]>
Commit | Line | Data |
---|---|---|
53b381b3 DW |
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> | |
d7011f5b | 34 | #include <linux/vmalloc.h> |
53b381b3 | 35 | #include <asm/div64.h> |
53b381b3 DW |
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 | ||
4ae10b3a CM |
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 | ||
4ae10b3a CM |
61 | #define RBIO_CACHE_SIZE 1024 |
62 | ||
1b94b556 MX |
63 | enum btrfs_rbio_ops { |
64 | BTRFS_RBIO_WRITE = 0, | |
65 | BTRFS_RBIO_READ_REBUILD = 1, | |
5a6ac9ea | 66 | BTRFS_RBIO_PARITY_SCRUB = 2, |
1b94b556 MX |
67 | }; |
68 | ||
53b381b3 DW |
69 | struct btrfs_raid_bio { |
70 | struct btrfs_fs_info *fs_info; | |
71 | struct btrfs_bio *bbio; | |
72 | ||
53b381b3 DW |
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 | ||
4ae10b3a CM |
80 | /* |
81 | * LRU list for the stripe cache | |
82 | */ | |
83 | struct list_head stripe_cache; | |
84 | ||
53b381b3 DW |
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 | ||
6ac0f488 CM |
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 | |
53b381b3 DW |
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 | ||
2c8cdd6e MX |
118 | int real_stripes; |
119 | ||
5a6ac9ea | 120 | int stripe_npages; |
53b381b3 DW |
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 | */ | |
1b94b556 | 127 | enum btrfs_rbio_ops operation; |
53b381b3 DW |
128 | |
129 | /* first bad stripe */ | |
130 | int faila; | |
131 | ||
132 | /* second bad stripe (for raid6 use) */ | |
133 | int failb; | |
134 | ||
5a6ac9ea | 135 | int scrubp; |
53b381b3 DW |
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 | ||
4245215d MX |
149 | int generic_bio_cnt; |
150 | ||
53b381b3 DW |
151 | atomic_t refs; |
152 | ||
b89e1b01 MX |
153 | atomic_t stripes_pending; |
154 | ||
155 | atomic_t error; | |
53b381b3 DW |
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; | |
5a6ac9ea MX |
172 | |
173 | /* | |
174 | * bitmap to record which horizontal stripe has data | |
175 | */ | |
176 | unsigned long *dbitmap; | |
53b381b3 DW |
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 | ||
5a6ac9ea MX |
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 | ||
53b381b3 DW |
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; | |
83c8266a | 207 | int table_size; |
53b381b3 DW |
208 | |
209 | if (info->stripe_hash_table) | |
210 | return 0; | |
211 | ||
83c8266a DS |
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 | } | |
53b381b3 | 226 | |
4ae10b3a CM |
227 | spin_lock_init(&table->cache_lock); |
228 | INIT_LIST_HEAD(&table->stripe_cache); | |
229 | ||
53b381b3 DW |
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); | |
f749303b WS |
240 | if (x) |
241 | kvfree(x); | |
53b381b3 DW |
242 | return 0; |
243 | } | |
244 | ||
4ae10b3a CM |
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 | ||
53b381b3 DW |
281 | /* |
282 | * we hash on the first logical address of the stripe | |
283 | */ | |
284 | static int rbio_bucket(struct btrfs_raid_bio *rbio) | |
285 | { | |
8e5cfb55 | 286 | u64 num = rbio->bbio->raid_map[0]; |
53b381b3 DW |
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 | ||
4ae10b3a CM |
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 | ||
53b381b3 DW |
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; | |
4245215d | 339 | dest->generic_bio_cnt += victim->generic_bio_cnt; |
53b381b3 DW |
340 | bio_list_init(&victim->bio_list); |
341 | } | |
342 | ||
343 | /* | |
4ae10b3a CM |
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 | */ | |
48a3b636 | 425 | static void btrfs_clear_rbio_cache(struct btrfs_fs_info *info) |
4ae10b3a CM |
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 | |
53b381b3 DW |
446 | */ |
447 | void btrfs_free_stripe_hash_table(struct btrfs_fs_info *info) | |
448 | { | |
449 | if (!info->stripe_hash_table) | |
450 | return; | |
4ae10b3a | 451 | btrfs_clear_rbio_cache(info); |
f749303b | 452 | kvfree(info->stripe_hash_table); |
53b381b3 DW |
453 | info->stripe_hash_table = NULL; |
454 | } | |
455 | ||
4ae10b3a CM |
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 | ||
53b381b3 DW |
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 | ||
4ae10b3a CM |
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 | ||
8e5cfb55 ZL |
586 | if (last->bbio->raid_map[0] != |
587 | cur->bbio->raid_map[0]) | |
53b381b3 DW |
588 | return 0; |
589 | ||
5a6ac9ea MX |
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) | |
53b381b3 | 603 | return 0; |
53b381b3 DW |
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 | { | |
2c8cdd6e | 623 | if (rbio->nr_data + 1 == rbio->real_stripes) |
53b381b3 DW |
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; | |
4ae10b3a | 662 | struct btrfs_raid_bio *cache_drop = NULL; |
53b381b3 DW |
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++; | |
8e5cfb55 | 669 | if (cur->bbio->raid_map[0] == rbio->bbio->raid_map[0]) { |
53b381b3 DW |
670 | spin_lock(&cur->bio_list_lock); |
671 | ||
4ae10b3a CM |
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 | ||
53b381b3 DW |
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 | ||
4ae10b3a | 696 | |
53b381b3 DW |
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 | } | |
4ae10b3a | 726 | lockit: |
53b381b3 DW |
727 | atomic_inc(&rbio->refs); |
728 | list_add(&rbio->hash_list, &h->hash_list); | |
729 | out: | |
730 | spin_unlock_irqrestore(&h->lock, flags); | |
4ae10b3a CM |
731 | if (cache_drop) |
732 | remove_rbio_from_cache(cache_drop); | |
53b381b3 DW |
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; | |
4ae10b3a | 747 | int keep_cache = 0; |
53b381b3 DW |
748 | |
749 | bucket = rbio_bucket(rbio); | |
750 | h = rbio->fs_info->stripe_hash_table->table + bucket; | |
751 | ||
4ae10b3a CM |
752 | if (list_empty(&rbio->plug_list)) |
753 | cache_rbio(rbio); | |
754 | ||
53b381b3 DW |
755 | spin_lock_irqsave(&h->lock, flags); |
756 | spin_lock(&rbio->bio_list_lock); | |
757 | ||
758 | if (!list_empty(&rbio->hash_list)) { | |
4ae10b3a CM |
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 | } | |
53b381b3 DW |
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 | ||
1b94b556 | 794 | if (next->operation == BTRFS_RBIO_READ_REBUILD) |
53b381b3 | 795 | async_read_rebuild(next); |
5a6ac9ea | 796 | else if (next->operation == BTRFS_RBIO_WRITE) { |
4ae10b3a | 797 | steal_rbio(rbio, next); |
53b381b3 | 798 | async_rmw_stripe(next); |
5a6ac9ea MX |
799 | } else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) { |
800 | steal_rbio(rbio, next); | |
801 | async_scrub_parity(next); | |
4ae10b3a | 802 | } |
53b381b3 DW |
803 | |
804 | goto done_nolock; | |
53b381b3 DW |
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 | } | |
4ae10b3a | 812 | done: |
53b381b3 DW |
813 | spin_unlock(&rbio->bio_list_lock); |
814 | spin_unlock_irqrestore(&h->lock, flags); | |
815 | ||
816 | done_nolock: | |
4ae10b3a CM |
817 | if (!keep_cache) |
818 | remove_rbio_from_cache(rbio); | |
53b381b3 DW |
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 | ||
4ae10b3a | 829 | WARN_ON(!list_empty(&rbio->stripe_cache)); |
53b381b3 DW |
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 | } | |
af8e2d1d | 839 | |
6e9606d2 | 840 | btrfs_put_bbio(rbio->bbio); |
53b381b3 DW |
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, int uptodate) | |
855 | { | |
856 | struct bio *cur = bio_list_get(&rbio->bio_list); | |
857 | struct bio *next; | |
4245215d MX |
858 | |
859 | if (rbio->generic_bio_cnt) | |
860 | btrfs_bio_counter_sub(rbio->fs_info, rbio->generic_bio_cnt); | |
861 | ||
53b381b3 DW |
862 | free_raid_bio(rbio); |
863 | ||
864 | while (cur) { | |
865 | next = cur->bi_next; | |
866 | cur->bi_next = NULL; | |
867 | if (uptodate) | |
868 | set_bit(BIO_UPTODATE, &cur->bi_flags); | |
869 | bio_endio(cur, err); | |
870 | cur = next; | |
871 | } | |
872 | } | |
873 | ||
874 | /* | |
875 | * end io function used by finish_rmw. When we finally | |
876 | * get here, we've written a full stripe | |
877 | */ | |
878 | static void raid_write_end_io(struct bio *bio, int err) | |
879 | { | |
880 | struct btrfs_raid_bio *rbio = bio->bi_private; | |
881 | ||
882 | if (err) | |
883 | fail_bio_stripe(rbio, bio); | |
884 | ||
885 | bio_put(bio); | |
886 | ||
b89e1b01 | 887 | if (!atomic_dec_and_test(&rbio->stripes_pending)) |
53b381b3 DW |
888 | return; |
