&rbio->error, 0); | |
2355 | ||
2356 | for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) { | |
2357 | struct page *p; | |
2358 | void *parity; | |
2359 | /* first collect one page from each data stripe */ | |
2360 | for (stripe = 0; stripe < nr_data; stripe++) { | |
2361 | p = page_in_rbio(rbio, stripe, pagenr, 0); | |
2362 | pointers[stripe] = kmap(p); | |
2363 | } | |
2364 | ||
2365 | /* then add the parity stripe */ | |
2366 | pointers[stripe++] = kmap(p_page); | |
2367 | ||
2368 | if (q_stripe != -1) { | |
2369 | ||
2370 | /* | |
2371 | * raid6, add the qstripe and call the | |
2372 | * library function to fill in our p/q | |
2373 | */ | |
2374 | pointers[stripe++] = kmap(q_page); | |
2375 | ||
2376 | raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE, | |
2377 | pointers); | |
2378 | } else { | |
2379 | /* raid5 */ | |
2380 | memcpy(pointers[nr_data], pointers[0], PAGE_SIZE); | |
2381 | run_xor(pointers + 1, nr_data - 1, PAGE_CACHE_SIZE); | |
2382 | } | |
2383 | ||
2384 | /* Check scrubbing pairty and repair it */ | |
2385 | p = rbio_stripe_page(rbio, rbio->scrubp, pagenr); | |
2386 | parity = kmap(p); | |
2387 | if (memcmp(parity, pointers[rbio->scrubp], PAGE_CACHE_SIZE)) | |
2388 | memcpy(parity, pointers[rbio->scrubp], PAGE_CACHE_SIZE); | |
2389 | else | |
2390 | /* Parity is right, needn't writeback */ | |
2391 | bitmap_clear(rbio->dbitmap, pagenr, 1); | |
2392 | kunmap(p); | |
2393 | ||
2394 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) | |
2395 | kunmap(page_in_rbio(rbio, stripe, pagenr, 0)); | |
2396 | } | |
2397 | ||
2398 | __free_page(p_page); | |
2399 | if (q_page) | |
2400 | __free_page(q_page); | |
2401 | ||
2402 | writeback: | |
2403 | /* | |
2404 | * time to start writing. Make bios for everything from the | |
2405 | * higher layers (the bio_list in our rbio) and our p/q. Ignore | |
2406 | * everything else. | |
2407 | */ | |
2408 | for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) { | |
2409 | struct page *page; | |
2410 | ||
2411 | page = rbio_stripe_page(rbio, rbio->scrubp, pagenr); | |
2412 | ret = rbio_add_io_page(rbio, &bio_list, | |
2413 | page, rbio->scrubp, pagenr, rbio->stripe_len); | |
2414 | if (ret) | |
2415 | goto cleanup; | |
2416 | } | |
2417 | ||
2418 | if (!is_replace) | |
2419 | goto submit_write; | |
2420 | ||
2421 | for_each_set_bit(pagenr, pbitmap, rbio->stripe_npages) { | |
2422 | struct page *page; | |
2423 | ||
2424 | page = rbio_stripe_page(rbio, rbio->scrubp, pagenr); | |
2425 | ret = rbio_add_io_page(rbio, &bio_list, page, | |
2426 | bbio->tgtdev_map[rbio->scrubp], | |
2427 | pagenr, rbio->stripe_len); | |
2428 | if (ret) | |
2429 | goto cleanup; | |
2430 | } | |
2431 | ||
2432 | submit_write: | |
2433 | nr_data = bio_list_size(&bio_list); | |
2434 | if (!nr_data) { | |
2435 | /* Every parity is right */ | |
2436 | rbio_orig_end_io(rbio, 0); | |
2437 | return; | |
2438 | } | |
2439 | ||
2440 | atomic_set(&rbio->stripes_pending, nr_data); | |
2441 | ||
2442 | while (1) { | |
2443 | bio = bio_list_pop(&bio_list); | |
2444 | if (!bio) | |
2445 | break; | |
2446 | ||
2447 | bio->bi_private = rbio; | |
2448 | bio->bi_end_io = raid_write_parity_end_io; | |
2449 | submit_bio(WRITE, bio); | |
2450 | } | |
2451 | return; | |
2452 | ||
2453 | cleanup: | |
2454 | rbio_orig_end_io(rbio, -EIO); | |
2455 | } | |
2456 | ||
2457 | static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe) | |
