]>
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 | ||
af8e2d1d MX |
61 | /* |
62 | * bbio and raid_map is managed by the caller, so we shouldn't free | |
63 | * them here. And besides that, all rbios with this flag should not | |
64 | * be cached, because we need raid_map to check the rbios' stripe | |
65 | * is the same or not, but it is very likely that the caller has | |
66 | * free raid_map, so don't cache those rbios. | |
67 | */ | |
68 | #define RBIO_HOLD_BBIO_MAP_BIT 4 | |
69 | ||
4ae10b3a CM |
70 | #define RBIO_CACHE_SIZE 1024 |
71 | ||
1b94b556 MX |
72 | enum btrfs_rbio_ops { |
73 | BTRFS_RBIO_WRITE = 0, | |
74 | BTRFS_RBIO_READ_REBUILD = 1, | |
5a6ac9ea | 75 | BTRFS_RBIO_PARITY_SCRUB = 2, |
1b94b556 MX |
76 | }; |
77 | ||
53b381b3 DW |
78 | struct btrfs_raid_bio { |
79 | struct btrfs_fs_info *fs_info; | |
80 | struct btrfs_bio *bbio; | |
81 | ||
82 | /* | |
83 | * logical block numbers for the start of each stripe | |
84 | * The last one or two are p/q. These are sorted, | |
85 | * so raid_map[0] is the start of our full stripe | |
86 | */ | |
87 | u64 *raid_map; | |
88 | ||
89 | /* while we're doing rmw on a stripe | |
90 | * we put it into a hash table so we can | |
91 | * lock the stripe and merge more rbios | |
92 | * into it. | |
93 | */ | |
94 | struct list_head hash_list; | |
95 | ||
4ae10b3a CM |
96 | /* |
97 | * LRU list for the stripe cache | |
98 | */ | |
99 | struct list_head stripe_cache; | |
100 | ||
53b381b3 DW |
101 | /* |
102 | * for scheduling work in the helper threads | |
103 | */ | |
104 | struct btrfs_work work; | |
105 | ||
106 | /* | |
107 | * bio list and bio_list_lock are used | |
108 | * to add more bios into the stripe | |
109 | * in hopes of avoiding the full rmw | |
110 | */ | |
111 | struct bio_list bio_list; | |
112 | spinlock_t bio_list_lock; | |
113 | ||
6ac0f488 CM |
114 | /* also protected by the bio_list_lock, the |
115 | * plug list is used by the plugging code | |
116 | * to collect partial bios while plugged. The | |
117 | * stripe locking code also uses it to hand off | |
53b381b3 DW |
118 | * the stripe lock to the next pending IO |
119 | */ | |
120 | struct list_head plug_list; | |
121 | ||
122 | /* | |
123 | * flags that tell us if it is safe to | |
124 | * merge with this bio | |
125 | */ | |
126 | unsigned long flags; | |
127 | ||
128 | /* size of each individual stripe on disk */ | |
129 | int stripe_len; | |
130 | ||
131 | /* number of data stripes (no p/q) */ | |
132 | int nr_data; | |
133 | ||
2c8cdd6e MX |
134 | int real_stripes; |
135 | ||
5a6ac9ea | 136 | int stripe_npages; |
53b381b3 DW |
137 | /* |
138 | * set if we're doing a parity rebuild | |
139 | * for a read from higher up, which is handled | |
140 | * differently from a parity rebuild as part of | |
141 | * rmw | |
142 | */ | |
1b94b556 | 143 | enum btrfs_rbio_ops operation; |
53b381b3 DW |
144 | |
145 | /* first bad stripe */ | |
146 | int faila; | |
147 | ||
148 | /* second bad stripe (for raid6 use) */ | |
149 | int failb; | |
150 | ||
5a6ac9ea | 151 | int scrubp; |
53b381b3 DW |
152 | /* |
153 | * number of pages needed to represent the full | |
154 | * stripe | |
155 | */ | |
156 | int nr_pages; | |
157 | ||
158 | /* | |
159 | * size of all the bios in the bio_list. This | |
160 | * helps us decide if the rbio maps to a full | |
161 | * stripe or not | |
162 | */ | |
163 | int bio_list_bytes; | |
164 | ||
165 | atomic_t refs; | |
166 | ||
b89e1b01 MX |
167 | atomic_t stripes_pending; |
168 | ||
169 | atomic_t error; | |
53b381b3 DW |
170 | /* |
171 | * these are two arrays of pointers. We allocate the | |
172 | * rbio big enough to hold them both and setup their | |
173 | * locations when the rbio is allocated | |
174 | */ | |
175 | ||
176 | /* pointers to pages that we allocated for | |
177 | * reading/writing stripes directly from the disk (including P/Q) | |
178 | */ | |
179 | struct page **stripe_pages; | |
180 | ||
181 | /* | |
182 | * pointers to the pages in the bio_list. Stored | |
183 | * here for faster lookup | |
184 | */ | |
185 | struct page **bio_pages; | |
5a6ac9ea MX |
186 | |
187 | /* | |
188 | * bitmap to record which horizontal stripe has data | |
189 | */ | |
190 | unsigned long *dbitmap; | |
53b381b3 DW |
191 | }; |
192 | ||
193 | static int __raid56_parity_recover(struct btrfs_raid_bio *rbio); | |
194 | static noinline void finish_rmw(struct btrfs_raid_bio *rbio); | |
195 | static void rmw_work(struct btrfs_work *work); | |
196 | static void read_rebuild_work(struct btrfs_work *work); | |
197 | static void async_rmw_stripe(struct btrfs_raid_bio *rbio); | |
198 | static void async_read_rebuild(struct btrfs_raid_bio *rbio); | |
199 | static int fail_bio_stripe(struct btrfs_raid_bio *rbio, struct bio *bio); | |
200 | static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed); | |
201 | static void __free_raid_bio(struct btrfs_raid_bio *rbio); | |
202 | static void index_rbio_pages(struct btrfs_raid_bio *rbio); | |
203 | static int alloc_rbio_pages(struct btrfs_raid_bio *rbio); | |
204 | ||
5a6ac9ea MX |
205 | static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio, |
206 | int need_check); | |
207 | static void async_scrub_parity(struct btrfs_raid_bio *rbio); | |
208 | ||
53b381b3 DW |
209 | /* |
210 | * the stripe hash table is used for locking, and to collect | |
211 | * bios in hopes of making a full stripe | |
212 | */ | |
213 | int btrfs_alloc_stripe_hash_table(struct btrfs_fs_info *info) | |
214 | { | |
215 | struct btrfs_stripe_hash_table *table; | |
216 | struct btrfs_stripe_hash_table *x; | |
217 | struct btrfs_stripe_hash *cur; | |
218 | struct btrfs_stripe_hash *h; | |
219 | int num_entries = 1 << BTRFS_STRIPE_HASH_TABLE_BITS; | |
220 | int i; | |
83c8266a | 221 | int table_size; |
53b381b3 DW |
222 | |
223 | if (info->stripe_hash_table) | |
224 | return 0; | |
225 | ||
83c8266a DS |
226 | /* |
227 | * The table is large, starting with order 4 and can go as high as | |
228 | * order 7 in case lock debugging is turned on. | |
229 | * | |
230 | * Try harder to allocate and fallback to vmalloc to lower the chance | |
231 | * of a failing mount. | |
232 | */ | |
233 | table_size = sizeof(*table) + sizeof(*h) * num_entries; | |
234 | table = kzalloc(table_size, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT); | |
235 | if (!table) { | |
236 | table = vzalloc(table_size); | |
237 | if (!table) | |
238 | return -ENOMEM; | |
239 | } | |
53b381b3 | 240 | |
4ae10b3a CM |
241 | spin_lock_init(&table->cache_lock); |
242 | INIT_LIST_HEAD(&table->stripe_cache); | |
243 | ||
53b381b3 DW |
244 | h = table->table; |
245 | ||
246 | for (i = 0; i < num_entries; i++) { | |
247 | cur = h + i; | |
248 | INIT_LIST_HEAD(&cur->hash_list); | |
249 | spin_lock_init(&cur->lock); | |
250 | init_waitqueue_head(&cur->wait); | |
251 | } | |
252 | ||
253 | x = cmpxchg(&info->stripe_hash_table, NULL, table); | |
83c8266a DS |
254 | if (x) { |
255 | if (is_vmalloc_addr(x)) | |
256 | vfree(x); | |
257 | else | |
258 | kfree(x); | |
259 | } | |
53b381b3 DW |
260 | return 0; |
261 | } | |
262 | ||
4ae10b3a CM |
263 | /* |
264 | * caching an rbio means to copy anything from the | |
265 | * bio_pages array into the stripe_pages array. We | |
266 | * use the page uptodate bit in the stripe cache array | |
267 | * to indicate if it has valid data | |
268 | * | |
269 | * once the caching is done, we set the cache ready | |
270 | * bit. | |
271 | */ | |
272 | static void cache_rbio_pages(struct btrfs_raid_bio *rbio) | |
273 | { | |
274 | int i; | |
275 | char *s; | |
276 | char *d; | |
277 | int ret; | |
278 | ||
279 | ret = alloc_rbio_pages(rbio); | |
280 | if (ret) | |
281 | return; | |
282 | ||
283 | for (i = 0; i < rbio->nr_pages; i++) { | |
284 | if (!rbio->bio_pages[i]) | |
285 | continue; | |
286 | ||
287 | s = kmap(rbio->bio_pages[i]); | |
288 | d = kmap(rbio->stripe_pages[i]); | |
289 | ||
290 | memcpy(d, s, PAGE_CACHE_SIZE); | |
291 | ||
292 | kunmap(rbio->bio_pages[i]); | |
293 | kunmap(rbio->stripe_pages[i]); | |
294 | SetPageUptodate(rbio->stripe_pages[i]); | |
295 | } | |
296 | set_bit(RBIO_CACHE_READY_BIT, &rbio->flags); | |
297 | } | |
298 | ||
53b381b3 DW |
299 | /* |
300 | * we hash on the first logical address of the stripe | |
301 | */ | |
302 | static int rbio_bucket(struct btrfs_raid_bio *rbio) | |
303 | { | |
304 | u64 num = rbio->raid_map[0]; | |
305 | ||
306 | /* | |
307 | * we shift down quite a bit. We're using byte | |
308 | * addressing, and most of the lower bits are zeros. | |
309 | * This tends to upset hash_64, and it consistently | |
310 | * returns just one or two different values. | |
311 | * | |
312 | * shifting off the lower bits fixes things. | |
313 | */ | |
314 | return hash_64(num >> 16, BTRFS_STRIPE_HASH_TABLE_BITS); | |
315 | } | |
316 | ||
4ae10b3a CM |
317 | /* |
318 | * stealing an rbio means taking all the uptodate pages from the stripe | |
319 | * array in the source rbio and putting them into the destination rbio | |
320 | */ | |
321 | static void steal_rbio(struct btrfs_raid_bio *src, struct btrfs_raid_bio *dest) | |
322 | { | |
323 | int i; | |
324 | struct page *s; | |
325 | struct page *d; | |
326 | ||
327 | if (!test_bit(RBIO_CACHE_READY_BIT, &src->flags)) | |
328 | return; | |
329 | ||
330 | for (i = 0; i < dest->nr_pages; i++) { | |
331 | s = src->stripe_pages[i]; | |
332 | if (!s || !PageUptodate(s)) { | |
333 | continue; | |
334 | } | |
335 | ||
336 | d = dest->stripe_pages[i]; | |
337 | if (d) | |
338 | __free_page(d); | |
339 | ||
340 | dest->stripe_pages[i] = s; | |
341 | src->stripe_pages[i] = NULL; | |
342 | } | |
343 | } | |
344 | ||
53b381b3 DW |
345 | /* |
346 | * merging means we take the bio_list from the victim and | |
347 | * splice it into the destination. The victim should | |
348 | * be discarded afterwards. | |
349 | * | |
350 | * must be called with dest->rbio_list_lock held | |
351 | */ | |
352 | static void merge_rbio(struct btrfs_raid_bio *dest, | |
353 | struct btrfs_raid_bio *victim) | |
354 | { | |
355 | bio_list_merge(&dest->bio_list, &victim->bio_list); | |
356 | dest->bio_list_bytes += victim->bio_list_bytes; | |
357 | bio_list_init(&victim->bio_list); | |
358 | } | |
359 | ||
360 | /* | |
4ae10b3a CM |
361 | * used to prune items that are in the cache. The caller |
362 | * must hold the hash table lock. | |
363 | */ | |
364 | static void __remove_rbio_from_cache(struct btrfs_raid_bio *rbio) | |
365 | { | |
366 | int bucket = rbio_bucket(rbio); | |
367 | struct btrfs_stripe_hash_table *table; | |
368 | struct btrfs_stripe_hash *h; | |
369 | int freeit = 0; | |
370 | ||
371 | /* | |
372 | * check the bit again under the hash table lock. | |
373 | */ | |
374 | if (!test_bit(RBIO_CACHE_BIT, &rbio->flags)) | |
375 | return; | |
376 | ||
377 | table = rbio->fs_info->stripe_hash_table; | |
378 | h = table->table + bucket; | |
379 | ||
380 | /* hold the lock for the bucket because we may be | |
381 | * removing it from the hash table | |
382 | */ | |
383 | spin_lock(&h->lock); | |
384 | ||
385 | /* | |
386 | * hold the lock for the bio list because we need | |
387 | * to make sure the bio list is empty | |
388 | */ | |
389 | spin_lock(&rbio->bio_list_lock); | |
390 | ||
391 | if (test_and_clear_bit(RBIO_CACHE_BIT, &rbio->flags)) { | |
392 | list_del_init(&rbio->stripe_cache); | |
393 | table->cache_size -= 1; | |
394 | freeit = 1; | |
395 | ||
396 | /* if the bio list isn't empty, this rbio is | |