889 | ||
890 | err = 0; | |
891 | ||
892 | /* OK, we have read all the stripes we need to. */ | |
b89e1b01 | 893 | if (atomic_read(&rbio->error) > rbio->bbio->max_errors) |
53b381b3 DW |
894 | err = -EIO; |
895 | ||
896 | rbio_orig_end_io(rbio, err, 0); | |
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; | |
ed6078f7 | 941 | return DIV_ROUND_UP(nr, PAGE_CACHE_SIZE); |
53b381b3 DW |
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, | |
8e5cfb55 | 949 | struct btrfs_bio *bbio, u64 stripe_len) |
53b381b3 DW |
950 | { |
951 | struct btrfs_raid_bio *rbio; | |
952 | int nr_data = 0; | |
2c8cdd6e MX |
953 | int real_stripes = bbio->num_stripes - bbio->num_tgtdevs; |
954 | int num_pages = rbio_nr_pages(stripe_len, real_stripes); | |
5a6ac9ea | 955 | int stripe_npages = DIV_ROUND_UP(stripe_len, PAGE_SIZE); |
53b381b3 DW |
956 | void *p; |
957 | ||
5a6ac9ea MX |
958 | rbio = kzalloc(sizeof(*rbio) + num_pages * sizeof(struct page *) * 2 + |
959 | DIV_ROUND_UP(stripe_npages, BITS_PER_LONG / 8), | |
53b381b3 | 960 | GFP_NOFS); |
af8e2d1d | 961 | if (!rbio) |
53b381b3 | 962 | return ERR_PTR(-ENOMEM); |
53b381b3 DW |
963 | |
964 | bio_list_init(&rbio->bio_list); | |
965 | INIT_LIST_HEAD(&rbio->plug_list); | |
966 | spin_lock_init(&rbio->bio_list_lock); | |
4ae10b3a | 967 | INIT_LIST_HEAD(&rbio->stripe_cache); |
53b381b3 DW |
968 | INIT_LIST_HEAD(&rbio->hash_list); |
969 | rbio->bbio = bbio; | |
53b381b3 DW |
970 | rbio->fs_info = root->fs_info; |
971 | rbio->stripe_len = stripe_len; | |
972 | rbio->nr_pages = num_pages; | |
2c8cdd6e | 973 | rbio->real_stripes = real_stripes; |
5a6ac9ea | 974 | rbio->stripe_npages = stripe_npages; |
53b381b3 DW |
975 | rbio->faila = -1; |
976 | rbio->failb = -1; | |
977 | atomic_set(&rbio->refs, 1); | |
b89e1b01 MX |
978 | atomic_set(&rbio->error, 0); |
979 | atomic_set(&rbio->stripes_pending, 0); | |
53b381b3 DW |
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; | |
5a6ac9ea | 988 | rbio->dbitmap = p + sizeof(struct page *) * num_pages * 2; |
53b381b3 | 989 | |
10f11900 ZL |
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) | |
2c8cdd6e | 993 | nr_data = real_stripes - 2; |
53b381b3 | 994 | else |
10f11900 | 995 | BUG(); |
53b381b3 DW |
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 | */ | |
48a3b636 ES |
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) | |
53b381b3 DW |
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) { | |
4f024f37 KO |
1066 | last_end = (u64)last->bi_iter.bi_sector << 9; |
1067 | last_end += last->bi_iter.bi_size; | |
53b381b3 DW |
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 | test_bit(BIO_UPTODATE, &last->bi_flags) && | |
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 */ | |
9be3395b | 1083 | bio = btrfs_io_bio_alloc(GFP_NOFS, bio_max_len >> PAGE_SHIFT?:1); |
53b381b3 DW |
1084 | if (!bio) |
1085 | return -ENOMEM; | |
1086 | ||
4f024f37 | 1087 | bio->bi_iter.bi_size = 0; |
53b381b3 | 1088 | bio->bi_bdev = stripe->dev->bdev; |
4f024f37 | 1089 | bio->bi_iter.bi_sector = disk_start >> 9; |
53b381b3 DW |
1090 | set_bit(BIO_UPTODATE, &bio->bi_flags); |
1091 | ||
1092 | bio_add_page(bio, page, PAGE_CACHE_SIZE, 0); | |
1093 | bio_list_add(bio_list, bio); | |
1094 | return 0; | |
1095 | } | |
1096 | ||
1097 | /* | |
1098 | * while we're doing the read/modify/write cycle, we could | |
1099 | * have errors in reading pages off the disk. This checks | |
1100 | * for errors and if we're not able to read the page it'll | |
1101 | * trigger parity reconstruction. The rmw will be finished | |
1102 | * after we've reconstructed the failed stripes | |
1103 | */ | |
1104 | static void validate_rbio_for_rmw(struct btrfs_raid_bio *rbio) | |
1105 | { | |
1106 | if (rbio->faila >= 0 || rbio->failb >= 0) { | |
2c8cdd6e | 1107 | BUG_ON(rbio->faila == rbio->real_stripes - 1); |
53b381b3 DW |
1108 | __raid56_parity_recover(rbio); |
1109 | } else { | |
1110 | finish_rmw(rbio); | |
1111 | } | |
1112 | } | |
1113 | ||
1114 | /* | |
1115 | * these are just the pages from the rbio array, not from anything | |
1116 | * the FS sent down to us | |
1117 | */ | |
1118 | static struct page *rbio_stripe_page(struct btrfs_raid_bio *rbio, int stripe, int page) | |
1119 | { | |
1120 | int index; | |
1121 | index = stripe * (rbio->stripe_len >> PAGE_CACHE_SHIFT); | |
1122 | index += page; | |
1123 | return rbio->stripe_pages[index]; | |
1124 | } | |
1125 | ||
1126 | /* | |
1127 | * helper function to walk our bio list and populate the bio_pages array with | |
1128 | * the result. This seems expensive, but it is faster than constantly | |
1129 | * searching through the bio list as we setup the IO in finish_rmw or stripe | |
1130 | * reconstruction. | |
1131 | * | |
1132 | * This must be called before you trust the answers from page_in_rbio | |
1133 | */ | |
1134 | static void index_rbio_pages(struct btrfs_raid_bio *rbio) | |
1135 | { | |
1136 | struct bio *bio; | |
1137 | u64 start; | |
1138 | unsigned long stripe_offset; | |
1139 | unsigned long page_index; | |
1140 | struct page *p; | |
1141 | int i; | |
1142 | ||
1143 | spin_lock_irq(&rbio->bio_list_lock); | |
1144 | bio_list_for_each(bio, &rbio->bio_list) { | |
4f024f37 | 1145 | start = (u64)bio->bi_iter.bi_sector << 9; |
8e5cfb55 | 1146 | stripe_offset = start - rbio->bbio->raid_map[0]; |
53b381b3 DW |
1147 | page_index = stripe_offset >> PAGE_CACHE_SHIFT; |
1148 | ||
1149 | for (i = 0; i < bio->bi_vcnt; i++) { | |
1150 | p = bio->bi_io_vec[i].bv_page; | |
1151 | rbio->bio_pages[page_index + i] = p; | |
1152 | } | |
1153 | } | |
1154 | spin_unlock_irq(&rbio->bio_list_lock); | |
1155 | } | |
1156 | ||
1157 | /* | |
1158 | * this is called from one of two situations. We either | |
1159 | * have a full stripe from the higher layers, or we've read all | |
1160 | * the missing bits off disk. | |
1161 | * | |
1162 | * This will calculate the parity and then send down any | |
1163 | * changed blocks. | |
1164 | */ | |
1165 | static noinline void finish_rmw(struct btrfs_raid_bio *rbio) | |
1166 | { | |
1167 | struct btrfs_bio *bbio = rbio->bbio; | |
2c8cdd6e | 1168 | void *pointers[rbio->real_stripes]; |
53b381b3 DW |
1169 | int stripe_len = rbio->stripe_len; |
1170 | int nr_data = rbio->nr_data; | |
1171 | int stripe; | |
1172 | int pagenr; | |
1173 | int p_stripe = -1; | |
1174 | int q_stripe = -1; | |
1175 | struct bio_list bio_list; | |
1176 | struct bio *bio; | |
1177 | int pages_per_stripe = stripe_len >> PAGE_CACHE_SHIFT; | |
1178 | int ret; | |
1179 | ||
1180 | bio_list_init(&bio_list); | |
1181 | ||
2c8cdd6e MX |
1182 | if (rbio->real_stripes - rbio->nr_data == 1) { |
1183 | p_stripe = rbio->real_stripes - 1; | |
1184 | } else if (rbio->real_stripes - rbio->nr_data == 2) { | |
1185 | p_stripe = rbio->real_stripes - 2; | |
1186 | q_stripe = rbio->real_stripes - 1; | |
53b381b3 DW |
1187 | } else { |
1188 | BUG(); | |
1189 | } | |
1190 | ||
1191 | /* at this point we either have a full stripe, | |
1192 | * or we've read the full stripe from the drive. | |
1193 | * recalculate the parity and write the new results. | |
1194 | * | |
1195 | * We're not allowed to add any new bios to the | |
1196 | * bio list here, anyone else that wants to | |
1197 | * change this stripe needs to do their own rmw. | |
1198 | */ | |
1199 | spin_lock_irq(&rbio->bio_list_lock); | |
1200 | set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags); | |
1201 | spin_unlock_irq(&rbio->bio_list_lock); | |
1202 | ||
b89e1b01 | 1203 | atomic_set(&rbio->error, 0); |
53b381b3 DW |
1204 | |
1205 | /* | |
1206 | * now that we've set rmw_locked, run through the | |
1207 | * bio list one last time and map the page pointers | |
4ae10b3a CM |
1208 | * |
1209 | * We don't cache full rbios because we're assuming | |
1210 | * the higher layers are unlikely to use this area of | |
1211 | * the disk again soon. If they do use it again, | |
1212 | * hopefully they will send another full bio. | |
53b381b3 DW |
1213 | */ |
1214 | index_rbio_pages(rbio); | |
4ae10b3a CM |
1215 | if (!rbio_is_full(rbio)) |
1216 | cache_rbio_pages(rbio); | |
1217 | else | |
1218 | clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags); | |
53b381b3 DW |
1219 | |
1220 | for (pagenr = 0; pagenr < pages_per_stripe; pagenr++) { | |
1221 | struct page *p; | |