2458 | { | |
2459 | if (stripe >= 0 && stripe < rbio->nr_data) | |
2460 | return 1; | |
2461 | return 0; | |
2462 | } | |
2463 | ||
2464 | /* | |
2465 | * While we're doing the parity check and repair, we could have errors | |
2466 | * in reading pages off the disk. This checks for errors and if we're | |
2467 | * not able to read the page it'll trigger parity reconstruction. The | |
2468 | * parity scrub will be finished after we've reconstructed the failed | |
2469 | * stripes | |
2470 | */ | |
2471 | static void validate_rbio_for_parity_scrub(struct btrfs_raid_bio *rbio) | |
2472 | { | |
2473 | if (atomic_read(&rbio->error) > rbio->bbio->max_errors) | |
2474 | goto cleanup; | |
2475 | ||
2476 | if (rbio->faila >= 0 || rbio->failb >= 0) { | |
2477 | int dfail = 0, failp = -1; | |
2478 | ||
2479 | if (is_data_stripe(rbio, rbio->faila)) | |
2480 | dfail++; | |
2481 | else if (is_parity_stripe(rbio->faila)) | |
2482 | failp = rbio->faila; | |
2483 | ||
2484 | if (is_data_stripe(rbio, rbio->failb)) | |
2485 | dfail++; | |
2486 | else if (is_parity_stripe(rbio->failb)) | |
2487 | failp = rbio->failb; | |
2488 | ||
2489 | /* | |
2490 | * Because we can not use a scrubbing parity to repair | |
2491 | * the data, so the capability of the repair is declined. | |
2492 | * (In the case of RAID5, we can not repair anything) | |
2493 | */ | |
2494 | if (dfail > rbio->bbio->max_errors - 1) | |
2495 | goto cleanup; | |
2496 | ||
2497 | /* | |
2498 | * If all data is good, only parity is correctly, just | |
2499 | * repair the parity. | |
2500 | */ | |
2501 | if (dfail == 0) { | |
2502 | finish_parity_scrub(rbio, 0); | |
2503 | return; | |
2504 | } | |
2505 | ||
2506 | /* | |
2507 | * Here means we got one corrupted data stripe and one | |
2508 | * corrupted parity on RAID6, if the corrupted parity | |
2509 | * is scrubbing parity, luckly, use the other one to repair | |
2510 | * the data, or we can not repair the data stripe. | |
2511 | */ | |
2512 | if (failp != rbio->scrubp) | |
2513 | goto cleanup; | |
2514 | ||
2515 | __raid_recover_end_io(rbio); | |
2516 | } else { | |
2517 | finish_parity_scrub(rbio, 1); | |
2518 | } | |
2519 | return; | |
2520 | ||
2521 | cleanup: | |
2522 | rbio_orig_end_io(rbio, -EIO); | |
2523 | } | |
2524 | ||
2525 | /* | |
2526 | * end io for the read phase of the rmw cycle. All the bios here are physical | |
2527 | * stripe bios we've read from the disk so we can recalculate the parity of the | |
2528 | * stripe. | |
2529 | * | |
2530 | * This will usually kick off finish_rmw once all the bios are read in, but it | |
2531 | * may trigger parity reconstruction if we had any errors along the way | |
2532 | */ | |
2533 | static void raid56_parity_scrub_end_io(struct bio *bio) | |
2534 | { | |
2535 | struct btrfs_raid_bio *rbio = bio->bi_private; | |
2536 | ||
2537 | if (bio->bi_error) | |
2538 | fail_bio_stripe(rbio, bio); | |
2539 | else | |
2540 | set_bio_pages_uptodate(bio); | |
2541 | ||
2542 | bio_put(bio); | |
2543 | ||
2544 | if (!atomic_dec_and_test(&rbio->stripes_pending)) | |
2545 | return; | |
2546 | ||
2547 | /* | |
2548 | * this will normally call finish_rmw to start our write | |
2549 | * but if there are any failed stripes we'll reconstruct | |
2550 | * from parity first | |
2551 | */ | |
2552 | validate_rbio_for_parity_scrub(rbio); | |
2553 | } | |
2554 | ||
2555 | static void raid56_parity_scrub_stripe(struct btrfs_raid_bio *rbio) | |