397 | * still involved in an IO. We take it out | |
398 | * of the cache list, and drop the ref that | |
399 | * was held for the list. | |
400 | * | |
401 | * If the bio_list was empty, we also remove | |
402 | * the rbio from the hash_table, and drop | |
403 | * the corresponding ref | |
404 | */ | |
405 | if (bio_list_empty(&rbio->bio_list)) { | |
406 | if (!list_empty(&rbio->hash_list)) { | |
407 | list_del_init(&rbio->hash_list); | |
408 | atomic_dec(&rbio->refs); | |
409 | BUG_ON(!list_empty(&rbio->plug_list)); | |
410 | } | |
411 | } | |
412 | } | |
413 | ||
414 | spin_unlock(&rbio->bio_list_lock); | |
415 | spin_unlock(&h->lock); | |
416 | ||
417 | if (freeit) | |
418 | __free_raid_bio(rbio); | |
419 | } | |
420 | ||
421 | /* | |
422 | * prune a given rbio from the cache | |
423 | */ | |
424 | static void remove_rbio_from_cache(struct btrfs_raid_bio *rbio) | |
425 | { | |
426 | struct btrfs_stripe_hash_table *table; | |
427 | unsigned long flags; | |
428 | ||
429 | if (!test_bit(RBIO_CACHE_BIT, &rbio->flags)) | |
430 | return; | |
431 | ||
432 | table = rbio->fs_info->stripe_hash_table; | |
433 | ||
434 | spin_lock_irqsave(&table->cache_lock, flags); | |
435 | __remove_rbio_from_cache(rbio); | |
436 | spin_unlock_irqrestore(&table->cache_lock, flags); | |
437 | } | |
438 | ||
439 | /* | |
440 | * remove everything in the cache | |
441 | */ | |
48a3b636 | 442 | static void btrfs_clear_rbio_cache(struct btrfs_fs_info *info) |
4ae10b3a CM |
443 | { |
444 | struct btrfs_stripe_hash_table *table; | |
445 | unsigned long flags; | |
446 | struct btrfs_raid_bio *rbio; | |
447 | ||
448 | table = info->stripe_hash_table; | |
449 | ||
450 | spin_lock_irqsave(&table->cache_lock, flags); | |
451 | while (!list_empty(&table->stripe_cache)) { | |
452 | rbio = list_entry(table->stripe_cache.next, | |
453 | struct btrfs_raid_bio, | |
454 | stripe_cache); | |
455 | __remove_rbio_from_cache(rbio); | |
456 | } | |
457 | spin_unlock_irqrestore(&table->cache_lock, flags); | |
458 | } | |
459 | ||
460 | /* | |
461 | * remove all cached entries and free the hash table | |
462 | * used by unmount | |
53b381b3 DW |
463 | */ |
464 | void btrfs_free_stripe_hash_table(struct btrfs_fs_info *info) | |
465 | { | |
466 | if (!info->stripe_hash_table) | |
467 | return; | |
4ae10b3a | 468 | btrfs_clear_rbio_cache(info); |
83c8266a DS |
469 | if (is_vmalloc_addr(info->stripe_hash_table)) |
470 | vfree(info->stripe_hash_table); | |
471 | else | |
472 | kfree(info->stripe_hash_table); | |
53b381b3 DW |
473 | info->stripe_hash_table = NULL; |
474 | } | |
475 | ||
4ae10b3a CM |
476 | /* |
477 | * insert an rbio into the stripe cache. It | |
478 | * must have already been prepared by calling | |
479 | * cache_rbio_pages | |
480 | * | |
481 | * If this rbio was already cached, it gets | |
482 | * moved to the front of the lru. | |
483 | * | |
484 | * If the size of the rbio cache is too big, we | |
485 | * prune an item. | |
486 | */ | |
487 | static void cache_rbio(struct btrfs_raid_bio *rbio) | |
488 | { | |
489 | struct btrfs_stripe_hash_table *table; | |
490 | unsigned long flags; | |
491 | ||
492 | if (!test_bit(RBIO_CACHE_READY_BIT, &rbio->flags)) | |
493 | return; | |
494 | ||
495 | table = rbio->fs_info->stripe_hash_table; | |
496 | ||
497 | spin_lock_irqsave(&table->cache_lock, flags); | |
498 | spin_lock(&rbio->bio_list_lock); | |
499 | ||
500 | /* bump our ref if we were not in the list before */ | |
501 | if (!test_and_set_bit(RBIO_CACHE_BIT, &rbio->flags)) | |
502 | atomic_inc(&rbio->refs); | |
503 | ||
504 | if (!list_empty(&rbio->stripe_cache)){ | |
505 | list_move(&rbio->stripe_cache, &table->stripe_cache); | |
506 | } else { | |
507 | list_add(&rbio->stripe_cache, &table->stripe_cache); | |
508 | table->cache_size += 1; | |
509 | } | |
510 | ||
511 | spin_unlock(&rbio->bio_list_lock); | |
512 | ||
513 | if (table->cache_size > RBIO_CACHE_SIZE) { | |
514 | struct btrfs_raid_bio *found; | |
515 | ||
516 | found = list_entry(table->stripe_cache.prev, | |
517 | struct btrfs_raid_bio, | |
518 | stripe_cache); | |
519 | ||
520 | if (found != rbio) | |
521 | __remove_rbio_from_cache(found); | |
522 | } | |
523 | ||
524 | spin_unlock_irqrestore(&table->cache_lock, flags); | |
525 | return; | |
526 | } | |
527 | ||
53b381b3 DW |
528 | /* |
529 | * helper function to run the xor_blocks api. It is only | |
530 | * able to do MAX_XOR_BLOCKS at a time, so we need to | |
531 | * loop through. | |
532 | */ | |
533 | static void run_xor(void **pages, int src_cnt, ssize_t len) | |
534 | { | |
535 | int src_off = 0; | |
536 | int xor_src_cnt = 0; | |
537 | void *dest = pages[src_cnt]; | |
538 | ||
539 | while(src_cnt > 0) { | |
540 | xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS); | |
541 | xor_blocks(xor_src_cnt, len, dest, pages + src_off); | |
542 | ||
543 | src_cnt -= xor_src_cnt; | |
544 | src_off += xor_src_cnt; | |
545 | } | |
546 | } | |
547 | ||
548 | /* | |
549 | * returns true if the bio list inside this rbio | |
550 | * covers an entire stripe (no rmw required). | |
551 | * Must be called with the bio list lock held, or | |
552 | * at a time when you know it is impossible to add | |
553 | * new bios into the list | |
554 | */ | |
555 | static int __rbio_is_full(struct btrfs_raid_bio *rbio) | |
556 | { | |
557 | unsigned long size = rbio->bio_list_bytes; | |
558 | int ret = 1; | |
559 | ||
560 | if (size != rbio->nr_data * rbio->stripe_len) | |
561 | ret = 0; | |
562 | ||
563 | BUG_ON(size > rbio->nr_data * rbio->stripe_len); | |
564 | return ret; | |
565 | } | |
566 | ||
567 | static int rbio_is_full(struct btrfs_raid_bio *rbio) | |
568 | { | |
569 | unsigned long flags; | |
570 | int ret; | |
571 | ||
572 | spin_lock_irqsave(&rbio->bio_list_lock, flags); | |
573 | ret = __rbio_is_full(rbio); | |
574 | spin_unlock_irqrestore(&rbio->bio_list_lock, flags); | |
575 | return ret; | |
576 | } | |
577 | ||
578 | /* | |
579 | * returns 1 if it is safe to merge two rbios together. | |
580 | * The merging is safe if the two rbios correspond to | |
581 | * the same stripe and if they are both going in the same | |
582 | * direction (read vs write), and if neither one is | |
583 | * locked for final IO | |
584 | * | |
585 | * The caller is responsible for locking such that | |
586 | * rmw_locked is safe to test | |
587 | */ | |
588 | static int rbio_can_merge(struct btrfs_raid_bio *last, | |
589 | struct btrfs_raid_bio *cur) | |
590 | { | |
591 | if (test_bit(RBIO_RMW_LOCKED_BIT, &last->flags) || | |
592 | test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) | |
593 | return 0; | |
594 | ||
4ae10b3a CM |
595 | /* |
596 | * we can't merge with cached rbios, since the | |
597 | * idea is that when we merge the destination | |
598 | * rbio is going to run our IO for us. We can | |
599 | * steal from cached rbio's though, other functions | |
600 | * handle that. | |
601 | */ | |
602 | if (test_bit(RBIO_CACHE_BIT, &last->flags) || | |
603 | test_bit(RBIO_CACHE_BIT, &cur->flags)) | |
604 | return 0; | |
605 | ||
53b381b3 DW |
606 | if (last->raid_map[0] != |
607 | cur->raid_map[0]) | |
608 | return 0; | |
609 | ||
5a6ac9ea MX |
610 | /* we can't merge with different operations */ |
611 | if (last->operation != cur->operation) | |
612 | return 0; | |
613 | /* | |
614 | * We've need read the full stripe from the drive. | |
615 | * check and repair the parity and write the new results. | |
616 | * | |
617 | * We're not allowed to add any new bios to the | |
618 | * bio list here, anyone else that wants to | |
619 | * change this stripe needs to do their own rmw. | |
620 | */ | |
621 | if (last->operation == BTRFS_RBIO_PARITY_SCRUB || | |
622 | cur->operation == BTRFS_RBIO_PARITY_SCRUB) | |
53b381b3 | 623 | return 0; |
53b381b3 DW |
624 | |
625 | return 1; | |
626 | } | |
627 | ||
628 | /* | |
629 | * helper to index into the pstripe | |
630 | */ | |
631 | static struct page *rbio_pstripe_page(struct btrfs_raid_bio *rbio, int index) | |
632 | { | |
633 | index += (rbio->nr_data * rbio->stripe_len) >> PAGE_CACHE_SHIFT; | |
634 | return rbio->stripe_pages[index]; | |
635 | } | |
636 | ||
637 | /* | |
638 | * helper to index into the qstripe, returns null | |
639 | * if there is no qstripe | |
640 | */ | |
641 | static struct page *rbio_qstripe_page(struct btrfs_raid_bio *rbio, int index) | |
642 | { | |
2c8cdd6e | 643 | if (rbio->nr_data + 1 == rbio->real_stripes) |
53b381b3 DW |
644 | return NULL; |
645 | ||
646 | index += ((rbio->nr_data + 1) * rbio->stripe_len) >> | |
647 | PAGE_CACHE_SHIFT; | |
648 | return rbio->stripe_pages[index]; | |
649 | } | |
650 | ||
651 | /* | |
652 | * The first stripe in the table for a logical address | |
653 | * has the lock. rbios are added in one of three ways: | |
654 | * | |
655 | * 1) Nobody has the stripe locked yet. The rbio is given | |
656 | * the lock and 0 is returned. The caller must start the IO | |
657 | * themselves. | |
658 | * | |
659 | * 2) Someone has the stripe locked, but we're able to merge | |
660 | * with the lock owner. The rbio is freed and the IO will | |
661 | * start automatically along with the existing rbio. 1 is returned. | |
662 | * | |
663 | * 3) Someone has the stripe locked, but we're not able to merge. | |
664 | * The rbio is added to the lock owner's plug list, or merged into | |
665 | * an rbio already on the plug list. When the lock owner unlocks, | |
666 | * the next rbio on the list is run and the IO is started automatically. | |
667 | * 1 is returned | |
668 | * | |
669 | * If we return 0, the caller still owns the rbio and must continue with | |
670 | * IO submission. If we return 1, the caller must assume the rbio has | |
671 | * already been freed. | |
672 | */ | |
673 | static noinline int lock_stripe_add(struct btrfs_raid_bio *rbio) | |
674 | { | |
675 | int bucket = rbio_bucket(rbio); | |
676 | struct btrfs_stripe_hash *h = rbio->fs_info->stripe_hash_table->table + bucket; | |
677 | struct btrfs_raid_bio *cur; | |
678 | struct btrfs_raid_bio *pending; | |
679 | unsigned long flags; | |
680 | DEFINE_WAIT(wait); | |
681 | struct btrfs_raid_bio *freeit = NULL; | |
4ae10b3a | 682 | struct btrfs_raid_bio *cache_drop = NULL; |
53b381b3 DW |
683 | int ret = 0; |
684 | int walk = 0; | |
685 | ||
686 | spin_lock_irqsave(&h->lock, flags); | |
687 | list_for_each_entry(cur, &h->hash_list, hash_list) { | |
688 | walk++; | |
689 | if (cur->raid_map[0] == rbio->raid_map[0]) { | |
690 | spin_lock(&cur->bio_list_lock); | |
691 | ||
4ae10b3a CM |
692 | /* can we steal this cached rbio's pages? */ |
693 | if (bio_list_empty(&cur->bio_list) && | |
694 | list_empty(&cur->plug_list) && | |
695 | test_bit(RBIO_CACHE_BIT, &cur->flags) && | |
696 | !test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) { | |
697 | list_del_init(&cur->hash_list); | |
698 | atomic_dec(&cur->refs); | |
699 | ||
700 | steal_rbio(cur, rbio); | |
701 | cache_drop = cur; | |
702 | spin_unlock(&cur->bio_list_lock); | |
703 | ||
704 | goto lockit; | |
705 | } | |
706 | ||
53b381b3 DW |
707 | /* can we merge into the lock owner? */ |
708 | if (rbio_can_merge(cur, rbio)) { | |
709 | merge_rbio(cur, rbio); | |
710 | spin_unlock(&cur->bio_list_lock); | |
711 | freeit = rbio; | |
712 | ret = 1; | |
713 | goto out; | |
714 | } | |
715 | ||
4ae10b3a | 716 | |
53b381b3 DW |
717 | /* |
718 | * we couldn't merge with the running | |
719 | * rbio, see if we can merge with the | |
720 | * pending ones. We don't have to | |