1222 | /* first collect one page from each data stripe */ | |
1223 | for (stripe = 0; stripe < nr_data; stripe++) { | |
1224 | p = page_in_rbio(rbio, stripe, pagenr, 0); | |
1225 | pointers[stripe] = kmap(p); | |
1226 | } | |
1227 | ||
1228 | /* then add the parity stripe */ | |
1229 | p = rbio_pstripe_page(rbio, pagenr); | |
1230 | SetPageUptodate(p); | |
1231 | pointers[stripe++] = kmap(p); | |
1232 | ||
1233 | if (q_stripe != -1) { | |
1234 | ||
1235 | /* | |
1236 | * raid6, add the qstripe and call the | |
1237 | * library function to fill in our p/q | |
1238 | */ | |
1239 | p = rbio_qstripe_page(rbio, pagenr); | |
1240 | SetPageUptodate(p); | |
1241 | pointers[stripe++] = kmap(p); | |
1242 | ||
2c8cdd6e | 1243 | raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE, |
53b381b3 DW |
1244 | pointers); |
1245 | } else { | |
1246 | /* raid5 */ | |
1247 | memcpy(pointers[nr_data], pointers[0], PAGE_SIZE); | |
1248 | run_xor(pointers + 1, nr_data - 1, PAGE_CACHE_SIZE); | |
1249 | } | |
1250 | ||
1251 | ||
2c8cdd6e | 1252 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) |
53b381b3 DW |
1253 | kunmap(page_in_rbio(rbio, stripe, pagenr, 0)); |
1254 | } | |
1255 | ||
1256 | /* | |
1257 | * time to start writing. Make bios for everything from the | |
1258 | * higher layers (the bio_list in our rbio) and our p/q. Ignore | |
1259 | * everything else. | |
1260 | */ | |
2c8cdd6e | 1261 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) { |
53b381b3 DW |
1262 | for (pagenr = 0; pagenr < pages_per_stripe; pagenr++) { |
1263 | struct page *page; | |
1264 | if (stripe < rbio->nr_data) { | |
1265 | page = page_in_rbio(rbio, stripe, pagenr, 1); | |
1266 | if (!page) | |
1267 | continue; | |
1268 | } else { | |
1269 | page = rbio_stripe_page(rbio, stripe, pagenr); | |
1270 | } | |
1271 | ||
1272 | ret = rbio_add_io_page(rbio, &bio_list, | |
1273 | page, stripe, pagenr, rbio->stripe_len); | |
1274 | if (ret) | |
1275 | goto cleanup; | |
1276 | } | |
1277 | } | |
1278 | ||
2c8cdd6e MX |
1279 | if (likely(!bbio->num_tgtdevs)) |
1280 | goto write_data; | |
1281 | ||
1282 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) { | |
1283 | if (!bbio->tgtdev_map[stripe]) | |
1284 | continue; | |
1285 | ||
1286 | for (pagenr = 0; pagenr < pages_per_stripe; pagenr++) { | |
1287 | struct page *page; | |
1288 | if (stripe < rbio->nr_data) { | |
1289 | page = page_in_rbio(rbio, stripe, pagenr, 1); | |
1290 | if (!page) | |
1291 | continue; | |
1292 | } else { | |
1293 | page = rbio_stripe_page(rbio, stripe, pagenr); | |
1294 | } | |
1295 | ||
1296 | ret = rbio_add_io_page(rbio, &bio_list, page, | |
1297 | rbio->bbio->tgtdev_map[stripe], | |
1298 | pagenr, rbio->stripe_len); | |
1299 | if (ret) | |
1300 | goto cleanup; | |
1301 | } | |
1302 | } | |
1303 | ||
1304 | write_data: | |
b89e1b01 MX |
1305 | atomic_set(&rbio->stripes_pending, bio_list_size(&bio_list)); |
1306 | BUG_ON(atomic_read(&rbio->stripes_pending) == 0); | |
53b381b3 DW |
1307 | |
1308 | while (1) { | |
1309 | bio = bio_list_pop(&bio_list); | |
1310 | if (!bio) | |
1311 | break; | |
1312 | ||
1313 | bio->bi_private = rbio; | |
1314 | bio->bi_end_io = raid_write_end_io; | |
1315 | BUG_ON(!test_bit(BIO_UPTODATE, &bio->bi_flags)); | |
1316 | submit_bio(WRITE, bio); | |
1317 | } | |
1318 | return; | |
1319 | ||
1320 | cleanup: | |
1321 | rbio_orig_end_io(rbio, -EIO, 0); | |
1322 | } | |
1323 | ||
1324 | /* | |
1325 | * helper to find the stripe number for a given bio. Used to figure out which | |
1326 | * stripe has failed. This expects the bio to correspond to a physical disk, | |
1327 | * so it looks up based on physical sector numbers. | |
1328 | */ | |
1329 | static int find_bio_stripe(struct btrfs_raid_bio *rbio, | |
1330 | struct bio *bio) | |
1331 | { | |
4f024f37 | 1332 | u64 physical = bio->bi_iter.bi_sector; |
53b381b3 DW |
1333 | u64 stripe_start; |
1334 | int i; | |
1335 | struct btrfs_bio_stripe *stripe; | |
1336 | ||
1337 | physical <<= 9; | |
1338 | ||
1339 | for (i = 0; i < rbio->bbio->num_stripes; i++) { | |
1340 | stripe = &rbio->bbio->stripes[i]; | |
1341 | stripe_start = stripe->physical; | |
1342 | if (physical >= stripe_start && | |
2c8cdd6e MX |
1343 | physical < stripe_start + rbio->stripe_len && |
1344 | bio->bi_bdev == stripe->dev->bdev) { | |
53b381b3 DW |
1345 | return i; |
1346 | } | |
1347 | } | |
1348 | return -1; | |
1349 | } | |
1350 | ||
1351 | /* | |
1352 | * helper to find the stripe number for a given | |
1353 | * bio (before mapping). Used to figure out which stripe has | |
1354 | * failed. This looks up based on logical block numbers. | |
1355 | */ | |
1356 | static int find_logical_bio_stripe(struct btrfs_raid_bio *rbio, | |
1357 | struct bio *bio) | |
1358 | { | |
4f024f37 | 1359 | u64 logical = bio->bi_iter.bi_sector; |
53b381b3 DW |
1360 | u64 stripe_start; |
1361 | int i; | |
1362 | ||
1363 | logical <<= 9; | |
1364 | ||
1365 | for (i = 0; i < rbio->nr_data; i++) { | |
8e5cfb55 | 1366 | stripe_start = rbio->bbio->raid_map[i]; |
53b381b3 DW |
1367 | if (logical >= stripe_start && |
1368 | logical < stripe_start + rbio->stripe_len) { | |
1369 | return i; | |
1370 | } | |
1371 | } | |
1372 | return -1; | |
1373 | } | |
1374 | ||
1375 | /* | |
1376 | * returns -EIO if we had too many failures | |
1377 | */ | |
1378 | static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed) | |
1379 | { | |
1380 | unsigned long flags; | |
1381 | int ret = 0; | |
1382 | ||
1383 | spin_lock_irqsave(&rbio->bio_list_lock, flags); | |
1384 | ||
1385 | /* we already know this stripe is bad, move on */ | |
1386 | if (rbio->faila == failed || rbio->failb == failed) | |
1387 | goto out; | |
1388 | ||
1389 | if (rbio->faila == -1) { | |
1390 | /* first failure on this rbio */ | |
1391 | rbio->faila = failed; | |
b89e1b01 | 1392 | atomic_inc(&rbio->error); |
53b381b3 DW |
1393 | } else if (rbio->failb == -1) { |
1394 | /* second failure on this rbio */ | |
1395 | rbio->failb = failed; | |
b89e1b01 | 1396 | atomic_inc(&rbio->error); |
53b381b3 DW |
1397 | } else { |
1398 | ret = -EIO; | |
1399 | } | |
1400 | out: | |
1401 | spin_unlock_irqrestore(&rbio->bio_list_lock, flags); | |
1402 | ||
1403 | return ret; | |
1404 | } | |
1405 | ||
1406 | /* | |
1407 | * helper to fail a stripe based on a physical disk | |
1408 | * bio. | |
1409 | */ | |
1410 | static int fail_bio_stripe(struct btrfs_raid_bio *rbio, | |
1411 | struct bio *bio) | |
1412 | { | |
1413 | int failed = find_bio_stripe(rbio, bio); | |
1414 | ||
1415 | if (failed < 0) | |
1416 | return -EIO; | |
1417 | ||
1418 | return fail_rbio_index(rbio, failed); | |
1419 | } | |
1420 | ||
1421 | /* | |
1422 | * this sets each page in the bio uptodate. It should only be used on private | |
1423 | * rbio pages, nothing that comes in from the higher layers | |
1424 | */ | |
1425 | static void set_bio_pages_uptodate(struct bio *bio) | |
1426 | { | |
1427 | int i; | |
1428 | struct page *p; | |
1429 | ||
1430 | for (i = 0; i < bio->bi_vcnt; i++) { | |
1431 | p = bio->bi_io_vec[i].bv_page; | |
1432 | SetPageUptodate(p); | |
1433 | } | |
1434 | } | |
1435 | ||
1436 | /* | |
1437 | * end io for the read phase of the rmw cycle. All the bios here are physical | |
1438 | * stripe bios we've read from the disk so we can recalculate the parity of the | |
1439 | * stripe. | |
1440 | * | |
1441 | * This will usually kick off finish_rmw once all the bios are read in, but it | |
1442 | * may trigger parity reconstruction if we had any errors along the way | |
1443 | */ | |
1444 | static void raid_rmw_end_io(struct bio *bio, int err) | |
1445 | { | |
1446 | struct btrfs_raid_bio *rbio = bio->bi_private; | |
1447 | ||
1448 | if (err) | |
1449 | fail_bio_stripe(rbio, bio); | |
1450 | else | |
1451 | set_bio_pages_uptodate(bio); | |
1452 | ||
1453 | bio_put(bio); | |
1454 | ||
b89e1b01 | 1455 | if (!atomic_dec_and_test(&rbio->stripes_pending)) |
53b381b3 DW |
1456 | return; |
1457 | ||
1458 | err = 0; | |
b89e1b01 | 1459 | if (atomic_read(&rbio->error) > rbio->bbio->max_errors) |
53b381b3 DW |
1460 | goto cleanup; |
1461 | ||
1462 | /* | |
1463 | * this will normally call finish_rmw to start our write | |
1464 | * but if there are any failed stripes we'll reconstruct | |
1465 | * from parity first | |
1466 | */ | |
1467 | validate_rbio_for_rmw(rbio); | |
1468 | return; | |