2556 | { | |
2557 | int bios_to_read = 0; | |
2558 | struct bio_list bio_list; | |
2559 | int ret; | |
2560 | int pagenr; | |
2561 | int stripe; | |
2562 | struct bio *bio; | |
2563 | ||
2564 | ret = alloc_rbio_essential_pages(rbio); | |
2565 | if (ret) | |
2566 | goto cleanup; | |
2567 | ||
2568 | bio_list_init(&bio_list); | |
2569 | ||
2570 | atomic_set(&rbio->error, 0); | |
2571 | /* | |
2572 | * build a list of bios to read all the missing parts of this | |
2573 | * stripe | |
2574 | */ | |
2575 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) { | |
2576 | for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) { | |
2577 | struct page *page; | |
2578 | /* | |
2579 | * we want to find all the pages missing from | |
2580 | * the rbio and read them from the disk. If | |
2581 | * page_in_rbio finds a page in the bio list | |
2582 | * we don't need to read it off the stripe. | |
2583 | */ | |
2584 | page = page_in_rbio(rbio, stripe, pagenr, 1); | |
2585 | if (page) | |
2586 | continue; | |
2587 | ||
2588 | page = rbio_stripe_page(rbio, stripe, pagenr); | |
2589 | /* | |
2590 | * the bio cache may have handed us an uptodate | |
2591 | * page. If so, be happy and use it | |
2592 | */ | |
2593 | if (PageUptodate(page)) | |
2594 | continue; | |
2595 | ||
2596 | ret = rbio_add_io_page(rbio, &bio_list, page, | |
2597 | stripe, pagenr, rbio->stripe_len); | |
2598 | if (ret) | |
2599 | goto cleanup; | |
2600 | } | |
2601 | } | |
2602 | ||
2603 | bios_to_read = bio_list_size(&bio_list); | |
2604 | if (!bios_to_read) { | |
2605 | /* | |
2606 | * this can happen if others have merged with | |
2607 | * us, it means there is nothing left to read. | |
2608 | * But if there are missing devices it may not be | |
2609 | * safe to do the full stripe write yet. | |
2610 | */ | |
2611 | goto finish; | |
2612 | } | |
2613 | ||
2614 | /* | |
2615 | * the bbio may be freed once we submit the last bio. Make sure | |
2616 | * not to touch it after that | |
2617 | */ | |
2618 | atomic_set(&rbio->stripes_pending, bios_to_read); | |
2619 | while (1) { | |
2620 | bio = bio_list_pop(&bio_list); | |
2621 | if (!bio) | |
2622 | break; | |
2623 | ||
2624 | bio->bi_private = rbio; | |
2625 | bio->bi_end_io = raid56_parity_scrub_end_io; | |
2626 | ||
2627 | btrfs_bio_wq_end_io(rbio->fs_info, bio, | |
2628 | BTRFS_WQ_ENDIO_RAID56); | |
2629 | ||
2630 | submit_bio(READ, bio); | |
2631 | } | |
2632 | /* the actual write will happen once the reads are done */ | |
2633 | return; | |
2634 | ||
2635 | cleanup: | |
2636 | rbio_orig_end_io(rbio, -EIO); | |
2637 | return; | |
2638 | ||
2639 | finish: | |
2640 | validate_rbio_for_parity_scrub(rbio); | |
2641 | } | |
2642 | ||
2643 | static void scrub_parity_work(struct btrfs_work *work) | |
2644 | { | |
2645 | struct btrfs_raid_bio *rbio; | |
2646 | ||
2647 | rbio = container_of(work, struct btrfs_raid_bio, work); | |
2648 | raid56_parity_scrub_stripe(rbio); | |
2649 | } | |
2650 | ||
2651 | static void async_scrub_parity(struct btrfs_raid_bio *rbio) | |
2652 | { | |
2653 | btrfs_init_work(&rbio->work, btrfs_rmw_helper, | |
2654 | scrub_parity_work, NULL, NULL); | |
2655 | ||
2656 | btrfs_queue_work(rbio->fs_info->rmw_workers, | |
2657 | &rbio->work); | |
2658 | } | |
2659 | ||
2660 | void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio) | |
2661 | { | |
2662 | if (!lock_stripe_add(rbio)) | |
2663 | async_scrub_parity(rbio); | |
2664 | } |