721 | * check for rmw_locked because there | |
722 | * is no way they are inside finish_rmw | |
723 | * right now | |
724 | */ | |
725 | list_for_each_entry(pending, &cur->plug_list, | |
726 | plug_list) { | |
727 | if (rbio_can_merge(pending, rbio)) { | |
728 | merge_rbio(pending, rbio); | |
729 | spin_unlock(&cur->bio_list_lock); | |
730 | freeit = rbio; | |
731 | ret = 1; | |
732 | goto out; | |
733 | } | |
734 | } | |
735 | ||
736 | /* no merging, put us on the tail of the plug list, | |
737 | * our rbio will be started with the currently | |
738 | * running rbio unlocks | |
739 | */ | |
740 | list_add_tail(&rbio->plug_list, &cur->plug_list); | |
741 | spin_unlock(&cur->bio_list_lock); | |
742 | ret = 1; | |
743 | goto out; | |
744 | } | |
745 | } | |
4ae10b3a | 746 | lockit: |
53b381b3 DW |
747 | atomic_inc(&rbio->refs); |
748 | list_add(&rbio->hash_list, &h->hash_list); | |
749 | out: | |
750 | spin_unlock_irqrestore(&h->lock, flags); | |
4ae10b3a CM |
751 | if (cache_drop) |
752 | remove_rbio_from_cache(cache_drop); | |
53b381b3 DW |
753 | if (freeit) |
754 | __free_raid_bio(freeit); | |
755 | return ret; | |
756 | } | |
757 | ||
758 | /* | |
759 | * called as rmw or parity rebuild is completed. If the plug list has more | |
760 | * rbios waiting for this stripe, the next one on the list will be started | |
761 | */ | |
762 | static noinline void unlock_stripe(struct btrfs_raid_bio *rbio) | |
763 | { | |
764 | int bucket; | |
765 | struct btrfs_stripe_hash *h; | |
766 | unsigned long flags; | |
4ae10b3a | 767 | int keep_cache = 0; |
53b381b3 DW |
768 | |
769 | bucket = rbio_bucket(rbio); | |
770 | h = rbio->fs_info->stripe_hash_table->table + bucket; | |
771 | ||
4ae10b3a CM |
772 | if (list_empty(&rbio->plug_list)) |
773 | cache_rbio(rbio); | |
774 | ||
53b381b3 DW |
775 | spin_lock_irqsave(&h->lock, flags); |
776 | spin_lock(&rbio->bio_list_lock); | |
777 | ||
778 | if (!list_empty(&rbio->hash_list)) { | |
4ae10b3a CM |
779 | /* |
780 | * if we're still cached and there is no other IO | |
781 | * to perform, just leave this rbio here for others | |
782 | * to steal from later | |
783 | */ | |
784 | if (list_empty(&rbio->plug_list) && | |
785 | test_bit(RBIO_CACHE_BIT, &rbio->flags)) { | |
786 | keep_cache = 1; | |
787 | clear_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags); | |
788 | BUG_ON(!bio_list_empty(&rbio->bio_list)); | |
789 | goto done; | |
790 | } | |
53b381b3 DW |
791 | |
792 | list_del_init(&rbio->hash_list); | |
793 | atomic_dec(&rbio->refs); | |
794 | ||
795 | /* | |
796 | * we use the plug list to hold all the rbios | |
797 | * waiting for the chance to lock this stripe. | |
798 | * hand the lock over to one of them. | |
799 | */ | |
800 | if (!list_empty(&rbio->plug_list)) { | |
801 | struct btrfs_raid_bio *next; | |
802 | struct list_head *head = rbio->plug_list.next; | |
803 | ||
804 | next = list_entry(head, struct btrfs_raid_bio, | |
805 | plug_list); | |
806 | ||
807 | list_del_init(&rbio->plug_list); | |
808 | ||
809 | list_add(&next->hash_list, &h->hash_list); | |
810 | atomic_inc(&next->refs); | |
811 | spin_unlock(&rbio->bio_list_lock); | |
812 | spin_unlock_irqrestore(&h->lock, flags); | |
813 | ||
1b94b556 | 814 | if (next->operation == BTRFS_RBIO_READ_REBUILD) |
53b381b3 | 815 | async_read_rebuild(next); |
5a6ac9ea | 816 | else if (next->operation == BTRFS_RBIO_WRITE) { |
4ae10b3a | 817 | steal_rbio(rbio, next); |
53b381b3 | 818 | async_rmw_stripe(next); |
5a6ac9ea MX |
819 | } else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) { |
820 | steal_rbio(rbio, next); | |
821 | async_scrub_parity(next); | |
4ae10b3a | 822 | } |
53b381b3 DW |
823 | |
824 | goto done_nolock; | |
53b381b3 DW |
825 | } else if (waitqueue_active(&h->wait)) { |
826 | spin_unlock(&rbio->bio_list_lock); | |
827 | spin_unlock_irqrestore(&h->lock, flags); | |
828 | wake_up(&h->wait); | |
829 | goto done_nolock; | |
830 | } | |
831 | } | |
4ae10b3a | 832 | done: |
53b381b3 DW |
833 | spin_unlock(&rbio->bio_list_lock); |
834 | spin_unlock_irqrestore(&h->lock, flags); | |
835 | ||
836 | done_nolock: | |
4ae10b3a CM |
837 | if (!keep_cache) |
838 | remove_rbio_from_cache(rbio); | |
53b381b3 DW |
839 | } |
840 | ||
af8e2d1d MX |
841 | static inline void |
842 | __free_bbio_and_raid_map(struct btrfs_bio *bbio, u64 *raid_map, int need) | |
843 | { | |
844 | if (need) { | |
845 | kfree(raid_map); | |
846 | kfree(bbio); | |
847 | } | |
848 | } | |
849 | ||
850 | static inline void free_bbio_and_raid_map(struct btrfs_raid_bio *rbio) | |
851 | { | |
852 | __free_bbio_and_raid_map(rbio->bbio, rbio->raid_map, | |
853 | !test_bit(RBIO_HOLD_BBIO_MAP_BIT, &rbio->flags)); | |
854 | } | |
855 | ||
53b381b3 DW |
856 | static void __free_raid_bio(struct btrfs_raid_bio *rbio) |
857 | { | |
858 | int i; | |
859 | ||
860 | WARN_ON(atomic_read(&rbio->refs) < 0); | |
861 | if (!atomic_dec_and_test(&rbio->refs)) | |
862 | return; | |
863 | ||
4ae10b3a | 864 | WARN_ON(!list_empty(&rbio->stripe_cache)); |
53b381b3 DW |
865 | WARN_ON(!list_empty(&rbio->hash_list)); |
866 | WARN_ON(!bio_list_empty(&rbio->bio_list)); | |
867 | ||
868 | for (i = 0; i < rbio->nr_pages; i++) { | |
869 | if (rbio->stripe_pages[i]) { | |
870 | __free_page(rbio->stripe_pages[i]); | |
871 | rbio->stripe_pages[i] = NULL; | |
872 | } | |
873 | } | |
af8e2d1d MX |
874 | |
875 | free_bbio_and_raid_map(rbio); | |
876 | ||
53b381b3 DW |
877 | kfree(rbio); |
878 | } | |
879 | ||
880 | static void free_raid_bio(struct btrfs_raid_bio *rbio) | |
881 | { | |
882 | unlock_stripe(rbio); | |
883 | __free_raid_bio(rbio); | |
884 | } | |
885 | ||
886 | /* | |
887 | * this frees the rbio and runs through all the bios in the | |
888 | * bio_list and calls end_io on them | |
889 | */ | |
890 | static void rbio_orig_end_io(struct btrfs_raid_bio *rbio, int err, int uptodate) | |
891 | { | |
892 | struct bio *cur = bio_list_get(&rbio->bio_list); | |
893 | struct bio *next; | |
894 | free_raid_bio(rbio); | |
895 | ||
896 | while (cur) { | |
897 | next = cur->bi_next; | |
898 | cur->bi_next = NULL; | |
899 | if (uptodate) | |
900 | set_bit(BIO_UPTODATE, &cur->bi_flags); | |
901 | bio_endio(cur, err); | |
902 | cur = next; | |
903 | } | |
904 | } | |
905 | ||
906 | /* | |
907 | * end io function used by finish_rmw. When we finally | |
908 | * get here, we've written a full stripe | |
909 | */ | |
910 | static void raid_write_end_io(struct bio *bio, int err) | |
911 | { | |
912 | struct btrfs_raid_bio *rbio = bio->bi_private; | |
913 | ||
914 | if (err) | |
915 | fail_bio_stripe(rbio, bio); | |
916 | ||
917 | bio_put(bio); | |
918 | ||
b89e1b01 | 919 | if (!atomic_dec_and_test(&rbio->stripes_pending)) |
53b381b3 DW |
920 | return; |
921 | ||
922 | err = 0; | |
923 | ||
924 | /* OK, we have read all the stripes we need to. */ | |
b89e1b01 | 925 | if (atomic_read(&rbio->error) > rbio->bbio->max_errors) |
53b381b3 DW |
926 | err = -EIO; |
927 | ||
928 | rbio_orig_end_io(rbio, err, 0); | |
929 | return; | |
930 | } | |
931 | ||
932 | /* | |
933 | * the read/modify/write code wants to use the original bio for | |
934 | * any pages it included, and then use the rbio for everything | |
935 | * else. This function decides if a given index (stripe number) | |
936 | * and page number in that stripe fall inside the original bio | |
937 | * or the rbio. | |
938 | * | |
939 | * if you set bio_list_only, you'll get a NULL back for any ranges | |
940 | * that are outside the bio_list | |
941 | * | |
942 | * This doesn't take any refs on anything, you get a bare page pointer | |
943 | * and the caller must bump refs as required. | |
944 | * | |
945 | * You must call index_rbio_pages once before you can trust | |
946 | * the answers from this function. | |
947 | */ | |
948 | static struct page *page_in_rbio(struct btrfs_raid_bio *rbio, | |
949 | int index, int pagenr, int bio_list_only) | |
950 | { | |
951 | int chunk_page; | |
952 | struct page *p = NULL; | |
953 | ||
954 | chunk_page = index * (rbio->stripe_len >> PAGE_SHIFT) + pagenr; | |
955 | ||
956 | spin_lock_irq(&rbio->bio_list_lock); | |
957 | p = rbio->bio_pages[chunk_page]; | |
958 | spin_unlock_irq(&rbio->bio_list_lock); | |
959 | ||
960 | if (p || bio_list_only) | |
961 | return p; | |
962 | ||
963 | return rbio->stripe_pages[chunk_page]; | |
964 | } | |
965 | ||
966 | /* | |
967 | * number of pages we need for the entire stripe across all the | |
968 | * drives | |
969 | */ | |
970 | static unsigned long rbio_nr_pages(unsigned long stripe_len, int nr_stripes) | |
971 | { | |
972 | unsigned long nr = stripe_len * nr_stripes; | |
ed6078f7 | 973 | return DIV_ROUND_UP(nr, PAGE_CACHE_SIZE); |
53b381b3 DW |
974 | } |
975 | ||
976 | /* | |
977 | * allocation and initial setup for the btrfs_raid_bio. Not | |
978 | * this does not allocate any pages for rbio->pages. | |
979 | */ | |
980 | static struct btrfs_raid_bio *alloc_rbio(struct btrfs_root *root, | |
981 | struct btrfs_bio *bbio, u64 *raid_map, | |
982 | u64 stripe_len) | |
983 | { | |
984 | struct btrfs_raid_bio *rbio; | |
985 | int nr_data = 0; | |
2c8cdd6e MX |
986 | int real_stripes = bbio->num_stripes - bbio->num_tgtdevs; |
987 | int num_pages = rbio_nr_pages(stripe_len, real_stripes); | |
5a6ac9ea | 988 | int stripe_npages = DIV_ROUND_UP(stripe_len, PAGE_SIZE); |
53b381b3 DW |
989 | void *p; |
990 | ||
5a6ac9ea MX |
991 | rbio = kzalloc(sizeof(*rbio) + num_pages * sizeof(struct page *) * 2 + |
992 | DIV_ROUND_UP(stripe_npages, BITS_PER_LONG / 8), | |
53b381b3 | 993 | GFP_NOFS); |
af8e2d1d | 994 | if (!rbio) |
53b381b3 | 995 | return ERR_PTR(-ENOMEM); |
53b381b3 DW |
996 | |
997 | bio_list_init(&rbio->bio_list); | |
998 | INIT_LIST_HEAD(&rbio->plug_list); | |
999 | spin_lock_init(&rbio->bio_list_lock); | |
4ae10b3a | 1000 | INIT_LIST_HEAD(&rbio->stripe_cache); |
53b381b3 DW |
1001 | INIT_LIST_HEAD(&rbio->hash_list); |
1002 | rbio->bbio = bbio; | |
1003 | rbio->raid_map = raid_map; | |
1004 | rbio->fs_info = root->fs_info; | |
1005 | rbio->stripe_len = stripe_len; | |
1006 | rbio->nr_pages = num_pages; | |
2c8cdd6e | 1007 | rbio->real_stripes = real_stripes; |
5a6ac9ea | 1008 | rbio->stripe_npages = stripe_npages; |
53b381b3 DW |
1009 | rbio->faila = -1; |
1010 | rbio->failb = -1; | |
1011 | atomic_set(&rbio->refs, 1); | |
b89e1b01 MX |
1012 | atomic_set(&rbio->error, 0); |
1013 | atomic_set(&rbio->stripes_pending, 0); | |
53b381b3 DW |
1014 | |
1015 | /* | |
1016 | * the stripe_pages and bio_pages array point to the extra | |
1017 | * memory we allocated past the end of the rbio | |
1018 | */ | |
1019 | p = rbio + 1; | |
1020 | rbio->stripe_pages = p; | |
1021 | rbio->bio_pages = p + sizeof(struct page *) * num_pages; | |
5a6ac9ea | 1022 | rbio->dbitmap = p + sizeof(struct page *) * num_pages * 2; |
53b381b3 | 1023 | |
2c8cdd6e MX |
1024 | if (raid_map[real_stripes - 1] == RAID6_Q_STRIPE) |
1025 | nr_data = real_stripes - 2; | |
53b381b3 | 1026 | else |
2c8cdd6e | 1027 | nr_data = real_stripes - 1; |
53b381b3 DW |
1028 | |
1029 | rbio->nr_data = nr_data; | |
1030 | return rbio; | |
1031 | } | |
1032 | ||
1033 | /* allocate pages for all the stripes in the bio, including parity */ | |
1034 | static int alloc_rbio_pages(struct btrfs_raid_bio *rbio) | |
1035 | { | |
1036 | int i; | |
1037 | struct page *page; | |
1038 | ||
1039 | for (i = 0; i < rbio->nr_pages; i++) { | |
1040 | if (rbio->stripe_pages[i]) | |
1041 | continue; | |
1042 | page = alloc_page(GFP_NOFS | __GFP_HIGHMEM); | |
1043 | if (!page) | |