1469 | ||
1470 | cleanup: | |
1471 | ||
1472 | rbio_orig_end_io(rbio, -EIO, 0); | |
1473 | } | |
1474 | ||
1475 | static void async_rmw_stripe(struct btrfs_raid_bio *rbio) | |
1476 | { | |
9e0af237 LB |
1477 | btrfs_init_work(&rbio->work, btrfs_rmw_helper, |
1478 | rmw_work, NULL, NULL); | |
53b381b3 | 1479 | |
d05a33ac QW |
1480 | btrfs_queue_work(rbio->fs_info->rmw_workers, |
1481 | &rbio->work); | |
53b381b3 DW |
1482 | } |
1483 | ||
1484 | static void async_read_rebuild(struct btrfs_raid_bio *rbio) | |
1485 | { | |
9e0af237 LB |
1486 | btrfs_init_work(&rbio->work, btrfs_rmw_helper, |
1487 | read_rebuild_work, NULL, NULL); | |
53b381b3 | 1488 | |
d05a33ac QW |
1489 | btrfs_queue_work(rbio->fs_info->rmw_workers, |
1490 | &rbio->work); | |
53b381b3 DW |
1491 | } |
1492 | ||
1493 | /* | |
1494 | * the stripe must be locked by the caller. It will | |
1495 | * unlock after all the writes are done | |
1496 | */ | |
1497 | static int raid56_rmw_stripe(struct btrfs_raid_bio *rbio) | |
1498 | { | |
1499 | int bios_to_read = 0; | |
53b381b3 DW |
1500 | struct bio_list bio_list; |
1501 | int ret; | |
ed6078f7 | 1502 | int nr_pages = DIV_ROUND_UP(rbio->stripe_len, PAGE_CACHE_SIZE); |
53b381b3 DW |
1503 | int pagenr; |
1504 | int stripe; | |
1505 | struct bio *bio; | |
1506 | ||
1507 | bio_list_init(&bio_list); | |
1508 | ||
1509 | ret = alloc_rbio_pages(rbio); | |
1510 | if (ret) | |
1511 | goto cleanup; | |
1512 | ||
1513 | index_rbio_pages(rbio); | |
1514 | ||
b89e1b01 | 1515 | atomic_set(&rbio->error, 0); |
53b381b3 DW |
1516 | /* |
1517 | * build a list of bios to read all the missing parts of this | |
1518 | * stripe | |
1519 | */ | |
1520 | for (stripe = 0; stripe < rbio->nr_data; stripe++) { | |
1521 | for (pagenr = 0; pagenr < nr_pages; pagenr++) { | |
1522 | struct page *page; | |
1523 | /* | |
1524 | * we want to find all the pages missing from | |
1525 | * the rbio and read them from the disk. If | |
1526 | * page_in_rbio finds a page in the bio list | |
1527 | * we don't need to read it off the stripe. | |
1528 | */ | |
1529 | page = page_in_rbio(rbio, stripe, pagenr, 1); | |
1530 | if (page) | |
1531 | continue; | |
1532 | ||
1533 | page = rbio_stripe_page(rbio, stripe, pagenr); | |
4ae10b3a CM |
1534 | /* |
1535 | * the bio cache may have handed us an uptodate | |
1536 | * page. If so, be happy and use it | |
1537 | */ | |
1538 | if (PageUptodate(page)) | |
1539 | continue; | |
1540 | ||
53b381b3 DW |
1541 | ret = rbio_add_io_page(rbio, &bio_list, page, |
1542 | stripe, pagenr, rbio->stripe_len); | |
1543 | if (ret) | |
1544 | goto cleanup; | |
1545 | } | |
1546 | } | |
1547 | ||
1548 | bios_to_read = bio_list_size(&bio_list); | |
1549 | if (!bios_to_read) { | |
1550 | /* | |
1551 | * this can happen if others have merged with | |
1552 | * us, it means there is nothing left to read. | |
1553 | * But if there are missing devices it may not be | |
1554 | * safe to do the full stripe write yet. | |
1555 | */ | |
1556 | goto finish; | |
1557 | } | |
1558 | ||
1559 | /* | |
1560 | * the bbio may be freed once we submit the last bio. Make sure | |
1561 | * not to touch it after that | |
1562 | */ | |
b89e1b01 | 1563 | atomic_set(&rbio->stripes_pending, bios_to_read); |
53b381b3 DW |
1564 | while (1) { |
1565 | bio = bio_list_pop(&bio_list); | |
1566 | if (!bio) | |
1567 | break; | |
1568 | ||
1569 | bio->bi_private = rbio; | |
1570 | bio->bi_end_io = raid_rmw_end_io; | |
1571 | ||
1572 | btrfs_bio_wq_end_io(rbio->fs_info, bio, | |
1573 | BTRFS_WQ_ENDIO_RAID56); | |
1574 | ||
1575 | BUG_ON(!test_bit(BIO_UPTODATE, &bio->bi_flags)); | |
1576 | submit_bio(READ, bio); | |
1577 | } | |
1578 | /* the actual write will happen once the reads are done */ | |
1579 | return 0; | |
1580 | ||
1581 | cleanup: | |
1582 | rbio_orig_end_io(rbio, -EIO, 0); | |
1583 | return -EIO; | |
1584 | ||
1585 | finish: | |
1586 | validate_rbio_for_rmw(rbio); | |
1587 | return 0; | |
1588 | } | |
1589 | ||
1590 | /* | |
1591 | * if the upper layers pass in a full stripe, we thank them by only allocating | |
1592 | * enough pages to hold the parity, and sending it all down quickly. | |
1593 | */ | |
1594 | static int full_stripe_write(struct btrfs_raid_bio *rbio) | |
1595 | { | |
1596 | int ret; | |
1597 | ||
1598 | ret = alloc_rbio_parity_pages(rbio); | |
3cd846d1 MX |
1599 | if (ret) { |
1600 | __free_raid_bio(rbio); | |
53b381b3 | 1601 | return ret; |
3cd846d1 | 1602 | } |
53b381b3 DW |
1603 | |
1604 | ret = lock_stripe_add(rbio); | |
1605 | if (ret == 0) | |
1606 | finish_rmw(rbio); | |
1607 | return 0; | |
1608 | } | |
1609 | ||
1610 | /* | |
1611 | * partial stripe writes get handed over to async helpers. | |
1612 | * We're really hoping to merge a few more writes into this | |
1613 | * rbio before calculating new parity | |
1614 | */ | |
1615 | static int partial_stripe_write(struct btrfs_raid_bio *rbio) | |
1616 | { | |
1617 | int ret; | |
1618 | ||
1619 | ret = lock_stripe_add(rbio); | |
1620 | if (ret == 0) | |
1621 | async_rmw_stripe(rbio); | |
1622 | return 0; | |
1623 | } | |
1624 | ||
1625 | /* | |
1626 | * sometimes while we were reading from the drive to | |
1627 | * recalculate parity, enough new bios come into create | |
1628 | * a full stripe. So we do a check here to see if we can | |
1629 | * go directly to finish_rmw | |
1630 | */ | |
1631 | static int __raid56_parity_write(struct btrfs_raid_bio *rbio) | |
1632 | { | |
1633 | /* head off into rmw land if we don't have a full stripe */ | |
1634 | if (!rbio_is_full(rbio)) | |
1635 | return partial_stripe_write(rbio); | |
1636 | return full_stripe_write(rbio); | |
1637 | } | |
1638 | ||
6ac0f488 CM |
1639 | /* |
1640 | * We use plugging call backs to collect full stripes. | |
1641 | * Any time we get a partial stripe write while plugged | |
1642 | * we collect it into a list. When the unplug comes down, | |
1643 | * we sort the list by logical block number and merge | |
1644 | * everything we can into the same rbios | |
1645 | */ | |
1646 | struct btrfs_plug_cb { | |
1647 | struct blk_plug_cb cb; | |
1648 | struct btrfs_fs_info *info; | |
1649 | struct list_head rbio_list; | |
1650 | struct btrfs_work work; | |
1651 | }; | |
1652 | ||
1653 | /* | |
1654 | * rbios on the plug list are sorted for easier merging. | |
1655 | */ | |
1656 | static int plug_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1657 | { | |
1658 | struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio, | |
1659 | plug_list); | |
1660 | struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio, | |
1661 | plug_list); | |
4f024f37 KO |
1662 | u64 a_sector = ra->bio_list.head->bi_iter.bi_sector; |
1663 | u64 b_sector = rb->bio_list.head->bi_iter.bi_sector; | |
6ac0f488 CM |
1664 | |
1665 | if (a_sector < b_sector) | |
1666 | return -1; | |
1667 | if (a_sector > b_sector) | |
1668 | return 1; | |
1669 | return 0; | |
1670 | } | |
1671 | ||
1672 | static void run_plug(struct btrfs_plug_cb *plug) | |
1673 | { | |
1674 | struct btrfs_raid_bio *cur; | |
1675 | struct btrfs_raid_bio *last = NULL; | |
1676 | ||
1677 | /* | |
1678 | * sort our plug list then try to merge | |
1679 | * everything we can in hopes of creating full | |
1680 | * stripes. | |
1681 | */ | |
1682 | list_sort(NULL, &plug->rbio_list, plug_cmp); | |
1683 | while (!list_empty(&plug->rbio_list)) { | |
1684 | cur = list_entry(plug->rbio_list.next, | |
1685 | struct btrfs_raid_bio, plug_list); | |
1686 | list_del_init(&cur->plug_list); | |
1687 | ||
1688 | if (rbio_is_full(cur)) { | |
1689 | /* we have a full stripe, send it down */ | |
1690 | full_stripe_write(cur); | |
1691 | continue; | |
1692 | } | |
1693 | if (last) { | |
1694 | if (rbio_can_merge(last, cur)) { | |
1695 | merge_rbio(last, cur); | |
1696 | __free_raid_bio(cur); | |
1697 | continue; | |
1698 | ||
1699 | } | |
1700 | __raid56_parity_write(last); | |
1701 | } | |
1702 | last = cur; | |
1703 | } | |
1704 | if (last) { | |
1705 | __raid56_parity_write(last); | |
1706 | } | |
1707 | kfree(plug); | |
1708 | } | |
1709 | ||
1710 | /* | |
1711 | * if the unplug comes from schedule, we have to push the | |
1712 | * work off to a helper thread | |
1713 | */ | |
1714 | static void unplug_work(struct btrfs_work *work) | |
1715 | { | |
1716 | struct btrfs_plug_cb *plug; | |