1044 | return -ENOMEM; | |
1045 | rbio->stripe_pages[i] = page; | |
1046 | ClearPageUptodate(page); | |
1047 | } | |
1048 | return 0; | |
1049 | } | |
1050 | ||
1051 | /* allocate pages for just the p/q stripes */ | |
1052 | static int alloc_rbio_parity_pages(struct btrfs_raid_bio *rbio) | |
1053 | { | |
1054 | int i; | |
1055 | struct page *page; | |
1056 | ||
1057 | i = (rbio->nr_data * rbio->stripe_len) >> PAGE_CACHE_SHIFT; | |
1058 | ||
1059 | for (; i < rbio->nr_pages; i++) { | |
1060 | if (rbio->stripe_pages[i]) | |
1061 | continue; | |
1062 | page = alloc_page(GFP_NOFS | __GFP_HIGHMEM); | |
1063 | if (!page) | |
1064 | return -ENOMEM; | |
1065 | rbio->stripe_pages[i] = page; | |
1066 | } | |
1067 | return 0; | |
1068 | } | |
1069 | ||
1070 | /* | |
1071 | * add a single page from a specific stripe into our list of bios for IO | |
1072 | * this will try to merge into existing bios if possible, and returns | |
1073 | * zero if all went well. | |
1074 | */ | |
48a3b636 ES |
1075 | static int rbio_add_io_page(struct btrfs_raid_bio *rbio, |
1076 | struct bio_list *bio_list, | |
1077 | struct page *page, | |
1078 | int stripe_nr, | |
1079 | unsigned long page_index, | |
1080 | unsigned long bio_max_len) | |
53b381b3 DW |
1081 | { |
1082 | struct bio *last = bio_list->tail; | |
1083 | u64 last_end = 0; | |
1084 | int ret; | |
1085 | struct bio *bio; | |
1086 | struct btrfs_bio_stripe *stripe; | |
1087 | u64 disk_start; | |
1088 | ||
1089 | stripe = &rbio->bbio->stripes[stripe_nr]; | |
1090 | disk_start = stripe->physical + (page_index << PAGE_CACHE_SHIFT); | |
1091 | ||
1092 | /* if the device is missing, just fail this stripe */ | |
1093 | if (!stripe->dev->bdev) | |
1094 | return fail_rbio_index(rbio, stripe_nr); | |
1095 | ||
1096 | /* see if we can add this page onto our existing bio */ | |
1097 | if (last) { | |
4f024f37 KO |
1098 | last_end = (u64)last->bi_iter.bi_sector << 9; |
1099 | last_end += last->bi_iter.bi_size; | |
53b381b3 DW |
1100 | |
1101 | /* | |
1102 | * we can't merge these if they are from different | |
1103 | * devices or if they are not contiguous | |
1104 | */ | |
1105 | if (last_end == disk_start && stripe->dev->bdev && | |
1106 | test_bit(BIO_UPTODATE, &last->bi_flags) && | |
1107 | last->bi_bdev == stripe->dev->bdev) { | |
1108 | ret = bio_add_page(last, page, PAGE_CACHE_SIZE, 0); | |
1109 | if (ret == PAGE_CACHE_SIZE) | |
1110 | return 0; | |
1111 | } | |
1112 | } | |
1113 | ||
1114 | /* put a new bio on the list */ | |
9be3395b | 1115 | bio = btrfs_io_bio_alloc(GFP_NOFS, bio_max_len >> PAGE_SHIFT?:1); |
53b381b3 DW |
1116 | if (!bio) |
1117 | return -ENOMEM; | |
1118 | ||
4f024f37 | 1119 | bio->bi_iter.bi_size = 0; |
53b381b3 | 1120 | bio->bi_bdev = stripe->dev->bdev; |
4f024f37 | 1121 | bio->bi_iter.bi_sector = disk_start >> 9; |
53b381b3 DW |
1122 | set_bit(BIO_UPTODATE, &bio->bi_flags); |
1123 | ||
1124 | bio_add_page(bio, page, PAGE_CACHE_SIZE, 0); | |
1125 | bio_list_add(bio_list, bio); | |
1126 | return 0; | |
1127 | } | |
1128 | ||
1129 | /* | |
1130 | * while we're doing the read/modify/write cycle, we could | |
1131 | * have errors in reading pages off the disk. This checks | |
1132 | * for errors and if we're not able to read the page it'll | |
1133 | * trigger parity reconstruction. The rmw will be finished | |
1134 | * after we've reconstructed the failed stripes | |
1135 | */ | |
1136 | static void validate_rbio_for_rmw(struct btrfs_raid_bio *rbio) | |
1137 | { | |
1138 | if (rbio->faila >= 0 || rbio->failb >= 0) { | |
2c8cdd6e | 1139 | BUG_ON(rbio->faila == rbio->real_stripes - 1); |
53b381b3 DW |
1140 | __raid56_parity_recover(rbio); |
1141 | } else { | |
1142 | finish_rmw(rbio); | |
1143 | } | |
1144 | } | |
1145 | ||
1146 | /* | |
1147 | * these are just the pages from the rbio array, not from anything | |
1148 | * the FS sent down to us | |
1149 | */ | |
1150 | static struct page *rbio_stripe_page(struct btrfs_raid_bio *rbio, int stripe, int page) | |
1151 | { | |
1152 | int index; | |
1153 | index = stripe * (rbio->stripe_len >> PAGE_CACHE_SHIFT); | |
1154 | index += page; | |
1155 | return rbio->stripe_pages[index]; | |
1156 | } | |
1157 | ||
1158 | /* | |
1159 | * helper function to walk our bio list and populate the bio_pages array with | |
1160 | * the result. This seems expensive, but it is faster than constantly | |
1161 | * searching through the bio list as we setup the IO in finish_rmw or stripe | |
1162 | * reconstruction. | |
1163 | * | |
1164 | * This must be called before you trust the answers from page_in_rbio | |
1165 | */ | |
1166 | static void index_rbio_pages(struct btrfs_raid_bio *rbio) | |
1167 | { | |
1168 | struct bio *bio; | |
1169 | u64 start; | |
1170 | unsigned long stripe_offset; | |
1171 | unsigned long page_index; | |
1172 | struct page *p; | |
1173 | int i; | |
1174 | ||
1175 | spin_lock_irq(&rbio->bio_list_lock); | |
1176 | bio_list_for_each(bio, &rbio->bio_list) { | |
4f024f37 | 1177 | start = (u64)bio->bi_iter.bi_sector << 9; |
53b381b3 DW |
1178 | stripe_offset = start - rbio->raid_map[0]; |
1179 | page_index = stripe_offset >> PAGE_CACHE_SHIFT; | |
1180 | ||
1181 | for (i = 0; i < bio->bi_vcnt; i++) { | |
1182 | p = bio->bi_io_vec[i].bv_page; | |
1183 | rbio->bio_pages[page_index + i] = p; | |
1184 | } | |
1185 | } | |
1186 | spin_unlock_irq(&rbio->bio_list_lock); | |
1187 | } | |
1188 | ||
1189 | /* | |
1190 | * this is called from one of two situations. We either | |
1191 | * have a full stripe from the higher layers, or we've read all | |
1192 | * the missing bits off disk. | |
1193 | * | |
1194 | * This will calculate the parity and then send down any | |
1195 | * changed blocks. | |
1196 | */ | |
1197 | static noinline void finish_rmw(struct btrfs_raid_bio *rbio) | |
1198 | { | |
1199 | struct btrfs_bio *bbio = rbio->bbio; | |
2c8cdd6e | 1200 | void *pointers[rbio->real_stripes]; |
53b381b3 DW |
1201 | int stripe_len = rbio->stripe_len; |
1202 | int nr_data = rbio->nr_data; | |
1203 | int stripe; | |
1204 | int pagenr; | |
1205 | int p_stripe = -1; | |
1206 | int q_stripe = -1; | |
1207 | struct bio_list bio_list; | |
1208 | struct bio *bio; | |
1209 | int pages_per_stripe = stripe_len >> PAGE_CACHE_SHIFT; | |
1210 | int ret; | |
1211 | ||
1212 | bio_list_init(&bio_list); | |
1213 | ||
2c8cdd6e MX |
1214 | if (rbio->real_stripes - rbio->nr_data == 1) { |
1215 | p_stripe = rbio->real_stripes - 1; | |
1216 | } else if (rbio->real_stripes - rbio->nr_data == 2) { | |
1217 | p_stripe = rbio->real_stripes - 2; | |
1218 | q_stripe = rbio->real_stripes - 1; | |
53b381b3 DW |
1219 | } else { |
1220 | BUG(); | |
1221 | } | |
1222 | ||
1223 | /* at this point we either have a full stripe, | |
1224 | * or we've read the full stripe from the drive. | |
1225 | * recalculate the parity and write the new results. | |
1226 | * | |
1227 | * We're not allowed to add any new bios to the | |
1228 | * bio list here, anyone else that wants to | |
1229 | * change this stripe needs to do their own rmw. | |
1230 | */ | |
1231 | spin_lock_irq(&rbio->bio_list_lock); | |
1232 | set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags); | |
1233 | spin_unlock_irq(&rbio->bio_list_lock); | |
1234 | ||
b89e1b01 | 1235 | atomic_set(&rbio->error, 0); |
53b381b3 DW |
1236 | |
1237 | /* | |
1238 | * now that we've set rmw_locked, run through the | |
1239 | * bio list one last time and map the page pointers | |
4ae10b3a CM |
1240 | * |
1241 | * We don't cache full rbios because we're assuming | |
1242 | * the higher layers are unlikely to use this area of | |
1243 | * the disk again soon. If they do use it again, | |
1244 | * hopefully they will send another full bio. | |
53b381b3 DW |
1245 | */ |
1246 | index_rbio_pages(rbio); | |
4ae10b3a CM |
1247 | if (!rbio_is_full(rbio)) |
1248 | cache_rbio_pages(rbio); | |
1249 | else | |
1250 | clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags); | |
53b381b3 DW |
1251 | |
1252 | for (pagenr = 0; pagenr < pages_per_stripe; pagenr++) { | |
1253 | struct page *p; | |
1254 | /* first collect one page from each data stripe */ | |
1255 | for (stripe = 0; stripe < nr_data; stripe++) { | |
1256 | p = page_in_rbio(rbio, stripe, pagenr, 0); | |
1257 | pointers[stripe] = kmap(p); | |
1258 | } | |
1259 | ||
1260 | /* then add the parity stripe */ | |
1261 | p = rbio_pstripe_page(rbio, pagenr); | |
1262 | SetPageUptodate(p); | |
1263 | pointers[stripe++] = kmap(p); | |
1264 | ||
1265 | if (q_stripe != -1) { | |
1266 | ||
1267 | /* | |
1268 | * raid6, add the qstripe and call the | |
1269 | * library function to fill in our p/q | |
1270 | */ | |
1271 | p = rbio_qstripe_page(rbio, pagenr); | |
1272 | SetPageUptodate(p); | |
1273 | pointers[stripe++] = kmap(p); | |
1274 | ||
2c8cdd6e | 1275 | raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE, |
53b381b3 DW |
1276 | pointers); |
1277 | } else { | |
1278 | /* raid5 */ | |
1279 | memcpy(pointers[nr_data], pointers[0], PAGE_SIZE); | |
1280 | run_xor(pointers + 1, nr_data - 1, PAGE_CACHE_SIZE); | |
1281 | } | |
1282 | ||
1283 | ||
2c8cdd6e | 1284 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) |
53b381b3 DW |
1285 | kunmap(page_in_rbio(rbio, stripe, pagenr, 0)); |
1286 | } | |
1287 | ||
1288 | /* | |
1289 | * time to start writing. Make bios for everything from the | |
1290 | * higher layers (the bio_list in our rbio) and our p/q. Ignore | |
1291 | * everything else. | |
1292 | */ | |
2c8cdd6e | 1293 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) { |
53b381b3 DW |
1294 | for (pagenr = 0; pagenr < pages_per_stripe; pagenr++) { |
1295 | struct page *page; | |
1296 | if (stripe < rbio->nr_data) { | |
1297 | page = page_in_rbio(rbio, stripe, pagenr, 1); | |
1298 | if (!page) | |
1299 | continue; | |
1300 | } else { | |
1301 | page = rbio_stripe_page(rbio, stripe, pagenr); | |
1302 | } | |
1303 | ||
1304 | ret = rbio_add_io_page(rbio, &bio_list, | |
1305 | page, stripe, pagenr, rbio->stripe_len); | |
1306 | if (ret) | |
1307 | goto cleanup; | |
1308 | } | |
1309 | } | |
1310 | ||
2c8cdd6e MX |
1311 | if (likely(!bbio->num_tgtdevs)) |
1312 | goto write_data; | |
1313 | ||
1314 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) { | |
1315 | if (!bbio->tgtdev_map[stripe]) | |
1316 | continue; | |
1317 | ||
1318 | for (pagenr = 0; pagenr < pages_per_stripe; pagenr++) { | |
1319 | struct page *page; | |
1320 | if (stripe < rbio->nr_data) { | |
1321 | page = page_in_rbio(rbio, stripe, pagenr, 1); | |
1322 | if (!page) | |
1323 | continue; | |
1324 | } else { | |
1325 | page = rbio_stripe_page(rbio, stripe, pagenr); | |
1326 | } | |
1327 | ||
1328 | ret = rbio_add_io_page(rbio, &bio_list, page, | |
1329 | rbio->bbio->tgtdev_map[stripe], | |
1330 | pagenr, rbio->stripe_len); | |
1331 | if (ret) | |
1332 | goto cleanup; | |
1333 | } | |
1334 | } | |
1335 | ||
1336 | write_data: | |
b89e1b01 MX |
1337 | atomic_set(&rbio->stripes_pending, bio_list_size(&bio_list)); |
1338 | BUG_ON(atomic_read(&rbio->stripes_pending) == 0); | |
53b381b3 DW |
1339 | |
1340 | while (1) { | |
1341 | bio = bio_list_pop(&bio_list); | |
1342 | if (!bio) | |
1343 | break; | |
1344 | ||
1345 | bio->bi_private = rbio; | |
1346 | bio->bi_end_io = raid_write_end_io; | |
1347 | BUG_ON(!test_bit(BIO_UPTODATE, &bio->bi_flags)); | |
1348 | submit_bio(WRITE, bio); | |
1349 | } | |
1350 | return; | |
1351 | ||
1352 | cleanup: | |
1353 | rbio_orig_end_io(rbio, -EIO, 0); | |
1354 | } | |
1355 | ||
1356 | /* | |
1357 | * helper to find the stripe number for a given bio. Used to figure out which | |
1358 | * stripe has failed. This expects the bio to correspond to a physical disk, | |
1359 | * so it looks up based on physical sector numbers. | |
1360 | */ | |
1361 | static int find_bio_stripe(struct btrfs_raid_bio *rbio, | |
1362 | struct bio *bio) | |
1363 | { | |
4f024f37 | 1364 | u64 physical = bio->bi_iter.bi_sector; |
53b381b3 DW |
1365 | u64 stripe_start; |
1366 | int i; | |
1367 | struct btrfs_bio_stripe *stripe; | |
1368 | ||
1369 | physical <<= 9; | |
1370 | ||
1371 | for (i = 0; i < rbio->bbio->num_stripes; i++) { | |
1372 | stripe = &rbio->bbio->stripes[i]; | |
1373 | stripe_start = stripe->physical; | |
1374 | if (physical >= stripe_start && | |
2c8cdd6e MX |
1375 | physical < stripe_start + rbio->stripe_len && |
1376 | bio->bi_bdev == stripe->dev->bdev) { | |
53b381b3 DW |
1377 | return i; |
1378 | } | |
1379 | } | |
1380 | return -1; | |
1381 | } | |
1382 | ||
1383 | /* | |
1384 | * helper to find the stripe number for a given | |
1385 | * bio (before mapping). Used to figure out which stripe has | |
1386 | * failed. This looks up based on logical block numbers. | |
1387 | */ | |
1388 | static int find_logical_bio_stripe(struct btrfs_raid_bio *rbio, | |
1389 | struct bio *bio) | |
1390 | { | |
4f024f37 | 1391 | u64 logical = bio->bi_iter.bi_sector; |
53b381b3 DW |
1392 | u64 stripe_start; |
1393 | int i; | |
1394 | ||
1395 | logical <<= 9; | |
1396 | ||
1397 | for (i = 0; i < rbio->nr_data; i++) { | |
1398 | stripe_start = rbio->raid_map[i]; | |
1399 | if (logical >= stripe_start && | |
1400 | logical < stripe_start + rbio->stripe_len) { | |
1401 | return i; | |
1402 | } | |
1403 | } | |
1404 | return -1; | |
1405 | } | |
1406 | ||
1407 | /* | |
1408 | * returns -EIO if we had too many failures | |
1409 | */ | |
1410 | static int fail_rbio_index(struct btrfs_raid_bio *rbio, int failed) | |
1411 | { | |
1412 | unsigned long flags; | |
1413 | int ret = 0; | |
1414 | ||
1415 | spin_lock_irqsave(&rbio->bio_list_lock, flags); | |
1416 | ||
1417 | /* we already know this stripe is bad, move on */ | |
1418 | if (rbio->faila == failed || rbio->failb == failed) | |
1419 | goto out; | |
1420 | ||
1421 | if (rbio->faila == -1) { | |
1422 | /* first failure on this rbio */ | |
1423 | rbio->faila = failed; | |
b89e1b01 | 1424 | atomic_inc(&rbio->error); |
53b381b3 DW |
1425 | } else if (rbio->failb == -1) { |
1426 | /* second failure on this rbio */ | |
1427 | rbio->failb = failed; | |
b89e1b01 | 1428 | atomic_inc(&rbio->error); |
53b381b3 DW |
1429 | } else { |
1430 | ret = -EIO; | |
1431 | } | |
1432 | out: | |
1433 | spin_unlock_irqrestore(&rbio->bio_list_lock, flags); | |
1434 | ||
1435 | return ret; | |
1436 | } | |
1437 | ||
1438 | /* | |
1439 | * helper to fail a stripe based on a physical disk | |
1440 | * bio. | |
1441 | */ | |
1442 | static int fail_bio_stripe(struct btrfs_raid_bio *rbio, | |
1443 | struct bio *bio) | |
1444 | { | |
1445 | int failed = find_bio_stripe(rbio, bio); | |
1446 | ||
1447 | if (failed < 0) | |
1448 | return -EIO; | |
1449 | ||
1450 | return fail_rbio_index(rbio, failed); | |
1451 | } | |
1452 | ||
1453 | /* | |
1454 | * this sets each page in the bio uptodate. It should only be used on private | |
1455 | * rbio pages, nothing that comes in from the higher layers | |
1456 | */ | |
1457 | static void set_bio_pages_uptodate(struct bio *bio) | |
1458 | { | |
1459 | int i; | |
1460 | struct page *p; | |
1461 | ||
1462 | for (i = 0; i < bio->bi_vcnt; i++) { | |
1463 | p = bio->bi_io_vec[i].bv_page; | |
1464 | SetPageUptodate(p); | |
1465 | } | |
1466 | } | |
1467 | ||
1468 | /* | |
1469 | * end io for the read phase of the rmw cycle. All the bios here are physical | |
1470 | * stripe bios we've read from the disk so we can recalculate the parity of the | |
1471 | * stripe. | |
1472 | * | |
1473 | * This will usually kick off finish_rmw once all the bios are read in, but it | |
1474 | * may trigger parity reconstruction if we had any errors along the way | |
1475 | */ | |
1476 | static void raid_rmw_end_io(struct bio *bio, int err) | |
1477 | { | |
1478 | struct btrfs_raid_bio *rbio = bio->bi_private; | |
1479 | ||
1480 | if (err) | |
1481 | fail_bio_stripe(rbio, bio); | |
1482 | else | |
1483 | set_bio_pages_uptodate(bio); | |
1484 | ||
1485 | bio_put(bio); | |
1486 | ||
b89e1b01 | 1487 | if (!atomic_dec_and_test(&rbio->stripes_pending)) |
53b381b3 DW |
1488 | return; |
1489 | ||
1490 | err = 0; | |
b89e1b01 | 1491 | if (atomic_read(&rbio->error) > rbio->bbio->max_errors) |
53b381b3 DW |
1492 | goto cleanup; |
1493 | ||
1494 | /* | |
1495 | * this will normally call finish_rmw to start our write | |
1496 | * but if there are any failed stripes we'll reconstruct | |
1497 | * from parity first | |
1498 | */ | |
1499 | validate_rbio_for_rmw(rbio); | |
1500 | return; | |
1501 | ||
1502 | cleanup: | |
1503 | ||
1504 | rbio_orig_end_io(rbio, -EIO, 0); | |
1505 | } | |
1506 | ||
1507 | static void async_rmw_stripe(struct btrfs_raid_bio *rbio) | |
1508 | { | |
9e0af237 LB |
1509 | btrfs_init_work(&rbio->work, btrfs_rmw_helper, |
1510 | rmw_work, NULL, NULL); | |
53b381b3 | 1511 | |
d05a33ac QW |
1512 | btrfs_queue_work(rbio->fs_info->rmw_workers, |
1513 | &rbio->work); | |
53b381b3 DW |
1514 | } |
1515 | ||
1516 | static void async_read_rebuild(struct btrfs_raid_bio *rbio) | |
1517 | { | |
9e0af237 LB |
1518 | btrfs_init_work(&rbio->work, btrfs_rmw_helper, |
1519 | read_rebuild_work, NULL, NULL); | |
53b381b3 | 1520 | |
d05a33ac QW |
1521 | btrfs_queue_work(rbio->fs_info->rmw_workers, |
1522 | &rbio->work); | |
53b381b3 DW |
1523 | } |
1524 | ||
1525 | /* | |
1526 | * the stripe must be locked by the caller. It will | |
1527 | * unlock after all the writes are done | |
1528 | */ | |
1529 | static int raid56_rmw_stripe(struct btrfs_raid_bio *rbio) | |
1530 | { | |
1531 | int bios_to_read = 0; | |
53b381b3 DW |
1532 | struct bio_list bio_list; |
1533 | int ret; | |
ed6078f7 | 1534 | int nr_pages = DIV_ROUND_UP(rbio->stripe_len, PAGE_CACHE_SIZE); |
53b381b3 DW |
1535 | int pagenr; |
1536 | int stripe; | |
1537 | struct bio *bio; | |
1538 | ||
1539 | bio_list_init(&bio_list); | |
1540 | ||
1541 | ret = alloc_rbio_pages(rbio); | |
1542 | if (ret) | |
1543 | goto cleanup; | |
1544 | ||
1545 | index_rbio_pages(rbio); | |
1546 | ||
b89e1b01 | 1547 | atomic_set(&rbio->error, 0); |
53b381b3 DW |
1548 | /* |
1549 | * build a list of bios to read all the missing parts of this | |
1550 | * stripe | |
1551 | */ | |
1552 | for (stripe = 0; stripe < rbio->nr_data; stripe++) { | |
1553 | for (pagenr = 0; pagenr < nr_pages; pagenr++) { | |
1554 | struct page *page; | |
1555 | /* | |
1556 | * we want to find all the pages missing from | |
1557 | * the rbio and read them from the disk. If | |
1558 | * page_in_rbio finds a page in the bio list | |
1559 | * we don't need to read it off the stripe. | |
1560 | */ | |
1561 | page = page_in_rbio(rbio, stripe, pagenr, 1); | |
1562 | if (page) | |
1563 | continue; | |
1564 | ||
1565 | page = rbio_stripe_page(rbio, stripe, pagenr); | |
4ae10b3a CM |
1566 | /* |
1567 | * the bio cache may have handed us an uptodate | |
1568 | * page. If so, be happy and use it | |
1569 | */ | |
1570 | if (PageUptodate(page)) | |
1571 | continue; | |
1572 | ||
53b381b3 DW |
1573 | ret = rbio_add_io_page(rbio, &bio_list, page, |
1574 | stripe, pagenr, rbio->stripe_len); | |
1575 | if (ret) | |
1576 | goto cleanup; | |
1577 | } | |
1578 | } | |
1579 | ||
1580 | bios_to_read = bio_list_size(&bio_list); | |
1581 | if (!bios_to_read) { | |
1582 | /* | |
1583 | * this can happen if others have merged with | |
1584 | * us, it means there is nothing left to read. | |
1585 | * But if there are missing devices it may not be | |
1586 | * safe to do the full stripe write yet. | |
1587 | */ | |
1588 | goto finish; | |
1589 | } | |
1590 | ||
1591 | /* | |
1592 | * the bbio may be freed once we submit the last bio. Make sure | |
1593 | * not to touch it after that | |
1594 | */ | |
b89e1b01 | 1595 | atomic_set(&rbio->stripes_pending, bios_to_read); |
53b381b3 DW |
1596 | while (1) { |
1597 | bio = bio_list_pop(&bio_list); | |
1598 | if (!bio) | |
1599 | break; | |
1600 | ||
1601 | bio->bi_private = rbio; | |
1602 | bio->bi_end_io = raid_rmw_end_io; | |
1603 | ||
1604 | btrfs_bio_wq_end_io(rbio->fs_info, bio, | |
1605 | BTRFS_WQ_ENDIO_RAID56); | |
1606 | ||
1607 | BUG_ON(!test_bit(BIO_UPTODATE, &bio->bi_flags)); | |
1608 | submit_bio(READ, bio); | |
1609 | } | |
1610 | /* the actual write will happen once the reads are done */ | |
1611 | return 0; | |
1612 | ||
1613 | cleanup: | |
1614 | rbio_orig_end_io(rbio, -EIO, 0); | |
1615 | return -EIO; | |
1616 | ||
1617 | finish: | |
1618 | validate_rbio_for_rmw(rbio); | |
1619 | return 0; | |
1620 | } | |
1621 | ||
1622 | /* | |
1623 | * if the upper layers pass in a full stripe, we thank them by only allocating | |
1624 | * enough pages to hold the parity, and sending it all down quickly. | |
1625 | */ | |
1626 | static int full_stripe_write(struct btrfs_raid_bio *rbio) | |
1627 | { | |
1628 | int ret; | |
1629 | ||
1630 | ret = alloc_rbio_parity_pages(rbio); | |
3cd846d1 MX |
1631 | if (ret) { |
1632 | __free_raid_bio(rbio); | |
53b381b3 | 1633 | return ret; |
3cd846d1 | 1634 | } |
53b381b3 DW |
1635 | |
1636 | ret = lock_stripe_add(rbio); | |
1637 | if (ret == 0) | |
1638 | finish_rmw(rbio); | |
1639 | return 0; | |
1640 | } | |
1641 | ||
1642 | /* | |
1643 | * partial stripe writes get handed over to async helpers. | |
1644 | * We're really hoping to merge a few more writes into this | |
1645 | * rbio before calculating new parity | |
1646 | */ | |
1647 | static int partial_stripe_write(struct btrfs_raid_bio *rbio) | |
1648 | { | |
1649 | int ret; | |
1650 | ||
1651 | ret = lock_stripe_add(rbio); | |
1652 | if (ret == 0) | |
1653 | async_rmw_stripe(rbio); | |
1654 | return 0; | |
1655 | } | |
1656 | ||
1657 | /* | |
1658 | * sometimes while we were reading from the drive to | |
1659 | * recalculate parity, enough new bios come into create | |
1660 | * a full stripe. So we do a check here to see if we can | |
1661 | * go directly to finish_rmw | |
1662 | */ | |
1663 | static int __raid56_parity_write(struct btrfs_raid_bio *rbio) | |
1664 | { | |
1665 | /* head off into rmw land if we don't have a full stripe */ | |
1666 | if (!rbio_is_full(rbio)) | |
1667 | return partial_stripe_write(rbio); | |
1668 | return full_stripe_write(rbio); | |
1669 | } | |
1670 | ||
6ac0f488 CM |
1671 | /* |
1672 | * We use plugging call backs to collect full stripes. | |
1673 | * Any time we get a partial stripe write while plugged | |
1674 | * we collect it into a list. When the unplug comes down, | |
1675 | * we sort the list by logical block number and merge | |
1676 | * everything we can into the same rbios | |
1677 | */ | |
1678 | struct btrfs_plug_cb { | |
1679 | struct blk_plug_cb cb; | |
1680 | struct btrfs_fs_info *info; | |
1681 | struct list_head rbio_list; | |
1682 | struct btrfs_work work; | |
1683 | }; | |
1684 | ||
1685 | /* | |
1686 | * rbios on the plug list are sorted for easier merging. | |
1687 | */ | |
1688 | static int plug_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1689 | { | |
1690 | struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio, | |
1691 | plug_list); | |
1692 | struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio, | |
1693 | plug_list); | |