1717 | plug = container_of(work, struct btrfs_plug_cb, work); | |
1718 | run_plug(plug); | |
1719 | } | |
1720 | ||
1721 | static void btrfs_raid_unplug(struct blk_plug_cb *cb, bool from_schedule) | |
1722 | { | |
1723 | struct btrfs_plug_cb *plug; | |
1724 | plug = container_of(cb, struct btrfs_plug_cb, cb); | |
1725 | ||
1726 | if (from_schedule) { | |
9e0af237 LB |
1727 | btrfs_init_work(&plug->work, btrfs_rmw_helper, |
1728 | unplug_work, NULL, NULL); | |
d05a33ac QW |
1729 | btrfs_queue_work(plug->info->rmw_workers, |
1730 | &plug->work); | |
6ac0f488 CM |
1731 | return; |
1732 | } | |
1733 | run_plug(plug); | |
1734 | } | |
1735 | ||
53b381b3 DW |
1736 | /* |
1737 | * our main entry point for writes from the rest of the FS. | |
1738 | */ | |
1739 | int raid56_parity_write(struct btrfs_root *root, struct bio *bio, | |
8e5cfb55 | 1740 | struct btrfs_bio *bbio, u64 stripe_len) |
53b381b3 DW |
1741 | { |
1742 | struct btrfs_raid_bio *rbio; | |
6ac0f488 CM |
1743 | struct btrfs_plug_cb *plug = NULL; |
1744 | struct blk_plug_cb *cb; | |
4245215d | 1745 | int ret; |
53b381b3 | 1746 | |
8e5cfb55 | 1747 | rbio = alloc_rbio(root, bbio, stripe_len); |
af8e2d1d | 1748 | if (IS_ERR(rbio)) { |
6e9606d2 | 1749 | btrfs_put_bbio(bbio); |
53b381b3 | 1750 | return PTR_ERR(rbio); |
af8e2d1d | 1751 | } |
53b381b3 | 1752 | bio_list_add(&rbio->bio_list, bio); |
4f024f37 | 1753 | rbio->bio_list_bytes = bio->bi_iter.bi_size; |
1b94b556 | 1754 | rbio->operation = BTRFS_RBIO_WRITE; |
6ac0f488 | 1755 | |
4245215d MX |
1756 | btrfs_bio_counter_inc_noblocked(root->fs_info); |
1757 | rbio->generic_bio_cnt = 1; | |
1758 | ||
6ac0f488 CM |
1759 | /* |
1760 | * don't plug on full rbios, just get them out the door | |
1761 | * as quickly as we can | |
1762 | */ | |
4245215d MX |
1763 | if (rbio_is_full(rbio)) { |
1764 | ret = full_stripe_write(rbio); | |
1765 | if (ret) | |
1766 | btrfs_bio_counter_dec(root->fs_info); | |
1767 | return ret; | |
1768 | } | |
6ac0f488 CM |
1769 | |
1770 | cb = blk_check_plugged(btrfs_raid_unplug, root->fs_info, | |
1771 | sizeof(*plug)); | |
1772 | if (cb) { | |
1773 | plug = container_of(cb, struct btrfs_plug_cb, cb); | |
1774 | if (!plug->info) { | |
1775 | plug->info = root->fs_info; | |
1776 | INIT_LIST_HEAD(&plug->rbio_list); | |
1777 | } | |
1778 | list_add_tail(&rbio->plug_list, &plug->rbio_list); | |
4245215d | 1779 | ret = 0; |
6ac0f488 | 1780 | } else { |
4245215d MX |
1781 | ret = __raid56_parity_write(rbio); |
1782 | if (ret) | |
1783 | btrfs_bio_counter_dec(root->fs_info); | |
6ac0f488 | 1784 | } |
4245215d | 1785 | return ret; |
53b381b3 DW |
1786 | } |
1787 | ||
1788 | /* | |
1789 | * all parity reconstruction happens here. We've read in everything | |
1790 | * we can find from the drives and this does the heavy lifting of | |
1791 | * sorting the good from the bad. | |
1792 | */ | |
1793 | static void __raid_recover_end_io(struct btrfs_raid_bio *rbio) | |
1794 | { | |
1795 | int pagenr, stripe; | |
1796 | void **pointers; | |
1797 | int faila = -1, failb = -1; | |
ed6078f7 | 1798 | int nr_pages = DIV_ROUND_UP(rbio->stripe_len, PAGE_CACHE_SIZE); |
53b381b3 DW |
1799 | struct page *page; |
1800 | int err; | |
1801 | int i; | |
1802 | ||
31e818fe | 1803 | pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS); |
53b381b3 DW |
1804 | if (!pointers) { |
1805 | err = -ENOMEM; | |
1806 | goto cleanup_io; | |
1807 | } | |
1808 | ||
1809 | faila = rbio->faila; | |
1810 | failb = rbio->failb; | |
1811 | ||
1b94b556 | 1812 | if (rbio->operation == BTRFS_RBIO_READ_REBUILD) { |
53b381b3 DW |
1813 | spin_lock_irq(&rbio->bio_list_lock); |
1814 | set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags); | |
1815 | spin_unlock_irq(&rbio->bio_list_lock); | |
1816 | } | |
1817 | ||
1818 | index_rbio_pages(rbio); | |
1819 | ||
1820 | for (pagenr = 0; pagenr < nr_pages; pagenr++) { | |
5a6ac9ea MX |
1821 | /* |
1822 | * Now we just use bitmap to mark the horizontal stripes in | |
1823 | * which we have data when doing parity scrub. | |
1824 | */ | |
1825 | if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB && | |
1826 | !test_bit(pagenr, rbio->dbitmap)) | |
1827 | continue; | |
1828 | ||
53b381b3 DW |
1829 | /* setup our array of pointers with pages |
1830 | * from each stripe | |
1831 | */ | |
2c8cdd6e | 1832 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) { |
53b381b3 DW |
1833 | /* |
1834 | * if we're rebuilding a read, we have to use | |
1835 | * pages from the bio list | |
1836 | */ | |
1b94b556 | 1837 | if (rbio->operation == BTRFS_RBIO_READ_REBUILD && |
53b381b3 DW |
1838 | (stripe == faila || stripe == failb)) { |
1839 | page = page_in_rbio(rbio, stripe, pagenr, 0); | |
1840 | } else { | |
1841 | page = rbio_stripe_page(rbio, stripe, pagenr); | |
1842 | } | |
1843 | pointers[stripe] = kmap(page); | |
1844 | } | |
1845 | ||
1846 | /* all raid6 handling here */ | |
10f11900 | 1847 | if (rbio->bbio->map_type & BTRFS_BLOCK_GROUP_RAID6) { |
53b381b3 DW |
1848 | /* |
1849 | * single failure, rebuild from parity raid5 | |
1850 | * style | |
1851 | */ | |
1852 | if (failb < 0) { | |
1853 | if (faila == rbio->nr_data) { | |
1854 | /* | |
1855 | * Just the P stripe has failed, without | |
1856 | * a bad data or Q stripe. | |
1857 | * TODO, we should redo the xor here. | |
1858 | */ | |
1859 | err = -EIO; | |
1860 | goto cleanup; | |
1861 | } | |
1862 | /* | |
1863 | * a single failure in raid6 is rebuilt | |
1864 | * in the pstripe code below | |
1865 | */ | |
1866 | goto pstripe; | |
1867 | } | |
1868 | ||
1869 | /* make sure our ps and qs are in order */ | |
1870 | if (faila > failb) { | |
1871 | int tmp = failb; | |
1872 | failb = faila; | |
1873 | faila = tmp; | |
1874 | } | |
1875 | ||
1876 | /* if the q stripe is failed, do a pstripe reconstruction | |
1877 | * from the xors. | |
1878 | * If both the q stripe and the P stripe are failed, we're | |
1879 | * here due to a crc mismatch and we can't give them the | |
1880 | * data they want | |
1881 | */ | |
8e5cfb55 ZL |
1882 | if (rbio->bbio->raid_map[failb] == RAID6_Q_STRIPE) { |
1883 | if (rbio->bbio->raid_map[faila] == | |
1884 | RAID5_P_STRIPE) { | |
53b381b3 DW |
1885 | err = -EIO; |
1886 | goto cleanup; | |
1887 | } | |
1888 | /* | |
1889 | * otherwise we have one bad data stripe and | |
1890 | * a good P stripe. raid5! | |
1891 | */ | |
1892 | goto pstripe; | |
1893 | } | |
1894 | ||
8e5cfb55 | 1895 | if (rbio->bbio->raid_map[failb] == RAID5_P_STRIPE) { |
2c8cdd6e | 1896 | raid6_datap_recov(rbio->real_stripes, |
53b381b3 DW |
1897 | PAGE_SIZE, faila, pointers); |
1898 | } else { | |
2c8cdd6e | 1899 | raid6_2data_recov(rbio->real_stripes, |
53b381b3 DW |
1900 | PAGE_SIZE, faila, failb, |
1901 | pointers); | |
1902 | } | |
1903 | } else { | |
1904 | void *p; | |
1905 | ||
1906 | /* rebuild from P stripe here (raid5 or raid6) */ | |
1907 | BUG_ON(failb != -1); | |
1908 | pstripe: | |
1909 | /* Copy parity block into failed block to start with */ | |
1910 | memcpy(pointers[faila], | |
1911 | pointers[rbio->nr_data], | |
1912 | PAGE_CACHE_SIZE); | |
1913 | ||
1914 | /* rearrange the pointer array */ | |
1915 | p = pointers[faila]; | |
1916 | for (stripe = faila; stripe < rbio->nr_data - 1; stripe++) | |
1917 | pointers[stripe] = pointers[stripe + 1]; | |
1918 | pointers[rbio->nr_data - 1] = p; | |
1919 | ||
1920 | /* xor in the rest */ | |
1921 | run_xor(pointers, rbio->nr_data - 1, PAGE_CACHE_SIZE); | |
1922 | } | |
1923 | /* if we're doing this rebuild as part of an rmw, go through | |
1924 | * and set all of our private rbio pages in the | |
1925 | * failed stripes as uptodate. This way finish_rmw will | |
1926 | * know they can be trusted. If this was a read reconstruction, | |
1927 | * other endio functions will fiddle the uptodate bits | |
1928 | */ | |
1b94b556 | 1929 | if (rbio->operation == BTRFS_RBIO_WRITE) { |
53b381b3 DW |
1930 | for (i = 0; i < nr_pages; i++) { |
1931 | if (faila != -1) { | |
1932 | page = rbio_stripe_page(rbio, faila, i); | |
1933 | SetPageUptodate(page); | |
1934 | } | |
1935 | if (failb != -1) { | |
1936 | page = rbio_stripe_page(rbio, failb, i); | |
1937 | SetPageUptodate(page); | |
1938 | } | |
1939 | } | |
1940 | } | |
2c8cdd6e | 1941 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) { |
53b381b3 DW |
1942 | /* |
1943 | * if we're rebuilding a read, we have to use | |
1944 | * pages from the bio list | |
1945 | */ | |
1b94b556 | 1946 | if (rbio->operation == BTRFS_RBIO_READ_REBUILD && |
53b381b3 DW |
1947 | (stripe == faila || stripe == failb)) { |
1948 | page = page_in_rbio(rbio, stripe, pagenr, 0); | |
1949 | } else { | |
1950 | page = rbio_stripe_page(rbio, stripe, pagenr); | |
1951 | } | |
1952 | kunmap(page); | |
1953 | } | |
1954 | } | |
1955 | ||
1956 | err = 0; | |
1957 | cleanup: | |
1958 | kfree(pointers); | |
1959 | ||
1960 | cleanup_io: | |
1b94b556 | 1961 | if (rbio->operation == BTRFS_RBIO_READ_REBUILD) { |
6e9606d2 | 1962 | if (err == 0) |
4ae10b3a CM |
1963 | cache_rbio_pages(rbio); |
1964 | else | |
1965 | clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags); | |
1966 | ||
53b381b3 DW |
1967 | rbio_orig_end_io(rbio, err, err == 0); |
1968 | } else if (err == 0) { | |
1969 | rbio->faila = -1; | |
1970 | rbio->failb = -1; | |
5a6ac9ea MX |
1971 | |
1972 | if (rbio->operation == BTRFS_RBIO_WRITE) | |
1973 | finish_rmw(rbio); | |
1974 | else if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB) | |
1975 | finish_parity_scrub(rbio, 0); | |
1976 | else | |
1977 | BUG(); | |
53b381b3 DW |
1978 | } else { |
1979 | rbio_orig_end_io(rbio, err, 0); | |
1980 | } | |
1981 | } | |
1982 | ||
1983 | /* | |
1984 | * This is called only for stripes we've read from disk to | |
1985 | * reconstruct the parity. | |
1986 | */ | |
1987 | static void raid_recover_end_io(struct bio *bio, int err) | |
1988 | { | |
1989 | struct btrfs_raid_bio *rbio = bio->bi_private; | |
1990 | ||
1991 | /* | |
1992 | * we only read stripe pages off the disk, set them | |
1993 | * up to date if there were no errors | |
1994 | */ | |
1995 | if (err) | |
1996 | fail_bio_stripe(rbio, bio); | |
1997 | else | |
1998 | set_bio_pages_uptodate(bio); | |
1999 | bio_put(bio); | |
2000 | ||
b89e1b01 | 2001 | if (!atomic_dec_and_test(&rbio->stripes_pending)) |
53b381b3 DW |
2002 | return; |
2003 | ||
b89e1b01 | 2004 | if (atomic_read(&rbio->error) > rbio->bbio->max_errors) |
53b381b3 DW |
2005 | rbio_orig_end_io(rbio, -EIO, 0); |
2006 | else | |
2007 | __raid_recover_end_io(rbio); | |
2008 | } | |
2009 | ||
2010 | /* | |
2011 | * reads everything we need off the disk to reconstruct | |
2012 | * the parity. endio handlers trigger final reconstruction | |
2013 | * when the IO is done. | |
2014 | * | |
2015 | * This is used both for reads from the higher layers and for | |
2016 | * parity construction required to finish a rmw cycle. | |
2017 | */ | |
2018 | static int __raid56_parity_recover(struct btrfs_raid_bio *rbio) | |
2019 | { | |
2020 | int bios_to_read = 0; | |
53b381b3 DW |
2021 | struct bio_list bio_list; |
2022 | int ret; | |
ed6078f7 | 2023 | int nr_pages = DIV_ROUND_UP(rbio->stripe_len, PAGE_CACHE_SIZE); |
53b381b3 DW |
2024 | int pagenr; |
2025 | int stripe; | |
2026 | struct bio *bio; | |
2027 | ||
2028 | bio_list_init(&bio_list); | |
2029 | ||
2030 | ret = alloc_rbio_pages(rbio); | |
2031 | if (ret) | |
2032 | goto cleanup; | |
2033 | ||
b89e1b01 | 2034 | atomic_set(&rbio->error, 0); |
53b381b3 DW |
2035 | |
2036 | /* | |
4ae10b3a CM |
2037 | * read everything that hasn't failed. Thanks to the |
2038 | * stripe cache, it is possible that some or all of these | |
2039 | * pages are going to be uptodate. | |
53b381b3 | 2040 | */ |
2c8cdd6e | 2041 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) { |
5588383e | 2042 | if (rbio->faila == stripe || rbio->failb == stripe) { |
b89e1b01 | 2043 | atomic_inc(&rbio->error); |
53b381b3 | 2044 | continue; |
5588383e | 2045 | } |
53b381b3 DW |
2046 | |
2047 | for (pagenr = 0; pagenr < nr_pages; pagenr++) { | |
2048 | struct page *p; | |
2049 | ||
2050 | /* | |
2051 | * the rmw code may have already read this | |
2052 | * page in | |
2053 | */ | |
2054 | p = rbio_stripe_page(rbio, stripe, pagenr); | |
2055 | if (PageUptodate(p)) | |
2056 | continue; | |
2057 | ||
2058 | ret = rbio_add_io_page(rbio, &bio_list, | |
2059 | rbio_stripe_page(rbio, stripe, pagenr), | |
2060 | stripe, pagenr, rbio->stripe_len); | |
2061 | if (ret < 0) | |
2062 | goto cleanup; | |
2063 | } | |
2064 | } | |
2065 | ||
2066 | bios_to_read = bio_list_size(&bio_list); | |
2067 | if (!bios_to_read) { | |
2068 | /* | |
2069 | * we might have no bios to read just because the pages | |
2070 | * were up to date, or we might have no bios to read because | |
2071 | * the devices were gone. | |
2072 | */ | |
b89e1b01 | 2073 | if (atomic_read(&rbio->error) <= rbio->bbio->max_errors) { |
53b381b3 DW |
2074 | __raid_recover_end_io(rbio); |
2075 | goto out; | |
2076 | } else { | |
2077 | goto cleanup; | |
2078 | } | |
2079 | } | |
2080 | ||
2081 | /* | |
2082 | * the bbio may be freed once we submit the last bio. Make sure | |
2083 | * not to touch it after that | |
2084 | */ | |
b89e1b01 | 2085 | atomic_set(&rbio->stripes_pending, bios_to_read); |
53b381b3 DW |
2086 | while (1) { |
2087 | bio = bio_list_pop(&bio_list); | |
2088 | if (!bio) | |
2089 | break; | |
2090 | ||
2091 | bio->bi_private = rbio; | |
2092 | bio->bi_end_io = raid_recover_end_io; | |
2093 | ||
2094 | btrfs_bio_wq_end_io(rbio->fs_info, bio, | |
2095 | BTRFS_WQ_ENDIO_RAID56); | |
2096 | ||
2097 | BUG_ON(!test_bit(BIO_UPTODATE, &bio->bi_flags)); | |
2098 | submit_bio(READ, bio); | |
2099 | } | |
2100 | out: | |
2101 | return 0; | |
2102 | ||
2103 | cleanup: | |
1b94b556 | 2104 | if (rbio->operation == BTRFS_RBIO_READ_REBUILD) |
53b381b3 DW |
2105 | rbio_orig_end_io(rbio, -EIO, 0); |
2106 | return -EIO; | |
2107 | } | |
2108 | ||
2109 | /* | |
2110 | * the main entry point for reads from the higher layers. This | |
2111 | * is really only called when the normal read path had a failure, | |
2112 | * so we assume the bio they send down corresponds to a failed part | |
2113 | * of the drive. | |
2114 | */ | |
2115 | int raid56_parity_recover(struct btrfs_root *root, struct bio *bio, | |
8e5cfb55 ZL |
2116 | struct btrfs_bio *bbio, u64 stripe_len, |
2117 | int mirror_num, int generic_io) | |
53b381b3 DW |
2118 | { |
2119 | struct btrfs_raid_bio *rbio; | |
2120 | int ret; | |
2121 | ||
8e5cfb55 | 2122 | rbio = alloc_rbio(root, bbio, stripe_len); |
af8e2d1d | 2123 | if (IS_ERR(rbio)) { |
6e9606d2 ZL |
2124 | if (generic_io) |
2125 | btrfs_put_bbio(bbio); | |
53b381b3 | 2126 | return PTR_ERR(rbio); |
af8e2d1d | 2127 | } |
53b381b3 | 2128 | |
1b94b556 | 2129 | rbio->operation = BTRFS_RBIO_READ_REBUILD; |
53b381b3 | 2130 | bio_list_add(&rbio->bio_list, bio); |
4f024f37 | 2131 | rbio->bio_list_bytes = bio->bi_iter.bi_size; |
53b381b3 DW |
2132 | |
2133 | rbio->faila = find_logical_bio_stripe(rbio, bio); | |
2134 | if (rbio->faila == -1) { | |
2135 | BUG(); | |
6e9606d2 ZL |
2136 | if (generic_io) |
2137 | btrfs_put_bbio(bbio); | |
53b381b3 DW |
2138 | kfree(rbio); |
2139 | return -EIO; | |
2140 | } | |
2141 | ||
4245215d MX |
2142 | if (generic_io) { |
2143 | btrfs_bio_counter_inc_noblocked(root->fs_info); | |
2144 | rbio->generic_bio_cnt = 1; | |
2145 | } else { | |
6e9606d2 | 2146 | btrfs_get_bbio(bbio); |
4245215d MX |
2147 | } |
2148 | ||
53b381b3 DW |
2149 | /* |
2150 | * reconstruct from the q stripe if they are | |
2151 | * asking for mirror 3 | |
2152 | */ | |
2153 | if (mirror_num == 3) | |
2c8cdd6e | 2154 | rbio->failb = rbio->real_stripes - 2; |
53b381b3 DW |
2155 | |
2156 | ret = lock_stripe_add(rbio); | |
2157 | ||
2158 | /* | |
2159 | * __raid56_parity_recover will end the bio with | |
2160 | * any errors it hits. We don't want to return | |
2161 | * its error value up the stack because our caller | |
2162 | * will end up calling bio_endio with any nonzero | |
2163 | * return | |
2164 | */ | |
2165 | if (ret == 0) | |
2166 | __raid56_parity_recover(rbio); | |
2167 | /* | |
2168 | * our rbio has been added to the list of | |
2169 | * rbios that will be handled after the | |
2170 | * currently lock owner is done | |
2171 | */ | |
2172 | return 0; | |
2173 | ||
2174 | } | |
2175 | ||
2176 | static void rmw_work(struct btrfs_work *work) | |
2177 | { | |
2178 | struct btrfs_raid_bio *rbio; | |
2179 | ||
2180 | rbio = container_of(work, struct btrfs_raid_bio, work); | |
2181 | raid56_rmw_stripe(rbio); | |
2182 | } | |
2183 | ||
2184 | static void read_rebuild_work(struct btrfs_work *work) | |
2185 | { | |