4f024f37 KO |
1694 | u64 a_sector = ra->bio_list.head->bi_iter.bi_sector; |
1695 | u64 b_sector = rb->bio_list.head->bi_iter.bi_sector; | |
6ac0f488 CM |
1696 | |
1697 | if (a_sector < b_sector) | |
1698 | return -1; | |
1699 | if (a_sector > b_sector) | |
1700 | return 1; | |
1701 | return 0; | |
1702 | } | |
1703 | ||
1704 | static void run_plug(struct btrfs_plug_cb *plug) | |
1705 | { | |
1706 | struct btrfs_raid_bio *cur; | |
1707 | struct btrfs_raid_bio *last = NULL; | |
1708 | ||
1709 | /* | |
1710 | * sort our plug list then try to merge | |
1711 | * everything we can in hopes of creating full | |
1712 | * stripes. | |
1713 | */ | |
1714 | list_sort(NULL, &plug->rbio_list, plug_cmp); | |
1715 | while (!list_empty(&plug->rbio_list)) { | |
1716 | cur = list_entry(plug->rbio_list.next, | |
1717 | struct btrfs_raid_bio, plug_list); | |
1718 | list_del_init(&cur->plug_list); | |
1719 | ||
1720 | if (rbio_is_full(cur)) { | |
1721 | /* we have a full stripe, send it down */ | |
1722 | full_stripe_write(cur); | |
1723 | continue; | |
1724 | } | |
1725 | if (last) { | |
1726 | if (rbio_can_merge(last, cur)) { | |
1727 | merge_rbio(last, cur); | |
1728 | __free_raid_bio(cur); | |
1729 | continue; | |
1730 | ||
1731 | } | |
1732 | __raid56_parity_write(last); | |
1733 | } | |
1734 | last = cur; | |
1735 | } | |
1736 | if (last) { | |
1737 | __raid56_parity_write(last); | |
1738 | } | |
1739 | kfree(plug); | |
1740 | } | |
1741 | ||
1742 | /* | |
1743 | * if the unplug comes from schedule, we have to push the | |
1744 | * work off to a helper thread | |
1745 | */ | |
1746 | static void unplug_work(struct btrfs_work *work) | |
1747 | { | |
1748 | struct btrfs_plug_cb *plug; | |
1749 | plug = container_of(work, struct btrfs_plug_cb, work); | |
1750 | run_plug(plug); | |
1751 | } | |
1752 | ||
1753 | static void btrfs_raid_unplug(struct blk_plug_cb *cb, bool from_schedule) | |
1754 | { | |
1755 | struct btrfs_plug_cb *plug; | |
1756 | plug = container_of(cb, struct btrfs_plug_cb, cb); | |
1757 | ||
1758 | if (from_schedule) { | |
9e0af237 LB |
1759 | btrfs_init_work(&plug->work, btrfs_rmw_helper, |
1760 | unplug_work, NULL, NULL); | |
d05a33ac QW |
1761 | btrfs_queue_work(plug->info->rmw_workers, |
1762 | &plug->work); | |
6ac0f488 CM |
1763 | return; |
1764 | } | |
1765 | run_plug(plug); | |
1766 | } | |
1767 | ||
53b381b3 DW |
1768 | /* |
1769 | * our main entry point for writes from the rest of the FS. | |
1770 | */ | |
1771 | int raid56_parity_write(struct btrfs_root *root, struct bio *bio, | |
1772 | struct btrfs_bio *bbio, u64 *raid_map, | |
1773 | u64 stripe_len) | |
1774 | { | |
1775 | struct btrfs_raid_bio *rbio; | |
6ac0f488 CM |
1776 | struct btrfs_plug_cb *plug = NULL; |
1777 | struct blk_plug_cb *cb; | |
53b381b3 DW |
1778 | |
1779 | rbio = alloc_rbio(root, bbio, raid_map, stripe_len); | |
af8e2d1d MX |
1780 | if (IS_ERR(rbio)) { |
1781 | __free_bbio_and_raid_map(bbio, raid_map, 1); | |
53b381b3 | 1782 | return PTR_ERR(rbio); |
af8e2d1d | 1783 | } |
53b381b3 | 1784 | bio_list_add(&rbio->bio_list, bio); |
4f024f37 | 1785 | rbio->bio_list_bytes = bio->bi_iter.bi_size; |
1b94b556 | 1786 | rbio->operation = BTRFS_RBIO_WRITE; |
6ac0f488 CM |
1787 | |
1788 | /* | |
1789 | * don't plug on full rbios, just get them out the door | |
1790 | * as quickly as we can | |
1791 | */ | |
1792 | if (rbio_is_full(rbio)) | |
1793 | return full_stripe_write(rbio); | |
1794 | ||
1795 | cb = blk_check_plugged(btrfs_raid_unplug, root->fs_info, | |
1796 | sizeof(*plug)); | |
1797 | if (cb) { | |
1798 | plug = container_of(cb, struct btrfs_plug_cb, cb); | |
1799 | if (!plug->info) { | |
1800 | plug->info = root->fs_info; | |
1801 | INIT_LIST_HEAD(&plug->rbio_list); | |
1802 | } | |
1803 | list_add_tail(&rbio->plug_list, &plug->rbio_list); | |
1804 | } else { | |
1805 | return __raid56_parity_write(rbio); | |
1806 | } | |
1807 | return 0; | |
53b381b3 DW |
1808 | } |
1809 | ||
1810 | /* | |
1811 | * all parity reconstruction happens here. We've read in everything | |
1812 | * we can find from the drives and this does the heavy lifting of | |
1813 | * sorting the good from the bad. | |
1814 | */ | |
1815 | static void __raid_recover_end_io(struct btrfs_raid_bio *rbio) | |
1816 | { | |
1817 | int pagenr, stripe; | |
1818 | void **pointers; | |
1819 | int faila = -1, failb = -1; | |
ed6078f7 | 1820 | int nr_pages = DIV_ROUND_UP(rbio->stripe_len, PAGE_CACHE_SIZE); |
53b381b3 DW |
1821 | struct page *page; |
1822 | int err; | |
1823 | int i; | |
1824 | ||
2c8cdd6e | 1825 | pointers = kzalloc(rbio->real_stripes * sizeof(void *), |
53b381b3 DW |
1826 | GFP_NOFS); |
1827 | if (!pointers) { | |
1828 | err = -ENOMEM; | |
1829 | goto cleanup_io; | |
1830 | } | |
1831 | ||
1832 | faila = rbio->faila; | |
1833 | failb = rbio->failb; | |
1834 | ||
1b94b556 | 1835 | if (rbio->operation == BTRFS_RBIO_READ_REBUILD) { |
53b381b3 DW |
1836 | spin_lock_irq(&rbio->bio_list_lock); |
1837 | set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags); | |
1838 | spin_unlock_irq(&rbio->bio_list_lock); | |
1839 | } | |
1840 | ||
1841 | index_rbio_pages(rbio); | |
1842 | ||
1843 | for (pagenr = 0; pagenr < nr_pages; pagenr++) { | |
5a6ac9ea MX |
1844 | /* |
1845 | * Now we just use bitmap to mark the horizontal stripes in | |
1846 | * which we have data when doing parity scrub. | |
1847 | */ | |
1848 | if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB && | |
1849 | !test_bit(pagenr, rbio->dbitmap)) | |
1850 | continue; | |
1851 | ||
53b381b3 DW |
1852 | /* setup our array of pointers with pages |
1853 | * from each stripe | |
1854 | */ | |
2c8cdd6e | 1855 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) { |
53b381b3 DW |
1856 | /* |
1857 | * if we're rebuilding a read, we have to use | |
1858 | * pages from the bio list | |
1859 | */ | |
1b94b556 | 1860 | if (rbio->operation == BTRFS_RBIO_READ_REBUILD && |
53b381b3 DW |
1861 | (stripe == faila || stripe == failb)) { |
1862 | page = page_in_rbio(rbio, stripe, pagenr, 0); | |
1863 | } else { | |
1864 | page = rbio_stripe_page(rbio, stripe, pagenr); | |
1865 | } | |
1866 | pointers[stripe] = kmap(page); | |
1867 | } | |
1868 | ||
1869 | /* all raid6 handling here */ | |
2c8cdd6e | 1870 | if (rbio->raid_map[rbio->real_stripes - 1] == |
53b381b3 DW |
1871 | RAID6_Q_STRIPE) { |
1872 | ||
1873 | /* | |
1874 | * single failure, rebuild from parity raid5 | |
1875 | * style | |
1876 | */ | |
1877 | if (failb < 0) { | |
1878 | if (faila == rbio->nr_data) { | |
1879 | /* | |
1880 | * Just the P stripe has failed, without | |
1881 | * a bad data or Q stripe. | |
1882 | * TODO, we should redo the xor here. | |
1883 | */ | |
1884 | err = -EIO; | |
1885 | goto cleanup; | |
1886 | } | |
1887 | /* | |
1888 | * a single failure in raid6 is rebuilt | |
1889 | * in the pstripe code below | |
1890 | */ | |
1891 | goto pstripe; | |
1892 | } | |
1893 | ||
1894 | /* make sure our ps and qs are in order */ | |
1895 | if (faila > failb) { | |
1896 | int tmp = failb; | |
1897 | failb = faila; | |
1898 | faila = tmp; | |
1899 | } | |
1900 | ||
1901 | /* if the q stripe is failed, do a pstripe reconstruction | |
1902 | * from the xors. | |
1903 | * If both the q stripe and the P stripe are failed, we're | |
1904 | * here due to a crc mismatch and we can't give them the | |
1905 | * data they want | |
1906 | */ | |
1907 | if (rbio->raid_map[failb] == RAID6_Q_STRIPE) { | |
1908 | if (rbio->raid_map[faila] == RAID5_P_STRIPE) { | |
1909 | err = -EIO; | |
1910 | goto cleanup; | |
1911 | } | |
1912 | /* | |
1913 | * otherwise we have one bad data stripe and | |
1914 | * a good P stripe. raid5! | |
1915 | */ | |
1916 | goto pstripe; | |
1917 | } | |
1918 | ||
1919 | if (rbio->raid_map[failb] == RAID5_P_STRIPE) { | |
2c8cdd6e | 1920 | raid6_datap_recov(rbio->real_stripes, |
53b381b3 DW |
1921 | PAGE_SIZE, faila, pointers); |
1922 | } else { | |
2c8cdd6e | 1923 | raid6_2data_recov(rbio->real_stripes, |
53b381b3 DW |
1924 | PAGE_SIZE, faila, failb, |
1925 | pointers); | |
1926 | } | |
1927 | } else { | |
1928 | void *p; | |
1929 | ||
1930 | /* rebuild from P stripe here (raid5 or raid6) */ | |
1931 | BUG_ON(failb != -1); | |
1932 | pstripe: | |
1933 | /* Copy parity block into failed block to start with */ | |
1934 | memcpy(pointers[faila], | |
1935 | pointers[rbio->nr_data], | |
1936 | PAGE_CACHE_SIZE); | |
1937 | ||
1938 | /* rearrange the pointer array */ | |
1939 | p = pointers[faila]; | |
1940 | for (stripe = faila; stripe < rbio->nr_data - 1; stripe++) | |
1941 | pointers[stripe] = pointers[stripe + 1]; | |
1942 | pointers[rbio->nr_data - 1] = p; | |
1943 | ||
1944 | /* xor in the rest */ | |
1945 | run_xor(pointers, rbio->nr_data - 1, PAGE_CACHE_SIZE); | |
1946 | } | |
1947 | /* if we're doing this rebuild as part of an rmw, go through | |
1948 | * and set all of our private rbio pages in the | |
1949 | * failed stripes as uptodate. This way finish_rmw will | |
1950 | * know they can be trusted. If this was a read reconstruction, | |
1951 | * other endio functions will fiddle the uptodate bits | |
1952 | */ | |
1b94b556 | 1953 | if (rbio->operation == BTRFS_RBIO_WRITE) { |
53b381b3 DW |
1954 | for (i = 0; i < nr_pages; i++) { |
1955 | if (faila != -1) { | |
1956 | page = rbio_stripe_page(rbio, faila, i); | |
1957 | SetPageUptodate(page); | |
1958 | } | |
1959 | if (failb != -1) { | |
1960 | page = rbio_stripe_page(rbio, failb, i); | |
1961 | SetPageUptodate(page); | |
1962 | } | |
1963 | } | |
1964 | } | |
2c8cdd6e | 1965 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) { |
53b381b3 DW |
1966 | /* |
1967 | * if we're rebuilding a read, we have to use | |
1968 | * pages from the bio list | |
1969 | */ | |
1b94b556 | 1970 | if (rbio->operation == BTRFS_RBIO_READ_REBUILD && |
53b381b3 DW |
1971 | (stripe == faila || stripe == failb)) { |
1972 | page = page_in_rbio(rbio, stripe, pagenr, 0); | |
1973 | } else { | |
1974 | page = rbio_stripe_page(rbio, stripe, pagenr); | |
1975 | } | |
1976 | kunmap(page); | |
1977 | } | |
1978 | } | |
1979 | ||
1980 | err = 0; | |
1981 | cleanup: | |
1982 | kfree(pointers); | |
1983 | ||
1984 | cleanup_io: | |
1b94b556 | 1985 | if (rbio->operation == BTRFS_RBIO_READ_REBUILD) { |
af8e2d1d MX |
1986 | if (err == 0 && |
1987 | !test_bit(RBIO_HOLD_BBIO_MAP_BIT, &rbio->flags)) | |
4ae10b3a CM |
1988 | cache_rbio_pages(rbio); |
1989 | else | |
1990 | clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags); | |
1991 | ||
53b381b3 DW |
1992 | rbio_orig_end_io(rbio, err, err == 0); |
1993 | } else if (err == 0) { | |
1994 | rbio->faila = -1; | |
1995 | rbio->failb = -1; | |
5a6ac9ea MX |
1996 | |
1997 | if (rbio->operation == BTRFS_RBIO_WRITE) | |
1998 | finish_rmw(rbio); | |
1999 | else if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB) | |
2000 | finish_parity_scrub(rbio, 0); | |
2001 | else | |
2002 | BUG(); | |
53b381b3 DW |
2003 | } else { |
2004 | rbio_orig_end_io(rbio, err, 0); | |
2005 | } | |
2006 | } | |
2007 | ||
2008 | /* | |
2009 | * This is called only for stripes we've read from disk to | |
2010 | * reconstruct the parity. | |
2011 | */ | |
2012 | static void raid_recover_end_io(struct bio *bio, int err) | |
2013 | { | |
2014 | struct btrfs_raid_bio *rbio = bio->bi_private; | |
2015 | ||
2016 | /* | |
2017 | * we only read stripe pages off the disk, set them | |
2018 | * up to date if there were no errors | |
2019 | */ | |
2020 | if (err) | |
2021 | fail_bio_stripe(rbio, bio); | |
2022 | else | |