2186 | struct btrfs_raid_bio *rbio; | |
2187 | ||
2188 | rbio = container_of(work, struct btrfs_raid_bio, work); | |
2189 | __raid56_parity_recover(rbio); | |
2190 | } | |
5a6ac9ea MX |
2191 | |
2192 | /* | |
2193 | * The following code is used to scrub/replace the parity stripe | |
2194 | * | |
2195 | * Note: We need make sure all the pages that add into the scrub/replace | |
2196 | * raid bio are correct and not be changed during the scrub/replace. That | |
2197 | * is those pages just hold metadata or file data with checksum. | |
2198 | */ | |
2199 | ||
2200 | struct btrfs_raid_bio * | |
2201 | raid56_parity_alloc_scrub_rbio(struct btrfs_root *root, struct bio *bio, | |
8e5cfb55 ZL |
2202 | struct btrfs_bio *bbio, u64 stripe_len, |
2203 | struct btrfs_device *scrub_dev, | |
5a6ac9ea MX |
2204 | unsigned long *dbitmap, int stripe_nsectors) |
2205 | { | |
2206 | struct btrfs_raid_bio *rbio; | |
2207 | int i; | |
2208 | ||
8e5cfb55 | 2209 | rbio = alloc_rbio(root, bbio, stripe_len); |
5a6ac9ea MX |
2210 | if (IS_ERR(rbio)) |
2211 | return NULL; | |
2212 | bio_list_add(&rbio->bio_list, bio); | |
2213 | /* | |
2214 | * This is a special bio which is used to hold the completion handler | |
2215 | * and make the scrub rbio is similar to the other types | |
2216 | */ | |
2217 | ASSERT(!bio->bi_iter.bi_size); | |
2218 | rbio->operation = BTRFS_RBIO_PARITY_SCRUB; | |
2219 | ||
2c8cdd6e | 2220 | for (i = 0; i < rbio->real_stripes; i++) { |
5a6ac9ea MX |
2221 | if (bbio->stripes[i].dev == scrub_dev) { |
2222 | rbio->scrubp = i; | |
2223 | break; | |
2224 | } | |
2225 | } | |
2226 | ||
2227 | /* Now we just support the sectorsize equals to page size */ | |
2228 | ASSERT(root->sectorsize == PAGE_SIZE); | |
2229 | ASSERT(rbio->stripe_npages == stripe_nsectors); | |
2230 | bitmap_copy(rbio->dbitmap, dbitmap, stripe_nsectors); | |
2231 | ||
2232 | return rbio; | |
2233 | } | |
2234 | ||
2235 | void raid56_parity_add_scrub_pages(struct btrfs_raid_bio *rbio, | |
2236 | struct page *page, u64 logical) | |
2237 | { | |
2238 | int stripe_offset; | |
2239 | int index; | |
2240 | ||
8e5cfb55 ZL |
2241 | ASSERT(logical >= rbio->bbio->raid_map[0]); |
2242 | ASSERT(logical + PAGE_SIZE <= rbio->bbio->raid_map[0] + | |
5a6ac9ea | 2243 | rbio->stripe_len * rbio->nr_data); |
8e5cfb55 | 2244 | stripe_offset = (int)(logical - rbio->bbio->raid_map[0]); |
5a6ac9ea MX |
2245 | index = stripe_offset >> PAGE_CACHE_SHIFT; |
2246 | rbio->bio_pages[index] = page; | |
2247 | } | |
2248 | ||
2249 | /* | |
2250 | * We just scrub the parity that we have correct data on the same horizontal, | |
2251 | * so we needn't allocate all pages for all the stripes. | |
2252 | */ | |
2253 | static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio) | |
2254 | { | |
2255 | int i; | |
2256 | int bit; | |
2257 | int index; | |
2258 | struct page *page; | |
2259 | ||
2260 | for_each_set_bit(bit, rbio->dbitmap, rbio->stripe_npages) { | |
2c8cdd6e | 2261 | for (i = 0; i < rbio->real_stripes; i++) { |
5a6ac9ea MX |
2262 | index = i * rbio->stripe_npages + bit; |
2263 | if (rbio->stripe_pages[index]) | |
2264 | continue; | |
2265 | ||
2266 | page = alloc_page(GFP_NOFS | __GFP_HIGHMEM); | |
2267 | if (!page) | |
2268 | return -ENOMEM; | |
2269 | rbio->stripe_pages[index] = page; | |
2270 | ClearPageUptodate(page); | |
2271 | } | |
2272 | } | |
2273 | return 0; | |
2274 | } | |
2275 | ||
2276 | /* | |
2277 | * end io function used by finish_rmw. When we finally | |
2278 | * get here, we've written a full stripe | |
2279 | */ | |
2280 | static void raid_write_parity_end_io(struct bio *bio, int err) | |
2281 | { | |
2282 | struct btrfs_raid_bio *rbio = bio->bi_private; | |
2283 | ||
2284 | if (err) | |
2285 | fail_bio_stripe(rbio, bio); | |
2286 | ||
2287 | bio_put(bio); | |
2288 | ||
2289 | if (!atomic_dec_and_test(&rbio->stripes_pending)) | |
2290 | return; | |
2291 | ||
2292 | err = 0; | |
2293 | ||
2294 | if (atomic_read(&rbio->error)) | |
2295 | err = -EIO; | |
2296 | ||
2297 | rbio_orig_end_io(rbio, err, 0); | |
2298 | } | |
2299 | ||
2300 | static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio, | |
2301 | int need_check) | |
2302 | { | |
76035976 | 2303 | struct btrfs_bio *bbio = rbio->bbio; |
2c8cdd6e | 2304 | void *pointers[rbio->real_stripes]; |
76035976 | 2305 | DECLARE_BITMAP(pbitmap, rbio->stripe_npages); |
5a6ac9ea MX |
2306 | int nr_data = rbio->nr_data; |
2307 | int stripe; | |
2308 | int pagenr; | |
2309 | int p_stripe = -1; | |
2310 | int q_stripe = -1; | |
2311 | struct page *p_page = NULL; | |
2312 | struct page *q_page = NULL; | |
2313 | struct bio_list bio_list; | |
2314 | struct bio *bio; | |
76035976 | 2315 | int is_replace = 0; |
5a6ac9ea MX |
2316 | int ret; |
2317 | ||
2318 | bio_list_init(&bio_list); | |
2319 | ||
2c8cdd6e MX |
2320 | if (rbio->real_stripes - rbio->nr_data == 1) { |
2321 | p_stripe = rbio->real_stripes - 1; | |
2322 | } else if (rbio->real_stripes - rbio->nr_data == 2) { | |
2323 | p_stripe = rbio->real_stripes - 2; | |
2324 | q_stripe = rbio->real_stripes - 1; | |
5a6ac9ea MX |
2325 | } else { |
2326 | BUG(); | |
2327 | } | |
2328 | ||
76035976 MX |
2329 | if (bbio->num_tgtdevs && bbio->tgtdev_map[rbio->scrubp]) { |
2330 | is_replace = 1; | |
2331 | bitmap_copy(pbitmap, rbio->dbitmap, rbio->stripe_npages); | |
2332 | } | |
2333 | ||
5a6ac9ea MX |
2334 | /* |
2335 | * Because the higher layers(scrubber) are unlikely to | |
2336 | * use this area of the disk again soon, so don't cache | |
2337 | * it. | |
2338 | */ | |
2339 | clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags); | |
2340 | ||
2341 | if (!need_check) | |
2342 | goto writeback; | |
2343 | ||
2344 | p_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM); | |
2345 | if (!p_page) | |
2346 | goto cleanup; | |
2347 | SetPageUptodate(p_page); | |
2348 | ||
2349 | if (q_stripe != -1) { | |
2350 | q_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM); | |
2351 | if (!q_page) { | |
2352 | __free_page(p_page); | |
2353 | goto cleanup; | |
2354 | } | |
2355 | SetPageUptodate(q_page); | |
2356 | } | |
2357 | ||
2358 | atomic_set(&rbio->error, 0); | |
2359 | ||
2360 | for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) { | |
2361 | struct page *p; | |
2362 | void *parity; | |
2363 | /* first collect one page from each data stripe */ | |
2364 | for (stripe = 0; stripe < nr_data; stripe++) { | |
2365 | p = page_in_rbio(rbio, stripe, pagenr, 0); | |
2366 | pointers[stripe] = kmap(p); | |
2367 | } | |
2368 | ||
2369 | /* then add the parity stripe */ | |
2370 | pointers[stripe++] = kmap(p_page); | |
2371 | ||
2372 | if (q_stripe != -1) { | |
2373 | ||
2374 | /* | |
2375 | * raid6, add the qstripe and call the | |
2376 | * library function to fill in our p/q | |
2377 | */ | |
2378 | pointers[stripe++] = kmap(q_page); | |
2379 | ||
2c8cdd6e | 2380 | raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE, |
5a6ac9ea MX |
2381 | pointers); |
2382 | } else { | |
2383 | /* raid5 */ | |
2384 | memcpy(pointers[nr_data], pointers[0], PAGE_SIZE); | |
2385 | run_xor(pointers + 1, nr_data - 1, PAGE_CACHE_SIZE); | |
2386 | } | |
2387 | ||
2388 | /* Check scrubbing pairty and repair it */ | |
2389 | p = rbio_stripe_page(rbio, rbio->scrubp, pagenr); | |
2390 | parity = kmap(p); | |
2391 | if (memcmp(parity, pointers[rbio->scrubp], PAGE_CACHE_SIZE)) | |
2392 | memcpy(parity, pointers[rbio->scrubp], PAGE_CACHE_SIZE); | |
2393 | else | |
2394 | /* Parity is right, needn't writeback */ | |
2395 | bitmap_clear(rbio->dbitmap, pagenr, 1); | |
2396 | kunmap(p); | |
2397 | ||
2c8cdd6e | 2398 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) |
5a6ac9ea MX |
2399 | kunmap(page_in_rbio(rbio, stripe, pagenr, 0)); |
2400 | } | |
2401 | ||
2402 | __free_page(p_page); | |
2403 | if (q_page) | |
2404 | __free_page(q_page); | |
2405 | ||
2406 | writeback: | |
2407 | /* | |
2408 | * time to start writing. Make bios for everything from the | |
2409 | * higher layers (the bio_list in our rbio) and our p/q. Ignore | |
2410 | * everything else. | |
2411 | */ | |
2412 | for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) { | |
2413 | struct page *page; | |
2414 | ||
2415 | page = rbio_stripe_page(rbio, rbio->scrubp, pagenr); | |
2416 | ret = rbio_add_io_page(rbio, &bio_list, | |
2417 | page, rbio->scrubp, pagenr, rbio->stripe_len); | |