2023 | set_bio_pages_uptodate(bio); | |
2024 | bio_put(bio); | |
2025 | ||
b89e1b01 | 2026 | if (!atomic_dec_and_test(&rbio->stripes_pending)) |
53b381b3 DW |
2027 | return; |
2028 | ||
b89e1b01 | 2029 | if (atomic_read(&rbio->error) > rbio->bbio->max_errors) |
53b381b3 DW |
2030 | rbio_orig_end_io(rbio, -EIO, 0); |
2031 | else | |
2032 | __raid_recover_end_io(rbio); | |
2033 | } | |
2034 | ||
2035 | /* | |
2036 | * reads everything we need off the disk to reconstruct | |
2037 | * the parity. endio handlers trigger final reconstruction | |
2038 | * when the IO is done. | |
2039 | * | |
2040 | * This is used both for reads from the higher layers and for | |
2041 | * parity construction required to finish a rmw cycle. | |
2042 | */ | |
2043 | static int __raid56_parity_recover(struct btrfs_raid_bio *rbio) | |
2044 | { | |
2045 | int bios_to_read = 0; | |
53b381b3 DW |
2046 | struct bio_list bio_list; |
2047 | int ret; | |
ed6078f7 | 2048 | int nr_pages = DIV_ROUND_UP(rbio->stripe_len, PAGE_CACHE_SIZE); |
53b381b3 DW |
2049 | int pagenr; |
2050 | int stripe; | |
2051 | struct bio *bio; | |
2052 | ||
2053 | bio_list_init(&bio_list); | |
2054 | ||
2055 | ret = alloc_rbio_pages(rbio); | |
2056 | if (ret) | |
2057 | goto cleanup; | |
2058 | ||
b89e1b01 | 2059 | atomic_set(&rbio->error, 0); |
53b381b3 DW |
2060 | |
2061 | /* | |
4ae10b3a CM |
2062 | * read everything that hasn't failed. Thanks to the |
2063 | * stripe cache, it is possible that some or all of these | |
2064 | * pages are going to be uptodate. | |
53b381b3 | 2065 | */ |
2c8cdd6e | 2066 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) { |
5588383e | 2067 | if (rbio->faila == stripe || rbio->failb == stripe) { |
b89e1b01 | 2068 | atomic_inc(&rbio->error); |
53b381b3 | 2069 | continue; |
5588383e | 2070 | } |
53b381b3 DW |
2071 | |
2072 | for (pagenr = 0; pagenr < nr_pages; pagenr++) { | |
2073 | struct page *p; | |
2074 | ||
2075 | /* | |
2076 | * the rmw code may have already read this | |
2077 | * page in | |
2078 | */ | |
2079 | p = rbio_stripe_page(rbio, stripe, pagenr); | |
2080 | if (PageUptodate(p)) | |
2081 | continue; | |
2082 | ||
2083 | ret = rbio_add_io_page(rbio, &bio_list, | |
2084 | rbio_stripe_page(rbio, stripe, pagenr), | |
2085 | stripe, pagenr, rbio->stripe_len); | |
2086 | if (ret < 0) | |
2087 | goto cleanup; | |
2088 | } | |
2089 | } | |
2090 | ||
2091 | bios_to_read = bio_list_size(&bio_list); | |
2092 | if (!bios_to_read) { | |
2093 | /* | |
2094 | * we might have no bios to read just because the pages | |
2095 | * were up to date, or we might have no bios to read because | |
2096 | * the devices were gone. | |
2097 | */ | |
b89e1b01 | 2098 | if (atomic_read(&rbio->error) <= rbio->bbio->max_errors) { |
53b381b3 DW |
2099 | __raid_recover_end_io(rbio); |
2100 | goto out; | |
2101 | } else { | |
2102 | goto cleanup; | |
2103 | } | |
2104 | } | |
2105 | ||
2106 | /* | |
2107 | * the bbio may be freed once we submit the last bio. Make sure | |
2108 | * not to touch it after that | |
2109 | */ | |
b89e1b01 | 2110 | atomic_set(&rbio->stripes_pending, bios_to_read); |
53b381b3 DW |
2111 | while (1) { |
2112 | bio = bio_list_pop(&bio_list); | |
2113 | if (!bio) | |
2114 | break; | |
2115 | ||
2116 | bio->bi_private = rbio; | |
2117 | bio->bi_end_io = raid_recover_end_io; | |
2118 | ||
2119 | btrfs_bio_wq_end_io(rbio->fs_info, bio, | |
2120 | BTRFS_WQ_ENDIO_RAID56); | |
2121 | ||
2122 | BUG_ON(!test_bit(BIO_UPTODATE, &bio->bi_flags)); | |
2123 | submit_bio(READ, bio); | |
2124 | } | |
2125 | out: | |
2126 | return 0; | |
2127 | ||
2128 | cleanup: | |
1b94b556 | 2129 | if (rbio->operation == BTRFS_RBIO_READ_REBUILD) |
53b381b3 DW |
2130 | rbio_orig_end_io(rbio, -EIO, 0); |
2131 | return -EIO; | |
2132 | } | |
2133 | ||
2134 | /* | |
2135 | * the main entry point for reads from the higher layers. This | |
2136 | * is really only called when the normal read path had a failure, | |
2137 | * so we assume the bio they send down corresponds to a failed part | |
2138 | * of the drive. | |
2139 | */ | |
2140 | int raid56_parity_recover(struct btrfs_root *root, struct bio *bio, | |
2141 | struct btrfs_bio *bbio, u64 *raid_map, | |
af8e2d1d | 2142 | u64 stripe_len, int mirror_num, int hold_bbio) |
53b381b3 DW |
2143 | { |
2144 | struct btrfs_raid_bio *rbio; | |
2145 | int ret; | |
2146 | ||
2147 | rbio = alloc_rbio(root, bbio, raid_map, stripe_len); | |
af8e2d1d MX |
2148 | if (IS_ERR(rbio)) { |
2149 | __free_bbio_and_raid_map(bbio, raid_map, !hold_bbio); | |
53b381b3 | 2150 | return PTR_ERR(rbio); |
af8e2d1d | 2151 | } |
53b381b3 | 2152 | |
af8e2d1d MX |
2153 | if (hold_bbio) |
2154 | set_bit(RBIO_HOLD_BBIO_MAP_BIT, &rbio->flags); | |
1b94b556 | 2155 | rbio->operation = BTRFS_RBIO_READ_REBUILD; |
53b381b3 | 2156 | bio_list_add(&rbio->bio_list, bio); |
4f024f37 | 2157 | rbio->bio_list_bytes = bio->bi_iter.bi_size; |
53b381b3 DW |
2158 | |
2159 | rbio->faila = find_logical_bio_stripe(rbio, bio); | |
2160 | if (rbio->faila == -1) { | |
2161 | BUG(); | |
af8e2d1d | 2162 | __free_bbio_and_raid_map(bbio, raid_map, !hold_bbio); |
53b381b3 DW |
2163 | kfree(rbio); |
2164 | return -EIO; | |
2165 | } | |
2166 | ||
2167 | /* | |
2168 | * reconstruct from the q stripe if they are | |
2169 | * asking for mirror 3 | |
2170 | */ | |
2171 | if (mirror_num == 3) | |
2c8cdd6e | 2172 | rbio->failb = rbio->real_stripes - 2; |
53b381b3 DW |
2173 | |
2174 | ret = lock_stripe_add(rbio); | |
2175 | ||
2176 | /* | |
2177 | * __raid56_parity_recover will end the bio with | |
2178 | * any errors it hits. We don't want to return | |
2179 | * its error value up the stack because our caller | |
2180 | * will end up calling bio_endio with any nonzero | |
2181 | * return | |
2182 | */ | |
2183 | if (ret == 0) | |
2184 | __raid56_parity_recover(rbio); | |
2185 | /* | |
2186 | * our rbio has been added to the list of | |
2187 | * rbios that will be handled after the | |
2188 | * currently lock owner is done | |
2189 | */ | |
2190 | return 0; | |
2191 | ||
2192 | } | |
2193 | ||
2194 | static void rmw_work(struct btrfs_work *work) | |
2195 | { | |
2196 | struct btrfs_raid_bio *rbio; | |
2197 | ||
2198 | rbio = container_of(work, struct btrfs_raid_bio, work); | |
2199 | raid56_rmw_stripe(rbio); | |
2200 | } | |
2201 | ||
2202 | static void read_rebuild_work(struct btrfs_work *work) | |
2203 | { | |
2204 | struct btrfs_raid_bio *rbio; | |
2205 | ||
2206 | rbio = container_of(work, struct btrfs_raid_bio, work); | |
2207 | __raid56_parity_recover(rbio); | |
2208 | } | |
5a6ac9ea MX |
2209 | |
2210 | /* | |
2211 | * The following code is used to scrub/replace the parity stripe | |
2212 | * | |
2213 | * Note: We need make sure all the pages that add into the scrub/replace | |
2214 | * raid bio are correct and not be changed during the scrub/replace. That | |
2215 | * is those pages just hold metadata or file data with checksum. | |
2216 | */ | |
2217 | ||
2218 | struct btrfs_raid_bio * | |
2219 | raid56_parity_alloc_scrub_rbio(struct btrfs_root *root, struct bio *bio, | |
2220 | struct btrfs_bio *bbio, u64 *raid_map, | |
2221 | u64 stripe_len, struct btrfs_device *scrub_dev, | |
2222 | unsigned long *dbitmap, int stripe_nsectors) | |
2223 | { | |
2224 | struct btrfs_raid_bio *rbio; | |
2225 | int i; | |
2226 | ||
2227 | rbio = alloc_rbio(root, bbio, raid_map, stripe_len); | |
2228 | if (IS_ERR(rbio)) | |
2229 | return NULL; | |
2230 | bio_list_add(&rbio->bio_list, bio); | |
2231 | /* | |
2232 | * This is a special bio which is used to hold the completion handler | |
2233 | * and make the scrub rbio is similar to the other types | |
2234 | */ | |
2235 | ASSERT(!bio->bi_iter.bi_size); | |
2236 | rbio->operation = BTRFS_RBIO_PARITY_SCRUB; | |
2237 | ||
2c8cdd6e | 2238 | for (i = 0; i < rbio->real_stripes; i++) { |
5a6ac9ea MX |
2239 | if (bbio->stripes[i].dev == scrub_dev) { |
2240 | rbio->scrubp = i; | |
2241 | break; | |
2242 | } | |
2243 | } | |
2244 | ||
2245 | /* Now we just support the sectorsize equals to page size */ | |
2246 | ASSERT(root->sectorsize == PAGE_SIZE); | |
2247 | ASSERT(rbio->stripe_npages == stripe_nsectors); | |
2248 | bitmap_copy(rbio->dbitmap, dbitmap, stripe_nsectors); | |
2249 | ||
2250 | return rbio; | |
2251 | } | |
2252 | ||
2253 | void raid56_parity_add_scrub_pages(struct btrfs_raid_bio *rbio, | |
2254 | struct page *page, u64 logical) | |
2255 | { | |
2256 | int stripe_offset; | |
2257 | int index; | |
2258 | ||
2259 | ASSERT(logical >= rbio->raid_map[0]); | |
2260 | ASSERT(logical + PAGE_SIZE <= rbio->raid_map[0] + | |
2261 | rbio->stripe_len * rbio->nr_data); | |
2262 | stripe_offset = (int)(logical - rbio->raid_map[0]); | |
2263 | index = stripe_offset >> PAGE_CACHE_SHIFT; | |
2264 | rbio->bio_pages[index] = page; | |
2265 | } | |
2266 | ||
2267 | /* | |
2268 | * We just scrub the parity that we have correct data on the same horizontal, | |
2269 | * so we needn't allocate all pages for all the stripes. | |
2270 | */ | |
2271 | static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio) | |
2272 | { | |
2273 | int i; | |
2274 | int bit; | |
2275 | int index; | |
2276 | struct page *page; | |
2277 | ||
2278 | for_each_set_bit(bit, rbio->dbitmap, rbio->stripe_npages) { | |
2c8cdd6e | 2279 | for (i = 0; i < rbio->real_stripes; i++) { |
5a6ac9ea MX |
2280 | index = i * rbio->stripe_npages + bit; |
2281 | if (rbio->stripe_pages[index]) | |
2282 | continue; | |
2283 | ||
2284 | page = alloc_page(GFP_NOFS | __GFP_HIGHMEM); | |
2285 | if (!page) | |
2286 | return -ENOMEM; | |
2287 | rbio->stripe_pages[index] = page; | |
2288 | ClearPageUptodate(page); | |
2289 | } | |
2290 | } | |
2291 | return 0; | |
2292 | } | |
2293 | ||
2294 | /* | |
2295 | * end io function used by finish_rmw. When we finally | |
2296 | * get here, we've written a full stripe | |
2297 | */ | |
2298 | static void raid_write_parity_end_io(struct bio *bio, int err) | |
2299 | { | |
2300 | struct btrfs_raid_bio *rbio = bio->bi_private; | |
2301 | ||
2302 | if (err) | |
2303 | fail_bio_stripe(rbio, bio); | |
2304 | ||
2305 | bio_put(bio); | |
2306 | ||
2307 | if (!atomic_dec_and_test(&rbio->stripes_pending)) | |
2308 | return; | |
2309 | ||
2310 | err = 0; | |
2311 | ||
2312 | if (atomic_read(&rbio->error)) | |
2313 | err = -EIO; | |
2314 | ||
2315 | rbio_orig_end_io(rbio, err, 0); | |
2316 | } | |
2317 | ||
2318 | static noinline void finish_parity_scrub(struct btrfs_raid_bio *rbio, | |
2319 | int need_check) | |
2320 | { | |
76035976 | 2321 | struct btrfs_bio *bbio = rbio->bbio; |
2c8cdd6e | 2322 | void *pointers[rbio->real_stripes]; |
76035976 | 2323 | DECLARE_BITMAP(pbitmap, rbio->stripe_npages); |
5a6ac9ea MX |
2324 | int nr_data = rbio->nr_data; |
2325 | int stripe; | |
2326 | int pagenr; | |
2327 | int p_stripe = -1; | |
2328 | int q_stripe = -1; | |
2329 | struct page *p_page = NULL; | |
2330 | struct page *q_page = NULL; | |
2331 | struct bio_list bio_list; | |
2332 | struct bio *bio; | |
76035976 | 2333 | int is_replace = 0; |
5a6ac9ea MX |
2334 | int ret; |
2335 | ||
2336 | bio_list_init(&bio_list); | |
2337 | ||
2c8cdd6e MX |
2338 | if (rbio->real_stripes - rbio->nr_data == 1) { |
2339 | p_stripe = rbio->real_stripes - 1; | |
2340 | } else if (rbio->real_stripes - rbio->nr_data == 2) { | |
2341 | p_stripe = rbio->real_stripes - 2; | |
2342 | q_stripe = rbio->real_stripes - 1; | |
5a6ac9ea MX |
2343 | } else { |
2344 | BUG(); | |
2345 | } | |
2346 | ||
76035976 MX |
2347 | if (bbio->num_tgtdevs && bbio->tgtdev_map[rbio->scrubp]) { |
2348 | is_replace = 1; | |
2349 | bitmap_copy(pbitmap, rbio->dbitmap, rbio->stripe_npages); | |
2350 | } | |
2351 | ||
5a6ac9ea MX |
2352 | /* |
2353 | * Because the higher layers(scrubber) are unlikely to | |
2354 | * use this area of the disk again soon, so don't cache | |
2355 | * it. | |
2356 | */ | |
2357 | clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags); | |
2358 | ||
2359 | if (!need_check) | |
2360 | goto writeback; | |
2361 | ||
2362 | p_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM); | |
2363 | if (!p_page) | |
2364 | goto cleanup; | |
2365 | SetPageUptodate(p_page); | |
2366 | ||
2367 | if (q_stripe != -1) { | |
2368 | q_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM); | |
2369 | if (!q_page) { | |
2370 | __free_page(p_page); | |
2371 | goto cleanup; | |
2372 | } | |
2373 | SetPageUptodate(q_page); | |
2374 | } | |
2375 | ||
2376 | atomic_set(&rbio->error, 0); | |
2377 | ||
2378 | for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) { | |
2379 | struct page *p; | |
2380 | void *parity; | |
2381 | /* first collect one page from each data stripe */ | |
2382 | for (stripe = 0; stripe < nr_data; stripe++) { | |
2383 | p = page_in_rbio(rbio, stripe, pagenr, 0); | |
2384 | pointers[stripe] = kmap(p); | |
2385 | } | |
2386 | ||
2387 | /* then add the parity stripe */ | |
2388 | pointers[stripe++] = kmap(p_page); | |
2389 | ||
2390 | if (q_stripe != -1) { | |
2391 | ||
2392 | /* | |
2393 | * raid6, add the qstripe and call the | |
2394 | * library function to fill in our p/q | |
2395 | */ | |
2396 | pointers[stripe++] = kmap(q_page); | |
2397 | ||
2c8cdd6e | 2398 | raid6_call.gen_syndrome(rbio->real_stripes, PAGE_SIZE, |
5a6ac9ea MX |
2399 | pointers); |
2400 | } else { | |
2401 | /* raid5 */ | |
2402 | memcpy(pointers[nr_data], pointers[0], PAGE_SIZE); | |
2403 | run_xor(pointers + 1, nr_data - 1, PAGE_CACHE_SIZE); | |
2404 | } | |
2405 | ||
2406 | /* Check scrubbing pairty and repair it */ | |
2407 | p = rbio_stripe_page(rbio, rbio->scrubp, pagenr); | |
2408 | parity = kmap(p); | |
2409 | if (memcmp(parity, pointers[rbio->scrubp], PAGE_CACHE_SIZE)) | |
2410 | memcpy(parity, pointers[rbio->scrubp], PAGE_CACHE_SIZE); | |
2411 | else | |
2412 | /* Parity is right, needn't writeback */ | |
2413 | bitmap_clear(rbio->dbitmap, pagenr, 1); | |
2414 | kunmap(p); | |
2415 | ||
2c8cdd6e | 2416 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) |
5a6ac9ea MX |
2417 | kunmap(page_in_rbio(rbio, stripe, pagenr, 0)); |
2418 | } | |
2419 | ||
2420 | __free_page(p_page); | |
2421 | if (q_page) | |
2422 | __free_page(q_page); | |
2423 | ||
2424 | writeback: | |
2425 | /* | |
2426 | * time to start writing. Make bios for everything from the | |
2427 | * higher layers (the bio_list in our rbio) and our p/q. Ignore | |
2428 | * everything else. | |
2429 | */ | |
2430 | for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) { | |
2431 | struct page *page; | |
2432 | ||
2433 | page = rbio_stripe_page(rbio, rbio->scrubp, pagenr); | |
2434 | ret = rbio_add_io_page(rbio, &bio_list, | |
2435 | page, rbio->scrubp, pagenr, rbio->stripe_len); | |
2436 | if (ret) | |
2437 | goto cleanup; | |
2438 | } | |
2439 | ||
76035976 MX |
2440 | if (!is_replace) |
2441 | goto submit_write; | |
2442 | ||
2443 | for_each_set_bit(pagenr, pbitmap, rbio->stripe_npages) { | |
2444 | struct page *page; | |
2445 | ||
2446 | page = rbio_stripe_page(rbio, rbio->scrubp, pagenr); | |
2447 | ret = rbio_add_io_page(rbio, &bio_list, page, | |
2448 | bbio->tgtdev_map[rbio->scrubp], | |
2449 | pagenr, rbio->stripe_len); | |
2450 | if (ret) | |
2451 | goto cleanup; | |
2452 | } | |
2453 | ||
2454 | submit_write: | |
5a6ac9ea MX |
2455 | nr_data = bio_list_size(&bio_list); |
2456 | if (!nr_data) { | |
2457 | /* Every parity is right */ | |
2458 | rbio_orig_end_io(rbio, 0, 0); | |
2459 | return; | |
2460 | } | |
2461 | ||
2462 | atomic_set(&rbio->stripes_pending, nr_data); | |
2463 | ||
2464 | while (1) { | |
2465 | bio = bio_list_pop(&bio_list); | |
2466 | if (!bio) | |
2467 | break; | |
2468 | ||
2469 | bio->bi_private = rbio; | |
2470 | bio->bi_end_io = raid_write_parity_end_io; | |
2471 | BUG_ON(!test_bit(BIO_UPTODATE, &bio->bi_flags)); | |
2472 | submit_bio(WRITE, bio); | |
2473 | } | |
2474 | return; | |
2475 | ||
2476 | cleanup: | |
2477 | rbio_orig_end_io(rbio, -EIO, 0); | |
2478 | } | |
2479 | ||
2480 | static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe) | |
2481 | { | |
2482 | if (stripe >= 0 && stripe < rbio->nr_data) | |
2483 | return 1; | |
2484 | return 0; | |
2485 | } | |
2486 | ||
2487 | /* | |
2488 | * While we're doing the parity check and repair, we could have errors | |
2489 | * in reading pages off the disk. This checks for errors and if we're | |
2490 | * not able to read the page it'll trigger parity reconstruction. The | |
2491 | * parity scrub will be finished after we've reconstructed the failed | |
2492 | * stripes | |
2493 | */ | |
2494 | static void validate_rbio_for_parity_scrub(struct btrfs_raid_bio *rbio) | |
2495 | { | |
2496 | if (atomic_read(&rbio->error) > rbio->bbio->max_errors) | |
2497 | goto cleanup; | |
2498 | ||
2499 | if (rbio->faila >= 0 || rbio->failb >= 0) { | |
2500 | int dfail = 0, failp = -1; | |
2501 | ||
2502 | if (is_data_stripe(rbio, rbio->faila)) | |
2503 | dfail++; | |
2504 | else if (is_parity_stripe(rbio->faila)) | |
2505 | failp = rbio->faila; | |
2506 | ||
2507 | if (is_data_stripe(rbio, rbio->failb)) | |
2508 | dfail++; | |
2509 | else if (is_parity_stripe(rbio->failb)) | |
2510 | failp = rbio->failb; | |
2511 | ||
2512 | /* | |
2513 | * Because we can not use a scrubbing parity to repair | |
2514 | * the data, so the capability of the repair is declined. | |
2515 | * (In the case of RAID5, we can not repair anything) | |
2516 | */ | |
2517 | if (dfail > rbio->bbio->max_errors - 1) | |
2518 | goto cleanup; | |
2519 | ||
2520 | /* | |
2521 | * If all data is good, only parity is correctly, just | |
2522 | * repair the parity. | |
2523 | */ | |
2524 | if (dfail == 0) { | |
2525 | finish_parity_scrub(rbio, 0); | |
2526 | return; | |
2527 | } | |
2528 | ||
2529 | /* | |
2530 | * Here means we got one corrupted data stripe and one | |
2531 | * corrupted parity on RAID6, if the corrupted parity | |
2532 | * is scrubbing parity, luckly, use the other one to repair | |
2533 | * the data, or we can not repair the data stripe. | |
2534 | */ | |
2535 | if (failp != rbio->scrubp) | |
2536 | goto cleanup; | |
2537 | ||
2538 | __raid_recover_end_io(rbio); | |
2539 | } else { | |
2540 | finish_parity_scrub(rbio, 1); | |
2541 | } | |
2542 | return; | |
2543 | ||
2544 | cleanup: | |
2545 | rbio_orig_end_io(rbio, -EIO, 0); | |
2546 | } | |
2547 | ||
2548 | /* | |
2549 | * end io for the read phase of the rmw cycle. All the bios here are physical | |
2550 | * stripe bios we've read from the disk so we can recalculate the parity of the | |
2551 | * stripe. | |
2552 | * | |
2553 | * This will usually kick off finish_rmw once all the bios are read in, but it | |
2554 | * may trigger parity reconstruction if we had any errors along the way | |
2555 | */ | |
2556 | static void raid56_parity_scrub_end_io(struct bio *bio, int err) | |
2557 | { | |
2558 | struct btrfs_raid_bio *rbio = bio->bi_private; | |
2559 | ||
2560 | if (err) | |
2561 | fail_bio_stripe(rbio, bio); | |
2562 | else | |
2563 | set_bio_pages_uptodate(bio); | |
2564 | ||
2565 | bio_put(bio); | |
2566 | ||
2567 | if (!atomic_dec_and_test(&rbio->stripes_pending)) | |
2568 | return; | |
2569 | ||
2570 | /* | |
2571 | * this will normally call finish_rmw to start our write | |
2572 | * but if there are any failed stripes we'll reconstruct | |
2573 | * from parity first | |
2574 | */ | |
2575 | validate_rbio_for_parity_scrub(rbio); | |
2576 | } | |
2577 | ||
2578 | static void raid56_parity_scrub_stripe(struct btrfs_raid_bio *rbio) | |
2579 | { | |
2580 | int bios_to_read = 0; | |
5a6ac9ea MX |
2581 | struct bio_list bio_list; |
2582 | int ret; | |
2583 | int pagenr; | |
2584 | int stripe; | |
2585 | struct bio *bio; | |
2586 | ||
2587 | ret = alloc_rbio_essential_pages(rbio); | |
2588 | if (ret) | |
2589 | goto cleanup; | |
2590 | ||
2591 | bio_list_init(&bio_list); | |
2592 | ||
2593 | atomic_set(&rbio->error, 0); | |
2594 | /* | |
2595 | * build a list of bios to read all the missing parts of this | |
2596 | * stripe | |
2597 | */ | |
2c8cdd6e | 2598 | for (stripe = 0; stripe < rbio->real_stripes; stripe++) { |
5a6ac9ea MX |
2599 | for_each_set_bit(pagenr, rbio->dbitmap, rbio->stripe_npages) { |
2600 | struct page *page; | |
2601 | /* | |
2602 | * we want to find all the pages missing from | |
2603 | * the rbio and read them from the disk. If | |
2604 | * page_in_rbio finds a page in the bio list | |
2605 | * we don't need to read it off the stripe. | |
2606 | */ | |
2607 | page = page_in_rbio(rbio, stripe, pagenr, 1); | |
2608 | if (page) | |
2609 | continue; | |
2610 | ||
2611 | page = rbio_stripe_page(rbio, stripe, pagenr); | |
2612 | /* | |
2613 | * the bio cache may have handed us an uptodate | |
2614 | * page. If so, be happy and use it | |
2615 | */ | |
2616 | if (PageUptodate(page)) | |
2617 | continue; | |
2618 | ||
2619 | ret = rbio_add_io_page(rbio, &bio_list, page, | |
2620 | stripe, pagenr, rbio->stripe_len); | |
2621 | if (ret) | |
2622 | goto cleanup; | |
2623 | } | |
2624 | } | |
2625 | ||
2626 | bios_to_read = bio_list_size(&bio_list); | |
2627 | if (!bios_to_read) { | |
2628 | /* | |
2629 | * this can happen if others have merged with | |
2630 | * us, it means there is nothing left to read. | |
2631 | * But if there are missing devices it may not be | |
2632 | * safe to do the full stripe write yet. | |
2633 | */ | |
2634 | goto finish; | |
2635 | } | |
2636 | ||
2637 | /* | |
2638 | * the bbio may be freed once we submit the last bio. Make sure | |
2639 | * not to touch it after that | |
2640 | */ | |
2641 | atomic_set(&rbio->stripes_pending, bios_to_read); | |
2642 | while (1) { | |
2643 | bio = bio_list_pop(&bio_list); | |
2644 | if (!bio) | |
2645 | break; | |
2646 | ||
2647 | bio->bi_private = rbio; | |
2648 | bio->bi_end_io = raid56_parity_scrub_end_io; | |
2649 | ||
2650 | btrfs_bio_wq_end_io(rbio->fs_info, bio, | |
2651 | BTRFS_WQ_ENDIO_RAID56); | |
2652 | ||
2653 | BUG_ON(!test_bit(BIO_UPTODATE, &bio->bi_flags)); | |
2654 | submit_bio(READ, bio); | |
2655 | } | |
2656 | /* the actual write will happen once the reads are done */ | |
2657 | return; | |
2658 | ||
2659 | cleanup: | |
2660 | rbio_orig_end_io(rbio, -EIO, 0); | |
2661 | return; | |
2662 | ||
2663 | finish: | |
2664 | validate_rbio_for_parity_scrub(rbio); | |
2665 | } | |
2666 | ||
2667 | static void scrub_parity_work(struct btrfs_work *work) | |
2668 | { | |
2669 | struct btrfs_raid_bio *rbio; | |
2670 | ||
2671 | rbio = container_of(work, struct btrfs_raid_bio, work); | |
2672 | raid56_parity_scrub_stripe(rbio); | |
2673 | } | |
2674 | ||
2675 | static void async_scrub_parity(struct btrfs_raid_bio *rbio) | |
2676 | { | |
2677 | btrfs_init_work(&rbio->work, btrfs_rmw_helper, | |
2678 | scrub_parity_work, NULL, NULL); | |
2679 | ||
2680 | btrfs_queue_work(rbio->fs_info->rmw_workers, | |
2681 | &rbio->work); | |
2682 | } | |
2683 | ||
2684 | void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio) | |
2685 | { | |
2686 | if (!lock_stripe_add(rbio)) | |
2687 | async_scrub_parity(rbio); | |
2688 | } |