2418 | if (ret) | |
2419 | goto cleanup; | |
2420 | } | |
2421 | ||
76035976 MX |
2422 | if (!is_replace) |
2423 | goto submit_write; | |
2424 | ||
2425 | for_each_set_bit(pagenr, pbitmap, rbio->stripe_npages) { | |
2426 | struct page *page; | |
2427 | ||
2428 | page = rbio_stripe_page(rbio, rbio->scrubp, pagenr); | |
2429 | ret = rbio_add_io_page(rbio, &bio_list, page, | |
2430 | bbio->tgtdev_map[rbio->scrubp], | |
2431 | pagenr, rbio->stripe_len); | |
2432 | if (ret) | |
2433 | goto cleanup; | |
2434 | } | |
2435 | ||
2436 | submit_write: | |
5a6ac9ea MX |
2437 | nr_data = bio_list_size(&bio_list); |
2438 | if (!nr_data) { | |
2439 | /* Every parity is right */ | |
2440 | rbio_orig_end_io(rbio, 0, 0); | |
2441 | return; | |
2442 | } | |
2443 | ||
2444 | atomic_set(&rbio->stripes_pending, nr_data); | |
2445 | ||
2446 | while (1) { | |
2447 | bio = bio_list_pop(&bio_list); | |
2448 | if (!bio) | |
2449 | break; | |
2450 | ||
2451 | bio->bi_private = rbio; | |
2452 | bio->bi_end_io = raid_write_parity_end_io; | |
2453 | BUG_ON(!test_bit(BIO_UPTODATE, &bio->bi_flags)); | |
2454 | submit_bio(WRITE, bio); | |
2455 | } | |
2456 | return; | |
2457 | ||
2458 | cleanup: | |
2459 | rbio_orig_end_io(rbio, -EIO, 0); | |
2460 | } | |
2461 | ||
2462 | static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe) | |
2463 | { | |
2464 | if (stripe >= 0 && stripe < rbio->nr_data) | |
2465 | return 1; | |
2466 | return 0; | |
2467 | } | |
2468 | ||
2469 | /* | |
2470 | * While we're doing the parity check and repair, we could have errors | |
2471 | * in reading pages off the disk. This checks for errors and if we're | |
2472 | * not able to read the page it'll trigger parity reconstruction. The | |
2473 | * parity scrub will be finished after we've reconstructed the failed | |
2474 | * stripes | |
2475 | */ | |
2476 | static void validate_rbio_for_parity_scrub(struct btrfs_raid_bio *rbio) | |
2477 | { | |
2478 | if (atomic_read(&rbio->error) > rbio->bbio->max_errors) | |
2479 | goto cleanup; | |
2480 | ||
2481 | if (rbio->faila >= 0 || rbio->failb >= 0) { | |
2482 | int dfail = 0, failp = -1; | |
2483 | ||
2484 | if (is_data_stripe(rbio, rbio->faila)) | |
2485 | dfail++; | |
2486 | else if (is_parity_stripe(rbio->faila)) | |
2487 | failp = rbio->faila; | |
2488 | ||
2489 | if (is_data_stripe(rbio, rbio->failb)) | |
2490 | dfail++; | |
2491 | else if (is_parity_stripe(rbio->failb)) | |
2492 | failp = rbio->failb; | |
2493 | ||
2494 | /* | |
2495 | * Because we can not use a scrubbing parity to repair | |
2496 | * the data, so the capability of the repair is declined. | |
2497 | * (In the case of RAID5, we can not repair anything) | |
2498 | */ | |
2499 | if (dfail > rbio->bbio->max_errors - 1) | |
2500 | goto cleanup; | |
2501 | ||
2502 | /* | |
2503 | * If all data is good, only parity is correctly, just | |
2504 | * repair the parity. | |
2505 | */ | |
2506 | if (dfail == 0) { | |
2507 | finish_parity_scrub(rbio, 0); | |
2508 | return; | |
2509 | } | |
2510 | ||
2511 | /* | |
2512 | * Here means we got one corrupted data stripe and one | |
2513 | * corrupted parity on RAID6, if the corrupted parity | |
2514 | * is scrubbing parity, luckly, use the other one to repair | |
2515 | * the data, or we can not repair the data stripe. | |
2516 | */ | |
2517 | if (failp != rbio->scrubp) | |
2518 | goto cleanup; | |
2519 | ||
2520 | __raid_recover_end_io(rbio); | |
2521 | } else { | |
2522 | finish_parity_scrub(rbio, 1); | |
2523 | } | |
2524 | return; | |
2525 | ||
2526 | cleanup: | |
2527 | rbio_orig_end_io(rbio, -EIO, 0); | |
2528 | } | |
2529 | ||
2530 | /* | |
2531 | * end io for the read phase of the rmw cycle. All the bios here are physical | |
2532 | * stripe bios we've read from the disk so we can recalculate the parity of the | |
2533 | * stripe. | |
2534 | * | |
2535 | * This will usually kick off finish_rmw once all the bios are read in, but it | |
2536 | * may trigger parity reconstruction if we had any errors along the way | |
2537 | */ | |
2538 | static void raid56_parity_scrub_end_io(struct bio *bio, int err) | |
2539 | { | |
2540 | struct btrfs_raid_bio *rbio = bio->bi_private; | |
2541 | ||
2542 | if (err) | |
2543 | fail_bio_stripe(rbio, bio); | |
2544 | else | |
2545 | set_bio_pages_uptodate(bio); | |
2546 | ||
2547 | bio_put(bio); | |
2548 | ||
2549 | if (!atomic_dec_and_test(&rbio->stripes_pending)) | |
2550 | return; | |
2551 | ||
2552 | /* | |
2553 | * this will normally call finish_rmw to start our write | |
2554 | * but if there are any failed stripes we'll reconstruct | |
2555 | * from parity first | |
2556 | */ | |
2557 | validate_rbio_for_parity_scrub(rbio); | |
2558 | } | |
2559 | ||
2560 | static void raid56_parity_scrub_stripe(struct btrfs_raid_bio *rbio) | |
2561 | { | |
2562 | int bios_to_read = 0; | |
5a6ac9ea MX |
2563 | struct bio_list bio_list; |
2564 | int ret; | |
2565 | int pagenr; | |
2566 | int stripe; | |
2567 | struct bio *bio; | |
2568 | ||
2569 | ret = alloc_rbio_essential_pages(rbio); | |
2570 | if (ret) | |
2571 | goto cleanup; | |
2572 | ||
2573 | bio_list_init(&bio_list); | |
2574 | ||
2575 | atomic_set(&rbio->error, 0); | |
2576 | /* | |
2577 | * build a list of bios to read all the missing parts of this | |
2578 | * stripe | |
2579 | */ | |
2c8cdd6e | 2580 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) { |
5a6ac9ea MX |
2581 | for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) { |
2582 | struct page *page; | |
2583 | /* | |
2584 | * we want to find all the pages missing from | |
2585 | * the rbio and read them from the disk. If | |
2586 | * page_in_rbio finds a page in the bio list | |
2587 | * we don't need to read it off the stripe. | |
2588 | */ | |
2589 | page = page_in_rbio(rbio, stripe, pagenr, 1); | |
2590 | if (page) | |
2591 | continue; | |
2592 | ||
2593 | page = rbio_stripe_page(rbio, stripe, pagenr); | |
2594 | /* | |
2595 | * the bio cache may have handed us an uptodate | |
2596 | * page. If so, be happy and use it | |
2597 | */ | |
2598 | if (PageUptodate(page)) | |
2599 | continue; | |
2600 | ||
2601 | ret = rbio_add_io_page(rbio, &bio_list, page, | |
2602 | stripe, pagenr, rbio->stripe_len); | |
2603 | if (ret) | |
2604 | goto cleanup; | |
2605 | } | |
2606 | } | |
2607 | ||
2608 | bios_to_read = bio_list_size(&bio_list); | |
2609 | if (!bios_to_read) { | |
2610 | /* | |
2611 | * this can happen if others have merged with | |
2612 | * us, it means there is nothing left to read. | |
2613 | * But if there are missing devices it may not be | |
2614 | * safe to do the full stripe write yet. | |
2615 | */ | |
2616 | goto finish; | |
2617 | } | |
2618 | ||
2619 | /* | |
2620 | * the bbio may be freed once we submit the last bio. Make sure | |
2621 | * not to touch it after that | |
2622 | */ | |
2623 | atomic_set(&rbio->stripes_pending, bios_to_read); | |
2624 | while (1) { | |
2625 | bio = bio_list_pop(&bio_list); | |
2626 | if (!bio) | |
2627 | break; | |
2628 | ||
2629 | bio->bi_private = rbio; | |
2630 | bio->bi_end_io = raid56_parity_scrub_end_io; | |
2631 | ||
2632 | btrfs_bio_wq_end_io(rbio->fs_info, bio, | |
2633 | BTRFS_WQ_ENDIO_RAID56); | |
2634 | ||
2635 | BUG_ON(!test_bit(BIO_UPTODATE, &bio->bi_flags)); | |
2636 | submit_bio(READ, bio); | |
2637 | } | |
2638 | /* the actual write will happen once the reads are done */ | |
2639 | return; | |
2640 | ||
2641 | cleanup: | |
2642 | rbio_orig_end_io(rbio, -EIO, 0); | |
2643 | return; | |
2644 | ||
2645 | finish: | |
2646 | validate_rbio_for_parity_scrub(rbio); | |
2647 | } | |
2648 | ||
2649 | static void scrub_parity_work(struct btrfs_work *work) | |
2650 | { | |
2651 | struct btrfs_raid_bio *rbio; | |
2652 | ||
2653 | rbio = container_of(work, struct btrfs_raid_bio, work); | |
2654 | raid56_parity_scrub_stripe(rbio); | |
2655 | } | |
2656 | ||
2657 | static void async_scrub_parity(struct btrfs_raid_bio *rbio) | |
2658 | { | |
2659 | btrfs_init_work(&rbio->work, btrfs_rmw_helper, | |
2660 | scrub_parity_work, NULL, NULL); | |
2661 | ||
2662 | btrfs_queue_work(rbio->fs_info->rmw_workers, | |
2663 | &rbio->work); | |
2664 | } | |
2665 | ||
2666 | void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio) | |
2667 | { | |
2668 | if (!lock_stripe_add(rbio)) | |
2669 | async_scrub_parity(rbio); | |
2670 | } |